xref: /freebsd/sys/vm/vm_page.c (revision da477bcdc0c335171bb0ed3813f570026de6df85)
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 struct vm_domain vm_dom[MAXMEMDOM];
117 
118 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
119 
120 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
121 
122 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
123 /* The following fields are protected by the domainset lock. */
124 domainset_t __exclusive_cache_line vm_min_domains;
125 domainset_t __exclusive_cache_line vm_severe_domains;
126 static int vm_min_waiters;
127 static int vm_severe_waiters;
128 static int vm_pageproc_waiters;
129 
130 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
131     "VM page statistics");
132 
133 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
134 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
135     CTLFLAG_RD, &pqstate_commit_retries,
136     "Number of failed per-page atomic queue state updates");
137 
138 static COUNTER_U64_DEFINE_EARLY(queue_ops);
139 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
140     CTLFLAG_RD, &queue_ops,
141     "Number of batched queue operations");
142 
143 static COUNTER_U64_DEFINE_EARLY(queue_nops);
144 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
145     CTLFLAG_RD, &queue_nops,
146     "Number of batched queue operations with no effects");
147 
148 /*
149  * bogus page -- for I/O to/from partially complete buffers,
150  * or for paging into sparsely invalid regions.
151  */
152 vm_page_t bogus_page;
153 
154 vm_page_t vm_page_array;
155 long vm_page_array_size;
156 long first_page;
157 
158 static TAILQ_HEAD(, vm_page) blacklist_head;
159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
161     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
162 
163 static uma_zone_t fakepg_zone;
164 
165 static void vm_page_alloc_check(vm_page_t m);
166 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
167     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
168 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
169 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
170 static bool vm_page_free_prep(vm_page_t m);
171 static void vm_page_free_toq(vm_page_t m);
172 static void vm_page_init(void *dummy);
173 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
174     vm_pindex_t pindex, vm_page_t mpred);
175 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
176     vm_page_t mpred);
177 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
178     const uint16_t nflag);
179 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
180     vm_page_t m_run, vm_paddr_t high);
181 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
182 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
183     int req);
184 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
185     int flags);
186 static void vm_page_zone_release(void *arg, void **store, int cnt);
187 
188 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
189 
190 static void
191 vm_page_init(void *dummy)
192 {
193 
194 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
195 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
196 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
197 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
198 }
199 
200 /*
201  * The cache page zone is initialized later since we need to be able to allocate
202  * pages before UMA is fully initialized.
203  */
204 static void
205 vm_page_init_cache_zones(void *dummy __unused)
206 {
207 	struct vm_domain *vmd;
208 	struct vm_pgcache *pgcache;
209 	int cache, domain, maxcache, pool;
210 
211 	maxcache = 0;
212 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
213 	maxcache *= mp_ncpus;
214 	for (domain = 0; domain < vm_ndomains; domain++) {
215 		vmd = VM_DOMAIN(domain);
216 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
217 			pgcache = &vmd->vmd_pgcache[pool];
218 			pgcache->domain = domain;
219 			pgcache->pool = pool;
220 			pgcache->zone = uma_zcache_create("vm pgcache",
221 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
222 			    vm_page_zone_import, vm_page_zone_release, pgcache,
223 			    UMA_ZONE_VM);
224 
225 			/*
226 			 * Limit each pool's zone to 0.1% of the pages in the
227 			 * domain.
228 			 */
229 			cache = maxcache != 0 ? maxcache :
230 			    vmd->vmd_page_count / 1000;
231 			uma_zone_set_maxcache(pgcache->zone, cache);
232 		}
233 	}
234 }
235 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
236 
237 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
238 #if PAGE_SIZE == 32768
239 #ifdef CTASSERT
240 CTASSERT(sizeof(u_long) >= 8);
241 #endif
242 #endif
243 
244 /*
245  *	vm_set_page_size:
246  *
247  *	Sets the page size, perhaps based upon the memory
248  *	size.  Must be called before any use of page-size
249  *	dependent functions.
250  */
251 void
252 vm_set_page_size(void)
253 {
254 	if (vm_cnt.v_page_size == 0)
255 		vm_cnt.v_page_size = PAGE_SIZE;
256 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
257 		panic("vm_set_page_size: page size not a power of two");
258 }
259 
260 /*
261  *	vm_page_blacklist_next:
262  *
263  *	Find the next entry in the provided string of blacklist
264  *	addresses.  Entries are separated by space, comma, or newline.
265  *	If an invalid integer is encountered then the rest of the
266  *	string is skipped.  Updates the list pointer to the next
267  *	character, or NULL if the string is exhausted or invalid.
268  */
269 static vm_paddr_t
270 vm_page_blacklist_next(char **list, char *end)
271 {
272 	vm_paddr_t bad;
273 	char *cp, *pos;
274 
275 	if (list == NULL || *list == NULL)
276 		return (0);
277 	if (**list =='\0') {
278 		*list = NULL;
279 		return (0);
280 	}
281 
282 	/*
283 	 * If there's no end pointer then the buffer is coming from
284 	 * the kenv and we know it's null-terminated.
285 	 */
286 	if (end == NULL)
287 		end = *list + strlen(*list);
288 
289 	/* Ensure that strtoq() won't walk off the end */
290 	if (*end != '\0') {
291 		if (*end == '\n' || *end == ' ' || *end  == ',')
292 			*end = '\0';
293 		else {
294 			printf("Blacklist not terminated, skipping\n");
295 			*list = NULL;
296 			return (0);
297 		}
298 	}
299 
300 	for (pos = *list; *pos != '\0'; pos = cp) {
301 		bad = strtoq(pos, &cp, 0);
302 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
303 			if (bad == 0) {
304 				if (++cp < end)
305 					continue;
306 				else
307 					break;
308 			}
309 		} else
310 			break;
311 		if (*cp == '\0' || ++cp >= end)
312 			*list = NULL;
313 		else
314 			*list = cp;
315 		return (trunc_page(bad));
316 	}
317 	printf("Garbage in RAM blacklist, skipping\n");
318 	*list = NULL;
319 	return (0);
320 }
321 
322 bool
323 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
324 {
325 	struct vm_domain *vmd;
326 	vm_page_t m;
327 	int ret;
328 
329 	m = vm_phys_paddr_to_vm_page(pa);
330 	if (m == NULL)
331 		return (true); /* page does not exist, no failure */
332 
333 	vmd = vm_pagequeue_domain(m);
334 	vm_domain_free_lock(vmd);
335 	ret = vm_phys_unfree_page(m);
336 	vm_domain_free_unlock(vmd);
337 	if (ret != 0) {
338 		vm_domain_freecnt_inc(vmd, -1);
339 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
340 		if (verbose)
341 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
342 	}
343 	return (ret);
344 }
345 
346 /*
347  *	vm_page_blacklist_check:
348  *
349  *	Iterate through the provided string of blacklist addresses, pulling
350  *	each entry out of the physical allocator free list and putting it
351  *	onto a list for reporting via the vm.page_blacklist sysctl.
352  */
353 static void
354 vm_page_blacklist_check(char *list, char *end)
355 {
356 	vm_paddr_t pa;
357 	char *next;
358 
359 	next = list;
360 	while (next != NULL) {
361 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
362 			continue;
363 		vm_page_blacklist_add(pa, bootverbose);
364 	}
365 }
366 
367 /*
368  *	vm_page_blacklist_load:
369  *
370  *	Search for a special module named "ram_blacklist".  It'll be a
371  *	plain text file provided by the user via the loader directive
372  *	of the same name.
373  */
374 static void
375 vm_page_blacklist_load(char **list, char **end)
376 {
377 	void *mod;
378 	u_char *ptr;
379 	u_int len;
380 
381 	mod = NULL;
382 	ptr = NULL;
383 
384 	mod = preload_search_by_type("ram_blacklist");
385 	if (mod != NULL) {
386 		ptr = preload_fetch_addr(mod);
387 		len = preload_fetch_size(mod);
388         }
389 	*list = ptr;
390 	if (ptr != NULL)
391 		*end = ptr + len;
392 	else
393 		*end = NULL;
394 	return;
395 }
396 
397 static int
398 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
399 {
400 	vm_page_t m;
401 	struct sbuf sbuf;
402 	int error, first;
403 
404 	first = 1;
405 	error = sysctl_wire_old_buffer(req, 0);
406 	if (error != 0)
407 		return (error);
408 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
409 	TAILQ_FOREACH(m, &blacklist_head, listq) {
410 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
411 		    (uintmax_t)m->phys_addr);
412 		first = 0;
413 	}
414 	error = sbuf_finish(&sbuf);
415 	sbuf_delete(&sbuf);
416 	return (error);
417 }
418 
419 /*
420  * Initialize a dummy page for use in scans of the specified paging queue.
421  * In principle, this function only needs to set the flag PG_MARKER.
422  * Nonetheless, it write busies the page as a safety precaution.
423  */
424 void
425 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
426 {
427 
428 	bzero(marker, sizeof(*marker));
429 	marker->flags = PG_MARKER;
430 	marker->a.flags = aflags;
431 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
432 	marker->a.queue = queue;
433 }
434 
435 static void
436 vm_page_domain_init(int domain)
437 {
438 	struct vm_domain *vmd;
439 	struct vm_pagequeue *pq;
440 	int i;
441 
442 	vmd = VM_DOMAIN(domain);
443 	bzero(vmd, sizeof(*vmd));
444 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
445 	    "vm inactive pagequeue";
446 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
447 	    "vm active pagequeue";
448 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
449 	    "vm laundry pagequeue";
450 	*__DECONST(const char **,
451 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
452 	    "vm unswappable pagequeue";
453 	vmd->vmd_domain = domain;
454 	vmd->vmd_page_count = 0;
455 	vmd->vmd_free_count = 0;
456 	vmd->vmd_segs = 0;
457 	vmd->vmd_oom = FALSE;
458 	for (i = 0; i < PQ_COUNT; i++) {
459 		pq = &vmd->vmd_pagequeues[i];
460 		TAILQ_INIT(&pq->pq_pl);
461 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
462 		    MTX_DEF | MTX_DUPOK);
463 		pq->pq_pdpages = 0;
464 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
465 	}
466 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
467 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
468 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
469 
470 	/*
471 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
472 	 * insertions.
473 	 */
474 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
475 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
476 	    &vmd->vmd_inacthead, plinks.q);
477 
478 	/*
479 	 * The clock pages are used to implement active queue scanning without
480 	 * requeues.  Scans start at clock[0], which is advanced after the scan
481 	 * ends.  When the two clock hands meet, they are reset and scanning
482 	 * resumes from the head of the queue.
483 	 */
484 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
485 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
486 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
487 	    &vmd->vmd_clock[0], plinks.q);
488 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
489 	    &vmd->vmd_clock[1], plinks.q);
490 }
491 
492 /*
493  * Initialize a physical page in preparation for adding it to the free
494  * lists.
495  */
496 static void
497 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
498 {
499 
500 	m->object = NULL;
501 	m->ref_count = 0;
502 	m->busy_lock = VPB_FREED;
503 	m->flags = m->a.flags = 0;
504 	m->phys_addr = pa;
505 	m->a.queue = PQ_NONE;
506 	m->psind = 0;
507 	m->segind = segind;
508 	m->order = VM_NFREEORDER;
509 	m->pool = VM_FREEPOOL_DEFAULT;
510 	m->valid = m->dirty = 0;
511 	pmap_page_init(m);
512 }
513 
514 #ifndef PMAP_HAS_PAGE_ARRAY
515 static vm_paddr_t
516 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
517 {
518 	vm_paddr_t new_end;
519 
520 	/*
521 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
522 	 * However, because this page is allocated from KVM, out-of-bounds
523 	 * accesses using the direct map will not be trapped.
524 	 */
525 	*vaddr += PAGE_SIZE;
526 
527 	/*
528 	 * Allocate physical memory for the page structures, and map it.
529 	 */
530 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
531 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
532 	    VM_PROT_READ | VM_PROT_WRITE);
533 	vm_page_array_size = page_range;
534 
535 	return (new_end);
536 }
537 #endif
538 
539 /*
540  *	vm_page_startup:
541  *
542  *	Initializes the resident memory module.  Allocates physical memory for
543  *	bootstrapping UMA and some data structures that are used to manage
544  *	physical pages.  Initializes these structures, and populates the free
545  *	page queues.
546  */
547 vm_offset_t
548 vm_page_startup(vm_offset_t vaddr)
549 {
550 	struct vm_phys_seg *seg;
551 	vm_page_t m;
552 	char *list, *listend;
553 	vm_paddr_t end, high_avail, low_avail, new_end, size;
554 	vm_paddr_t page_range __unused;
555 	vm_paddr_t last_pa, pa;
556 	u_long pagecount;
557 	int biggestone, i, segind;
558 #ifdef WITNESS
559 	vm_offset_t mapped;
560 	int witness_size;
561 #endif
562 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
563 	long ii;
564 #endif
565 
566 	vaddr = round_page(vaddr);
567 
568 	vm_phys_early_startup();
569 	biggestone = vm_phys_avail_largest();
570 	end = phys_avail[biggestone+1];
571 
572 	/*
573 	 * Initialize the page and queue locks.
574 	 */
575 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
576 	for (i = 0; i < PA_LOCK_COUNT; i++)
577 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
578 	for (i = 0; i < vm_ndomains; i++)
579 		vm_page_domain_init(i);
580 
581 	new_end = end;
582 #ifdef WITNESS
583 	witness_size = round_page(witness_startup_count());
584 	new_end -= witness_size;
585 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
586 	    VM_PROT_READ | VM_PROT_WRITE);
587 	bzero((void *)mapped, witness_size);
588 	witness_startup((void *)mapped);
589 #endif
590 
591 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
592     defined(__i386__) || defined(__mips__) || defined(__riscv) || \
593     defined(__powerpc64__)
594 	/*
595 	 * Allocate a bitmap to indicate that a random physical page
596 	 * needs to be included in a minidump.
597 	 *
598 	 * The amd64 port needs this to indicate which direct map pages
599 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
600 	 *
601 	 * However, i386 still needs this workspace internally within the
602 	 * minidump code.  In theory, they are not needed on i386, but are
603 	 * included should the sf_buf code decide to use them.
604 	 */
605 	last_pa = 0;
606 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
607 		if (dump_avail[i + 1] > last_pa)
608 			last_pa = dump_avail[i + 1];
609 	page_range = last_pa / PAGE_SIZE;
610 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
611 	new_end -= vm_page_dump_size;
612 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
613 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
614 	bzero((void *)vm_page_dump, vm_page_dump_size);
615 #else
616 	(void)last_pa;
617 #endif
618 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
619     defined(__riscv) || defined(__powerpc64__)
620 	/*
621 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
622 	 * in a crash dump.  When pmap_map() uses the direct map, they are
623 	 * not automatically included.
624 	 */
625 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
626 		dump_add_page(pa);
627 #endif
628 	phys_avail[biggestone + 1] = new_end;
629 #ifdef __amd64__
630 	/*
631 	 * Request that the physical pages underlying the message buffer be
632 	 * included in a crash dump.  Since the message buffer is accessed
633 	 * through the direct map, they are not automatically included.
634 	 */
635 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
636 	last_pa = pa + round_page(msgbufsize);
637 	while (pa < last_pa) {
638 		dump_add_page(pa);
639 		pa += PAGE_SIZE;
640 	}
641 #endif
642 	/*
643 	 * Compute the number of pages of memory that will be available for
644 	 * use, taking into account the overhead of a page structure per page.
645 	 * In other words, solve
646 	 *	"available physical memory" - round_page(page_range *
647 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
648 	 * for page_range.
649 	 */
650 	low_avail = phys_avail[0];
651 	high_avail = phys_avail[1];
652 	for (i = 0; i < vm_phys_nsegs; i++) {
653 		if (vm_phys_segs[i].start < low_avail)
654 			low_avail = vm_phys_segs[i].start;
655 		if (vm_phys_segs[i].end > high_avail)
656 			high_avail = vm_phys_segs[i].end;
657 	}
658 	/* Skip the first chunk.  It is already accounted for. */
659 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
660 		if (phys_avail[i] < low_avail)
661 			low_avail = phys_avail[i];
662 		if (phys_avail[i + 1] > high_avail)
663 			high_avail = phys_avail[i + 1];
664 	}
665 	first_page = low_avail / PAGE_SIZE;
666 #ifdef VM_PHYSSEG_SPARSE
667 	size = 0;
668 	for (i = 0; i < vm_phys_nsegs; i++)
669 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
670 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
671 		size += phys_avail[i + 1] - phys_avail[i];
672 #elif defined(VM_PHYSSEG_DENSE)
673 	size = high_avail - low_avail;
674 #else
675 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
676 #endif
677 
678 #ifdef PMAP_HAS_PAGE_ARRAY
679 	pmap_page_array_startup(size / PAGE_SIZE);
680 	biggestone = vm_phys_avail_largest();
681 	end = new_end = phys_avail[biggestone + 1];
682 #else
683 #ifdef VM_PHYSSEG_DENSE
684 	/*
685 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
686 	 * the overhead of a page structure per page only if vm_page_array is
687 	 * allocated from the last physical memory chunk.  Otherwise, we must
688 	 * allocate page structures representing the physical memory
689 	 * underlying vm_page_array, even though they will not be used.
690 	 */
691 	if (new_end != high_avail)
692 		page_range = size / PAGE_SIZE;
693 	else
694 #endif
695 	{
696 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
697 
698 		/*
699 		 * If the partial bytes remaining are large enough for
700 		 * a page (PAGE_SIZE) without a corresponding
701 		 * 'struct vm_page', then new_end will contain an
702 		 * extra page after subtracting the length of the VM
703 		 * page array.  Compensate by subtracting an extra
704 		 * page from new_end.
705 		 */
706 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
707 			if (new_end == high_avail)
708 				high_avail -= PAGE_SIZE;
709 			new_end -= PAGE_SIZE;
710 		}
711 	}
712 	end = new_end;
713 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
714 #endif
715 
716 #if VM_NRESERVLEVEL > 0
717 	/*
718 	 * Allocate physical memory for the reservation management system's
719 	 * data structures, and map it.
720 	 */
721 	new_end = vm_reserv_startup(&vaddr, new_end);
722 #endif
723 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
724     defined(__riscv) || defined(__powerpc64__)
725 	/*
726 	 * Include vm_page_array and vm_reserv_array in a crash dump.
727 	 */
728 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
729 		dump_add_page(pa);
730 #endif
731 	phys_avail[biggestone + 1] = new_end;
732 
733 	/*
734 	 * Add physical memory segments corresponding to the available
735 	 * physical pages.
736 	 */
737 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
738 		if (vm_phys_avail_size(i) != 0)
739 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
740 
741 	/*
742 	 * Initialize the physical memory allocator.
743 	 */
744 	vm_phys_init();
745 
746 	/*
747 	 * Initialize the page structures and add every available page to the
748 	 * physical memory allocator's free lists.
749 	 */
750 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
751 	for (ii = 0; ii < vm_page_array_size; ii++) {
752 		m = &vm_page_array[ii];
753 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
754 		m->flags = PG_FICTITIOUS;
755 	}
756 #endif
757 	vm_cnt.v_page_count = 0;
758 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
759 		seg = &vm_phys_segs[segind];
760 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
761 		    m++, pa += PAGE_SIZE)
762 			vm_page_init_page(m, pa, segind);
763 
764 		/*
765 		 * Add the segment to the free lists only if it is covered by
766 		 * one of the ranges in phys_avail.  Because we've added the
767 		 * ranges to the vm_phys_segs array, we can assume that each
768 		 * segment is either entirely contained in one of the ranges,
769 		 * or doesn't overlap any of them.
770 		 */
771 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
772 			struct vm_domain *vmd;
773 
774 			if (seg->start < phys_avail[i] ||
775 			    seg->end > phys_avail[i + 1])
776 				continue;
777 
778 			m = seg->first_page;
779 			pagecount = (u_long)atop(seg->end - seg->start);
780 
781 			vmd = VM_DOMAIN(seg->domain);
782 			vm_domain_free_lock(vmd);
783 			vm_phys_enqueue_contig(m, pagecount);
784 			vm_domain_free_unlock(vmd);
785 			vm_domain_freecnt_inc(vmd, pagecount);
786 			vm_cnt.v_page_count += (u_int)pagecount;
787 
788 			vmd = VM_DOMAIN(seg->domain);
789 			vmd->vmd_page_count += (u_int)pagecount;
790 			vmd->vmd_segs |= 1UL << m->segind;
791 			break;
792 		}
793 	}
794 
795 	/*
796 	 * Remove blacklisted pages from the physical memory allocator.
797 	 */
798 	TAILQ_INIT(&blacklist_head);
799 	vm_page_blacklist_load(&list, &listend);
800 	vm_page_blacklist_check(list, listend);
801 
802 	list = kern_getenv("vm.blacklist");
803 	vm_page_blacklist_check(list, NULL);
804 
805 	freeenv(list);
806 #if VM_NRESERVLEVEL > 0
807 	/*
808 	 * Initialize the reservation management system.
809 	 */
810 	vm_reserv_init();
811 #endif
812 
813 	return (vaddr);
814 }
815 
816 void
817 vm_page_reference(vm_page_t m)
818 {
819 
820 	vm_page_aflag_set(m, PGA_REFERENCED);
821 }
822 
823 /*
824  *	vm_page_trybusy
825  *
826  *	Helper routine for grab functions to trylock busy.
827  *
828  *	Returns true on success and false on failure.
829  */
830 static bool
831 vm_page_trybusy(vm_page_t m, int allocflags)
832 {
833 
834 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
835 		return (vm_page_trysbusy(m));
836 	else
837 		return (vm_page_tryxbusy(m));
838 }
839 
840 /*
841  *	vm_page_tryacquire
842  *
843  *	Helper routine for grab functions to trylock busy and wire.
844  *
845  *	Returns true on success and false on failure.
846  */
847 static inline bool
848 vm_page_tryacquire(vm_page_t m, int allocflags)
849 {
850 	bool locked;
851 
852 	locked = vm_page_trybusy(m, allocflags);
853 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
854 		vm_page_wire(m);
855 	return (locked);
856 }
857 
858 /*
859  *	vm_page_busy_acquire:
860  *
861  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
862  *	and drop the object lock if necessary.
863  */
864 bool
865 vm_page_busy_acquire(vm_page_t m, int allocflags)
866 {
867 	vm_object_t obj;
868 	bool locked;
869 
870 	/*
871 	 * The page-specific object must be cached because page
872 	 * identity can change during the sleep, causing the
873 	 * re-lock of a different object.
874 	 * It is assumed that a reference to the object is already
875 	 * held by the callers.
876 	 */
877 	obj = m->object;
878 	for (;;) {
879 		if (vm_page_tryacquire(m, allocflags))
880 			return (true);
881 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
882 			return (false);
883 		if (obj != NULL)
884 			locked = VM_OBJECT_WOWNED(obj);
885 		else
886 			locked = false;
887 		MPASS(locked || vm_page_wired(m));
888 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
889 		    locked) && locked)
890 			VM_OBJECT_WLOCK(obj);
891 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
892 			return (false);
893 		KASSERT(m->object == obj || m->object == NULL,
894 		    ("vm_page_busy_acquire: page %p does not belong to %p",
895 		    m, obj));
896 	}
897 }
898 
899 /*
900  *	vm_page_busy_downgrade:
901  *
902  *	Downgrade an exclusive busy page into a single shared busy page.
903  */
904 void
905 vm_page_busy_downgrade(vm_page_t m)
906 {
907 	u_int x;
908 
909 	vm_page_assert_xbusied(m);
910 
911 	x = vm_page_busy_fetch(m);
912 	for (;;) {
913 		if (atomic_fcmpset_rel_int(&m->busy_lock,
914 		    &x, VPB_SHARERS_WORD(1)))
915 			break;
916 	}
917 	if ((x & VPB_BIT_WAITERS) != 0)
918 		wakeup(m);
919 }
920 
921 /*
922  *
923  *	vm_page_busy_tryupgrade:
924  *
925  *	Attempt to upgrade a single shared busy into an exclusive busy.
926  */
927 int
928 vm_page_busy_tryupgrade(vm_page_t m)
929 {
930 	u_int ce, x;
931 
932 	vm_page_assert_sbusied(m);
933 
934 	x = vm_page_busy_fetch(m);
935 	ce = VPB_CURTHREAD_EXCLUSIVE;
936 	for (;;) {
937 		if (VPB_SHARERS(x) > 1)
938 			return (0);
939 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
940 		    ("vm_page_busy_tryupgrade: invalid lock state"));
941 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
942 		    ce | (x & VPB_BIT_WAITERS)))
943 			continue;
944 		return (1);
945 	}
946 }
947 
948 /*
949  *	vm_page_sbusied:
950  *
951  *	Return a positive value if the page is shared busied, 0 otherwise.
952  */
953 int
954 vm_page_sbusied(vm_page_t m)
955 {
956 	u_int x;
957 
958 	x = vm_page_busy_fetch(m);
959 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
960 }
961 
962 /*
963  *	vm_page_sunbusy:
964  *
965  *	Shared unbusy a page.
966  */
967 void
968 vm_page_sunbusy(vm_page_t m)
969 {
970 	u_int x;
971 
972 	vm_page_assert_sbusied(m);
973 
974 	x = vm_page_busy_fetch(m);
975 	for (;;) {
976 		KASSERT(x != VPB_FREED,
977 		    ("vm_page_sunbusy: Unlocking freed page."));
978 		if (VPB_SHARERS(x) > 1) {
979 			if (atomic_fcmpset_int(&m->busy_lock, &x,
980 			    x - VPB_ONE_SHARER))
981 				break;
982 			continue;
983 		}
984 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
985 		    ("vm_page_sunbusy: invalid lock state"));
986 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
987 			continue;
988 		if ((x & VPB_BIT_WAITERS) == 0)
989 			break;
990 		wakeup(m);
991 		break;
992 	}
993 }
994 
995 /*
996  *	vm_page_busy_sleep:
997  *
998  *	Sleep if the page is busy, using the page pointer as wchan.
999  *	This is used to implement the hard-path of busying mechanism.
1000  *
1001  *	If nonshared is true, sleep only if the page is xbusy.
1002  *
1003  *	The object lock must be held on entry and will be released on exit.
1004  */
1005 void
1006 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1007 {
1008 	vm_object_t obj;
1009 
1010 	obj = m->object;
1011 	VM_OBJECT_ASSERT_LOCKED(obj);
1012 	vm_page_lock_assert(m, MA_NOTOWNED);
1013 
1014 	if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1015 	    nonshared ? VM_ALLOC_SBUSY : 0 , true))
1016 		VM_OBJECT_DROP(obj);
1017 }
1018 
1019 /*
1020  *	vm_page_busy_sleep_unlocked:
1021  *
1022  *	Sleep if the page is busy, using the page pointer as wchan.
1023  *	This is used to implement the hard-path of busying mechanism.
1024  *
1025  *	If nonshared is true, sleep only if the page is xbusy.
1026  *
1027  *	The object lock must not be held on entry.  The operation will
1028  *	return if the page changes identity.
1029  */
1030 void
1031 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1032     const char *wmesg, bool nonshared)
1033 {
1034 
1035 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1036 	vm_page_lock_assert(m, MA_NOTOWNED);
1037 
1038 	_vm_page_busy_sleep(obj, m, pindex, wmesg,
1039 	    nonshared ? VM_ALLOC_SBUSY : 0, false);
1040 }
1041 
1042 /*
1043  *	_vm_page_busy_sleep:
1044  *
1045  *	Internal busy sleep function.  Verifies the page identity and
1046  *	lockstate against parameters.  Returns true if it sleeps and
1047  *	false otherwise.
1048  *
1049  *	If locked is true the lock will be dropped for any true returns
1050  *	and held for any false returns.
1051  */
1052 static bool
1053 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1054     const char *wmesg, int allocflags, bool locked)
1055 {
1056 	bool xsleep;
1057 	u_int x;
1058 
1059 	/*
1060 	 * If the object is busy we must wait for that to drain to zero
1061 	 * before trying the page again.
1062 	 */
1063 	if (obj != NULL && vm_object_busied(obj)) {
1064 		if (locked)
1065 			VM_OBJECT_DROP(obj);
1066 		vm_object_busy_wait(obj, wmesg);
1067 		return (true);
1068 	}
1069 
1070 	if (!vm_page_busied(m))
1071 		return (false);
1072 
1073 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1074 	sleepq_lock(m);
1075 	x = vm_page_busy_fetch(m);
1076 	do {
1077 		/*
1078 		 * If the page changes objects or becomes unlocked we can
1079 		 * simply return.
1080 		 */
1081 		if (x == VPB_UNBUSIED ||
1082 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1083 		    m->object != obj || m->pindex != pindex) {
1084 			sleepq_release(m);
1085 			return (false);
1086 		}
1087 		if ((x & VPB_BIT_WAITERS) != 0)
1088 			break;
1089 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1090 	if (locked)
1091 		VM_OBJECT_DROP(obj);
1092 	DROP_GIANT();
1093 	sleepq_add(m, NULL, wmesg, 0, 0);
1094 	sleepq_wait(m, PVM);
1095 	PICKUP_GIANT();
1096 	return (true);
1097 }
1098 
1099 /*
1100  *	vm_page_trysbusy:
1101  *
1102  *	Try to shared busy a page.
1103  *	If the operation succeeds 1 is returned otherwise 0.
1104  *	The operation never sleeps.
1105  */
1106 int
1107 vm_page_trysbusy(vm_page_t m)
1108 {
1109 	vm_object_t obj;
1110 	u_int x;
1111 
1112 	obj = m->object;
1113 	x = vm_page_busy_fetch(m);
1114 	for (;;) {
1115 		if ((x & VPB_BIT_SHARED) == 0)
1116 			return (0);
1117 		/*
1118 		 * Reduce the window for transient busies that will trigger
1119 		 * false negatives in vm_page_ps_test().
1120 		 */
1121 		if (obj != NULL && vm_object_busied(obj))
1122 			return (0);
1123 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1124 		    x + VPB_ONE_SHARER))
1125 			break;
1126 	}
1127 
1128 	/* Refetch the object now that we're guaranteed that it is stable. */
1129 	obj = m->object;
1130 	if (obj != NULL && vm_object_busied(obj)) {
1131 		vm_page_sunbusy(m);
1132 		return (0);
1133 	}
1134 	return (1);
1135 }
1136 
1137 /*
1138  *	vm_page_tryxbusy:
1139  *
1140  *	Try to exclusive busy a page.
1141  *	If the operation succeeds 1 is returned otherwise 0.
1142  *	The operation never sleeps.
1143  */
1144 int
1145 vm_page_tryxbusy(vm_page_t m)
1146 {
1147 	vm_object_t obj;
1148 
1149         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1150             VPB_CURTHREAD_EXCLUSIVE) == 0)
1151 		return (0);
1152 
1153 	obj = m->object;
1154 	if (obj != NULL && vm_object_busied(obj)) {
1155 		vm_page_xunbusy(m);
1156 		return (0);
1157 	}
1158 	return (1);
1159 }
1160 
1161 static void
1162 vm_page_xunbusy_hard_tail(vm_page_t m)
1163 {
1164 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1165 	/* Wake the waiter. */
1166 	wakeup(m);
1167 }
1168 
1169 /*
1170  *	vm_page_xunbusy_hard:
1171  *
1172  *	Called when unbusy has failed because there is a waiter.
1173  */
1174 void
1175 vm_page_xunbusy_hard(vm_page_t m)
1176 {
1177 	vm_page_assert_xbusied(m);
1178 	vm_page_xunbusy_hard_tail(m);
1179 }
1180 
1181 void
1182 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1183 {
1184 	vm_page_assert_xbusied_unchecked(m);
1185 	vm_page_xunbusy_hard_tail(m);
1186 }
1187 
1188 static void
1189 vm_page_busy_free(vm_page_t m)
1190 {
1191 	u_int x;
1192 
1193 	atomic_thread_fence_rel();
1194 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1195 	if ((x & VPB_BIT_WAITERS) != 0)
1196 		wakeup(m);
1197 }
1198 
1199 /*
1200  *	vm_page_unhold_pages:
1201  *
1202  *	Unhold each of the pages that is referenced by the given array.
1203  */
1204 void
1205 vm_page_unhold_pages(vm_page_t *ma, int count)
1206 {
1207 
1208 	for (; count != 0; count--) {
1209 		vm_page_unwire(*ma, PQ_ACTIVE);
1210 		ma++;
1211 	}
1212 }
1213 
1214 vm_page_t
1215 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1216 {
1217 	vm_page_t m;
1218 
1219 #ifdef VM_PHYSSEG_SPARSE
1220 	m = vm_phys_paddr_to_vm_page(pa);
1221 	if (m == NULL)
1222 		m = vm_phys_fictitious_to_vm_page(pa);
1223 	return (m);
1224 #elif defined(VM_PHYSSEG_DENSE)
1225 	long pi;
1226 
1227 	pi = atop(pa);
1228 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1229 		m = &vm_page_array[pi - first_page];
1230 		return (m);
1231 	}
1232 	return (vm_phys_fictitious_to_vm_page(pa));
1233 #else
1234 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1235 #endif
1236 }
1237 
1238 /*
1239  *	vm_page_getfake:
1240  *
1241  *	Create a fictitious page with the specified physical address and
1242  *	memory attribute.  The memory attribute is the only the machine-
1243  *	dependent aspect of a fictitious page that must be initialized.
1244  */
1245 vm_page_t
1246 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1247 {
1248 	vm_page_t m;
1249 
1250 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1251 	vm_page_initfake(m, paddr, memattr);
1252 	return (m);
1253 }
1254 
1255 void
1256 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1257 {
1258 
1259 	if ((m->flags & PG_FICTITIOUS) != 0) {
1260 		/*
1261 		 * The page's memattr might have changed since the
1262 		 * previous initialization.  Update the pmap to the
1263 		 * new memattr.
1264 		 */
1265 		goto memattr;
1266 	}
1267 	m->phys_addr = paddr;
1268 	m->a.queue = PQ_NONE;
1269 	/* Fictitious pages don't use "segind". */
1270 	m->flags = PG_FICTITIOUS;
1271 	/* Fictitious pages don't use "order" or "pool". */
1272 	m->oflags = VPO_UNMANAGED;
1273 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1274 	/* Fictitious pages are unevictable. */
1275 	m->ref_count = 1;
1276 	pmap_page_init(m);
1277 memattr:
1278 	pmap_page_set_memattr(m, memattr);
1279 }
1280 
1281 /*
1282  *	vm_page_putfake:
1283  *
1284  *	Release a fictitious page.
1285  */
1286 void
1287 vm_page_putfake(vm_page_t m)
1288 {
1289 
1290 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1291 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1292 	    ("vm_page_putfake: bad page %p", m));
1293 	vm_page_assert_xbusied(m);
1294 	vm_page_busy_free(m);
1295 	uma_zfree(fakepg_zone, m);
1296 }
1297 
1298 /*
1299  *	vm_page_updatefake:
1300  *
1301  *	Update the given fictitious page to the specified physical address and
1302  *	memory attribute.
1303  */
1304 void
1305 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1306 {
1307 
1308 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1309 	    ("vm_page_updatefake: bad page %p", m));
1310 	m->phys_addr = paddr;
1311 	pmap_page_set_memattr(m, memattr);
1312 }
1313 
1314 /*
1315  *	vm_page_free:
1316  *
1317  *	Free a page.
1318  */
1319 void
1320 vm_page_free(vm_page_t m)
1321 {
1322 
1323 	m->flags &= ~PG_ZERO;
1324 	vm_page_free_toq(m);
1325 }
1326 
1327 /*
1328  *	vm_page_free_zero:
1329  *
1330  *	Free a page to the zerod-pages queue
1331  */
1332 void
1333 vm_page_free_zero(vm_page_t m)
1334 {
1335 
1336 	m->flags |= PG_ZERO;
1337 	vm_page_free_toq(m);
1338 }
1339 
1340 /*
1341  * Unbusy and handle the page queueing for a page from a getpages request that
1342  * was optionally read ahead or behind.
1343  */
1344 void
1345 vm_page_readahead_finish(vm_page_t m)
1346 {
1347 
1348 	/* We shouldn't put invalid pages on queues. */
1349 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1350 
1351 	/*
1352 	 * Since the page is not the actually needed one, whether it should
1353 	 * be activated or deactivated is not obvious.  Empirical results
1354 	 * have shown that deactivating the page is usually the best choice,
1355 	 * unless the page is wanted by another thread.
1356 	 */
1357 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1358 		vm_page_activate(m);
1359 	else
1360 		vm_page_deactivate(m);
1361 	vm_page_xunbusy_unchecked(m);
1362 }
1363 
1364 /*
1365  * Destroy the identity of an invalid page and free it if possible.
1366  * This is intended to be used when reading a page from backing store fails.
1367  */
1368 void
1369 vm_page_free_invalid(vm_page_t m)
1370 {
1371 
1372 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1373 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1374 	KASSERT(m->object != NULL, ("page %p has no object", m));
1375 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1376 
1377 	/*
1378 	 * We may be attempting to free the page as part of the handling for an
1379 	 * I/O error, in which case the page was xbusied by a different thread.
1380 	 */
1381 	vm_page_xbusy_claim(m);
1382 
1383 	/*
1384 	 * If someone has wired this page while the object lock
1385 	 * was not held, then the thread that unwires is responsible
1386 	 * for freeing the page.  Otherwise just free the page now.
1387 	 * The wire count of this unmapped page cannot change while
1388 	 * we have the page xbusy and the page's object wlocked.
1389 	 */
1390 	if (vm_page_remove(m))
1391 		vm_page_free(m);
1392 }
1393 
1394 /*
1395  *	vm_page_sleep_if_busy:
1396  *
1397  *	Sleep and release the object lock if the page is busied.
1398  *	Returns TRUE if the thread slept.
1399  *
1400  *	The given page must be unlocked and object containing it must
1401  *	be locked.
1402  */
1403 int
1404 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1405 {
1406 	vm_object_t obj;
1407 
1408 	vm_page_lock_assert(m, MA_NOTOWNED);
1409 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1410 
1411 	/*
1412 	 * The page-specific object must be cached because page
1413 	 * identity can change during the sleep, causing the
1414 	 * re-lock of a different object.
1415 	 * It is assumed that a reference to the object is already
1416 	 * held by the callers.
1417 	 */
1418 	obj = m->object;
1419 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1420 		VM_OBJECT_WLOCK(obj);
1421 		return (TRUE);
1422 	}
1423 	return (FALSE);
1424 }
1425 
1426 /*
1427  *	vm_page_sleep_if_xbusy:
1428  *
1429  *	Sleep and release the object lock if the page is xbusied.
1430  *	Returns TRUE if the thread slept.
1431  *
1432  *	The given page must be unlocked and object containing it must
1433  *	be locked.
1434  */
1435 int
1436 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1437 {
1438 	vm_object_t obj;
1439 
1440 	vm_page_lock_assert(m, MA_NOTOWNED);
1441 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1442 
1443 	/*
1444 	 * The page-specific object must be cached because page
1445 	 * identity can change during the sleep, causing the
1446 	 * re-lock of a different object.
1447 	 * It is assumed that a reference to the object is already
1448 	 * held by the callers.
1449 	 */
1450 	obj = m->object;
1451 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1452 	    true)) {
1453 		VM_OBJECT_WLOCK(obj);
1454 		return (TRUE);
1455 	}
1456 	return (FALSE);
1457 }
1458 
1459 /*
1460  *	vm_page_dirty_KBI:		[ internal use only ]
1461  *
1462  *	Set all bits in the page's dirty field.
1463  *
1464  *	The object containing the specified page must be locked if the
1465  *	call is made from the machine-independent layer.
1466  *
1467  *	See vm_page_clear_dirty_mask().
1468  *
1469  *	This function should only be called by vm_page_dirty().
1470  */
1471 void
1472 vm_page_dirty_KBI(vm_page_t m)
1473 {
1474 
1475 	/* Refer to this operation by its public name. */
1476 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1477 	m->dirty = VM_PAGE_BITS_ALL;
1478 }
1479 
1480 /*
1481  *	vm_page_insert:		[ internal use only ]
1482  *
1483  *	Inserts the given mem entry into the object and object list.
1484  *
1485  *	The object must be locked.
1486  */
1487 int
1488 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1489 {
1490 	vm_page_t mpred;
1491 
1492 	VM_OBJECT_ASSERT_WLOCKED(object);
1493 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1494 	return (vm_page_insert_after(m, object, pindex, mpred));
1495 }
1496 
1497 /*
1498  *	vm_page_insert_after:
1499  *
1500  *	Inserts the page "m" into the specified object at offset "pindex".
1501  *
1502  *	The page "mpred" must immediately precede the offset "pindex" within
1503  *	the specified object.
1504  *
1505  *	The object must be locked.
1506  */
1507 static int
1508 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1509     vm_page_t mpred)
1510 {
1511 	vm_page_t msucc;
1512 
1513 	VM_OBJECT_ASSERT_WLOCKED(object);
1514 	KASSERT(m->object == NULL,
1515 	    ("vm_page_insert_after: page already inserted"));
1516 	if (mpred != NULL) {
1517 		KASSERT(mpred->object == object,
1518 		    ("vm_page_insert_after: object doesn't contain mpred"));
1519 		KASSERT(mpred->pindex < pindex,
1520 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1521 		msucc = TAILQ_NEXT(mpred, listq);
1522 	} else
1523 		msucc = TAILQ_FIRST(&object->memq);
1524 	if (msucc != NULL)
1525 		KASSERT(msucc->pindex > pindex,
1526 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1527 
1528 	/*
1529 	 * Record the object/offset pair in this page.
1530 	 */
1531 	m->object = object;
1532 	m->pindex = pindex;
1533 	m->ref_count |= VPRC_OBJREF;
1534 
1535 	/*
1536 	 * Now link into the object's ordered list of backed pages.
1537 	 */
1538 	if (vm_radix_insert(&object->rtree, m)) {
1539 		m->object = NULL;
1540 		m->pindex = 0;
1541 		m->ref_count &= ~VPRC_OBJREF;
1542 		return (1);
1543 	}
1544 	vm_page_insert_radixdone(m, object, mpred);
1545 	return (0);
1546 }
1547 
1548 /*
1549  *	vm_page_insert_radixdone:
1550  *
1551  *	Complete page "m" insertion into the specified object after the
1552  *	radix trie hooking.
1553  *
1554  *	The page "mpred" must precede the offset "m->pindex" within the
1555  *	specified object.
1556  *
1557  *	The object must be locked.
1558  */
1559 static void
1560 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1561 {
1562 
1563 	VM_OBJECT_ASSERT_WLOCKED(object);
1564 	KASSERT(object != NULL && m->object == object,
1565 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1566 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1567 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1568 	if (mpred != NULL) {
1569 		KASSERT(mpred->object == object,
1570 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1571 		KASSERT(mpred->pindex < m->pindex,
1572 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1573 	}
1574 
1575 	if (mpred != NULL)
1576 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1577 	else
1578 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1579 
1580 	/*
1581 	 * Show that the object has one more resident page.
1582 	 */
1583 	object->resident_page_count++;
1584 
1585 	/*
1586 	 * Hold the vnode until the last page is released.
1587 	 */
1588 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1589 		vhold(object->handle);
1590 
1591 	/*
1592 	 * Since we are inserting a new and possibly dirty page,
1593 	 * update the object's generation count.
1594 	 */
1595 	if (pmap_page_is_write_mapped(m))
1596 		vm_object_set_writeable_dirty(object);
1597 }
1598 
1599 /*
1600  * Do the work to remove a page from its object.  The caller is responsible for
1601  * updating the page's fields to reflect this removal.
1602  */
1603 static void
1604 vm_page_object_remove(vm_page_t m)
1605 {
1606 	vm_object_t object;
1607 	vm_page_t mrem;
1608 
1609 	vm_page_assert_xbusied(m);
1610 	object = m->object;
1611 	VM_OBJECT_ASSERT_WLOCKED(object);
1612 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1613 	    ("page %p is missing its object ref", m));
1614 
1615 	/* Deferred free of swap space. */
1616 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1617 		vm_pager_page_unswapped(m);
1618 
1619 	m->object = NULL;
1620 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1621 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1622 
1623 	/*
1624 	 * Now remove from the object's list of backed pages.
1625 	 */
1626 	TAILQ_REMOVE(&object->memq, m, listq);
1627 
1628 	/*
1629 	 * And show that the object has one fewer resident page.
1630 	 */
1631 	object->resident_page_count--;
1632 
1633 	/*
1634 	 * The vnode may now be recycled.
1635 	 */
1636 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1637 		vdrop(object->handle);
1638 }
1639 
1640 /*
1641  *	vm_page_remove:
1642  *
1643  *	Removes the specified page from its containing object, but does not
1644  *	invalidate any backing storage.  Returns true if the object's reference
1645  *	was the last reference to the page, and false otherwise.
1646  *
1647  *	The object must be locked and the page must be exclusively busied.
1648  *	The exclusive busy will be released on return.  If this is not the
1649  *	final ref and the caller does not hold a wire reference it may not
1650  *	continue to access the page.
1651  */
1652 bool
1653 vm_page_remove(vm_page_t m)
1654 {
1655 	bool dropped;
1656 
1657 	dropped = vm_page_remove_xbusy(m);
1658 	vm_page_xunbusy(m);
1659 
1660 	return (dropped);
1661 }
1662 
1663 /*
1664  *	vm_page_remove_xbusy
1665  *
1666  *	Removes the page but leaves the xbusy held.  Returns true if this
1667  *	removed the final ref and false otherwise.
1668  */
1669 bool
1670 vm_page_remove_xbusy(vm_page_t m)
1671 {
1672 
1673 	vm_page_object_remove(m);
1674 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1675 }
1676 
1677 /*
1678  *	vm_page_lookup:
1679  *
1680  *	Returns the page associated with the object/offset
1681  *	pair specified; if none is found, NULL is returned.
1682  *
1683  *	The object must be locked.
1684  */
1685 vm_page_t
1686 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1687 {
1688 
1689 	VM_OBJECT_ASSERT_LOCKED(object);
1690 	return (vm_radix_lookup(&object->rtree, pindex));
1691 }
1692 
1693 /*
1694  *	vm_page_relookup:
1695  *
1696  *	Returns a page that must already have been busied by
1697  *	the caller.  Used for bogus page replacement.
1698  */
1699 vm_page_t
1700 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1701 {
1702 	vm_page_t m;
1703 
1704 	m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1705 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1706 	    m->object == object && m->pindex == pindex,
1707 	    ("vm_page_relookup: Invalid page %p", m));
1708 	return (m);
1709 }
1710 
1711 /*
1712  * This should only be used by lockless functions for releasing transient
1713  * incorrect acquires.  The page may have been freed after we acquired a
1714  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1715  * further to do.
1716  */
1717 static void
1718 vm_page_busy_release(vm_page_t m)
1719 {
1720 	u_int x;
1721 
1722 	x = vm_page_busy_fetch(m);
1723 	for (;;) {
1724 		if (x == VPB_FREED)
1725 			break;
1726 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1727 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1728 			    x - VPB_ONE_SHARER))
1729 				break;
1730 			continue;
1731 		}
1732 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1733 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1734 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1735 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1736 			continue;
1737 		if ((x & VPB_BIT_WAITERS) != 0)
1738 			wakeup(m);
1739 		break;
1740 	}
1741 }
1742 
1743 /*
1744  *	vm_page_find_least:
1745  *
1746  *	Returns the page associated with the object with least pindex
1747  *	greater than or equal to the parameter pindex, or NULL.
1748  *
1749  *	The object must be locked.
1750  */
1751 vm_page_t
1752 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1753 {
1754 	vm_page_t m;
1755 
1756 	VM_OBJECT_ASSERT_LOCKED(object);
1757 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1758 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1759 	return (m);
1760 }
1761 
1762 /*
1763  * Returns the given page's successor (by pindex) within the object if it is
1764  * resident; if none is found, NULL is returned.
1765  *
1766  * The object must be locked.
1767  */
1768 vm_page_t
1769 vm_page_next(vm_page_t m)
1770 {
1771 	vm_page_t next;
1772 
1773 	VM_OBJECT_ASSERT_LOCKED(m->object);
1774 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1775 		MPASS(next->object == m->object);
1776 		if (next->pindex != m->pindex + 1)
1777 			next = NULL;
1778 	}
1779 	return (next);
1780 }
1781 
1782 /*
1783  * Returns the given page's predecessor (by pindex) within the object if it is
1784  * resident; if none is found, NULL is returned.
1785  *
1786  * The object must be locked.
1787  */
1788 vm_page_t
1789 vm_page_prev(vm_page_t m)
1790 {
1791 	vm_page_t prev;
1792 
1793 	VM_OBJECT_ASSERT_LOCKED(m->object);
1794 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1795 		MPASS(prev->object == m->object);
1796 		if (prev->pindex != m->pindex - 1)
1797 			prev = NULL;
1798 	}
1799 	return (prev);
1800 }
1801 
1802 /*
1803  * Uses the page mnew as a replacement for an existing page at index
1804  * pindex which must be already present in the object.
1805  *
1806  * Both pages must be exclusively busied on enter.  The old page is
1807  * unbusied on exit.
1808  *
1809  * A return value of true means mold is now free.  If this is not the
1810  * final ref and the caller does not hold a wire reference it may not
1811  * continue to access the page.
1812  */
1813 static bool
1814 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1815     vm_page_t mold)
1816 {
1817 	vm_page_t mret;
1818 	bool dropped;
1819 
1820 	VM_OBJECT_ASSERT_WLOCKED(object);
1821 	vm_page_assert_xbusied(mold);
1822 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1823 	    ("vm_page_replace: page %p already in object", mnew));
1824 
1825 	/*
1826 	 * This function mostly follows vm_page_insert() and
1827 	 * vm_page_remove() without the radix, object count and vnode
1828 	 * dance.  Double check such functions for more comments.
1829 	 */
1830 
1831 	mnew->object = object;
1832 	mnew->pindex = pindex;
1833 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1834 	mret = vm_radix_replace(&object->rtree, mnew);
1835 	KASSERT(mret == mold,
1836 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1837 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1838 	    (mnew->oflags & VPO_UNMANAGED),
1839 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1840 
1841 	/* Keep the resident page list in sorted order. */
1842 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1843 	TAILQ_REMOVE(&object->memq, mold, listq);
1844 	mold->object = NULL;
1845 
1846 	/*
1847 	 * The object's resident_page_count does not change because we have
1848 	 * swapped one page for another, but the generation count should
1849 	 * change if the page is dirty.
1850 	 */
1851 	if (pmap_page_is_write_mapped(mnew))
1852 		vm_object_set_writeable_dirty(object);
1853 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1854 	vm_page_xunbusy(mold);
1855 
1856 	return (dropped);
1857 }
1858 
1859 void
1860 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1861     vm_page_t mold)
1862 {
1863 
1864 	vm_page_assert_xbusied(mnew);
1865 
1866 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1867 		vm_page_free(mold);
1868 }
1869 
1870 /*
1871  *	vm_page_rename:
1872  *
1873  *	Move the given memory entry from its
1874  *	current object to the specified target object/offset.
1875  *
1876  *	Note: swap associated with the page must be invalidated by the move.  We
1877  *	      have to do this for several reasons:  (1) we aren't freeing the
1878  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1879  *	      moving the page from object A to B, and will then later move
1880  *	      the backing store from A to B and we can't have a conflict.
1881  *
1882  *	Note: we *always* dirty the page.  It is necessary both for the
1883  *	      fact that we moved it, and because we may be invalidating
1884  *	      swap.
1885  *
1886  *	The objects must be locked.
1887  */
1888 int
1889 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1890 {
1891 	vm_page_t mpred;
1892 	vm_pindex_t opidx;
1893 
1894 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1895 
1896 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1897 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1898 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1899 	    ("vm_page_rename: pindex already renamed"));
1900 
1901 	/*
1902 	 * Create a custom version of vm_page_insert() which does not depend
1903 	 * by m_prev and can cheat on the implementation aspects of the
1904 	 * function.
1905 	 */
1906 	opidx = m->pindex;
1907 	m->pindex = new_pindex;
1908 	if (vm_radix_insert(&new_object->rtree, m)) {
1909 		m->pindex = opidx;
1910 		return (1);
1911 	}
1912 
1913 	/*
1914 	 * The operation cannot fail anymore.  The removal must happen before
1915 	 * the listq iterator is tainted.
1916 	 */
1917 	m->pindex = opidx;
1918 	vm_page_object_remove(m);
1919 
1920 	/* Return back to the new pindex to complete vm_page_insert(). */
1921 	m->pindex = new_pindex;
1922 	m->object = new_object;
1923 
1924 	vm_page_insert_radixdone(m, new_object, mpred);
1925 	vm_page_dirty(m);
1926 	return (0);
1927 }
1928 
1929 /*
1930  *	vm_page_alloc:
1931  *
1932  *	Allocate and return a page that is associated with the specified
1933  *	object and offset pair.  By default, this page is exclusive busied.
1934  *
1935  *	The caller must always specify an allocation class.
1936  *
1937  *	allocation classes:
1938  *	VM_ALLOC_NORMAL		normal process request
1939  *	VM_ALLOC_SYSTEM		system *really* needs a page
1940  *	VM_ALLOC_INTERRUPT	interrupt time request
1941  *
1942  *	optional allocation flags:
1943  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1944  *				intends to allocate
1945  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1946  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1947  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1948  *				should not be exclusive busy
1949  *	VM_ALLOC_SBUSY		shared busy the allocated page
1950  *	VM_ALLOC_WIRED		wire the allocated page
1951  *	VM_ALLOC_ZERO		prefer a zeroed page
1952  */
1953 vm_page_t
1954 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1955 {
1956 
1957 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1958 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1959 }
1960 
1961 vm_page_t
1962 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1963     int req)
1964 {
1965 
1966 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1967 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1968 	    NULL));
1969 }
1970 
1971 /*
1972  * Allocate a page in the specified object with the given page index.  To
1973  * optimize insertion of the page into the object, the caller must also specifiy
1974  * the resident page in the object with largest index smaller than the given
1975  * page index, or NULL if no such page exists.
1976  */
1977 vm_page_t
1978 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1979     int req, vm_page_t mpred)
1980 {
1981 	struct vm_domainset_iter di;
1982 	vm_page_t m;
1983 	int domain;
1984 
1985 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1986 	do {
1987 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1988 		    mpred);
1989 		if (m != NULL)
1990 			break;
1991 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1992 
1993 	return (m);
1994 }
1995 
1996 /*
1997  * Returns true if the number of free pages exceeds the minimum
1998  * for the request class and false otherwise.
1999  */
2000 static int
2001 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2002 {
2003 	u_int limit, old, new;
2004 
2005 	if (req_class == VM_ALLOC_INTERRUPT)
2006 		limit = 0;
2007 	else if (req_class == VM_ALLOC_SYSTEM)
2008 		limit = vmd->vmd_interrupt_free_min;
2009 	else
2010 		limit = vmd->vmd_free_reserved;
2011 
2012 	/*
2013 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2014 	 */
2015 	limit += npages;
2016 	old = vmd->vmd_free_count;
2017 	do {
2018 		if (old < limit)
2019 			return (0);
2020 		new = old - npages;
2021 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2022 
2023 	/* Wake the page daemon if we've crossed the threshold. */
2024 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2025 		pagedaemon_wakeup(vmd->vmd_domain);
2026 
2027 	/* Only update bitsets on transitions. */
2028 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2029 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2030 		vm_domain_set(vmd);
2031 
2032 	return (1);
2033 }
2034 
2035 int
2036 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2037 {
2038 	int req_class;
2039 
2040 	/*
2041 	 * The page daemon is allowed to dig deeper into the free page list.
2042 	 */
2043 	req_class = req & VM_ALLOC_CLASS_MASK;
2044 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2045 		req_class = VM_ALLOC_SYSTEM;
2046 	return (_vm_domain_allocate(vmd, req_class, npages));
2047 }
2048 
2049 vm_page_t
2050 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2051     int req, vm_page_t mpred)
2052 {
2053 	struct vm_domain *vmd;
2054 	vm_page_t m;
2055 	int flags, pool;
2056 
2057 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2058 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2059 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2060 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2061 	    ("inconsistent object(%p)/req(%x)", object, req));
2062 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2063 	    ("Can't sleep and retry object insertion."));
2064 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2065 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2066 	    (uintmax_t)pindex));
2067 	if (object != NULL)
2068 		VM_OBJECT_ASSERT_WLOCKED(object);
2069 
2070 	flags = 0;
2071 	m = NULL;
2072 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2073 again:
2074 #if VM_NRESERVLEVEL > 0
2075 	/*
2076 	 * Can we allocate the page from a reservation?
2077 	 */
2078 	if (vm_object_reserv(object) &&
2079 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2080 	    NULL) {
2081 		goto found;
2082 	}
2083 #endif
2084 	vmd = VM_DOMAIN(domain);
2085 	if (vmd->vmd_pgcache[pool].zone != NULL) {
2086 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM);
2087 		if (m != NULL) {
2088 			flags |= PG_PCPU_CACHE;
2089 			goto found;
2090 		}
2091 	}
2092 	if (vm_domain_allocate(vmd, req, 1)) {
2093 		/*
2094 		 * If not, allocate it from the free page queues.
2095 		 */
2096 		vm_domain_free_lock(vmd);
2097 		m = vm_phys_alloc_pages(domain, pool, 0);
2098 		vm_domain_free_unlock(vmd);
2099 		if (m == NULL) {
2100 			vm_domain_freecnt_inc(vmd, 1);
2101 #if VM_NRESERVLEVEL > 0
2102 			if (vm_reserv_reclaim_inactive(domain))
2103 				goto again;
2104 #endif
2105 		}
2106 	}
2107 	if (m == NULL) {
2108 		/*
2109 		 * Not allocatable, give up.
2110 		 */
2111 		if (vm_domain_alloc_fail(vmd, object, req))
2112 			goto again;
2113 		return (NULL);
2114 	}
2115 
2116 	/*
2117 	 * At this point we had better have found a good page.
2118 	 */
2119 found:
2120 	vm_page_dequeue(m);
2121 	vm_page_alloc_check(m);
2122 
2123 	/*
2124 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2125 	 */
2126 	if ((req & VM_ALLOC_ZERO) != 0)
2127 		flags |= (m->flags & PG_ZERO);
2128 	if ((req & VM_ALLOC_NODUMP) != 0)
2129 		flags |= PG_NODUMP;
2130 	m->flags = flags;
2131 	m->a.flags = 0;
2132 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2133 	    VPO_UNMANAGED : 0;
2134 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2135 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2136 	else if ((req & VM_ALLOC_SBUSY) != 0)
2137 		m->busy_lock = VPB_SHARERS_WORD(1);
2138 	else
2139 		m->busy_lock = VPB_UNBUSIED;
2140 	if (req & VM_ALLOC_WIRED) {
2141 		vm_wire_add(1);
2142 		m->ref_count = 1;
2143 	}
2144 	m->a.act_count = 0;
2145 
2146 	if (object != NULL) {
2147 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2148 			if (req & VM_ALLOC_WIRED) {
2149 				vm_wire_sub(1);
2150 				m->ref_count = 0;
2151 			}
2152 			KASSERT(m->object == NULL, ("page %p has object", m));
2153 			m->oflags = VPO_UNMANAGED;
2154 			m->busy_lock = VPB_UNBUSIED;
2155 			/* Don't change PG_ZERO. */
2156 			vm_page_free_toq(m);
2157 			if (req & VM_ALLOC_WAITFAIL) {
2158 				VM_OBJECT_WUNLOCK(object);
2159 				vm_radix_wait();
2160 				VM_OBJECT_WLOCK(object);
2161 			}
2162 			return (NULL);
2163 		}
2164 
2165 		/* Ignore device objects; the pager sets "memattr" for them. */
2166 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2167 		    (object->flags & OBJ_FICTITIOUS) == 0)
2168 			pmap_page_set_memattr(m, object->memattr);
2169 	} else
2170 		m->pindex = pindex;
2171 
2172 	return (m);
2173 }
2174 
2175 /*
2176  *	vm_page_alloc_contig:
2177  *
2178  *	Allocate a contiguous set of physical pages of the given size "npages"
2179  *	from the free lists.  All of the physical pages must be at or above
2180  *	the given physical address "low" and below the given physical address
2181  *	"high".  The given value "alignment" determines the alignment of the
2182  *	first physical page in the set.  If the given value "boundary" is
2183  *	non-zero, then the set of physical pages cannot cross any physical
2184  *	address boundary that is a multiple of that value.  Both "alignment"
2185  *	and "boundary" must be a power of two.
2186  *
2187  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2188  *	then the memory attribute setting for the physical pages is configured
2189  *	to the object's memory attribute setting.  Otherwise, the memory
2190  *	attribute setting for the physical pages is configured to "memattr",
2191  *	overriding the object's memory attribute setting.  However, if the
2192  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2193  *	memory attribute setting for the physical pages cannot be configured
2194  *	to VM_MEMATTR_DEFAULT.
2195  *
2196  *	The specified object may not contain fictitious pages.
2197  *
2198  *	The caller must always specify an allocation class.
2199  *
2200  *	allocation classes:
2201  *	VM_ALLOC_NORMAL		normal process request
2202  *	VM_ALLOC_SYSTEM		system *really* needs a page
2203  *	VM_ALLOC_INTERRUPT	interrupt time request
2204  *
2205  *	optional allocation flags:
2206  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2207  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2208  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2209  *				should not be exclusive busy
2210  *	VM_ALLOC_SBUSY		shared busy the allocated page
2211  *	VM_ALLOC_WIRED		wire the allocated page
2212  *	VM_ALLOC_ZERO		prefer a zeroed page
2213  */
2214 vm_page_t
2215 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2216     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2217     vm_paddr_t boundary, vm_memattr_t memattr)
2218 {
2219 	struct vm_domainset_iter di;
2220 	vm_page_t m;
2221 	int domain;
2222 
2223 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2224 	do {
2225 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2226 		    npages, low, high, alignment, boundary, memattr);
2227 		if (m != NULL)
2228 			break;
2229 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2230 
2231 	return (m);
2232 }
2233 
2234 vm_page_t
2235 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2236     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2237     vm_paddr_t boundary, vm_memattr_t memattr)
2238 {
2239 	struct vm_domain *vmd;
2240 	vm_page_t m, m_ret, mpred;
2241 	u_int busy_lock, flags, oflags;
2242 
2243 	mpred = NULL;	/* XXX: pacify gcc */
2244 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2245 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2246 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2247 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2248 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2249 	    req));
2250 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2251 	    ("Can't sleep and retry object insertion."));
2252 	if (object != NULL) {
2253 		VM_OBJECT_ASSERT_WLOCKED(object);
2254 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2255 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2256 		    object));
2257 	}
2258 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2259 
2260 	if (object != NULL) {
2261 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2262 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2263 		    ("vm_page_alloc_contig: pindex already allocated"));
2264 	}
2265 
2266 	/*
2267 	 * Can we allocate the pages without the number of free pages falling
2268 	 * below the lower bound for the allocation class?
2269 	 */
2270 	m_ret = NULL;
2271 again:
2272 #if VM_NRESERVLEVEL > 0
2273 	/*
2274 	 * Can we allocate the pages from a reservation?
2275 	 */
2276 	if (vm_object_reserv(object) &&
2277 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2278 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2279 		goto found;
2280 	}
2281 #endif
2282 	vmd = VM_DOMAIN(domain);
2283 	if (vm_domain_allocate(vmd, req, npages)) {
2284 		/*
2285 		 * allocate them from the free page queues.
2286 		 */
2287 		vm_domain_free_lock(vmd);
2288 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2289 		    alignment, boundary);
2290 		vm_domain_free_unlock(vmd);
2291 		if (m_ret == NULL) {
2292 			vm_domain_freecnt_inc(vmd, npages);
2293 #if VM_NRESERVLEVEL > 0
2294 			if (vm_reserv_reclaim_contig(domain, npages, low,
2295 			    high, alignment, boundary))
2296 				goto again;
2297 #endif
2298 		}
2299 	}
2300 	if (m_ret == NULL) {
2301 		if (vm_domain_alloc_fail(vmd, object, req))
2302 			goto again;
2303 		return (NULL);
2304 	}
2305 #if VM_NRESERVLEVEL > 0
2306 found:
2307 #endif
2308 	for (m = m_ret; m < &m_ret[npages]; m++) {
2309 		vm_page_dequeue(m);
2310 		vm_page_alloc_check(m);
2311 	}
2312 
2313 	/*
2314 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2315 	 */
2316 	flags = 0;
2317 	if ((req & VM_ALLOC_ZERO) != 0)
2318 		flags = PG_ZERO;
2319 	if ((req & VM_ALLOC_NODUMP) != 0)
2320 		flags |= PG_NODUMP;
2321 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2322 	    VPO_UNMANAGED : 0;
2323 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2324 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2325 	else if ((req & VM_ALLOC_SBUSY) != 0)
2326 		busy_lock = VPB_SHARERS_WORD(1);
2327 	else
2328 		busy_lock = VPB_UNBUSIED;
2329 	if ((req & VM_ALLOC_WIRED) != 0)
2330 		vm_wire_add(npages);
2331 	if (object != NULL) {
2332 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2333 		    memattr == VM_MEMATTR_DEFAULT)
2334 			memattr = object->memattr;
2335 	}
2336 	for (m = m_ret; m < &m_ret[npages]; m++) {
2337 		m->a.flags = 0;
2338 		m->flags = (m->flags | PG_NODUMP) & flags;
2339 		m->busy_lock = busy_lock;
2340 		if ((req & VM_ALLOC_WIRED) != 0)
2341 			m->ref_count = 1;
2342 		m->a.act_count = 0;
2343 		m->oflags = oflags;
2344 		if (object != NULL) {
2345 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2346 				if ((req & VM_ALLOC_WIRED) != 0)
2347 					vm_wire_sub(npages);
2348 				KASSERT(m->object == NULL,
2349 				    ("page %p has object", m));
2350 				mpred = m;
2351 				for (m = m_ret; m < &m_ret[npages]; m++) {
2352 					if (m <= mpred &&
2353 					    (req & VM_ALLOC_WIRED) != 0)
2354 						m->ref_count = 0;
2355 					m->oflags = VPO_UNMANAGED;
2356 					m->busy_lock = VPB_UNBUSIED;
2357 					/* Don't change PG_ZERO. */
2358 					vm_page_free_toq(m);
2359 				}
2360 				if (req & VM_ALLOC_WAITFAIL) {
2361 					VM_OBJECT_WUNLOCK(object);
2362 					vm_radix_wait();
2363 					VM_OBJECT_WLOCK(object);
2364 				}
2365 				return (NULL);
2366 			}
2367 			mpred = m;
2368 		} else
2369 			m->pindex = pindex;
2370 		if (memattr != VM_MEMATTR_DEFAULT)
2371 			pmap_page_set_memattr(m, memattr);
2372 		pindex++;
2373 	}
2374 	return (m_ret);
2375 }
2376 
2377 /*
2378  * Check a page that has been freshly dequeued from a freelist.
2379  */
2380 static void
2381 vm_page_alloc_check(vm_page_t m)
2382 {
2383 
2384 	KASSERT(m->object == NULL, ("page %p has object", m));
2385 	KASSERT(m->a.queue == PQ_NONE &&
2386 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2387 	    ("page %p has unexpected queue %d, flags %#x",
2388 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2389 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2390 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2391 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2392 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2393 	    ("page %p has unexpected memattr %d",
2394 	    m, pmap_page_get_memattr(m)));
2395 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2396 }
2397 
2398 /*
2399  * 	vm_page_alloc_freelist:
2400  *
2401  *	Allocate a physical page from the specified free page list.
2402  *
2403  *	The caller must always specify an allocation class.
2404  *
2405  *	allocation classes:
2406  *	VM_ALLOC_NORMAL		normal process request
2407  *	VM_ALLOC_SYSTEM		system *really* needs a page
2408  *	VM_ALLOC_INTERRUPT	interrupt time request
2409  *
2410  *	optional allocation flags:
2411  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2412  *				intends to allocate
2413  *	VM_ALLOC_WIRED		wire the allocated page
2414  *	VM_ALLOC_ZERO		prefer a zeroed page
2415  */
2416 vm_page_t
2417 vm_page_alloc_freelist(int freelist, int req)
2418 {
2419 	struct vm_domainset_iter di;
2420 	vm_page_t m;
2421 	int domain;
2422 
2423 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2424 	do {
2425 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2426 		if (m != NULL)
2427 			break;
2428 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2429 
2430 	return (m);
2431 }
2432 
2433 vm_page_t
2434 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2435 {
2436 	struct vm_domain *vmd;
2437 	vm_page_t m;
2438 	u_int flags;
2439 
2440 	m = NULL;
2441 	vmd = VM_DOMAIN(domain);
2442 again:
2443 	if (vm_domain_allocate(vmd, req, 1)) {
2444 		vm_domain_free_lock(vmd);
2445 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2446 		    VM_FREEPOOL_DIRECT, 0);
2447 		vm_domain_free_unlock(vmd);
2448 		if (m == NULL)
2449 			vm_domain_freecnt_inc(vmd, 1);
2450 	}
2451 	if (m == NULL) {
2452 		if (vm_domain_alloc_fail(vmd, NULL, req))
2453 			goto again;
2454 		return (NULL);
2455 	}
2456 	vm_page_dequeue(m);
2457 	vm_page_alloc_check(m);
2458 
2459 	/*
2460 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2461 	 */
2462 	m->a.flags = 0;
2463 	flags = 0;
2464 	if ((req & VM_ALLOC_ZERO) != 0)
2465 		flags = PG_ZERO;
2466 	m->flags &= flags;
2467 	if ((req & VM_ALLOC_WIRED) != 0) {
2468 		vm_wire_add(1);
2469 		m->ref_count = 1;
2470 	}
2471 	/* Unmanaged pages don't use "act_count". */
2472 	m->oflags = VPO_UNMANAGED;
2473 	return (m);
2474 }
2475 
2476 static int
2477 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2478 {
2479 	struct vm_domain *vmd;
2480 	struct vm_pgcache *pgcache;
2481 	int i;
2482 
2483 	pgcache = arg;
2484 	vmd = VM_DOMAIN(pgcache->domain);
2485 
2486 	/*
2487 	 * The page daemon should avoid creating extra memory pressure since its
2488 	 * main purpose is to replenish the store of free pages.
2489 	 */
2490 	if (vmd->vmd_severeset || curproc == pageproc ||
2491 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2492 		return (0);
2493 	domain = vmd->vmd_domain;
2494 	vm_domain_free_lock(vmd);
2495 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2496 	    (vm_page_t *)store);
2497 	vm_domain_free_unlock(vmd);
2498 	if (cnt != i)
2499 		vm_domain_freecnt_inc(vmd, cnt - i);
2500 
2501 	return (i);
2502 }
2503 
2504 static void
2505 vm_page_zone_release(void *arg, void **store, int cnt)
2506 {
2507 	struct vm_domain *vmd;
2508 	struct vm_pgcache *pgcache;
2509 	vm_page_t m;
2510 	int i;
2511 
2512 	pgcache = arg;
2513 	vmd = VM_DOMAIN(pgcache->domain);
2514 	vm_domain_free_lock(vmd);
2515 	for (i = 0; i < cnt; i++) {
2516 		m = (vm_page_t)store[i];
2517 		vm_phys_free_pages(m, 0);
2518 	}
2519 	vm_domain_free_unlock(vmd);
2520 	vm_domain_freecnt_inc(vmd, cnt);
2521 }
2522 
2523 #define	VPSC_ANY	0	/* No restrictions. */
2524 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2525 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2526 
2527 /*
2528  *	vm_page_scan_contig:
2529  *
2530  *	Scan vm_page_array[] between the specified entries "m_start" and
2531  *	"m_end" for a run of contiguous physical pages that satisfy the
2532  *	specified conditions, and return the lowest page in the run.  The
2533  *	specified "alignment" determines the alignment of the lowest physical
2534  *	page in the run.  If the specified "boundary" is non-zero, then the
2535  *	run of physical pages cannot span a physical address that is a
2536  *	multiple of "boundary".
2537  *
2538  *	"m_end" is never dereferenced, so it need not point to a vm_page
2539  *	structure within vm_page_array[].
2540  *
2541  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2542  *	span a hole (or discontiguity) in the physical address space.  Both
2543  *	"alignment" and "boundary" must be a power of two.
2544  */
2545 vm_page_t
2546 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2547     u_long alignment, vm_paddr_t boundary, int options)
2548 {
2549 	vm_object_t object;
2550 	vm_paddr_t pa;
2551 	vm_page_t m, m_run;
2552 #if VM_NRESERVLEVEL > 0
2553 	int level;
2554 #endif
2555 	int m_inc, order, run_ext, run_len;
2556 
2557 	KASSERT(npages > 0, ("npages is 0"));
2558 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2559 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2560 	m_run = NULL;
2561 	run_len = 0;
2562 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2563 		KASSERT((m->flags & PG_MARKER) == 0,
2564 		    ("page %p is PG_MARKER", m));
2565 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2566 		    ("fictitious page %p has invalid ref count", m));
2567 
2568 		/*
2569 		 * If the current page would be the start of a run, check its
2570 		 * physical address against the end, alignment, and boundary
2571 		 * conditions.  If it doesn't satisfy these conditions, either
2572 		 * terminate the scan or advance to the next page that
2573 		 * satisfies the failed condition.
2574 		 */
2575 		if (run_len == 0) {
2576 			KASSERT(m_run == NULL, ("m_run != NULL"));
2577 			if (m + npages > m_end)
2578 				break;
2579 			pa = VM_PAGE_TO_PHYS(m);
2580 			if ((pa & (alignment - 1)) != 0) {
2581 				m_inc = atop(roundup2(pa, alignment) - pa);
2582 				continue;
2583 			}
2584 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2585 			    boundary) != 0) {
2586 				m_inc = atop(roundup2(pa, boundary) - pa);
2587 				continue;
2588 			}
2589 		} else
2590 			KASSERT(m_run != NULL, ("m_run == NULL"));
2591 
2592 retry:
2593 		m_inc = 1;
2594 		if (vm_page_wired(m))
2595 			run_ext = 0;
2596 #if VM_NRESERVLEVEL > 0
2597 		else if ((level = vm_reserv_level(m)) >= 0 &&
2598 		    (options & VPSC_NORESERV) != 0) {
2599 			run_ext = 0;
2600 			/* Advance to the end of the reservation. */
2601 			pa = VM_PAGE_TO_PHYS(m);
2602 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2603 			    pa);
2604 		}
2605 #endif
2606 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2607 			/*
2608 			 * The page is considered eligible for relocation if
2609 			 * and only if it could be laundered or reclaimed by
2610 			 * the page daemon.
2611 			 */
2612 			VM_OBJECT_RLOCK(object);
2613 			if (object != m->object) {
2614 				VM_OBJECT_RUNLOCK(object);
2615 				goto retry;
2616 			}
2617 			/* Don't care: PG_NODUMP, PG_ZERO. */
2618 			if (object->type != OBJT_DEFAULT &&
2619 			    object->type != OBJT_SWAP &&
2620 			    object->type != OBJT_VNODE) {
2621 				run_ext = 0;
2622 #if VM_NRESERVLEVEL > 0
2623 			} else if ((options & VPSC_NOSUPER) != 0 &&
2624 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2625 				run_ext = 0;
2626 				/* Advance to the end of the superpage. */
2627 				pa = VM_PAGE_TO_PHYS(m);
2628 				m_inc = atop(roundup2(pa + 1,
2629 				    vm_reserv_size(level)) - pa);
2630 #endif
2631 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2632 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2633 				/*
2634 				 * The page is allocated but eligible for
2635 				 * relocation.  Extend the current run by one
2636 				 * page.
2637 				 */
2638 				KASSERT(pmap_page_get_memattr(m) ==
2639 				    VM_MEMATTR_DEFAULT,
2640 				    ("page %p has an unexpected memattr", m));
2641 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2642 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2643 				    ("page %p has unexpected oflags", m));
2644 				/* Don't care: PGA_NOSYNC. */
2645 				run_ext = 1;
2646 			} else
2647 				run_ext = 0;
2648 			VM_OBJECT_RUNLOCK(object);
2649 #if VM_NRESERVLEVEL > 0
2650 		} else if (level >= 0) {
2651 			/*
2652 			 * The page is reserved but not yet allocated.  In
2653 			 * other words, it is still free.  Extend the current
2654 			 * run by one page.
2655 			 */
2656 			run_ext = 1;
2657 #endif
2658 		} else if ((order = m->order) < VM_NFREEORDER) {
2659 			/*
2660 			 * The page is enqueued in the physical memory
2661 			 * allocator's free page queues.  Moreover, it is the
2662 			 * first page in a power-of-two-sized run of
2663 			 * contiguous free pages.  Add these pages to the end
2664 			 * of the current run, and jump ahead.
2665 			 */
2666 			run_ext = 1 << order;
2667 			m_inc = 1 << order;
2668 		} else {
2669 			/*
2670 			 * Skip the page for one of the following reasons: (1)
2671 			 * It is enqueued in the physical memory allocator's
2672 			 * free page queues.  However, it is not the first
2673 			 * page in a run of contiguous free pages.  (This case
2674 			 * rarely occurs because the scan is performed in
2675 			 * ascending order.) (2) It is not reserved, and it is
2676 			 * transitioning from free to allocated.  (Conversely,
2677 			 * the transition from allocated to free for managed
2678 			 * pages is blocked by the page busy lock.) (3) It is
2679 			 * allocated but not contained by an object and not
2680 			 * wired, e.g., allocated by Xen's balloon driver.
2681 			 */
2682 			run_ext = 0;
2683 		}
2684 
2685 		/*
2686 		 * Extend or reset the current run of pages.
2687 		 */
2688 		if (run_ext > 0) {
2689 			if (run_len == 0)
2690 				m_run = m;
2691 			run_len += run_ext;
2692 		} else {
2693 			if (run_len > 0) {
2694 				m_run = NULL;
2695 				run_len = 0;
2696 			}
2697 		}
2698 	}
2699 	if (run_len >= npages)
2700 		return (m_run);
2701 	return (NULL);
2702 }
2703 
2704 /*
2705  *	vm_page_reclaim_run:
2706  *
2707  *	Try to relocate each of the allocated virtual pages within the
2708  *	specified run of physical pages to a new physical address.  Free the
2709  *	physical pages underlying the relocated virtual pages.  A virtual page
2710  *	is relocatable if and only if it could be laundered or reclaimed by
2711  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2712  *	physical address above "high".
2713  *
2714  *	Returns 0 if every physical page within the run was already free or
2715  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2716  *	value indicating why the last attempt to relocate a virtual page was
2717  *	unsuccessful.
2718  *
2719  *	"req_class" must be an allocation class.
2720  */
2721 static int
2722 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2723     vm_paddr_t high)
2724 {
2725 	struct vm_domain *vmd;
2726 	struct spglist free;
2727 	vm_object_t object;
2728 	vm_paddr_t pa;
2729 	vm_page_t m, m_end, m_new;
2730 	int error, order, req;
2731 
2732 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2733 	    ("req_class is not an allocation class"));
2734 	SLIST_INIT(&free);
2735 	error = 0;
2736 	m = m_run;
2737 	m_end = m_run + npages;
2738 	for (; error == 0 && m < m_end; m++) {
2739 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2740 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2741 
2742 		/*
2743 		 * Racily check for wirings.  Races are handled once the object
2744 		 * lock is held and the page is unmapped.
2745 		 */
2746 		if (vm_page_wired(m))
2747 			error = EBUSY;
2748 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2749 			/*
2750 			 * The page is relocated if and only if it could be
2751 			 * laundered or reclaimed by the page daemon.
2752 			 */
2753 			VM_OBJECT_WLOCK(object);
2754 			/* Don't care: PG_NODUMP, PG_ZERO. */
2755 			if (m->object != object ||
2756 			    (object->type != OBJT_DEFAULT &&
2757 			    object->type != OBJT_SWAP &&
2758 			    object->type != OBJT_VNODE))
2759 				error = EINVAL;
2760 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2761 				error = EINVAL;
2762 			else if (vm_page_queue(m) != PQ_NONE &&
2763 			    vm_page_tryxbusy(m) != 0) {
2764 				if (vm_page_wired(m)) {
2765 					vm_page_xunbusy(m);
2766 					error = EBUSY;
2767 					goto unlock;
2768 				}
2769 				KASSERT(pmap_page_get_memattr(m) ==
2770 				    VM_MEMATTR_DEFAULT,
2771 				    ("page %p has an unexpected memattr", m));
2772 				KASSERT(m->oflags == 0,
2773 				    ("page %p has unexpected oflags", m));
2774 				/* Don't care: PGA_NOSYNC. */
2775 				if (!vm_page_none_valid(m)) {
2776 					/*
2777 					 * First, try to allocate a new page
2778 					 * that is above "high".  Failing
2779 					 * that, try to allocate a new page
2780 					 * that is below "m_run".  Allocate
2781 					 * the new page between the end of
2782 					 * "m_run" and "high" only as a last
2783 					 * resort.
2784 					 */
2785 					req = req_class | VM_ALLOC_NOOBJ;
2786 					if ((m->flags & PG_NODUMP) != 0)
2787 						req |= VM_ALLOC_NODUMP;
2788 					if (trunc_page(high) !=
2789 					    ~(vm_paddr_t)PAGE_MASK) {
2790 						m_new = vm_page_alloc_contig(
2791 						    NULL, 0, req, 1,
2792 						    round_page(high),
2793 						    ~(vm_paddr_t)0,
2794 						    PAGE_SIZE, 0,
2795 						    VM_MEMATTR_DEFAULT);
2796 					} else
2797 						m_new = NULL;
2798 					if (m_new == NULL) {
2799 						pa = VM_PAGE_TO_PHYS(m_run);
2800 						m_new = vm_page_alloc_contig(
2801 						    NULL, 0, req, 1,
2802 						    0, pa - 1, PAGE_SIZE, 0,
2803 						    VM_MEMATTR_DEFAULT);
2804 					}
2805 					if (m_new == NULL) {
2806 						pa += ptoa(npages);
2807 						m_new = vm_page_alloc_contig(
2808 						    NULL, 0, req, 1,
2809 						    pa, high, PAGE_SIZE, 0,
2810 						    VM_MEMATTR_DEFAULT);
2811 					}
2812 					if (m_new == NULL) {
2813 						vm_page_xunbusy(m);
2814 						error = ENOMEM;
2815 						goto unlock;
2816 					}
2817 
2818 					/*
2819 					 * Unmap the page and check for new
2820 					 * wirings that may have been acquired
2821 					 * through a pmap lookup.
2822 					 */
2823 					if (object->ref_count != 0 &&
2824 					    !vm_page_try_remove_all(m)) {
2825 						vm_page_xunbusy(m);
2826 						vm_page_free(m_new);
2827 						error = EBUSY;
2828 						goto unlock;
2829 					}
2830 
2831 					/*
2832 					 * Replace "m" with the new page.  For
2833 					 * vm_page_replace(), "m" must be busy
2834 					 * and dequeued.  Finally, change "m"
2835 					 * as if vm_page_free() was called.
2836 					 */
2837 					m_new->a.flags = m->a.flags &
2838 					    ~PGA_QUEUE_STATE_MASK;
2839 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2840 					    ("page %p is managed", m_new));
2841 					m_new->oflags = 0;
2842 					pmap_copy_page(m, m_new);
2843 					m_new->valid = m->valid;
2844 					m_new->dirty = m->dirty;
2845 					m->flags &= ~PG_ZERO;
2846 					vm_page_dequeue(m);
2847 					if (vm_page_replace_hold(m_new, object,
2848 					    m->pindex, m) &&
2849 					    vm_page_free_prep(m))
2850 						SLIST_INSERT_HEAD(&free, m,
2851 						    plinks.s.ss);
2852 
2853 					/*
2854 					 * The new page must be deactivated
2855 					 * before the object is unlocked.
2856 					 */
2857 					vm_page_deactivate(m_new);
2858 				} else {
2859 					m->flags &= ~PG_ZERO;
2860 					vm_page_dequeue(m);
2861 					if (vm_page_free_prep(m))
2862 						SLIST_INSERT_HEAD(&free, m,
2863 						    plinks.s.ss);
2864 					KASSERT(m->dirty == 0,
2865 					    ("page %p is dirty", m));
2866 				}
2867 			} else
2868 				error = EBUSY;
2869 unlock:
2870 			VM_OBJECT_WUNLOCK(object);
2871 		} else {
2872 			MPASS(vm_phys_domain(m) == domain);
2873 			vmd = VM_DOMAIN(domain);
2874 			vm_domain_free_lock(vmd);
2875 			order = m->order;
2876 			if (order < VM_NFREEORDER) {
2877 				/*
2878 				 * The page is enqueued in the physical memory
2879 				 * allocator's free page queues.  Moreover, it
2880 				 * is the first page in a power-of-two-sized
2881 				 * run of contiguous free pages.  Jump ahead
2882 				 * to the last page within that run, and
2883 				 * continue from there.
2884 				 */
2885 				m += (1 << order) - 1;
2886 			}
2887 #if VM_NRESERVLEVEL > 0
2888 			else if (vm_reserv_is_page_free(m))
2889 				order = 0;
2890 #endif
2891 			vm_domain_free_unlock(vmd);
2892 			if (order == VM_NFREEORDER)
2893 				error = EINVAL;
2894 		}
2895 	}
2896 	if ((m = SLIST_FIRST(&free)) != NULL) {
2897 		int cnt;
2898 
2899 		vmd = VM_DOMAIN(domain);
2900 		cnt = 0;
2901 		vm_domain_free_lock(vmd);
2902 		do {
2903 			MPASS(vm_phys_domain(m) == domain);
2904 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2905 			vm_phys_free_pages(m, 0);
2906 			cnt++;
2907 		} while ((m = SLIST_FIRST(&free)) != NULL);
2908 		vm_domain_free_unlock(vmd);
2909 		vm_domain_freecnt_inc(vmd, cnt);
2910 	}
2911 	return (error);
2912 }
2913 
2914 #define	NRUNS	16
2915 
2916 CTASSERT(powerof2(NRUNS));
2917 
2918 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2919 
2920 #define	MIN_RECLAIM	8
2921 
2922 /*
2923  *	vm_page_reclaim_contig:
2924  *
2925  *	Reclaim allocated, contiguous physical memory satisfying the specified
2926  *	conditions by relocating the virtual pages using that physical memory.
2927  *	Returns true if reclamation is successful and false otherwise.  Since
2928  *	relocation requires the allocation of physical pages, reclamation may
2929  *	fail due to a shortage of free pages.  When reclamation fails, callers
2930  *	are expected to perform vm_wait() before retrying a failed allocation
2931  *	operation, e.g., vm_page_alloc_contig().
2932  *
2933  *	The caller must always specify an allocation class through "req".
2934  *
2935  *	allocation classes:
2936  *	VM_ALLOC_NORMAL		normal process request
2937  *	VM_ALLOC_SYSTEM		system *really* needs a page
2938  *	VM_ALLOC_INTERRUPT	interrupt time request
2939  *
2940  *	The optional allocation flags are ignored.
2941  *
2942  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2943  *	must be a power of two.
2944  */
2945 bool
2946 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2947     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2948 {
2949 	struct vm_domain *vmd;
2950 	vm_paddr_t curr_low;
2951 	vm_page_t m_run, m_runs[NRUNS];
2952 	u_long count, reclaimed;
2953 	int error, i, options, req_class;
2954 
2955 	KASSERT(npages > 0, ("npages is 0"));
2956 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2957 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2958 	req_class = req & VM_ALLOC_CLASS_MASK;
2959 
2960 	/*
2961 	 * The page daemon is allowed to dig deeper into the free page list.
2962 	 */
2963 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2964 		req_class = VM_ALLOC_SYSTEM;
2965 
2966 	/*
2967 	 * Return if the number of free pages cannot satisfy the requested
2968 	 * allocation.
2969 	 */
2970 	vmd = VM_DOMAIN(domain);
2971 	count = vmd->vmd_free_count;
2972 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2973 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2974 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2975 		return (false);
2976 
2977 	/*
2978 	 * Scan up to three times, relaxing the restrictions ("options") on
2979 	 * the reclamation of reservations and superpages each time.
2980 	 */
2981 	for (options = VPSC_NORESERV;;) {
2982 		/*
2983 		 * Find the highest runs that satisfy the given constraints
2984 		 * and restrictions, and record them in "m_runs".
2985 		 */
2986 		curr_low = low;
2987 		count = 0;
2988 		for (;;) {
2989 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2990 			    high, alignment, boundary, options);
2991 			if (m_run == NULL)
2992 				break;
2993 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2994 			m_runs[RUN_INDEX(count)] = m_run;
2995 			count++;
2996 		}
2997 
2998 		/*
2999 		 * Reclaim the highest runs in LIFO (descending) order until
3000 		 * the number of reclaimed pages, "reclaimed", is at least
3001 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
3002 		 * reclamation is idempotent, and runs will (likely) recur
3003 		 * from one scan to the next as restrictions are relaxed.
3004 		 */
3005 		reclaimed = 0;
3006 		for (i = 0; count > 0 && i < NRUNS; i++) {
3007 			count--;
3008 			m_run = m_runs[RUN_INDEX(count)];
3009 			error = vm_page_reclaim_run(req_class, domain, npages,
3010 			    m_run, high);
3011 			if (error == 0) {
3012 				reclaimed += npages;
3013 				if (reclaimed >= MIN_RECLAIM)
3014 					return (true);
3015 			}
3016 		}
3017 
3018 		/*
3019 		 * Either relax the restrictions on the next scan or return if
3020 		 * the last scan had no restrictions.
3021 		 */
3022 		if (options == VPSC_NORESERV)
3023 			options = VPSC_NOSUPER;
3024 		else if (options == VPSC_NOSUPER)
3025 			options = VPSC_ANY;
3026 		else if (options == VPSC_ANY)
3027 			return (reclaimed != 0);
3028 	}
3029 }
3030 
3031 bool
3032 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3033     u_long alignment, vm_paddr_t boundary)
3034 {
3035 	struct vm_domainset_iter di;
3036 	int domain;
3037 	bool ret;
3038 
3039 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3040 	do {
3041 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3042 		    high, alignment, boundary);
3043 		if (ret)
3044 			break;
3045 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3046 
3047 	return (ret);
3048 }
3049 
3050 /*
3051  * Set the domain in the appropriate page level domainset.
3052  */
3053 void
3054 vm_domain_set(struct vm_domain *vmd)
3055 {
3056 
3057 	mtx_lock(&vm_domainset_lock);
3058 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3059 		vmd->vmd_minset = 1;
3060 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3061 	}
3062 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3063 		vmd->vmd_severeset = 1;
3064 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3065 	}
3066 	mtx_unlock(&vm_domainset_lock);
3067 }
3068 
3069 /*
3070  * Clear the domain from the appropriate page level domainset.
3071  */
3072 void
3073 vm_domain_clear(struct vm_domain *vmd)
3074 {
3075 
3076 	mtx_lock(&vm_domainset_lock);
3077 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3078 		vmd->vmd_minset = 0;
3079 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3080 		if (vm_min_waiters != 0) {
3081 			vm_min_waiters = 0;
3082 			wakeup(&vm_min_domains);
3083 		}
3084 	}
3085 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3086 		vmd->vmd_severeset = 0;
3087 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3088 		if (vm_severe_waiters != 0) {
3089 			vm_severe_waiters = 0;
3090 			wakeup(&vm_severe_domains);
3091 		}
3092 	}
3093 
3094 	/*
3095 	 * If pageout daemon needs pages, then tell it that there are
3096 	 * some free.
3097 	 */
3098 	if (vmd->vmd_pageout_pages_needed &&
3099 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3100 		wakeup(&vmd->vmd_pageout_pages_needed);
3101 		vmd->vmd_pageout_pages_needed = 0;
3102 	}
3103 
3104 	/* See comments in vm_wait_doms(). */
3105 	if (vm_pageproc_waiters) {
3106 		vm_pageproc_waiters = 0;
3107 		wakeup(&vm_pageproc_waiters);
3108 	}
3109 	mtx_unlock(&vm_domainset_lock);
3110 }
3111 
3112 /*
3113  * Wait for free pages to exceed the min threshold globally.
3114  */
3115 void
3116 vm_wait_min(void)
3117 {
3118 
3119 	mtx_lock(&vm_domainset_lock);
3120 	while (vm_page_count_min()) {
3121 		vm_min_waiters++;
3122 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3123 	}
3124 	mtx_unlock(&vm_domainset_lock);
3125 }
3126 
3127 /*
3128  * Wait for free pages to exceed the severe threshold globally.
3129  */
3130 void
3131 vm_wait_severe(void)
3132 {
3133 
3134 	mtx_lock(&vm_domainset_lock);
3135 	while (vm_page_count_severe()) {
3136 		vm_severe_waiters++;
3137 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3138 		    "vmwait", 0);
3139 	}
3140 	mtx_unlock(&vm_domainset_lock);
3141 }
3142 
3143 u_int
3144 vm_wait_count(void)
3145 {
3146 
3147 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3148 }
3149 
3150 void
3151 vm_wait_doms(const domainset_t *wdoms)
3152 {
3153 
3154 	/*
3155 	 * We use racey wakeup synchronization to avoid expensive global
3156 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3157 	 * To handle this, we only sleep for one tick in this instance.  It
3158 	 * is expected that most allocations for the pageproc will come from
3159 	 * kmem or vm_page_grab* which will use the more specific and
3160 	 * race-free vm_wait_domain().
3161 	 */
3162 	if (curproc == pageproc) {
3163 		mtx_lock(&vm_domainset_lock);
3164 		vm_pageproc_waiters++;
3165 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3166 		    "pageprocwait", 1);
3167 	} else {
3168 		/*
3169 		 * XXX Ideally we would wait only until the allocation could
3170 		 * be satisfied.  This condition can cause new allocators to
3171 		 * consume all freed pages while old allocators wait.
3172 		 */
3173 		mtx_lock(&vm_domainset_lock);
3174 		if (vm_page_count_min_set(wdoms)) {
3175 			vm_min_waiters++;
3176 			msleep(&vm_min_domains, &vm_domainset_lock,
3177 			    PVM | PDROP, "vmwait", 0);
3178 		} else
3179 			mtx_unlock(&vm_domainset_lock);
3180 	}
3181 }
3182 
3183 /*
3184  *	vm_wait_domain:
3185  *
3186  *	Sleep until free pages are available for allocation.
3187  *	- Called in various places after failed memory allocations.
3188  */
3189 void
3190 vm_wait_domain(int domain)
3191 {
3192 	struct vm_domain *vmd;
3193 	domainset_t wdom;
3194 
3195 	vmd = VM_DOMAIN(domain);
3196 	vm_domain_free_assert_unlocked(vmd);
3197 
3198 	if (curproc == pageproc) {
3199 		mtx_lock(&vm_domainset_lock);
3200 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3201 			vmd->vmd_pageout_pages_needed = 1;
3202 			msleep(&vmd->vmd_pageout_pages_needed,
3203 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3204 		} else
3205 			mtx_unlock(&vm_domainset_lock);
3206 	} else {
3207 		if (pageproc == NULL)
3208 			panic("vm_wait in early boot");
3209 		DOMAINSET_ZERO(&wdom);
3210 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3211 		vm_wait_doms(&wdom);
3212 	}
3213 }
3214 
3215 /*
3216  *	vm_wait:
3217  *
3218  *	Sleep until free pages are available for allocation in the
3219  *	affinity domains of the obj.  If obj is NULL, the domain set
3220  *	for the calling thread is used.
3221  *	Called in various places after failed memory allocations.
3222  */
3223 void
3224 vm_wait(vm_object_t obj)
3225 {
3226 	struct domainset *d;
3227 
3228 	d = NULL;
3229 
3230 	/*
3231 	 * Carefully fetch pointers only once: the struct domainset
3232 	 * itself is ummutable but the pointer might change.
3233 	 */
3234 	if (obj != NULL)
3235 		d = obj->domain.dr_policy;
3236 	if (d == NULL)
3237 		d = curthread->td_domain.dr_policy;
3238 
3239 	vm_wait_doms(&d->ds_mask);
3240 }
3241 
3242 /*
3243  *	vm_domain_alloc_fail:
3244  *
3245  *	Called when a page allocation function fails.  Informs the
3246  *	pagedaemon and performs the requested wait.  Requires the
3247  *	domain_free and object lock on entry.  Returns with the
3248  *	object lock held and free lock released.  Returns an error when
3249  *	retry is necessary.
3250  *
3251  */
3252 static int
3253 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3254 {
3255 
3256 	vm_domain_free_assert_unlocked(vmd);
3257 
3258 	atomic_add_int(&vmd->vmd_pageout_deficit,
3259 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3260 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3261 		if (object != NULL)
3262 			VM_OBJECT_WUNLOCK(object);
3263 		vm_wait_domain(vmd->vmd_domain);
3264 		if (object != NULL)
3265 			VM_OBJECT_WLOCK(object);
3266 		if (req & VM_ALLOC_WAITOK)
3267 			return (EAGAIN);
3268 	}
3269 
3270 	return (0);
3271 }
3272 
3273 /*
3274  *	vm_waitpfault:
3275  *
3276  *	Sleep until free pages are available for allocation.
3277  *	- Called only in vm_fault so that processes page faulting
3278  *	  can be easily tracked.
3279  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3280  *	  processes will be able to grab memory first.  Do not change
3281  *	  this balance without careful testing first.
3282  */
3283 void
3284 vm_waitpfault(struct domainset *dset, int timo)
3285 {
3286 
3287 	/*
3288 	 * XXX Ideally we would wait only until the allocation could
3289 	 * be satisfied.  This condition can cause new allocators to
3290 	 * consume all freed pages while old allocators wait.
3291 	 */
3292 	mtx_lock(&vm_domainset_lock);
3293 	if (vm_page_count_min_set(&dset->ds_mask)) {
3294 		vm_min_waiters++;
3295 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3296 		    "pfault", timo);
3297 	} else
3298 		mtx_unlock(&vm_domainset_lock);
3299 }
3300 
3301 static struct vm_pagequeue *
3302 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3303 {
3304 
3305 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3306 }
3307 
3308 #ifdef INVARIANTS
3309 static struct vm_pagequeue *
3310 vm_page_pagequeue(vm_page_t m)
3311 {
3312 
3313 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3314 }
3315 #endif
3316 
3317 static __always_inline bool
3318 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3319 {
3320 	vm_page_astate_t tmp;
3321 
3322 	tmp = *old;
3323 	do {
3324 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3325 			return (true);
3326 		counter_u64_add(pqstate_commit_retries, 1);
3327 	} while (old->_bits == tmp._bits);
3328 
3329 	return (false);
3330 }
3331 
3332 /*
3333  * Do the work of committing a queue state update that moves the page out of
3334  * its current queue.
3335  */
3336 static bool
3337 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3338     vm_page_astate_t *old, vm_page_astate_t new)
3339 {
3340 	vm_page_t next;
3341 
3342 	vm_pagequeue_assert_locked(pq);
3343 	KASSERT(vm_page_pagequeue(m) == pq,
3344 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3345 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3346 	    ("%s: invalid queue indices %d %d",
3347 	    __func__, old->queue, new.queue));
3348 
3349 	/*
3350 	 * Once the queue index of the page changes there is nothing
3351 	 * synchronizing with further updates to the page's physical
3352 	 * queue state.  Therefore we must speculatively remove the page
3353 	 * from the queue now and be prepared to roll back if the queue
3354 	 * state update fails.  If the page is not physically enqueued then
3355 	 * we just update its queue index.
3356 	 */
3357 	if ((old->flags & PGA_ENQUEUED) != 0) {
3358 		new.flags &= ~PGA_ENQUEUED;
3359 		next = TAILQ_NEXT(m, plinks.q);
3360 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3361 		vm_pagequeue_cnt_dec(pq);
3362 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3363 			if (next == NULL)
3364 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3365 			else
3366 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3367 			vm_pagequeue_cnt_inc(pq);
3368 			return (false);
3369 		} else {
3370 			return (true);
3371 		}
3372 	} else {
3373 		return (vm_page_pqstate_fcmpset(m, old, new));
3374 	}
3375 }
3376 
3377 static bool
3378 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3379     vm_page_astate_t new)
3380 {
3381 	struct vm_pagequeue *pq;
3382 	vm_page_astate_t as;
3383 	bool ret;
3384 
3385 	pq = _vm_page_pagequeue(m, old->queue);
3386 
3387 	/*
3388 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3389 	 * corresponding page queue lock is held.
3390 	 */
3391 	vm_pagequeue_lock(pq);
3392 	as = vm_page_astate_load(m);
3393 	if (__predict_false(as._bits != old->_bits)) {
3394 		*old = as;
3395 		ret = false;
3396 	} else {
3397 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3398 	}
3399 	vm_pagequeue_unlock(pq);
3400 	return (ret);
3401 }
3402 
3403 /*
3404  * Commit a queue state update that enqueues or requeues a page.
3405  */
3406 static bool
3407 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3408     vm_page_astate_t *old, vm_page_astate_t new)
3409 {
3410 	struct vm_domain *vmd;
3411 
3412 	vm_pagequeue_assert_locked(pq);
3413 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3414 	    ("%s: invalid queue indices %d %d",
3415 	    __func__, old->queue, new.queue));
3416 
3417 	new.flags |= PGA_ENQUEUED;
3418 	if (!vm_page_pqstate_fcmpset(m, old, new))
3419 		return (false);
3420 
3421 	if ((old->flags & PGA_ENQUEUED) != 0)
3422 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3423 	else
3424 		vm_pagequeue_cnt_inc(pq);
3425 
3426 	/*
3427 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3428 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3429 	 * applied, even if it was set first.
3430 	 */
3431 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3432 		vmd = vm_pagequeue_domain(m);
3433 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3434 		    ("%s: invalid page queue for page %p", __func__, m));
3435 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3436 	} else {
3437 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3438 	}
3439 	return (true);
3440 }
3441 
3442 /*
3443  * Commit a queue state update that encodes a request for a deferred queue
3444  * operation.
3445  */
3446 static bool
3447 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3448     vm_page_astate_t new)
3449 {
3450 
3451 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3452 	    ("%s: invalid state, queue %d flags %x",
3453 	    __func__, new.queue, new.flags));
3454 
3455 	if (old->_bits != new._bits &&
3456 	    !vm_page_pqstate_fcmpset(m, old, new))
3457 		return (false);
3458 	vm_page_pqbatch_submit(m, new.queue);
3459 	return (true);
3460 }
3461 
3462 /*
3463  * A generic queue state update function.  This handles more cases than the
3464  * specialized functions above.
3465  */
3466 bool
3467 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3468 {
3469 
3470 	if (old->_bits == new._bits)
3471 		return (true);
3472 
3473 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3474 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3475 			return (false);
3476 		if (new.queue != PQ_NONE)
3477 			vm_page_pqbatch_submit(m, new.queue);
3478 	} else {
3479 		if (!vm_page_pqstate_fcmpset(m, old, new))
3480 			return (false);
3481 		if (new.queue != PQ_NONE &&
3482 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3483 			vm_page_pqbatch_submit(m, new.queue);
3484 	}
3485 	return (true);
3486 }
3487 
3488 /*
3489  * Apply deferred queue state updates to a page.
3490  */
3491 static inline void
3492 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3493 {
3494 	vm_page_astate_t new, old;
3495 
3496 	CRITICAL_ASSERT(curthread);
3497 	vm_pagequeue_assert_locked(pq);
3498 	KASSERT(queue < PQ_COUNT,
3499 	    ("%s: invalid queue index %d", __func__, queue));
3500 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3501 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3502 
3503 	for (old = vm_page_astate_load(m);;) {
3504 		if (__predict_false(old.queue != queue ||
3505 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3506 			counter_u64_add(queue_nops, 1);
3507 			break;
3508 		}
3509 		KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3510 		    ("%s: page %p has unexpected queue state", __func__, m));
3511 
3512 		new = old;
3513 		if ((old.flags & PGA_DEQUEUE) != 0) {
3514 			new.flags &= ~PGA_QUEUE_OP_MASK;
3515 			new.queue = PQ_NONE;
3516 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3517 			    m, &old, new))) {
3518 				counter_u64_add(queue_ops, 1);
3519 				break;
3520 			}
3521 		} else {
3522 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3523 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3524 			    m, &old, new))) {
3525 				counter_u64_add(queue_ops, 1);
3526 				break;
3527 			}
3528 		}
3529 	}
3530 }
3531 
3532 static void
3533 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3534     uint8_t queue)
3535 {
3536 	int i;
3537 
3538 	for (i = 0; i < bq->bq_cnt; i++)
3539 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3540 	vm_batchqueue_init(bq);
3541 }
3542 
3543 /*
3544  *	vm_page_pqbatch_submit:		[ internal use only ]
3545  *
3546  *	Enqueue a page in the specified page queue's batched work queue.
3547  *	The caller must have encoded the requested operation in the page
3548  *	structure's a.flags field.
3549  */
3550 void
3551 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3552 {
3553 	struct vm_batchqueue *bq;
3554 	struct vm_pagequeue *pq;
3555 	int domain;
3556 
3557 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3558 	    ("page %p is unmanaged", m));
3559 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3560 
3561 	domain = vm_phys_domain(m);
3562 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3563 
3564 	critical_enter();
3565 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3566 	if (vm_batchqueue_insert(bq, m)) {
3567 		critical_exit();
3568 		return;
3569 	}
3570 	critical_exit();
3571 	vm_pagequeue_lock(pq);
3572 	critical_enter();
3573 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3574 	vm_pqbatch_process(pq, bq, queue);
3575 	vm_pqbatch_process_page(pq, m, queue);
3576 	vm_pagequeue_unlock(pq);
3577 	critical_exit();
3578 }
3579 
3580 /*
3581  *	vm_page_pqbatch_drain:		[ internal use only ]
3582  *
3583  *	Force all per-CPU page queue batch queues to be drained.  This is
3584  *	intended for use in severe memory shortages, to ensure that pages
3585  *	do not remain stuck in the batch queues.
3586  */
3587 void
3588 vm_page_pqbatch_drain(void)
3589 {
3590 	struct thread *td;
3591 	struct vm_domain *vmd;
3592 	struct vm_pagequeue *pq;
3593 	int cpu, domain, queue;
3594 
3595 	td = curthread;
3596 	CPU_FOREACH(cpu) {
3597 		thread_lock(td);
3598 		sched_bind(td, cpu);
3599 		thread_unlock(td);
3600 
3601 		for (domain = 0; domain < vm_ndomains; domain++) {
3602 			vmd = VM_DOMAIN(domain);
3603 			for (queue = 0; queue < PQ_COUNT; queue++) {
3604 				pq = &vmd->vmd_pagequeues[queue];
3605 				vm_pagequeue_lock(pq);
3606 				critical_enter();
3607 				vm_pqbatch_process(pq,
3608 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3609 				critical_exit();
3610 				vm_pagequeue_unlock(pq);
3611 			}
3612 		}
3613 	}
3614 	thread_lock(td);
3615 	sched_unbind(td);
3616 	thread_unlock(td);
3617 }
3618 
3619 /*
3620  *	vm_page_dequeue_deferred:	[ internal use only ]
3621  *
3622  *	Request removal of the given page from its current page
3623  *	queue.  Physical removal from the queue may be deferred
3624  *	indefinitely.
3625  */
3626 void
3627 vm_page_dequeue_deferred(vm_page_t m)
3628 {
3629 	vm_page_astate_t new, old;
3630 
3631 	old = vm_page_astate_load(m);
3632 	do {
3633 		if (old.queue == PQ_NONE) {
3634 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3635 			    ("%s: page %p has unexpected queue state",
3636 			    __func__, m));
3637 			break;
3638 		}
3639 		new = old;
3640 		new.flags |= PGA_DEQUEUE;
3641 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3642 }
3643 
3644 /*
3645  *	vm_page_dequeue:
3646  *
3647  *	Remove the page from whichever page queue it's in, if any, before
3648  *	returning.
3649  */
3650 void
3651 vm_page_dequeue(vm_page_t m)
3652 {
3653 	vm_page_astate_t new, old;
3654 
3655 	old = vm_page_astate_load(m);
3656 	do {
3657 		if (old.queue == PQ_NONE) {
3658 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3659 			    ("%s: page %p has unexpected queue state",
3660 			    __func__, m));
3661 			break;
3662 		}
3663 		new = old;
3664 		new.flags &= ~PGA_QUEUE_OP_MASK;
3665 		new.queue = PQ_NONE;
3666 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3667 
3668 }
3669 
3670 /*
3671  * Schedule the given page for insertion into the specified page queue.
3672  * Physical insertion of the page may be deferred indefinitely.
3673  */
3674 static void
3675 vm_page_enqueue(vm_page_t m, uint8_t queue)
3676 {
3677 
3678 	KASSERT(m->a.queue == PQ_NONE &&
3679 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3680 	    ("%s: page %p is already enqueued", __func__, m));
3681 	KASSERT(m->ref_count > 0,
3682 	    ("%s: page %p does not carry any references", __func__, m));
3683 
3684 	m->a.queue = queue;
3685 	if ((m->a.flags & PGA_REQUEUE) == 0)
3686 		vm_page_aflag_set(m, PGA_REQUEUE);
3687 	vm_page_pqbatch_submit(m, queue);
3688 }
3689 
3690 /*
3691  *	vm_page_free_prep:
3692  *
3693  *	Prepares the given page to be put on the free list,
3694  *	disassociating it from any VM object. The caller may return
3695  *	the page to the free list only if this function returns true.
3696  *
3697  *	The object, if it exists, must be locked, and then the page must
3698  *	be xbusy.  Otherwise the page must be not busied.  A managed
3699  *	page must be unmapped.
3700  */
3701 static bool
3702 vm_page_free_prep(vm_page_t m)
3703 {
3704 
3705 	/*
3706 	 * Synchronize with threads that have dropped a reference to this
3707 	 * page.
3708 	 */
3709 	atomic_thread_fence_acq();
3710 
3711 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3712 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3713 		uint64_t *p;
3714 		int i;
3715 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3716 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3717 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3718 			    m, i, (uintmax_t)*p));
3719 	}
3720 #endif
3721 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3722 		KASSERT(!pmap_page_is_mapped(m),
3723 		    ("vm_page_free_prep: freeing mapped page %p", m));
3724 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3725 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3726 	} else {
3727 		KASSERT(m->a.queue == PQ_NONE,
3728 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3729 	}
3730 	VM_CNT_INC(v_tfree);
3731 
3732 	if (m->object != NULL) {
3733 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3734 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3735 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3736 		    m));
3737 		vm_page_assert_xbusied(m);
3738 
3739 		/*
3740 		 * The object reference can be released without an atomic
3741 		 * operation.
3742 		 */
3743 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3744 		    m->ref_count == VPRC_OBJREF,
3745 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3746 		    m, m->ref_count));
3747 		vm_page_object_remove(m);
3748 		m->ref_count -= VPRC_OBJREF;
3749 	} else
3750 		vm_page_assert_unbusied(m);
3751 
3752 	vm_page_busy_free(m);
3753 
3754 	/*
3755 	 * If fictitious remove object association and
3756 	 * return.
3757 	 */
3758 	if ((m->flags & PG_FICTITIOUS) != 0) {
3759 		KASSERT(m->ref_count == 1,
3760 		    ("fictitious page %p is referenced", m));
3761 		KASSERT(m->a.queue == PQ_NONE,
3762 		    ("fictitious page %p is queued", m));
3763 		return (false);
3764 	}
3765 
3766 	/*
3767 	 * Pages need not be dequeued before they are returned to the physical
3768 	 * memory allocator, but they must at least be marked for a deferred
3769 	 * dequeue.
3770 	 */
3771 	if ((m->oflags & VPO_UNMANAGED) == 0)
3772 		vm_page_dequeue_deferred(m);
3773 
3774 	m->valid = 0;
3775 	vm_page_undirty(m);
3776 
3777 	if (m->ref_count != 0)
3778 		panic("vm_page_free_prep: page %p has references", m);
3779 
3780 	/*
3781 	 * Restore the default memory attribute to the page.
3782 	 */
3783 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3784 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3785 
3786 #if VM_NRESERVLEVEL > 0
3787 	/*
3788 	 * Determine whether the page belongs to a reservation.  If the page was
3789 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3790 	 * as an optimization, we avoid the check in that case.
3791 	 */
3792 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3793 		return (false);
3794 #endif
3795 
3796 	return (true);
3797 }
3798 
3799 /*
3800  *	vm_page_free_toq:
3801  *
3802  *	Returns the given page to the free list, disassociating it
3803  *	from any VM object.
3804  *
3805  *	The object must be locked.  The page must be exclusively busied if it
3806  *	belongs to an object.
3807  */
3808 static void
3809 vm_page_free_toq(vm_page_t m)
3810 {
3811 	struct vm_domain *vmd;
3812 	uma_zone_t zone;
3813 
3814 	if (!vm_page_free_prep(m))
3815 		return;
3816 
3817 	vmd = vm_pagequeue_domain(m);
3818 	zone = vmd->vmd_pgcache[m->pool].zone;
3819 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3820 		uma_zfree(zone, m);
3821 		return;
3822 	}
3823 	vm_domain_free_lock(vmd);
3824 	vm_phys_free_pages(m, 0);
3825 	vm_domain_free_unlock(vmd);
3826 	vm_domain_freecnt_inc(vmd, 1);
3827 }
3828 
3829 /*
3830  *	vm_page_free_pages_toq:
3831  *
3832  *	Returns a list of pages to the free list, disassociating it
3833  *	from any VM object.  In other words, this is equivalent to
3834  *	calling vm_page_free_toq() for each page of a list of VM objects.
3835  */
3836 void
3837 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3838 {
3839 	vm_page_t m;
3840 	int count;
3841 
3842 	if (SLIST_EMPTY(free))
3843 		return;
3844 
3845 	count = 0;
3846 	while ((m = SLIST_FIRST(free)) != NULL) {
3847 		count++;
3848 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3849 		vm_page_free_toq(m);
3850 	}
3851 
3852 	if (update_wire_count)
3853 		vm_wire_sub(count);
3854 }
3855 
3856 /*
3857  * Mark this page as wired down, preventing reclamation by the page daemon
3858  * or when the containing object is destroyed.
3859  */
3860 void
3861 vm_page_wire(vm_page_t m)
3862 {
3863 	u_int old;
3864 
3865 	KASSERT(m->object != NULL,
3866 	    ("vm_page_wire: page %p does not belong to an object", m));
3867 	if (!vm_page_busied(m) && !vm_object_busied(m->object))
3868 		VM_OBJECT_ASSERT_LOCKED(m->object);
3869 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3870 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3871 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3872 
3873 	old = atomic_fetchadd_int(&m->ref_count, 1);
3874 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3875 	    ("vm_page_wire: counter overflow for page %p", m));
3876 	if (VPRC_WIRE_COUNT(old) == 0) {
3877 		if ((m->oflags & VPO_UNMANAGED) == 0)
3878 			vm_page_aflag_set(m, PGA_DEQUEUE);
3879 		vm_wire_add(1);
3880 	}
3881 }
3882 
3883 /*
3884  * Attempt to wire a mapped page following a pmap lookup of that page.
3885  * This may fail if a thread is concurrently tearing down mappings of the page.
3886  * The transient failure is acceptable because it translates to the
3887  * failure of the caller pmap_extract_and_hold(), which should be then
3888  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3889  */
3890 bool
3891 vm_page_wire_mapped(vm_page_t m)
3892 {
3893 	u_int old;
3894 
3895 	old = m->ref_count;
3896 	do {
3897 		KASSERT(old > 0,
3898 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3899 		if ((old & VPRC_BLOCKED) != 0)
3900 			return (false);
3901 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3902 
3903 	if (VPRC_WIRE_COUNT(old) == 0) {
3904 		if ((m->oflags & VPO_UNMANAGED) == 0)
3905 			vm_page_aflag_set(m, PGA_DEQUEUE);
3906 		vm_wire_add(1);
3907 	}
3908 	return (true);
3909 }
3910 
3911 /*
3912  * Release a wiring reference to a managed page.  If the page still belongs to
3913  * an object, update its position in the page queues to reflect the reference.
3914  * If the wiring was the last reference to the page, free the page.
3915  */
3916 static void
3917 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3918 {
3919 	u_int old;
3920 
3921 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3922 	    ("%s: page %p is unmanaged", __func__, m));
3923 
3924 	/*
3925 	 * Update LRU state before releasing the wiring reference.
3926 	 * Use a release store when updating the reference count to
3927 	 * synchronize with vm_page_free_prep().
3928 	 */
3929 	old = m->ref_count;
3930 	do {
3931 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3932 		    ("vm_page_unwire: wire count underflow for page %p", m));
3933 
3934 		if (old > VPRC_OBJREF + 1) {
3935 			/*
3936 			 * The page has at least one other wiring reference.  An
3937 			 * earlier iteration of this loop may have called
3938 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3939 			 * re-set it if necessary.
3940 			 */
3941 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3942 				vm_page_aflag_set(m, PGA_DEQUEUE);
3943 		} else if (old == VPRC_OBJREF + 1) {
3944 			/*
3945 			 * This is the last wiring.  Clear PGA_DEQUEUE and
3946 			 * update the page's queue state to reflect the
3947 			 * reference.  If the page does not belong to an object
3948 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3949 			 * clear leftover queue state.
3950 			 */
3951 			vm_page_release_toq(m, nqueue, false);
3952 		} else if (old == 1) {
3953 			vm_page_aflag_clear(m, PGA_DEQUEUE);
3954 		}
3955 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3956 
3957 	if (VPRC_WIRE_COUNT(old) == 1) {
3958 		vm_wire_sub(1);
3959 		if (old == 1)
3960 			vm_page_free(m);
3961 	}
3962 }
3963 
3964 /*
3965  * Release one wiring of the specified page, potentially allowing it to be
3966  * paged out.
3967  *
3968  * Only managed pages belonging to an object can be paged out.  If the number
3969  * of wirings transitions to zero and the page is eligible for page out, then
3970  * the page is added to the specified paging queue.  If the released wiring
3971  * represented the last reference to the page, the page is freed.
3972  */
3973 void
3974 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3975 {
3976 
3977 	KASSERT(nqueue < PQ_COUNT,
3978 	    ("vm_page_unwire: invalid queue %u request for page %p",
3979 	    nqueue, m));
3980 
3981 	if ((m->oflags & VPO_UNMANAGED) != 0) {
3982 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
3983 			vm_page_free(m);
3984 		return;
3985 	}
3986 	vm_page_unwire_managed(m, nqueue, false);
3987 }
3988 
3989 /*
3990  * Unwire a page without (re-)inserting it into a page queue.  It is up
3991  * to the caller to enqueue, requeue, or free the page as appropriate.
3992  * In most cases involving managed pages, vm_page_unwire() should be used
3993  * instead.
3994  */
3995 bool
3996 vm_page_unwire_noq(vm_page_t m)
3997 {
3998 	u_int old;
3999 
4000 	old = vm_page_drop(m, 1);
4001 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4002 	    ("vm_page_unref: counter underflow for page %p", m));
4003 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4004 	    ("vm_page_unref: missing ref on fictitious page %p", m));
4005 
4006 	if (VPRC_WIRE_COUNT(old) > 1)
4007 		return (false);
4008 	if ((m->oflags & VPO_UNMANAGED) == 0)
4009 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4010 	vm_wire_sub(1);
4011 	return (true);
4012 }
4013 
4014 /*
4015  * Ensure that the page ends up in the specified page queue.  If the page is
4016  * active or being moved to the active queue, ensure that its act_count is
4017  * at least ACT_INIT but do not otherwise mess with it.
4018  */
4019 static __always_inline void
4020 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4021 {
4022 	vm_page_astate_t old, new;
4023 
4024 	KASSERT(m->ref_count > 0,
4025 	    ("%s: page %p does not carry any references", __func__, m));
4026 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4027 	    ("%s: invalid flags %x", __func__, nflag));
4028 
4029 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4030 		return;
4031 
4032 	old = vm_page_astate_load(m);
4033 	do {
4034 		if ((old.flags & PGA_DEQUEUE) != 0)
4035 			break;
4036 		new = old;
4037 		new.flags &= ~PGA_QUEUE_OP_MASK;
4038 		if (nqueue == PQ_ACTIVE)
4039 			new.act_count = max(old.act_count, ACT_INIT);
4040 		if (old.queue == nqueue) {
4041 			if (nqueue != PQ_ACTIVE)
4042 				new.flags |= nflag;
4043 		} else {
4044 			new.flags |= nflag;
4045 			new.queue = nqueue;
4046 		}
4047 	} while (!vm_page_pqstate_commit(m, &old, new));
4048 }
4049 
4050 /*
4051  * Put the specified page on the active list (if appropriate).
4052  */
4053 void
4054 vm_page_activate(vm_page_t m)
4055 {
4056 
4057 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4058 }
4059 
4060 /*
4061  * Move the specified page to the tail of the inactive queue, or requeue
4062  * the page if it is already in the inactive queue.
4063  */
4064 void
4065 vm_page_deactivate(vm_page_t m)
4066 {
4067 
4068 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4069 }
4070 
4071 void
4072 vm_page_deactivate_noreuse(vm_page_t m)
4073 {
4074 
4075 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4076 }
4077 
4078 /*
4079  * Put a page in the laundry, or requeue it if it is already there.
4080  */
4081 void
4082 vm_page_launder(vm_page_t m)
4083 {
4084 
4085 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4086 }
4087 
4088 /*
4089  * Put a page in the PQ_UNSWAPPABLE holding queue.
4090  */
4091 void
4092 vm_page_unswappable(vm_page_t m)
4093 {
4094 
4095 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4096 	    ("page %p already unswappable", m));
4097 
4098 	vm_page_dequeue(m);
4099 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4100 }
4101 
4102 /*
4103  * Release a page back to the page queues in preparation for unwiring.
4104  */
4105 static void
4106 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4107 {
4108 	vm_page_astate_t old, new;
4109 	uint16_t nflag;
4110 
4111 	/*
4112 	 * Use a check of the valid bits to determine whether we should
4113 	 * accelerate reclamation of the page.  The object lock might not be
4114 	 * held here, in which case the check is racy.  At worst we will either
4115 	 * accelerate reclamation of a valid page and violate LRU, or
4116 	 * unnecessarily defer reclamation of an invalid page.
4117 	 *
4118 	 * If we were asked to not cache the page, place it near the head of the
4119 	 * inactive queue so that is reclaimed sooner.
4120 	 */
4121 	if (noreuse || m->valid == 0) {
4122 		nqueue = PQ_INACTIVE;
4123 		nflag = PGA_REQUEUE_HEAD;
4124 	} else {
4125 		nflag = PGA_REQUEUE;
4126 	}
4127 
4128 	old = vm_page_astate_load(m);
4129 	do {
4130 		new = old;
4131 
4132 		/*
4133 		 * If the page is already in the active queue and we are not
4134 		 * trying to accelerate reclamation, simply mark it as
4135 		 * referenced and avoid any queue operations.
4136 		 */
4137 		new.flags &= ~PGA_QUEUE_OP_MASK;
4138 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4139 			new.flags |= PGA_REFERENCED;
4140 		else {
4141 			new.flags |= nflag;
4142 			new.queue = nqueue;
4143 		}
4144 	} while (!vm_page_pqstate_commit(m, &old, new));
4145 }
4146 
4147 /*
4148  * Unwire a page and either attempt to free it or re-add it to the page queues.
4149  */
4150 void
4151 vm_page_release(vm_page_t m, int flags)
4152 {
4153 	vm_object_t object;
4154 
4155 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4156 	    ("vm_page_release: page %p is unmanaged", m));
4157 
4158 	if ((flags & VPR_TRYFREE) != 0) {
4159 		for (;;) {
4160 			object = atomic_load_ptr(&m->object);
4161 			if (object == NULL)
4162 				break;
4163 			/* Depends on type-stability. */
4164 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4165 				break;
4166 			if (object == m->object) {
4167 				vm_page_release_locked(m, flags);
4168 				VM_OBJECT_WUNLOCK(object);
4169 				return;
4170 			}
4171 			VM_OBJECT_WUNLOCK(object);
4172 		}
4173 	}
4174 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4175 }
4176 
4177 /* See vm_page_release(). */
4178 void
4179 vm_page_release_locked(vm_page_t m, int flags)
4180 {
4181 
4182 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4183 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4184 	    ("vm_page_release_locked: page %p is unmanaged", m));
4185 
4186 	if (vm_page_unwire_noq(m)) {
4187 		if ((flags & VPR_TRYFREE) != 0 &&
4188 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4189 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4190 			/*
4191 			 * An unlocked lookup may have wired the page before the
4192 			 * busy lock was acquired, in which case the page must
4193 			 * not be freed.
4194 			 */
4195 			if (__predict_true(!vm_page_wired(m))) {
4196 				vm_page_free(m);
4197 				return;
4198 			}
4199 			vm_page_xunbusy(m);
4200 		} else {
4201 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4202 		}
4203 	}
4204 }
4205 
4206 static bool
4207 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4208 {
4209 	u_int old;
4210 
4211 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4212 	    ("vm_page_try_blocked_op: page %p has no object", m));
4213 	KASSERT(vm_page_busied(m),
4214 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4215 	VM_OBJECT_ASSERT_LOCKED(m->object);
4216 
4217 	old = m->ref_count;
4218 	do {
4219 		KASSERT(old != 0,
4220 		    ("vm_page_try_blocked_op: page %p has no references", m));
4221 		if (VPRC_WIRE_COUNT(old) != 0)
4222 			return (false);
4223 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4224 
4225 	(op)(m);
4226 
4227 	/*
4228 	 * If the object is read-locked, new wirings may be created via an
4229 	 * object lookup.
4230 	 */
4231 	old = vm_page_drop(m, VPRC_BLOCKED);
4232 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4233 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4234 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4235 	    old, m));
4236 	return (true);
4237 }
4238 
4239 /*
4240  * Atomically check for wirings and remove all mappings of the page.
4241  */
4242 bool
4243 vm_page_try_remove_all(vm_page_t m)
4244 {
4245 
4246 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4247 }
4248 
4249 /*
4250  * Atomically check for wirings and remove all writeable mappings of the page.
4251  */
4252 bool
4253 vm_page_try_remove_write(vm_page_t m)
4254 {
4255 
4256 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4257 }
4258 
4259 /*
4260  * vm_page_advise
4261  *
4262  * 	Apply the specified advice to the given page.
4263  */
4264 void
4265 vm_page_advise(vm_page_t m, int advice)
4266 {
4267 
4268 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4269 	vm_page_assert_xbusied(m);
4270 
4271 	if (advice == MADV_FREE)
4272 		/*
4273 		 * Mark the page clean.  This will allow the page to be freed
4274 		 * without first paging it out.  MADV_FREE pages are often
4275 		 * quickly reused by malloc(3), so we do not do anything that
4276 		 * would result in a page fault on a later access.
4277 		 */
4278 		vm_page_undirty(m);
4279 	else if (advice != MADV_DONTNEED) {
4280 		if (advice == MADV_WILLNEED)
4281 			vm_page_activate(m);
4282 		return;
4283 	}
4284 
4285 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4286 		vm_page_dirty(m);
4287 
4288 	/*
4289 	 * Clear any references to the page.  Otherwise, the page daemon will
4290 	 * immediately reactivate the page.
4291 	 */
4292 	vm_page_aflag_clear(m, PGA_REFERENCED);
4293 
4294 	/*
4295 	 * Place clean pages near the head of the inactive queue rather than
4296 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4297 	 * the page will be reused quickly.  Dirty pages not already in the
4298 	 * laundry are moved there.
4299 	 */
4300 	if (m->dirty == 0)
4301 		vm_page_deactivate_noreuse(m);
4302 	else if (!vm_page_in_laundry(m))
4303 		vm_page_launder(m);
4304 }
4305 
4306 /*
4307  *	vm_page_grab_release
4308  *
4309  *	Helper routine for grab functions to release busy on return.
4310  */
4311 static inline void
4312 vm_page_grab_release(vm_page_t m, int allocflags)
4313 {
4314 
4315 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4316 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4317 			vm_page_sunbusy(m);
4318 		else
4319 			vm_page_xunbusy(m);
4320 	}
4321 }
4322 
4323 /*
4324  *	vm_page_grab_sleep
4325  *
4326  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4327  *	if the caller should retry and false otherwise.
4328  *
4329  *	If the object is locked on entry the object will be unlocked with
4330  *	false returns and still locked but possibly having been dropped
4331  *	with true returns.
4332  */
4333 static bool
4334 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4335     const char *wmesg, int allocflags, bool locked)
4336 {
4337 
4338 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4339 		return (false);
4340 
4341 	/*
4342 	 * Reference the page before unlocking and sleeping so that
4343 	 * the page daemon is less likely to reclaim it.
4344 	 */
4345 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4346 		vm_page_reference(m);
4347 
4348 	if (_vm_page_busy_sleep(object, m, m->pindex, wmesg, allocflags,
4349 	    locked) && locked)
4350 		VM_OBJECT_WLOCK(object);
4351 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4352 		return (false);
4353 
4354 	return (true);
4355 }
4356 
4357 /*
4358  * Assert that the grab flags are valid.
4359  */
4360 static inline void
4361 vm_page_grab_check(int allocflags)
4362 {
4363 
4364 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4365 	    (allocflags & VM_ALLOC_WIRED) != 0,
4366 	    ("vm_page_grab*: the pages must be busied or wired"));
4367 
4368 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4369 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4370 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4371 }
4372 
4373 /*
4374  * Calculate the page allocation flags for grab.
4375  */
4376 static inline int
4377 vm_page_grab_pflags(int allocflags)
4378 {
4379 	int pflags;
4380 
4381 	pflags = allocflags &
4382 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4383 	    VM_ALLOC_NOBUSY);
4384 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4385 		pflags |= VM_ALLOC_WAITFAIL;
4386 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4387 		pflags |= VM_ALLOC_SBUSY;
4388 
4389 	return (pflags);
4390 }
4391 
4392 /*
4393  * Grab a page, waiting until we are waken up due to the page
4394  * changing state.  We keep on waiting, if the page continues
4395  * to be in the object.  If the page doesn't exist, first allocate it
4396  * and then conditionally zero it.
4397  *
4398  * This routine may sleep.
4399  *
4400  * The object must be locked on entry.  The lock will, however, be released
4401  * and reacquired if the routine sleeps.
4402  */
4403 vm_page_t
4404 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4405 {
4406 	vm_page_t m;
4407 
4408 	VM_OBJECT_ASSERT_WLOCKED(object);
4409 	vm_page_grab_check(allocflags);
4410 
4411 retrylookup:
4412 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4413 		if (!vm_page_tryacquire(m, allocflags)) {
4414 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4415 			    allocflags, true))
4416 				goto retrylookup;
4417 			return (NULL);
4418 		}
4419 		goto out;
4420 	}
4421 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4422 		return (NULL);
4423 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4424 	if (m == NULL) {
4425 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4426 			return (NULL);
4427 		goto retrylookup;
4428 	}
4429 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4430 		pmap_zero_page(m);
4431 
4432 out:
4433 	vm_page_grab_release(m, allocflags);
4434 
4435 	return (m);
4436 }
4437 
4438 /*
4439  * Locklessly attempt to acquire a page given a (object, pindex) tuple
4440  * and an optional previous page to avoid the radix lookup.  The resulting
4441  * page will be validated against the identity tuple and busied or wired
4442  * as requested.  A NULL *mp return guarantees that the page was not in
4443  * radix at the time of the call but callers must perform higher level
4444  * synchronization or retry the operation under a lock if they require
4445  * an atomic answer.  This is the only lock free validation routine,
4446  * other routines can depend on the resulting page state.
4447  *
4448  * The return value indicates whether the operation failed due to caller
4449  * flags.  The return is tri-state with mp:
4450  *
4451  * (true, *mp != NULL) - The operation was successful.
4452  * (true, *mp == NULL) - The page was not found in tree.
4453  * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4454  */
4455 static bool
4456 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4457     vm_page_t prev, vm_page_t *mp, int allocflags)
4458 {
4459 	vm_page_t m;
4460 
4461 	vm_page_grab_check(allocflags);
4462 	MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4463 
4464 	*mp = NULL;
4465 	for (;;) {
4466 		/*
4467 		 * We may see a false NULL here because the previous page
4468 		 * has been removed or just inserted and the list is loaded
4469 		 * without barriers.  Switch to radix to verify.
4470 		 */
4471 		if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4472 		    QMD_IS_TRASHED(m) || m->pindex != pindex ||
4473 		    atomic_load_ptr(&m->object) != object) {
4474 			prev = NULL;
4475 			/*
4476 			 * This guarantees the result is instantaneously
4477 			 * correct.
4478 			 */
4479 			m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4480 		}
4481 		if (m == NULL)
4482 			return (true);
4483 		if (vm_page_trybusy(m, allocflags)) {
4484 			if (m->object == object && m->pindex == pindex)
4485 				break;
4486 			/* relookup. */
4487 			vm_page_busy_release(m);
4488 			cpu_spinwait();
4489 			continue;
4490 		}
4491 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4492 		    allocflags, false))
4493 			return (false);
4494 	}
4495 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4496 		vm_page_wire(m);
4497 	vm_page_grab_release(m, allocflags);
4498 	*mp = m;
4499 	return (true);
4500 }
4501 
4502 /*
4503  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4504  * is not set.
4505  */
4506 vm_page_t
4507 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4508 {
4509 	vm_page_t m;
4510 
4511 	vm_page_grab_check(allocflags);
4512 
4513 	if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4514 		return (NULL);
4515 	if (m != NULL)
4516 		return (m);
4517 
4518 	/*
4519 	 * The radix lockless lookup should never return a false negative
4520 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4521 	 * was no page present at the instant of the call.  A NOCREAT caller
4522 	 * must handle create races gracefully.
4523 	 */
4524 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4525 		return (NULL);
4526 
4527 	VM_OBJECT_WLOCK(object);
4528 	m = vm_page_grab(object, pindex, allocflags);
4529 	VM_OBJECT_WUNLOCK(object);
4530 
4531 	return (m);
4532 }
4533 
4534 /*
4535  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4536  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4537  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4538  * in simultaneously.  Additional pages will be left on a paging queue but
4539  * will neither be wired nor busy regardless of allocflags.
4540  */
4541 int
4542 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4543 {
4544 	vm_page_t m;
4545 	vm_page_t ma[VM_INITIAL_PAGEIN];
4546 	int after, i, pflags, rv;
4547 
4548 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4549 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4550 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4551 	KASSERT((allocflags &
4552 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4553 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4554 	VM_OBJECT_ASSERT_WLOCKED(object);
4555 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4556 	    VM_ALLOC_WIRED);
4557 	pflags |= VM_ALLOC_WAITFAIL;
4558 
4559 retrylookup:
4560 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4561 		/*
4562 		 * If the page is fully valid it can only become invalid
4563 		 * with the object lock held.  If it is not valid it can
4564 		 * become valid with the busy lock held.  Therefore, we
4565 		 * may unnecessarily lock the exclusive busy here if we
4566 		 * race with I/O completion not using the object lock.
4567 		 * However, we will not end up with an invalid page and a
4568 		 * shared lock.
4569 		 */
4570 		if (!vm_page_trybusy(m,
4571 		    vm_page_all_valid(m) ? allocflags : 0)) {
4572 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4573 			    allocflags, true);
4574 			goto retrylookup;
4575 		}
4576 		if (vm_page_all_valid(m))
4577 			goto out;
4578 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4579 			vm_page_busy_release(m);
4580 			*mp = NULL;
4581 			return (VM_PAGER_FAIL);
4582 		}
4583 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4584 		*mp = NULL;
4585 		return (VM_PAGER_FAIL);
4586 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4587 		goto retrylookup;
4588 	}
4589 
4590 	vm_page_assert_xbusied(m);
4591 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4592 		after = MIN(after, VM_INITIAL_PAGEIN);
4593 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4594 		after = MAX(after, 1);
4595 		ma[0] = m;
4596 		for (i = 1; i < after; i++) {
4597 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4598 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4599 					break;
4600 			} else {
4601 				ma[i] = vm_page_alloc(object, m->pindex + i,
4602 				    VM_ALLOC_NORMAL);
4603 				if (ma[i] == NULL)
4604 					break;
4605 			}
4606 		}
4607 		after = i;
4608 		vm_object_pip_add(object, after);
4609 		VM_OBJECT_WUNLOCK(object);
4610 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4611 		VM_OBJECT_WLOCK(object);
4612 		vm_object_pip_wakeupn(object, after);
4613 		/* Pager may have replaced a page. */
4614 		m = ma[0];
4615 		if (rv != VM_PAGER_OK) {
4616 			for (i = 0; i < after; i++) {
4617 				if (!vm_page_wired(ma[i]))
4618 					vm_page_free(ma[i]);
4619 				else
4620 					vm_page_xunbusy(ma[i]);
4621 			}
4622 			*mp = NULL;
4623 			return (rv);
4624 		}
4625 		for (i = 1; i < after; i++)
4626 			vm_page_readahead_finish(ma[i]);
4627 		MPASS(vm_page_all_valid(m));
4628 	} else {
4629 		vm_page_zero_invalid(m, TRUE);
4630 	}
4631 out:
4632 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4633 		vm_page_wire(m);
4634 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4635 		vm_page_busy_downgrade(m);
4636 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4637 		vm_page_busy_release(m);
4638 	*mp = m;
4639 	return (VM_PAGER_OK);
4640 }
4641 
4642 /*
4643  * Locklessly grab a valid page.  If the page is not valid or not yet
4644  * allocated this will fall back to the object lock method.
4645  */
4646 int
4647 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4648     vm_pindex_t pindex, int allocflags)
4649 {
4650 	vm_page_t m;
4651 	int flags;
4652 	int error;
4653 
4654 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4655 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4656 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4657 	    "mismatch"));
4658 	KASSERT((allocflags &
4659 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4660 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4661 
4662 	/*
4663 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
4664 	 * before we can inspect the valid field and return a wired page.
4665 	 */
4666 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4667 	if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4668 		return (VM_PAGER_FAIL);
4669 	if ((m = *mp) != NULL) {
4670 		if (vm_page_all_valid(m)) {
4671 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4672 				vm_page_wire(m);
4673 			vm_page_grab_release(m, allocflags);
4674 			return (VM_PAGER_OK);
4675 		}
4676 		vm_page_busy_release(m);
4677 	}
4678 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4679 		*mp = NULL;
4680 		return (VM_PAGER_FAIL);
4681 	}
4682 	VM_OBJECT_WLOCK(object);
4683 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
4684 	VM_OBJECT_WUNLOCK(object);
4685 
4686 	return (error);
4687 }
4688 
4689 /*
4690  * Return the specified range of pages from the given object.  For each
4691  * page offset within the range, if a page already exists within the object
4692  * at that offset and it is busy, then wait for it to change state.  If,
4693  * instead, the page doesn't exist, then allocate it.
4694  *
4695  * The caller must always specify an allocation class.
4696  *
4697  * allocation classes:
4698  *	VM_ALLOC_NORMAL		normal process request
4699  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4700  *
4701  * The caller must always specify that the pages are to be busied and/or
4702  * wired.
4703  *
4704  * optional allocation flags:
4705  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4706  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4707  *	VM_ALLOC_NOWAIT		do not sleep
4708  *	VM_ALLOC_SBUSY		set page to sbusy state
4709  *	VM_ALLOC_WIRED		wire the pages
4710  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4711  *
4712  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4713  * may return a partial prefix of the requested range.
4714  */
4715 int
4716 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4717     vm_page_t *ma, int count)
4718 {
4719 	vm_page_t m, mpred;
4720 	int pflags;
4721 	int i;
4722 
4723 	VM_OBJECT_ASSERT_WLOCKED(object);
4724 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4725 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4726 	vm_page_grab_check(allocflags);
4727 
4728 	pflags = vm_page_grab_pflags(allocflags);
4729 	if (count == 0)
4730 		return (0);
4731 
4732 	i = 0;
4733 retrylookup:
4734 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4735 	if (m == NULL || m->pindex != pindex + i) {
4736 		mpred = m;
4737 		m = NULL;
4738 	} else
4739 		mpred = TAILQ_PREV(m, pglist, listq);
4740 	for (; i < count; i++) {
4741 		if (m != NULL) {
4742 			if (!vm_page_tryacquire(m, allocflags)) {
4743 				if (vm_page_grab_sleep(object, m, pindex,
4744 				    "grbmaw", allocflags, true))
4745 					goto retrylookup;
4746 				break;
4747 			}
4748 		} else {
4749 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4750 				break;
4751 			m = vm_page_alloc_after(object, pindex + i,
4752 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4753 			if (m == NULL) {
4754 				if ((allocflags & (VM_ALLOC_NOWAIT |
4755 				    VM_ALLOC_WAITFAIL)) != 0)
4756 					break;
4757 				goto retrylookup;
4758 			}
4759 		}
4760 		if (vm_page_none_valid(m) &&
4761 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4762 			if ((m->flags & PG_ZERO) == 0)
4763 				pmap_zero_page(m);
4764 			vm_page_valid(m);
4765 		}
4766 		vm_page_grab_release(m, allocflags);
4767 		ma[i] = mpred = m;
4768 		m = vm_page_next(m);
4769 	}
4770 	return (i);
4771 }
4772 
4773 /*
4774  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
4775  * and will fall back to the locked variant to handle allocation.
4776  */
4777 int
4778 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4779     int allocflags, vm_page_t *ma, int count)
4780 {
4781 	vm_page_t m, pred;
4782 	int flags;
4783 	int i;
4784 
4785 	vm_page_grab_check(allocflags);
4786 
4787 	/*
4788 	 * Modify flags for lockless acquire to hold the page until we
4789 	 * set it valid if necessary.
4790 	 */
4791 	flags = allocflags & ~VM_ALLOC_NOBUSY;
4792 	pred = NULL;
4793 	for (i = 0; i < count; i++, pindex++) {
4794 		if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4795 			return (i);
4796 		if (m == NULL)
4797 			break;
4798 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4799 			if ((m->flags & PG_ZERO) == 0)
4800 				pmap_zero_page(m);
4801 			vm_page_valid(m);
4802 		}
4803 		/* m will still be wired or busy according to flags. */
4804 		vm_page_grab_release(m, allocflags);
4805 		pred = ma[i] = m;
4806 	}
4807 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4808 		return (i);
4809 	count -= i;
4810 	VM_OBJECT_WLOCK(object);
4811 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4812 	VM_OBJECT_WUNLOCK(object);
4813 
4814 	return (i);
4815 }
4816 
4817 /*
4818  * Mapping function for valid or dirty bits in a page.
4819  *
4820  * Inputs are required to range within a page.
4821  */
4822 vm_page_bits_t
4823 vm_page_bits(int base, int size)
4824 {
4825 	int first_bit;
4826 	int last_bit;
4827 
4828 	KASSERT(
4829 	    base + size <= PAGE_SIZE,
4830 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4831 	);
4832 
4833 	if (size == 0)		/* handle degenerate case */
4834 		return (0);
4835 
4836 	first_bit = base >> DEV_BSHIFT;
4837 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4838 
4839 	return (((vm_page_bits_t)2 << last_bit) -
4840 	    ((vm_page_bits_t)1 << first_bit));
4841 }
4842 
4843 void
4844 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4845 {
4846 
4847 #if PAGE_SIZE == 32768
4848 	atomic_set_64((uint64_t *)bits, set);
4849 #elif PAGE_SIZE == 16384
4850 	atomic_set_32((uint32_t *)bits, set);
4851 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4852 	atomic_set_16((uint16_t *)bits, set);
4853 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4854 	atomic_set_8((uint8_t *)bits, set);
4855 #else		/* PAGE_SIZE <= 8192 */
4856 	uintptr_t addr;
4857 	int shift;
4858 
4859 	addr = (uintptr_t)bits;
4860 	/*
4861 	 * Use a trick to perform a 32-bit atomic on the
4862 	 * containing aligned word, to not depend on the existence
4863 	 * of atomic_{set, clear}_{8, 16}.
4864 	 */
4865 	shift = addr & (sizeof(uint32_t) - 1);
4866 #if BYTE_ORDER == BIG_ENDIAN
4867 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4868 #else
4869 	shift *= NBBY;
4870 #endif
4871 	addr &= ~(sizeof(uint32_t) - 1);
4872 	atomic_set_32((uint32_t *)addr, set << shift);
4873 #endif		/* PAGE_SIZE */
4874 }
4875 
4876 static inline void
4877 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4878 {
4879 
4880 #if PAGE_SIZE == 32768
4881 	atomic_clear_64((uint64_t *)bits, clear);
4882 #elif PAGE_SIZE == 16384
4883 	atomic_clear_32((uint32_t *)bits, clear);
4884 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4885 	atomic_clear_16((uint16_t *)bits, clear);
4886 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4887 	atomic_clear_8((uint8_t *)bits, clear);
4888 #else		/* PAGE_SIZE <= 8192 */
4889 	uintptr_t addr;
4890 	int shift;
4891 
4892 	addr = (uintptr_t)bits;
4893 	/*
4894 	 * Use a trick to perform a 32-bit atomic on the
4895 	 * containing aligned word, to not depend on the existence
4896 	 * of atomic_{set, clear}_{8, 16}.
4897 	 */
4898 	shift = addr & (sizeof(uint32_t) - 1);
4899 #if BYTE_ORDER == BIG_ENDIAN
4900 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4901 #else
4902 	shift *= NBBY;
4903 #endif
4904 	addr &= ~(sizeof(uint32_t) - 1);
4905 	atomic_clear_32((uint32_t *)addr, clear << shift);
4906 #endif		/* PAGE_SIZE */
4907 }
4908 
4909 static inline vm_page_bits_t
4910 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4911 {
4912 #if PAGE_SIZE == 32768
4913 	uint64_t old;
4914 
4915 	old = *bits;
4916 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4917 	return (old);
4918 #elif PAGE_SIZE == 16384
4919 	uint32_t old;
4920 
4921 	old = *bits;
4922 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4923 	return (old);
4924 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4925 	uint16_t old;
4926 
4927 	old = *bits;
4928 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4929 	return (old);
4930 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4931 	uint8_t old;
4932 
4933 	old = *bits;
4934 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4935 	return (old);
4936 #else		/* PAGE_SIZE <= 4096*/
4937 	uintptr_t addr;
4938 	uint32_t old, new, mask;
4939 	int shift;
4940 
4941 	addr = (uintptr_t)bits;
4942 	/*
4943 	 * Use a trick to perform a 32-bit atomic on the
4944 	 * containing aligned word, to not depend on the existence
4945 	 * of atomic_{set, swap, clear}_{8, 16}.
4946 	 */
4947 	shift = addr & (sizeof(uint32_t) - 1);
4948 #if BYTE_ORDER == BIG_ENDIAN
4949 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4950 #else
4951 	shift *= NBBY;
4952 #endif
4953 	addr &= ~(sizeof(uint32_t) - 1);
4954 	mask = VM_PAGE_BITS_ALL << shift;
4955 
4956 	old = *bits;
4957 	do {
4958 		new = old & ~mask;
4959 		new |= newbits << shift;
4960 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4961 	return (old >> shift);
4962 #endif		/* PAGE_SIZE */
4963 }
4964 
4965 /*
4966  *	vm_page_set_valid_range:
4967  *
4968  *	Sets portions of a page valid.  The arguments are expected
4969  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4970  *	of any partial chunks touched by the range.  The invalid portion of
4971  *	such chunks will be zeroed.
4972  *
4973  *	(base + size) must be less then or equal to PAGE_SIZE.
4974  */
4975 void
4976 vm_page_set_valid_range(vm_page_t m, int base, int size)
4977 {
4978 	int endoff, frag;
4979 	vm_page_bits_t pagebits;
4980 
4981 	vm_page_assert_busied(m);
4982 	if (size == 0)	/* handle degenerate case */
4983 		return;
4984 
4985 	/*
4986 	 * If the base is not DEV_BSIZE aligned and the valid
4987 	 * bit is clear, we have to zero out a portion of the
4988 	 * first block.
4989 	 */
4990 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4991 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4992 		pmap_zero_page_area(m, frag, base - frag);
4993 
4994 	/*
4995 	 * If the ending offset is not DEV_BSIZE aligned and the
4996 	 * valid bit is clear, we have to zero out a portion of
4997 	 * the last block.
4998 	 */
4999 	endoff = base + size;
5000 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5001 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5002 		pmap_zero_page_area(m, endoff,
5003 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5004 
5005 	/*
5006 	 * Assert that no previously invalid block that is now being validated
5007 	 * is already dirty.
5008 	 */
5009 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5010 	    ("vm_page_set_valid_range: page %p is dirty", m));
5011 
5012 	/*
5013 	 * Set valid bits inclusive of any overlap.
5014 	 */
5015 	pagebits = vm_page_bits(base, size);
5016 	if (vm_page_xbusied(m))
5017 		m->valid |= pagebits;
5018 	else
5019 		vm_page_bits_set(m, &m->valid, pagebits);
5020 }
5021 
5022 /*
5023  * Set the page dirty bits and free the invalid swap space if
5024  * present.  Returns the previous dirty bits.
5025  */
5026 vm_page_bits_t
5027 vm_page_set_dirty(vm_page_t m)
5028 {
5029 	vm_page_bits_t old;
5030 
5031 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5032 
5033 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5034 		old = m->dirty;
5035 		m->dirty = VM_PAGE_BITS_ALL;
5036 	} else
5037 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5038 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5039 		vm_pager_page_unswapped(m);
5040 
5041 	return (old);
5042 }
5043 
5044 /*
5045  * Clear the given bits from the specified page's dirty field.
5046  */
5047 static __inline void
5048 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5049 {
5050 
5051 	vm_page_assert_busied(m);
5052 
5053 	/*
5054 	 * If the page is xbusied and not write mapped we are the
5055 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5056 	 * layer can call vm_page_dirty() without holding a distinguished
5057 	 * lock.  The combination of page busy and atomic operations
5058 	 * suffice to guarantee consistency of the page dirty field.
5059 	 */
5060 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5061 		m->dirty &= ~pagebits;
5062 	else
5063 		vm_page_bits_clear(m, &m->dirty, pagebits);
5064 }
5065 
5066 /*
5067  *	vm_page_set_validclean:
5068  *
5069  *	Sets portions of a page valid and clean.  The arguments are expected
5070  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5071  *	of any partial chunks touched by the range.  The invalid portion of
5072  *	such chunks will be zero'd.
5073  *
5074  *	(base + size) must be less then or equal to PAGE_SIZE.
5075  */
5076 void
5077 vm_page_set_validclean(vm_page_t m, int base, int size)
5078 {
5079 	vm_page_bits_t oldvalid, pagebits;
5080 	int endoff, frag;
5081 
5082 	vm_page_assert_busied(m);
5083 	if (size == 0)	/* handle degenerate case */
5084 		return;
5085 
5086 	/*
5087 	 * If the base is not DEV_BSIZE aligned and the valid
5088 	 * bit is clear, we have to zero out a portion of the
5089 	 * first block.
5090 	 */
5091 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5092 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5093 		pmap_zero_page_area(m, frag, base - frag);
5094 
5095 	/*
5096 	 * If the ending offset is not DEV_BSIZE aligned and the
5097 	 * valid bit is clear, we have to zero out a portion of
5098 	 * the last block.
5099 	 */
5100 	endoff = base + size;
5101 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5102 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5103 		pmap_zero_page_area(m, endoff,
5104 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5105 
5106 	/*
5107 	 * Set valid, clear dirty bits.  If validating the entire
5108 	 * page we can safely clear the pmap modify bit.  We also
5109 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5110 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5111 	 * be set again.
5112 	 *
5113 	 * We set valid bits inclusive of any overlap, but we can only
5114 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5115 	 * the range.
5116 	 */
5117 	oldvalid = m->valid;
5118 	pagebits = vm_page_bits(base, size);
5119 	if (vm_page_xbusied(m))
5120 		m->valid |= pagebits;
5121 	else
5122 		vm_page_bits_set(m, &m->valid, pagebits);
5123 #if 0	/* NOT YET */
5124 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5125 		frag = DEV_BSIZE - frag;
5126 		base += frag;
5127 		size -= frag;
5128 		if (size < 0)
5129 			size = 0;
5130 	}
5131 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5132 #endif
5133 	if (base == 0 && size == PAGE_SIZE) {
5134 		/*
5135 		 * The page can only be modified within the pmap if it is
5136 		 * mapped, and it can only be mapped if it was previously
5137 		 * fully valid.
5138 		 */
5139 		if (oldvalid == VM_PAGE_BITS_ALL)
5140 			/*
5141 			 * Perform the pmap_clear_modify() first.  Otherwise,
5142 			 * a concurrent pmap operation, such as
5143 			 * pmap_protect(), could clear a modification in the
5144 			 * pmap and set the dirty field on the page before
5145 			 * pmap_clear_modify() had begun and after the dirty
5146 			 * field was cleared here.
5147 			 */
5148 			pmap_clear_modify(m);
5149 		m->dirty = 0;
5150 		vm_page_aflag_clear(m, PGA_NOSYNC);
5151 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5152 		m->dirty &= ~pagebits;
5153 	else
5154 		vm_page_clear_dirty_mask(m, pagebits);
5155 }
5156 
5157 void
5158 vm_page_clear_dirty(vm_page_t m, int base, int size)
5159 {
5160 
5161 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5162 }
5163 
5164 /*
5165  *	vm_page_set_invalid:
5166  *
5167  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5168  *	valid and dirty bits for the effected areas are cleared.
5169  */
5170 void
5171 vm_page_set_invalid(vm_page_t m, int base, int size)
5172 {
5173 	vm_page_bits_t bits;
5174 	vm_object_t object;
5175 
5176 	/*
5177 	 * The object lock is required so that pages can't be mapped
5178 	 * read-only while we're in the process of invalidating them.
5179 	 */
5180 	object = m->object;
5181 	VM_OBJECT_ASSERT_WLOCKED(object);
5182 	vm_page_assert_busied(m);
5183 
5184 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5185 	    size >= object->un_pager.vnp.vnp_size)
5186 		bits = VM_PAGE_BITS_ALL;
5187 	else
5188 		bits = vm_page_bits(base, size);
5189 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5190 		pmap_remove_all(m);
5191 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5192 	    !pmap_page_is_mapped(m),
5193 	    ("vm_page_set_invalid: page %p is mapped", m));
5194 	if (vm_page_xbusied(m)) {
5195 		m->valid &= ~bits;
5196 		m->dirty &= ~bits;
5197 	} else {
5198 		vm_page_bits_clear(m, &m->valid, bits);
5199 		vm_page_bits_clear(m, &m->dirty, bits);
5200 	}
5201 }
5202 
5203 /*
5204  *	vm_page_invalid:
5205  *
5206  *	Invalidates the entire page.  The page must be busy, unmapped, and
5207  *	the enclosing object must be locked.  The object locks protects
5208  *	against concurrent read-only pmap enter which is done without
5209  *	busy.
5210  */
5211 void
5212 vm_page_invalid(vm_page_t m)
5213 {
5214 
5215 	vm_page_assert_busied(m);
5216 	VM_OBJECT_ASSERT_LOCKED(m->object);
5217 	MPASS(!pmap_page_is_mapped(m));
5218 
5219 	if (vm_page_xbusied(m))
5220 		m->valid = 0;
5221 	else
5222 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5223 }
5224 
5225 /*
5226  * vm_page_zero_invalid()
5227  *
5228  *	The kernel assumes that the invalid portions of a page contain
5229  *	garbage, but such pages can be mapped into memory by user code.
5230  *	When this occurs, we must zero out the non-valid portions of the
5231  *	page so user code sees what it expects.
5232  *
5233  *	Pages are most often semi-valid when the end of a file is mapped
5234  *	into memory and the file's size is not page aligned.
5235  */
5236 void
5237 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5238 {
5239 	int b;
5240 	int i;
5241 
5242 	/*
5243 	 * Scan the valid bits looking for invalid sections that
5244 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5245 	 * valid bit may be set ) have already been zeroed by
5246 	 * vm_page_set_validclean().
5247 	 */
5248 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5249 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5250 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5251 			if (i > b) {
5252 				pmap_zero_page_area(m,
5253 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5254 			}
5255 			b = i + 1;
5256 		}
5257 	}
5258 
5259 	/*
5260 	 * setvalid is TRUE when we can safely set the zero'd areas
5261 	 * as being valid.  We can do this if there are no cache consistancy
5262 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5263 	 */
5264 	if (setvalid)
5265 		vm_page_valid(m);
5266 }
5267 
5268 /*
5269  *	vm_page_is_valid:
5270  *
5271  *	Is (partial) page valid?  Note that the case where size == 0
5272  *	will return FALSE in the degenerate case where the page is
5273  *	entirely invalid, and TRUE otherwise.
5274  *
5275  *	Some callers envoke this routine without the busy lock held and
5276  *	handle races via higher level locks.  Typical callers should
5277  *	hold a busy lock to prevent invalidation.
5278  */
5279 int
5280 vm_page_is_valid(vm_page_t m, int base, int size)
5281 {
5282 	vm_page_bits_t bits;
5283 
5284 	bits = vm_page_bits(base, size);
5285 	return (m->valid != 0 && (m->valid & bits) == bits);
5286 }
5287 
5288 /*
5289  * Returns true if all of the specified predicates are true for the entire
5290  * (super)page and false otherwise.
5291  */
5292 bool
5293 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5294 {
5295 	vm_object_t object;
5296 	int i, npages;
5297 
5298 	object = m->object;
5299 	if (skip_m != NULL && skip_m->object != object)
5300 		return (false);
5301 	VM_OBJECT_ASSERT_LOCKED(object);
5302 	npages = atop(pagesizes[m->psind]);
5303 
5304 	/*
5305 	 * The physically contiguous pages that make up a superpage, i.e., a
5306 	 * page with a page size index ("psind") greater than zero, will
5307 	 * occupy adjacent entries in vm_page_array[].
5308 	 */
5309 	for (i = 0; i < npages; i++) {
5310 		/* Always test object consistency, including "skip_m". */
5311 		if (m[i].object != object)
5312 			return (false);
5313 		if (&m[i] == skip_m)
5314 			continue;
5315 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5316 			return (false);
5317 		if ((flags & PS_ALL_DIRTY) != 0) {
5318 			/*
5319 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5320 			 * might stop this case from spuriously returning
5321 			 * "false".  However, that would require a write lock
5322 			 * on the object containing "m[i]".
5323 			 */
5324 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5325 				return (false);
5326 		}
5327 		if ((flags & PS_ALL_VALID) != 0 &&
5328 		    m[i].valid != VM_PAGE_BITS_ALL)
5329 			return (false);
5330 	}
5331 	return (true);
5332 }
5333 
5334 /*
5335  * Set the page's dirty bits if the page is modified.
5336  */
5337 void
5338 vm_page_test_dirty(vm_page_t m)
5339 {
5340 
5341 	vm_page_assert_busied(m);
5342 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5343 		vm_page_dirty(m);
5344 }
5345 
5346 void
5347 vm_page_valid(vm_page_t m)
5348 {
5349 
5350 	vm_page_assert_busied(m);
5351 	if (vm_page_xbusied(m))
5352 		m->valid = VM_PAGE_BITS_ALL;
5353 	else
5354 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5355 }
5356 
5357 void
5358 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5359 {
5360 
5361 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5362 }
5363 
5364 void
5365 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5366 {
5367 
5368 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5369 }
5370 
5371 int
5372 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5373 {
5374 
5375 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5376 }
5377 
5378 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5379 void
5380 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5381 {
5382 
5383 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5384 }
5385 
5386 void
5387 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5388 {
5389 
5390 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5391 }
5392 #endif
5393 
5394 #ifdef INVARIANTS
5395 void
5396 vm_page_object_busy_assert(vm_page_t m)
5397 {
5398 
5399 	/*
5400 	 * Certain of the page's fields may only be modified by the
5401 	 * holder of a page or object busy.
5402 	 */
5403 	if (m->object != NULL && !vm_page_busied(m))
5404 		VM_OBJECT_ASSERT_BUSY(m->object);
5405 }
5406 
5407 void
5408 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5409 {
5410 
5411 	if ((bits & PGA_WRITEABLE) == 0)
5412 		return;
5413 
5414 	/*
5415 	 * The PGA_WRITEABLE flag can only be set if the page is
5416 	 * managed, is exclusively busied or the object is locked.
5417 	 * Currently, this flag is only set by pmap_enter().
5418 	 */
5419 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5420 	    ("PGA_WRITEABLE on unmanaged page"));
5421 	if (!vm_page_xbusied(m))
5422 		VM_OBJECT_ASSERT_BUSY(m->object);
5423 }
5424 #endif
5425 
5426 #include "opt_ddb.h"
5427 #ifdef DDB
5428 #include <sys/kernel.h>
5429 
5430 #include <ddb/ddb.h>
5431 
5432 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5433 {
5434 
5435 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5436 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5437 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5438 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5439 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5440 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5441 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5442 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5443 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5444 }
5445 
5446 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5447 {
5448 	int dom;
5449 
5450 	db_printf("pq_free %d\n", vm_free_count());
5451 	for (dom = 0; dom < vm_ndomains; dom++) {
5452 		db_printf(
5453     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5454 		    dom,
5455 		    vm_dom[dom].vmd_page_count,
5456 		    vm_dom[dom].vmd_free_count,
5457 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5458 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5459 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5460 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5461 	}
5462 }
5463 
5464 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5465 {
5466 	vm_page_t m;
5467 	boolean_t phys, virt;
5468 
5469 	if (!have_addr) {
5470 		db_printf("show pginfo addr\n");
5471 		return;
5472 	}
5473 
5474 	phys = strchr(modif, 'p') != NULL;
5475 	virt = strchr(modif, 'v') != NULL;
5476 	if (virt)
5477 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5478 	else if (phys)
5479 		m = PHYS_TO_VM_PAGE(addr);
5480 	else
5481 		m = (vm_page_t)addr;
5482 	db_printf(
5483     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5484     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5485 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5486 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5487 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5488 }
5489 #endif /* DDB */
5490