xref: /freebsd/sys/vm/vm_page.c (revision 5036d9652a5701d00e9e40ea942c278e9f77d33d)
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 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Resident memory management module.
65  */
66 
67 #include <sys/cdefs.h>
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/counter.h>
73 #include <sys/domainset.h>
74 #include <sys/kernel.h>
75 #include <sys/limits.h>
76 #include <sys/linker.h>
77 #include <sys/lock.h>
78 #include <sys/malloc.h>
79 #include <sys/mman.h>
80 #include <sys/msgbuf.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/rwlock.h>
84 #include <sys/sleepqueue.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/smp.h>
88 #include <sys/sysctl.h>
89 #include <sys/vmmeter.h>
90 #include <sys/vnode.h>
91 
92 #include <vm/vm.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_param.h>
95 #include <vm/vm_domainset.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_dumpset.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 struct vm_domain vm_dom[MAXMEMDOM];
114 
115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
116 
117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
118 
119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
120 /* The following fields are protected by the domainset lock. */
121 domainset_t __exclusive_cache_line vm_min_domains;
122 domainset_t __exclusive_cache_line vm_severe_domains;
123 static int vm_min_waiters;
124 static int vm_severe_waiters;
125 static int vm_pageproc_waiters;
126 
127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
128     "VM page statistics");
129 
130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
132     CTLFLAG_RD, &pqstate_commit_retries,
133     "Number of failed per-page atomic queue state updates");
134 
135 static COUNTER_U64_DEFINE_EARLY(queue_ops);
136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
137     CTLFLAG_RD, &queue_ops,
138     "Number of batched queue operations");
139 
140 static COUNTER_U64_DEFINE_EARLY(queue_nops);
141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
142     CTLFLAG_RD, &queue_nops,
143     "Number of batched queue operations with no effects");
144 
145 /*
146  * bogus page -- for I/O to/from partially complete buffers,
147  * or for paging into sparsely invalid regions.
148  */
149 vm_page_t bogus_page;
150 
151 vm_page_t vm_page_array;
152 long vm_page_array_size;
153 long first_page;
154 
155 struct bitset *vm_page_dump;
156 long vm_page_dump_pages;
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 vm_page_t vm_page_alloc_nofree_domain(int domain, int req);
167 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
168     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
169 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
170 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
171 static bool vm_page_free_prep(vm_page_t m);
172 static void vm_page_free_toq(vm_page_t m);
173 static void vm_page_init(void *dummy);
174 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
175     vm_pindex_t pindex, vm_page_t mpred);
176 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
177     vm_page_t mpred);
178 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
179     const uint16_t nflag);
180 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
181     vm_page_t m_run, vm_paddr_t high);
182 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
183 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
184     int req);
185 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
186     int flags);
187 static void vm_page_zone_release(void *arg, void **store, int cnt);
188 
189 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
190 
191 static void
192 vm_page_init(void *dummy)
193 {
194 
195 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
196 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
197 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE);
198 }
199 
200 static int pgcache_zone_max_pcpu;
201 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
202     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
203     "Per-CPU page cache size");
204 
205 /*
206  * The cache page zone is initialized later since we need to be able to allocate
207  * pages before UMA is fully initialized.
208  */
209 static void
210 vm_page_init_cache_zones(void *dummy __unused)
211 {
212 	struct vm_domain *vmd;
213 	struct vm_pgcache *pgcache;
214 	int cache, domain, maxcache, pool;
215 
216 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
217 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
218 	for (domain = 0; domain < vm_ndomains; domain++) {
219 		vmd = VM_DOMAIN(domain);
220 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
221 			pgcache = &vmd->vmd_pgcache[pool];
222 			pgcache->domain = domain;
223 			pgcache->pool = pool;
224 			pgcache->zone = uma_zcache_create("vm pgcache",
225 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
226 			    vm_page_zone_import, vm_page_zone_release, pgcache,
227 			    UMA_ZONE_VM);
228 
229 			/*
230 			 * Limit each pool's zone to 0.1% of the pages in the
231 			 * domain.
232 			 */
233 			cache = maxcache != 0 ? maxcache :
234 			    vmd->vmd_page_count / 1000;
235 			uma_zone_set_maxcache(pgcache->zone, cache);
236 		}
237 	}
238 }
239 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
240 
241 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
242 #if PAGE_SIZE == 32768
243 #ifdef CTASSERT
244 CTASSERT(sizeof(u_long) >= 8);
245 #endif
246 #endif
247 
248 /*
249  *	vm_set_page_size:
250  *
251  *	Sets the page size, perhaps based upon the memory
252  *	size.  Must be called before any use of page-size
253  *	dependent functions.
254  */
255 void
256 vm_set_page_size(void)
257 {
258 	if (vm_cnt.v_page_size == 0)
259 		vm_cnt.v_page_size = PAGE_SIZE;
260 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
261 		panic("vm_set_page_size: page size not a power of two");
262 }
263 
264 /*
265  *	vm_page_blacklist_next:
266  *
267  *	Find the next entry in the provided string of blacklist
268  *	addresses.  Entries are separated by space, comma, or newline.
269  *	If an invalid integer is encountered then the rest of the
270  *	string is skipped.  Updates the list pointer to the next
271  *	character, or NULL if the string is exhausted or invalid.
272  */
273 static vm_paddr_t
274 vm_page_blacklist_next(char **list, char *end)
275 {
276 	vm_paddr_t bad;
277 	char *cp, *pos;
278 
279 	if (list == NULL || *list == NULL)
280 		return (0);
281 	if (**list =='\0') {
282 		*list = NULL;
283 		return (0);
284 	}
285 
286 	/*
287 	 * If there's no end pointer then the buffer is coming from
288 	 * the kenv and we know it's null-terminated.
289 	 */
290 	if (end == NULL)
291 		end = *list + strlen(*list);
292 
293 	/* Ensure that strtoq() won't walk off the end */
294 	if (*end != '\0') {
295 		if (*end == '\n' || *end == ' ' || *end  == ',')
296 			*end = '\0';
297 		else {
298 			printf("Blacklist not terminated, skipping\n");
299 			*list = NULL;
300 			return (0);
301 		}
302 	}
303 
304 	for (pos = *list; *pos != '\0'; pos = cp) {
305 		bad = strtoq(pos, &cp, 0);
306 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
307 			if (bad == 0) {
308 				if (++cp < end)
309 					continue;
310 				else
311 					break;
312 			}
313 		} else
314 			break;
315 		if (*cp == '\0' || ++cp >= end)
316 			*list = NULL;
317 		else
318 			*list = cp;
319 		return (trunc_page(bad));
320 	}
321 	printf("Garbage in RAM blacklist, skipping\n");
322 	*list = NULL;
323 	return (0);
324 }
325 
326 bool
327 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
328 {
329 	struct vm_domain *vmd;
330 	vm_page_t m;
331 	bool found;
332 
333 	m = vm_phys_paddr_to_vm_page(pa);
334 	if (m == NULL)
335 		return (true); /* page does not exist, no failure */
336 
337 	vmd = VM_DOMAIN(vm_phys_domain(pa));
338 	vm_domain_free_lock(vmd);
339 	found = vm_phys_unfree_page(pa);
340 	vm_domain_free_unlock(vmd);
341 	if (found) {
342 		vm_domain_freecnt_inc(vmd, -1);
343 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
344 		if (verbose)
345 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
346 	}
347 	return (found);
348 }
349 
350 /*
351  *	vm_page_blacklist_check:
352  *
353  *	Iterate through the provided string of blacklist addresses, pulling
354  *	each entry out of the physical allocator free list and putting it
355  *	onto a list for reporting via the vm.page_blacklist sysctl.
356  */
357 static void
358 vm_page_blacklist_check(char *list, char *end)
359 {
360 	vm_paddr_t pa;
361 	char *next;
362 
363 	next = list;
364 	while (next != NULL) {
365 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
366 			continue;
367 		vm_page_blacklist_add(pa, bootverbose);
368 	}
369 }
370 
371 /*
372  *	vm_page_blacklist_load:
373  *
374  *	Search for a special module named "ram_blacklist".  It'll be a
375  *	plain text file provided by the user via the loader directive
376  *	of the same name.
377  */
378 static void
379 vm_page_blacklist_load(char **list, char **end)
380 {
381 	void *mod;
382 	u_char *ptr;
383 	u_int len;
384 
385 	mod = NULL;
386 	ptr = NULL;
387 
388 	mod = preload_search_by_type("ram_blacklist");
389 	if (mod != NULL) {
390 		ptr = preload_fetch_addr(mod);
391 		len = preload_fetch_size(mod);
392         }
393 	*list = ptr;
394 	if (ptr != NULL)
395 		*end = ptr + len;
396 	else
397 		*end = NULL;
398 	return;
399 }
400 
401 static int
402 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
403 {
404 	vm_page_t m;
405 	struct sbuf sbuf;
406 	int error, first;
407 
408 	first = 1;
409 	error = sysctl_wire_old_buffer(req, 0);
410 	if (error != 0)
411 		return (error);
412 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
413 	TAILQ_FOREACH(m, &blacklist_head, listq) {
414 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
415 		    (uintmax_t)m->phys_addr);
416 		first = 0;
417 	}
418 	error = sbuf_finish(&sbuf);
419 	sbuf_delete(&sbuf);
420 	return (error);
421 }
422 
423 /*
424  * Initialize a dummy page for use in scans of the specified paging queue.
425  * In principle, this function only needs to set the flag PG_MARKER.
426  * Nonetheless, it write busies the page as a safety precaution.
427  */
428 void
429 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
430 {
431 
432 	bzero(marker, sizeof(*marker));
433 	marker->flags = PG_MARKER;
434 	marker->a.flags = aflags;
435 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
436 	marker->a.queue = queue;
437 }
438 
439 static void
440 vm_page_domain_init(int domain)
441 {
442 	struct vm_domain *vmd;
443 	struct vm_pagequeue *pq;
444 	int i;
445 
446 	vmd = VM_DOMAIN(domain);
447 	bzero(vmd, sizeof(*vmd));
448 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
449 	    "vm inactive pagequeue";
450 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
451 	    "vm active pagequeue";
452 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
453 	    "vm laundry pagequeue";
454 	*__DECONST(const char **,
455 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
456 	    "vm unswappable pagequeue";
457 	vmd->vmd_domain = domain;
458 	vmd->vmd_page_count = 0;
459 	vmd->vmd_free_count = 0;
460 	vmd->vmd_segs = 0;
461 	vmd->vmd_oom = FALSE;
462 	for (i = 0; i < PQ_COUNT; i++) {
463 		pq = &vmd->vmd_pagequeues[i];
464 		TAILQ_INIT(&pq->pq_pl);
465 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
466 		    MTX_DEF | MTX_DUPOK);
467 		pq->pq_pdpages = 0;
468 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
469 	}
470 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
471 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
472 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
473 
474 	/*
475 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
476 	 * insertions.
477 	 */
478 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
479 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
480 	    &vmd->vmd_inacthead, plinks.q);
481 
482 	/*
483 	 * The clock pages are used to implement active queue scanning without
484 	 * requeues.  Scans start at clock[0], which is advanced after the scan
485 	 * ends.  When the two clock hands meet, they are reset and scanning
486 	 * resumes from the head of the queue.
487 	 */
488 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
489 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
490 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
491 	    &vmd->vmd_clock[0], plinks.q);
492 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
493 	    &vmd->vmd_clock[1], plinks.q);
494 }
495 
496 /*
497  * Initialize a physical page in preparation for adding it to the free
498  * lists.
499  */
500 void
501 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool)
502 {
503 	m->object = NULL;
504 	m->ref_count = 0;
505 	m->busy_lock = VPB_FREED;
506 	m->flags = m->a.flags = 0;
507 	m->phys_addr = pa;
508 	m->a.queue = PQ_NONE;
509 	m->psind = 0;
510 	m->segind = segind;
511 	m->order = VM_NFREEORDER;
512 	m->pool = pool;
513 	m->valid = m->dirty = 0;
514 	pmap_page_init(m);
515 }
516 
517 #ifndef PMAP_HAS_PAGE_ARRAY
518 static vm_paddr_t
519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
520 {
521 	vm_paddr_t new_end;
522 
523 	/*
524 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
525 	 * However, because this page is allocated from KVM, out-of-bounds
526 	 * accesses using the direct map will not be trapped.
527 	 */
528 	*vaddr += PAGE_SIZE;
529 
530 	/*
531 	 * Allocate physical memory for the page structures, and map it.
532 	 */
533 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
534 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
535 	    VM_PROT_READ | VM_PROT_WRITE);
536 	vm_page_array_size = page_range;
537 
538 	return (new_end);
539 }
540 #endif
541 
542 /*
543  *	vm_page_startup:
544  *
545  *	Initializes the resident memory module.  Allocates physical memory for
546  *	bootstrapping UMA and some data structures that are used to manage
547  *	physical pages.  Initializes these structures, and populates the free
548  *	page queues.
549  */
550 vm_offset_t
551 vm_page_startup(vm_offset_t vaddr)
552 {
553 	struct vm_phys_seg *seg;
554 	struct vm_domain *vmd;
555 	vm_page_t m;
556 	char *list, *listend;
557 	vm_paddr_t end, high_avail, low_avail, new_end, size;
558 	vm_paddr_t page_range __unused;
559 	vm_paddr_t last_pa, pa, startp, endp;
560 	u_long pagecount;
561 #if MINIDUMP_PAGE_TRACKING
562 	u_long vm_page_dump_size;
563 #endif
564 	int biggestone, i, segind;
565 #ifdef WITNESS
566 	vm_offset_t mapped;
567 	int witness_size;
568 #endif
569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
570 	long ii;
571 #endif
572 #ifdef VM_FREEPOOL_LAZYINIT
573 	int lazyinit;
574 #endif
575 
576 	vaddr = round_page(vaddr);
577 
578 	vm_phys_early_startup();
579 	biggestone = vm_phys_avail_largest();
580 	end = phys_avail[biggestone+1];
581 
582 	/*
583 	 * Initialize the page and queue locks.
584 	 */
585 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
586 	for (i = 0; i < PA_LOCK_COUNT; i++)
587 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
588 	for (i = 0; i < vm_ndomains; i++)
589 		vm_page_domain_init(i);
590 
591 	new_end = end;
592 #ifdef WITNESS
593 	witness_size = round_page(witness_startup_count());
594 	new_end -= witness_size;
595 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
596 	    VM_PROT_READ | VM_PROT_WRITE);
597 	bzero((void *)mapped, witness_size);
598 	witness_startup((void *)mapped);
599 #endif
600 
601 #if MINIDUMP_PAGE_TRACKING
602 	/*
603 	 * Allocate a bitmap to indicate that a random physical page
604 	 * needs to be included in a minidump.
605 	 *
606 	 * The amd64 port needs this to indicate which direct map pages
607 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
608 	 *
609 	 * However, i386 still needs this workspace internally within the
610 	 * minidump code.  In theory, they are not needed on i386, but are
611 	 * included should the sf_buf code decide to use them.
612 	 */
613 	last_pa = 0;
614 	vm_page_dump_pages = 0;
615 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
616 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
617 		    dump_avail[i] / PAGE_SIZE;
618 		if (dump_avail[i + 1] > last_pa)
619 			last_pa = dump_avail[i + 1];
620 	}
621 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
622 	new_end -= vm_page_dump_size;
623 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
624 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
625 	bzero((void *)vm_page_dump, vm_page_dump_size);
626 #if MINIDUMP_STARTUP_PAGE_TRACKING
627 	/*
628 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
629 	 * in a crash dump.  When pmap_map() uses the direct map, they are
630 	 * not automatically included.
631 	 */
632 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
633 		dump_add_page(pa);
634 #endif
635 #else
636 	(void)last_pa;
637 #endif
638 	phys_avail[biggestone + 1] = new_end;
639 #ifdef __amd64__
640 	/*
641 	 * Request that the physical pages underlying the message buffer be
642 	 * included in a crash dump.  Since the message buffer is accessed
643 	 * through the direct map, they are not automatically included.
644 	 */
645 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
646 	last_pa = pa + round_page(msgbufsize);
647 	while (pa < last_pa) {
648 		dump_add_page(pa);
649 		pa += PAGE_SIZE;
650 	}
651 #endif
652 	/*
653 	 * Compute the number of pages of memory that will be available for
654 	 * use, taking into account the overhead of a page structure per page.
655 	 * In other words, solve
656 	 *	"available physical memory" - round_page(page_range *
657 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
658 	 * for page_range.
659 	 */
660 	low_avail = phys_avail[0];
661 	high_avail = phys_avail[1];
662 	for (i = 0; i < vm_phys_nsegs; i++) {
663 		if (vm_phys_segs[i].start < low_avail)
664 			low_avail = vm_phys_segs[i].start;
665 		if (vm_phys_segs[i].end > high_avail)
666 			high_avail = vm_phys_segs[i].end;
667 	}
668 	/* Skip the first chunk.  It is already accounted for. */
669 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
670 		if (phys_avail[i] < low_avail)
671 			low_avail = phys_avail[i];
672 		if (phys_avail[i + 1] > high_avail)
673 			high_avail = phys_avail[i + 1];
674 	}
675 	first_page = low_avail / PAGE_SIZE;
676 #ifdef VM_PHYSSEG_SPARSE
677 	size = 0;
678 	for (i = 0; i < vm_phys_nsegs; i++)
679 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
680 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
681 		size += phys_avail[i + 1] - phys_avail[i];
682 #elif defined(VM_PHYSSEG_DENSE)
683 	size = high_avail - low_avail;
684 #else
685 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
686 #endif
687 
688 #ifdef PMAP_HAS_PAGE_ARRAY
689 	pmap_page_array_startup(size / PAGE_SIZE);
690 	biggestone = vm_phys_avail_largest();
691 	end = new_end = phys_avail[biggestone + 1];
692 #else
693 #ifdef VM_PHYSSEG_DENSE
694 	/*
695 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
696 	 * the overhead of a page structure per page only if vm_page_array is
697 	 * allocated from the last physical memory chunk.  Otherwise, we must
698 	 * allocate page structures representing the physical memory
699 	 * underlying vm_page_array, even though they will not be used.
700 	 */
701 	if (new_end != high_avail)
702 		page_range = size / PAGE_SIZE;
703 	else
704 #endif
705 	{
706 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
707 
708 		/*
709 		 * If the partial bytes remaining are large enough for
710 		 * a page (PAGE_SIZE) without a corresponding
711 		 * 'struct vm_page', then new_end will contain an
712 		 * extra page after subtracting the length of the VM
713 		 * page array.  Compensate by subtracting an extra
714 		 * page from new_end.
715 		 */
716 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
717 			if (new_end == high_avail)
718 				high_avail -= PAGE_SIZE;
719 			new_end -= PAGE_SIZE;
720 		}
721 	}
722 	end = new_end;
723 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
724 #endif
725 
726 #if VM_NRESERVLEVEL > 0
727 	/*
728 	 * Allocate physical memory for the reservation management system's
729 	 * data structures, and map it.
730 	 */
731 	new_end = vm_reserv_startup(&vaddr, new_end);
732 #endif
733 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
734 	/*
735 	 * Include vm_page_array and vm_reserv_array in a crash dump.
736 	 */
737 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
738 		dump_add_page(pa);
739 #endif
740 	phys_avail[biggestone + 1] = new_end;
741 
742 	/*
743 	 * Add physical memory segments corresponding to the available
744 	 * physical pages.
745 	 */
746 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
747 		if (vm_phys_avail_size(i) != 0)
748 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
749 
750 	/*
751 	 * Initialize the physical memory allocator.
752 	 */
753 	vm_phys_init();
754 
755 #ifdef VM_FREEPOOL_LAZYINIT
756 	lazyinit = 1;
757 	TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit);
758 #endif
759 
760 	/*
761 	 * Initialize the page structures and add every available page to the
762 	 * physical memory allocator's free lists.
763 	 */
764 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
765 	for (ii = 0; ii < vm_page_array_size; ii++) {
766 		m = &vm_page_array[ii];
767 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0,
768 		    VM_FREEPOOL_DEFAULT);
769 		m->flags = PG_FICTITIOUS;
770 	}
771 #endif
772 	vm_cnt.v_page_count = 0;
773 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
774 		seg = &vm_phys_segs[segind];
775 
776 		/*
777 		 * If lazy vm_page initialization is not enabled, simply
778 		 * initialize all of the pages in the segment.  Otherwise, we
779 		 * only initialize:
780 		 * 1. Pages not covered by phys_avail[], since they might be
781 		 *    freed to the allocator at some future point, e.g., by
782 		 *    kmem_bootstrap_free().
783 		 * 2. The first page of each run of free pages handed to the
784 		 *    vm_phys allocator, which in turn defers initialization
785 		 *    of pages until they are needed.
786 		 * This avoids blocking the boot process for long periods, which
787 		 * may be relevant for VMs (which ought to boot as quickly as
788 		 * possible) and/or systems with large amounts of physical
789 		 * memory.
790 		 */
791 #ifdef VM_FREEPOOL_LAZYINIT
792 		if (lazyinit) {
793 			startp = seg->start;
794 			for (i = 0; phys_avail[i + 1] != 0; i += 2) {
795 				if (startp >= seg->end)
796 					break;
797 
798 				if (phys_avail[i + 1] < startp)
799 					continue;
800 				if (phys_avail[i] <= startp) {
801 					startp = phys_avail[i + 1];
802 					continue;
803 				}
804 
805 				m = vm_phys_seg_paddr_to_vm_page(seg, startp);
806 				for (endp = MIN(phys_avail[i], seg->end);
807 				    startp < endp; startp += PAGE_SIZE, m++) {
808 					vm_page_init_page(m, startp, segind,
809 					    VM_FREEPOOL_DEFAULT);
810 				}
811 			}
812 		} else
813 #endif
814 			for (m = seg->first_page, pa = seg->start;
815 			    pa < seg->end; m++, pa += PAGE_SIZE) {
816 				vm_page_init_page(m, pa, segind,
817 				    VM_FREEPOOL_DEFAULT);
818 			}
819 
820 		/*
821 		 * Add the segment's pages that are covered by one of
822 		 * phys_avail's ranges to the free lists.
823 		 */
824 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
825 			if (seg->end <= phys_avail[i] ||
826 			    seg->start >= phys_avail[i + 1])
827 				continue;
828 
829 			startp = MAX(seg->start, phys_avail[i]);
830 			endp = MIN(seg->end, phys_avail[i + 1]);
831 			pagecount = (u_long)atop(endp - startp);
832 			if (pagecount == 0)
833 				continue;
834 
835 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
836 #ifdef VM_FREEPOOL_LAZYINIT
837 			if (lazyinit) {
838 				vm_page_init_page(m, startp, segind,
839 				    VM_FREEPOOL_LAZYINIT);
840 			}
841 #endif
842 			vmd = VM_DOMAIN(seg->domain);
843 			vm_domain_free_lock(vmd);
844 			vm_phys_enqueue_contig(m, pagecount);
845 			vm_domain_free_unlock(vmd);
846 			vm_domain_freecnt_inc(vmd, pagecount);
847 			vm_cnt.v_page_count += (u_int)pagecount;
848 			vmd->vmd_page_count += (u_int)pagecount;
849 			vmd->vmd_segs |= 1UL << segind;
850 		}
851 	}
852 
853 	/*
854 	 * Remove blacklisted pages from the physical memory allocator.
855 	 */
856 	TAILQ_INIT(&blacklist_head);
857 	vm_page_blacklist_load(&list, &listend);
858 	vm_page_blacklist_check(list, listend);
859 
860 	list = kern_getenv("vm.blacklist");
861 	vm_page_blacklist_check(list, NULL);
862 
863 	freeenv(list);
864 #if VM_NRESERVLEVEL > 0
865 	/*
866 	 * Initialize the reservation management system.
867 	 */
868 	vm_reserv_init();
869 #endif
870 
871 	return (vaddr);
872 }
873 
874 void
875 vm_page_reference(vm_page_t m)
876 {
877 
878 	vm_page_aflag_set(m, PGA_REFERENCED);
879 }
880 
881 /*
882  *	vm_page_trybusy
883  *
884  *	Helper routine for grab functions to trylock busy.
885  *
886  *	Returns true on success and false on failure.
887  */
888 static bool
889 vm_page_trybusy(vm_page_t m, int allocflags)
890 {
891 
892 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
893 		return (vm_page_trysbusy(m));
894 	else
895 		return (vm_page_tryxbusy(m));
896 }
897 
898 /*
899  *	vm_page_tryacquire
900  *
901  *	Helper routine for grab functions to trylock busy and wire.
902  *
903  *	Returns true on success and false on failure.
904  */
905 static inline bool
906 vm_page_tryacquire(vm_page_t m, int allocflags)
907 {
908 	bool locked;
909 
910 	locked = vm_page_trybusy(m, allocflags);
911 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
912 		vm_page_wire(m);
913 	return (locked);
914 }
915 
916 /*
917  *	vm_page_busy_acquire:
918  *
919  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
920  *	and drop the object lock if necessary.
921  */
922 bool
923 vm_page_busy_acquire(vm_page_t m, int allocflags)
924 {
925 	vm_object_t obj;
926 	bool locked;
927 
928 	/*
929 	 * The page-specific object must be cached because page
930 	 * identity can change during the sleep, causing the
931 	 * re-lock of a different object.
932 	 * It is assumed that a reference to the object is already
933 	 * held by the callers.
934 	 */
935 	obj = atomic_load_ptr(&m->object);
936 	for (;;) {
937 		if (vm_page_tryacquire(m, allocflags))
938 			return (true);
939 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
940 			return (false);
941 		if (obj != NULL)
942 			locked = VM_OBJECT_WOWNED(obj);
943 		else
944 			locked = false;
945 		MPASS(locked || vm_page_wired(m));
946 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
947 		    locked) && locked)
948 			VM_OBJECT_WLOCK(obj);
949 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
950 			return (false);
951 		KASSERT(m->object == obj || m->object == NULL,
952 		    ("vm_page_busy_acquire: page %p does not belong to %p",
953 		    m, obj));
954 	}
955 }
956 
957 /*
958  *	vm_page_busy_downgrade:
959  *
960  *	Downgrade an exclusive busy page into a single shared busy page.
961  */
962 void
963 vm_page_busy_downgrade(vm_page_t m)
964 {
965 	u_int x;
966 
967 	vm_page_assert_xbusied(m);
968 
969 	x = vm_page_busy_fetch(m);
970 	for (;;) {
971 		if (atomic_fcmpset_rel_int(&m->busy_lock,
972 		    &x, VPB_SHARERS_WORD(1)))
973 			break;
974 	}
975 	if ((x & VPB_BIT_WAITERS) != 0)
976 		wakeup(m);
977 }
978 
979 /*
980  *
981  *	vm_page_busy_tryupgrade:
982  *
983  *	Attempt to upgrade a single shared busy into an exclusive busy.
984  */
985 int
986 vm_page_busy_tryupgrade(vm_page_t m)
987 {
988 	u_int ce, x;
989 
990 	vm_page_assert_sbusied(m);
991 
992 	x = vm_page_busy_fetch(m);
993 	ce = VPB_CURTHREAD_EXCLUSIVE;
994 	for (;;) {
995 		if (VPB_SHARERS(x) > 1)
996 			return (0);
997 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
998 		    ("vm_page_busy_tryupgrade: invalid lock state"));
999 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1000 		    ce | (x & VPB_BIT_WAITERS)))
1001 			continue;
1002 		return (1);
1003 	}
1004 }
1005 
1006 /*
1007  *	vm_page_sbusied:
1008  *
1009  *	Return a positive value if the page is shared busied, 0 otherwise.
1010  */
1011 int
1012 vm_page_sbusied(vm_page_t m)
1013 {
1014 	u_int x;
1015 
1016 	x = vm_page_busy_fetch(m);
1017 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1018 }
1019 
1020 /*
1021  *	vm_page_sunbusy:
1022  *
1023  *	Shared unbusy a page.
1024  */
1025 void
1026 vm_page_sunbusy(vm_page_t m)
1027 {
1028 	u_int x;
1029 
1030 	vm_page_assert_sbusied(m);
1031 
1032 	x = vm_page_busy_fetch(m);
1033 	for (;;) {
1034 		KASSERT(x != VPB_FREED,
1035 		    ("vm_page_sunbusy: Unlocking freed page."));
1036 		if (VPB_SHARERS(x) > 1) {
1037 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1038 			    x - VPB_ONE_SHARER))
1039 				break;
1040 			continue;
1041 		}
1042 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1043 		    ("vm_page_sunbusy: invalid lock state"));
1044 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1045 			continue;
1046 		if ((x & VPB_BIT_WAITERS) == 0)
1047 			break;
1048 		wakeup(m);
1049 		break;
1050 	}
1051 }
1052 
1053 /*
1054  *	vm_page_busy_sleep:
1055  *
1056  *	Sleep if the page is busy, using the page pointer as wchan.
1057  *	This is used to implement the hard-path of the busying mechanism.
1058  *
1059  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1060  *	will not sleep if the page is shared-busy.
1061  *
1062  *	The object lock must be held on entry.
1063  *
1064  *	Returns true if it slept and dropped the object lock, or false
1065  *	if there was no sleep and the lock is still held.
1066  */
1067 bool
1068 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1069 {
1070 	vm_object_t obj;
1071 
1072 	obj = m->object;
1073 	VM_OBJECT_ASSERT_LOCKED(obj);
1074 
1075 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1076 	    true));
1077 }
1078 
1079 /*
1080  *	vm_page_busy_sleep_unlocked:
1081  *
1082  *	Sleep if the page is busy, using the page pointer as wchan.
1083  *	This is used to implement the hard-path of busying mechanism.
1084  *
1085  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1086  *	will not sleep if the page is shared-busy.
1087  *
1088  *	The object lock must not be held on entry.  The operation will
1089  *	return if the page changes identity.
1090  */
1091 void
1092 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1093     const char *wmesg, int allocflags)
1094 {
1095 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1096 
1097 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1098 }
1099 
1100 /*
1101  *	_vm_page_busy_sleep:
1102  *
1103  *	Internal busy sleep function.  Verifies the page identity and
1104  *	lockstate against parameters.  Returns true if it sleeps and
1105  *	false otherwise.
1106  *
1107  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1108  *
1109  *	If locked is true the lock will be dropped for any true returns
1110  *	and held for any false returns.
1111  */
1112 static bool
1113 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1114     const char *wmesg, int allocflags, bool locked)
1115 {
1116 	bool xsleep;
1117 	u_int x;
1118 
1119 	/*
1120 	 * If the object is busy we must wait for that to drain to zero
1121 	 * before trying the page again.
1122 	 */
1123 	if (obj != NULL && vm_object_busied(obj)) {
1124 		if (locked)
1125 			VM_OBJECT_DROP(obj);
1126 		vm_object_busy_wait(obj, wmesg);
1127 		return (true);
1128 	}
1129 
1130 	if (!vm_page_busied(m))
1131 		return (false);
1132 
1133 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1134 	sleepq_lock(m);
1135 	x = vm_page_busy_fetch(m);
1136 	do {
1137 		/*
1138 		 * If the page changes objects or becomes unlocked we can
1139 		 * simply return.
1140 		 */
1141 		if (x == VPB_UNBUSIED ||
1142 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1143 		    m->object != obj || m->pindex != pindex) {
1144 			sleepq_release(m);
1145 			return (false);
1146 		}
1147 		if ((x & VPB_BIT_WAITERS) != 0)
1148 			break;
1149 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1150 	if (locked)
1151 		VM_OBJECT_DROP(obj);
1152 	DROP_GIANT();
1153 	sleepq_add(m, NULL, wmesg, 0, 0);
1154 	sleepq_wait(m, PVM);
1155 	PICKUP_GIANT();
1156 	return (true);
1157 }
1158 
1159 /*
1160  *	vm_page_trysbusy:
1161  *
1162  *	Try to shared busy a page.
1163  *	If the operation succeeds 1 is returned otherwise 0.
1164  *	The operation never sleeps.
1165  */
1166 int
1167 vm_page_trysbusy(vm_page_t m)
1168 {
1169 	vm_object_t obj;
1170 	u_int x;
1171 
1172 	obj = m->object;
1173 	x = vm_page_busy_fetch(m);
1174 	for (;;) {
1175 		if ((x & VPB_BIT_SHARED) == 0)
1176 			return (0);
1177 		/*
1178 		 * Reduce the window for transient busies that will trigger
1179 		 * false negatives in vm_page_ps_test().
1180 		 */
1181 		if (obj != NULL && vm_object_busied(obj))
1182 			return (0);
1183 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1184 		    x + VPB_ONE_SHARER))
1185 			break;
1186 	}
1187 
1188 	/* Refetch the object now that we're guaranteed that it is stable. */
1189 	obj = m->object;
1190 	if (obj != NULL && vm_object_busied(obj)) {
1191 		vm_page_sunbusy(m);
1192 		return (0);
1193 	}
1194 	return (1);
1195 }
1196 
1197 /*
1198  *	vm_page_tryxbusy:
1199  *
1200  *	Try to exclusive busy a page.
1201  *	If the operation succeeds 1 is returned otherwise 0.
1202  *	The operation never sleeps.
1203  */
1204 int
1205 vm_page_tryxbusy(vm_page_t m)
1206 {
1207 	vm_object_t obj;
1208 
1209         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1210             VPB_CURTHREAD_EXCLUSIVE) == 0)
1211 		return (0);
1212 
1213 	obj = m->object;
1214 	if (obj != NULL && vm_object_busied(obj)) {
1215 		vm_page_xunbusy(m);
1216 		return (0);
1217 	}
1218 	return (1);
1219 }
1220 
1221 static void
1222 vm_page_xunbusy_hard_tail(vm_page_t m)
1223 {
1224 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1225 	/* Wake the waiter. */
1226 	wakeup(m);
1227 }
1228 
1229 /*
1230  *	vm_page_xunbusy_hard:
1231  *
1232  *	Called when unbusy has failed because there is a waiter.
1233  */
1234 void
1235 vm_page_xunbusy_hard(vm_page_t m)
1236 {
1237 	vm_page_assert_xbusied(m);
1238 	vm_page_xunbusy_hard_tail(m);
1239 }
1240 
1241 void
1242 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1243 {
1244 	vm_page_assert_xbusied_unchecked(m);
1245 	vm_page_xunbusy_hard_tail(m);
1246 }
1247 
1248 static void
1249 vm_page_busy_free(vm_page_t m)
1250 {
1251 	u_int x;
1252 
1253 	atomic_thread_fence_rel();
1254 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1255 	if ((x & VPB_BIT_WAITERS) != 0)
1256 		wakeup(m);
1257 }
1258 
1259 /*
1260  *	vm_page_unhold_pages:
1261  *
1262  *	Unhold each of the pages that is referenced by the given array.
1263  */
1264 void
1265 vm_page_unhold_pages(vm_page_t *ma, int count)
1266 {
1267 
1268 	for (; count != 0; count--) {
1269 		vm_page_unwire(*ma, PQ_ACTIVE);
1270 		ma++;
1271 	}
1272 }
1273 
1274 vm_page_t
1275 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1276 {
1277 	vm_page_t m;
1278 
1279 #ifdef VM_PHYSSEG_SPARSE
1280 	m = vm_phys_paddr_to_vm_page(pa);
1281 	if (m == NULL)
1282 		m = vm_phys_fictitious_to_vm_page(pa);
1283 	return (m);
1284 #elif defined(VM_PHYSSEG_DENSE)
1285 	long pi;
1286 
1287 	pi = atop(pa);
1288 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1289 		m = &vm_page_array[pi - first_page];
1290 		return (m);
1291 	}
1292 	return (vm_phys_fictitious_to_vm_page(pa));
1293 #else
1294 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1295 #endif
1296 }
1297 
1298 /*
1299  *	vm_page_getfake:
1300  *
1301  *	Create a fictitious page with the specified physical address and
1302  *	memory attribute.  The memory attribute is the only the machine-
1303  *	dependent aspect of a fictitious page that must be initialized.
1304  */
1305 vm_page_t
1306 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1307 {
1308 	vm_page_t m;
1309 
1310 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1311 	vm_page_initfake(m, paddr, memattr);
1312 	return (m);
1313 }
1314 
1315 void
1316 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1317 {
1318 
1319 	if ((m->flags & PG_FICTITIOUS) != 0) {
1320 		/*
1321 		 * The page's memattr might have changed since the
1322 		 * previous initialization.  Update the pmap to the
1323 		 * new memattr.
1324 		 */
1325 		goto memattr;
1326 	}
1327 	m->phys_addr = paddr;
1328 	m->a.queue = PQ_NONE;
1329 	/* Fictitious pages don't use "segind". */
1330 	m->flags = PG_FICTITIOUS;
1331 	/* Fictitious pages don't use "order" or "pool". */
1332 	m->oflags = VPO_UNMANAGED;
1333 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1334 	/* Fictitious pages are unevictable. */
1335 	m->ref_count = 1;
1336 	pmap_page_init(m);
1337 memattr:
1338 	pmap_page_set_memattr(m, memattr);
1339 }
1340 
1341 /*
1342  *	vm_page_putfake:
1343  *
1344  *	Release a fictitious page.
1345  */
1346 void
1347 vm_page_putfake(vm_page_t m)
1348 {
1349 
1350 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1351 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1352 	    ("vm_page_putfake: bad page %p", m));
1353 	vm_page_assert_xbusied(m);
1354 	vm_page_busy_free(m);
1355 	uma_zfree(fakepg_zone, m);
1356 }
1357 
1358 /*
1359  *	vm_page_updatefake:
1360  *
1361  *	Update the given fictitious page to the specified physical address and
1362  *	memory attribute.
1363  */
1364 void
1365 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1366 {
1367 
1368 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1369 	    ("vm_page_updatefake: bad page %p", m));
1370 	m->phys_addr = paddr;
1371 	pmap_page_set_memattr(m, memattr);
1372 }
1373 
1374 /*
1375  *	vm_page_free:
1376  *
1377  *	Free a page.
1378  */
1379 void
1380 vm_page_free(vm_page_t m)
1381 {
1382 
1383 	m->flags &= ~PG_ZERO;
1384 	vm_page_free_toq(m);
1385 }
1386 
1387 /*
1388  *	vm_page_free_zero:
1389  *
1390  *	Free a page to the zerod-pages queue
1391  */
1392 void
1393 vm_page_free_zero(vm_page_t m)
1394 {
1395 
1396 	m->flags |= PG_ZERO;
1397 	vm_page_free_toq(m);
1398 }
1399 
1400 /*
1401  * Unbusy and handle the page queueing for a page from a getpages request that
1402  * was optionally read ahead or behind.
1403  */
1404 void
1405 vm_page_readahead_finish(vm_page_t m)
1406 {
1407 
1408 	/* We shouldn't put invalid pages on queues. */
1409 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1410 
1411 	/*
1412 	 * Since the page is not the actually needed one, whether it should
1413 	 * be activated or deactivated is not obvious.  Empirical results
1414 	 * have shown that deactivating the page is usually the best choice,
1415 	 * unless the page is wanted by another thread.
1416 	 */
1417 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1418 		vm_page_activate(m);
1419 	else
1420 		vm_page_deactivate(m);
1421 	vm_page_xunbusy_unchecked(m);
1422 }
1423 
1424 /*
1425  * Destroy the identity of an invalid page and free it if possible.
1426  * This is intended to be used when reading a page from backing store fails.
1427  */
1428 void
1429 vm_page_free_invalid(vm_page_t m)
1430 {
1431 
1432 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1433 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1434 	KASSERT(m->object != NULL, ("page %p has no object", m));
1435 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1436 
1437 	/*
1438 	 * We may be attempting to free the page as part of the handling for an
1439 	 * I/O error, in which case the page was xbusied by a different thread.
1440 	 */
1441 	vm_page_xbusy_claim(m);
1442 
1443 	/*
1444 	 * If someone has wired this page while the object lock
1445 	 * was not held, then the thread that unwires is responsible
1446 	 * for freeing the page.  Otherwise just free the page now.
1447 	 * The wire count of this unmapped page cannot change while
1448 	 * we have the page xbusy and the page's object wlocked.
1449 	 */
1450 	if (vm_page_remove(m))
1451 		vm_page_free(m);
1452 }
1453 
1454 /*
1455  *	vm_page_dirty_KBI:		[ internal use only ]
1456  *
1457  *	Set all bits in the page's dirty field.
1458  *
1459  *	The object containing the specified page must be locked if the
1460  *	call is made from the machine-independent layer.
1461  *
1462  *	See vm_page_clear_dirty_mask().
1463  *
1464  *	This function should only be called by vm_page_dirty().
1465  */
1466 void
1467 vm_page_dirty_KBI(vm_page_t m)
1468 {
1469 
1470 	/* Refer to this operation by its public name. */
1471 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1472 	m->dirty = VM_PAGE_BITS_ALL;
1473 }
1474 
1475 /*
1476  * Insert the given page into the given object at the given pindex.  mpred is
1477  * used for memq linkage.  From vm_page_insert, lookup is true, mpred is
1478  * initially NULL, and this procedure looks it up.  From vm_page_insert_after,
1479  * lookup is false and mpred is known to the caller to be valid, and may be
1480  * NULL if this will be the page with the lowest pindex.
1481  *
1482  * The procedure is marked __always_inline to suggest to the compiler to
1483  * eliminate the lookup parameter and the associated alternate branch.
1484  */
1485 static __always_inline int
1486 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1487     vm_page_t mpred, bool lookup)
1488 {
1489 	int error;
1490 
1491 	VM_OBJECT_ASSERT_WLOCKED(object);
1492 	KASSERT(m->object == NULL,
1493 	    ("vm_page_insert: page %p already inserted", m));
1494 
1495 	/*
1496 	 * Record the object/offset pair in this page.
1497 	 */
1498 	m->object = object;
1499 	m->pindex = pindex;
1500 	m->ref_count |= VPRC_OBJREF;
1501 
1502 	/*
1503 	 * Add this page to the object's radix tree, and look up mpred if
1504 	 * needed.
1505 	 */
1506 	if (lookup)
1507 		error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred);
1508 	else
1509 		error = vm_radix_insert(&object->rtree, m);
1510 	if (__predict_false(error != 0)) {
1511 		m->object = NULL;
1512 		m->pindex = 0;
1513 		m->ref_count &= ~VPRC_OBJREF;
1514 		return (1);
1515 	}
1516 
1517 	/*
1518 	 * Now link into the object's ordered list of backed pages.
1519 	 */
1520 	vm_page_insert_radixdone(m, object, mpred);
1521 	vm_pager_page_inserted(object, m);
1522 	return (0);
1523 }
1524 
1525 /*
1526  *	vm_page_insert:		[ internal use only ]
1527  *
1528  *	Inserts the given mem entry into the object and object list.
1529  *
1530  *	The object must be locked.
1531  */
1532 int
1533 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1534 {
1535 	return (vm_page_insert_lookup(m, object, pindex, NULL, true));
1536 }
1537 
1538 /*
1539  *	vm_page_insert_after:
1540  *
1541  *	Inserts the page "m" into the specified object at offset "pindex".
1542  *
1543  *	The page "mpred" must immediately precede the offset "pindex" within
1544  *	the specified object.
1545  *
1546  *	The object must be locked.
1547  */
1548 static int
1549 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1550     vm_page_t mpred)
1551 {
1552 	return (vm_page_insert_lookup(m, object, pindex, mpred, false));
1553 }
1554 
1555 /*
1556  *	vm_page_insert_radixdone:
1557  *
1558  *	Complete page "m" insertion into the specified object after the
1559  *	radix trie hooking.
1560  *
1561  *	The page "mpred" must precede the offset "m->pindex" within the
1562  *	specified object.
1563  *
1564  *	The object must be locked.
1565  */
1566 static void
1567 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1568 {
1569 
1570 	VM_OBJECT_ASSERT_WLOCKED(object);
1571 	KASSERT(object != NULL && m->object == object,
1572 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1573 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1574 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1575 	if (mpred != NULL) {
1576 		KASSERT(mpred->object == object,
1577 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1578 		KASSERT(mpred->pindex < m->pindex,
1579 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1580 		KASSERT(TAILQ_NEXT(mpred, listq) == NULL ||
1581 		    m->pindex < TAILQ_NEXT(mpred, listq)->pindex,
1582 		    ("vm_page_insert_radixdone: pindex doesn't precede msucc"));
1583 	} else {
1584 		KASSERT(TAILQ_EMPTY(&object->memq) ||
1585 		    m->pindex < TAILQ_FIRST(&object->memq)->pindex,
1586 		    ("vm_page_insert_radixdone: no mpred but not first page"));
1587 	}
1588 
1589 	if (mpred != NULL)
1590 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1591 	else
1592 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1593 
1594 	/*
1595 	 * Show that the object has one more resident page.
1596 	 */
1597 	object->resident_page_count++;
1598 
1599 	/*
1600 	 * Hold the vnode until the last page is released.
1601 	 */
1602 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1603 		vhold(object->handle);
1604 
1605 	/*
1606 	 * Since we are inserting a new and possibly dirty page,
1607 	 * update the object's generation count.
1608 	 */
1609 	if (pmap_page_is_write_mapped(m))
1610 		vm_object_set_writeable_dirty(object);
1611 }
1612 
1613 /*
1614  * Do the work to remove a page from its object.  The caller is responsible for
1615  * updating the page's fields to reflect this removal.
1616  */
1617 static void
1618 vm_page_object_remove(vm_page_t m)
1619 {
1620 	vm_object_t object;
1621 	vm_page_t mrem __diagused;
1622 
1623 	vm_page_assert_xbusied(m);
1624 	object = m->object;
1625 	VM_OBJECT_ASSERT_WLOCKED(object);
1626 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1627 	    ("page %p is missing its object ref", m));
1628 
1629 	/* Deferred free of swap space. */
1630 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1631 		vm_pager_page_unswapped(m);
1632 
1633 	vm_pager_page_removed(object, m);
1634 
1635 	m->object = NULL;
1636 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1637 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1638 
1639 	/*
1640 	 * Now remove from the object's list of backed pages.
1641 	 */
1642 	TAILQ_REMOVE(&object->memq, m, listq);
1643 
1644 	/*
1645 	 * And show that the object has one fewer resident page.
1646 	 */
1647 	object->resident_page_count--;
1648 
1649 	/*
1650 	 * The vnode may now be recycled.
1651 	 */
1652 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1653 		vdrop(object->handle);
1654 }
1655 
1656 /*
1657  *	vm_page_remove:
1658  *
1659  *	Removes the specified page from its containing object, but does not
1660  *	invalidate any backing storage.  Returns true if the object's reference
1661  *	was the last reference to the page, and false otherwise.
1662  *
1663  *	The object must be locked and the page must be exclusively busied.
1664  *	The exclusive busy will be released on return.  If this is not the
1665  *	final ref and the caller does not hold a wire reference it may not
1666  *	continue to access the page.
1667  */
1668 bool
1669 vm_page_remove(vm_page_t m)
1670 {
1671 	bool dropped;
1672 
1673 	dropped = vm_page_remove_xbusy(m);
1674 	vm_page_xunbusy(m);
1675 
1676 	return (dropped);
1677 }
1678 
1679 /*
1680  *	vm_page_remove_xbusy
1681  *
1682  *	Removes the page but leaves the xbusy held.  Returns true if this
1683  *	removed the final ref and false otherwise.
1684  */
1685 bool
1686 vm_page_remove_xbusy(vm_page_t m)
1687 {
1688 
1689 	vm_page_object_remove(m);
1690 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1691 }
1692 
1693 /*
1694  *	vm_page_lookup:
1695  *
1696  *	Returns the page associated with the object/offset
1697  *	pair specified; if none is found, NULL is returned.
1698  *
1699  *	The object must be locked.
1700  */
1701 vm_page_t
1702 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1703 {
1704 
1705 	VM_OBJECT_ASSERT_LOCKED(object);
1706 	return (vm_radix_lookup(&object->rtree, pindex));
1707 }
1708 
1709 /*
1710  *	vm_page_iter_init:
1711  *
1712  *	Initialize iterator for vm pages.
1713  */
1714 void
1715 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object)
1716 {
1717 
1718 	vm_radix_iter_init(pages, &object->rtree);
1719 }
1720 
1721 /*
1722  *	vm_page_iter_init:
1723  *
1724  *	Initialize iterator for vm pages.
1725  */
1726 void
1727 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object,
1728     vm_pindex_t limit)
1729 {
1730 
1731 	vm_radix_iter_limit_init(pages, &object->rtree, limit);
1732 }
1733 
1734 /*
1735  *	vm_page_iter_lookup:
1736  *
1737  *	Returns the page associated with the object/offset pair specified, and
1738  *	stores the path to its position; if none is found, NULL is returned.
1739  *
1740  *	The iter pctrie must be locked.
1741  */
1742 vm_page_t
1743 vm_page_iter_lookup(struct pctrie_iter *pages, vm_pindex_t pindex)
1744 {
1745 
1746 	return (vm_radix_iter_lookup(pages, pindex));
1747 }
1748 
1749 /*
1750  *	vm_page_lookup_unlocked:
1751  *
1752  *	Returns the page associated with the object/offset pair specified;
1753  *	if none is found, NULL is returned.  The page may be no longer be
1754  *	present in the object at the time that this function returns.  Only
1755  *	useful for opportunistic checks such as inmem().
1756  */
1757 vm_page_t
1758 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1759 {
1760 
1761 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1762 }
1763 
1764 /*
1765  *	vm_page_relookup:
1766  *
1767  *	Returns a page that must already have been busied by
1768  *	the caller.  Used for bogus page replacement.
1769  */
1770 vm_page_t
1771 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1772 {
1773 	vm_page_t m;
1774 
1775 	m = vm_page_lookup_unlocked(object, pindex);
1776 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1777 	    m->object == object && m->pindex == pindex,
1778 	    ("vm_page_relookup: Invalid page %p", m));
1779 	return (m);
1780 }
1781 
1782 /*
1783  * This should only be used by lockless functions for releasing transient
1784  * incorrect acquires.  The page may have been freed after we acquired a
1785  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1786  * further to do.
1787  */
1788 static void
1789 vm_page_busy_release(vm_page_t m)
1790 {
1791 	u_int x;
1792 
1793 	x = vm_page_busy_fetch(m);
1794 	for (;;) {
1795 		if (x == VPB_FREED)
1796 			break;
1797 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1798 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1799 			    x - VPB_ONE_SHARER))
1800 				break;
1801 			continue;
1802 		}
1803 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1804 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1805 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1806 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1807 			continue;
1808 		if ((x & VPB_BIT_WAITERS) != 0)
1809 			wakeup(m);
1810 		break;
1811 	}
1812 }
1813 
1814 /*
1815  *	vm_page_find_least:
1816  *
1817  *	Returns the page associated with the object with least pindex
1818  *	greater than or equal to the parameter pindex, or NULL.
1819  *
1820  *	The object must be locked.
1821  */
1822 vm_page_t
1823 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1824 {
1825 	vm_page_t m;
1826 
1827 	VM_OBJECT_ASSERT_LOCKED(object);
1828 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1829 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1830 	return (m);
1831 }
1832 
1833 /*
1834  *	vm_page_iter_lookup_ge:
1835  *
1836  *	Returns the page associated with the object with least pindex
1837  *	greater than or equal to the parameter pindex, or NULL.  Initializes the
1838  *	iterator to point to that page.
1839  *
1840  *	The iter pctrie must be locked.
1841  */
1842 vm_page_t
1843 vm_page_iter_lookup_ge(struct pctrie_iter *pages, vm_pindex_t pindex)
1844 {
1845 
1846 	return (vm_radix_iter_lookup_ge(pages, pindex));
1847 }
1848 
1849 /*
1850  * Returns the given page's successor (by pindex) within the object if it is
1851  * resident; if none is found, NULL is returned.
1852  *
1853  * The object must be locked.
1854  */
1855 vm_page_t
1856 vm_page_next(vm_page_t m)
1857 {
1858 	vm_page_t next;
1859 
1860 	VM_OBJECT_ASSERT_LOCKED(m->object);
1861 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1862 		MPASS(next->object == m->object);
1863 		if (next->pindex != m->pindex + 1)
1864 			next = NULL;
1865 	}
1866 	return (next);
1867 }
1868 
1869 /*
1870  * Returns the given page's predecessor (by pindex) within the object if it is
1871  * resident; if none is found, NULL is returned.
1872  *
1873  * The object must be locked.
1874  */
1875 vm_page_t
1876 vm_page_prev(vm_page_t m)
1877 {
1878 	vm_page_t prev;
1879 
1880 	VM_OBJECT_ASSERT_LOCKED(m->object);
1881 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1882 		MPASS(prev->object == m->object);
1883 		if (prev->pindex != m->pindex - 1)
1884 			prev = NULL;
1885 	}
1886 	return (prev);
1887 }
1888 
1889 /*
1890  * Uses the page mnew as a replacement for an existing page at index
1891  * pindex which must be already present in the object.
1892  *
1893  * Both pages must be exclusively busied on enter.  The old page is
1894  * unbusied on exit.
1895  *
1896  * A return value of true means mold is now free.  If this is not the
1897  * final ref and the caller does not hold a wire reference it may not
1898  * continue to access the page.
1899  */
1900 static bool
1901 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1902     vm_page_t mold)
1903 {
1904 	vm_page_t mret __diagused;
1905 	bool dropped;
1906 
1907 	VM_OBJECT_ASSERT_WLOCKED(object);
1908 	vm_page_assert_xbusied(mold);
1909 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1910 	    ("vm_page_replace: page %p already in object", mnew));
1911 
1912 	/*
1913 	 * This function mostly follows vm_page_insert() and
1914 	 * vm_page_remove() without the radix, object count and vnode
1915 	 * dance.  Double check such functions for more comments.
1916 	 */
1917 
1918 	mnew->object = object;
1919 	mnew->pindex = pindex;
1920 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1921 	mret = vm_radix_replace(&object->rtree, mnew);
1922 	KASSERT(mret == mold,
1923 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1924 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1925 	    (mnew->oflags & VPO_UNMANAGED),
1926 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1927 
1928 	/* Keep the resident page list in sorted order. */
1929 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1930 	TAILQ_REMOVE(&object->memq, mold, listq);
1931 	mold->object = NULL;
1932 
1933 	/*
1934 	 * The object's resident_page_count does not change because we have
1935 	 * swapped one page for another, but the generation count should
1936 	 * change if the page is dirty.
1937 	 */
1938 	if (pmap_page_is_write_mapped(mnew))
1939 		vm_object_set_writeable_dirty(object);
1940 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1941 	vm_page_xunbusy(mold);
1942 
1943 	return (dropped);
1944 }
1945 
1946 void
1947 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1948     vm_page_t mold)
1949 {
1950 
1951 	vm_page_assert_xbusied(mnew);
1952 
1953 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1954 		vm_page_free(mold);
1955 }
1956 
1957 /*
1958  *	vm_page_rename:
1959  *
1960  *	Move the given memory entry from its
1961  *	current object to the specified target object/offset.
1962  *
1963  *	Note: swap associated with the page must be invalidated by the move.  We
1964  *	      have to do this for several reasons:  (1) we aren't freeing the
1965  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1966  *	      moving the page from object A to B, and will then later move
1967  *	      the backing store from A to B and we can't have a conflict.
1968  *
1969  *	Note: we *always* dirty the page.  It is necessary both for the
1970  *	      fact that we moved it, and because we may be invalidating
1971  *	      swap.
1972  *
1973  *	The objects must be locked.
1974  */
1975 int
1976 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1977 {
1978 	vm_page_t mpred;
1979 	vm_pindex_t opidx;
1980 
1981 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1982 
1983 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1984 
1985 	/*
1986 	 * Create a custom version of vm_page_insert() which does not depend
1987 	 * by m_prev and can cheat on the implementation aspects of the
1988 	 * function.
1989 	 */
1990 	opidx = m->pindex;
1991 	m->pindex = new_pindex;
1992 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
1993 		m->pindex = opidx;
1994 		return (1);
1995 	}
1996 
1997 	/*
1998 	 * The operation cannot fail anymore.  The removal must happen before
1999 	 * the listq iterator is tainted.
2000 	 */
2001 	m->pindex = opidx;
2002 	vm_page_object_remove(m);
2003 
2004 	/* Return back to the new pindex to complete vm_page_insert(). */
2005 	m->pindex = new_pindex;
2006 	m->object = new_object;
2007 
2008 	vm_page_insert_radixdone(m, new_object, mpred);
2009 	vm_page_dirty(m);
2010 	vm_pager_page_inserted(new_object, m);
2011 	return (0);
2012 }
2013 
2014 /*
2015  *	vm_page_mpred:
2016  *
2017  *	Return the greatest page of the object with index <= pindex,
2018  *	or NULL, if there is none.  Assumes object lock is held.
2019  */
2020 vm_page_t
2021 vm_page_mpred(vm_object_t object, vm_pindex_t pindex)
2022 {
2023 	return (vm_radix_lookup_le(&object->rtree, pindex));
2024 }
2025 
2026 /*
2027  *	vm_page_alloc:
2028  *
2029  *	Allocate and return a page that is associated with the specified
2030  *	object and offset pair.  By default, this page is exclusive busied.
2031  *
2032  *	The caller must always specify an allocation class.
2033  *
2034  *	allocation classes:
2035  *	VM_ALLOC_NORMAL		normal process request
2036  *	VM_ALLOC_SYSTEM		system *really* needs a page
2037  *	VM_ALLOC_INTERRUPT	interrupt time request
2038  *
2039  *	optional allocation flags:
2040  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2041  *				intends to allocate
2042  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2043  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2044  *	VM_ALLOC_SBUSY		shared busy the allocated page
2045  *	VM_ALLOC_WIRED		wire the allocated page
2046  *	VM_ALLOC_ZERO		prefer a zeroed page
2047  */
2048 vm_page_t
2049 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
2050 {
2051 
2052 	return (vm_page_alloc_after(object, pindex, req,
2053 	    vm_page_mpred(object, pindex)));
2054 }
2055 
2056 /*
2057  * Allocate a page in the specified object with the given page index.  To
2058  * optimize insertion of the page into the object, the caller must also specify
2059  * the resident page in the object with largest index smaller than the given
2060  * page index, or NULL if no such page exists.
2061  */
2062 vm_page_t
2063 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
2064     int req, vm_page_t mpred)
2065 {
2066 	struct vm_domainset_iter di;
2067 	vm_page_t m;
2068 	int domain;
2069 
2070 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2071 	do {
2072 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2073 		    mpred);
2074 		if (m != NULL)
2075 			break;
2076 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2077 
2078 	return (m);
2079 }
2080 
2081 /*
2082  * Returns true if the number of free pages exceeds the minimum
2083  * for the request class and false otherwise.
2084  */
2085 static int
2086 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2087 {
2088 	u_int limit, old, new;
2089 
2090 	if (req_class == VM_ALLOC_INTERRUPT)
2091 		limit = 0;
2092 	else if (req_class == VM_ALLOC_SYSTEM)
2093 		limit = vmd->vmd_interrupt_free_min;
2094 	else
2095 		limit = vmd->vmd_free_reserved;
2096 
2097 	/*
2098 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2099 	 */
2100 	limit += npages;
2101 	old = atomic_load_int(&vmd->vmd_free_count);
2102 	do {
2103 		if (old < limit)
2104 			return (0);
2105 		new = old - npages;
2106 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2107 
2108 	/* Wake the page daemon if we've crossed the threshold. */
2109 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2110 		pagedaemon_wakeup(vmd->vmd_domain);
2111 
2112 	/* Only update bitsets on transitions. */
2113 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2114 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2115 		vm_domain_set(vmd);
2116 
2117 	return (1);
2118 }
2119 
2120 int
2121 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2122 {
2123 	int req_class;
2124 
2125 	/*
2126 	 * The page daemon is allowed to dig deeper into the free page list.
2127 	 */
2128 	req_class = req & VM_ALLOC_CLASS_MASK;
2129 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2130 		req_class = VM_ALLOC_SYSTEM;
2131 	return (_vm_domain_allocate(vmd, req_class, npages));
2132 }
2133 
2134 vm_page_t
2135 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2136     int req, vm_page_t mpred)
2137 {
2138 	struct vm_domain *vmd;
2139 	vm_page_t m;
2140 	int flags;
2141 
2142 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2143 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2144 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2145 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2146 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2147 	KASSERT((req & ~VPA_FLAGS) == 0,
2148 	    ("invalid request %#x", req));
2149 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2150 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2151 	    ("invalid request %#x", req));
2152 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2153 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2154 	    (uintmax_t)pindex));
2155 	VM_OBJECT_ASSERT_WLOCKED(object);
2156 
2157 	flags = 0;
2158 	m = NULL;
2159 	if (!vm_pager_can_alloc_page(object, pindex))
2160 		return (NULL);
2161 again:
2162 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2163 		m = vm_page_alloc_nofree_domain(domain, req);
2164 		if (m != NULL)
2165 			goto found;
2166 	}
2167 #if VM_NRESERVLEVEL > 0
2168 	/*
2169 	 * Can we allocate the page from a reservation?
2170 	 */
2171 	if (vm_object_reserv(object) &&
2172 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2173 	    NULL) {
2174 		goto found;
2175 	}
2176 #endif
2177 	vmd = VM_DOMAIN(domain);
2178 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2179 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2180 		    M_NOWAIT | M_NOVM);
2181 		if (m != NULL) {
2182 			flags |= PG_PCPU_CACHE;
2183 			goto found;
2184 		}
2185 	}
2186 	if (vm_domain_allocate(vmd, req, 1)) {
2187 		/*
2188 		 * If not, allocate it from the free page queues.
2189 		 */
2190 		vm_domain_free_lock(vmd);
2191 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2192 		vm_domain_free_unlock(vmd);
2193 		if (m == NULL) {
2194 			vm_domain_freecnt_inc(vmd, 1);
2195 #if VM_NRESERVLEVEL > 0
2196 			if (vm_reserv_reclaim_inactive(domain))
2197 				goto again;
2198 #endif
2199 		}
2200 	}
2201 	if (m == NULL) {
2202 		/*
2203 		 * Not allocatable, give up.
2204 		 */
2205 		if (vm_domain_alloc_fail(vmd, object, req))
2206 			goto again;
2207 		return (NULL);
2208 	}
2209 
2210 	/*
2211 	 * At this point we had better have found a good page.
2212 	 */
2213 found:
2214 	vm_page_dequeue(m);
2215 	vm_page_alloc_check(m);
2216 
2217 	/*
2218 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2219 	 */
2220 	flags |= m->flags & PG_ZERO;
2221 	if ((req & VM_ALLOC_NODUMP) != 0)
2222 		flags |= PG_NODUMP;
2223 	if ((req & VM_ALLOC_NOFREE) != 0)
2224 		flags |= PG_NOFREE;
2225 	m->flags = flags;
2226 	m->a.flags = 0;
2227 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2228 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2229 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2230 	else if ((req & VM_ALLOC_SBUSY) != 0)
2231 		m->busy_lock = VPB_SHARERS_WORD(1);
2232 	else
2233 		m->busy_lock = VPB_UNBUSIED;
2234 	if (req & VM_ALLOC_WIRED) {
2235 		vm_wire_add(1);
2236 		m->ref_count = 1;
2237 	}
2238 	m->a.act_count = 0;
2239 
2240 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2241 		if (req & VM_ALLOC_WIRED) {
2242 			vm_wire_sub(1);
2243 			m->ref_count = 0;
2244 		}
2245 		KASSERT(m->object == NULL, ("page %p has object", m));
2246 		m->oflags = VPO_UNMANAGED;
2247 		m->busy_lock = VPB_UNBUSIED;
2248 		/* Don't change PG_ZERO. */
2249 		vm_page_free_toq(m);
2250 		if (req & VM_ALLOC_WAITFAIL) {
2251 			VM_OBJECT_WUNLOCK(object);
2252 			vm_radix_wait();
2253 			VM_OBJECT_WLOCK(object);
2254 		}
2255 		return (NULL);
2256 	}
2257 
2258 	/* Ignore device objects; the pager sets "memattr" for them. */
2259 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2260 	    (object->flags & OBJ_FICTITIOUS) == 0)
2261 		pmap_page_set_memattr(m, object->memattr);
2262 
2263 	return (m);
2264 }
2265 
2266 /*
2267  *	vm_page_alloc_contig:
2268  *
2269  *	Allocate a contiguous set of physical pages of the given size "npages"
2270  *	from the free lists.  All of the physical pages must be at or above
2271  *	the given physical address "low" and below the given physical address
2272  *	"high".  The given value "alignment" determines the alignment of the
2273  *	first physical page in the set.  If the given value "boundary" is
2274  *	non-zero, then the set of physical pages cannot cross any physical
2275  *	address boundary that is a multiple of that value.  Both "alignment"
2276  *	and "boundary" must be a power of two.
2277  *
2278  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2279  *	then the memory attribute setting for the physical pages is configured
2280  *	to the object's memory attribute setting.  Otherwise, the memory
2281  *	attribute setting for the physical pages is configured to "memattr",
2282  *	overriding the object's memory attribute setting.  However, if the
2283  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2284  *	memory attribute setting for the physical pages cannot be configured
2285  *	to VM_MEMATTR_DEFAULT.
2286  *
2287  *	The specified object may not contain fictitious pages.
2288  *
2289  *	The caller must always specify an allocation class.
2290  *
2291  *	allocation classes:
2292  *	VM_ALLOC_NORMAL		normal process request
2293  *	VM_ALLOC_SYSTEM		system *really* needs a page
2294  *	VM_ALLOC_INTERRUPT	interrupt time request
2295  *
2296  *	optional allocation flags:
2297  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2298  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2299  *	VM_ALLOC_SBUSY		shared busy the allocated page
2300  *	VM_ALLOC_WIRED		wire the allocated page
2301  *	VM_ALLOC_ZERO		prefer a zeroed page
2302  */
2303 vm_page_t
2304 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2305     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2306     vm_paddr_t boundary, vm_memattr_t memattr)
2307 {
2308 	struct vm_domainset_iter di;
2309 	vm_page_t bounds[2];
2310 	vm_page_t m;
2311 	int domain;
2312 	int start_segind;
2313 
2314 	start_segind = -1;
2315 
2316 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2317 	do {
2318 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2319 		    npages, low, high, alignment, boundary, memattr);
2320 		if (m != NULL)
2321 			break;
2322 		if (start_segind == -1)
2323 			start_segind = vm_phys_lookup_segind(low);
2324 		if (vm_phys_find_range(bounds, start_segind, domain,
2325 		    npages, low, high) == -1) {
2326 			vm_domainset_iter_ignore(&di, domain);
2327 		}
2328 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2329 
2330 	return (m);
2331 }
2332 
2333 static vm_page_t
2334 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2335     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2336 {
2337 	struct vm_domain *vmd;
2338 	vm_page_t m_ret;
2339 
2340 	/*
2341 	 * Can we allocate the pages without the number of free pages falling
2342 	 * below the lower bound for the allocation class?
2343 	 */
2344 	vmd = VM_DOMAIN(domain);
2345 	if (!vm_domain_allocate(vmd, req, npages))
2346 		return (NULL);
2347 	/*
2348 	 * Try to allocate the pages from the free page queues.
2349 	 */
2350 	vm_domain_free_lock(vmd);
2351 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2352 	    alignment, boundary);
2353 	vm_domain_free_unlock(vmd);
2354 	if (m_ret != NULL)
2355 		return (m_ret);
2356 #if VM_NRESERVLEVEL > 0
2357 	/*
2358 	 * Try to break a reservation to allocate the pages.
2359 	 */
2360 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2361 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2362 	            high, alignment, boundary);
2363 		if (m_ret != NULL)
2364 			return (m_ret);
2365 	}
2366 #endif
2367 	vm_domain_freecnt_inc(vmd, npages);
2368 	return (NULL);
2369 }
2370 
2371 vm_page_t
2372 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2373     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2374     vm_paddr_t boundary, vm_memattr_t memattr)
2375 {
2376 	vm_page_t m, m_ret, mpred;
2377 	u_int busy_lock, flags, oflags;
2378 
2379 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2380 	KASSERT((req & ~VPAC_FLAGS) == 0,
2381 	    ("invalid request %#x", req));
2382 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2383 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2384 	    ("invalid request %#x", req));
2385 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2386 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2387 	    ("invalid request %#x", req));
2388 	VM_OBJECT_ASSERT_WLOCKED(object);
2389 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2390 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2391 	    object));
2392 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2393 
2394 	mpred = vm_page_mpred(object, pindex);
2395 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2396 	    ("vm_page_alloc_contig: pindex already allocated"));
2397 	for (;;) {
2398 #if VM_NRESERVLEVEL > 0
2399 		/*
2400 		 * Can we allocate the pages from a reservation?
2401 		 */
2402 		if (vm_object_reserv(object) &&
2403 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2404 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2405 			break;
2406 		}
2407 #endif
2408 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2409 		    low, high, alignment, boundary)) != NULL)
2410 			break;
2411 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2412 			return (NULL);
2413 	}
2414 
2415 	/*
2416 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2417 	 */
2418 	flags = PG_ZERO;
2419 	if ((req & VM_ALLOC_NODUMP) != 0)
2420 		flags |= PG_NODUMP;
2421 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2422 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2423 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2424 	else if ((req & VM_ALLOC_SBUSY) != 0)
2425 		busy_lock = VPB_SHARERS_WORD(1);
2426 	else
2427 		busy_lock = VPB_UNBUSIED;
2428 	if ((req & VM_ALLOC_WIRED) != 0)
2429 		vm_wire_add(npages);
2430 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2431 	    memattr == VM_MEMATTR_DEFAULT)
2432 		memattr = object->memattr;
2433 	for (m = m_ret; m < &m_ret[npages]; m++) {
2434 		vm_page_dequeue(m);
2435 		vm_page_alloc_check(m);
2436 		m->a.flags = 0;
2437 		m->flags = (m->flags | PG_NODUMP) & flags;
2438 		m->busy_lock = busy_lock;
2439 		if ((req & VM_ALLOC_WIRED) != 0)
2440 			m->ref_count = 1;
2441 		m->a.act_count = 0;
2442 		m->oflags = oflags;
2443 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2444 			if ((req & VM_ALLOC_WIRED) != 0)
2445 				vm_wire_sub(npages);
2446 			KASSERT(m->object == NULL,
2447 			    ("page %p has object", m));
2448 			mpred = m;
2449 			for (m = m_ret; m < &m_ret[npages]; m++) {
2450 				if (m <= mpred &&
2451 				    (req & VM_ALLOC_WIRED) != 0)
2452 					m->ref_count = 0;
2453 				m->oflags = VPO_UNMANAGED;
2454 				m->busy_lock = VPB_UNBUSIED;
2455 				/* Don't change PG_ZERO. */
2456 				vm_page_free_toq(m);
2457 			}
2458 			if (req & VM_ALLOC_WAITFAIL) {
2459 				VM_OBJECT_WUNLOCK(object);
2460 				vm_radix_wait();
2461 				VM_OBJECT_WLOCK(object);
2462 			}
2463 			return (NULL);
2464 		}
2465 		mpred = m;
2466 		if (memattr != VM_MEMATTR_DEFAULT)
2467 			pmap_page_set_memattr(m, memattr);
2468 		pindex++;
2469 	}
2470 	return (m_ret);
2471 }
2472 
2473 /*
2474  * Allocate a physical page that is not intended to be inserted into a VM
2475  * object.
2476  */
2477 vm_page_t
2478 vm_page_alloc_noobj_domain(int domain, int req)
2479 {
2480 	struct vm_domain *vmd;
2481 	vm_page_t m;
2482 	int flags;
2483 
2484 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2485 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2486 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2487 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2488 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2489 	KASSERT((req & ~VPAN_FLAGS) == 0,
2490 	    ("invalid request %#x", req));
2491 
2492 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2493 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2494 	vmd = VM_DOMAIN(domain);
2495 again:
2496 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2497 		m = vm_page_alloc_nofree_domain(domain, req);
2498 		if (m != NULL)
2499 			goto found;
2500 	}
2501 
2502 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2503 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2504 		    M_NOWAIT | M_NOVM);
2505 		if (m != NULL) {
2506 			flags |= PG_PCPU_CACHE;
2507 			goto found;
2508 		}
2509 	}
2510 
2511 	if (vm_domain_allocate(vmd, req, 1)) {
2512 		vm_domain_free_lock(vmd);
2513 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2514 		vm_domain_free_unlock(vmd);
2515 		if (m == NULL) {
2516 			vm_domain_freecnt_inc(vmd, 1);
2517 #if VM_NRESERVLEVEL > 0
2518 			if (vm_reserv_reclaim_inactive(domain))
2519 				goto again;
2520 #endif
2521 		}
2522 	}
2523 	if (m == NULL) {
2524 		if (vm_domain_alloc_fail(vmd, NULL, req))
2525 			goto again;
2526 		return (NULL);
2527 	}
2528 
2529 found:
2530 	vm_page_dequeue(m);
2531 	vm_page_alloc_check(m);
2532 
2533 	/*
2534 	 * Consumers should not rely on a useful default pindex value.
2535 	 */
2536 	m->pindex = 0xdeadc0dedeadc0de;
2537 	m->flags = (m->flags & PG_ZERO) | flags;
2538 	m->a.flags = 0;
2539 	m->oflags = VPO_UNMANAGED;
2540 	m->busy_lock = VPB_UNBUSIED;
2541 	if ((req & VM_ALLOC_WIRED) != 0) {
2542 		vm_wire_add(1);
2543 		m->ref_count = 1;
2544 	}
2545 
2546 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2547 		pmap_zero_page(m);
2548 
2549 	return (m);
2550 }
2551 
2552 #if VM_NRESERVLEVEL > 1
2553 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2554 #elif VM_NRESERVLEVEL > 0
2555 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2556 #else
2557 #define	VM_NOFREE_IMPORT_ORDER	8
2558 #endif
2559 
2560 /*
2561  * Allocate a single NOFREE page.
2562  *
2563  * This routine hands out NOFREE pages from higher-order
2564  * physical memory blocks in order to reduce memory fragmentation.
2565  * When a NOFREE for a given domain chunk is used up,
2566  * the routine will try to fetch a new one from the freelists
2567  * and discard the old one.
2568  */
2569 static vm_page_t
2570 vm_page_alloc_nofree_domain(int domain, int req)
2571 {
2572 	vm_page_t m;
2573 	struct vm_domain *vmd;
2574 	struct vm_nofreeq *nqp;
2575 
2576 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2577 
2578 	vmd = VM_DOMAIN(domain);
2579 	nqp = &vmd->vmd_nofreeq;
2580 	vm_domain_free_lock(vmd);
2581 	if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) {
2582 		if (!vm_domain_allocate(vmd, req,
2583 		    1 << VM_NOFREE_IMPORT_ORDER)) {
2584 			vm_domain_free_unlock(vmd);
2585 			return (NULL);
2586 		}
2587 		nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2588 		    VM_NOFREE_IMPORT_ORDER);
2589 		if (nqp->ma == NULL) {
2590 			vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER);
2591 			vm_domain_free_unlock(vmd);
2592 			return (NULL);
2593 		}
2594 		nqp->offs = 0;
2595 	}
2596 	m = &nqp->ma[nqp->offs++];
2597 	vm_domain_free_unlock(vmd);
2598 	VM_CNT_ADD(v_nofree_count, 1);
2599 
2600 	return (m);
2601 }
2602 
2603 vm_page_t
2604 vm_page_alloc_noobj(int req)
2605 {
2606 	struct vm_domainset_iter di;
2607 	vm_page_t m;
2608 	int domain;
2609 
2610 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2611 	do {
2612 		m = vm_page_alloc_noobj_domain(domain, req);
2613 		if (m != NULL)
2614 			break;
2615 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2616 
2617 	return (m);
2618 }
2619 
2620 vm_page_t
2621 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2622     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2623     vm_memattr_t memattr)
2624 {
2625 	struct vm_domainset_iter di;
2626 	vm_page_t m;
2627 	int domain;
2628 
2629 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2630 	do {
2631 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2632 		    high, alignment, boundary, memattr);
2633 		if (m != NULL)
2634 			break;
2635 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2636 
2637 	return (m);
2638 }
2639 
2640 vm_page_t
2641 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2642     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2643     vm_memattr_t memattr)
2644 {
2645 	vm_page_t m, m_ret;
2646 	u_int flags;
2647 
2648 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2649 	KASSERT((req & ~VPANC_FLAGS) == 0,
2650 	    ("invalid request %#x", req));
2651 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2652 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2653 	    ("invalid request %#x", req));
2654 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2655 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2656 	    ("invalid request %#x", req));
2657 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2658 
2659 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2660 	    low, high, alignment, boundary)) == NULL) {
2661 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2662 			return (NULL);
2663 	}
2664 
2665 	/*
2666 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2667 	 */
2668 	flags = PG_ZERO;
2669 	if ((req & VM_ALLOC_NODUMP) != 0)
2670 		flags |= PG_NODUMP;
2671 	if ((req & VM_ALLOC_WIRED) != 0)
2672 		vm_wire_add(npages);
2673 	for (m = m_ret; m < &m_ret[npages]; m++) {
2674 		vm_page_dequeue(m);
2675 		vm_page_alloc_check(m);
2676 
2677 		/*
2678 		 * Consumers should not rely on a useful default pindex value.
2679 		 */
2680 		m->pindex = 0xdeadc0dedeadc0de;
2681 		m->a.flags = 0;
2682 		m->flags = (m->flags | PG_NODUMP) & flags;
2683 		m->busy_lock = VPB_UNBUSIED;
2684 		if ((req & VM_ALLOC_WIRED) != 0)
2685 			m->ref_count = 1;
2686 		m->a.act_count = 0;
2687 		m->oflags = VPO_UNMANAGED;
2688 
2689 		/*
2690 		 * Zero the page before updating any mappings since the page is
2691 		 * not yet shared with any devices which might require the
2692 		 * non-default memory attribute.  pmap_page_set_memattr()
2693 		 * flushes data caches before returning.
2694 		 */
2695 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2696 			pmap_zero_page(m);
2697 		if (memattr != VM_MEMATTR_DEFAULT)
2698 			pmap_page_set_memattr(m, memattr);
2699 	}
2700 	return (m_ret);
2701 }
2702 
2703 /*
2704  * Check a page that has been freshly dequeued from a freelist.
2705  */
2706 static void
2707 vm_page_alloc_check(vm_page_t m)
2708 {
2709 
2710 	KASSERT(m->object == NULL, ("page %p has object", m));
2711 	KASSERT(m->a.queue == PQ_NONE &&
2712 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2713 	    ("page %p has unexpected queue %d, flags %#x",
2714 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2715 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2716 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2717 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2718 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2719 	    ("page %p has unexpected memattr %d",
2720 	    m, pmap_page_get_memattr(m)));
2721 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2722 	pmap_vm_page_alloc_check(m);
2723 }
2724 
2725 static int
2726 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2727 {
2728 	struct vm_domain *vmd;
2729 	struct vm_pgcache *pgcache;
2730 	int i;
2731 
2732 	pgcache = arg;
2733 	vmd = VM_DOMAIN(pgcache->domain);
2734 
2735 	/*
2736 	 * The page daemon should avoid creating extra memory pressure since its
2737 	 * main purpose is to replenish the store of free pages.
2738 	 */
2739 	if (vmd->vmd_severeset || curproc == pageproc ||
2740 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2741 		return (0);
2742 	domain = vmd->vmd_domain;
2743 	vm_domain_free_lock(vmd);
2744 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2745 	    (vm_page_t *)store);
2746 	vm_domain_free_unlock(vmd);
2747 	if (cnt != i)
2748 		vm_domain_freecnt_inc(vmd, cnt - i);
2749 
2750 	return (i);
2751 }
2752 
2753 static void
2754 vm_page_zone_release(void *arg, void **store, int cnt)
2755 {
2756 	struct vm_domain *vmd;
2757 	struct vm_pgcache *pgcache;
2758 	vm_page_t m;
2759 	int i;
2760 
2761 	pgcache = arg;
2762 	vmd = VM_DOMAIN(pgcache->domain);
2763 	vm_domain_free_lock(vmd);
2764 	for (i = 0; i < cnt; i++) {
2765 		m = (vm_page_t)store[i];
2766 		vm_phys_free_pages(m, 0);
2767 	}
2768 	vm_domain_free_unlock(vmd);
2769 	vm_domain_freecnt_inc(vmd, cnt);
2770 }
2771 
2772 #define	VPSC_ANY	0	/* No restrictions. */
2773 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2774 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2775 
2776 /*
2777  *	vm_page_scan_contig:
2778  *
2779  *	Scan vm_page_array[] between the specified entries "m_start" and
2780  *	"m_end" for a run of contiguous physical pages that satisfy the
2781  *	specified conditions, and return the lowest page in the run.  The
2782  *	specified "alignment" determines the alignment of the lowest physical
2783  *	page in the run.  If the specified "boundary" is non-zero, then the
2784  *	run of physical pages cannot span a physical address that is a
2785  *	multiple of "boundary".
2786  *
2787  *	"m_end" is never dereferenced, so it need not point to a vm_page
2788  *	structure within vm_page_array[].
2789  *
2790  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2791  *	span a hole (or discontiguity) in the physical address space.  Both
2792  *	"alignment" and "boundary" must be a power of two.
2793  */
2794 static vm_page_t
2795 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2796     u_long alignment, vm_paddr_t boundary, int options)
2797 {
2798 	vm_object_t object;
2799 	vm_paddr_t pa;
2800 	vm_page_t m, m_run;
2801 #if VM_NRESERVLEVEL > 0
2802 	int level;
2803 #endif
2804 	int m_inc, order, run_ext, run_len;
2805 
2806 	KASSERT(npages > 0, ("npages is 0"));
2807 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2808 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2809 	m_run = NULL;
2810 	run_len = 0;
2811 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2812 		KASSERT((m->flags & PG_MARKER) == 0,
2813 		    ("page %p is PG_MARKER", m));
2814 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2815 		    ("fictitious page %p has invalid ref count", m));
2816 
2817 		/*
2818 		 * If the current page would be the start of a run, check its
2819 		 * physical address against the end, alignment, and boundary
2820 		 * conditions.  If it doesn't satisfy these conditions, either
2821 		 * terminate the scan or advance to the next page that
2822 		 * satisfies the failed condition.
2823 		 */
2824 		if (run_len == 0) {
2825 			KASSERT(m_run == NULL, ("m_run != NULL"));
2826 			if (m + npages > m_end)
2827 				break;
2828 			pa = VM_PAGE_TO_PHYS(m);
2829 			if (!vm_addr_align_ok(pa, alignment)) {
2830 				m_inc = atop(roundup2(pa, alignment) - pa);
2831 				continue;
2832 			}
2833 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2834 				m_inc = atop(roundup2(pa, boundary) - pa);
2835 				continue;
2836 			}
2837 		} else
2838 			KASSERT(m_run != NULL, ("m_run == NULL"));
2839 
2840 retry:
2841 		m_inc = 1;
2842 		if (vm_page_wired(m))
2843 			run_ext = 0;
2844 #if VM_NRESERVLEVEL > 0
2845 		else if ((level = vm_reserv_level(m)) >= 0 &&
2846 		    (options & VPSC_NORESERV) != 0) {
2847 			run_ext = 0;
2848 			/* Advance to the end of the reservation. */
2849 			pa = VM_PAGE_TO_PHYS(m);
2850 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2851 			    pa);
2852 		}
2853 #endif
2854 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2855 			/*
2856 			 * The page is considered eligible for relocation if
2857 			 * and only if it could be laundered or reclaimed by
2858 			 * the page daemon.
2859 			 */
2860 			VM_OBJECT_RLOCK(object);
2861 			if (object != m->object) {
2862 				VM_OBJECT_RUNLOCK(object);
2863 				goto retry;
2864 			}
2865 			/* Don't care: PG_NODUMP, PG_ZERO. */
2866 			if ((object->flags & OBJ_SWAP) == 0 &&
2867 			    object->type != OBJT_VNODE) {
2868 				run_ext = 0;
2869 #if VM_NRESERVLEVEL > 0
2870 			} else if ((options & VPSC_NOSUPER) != 0 &&
2871 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2872 				run_ext = 0;
2873 				/* Advance to the end of the superpage. */
2874 				pa = VM_PAGE_TO_PHYS(m);
2875 				m_inc = atop(roundup2(pa + 1,
2876 				    vm_reserv_size(level)) - pa);
2877 #endif
2878 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2879 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2880 				/*
2881 				 * The page is allocated but eligible for
2882 				 * relocation.  Extend the current run by one
2883 				 * page.
2884 				 */
2885 				KASSERT(pmap_page_get_memattr(m) ==
2886 				    VM_MEMATTR_DEFAULT,
2887 				    ("page %p has an unexpected memattr", m));
2888 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2889 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2890 				    ("page %p has unexpected oflags", m));
2891 				/* Don't care: PGA_NOSYNC. */
2892 				run_ext = 1;
2893 			} else
2894 				run_ext = 0;
2895 			VM_OBJECT_RUNLOCK(object);
2896 #if VM_NRESERVLEVEL > 0
2897 		} else if (level >= 0) {
2898 			/*
2899 			 * The page is reserved but not yet allocated.  In
2900 			 * other words, it is still free.  Extend the current
2901 			 * run by one page.
2902 			 */
2903 			run_ext = 1;
2904 #endif
2905 		} else if ((order = m->order) < VM_NFREEORDER) {
2906 			/*
2907 			 * The page is enqueued in the physical memory
2908 			 * allocator's free page queues.  Moreover, it is the
2909 			 * first page in a power-of-two-sized run of
2910 			 * contiguous free pages.  Add these pages to the end
2911 			 * of the current run, and jump ahead.
2912 			 */
2913 			run_ext = 1 << order;
2914 			m_inc = 1 << order;
2915 		} else {
2916 			/*
2917 			 * Skip the page for one of the following reasons: (1)
2918 			 * It is enqueued in the physical memory allocator's
2919 			 * free page queues.  However, it is not the first
2920 			 * page in a run of contiguous free pages.  (This case
2921 			 * rarely occurs because the scan is performed in
2922 			 * ascending order.) (2) It is not reserved, and it is
2923 			 * transitioning from free to allocated.  (Conversely,
2924 			 * the transition from allocated to free for managed
2925 			 * pages is blocked by the page busy lock.) (3) It is
2926 			 * allocated but not contained by an object and not
2927 			 * wired, e.g., allocated by Xen's balloon driver.
2928 			 */
2929 			run_ext = 0;
2930 		}
2931 
2932 		/*
2933 		 * Extend or reset the current run of pages.
2934 		 */
2935 		if (run_ext > 0) {
2936 			if (run_len == 0)
2937 				m_run = m;
2938 			run_len += run_ext;
2939 		} else {
2940 			if (run_len > 0) {
2941 				m_run = NULL;
2942 				run_len = 0;
2943 			}
2944 		}
2945 	}
2946 	if (run_len >= npages)
2947 		return (m_run);
2948 	return (NULL);
2949 }
2950 
2951 /*
2952  *	vm_page_reclaim_run:
2953  *
2954  *	Try to relocate each of the allocated virtual pages within the
2955  *	specified run of physical pages to a new physical address.  Free the
2956  *	physical pages underlying the relocated virtual pages.  A virtual page
2957  *	is relocatable if and only if it could be laundered or reclaimed by
2958  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2959  *	physical address above "high".
2960  *
2961  *	Returns 0 if every physical page within the run was already free or
2962  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2963  *	value indicating why the last attempt to relocate a virtual page was
2964  *	unsuccessful.
2965  *
2966  *	"req_class" must be an allocation class.
2967  */
2968 static int
2969 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2970     vm_paddr_t high)
2971 {
2972 	struct vm_domain *vmd;
2973 	struct spglist free;
2974 	vm_object_t object;
2975 	vm_paddr_t pa;
2976 	vm_page_t m, m_end, m_new;
2977 	int error, order, req;
2978 
2979 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2980 	    ("req_class is not an allocation class"));
2981 	SLIST_INIT(&free);
2982 	error = 0;
2983 	m = m_run;
2984 	m_end = m_run + npages;
2985 	for (; error == 0 && m < m_end; m++) {
2986 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2987 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2988 
2989 		/*
2990 		 * Racily check for wirings.  Races are handled once the object
2991 		 * lock is held and the page is unmapped.
2992 		 */
2993 		if (vm_page_wired(m))
2994 			error = EBUSY;
2995 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2996 			/*
2997 			 * The page is relocated if and only if it could be
2998 			 * laundered or reclaimed by the page daemon.
2999 			 */
3000 			VM_OBJECT_WLOCK(object);
3001 			/* Don't care: PG_NODUMP, PG_ZERO. */
3002 			if (m->object != object ||
3003 			    ((object->flags & OBJ_SWAP) == 0 &&
3004 			    object->type != OBJT_VNODE))
3005 				error = EINVAL;
3006 			else if (object->memattr != VM_MEMATTR_DEFAULT)
3007 				error = EINVAL;
3008 			else if (vm_page_queue(m) != PQ_NONE &&
3009 			    vm_page_tryxbusy(m) != 0) {
3010 				if (vm_page_wired(m)) {
3011 					vm_page_xunbusy(m);
3012 					error = EBUSY;
3013 					goto unlock;
3014 				}
3015 				KASSERT(pmap_page_get_memattr(m) ==
3016 				    VM_MEMATTR_DEFAULT,
3017 				    ("page %p has an unexpected memattr", m));
3018 				KASSERT(m->oflags == 0,
3019 				    ("page %p has unexpected oflags", m));
3020 				/* Don't care: PGA_NOSYNC. */
3021 				if (!vm_page_none_valid(m)) {
3022 					/*
3023 					 * First, try to allocate a new page
3024 					 * that is above "high".  Failing
3025 					 * that, try to allocate a new page
3026 					 * that is below "m_run".  Allocate
3027 					 * the new page between the end of
3028 					 * "m_run" and "high" only as a last
3029 					 * resort.
3030 					 */
3031 					req = req_class;
3032 					if ((m->flags & PG_NODUMP) != 0)
3033 						req |= VM_ALLOC_NODUMP;
3034 					if (trunc_page(high) !=
3035 					    ~(vm_paddr_t)PAGE_MASK) {
3036 						m_new =
3037 						    vm_page_alloc_noobj_contig(
3038 						    req, 1, round_page(high),
3039 						    ~(vm_paddr_t)0, PAGE_SIZE,
3040 						    0, VM_MEMATTR_DEFAULT);
3041 					} else
3042 						m_new = NULL;
3043 					if (m_new == NULL) {
3044 						pa = VM_PAGE_TO_PHYS(m_run);
3045 						m_new =
3046 						    vm_page_alloc_noobj_contig(
3047 						    req, 1, 0, pa - 1,
3048 						    PAGE_SIZE, 0,
3049 						    VM_MEMATTR_DEFAULT);
3050 					}
3051 					if (m_new == NULL) {
3052 						pa += ptoa(npages);
3053 						m_new =
3054 						    vm_page_alloc_noobj_contig(
3055 						    req, 1, pa, high, PAGE_SIZE,
3056 						    0, VM_MEMATTR_DEFAULT);
3057 					}
3058 					if (m_new == NULL) {
3059 						vm_page_xunbusy(m);
3060 						error = ENOMEM;
3061 						goto unlock;
3062 					}
3063 
3064 					/*
3065 					 * Unmap the page and check for new
3066 					 * wirings that may have been acquired
3067 					 * through a pmap lookup.
3068 					 */
3069 					if (object->ref_count != 0 &&
3070 					    !vm_page_try_remove_all(m)) {
3071 						vm_page_xunbusy(m);
3072 						vm_page_free(m_new);
3073 						error = EBUSY;
3074 						goto unlock;
3075 					}
3076 
3077 					/*
3078 					 * Replace "m" with the new page.  For
3079 					 * vm_page_replace(), "m" must be busy
3080 					 * and dequeued.  Finally, change "m"
3081 					 * as if vm_page_free() was called.
3082 					 */
3083 					m_new->a.flags = m->a.flags &
3084 					    ~PGA_QUEUE_STATE_MASK;
3085 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3086 					    ("page %p is managed", m_new));
3087 					m_new->oflags = 0;
3088 					pmap_copy_page(m, m_new);
3089 					m_new->valid = m->valid;
3090 					m_new->dirty = m->dirty;
3091 					m->flags &= ~PG_ZERO;
3092 					vm_page_dequeue(m);
3093 					if (vm_page_replace_hold(m_new, object,
3094 					    m->pindex, m) &&
3095 					    vm_page_free_prep(m))
3096 						SLIST_INSERT_HEAD(&free, m,
3097 						    plinks.s.ss);
3098 
3099 					/*
3100 					 * The new page must be deactivated
3101 					 * before the object is unlocked.
3102 					 */
3103 					vm_page_deactivate(m_new);
3104 				} else {
3105 					m->flags &= ~PG_ZERO;
3106 					vm_page_dequeue(m);
3107 					if (vm_page_free_prep(m))
3108 						SLIST_INSERT_HEAD(&free, m,
3109 						    plinks.s.ss);
3110 					KASSERT(m->dirty == 0,
3111 					    ("page %p is dirty", m));
3112 				}
3113 			} else
3114 				error = EBUSY;
3115 unlock:
3116 			VM_OBJECT_WUNLOCK(object);
3117 		} else {
3118 			MPASS(vm_page_domain(m) == domain);
3119 			vmd = VM_DOMAIN(domain);
3120 			vm_domain_free_lock(vmd);
3121 			order = m->order;
3122 			if (order < VM_NFREEORDER) {
3123 				/*
3124 				 * The page is enqueued in the physical memory
3125 				 * allocator's free page queues.  Moreover, it
3126 				 * is the first page in a power-of-two-sized
3127 				 * run of contiguous free pages.  Jump ahead
3128 				 * to the last page within that run, and
3129 				 * continue from there.
3130 				 */
3131 				m += (1 << order) - 1;
3132 			}
3133 #if VM_NRESERVLEVEL > 0
3134 			else if (vm_reserv_is_page_free(m))
3135 				order = 0;
3136 #endif
3137 			vm_domain_free_unlock(vmd);
3138 			if (order == VM_NFREEORDER)
3139 				error = EINVAL;
3140 		}
3141 	}
3142 	if ((m = SLIST_FIRST(&free)) != NULL) {
3143 		int cnt;
3144 
3145 		vmd = VM_DOMAIN(domain);
3146 		cnt = 0;
3147 		vm_domain_free_lock(vmd);
3148 		do {
3149 			MPASS(vm_page_domain(m) == domain);
3150 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3151 			vm_phys_free_pages(m, 0);
3152 			cnt++;
3153 		} while ((m = SLIST_FIRST(&free)) != NULL);
3154 		vm_domain_free_unlock(vmd);
3155 		vm_domain_freecnt_inc(vmd, cnt);
3156 	}
3157 	return (error);
3158 }
3159 
3160 #define	NRUNS	16
3161 
3162 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3163 
3164 #define	MIN_RECLAIM	8
3165 
3166 /*
3167  *	vm_page_reclaim_contig:
3168  *
3169  *	Reclaim allocated, contiguous physical memory satisfying the specified
3170  *	conditions by relocating the virtual pages using that physical memory.
3171  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3172  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3173  *	currently able to reclaim the requested number of pages.  Since
3174  *	relocation requires the allocation of physical pages, reclamation may
3175  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3176  *	fails in this manner, callers are expected to perform vm_wait() before
3177  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3178  *
3179  *	The caller must always specify an allocation class through "req".
3180  *
3181  *	allocation classes:
3182  *	VM_ALLOC_NORMAL		normal process request
3183  *	VM_ALLOC_SYSTEM		system *really* needs a page
3184  *	VM_ALLOC_INTERRUPT	interrupt time request
3185  *
3186  *	The optional allocation flags are ignored.
3187  *
3188  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3189  *	must be a power of two.
3190  */
3191 int
3192 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3193     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3194     int desired_runs)
3195 {
3196 	struct vm_domain *vmd;
3197 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3198 	u_long count, minalign, reclaimed;
3199 	int error, i, min_reclaim, nruns, options, req_class;
3200 	int segind, start_segind;
3201 	int ret;
3202 
3203 	KASSERT(npages > 0, ("npages is 0"));
3204 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3205 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3206 
3207 	ret = ENOMEM;
3208 
3209 	/*
3210 	 * If the caller wants to reclaim multiple runs, try to allocate
3211 	 * space to store the runs.  If that fails, fall back to the old
3212 	 * behavior of just reclaiming MIN_RECLAIM pages.
3213 	 */
3214 	if (desired_runs > 1)
3215 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3216 		    M_TEMP, M_NOWAIT);
3217 	else
3218 		m_runs = NULL;
3219 
3220 	if (m_runs == NULL) {
3221 		m_runs = _m_runs;
3222 		nruns = NRUNS;
3223 	} else {
3224 		nruns = NRUNS + desired_runs - 1;
3225 	}
3226 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3227 
3228 	/*
3229 	 * The caller will attempt an allocation after some runs have been
3230 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3231 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3232 	 * power of two or maximum chunk size, and ensure that each run is
3233 	 * suitably aligned.
3234 	 */
3235 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3236 	npages = roundup2(npages, minalign);
3237 	if (alignment < ptoa(minalign))
3238 		alignment = ptoa(minalign);
3239 
3240 	/*
3241 	 * The page daemon is allowed to dig deeper into the free page list.
3242 	 */
3243 	req_class = req & VM_ALLOC_CLASS_MASK;
3244 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3245 		req_class = VM_ALLOC_SYSTEM;
3246 
3247 	start_segind = vm_phys_lookup_segind(low);
3248 
3249 	/*
3250 	 * Return if the number of free pages cannot satisfy the requested
3251 	 * allocation.
3252 	 */
3253 	vmd = VM_DOMAIN(domain);
3254 	count = vmd->vmd_free_count;
3255 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3256 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3257 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3258 		goto done;
3259 
3260 	/*
3261 	 * Scan up to three times, relaxing the restrictions ("options") on
3262 	 * the reclamation of reservations and superpages each time.
3263 	 */
3264 	for (options = VPSC_NORESERV;;) {
3265 		bool phys_range_exists = false;
3266 
3267 		/*
3268 		 * Find the highest runs that satisfy the given constraints
3269 		 * and restrictions, and record them in "m_runs".
3270 		 */
3271 		count = 0;
3272 		segind = start_segind;
3273 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3274 		    npages, low, high)) != -1) {
3275 			phys_range_exists = true;
3276 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3277 			    bounds[1], alignment, boundary, options))) {
3278 				bounds[0] = m_run + npages;
3279 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3280 				count++;
3281 			}
3282 			segind++;
3283 		}
3284 
3285 		if (!phys_range_exists) {
3286 			ret = ERANGE;
3287 			goto done;
3288 		}
3289 
3290 		/*
3291 		 * Reclaim the highest runs in LIFO (descending) order until
3292 		 * the number of reclaimed pages, "reclaimed", is at least
3293 		 * "min_reclaim".  Reset "reclaimed" each time because each
3294 		 * reclamation is idempotent, and runs will (likely) recur
3295 		 * from one scan to the next as restrictions are relaxed.
3296 		 */
3297 		reclaimed = 0;
3298 		for (i = 0; count > 0 && i < nruns; i++) {
3299 			count--;
3300 			m_run = m_runs[RUN_INDEX(count, nruns)];
3301 			error = vm_page_reclaim_run(req_class, domain, npages,
3302 			    m_run, high);
3303 			if (error == 0) {
3304 				reclaimed += npages;
3305 				if (reclaimed >= min_reclaim) {
3306 					ret = 0;
3307 					goto done;
3308 				}
3309 			}
3310 		}
3311 
3312 		/*
3313 		 * Either relax the restrictions on the next scan or return if
3314 		 * the last scan had no restrictions.
3315 		 */
3316 		if (options == VPSC_NORESERV)
3317 			options = VPSC_NOSUPER;
3318 		else if (options == VPSC_NOSUPER)
3319 			options = VPSC_ANY;
3320 		else if (options == VPSC_ANY) {
3321 			if (reclaimed != 0)
3322 				ret = 0;
3323 			goto done;
3324 		}
3325 	}
3326 done:
3327 	if (m_runs != _m_runs)
3328 		free(m_runs, M_TEMP);
3329 	return (ret);
3330 }
3331 
3332 int
3333 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3334     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3335 {
3336 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3337 	    alignment, boundary, 1));
3338 }
3339 
3340 int
3341 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3342     u_long alignment, vm_paddr_t boundary)
3343 {
3344 	struct vm_domainset_iter di;
3345 	int domain, ret, status;
3346 
3347 	ret = ERANGE;
3348 
3349 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3350 	do {
3351 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3352 		    high, alignment, boundary);
3353 		if (status == 0)
3354 			return (0);
3355 		else if (status == ERANGE)
3356 			vm_domainset_iter_ignore(&di, domain);
3357 		else {
3358 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3359 			    "from vm_page_reclaim_contig_domain()", status));
3360 			ret = ENOMEM;
3361 		}
3362 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3363 
3364 	return (ret);
3365 }
3366 
3367 /*
3368  * Set the domain in the appropriate page level domainset.
3369  */
3370 void
3371 vm_domain_set(struct vm_domain *vmd)
3372 {
3373 
3374 	mtx_lock(&vm_domainset_lock);
3375 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3376 		vmd->vmd_minset = 1;
3377 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3378 	}
3379 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3380 		vmd->vmd_severeset = 1;
3381 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3382 	}
3383 	mtx_unlock(&vm_domainset_lock);
3384 }
3385 
3386 /*
3387  * Clear the domain from the appropriate page level domainset.
3388  */
3389 void
3390 vm_domain_clear(struct vm_domain *vmd)
3391 {
3392 
3393 	mtx_lock(&vm_domainset_lock);
3394 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3395 		vmd->vmd_minset = 0;
3396 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3397 		if (vm_min_waiters != 0) {
3398 			vm_min_waiters = 0;
3399 			wakeup(&vm_min_domains);
3400 		}
3401 	}
3402 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3403 		vmd->vmd_severeset = 0;
3404 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3405 		if (vm_severe_waiters != 0) {
3406 			vm_severe_waiters = 0;
3407 			wakeup(&vm_severe_domains);
3408 		}
3409 	}
3410 
3411 	/*
3412 	 * If pageout daemon needs pages, then tell it that there are
3413 	 * some free.
3414 	 */
3415 	if (vmd->vmd_pageout_pages_needed &&
3416 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3417 		wakeup(&vmd->vmd_pageout_pages_needed);
3418 		vmd->vmd_pageout_pages_needed = 0;
3419 	}
3420 
3421 	/* See comments in vm_wait_doms(). */
3422 	if (vm_pageproc_waiters) {
3423 		vm_pageproc_waiters = 0;
3424 		wakeup(&vm_pageproc_waiters);
3425 	}
3426 	mtx_unlock(&vm_domainset_lock);
3427 }
3428 
3429 /*
3430  * Wait for free pages to exceed the min threshold globally.
3431  */
3432 void
3433 vm_wait_min(void)
3434 {
3435 
3436 	mtx_lock(&vm_domainset_lock);
3437 	while (vm_page_count_min()) {
3438 		vm_min_waiters++;
3439 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3440 	}
3441 	mtx_unlock(&vm_domainset_lock);
3442 }
3443 
3444 /*
3445  * Wait for free pages to exceed the severe threshold globally.
3446  */
3447 void
3448 vm_wait_severe(void)
3449 {
3450 
3451 	mtx_lock(&vm_domainset_lock);
3452 	while (vm_page_count_severe()) {
3453 		vm_severe_waiters++;
3454 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3455 		    "vmwait", 0);
3456 	}
3457 	mtx_unlock(&vm_domainset_lock);
3458 }
3459 
3460 u_int
3461 vm_wait_count(void)
3462 {
3463 
3464 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3465 }
3466 
3467 int
3468 vm_wait_doms(const domainset_t *wdoms, int mflags)
3469 {
3470 	int error;
3471 
3472 	error = 0;
3473 
3474 	/*
3475 	 * We use racey wakeup synchronization to avoid expensive global
3476 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3477 	 * To handle this, we only sleep for one tick in this instance.  It
3478 	 * is expected that most allocations for the pageproc will come from
3479 	 * kmem or vm_page_grab* which will use the more specific and
3480 	 * race-free vm_wait_domain().
3481 	 */
3482 	if (curproc == pageproc) {
3483 		mtx_lock(&vm_domainset_lock);
3484 		vm_pageproc_waiters++;
3485 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3486 		    PVM | PDROP | mflags, "pageprocwait", 1);
3487 	} else {
3488 		/*
3489 		 * XXX Ideally we would wait only until the allocation could
3490 		 * be satisfied.  This condition can cause new allocators to
3491 		 * consume all freed pages while old allocators wait.
3492 		 */
3493 		mtx_lock(&vm_domainset_lock);
3494 		if (vm_page_count_min_set(wdoms)) {
3495 			if (pageproc == NULL)
3496 				panic("vm_wait in early boot");
3497 			vm_min_waiters++;
3498 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3499 			    PVM | PDROP | mflags, "vmwait", 0);
3500 		} else
3501 			mtx_unlock(&vm_domainset_lock);
3502 	}
3503 	return (error);
3504 }
3505 
3506 /*
3507  *	vm_wait_domain:
3508  *
3509  *	Sleep until free pages are available for allocation.
3510  *	- Called in various places after failed memory allocations.
3511  */
3512 void
3513 vm_wait_domain(int domain)
3514 {
3515 	struct vm_domain *vmd;
3516 	domainset_t wdom;
3517 
3518 	vmd = VM_DOMAIN(domain);
3519 	vm_domain_free_assert_unlocked(vmd);
3520 
3521 	if (curproc == pageproc) {
3522 		mtx_lock(&vm_domainset_lock);
3523 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3524 			vmd->vmd_pageout_pages_needed = 1;
3525 			msleep(&vmd->vmd_pageout_pages_needed,
3526 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3527 		} else
3528 			mtx_unlock(&vm_domainset_lock);
3529 	} else {
3530 		DOMAINSET_ZERO(&wdom);
3531 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3532 		vm_wait_doms(&wdom, 0);
3533 	}
3534 }
3535 
3536 static int
3537 vm_wait_flags(vm_object_t obj, int mflags)
3538 {
3539 	struct domainset *d;
3540 
3541 	d = NULL;
3542 
3543 	/*
3544 	 * Carefully fetch pointers only once: the struct domainset
3545 	 * itself is ummutable but the pointer might change.
3546 	 */
3547 	if (obj != NULL)
3548 		d = obj->domain.dr_policy;
3549 	if (d == NULL)
3550 		d = curthread->td_domain.dr_policy;
3551 
3552 	return (vm_wait_doms(&d->ds_mask, mflags));
3553 }
3554 
3555 /*
3556  *	vm_wait:
3557  *
3558  *	Sleep until free pages are available for allocation in the
3559  *	affinity domains of the obj.  If obj is NULL, the domain set
3560  *	for the calling thread is used.
3561  *	Called in various places after failed memory allocations.
3562  */
3563 void
3564 vm_wait(vm_object_t obj)
3565 {
3566 	(void)vm_wait_flags(obj, 0);
3567 }
3568 
3569 int
3570 vm_wait_intr(vm_object_t obj)
3571 {
3572 	return (vm_wait_flags(obj, PCATCH));
3573 }
3574 
3575 /*
3576  *	vm_domain_alloc_fail:
3577  *
3578  *	Called when a page allocation function fails.  Informs the
3579  *	pagedaemon and performs the requested wait.  Requires the
3580  *	domain_free and object lock on entry.  Returns with the
3581  *	object lock held and free lock released.  Returns an error when
3582  *	retry is necessary.
3583  *
3584  */
3585 static int
3586 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3587 {
3588 
3589 	vm_domain_free_assert_unlocked(vmd);
3590 
3591 	atomic_add_int(&vmd->vmd_pageout_deficit,
3592 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3593 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3594 		if (object != NULL)
3595 			VM_OBJECT_WUNLOCK(object);
3596 		vm_wait_domain(vmd->vmd_domain);
3597 		if (object != NULL)
3598 			VM_OBJECT_WLOCK(object);
3599 		if (req & VM_ALLOC_WAITOK)
3600 			return (EAGAIN);
3601 	}
3602 
3603 	return (0);
3604 }
3605 
3606 /*
3607  *	vm_waitpfault:
3608  *
3609  *	Sleep until free pages are available for allocation.
3610  *	- Called only in vm_fault so that processes page faulting
3611  *	  can be easily tracked.
3612  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3613  *	  processes will be able to grab memory first.  Do not change
3614  *	  this balance without careful testing first.
3615  */
3616 void
3617 vm_waitpfault(struct domainset *dset, int timo)
3618 {
3619 
3620 	/*
3621 	 * XXX Ideally we would wait only until the allocation could
3622 	 * be satisfied.  This condition can cause new allocators to
3623 	 * consume all freed pages while old allocators wait.
3624 	 */
3625 	mtx_lock(&vm_domainset_lock);
3626 	if (vm_page_count_min_set(&dset->ds_mask)) {
3627 		vm_min_waiters++;
3628 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3629 		    "pfault", timo);
3630 	} else
3631 		mtx_unlock(&vm_domainset_lock);
3632 }
3633 
3634 static struct vm_pagequeue *
3635 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3636 {
3637 
3638 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3639 }
3640 
3641 #ifdef INVARIANTS
3642 static struct vm_pagequeue *
3643 vm_page_pagequeue(vm_page_t m)
3644 {
3645 
3646 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3647 }
3648 #endif
3649 
3650 static __always_inline bool
3651 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3652 {
3653 	vm_page_astate_t tmp;
3654 
3655 	tmp = *old;
3656 	do {
3657 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3658 			return (true);
3659 		counter_u64_add(pqstate_commit_retries, 1);
3660 	} while (old->_bits == tmp._bits);
3661 
3662 	return (false);
3663 }
3664 
3665 /*
3666  * Do the work of committing a queue state update that moves the page out of
3667  * its current queue.
3668  */
3669 static bool
3670 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3671     vm_page_astate_t *old, vm_page_astate_t new)
3672 {
3673 	vm_page_t next;
3674 
3675 	vm_pagequeue_assert_locked(pq);
3676 	KASSERT(vm_page_pagequeue(m) == pq,
3677 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3678 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3679 	    ("%s: invalid queue indices %d %d",
3680 	    __func__, old->queue, new.queue));
3681 
3682 	/*
3683 	 * Once the queue index of the page changes there is nothing
3684 	 * synchronizing with further updates to the page's physical
3685 	 * queue state.  Therefore we must speculatively remove the page
3686 	 * from the queue now and be prepared to roll back if the queue
3687 	 * state update fails.  If the page is not physically enqueued then
3688 	 * we just update its queue index.
3689 	 */
3690 	if ((old->flags & PGA_ENQUEUED) != 0) {
3691 		new.flags &= ~PGA_ENQUEUED;
3692 		next = TAILQ_NEXT(m, plinks.q);
3693 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3694 		vm_pagequeue_cnt_dec(pq);
3695 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3696 			if (next == NULL)
3697 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3698 			else
3699 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3700 			vm_pagequeue_cnt_inc(pq);
3701 			return (false);
3702 		} else {
3703 			return (true);
3704 		}
3705 	} else {
3706 		return (vm_page_pqstate_fcmpset(m, old, new));
3707 	}
3708 }
3709 
3710 static bool
3711 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3712     vm_page_astate_t new)
3713 {
3714 	struct vm_pagequeue *pq;
3715 	vm_page_astate_t as;
3716 	bool ret;
3717 
3718 	pq = _vm_page_pagequeue(m, old->queue);
3719 
3720 	/*
3721 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3722 	 * corresponding page queue lock is held.
3723 	 */
3724 	vm_pagequeue_lock(pq);
3725 	as = vm_page_astate_load(m);
3726 	if (__predict_false(as._bits != old->_bits)) {
3727 		*old = as;
3728 		ret = false;
3729 	} else {
3730 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3731 	}
3732 	vm_pagequeue_unlock(pq);
3733 	return (ret);
3734 }
3735 
3736 /*
3737  * Commit a queue state update that enqueues or requeues a page.
3738  */
3739 static bool
3740 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3741     vm_page_astate_t *old, vm_page_astate_t new)
3742 {
3743 	struct vm_domain *vmd;
3744 
3745 	vm_pagequeue_assert_locked(pq);
3746 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3747 	    ("%s: invalid queue indices %d %d",
3748 	    __func__, old->queue, new.queue));
3749 
3750 	new.flags |= PGA_ENQUEUED;
3751 	if (!vm_page_pqstate_fcmpset(m, old, new))
3752 		return (false);
3753 
3754 	if ((old->flags & PGA_ENQUEUED) != 0)
3755 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3756 	else
3757 		vm_pagequeue_cnt_inc(pq);
3758 
3759 	/*
3760 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3761 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3762 	 * applied, even if it was set first.
3763 	 */
3764 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3765 		vmd = vm_pagequeue_domain(m);
3766 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3767 		    ("%s: invalid page queue for page %p", __func__, m));
3768 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3769 	} else {
3770 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3771 	}
3772 	return (true);
3773 }
3774 
3775 /*
3776  * Commit a queue state update that encodes a request for a deferred queue
3777  * operation.
3778  */
3779 static bool
3780 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3781     vm_page_astate_t new)
3782 {
3783 
3784 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3785 	    ("%s: invalid state, queue %d flags %x",
3786 	    __func__, new.queue, new.flags));
3787 
3788 	if (old->_bits != new._bits &&
3789 	    !vm_page_pqstate_fcmpset(m, old, new))
3790 		return (false);
3791 	vm_page_pqbatch_submit(m, new.queue);
3792 	return (true);
3793 }
3794 
3795 /*
3796  * A generic queue state update function.  This handles more cases than the
3797  * specialized functions above.
3798  */
3799 bool
3800 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3801 {
3802 
3803 	if (old->_bits == new._bits)
3804 		return (true);
3805 
3806 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3807 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3808 			return (false);
3809 		if (new.queue != PQ_NONE)
3810 			vm_page_pqbatch_submit(m, new.queue);
3811 	} else {
3812 		if (!vm_page_pqstate_fcmpset(m, old, new))
3813 			return (false);
3814 		if (new.queue != PQ_NONE &&
3815 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3816 			vm_page_pqbatch_submit(m, new.queue);
3817 	}
3818 	return (true);
3819 }
3820 
3821 /*
3822  * Apply deferred queue state updates to a page.
3823  */
3824 static inline void
3825 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3826 {
3827 	vm_page_astate_t new, old;
3828 
3829 	CRITICAL_ASSERT(curthread);
3830 	vm_pagequeue_assert_locked(pq);
3831 	KASSERT(queue < PQ_COUNT,
3832 	    ("%s: invalid queue index %d", __func__, queue));
3833 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3834 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3835 
3836 	for (old = vm_page_astate_load(m);;) {
3837 		if (__predict_false(old.queue != queue ||
3838 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3839 			counter_u64_add(queue_nops, 1);
3840 			break;
3841 		}
3842 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3843 		    ("%s: page %p is unmanaged", __func__, m));
3844 
3845 		new = old;
3846 		if ((old.flags & PGA_DEQUEUE) != 0) {
3847 			new.flags &= ~PGA_QUEUE_OP_MASK;
3848 			new.queue = PQ_NONE;
3849 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3850 			    m, &old, new))) {
3851 				counter_u64_add(queue_ops, 1);
3852 				break;
3853 			}
3854 		} else {
3855 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3856 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3857 			    m, &old, new))) {
3858 				counter_u64_add(queue_ops, 1);
3859 				break;
3860 			}
3861 		}
3862 	}
3863 }
3864 
3865 static void
3866 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3867     uint8_t queue)
3868 {
3869 	int i;
3870 
3871 	for (i = 0; i < bq->bq_cnt; i++)
3872 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3873 	vm_batchqueue_init(bq);
3874 }
3875 
3876 /*
3877  *	vm_page_pqbatch_submit:		[ internal use only ]
3878  *
3879  *	Enqueue a page in the specified page queue's batched work queue.
3880  *	The caller must have encoded the requested operation in the page
3881  *	structure's a.flags field.
3882  */
3883 void
3884 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3885 {
3886 	struct vm_batchqueue *bq;
3887 	struct vm_pagequeue *pq;
3888 	int domain, slots_remaining;
3889 
3890 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3891 
3892 	domain = vm_page_domain(m);
3893 	critical_enter();
3894 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3895 	slots_remaining = vm_batchqueue_insert(bq, m);
3896 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3897 		/* keep building the bq */
3898 		critical_exit();
3899 		return;
3900 	} else if (slots_remaining > 0 ) {
3901 		/* Try to process the bq if we can get the lock */
3902 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3903 		if (vm_pagequeue_trylock(pq)) {
3904 			vm_pqbatch_process(pq, bq, queue);
3905 			vm_pagequeue_unlock(pq);
3906 		}
3907 		critical_exit();
3908 		return;
3909 	}
3910 	critical_exit();
3911 
3912 	/* if we make it here, the bq is full so wait for the lock */
3913 
3914 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3915 	vm_pagequeue_lock(pq);
3916 	critical_enter();
3917 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3918 	vm_pqbatch_process(pq, bq, queue);
3919 	vm_pqbatch_process_page(pq, m, queue);
3920 	vm_pagequeue_unlock(pq);
3921 	critical_exit();
3922 }
3923 
3924 /*
3925  *	vm_page_pqbatch_drain:		[ internal use only ]
3926  *
3927  *	Force all per-CPU page queue batch queues to be drained.  This is
3928  *	intended for use in severe memory shortages, to ensure that pages
3929  *	do not remain stuck in the batch queues.
3930  */
3931 void
3932 vm_page_pqbatch_drain(void)
3933 {
3934 	struct thread *td;
3935 	struct vm_domain *vmd;
3936 	struct vm_pagequeue *pq;
3937 	int cpu, domain, queue;
3938 
3939 	td = curthread;
3940 	CPU_FOREACH(cpu) {
3941 		thread_lock(td);
3942 		sched_bind(td, cpu);
3943 		thread_unlock(td);
3944 
3945 		for (domain = 0; domain < vm_ndomains; domain++) {
3946 			vmd = VM_DOMAIN(domain);
3947 			for (queue = 0; queue < PQ_COUNT; queue++) {
3948 				pq = &vmd->vmd_pagequeues[queue];
3949 				vm_pagequeue_lock(pq);
3950 				critical_enter();
3951 				vm_pqbatch_process(pq,
3952 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3953 				critical_exit();
3954 				vm_pagequeue_unlock(pq);
3955 			}
3956 		}
3957 	}
3958 	thread_lock(td);
3959 	sched_unbind(td);
3960 	thread_unlock(td);
3961 }
3962 
3963 /*
3964  *	vm_page_dequeue_deferred:	[ internal use only ]
3965  *
3966  *	Request removal of the given page from its current page
3967  *	queue.  Physical removal from the queue may be deferred
3968  *	indefinitely.
3969  */
3970 void
3971 vm_page_dequeue_deferred(vm_page_t m)
3972 {
3973 	vm_page_astate_t new, old;
3974 
3975 	old = vm_page_astate_load(m);
3976 	do {
3977 		if (old.queue == PQ_NONE) {
3978 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3979 			    ("%s: page %p has unexpected queue state",
3980 			    __func__, m));
3981 			break;
3982 		}
3983 		new = old;
3984 		new.flags |= PGA_DEQUEUE;
3985 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3986 }
3987 
3988 /*
3989  *	vm_page_dequeue:
3990  *
3991  *	Remove the page from whichever page queue it's in, if any, before
3992  *	returning.
3993  */
3994 void
3995 vm_page_dequeue(vm_page_t m)
3996 {
3997 	vm_page_astate_t new, old;
3998 
3999 	old = vm_page_astate_load(m);
4000 	do {
4001 		if (old.queue == PQ_NONE) {
4002 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4003 			    ("%s: page %p has unexpected queue state",
4004 			    __func__, m));
4005 			break;
4006 		}
4007 		new = old;
4008 		new.flags &= ~PGA_QUEUE_OP_MASK;
4009 		new.queue = PQ_NONE;
4010 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
4011 
4012 }
4013 
4014 /*
4015  * Schedule the given page for insertion into the specified page queue.
4016  * Physical insertion of the page may be deferred indefinitely.
4017  */
4018 static void
4019 vm_page_enqueue(vm_page_t m, uint8_t queue)
4020 {
4021 
4022 	KASSERT(m->a.queue == PQ_NONE &&
4023 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
4024 	    ("%s: page %p is already enqueued", __func__, m));
4025 	KASSERT(m->ref_count > 0,
4026 	    ("%s: page %p does not carry any references", __func__, m));
4027 
4028 	m->a.queue = queue;
4029 	if ((m->a.flags & PGA_REQUEUE) == 0)
4030 		vm_page_aflag_set(m, PGA_REQUEUE);
4031 	vm_page_pqbatch_submit(m, queue);
4032 }
4033 
4034 /*
4035  *	vm_page_free_prep:
4036  *
4037  *	Prepares the given page to be put on the free list,
4038  *	disassociating it from any VM object. The caller may return
4039  *	the page to the free list only if this function returns true.
4040  *
4041  *	The object, if it exists, must be locked, and then the page must
4042  *	be xbusy.  Otherwise the page must be not busied.  A managed
4043  *	page must be unmapped.
4044  */
4045 static bool
4046 vm_page_free_prep(vm_page_t m)
4047 {
4048 
4049 	/*
4050 	 * Synchronize with threads that have dropped a reference to this
4051 	 * page.
4052 	 */
4053 	atomic_thread_fence_acq();
4054 
4055 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
4056 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
4057 		uint64_t *p;
4058 		int i;
4059 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4060 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4061 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4062 			    m, i, (uintmax_t)*p));
4063 	}
4064 #endif
4065 	KASSERT((m->flags & PG_NOFREE) == 0,
4066 	    ("%s: attempting to free a PG_NOFREE page", __func__));
4067 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4068 		KASSERT(!pmap_page_is_mapped(m),
4069 		    ("vm_page_free_prep: freeing mapped page %p", m));
4070 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4071 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4072 	} else {
4073 		KASSERT(m->a.queue == PQ_NONE,
4074 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4075 	}
4076 	VM_CNT_INC(v_tfree);
4077 
4078 	if (m->object != NULL) {
4079 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
4080 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
4081 		    ("vm_page_free_prep: managed flag mismatch for page %p",
4082 		    m));
4083 		vm_page_assert_xbusied(m);
4084 
4085 		/*
4086 		 * The object reference can be released without an atomic
4087 		 * operation.
4088 		 */
4089 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4090 		    m->ref_count == VPRC_OBJREF,
4091 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
4092 		    m, m->ref_count));
4093 		vm_page_object_remove(m);
4094 		m->ref_count -= VPRC_OBJREF;
4095 	} else
4096 		vm_page_assert_unbusied(m);
4097 
4098 	vm_page_busy_free(m);
4099 
4100 	/*
4101 	 * If fictitious remove object association and
4102 	 * return.
4103 	 */
4104 	if ((m->flags & PG_FICTITIOUS) != 0) {
4105 		KASSERT(m->ref_count == 1,
4106 		    ("fictitious page %p is referenced", m));
4107 		KASSERT(m->a.queue == PQ_NONE,
4108 		    ("fictitious page %p is queued", m));
4109 		return (false);
4110 	}
4111 
4112 	/*
4113 	 * Pages need not be dequeued before they are returned to the physical
4114 	 * memory allocator, but they must at least be marked for a deferred
4115 	 * dequeue.
4116 	 */
4117 	if ((m->oflags & VPO_UNMANAGED) == 0)
4118 		vm_page_dequeue_deferred(m);
4119 
4120 	m->valid = 0;
4121 	vm_page_undirty(m);
4122 
4123 	if (m->ref_count != 0)
4124 		panic("vm_page_free_prep: page %p has references", m);
4125 
4126 	/*
4127 	 * Restore the default memory attribute to the page.
4128 	 */
4129 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4130 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4131 
4132 #if VM_NRESERVLEVEL > 0
4133 	/*
4134 	 * Determine whether the page belongs to a reservation.  If the page was
4135 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4136 	 * as an optimization, we avoid the check in that case.
4137 	 */
4138 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4139 		return (false);
4140 #endif
4141 
4142 	return (true);
4143 }
4144 
4145 /*
4146  *	vm_page_free_toq:
4147  *
4148  *	Returns the given page to the free list, disassociating it
4149  *	from any VM object.
4150  *
4151  *	The object must be locked.  The page must be exclusively busied if it
4152  *	belongs to an object.
4153  */
4154 static void
4155 vm_page_free_toq(vm_page_t m)
4156 {
4157 	struct vm_domain *vmd;
4158 	uma_zone_t zone;
4159 
4160 	if (!vm_page_free_prep(m))
4161 		return;
4162 
4163 	vmd = vm_pagequeue_domain(m);
4164 	zone = vmd->vmd_pgcache[m->pool].zone;
4165 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4166 		uma_zfree(zone, m);
4167 		return;
4168 	}
4169 	vm_domain_free_lock(vmd);
4170 	vm_phys_free_pages(m, 0);
4171 	vm_domain_free_unlock(vmd);
4172 	vm_domain_freecnt_inc(vmd, 1);
4173 }
4174 
4175 /*
4176  *	vm_page_free_pages_toq:
4177  *
4178  *	Returns a list of pages to the free list, disassociating it
4179  *	from any VM object.  In other words, this is equivalent to
4180  *	calling vm_page_free_toq() for each page of a list of VM objects.
4181  */
4182 int
4183 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4184 {
4185 	vm_page_t m;
4186 	int count;
4187 
4188 	if (SLIST_EMPTY(free))
4189 		return (0);
4190 
4191 	count = 0;
4192 	while ((m = SLIST_FIRST(free)) != NULL) {
4193 		count++;
4194 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4195 		vm_page_free_toq(m);
4196 	}
4197 
4198 	if (update_wire_count)
4199 		vm_wire_sub(count);
4200 	return (count);
4201 }
4202 
4203 /*
4204  * Mark this page as wired down.  For managed pages, this prevents reclamation
4205  * by the page daemon, or when the containing object, if any, is destroyed.
4206  */
4207 void
4208 vm_page_wire(vm_page_t m)
4209 {
4210 	u_int old;
4211 
4212 #ifdef INVARIANTS
4213 	if (m->object != NULL && !vm_page_busied(m) &&
4214 	    !vm_object_busied(m->object))
4215 		VM_OBJECT_ASSERT_LOCKED(m->object);
4216 #endif
4217 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4218 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4219 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4220 
4221 	old = atomic_fetchadd_int(&m->ref_count, 1);
4222 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4223 	    ("vm_page_wire: counter overflow for page %p", m));
4224 	if (VPRC_WIRE_COUNT(old) == 0) {
4225 		if ((m->oflags & VPO_UNMANAGED) == 0)
4226 			vm_page_aflag_set(m, PGA_DEQUEUE);
4227 		vm_wire_add(1);
4228 	}
4229 }
4230 
4231 /*
4232  * Attempt to wire a mapped page following a pmap lookup of that page.
4233  * This may fail if a thread is concurrently tearing down mappings of the page.
4234  * The transient failure is acceptable because it translates to the
4235  * failure of the caller pmap_extract_and_hold(), which should be then
4236  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4237  */
4238 bool
4239 vm_page_wire_mapped(vm_page_t m)
4240 {
4241 	u_int old;
4242 
4243 	old = atomic_load_int(&m->ref_count);
4244 	do {
4245 		KASSERT(old > 0,
4246 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4247 		if ((old & VPRC_BLOCKED) != 0)
4248 			return (false);
4249 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4250 
4251 	if (VPRC_WIRE_COUNT(old) == 0) {
4252 		if ((m->oflags & VPO_UNMANAGED) == 0)
4253 			vm_page_aflag_set(m, PGA_DEQUEUE);
4254 		vm_wire_add(1);
4255 	}
4256 	return (true);
4257 }
4258 
4259 /*
4260  * Release a wiring reference to a managed page.  If the page still belongs to
4261  * an object, update its position in the page queues to reflect the reference.
4262  * If the wiring was the last reference to the page, free the page.
4263  */
4264 static void
4265 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4266 {
4267 	u_int old;
4268 
4269 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4270 	    ("%s: page %p is unmanaged", __func__, m));
4271 
4272 	/*
4273 	 * Update LRU state before releasing the wiring reference.
4274 	 * Use a release store when updating the reference count to
4275 	 * synchronize with vm_page_free_prep().
4276 	 */
4277 	old = atomic_load_int(&m->ref_count);
4278 	do {
4279 		u_int count;
4280 
4281 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4282 		    ("vm_page_unwire: wire count underflow for page %p", m));
4283 
4284 		count = old & ~VPRC_BLOCKED;
4285 		if (count > VPRC_OBJREF + 1) {
4286 			/*
4287 			 * The page has at least one other wiring reference.  An
4288 			 * earlier iteration of this loop may have called
4289 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4290 			 * re-set it if necessary.
4291 			 */
4292 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4293 				vm_page_aflag_set(m, PGA_DEQUEUE);
4294 		} else if (count == VPRC_OBJREF + 1) {
4295 			/*
4296 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4297 			 * update the page's queue state to reflect the
4298 			 * reference.  If the page does not belong to an object
4299 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4300 			 * clear leftover queue state.
4301 			 */
4302 			vm_page_release_toq(m, nqueue, noreuse);
4303 		} else if (count == 1) {
4304 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4305 		}
4306 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4307 
4308 	if (VPRC_WIRE_COUNT(old) == 1) {
4309 		vm_wire_sub(1);
4310 		if (old == 1)
4311 			vm_page_free(m);
4312 	}
4313 }
4314 
4315 /*
4316  * Release one wiring of the specified page, potentially allowing it to be
4317  * paged out.
4318  *
4319  * Only managed pages belonging to an object can be paged out.  If the number
4320  * of wirings transitions to zero and the page is eligible for page out, then
4321  * the page is added to the specified paging queue.  If the released wiring
4322  * represented the last reference to the page, the page is freed.
4323  */
4324 void
4325 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4326 {
4327 
4328 	KASSERT(nqueue < PQ_COUNT,
4329 	    ("vm_page_unwire: invalid queue %u request for page %p",
4330 	    nqueue, m));
4331 
4332 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4333 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4334 			vm_page_free(m);
4335 		return;
4336 	}
4337 	vm_page_unwire_managed(m, nqueue, false);
4338 }
4339 
4340 /*
4341  * Unwire a page without (re-)inserting it into a page queue.  It is up
4342  * to the caller to enqueue, requeue, or free the page as appropriate.
4343  * In most cases involving managed pages, vm_page_unwire() should be used
4344  * instead.
4345  */
4346 bool
4347 vm_page_unwire_noq(vm_page_t m)
4348 {
4349 	u_int old;
4350 
4351 	old = vm_page_drop(m, 1);
4352 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4353 	    ("%s: counter underflow for page %p", __func__,  m));
4354 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4355 	    ("%s: missing ref on fictitious page %p", __func__, m));
4356 
4357 	if (VPRC_WIRE_COUNT(old) > 1)
4358 		return (false);
4359 	if ((m->oflags & VPO_UNMANAGED) == 0)
4360 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4361 	vm_wire_sub(1);
4362 	return (true);
4363 }
4364 
4365 /*
4366  * Ensure that the page ends up in the specified page queue.  If the page is
4367  * active or being moved to the active queue, ensure that its act_count is
4368  * at least ACT_INIT but do not otherwise mess with it.
4369  */
4370 static __always_inline void
4371 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4372 {
4373 	vm_page_astate_t old, new;
4374 
4375 	KASSERT(m->ref_count > 0,
4376 	    ("%s: page %p does not carry any references", __func__, m));
4377 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4378 	    ("%s: invalid flags %x", __func__, nflag));
4379 
4380 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4381 		return;
4382 
4383 	old = vm_page_astate_load(m);
4384 	do {
4385 		if ((old.flags & PGA_DEQUEUE) != 0)
4386 			break;
4387 		new = old;
4388 		new.flags &= ~PGA_QUEUE_OP_MASK;
4389 		if (nqueue == PQ_ACTIVE)
4390 			new.act_count = max(old.act_count, ACT_INIT);
4391 		if (old.queue == nqueue) {
4392 			/*
4393 			 * There is no need to requeue pages already in the
4394 			 * active queue.
4395 			 */
4396 			if (nqueue != PQ_ACTIVE ||
4397 			    (old.flags & PGA_ENQUEUED) == 0)
4398 				new.flags |= nflag;
4399 		} else {
4400 			new.flags |= nflag;
4401 			new.queue = nqueue;
4402 		}
4403 	} while (!vm_page_pqstate_commit(m, &old, new));
4404 }
4405 
4406 /*
4407  * Put the specified page on the active list (if appropriate).
4408  */
4409 void
4410 vm_page_activate(vm_page_t m)
4411 {
4412 
4413 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4414 }
4415 
4416 /*
4417  * Move the specified page to the tail of the inactive queue, or requeue
4418  * the page if it is already in the inactive queue.
4419  */
4420 void
4421 vm_page_deactivate(vm_page_t m)
4422 {
4423 
4424 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4425 }
4426 
4427 void
4428 vm_page_deactivate_noreuse(vm_page_t m)
4429 {
4430 
4431 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4432 }
4433 
4434 /*
4435  * Put a page in the laundry, or requeue it if it is already there.
4436  */
4437 void
4438 vm_page_launder(vm_page_t m)
4439 {
4440 
4441 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4442 }
4443 
4444 /*
4445  * Put a page in the PQ_UNSWAPPABLE holding queue.
4446  */
4447 void
4448 vm_page_unswappable(vm_page_t m)
4449 {
4450 
4451 	VM_OBJECT_ASSERT_LOCKED(m->object);
4452 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4453 	    ("page %p already unswappable", m));
4454 
4455 	vm_page_dequeue(m);
4456 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4457 }
4458 
4459 /*
4460  * Release a page back to the page queues in preparation for unwiring.
4461  */
4462 static void
4463 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4464 {
4465 	vm_page_astate_t old, new;
4466 	uint16_t nflag;
4467 
4468 	/*
4469 	 * Use a check of the valid bits to determine whether we should
4470 	 * accelerate reclamation of the page.  The object lock might not be
4471 	 * held here, in which case the check is racy.  At worst we will either
4472 	 * accelerate reclamation of a valid page and violate LRU, or
4473 	 * unnecessarily defer reclamation of an invalid page.
4474 	 *
4475 	 * If we were asked to not cache the page, place it near the head of the
4476 	 * inactive queue so that is reclaimed sooner.
4477 	 */
4478 	if (noreuse || vm_page_none_valid(m)) {
4479 		nqueue = PQ_INACTIVE;
4480 		nflag = PGA_REQUEUE_HEAD;
4481 	} else {
4482 		nflag = PGA_REQUEUE;
4483 	}
4484 
4485 	old = vm_page_astate_load(m);
4486 	do {
4487 		new = old;
4488 
4489 		/*
4490 		 * If the page is already in the active queue and we are not
4491 		 * trying to accelerate reclamation, simply mark it as
4492 		 * referenced and avoid any queue operations.
4493 		 */
4494 		new.flags &= ~PGA_QUEUE_OP_MASK;
4495 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4496 		    (old.flags & PGA_ENQUEUED) != 0)
4497 			new.flags |= PGA_REFERENCED;
4498 		else {
4499 			new.flags |= nflag;
4500 			new.queue = nqueue;
4501 		}
4502 	} while (!vm_page_pqstate_commit(m, &old, new));
4503 }
4504 
4505 /*
4506  * Unwire a page and either attempt to free it or re-add it to the page queues.
4507  */
4508 void
4509 vm_page_release(vm_page_t m, int flags)
4510 {
4511 	vm_object_t object;
4512 
4513 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4514 	    ("vm_page_release: page %p is unmanaged", m));
4515 
4516 	if ((flags & VPR_TRYFREE) != 0) {
4517 		for (;;) {
4518 			object = atomic_load_ptr(&m->object);
4519 			if (object == NULL)
4520 				break;
4521 			/* Depends on type-stability. */
4522 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4523 				break;
4524 			if (object == m->object) {
4525 				vm_page_release_locked(m, flags);
4526 				VM_OBJECT_WUNLOCK(object);
4527 				return;
4528 			}
4529 			VM_OBJECT_WUNLOCK(object);
4530 		}
4531 	}
4532 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4533 }
4534 
4535 /* See vm_page_release(). */
4536 void
4537 vm_page_release_locked(vm_page_t m, int flags)
4538 {
4539 
4540 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4541 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4542 	    ("vm_page_release_locked: page %p is unmanaged", m));
4543 
4544 	if (vm_page_unwire_noq(m)) {
4545 		if ((flags & VPR_TRYFREE) != 0 &&
4546 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4547 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4548 			/*
4549 			 * An unlocked lookup may have wired the page before the
4550 			 * busy lock was acquired, in which case the page must
4551 			 * not be freed.
4552 			 */
4553 			if (__predict_true(!vm_page_wired(m))) {
4554 				vm_page_free(m);
4555 				return;
4556 			}
4557 			vm_page_xunbusy(m);
4558 		} else {
4559 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4560 		}
4561 	}
4562 }
4563 
4564 static bool
4565 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4566 {
4567 	u_int old;
4568 
4569 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4570 	    ("vm_page_try_blocked_op: page %p has no object", m));
4571 	KASSERT(vm_page_busied(m),
4572 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4573 	VM_OBJECT_ASSERT_LOCKED(m->object);
4574 
4575 	old = atomic_load_int(&m->ref_count);
4576 	do {
4577 		KASSERT(old != 0,
4578 		    ("vm_page_try_blocked_op: page %p has no references", m));
4579 		KASSERT((old & VPRC_BLOCKED) == 0,
4580 		    ("vm_page_try_blocked_op: page %p blocks wirings", m));
4581 		if (VPRC_WIRE_COUNT(old) != 0)
4582 			return (false);
4583 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4584 
4585 	(op)(m);
4586 
4587 	/*
4588 	 * If the object is read-locked, new wirings may be created via an
4589 	 * object lookup.
4590 	 */
4591 	old = vm_page_drop(m, VPRC_BLOCKED);
4592 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4593 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4594 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4595 	    old, m));
4596 	return (true);
4597 }
4598 
4599 /*
4600  * Atomically check for wirings and remove all mappings of the page.
4601  */
4602 bool
4603 vm_page_try_remove_all(vm_page_t m)
4604 {
4605 
4606 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4607 }
4608 
4609 /*
4610  * Atomically check for wirings and remove all writeable mappings of the page.
4611  */
4612 bool
4613 vm_page_try_remove_write(vm_page_t m)
4614 {
4615 
4616 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4617 }
4618 
4619 /*
4620  * vm_page_advise
4621  *
4622  * 	Apply the specified advice to the given page.
4623  */
4624 void
4625 vm_page_advise(vm_page_t m, int advice)
4626 {
4627 
4628 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4629 	vm_page_assert_xbusied(m);
4630 
4631 	if (advice == MADV_FREE)
4632 		/*
4633 		 * Mark the page clean.  This will allow the page to be freed
4634 		 * without first paging it out.  MADV_FREE pages are often
4635 		 * quickly reused by malloc(3), so we do not do anything that
4636 		 * would result in a page fault on a later access.
4637 		 */
4638 		vm_page_undirty(m);
4639 	else if (advice != MADV_DONTNEED) {
4640 		if (advice == MADV_WILLNEED)
4641 			vm_page_activate(m);
4642 		return;
4643 	}
4644 
4645 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4646 		vm_page_dirty(m);
4647 
4648 	/*
4649 	 * Clear any references to the page.  Otherwise, the page daemon will
4650 	 * immediately reactivate the page.
4651 	 */
4652 	vm_page_aflag_clear(m, PGA_REFERENCED);
4653 
4654 	/*
4655 	 * Place clean pages near the head of the inactive queue rather than
4656 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4657 	 * the page will be reused quickly.  Dirty pages not already in the
4658 	 * laundry are moved there.
4659 	 */
4660 	if (m->dirty == 0)
4661 		vm_page_deactivate_noreuse(m);
4662 	else if (!vm_page_in_laundry(m))
4663 		vm_page_launder(m);
4664 }
4665 
4666 /*
4667  *	vm_page_grab_release
4668  *
4669  *	Helper routine for grab functions to release busy on return.
4670  */
4671 static inline void
4672 vm_page_grab_release(vm_page_t m, int allocflags)
4673 {
4674 
4675 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4676 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4677 			vm_page_sunbusy(m);
4678 		else
4679 			vm_page_xunbusy(m);
4680 	}
4681 }
4682 
4683 /*
4684  *	vm_page_grab_sleep
4685  *
4686  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4687  *	if the caller should retry and false otherwise.
4688  *
4689  *	If the object is locked on entry the object will be unlocked with
4690  *	false returns and still locked but possibly having been dropped
4691  *	with true returns.
4692  */
4693 static bool
4694 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4695     const char *wmesg, int allocflags, bool locked)
4696 {
4697 
4698 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4699 		return (false);
4700 
4701 	/*
4702 	 * Reference the page before unlocking and sleeping so that
4703 	 * the page daemon is less likely to reclaim it.
4704 	 */
4705 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4706 		vm_page_reference(m);
4707 
4708 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4709 	    locked)
4710 		VM_OBJECT_WLOCK(object);
4711 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4712 		return (false);
4713 
4714 	return (true);
4715 }
4716 
4717 /*
4718  * Assert that the grab flags are valid.
4719  */
4720 static inline void
4721 vm_page_grab_check(int allocflags)
4722 {
4723 
4724 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4725 	    (allocflags & VM_ALLOC_WIRED) != 0,
4726 	    ("vm_page_grab*: the pages must be busied or wired"));
4727 
4728 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4729 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4730 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4731 }
4732 
4733 /*
4734  * Calculate the page allocation flags for grab.
4735  */
4736 static inline int
4737 vm_page_grab_pflags(int allocflags)
4738 {
4739 	int pflags;
4740 
4741 	pflags = allocflags &
4742 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4743 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4744 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4745 		pflags |= VM_ALLOC_WAITFAIL;
4746 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4747 		pflags |= VM_ALLOC_SBUSY;
4748 
4749 	return (pflags);
4750 }
4751 
4752 /*
4753  * Grab a page, waiting until we are waken up due to the page
4754  * changing state.  We keep on waiting, if the page continues
4755  * to be in the object.  If the page doesn't exist, first allocate it
4756  * and then conditionally zero it.
4757  *
4758  * This routine may sleep.
4759  *
4760  * The object must be locked on entry.  The lock will, however, be released
4761  * and reacquired if the routine sleeps.
4762  */
4763 vm_page_t
4764 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4765 {
4766 	vm_page_t m;
4767 
4768 	VM_OBJECT_ASSERT_WLOCKED(object);
4769 	vm_page_grab_check(allocflags);
4770 
4771 retrylookup:
4772 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4773 		if (!vm_page_tryacquire(m, allocflags)) {
4774 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4775 			    allocflags, true))
4776 				goto retrylookup;
4777 			return (NULL);
4778 		}
4779 		goto out;
4780 	}
4781 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4782 		return (NULL);
4783 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4784 	if (m == NULL) {
4785 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4786 			return (NULL);
4787 		goto retrylookup;
4788 	}
4789 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4790 		pmap_zero_page(m);
4791 
4792 out:
4793 	vm_page_grab_release(m, allocflags);
4794 
4795 	return (m);
4796 }
4797 
4798 /*
4799  * Attempt to validate a page, locklessly acquiring it if necessary, given a
4800  * (object, pindex) tuple and either an invalided page or NULL.  The resulting
4801  * page will be validated against the identity tuple, and busied or wired as
4802  * requested.  A NULL page returned guarantees that the page was not in radix at
4803  * the time of the call but callers must perform higher level synchronization or
4804  * retry the operation under a lock if they require an atomic answer.  This is
4805  * the only lock free validation routine, other routines can depend on the
4806  * resulting page state.
4807  *
4808  * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to
4809  * caller flags.
4810  */
4811 #define PAGE_NOT_ACQUIRED ((vm_page_t)1)
4812 static vm_page_t
4813 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m,
4814     int allocflags)
4815 {
4816 	if (m == NULL)
4817 		m = vm_page_lookup_unlocked(object, pindex);
4818 	for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) {
4819 		if (vm_page_trybusy(m, allocflags)) {
4820 			if (m->object == object && m->pindex == pindex) {
4821 				if ((allocflags & VM_ALLOC_WIRED) != 0)
4822 					vm_page_wire(m);
4823 				vm_page_grab_release(m, allocflags);
4824 				break;
4825 			}
4826 			/* relookup. */
4827 			vm_page_busy_release(m);
4828 			cpu_spinwait();
4829 			continue;
4830 		}
4831 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4832 		    allocflags, false))
4833 			return (PAGE_NOT_ACQUIRED);
4834 	}
4835 	return (m);
4836 }
4837 
4838 /*
4839  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4840  * is not set.
4841  */
4842 vm_page_t
4843 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4844 {
4845 	vm_page_t m;
4846 
4847 	vm_page_grab_check(allocflags);
4848 	m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags);
4849 	if (m == PAGE_NOT_ACQUIRED)
4850 		return (NULL);
4851 	if (m != NULL)
4852 		return (m);
4853 
4854 	/*
4855 	 * The radix lockless lookup should never return a false negative
4856 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4857 	 * was no page present at the instant of the call.  A NOCREAT caller
4858 	 * must handle create races gracefully.
4859 	 */
4860 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4861 		return (NULL);
4862 
4863 	VM_OBJECT_WLOCK(object);
4864 	m = vm_page_grab(object, pindex, allocflags);
4865 	VM_OBJECT_WUNLOCK(object);
4866 
4867 	return (m);
4868 }
4869 
4870 /*
4871  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4872  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4873  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4874  * in simultaneously.  Additional pages will be left on a paging queue but
4875  * will neither be wired nor busy regardless of allocflags.
4876  */
4877 int
4878 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4879 {
4880 	vm_page_t m;
4881 	vm_page_t ma[VM_INITIAL_PAGEIN];
4882 	int after, i, pflags, rv;
4883 
4884 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4885 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4886 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4887 	KASSERT((allocflags &
4888 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4889 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4890 	VM_OBJECT_ASSERT_WLOCKED(object);
4891 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4892 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4893 	pflags |= VM_ALLOC_WAITFAIL;
4894 
4895 retrylookup:
4896 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4897 		/*
4898 		 * If the page is fully valid it can only become invalid
4899 		 * with the object lock held.  If it is not valid it can
4900 		 * become valid with the busy lock held.  Therefore, we
4901 		 * may unnecessarily lock the exclusive busy here if we
4902 		 * race with I/O completion not using the object lock.
4903 		 * However, we will not end up with an invalid page and a
4904 		 * shared lock.
4905 		 */
4906 		if (!vm_page_trybusy(m,
4907 		    vm_page_all_valid(m) ? allocflags : 0)) {
4908 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4909 			    allocflags, true);
4910 			goto retrylookup;
4911 		}
4912 		if (vm_page_all_valid(m))
4913 			goto out;
4914 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4915 			vm_page_busy_release(m);
4916 			*mp = NULL;
4917 			return (VM_PAGER_FAIL);
4918 		}
4919 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4920 		*mp = NULL;
4921 		return (VM_PAGER_FAIL);
4922 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4923 		if (!vm_pager_can_alloc_page(object, pindex)) {
4924 			*mp = NULL;
4925 			return (VM_PAGER_AGAIN);
4926 		}
4927 		goto retrylookup;
4928 	}
4929 
4930 	vm_page_assert_xbusied(m);
4931 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4932 		after = MIN(after, VM_INITIAL_PAGEIN);
4933 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4934 		after = MAX(after, 1);
4935 		ma[0] = m;
4936 		for (i = 1; i < after; i++) {
4937 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4938 				if (vm_page_any_valid(ma[i]) ||
4939 				    !vm_page_tryxbusy(ma[i]))
4940 					break;
4941 			} else {
4942 				ma[i] = vm_page_alloc(object, m->pindex + i,
4943 				    VM_ALLOC_NORMAL);
4944 				if (ma[i] == NULL)
4945 					break;
4946 			}
4947 		}
4948 		after = i;
4949 		vm_object_pip_add(object, after);
4950 		VM_OBJECT_WUNLOCK(object);
4951 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4952 		VM_OBJECT_WLOCK(object);
4953 		vm_object_pip_wakeupn(object, after);
4954 		/* Pager may have replaced a page. */
4955 		m = ma[0];
4956 		if (rv != VM_PAGER_OK) {
4957 			for (i = 0; i < after; i++) {
4958 				if (!vm_page_wired(ma[i]))
4959 					vm_page_free(ma[i]);
4960 				else
4961 					vm_page_xunbusy(ma[i]);
4962 			}
4963 			*mp = NULL;
4964 			return (rv);
4965 		}
4966 		for (i = 1; i < after; i++)
4967 			vm_page_readahead_finish(ma[i]);
4968 		MPASS(vm_page_all_valid(m));
4969 	} else {
4970 		vm_page_zero_invalid(m, TRUE);
4971 	}
4972 out:
4973 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4974 		vm_page_wire(m);
4975 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4976 		vm_page_busy_downgrade(m);
4977 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4978 		vm_page_busy_release(m);
4979 	*mp = m;
4980 	return (VM_PAGER_OK);
4981 }
4982 
4983 /*
4984  * Locklessly grab a valid page.  If the page is not valid or not yet
4985  * allocated this will fall back to the object lock method.
4986  */
4987 int
4988 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4989     vm_pindex_t pindex, int allocflags)
4990 {
4991 	vm_page_t m;
4992 	int flags;
4993 	int error;
4994 
4995 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4996 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4997 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4998 	    "mismatch"));
4999 	KASSERT((allocflags &
5000 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
5001 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
5002 
5003 	/*
5004 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
5005 	 * before we can inspect the valid field and return a wired page.
5006 	 */
5007 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
5008 	vm_page_grab_check(flags);
5009 	m = vm_page_acquire_unlocked(object, pindex, NULL, flags);
5010 	if (m == PAGE_NOT_ACQUIRED)
5011 		return (VM_PAGER_FAIL);
5012 	if (m != NULL) {
5013 		if (vm_page_all_valid(m)) {
5014 			if ((allocflags & VM_ALLOC_WIRED) != 0)
5015 				vm_page_wire(m);
5016 			vm_page_grab_release(m, allocflags);
5017 			*mp = m;
5018 			return (VM_PAGER_OK);
5019 		}
5020 		vm_page_busy_release(m);
5021 	}
5022 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5023 		*mp = NULL;
5024 		return (VM_PAGER_FAIL);
5025 	}
5026 	VM_OBJECT_WLOCK(object);
5027 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
5028 	VM_OBJECT_WUNLOCK(object);
5029 
5030 	return (error);
5031 }
5032 
5033 /*
5034  * Return the specified range of pages from the given object.  For each
5035  * page offset within the range, if a page already exists within the object
5036  * at that offset and it is busy, then wait for it to change state.  If,
5037  * instead, the page doesn't exist, then allocate it.
5038  *
5039  * The caller must always specify an allocation class.
5040  *
5041  * allocation classes:
5042  *	VM_ALLOC_NORMAL		normal process request
5043  *	VM_ALLOC_SYSTEM		system *really* needs the pages
5044  *
5045  * The caller must always specify that the pages are to be busied and/or
5046  * wired.
5047  *
5048  * optional allocation flags:
5049  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5050  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5051  *	VM_ALLOC_NOWAIT		do not sleep
5052  *	VM_ALLOC_SBUSY		set page to sbusy state
5053  *	VM_ALLOC_WIRED		wire the pages
5054  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5055  *
5056  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5057  * may return a partial prefix of the requested range.
5058  */
5059 int
5060 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5061     vm_page_t *ma, int count)
5062 {
5063 	vm_page_t m, mpred;
5064 	int pflags;
5065 	int i;
5066 
5067 	VM_OBJECT_ASSERT_WLOCKED(object);
5068 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5069 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5070 	KASSERT(count > 0,
5071 	    ("vm_page_grab_pages: invalid page count %d", count));
5072 	vm_page_grab_check(allocflags);
5073 
5074 	pflags = vm_page_grab_pflags(allocflags);
5075 	i = 0;
5076 retrylookup:
5077 	m = vm_page_mpred(object, pindex + i);
5078 	if (m == NULL || m->pindex != pindex + i) {
5079 		mpred = m;
5080 		m = NULL;
5081 	} else
5082 		mpred = TAILQ_PREV(m, pglist, listq);
5083 	for (; i < count; i++) {
5084 		if (m != NULL) {
5085 			if (!vm_page_tryacquire(m, allocflags)) {
5086 				if (vm_page_grab_sleep(object, m, pindex + i,
5087 				    "grbmaw", allocflags, true))
5088 					goto retrylookup;
5089 				break;
5090 			}
5091 		} else {
5092 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5093 				break;
5094 			m = vm_page_alloc_after(object, pindex + i,
5095 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5096 			if (m == NULL) {
5097 				if ((allocflags & (VM_ALLOC_NOWAIT |
5098 				    VM_ALLOC_WAITFAIL)) != 0)
5099 					break;
5100 				goto retrylookup;
5101 			}
5102 		}
5103 		if (vm_page_none_valid(m) &&
5104 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5105 			if ((m->flags & PG_ZERO) == 0)
5106 				pmap_zero_page(m);
5107 			vm_page_valid(m);
5108 		}
5109 		vm_page_grab_release(m, allocflags);
5110 		ma[i] = mpred = m;
5111 		m = vm_page_next(m);
5112 	}
5113 	return (i);
5114 }
5115 
5116 /*
5117  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5118  * and will fall back to the locked variant to handle allocation.
5119  */
5120 int
5121 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5122     int allocflags, vm_page_t *ma, int count)
5123 {
5124 	vm_page_t m;
5125 	int flags;
5126 	int i;
5127 
5128 	KASSERT(count > 0,
5129 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5130 	vm_page_grab_check(allocflags);
5131 
5132 	/*
5133 	 * Modify flags for lockless acquire to hold the page until we
5134 	 * set it valid if necessary.
5135 	 */
5136 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5137 	vm_page_grab_check(flags);
5138 	m = NULL;
5139 	for (i = 0; i < count; i++, pindex++) {
5140 		/*
5141 		 * We may see a false NULL here because the previous page has
5142 		 * been removed or just inserted and the list is loaded without
5143 		 * barriers.  Switch to radix to verify.
5144 		 */
5145 		if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex ||
5146 		    atomic_load_ptr(&m->object) != object) {
5147 			/*
5148 			 * This guarantees the result is instantaneously
5149 			 * correct.
5150 			 */
5151 			m = NULL;
5152 		}
5153 		m = vm_page_acquire_unlocked(object, pindex, m, flags);
5154 		if (m == PAGE_NOT_ACQUIRED)
5155 			return (i);
5156 		if (m == NULL)
5157 			break;
5158 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5159 			if ((m->flags & PG_ZERO) == 0)
5160 				pmap_zero_page(m);
5161 			vm_page_valid(m);
5162 		}
5163 		/* m will still be wired or busy according to flags. */
5164 		vm_page_grab_release(m, allocflags);
5165 		ma[i] = m;
5166 		m = TAILQ_NEXT(m, listq);
5167 	}
5168 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5169 		return (i);
5170 	count -= i;
5171 	VM_OBJECT_WLOCK(object);
5172 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5173 	VM_OBJECT_WUNLOCK(object);
5174 
5175 	return (i);
5176 }
5177 
5178 /*
5179  * Mapping function for valid or dirty bits in a page.
5180  *
5181  * Inputs are required to range within a page.
5182  */
5183 vm_page_bits_t
5184 vm_page_bits(int base, int size)
5185 {
5186 	int first_bit;
5187 	int last_bit;
5188 
5189 	KASSERT(
5190 	    base + size <= PAGE_SIZE,
5191 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5192 	);
5193 
5194 	if (size == 0)		/* handle degenerate case */
5195 		return (0);
5196 
5197 	first_bit = base >> DEV_BSHIFT;
5198 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5199 
5200 	return (((vm_page_bits_t)2 << last_bit) -
5201 	    ((vm_page_bits_t)1 << first_bit));
5202 }
5203 
5204 void
5205 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5206 {
5207 
5208 #if PAGE_SIZE == 32768
5209 	atomic_set_64((uint64_t *)bits, set);
5210 #elif PAGE_SIZE == 16384
5211 	atomic_set_32((uint32_t *)bits, set);
5212 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5213 	atomic_set_16((uint16_t *)bits, set);
5214 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5215 	atomic_set_8((uint8_t *)bits, set);
5216 #else		/* PAGE_SIZE <= 8192 */
5217 	uintptr_t addr;
5218 	int shift;
5219 
5220 	addr = (uintptr_t)bits;
5221 	/*
5222 	 * Use a trick to perform a 32-bit atomic on the
5223 	 * containing aligned word, to not depend on the existence
5224 	 * of atomic_{set, clear}_{8, 16}.
5225 	 */
5226 	shift = addr & (sizeof(uint32_t) - 1);
5227 #if BYTE_ORDER == BIG_ENDIAN
5228 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5229 #else
5230 	shift *= NBBY;
5231 #endif
5232 	addr &= ~(sizeof(uint32_t) - 1);
5233 	atomic_set_32((uint32_t *)addr, set << shift);
5234 #endif		/* PAGE_SIZE */
5235 }
5236 
5237 static inline void
5238 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5239 {
5240 
5241 #if PAGE_SIZE == 32768
5242 	atomic_clear_64((uint64_t *)bits, clear);
5243 #elif PAGE_SIZE == 16384
5244 	atomic_clear_32((uint32_t *)bits, clear);
5245 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5246 	atomic_clear_16((uint16_t *)bits, clear);
5247 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5248 	atomic_clear_8((uint8_t *)bits, clear);
5249 #else		/* PAGE_SIZE <= 8192 */
5250 	uintptr_t addr;
5251 	int shift;
5252 
5253 	addr = (uintptr_t)bits;
5254 	/*
5255 	 * Use a trick to perform a 32-bit atomic on the
5256 	 * containing aligned word, to not depend on the existence
5257 	 * of atomic_{set, clear}_{8, 16}.
5258 	 */
5259 	shift = addr & (sizeof(uint32_t) - 1);
5260 #if BYTE_ORDER == BIG_ENDIAN
5261 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5262 #else
5263 	shift *= NBBY;
5264 #endif
5265 	addr &= ~(sizeof(uint32_t) - 1);
5266 	atomic_clear_32((uint32_t *)addr, clear << shift);
5267 #endif		/* PAGE_SIZE */
5268 }
5269 
5270 static inline vm_page_bits_t
5271 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5272 {
5273 #if PAGE_SIZE == 32768
5274 	uint64_t old;
5275 
5276 	old = *bits;
5277 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5278 	return (old);
5279 #elif PAGE_SIZE == 16384
5280 	uint32_t old;
5281 
5282 	old = *bits;
5283 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5284 	return (old);
5285 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5286 	uint16_t old;
5287 
5288 	old = *bits;
5289 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5290 	return (old);
5291 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5292 	uint8_t old;
5293 
5294 	old = *bits;
5295 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5296 	return (old);
5297 #else		/* PAGE_SIZE <= 4096*/
5298 	uintptr_t addr;
5299 	uint32_t old, new, mask;
5300 	int shift;
5301 
5302 	addr = (uintptr_t)bits;
5303 	/*
5304 	 * Use a trick to perform a 32-bit atomic on the
5305 	 * containing aligned word, to not depend on the existence
5306 	 * of atomic_{set, swap, clear}_{8, 16}.
5307 	 */
5308 	shift = addr & (sizeof(uint32_t) - 1);
5309 #if BYTE_ORDER == BIG_ENDIAN
5310 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5311 #else
5312 	shift *= NBBY;
5313 #endif
5314 	addr &= ~(sizeof(uint32_t) - 1);
5315 	mask = VM_PAGE_BITS_ALL << shift;
5316 
5317 	old = *bits;
5318 	do {
5319 		new = old & ~mask;
5320 		new |= newbits << shift;
5321 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5322 	return (old >> shift);
5323 #endif		/* PAGE_SIZE */
5324 }
5325 
5326 /*
5327  *	vm_page_set_valid_range:
5328  *
5329  *	Sets portions of a page valid.  The arguments are expected
5330  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5331  *	of any partial chunks touched by the range.  The invalid portion of
5332  *	such chunks will be zeroed.
5333  *
5334  *	(base + size) must be less then or equal to PAGE_SIZE.
5335  */
5336 void
5337 vm_page_set_valid_range(vm_page_t m, int base, int size)
5338 {
5339 	int endoff, frag;
5340 	vm_page_bits_t pagebits;
5341 
5342 	vm_page_assert_busied(m);
5343 	if (size == 0)	/* handle degenerate case */
5344 		return;
5345 
5346 	/*
5347 	 * If the base is not DEV_BSIZE aligned and the valid
5348 	 * bit is clear, we have to zero out a portion of the
5349 	 * first block.
5350 	 */
5351 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5352 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5353 		pmap_zero_page_area(m, frag, base - frag);
5354 
5355 	/*
5356 	 * If the ending offset is not DEV_BSIZE aligned and the
5357 	 * valid bit is clear, we have to zero out a portion of
5358 	 * the last block.
5359 	 */
5360 	endoff = base + size;
5361 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5362 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5363 		pmap_zero_page_area(m, endoff,
5364 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5365 
5366 	/*
5367 	 * Assert that no previously invalid block that is now being validated
5368 	 * is already dirty.
5369 	 */
5370 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5371 	    ("vm_page_set_valid_range: page %p is dirty", m));
5372 
5373 	/*
5374 	 * Set valid bits inclusive of any overlap.
5375 	 */
5376 	pagebits = vm_page_bits(base, size);
5377 	if (vm_page_xbusied(m))
5378 		m->valid |= pagebits;
5379 	else
5380 		vm_page_bits_set(m, &m->valid, pagebits);
5381 }
5382 
5383 /*
5384  * Set the page dirty bits and free the invalid swap space if
5385  * present.  Returns the previous dirty bits.
5386  */
5387 vm_page_bits_t
5388 vm_page_set_dirty(vm_page_t m)
5389 {
5390 	vm_page_bits_t old;
5391 
5392 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5393 
5394 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5395 		old = m->dirty;
5396 		m->dirty = VM_PAGE_BITS_ALL;
5397 	} else
5398 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5399 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5400 		vm_pager_page_unswapped(m);
5401 
5402 	return (old);
5403 }
5404 
5405 /*
5406  * Clear the given bits from the specified page's dirty field.
5407  */
5408 static __inline void
5409 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5410 {
5411 
5412 	vm_page_assert_busied(m);
5413 
5414 	/*
5415 	 * If the page is xbusied and not write mapped we are the
5416 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5417 	 * layer can call vm_page_dirty() without holding a distinguished
5418 	 * lock.  The combination of page busy and atomic operations
5419 	 * suffice to guarantee consistency of the page dirty field.
5420 	 */
5421 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5422 		m->dirty &= ~pagebits;
5423 	else
5424 		vm_page_bits_clear(m, &m->dirty, pagebits);
5425 }
5426 
5427 /*
5428  *	vm_page_set_validclean:
5429  *
5430  *	Sets portions of a page valid and clean.  The arguments are expected
5431  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5432  *	of any partial chunks touched by the range.  The invalid portion of
5433  *	such chunks will be zero'd.
5434  *
5435  *	(base + size) must be less then or equal to PAGE_SIZE.
5436  */
5437 void
5438 vm_page_set_validclean(vm_page_t m, int base, int size)
5439 {
5440 	vm_page_bits_t oldvalid, pagebits;
5441 	int endoff, frag;
5442 
5443 	vm_page_assert_busied(m);
5444 	if (size == 0)	/* handle degenerate case */
5445 		return;
5446 
5447 	/*
5448 	 * If the base is not DEV_BSIZE aligned and the valid
5449 	 * bit is clear, we have to zero out a portion of the
5450 	 * first block.
5451 	 */
5452 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5453 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5454 		pmap_zero_page_area(m, frag, base - frag);
5455 
5456 	/*
5457 	 * If the ending offset is not DEV_BSIZE aligned and the
5458 	 * valid bit is clear, we have to zero out a portion of
5459 	 * the last block.
5460 	 */
5461 	endoff = base + size;
5462 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5463 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5464 		pmap_zero_page_area(m, endoff,
5465 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5466 
5467 	/*
5468 	 * Set valid, clear dirty bits.  If validating the entire
5469 	 * page we can safely clear the pmap modify bit.  We also
5470 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5471 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5472 	 * be set again.
5473 	 *
5474 	 * We set valid bits inclusive of any overlap, but we can only
5475 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5476 	 * the range.
5477 	 */
5478 	oldvalid = m->valid;
5479 	pagebits = vm_page_bits(base, size);
5480 	if (vm_page_xbusied(m))
5481 		m->valid |= pagebits;
5482 	else
5483 		vm_page_bits_set(m, &m->valid, pagebits);
5484 #if 0	/* NOT YET */
5485 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5486 		frag = DEV_BSIZE - frag;
5487 		base += frag;
5488 		size -= frag;
5489 		if (size < 0)
5490 			size = 0;
5491 	}
5492 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5493 #endif
5494 	if (base == 0 && size == PAGE_SIZE) {
5495 		/*
5496 		 * The page can only be modified within the pmap if it is
5497 		 * mapped, and it can only be mapped if it was previously
5498 		 * fully valid.
5499 		 */
5500 		if (oldvalid == VM_PAGE_BITS_ALL)
5501 			/*
5502 			 * Perform the pmap_clear_modify() first.  Otherwise,
5503 			 * a concurrent pmap operation, such as
5504 			 * pmap_protect(), could clear a modification in the
5505 			 * pmap and set the dirty field on the page before
5506 			 * pmap_clear_modify() had begun and after the dirty
5507 			 * field was cleared here.
5508 			 */
5509 			pmap_clear_modify(m);
5510 		m->dirty = 0;
5511 		vm_page_aflag_clear(m, PGA_NOSYNC);
5512 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5513 		m->dirty &= ~pagebits;
5514 	else
5515 		vm_page_clear_dirty_mask(m, pagebits);
5516 }
5517 
5518 void
5519 vm_page_clear_dirty(vm_page_t m, int base, int size)
5520 {
5521 
5522 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5523 }
5524 
5525 /*
5526  *	vm_page_set_invalid:
5527  *
5528  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5529  *	valid and dirty bits for the effected areas are cleared.
5530  */
5531 void
5532 vm_page_set_invalid(vm_page_t m, int base, int size)
5533 {
5534 	vm_page_bits_t bits;
5535 	vm_object_t object;
5536 
5537 	/*
5538 	 * The object lock is required so that pages can't be mapped
5539 	 * read-only while we're in the process of invalidating them.
5540 	 */
5541 	object = m->object;
5542 	VM_OBJECT_ASSERT_WLOCKED(object);
5543 	vm_page_assert_busied(m);
5544 
5545 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5546 	    size >= object->un_pager.vnp.vnp_size)
5547 		bits = VM_PAGE_BITS_ALL;
5548 	else
5549 		bits = vm_page_bits(base, size);
5550 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5551 		pmap_remove_all(m);
5552 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5553 	    !pmap_page_is_mapped(m),
5554 	    ("vm_page_set_invalid: page %p is mapped", m));
5555 	if (vm_page_xbusied(m)) {
5556 		m->valid &= ~bits;
5557 		m->dirty &= ~bits;
5558 	} else {
5559 		vm_page_bits_clear(m, &m->valid, bits);
5560 		vm_page_bits_clear(m, &m->dirty, bits);
5561 	}
5562 }
5563 
5564 /*
5565  *	vm_page_invalid:
5566  *
5567  *	Invalidates the entire page.  The page must be busy, unmapped, and
5568  *	the enclosing object must be locked.  The object locks protects
5569  *	against concurrent read-only pmap enter which is done without
5570  *	busy.
5571  */
5572 void
5573 vm_page_invalid(vm_page_t m)
5574 {
5575 
5576 	vm_page_assert_busied(m);
5577 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5578 	MPASS(!pmap_page_is_mapped(m));
5579 
5580 	if (vm_page_xbusied(m))
5581 		m->valid = 0;
5582 	else
5583 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5584 }
5585 
5586 /*
5587  * vm_page_zero_invalid()
5588  *
5589  *	The kernel assumes that the invalid portions of a page contain
5590  *	garbage, but such pages can be mapped into memory by user code.
5591  *	When this occurs, we must zero out the non-valid portions of the
5592  *	page so user code sees what it expects.
5593  *
5594  *	Pages are most often semi-valid when the end of a file is mapped
5595  *	into memory and the file's size is not page aligned.
5596  */
5597 void
5598 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5599 {
5600 	int b;
5601 	int i;
5602 
5603 	/*
5604 	 * Scan the valid bits looking for invalid sections that
5605 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5606 	 * valid bit may be set ) have already been zeroed by
5607 	 * vm_page_set_validclean().
5608 	 */
5609 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5610 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5611 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5612 			if (i > b) {
5613 				pmap_zero_page_area(m,
5614 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5615 			}
5616 			b = i + 1;
5617 		}
5618 	}
5619 
5620 	/*
5621 	 * setvalid is TRUE when we can safely set the zero'd areas
5622 	 * as being valid.  We can do this if there are no cache consistency
5623 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5624 	 */
5625 	if (setvalid)
5626 		vm_page_valid(m);
5627 }
5628 
5629 /*
5630  *	vm_page_is_valid:
5631  *
5632  *	Is (partial) page valid?  Note that the case where size == 0
5633  *	will return FALSE in the degenerate case where the page is
5634  *	entirely invalid, and TRUE otherwise.
5635  *
5636  *	Some callers envoke this routine without the busy lock held and
5637  *	handle races via higher level locks.  Typical callers should
5638  *	hold a busy lock to prevent invalidation.
5639  */
5640 int
5641 vm_page_is_valid(vm_page_t m, int base, int size)
5642 {
5643 	vm_page_bits_t bits;
5644 
5645 	bits = vm_page_bits(base, size);
5646 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5647 }
5648 
5649 /*
5650  * Returns true if all of the specified predicates are true for the entire
5651  * (super)page and false otherwise.
5652  */
5653 bool
5654 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5655 {
5656 	vm_object_t object;
5657 	int i, npages;
5658 
5659 	object = m->object;
5660 	if (skip_m != NULL && skip_m->object != object)
5661 		return (false);
5662 	VM_OBJECT_ASSERT_LOCKED(object);
5663 	KASSERT(psind <= m->psind,
5664 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5665 	npages = atop(pagesizes[psind]);
5666 
5667 	/*
5668 	 * The physically contiguous pages that make up a superpage, i.e., a
5669 	 * page with a page size index ("psind") greater than zero, will
5670 	 * occupy adjacent entries in vm_page_array[].
5671 	 */
5672 	for (i = 0; i < npages; i++) {
5673 		/* Always test object consistency, including "skip_m". */
5674 		if (m[i].object != object)
5675 			return (false);
5676 		if (&m[i] == skip_m)
5677 			continue;
5678 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5679 			return (false);
5680 		if ((flags & PS_ALL_DIRTY) != 0) {
5681 			/*
5682 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5683 			 * might stop this case from spuriously returning
5684 			 * "false".  However, that would require a write lock
5685 			 * on the object containing "m[i]".
5686 			 */
5687 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5688 				return (false);
5689 		}
5690 		if ((flags & PS_ALL_VALID) != 0 &&
5691 		    m[i].valid != VM_PAGE_BITS_ALL)
5692 			return (false);
5693 	}
5694 	return (true);
5695 }
5696 
5697 /*
5698  * Set the page's dirty bits if the page is modified.
5699  */
5700 void
5701 vm_page_test_dirty(vm_page_t m)
5702 {
5703 
5704 	vm_page_assert_busied(m);
5705 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5706 		vm_page_dirty(m);
5707 }
5708 
5709 void
5710 vm_page_valid(vm_page_t m)
5711 {
5712 
5713 	vm_page_assert_busied(m);
5714 	if (vm_page_xbusied(m))
5715 		m->valid = VM_PAGE_BITS_ALL;
5716 	else
5717 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5718 }
5719 
5720 void
5721 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5722 {
5723 
5724 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5725 }
5726 
5727 void
5728 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5729 {
5730 
5731 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5732 }
5733 
5734 int
5735 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5736 {
5737 
5738 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5739 }
5740 
5741 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5742 void
5743 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5744 {
5745 
5746 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5747 }
5748 
5749 void
5750 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5751 {
5752 
5753 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5754 }
5755 #endif
5756 
5757 #ifdef INVARIANTS
5758 void
5759 vm_page_object_busy_assert(vm_page_t m)
5760 {
5761 
5762 	/*
5763 	 * Certain of the page's fields may only be modified by the
5764 	 * holder of a page or object busy.
5765 	 */
5766 	if (m->object != NULL && !vm_page_busied(m))
5767 		VM_OBJECT_ASSERT_BUSY(m->object);
5768 }
5769 
5770 void
5771 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5772 {
5773 
5774 	if ((bits & PGA_WRITEABLE) == 0)
5775 		return;
5776 
5777 	/*
5778 	 * The PGA_WRITEABLE flag can only be set if the page is
5779 	 * managed, is exclusively busied or the object is locked.
5780 	 * Currently, this flag is only set by pmap_enter().
5781 	 */
5782 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5783 	    ("PGA_WRITEABLE on unmanaged page"));
5784 	if (!vm_page_xbusied(m))
5785 		VM_OBJECT_ASSERT_BUSY(m->object);
5786 }
5787 #endif
5788 
5789 #include "opt_ddb.h"
5790 #ifdef DDB
5791 #include <sys/kernel.h>
5792 
5793 #include <ddb/ddb.h>
5794 
5795 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5796 {
5797 
5798 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5799 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5800 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5801 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5802 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5803 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5804 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5805 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5806 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5807 }
5808 
5809 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5810 {
5811 	int dom;
5812 
5813 	db_printf("pq_free %d\n", vm_free_count());
5814 	for (dom = 0; dom < vm_ndomains; dom++) {
5815 		db_printf(
5816     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5817 		    dom,
5818 		    vm_dom[dom].vmd_page_count,
5819 		    vm_dom[dom].vmd_free_count,
5820 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5821 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5822 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5823 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5824 	}
5825 }
5826 
5827 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5828 {
5829 	vm_page_t m;
5830 	boolean_t phys, virt;
5831 
5832 	if (!have_addr) {
5833 		db_printf("show pginfo addr\n");
5834 		return;
5835 	}
5836 
5837 	phys = strchr(modif, 'p') != NULL;
5838 	virt = strchr(modif, 'v') != NULL;
5839 	if (virt)
5840 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5841 	else if (phys)
5842 		m = PHYS_TO_VM_PAGE(addr);
5843 	else
5844 		m = (vm_page_t)addr;
5845 	db_printf(
5846     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5847     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5848 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5849 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5850 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5851 }
5852 #endif /* DDB */
5853