xref: /freebsd/sys/vm/vm_page.c (revision 92f340d137ba5d6db7610ba1dae35842e2c9c8ea)
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 void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
175     vm_page_t mpred);
176 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
177     const uint16_t nflag);
178 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
179     vm_page_t m_run, vm_paddr_t high);
180 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
181 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
182     int req);
183 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
184     int flags);
185 static void vm_page_zone_release(void *arg, void **store, int cnt);
186 
187 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
188 
189 static void
190 vm_page_init(void *dummy)
191 {
192 
193 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
194 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
195 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE);
196 }
197 
198 static int pgcache_zone_max_pcpu;
199 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
200     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
201     "Per-CPU page cache size");
202 
203 /*
204  * The cache page zone is initialized later since we need to be able to allocate
205  * pages before UMA is fully initialized.
206  */
207 static void
208 vm_page_init_cache_zones(void *dummy __unused)
209 {
210 	struct vm_domain *vmd;
211 	struct vm_pgcache *pgcache;
212 	int cache, domain, maxcache, pool;
213 
214 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
215 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
216 	for (domain = 0; domain < vm_ndomains; domain++) {
217 		vmd = VM_DOMAIN(domain);
218 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
219 			pgcache = &vmd->vmd_pgcache[pool];
220 			pgcache->domain = domain;
221 			pgcache->pool = pool;
222 			pgcache->zone = uma_zcache_create("vm pgcache",
223 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
224 			    vm_page_zone_import, vm_page_zone_release, pgcache,
225 			    UMA_ZONE_VM);
226 
227 			/*
228 			 * Limit each pool's zone to 0.1% of the pages in the
229 			 * domain.
230 			 */
231 			cache = maxcache != 0 ? maxcache :
232 			    vmd->vmd_page_count / 1000;
233 			uma_zone_set_maxcache(pgcache->zone, cache);
234 		}
235 	}
236 }
237 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
238 
239 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
240 #if PAGE_SIZE == 32768
241 #ifdef CTASSERT
242 CTASSERT(sizeof(u_long) >= 8);
243 #endif
244 #endif
245 
246 /*
247  *	vm_set_page_size:
248  *
249  *	Sets the page size, perhaps based upon the memory
250  *	size.  Must be called before any use of page-size
251  *	dependent functions.
252  */
253 void
254 vm_set_page_size(void)
255 {
256 	if (vm_cnt.v_page_size == 0)
257 		vm_cnt.v_page_size = PAGE_SIZE;
258 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
259 		panic("vm_set_page_size: page size not a power of two");
260 }
261 
262 /*
263  *	vm_page_blacklist_next:
264  *
265  *	Find the next entry in the provided string of blacklist
266  *	addresses.  Entries are separated by space, comma, or newline.
267  *	If an invalid integer is encountered then the rest of the
268  *	string is skipped.  Updates the list pointer to the next
269  *	character, or NULL if the string is exhausted or invalid.
270  */
271 static vm_paddr_t
272 vm_page_blacklist_next(char **list, char *end)
273 {
274 	vm_paddr_t bad;
275 	char *cp, *pos;
276 
277 	if (list == NULL || *list == NULL)
278 		return (0);
279 	if (**list =='\0') {
280 		*list = NULL;
281 		return (0);
282 	}
283 
284 	/*
285 	 * If there's no end pointer then the buffer is coming from
286 	 * the kenv and we know it's null-terminated.
287 	 */
288 	if (end == NULL)
289 		end = *list + strlen(*list);
290 
291 	/* Ensure that strtoq() won't walk off the end */
292 	if (*end != '\0') {
293 		if (*end == '\n' || *end == ' ' || *end  == ',')
294 			*end = '\0';
295 		else {
296 			printf("Blacklist not terminated, skipping\n");
297 			*list = NULL;
298 			return (0);
299 		}
300 	}
301 
302 	for (pos = *list; *pos != '\0'; pos = cp) {
303 		bad = strtoq(pos, &cp, 0);
304 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
305 			if (bad == 0) {
306 				if (++cp < end)
307 					continue;
308 				else
309 					break;
310 			}
311 		} else
312 			break;
313 		if (*cp == '\0' || ++cp >= end)
314 			*list = NULL;
315 		else
316 			*list = cp;
317 		return (trunc_page(bad));
318 	}
319 	printf("Garbage in RAM blacklist, skipping\n");
320 	*list = NULL;
321 	return (0);
322 }
323 
324 bool
325 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
326 {
327 	struct vm_domain *vmd;
328 	vm_page_t m;
329 	bool found;
330 
331 	m = vm_phys_paddr_to_vm_page(pa);
332 	if (m == NULL)
333 		return (true); /* page does not exist, no failure */
334 
335 	vmd = VM_DOMAIN(vm_phys_domain(pa));
336 	vm_domain_free_lock(vmd);
337 	found = vm_phys_unfree_page(pa);
338 	vm_domain_free_unlock(vmd);
339 	if (found) {
340 		vm_domain_freecnt_inc(vmd, -1);
341 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
342 		if (verbose)
343 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
344 	}
345 	return (found);
346 }
347 
348 /*
349  *	vm_page_blacklist_check:
350  *
351  *	Iterate through the provided string of blacklist addresses, pulling
352  *	each entry out of the physical allocator free list and putting it
353  *	onto a list for reporting via the vm.page_blacklist sysctl.
354  */
355 static void
356 vm_page_blacklist_check(char *list, char *end)
357 {
358 	vm_paddr_t pa;
359 	char *next;
360 
361 	next = list;
362 	while (next != NULL) {
363 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
364 			continue;
365 		vm_page_blacklist_add(pa, bootverbose);
366 	}
367 }
368 
369 /*
370  *	vm_page_blacklist_load:
371  *
372  *	Search for a special module named "ram_blacklist".  It'll be a
373  *	plain text file provided by the user via the loader directive
374  *	of the same name.
375  */
376 static void
377 vm_page_blacklist_load(char **list, char **end)
378 {
379 	void *mod;
380 	u_char *ptr;
381 	u_int len;
382 
383 	mod = NULL;
384 	ptr = NULL;
385 
386 	mod = preload_search_by_type("ram_blacklist");
387 	if (mod != NULL) {
388 		ptr = preload_fetch_addr(mod);
389 		len = preload_fetch_size(mod);
390         }
391 	*list = ptr;
392 	if (ptr != NULL)
393 		*end = ptr + len;
394 	else
395 		*end = NULL;
396 	return;
397 }
398 
399 static int
400 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
401 {
402 	vm_page_t m;
403 	struct sbuf sbuf;
404 	int error, first;
405 
406 	first = 1;
407 	error = sysctl_wire_old_buffer(req, 0);
408 	if (error != 0)
409 		return (error);
410 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
411 	TAILQ_FOREACH(m, &blacklist_head, listq) {
412 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
413 		    (uintmax_t)m->phys_addr);
414 		first = 0;
415 	}
416 	error = sbuf_finish(&sbuf);
417 	sbuf_delete(&sbuf);
418 	return (error);
419 }
420 
421 /*
422  * Initialize a dummy page for use in scans of the specified paging queue.
423  * In principle, this function only needs to set the flag PG_MARKER.
424  * Nonetheless, it write busies the page as a safety precaution.
425  */
426 void
427 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
428 {
429 
430 	bzero(marker, sizeof(*marker));
431 	marker->flags = PG_MARKER;
432 	marker->a.flags = aflags;
433 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
434 	marker->a.queue = queue;
435 }
436 
437 static void
438 vm_page_domain_init(int domain)
439 {
440 	struct vm_domain *vmd;
441 	struct vm_pagequeue *pq;
442 	int i;
443 
444 	vmd = VM_DOMAIN(domain);
445 	bzero(vmd, sizeof(*vmd));
446 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
447 	    "vm inactive pagequeue";
448 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
449 	    "vm active pagequeue";
450 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
451 	    "vm laundry pagequeue";
452 	*__DECONST(const char **,
453 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
454 	    "vm unswappable pagequeue";
455 	vmd->vmd_domain = domain;
456 	vmd->vmd_page_count = 0;
457 	vmd->vmd_free_count = 0;
458 	vmd->vmd_segs = 0;
459 	vmd->vmd_oom = false;
460 	vmd->vmd_helper_threads_enabled = true;
461 	for (i = 0; i < PQ_COUNT; i++) {
462 		pq = &vmd->vmd_pagequeues[i];
463 		TAILQ_INIT(&pq->pq_pl);
464 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
465 		    MTX_DEF | MTX_DUPOK);
466 		pq->pq_pdpages = 0;
467 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
468 	}
469 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
470 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
471 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
472 
473 	/*
474 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
475 	 * insertions.
476 	 */
477 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
478 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
479 	    &vmd->vmd_inacthead, plinks.q);
480 
481 	/*
482 	 * The clock pages are used to implement active queue scanning without
483 	 * requeues.  Scans start at clock[0], which is advanced after the scan
484 	 * ends.  When the two clock hands meet, they are reset and scanning
485 	 * resumes from the head of the queue.
486 	 */
487 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
488 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
489 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
490 	    &vmd->vmd_clock[0], plinks.q);
491 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
492 	    &vmd->vmd_clock[1], plinks.q);
493 }
494 
495 /*
496  * Initialize a physical page in preparation for adding it to the free
497  * lists.
498  */
499 void
500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool)
501 {
502 	m->object = NULL;
503 	m->ref_count = 0;
504 	m->busy_lock = VPB_FREED;
505 	m->flags = m->a.flags = 0;
506 	m->phys_addr = pa;
507 	m->a.queue = PQ_NONE;
508 	m->psind = 0;
509 	m->segind = segind;
510 	m->order = VM_NFREEORDER;
511 	m->pool = pool;
512 	m->valid = m->dirty = 0;
513 	pmap_page_init(m);
514 }
515 
516 #ifndef PMAP_HAS_PAGE_ARRAY
517 static vm_paddr_t
518 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
519 {
520 	vm_paddr_t new_end;
521 
522 	/*
523 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
524 	 * However, because this page is allocated from KVM, out-of-bounds
525 	 * accesses using the direct map will not be trapped.
526 	 */
527 	*vaddr += PAGE_SIZE;
528 
529 	/*
530 	 * Allocate physical memory for the page structures, and map it.
531 	 */
532 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
533 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
534 	    VM_PROT_READ | VM_PROT_WRITE);
535 	vm_page_array_size = page_range;
536 
537 	return (new_end);
538 }
539 #endif
540 
541 /*
542  *	vm_page_startup:
543  *
544  *	Initializes the resident memory module.  Allocates physical memory for
545  *	bootstrapping UMA and some data structures that are used to manage
546  *	physical pages.  Initializes these structures, and populates the free
547  *	page queues.
548  */
549 vm_offset_t
550 vm_page_startup(vm_offset_t vaddr)
551 {
552 	struct vm_phys_seg *seg;
553 	struct vm_domain *vmd;
554 	vm_page_t m;
555 	char *list, *listend;
556 	vm_paddr_t end, high_avail, low_avail, new_end, size;
557 	vm_paddr_t page_range __unused;
558 	vm_paddr_t last_pa, pa, startp, endp;
559 	u_long pagecount;
560 #if MINIDUMP_PAGE_TRACKING
561 	u_long vm_page_dump_size;
562 #endif
563 	int biggestone, i, segind;
564 #ifdef WITNESS
565 	vm_offset_t mapped;
566 	int witness_size;
567 #endif
568 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
569 	long ii;
570 #endif
571 	int pool;
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 #else
652 	(void)pa;
653 #endif
654 
655 	/*
656 	 * Determine the lowest and highest physical addresses and, in the case
657 	 * of VM_PHYSSEG_SPARSE, the exact size of the available physical
658 	 * memory.  vm_phys_early_startup() already checked that phys_avail[]
659 	 * has at least one element.
660 	 */
661 #ifdef VM_PHYSSEG_SPARSE
662 	size = phys_avail[1] - phys_avail[0];
663 #endif
664 	low_avail = phys_avail[0];
665 	high_avail = phys_avail[1];
666 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
667 #ifdef VM_PHYSSEG_SPARSE
668 		size += phys_avail[i + 1] - phys_avail[i];
669 #endif
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 	for (i = 0; i < vm_phys_nsegs; i++) {
676 #ifdef VM_PHYSSEG_SPARSE
677 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
678 #endif
679 		if (vm_phys_segs[i].start < low_avail)
680 			low_avail = vm_phys_segs[i].start;
681 		if (vm_phys_segs[i].end > high_avail)
682 			high_avail = vm_phys_segs[i].end;
683 	}
684 	first_page = low_avail / PAGE_SIZE;
685 #ifdef VM_PHYSSEG_DENSE
686 	size = high_avail - low_avail;
687 #endif
688 
689 #ifdef PMAP_HAS_PAGE_ARRAY
690 	pmap_page_array_startup(size / PAGE_SIZE);
691 	biggestone = vm_phys_avail_largest();
692 	end = new_end = phys_avail[biggestone + 1];
693 #else
694 #ifdef VM_PHYSSEG_DENSE
695 	/*
696 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
697 	 * the overhead of a page structure per page only if vm_page_array is
698 	 * allocated from the last physical memory chunk.  Otherwise, we must
699 	 * allocate page structures representing the physical memory
700 	 * underlying vm_page_array, even though they will not be used.
701 	 */
702 	if (new_end != high_avail)
703 		page_range = size / PAGE_SIZE;
704 	else
705 #endif
706 	{
707 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
708 
709 		/*
710 		 * If the partial bytes remaining are large enough for
711 		 * a page (PAGE_SIZE) without a corresponding
712 		 * 'struct vm_page', then new_end will contain an
713 		 * extra page after subtracting the length of the VM
714 		 * page array.  Compensate by subtracting an extra
715 		 * page from new_end.
716 		 */
717 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
718 			if (new_end == high_avail)
719 				high_avail -= PAGE_SIZE;
720 			new_end -= PAGE_SIZE;
721 		}
722 	}
723 	end = new_end;
724 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
725 #endif
726 
727 #if VM_NRESERVLEVEL > 0
728 	/*
729 	 * Allocate physical memory for the reservation management system's
730 	 * data structures, and map it.
731 	 */
732 	new_end = vm_reserv_startup(&vaddr, new_end);
733 #endif
734 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
735 	/*
736 	 * Include vm_page_array and vm_reserv_array in a crash dump.
737 	 */
738 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
739 		dump_add_page(pa);
740 #endif
741 	phys_avail[biggestone + 1] = new_end;
742 
743 	/*
744 	 * Add physical memory segments corresponding to the available
745 	 * physical pages.
746 	 */
747 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
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 	pool = VM_FREEPOOL_DEFAULT;
756 #ifdef VM_FREEPOOL_LAZYINIT
757 	lazyinit = 1;
758 	TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit);
759 	if (lazyinit)
760 		pool = VM_FREEPOOL_LAZYINIT;
761 #endif
762 
763 	/*
764 	 * Initialize the page structures and add every available page to the
765 	 * physical memory allocator's free lists.
766 	 */
767 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
768 	for (ii = 0; ii < vm_page_array_size; ii++) {
769 		m = &vm_page_array[ii];
770 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0,
771 		    VM_FREEPOOL_DEFAULT);
772 		m->flags = PG_FICTITIOUS;
773 	}
774 #endif
775 	vm_cnt.v_page_count = 0;
776 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
777 		seg = &vm_phys_segs[segind];
778 
779 		/*
780 		 * Initialize pages not covered by phys_avail[], since they
781 		 * might be freed to the allocator at some future point, e.g.,
782 		 * by kmem_bootstrap_free().
783 		 */
784 		startp = seg->start;
785 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
786 			if (startp >= seg->end)
787 				break;
788 			if (phys_avail[i + 1] < startp)
789 				continue;
790 			if (phys_avail[i] <= startp) {
791 				startp = phys_avail[i + 1];
792 				continue;
793 			}
794 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
795 			for (endp = MIN(phys_avail[i], seg->end);
796 			    startp < endp; startp += PAGE_SIZE, m++) {
797 				vm_page_init_page(m, startp, segind,
798 				    VM_FREEPOOL_DEFAULT);
799 			}
800 		}
801 
802 		/*
803 		 * Add the segment's pages that are covered by one of
804 		 * phys_avail's ranges to the free lists.
805 		 */
806 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
807 			if (seg->end <= phys_avail[i] ||
808 			    seg->start >= phys_avail[i + 1])
809 				continue;
810 
811 			startp = MAX(seg->start, phys_avail[i]);
812 			endp = MIN(seg->end, phys_avail[i + 1]);
813 			pagecount = (u_long)atop(endp - startp);
814 			if (pagecount == 0)
815 				continue;
816 
817 			/*
818 			 * If lazy vm_page initialization is not enabled, simply
819 			 * initialize all of the pages in the segment covered by
820 			 * phys_avail.  Otherwise, initialize only the first
821 			 * page of each run of free pages handed to the vm_phys
822 			 * allocator, which in turn defers initialization of
823 			 * pages until they are needed.
824 			 *
825 			 * This avoids blocking the boot process for long
826 			 * periods, which may be relevant for VMs (which ought
827 			 * to boot as quickly as possible) and/or systems with
828 			 * large amounts of physical memory.
829 			 */
830 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
831 			vm_page_init_page(m, startp, segind, pool);
832 			if (pool == VM_FREEPOOL_DEFAULT) {
833 				for (u_long j = 1; j < pagecount; j++) {
834 					vm_page_init_page(&m[j],
835 					    startp + ptoa((vm_paddr_t)j),
836 					    segind, pool);
837 				}
838 			}
839 			vmd = VM_DOMAIN(seg->domain);
840 			vm_domain_free_lock(vmd);
841 			vm_phys_enqueue_contig(m, pool, pagecount);
842 			vm_domain_free_unlock(vmd);
843 			vm_domain_freecnt_inc(vmd, pagecount);
844 			vm_cnt.v_page_count += (u_int)pagecount;
845 			vmd->vmd_page_count += (u_int)pagecount;
846 			vmd->vmd_segs |= 1UL << segind;
847 		}
848 	}
849 
850 	/*
851 	 * Remove blacklisted pages from the physical memory allocator.
852 	 */
853 	TAILQ_INIT(&blacklist_head);
854 	vm_page_blacklist_load(&list, &listend);
855 	vm_page_blacklist_check(list, listend);
856 
857 	list = kern_getenv("vm.blacklist");
858 	vm_page_blacklist_check(list, NULL);
859 
860 	freeenv(list);
861 #if VM_NRESERVLEVEL > 0
862 	/*
863 	 * Initialize the reservation management system.
864 	 */
865 	vm_reserv_init();
866 #endif
867 
868 	return (vaddr);
869 }
870 
871 void
872 vm_page_reference(vm_page_t m)
873 {
874 
875 	vm_page_aflag_set(m, PGA_REFERENCED);
876 }
877 
878 /*
879  *	vm_page_trybusy
880  *
881  *	Helper routine for grab functions to trylock busy.
882  *
883  *	Returns true on success and false on failure.
884  */
885 static bool
886 vm_page_trybusy(vm_page_t m, int allocflags)
887 {
888 
889 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
890 		return (vm_page_trysbusy(m));
891 	else
892 		return (vm_page_tryxbusy(m));
893 }
894 
895 /*
896  *	vm_page_tryacquire
897  *
898  *	Helper routine for grab functions to trylock busy and wire.
899  *
900  *	Returns true on success and false on failure.
901  */
902 static inline bool
903 vm_page_tryacquire(vm_page_t m, int allocflags)
904 {
905 	bool locked;
906 
907 	locked = vm_page_trybusy(m, allocflags);
908 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
909 		vm_page_wire(m);
910 	return (locked);
911 }
912 
913 /*
914  *	vm_page_busy_acquire:
915  *
916  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
917  *	and drop the object lock if necessary.
918  */
919 bool
920 vm_page_busy_acquire(vm_page_t m, int allocflags)
921 {
922 	vm_object_t obj;
923 	bool locked;
924 
925 	/*
926 	 * The page-specific object must be cached because page
927 	 * identity can change during the sleep, causing the
928 	 * re-lock of a different object.
929 	 * It is assumed that a reference to the object is already
930 	 * held by the callers.
931 	 */
932 	obj = atomic_load_ptr(&m->object);
933 	for (;;) {
934 		if (vm_page_tryacquire(m, allocflags))
935 			return (true);
936 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
937 			return (false);
938 		if (obj != NULL)
939 			locked = VM_OBJECT_WOWNED(obj);
940 		else
941 			locked = false;
942 		MPASS(locked || vm_page_wired(m));
943 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
944 		    locked) && locked)
945 			VM_OBJECT_WLOCK(obj);
946 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
947 			return (false);
948 		KASSERT(m->object == obj || m->object == NULL,
949 		    ("vm_page_busy_acquire: page %p does not belong to %p",
950 		    m, obj));
951 	}
952 }
953 
954 /*
955  *	vm_page_busy_downgrade:
956  *
957  *	Downgrade an exclusive busy page into a single shared busy page.
958  */
959 void
960 vm_page_busy_downgrade(vm_page_t m)
961 {
962 	u_int x;
963 
964 	vm_page_assert_xbusied(m);
965 
966 	x = vm_page_busy_fetch(m);
967 	for (;;) {
968 		if (atomic_fcmpset_rel_int(&m->busy_lock,
969 		    &x, VPB_SHARERS_WORD(1)))
970 			break;
971 	}
972 	if ((x & VPB_BIT_WAITERS) != 0)
973 		wakeup(m);
974 }
975 
976 /*
977  *
978  *	vm_page_busy_tryupgrade:
979  *
980  *	Attempt to upgrade a single shared busy into an exclusive busy.
981  */
982 int
983 vm_page_busy_tryupgrade(vm_page_t m)
984 {
985 	u_int ce, x;
986 
987 	vm_page_assert_sbusied(m);
988 
989 	x = vm_page_busy_fetch(m);
990 	ce = VPB_CURTHREAD_EXCLUSIVE;
991 	for (;;) {
992 		if (VPB_SHARERS(x) > 1)
993 			return (0);
994 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
995 		    ("vm_page_busy_tryupgrade: invalid lock state"));
996 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
997 		    ce | (x & VPB_BIT_WAITERS)))
998 			continue;
999 		return (1);
1000 	}
1001 }
1002 
1003 /*
1004  *	vm_page_sbusied:
1005  *
1006  *	Return a positive value if the page is shared busied, 0 otherwise.
1007  */
1008 int
1009 vm_page_sbusied(vm_page_t m)
1010 {
1011 	u_int x;
1012 
1013 	x = vm_page_busy_fetch(m);
1014 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1015 }
1016 
1017 /*
1018  *	vm_page_sunbusy:
1019  *
1020  *	Shared unbusy a page.
1021  */
1022 void
1023 vm_page_sunbusy(vm_page_t m)
1024 {
1025 	u_int x;
1026 
1027 	vm_page_assert_sbusied(m);
1028 
1029 	x = vm_page_busy_fetch(m);
1030 	for (;;) {
1031 		KASSERT(x != VPB_FREED,
1032 		    ("vm_page_sunbusy: Unlocking freed page."));
1033 		if (VPB_SHARERS(x) > 1) {
1034 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1035 			    x - VPB_ONE_SHARER))
1036 				break;
1037 			continue;
1038 		}
1039 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1040 		    ("vm_page_sunbusy: invalid lock state"));
1041 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1042 			continue;
1043 		if ((x & VPB_BIT_WAITERS) == 0)
1044 			break;
1045 		wakeup(m);
1046 		break;
1047 	}
1048 }
1049 
1050 /*
1051  *	vm_page_busy_sleep:
1052  *
1053  *	Sleep if the page is busy, using the page pointer as wchan.
1054  *	This is used to implement the hard-path of the busying mechanism.
1055  *
1056  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1057  *	will not sleep if the page is shared-busy.
1058  *
1059  *	The object lock must be held on entry.
1060  *
1061  *	Returns true if it slept and dropped the object lock, or false
1062  *	if there was no sleep and the lock is still held.
1063  */
1064 bool
1065 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1066 {
1067 	vm_object_t obj;
1068 
1069 	obj = m->object;
1070 	VM_OBJECT_ASSERT_LOCKED(obj);
1071 
1072 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1073 	    true));
1074 }
1075 
1076 /*
1077  *	vm_page_busy_sleep_unlocked:
1078  *
1079  *	Sleep if the page is busy, using the page pointer as wchan.
1080  *	This is used to implement the hard-path of busying mechanism.
1081  *
1082  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1083  *	will not sleep if the page is shared-busy.
1084  *
1085  *	The object lock must not be held on entry.  The operation will
1086  *	return if the page changes identity.
1087  */
1088 void
1089 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1090     const char *wmesg, int allocflags)
1091 {
1092 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1093 
1094 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1095 }
1096 
1097 /*
1098  *	_vm_page_busy_sleep:
1099  *
1100  *	Internal busy sleep function.  Verifies the page identity and
1101  *	lockstate against parameters.  Returns true if it sleeps and
1102  *	false otherwise.
1103  *
1104  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1105  *
1106  *	If locked is true the lock will be dropped for any true returns
1107  *	and held for any false returns.
1108  */
1109 static bool
1110 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1111     const char *wmesg, int allocflags, bool locked)
1112 {
1113 	bool xsleep;
1114 	u_int x;
1115 
1116 	/*
1117 	 * If the object is busy we must wait for that to drain to zero
1118 	 * before trying the page again.
1119 	 */
1120 	if (obj != NULL && vm_object_busied(obj)) {
1121 		if (locked)
1122 			VM_OBJECT_DROP(obj);
1123 		vm_object_busy_wait(obj, wmesg);
1124 		return (true);
1125 	}
1126 
1127 	if (!vm_page_busied(m))
1128 		return (false);
1129 
1130 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1131 	sleepq_lock(m);
1132 	x = vm_page_busy_fetch(m);
1133 	do {
1134 		/*
1135 		 * If the page changes objects or becomes unlocked we can
1136 		 * simply return.
1137 		 */
1138 		if (x == VPB_UNBUSIED ||
1139 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1140 		    m->object != obj || m->pindex != pindex) {
1141 			sleepq_release(m);
1142 			return (false);
1143 		}
1144 		if ((x & VPB_BIT_WAITERS) != 0)
1145 			break;
1146 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1147 	if (locked)
1148 		VM_OBJECT_DROP(obj);
1149 	DROP_GIANT();
1150 	sleepq_add(m, NULL, wmesg, 0, 0);
1151 	sleepq_wait(m, PVM);
1152 	PICKUP_GIANT();
1153 	return (true);
1154 }
1155 
1156 /*
1157  *	vm_page_trysbusy:
1158  *
1159  *	Try to shared busy a page.
1160  *	If the operation succeeds 1 is returned otherwise 0.
1161  *	The operation never sleeps.
1162  */
1163 int
1164 vm_page_trysbusy(vm_page_t m)
1165 {
1166 	vm_object_t obj;
1167 	u_int x;
1168 
1169 	obj = m->object;
1170 	x = vm_page_busy_fetch(m);
1171 	for (;;) {
1172 		if ((x & VPB_BIT_SHARED) == 0)
1173 			return (0);
1174 		/*
1175 		 * Reduce the window for transient busies that will trigger
1176 		 * false negatives in vm_page_ps_test().
1177 		 */
1178 		if (obj != NULL && vm_object_busied(obj))
1179 			return (0);
1180 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1181 		    x + VPB_ONE_SHARER))
1182 			break;
1183 	}
1184 
1185 	/* Refetch the object now that we're guaranteed that it is stable. */
1186 	obj = m->object;
1187 	if (obj != NULL && vm_object_busied(obj)) {
1188 		vm_page_sunbusy(m);
1189 		return (0);
1190 	}
1191 	return (1);
1192 }
1193 
1194 /*
1195  *	vm_page_tryxbusy:
1196  *
1197  *	Try to exclusive busy a page.
1198  *	If the operation succeeds 1 is returned otherwise 0.
1199  *	The operation never sleeps.
1200  */
1201 int
1202 vm_page_tryxbusy(vm_page_t m)
1203 {
1204 	vm_object_t obj;
1205 
1206         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1207             VPB_CURTHREAD_EXCLUSIVE) == 0)
1208 		return (0);
1209 
1210 	obj = m->object;
1211 	if (obj != NULL && vm_object_busied(obj)) {
1212 		vm_page_xunbusy(m);
1213 		return (0);
1214 	}
1215 	return (1);
1216 }
1217 
1218 static void
1219 vm_page_xunbusy_hard_tail(vm_page_t m)
1220 {
1221 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1222 	/* Wake the waiter. */
1223 	wakeup(m);
1224 }
1225 
1226 /*
1227  *	vm_page_xunbusy_hard:
1228  *
1229  *	Called when unbusy has failed because there is a waiter.
1230  */
1231 void
1232 vm_page_xunbusy_hard(vm_page_t m)
1233 {
1234 	vm_page_assert_xbusied(m);
1235 	vm_page_xunbusy_hard_tail(m);
1236 }
1237 
1238 void
1239 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1240 {
1241 	vm_page_assert_xbusied_unchecked(m);
1242 	vm_page_xunbusy_hard_tail(m);
1243 }
1244 
1245 static void
1246 vm_page_busy_free(vm_page_t m)
1247 {
1248 	u_int x;
1249 
1250 	atomic_thread_fence_rel();
1251 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1252 	if ((x & VPB_BIT_WAITERS) != 0)
1253 		wakeup(m);
1254 }
1255 
1256 /*
1257  *	vm_page_unhold_pages:
1258  *
1259  *	Unhold each of the pages that is referenced by the given array.
1260  */
1261 void
1262 vm_page_unhold_pages(vm_page_t *ma, int count)
1263 {
1264 
1265 	for (; count != 0; count--) {
1266 		vm_page_unwire(*ma, PQ_ACTIVE);
1267 		ma++;
1268 	}
1269 }
1270 
1271 vm_page_t
1272 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1273 {
1274 	vm_page_t m;
1275 
1276 #ifdef VM_PHYSSEG_SPARSE
1277 	m = vm_phys_paddr_to_vm_page(pa);
1278 	if (m == NULL)
1279 		m = vm_phys_fictitious_to_vm_page(pa);
1280 	return (m);
1281 #elif defined(VM_PHYSSEG_DENSE)
1282 	long pi;
1283 
1284 	pi = atop(pa);
1285 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1286 		m = &vm_page_array[pi - first_page];
1287 		return (m);
1288 	}
1289 	return (vm_phys_fictitious_to_vm_page(pa));
1290 #else
1291 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1292 #endif
1293 }
1294 
1295 /*
1296  *	vm_page_getfake:
1297  *
1298  *	Create a fictitious page with the specified physical address and
1299  *	memory attribute.  The memory attribute is the only the machine-
1300  *	dependent aspect of a fictitious page that must be initialized.
1301  */
1302 vm_page_t
1303 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1304 {
1305 	vm_page_t m;
1306 
1307 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1308 	vm_page_initfake(m, paddr, memattr);
1309 	return (m);
1310 }
1311 
1312 void
1313 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1314 {
1315 
1316 	if ((m->flags & PG_FICTITIOUS) != 0) {
1317 		/*
1318 		 * The page's memattr might have changed since the
1319 		 * previous initialization.  Update the pmap to the
1320 		 * new memattr.
1321 		 */
1322 		goto memattr;
1323 	}
1324 	m->phys_addr = paddr;
1325 	m->a.queue = PQ_NONE;
1326 	/* Fictitious pages don't use "segind". */
1327 	m->flags = PG_FICTITIOUS;
1328 	/* Fictitious pages don't use "order" or "pool". */
1329 	m->oflags = VPO_UNMANAGED;
1330 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1331 	/* Fictitious pages are unevictable. */
1332 	m->ref_count = 1;
1333 	pmap_page_init(m);
1334 memattr:
1335 	pmap_page_set_memattr(m, memattr);
1336 }
1337 
1338 /*
1339  *	vm_page_putfake:
1340  *
1341  *	Release a fictitious page.
1342  */
1343 void
1344 vm_page_putfake(vm_page_t m)
1345 {
1346 
1347 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1348 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1349 	    ("vm_page_putfake: bad page %p", m));
1350 	vm_page_assert_xbusied(m);
1351 	vm_page_busy_free(m);
1352 	uma_zfree(fakepg_zone, m);
1353 }
1354 
1355 /*
1356  *	vm_page_updatefake:
1357  *
1358  *	Update the given fictitious page to the specified physical address and
1359  *	memory attribute.
1360  */
1361 void
1362 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1363 {
1364 
1365 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1366 	    ("vm_page_updatefake: bad page %p", m));
1367 	m->phys_addr = paddr;
1368 	pmap_page_set_memattr(m, memattr);
1369 }
1370 
1371 /*
1372  *	vm_page_free:
1373  *
1374  *	Free a page.
1375  */
1376 void
1377 vm_page_free(vm_page_t m)
1378 {
1379 
1380 	m->flags &= ~PG_ZERO;
1381 	vm_page_free_toq(m);
1382 }
1383 
1384 /*
1385  *	vm_page_free_zero:
1386  *
1387  *	Free a page to the zerod-pages queue
1388  */
1389 void
1390 vm_page_free_zero(vm_page_t m)
1391 {
1392 
1393 	m->flags |= PG_ZERO;
1394 	vm_page_free_toq(m);
1395 }
1396 
1397 /*
1398  * Unbusy and handle the page queueing for a page from a getpages request that
1399  * was optionally read ahead or behind.
1400  */
1401 void
1402 vm_page_readahead_finish(vm_page_t m)
1403 {
1404 
1405 	/* We shouldn't put invalid pages on queues. */
1406 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1407 
1408 	/*
1409 	 * Since the page is not the actually needed one, whether it should
1410 	 * be activated or deactivated is not obvious.  Empirical results
1411 	 * have shown that deactivating the page is usually the best choice,
1412 	 * unless the page is wanted by another thread.
1413 	 */
1414 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1415 		vm_page_activate(m);
1416 	else
1417 		vm_page_deactivate(m);
1418 	vm_page_xunbusy_unchecked(m);
1419 }
1420 
1421 /*
1422  * Destroy the identity of an invalid page and free it if possible.
1423  * This is intended to be used when reading a page from backing store fails.
1424  */
1425 void
1426 vm_page_free_invalid(vm_page_t m)
1427 {
1428 
1429 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1430 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1431 	KASSERT(m->object != NULL, ("page %p has no object", m));
1432 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1433 
1434 	/*
1435 	 * We may be attempting to free the page as part of the handling for an
1436 	 * I/O error, in which case the page was xbusied by a different thread.
1437 	 */
1438 	vm_page_xbusy_claim(m);
1439 
1440 	/*
1441 	 * If someone has wired this page while the object lock
1442 	 * was not held, then the thread that unwires is responsible
1443 	 * for freeing the page.  Otherwise just free the page now.
1444 	 * The wire count of this unmapped page cannot change while
1445 	 * we have the page xbusy and the page's object wlocked.
1446 	 */
1447 	if (vm_page_remove(m))
1448 		vm_page_free(m);
1449 }
1450 
1451 /*
1452  *	vm_page_dirty_KBI:		[ internal use only ]
1453  *
1454  *	Set all bits in the page's dirty field.
1455  *
1456  *	The object containing the specified page must be locked if the
1457  *	call is made from the machine-independent layer.
1458  *
1459  *	See vm_page_clear_dirty_mask().
1460  *
1461  *	This function should only be called by vm_page_dirty().
1462  */
1463 void
1464 vm_page_dirty_KBI(vm_page_t m)
1465 {
1466 
1467 	/* Refer to this operation by its public name. */
1468 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1469 	m->dirty = VM_PAGE_BITS_ALL;
1470 }
1471 
1472 /*
1473  * Insert the given page into the given object at the given pindex.  mpred is
1474  * used for memq linkage.  From vm_page_insert, iter is false, mpred is
1475  * initially NULL, and this procedure looks it up.  From vm_page_iter_insert,
1476  * iter is true and mpred is known to the caller to be valid, and may be NULL if
1477  * this will be the page with the lowest pindex.
1478  *
1479  * The procedure is marked __always_inline to suggest to the compiler to
1480  * eliminate the lookup parameter and the associated alternate branch.
1481  */
1482 static __always_inline int
1483 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1484     struct pctrie_iter *pages, bool iter, vm_page_t mpred)
1485 {
1486 	int error;
1487 
1488 	VM_OBJECT_ASSERT_WLOCKED(object);
1489 	KASSERT(m->object == NULL,
1490 	    ("vm_page_insert: page %p already inserted", m));
1491 
1492 	/*
1493 	 * Record the object/offset pair in this page.
1494 	 */
1495 	m->object = object;
1496 	m->pindex = pindex;
1497 	m->ref_count |= VPRC_OBJREF;
1498 
1499 	/*
1500 	 * Add this page to the object's radix tree, and look up mpred if
1501 	 * needed.
1502 	 */
1503 	if (iter)
1504 		error = vm_radix_iter_insert(pages, m);
1505 	else
1506 		error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred);
1507 	if (__predict_false(error != 0)) {
1508 		m->object = NULL;
1509 		m->pindex = 0;
1510 		m->ref_count &= ~VPRC_OBJREF;
1511 		return (1);
1512 	}
1513 
1514 	/*
1515 	 * Now link into the object's ordered list of backed pages.
1516 	 */
1517 	vm_page_insert_radixdone(m, object, mpred);
1518 	vm_pager_page_inserted(object, m);
1519 	return (0);
1520 }
1521 
1522 /*
1523  *	vm_page_insert:		[ internal use only ]
1524  *
1525  *	Inserts the given mem entry into the object and object list.
1526  *
1527  *	The object must be locked.
1528  */
1529 int
1530 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1531 {
1532 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL));
1533 }
1534 
1535 /*
1536  *	vm_page_iter_insert:
1537  *
1538  *	Tries to insert the page "m" into the specified object at offset
1539  *	"pindex" using the iterator "pages".  Returns 0 if the insertion was
1540  *	successful.
1541  *
1542  *	The page "mpred" must immediately precede the offset "pindex" within
1543  *	the specified object.
1544  *
1545  *	The object must be locked.
1546  */
1547 static int
1548 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object,
1549     vm_pindex_t pindex, vm_page_t mpred)
1550 {
1551 	return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred));
1552 }
1553 
1554 /*
1555  *	vm_page_insert_radixdone:
1556  *
1557  *	Complete page "m" insertion into the specified object after the
1558  *	radix trie hooking.
1559  *
1560  *	The page "mpred" must precede the offset "m->pindex" within the
1561  *	specified object.
1562  *
1563  *	The object must be locked.
1564  */
1565 static void
1566 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1567 {
1568 
1569 	VM_OBJECT_ASSERT_WLOCKED(object);
1570 	KASSERT(object != NULL && m->object == object,
1571 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1572 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1573 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1574 	if (mpred != NULL) {
1575 		KASSERT(mpred->object == object,
1576 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1577 		KASSERT(mpred->pindex < m->pindex,
1578 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1579 		KASSERT(TAILQ_NEXT(mpred, listq) == NULL ||
1580 		    m->pindex < TAILQ_NEXT(mpred, listq)->pindex,
1581 		    ("vm_page_insert_radixdone: pindex doesn't precede msucc"));
1582 	} else {
1583 		KASSERT(TAILQ_EMPTY(&object->memq) ||
1584 		    m->pindex < TAILQ_FIRST(&object->memq)->pindex,
1585 		    ("vm_page_insert_radixdone: no mpred but not first page"));
1586 	}
1587 
1588 	if (mpred != NULL)
1589 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1590 	else
1591 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1592 
1593 	/*
1594 	 * Show that the object has one more resident page.
1595 	 */
1596 	object->resident_page_count++;
1597 
1598 	/*
1599 	 * Hold the vnode until the last page is released.
1600 	 */
1601 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1602 		vhold(object->handle);
1603 
1604 	/*
1605 	 * Since we are inserting a new and possibly dirty page,
1606 	 * update the object's generation count.
1607 	 */
1608 	if (pmap_page_is_write_mapped(m))
1609 		vm_object_set_writeable_dirty(object);
1610 }
1611 
1612 /*
1613  *	vm_page_remove_radixdone
1614  *
1615  *	Complete page "m" removal from the specified object after the radix trie
1616  *	unhooking.
1617  *
1618  *	The caller is responsible for updating the page's fields to reflect this
1619  *	removal.
1620  */
1621 static void
1622 vm_page_remove_radixdone(vm_page_t m)
1623 {
1624 	vm_object_t object;
1625 
1626 	vm_page_assert_xbusied(m);
1627 	object = m->object;
1628 	VM_OBJECT_ASSERT_WLOCKED(object);
1629 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1630 	    ("page %p is missing its object ref", m));
1631 
1632 	/* Deferred free of swap space. */
1633 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1634 		vm_pager_page_unswapped(m);
1635 
1636 	vm_pager_page_removed(object, m);
1637 	m->object = NULL;
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_free_object_prep:
1658  *
1659  *	Disassociates the given page from its VM object.
1660  *
1661  *	The object must be locked, and the page must be xbusy.
1662  */
1663 static void
1664 vm_page_free_object_prep(vm_page_t m)
1665 {
1666 	KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
1667 	    ((m->object->flags & OBJ_UNMANAGED) != 0),
1668 	    ("%s: managed flag mismatch for page %p",
1669 	     __func__, m));
1670 	vm_page_assert_xbusied(m);
1671 
1672 	/*
1673 	 * The object reference can be released without an atomic
1674 	 * operation.
1675 	 */
1676 	KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
1677 	    m->ref_count == VPRC_OBJREF,
1678 	    ("%s: page %p has unexpected ref_count %u",
1679 	    __func__, m, m->ref_count));
1680 	vm_page_remove_radixdone(m);
1681 	m->ref_count -= VPRC_OBJREF;
1682 }
1683 
1684 /*
1685  *	vm_page_iter_free:
1686  *
1687  *	Free the given page, and use the iterator to remove it from the radix
1688  *	tree.
1689  */
1690 void
1691 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m)
1692 {
1693 	vm_radix_iter_remove(pages);
1694 	vm_page_free_object_prep(m);
1695 	vm_page_xunbusy(m);
1696 	m->flags &= ~PG_ZERO;
1697 	vm_page_free_toq(m);
1698 }
1699 
1700 /*
1701  *	vm_page_remove:
1702  *
1703  *	Removes the specified page from its containing object, but does not
1704  *	invalidate any backing storage.  Returns true if the object's reference
1705  *	was the last reference to the page, and false otherwise.
1706  *
1707  *	The object must be locked and the page must be exclusively busied.
1708  *	The exclusive busy will be released on return.  If this is not the
1709  *	final ref and the caller does not hold a wire reference it may not
1710  *	continue to access the page.
1711  */
1712 bool
1713 vm_page_remove(vm_page_t m)
1714 {
1715 	bool dropped;
1716 
1717 	dropped = vm_page_remove_xbusy(m);
1718 	vm_page_xunbusy(m);
1719 
1720 	return (dropped);
1721 }
1722 
1723 /*
1724  *	vm_page_iter_remove:
1725  *
1726  *	Remove the current page, and use the iterator to remove it from the
1727  *	radix tree.
1728  */
1729 bool
1730 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m)
1731 {
1732 	bool dropped;
1733 
1734 	vm_radix_iter_remove(pages);
1735 	vm_page_remove_radixdone(m);
1736 	dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1737 	vm_page_xunbusy(m);
1738 
1739 	return (dropped);
1740 }
1741 
1742 /*
1743  *	vm_page_radix_remove
1744  *
1745  *	Removes the specified page from the radix tree.
1746  */
1747 static void
1748 vm_page_radix_remove(vm_page_t m)
1749 {
1750 	vm_page_t mrem __diagused;
1751 
1752 	mrem = vm_radix_remove(&m->object->rtree, m->pindex);
1753 	KASSERT(mrem == m,
1754 	    ("removed page %p, expected page %p", mrem, m));
1755 }
1756 
1757 /*
1758  *	vm_page_remove_xbusy
1759  *
1760  *	Removes the page but leaves the xbusy held.  Returns true if this
1761  *	removed the final ref and false otherwise.
1762  */
1763 bool
1764 vm_page_remove_xbusy(vm_page_t m)
1765 {
1766 
1767 	vm_page_radix_remove(m);
1768 	vm_page_remove_radixdone(m);
1769 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1770 }
1771 
1772 /*
1773  *	vm_page_lookup:
1774  *
1775  *	Returns the page associated with the object/offset
1776  *	pair specified; if none is found, NULL is returned.
1777  *
1778  *	The object must be locked.
1779  */
1780 vm_page_t
1781 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1782 {
1783 
1784 	VM_OBJECT_ASSERT_LOCKED(object);
1785 	return (vm_radix_lookup(&object->rtree, pindex));
1786 }
1787 
1788 /*
1789  *	vm_page_iter_init:
1790  *
1791  *	Initialize iterator for vm pages.
1792  */
1793 void
1794 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object)
1795 {
1796 
1797 	vm_radix_iter_init(pages, &object->rtree);
1798 }
1799 
1800 /*
1801  *	vm_page_iter_init:
1802  *
1803  *	Initialize iterator for vm pages.
1804  */
1805 void
1806 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object,
1807     vm_pindex_t limit)
1808 {
1809 
1810 	vm_radix_iter_limit_init(pages, &object->rtree, limit);
1811 }
1812 
1813 /*
1814  *	vm_page_lookup_unlocked:
1815  *
1816  *	Returns the page associated with the object/offset pair specified;
1817  *	if none is found, NULL is returned.  The page may be no longer be
1818  *	present in the object at the time that this function returns.  Only
1819  *	useful for opportunistic checks such as inmem().
1820  */
1821 vm_page_t
1822 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1823 {
1824 
1825 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1826 }
1827 
1828 /*
1829  *	vm_page_relookup:
1830  *
1831  *	Returns a page that must already have been busied by
1832  *	the caller.  Used for bogus page replacement.
1833  */
1834 vm_page_t
1835 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1836 {
1837 	vm_page_t m;
1838 
1839 	m = vm_page_lookup_unlocked(object, pindex);
1840 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1841 	    m->object == object && m->pindex == pindex,
1842 	    ("vm_page_relookup: Invalid page %p", m));
1843 	return (m);
1844 }
1845 
1846 /*
1847  * This should only be used by lockless functions for releasing transient
1848  * incorrect acquires.  The page may have been freed after we acquired a
1849  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1850  * further to do.
1851  */
1852 static void
1853 vm_page_busy_release(vm_page_t m)
1854 {
1855 	u_int x;
1856 
1857 	x = vm_page_busy_fetch(m);
1858 	for (;;) {
1859 		if (x == VPB_FREED)
1860 			break;
1861 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1862 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1863 			    x - VPB_ONE_SHARER))
1864 				break;
1865 			continue;
1866 		}
1867 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1868 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1869 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1870 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1871 			continue;
1872 		if ((x & VPB_BIT_WAITERS) != 0)
1873 			wakeup(m);
1874 		break;
1875 	}
1876 }
1877 
1878 /*
1879  * Uses the page mnew as a replacement for an existing page at index
1880  * pindex which must be already present in the object.
1881  *
1882  * Both pages must be exclusively busied on enter.  The old page is
1883  * unbusied on exit.
1884  *
1885  * A return value of true means mold is now free.  If this is not the
1886  * final ref and the caller does not hold a wire reference it may not
1887  * continue to access the page.
1888  */
1889 static bool
1890 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1891     vm_page_t mold)
1892 {
1893 	vm_page_t mret __diagused;
1894 	bool dropped;
1895 
1896 	VM_OBJECT_ASSERT_WLOCKED(object);
1897 	vm_page_assert_xbusied(mold);
1898 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1899 	    ("vm_page_replace: page %p already in object", mnew));
1900 
1901 	/*
1902 	 * This function mostly follows vm_page_insert() and
1903 	 * vm_page_remove() without the radix, object count and vnode
1904 	 * dance.  Double check such functions for more comments.
1905 	 */
1906 
1907 	mnew->object = object;
1908 	mnew->pindex = pindex;
1909 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1910 	mret = vm_radix_replace(&object->rtree, mnew);
1911 	KASSERT(mret == mold,
1912 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1913 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1914 	    (mnew->oflags & VPO_UNMANAGED),
1915 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1916 
1917 	/* Keep the resident page list in sorted order. */
1918 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1919 	TAILQ_REMOVE(&object->memq, mold, listq);
1920 	mold->object = NULL;
1921 
1922 	/*
1923 	 * The object's resident_page_count does not change because we have
1924 	 * swapped one page for another, but the generation count should
1925 	 * change if the page is dirty.
1926 	 */
1927 	if (pmap_page_is_write_mapped(mnew))
1928 		vm_object_set_writeable_dirty(object);
1929 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1930 	vm_page_xunbusy(mold);
1931 
1932 	return (dropped);
1933 }
1934 
1935 void
1936 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1937     vm_page_t mold)
1938 {
1939 
1940 	vm_page_assert_xbusied(mnew);
1941 
1942 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1943 		vm_page_free(mold);
1944 }
1945 
1946 /*
1947  *	vm_page_iter_rename:
1948  *
1949  *	Tries to move the specified page from its current object to a new object
1950  *	and pindex, using the given iterator to remove the page from its current
1951  *	object.  Returns true if the move was successful, and false if the move
1952  *	was aborted due to a failed memory allocation.
1953  *
1954  *	Panics if a page already resides in the new object at the new pindex.
1955  *
1956  *	This routine dirties the page if it is valid, as callers are expected to
1957  *	transfer backing storage only after moving the page.  Dirtying the page
1958  *	ensures that the destination object retains the most recent copy of the
1959  *	page.
1960  *
1961  *	The objects must be locked.
1962  */
1963 bool
1964 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m,
1965     vm_object_t new_object, vm_pindex_t new_pindex)
1966 {
1967 	vm_page_t mpred;
1968 	vm_pindex_t opidx;
1969 
1970 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1971 	    ("%s: page %p is missing object ref", __func__, m));
1972 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1973 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1974 
1975 	/*
1976 	 * Create a custom version of vm_page_insert() which does not depend
1977 	 * by m_prev and can cheat on the implementation aspects of the
1978 	 * function.
1979 	 */
1980 	opidx = m->pindex;
1981 	m->pindex = new_pindex;
1982 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
1983 		m->pindex = opidx;
1984 		return (false);
1985 	}
1986 
1987 	/*
1988 	 * The operation cannot fail anymore.  The removal must happen before
1989 	 * the listq iterator is tainted.
1990 	 */
1991 	m->pindex = opidx;
1992 	vm_radix_iter_remove(old_pages);
1993 	vm_page_remove_radixdone(m);
1994 
1995 	/* Return back to the new pindex to complete vm_page_insert(). */
1996 	m->pindex = new_pindex;
1997 	m->object = new_object;
1998 
1999 	vm_page_insert_radixdone(m, new_object, mpred);
2000 	if (vm_page_any_valid(m))
2001 		vm_page_dirty(m);
2002 	vm_pager_page_inserted(new_object, m);
2003 	return (true);
2004 }
2005 
2006 /*
2007  *	vm_page_mpred:
2008  *
2009  *	Return the greatest page of the object with index <= pindex,
2010  *	or NULL, if there is none.  Assumes object lock is held.
2011  */
2012 vm_page_t
2013 vm_page_mpred(vm_object_t object, vm_pindex_t pindex)
2014 {
2015 	return (vm_radix_lookup_le(&object->rtree, pindex));
2016 }
2017 
2018 /*
2019  *	vm_page_alloc:
2020  *
2021  *	Allocate and return a page that is associated with the specified
2022  *	object and offset pair.  By default, this page is exclusive busied.
2023  *
2024  *	The caller must always specify an allocation class.
2025  *
2026  *	allocation classes:
2027  *	VM_ALLOC_NORMAL		normal process request
2028  *	VM_ALLOC_SYSTEM		system *really* needs a page
2029  *	VM_ALLOC_INTERRUPT	interrupt time request
2030  *
2031  *	optional allocation flags:
2032  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2033  *				intends to allocate
2034  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2035  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2036  *	VM_ALLOC_SBUSY		shared busy the allocated page
2037  *	VM_ALLOC_WIRED		wire the allocated page
2038  *	VM_ALLOC_ZERO		prefer a zeroed page
2039  */
2040 vm_page_t
2041 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
2042 {
2043 	struct pctrie_iter pages;
2044 
2045 	vm_page_iter_init(&pages, object);
2046 	return (vm_page_alloc_after(object, &pages, pindex, req,
2047 	    vm_page_mpred(object, pindex)));
2048 }
2049 
2050 /*
2051  * Allocate a page in the specified object with the given page index.  To
2052  * optimize insertion of the page into the object, the caller must also specify
2053  * the resident page in the object with largest index smaller than the given
2054  * page index, or NULL if no such page exists.
2055  */
2056 vm_page_t
2057 vm_page_alloc_after(vm_object_t object, struct pctrie_iter *pages,
2058     vm_pindex_t pindex, int req, vm_page_t mpred)
2059 {
2060 	struct vm_domainset_iter di;
2061 	vm_page_t m;
2062 	int domain;
2063 
2064 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2065 	do {
2066 		m = vm_page_alloc_domain_after(object, pages, pindex, domain,
2067 		    req, mpred);
2068 		if (m != NULL)
2069 			break;
2070 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2071 
2072 	return (m);
2073 }
2074 
2075 /*
2076  * Returns true if the number of free pages exceeds the minimum
2077  * for the request class and false otherwise.
2078  */
2079 static int
2080 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2081 {
2082 	u_int limit, old, new;
2083 
2084 	if (req_class == VM_ALLOC_INTERRUPT)
2085 		limit = 0;
2086 	else if (req_class == VM_ALLOC_SYSTEM)
2087 		limit = vmd->vmd_interrupt_free_min;
2088 	else
2089 		limit = vmd->vmd_free_reserved;
2090 
2091 	/*
2092 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2093 	 */
2094 	limit += npages;
2095 	old = atomic_load_int(&vmd->vmd_free_count);
2096 	do {
2097 		if (old < limit)
2098 			return (0);
2099 		new = old - npages;
2100 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2101 
2102 	/* Wake the page daemon if we've crossed the threshold. */
2103 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2104 		pagedaemon_wakeup(vmd->vmd_domain);
2105 
2106 	/* Only update bitsets on transitions. */
2107 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2108 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2109 		vm_domain_set(vmd);
2110 
2111 	return (1);
2112 }
2113 
2114 int
2115 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2116 {
2117 	int req_class;
2118 
2119 	/*
2120 	 * The page daemon is allowed to dig deeper into the free page list.
2121 	 */
2122 	req_class = req & VM_ALLOC_CLASS_MASK;
2123 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2124 		req_class = VM_ALLOC_SYSTEM;
2125 	return (_vm_domain_allocate(vmd, req_class, npages));
2126 }
2127 
2128 vm_page_t
2129 vm_page_alloc_domain_after(vm_object_t object, struct pctrie_iter *pages,
2130     vm_pindex_t pindex, int domain, int req, vm_page_t mpred)
2131 {
2132 	struct vm_domain *vmd;
2133 	vm_page_t m;
2134 	int flags;
2135 
2136 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2137 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2138 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2139 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2140 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2141 	KASSERT((req & ~VPA_FLAGS) == 0,
2142 	    ("invalid request %#x", req));
2143 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2144 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2145 	    ("invalid request %#x", req));
2146 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2147 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2148 	    (uintmax_t)pindex));
2149 	VM_OBJECT_ASSERT_WLOCKED(object);
2150 
2151 	flags = 0;
2152 	m = NULL;
2153 	if (!vm_pager_can_alloc_page(object, pindex))
2154 		return (NULL);
2155 again:
2156 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2157 		m = vm_page_alloc_nofree_domain(domain, req);
2158 		if (m != NULL)
2159 			goto found;
2160 	}
2161 #if VM_NRESERVLEVEL > 0
2162 	/*
2163 	 * Can we allocate the page from a reservation?
2164 	 */
2165 	if (vm_object_reserv(object) &&
2166 	    (m = vm_reserv_alloc_page(object, pages, pindex, domain, req)) !=
2167 	    NULL) {
2168 		goto found;
2169 	}
2170 #endif
2171 	vmd = VM_DOMAIN(domain);
2172 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2173 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2174 		    M_NOWAIT | M_NOVM);
2175 		if (m != NULL) {
2176 			flags |= PG_PCPU_CACHE;
2177 			goto found;
2178 		}
2179 	}
2180 	if (vm_domain_allocate(vmd, req, 1)) {
2181 		/*
2182 		 * If not, allocate it from the free page queues.
2183 		 */
2184 		vm_domain_free_lock(vmd);
2185 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2186 		vm_domain_free_unlock(vmd);
2187 		if (m == NULL) {
2188 			vm_domain_freecnt_inc(vmd, 1);
2189 #if VM_NRESERVLEVEL > 0
2190 			if (vm_reserv_reclaim_inactive(domain))
2191 				goto again;
2192 #endif
2193 		}
2194 	}
2195 	if (m == NULL) {
2196 		/*
2197 		 * Not allocatable, give up.
2198 		 */
2199 		pctrie_iter_reset(pages);
2200 		if (vm_domain_alloc_fail(vmd, object, req))
2201 			goto again;
2202 		return (NULL);
2203 	}
2204 
2205 	/*
2206 	 * At this point we had better have found a good page.
2207 	 */
2208 found:
2209 	vm_page_dequeue(m);
2210 	vm_page_alloc_check(m);
2211 
2212 	/*
2213 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2214 	 */
2215 	flags |= m->flags & PG_ZERO;
2216 	if ((req & VM_ALLOC_NODUMP) != 0)
2217 		flags |= PG_NODUMP;
2218 	if ((req & VM_ALLOC_NOFREE) != 0)
2219 		flags |= PG_NOFREE;
2220 	m->flags = flags;
2221 	m->a.flags = 0;
2222 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2223 	m->pool = VM_FREEPOOL_DEFAULT;
2224 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2225 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2226 	else if ((req & VM_ALLOC_SBUSY) != 0)
2227 		m->busy_lock = VPB_SHARERS_WORD(1);
2228 	else
2229 		m->busy_lock = VPB_UNBUSIED;
2230 	if (req & VM_ALLOC_WIRED) {
2231 		vm_wire_add(1);
2232 		m->ref_count = 1;
2233 	}
2234 	m->a.act_count = 0;
2235 
2236 	if (vm_page_insert_lookup(m, object, pindex, pages, true, mpred)) {
2237 		if (req & VM_ALLOC_WIRED) {
2238 			vm_wire_sub(1);
2239 			m->ref_count = 0;
2240 		}
2241 		KASSERT(m->object == NULL, ("page %p has object", m));
2242 		m->oflags = VPO_UNMANAGED;
2243 		m->busy_lock = VPB_UNBUSIED;
2244 		/* Don't change PG_ZERO. */
2245 		vm_page_free_toq(m);
2246 		if (req & VM_ALLOC_WAITFAIL) {
2247 			VM_OBJECT_WUNLOCK(object);
2248 			vm_radix_wait();
2249 			pctrie_iter_reset(pages);
2250 			VM_OBJECT_WLOCK(object);
2251 		}
2252 		return (NULL);
2253 	}
2254 
2255 	/* Ignore device objects; the pager sets "memattr" for them. */
2256 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2257 	    (object->flags & OBJ_FICTITIOUS) == 0)
2258 		pmap_page_set_memattr(m, object->memattr);
2259 
2260 	return (m);
2261 }
2262 
2263 /*
2264  *	vm_page_alloc_contig:
2265  *
2266  *	Allocate a contiguous set of physical pages of the given size "npages"
2267  *	from the free lists.  All of the physical pages must be at or above
2268  *	the given physical address "low" and below the given physical address
2269  *	"high".  The given value "alignment" determines the alignment of the
2270  *	first physical page in the set.  If the given value "boundary" is
2271  *	non-zero, then the set of physical pages cannot cross any physical
2272  *	address boundary that is a multiple of that value.  Both "alignment"
2273  *	and "boundary" must be a power of two.
2274  *
2275  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2276  *	then the memory attribute setting for the physical pages is configured
2277  *	to the object's memory attribute setting.  Otherwise, the memory
2278  *	attribute setting for the physical pages is configured to "memattr",
2279  *	overriding the object's memory attribute setting.  However, if the
2280  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2281  *	memory attribute setting for the physical pages cannot be configured
2282  *	to VM_MEMATTR_DEFAULT.
2283  *
2284  *	The specified object may not contain fictitious pages.
2285  *
2286  *	The caller must always specify an allocation class.
2287  *
2288  *	allocation classes:
2289  *	VM_ALLOC_NORMAL		normal process request
2290  *	VM_ALLOC_SYSTEM		system *really* needs a page
2291  *	VM_ALLOC_INTERRUPT	interrupt time request
2292  *
2293  *	optional allocation flags:
2294  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2295  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2296  *	VM_ALLOC_SBUSY		shared busy the allocated page
2297  *	VM_ALLOC_WIRED		wire the allocated page
2298  *	VM_ALLOC_ZERO		prefer a zeroed page
2299  */
2300 vm_page_t
2301 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2302     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2303     vm_paddr_t boundary, vm_memattr_t memattr)
2304 {
2305 	struct vm_domainset_iter di;
2306 	vm_page_t bounds[2];
2307 	vm_page_t m;
2308 	int domain;
2309 	int start_segind;
2310 
2311 	start_segind = -1;
2312 
2313 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2314 	do {
2315 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2316 		    npages, low, high, alignment, boundary, memattr);
2317 		if (m != NULL)
2318 			break;
2319 		if (start_segind == -1)
2320 			start_segind = vm_phys_lookup_segind(low);
2321 		if (vm_phys_find_range(bounds, start_segind, domain,
2322 		    npages, low, high) == -1) {
2323 			vm_domainset_iter_ignore(&di, domain);
2324 		}
2325 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2326 
2327 	return (m);
2328 }
2329 
2330 static vm_page_t
2331 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2332     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2333 {
2334 	struct vm_domain *vmd;
2335 	vm_page_t m_ret;
2336 
2337 	/*
2338 	 * Can we allocate the pages without the number of free pages falling
2339 	 * below the lower bound for the allocation class?
2340 	 */
2341 	vmd = VM_DOMAIN(domain);
2342 	if (!vm_domain_allocate(vmd, req, npages))
2343 		return (NULL);
2344 	/*
2345 	 * Try to allocate the pages from the free page queues.
2346 	 */
2347 	vm_domain_free_lock(vmd);
2348 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2349 	    alignment, boundary);
2350 	vm_domain_free_unlock(vmd);
2351 	if (m_ret != NULL)
2352 		return (m_ret);
2353 #if VM_NRESERVLEVEL > 0
2354 	/*
2355 	 * Try to break a reservation to allocate the pages.
2356 	 */
2357 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2358 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2359 	            high, alignment, boundary);
2360 		if (m_ret != NULL)
2361 			return (m_ret);
2362 	}
2363 #endif
2364 	vm_domain_freecnt_inc(vmd, npages);
2365 	return (NULL);
2366 }
2367 
2368 vm_page_t
2369 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2370     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2371     vm_paddr_t boundary, vm_memattr_t memattr)
2372 {
2373 	struct pctrie_iter pages;
2374 	vm_page_t m, m_ret, mpred;
2375 	u_int busy_lock, flags, oflags;
2376 
2377 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2378 	KASSERT((req & ~VPAC_FLAGS) == 0,
2379 	    ("invalid request %#x", req));
2380 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2381 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2382 	    ("invalid request %#x", req));
2383 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2384 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2385 	    ("invalid request %#x", req));
2386 	VM_OBJECT_ASSERT_WLOCKED(object);
2387 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2388 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2389 	    object));
2390 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2391 
2392 	vm_page_iter_init(&pages, object);
2393 	mpred = vm_radix_iter_lookup_lt(&pages, pindex);
2394 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2395 	    ("vm_page_alloc_contig: pindex already allocated"));
2396 	for (;;) {
2397 #if VM_NRESERVLEVEL > 0
2398 		/*
2399 		 * Can we allocate the pages from a reservation?
2400 		 */
2401 		if (vm_object_reserv(object) &&
2402 		    (m_ret = vm_reserv_alloc_contig(object, &pages, pindex,
2403 		    domain, req, npages, low, high, alignment, boundary)) !=
2404 		    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 		m->pool = VM_FREEPOOL_DEFAULT;
2444 		if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) {
2445 			if ((req & VM_ALLOC_WIRED) != 0)
2446 				vm_wire_sub(npages);
2447 			KASSERT(m->object == NULL,
2448 			    ("page %p has object", m));
2449 			mpred = m;
2450 			for (m = m_ret; m < &m_ret[npages]; m++) {
2451 				if (m <= mpred &&
2452 				    (req & VM_ALLOC_WIRED) != 0)
2453 					m->ref_count = 0;
2454 				m->oflags = VPO_UNMANAGED;
2455 				m->busy_lock = VPB_UNBUSIED;
2456 				/* Don't change PG_ZERO. */
2457 				vm_page_free_toq(m);
2458 			}
2459 			if (req & VM_ALLOC_WAITFAIL) {
2460 				VM_OBJECT_WUNLOCK(object);
2461 				vm_radix_wait();
2462 				VM_OBJECT_WLOCK(object);
2463 			}
2464 			return (NULL);
2465 		}
2466 		mpred = m;
2467 		if (memattr != VM_MEMATTR_DEFAULT)
2468 			pmap_page_set_memattr(m, memattr);
2469 		pindex++;
2470 	}
2471 	return (m_ret);
2472 }
2473 
2474 /*
2475  * Allocate a physical page that is not intended to be inserted into a VM
2476  * object.
2477  */
2478 vm_page_t
2479 vm_page_alloc_noobj_domain(int domain, int req)
2480 {
2481 	struct vm_domain *vmd;
2482 	vm_page_t m;
2483 	int flags;
2484 
2485 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2486 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2487 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2488 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2489 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2490 	KASSERT((req & ~VPAN_FLAGS) == 0,
2491 	    ("invalid request %#x", req));
2492 
2493 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2494 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2495 	vmd = VM_DOMAIN(domain);
2496 again:
2497 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2498 		m = vm_page_alloc_nofree_domain(domain, req);
2499 		if (m != NULL)
2500 			goto found;
2501 	}
2502 
2503 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2504 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2505 		    M_NOWAIT | M_NOVM);
2506 		if (m != NULL) {
2507 			flags |= PG_PCPU_CACHE;
2508 			goto found;
2509 		}
2510 	}
2511 
2512 	if (vm_domain_allocate(vmd, req, 1)) {
2513 		vm_domain_free_lock(vmd);
2514 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2515 		vm_domain_free_unlock(vmd);
2516 		if (m == NULL) {
2517 			vm_domain_freecnt_inc(vmd, 1);
2518 #if VM_NRESERVLEVEL > 0
2519 			if (vm_reserv_reclaim_inactive(domain))
2520 				goto again;
2521 #endif
2522 		}
2523 	}
2524 	if (m == NULL) {
2525 		if (vm_domain_alloc_fail(vmd, NULL, req))
2526 			goto again;
2527 		return (NULL);
2528 	}
2529 
2530 found:
2531 	vm_page_dequeue(m);
2532 	vm_page_alloc_check(m);
2533 
2534 	/*
2535 	 * Consumers should not rely on a useful default pindex value.
2536 	 */
2537 	m->pindex = 0xdeadc0dedeadc0de;
2538 	m->flags = (m->flags & PG_ZERO) | flags;
2539 	m->a.flags = 0;
2540 	m->oflags = VPO_UNMANAGED;
2541 	m->pool = VM_FREEPOOL_DIRECT;
2542 	m->busy_lock = VPB_UNBUSIED;
2543 	if ((req & VM_ALLOC_WIRED) != 0) {
2544 		vm_wire_add(1);
2545 		m->ref_count = 1;
2546 	}
2547 
2548 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2549 		pmap_zero_page(m);
2550 
2551 	return (m);
2552 }
2553 
2554 #if VM_NRESERVLEVEL > 1
2555 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2556 #elif VM_NRESERVLEVEL > 0
2557 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2558 #else
2559 #define	VM_NOFREE_IMPORT_ORDER	8
2560 #endif
2561 
2562 /*
2563  * Allocate a single NOFREE page.
2564  *
2565  * This routine hands out NOFREE pages from higher-order
2566  * physical memory blocks in order to reduce memory fragmentation.
2567  * When a NOFREE for a given domain chunk is used up,
2568  * the routine will try to fetch a new one from the freelists
2569  * and discard the old one.
2570  */
2571 static vm_page_t __noinline
2572 vm_page_alloc_nofree_domain(int domain, int req)
2573 {
2574 	vm_page_t m;
2575 	struct vm_domain *vmd;
2576 
2577 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2578 
2579 	vmd = VM_DOMAIN(domain);
2580 	vm_domain_free_lock(vmd);
2581 	if (TAILQ_EMPTY(&vmd->vmd_nofreeq)) {
2582 		int count;
2583 
2584 		count = 1 << VM_NOFREE_IMPORT_ORDER;
2585 		if (!vm_domain_allocate(vmd, req, count)) {
2586 			vm_domain_free_unlock(vmd);
2587 			return (NULL);
2588 		}
2589 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2590 		    VM_NOFREE_IMPORT_ORDER);
2591 		if (m == NULL) {
2592 			vm_domain_freecnt_inc(vmd, count);
2593 			vm_domain_free_unlock(vmd);
2594 			return (NULL);
2595 		}
2596 		m->pindex = count;
2597 		TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, listq);
2598 		VM_CNT_ADD(v_nofree_count, count);
2599 	}
2600 	m = TAILQ_FIRST(&vmd->vmd_nofreeq);
2601 	TAILQ_REMOVE(&vmd->vmd_nofreeq, m, listq);
2602 	if (m->pindex > 1) {
2603 		vm_page_t m_next;
2604 
2605 		m_next = &m[1];
2606 		m_next->pindex = m->pindex - 1;
2607 		TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m_next, listq);
2608 	}
2609 	vm_domain_free_unlock(vmd);
2610 	VM_CNT_ADD(v_nofree_count, -1);
2611 
2612 	return (m);
2613 }
2614 
2615 /*
2616  * Though a NOFREE page by definition should not be freed, we support putting
2617  * them aside for future NOFREE allocations.  This enables code which allocates
2618  * NOFREE pages for some purpose but then encounters an error and releases
2619  * resources.
2620  */
2621 static void __noinline
2622 vm_page_free_nofree(struct vm_domain *vmd, vm_page_t m)
2623 {
2624 	vm_domain_free_lock(vmd);
2625 	m->pindex = 1;
2626 	TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, listq);
2627 	vm_domain_free_unlock(vmd);
2628 	VM_CNT_ADD(v_nofree_count, 1);
2629 }
2630 
2631 vm_page_t
2632 vm_page_alloc_noobj(int req)
2633 {
2634 	struct vm_domainset_iter di;
2635 	vm_page_t m;
2636 	int domain;
2637 
2638 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2639 	do {
2640 		m = vm_page_alloc_noobj_domain(domain, req);
2641 		if (m != NULL)
2642 			break;
2643 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2644 
2645 	return (m);
2646 }
2647 
2648 vm_page_t
2649 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2650     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2651     vm_memattr_t memattr)
2652 {
2653 	struct vm_domainset_iter di;
2654 	vm_page_t m;
2655 	int domain;
2656 
2657 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2658 	do {
2659 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2660 		    high, alignment, boundary, memattr);
2661 		if (m != NULL)
2662 			break;
2663 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2664 
2665 	return (m);
2666 }
2667 
2668 vm_page_t
2669 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2670     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2671     vm_memattr_t memattr)
2672 {
2673 	vm_page_t m, m_ret;
2674 	u_int flags;
2675 
2676 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2677 	KASSERT((req & ~VPANC_FLAGS) == 0,
2678 	    ("invalid request %#x", req));
2679 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2680 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2681 	    ("invalid request %#x", req));
2682 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2683 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2684 	    ("invalid request %#x", req));
2685 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2686 
2687 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2688 	    low, high, alignment, boundary)) == NULL) {
2689 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2690 			return (NULL);
2691 	}
2692 
2693 	/*
2694 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2695 	 */
2696 	flags = PG_ZERO;
2697 	if ((req & VM_ALLOC_NODUMP) != 0)
2698 		flags |= PG_NODUMP;
2699 	if ((req & VM_ALLOC_WIRED) != 0)
2700 		vm_wire_add(npages);
2701 	for (m = m_ret; m < &m_ret[npages]; m++) {
2702 		vm_page_dequeue(m);
2703 		vm_page_alloc_check(m);
2704 
2705 		/*
2706 		 * Consumers should not rely on a useful default pindex value.
2707 		 */
2708 		m->pindex = 0xdeadc0dedeadc0de;
2709 		m->a.flags = 0;
2710 		m->flags = (m->flags | PG_NODUMP) & flags;
2711 		m->busy_lock = VPB_UNBUSIED;
2712 		if ((req & VM_ALLOC_WIRED) != 0)
2713 			m->ref_count = 1;
2714 		m->a.act_count = 0;
2715 		m->oflags = VPO_UNMANAGED;
2716 		m->pool = VM_FREEPOOL_DIRECT;
2717 
2718 		/*
2719 		 * Zero the page before updating any mappings since the page is
2720 		 * not yet shared with any devices which might require the
2721 		 * non-default memory attribute.  pmap_page_set_memattr()
2722 		 * flushes data caches before returning.
2723 		 */
2724 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2725 			pmap_zero_page(m);
2726 		if (memattr != VM_MEMATTR_DEFAULT)
2727 			pmap_page_set_memattr(m, memattr);
2728 	}
2729 	return (m_ret);
2730 }
2731 
2732 /*
2733  * Check a page that has been freshly dequeued from a freelist.
2734  */
2735 static void
2736 vm_page_alloc_check(vm_page_t m)
2737 {
2738 
2739 	KASSERT(m->object == NULL, ("page %p has object", m));
2740 	KASSERT(m->a.queue == PQ_NONE &&
2741 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2742 	    ("page %p has unexpected queue %d, flags %#x",
2743 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2744 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2745 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2746 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2747 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2748 	    ("page %p has unexpected memattr %d",
2749 	    m, pmap_page_get_memattr(m)));
2750 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2751 	pmap_vm_page_alloc_check(m);
2752 }
2753 
2754 static int
2755 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2756 {
2757 	struct vm_domain *vmd;
2758 	struct vm_pgcache *pgcache;
2759 	int i;
2760 
2761 	pgcache = arg;
2762 	vmd = VM_DOMAIN(pgcache->domain);
2763 
2764 	/*
2765 	 * The page daemon should avoid creating extra memory pressure since its
2766 	 * main purpose is to replenish the store of free pages.
2767 	 */
2768 	if (vmd->vmd_severeset || curproc == pageproc ||
2769 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2770 		return (0);
2771 	domain = vmd->vmd_domain;
2772 	vm_domain_free_lock(vmd);
2773 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2774 	    (vm_page_t *)store);
2775 	vm_domain_free_unlock(vmd);
2776 	if (cnt != i)
2777 		vm_domain_freecnt_inc(vmd, cnt - i);
2778 
2779 	return (i);
2780 }
2781 
2782 static void
2783 vm_page_zone_release(void *arg, void **store, int cnt)
2784 {
2785 	struct vm_domain *vmd;
2786 	struct vm_pgcache *pgcache;
2787 	vm_page_t m;
2788 	int i;
2789 
2790 	pgcache = arg;
2791 	vmd = VM_DOMAIN(pgcache->domain);
2792 	vm_domain_free_lock(vmd);
2793 	for (i = 0; i < cnt; i++) {
2794 		m = (vm_page_t)store[i];
2795 		vm_phys_free_pages(m, pgcache->pool, 0);
2796 	}
2797 	vm_domain_free_unlock(vmd);
2798 	vm_domain_freecnt_inc(vmd, cnt);
2799 }
2800 
2801 #define	VPSC_ANY	0	/* No restrictions. */
2802 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2803 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2804 
2805 /*
2806  *	vm_page_scan_contig:
2807  *
2808  *	Scan vm_page_array[] between the specified entries "m_start" and
2809  *	"m_end" for a run of contiguous physical pages that satisfy the
2810  *	specified conditions, and return the lowest page in the run.  The
2811  *	specified "alignment" determines the alignment of the lowest physical
2812  *	page in the run.  If the specified "boundary" is non-zero, then the
2813  *	run of physical pages cannot span a physical address that is a
2814  *	multiple of "boundary".
2815  *
2816  *	"m_end" is never dereferenced, so it need not point to a vm_page
2817  *	structure within vm_page_array[].
2818  *
2819  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2820  *	span a hole (or discontiguity) in the physical address space.  Both
2821  *	"alignment" and "boundary" must be a power of two.
2822  */
2823 static vm_page_t
2824 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2825     u_long alignment, vm_paddr_t boundary, int options)
2826 {
2827 	vm_object_t object;
2828 	vm_paddr_t pa;
2829 	vm_page_t m, m_run;
2830 #if VM_NRESERVLEVEL > 0
2831 	int level;
2832 #endif
2833 	int m_inc, order, run_ext, run_len;
2834 
2835 	KASSERT(npages > 0, ("npages is 0"));
2836 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2837 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2838 	m_run = NULL;
2839 	run_len = 0;
2840 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2841 		KASSERT((m->flags & PG_MARKER) == 0,
2842 		    ("page %p is PG_MARKER", m));
2843 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2844 		    ("fictitious page %p has invalid ref count", m));
2845 
2846 		/*
2847 		 * If the current page would be the start of a run, check its
2848 		 * physical address against the end, alignment, and boundary
2849 		 * conditions.  If it doesn't satisfy these conditions, either
2850 		 * terminate the scan or advance to the next page that
2851 		 * satisfies the failed condition.
2852 		 */
2853 		if (run_len == 0) {
2854 			KASSERT(m_run == NULL, ("m_run != NULL"));
2855 			if (m + npages > m_end)
2856 				break;
2857 			pa = VM_PAGE_TO_PHYS(m);
2858 			if (!vm_addr_align_ok(pa, alignment)) {
2859 				m_inc = atop(roundup2(pa, alignment) - pa);
2860 				continue;
2861 			}
2862 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2863 				m_inc = atop(roundup2(pa, boundary) - pa);
2864 				continue;
2865 			}
2866 		} else
2867 			KASSERT(m_run != NULL, ("m_run == NULL"));
2868 
2869 retry:
2870 		m_inc = 1;
2871 		if (vm_page_wired(m))
2872 			run_ext = 0;
2873 #if VM_NRESERVLEVEL > 0
2874 		else if ((level = vm_reserv_level(m)) >= 0 &&
2875 		    (options & VPSC_NORESERV) != 0) {
2876 			run_ext = 0;
2877 			/* Advance to the end of the reservation. */
2878 			pa = VM_PAGE_TO_PHYS(m);
2879 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2880 			    pa);
2881 		}
2882 #endif
2883 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2884 			/*
2885 			 * The page is considered eligible for relocation if
2886 			 * and only if it could be laundered or reclaimed by
2887 			 * the page daemon.
2888 			 */
2889 			VM_OBJECT_RLOCK(object);
2890 			if (object != m->object) {
2891 				VM_OBJECT_RUNLOCK(object);
2892 				goto retry;
2893 			}
2894 			/* Don't care: PG_NODUMP, PG_ZERO. */
2895 			if ((object->flags & OBJ_SWAP) == 0 &&
2896 			    object->type != OBJT_VNODE) {
2897 				run_ext = 0;
2898 #if VM_NRESERVLEVEL > 0
2899 			} else if ((options & VPSC_NOSUPER) != 0 &&
2900 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2901 				run_ext = 0;
2902 				/* Advance to the end of the superpage. */
2903 				pa = VM_PAGE_TO_PHYS(m);
2904 				m_inc = atop(roundup2(pa + 1,
2905 				    vm_reserv_size(level)) - pa);
2906 #endif
2907 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2908 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2909 				/*
2910 				 * The page is allocated but eligible for
2911 				 * relocation.  Extend the current run by one
2912 				 * page.
2913 				 */
2914 				KASSERT(pmap_page_get_memattr(m) ==
2915 				    VM_MEMATTR_DEFAULT,
2916 				    ("page %p has an unexpected memattr", m));
2917 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2918 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2919 				    ("page %p has unexpected oflags", m));
2920 				/* Don't care: PGA_NOSYNC. */
2921 				run_ext = 1;
2922 			} else
2923 				run_ext = 0;
2924 			VM_OBJECT_RUNLOCK(object);
2925 #if VM_NRESERVLEVEL > 0
2926 		} else if (level >= 0) {
2927 			/*
2928 			 * The page is reserved but not yet allocated.  In
2929 			 * other words, it is still free.  Extend the current
2930 			 * run by one page.
2931 			 */
2932 			run_ext = 1;
2933 #endif
2934 		} else if ((order = m->order) < VM_NFREEORDER) {
2935 			/*
2936 			 * The page is enqueued in the physical memory
2937 			 * allocator's free page queues.  Moreover, it is the
2938 			 * first page in a power-of-two-sized run of
2939 			 * contiguous free pages.  Add these pages to the end
2940 			 * of the current run, and jump ahead.
2941 			 */
2942 			run_ext = 1 << order;
2943 			m_inc = 1 << order;
2944 		} else {
2945 			/*
2946 			 * Skip the page for one of the following reasons: (1)
2947 			 * It is enqueued in the physical memory allocator's
2948 			 * free page queues.  However, it is not the first
2949 			 * page in a run of contiguous free pages.  (This case
2950 			 * rarely occurs because the scan is performed in
2951 			 * ascending order.) (2) It is not reserved, and it is
2952 			 * transitioning from free to allocated.  (Conversely,
2953 			 * the transition from allocated to free for managed
2954 			 * pages is blocked by the page busy lock.) (3) It is
2955 			 * allocated but not contained by an object and not
2956 			 * wired, e.g., allocated by Xen's balloon driver.
2957 			 */
2958 			run_ext = 0;
2959 		}
2960 
2961 		/*
2962 		 * Extend or reset the current run of pages.
2963 		 */
2964 		if (run_ext > 0) {
2965 			if (run_len == 0)
2966 				m_run = m;
2967 			run_len += run_ext;
2968 		} else {
2969 			if (run_len > 0) {
2970 				m_run = NULL;
2971 				run_len = 0;
2972 			}
2973 		}
2974 	}
2975 	if (run_len >= npages)
2976 		return (m_run);
2977 	return (NULL);
2978 }
2979 
2980 /*
2981  *	vm_page_reclaim_run:
2982  *
2983  *	Try to relocate each of the allocated virtual pages within the
2984  *	specified run of physical pages to a new physical address.  Free the
2985  *	physical pages underlying the relocated virtual pages.  A virtual page
2986  *	is relocatable if and only if it could be laundered or reclaimed by
2987  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2988  *	physical address above "high".
2989  *
2990  *	Returns 0 if every physical page within the run was already free or
2991  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2992  *	value indicating why the last attempt to relocate a virtual page was
2993  *	unsuccessful.
2994  *
2995  *	"req_class" must be an allocation class.
2996  */
2997 static int
2998 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2999     vm_paddr_t high)
3000 {
3001 	struct vm_domain *vmd;
3002 	struct spglist free;
3003 	vm_object_t object;
3004 	vm_paddr_t pa;
3005 	vm_page_t m, m_end, m_new;
3006 	int error, order, req;
3007 
3008 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
3009 	    ("req_class is not an allocation class"));
3010 	SLIST_INIT(&free);
3011 	error = 0;
3012 	m = m_run;
3013 	m_end = m_run + npages;
3014 	for (; error == 0 && m < m_end; m++) {
3015 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
3016 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
3017 
3018 		/*
3019 		 * Racily check for wirings.  Races are handled once the object
3020 		 * lock is held and the page is unmapped.
3021 		 */
3022 		if (vm_page_wired(m))
3023 			error = EBUSY;
3024 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
3025 			/*
3026 			 * The page is relocated if and only if it could be
3027 			 * laundered or reclaimed by the page daemon.
3028 			 */
3029 			VM_OBJECT_WLOCK(object);
3030 			/* Don't care: PG_NODUMP, PG_ZERO. */
3031 			if (m->object != object ||
3032 			    ((object->flags & OBJ_SWAP) == 0 &&
3033 			    object->type != OBJT_VNODE))
3034 				error = EINVAL;
3035 			else if (object->memattr != VM_MEMATTR_DEFAULT)
3036 				error = EINVAL;
3037 			else if (vm_page_queue(m) != PQ_NONE &&
3038 			    vm_page_tryxbusy(m) != 0) {
3039 				if (vm_page_wired(m)) {
3040 					vm_page_xunbusy(m);
3041 					error = EBUSY;
3042 					goto unlock;
3043 				}
3044 				KASSERT(pmap_page_get_memattr(m) ==
3045 				    VM_MEMATTR_DEFAULT,
3046 				    ("page %p has an unexpected memattr", m));
3047 				KASSERT(m->oflags == 0,
3048 				    ("page %p has unexpected oflags", m));
3049 				/* Don't care: PGA_NOSYNC. */
3050 				if (!vm_page_none_valid(m)) {
3051 					/*
3052 					 * First, try to allocate a new page
3053 					 * that is above "high".  Failing
3054 					 * that, try to allocate a new page
3055 					 * that is below "m_run".  Allocate
3056 					 * the new page between the end of
3057 					 * "m_run" and "high" only as a last
3058 					 * resort.
3059 					 */
3060 					req = req_class;
3061 					if ((m->flags & PG_NODUMP) != 0)
3062 						req |= VM_ALLOC_NODUMP;
3063 					if (trunc_page(high) !=
3064 					    ~(vm_paddr_t)PAGE_MASK) {
3065 						m_new =
3066 						    vm_page_alloc_noobj_contig(
3067 						    req, 1, round_page(high),
3068 						    ~(vm_paddr_t)0, PAGE_SIZE,
3069 						    0, VM_MEMATTR_DEFAULT);
3070 					} else
3071 						m_new = NULL;
3072 					if (m_new == NULL) {
3073 						pa = VM_PAGE_TO_PHYS(m_run);
3074 						m_new =
3075 						    vm_page_alloc_noobj_contig(
3076 						    req, 1, 0, pa - 1,
3077 						    PAGE_SIZE, 0,
3078 						    VM_MEMATTR_DEFAULT);
3079 					}
3080 					if (m_new == NULL) {
3081 						pa += ptoa(npages);
3082 						m_new =
3083 						    vm_page_alloc_noobj_contig(
3084 						    req, 1, pa, high, PAGE_SIZE,
3085 						    0, VM_MEMATTR_DEFAULT);
3086 					}
3087 					if (m_new == NULL) {
3088 						vm_page_xunbusy(m);
3089 						error = ENOMEM;
3090 						goto unlock;
3091 					}
3092 
3093 					/*
3094 					 * Unmap the page and check for new
3095 					 * wirings that may have been acquired
3096 					 * through a pmap lookup.
3097 					 */
3098 					if (object->ref_count != 0 &&
3099 					    !vm_page_try_remove_all(m)) {
3100 						vm_page_xunbusy(m);
3101 						vm_page_free(m_new);
3102 						error = EBUSY;
3103 						goto unlock;
3104 					}
3105 
3106 					/*
3107 					 * Replace "m" with the new page.  For
3108 					 * vm_page_replace(), "m" must be busy
3109 					 * and dequeued.  Finally, change "m"
3110 					 * as if vm_page_free() was called.
3111 					 */
3112 					m_new->a.flags = m->a.flags &
3113 					    ~PGA_QUEUE_STATE_MASK;
3114 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3115 					    ("page %p is managed", m_new));
3116 					m_new->oflags = 0;
3117 					pmap_copy_page(m, m_new);
3118 					m_new->valid = m->valid;
3119 					m_new->dirty = m->dirty;
3120 					m->flags &= ~PG_ZERO;
3121 					vm_page_dequeue(m);
3122 					if (vm_page_replace_hold(m_new, object,
3123 					    m->pindex, m) &&
3124 					    vm_page_free_prep(m))
3125 						SLIST_INSERT_HEAD(&free, m,
3126 						    plinks.s.ss);
3127 
3128 					/*
3129 					 * The new page must be deactivated
3130 					 * before the object is unlocked.
3131 					 */
3132 					vm_page_deactivate(m_new);
3133 				} else {
3134 					m->flags &= ~PG_ZERO;
3135 					vm_page_dequeue(m);
3136 					if (vm_page_free_prep(m))
3137 						SLIST_INSERT_HEAD(&free, m,
3138 						    plinks.s.ss);
3139 					KASSERT(m->dirty == 0,
3140 					    ("page %p is dirty", m));
3141 				}
3142 			} else
3143 				error = EBUSY;
3144 unlock:
3145 			VM_OBJECT_WUNLOCK(object);
3146 		} else {
3147 			MPASS(vm_page_domain(m) == domain);
3148 			vmd = VM_DOMAIN(domain);
3149 			vm_domain_free_lock(vmd);
3150 			order = m->order;
3151 			if (order < VM_NFREEORDER) {
3152 				/*
3153 				 * The page is enqueued in the physical memory
3154 				 * allocator's free page queues.  Moreover, it
3155 				 * is the first page in a power-of-two-sized
3156 				 * run of contiguous free pages.  Jump ahead
3157 				 * to the last page within that run, and
3158 				 * continue from there.
3159 				 */
3160 				m += (1 << order) - 1;
3161 			}
3162 #if VM_NRESERVLEVEL > 0
3163 			else if (vm_reserv_is_page_free(m))
3164 				order = 0;
3165 #endif
3166 			vm_domain_free_unlock(vmd);
3167 			if (order == VM_NFREEORDER)
3168 				error = EINVAL;
3169 		}
3170 	}
3171 	if ((m = SLIST_FIRST(&free)) != NULL) {
3172 		int cnt;
3173 
3174 		vmd = VM_DOMAIN(domain);
3175 		cnt = 0;
3176 		vm_domain_free_lock(vmd);
3177 		do {
3178 			MPASS(vm_page_domain(m) == domain);
3179 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3180 			vm_phys_free_pages(m, m->pool, 0);
3181 			cnt++;
3182 		} while ((m = SLIST_FIRST(&free)) != NULL);
3183 		vm_domain_free_unlock(vmd);
3184 		vm_domain_freecnt_inc(vmd, cnt);
3185 	}
3186 	return (error);
3187 }
3188 
3189 #define	NRUNS	16
3190 
3191 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3192 
3193 #define	MIN_RECLAIM	8
3194 
3195 /*
3196  *	vm_page_reclaim_contig:
3197  *
3198  *	Reclaim allocated, contiguous physical memory satisfying the specified
3199  *	conditions by relocating the virtual pages using that physical memory.
3200  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3201  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3202  *	currently able to reclaim the requested number of pages.  Since
3203  *	relocation requires the allocation of physical pages, reclamation may
3204  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3205  *	fails in this manner, callers are expected to perform vm_wait() before
3206  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3207  *
3208  *	The caller must always specify an allocation class through "req".
3209  *
3210  *	allocation classes:
3211  *	VM_ALLOC_NORMAL		normal process request
3212  *	VM_ALLOC_SYSTEM		system *really* needs a page
3213  *	VM_ALLOC_INTERRUPT	interrupt time request
3214  *
3215  *	The optional allocation flags are ignored.
3216  *
3217  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3218  *	must be a power of two.
3219  */
3220 int
3221 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3222     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3223     int desired_runs)
3224 {
3225 	struct vm_domain *vmd;
3226 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3227 	u_long count, minalign, reclaimed;
3228 	int error, i, min_reclaim, nruns, options, req_class;
3229 	int segind, start_segind;
3230 	int ret;
3231 
3232 	KASSERT(npages > 0, ("npages is 0"));
3233 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3234 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3235 
3236 	ret = ENOMEM;
3237 
3238 	/*
3239 	 * If the caller wants to reclaim multiple runs, try to allocate
3240 	 * space to store the runs.  If that fails, fall back to the old
3241 	 * behavior of just reclaiming MIN_RECLAIM pages.
3242 	 */
3243 	if (desired_runs > 1)
3244 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3245 		    M_TEMP, M_NOWAIT);
3246 	else
3247 		m_runs = NULL;
3248 
3249 	if (m_runs == NULL) {
3250 		m_runs = _m_runs;
3251 		nruns = NRUNS;
3252 	} else {
3253 		nruns = NRUNS + desired_runs - 1;
3254 	}
3255 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3256 
3257 	/*
3258 	 * The caller will attempt an allocation after some runs have been
3259 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3260 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3261 	 * power of two or maximum chunk size, and ensure that each run is
3262 	 * suitably aligned.
3263 	 */
3264 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3265 	npages = roundup2(npages, minalign);
3266 	if (alignment < ptoa(minalign))
3267 		alignment = ptoa(minalign);
3268 
3269 	/*
3270 	 * The page daemon is allowed to dig deeper into the free page list.
3271 	 */
3272 	req_class = req & VM_ALLOC_CLASS_MASK;
3273 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3274 		req_class = VM_ALLOC_SYSTEM;
3275 
3276 	start_segind = vm_phys_lookup_segind(low);
3277 
3278 	/*
3279 	 * Return if the number of free pages cannot satisfy the requested
3280 	 * allocation.
3281 	 */
3282 	vmd = VM_DOMAIN(domain);
3283 	count = vmd->vmd_free_count;
3284 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3285 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3286 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3287 		goto done;
3288 
3289 	/*
3290 	 * Scan up to three times, relaxing the restrictions ("options") on
3291 	 * the reclamation of reservations and superpages each time.
3292 	 */
3293 	for (options = VPSC_NORESERV;;) {
3294 		bool phys_range_exists = false;
3295 
3296 		/*
3297 		 * Find the highest runs that satisfy the given constraints
3298 		 * and restrictions, and record them in "m_runs".
3299 		 */
3300 		count = 0;
3301 		segind = start_segind;
3302 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3303 		    npages, low, high)) != -1) {
3304 			phys_range_exists = true;
3305 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3306 			    bounds[1], alignment, boundary, options))) {
3307 				bounds[0] = m_run + npages;
3308 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3309 				count++;
3310 			}
3311 			segind++;
3312 		}
3313 
3314 		if (!phys_range_exists) {
3315 			ret = ERANGE;
3316 			goto done;
3317 		}
3318 
3319 		/*
3320 		 * Reclaim the highest runs in LIFO (descending) order until
3321 		 * the number of reclaimed pages, "reclaimed", is at least
3322 		 * "min_reclaim".  Reset "reclaimed" each time because each
3323 		 * reclamation is idempotent, and runs will (likely) recur
3324 		 * from one scan to the next as restrictions are relaxed.
3325 		 */
3326 		reclaimed = 0;
3327 		for (i = 0; count > 0 && i < nruns; i++) {
3328 			count--;
3329 			m_run = m_runs[RUN_INDEX(count, nruns)];
3330 			error = vm_page_reclaim_run(req_class, domain, npages,
3331 			    m_run, high);
3332 			if (error == 0) {
3333 				reclaimed += npages;
3334 				if (reclaimed >= min_reclaim) {
3335 					ret = 0;
3336 					goto done;
3337 				}
3338 			}
3339 		}
3340 
3341 		/*
3342 		 * Either relax the restrictions on the next scan or return if
3343 		 * the last scan had no restrictions.
3344 		 */
3345 		if (options == VPSC_NORESERV)
3346 			options = VPSC_NOSUPER;
3347 		else if (options == VPSC_NOSUPER)
3348 			options = VPSC_ANY;
3349 		else if (options == VPSC_ANY) {
3350 			if (reclaimed != 0)
3351 				ret = 0;
3352 			goto done;
3353 		}
3354 	}
3355 done:
3356 	if (m_runs != _m_runs)
3357 		free(m_runs, M_TEMP);
3358 	return (ret);
3359 }
3360 
3361 int
3362 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3363     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3364 {
3365 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low,
3366 	    high, alignment, boundary, 1));
3367 }
3368 
3369 int
3370 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3371     u_long alignment, vm_paddr_t boundary)
3372 {
3373 	struct vm_domainset_iter di;
3374 	int domain, ret, status;
3375 
3376 	ret = ERANGE;
3377 
3378 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3379 	do {
3380 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3381 		    high, alignment, boundary);
3382 		if (status == 0)
3383 			return (0);
3384 		else if (status == ERANGE)
3385 			vm_domainset_iter_ignore(&di, domain);
3386 		else {
3387 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3388 			    "from vm_page_reclaim_contig_domain()", status));
3389 			ret = ENOMEM;
3390 		}
3391 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3392 
3393 	return (ret);
3394 }
3395 
3396 /*
3397  * Set the domain in the appropriate page level domainset.
3398  */
3399 void
3400 vm_domain_set(struct vm_domain *vmd)
3401 {
3402 
3403 	mtx_lock(&vm_domainset_lock);
3404 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3405 		vmd->vmd_minset = 1;
3406 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3407 	}
3408 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3409 		vmd->vmd_severeset = 1;
3410 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3411 	}
3412 	mtx_unlock(&vm_domainset_lock);
3413 }
3414 
3415 /*
3416  * Clear the domain from the appropriate page level domainset.
3417  */
3418 void
3419 vm_domain_clear(struct vm_domain *vmd)
3420 {
3421 
3422 	mtx_lock(&vm_domainset_lock);
3423 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3424 		vmd->vmd_minset = 0;
3425 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3426 		if (vm_min_waiters != 0) {
3427 			vm_min_waiters = 0;
3428 			wakeup(&vm_min_domains);
3429 		}
3430 	}
3431 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3432 		vmd->vmd_severeset = 0;
3433 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3434 		if (vm_severe_waiters != 0) {
3435 			vm_severe_waiters = 0;
3436 			wakeup(&vm_severe_domains);
3437 		}
3438 	}
3439 
3440 	/*
3441 	 * If pageout daemon needs pages, then tell it that there are
3442 	 * some free.
3443 	 */
3444 	if (vmd->vmd_pageout_pages_needed &&
3445 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3446 		wakeup(&vmd->vmd_pageout_pages_needed);
3447 		vmd->vmd_pageout_pages_needed = 0;
3448 	}
3449 
3450 	/* See comments in vm_wait_doms(). */
3451 	if (vm_pageproc_waiters) {
3452 		vm_pageproc_waiters = 0;
3453 		wakeup(&vm_pageproc_waiters);
3454 	}
3455 	mtx_unlock(&vm_domainset_lock);
3456 }
3457 
3458 /*
3459  * Wait for free pages to exceed the min threshold globally.
3460  */
3461 void
3462 vm_wait_min(void)
3463 {
3464 
3465 	mtx_lock(&vm_domainset_lock);
3466 	while (vm_page_count_min()) {
3467 		vm_min_waiters++;
3468 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3469 	}
3470 	mtx_unlock(&vm_domainset_lock);
3471 }
3472 
3473 /*
3474  * Wait for free pages to exceed the severe threshold globally.
3475  */
3476 void
3477 vm_wait_severe(void)
3478 {
3479 
3480 	mtx_lock(&vm_domainset_lock);
3481 	while (vm_page_count_severe()) {
3482 		vm_severe_waiters++;
3483 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3484 		    "vmwait", 0);
3485 	}
3486 	mtx_unlock(&vm_domainset_lock);
3487 }
3488 
3489 u_int
3490 vm_wait_count(void)
3491 {
3492 
3493 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3494 }
3495 
3496 int
3497 vm_wait_doms(const domainset_t *wdoms, int mflags)
3498 {
3499 	int error;
3500 
3501 	error = 0;
3502 
3503 	/*
3504 	 * We use racey wakeup synchronization to avoid expensive global
3505 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3506 	 * To handle this, we only sleep for one tick in this instance.  It
3507 	 * is expected that most allocations for the pageproc will come from
3508 	 * kmem or vm_page_grab* which will use the more specific and
3509 	 * race-free vm_wait_domain().
3510 	 */
3511 	if (curproc == pageproc) {
3512 		mtx_lock(&vm_domainset_lock);
3513 		vm_pageproc_waiters++;
3514 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3515 		    PVM | PDROP | mflags, "pageprocwait", 1);
3516 	} else {
3517 		/*
3518 		 * XXX Ideally we would wait only until the allocation could
3519 		 * be satisfied.  This condition can cause new allocators to
3520 		 * consume all freed pages while old allocators wait.
3521 		 */
3522 		mtx_lock(&vm_domainset_lock);
3523 		if (vm_page_count_min_set(wdoms)) {
3524 			if (pageproc == NULL)
3525 				panic("vm_wait in early boot");
3526 			vm_min_waiters++;
3527 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3528 			    PVM | PDROP | mflags, "vmwait", 0);
3529 		} else
3530 			mtx_unlock(&vm_domainset_lock);
3531 	}
3532 	return (error);
3533 }
3534 
3535 /*
3536  *	vm_wait_domain:
3537  *
3538  *	Sleep until free pages are available for allocation.
3539  *	- Called in various places after failed memory allocations.
3540  */
3541 void
3542 vm_wait_domain(int domain)
3543 {
3544 	struct vm_domain *vmd;
3545 	domainset_t wdom;
3546 
3547 	vmd = VM_DOMAIN(domain);
3548 	vm_domain_free_assert_unlocked(vmd);
3549 
3550 	if (curproc == pageproc) {
3551 		mtx_lock(&vm_domainset_lock);
3552 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3553 			vmd->vmd_pageout_pages_needed = 1;
3554 			msleep(&vmd->vmd_pageout_pages_needed,
3555 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3556 		} else
3557 			mtx_unlock(&vm_domainset_lock);
3558 	} else {
3559 		DOMAINSET_ZERO(&wdom);
3560 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3561 		vm_wait_doms(&wdom, 0);
3562 	}
3563 }
3564 
3565 static int
3566 vm_wait_flags(vm_object_t obj, int mflags)
3567 {
3568 	struct domainset *d;
3569 
3570 	d = NULL;
3571 
3572 	/*
3573 	 * Carefully fetch pointers only once: the struct domainset
3574 	 * itself is ummutable but the pointer might change.
3575 	 */
3576 	if (obj != NULL)
3577 		d = obj->domain.dr_policy;
3578 	if (d == NULL)
3579 		d = curthread->td_domain.dr_policy;
3580 
3581 	return (vm_wait_doms(&d->ds_mask, mflags));
3582 }
3583 
3584 /*
3585  *	vm_wait:
3586  *
3587  *	Sleep until free pages are available for allocation in the
3588  *	affinity domains of the obj.  If obj is NULL, the domain set
3589  *	for the calling thread is used.
3590  *	Called in various places after failed memory allocations.
3591  */
3592 void
3593 vm_wait(vm_object_t obj)
3594 {
3595 	(void)vm_wait_flags(obj, 0);
3596 }
3597 
3598 int
3599 vm_wait_intr(vm_object_t obj)
3600 {
3601 	return (vm_wait_flags(obj, PCATCH));
3602 }
3603 
3604 /*
3605  *	vm_domain_alloc_fail:
3606  *
3607  *	Called when a page allocation function fails.  Informs the
3608  *	pagedaemon and performs the requested wait.  Requires the
3609  *	domain_free and object lock on entry.  Returns with the
3610  *	object lock held and free lock released.  Returns an error when
3611  *	retry is necessary.
3612  *
3613  */
3614 static int
3615 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3616 {
3617 
3618 	vm_domain_free_assert_unlocked(vmd);
3619 
3620 	atomic_add_int(&vmd->vmd_pageout_deficit,
3621 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3622 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3623 		if (object != NULL)
3624 			VM_OBJECT_WUNLOCK(object);
3625 		vm_wait_domain(vmd->vmd_domain);
3626 		if (object != NULL)
3627 			VM_OBJECT_WLOCK(object);
3628 		if (req & VM_ALLOC_WAITOK)
3629 			return (EAGAIN);
3630 	}
3631 
3632 	return (0);
3633 }
3634 
3635 /*
3636  *	vm_waitpfault:
3637  *
3638  *	Sleep until free pages are available for allocation.
3639  *	- Called only in vm_fault so that processes page faulting
3640  *	  can be easily tracked.
3641  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3642  *	  processes will be able to grab memory first.  Do not change
3643  *	  this balance without careful testing first.
3644  */
3645 void
3646 vm_waitpfault(struct domainset *dset, int timo)
3647 {
3648 
3649 	/*
3650 	 * XXX Ideally we would wait only until the allocation could
3651 	 * be satisfied.  This condition can cause new allocators to
3652 	 * consume all freed pages while old allocators wait.
3653 	 */
3654 	mtx_lock(&vm_domainset_lock);
3655 	if (vm_page_count_min_set(&dset->ds_mask)) {
3656 		vm_min_waiters++;
3657 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3658 		    "pfault", timo);
3659 	} else
3660 		mtx_unlock(&vm_domainset_lock);
3661 }
3662 
3663 static struct vm_pagequeue *
3664 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3665 {
3666 
3667 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3668 }
3669 
3670 #ifdef INVARIANTS
3671 static struct vm_pagequeue *
3672 vm_page_pagequeue(vm_page_t m)
3673 {
3674 
3675 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3676 }
3677 #endif
3678 
3679 static __always_inline bool
3680 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old,
3681     vm_page_astate_t new)
3682 {
3683 	vm_page_astate_t tmp;
3684 
3685 	tmp = *old;
3686 	do {
3687 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3688 			return (true);
3689 		counter_u64_add(pqstate_commit_retries, 1);
3690 	} while (old->_bits == tmp._bits);
3691 
3692 	return (false);
3693 }
3694 
3695 /*
3696  * Do the work of committing a queue state update that moves the page out of
3697  * its current queue.
3698  */
3699 static bool
3700 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3701     vm_page_astate_t *old, vm_page_astate_t new)
3702 {
3703 	vm_page_t next;
3704 
3705 	vm_pagequeue_assert_locked(pq);
3706 	KASSERT(vm_page_pagequeue(m) == pq,
3707 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3708 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3709 	    ("%s: invalid queue indices %d %d",
3710 	    __func__, old->queue, new.queue));
3711 
3712 	/*
3713 	 * Once the queue index of the page changes there is nothing
3714 	 * synchronizing with further updates to the page's physical
3715 	 * queue state.  Therefore we must speculatively remove the page
3716 	 * from the queue now and be prepared to roll back if the queue
3717 	 * state update fails.  If the page is not physically enqueued then
3718 	 * we just update its queue index.
3719 	 */
3720 	if ((old->flags & PGA_ENQUEUED) != 0) {
3721 		new.flags &= ~PGA_ENQUEUED;
3722 		next = TAILQ_NEXT(m, plinks.q);
3723 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3724 		vm_pagequeue_cnt_dec(pq);
3725 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3726 			if (next == NULL)
3727 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3728 			else
3729 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3730 			vm_pagequeue_cnt_inc(pq);
3731 			return (false);
3732 		} else {
3733 			return (true);
3734 		}
3735 	} else {
3736 		return (vm_page_pqstate_fcmpset(m, old, new));
3737 	}
3738 }
3739 
3740 static bool
3741 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3742     vm_page_astate_t new)
3743 {
3744 	struct vm_pagequeue *pq;
3745 	vm_page_astate_t as;
3746 	bool ret;
3747 
3748 	pq = _vm_page_pagequeue(m, old->queue);
3749 
3750 	/*
3751 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3752 	 * corresponding page queue lock is held.
3753 	 */
3754 	vm_pagequeue_lock(pq);
3755 	as = vm_page_astate_load(m);
3756 	if (__predict_false(as._bits != old->_bits)) {
3757 		*old = as;
3758 		ret = false;
3759 	} else {
3760 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3761 	}
3762 	vm_pagequeue_unlock(pq);
3763 	return (ret);
3764 }
3765 
3766 /*
3767  * Commit a queue state update that enqueues or requeues a page.
3768  */
3769 static bool
3770 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3771     vm_page_astate_t *old, vm_page_astate_t new)
3772 {
3773 	struct vm_domain *vmd;
3774 
3775 	vm_pagequeue_assert_locked(pq);
3776 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3777 	    ("%s: invalid queue indices %d %d",
3778 	    __func__, old->queue, new.queue));
3779 
3780 	new.flags |= PGA_ENQUEUED;
3781 	if (!vm_page_pqstate_fcmpset(m, old, new))
3782 		return (false);
3783 
3784 	if ((old->flags & PGA_ENQUEUED) != 0)
3785 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3786 	else
3787 		vm_pagequeue_cnt_inc(pq);
3788 
3789 	/*
3790 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3791 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3792 	 * applied, even if it was set first.
3793 	 */
3794 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3795 		vmd = vm_pagequeue_domain(m);
3796 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3797 		    ("%s: invalid page queue for page %p", __func__, m));
3798 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3799 	} else {
3800 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3801 	}
3802 	return (true);
3803 }
3804 
3805 /*
3806  * Commit a queue state update that encodes a request for a deferred queue
3807  * operation.
3808  */
3809 static bool
3810 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3811     vm_page_astate_t new)
3812 {
3813 
3814 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3815 	    ("%s: invalid state, queue %d flags %x",
3816 	    __func__, new.queue, new.flags));
3817 
3818 	if (old->_bits != new._bits &&
3819 	    !vm_page_pqstate_fcmpset(m, old, new))
3820 		return (false);
3821 	vm_page_pqbatch_submit(m, new.queue);
3822 	return (true);
3823 }
3824 
3825 /*
3826  * A generic queue state update function.  This handles more cases than the
3827  * specialized functions above.
3828  */
3829 bool
3830 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3831 {
3832 
3833 	if (old->_bits == new._bits)
3834 		return (true);
3835 
3836 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3837 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3838 			return (false);
3839 		if (new.queue != PQ_NONE)
3840 			vm_page_pqbatch_submit(m, new.queue);
3841 	} else {
3842 		if (!vm_page_pqstate_fcmpset(m, old, new))
3843 			return (false);
3844 		if (new.queue != PQ_NONE &&
3845 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3846 			vm_page_pqbatch_submit(m, new.queue);
3847 	}
3848 	return (true);
3849 }
3850 
3851 /*
3852  * Apply deferred queue state updates to a page.
3853  */
3854 static inline void
3855 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3856 {
3857 	vm_page_astate_t new, old;
3858 
3859 	CRITICAL_ASSERT(curthread);
3860 	vm_pagequeue_assert_locked(pq);
3861 	KASSERT(queue < PQ_COUNT,
3862 	    ("%s: invalid queue index %d", __func__, queue));
3863 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3864 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3865 
3866 	for (old = vm_page_astate_load(m);;) {
3867 		if (__predict_false(old.queue != queue ||
3868 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3869 			counter_u64_add(queue_nops, 1);
3870 			break;
3871 		}
3872 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3873 		    ("%s: page %p is unmanaged", __func__, m));
3874 
3875 		new = old;
3876 		if ((old.flags & PGA_DEQUEUE) != 0) {
3877 			new.flags &= ~PGA_QUEUE_OP_MASK;
3878 			new.queue = PQ_NONE;
3879 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3880 			    m, &old, new))) {
3881 				counter_u64_add(queue_ops, 1);
3882 				break;
3883 			}
3884 		} else {
3885 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3886 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3887 			    m, &old, new))) {
3888 				counter_u64_add(queue_ops, 1);
3889 				break;
3890 			}
3891 		}
3892 	}
3893 }
3894 
3895 static void
3896 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3897     uint8_t queue)
3898 {
3899 	int i;
3900 
3901 	for (i = 0; i < bq->bq_cnt; i++)
3902 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3903 	vm_batchqueue_init(bq);
3904 }
3905 
3906 /*
3907  *	vm_page_pqbatch_submit:		[ internal use only ]
3908  *
3909  *	Enqueue a page in the specified page queue's batched work queue.
3910  *	The caller must have encoded the requested operation in the page
3911  *	structure's a.flags field.
3912  */
3913 void
3914 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3915 {
3916 	struct vm_batchqueue *bq;
3917 	struct vm_pagequeue *pq;
3918 	int domain, slots_remaining;
3919 
3920 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3921 
3922 	domain = vm_page_domain(m);
3923 	critical_enter();
3924 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3925 	slots_remaining = vm_batchqueue_insert(bq, m);
3926 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3927 		/* keep building the bq */
3928 		critical_exit();
3929 		return;
3930 	} else if (slots_remaining > 0 ) {
3931 		/* Try to process the bq if we can get the lock */
3932 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3933 		if (vm_pagequeue_trylock(pq)) {
3934 			vm_pqbatch_process(pq, bq, queue);
3935 			vm_pagequeue_unlock(pq);
3936 		}
3937 		critical_exit();
3938 		return;
3939 	}
3940 	critical_exit();
3941 
3942 	/* if we make it here, the bq is full so wait for the lock */
3943 
3944 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3945 	vm_pagequeue_lock(pq);
3946 	critical_enter();
3947 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3948 	vm_pqbatch_process(pq, bq, queue);
3949 	vm_pqbatch_process_page(pq, m, queue);
3950 	vm_pagequeue_unlock(pq);
3951 	critical_exit();
3952 }
3953 
3954 /*
3955  *	vm_page_pqbatch_drain:		[ internal use only ]
3956  *
3957  *	Force all per-CPU page queue batch queues to be drained.  This is
3958  *	intended for use in severe memory shortages, to ensure that pages
3959  *	do not remain stuck in the batch queues.
3960  */
3961 void
3962 vm_page_pqbatch_drain(void)
3963 {
3964 	struct thread *td;
3965 	struct vm_domain *vmd;
3966 	struct vm_pagequeue *pq;
3967 	int cpu, domain, queue;
3968 
3969 	td = curthread;
3970 	CPU_FOREACH(cpu) {
3971 		thread_lock(td);
3972 		sched_bind(td, cpu);
3973 		thread_unlock(td);
3974 
3975 		for (domain = 0; domain < vm_ndomains; domain++) {
3976 			vmd = VM_DOMAIN(domain);
3977 			for (queue = 0; queue < PQ_COUNT; queue++) {
3978 				pq = &vmd->vmd_pagequeues[queue];
3979 				vm_pagequeue_lock(pq);
3980 				critical_enter();
3981 				vm_pqbatch_process(pq,
3982 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3983 				critical_exit();
3984 				vm_pagequeue_unlock(pq);
3985 			}
3986 		}
3987 	}
3988 	thread_lock(td);
3989 	sched_unbind(td);
3990 	thread_unlock(td);
3991 }
3992 
3993 /*
3994  *	vm_page_dequeue_deferred:	[ internal use only ]
3995  *
3996  *	Request removal of the given page from its current page
3997  *	queue.  Physical removal from the queue may be deferred
3998  *	indefinitely.
3999  */
4000 void
4001 vm_page_dequeue_deferred(vm_page_t m)
4002 {
4003 	vm_page_astate_t new, old;
4004 
4005 	old = vm_page_astate_load(m);
4006 	do {
4007 		if (old.queue == PQ_NONE) {
4008 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4009 			    ("%s: page %p has unexpected queue state",
4010 			    __func__, m));
4011 			break;
4012 		}
4013 		new = old;
4014 		new.flags |= PGA_DEQUEUE;
4015 	} while (!vm_page_pqstate_commit_request(m, &old, new));
4016 }
4017 
4018 /*
4019  *	vm_page_dequeue:
4020  *
4021  *	Remove the page from whichever page queue it's in, if any, before
4022  *	returning.
4023  */
4024 void
4025 vm_page_dequeue(vm_page_t m)
4026 {
4027 	vm_page_astate_t new, old;
4028 
4029 	old = vm_page_astate_load(m);
4030 	do {
4031 		if (old.queue == PQ_NONE) {
4032 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4033 			    ("%s: page %p has unexpected queue state",
4034 			    __func__, m));
4035 			break;
4036 		}
4037 		new = old;
4038 		new.flags &= ~PGA_QUEUE_OP_MASK;
4039 		new.queue = PQ_NONE;
4040 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
4041 
4042 }
4043 
4044 /*
4045  * Schedule the given page for insertion into the specified page queue.
4046  * Physical insertion of the page may be deferred indefinitely.
4047  */
4048 static void
4049 vm_page_enqueue(vm_page_t m, uint8_t queue)
4050 {
4051 
4052 	KASSERT(m->a.queue == PQ_NONE &&
4053 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
4054 	    ("%s: page %p is already enqueued", __func__, m));
4055 	KASSERT(m->ref_count > 0,
4056 	    ("%s: page %p does not carry any references", __func__, m));
4057 
4058 	m->a.queue = queue;
4059 	if ((m->a.flags & PGA_REQUEUE) == 0)
4060 		vm_page_aflag_set(m, PGA_REQUEUE);
4061 	vm_page_pqbatch_submit(m, queue);
4062 }
4063 
4064 /*
4065  *	vm_page_free_prep:
4066  *
4067  *	Prepares the given page to be put on the free list,
4068  *	disassociating it from any VM object. The caller may return
4069  *	the page to the free list only if this function returns true.
4070  *
4071  *	The object, if it exists, must be locked, and then the page must
4072  *	be xbusy.  Otherwise the page must be not busied.  A managed
4073  *	page must be unmapped.
4074  */
4075 static bool
4076 vm_page_free_prep(vm_page_t m)
4077 {
4078 
4079 	/*
4080 	 * Synchronize with threads that have dropped a reference to this
4081 	 * page.
4082 	 */
4083 	atomic_thread_fence_acq();
4084 
4085 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
4086 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
4087 		uint64_t *p;
4088 		int i;
4089 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4090 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4091 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4092 			    m, i, (uintmax_t)*p));
4093 	}
4094 #endif
4095 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4096 		KASSERT(!pmap_page_is_mapped(m),
4097 		    ("vm_page_free_prep: freeing mapped page %p", m));
4098 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4099 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4100 	} else {
4101 		KASSERT(m->a.queue == PQ_NONE,
4102 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4103 	}
4104 	VM_CNT_INC(v_tfree);
4105 
4106 	if (m->object != NULL) {
4107 		vm_page_radix_remove(m);
4108 		vm_page_free_object_prep(m);
4109 	} else
4110 		vm_page_assert_unbusied(m);
4111 
4112 	vm_page_busy_free(m);
4113 
4114 	/*
4115 	 * If fictitious remove object association and
4116 	 * return.
4117 	 */
4118 	if ((m->flags & PG_FICTITIOUS) != 0) {
4119 		KASSERT(m->ref_count == 1,
4120 		    ("fictitious page %p is referenced", m));
4121 		KASSERT(m->a.queue == PQ_NONE,
4122 		    ("fictitious page %p is queued", m));
4123 		return (false);
4124 	}
4125 
4126 	/*
4127 	 * Pages need not be dequeued before they are returned to the physical
4128 	 * memory allocator, but they must at least be marked for a deferred
4129 	 * dequeue.
4130 	 */
4131 	if ((m->oflags & VPO_UNMANAGED) == 0)
4132 		vm_page_dequeue_deferred(m);
4133 
4134 	m->valid = 0;
4135 	vm_page_undirty(m);
4136 
4137 	if (m->ref_count != 0)
4138 		panic("vm_page_free_prep: page %p has references", m);
4139 
4140 	/*
4141 	 * Restore the default memory attribute to the page.
4142 	 */
4143 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4144 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4145 
4146 #if VM_NRESERVLEVEL > 0
4147 	/*
4148 	 * Determine whether the page belongs to a reservation.  If the page was
4149 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4150 	 * as an optimization, we avoid the check in that case.
4151 	 */
4152 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4153 		return (false);
4154 #endif
4155 
4156 	return (true);
4157 }
4158 
4159 /*
4160  *	vm_page_free_toq:
4161  *
4162  *	Returns the given page to the free list, disassociating it
4163  *	from any VM object.
4164  *
4165  *	The object must be locked.  The page must be exclusively busied if it
4166  *	belongs to an object.
4167  */
4168 static void
4169 vm_page_free_toq(vm_page_t m)
4170 {
4171 	struct vm_domain *vmd;
4172 	uma_zone_t zone;
4173 
4174 	if (!vm_page_free_prep(m))
4175 		return;
4176 
4177 	vmd = vm_pagequeue_domain(m);
4178 	if (__predict_false((m->flags & PG_NOFREE) != 0)) {
4179 		vm_page_free_nofree(vmd, m);
4180 		return;
4181 	}
4182 	zone = vmd->vmd_pgcache[m->pool].zone;
4183 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4184 		uma_zfree(zone, m);
4185 		return;
4186 	}
4187 	vm_domain_free_lock(vmd);
4188 	vm_phys_free_pages(m, m->pool, 0);
4189 	vm_domain_free_unlock(vmd);
4190 	vm_domain_freecnt_inc(vmd, 1);
4191 }
4192 
4193 /*
4194  *	vm_page_free_pages_toq:
4195  *
4196  *	Returns a list of pages to the free list, disassociating it
4197  *	from any VM object.  In other words, this is equivalent to
4198  *	calling vm_page_free_toq() for each page of a list of VM objects.
4199  */
4200 int
4201 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4202 {
4203 	vm_page_t m;
4204 	int count;
4205 
4206 	if (SLIST_EMPTY(free))
4207 		return (0);
4208 
4209 	count = 0;
4210 	while ((m = SLIST_FIRST(free)) != NULL) {
4211 		count++;
4212 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4213 		vm_page_free_toq(m);
4214 	}
4215 
4216 	if (update_wire_count)
4217 		vm_wire_sub(count);
4218 	return (count);
4219 }
4220 
4221 /*
4222  * Mark this page as wired down.  For managed pages, this prevents reclamation
4223  * by the page daemon, or when the containing object, if any, is destroyed.
4224  */
4225 void
4226 vm_page_wire(vm_page_t m)
4227 {
4228 	u_int old;
4229 
4230 #ifdef INVARIANTS
4231 	if (m->object != NULL && !vm_page_busied(m) &&
4232 	    !vm_object_busied(m->object))
4233 		VM_OBJECT_ASSERT_LOCKED(m->object);
4234 #endif
4235 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4236 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4237 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4238 
4239 	old = atomic_fetchadd_int(&m->ref_count, 1);
4240 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4241 	    ("vm_page_wire: counter overflow for page %p", m));
4242 	if (VPRC_WIRE_COUNT(old) == 0) {
4243 		if ((m->oflags & VPO_UNMANAGED) == 0)
4244 			vm_page_aflag_set(m, PGA_DEQUEUE);
4245 		vm_wire_add(1);
4246 	}
4247 }
4248 
4249 /*
4250  * Attempt to wire a mapped page following a pmap lookup of that page.
4251  * This may fail if a thread is concurrently tearing down mappings of the page.
4252  * The transient failure is acceptable because it translates to the
4253  * failure of the caller pmap_extract_and_hold(), which should be then
4254  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4255  */
4256 bool
4257 vm_page_wire_mapped(vm_page_t m)
4258 {
4259 	u_int old;
4260 
4261 	old = atomic_load_int(&m->ref_count);
4262 	do {
4263 		KASSERT(old > 0,
4264 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4265 		if ((old & VPRC_BLOCKED) != 0)
4266 			return (false);
4267 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4268 
4269 	if (VPRC_WIRE_COUNT(old) == 0) {
4270 		if ((m->oflags & VPO_UNMANAGED) == 0)
4271 			vm_page_aflag_set(m, PGA_DEQUEUE);
4272 		vm_wire_add(1);
4273 	}
4274 	return (true);
4275 }
4276 
4277 /*
4278  * Release a wiring reference to a managed page.  If the page still belongs to
4279  * an object, update its position in the page queues to reflect the reference.
4280  * If the wiring was the last reference to the page, free the page.
4281  */
4282 static void
4283 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4284 {
4285 	u_int old;
4286 
4287 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4288 	    ("%s: page %p is unmanaged", __func__, m));
4289 
4290 	/*
4291 	 * Update LRU state before releasing the wiring reference.
4292 	 * Use a release store when updating the reference count to
4293 	 * synchronize with vm_page_free_prep().
4294 	 */
4295 	old = atomic_load_int(&m->ref_count);
4296 	do {
4297 		u_int count;
4298 
4299 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4300 		    ("vm_page_unwire: wire count underflow for page %p", m));
4301 
4302 		count = old & ~VPRC_BLOCKED;
4303 		if (count > VPRC_OBJREF + 1) {
4304 			/*
4305 			 * The page has at least one other wiring reference.  An
4306 			 * earlier iteration of this loop may have called
4307 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4308 			 * re-set it if necessary.
4309 			 */
4310 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4311 				vm_page_aflag_set(m, PGA_DEQUEUE);
4312 		} else if (count == VPRC_OBJREF + 1) {
4313 			/*
4314 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4315 			 * update the page's queue state to reflect the
4316 			 * reference.  If the page does not belong to an object
4317 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4318 			 * clear leftover queue state.
4319 			 */
4320 			vm_page_release_toq(m, nqueue, noreuse);
4321 		} else if (count == 1) {
4322 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4323 		}
4324 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4325 
4326 	if (VPRC_WIRE_COUNT(old) == 1) {
4327 		vm_wire_sub(1);
4328 		if (old == 1)
4329 			vm_page_free(m);
4330 	}
4331 }
4332 
4333 /*
4334  * Release one wiring of the specified page, potentially allowing it to be
4335  * paged out.
4336  *
4337  * Only managed pages belonging to an object can be paged out.  If the number
4338  * of wirings transitions to zero and the page is eligible for page out, then
4339  * the page is added to the specified paging queue.  If the released wiring
4340  * represented the last reference to the page, the page is freed.
4341  */
4342 void
4343 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4344 {
4345 
4346 	KASSERT(nqueue < PQ_COUNT,
4347 	    ("vm_page_unwire: invalid queue %u request for page %p",
4348 	    nqueue, m));
4349 
4350 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4351 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4352 			vm_page_free(m);
4353 		return;
4354 	}
4355 	vm_page_unwire_managed(m, nqueue, false);
4356 }
4357 
4358 /*
4359  * Unwire a page without (re-)inserting it into a page queue.  It is up
4360  * to the caller to enqueue, requeue, or free the page as appropriate.
4361  * In most cases involving managed pages, vm_page_unwire() should be used
4362  * instead.
4363  */
4364 bool
4365 vm_page_unwire_noq(vm_page_t m)
4366 {
4367 	u_int old;
4368 
4369 	old = vm_page_drop(m, 1);
4370 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4371 	    ("%s: counter underflow for page %p", __func__,  m));
4372 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4373 	    ("%s: missing ref on fictitious page %p", __func__, m));
4374 
4375 	if (VPRC_WIRE_COUNT(old) > 1)
4376 		return (false);
4377 	if ((m->oflags & VPO_UNMANAGED) == 0)
4378 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4379 	vm_wire_sub(1);
4380 	return (true);
4381 }
4382 
4383 /*
4384  * Ensure that the page ends up in the specified page queue.  If the page is
4385  * active or being moved to the active queue, ensure that its act_count is
4386  * at least ACT_INIT but do not otherwise mess with it.
4387  */
4388 static __always_inline void
4389 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4390 {
4391 	vm_page_astate_t old, new;
4392 
4393 	KASSERT(m->ref_count > 0,
4394 	    ("%s: page %p does not carry any references", __func__, m));
4395 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4396 	    ("%s: invalid flags %x", __func__, nflag));
4397 
4398 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4399 		return;
4400 
4401 	old = vm_page_astate_load(m);
4402 	do {
4403 		if ((old.flags & PGA_DEQUEUE) != 0)
4404 			break;
4405 		new = old;
4406 		new.flags &= ~PGA_QUEUE_OP_MASK;
4407 		if (nqueue == PQ_ACTIVE)
4408 			new.act_count = max(old.act_count, ACT_INIT);
4409 		if (old.queue == nqueue) {
4410 			/*
4411 			 * There is no need to requeue pages already in the
4412 			 * active queue.
4413 			 */
4414 			if (nqueue != PQ_ACTIVE ||
4415 			    (old.flags & PGA_ENQUEUED) == 0)
4416 				new.flags |= nflag;
4417 		} else {
4418 			new.flags |= nflag;
4419 			new.queue = nqueue;
4420 		}
4421 	} while (!vm_page_pqstate_commit(m, &old, new));
4422 }
4423 
4424 /*
4425  * Put the specified page on the active list (if appropriate).
4426  */
4427 void
4428 vm_page_activate(vm_page_t m)
4429 {
4430 
4431 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4432 }
4433 
4434 /*
4435  * Move the specified page to the tail of the inactive queue, or requeue
4436  * the page if it is already in the inactive queue.
4437  */
4438 void
4439 vm_page_deactivate(vm_page_t m)
4440 {
4441 
4442 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4443 }
4444 
4445 void
4446 vm_page_deactivate_noreuse(vm_page_t m)
4447 {
4448 
4449 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4450 }
4451 
4452 /*
4453  * Put a page in the laundry, or requeue it if it is already there.
4454  */
4455 void
4456 vm_page_launder(vm_page_t m)
4457 {
4458 
4459 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4460 }
4461 
4462 /*
4463  * Put a page in the PQ_UNSWAPPABLE holding queue.
4464  */
4465 void
4466 vm_page_unswappable(vm_page_t m)
4467 {
4468 
4469 	VM_OBJECT_ASSERT_LOCKED(m->object);
4470 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4471 	    ("page %p already unswappable", m));
4472 
4473 	vm_page_dequeue(m);
4474 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4475 }
4476 
4477 /*
4478  * Release a page back to the page queues in preparation for unwiring.
4479  */
4480 static void
4481 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4482 {
4483 	vm_page_astate_t old, new;
4484 	uint16_t nflag;
4485 
4486 	/*
4487 	 * Use a check of the valid bits to determine whether we should
4488 	 * accelerate reclamation of the page.  The object lock might not be
4489 	 * held here, in which case the check is racy.  At worst we will either
4490 	 * accelerate reclamation of a valid page and violate LRU, or
4491 	 * unnecessarily defer reclamation of an invalid page.
4492 	 *
4493 	 * If we were asked to not cache the page, place it near the head of the
4494 	 * inactive queue so that is reclaimed sooner.
4495 	 */
4496 	if (noreuse || vm_page_none_valid(m)) {
4497 		nqueue = PQ_INACTIVE;
4498 		nflag = PGA_REQUEUE_HEAD;
4499 	} else {
4500 		nflag = PGA_REQUEUE;
4501 	}
4502 
4503 	old = vm_page_astate_load(m);
4504 	do {
4505 		new = old;
4506 
4507 		/*
4508 		 * If the page is already in the active queue and we are not
4509 		 * trying to accelerate reclamation, simply mark it as
4510 		 * referenced and avoid any queue operations.
4511 		 */
4512 		new.flags &= ~PGA_QUEUE_OP_MASK;
4513 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4514 		    (old.flags & PGA_ENQUEUED) != 0)
4515 			new.flags |= PGA_REFERENCED;
4516 		else {
4517 			new.flags |= nflag;
4518 			new.queue = nqueue;
4519 		}
4520 	} while (!vm_page_pqstate_commit(m, &old, new));
4521 }
4522 
4523 /*
4524  * Unwire a page and either attempt to free it or re-add it to the page queues.
4525  */
4526 void
4527 vm_page_release(vm_page_t m, int flags)
4528 {
4529 	vm_object_t object;
4530 
4531 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4532 	    ("vm_page_release: page %p is unmanaged", m));
4533 
4534 	if ((flags & VPR_TRYFREE) != 0) {
4535 		for (;;) {
4536 			object = atomic_load_ptr(&m->object);
4537 			if (object == NULL)
4538 				break;
4539 			/* Depends on type-stability. */
4540 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4541 				break;
4542 			if (object == m->object) {
4543 				vm_page_release_locked(m, flags);
4544 				VM_OBJECT_WUNLOCK(object);
4545 				return;
4546 			}
4547 			VM_OBJECT_WUNLOCK(object);
4548 		}
4549 	}
4550 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4551 }
4552 
4553 /* See vm_page_release(). */
4554 void
4555 vm_page_release_locked(vm_page_t m, int flags)
4556 {
4557 
4558 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4559 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4560 	    ("vm_page_release_locked: page %p is unmanaged", m));
4561 
4562 	if (vm_page_unwire_noq(m)) {
4563 		if ((flags & VPR_TRYFREE) != 0 &&
4564 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4565 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4566 			/*
4567 			 * An unlocked lookup may have wired the page before the
4568 			 * busy lock was acquired, in which case the page must
4569 			 * not be freed.
4570 			 */
4571 			if (__predict_true(!vm_page_wired(m))) {
4572 				vm_page_free(m);
4573 				return;
4574 			}
4575 			vm_page_xunbusy(m);
4576 		} else {
4577 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4578 		}
4579 	}
4580 }
4581 
4582 static bool
4583 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4584 {
4585 	u_int old;
4586 
4587 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4588 	    ("vm_page_try_blocked_op: page %p has no object", m));
4589 	KASSERT(vm_page_busied(m),
4590 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4591 	VM_OBJECT_ASSERT_LOCKED(m->object);
4592 
4593 	old = atomic_load_int(&m->ref_count);
4594 	do {
4595 		KASSERT(old != 0,
4596 		    ("vm_page_try_blocked_op: page %p has no references", m));
4597 		KASSERT((old & VPRC_BLOCKED) == 0,
4598 		    ("vm_page_try_blocked_op: page %p blocks wirings", m));
4599 		if (VPRC_WIRE_COUNT(old) != 0)
4600 			return (false);
4601 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4602 
4603 	(op)(m);
4604 
4605 	/*
4606 	 * If the object is read-locked, new wirings may be created via an
4607 	 * object lookup.
4608 	 */
4609 	old = vm_page_drop(m, VPRC_BLOCKED);
4610 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4611 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4612 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4613 	    old, m));
4614 	return (true);
4615 }
4616 
4617 /*
4618  * Atomically check for wirings and remove all mappings of the page.
4619  */
4620 bool
4621 vm_page_try_remove_all(vm_page_t m)
4622 {
4623 
4624 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4625 }
4626 
4627 /*
4628  * Atomically check for wirings and remove all writeable mappings of the page.
4629  */
4630 bool
4631 vm_page_try_remove_write(vm_page_t m)
4632 {
4633 
4634 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4635 }
4636 
4637 /*
4638  * vm_page_advise
4639  *
4640  * 	Apply the specified advice to the given page.
4641  */
4642 void
4643 vm_page_advise(vm_page_t m, int advice)
4644 {
4645 
4646 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4647 	vm_page_assert_xbusied(m);
4648 
4649 	if (advice == MADV_FREE)
4650 		/*
4651 		 * Mark the page clean.  This will allow the page to be freed
4652 		 * without first paging it out.  MADV_FREE pages are often
4653 		 * quickly reused by malloc(3), so we do not do anything that
4654 		 * would result in a page fault on a later access.
4655 		 */
4656 		vm_page_undirty(m);
4657 	else if (advice != MADV_DONTNEED) {
4658 		if (advice == MADV_WILLNEED)
4659 			vm_page_activate(m);
4660 		return;
4661 	}
4662 
4663 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4664 		vm_page_dirty(m);
4665 
4666 	/*
4667 	 * Clear any references to the page.  Otherwise, the page daemon will
4668 	 * immediately reactivate the page.
4669 	 */
4670 	vm_page_aflag_clear(m, PGA_REFERENCED);
4671 
4672 	/*
4673 	 * Place clean pages near the head of the inactive queue rather than
4674 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4675 	 * the page will be reused quickly.  Dirty pages not already in the
4676 	 * laundry are moved there.
4677 	 */
4678 	if (m->dirty == 0)
4679 		vm_page_deactivate_noreuse(m);
4680 	else if (!vm_page_in_laundry(m))
4681 		vm_page_launder(m);
4682 }
4683 
4684 /*
4685  *	vm_page_grab_release
4686  *
4687  *	Helper routine for grab functions to release busy on return.
4688  */
4689 static inline void
4690 vm_page_grab_release(vm_page_t m, int allocflags)
4691 {
4692 
4693 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4694 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4695 			vm_page_sunbusy(m);
4696 		else
4697 			vm_page_xunbusy(m);
4698 	}
4699 }
4700 
4701 /*
4702  *	vm_page_grab_sleep
4703  *
4704  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4705  *	if the caller should retry and false otherwise.
4706  *
4707  *	If the object is locked on entry the object will be unlocked with
4708  *	false returns and still locked but possibly having been dropped
4709  *	with true returns.
4710  */
4711 static bool
4712 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4713     const char *wmesg, int allocflags, bool locked)
4714 {
4715 
4716 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4717 		return (false);
4718 
4719 	/*
4720 	 * Reference the page before unlocking and sleeping so that
4721 	 * the page daemon is less likely to reclaim it.
4722 	 */
4723 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4724 		vm_page_reference(m);
4725 
4726 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4727 	    locked)
4728 		VM_OBJECT_WLOCK(object);
4729 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4730 		return (false);
4731 
4732 	return (true);
4733 }
4734 
4735 /*
4736  * Assert that the grab flags are valid.
4737  */
4738 static inline void
4739 vm_page_grab_check(int allocflags)
4740 {
4741 
4742 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4743 	    (allocflags & VM_ALLOC_WIRED) != 0,
4744 	    ("vm_page_grab*: the pages must be busied or wired"));
4745 
4746 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4747 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4748 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4749 }
4750 
4751 /*
4752  * Calculate the page allocation flags for grab.
4753  */
4754 static inline int
4755 vm_page_grab_pflags(int allocflags)
4756 {
4757 	int pflags;
4758 
4759 	pflags = allocflags &
4760 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4761 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4762 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4763 		pflags |= VM_ALLOC_WAITFAIL;
4764 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4765 		pflags |= VM_ALLOC_SBUSY;
4766 
4767 	return (pflags);
4768 }
4769 
4770 /*
4771  * Grab a page, waiting until we are woken up due to the page changing state.
4772  * We keep on waiting, if the page continues to be in the object, unless
4773  * allocflags forbid waiting.
4774  *
4775  * The object must be locked on entry.  This routine may sleep.  The lock will,
4776  * however, be released and reacquired if the routine sleeps.
4777  *
4778  *  Return a grabbed page, or NULL.  Set *found if a page was found, whether or
4779  *  not it was grabbed.
4780  */
4781 static inline vm_page_t
4782 vm_page_grab_lookup(struct pctrie_iter *pages, vm_object_t object,
4783     vm_pindex_t pindex, int allocflags, bool *found)
4784 {
4785 	vm_page_t m;
4786 
4787 	while ((*found = (m = vm_radix_iter_lookup(pages, pindex)) != NULL) &&
4788 	    !vm_page_tryacquire(m, allocflags)) {
4789 		if (!vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4790 		    allocflags, true))
4791 			return (NULL);
4792 		pctrie_iter_reset(pages);
4793 	}
4794 	return (m);
4795 }
4796 
4797 /*
4798  * Grab a page.  Use an iterator parameter. Keep on waiting, as long as the page
4799  * exists in the object.  If the page doesn't exist, first allocate it and then
4800  * conditionally zero it.
4801  *
4802  * The object must be locked on entry.  This routine may sleep.  The lock will,
4803  * however, be released and reacquired if the routine sleeps.
4804  */
4805 vm_page_t
4806 vm_page_grab_iter(vm_object_t object, struct pctrie_iter *pages,
4807     vm_pindex_t pindex, int allocflags)
4808 {
4809 	vm_page_t m, mpred;
4810 	bool found;
4811 
4812 	VM_OBJECT_ASSERT_WLOCKED(object);
4813 	vm_page_grab_check(allocflags);
4814 
4815 	while ((m = vm_page_grab_lookup(
4816 	    pages, object, pindex, allocflags, &found)) == NULL) {
4817 		if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4818 			return (NULL);
4819 		if (found &&
4820 		    (allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4821 			return (NULL);
4822 		mpred = vm_radix_iter_lookup_lt(pages, pindex);
4823 		m = vm_page_alloc_after(object, pages, pindex,
4824 		    vm_page_grab_pflags(allocflags), mpred);
4825 		if (m != NULL) {
4826 			if ((allocflags & VM_ALLOC_ZERO) != 0 &&
4827 			    (m->flags & PG_ZERO) == 0)
4828 				pmap_zero_page(m);
4829 			break;
4830 		}
4831 		if ((allocflags &
4832 		    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4833 			return (NULL);
4834 	}
4835 	vm_page_grab_release(m, allocflags);
4836 
4837 	return (m);
4838 }
4839 
4840 /*
4841  * Grab a page.  Keep on waiting, as long as the page exists in the object.  If
4842  * the page doesn't exist, first allocate it and then conditionally zero it.
4843  *
4844  * The object must be locked on entry.  This routine may sleep.  The lock will,
4845  * however, be released and reacquired if the routine sleeps.
4846  */
4847 vm_page_t
4848 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4849 {
4850 	struct pctrie_iter pages;
4851 
4852 	VM_OBJECT_ASSERT_WLOCKED(object);
4853 	vm_page_iter_init(&pages, object);
4854 	return (vm_page_grab_iter(object, &pages, pindex, allocflags));
4855 }
4856 
4857 /*
4858  * Attempt to validate a page, locklessly acquiring it if necessary, given a
4859  * (object, pindex) tuple and either an invalided page or NULL.  The resulting
4860  * page will be validated against the identity tuple, and busied or wired as
4861  * requested.  A NULL page returned guarantees that the page was not in radix at
4862  * the time of the call but callers must perform higher level synchronization or
4863  * retry the operation under a lock if they require an atomic answer.  This is
4864  * the only lock free validation routine, other routines can depend on the
4865  * resulting page state.
4866  *
4867  * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to
4868  * caller flags.
4869  */
4870 #define PAGE_NOT_ACQUIRED ((vm_page_t)1)
4871 static vm_page_t
4872 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m,
4873     int allocflags)
4874 {
4875 	if (m == NULL)
4876 		m = vm_page_lookup_unlocked(object, pindex);
4877 	for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) {
4878 		if (vm_page_trybusy(m, allocflags)) {
4879 			if (m->object == object && m->pindex == pindex) {
4880 				if ((allocflags & VM_ALLOC_WIRED) != 0)
4881 					vm_page_wire(m);
4882 				vm_page_grab_release(m, allocflags);
4883 				break;
4884 			}
4885 			/* relookup. */
4886 			vm_page_busy_release(m);
4887 			cpu_spinwait();
4888 			continue;
4889 		}
4890 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4891 		    allocflags, false))
4892 			return (PAGE_NOT_ACQUIRED);
4893 	}
4894 	return (m);
4895 }
4896 
4897 /*
4898  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4899  * is not set.
4900  */
4901 vm_page_t
4902 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4903 {
4904 	vm_page_t m;
4905 
4906 	vm_page_grab_check(allocflags);
4907 	m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags);
4908 	if (m == PAGE_NOT_ACQUIRED)
4909 		return (NULL);
4910 	if (m != NULL)
4911 		return (m);
4912 
4913 	/*
4914 	 * The radix lockless lookup should never return a false negative
4915 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4916 	 * was no page present at the instant of the call.  A NOCREAT caller
4917 	 * must handle create races gracefully.
4918 	 */
4919 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4920 		return (NULL);
4921 
4922 	VM_OBJECT_WLOCK(object);
4923 	m = vm_page_grab(object, pindex, allocflags);
4924 	VM_OBJECT_WUNLOCK(object);
4925 
4926 	return (m);
4927 }
4928 
4929 /*
4930  * Grab a page and make it valid, paging in if necessary.  Use an iterator
4931  * parameter. Pages missing from their pager are zero filled and validated.  If
4932  * a VM_ALLOC_COUNT is supplied and the page is not valid as many as
4933  * VM_INITIAL_PAGEIN pages can be brought in simultaneously.  Additional pages
4934  * will be left on a paging queue but will neither be wired nor busy regardless
4935  * of allocflags.
4936  */
4937 int
4938 vm_page_grab_valid_iter(vm_page_t *mp, vm_object_t object,
4939     struct pctrie_iter *pages, vm_pindex_t pindex, int allocflags)
4940 {
4941 	vm_page_t m, mpred;
4942 	vm_page_t ma[VM_INITIAL_PAGEIN];
4943 	int after, i, pflags, rv;
4944 
4945 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4946 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4947 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4948 	KASSERT((allocflags &
4949 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4950 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4951 	VM_OBJECT_ASSERT_WLOCKED(object);
4952 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4953 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4954 	pflags |= VM_ALLOC_WAITFAIL;
4955 
4956 retrylookup:
4957 	if ((m = vm_radix_iter_lookup(pages, pindex)) != NULL) {
4958 		/*
4959 		 * If the page is fully valid it can only become invalid
4960 		 * with the object lock held.  If it is not valid it can
4961 		 * become valid with the busy lock held.  Therefore, we
4962 		 * may unnecessarily lock the exclusive busy here if we
4963 		 * race with I/O completion not using the object lock.
4964 		 * However, we will not end up with an invalid page and a
4965 		 * shared lock.
4966 		 */
4967 		if (!vm_page_trybusy(m,
4968 		    vm_page_all_valid(m) ? allocflags : 0)) {
4969 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4970 			    allocflags, true);
4971 			pctrie_iter_reset(pages);
4972 			goto retrylookup;
4973 		}
4974 		if (vm_page_all_valid(m))
4975 			goto out;
4976 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4977 			vm_page_busy_release(m);
4978 			*mp = NULL;
4979 			return (VM_PAGER_FAIL);
4980 		}
4981 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4982 		*mp = NULL;
4983 		return (VM_PAGER_FAIL);
4984 	} else {
4985 		mpred = vm_radix_iter_lookup_lt(pages, pindex);
4986 		m = vm_page_alloc_after(object, pages, pindex, pflags, mpred);
4987 		if (m == NULL) {
4988 			if (!vm_pager_can_alloc_page(object, pindex)) {
4989 				*mp = NULL;
4990 				return (VM_PAGER_AGAIN);
4991 			}
4992 			pctrie_iter_reset(pages);
4993 			goto retrylookup;
4994 		}
4995 	}
4996 
4997 	vm_page_assert_xbusied(m);
4998 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4999 		after = MIN(after, VM_INITIAL_PAGEIN);
5000 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
5001 		after = MAX(after, 1);
5002 		ma[0] = mpred = m;
5003 		pctrie_iter_reset(pages);
5004 		for (i = 1; i < after; i++) {
5005 			m = vm_radix_iter_lookup(pages, pindex + i);
5006 			if (m == NULL) {
5007 				m = vm_page_alloc_after(object, pages,
5008 				    pindex + i, VM_ALLOC_NORMAL, mpred);
5009 				if (m == NULL)
5010 					break;
5011 			} else if (vm_page_any_valid(m) || !vm_page_tryxbusy(m))
5012 				break;
5013 			mpred = ma[i] = m;
5014 		}
5015 		after = i;
5016 		vm_object_pip_add(object, after);
5017 		VM_OBJECT_WUNLOCK(object);
5018 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
5019 		VM_OBJECT_WLOCK(object);
5020 		vm_object_pip_wakeupn(object, after);
5021 		/* Pager may have replaced a page. */
5022 		m = ma[0];
5023 		if (rv != VM_PAGER_OK) {
5024 			for (i = 0; i < after; i++) {
5025 				if (!vm_page_wired(ma[i]))
5026 					vm_page_free(ma[i]);
5027 				else
5028 					vm_page_xunbusy(ma[i]);
5029 			}
5030 			*mp = NULL;
5031 			return (rv);
5032 		}
5033 		for (i = 1; i < after; i++)
5034 			vm_page_readahead_finish(ma[i]);
5035 		MPASS(vm_page_all_valid(m));
5036 	} else {
5037 		vm_page_zero_invalid(m, TRUE);
5038 	}
5039 	pctrie_iter_reset(pages);
5040 out:
5041 	if ((allocflags & VM_ALLOC_WIRED) != 0)
5042 		vm_page_wire(m);
5043 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
5044 		vm_page_busy_downgrade(m);
5045 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
5046 		vm_page_busy_release(m);
5047 	*mp = m;
5048 	return (VM_PAGER_OK);
5049 }
5050 
5051 /*
5052  * Grab a page and make it valid, paging in if necessary.  Pages missing from
5053  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
5054  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
5055  * in simultaneously.  Additional pages will be left on a paging queue but
5056  * will neither be wired nor busy regardless of allocflags.
5057  */
5058 int
5059 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex,
5060     int allocflags)
5061 {
5062 	struct pctrie_iter pages;
5063 
5064 	VM_OBJECT_ASSERT_WLOCKED(object);
5065 	vm_page_iter_init(&pages, object);
5066 	return (vm_page_grab_valid_iter(mp, object, &pages, pindex,
5067 	    allocflags));
5068 }
5069 
5070 /*
5071  * Grab a page.  Keep on waiting, as long as the page exists in the object.  If
5072  * the page doesn't exist, and the pager has it, allocate it and zero part of
5073  * it.
5074  *
5075  * The object must be locked on entry.  This routine may sleep.  The lock will,
5076  * however, be released and reacquired if the routine sleeps.
5077  */
5078 int
5079 vm_page_grab_zero_partial(vm_object_t object, vm_pindex_t pindex, int base,
5080     int end)
5081 {
5082 	struct pctrie_iter pages;
5083 	vm_page_t m, mpred;
5084 	int allocflags, rv;
5085 	bool found;
5086 
5087 	VM_OBJECT_ASSERT_WLOCKED(object);
5088 	KASSERT(base >= 0, ("%s: base %d", __func__, base));
5089 	KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base,
5090 	    end));
5091 
5092 	allocflags = VM_ALLOC_NOCREAT | VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL;
5093 	vm_page_iter_init(&pages, object);
5094 	while ((m = vm_page_grab_lookup(
5095 	    &pages, object, pindex, allocflags, &found)) == NULL) {
5096 		if (!vm_pager_has_page(object, pindex, NULL, NULL))
5097 			return (0);
5098 		mpred = vm_radix_iter_lookup_lt(&pages, pindex);
5099 		m = vm_page_alloc_after(object, &pages, pindex,
5100 		    vm_page_grab_pflags(allocflags), mpred);
5101 		if (m != NULL) {
5102 			vm_object_pip_add(object, 1);
5103 			VM_OBJECT_WUNLOCK(object);
5104 			rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
5105 			VM_OBJECT_WLOCK(object);
5106 			vm_object_pip_wakeup(object);
5107 			if (rv != VM_PAGER_OK) {
5108 				vm_page_free(m);
5109 				return (EIO);
5110 			}
5111 
5112 			/*
5113 			 * Since the page was not resident, and therefore not
5114 			 * recently accessed, immediately enqueue it for
5115 			 * asynchronous laundering.  The current operation is
5116 			 * not regarded as an access.
5117 			 */
5118 			vm_page_launder(m);
5119 			break;
5120 		}
5121 	}
5122 
5123 	pmap_zero_page_area(m, base, end - base);
5124 	KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m));
5125 	vm_page_set_dirty(m);
5126 	vm_page_xunbusy(m);
5127 	return (0);
5128 }
5129 
5130 /*
5131  * Locklessly grab a valid page.  If the page is not valid or not yet
5132  * allocated this will fall back to the object lock method.
5133  */
5134 int
5135 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
5136     vm_pindex_t pindex, int allocflags)
5137 {
5138 	vm_page_t m;
5139 	int flags;
5140 	int error;
5141 
5142 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
5143 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
5144 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
5145 	    "mismatch"));
5146 	KASSERT((allocflags &
5147 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
5148 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
5149 
5150 	/*
5151 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
5152 	 * before we can inspect the valid field and return a wired page.
5153 	 */
5154 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
5155 	vm_page_grab_check(flags);
5156 	m = vm_page_acquire_unlocked(object, pindex, NULL, flags);
5157 	if (m == PAGE_NOT_ACQUIRED)
5158 		return (VM_PAGER_FAIL);
5159 	if (m != NULL) {
5160 		if (vm_page_all_valid(m)) {
5161 			if ((allocflags & VM_ALLOC_WIRED) != 0)
5162 				vm_page_wire(m);
5163 			vm_page_grab_release(m, allocflags);
5164 			*mp = m;
5165 			return (VM_PAGER_OK);
5166 		}
5167 		vm_page_busy_release(m);
5168 	}
5169 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5170 		*mp = NULL;
5171 		return (VM_PAGER_FAIL);
5172 	}
5173 	VM_OBJECT_WLOCK(object);
5174 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
5175 	VM_OBJECT_WUNLOCK(object);
5176 
5177 	return (error);
5178 }
5179 
5180 /*
5181  * Return the specified range of pages from the given object.  For each
5182  * page offset within the range, if a page already exists within the object
5183  * at that offset and it is busy, then wait for it to change state.  If,
5184  * instead, the page doesn't exist, then allocate it.
5185  *
5186  * The caller must always specify an allocation class.
5187  *
5188  * allocation classes:
5189  *	VM_ALLOC_NORMAL		normal process request
5190  *	VM_ALLOC_SYSTEM		system *really* needs the pages
5191  *
5192  * The caller must always specify that the pages are to be busied and/or
5193  * wired.
5194  *
5195  * optional allocation flags:
5196  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5197  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5198  *	VM_ALLOC_NOWAIT		do not sleep
5199  *	VM_ALLOC_SBUSY		set page to sbusy state
5200  *	VM_ALLOC_WIRED		wire the pages
5201  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5202  *
5203  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5204  * may return a partial prefix of the requested range.
5205  */
5206 int
5207 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5208     vm_page_t *ma, int count)
5209 {
5210 	struct pctrie_iter pages;
5211 	vm_page_t m, mpred;
5212 	int pflags;
5213 	int i;
5214 
5215 	VM_OBJECT_ASSERT_WLOCKED(object);
5216 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5217 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5218 	KASSERT(count > 0,
5219 	    ("vm_page_grab_pages: invalid page count %d", count));
5220 	vm_page_grab_check(allocflags);
5221 
5222 	pflags = vm_page_grab_pflags(allocflags);
5223 	i = 0;
5224 	vm_page_iter_init(&pages, object);
5225 retrylookup:
5226 	mpred = vm_radix_iter_lookup_lt(&pages, pindex + i);
5227 	for (; i < count; i++) {
5228 		m = vm_radix_iter_lookup(&pages, pindex + i);
5229 		if (m != NULL) {
5230 			if (!vm_page_tryacquire(m, allocflags)) {
5231 				if (vm_page_grab_sleep(object, m, pindex + i,
5232 				    "grbmaw", allocflags, true)) {
5233 					pctrie_iter_reset(&pages);
5234 					goto retrylookup;
5235 				}
5236 				break;
5237 			}
5238 		} else {
5239 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5240 				break;
5241 			m = vm_page_alloc_after(object, &pages, pindex + i,
5242 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5243 			if (m == NULL) {
5244 				if ((allocflags & (VM_ALLOC_NOWAIT |
5245 				    VM_ALLOC_WAITFAIL)) != 0)
5246 					break;
5247 				goto retrylookup;
5248 			}
5249 		}
5250 		if (vm_page_none_valid(m) &&
5251 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5252 			if ((m->flags & PG_ZERO) == 0)
5253 				pmap_zero_page(m);
5254 			vm_page_valid(m);
5255 		}
5256 		vm_page_grab_release(m, allocflags);
5257 		ma[i] = mpred = m;
5258 	}
5259 	return (i);
5260 }
5261 
5262 /*
5263  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5264  * and will fall back to the locked variant to handle allocation.
5265  */
5266 int
5267 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5268     int allocflags, vm_page_t *ma, int count)
5269 {
5270 	vm_page_t m;
5271 	int flags;
5272 	int i;
5273 
5274 	KASSERT(count > 0,
5275 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5276 	vm_page_grab_check(allocflags);
5277 
5278 	/*
5279 	 * Modify flags for lockless acquire to hold the page until we
5280 	 * set it valid if necessary.
5281 	 */
5282 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5283 	vm_page_grab_check(flags);
5284 	m = NULL;
5285 	for (i = 0; i < count; i++, pindex++) {
5286 		/*
5287 		 * We may see a false NULL here because the previous page has
5288 		 * been removed or just inserted and the list is loaded without
5289 		 * barriers.  Switch to radix to verify.
5290 		 */
5291 		if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex ||
5292 		    atomic_load_ptr(&m->object) != object) {
5293 			/*
5294 			 * This guarantees the result is instantaneously
5295 			 * correct.
5296 			 */
5297 			m = NULL;
5298 		}
5299 		m = vm_page_acquire_unlocked(object, pindex, m, flags);
5300 		if (m == PAGE_NOT_ACQUIRED)
5301 			return (i);
5302 		if (m == NULL)
5303 			break;
5304 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5305 			if ((m->flags & PG_ZERO) == 0)
5306 				pmap_zero_page(m);
5307 			vm_page_valid(m);
5308 		}
5309 		/* m will still be wired or busy according to flags. */
5310 		vm_page_grab_release(m, allocflags);
5311 		ma[i] = m;
5312 		m = TAILQ_NEXT(m, listq);
5313 	}
5314 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5315 		return (i);
5316 	count -= i;
5317 	VM_OBJECT_WLOCK(object);
5318 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5319 	VM_OBJECT_WUNLOCK(object);
5320 
5321 	return (i);
5322 }
5323 
5324 /*
5325  * Mapping function for valid or dirty bits in a page.
5326  *
5327  * Inputs are required to range within a page.
5328  */
5329 vm_page_bits_t
5330 vm_page_bits(int base, int size)
5331 {
5332 	int first_bit;
5333 	int last_bit;
5334 
5335 	KASSERT(
5336 	    base + size <= PAGE_SIZE,
5337 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5338 	);
5339 
5340 	if (size == 0)		/* handle degenerate case */
5341 		return (0);
5342 
5343 	first_bit = base >> DEV_BSHIFT;
5344 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5345 
5346 	return (((vm_page_bits_t)2 << last_bit) -
5347 	    ((vm_page_bits_t)1 << first_bit));
5348 }
5349 
5350 void
5351 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5352 {
5353 
5354 #if PAGE_SIZE == 32768
5355 	atomic_set_64((uint64_t *)bits, set);
5356 #elif PAGE_SIZE == 16384
5357 	atomic_set_32((uint32_t *)bits, set);
5358 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5359 	atomic_set_16((uint16_t *)bits, set);
5360 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5361 	atomic_set_8((uint8_t *)bits, set);
5362 #else		/* PAGE_SIZE <= 8192 */
5363 	uintptr_t addr;
5364 	int shift;
5365 
5366 	addr = (uintptr_t)bits;
5367 	/*
5368 	 * Use a trick to perform a 32-bit atomic on the
5369 	 * containing aligned word, to not depend on the existence
5370 	 * of atomic_{set, clear}_{8, 16}.
5371 	 */
5372 	shift = addr & (sizeof(uint32_t) - 1);
5373 #if BYTE_ORDER == BIG_ENDIAN
5374 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5375 #else
5376 	shift *= NBBY;
5377 #endif
5378 	addr &= ~(sizeof(uint32_t) - 1);
5379 	atomic_set_32((uint32_t *)addr, set << shift);
5380 #endif		/* PAGE_SIZE */
5381 }
5382 
5383 static inline void
5384 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5385 {
5386 
5387 #if PAGE_SIZE == 32768
5388 	atomic_clear_64((uint64_t *)bits, clear);
5389 #elif PAGE_SIZE == 16384
5390 	atomic_clear_32((uint32_t *)bits, clear);
5391 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5392 	atomic_clear_16((uint16_t *)bits, clear);
5393 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5394 	atomic_clear_8((uint8_t *)bits, clear);
5395 #else		/* PAGE_SIZE <= 8192 */
5396 	uintptr_t addr;
5397 	int shift;
5398 
5399 	addr = (uintptr_t)bits;
5400 	/*
5401 	 * Use a trick to perform a 32-bit atomic on the
5402 	 * containing aligned word, to not depend on the existence
5403 	 * of atomic_{set, clear}_{8, 16}.
5404 	 */
5405 	shift = addr & (sizeof(uint32_t) - 1);
5406 #if BYTE_ORDER == BIG_ENDIAN
5407 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5408 #else
5409 	shift *= NBBY;
5410 #endif
5411 	addr &= ~(sizeof(uint32_t) - 1);
5412 	atomic_clear_32((uint32_t *)addr, clear << shift);
5413 #endif		/* PAGE_SIZE */
5414 }
5415 
5416 static inline vm_page_bits_t
5417 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5418 {
5419 #if PAGE_SIZE == 32768
5420 	uint64_t old;
5421 
5422 	old = *bits;
5423 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5424 	return (old);
5425 #elif PAGE_SIZE == 16384
5426 	uint32_t old;
5427 
5428 	old = *bits;
5429 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5430 	return (old);
5431 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5432 	uint16_t old;
5433 
5434 	old = *bits;
5435 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5436 	return (old);
5437 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5438 	uint8_t old;
5439 
5440 	old = *bits;
5441 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5442 	return (old);
5443 #else		/* PAGE_SIZE <= 4096*/
5444 	uintptr_t addr;
5445 	uint32_t old, new, mask;
5446 	int shift;
5447 
5448 	addr = (uintptr_t)bits;
5449 	/*
5450 	 * Use a trick to perform a 32-bit atomic on the
5451 	 * containing aligned word, to not depend on the existence
5452 	 * of atomic_{set, swap, clear}_{8, 16}.
5453 	 */
5454 	shift = addr & (sizeof(uint32_t) - 1);
5455 #if BYTE_ORDER == BIG_ENDIAN
5456 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5457 #else
5458 	shift *= NBBY;
5459 #endif
5460 	addr &= ~(sizeof(uint32_t) - 1);
5461 	mask = VM_PAGE_BITS_ALL << shift;
5462 
5463 	old = *bits;
5464 	do {
5465 		new = old & ~mask;
5466 		new |= newbits << shift;
5467 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5468 	return (old >> shift);
5469 #endif		/* PAGE_SIZE */
5470 }
5471 
5472 /*
5473  *	vm_page_set_valid_range:
5474  *
5475  *	Sets portions of a page valid.  The arguments are expected
5476  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5477  *	of any partial chunks touched by the range.  The invalid portion of
5478  *	such chunks will be zeroed.
5479  *
5480  *	(base + size) must be less then or equal to PAGE_SIZE.
5481  */
5482 void
5483 vm_page_set_valid_range(vm_page_t m, int base, int size)
5484 {
5485 	int endoff, frag;
5486 	vm_page_bits_t pagebits;
5487 
5488 	vm_page_assert_busied(m);
5489 	if (size == 0)	/* handle degenerate case */
5490 		return;
5491 
5492 	/*
5493 	 * If the base is not DEV_BSIZE aligned and the valid
5494 	 * bit is clear, we have to zero out a portion of the
5495 	 * first block.
5496 	 */
5497 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5498 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5499 		pmap_zero_page_area(m, frag, base - frag);
5500 
5501 	/*
5502 	 * If the ending offset is not DEV_BSIZE aligned and the
5503 	 * valid bit is clear, we have to zero out a portion of
5504 	 * the last block.
5505 	 */
5506 	endoff = base + size;
5507 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5508 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5509 		pmap_zero_page_area(m, endoff,
5510 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5511 
5512 	/*
5513 	 * Assert that no previously invalid block that is now being validated
5514 	 * is already dirty.
5515 	 */
5516 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5517 	    ("vm_page_set_valid_range: page %p is dirty", m));
5518 
5519 	/*
5520 	 * Set valid bits inclusive of any overlap.
5521 	 */
5522 	pagebits = vm_page_bits(base, size);
5523 	if (vm_page_xbusied(m))
5524 		m->valid |= pagebits;
5525 	else
5526 		vm_page_bits_set(m, &m->valid, pagebits);
5527 }
5528 
5529 /*
5530  * Set the page dirty bits and free the invalid swap space if
5531  * present.  Returns the previous dirty bits.
5532  */
5533 vm_page_bits_t
5534 vm_page_set_dirty(vm_page_t m)
5535 {
5536 	vm_page_bits_t old;
5537 
5538 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5539 
5540 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5541 		old = m->dirty;
5542 		m->dirty = VM_PAGE_BITS_ALL;
5543 	} else
5544 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5545 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5546 		vm_pager_page_unswapped(m);
5547 
5548 	return (old);
5549 }
5550 
5551 /*
5552  * Clear the given bits from the specified page's dirty field.
5553  */
5554 static __inline void
5555 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5556 {
5557 
5558 	vm_page_assert_busied(m);
5559 
5560 	/*
5561 	 * If the page is xbusied and not write mapped we are the
5562 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5563 	 * layer can call vm_page_dirty() without holding a distinguished
5564 	 * lock.  The combination of page busy and atomic operations
5565 	 * suffice to guarantee consistency of the page dirty field.
5566 	 */
5567 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5568 		m->dirty &= ~pagebits;
5569 	else
5570 		vm_page_bits_clear(m, &m->dirty, pagebits);
5571 }
5572 
5573 /*
5574  *	vm_page_set_validclean:
5575  *
5576  *	Sets portions of a page valid and clean.  The arguments are expected
5577  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5578  *	of any partial chunks touched by the range.  The invalid portion of
5579  *	such chunks will be zero'd.
5580  *
5581  *	(base + size) must be less then or equal to PAGE_SIZE.
5582  */
5583 void
5584 vm_page_set_validclean(vm_page_t m, int base, int size)
5585 {
5586 	vm_page_bits_t oldvalid, pagebits;
5587 	int endoff, frag;
5588 
5589 	vm_page_assert_busied(m);
5590 	if (size == 0)	/* handle degenerate case */
5591 		return;
5592 
5593 	/*
5594 	 * If the base is not DEV_BSIZE aligned and the valid
5595 	 * bit is clear, we have to zero out a portion of the
5596 	 * first block.
5597 	 */
5598 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5599 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5600 		pmap_zero_page_area(m, frag, base - frag);
5601 
5602 	/*
5603 	 * If the ending offset is not DEV_BSIZE aligned and the
5604 	 * valid bit is clear, we have to zero out a portion of
5605 	 * the last block.
5606 	 */
5607 	endoff = base + size;
5608 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5609 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5610 		pmap_zero_page_area(m, endoff,
5611 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5612 
5613 	/*
5614 	 * Set valid, clear dirty bits.  If validating the entire
5615 	 * page we can safely clear the pmap modify bit.  We also
5616 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5617 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5618 	 * be set again.
5619 	 *
5620 	 * We set valid bits inclusive of any overlap, but we can only
5621 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5622 	 * the range.
5623 	 */
5624 	oldvalid = m->valid;
5625 	pagebits = vm_page_bits(base, size);
5626 	if (vm_page_xbusied(m))
5627 		m->valid |= pagebits;
5628 	else
5629 		vm_page_bits_set(m, &m->valid, pagebits);
5630 #if 0	/* NOT YET */
5631 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5632 		frag = DEV_BSIZE - frag;
5633 		base += frag;
5634 		size -= frag;
5635 		if (size < 0)
5636 			size = 0;
5637 	}
5638 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5639 #endif
5640 	if (base == 0 && size == PAGE_SIZE) {
5641 		/*
5642 		 * The page can only be modified within the pmap if it is
5643 		 * mapped, and it can only be mapped if it was previously
5644 		 * fully valid.
5645 		 */
5646 		if (oldvalid == VM_PAGE_BITS_ALL)
5647 			/*
5648 			 * Perform the pmap_clear_modify() first.  Otherwise,
5649 			 * a concurrent pmap operation, such as
5650 			 * pmap_protect(), could clear a modification in the
5651 			 * pmap and set the dirty field on the page before
5652 			 * pmap_clear_modify() had begun and after the dirty
5653 			 * field was cleared here.
5654 			 */
5655 			pmap_clear_modify(m);
5656 		m->dirty = 0;
5657 		vm_page_aflag_clear(m, PGA_NOSYNC);
5658 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5659 		m->dirty &= ~pagebits;
5660 	else
5661 		vm_page_clear_dirty_mask(m, pagebits);
5662 }
5663 
5664 void
5665 vm_page_clear_dirty(vm_page_t m, int base, int size)
5666 {
5667 
5668 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5669 }
5670 
5671 /*
5672  *	vm_page_set_invalid:
5673  *
5674  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5675  *	valid and dirty bits for the effected areas are cleared.
5676  */
5677 void
5678 vm_page_set_invalid(vm_page_t m, int base, int size)
5679 {
5680 	vm_page_bits_t bits;
5681 	vm_object_t object;
5682 
5683 	/*
5684 	 * The object lock is required so that pages can't be mapped
5685 	 * read-only while we're in the process of invalidating them.
5686 	 */
5687 	object = m->object;
5688 	VM_OBJECT_ASSERT_WLOCKED(object);
5689 	vm_page_assert_busied(m);
5690 
5691 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5692 	    size >= object->un_pager.vnp.vnp_size)
5693 		bits = VM_PAGE_BITS_ALL;
5694 	else
5695 		bits = vm_page_bits(base, size);
5696 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5697 		pmap_remove_all(m);
5698 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5699 	    !pmap_page_is_mapped(m),
5700 	    ("vm_page_set_invalid: page %p is mapped", m));
5701 	if (vm_page_xbusied(m)) {
5702 		m->valid &= ~bits;
5703 		m->dirty &= ~bits;
5704 	} else {
5705 		vm_page_bits_clear(m, &m->valid, bits);
5706 		vm_page_bits_clear(m, &m->dirty, bits);
5707 	}
5708 }
5709 
5710 /*
5711  *	vm_page_invalid:
5712  *
5713  *	Invalidates the entire page.  The page must be busy, unmapped, and
5714  *	the enclosing object must be locked.  The object locks protects
5715  *	against concurrent read-only pmap enter which is done without
5716  *	busy.
5717  */
5718 void
5719 vm_page_invalid(vm_page_t m)
5720 {
5721 
5722 	vm_page_assert_busied(m);
5723 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5724 	MPASS(!pmap_page_is_mapped(m));
5725 
5726 	if (vm_page_xbusied(m))
5727 		m->valid = 0;
5728 	else
5729 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5730 }
5731 
5732 /*
5733  * vm_page_zero_invalid()
5734  *
5735  *	The kernel assumes that the invalid portions of a page contain
5736  *	garbage, but such pages can be mapped into memory by user code.
5737  *	When this occurs, we must zero out the non-valid portions of the
5738  *	page so user code sees what it expects.
5739  *
5740  *	Pages are most often semi-valid when the end of a file is mapped
5741  *	into memory and the file's size is not page aligned.
5742  */
5743 void
5744 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5745 {
5746 	int b;
5747 	int i;
5748 
5749 	/*
5750 	 * Scan the valid bits looking for invalid sections that
5751 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5752 	 * valid bit may be set ) have already been zeroed by
5753 	 * vm_page_set_validclean().
5754 	 */
5755 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5756 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5757 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5758 			if (i > b) {
5759 				pmap_zero_page_area(m,
5760 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5761 			}
5762 			b = i + 1;
5763 		}
5764 	}
5765 
5766 	/*
5767 	 * setvalid is TRUE when we can safely set the zero'd areas
5768 	 * as being valid.  We can do this if there are no cache consistency
5769 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5770 	 */
5771 	if (setvalid)
5772 		vm_page_valid(m);
5773 }
5774 
5775 /*
5776  *	vm_page_is_valid:
5777  *
5778  *	Is (partial) page valid?  Note that the case where size == 0
5779  *	will return FALSE in the degenerate case where the page is
5780  *	entirely invalid, and TRUE otherwise.
5781  *
5782  *	Some callers envoke this routine without the busy lock held and
5783  *	handle races via higher level locks.  Typical callers should
5784  *	hold a busy lock to prevent invalidation.
5785  */
5786 int
5787 vm_page_is_valid(vm_page_t m, int base, int size)
5788 {
5789 	vm_page_bits_t bits;
5790 
5791 	bits = vm_page_bits(base, size);
5792 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5793 }
5794 
5795 /*
5796  * Returns true if all of the specified predicates are true for the entire
5797  * (super)page and false otherwise.
5798  */
5799 bool
5800 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5801 {
5802 	vm_object_t object;
5803 	int i, npages;
5804 
5805 	object = m->object;
5806 	if (skip_m != NULL && skip_m->object != object)
5807 		return (false);
5808 	VM_OBJECT_ASSERT_LOCKED(object);
5809 	KASSERT(psind <= m->psind,
5810 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5811 	npages = atop(pagesizes[psind]);
5812 
5813 	/*
5814 	 * The physically contiguous pages that make up a superpage, i.e., a
5815 	 * page with a page size index ("psind") greater than zero, will
5816 	 * occupy adjacent entries in vm_page_array[].
5817 	 */
5818 	for (i = 0; i < npages; i++) {
5819 		/* Always test object consistency, including "skip_m". */
5820 		if (m[i].object != object)
5821 			return (false);
5822 		if (&m[i] == skip_m)
5823 			continue;
5824 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5825 			return (false);
5826 		if ((flags & PS_ALL_DIRTY) != 0) {
5827 			/*
5828 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5829 			 * might stop this case from spuriously returning
5830 			 * "false".  However, that would require a write lock
5831 			 * on the object containing "m[i]".
5832 			 */
5833 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5834 				return (false);
5835 		}
5836 		if ((flags & PS_ALL_VALID) != 0 &&
5837 		    m[i].valid != VM_PAGE_BITS_ALL)
5838 			return (false);
5839 	}
5840 	return (true);
5841 }
5842 
5843 /*
5844  * Set the page's dirty bits if the page is modified.
5845  */
5846 void
5847 vm_page_test_dirty(vm_page_t m)
5848 {
5849 
5850 	vm_page_assert_busied(m);
5851 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5852 		vm_page_dirty(m);
5853 }
5854 
5855 void
5856 vm_page_valid(vm_page_t m)
5857 {
5858 
5859 	vm_page_assert_busied(m);
5860 	if (vm_page_xbusied(m))
5861 		m->valid = VM_PAGE_BITS_ALL;
5862 	else
5863 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5864 }
5865 
5866 void
5867 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5868 {
5869 
5870 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5871 }
5872 
5873 void
5874 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5875 {
5876 
5877 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5878 }
5879 
5880 int
5881 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5882 {
5883 
5884 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5885 }
5886 
5887 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5888 void
5889 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5890 {
5891 
5892 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5893 }
5894 
5895 void
5896 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5897 {
5898 
5899 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5900 }
5901 #endif
5902 
5903 #ifdef INVARIANTS
5904 void
5905 vm_page_object_busy_assert(vm_page_t m)
5906 {
5907 
5908 	/*
5909 	 * Certain of the page's fields may only be modified by the
5910 	 * holder of a page or object busy.
5911 	 */
5912 	if (m->object != NULL && !vm_page_busied(m))
5913 		VM_OBJECT_ASSERT_BUSY(m->object);
5914 }
5915 
5916 void
5917 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5918 {
5919 
5920 	if ((bits & PGA_WRITEABLE) == 0)
5921 		return;
5922 
5923 	/*
5924 	 * The PGA_WRITEABLE flag can only be set if the page is
5925 	 * managed, is exclusively busied or the object is locked.
5926 	 * Currently, this flag is only set by pmap_enter().
5927 	 */
5928 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5929 	    ("PGA_WRITEABLE on unmanaged page"));
5930 	if (!vm_page_xbusied(m))
5931 		VM_OBJECT_ASSERT_BUSY(m->object);
5932 }
5933 #endif
5934 
5935 #include "opt_ddb.h"
5936 #ifdef DDB
5937 #include <sys/kernel.h>
5938 
5939 #include <ddb/ddb.h>
5940 
5941 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5942 {
5943 
5944 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5945 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5946 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5947 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5948 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5949 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5950 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5951 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5952 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5953 }
5954 
5955 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5956 {
5957 	int dom;
5958 
5959 	db_printf("pq_free %d\n", vm_free_count());
5960 	for (dom = 0; dom < vm_ndomains; dom++) {
5961 		db_printf(
5962     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5963 		    dom,
5964 		    vm_dom[dom].vmd_page_count,
5965 		    vm_dom[dom].vmd_free_count,
5966 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5967 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5968 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5969 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5970 	}
5971 }
5972 
5973 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5974 {
5975 	vm_page_t m;
5976 	boolean_t phys, virt;
5977 
5978 	if (!have_addr) {
5979 		db_printf("show pginfo addr\n");
5980 		return;
5981 	}
5982 
5983 	phys = strchr(modif, 'p') != NULL;
5984 	virt = strchr(modif, 'v') != NULL;
5985 	if (virt)
5986 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5987 	else if (phys)
5988 		m = PHYS_TO_VM_PAGE(addr);
5989 	else
5990 		m = (vm_page_t)addr;
5991 	db_printf(
5992     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5993     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5994 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5995 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5996 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5997 }
5998 #endif /* DDB */
5999