xref: /freebsd/sys/vm/vm_page.c (revision 160a2ba804973e4b258c24247fa7c0cdc230dfb4)
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  *	vm_page_find_least:
1880  *
1881  *	Returns the page associated with the object with least pindex
1882  *	greater than or equal to the parameter pindex, or NULL.
1883  *
1884  *	The object must be locked.
1885  */
1886 vm_page_t
1887 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1888 {
1889 	vm_page_t m;
1890 
1891 	VM_OBJECT_ASSERT_LOCKED(object);
1892 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1893 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1894 	return (m);
1895 }
1896 
1897 /*
1898  * Uses the page mnew as a replacement for an existing page at index
1899  * pindex which must be already present in the object.
1900  *
1901  * Both pages must be exclusively busied on enter.  The old page is
1902  * unbusied on exit.
1903  *
1904  * A return value of true means mold is now free.  If this is not the
1905  * final ref and the caller does not hold a wire reference it may not
1906  * continue to access the page.
1907  */
1908 static bool
1909 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1910     vm_page_t mold)
1911 {
1912 	vm_page_t mret __diagused;
1913 	bool dropped;
1914 
1915 	VM_OBJECT_ASSERT_WLOCKED(object);
1916 	vm_page_assert_xbusied(mold);
1917 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1918 	    ("vm_page_replace: page %p already in object", mnew));
1919 
1920 	/*
1921 	 * This function mostly follows vm_page_insert() and
1922 	 * vm_page_remove() without the radix, object count and vnode
1923 	 * dance.  Double check such functions for more comments.
1924 	 */
1925 
1926 	mnew->object = object;
1927 	mnew->pindex = pindex;
1928 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1929 	mret = vm_radix_replace(&object->rtree, mnew);
1930 	KASSERT(mret == mold,
1931 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1932 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1933 	    (mnew->oflags & VPO_UNMANAGED),
1934 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1935 
1936 	/* Keep the resident page list in sorted order. */
1937 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1938 	TAILQ_REMOVE(&object->memq, mold, listq);
1939 	mold->object = NULL;
1940 
1941 	/*
1942 	 * The object's resident_page_count does not change because we have
1943 	 * swapped one page for another, but the generation count should
1944 	 * change if the page is dirty.
1945 	 */
1946 	if (pmap_page_is_write_mapped(mnew))
1947 		vm_object_set_writeable_dirty(object);
1948 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1949 	vm_page_xunbusy(mold);
1950 
1951 	return (dropped);
1952 }
1953 
1954 void
1955 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1956     vm_page_t mold)
1957 {
1958 
1959 	vm_page_assert_xbusied(mnew);
1960 
1961 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1962 		vm_page_free(mold);
1963 }
1964 
1965 /*
1966  *	vm_page_iter_rename:
1967  *
1968  *	Tries to move the specified page from its current object to a new object
1969  *	and pindex, using the given iterator to remove the page from its current
1970  *	object.  Returns true if the move was successful, and false if the move
1971  *	was aborted due to a failed memory allocation.
1972  *
1973  *	Panics if a page already resides in the new object at the new pindex.
1974  *
1975  *	This routine dirties the page if it is valid, as callers are expected to
1976  *	transfer backing storage only after moving the page.  Dirtying the page
1977  *	ensures that the destination object retains the most recent copy of the
1978  *	page.
1979  *
1980  *	The objects must be locked.
1981  */
1982 bool
1983 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m,
1984     vm_object_t new_object, vm_pindex_t new_pindex)
1985 {
1986 	vm_page_t mpred;
1987 	vm_pindex_t opidx;
1988 
1989 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1990 	    ("%s: page %p is missing object ref", __func__, m));
1991 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1992 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1993 
1994 	/*
1995 	 * Create a custom version of vm_page_insert() which does not depend
1996 	 * by m_prev and can cheat on the implementation aspects of the
1997 	 * function.
1998 	 */
1999 	opidx = m->pindex;
2000 	m->pindex = new_pindex;
2001 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
2002 		m->pindex = opidx;
2003 		return (false);
2004 	}
2005 
2006 	/*
2007 	 * The operation cannot fail anymore.  The removal must happen before
2008 	 * the listq iterator is tainted.
2009 	 */
2010 	m->pindex = opidx;
2011 	vm_radix_iter_remove(old_pages);
2012 	vm_page_remove_radixdone(m);
2013 
2014 	/* Return back to the new pindex to complete vm_page_insert(). */
2015 	m->pindex = new_pindex;
2016 	m->object = new_object;
2017 
2018 	vm_page_insert_radixdone(m, new_object, mpred);
2019 	if (vm_page_any_valid(m))
2020 		vm_page_dirty(m);
2021 	vm_pager_page_inserted(new_object, m);
2022 	return (true);
2023 }
2024 
2025 /*
2026  *	vm_page_mpred:
2027  *
2028  *	Return the greatest page of the object with index <= pindex,
2029  *	or NULL, if there is none.  Assumes object lock is held.
2030  */
2031 vm_page_t
2032 vm_page_mpred(vm_object_t object, vm_pindex_t pindex)
2033 {
2034 	return (vm_radix_lookup_le(&object->rtree, pindex));
2035 }
2036 
2037 /*
2038  *	vm_page_alloc:
2039  *
2040  *	Allocate and return a page that is associated with the specified
2041  *	object and offset pair.  By default, this page is exclusive busied.
2042  *
2043  *	The caller must always specify an allocation class.
2044  *
2045  *	allocation classes:
2046  *	VM_ALLOC_NORMAL		normal process request
2047  *	VM_ALLOC_SYSTEM		system *really* needs a page
2048  *	VM_ALLOC_INTERRUPT	interrupt time request
2049  *
2050  *	optional allocation flags:
2051  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2052  *				intends to allocate
2053  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2054  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2055  *	VM_ALLOC_SBUSY		shared busy the allocated page
2056  *	VM_ALLOC_WIRED		wire the allocated page
2057  *	VM_ALLOC_ZERO		prefer a zeroed page
2058  */
2059 vm_page_t
2060 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
2061 {
2062 	struct pctrie_iter pages;
2063 
2064 	vm_page_iter_init(&pages, object);
2065 	return (vm_page_alloc_after(object, &pages, pindex, req,
2066 	    vm_page_mpred(object, pindex)));
2067 }
2068 
2069 /*
2070  * Allocate a page in the specified object with the given page index.  To
2071  * optimize insertion of the page into the object, the caller must also specify
2072  * the resident page in the object with largest index smaller than the given
2073  * page index, or NULL if no such page exists.
2074  */
2075 vm_page_t
2076 vm_page_alloc_after(vm_object_t object, struct pctrie_iter *pages,
2077     vm_pindex_t pindex, int req, vm_page_t mpred)
2078 {
2079 	struct vm_domainset_iter di;
2080 	vm_page_t m;
2081 	int domain;
2082 
2083 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2084 	do {
2085 		m = vm_page_alloc_domain_after(object, pages, pindex, domain,
2086 		    req, mpred);
2087 		if (m != NULL)
2088 			break;
2089 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2090 
2091 	return (m);
2092 }
2093 
2094 /*
2095  * Returns true if the number of free pages exceeds the minimum
2096  * for the request class and false otherwise.
2097  */
2098 static int
2099 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2100 {
2101 	u_int limit, old, new;
2102 
2103 	if (req_class == VM_ALLOC_INTERRUPT)
2104 		limit = 0;
2105 	else if (req_class == VM_ALLOC_SYSTEM)
2106 		limit = vmd->vmd_interrupt_free_min;
2107 	else
2108 		limit = vmd->vmd_free_reserved;
2109 
2110 	/*
2111 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2112 	 */
2113 	limit += npages;
2114 	old = atomic_load_int(&vmd->vmd_free_count);
2115 	do {
2116 		if (old < limit)
2117 			return (0);
2118 		new = old - npages;
2119 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2120 
2121 	/* Wake the page daemon if we've crossed the threshold. */
2122 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2123 		pagedaemon_wakeup(vmd->vmd_domain);
2124 
2125 	/* Only update bitsets on transitions. */
2126 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2127 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2128 		vm_domain_set(vmd);
2129 
2130 	return (1);
2131 }
2132 
2133 int
2134 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2135 {
2136 	int req_class;
2137 
2138 	/*
2139 	 * The page daemon is allowed to dig deeper into the free page list.
2140 	 */
2141 	req_class = req & VM_ALLOC_CLASS_MASK;
2142 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2143 		req_class = VM_ALLOC_SYSTEM;
2144 	return (_vm_domain_allocate(vmd, req_class, npages));
2145 }
2146 
2147 vm_page_t
2148 vm_page_alloc_domain_after(vm_object_t object, struct pctrie_iter *pages,
2149     vm_pindex_t pindex, int domain, int req, vm_page_t mpred)
2150 {
2151 	struct vm_domain *vmd;
2152 	vm_page_t m;
2153 	int flags;
2154 
2155 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2156 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2157 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2158 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2159 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2160 	KASSERT((req & ~VPA_FLAGS) == 0,
2161 	    ("invalid request %#x", req));
2162 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2163 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2164 	    ("invalid request %#x", req));
2165 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2166 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2167 	    (uintmax_t)pindex));
2168 	VM_OBJECT_ASSERT_WLOCKED(object);
2169 
2170 	flags = 0;
2171 	m = NULL;
2172 	if (!vm_pager_can_alloc_page(object, pindex))
2173 		return (NULL);
2174 again:
2175 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2176 		m = vm_page_alloc_nofree_domain(domain, req);
2177 		if (m != NULL)
2178 			goto found;
2179 	}
2180 #if VM_NRESERVLEVEL > 0
2181 	/*
2182 	 * Can we allocate the page from a reservation?
2183 	 */
2184 	if (vm_object_reserv(object) &&
2185 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2186 	    NULL) {
2187 		goto found;
2188 	}
2189 #endif
2190 	vmd = VM_DOMAIN(domain);
2191 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2192 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2193 		    M_NOWAIT | M_NOVM);
2194 		if (m != NULL) {
2195 			flags |= PG_PCPU_CACHE;
2196 			goto found;
2197 		}
2198 	}
2199 	if (vm_domain_allocate(vmd, req, 1)) {
2200 		/*
2201 		 * If not, allocate it from the free page queues.
2202 		 */
2203 		vm_domain_free_lock(vmd);
2204 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2205 		vm_domain_free_unlock(vmd);
2206 		if (m == NULL) {
2207 			vm_domain_freecnt_inc(vmd, 1);
2208 #if VM_NRESERVLEVEL > 0
2209 			if (vm_reserv_reclaim_inactive(domain))
2210 				goto again;
2211 #endif
2212 		}
2213 	}
2214 	if (m == NULL) {
2215 		/*
2216 		 * Not allocatable, give up.
2217 		 */
2218 		pctrie_iter_reset(pages);
2219 		if (vm_domain_alloc_fail(vmd, object, req))
2220 			goto again;
2221 		return (NULL);
2222 	}
2223 
2224 	/*
2225 	 * At this point we had better have found a good page.
2226 	 */
2227 found:
2228 	vm_page_dequeue(m);
2229 	vm_page_alloc_check(m);
2230 
2231 	/*
2232 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2233 	 */
2234 	flags |= m->flags & PG_ZERO;
2235 	if ((req & VM_ALLOC_NODUMP) != 0)
2236 		flags |= PG_NODUMP;
2237 	if ((req & VM_ALLOC_NOFREE) != 0)
2238 		flags |= PG_NOFREE;
2239 	m->flags = flags;
2240 	m->a.flags = 0;
2241 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2242 	m->pool = VM_FREEPOOL_DEFAULT;
2243 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2244 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2245 	else if ((req & VM_ALLOC_SBUSY) != 0)
2246 		m->busy_lock = VPB_SHARERS_WORD(1);
2247 	else
2248 		m->busy_lock = VPB_UNBUSIED;
2249 	if (req & VM_ALLOC_WIRED) {
2250 		vm_wire_add(1);
2251 		m->ref_count = 1;
2252 	}
2253 	m->a.act_count = 0;
2254 
2255 	if (vm_page_insert_lookup(m, object, pindex, pages, true, mpred)) {
2256 		if (req & VM_ALLOC_WIRED) {
2257 			vm_wire_sub(1);
2258 			m->ref_count = 0;
2259 		}
2260 		KASSERT(m->object == NULL, ("page %p has object", m));
2261 		m->oflags = VPO_UNMANAGED;
2262 		m->busy_lock = VPB_UNBUSIED;
2263 		/* Don't change PG_ZERO. */
2264 		vm_page_free_toq(m);
2265 		if (req & VM_ALLOC_WAITFAIL) {
2266 			VM_OBJECT_WUNLOCK(object);
2267 			vm_radix_wait();
2268 			pctrie_iter_reset(pages);
2269 			VM_OBJECT_WLOCK(object);
2270 		}
2271 		return (NULL);
2272 	}
2273 
2274 	/* Ignore device objects; the pager sets "memattr" for them. */
2275 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2276 	    (object->flags & OBJ_FICTITIOUS) == 0)
2277 		pmap_page_set_memattr(m, object->memattr);
2278 
2279 	return (m);
2280 }
2281 
2282 /*
2283  *	vm_page_alloc_contig:
2284  *
2285  *	Allocate a contiguous set of physical pages of the given size "npages"
2286  *	from the free lists.  All of the physical pages must be at or above
2287  *	the given physical address "low" and below the given physical address
2288  *	"high".  The given value "alignment" determines the alignment of the
2289  *	first physical page in the set.  If the given value "boundary" is
2290  *	non-zero, then the set of physical pages cannot cross any physical
2291  *	address boundary that is a multiple of that value.  Both "alignment"
2292  *	and "boundary" must be a power of two.
2293  *
2294  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2295  *	then the memory attribute setting for the physical pages is configured
2296  *	to the object's memory attribute setting.  Otherwise, the memory
2297  *	attribute setting for the physical pages is configured to "memattr",
2298  *	overriding the object's memory attribute setting.  However, if the
2299  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2300  *	memory attribute setting for the physical pages cannot be configured
2301  *	to VM_MEMATTR_DEFAULT.
2302  *
2303  *	The specified object may not contain fictitious pages.
2304  *
2305  *	The caller must always specify an allocation class.
2306  *
2307  *	allocation classes:
2308  *	VM_ALLOC_NORMAL		normal process request
2309  *	VM_ALLOC_SYSTEM		system *really* needs a page
2310  *	VM_ALLOC_INTERRUPT	interrupt time request
2311  *
2312  *	optional allocation flags:
2313  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2314  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2315  *	VM_ALLOC_SBUSY		shared busy the allocated page
2316  *	VM_ALLOC_WIRED		wire the allocated page
2317  *	VM_ALLOC_ZERO		prefer a zeroed page
2318  */
2319 vm_page_t
2320 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2321     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2322     vm_paddr_t boundary, vm_memattr_t memattr)
2323 {
2324 	struct vm_domainset_iter di;
2325 	vm_page_t bounds[2];
2326 	vm_page_t m;
2327 	int domain;
2328 	int start_segind;
2329 
2330 	start_segind = -1;
2331 
2332 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2333 	do {
2334 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2335 		    npages, low, high, alignment, boundary, memattr);
2336 		if (m != NULL)
2337 			break;
2338 		if (start_segind == -1)
2339 			start_segind = vm_phys_lookup_segind(low);
2340 		if (vm_phys_find_range(bounds, start_segind, domain,
2341 		    npages, low, high) == -1) {
2342 			vm_domainset_iter_ignore(&di, domain);
2343 		}
2344 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2345 
2346 	return (m);
2347 }
2348 
2349 static vm_page_t
2350 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2351     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2352 {
2353 	struct vm_domain *vmd;
2354 	vm_page_t m_ret;
2355 
2356 	/*
2357 	 * Can we allocate the pages without the number of free pages falling
2358 	 * below the lower bound for the allocation class?
2359 	 */
2360 	vmd = VM_DOMAIN(domain);
2361 	if (!vm_domain_allocate(vmd, req, npages))
2362 		return (NULL);
2363 	/*
2364 	 * Try to allocate the pages from the free page queues.
2365 	 */
2366 	vm_domain_free_lock(vmd);
2367 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2368 	    alignment, boundary);
2369 	vm_domain_free_unlock(vmd);
2370 	if (m_ret != NULL)
2371 		return (m_ret);
2372 #if VM_NRESERVLEVEL > 0
2373 	/*
2374 	 * Try to break a reservation to allocate the pages.
2375 	 */
2376 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2377 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2378 	            high, alignment, boundary);
2379 		if (m_ret != NULL)
2380 			return (m_ret);
2381 	}
2382 #endif
2383 	vm_domain_freecnt_inc(vmd, npages);
2384 	return (NULL);
2385 }
2386 
2387 vm_page_t
2388 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2389     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2390     vm_paddr_t boundary, vm_memattr_t memattr)
2391 {
2392 	struct pctrie_iter pages;
2393 	vm_page_t m, m_ret, mpred;
2394 	u_int busy_lock, flags, oflags;
2395 
2396 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2397 	KASSERT((req & ~VPAC_FLAGS) == 0,
2398 	    ("invalid request %#x", req));
2399 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2400 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2401 	    ("invalid request %#x", req));
2402 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2403 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2404 	    ("invalid request %#x", req));
2405 	VM_OBJECT_ASSERT_WLOCKED(object);
2406 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2407 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2408 	    object));
2409 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2410 
2411 	vm_page_iter_init(&pages, object);
2412 	mpred = vm_radix_iter_lookup_lt(&pages, pindex);
2413 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2414 	    ("vm_page_alloc_contig: pindex already allocated"));
2415 	for (;;) {
2416 #if VM_NRESERVLEVEL > 0
2417 		/*
2418 		 * Can we allocate the pages from a reservation?
2419 		 */
2420 		if (vm_object_reserv(object) &&
2421 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2422 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2423 			break;
2424 		}
2425 #endif
2426 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2427 		    low, high, alignment, boundary)) != NULL)
2428 			break;
2429 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2430 			return (NULL);
2431 	}
2432 
2433 	/*
2434 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2435 	 */
2436 	flags = PG_ZERO;
2437 	if ((req & VM_ALLOC_NODUMP) != 0)
2438 		flags |= PG_NODUMP;
2439 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2440 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2441 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2442 	else if ((req & VM_ALLOC_SBUSY) != 0)
2443 		busy_lock = VPB_SHARERS_WORD(1);
2444 	else
2445 		busy_lock = VPB_UNBUSIED;
2446 	if ((req & VM_ALLOC_WIRED) != 0)
2447 		vm_wire_add(npages);
2448 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2449 	    memattr == VM_MEMATTR_DEFAULT)
2450 		memattr = object->memattr;
2451 	for (m = m_ret; m < &m_ret[npages]; m++) {
2452 		vm_page_dequeue(m);
2453 		vm_page_alloc_check(m);
2454 		m->a.flags = 0;
2455 		m->flags = (m->flags | PG_NODUMP) & flags;
2456 		m->busy_lock = busy_lock;
2457 		if ((req & VM_ALLOC_WIRED) != 0)
2458 			m->ref_count = 1;
2459 		m->a.act_count = 0;
2460 		m->oflags = oflags;
2461 		m->pool = VM_FREEPOOL_DEFAULT;
2462 		if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) {
2463 			if ((req & VM_ALLOC_WIRED) != 0)
2464 				vm_wire_sub(npages);
2465 			KASSERT(m->object == NULL,
2466 			    ("page %p has object", m));
2467 			mpred = m;
2468 			for (m = m_ret; m < &m_ret[npages]; m++) {
2469 				if (m <= mpred &&
2470 				    (req & VM_ALLOC_WIRED) != 0)
2471 					m->ref_count = 0;
2472 				m->oflags = VPO_UNMANAGED;
2473 				m->busy_lock = VPB_UNBUSIED;
2474 				/* Don't change PG_ZERO. */
2475 				vm_page_free_toq(m);
2476 			}
2477 			if (req & VM_ALLOC_WAITFAIL) {
2478 				VM_OBJECT_WUNLOCK(object);
2479 				vm_radix_wait();
2480 				VM_OBJECT_WLOCK(object);
2481 			}
2482 			return (NULL);
2483 		}
2484 		mpred = m;
2485 		if (memattr != VM_MEMATTR_DEFAULT)
2486 			pmap_page_set_memattr(m, memattr);
2487 		pindex++;
2488 	}
2489 	return (m_ret);
2490 }
2491 
2492 /*
2493  * Allocate a physical page that is not intended to be inserted into a VM
2494  * object.
2495  */
2496 vm_page_t
2497 vm_page_alloc_noobj_domain(int domain, int req)
2498 {
2499 	struct vm_domain *vmd;
2500 	vm_page_t m;
2501 	int flags;
2502 
2503 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2504 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2505 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2506 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2507 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2508 	KASSERT((req & ~VPAN_FLAGS) == 0,
2509 	    ("invalid request %#x", req));
2510 
2511 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2512 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2513 	vmd = VM_DOMAIN(domain);
2514 again:
2515 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2516 		m = vm_page_alloc_nofree_domain(domain, req);
2517 		if (m != NULL)
2518 			goto found;
2519 	}
2520 
2521 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2522 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2523 		    M_NOWAIT | M_NOVM);
2524 		if (m != NULL) {
2525 			flags |= PG_PCPU_CACHE;
2526 			goto found;
2527 		}
2528 	}
2529 
2530 	if (vm_domain_allocate(vmd, req, 1)) {
2531 		vm_domain_free_lock(vmd);
2532 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2533 		vm_domain_free_unlock(vmd);
2534 		if (m == NULL) {
2535 			vm_domain_freecnt_inc(vmd, 1);
2536 #if VM_NRESERVLEVEL > 0
2537 			if (vm_reserv_reclaim_inactive(domain))
2538 				goto again;
2539 #endif
2540 		}
2541 	}
2542 	if (m == NULL) {
2543 		if (vm_domain_alloc_fail(vmd, NULL, req))
2544 			goto again;
2545 		return (NULL);
2546 	}
2547 
2548 found:
2549 	vm_page_dequeue(m);
2550 	vm_page_alloc_check(m);
2551 
2552 	/*
2553 	 * Consumers should not rely on a useful default pindex value.
2554 	 */
2555 	m->pindex = 0xdeadc0dedeadc0de;
2556 	m->flags = (m->flags & PG_ZERO) | flags;
2557 	m->a.flags = 0;
2558 	m->oflags = VPO_UNMANAGED;
2559 	m->pool = VM_FREEPOOL_DIRECT;
2560 	m->busy_lock = VPB_UNBUSIED;
2561 	if ((req & VM_ALLOC_WIRED) != 0) {
2562 		vm_wire_add(1);
2563 		m->ref_count = 1;
2564 	}
2565 
2566 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2567 		pmap_zero_page(m);
2568 
2569 	return (m);
2570 }
2571 
2572 #if VM_NRESERVLEVEL > 1
2573 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2574 #elif VM_NRESERVLEVEL > 0
2575 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2576 #else
2577 #define	VM_NOFREE_IMPORT_ORDER	8
2578 #endif
2579 
2580 /*
2581  * Allocate a single NOFREE page.
2582  *
2583  * This routine hands out NOFREE pages from higher-order
2584  * physical memory blocks in order to reduce memory fragmentation.
2585  * When a NOFREE for a given domain chunk is used up,
2586  * the routine will try to fetch a new one from the freelists
2587  * and discard the old one.
2588  */
2589 static vm_page_t __noinline
2590 vm_page_alloc_nofree_domain(int domain, int req)
2591 {
2592 	vm_page_t m;
2593 	struct vm_domain *vmd;
2594 
2595 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2596 
2597 	vmd = VM_DOMAIN(domain);
2598 	vm_domain_free_lock(vmd);
2599 	if (TAILQ_EMPTY(&vmd->vmd_nofreeq)) {
2600 		int count;
2601 
2602 		count = 1 << VM_NOFREE_IMPORT_ORDER;
2603 		if (!vm_domain_allocate(vmd, req, count)) {
2604 			vm_domain_free_unlock(vmd);
2605 			return (NULL);
2606 		}
2607 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2608 		    VM_NOFREE_IMPORT_ORDER);
2609 		if (m == NULL) {
2610 			vm_domain_freecnt_inc(vmd, count);
2611 			vm_domain_free_unlock(vmd);
2612 			return (NULL);
2613 		}
2614 		m->pindex = count;
2615 		TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, listq);
2616 		VM_CNT_ADD(v_nofree_count, count);
2617 	}
2618 	m = TAILQ_FIRST(&vmd->vmd_nofreeq);
2619 	TAILQ_REMOVE(&vmd->vmd_nofreeq, m, listq);
2620 	if (m->pindex > 1) {
2621 		vm_page_t m_next;
2622 
2623 		m_next = &m[1];
2624 		m_next->pindex = m->pindex - 1;
2625 		TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m_next, listq);
2626 	}
2627 	vm_domain_free_unlock(vmd);
2628 	VM_CNT_ADD(v_nofree_count, -1);
2629 
2630 	return (m);
2631 }
2632 
2633 /*
2634  * Though a NOFREE page by definition should not be freed, we support putting
2635  * them aside for future NOFREE allocations.  This enables code which allocates
2636  * NOFREE pages for some purpose but then encounters an error and releases
2637  * resources.
2638  */
2639 static void __noinline
2640 vm_page_free_nofree(struct vm_domain *vmd, vm_page_t m)
2641 {
2642 	vm_domain_free_lock(vmd);
2643 	m->pindex = 1;
2644 	TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, listq);
2645 	vm_domain_free_unlock(vmd);
2646 	VM_CNT_ADD(v_nofree_count, 1);
2647 }
2648 
2649 vm_page_t
2650 vm_page_alloc_noobj(int req)
2651 {
2652 	struct vm_domainset_iter di;
2653 	vm_page_t m;
2654 	int domain;
2655 
2656 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2657 	do {
2658 		m = vm_page_alloc_noobj_domain(domain, req);
2659 		if (m != NULL)
2660 			break;
2661 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2662 
2663 	return (m);
2664 }
2665 
2666 vm_page_t
2667 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2668     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2669     vm_memattr_t memattr)
2670 {
2671 	struct vm_domainset_iter di;
2672 	vm_page_t m;
2673 	int domain;
2674 
2675 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2676 	do {
2677 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2678 		    high, alignment, boundary, memattr);
2679 		if (m != NULL)
2680 			break;
2681 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2682 
2683 	return (m);
2684 }
2685 
2686 vm_page_t
2687 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2688     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2689     vm_memattr_t memattr)
2690 {
2691 	vm_page_t m, m_ret;
2692 	u_int flags;
2693 
2694 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2695 	KASSERT((req & ~VPANC_FLAGS) == 0,
2696 	    ("invalid request %#x", req));
2697 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2698 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2699 	    ("invalid request %#x", req));
2700 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2701 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2702 	    ("invalid request %#x", req));
2703 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2704 
2705 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2706 	    low, high, alignment, boundary)) == NULL) {
2707 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2708 			return (NULL);
2709 	}
2710 
2711 	/*
2712 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2713 	 */
2714 	flags = PG_ZERO;
2715 	if ((req & VM_ALLOC_NODUMP) != 0)
2716 		flags |= PG_NODUMP;
2717 	if ((req & VM_ALLOC_WIRED) != 0)
2718 		vm_wire_add(npages);
2719 	for (m = m_ret; m < &m_ret[npages]; m++) {
2720 		vm_page_dequeue(m);
2721 		vm_page_alloc_check(m);
2722 
2723 		/*
2724 		 * Consumers should not rely on a useful default pindex value.
2725 		 */
2726 		m->pindex = 0xdeadc0dedeadc0de;
2727 		m->a.flags = 0;
2728 		m->flags = (m->flags | PG_NODUMP) & flags;
2729 		m->busy_lock = VPB_UNBUSIED;
2730 		if ((req & VM_ALLOC_WIRED) != 0)
2731 			m->ref_count = 1;
2732 		m->a.act_count = 0;
2733 		m->oflags = VPO_UNMANAGED;
2734 		m->pool = VM_FREEPOOL_DIRECT;
2735 
2736 		/*
2737 		 * Zero the page before updating any mappings since the page is
2738 		 * not yet shared with any devices which might require the
2739 		 * non-default memory attribute.  pmap_page_set_memattr()
2740 		 * flushes data caches before returning.
2741 		 */
2742 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2743 			pmap_zero_page(m);
2744 		if (memattr != VM_MEMATTR_DEFAULT)
2745 			pmap_page_set_memattr(m, memattr);
2746 	}
2747 	return (m_ret);
2748 }
2749 
2750 /*
2751  * Check a page that has been freshly dequeued from a freelist.
2752  */
2753 static void
2754 vm_page_alloc_check(vm_page_t m)
2755 {
2756 
2757 	KASSERT(m->object == NULL, ("page %p has object", m));
2758 	KASSERT(m->a.queue == PQ_NONE &&
2759 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2760 	    ("page %p has unexpected queue %d, flags %#x",
2761 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2762 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2763 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2764 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2765 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2766 	    ("page %p has unexpected memattr %d",
2767 	    m, pmap_page_get_memattr(m)));
2768 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2769 	pmap_vm_page_alloc_check(m);
2770 }
2771 
2772 static int
2773 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2774 {
2775 	struct vm_domain *vmd;
2776 	struct vm_pgcache *pgcache;
2777 	int i;
2778 
2779 	pgcache = arg;
2780 	vmd = VM_DOMAIN(pgcache->domain);
2781 
2782 	/*
2783 	 * The page daemon should avoid creating extra memory pressure since its
2784 	 * main purpose is to replenish the store of free pages.
2785 	 */
2786 	if (vmd->vmd_severeset || curproc == pageproc ||
2787 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2788 		return (0);
2789 	domain = vmd->vmd_domain;
2790 	vm_domain_free_lock(vmd);
2791 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2792 	    (vm_page_t *)store);
2793 	vm_domain_free_unlock(vmd);
2794 	if (cnt != i)
2795 		vm_domain_freecnt_inc(vmd, cnt - i);
2796 
2797 	return (i);
2798 }
2799 
2800 static void
2801 vm_page_zone_release(void *arg, void **store, int cnt)
2802 {
2803 	struct vm_domain *vmd;
2804 	struct vm_pgcache *pgcache;
2805 	vm_page_t m;
2806 	int i;
2807 
2808 	pgcache = arg;
2809 	vmd = VM_DOMAIN(pgcache->domain);
2810 	vm_domain_free_lock(vmd);
2811 	for (i = 0; i < cnt; i++) {
2812 		m = (vm_page_t)store[i];
2813 		vm_phys_free_pages(m, pgcache->pool, 0);
2814 	}
2815 	vm_domain_free_unlock(vmd);
2816 	vm_domain_freecnt_inc(vmd, cnt);
2817 }
2818 
2819 #define	VPSC_ANY	0	/* No restrictions. */
2820 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2821 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2822 
2823 /*
2824  *	vm_page_scan_contig:
2825  *
2826  *	Scan vm_page_array[] between the specified entries "m_start" and
2827  *	"m_end" for a run of contiguous physical pages that satisfy the
2828  *	specified conditions, and return the lowest page in the run.  The
2829  *	specified "alignment" determines the alignment of the lowest physical
2830  *	page in the run.  If the specified "boundary" is non-zero, then the
2831  *	run of physical pages cannot span a physical address that is a
2832  *	multiple of "boundary".
2833  *
2834  *	"m_end" is never dereferenced, so it need not point to a vm_page
2835  *	structure within vm_page_array[].
2836  *
2837  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2838  *	span a hole (or discontiguity) in the physical address space.  Both
2839  *	"alignment" and "boundary" must be a power of two.
2840  */
2841 static vm_page_t
2842 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2843     u_long alignment, vm_paddr_t boundary, int options)
2844 {
2845 	vm_object_t object;
2846 	vm_paddr_t pa;
2847 	vm_page_t m, m_run;
2848 #if VM_NRESERVLEVEL > 0
2849 	int level;
2850 #endif
2851 	int m_inc, order, run_ext, run_len;
2852 
2853 	KASSERT(npages > 0, ("npages is 0"));
2854 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2855 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2856 	m_run = NULL;
2857 	run_len = 0;
2858 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2859 		KASSERT((m->flags & PG_MARKER) == 0,
2860 		    ("page %p is PG_MARKER", m));
2861 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2862 		    ("fictitious page %p has invalid ref count", m));
2863 
2864 		/*
2865 		 * If the current page would be the start of a run, check its
2866 		 * physical address against the end, alignment, and boundary
2867 		 * conditions.  If it doesn't satisfy these conditions, either
2868 		 * terminate the scan or advance to the next page that
2869 		 * satisfies the failed condition.
2870 		 */
2871 		if (run_len == 0) {
2872 			KASSERT(m_run == NULL, ("m_run != NULL"));
2873 			if (m + npages > m_end)
2874 				break;
2875 			pa = VM_PAGE_TO_PHYS(m);
2876 			if (!vm_addr_align_ok(pa, alignment)) {
2877 				m_inc = atop(roundup2(pa, alignment) - pa);
2878 				continue;
2879 			}
2880 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2881 				m_inc = atop(roundup2(pa, boundary) - pa);
2882 				continue;
2883 			}
2884 		} else
2885 			KASSERT(m_run != NULL, ("m_run == NULL"));
2886 
2887 retry:
2888 		m_inc = 1;
2889 		if (vm_page_wired(m))
2890 			run_ext = 0;
2891 #if VM_NRESERVLEVEL > 0
2892 		else if ((level = vm_reserv_level(m)) >= 0 &&
2893 		    (options & VPSC_NORESERV) != 0) {
2894 			run_ext = 0;
2895 			/* Advance to the end of the reservation. */
2896 			pa = VM_PAGE_TO_PHYS(m);
2897 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2898 			    pa);
2899 		}
2900 #endif
2901 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2902 			/*
2903 			 * The page is considered eligible for relocation if
2904 			 * and only if it could be laundered or reclaimed by
2905 			 * the page daemon.
2906 			 */
2907 			VM_OBJECT_RLOCK(object);
2908 			if (object != m->object) {
2909 				VM_OBJECT_RUNLOCK(object);
2910 				goto retry;
2911 			}
2912 			/* Don't care: PG_NODUMP, PG_ZERO. */
2913 			if ((object->flags & OBJ_SWAP) == 0 &&
2914 			    object->type != OBJT_VNODE) {
2915 				run_ext = 0;
2916 #if VM_NRESERVLEVEL > 0
2917 			} else if ((options & VPSC_NOSUPER) != 0 &&
2918 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2919 				run_ext = 0;
2920 				/* Advance to the end of the superpage. */
2921 				pa = VM_PAGE_TO_PHYS(m);
2922 				m_inc = atop(roundup2(pa + 1,
2923 				    vm_reserv_size(level)) - pa);
2924 #endif
2925 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2926 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2927 				/*
2928 				 * The page is allocated but eligible for
2929 				 * relocation.  Extend the current run by one
2930 				 * page.
2931 				 */
2932 				KASSERT(pmap_page_get_memattr(m) ==
2933 				    VM_MEMATTR_DEFAULT,
2934 				    ("page %p has an unexpected memattr", m));
2935 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2936 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2937 				    ("page %p has unexpected oflags", m));
2938 				/* Don't care: PGA_NOSYNC. */
2939 				run_ext = 1;
2940 			} else
2941 				run_ext = 0;
2942 			VM_OBJECT_RUNLOCK(object);
2943 #if VM_NRESERVLEVEL > 0
2944 		} else if (level >= 0) {
2945 			/*
2946 			 * The page is reserved but not yet allocated.  In
2947 			 * other words, it is still free.  Extend the current
2948 			 * run by one page.
2949 			 */
2950 			run_ext = 1;
2951 #endif
2952 		} else if ((order = m->order) < VM_NFREEORDER) {
2953 			/*
2954 			 * The page is enqueued in the physical memory
2955 			 * allocator's free page queues.  Moreover, it is the
2956 			 * first page in a power-of-two-sized run of
2957 			 * contiguous free pages.  Add these pages to the end
2958 			 * of the current run, and jump ahead.
2959 			 */
2960 			run_ext = 1 << order;
2961 			m_inc = 1 << order;
2962 		} else {
2963 			/*
2964 			 * Skip the page for one of the following reasons: (1)
2965 			 * It is enqueued in the physical memory allocator's
2966 			 * free page queues.  However, it is not the first
2967 			 * page in a run of contiguous free pages.  (This case
2968 			 * rarely occurs because the scan is performed in
2969 			 * ascending order.) (2) It is not reserved, and it is
2970 			 * transitioning from free to allocated.  (Conversely,
2971 			 * the transition from allocated to free for managed
2972 			 * pages is blocked by the page busy lock.) (3) It is
2973 			 * allocated but not contained by an object and not
2974 			 * wired, e.g., allocated by Xen's balloon driver.
2975 			 */
2976 			run_ext = 0;
2977 		}
2978 
2979 		/*
2980 		 * Extend or reset the current run of pages.
2981 		 */
2982 		if (run_ext > 0) {
2983 			if (run_len == 0)
2984 				m_run = m;
2985 			run_len += run_ext;
2986 		} else {
2987 			if (run_len > 0) {
2988 				m_run = NULL;
2989 				run_len = 0;
2990 			}
2991 		}
2992 	}
2993 	if (run_len >= npages)
2994 		return (m_run);
2995 	return (NULL);
2996 }
2997 
2998 /*
2999  *	vm_page_reclaim_run:
3000  *
3001  *	Try to relocate each of the allocated virtual pages within the
3002  *	specified run of physical pages to a new physical address.  Free the
3003  *	physical pages underlying the relocated virtual pages.  A virtual page
3004  *	is relocatable if and only if it could be laundered or reclaimed by
3005  *	the page daemon.  Whenever possible, a virtual page is relocated to a
3006  *	physical address above "high".
3007  *
3008  *	Returns 0 if every physical page within the run was already free or
3009  *	just freed by a successful relocation.  Otherwise, returns a non-zero
3010  *	value indicating why the last attempt to relocate a virtual page was
3011  *	unsuccessful.
3012  *
3013  *	"req_class" must be an allocation class.
3014  */
3015 static int
3016 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
3017     vm_paddr_t high)
3018 {
3019 	struct vm_domain *vmd;
3020 	struct spglist free;
3021 	vm_object_t object;
3022 	vm_paddr_t pa;
3023 	vm_page_t m, m_end, m_new;
3024 	int error, order, req;
3025 
3026 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
3027 	    ("req_class is not an allocation class"));
3028 	SLIST_INIT(&free);
3029 	error = 0;
3030 	m = m_run;
3031 	m_end = m_run + npages;
3032 	for (; error == 0 && m < m_end; m++) {
3033 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
3034 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
3035 
3036 		/*
3037 		 * Racily check for wirings.  Races are handled once the object
3038 		 * lock is held and the page is unmapped.
3039 		 */
3040 		if (vm_page_wired(m))
3041 			error = EBUSY;
3042 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
3043 			/*
3044 			 * The page is relocated if and only if it could be
3045 			 * laundered or reclaimed by the page daemon.
3046 			 */
3047 			VM_OBJECT_WLOCK(object);
3048 			/* Don't care: PG_NODUMP, PG_ZERO. */
3049 			if (m->object != object ||
3050 			    ((object->flags & OBJ_SWAP) == 0 &&
3051 			    object->type != OBJT_VNODE))
3052 				error = EINVAL;
3053 			else if (object->memattr != VM_MEMATTR_DEFAULT)
3054 				error = EINVAL;
3055 			else if (vm_page_queue(m) != PQ_NONE &&
3056 			    vm_page_tryxbusy(m) != 0) {
3057 				if (vm_page_wired(m)) {
3058 					vm_page_xunbusy(m);
3059 					error = EBUSY;
3060 					goto unlock;
3061 				}
3062 				KASSERT(pmap_page_get_memattr(m) ==
3063 				    VM_MEMATTR_DEFAULT,
3064 				    ("page %p has an unexpected memattr", m));
3065 				KASSERT(m->oflags == 0,
3066 				    ("page %p has unexpected oflags", m));
3067 				/* Don't care: PGA_NOSYNC. */
3068 				if (!vm_page_none_valid(m)) {
3069 					/*
3070 					 * First, try to allocate a new page
3071 					 * that is above "high".  Failing
3072 					 * that, try to allocate a new page
3073 					 * that is below "m_run".  Allocate
3074 					 * the new page between the end of
3075 					 * "m_run" and "high" only as a last
3076 					 * resort.
3077 					 */
3078 					req = req_class;
3079 					if ((m->flags & PG_NODUMP) != 0)
3080 						req |= VM_ALLOC_NODUMP;
3081 					if (trunc_page(high) !=
3082 					    ~(vm_paddr_t)PAGE_MASK) {
3083 						m_new =
3084 						    vm_page_alloc_noobj_contig(
3085 						    req, 1, round_page(high),
3086 						    ~(vm_paddr_t)0, PAGE_SIZE,
3087 						    0, VM_MEMATTR_DEFAULT);
3088 					} else
3089 						m_new = NULL;
3090 					if (m_new == NULL) {
3091 						pa = VM_PAGE_TO_PHYS(m_run);
3092 						m_new =
3093 						    vm_page_alloc_noobj_contig(
3094 						    req, 1, 0, pa - 1,
3095 						    PAGE_SIZE, 0,
3096 						    VM_MEMATTR_DEFAULT);
3097 					}
3098 					if (m_new == NULL) {
3099 						pa += ptoa(npages);
3100 						m_new =
3101 						    vm_page_alloc_noobj_contig(
3102 						    req, 1, pa, high, PAGE_SIZE,
3103 						    0, VM_MEMATTR_DEFAULT);
3104 					}
3105 					if (m_new == NULL) {
3106 						vm_page_xunbusy(m);
3107 						error = ENOMEM;
3108 						goto unlock;
3109 					}
3110 
3111 					/*
3112 					 * Unmap the page and check for new
3113 					 * wirings that may have been acquired
3114 					 * through a pmap lookup.
3115 					 */
3116 					if (object->ref_count != 0 &&
3117 					    !vm_page_try_remove_all(m)) {
3118 						vm_page_xunbusy(m);
3119 						vm_page_free(m_new);
3120 						error = EBUSY;
3121 						goto unlock;
3122 					}
3123 
3124 					/*
3125 					 * Replace "m" with the new page.  For
3126 					 * vm_page_replace(), "m" must be busy
3127 					 * and dequeued.  Finally, change "m"
3128 					 * as if vm_page_free() was called.
3129 					 */
3130 					m_new->a.flags = m->a.flags &
3131 					    ~PGA_QUEUE_STATE_MASK;
3132 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3133 					    ("page %p is managed", m_new));
3134 					m_new->oflags = 0;
3135 					pmap_copy_page(m, m_new);
3136 					m_new->valid = m->valid;
3137 					m_new->dirty = m->dirty;
3138 					m->flags &= ~PG_ZERO;
3139 					vm_page_dequeue(m);
3140 					if (vm_page_replace_hold(m_new, object,
3141 					    m->pindex, m) &&
3142 					    vm_page_free_prep(m))
3143 						SLIST_INSERT_HEAD(&free, m,
3144 						    plinks.s.ss);
3145 
3146 					/*
3147 					 * The new page must be deactivated
3148 					 * before the object is unlocked.
3149 					 */
3150 					vm_page_deactivate(m_new);
3151 				} else {
3152 					m->flags &= ~PG_ZERO;
3153 					vm_page_dequeue(m);
3154 					if (vm_page_free_prep(m))
3155 						SLIST_INSERT_HEAD(&free, m,
3156 						    plinks.s.ss);
3157 					KASSERT(m->dirty == 0,
3158 					    ("page %p is dirty", m));
3159 				}
3160 			} else
3161 				error = EBUSY;
3162 unlock:
3163 			VM_OBJECT_WUNLOCK(object);
3164 		} else {
3165 			MPASS(vm_page_domain(m) == domain);
3166 			vmd = VM_DOMAIN(domain);
3167 			vm_domain_free_lock(vmd);
3168 			order = m->order;
3169 			if (order < VM_NFREEORDER) {
3170 				/*
3171 				 * The page is enqueued in the physical memory
3172 				 * allocator's free page queues.  Moreover, it
3173 				 * is the first page in a power-of-two-sized
3174 				 * run of contiguous free pages.  Jump ahead
3175 				 * to the last page within that run, and
3176 				 * continue from there.
3177 				 */
3178 				m += (1 << order) - 1;
3179 			}
3180 #if VM_NRESERVLEVEL > 0
3181 			else if (vm_reserv_is_page_free(m))
3182 				order = 0;
3183 #endif
3184 			vm_domain_free_unlock(vmd);
3185 			if (order == VM_NFREEORDER)
3186 				error = EINVAL;
3187 		}
3188 	}
3189 	if ((m = SLIST_FIRST(&free)) != NULL) {
3190 		int cnt;
3191 
3192 		vmd = VM_DOMAIN(domain);
3193 		cnt = 0;
3194 		vm_domain_free_lock(vmd);
3195 		do {
3196 			MPASS(vm_page_domain(m) == domain);
3197 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3198 			vm_phys_free_pages(m, m->pool, 0);
3199 			cnt++;
3200 		} while ((m = SLIST_FIRST(&free)) != NULL);
3201 		vm_domain_free_unlock(vmd);
3202 		vm_domain_freecnt_inc(vmd, cnt);
3203 	}
3204 	return (error);
3205 }
3206 
3207 #define	NRUNS	16
3208 
3209 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3210 
3211 #define	MIN_RECLAIM	8
3212 
3213 /*
3214  *	vm_page_reclaim_contig:
3215  *
3216  *	Reclaim allocated, contiguous physical memory satisfying the specified
3217  *	conditions by relocating the virtual pages using that physical memory.
3218  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3219  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3220  *	currently able to reclaim the requested number of pages.  Since
3221  *	relocation requires the allocation of physical pages, reclamation may
3222  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3223  *	fails in this manner, callers are expected to perform vm_wait() before
3224  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3225  *
3226  *	The caller must always specify an allocation class through "req".
3227  *
3228  *	allocation classes:
3229  *	VM_ALLOC_NORMAL		normal process request
3230  *	VM_ALLOC_SYSTEM		system *really* needs a page
3231  *	VM_ALLOC_INTERRUPT	interrupt time request
3232  *
3233  *	The optional allocation flags are ignored.
3234  *
3235  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3236  *	must be a power of two.
3237  */
3238 int
3239 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3240     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3241     int desired_runs)
3242 {
3243 	struct vm_domain *vmd;
3244 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3245 	u_long count, minalign, reclaimed;
3246 	int error, i, min_reclaim, nruns, options, req_class;
3247 	int segind, start_segind;
3248 	int ret;
3249 
3250 	KASSERT(npages > 0, ("npages is 0"));
3251 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3252 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3253 
3254 	ret = ENOMEM;
3255 
3256 	/*
3257 	 * If the caller wants to reclaim multiple runs, try to allocate
3258 	 * space to store the runs.  If that fails, fall back to the old
3259 	 * behavior of just reclaiming MIN_RECLAIM pages.
3260 	 */
3261 	if (desired_runs > 1)
3262 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3263 		    M_TEMP, M_NOWAIT);
3264 	else
3265 		m_runs = NULL;
3266 
3267 	if (m_runs == NULL) {
3268 		m_runs = _m_runs;
3269 		nruns = NRUNS;
3270 	} else {
3271 		nruns = NRUNS + desired_runs - 1;
3272 	}
3273 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3274 
3275 	/*
3276 	 * The caller will attempt an allocation after some runs have been
3277 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3278 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3279 	 * power of two or maximum chunk size, and ensure that each run is
3280 	 * suitably aligned.
3281 	 */
3282 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3283 	npages = roundup2(npages, minalign);
3284 	if (alignment < ptoa(minalign))
3285 		alignment = ptoa(minalign);
3286 
3287 	/*
3288 	 * The page daemon is allowed to dig deeper into the free page list.
3289 	 */
3290 	req_class = req & VM_ALLOC_CLASS_MASK;
3291 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3292 		req_class = VM_ALLOC_SYSTEM;
3293 
3294 	start_segind = vm_phys_lookup_segind(low);
3295 
3296 	/*
3297 	 * Return if the number of free pages cannot satisfy the requested
3298 	 * allocation.
3299 	 */
3300 	vmd = VM_DOMAIN(domain);
3301 	count = vmd->vmd_free_count;
3302 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3303 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3304 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3305 		goto done;
3306 
3307 	/*
3308 	 * Scan up to three times, relaxing the restrictions ("options") on
3309 	 * the reclamation of reservations and superpages each time.
3310 	 */
3311 	for (options = VPSC_NORESERV;;) {
3312 		bool phys_range_exists = false;
3313 
3314 		/*
3315 		 * Find the highest runs that satisfy the given constraints
3316 		 * and restrictions, and record them in "m_runs".
3317 		 */
3318 		count = 0;
3319 		segind = start_segind;
3320 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3321 		    npages, low, high)) != -1) {
3322 			phys_range_exists = true;
3323 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3324 			    bounds[1], alignment, boundary, options))) {
3325 				bounds[0] = m_run + npages;
3326 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3327 				count++;
3328 			}
3329 			segind++;
3330 		}
3331 
3332 		if (!phys_range_exists) {
3333 			ret = ERANGE;
3334 			goto done;
3335 		}
3336 
3337 		/*
3338 		 * Reclaim the highest runs in LIFO (descending) order until
3339 		 * the number of reclaimed pages, "reclaimed", is at least
3340 		 * "min_reclaim".  Reset "reclaimed" each time because each
3341 		 * reclamation is idempotent, and runs will (likely) recur
3342 		 * from one scan to the next as restrictions are relaxed.
3343 		 */
3344 		reclaimed = 0;
3345 		for (i = 0; count > 0 && i < nruns; i++) {
3346 			count--;
3347 			m_run = m_runs[RUN_INDEX(count, nruns)];
3348 			error = vm_page_reclaim_run(req_class, domain, npages,
3349 			    m_run, high);
3350 			if (error == 0) {
3351 				reclaimed += npages;
3352 				if (reclaimed >= min_reclaim) {
3353 					ret = 0;
3354 					goto done;
3355 				}
3356 			}
3357 		}
3358 
3359 		/*
3360 		 * Either relax the restrictions on the next scan or return if
3361 		 * the last scan had no restrictions.
3362 		 */
3363 		if (options == VPSC_NORESERV)
3364 			options = VPSC_NOSUPER;
3365 		else if (options == VPSC_NOSUPER)
3366 			options = VPSC_ANY;
3367 		else if (options == VPSC_ANY) {
3368 			if (reclaimed != 0)
3369 				ret = 0;
3370 			goto done;
3371 		}
3372 	}
3373 done:
3374 	if (m_runs != _m_runs)
3375 		free(m_runs, M_TEMP);
3376 	return (ret);
3377 }
3378 
3379 int
3380 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3381     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3382 {
3383 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low,
3384 	    high, alignment, boundary, 1));
3385 }
3386 
3387 int
3388 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3389     u_long alignment, vm_paddr_t boundary)
3390 {
3391 	struct vm_domainset_iter di;
3392 	int domain, ret, status;
3393 
3394 	ret = ERANGE;
3395 
3396 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3397 	do {
3398 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3399 		    high, alignment, boundary);
3400 		if (status == 0)
3401 			return (0);
3402 		else if (status == ERANGE)
3403 			vm_domainset_iter_ignore(&di, domain);
3404 		else {
3405 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3406 			    "from vm_page_reclaim_contig_domain()", status));
3407 			ret = ENOMEM;
3408 		}
3409 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3410 
3411 	return (ret);
3412 }
3413 
3414 /*
3415  * Set the domain in the appropriate page level domainset.
3416  */
3417 void
3418 vm_domain_set(struct vm_domain *vmd)
3419 {
3420 
3421 	mtx_lock(&vm_domainset_lock);
3422 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3423 		vmd->vmd_minset = 1;
3424 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3425 	}
3426 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3427 		vmd->vmd_severeset = 1;
3428 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3429 	}
3430 	mtx_unlock(&vm_domainset_lock);
3431 }
3432 
3433 /*
3434  * Clear the domain from the appropriate page level domainset.
3435  */
3436 void
3437 vm_domain_clear(struct vm_domain *vmd)
3438 {
3439 
3440 	mtx_lock(&vm_domainset_lock);
3441 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3442 		vmd->vmd_minset = 0;
3443 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3444 		if (vm_min_waiters != 0) {
3445 			vm_min_waiters = 0;
3446 			wakeup(&vm_min_domains);
3447 		}
3448 	}
3449 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3450 		vmd->vmd_severeset = 0;
3451 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3452 		if (vm_severe_waiters != 0) {
3453 			vm_severe_waiters = 0;
3454 			wakeup(&vm_severe_domains);
3455 		}
3456 	}
3457 
3458 	/*
3459 	 * If pageout daemon needs pages, then tell it that there are
3460 	 * some free.
3461 	 */
3462 	if (vmd->vmd_pageout_pages_needed &&
3463 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3464 		wakeup(&vmd->vmd_pageout_pages_needed);
3465 		vmd->vmd_pageout_pages_needed = 0;
3466 	}
3467 
3468 	/* See comments in vm_wait_doms(). */
3469 	if (vm_pageproc_waiters) {
3470 		vm_pageproc_waiters = 0;
3471 		wakeup(&vm_pageproc_waiters);
3472 	}
3473 	mtx_unlock(&vm_domainset_lock);
3474 }
3475 
3476 /*
3477  * Wait for free pages to exceed the min threshold globally.
3478  */
3479 void
3480 vm_wait_min(void)
3481 {
3482 
3483 	mtx_lock(&vm_domainset_lock);
3484 	while (vm_page_count_min()) {
3485 		vm_min_waiters++;
3486 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3487 	}
3488 	mtx_unlock(&vm_domainset_lock);
3489 }
3490 
3491 /*
3492  * Wait for free pages to exceed the severe threshold globally.
3493  */
3494 void
3495 vm_wait_severe(void)
3496 {
3497 
3498 	mtx_lock(&vm_domainset_lock);
3499 	while (vm_page_count_severe()) {
3500 		vm_severe_waiters++;
3501 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3502 		    "vmwait", 0);
3503 	}
3504 	mtx_unlock(&vm_domainset_lock);
3505 }
3506 
3507 u_int
3508 vm_wait_count(void)
3509 {
3510 
3511 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3512 }
3513 
3514 int
3515 vm_wait_doms(const domainset_t *wdoms, int mflags)
3516 {
3517 	int error;
3518 
3519 	error = 0;
3520 
3521 	/*
3522 	 * We use racey wakeup synchronization to avoid expensive global
3523 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3524 	 * To handle this, we only sleep for one tick in this instance.  It
3525 	 * is expected that most allocations for the pageproc will come from
3526 	 * kmem or vm_page_grab* which will use the more specific and
3527 	 * race-free vm_wait_domain().
3528 	 */
3529 	if (curproc == pageproc) {
3530 		mtx_lock(&vm_domainset_lock);
3531 		vm_pageproc_waiters++;
3532 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3533 		    PVM | PDROP | mflags, "pageprocwait", 1);
3534 	} else {
3535 		/*
3536 		 * XXX Ideally we would wait only until the allocation could
3537 		 * be satisfied.  This condition can cause new allocators to
3538 		 * consume all freed pages while old allocators wait.
3539 		 */
3540 		mtx_lock(&vm_domainset_lock);
3541 		if (vm_page_count_min_set(wdoms)) {
3542 			if (pageproc == NULL)
3543 				panic("vm_wait in early boot");
3544 			vm_min_waiters++;
3545 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3546 			    PVM | PDROP | mflags, "vmwait", 0);
3547 		} else
3548 			mtx_unlock(&vm_domainset_lock);
3549 	}
3550 	return (error);
3551 }
3552 
3553 /*
3554  *	vm_wait_domain:
3555  *
3556  *	Sleep until free pages are available for allocation.
3557  *	- Called in various places after failed memory allocations.
3558  */
3559 void
3560 vm_wait_domain(int domain)
3561 {
3562 	struct vm_domain *vmd;
3563 	domainset_t wdom;
3564 
3565 	vmd = VM_DOMAIN(domain);
3566 	vm_domain_free_assert_unlocked(vmd);
3567 
3568 	if (curproc == pageproc) {
3569 		mtx_lock(&vm_domainset_lock);
3570 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3571 			vmd->vmd_pageout_pages_needed = 1;
3572 			msleep(&vmd->vmd_pageout_pages_needed,
3573 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3574 		} else
3575 			mtx_unlock(&vm_domainset_lock);
3576 	} else {
3577 		DOMAINSET_ZERO(&wdom);
3578 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3579 		vm_wait_doms(&wdom, 0);
3580 	}
3581 }
3582 
3583 static int
3584 vm_wait_flags(vm_object_t obj, int mflags)
3585 {
3586 	struct domainset *d;
3587 
3588 	d = NULL;
3589 
3590 	/*
3591 	 * Carefully fetch pointers only once: the struct domainset
3592 	 * itself is ummutable but the pointer might change.
3593 	 */
3594 	if (obj != NULL)
3595 		d = obj->domain.dr_policy;
3596 	if (d == NULL)
3597 		d = curthread->td_domain.dr_policy;
3598 
3599 	return (vm_wait_doms(&d->ds_mask, mflags));
3600 }
3601 
3602 /*
3603  *	vm_wait:
3604  *
3605  *	Sleep until free pages are available for allocation in the
3606  *	affinity domains of the obj.  If obj is NULL, the domain set
3607  *	for the calling thread is used.
3608  *	Called in various places after failed memory allocations.
3609  */
3610 void
3611 vm_wait(vm_object_t obj)
3612 {
3613 	(void)vm_wait_flags(obj, 0);
3614 }
3615 
3616 int
3617 vm_wait_intr(vm_object_t obj)
3618 {
3619 	return (vm_wait_flags(obj, PCATCH));
3620 }
3621 
3622 /*
3623  *	vm_domain_alloc_fail:
3624  *
3625  *	Called when a page allocation function fails.  Informs the
3626  *	pagedaemon and performs the requested wait.  Requires the
3627  *	domain_free and object lock on entry.  Returns with the
3628  *	object lock held and free lock released.  Returns an error when
3629  *	retry is necessary.
3630  *
3631  */
3632 static int
3633 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3634 {
3635 
3636 	vm_domain_free_assert_unlocked(vmd);
3637 
3638 	atomic_add_int(&vmd->vmd_pageout_deficit,
3639 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3640 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3641 		if (object != NULL)
3642 			VM_OBJECT_WUNLOCK(object);
3643 		vm_wait_domain(vmd->vmd_domain);
3644 		if (object != NULL)
3645 			VM_OBJECT_WLOCK(object);
3646 		if (req & VM_ALLOC_WAITOK)
3647 			return (EAGAIN);
3648 	}
3649 
3650 	return (0);
3651 }
3652 
3653 /*
3654  *	vm_waitpfault:
3655  *
3656  *	Sleep until free pages are available for allocation.
3657  *	- Called only in vm_fault so that processes page faulting
3658  *	  can be easily tracked.
3659  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3660  *	  processes will be able to grab memory first.  Do not change
3661  *	  this balance without careful testing first.
3662  */
3663 void
3664 vm_waitpfault(struct domainset *dset, int timo)
3665 {
3666 
3667 	/*
3668 	 * XXX Ideally we would wait only until the allocation could
3669 	 * be satisfied.  This condition can cause new allocators to
3670 	 * consume all freed pages while old allocators wait.
3671 	 */
3672 	mtx_lock(&vm_domainset_lock);
3673 	if (vm_page_count_min_set(&dset->ds_mask)) {
3674 		vm_min_waiters++;
3675 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3676 		    "pfault", timo);
3677 	} else
3678 		mtx_unlock(&vm_domainset_lock);
3679 }
3680 
3681 static struct vm_pagequeue *
3682 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3683 {
3684 
3685 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3686 }
3687 
3688 #ifdef INVARIANTS
3689 static struct vm_pagequeue *
3690 vm_page_pagequeue(vm_page_t m)
3691 {
3692 
3693 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3694 }
3695 #endif
3696 
3697 static __always_inline bool
3698 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old,
3699     vm_page_astate_t new)
3700 {
3701 	vm_page_astate_t tmp;
3702 
3703 	tmp = *old;
3704 	do {
3705 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3706 			return (true);
3707 		counter_u64_add(pqstate_commit_retries, 1);
3708 	} while (old->_bits == tmp._bits);
3709 
3710 	return (false);
3711 }
3712 
3713 /*
3714  * Do the work of committing a queue state update that moves the page out of
3715  * its current queue.
3716  */
3717 static bool
3718 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3719     vm_page_astate_t *old, vm_page_astate_t new)
3720 {
3721 	vm_page_t next;
3722 
3723 	vm_pagequeue_assert_locked(pq);
3724 	KASSERT(vm_page_pagequeue(m) == pq,
3725 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3726 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3727 	    ("%s: invalid queue indices %d %d",
3728 	    __func__, old->queue, new.queue));
3729 
3730 	/*
3731 	 * Once the queue index of the page changes there is nothing
3732 	 * synchronizing with further updates to the page's physical
3733 	 * queue state.  Therefore we must speculatively remove the page
3734 	 * from the queue now and be prepared to roll back if the queue
3735 	 * state update fails.  If the page is not physically enqueued then
3736 	 * we just update its queue index.
3737 	 */
3738 	if ((old->flags & PGA_ENQUEUED) != 0) {
3739 		new.flags &= ~PGA_ENQUEUED;
3740 		next = TAILQ_NEXT(m, plinks.q);
3741 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3742 		vm_pagequeue_cnt_dec(pq);
3743 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3744 			if (next == NULL)
3745 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3746 			else
3747 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3748 			vm_pagequeue_cnt_inc(pq);
3749 			return (false);
3750 		} else {
3751 			return (true);
3752 		}
3753 	} else {
3754 		return (vm_page_pqstate_fcmpset(m, old, new));
3755 	}
3756 }
3757 
3758 static bool
3759 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3760     vm_page_astate_t new)
3761 {
3762 	struct vm_pagequeue *pq;
3763 	vm_page_astate_t as;
3764 	bool ret;
3765 
3766 	pq = _vm_page_pagequeue(m, old->queue);
3767 
3768 	/*
3769 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3770 	 * corresponding page queue lock is held.
3771 	 */
3772 	vm_pagequeue_lock(pq);
3773 	as = vm_page_astate_load(m);
3774 	if (__predict_false(as._bits != old->_bits)) {
3775 		*old = as;
3776 		ret = false;
3777 	} else {
3778 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3779 	}
3780 	vm_pagequeue_unlock(pq);
3781 	return (ret);
3782 }
3783 
3784 /*
3785  * Commit a queue state update that enqueues or requeues a page.
3786  */
3787 static bool
3788 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3789     vm_page_astate_t *old, vm_page_astate_t new)
3790 {
3791 	struct vm_domain *vmd;
3792 
3793 	vm_pagequeue_assert_locked(pq);
3794 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3795 	    ("%s: invalid queue indices %d %d",
3796 	    __func__, old->queue, new.queue));
3797 
3798 	new.flags |= PGA_ENQUEUED;
3799 	if (!vm_page_pqstate_fcmpset(m, old, new))
3800 		return (false);
3801 
3802 	if ((old->flags & PGA_ENQUEUED) != 0)
3803 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3804 	else
3805 		vm_pagequeue_cnt_inc(pq);
3806 
3807 	/*
3808 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3809 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3810 	 * applied, even if it was set first.
3811 	 */
3812 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3813 		vmd = vm_pagequeue_domain(m);
3814 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3815 		    ("%s: invalid page queue for page %p", __func__, m));
3816 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3817 	} else {
3818 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3819 	}
3820 	return (true);
3821 }
3822 
3823 /*
3824  * Commit a queue state update that encodes a request for a deferred queue
3825  * operation.
3826  */
3827 static bool
3828 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3829     vm_page_astate_t new)
3830 {
3831 
3832 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3833 	    ("%s: invalid state, queue %d flags %x",
3834 	    __func__, new.queue, new.flags));
3835 
3836 	if (old->_bits != new._bits &&
3837 	    !vm_page_pqstate_fcmpset(m, old, new))
3838 		return (false);
3839 	vm_page_pqbatch_submit(m, new.queue);
3840 	return (true);
3841 }
3842 
3843 /*
3844  * A generic queue state update function.  This handles more cases than the
3845  * specialized functions above.
3846  */
3847 bool
3848 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3849 {
3850 
3851 	if (old->_bits == new._bits)
3852 		return (true);
3853 
3854 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3855 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3856 			return (false);
3857 		if (new.queue != PQ_NONE)
3858 			vm_page_pqbatch_submit(m, new.queue);
3859 	} else {
3860 		if (!vm_page_pqstate_fcmpset(m, old, new))
3861 			return (false);
3862 		if (new.queue != PQ_NONE &&
3863 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3864 			vm_page_pqbatch_submit(m, new.queue);
3865 	}
3866 	return (true);
3867 }
3868 
3869 /*
3870  * Apply deferred queue state updates to a page.
3871  */
3872 static inline void
3873 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3874 {
3875 	vm_page_astate_t new, old;
3876 
3877 	CRITICAL_ASSERT(curthread);
3878 	vm_pagequeue_assert_locked(pq);
3879 	KASSERT(queue < PQ_COUNT,
3880 	    ("%s: invalid queue index %d", __func__, queue));
3881 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3882 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3883 
3884 	for (old = vm_page_astate_load(m);;) {
3885 		if (__predict_false(old.queue != queue ||
3886 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3887 			counter_u64_add(queue_nops, 1);
3888 			break;
3889 		}
3890 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3891 		    ("%s: page %p is unmanaged", __func__, m));
3892 
3893 		new = old;
3894 		if ((old.flags & PGA_DEQUEUE) != 0) {
3895 			new.flags &= ~PGA_QUEUE_OP_MASK;
3896 			new.queue = PQ_NONE;
3897 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3898 			    m, &old, new))) {
3899 				counter_u64_add(queue_ops, 1);
3900 				break;
3901 			}
3902 		} else {
3903 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3904 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3905 			    m, &old, new))) {
3906 				counter_u64_add(queue_ops, 1);
3907 				break;
3908 			}
3909 		}
3910 	}
3911 }
3912 
3913 static void
3914 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3915     uint8_t queue)
3916 {
3917 	int i;
3918 
3919 	for (i = 0; i < bq->bq_cnt; i++)
3920 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3921 	vm_batchqueue_init(bq);
3922 }
3923 
3924 /*
3925  *	vm_page_pqbatch_submit:		[ internal use only ]
3926  *
3927  *	Enqueue a page in the specified page queue's batched work queue.
3928  *	The caller must have encoded the requested operation in the page
3929  *	structure's a.flags field.
3930  */
3931 void
3932 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3933 {
3934 	struct vm_batchqueue *bq;
3935 	struct vm_pagequeue *pq;
3936 	int domain, slots_remaining;
3937 
3938 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3939 
3940 	domain = vm_page_domain(m);
3941 	critical_enter();
3942 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3943 	slots_remaining = vm_batchqueue_insert(bq, m);
3944 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3945 		/* keep building the bq */
3946 		critical_exit();
3947 		return;
3948 	} else if (slots_remaining > 0 ) {
3949 		/* Try to process the bq if we can get the lock */
3950 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3951 		if (vm_pagequeue_trylock(pq)) {
3952 			vm_pqbatch_process(pq, bq, queue);
3953 			vm_pagequeue_unlock(pq);
3954 		}
3955 		critical_exit();
3956 		return;
3957 	}
3958 	critical_exit();
3959 
3960 	/* if we make it here, the bq is full so wait for the lock */
3961 
3962 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3963 	vm_pagequeue_lock(pq);
3964 	critical_enter();
3965 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3966 	vm_pqbatch_process(pq, bq, queue);
3967 	vm_pqbatch_process_page(pq, m, queue);
3968 	vm_pagequeue_unlock(pq);
3969 	critical_exit();
3970 }
3971 
3972 /*
3973  *	vm_page_pqbatch_drain:		[ internal use only ]
3974  *
3975  *	Force all per-CPU page queue batch queues to be drained.  This is
3976  *	intended for use in severe memory shortages, to ensure that pages
3977  *	do not remain stuck in the batch queues.
3978  */
3979 void
3980 vm_page_pqbatch_drain(void)
3981 {
3982 	struct thread *td;
3983 	struct vm_domain *vmd;
3984 	struct vm_pagequeue *pq;
3985 	int cpu, domain, queue;
3986 
3987 	td = curthread;
3988 	CPU_FOREACH(cpu) {
3989 		thread_lock(td);
3990 		sched_bind(td, cpu);
3991 		thread_unlock(td);
3992 
3993 		for (domain = 0; domain < vm_ndomains; domain++) {
3994 			vmd = VM_DOMAIN(domain);
3995 			for (queue = 0; queue < PQ_COUNT; queue++) {
3996 				pq = &vmd->vmd_pagequeues[queue];
3997 				vm_pagequeue_lock(pq);
3998 				critical_enter();
3999 				vm_pqbatch_process(pq,
4000 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
4001 				critical_exit();
4002 				vm_pagequeue_unlock(pq);
4003 			}
4004 		}
4005 	}
4006 	thread_lock(td);
4007 	sched_unbind(td);
4008 	thread_unlock(td);
4009 }
4010 
4011 /*
4012  *	vm_page_dequeue_deferred:	[ internal use only ]
4013  *
4014  *	Request removal of the given page from its current page
4015  *	queue.  Physical removal from the queue may be deferred
4016  *	indefinitely.
4017  */
4018 void
4019 vm_page_dequeue_deferred(vm_page_t m)
4020 {
4021 	vm_page_astate_t new, old;
4022 
4023 	old = vm_page_astate_load(m);
4024 	do {
4025 		if (old.queue == PQ_NONE) {
4026 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4027 			    ("%s: page %p has unexpected queue state",
4028 			    __func__, m));
4029 			break;
4030 		}
4031 		new = old;
4032 		new.flags |= PGA_DEQUEUE;
4033 	} while (!vm_page_pqstate_commit_request(m, &old, new));
4034 }
4035 
4036 /*
4037  *	vm_page_dequeue:
4038  *
4039  *	Remove the page from whichever page queue it's in, if any, before
4040  *	returning.
4041  */
4042 void
4043 vm_page_dequeue(vm_page_t m)
4044 {
4045 	vm_page_astate_t new, old;
4046 
4047 	old = vm_page_astate_load(m);
4048 	do {
4049 		if (old.queue == PQ_NONE) {
4050 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4051 			    ("%s: page %p has unexpected queue state",
4052 			    __func__, m));
4053 			break;
4054 		}
4055 		new = old;
4056 		new.flags &= ~PGA_QUEUE_OP_MASK;
4057 		new.queue = PQ_NONE;
4058 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
4059 
4060 }
4061 
4062 /*
4063  * Schedule the given page for insertion into the specified page queue.
4064  * Physical insertion of the page may be deferred indefinitely.
4065  */
4066 static void
4067 vm_page_enqueue(vm_page_t m, uint8_t queue)
4068 {
4069 
4070 	KASSERT(m->a.queue == PQ_NONE &&
4071 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
4072 	    ("%s: page %p is already enqueued", __func__, m));
4073 	KASSERT(m->ref_count > 0,
4074 	    ("%s: page %p does not carry any references", __func__, m));
4075 
4076 	m->a.queue = queue;
4077 	if ((m->a.flags & PGA_REQUEUE) == 0)
4078 		vm_page_aflag_set(m, PGA_REQUEUE);
4079 	vm_page_pqbatch_submit(m, queue);
4080 }
4081 
4082 /*
4083  *	vm_page_free_prep:
4084  *
4085  *	Prepares the given page to be put on the free list,
4086  *	disassociating it from any VM object. The caller may return
4087  *	the page to the free list only if this function returns true.
4088  *
4089  *	The object, if it exists, must be locked, and then the page must
4090  *	be xbusy.  Otherwise the page must be not busied.  A managed
4091  *	page must be unmapped.
4092  */
4093 static bool
4094 vm_page_free_prep(vm_page_t m)
4095 {
4096 
4097 	/*
4098 	 * Synchronize with threads that have dropped a reference to this
4099 	 * page.
4100 	 */
4101 	atomic_thread_fence_acq();
4102 
4103 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
4104 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
4105 		uint64_t *p;
4106 		int i;
4107 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4108 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4109 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4110 			    m, i, (uintmax_t)*p));
4111 	}
4112 #endif
4113 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4114 		KASSERT(!pmap_page_is_mapped(m),
4115 		    ("vm_page_free_prep: freeing mapped page %p", m));
4116 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4117 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4118 	} else {
4119 		KASSERT(m->a.queue == PQ_NONE,
4120 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4121 	}
4122 	VM_CNT_INC(v_tfree);
4123 
4124 	if (m->object != NULL) {
4125 		vm_page_radix_remove(m);
4126 		vm_page_free_object_prep(m);
4127 	} else
4128 		vm_page_assert_unbusied(m);
4129 
4130 	vm_page_busy_free(m);
4131 
4132 	/*
4133 	 * If fictitious remove object association and
4134 	 * return.
4135 	 */
4136 	if ((m->flags & PG_FICTITIOUS) != 0) {
4137 		KASSERT(m->ref_count == 1,
4138 		    ("fictitious page %p is referenced", m));
4139 		KASSERT(m->a.queue == PQ_NONE,
4140 		    ("fictitious page %p is queued", m));
4141 		return (false);
4142 	}
4143 
4144 	/*
4145 	 * Pages need not be dequeued before they are returned to the physical
4146 	 * memory allocator, but they must at least be marked for a deferred
4147 	 * dequeue.
4148 	 */
4149 	if ((m->oflags & VPO_UNMANAGED) == 0)
4150 		vm_page_dequeue_deferred(m);
4151 
4152 	m->valid = 0;
4153 	vm_page_undirty(m);
4154 
4155 	if (m->ref_count != 0)
4156 		panic("vm_page_free_prep: page %p has references", m);
4157 
4158 	/*
4159 	 * Restore the default memory attribute to the page.
4160 	 */
4161 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4162 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4163 
4164 #if VM_NRESERVLEVEL > 0
4165 	/*
4166 	 * Determine whether the page belongs to a reservation.  If the page was
4167 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4168 	 * as an optimization, we avoid the check in that case.
4169 	 */
4170 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4171 		return (false);
4172 #endif
4173 
4174 	return (true);
4175 }
4176 
4177 /*
4178  *	vm_page_free_toq:
4179  *
4180  *	Returns the given page to the free list, disassociating it
4181  *	from any VM object.
4182  *
4183  *	The object must be locked.  The page must be exclusively busied if it
4184  *	belongs to an object.
4185  */
4186 static void
4187 vm_page_free_toq(vm_page_t m)
4188 {
4189 	struct vm_domain *vmd;
4190 	uma_zone_t zone;
4191 
4192 	if (!vm_page_free_prep(m))
4193 		return;
4194 
4195 	vmd = vm_pagequeue_domain(m);
4196 	if (__predict_false((m->flags & PG_NOFREE) != 0)) {
4197 		vm_page_free_nofree(vmd, m);
4198 		return;
4199 	}
4200 	zone = vmd->vmd_pgcache[m->pool].zone;
4201 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4202 		uma_zfree(zone, m);
4203 		return;
4204 	}
4205 	vm_domain_free_lock(vmd);
4206 	vm_phys_free_pages(m, m->pool, 0);
4207 	vm_domain_free_unlock(vmd);
4208 	vm_domain_freecnt_inc(vmd, 1);
4209 }
4210 
4211 /*
4212  *	vm_page_free_pages_toq:
4213  *
4214  *	Returns a list of pages to the free list, disassociating it
4215  *	from any VM object.  In other words, this is equivalent to
4216  *	calling vm_page_free_toq() for each page of a list of VM objects.
4217  */
4218 int
4219 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4220 {
4221 	vm_page_t m;
4222 	int count;
4223 
4224 	if (SLIST_EMPTY(free))
4225 		return (0);
4226 
4227 	count = 0;
4228 	while ((m = SLIST_FIRST(free)) != NULL) {
4229 		count++;
4230 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4231 		vm_page_free_toq(m);
4232 	}
4233 
4234 	if (update_wire_count)
4235 		vm_wire_sub(count);
4236 	return (count);
4237 }
4238 
4239 /*
4240  * Mark this page as wired down.  For managed pages, this prevents reclamation
4241  * by the page daemon, or when the containing object, if any, is destroyed.
4242  */
4243 void
4244 vm_page_wire(vm_page_t m)
4245 {
4246 	u_int old;
4247 
4248 #ifdef INVARIANTS
4249 	if (m->object != NULL && !vm_page_busied(m) &&
4250 	    !vm_object_busied(m->object))
4251 		VM_OBJECT_ASSERT_LOCKED(m->object);
4252 #endif
4253 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4254 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4255 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4256 
4257 	old = atomic_fetchadd_int(&m->ref_count, 1);
4258 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4259 	    ("vm_page_wire: counter overflow for page %p", m));
4260 	if (VPRC_WIRE_COUNT(old) == 0) {
4261 		if ((m->oflags & VPO_UNMANAGED) == 0)
4262 			vm_page_aflag_set(m, PGA_DEQUEUE);
4263 		vm_wire_add(1);
4264 	}
4265 }
4266 
4267 /*
4268  * Attempt to wire a mapped page following a pmap lookup of that page.
4269  * This may fail if a thread is concurrently tearing down mappings of the page.
4270  * The transient failure is acceptable because it translates to the
4271  * failure of the caller pmap_extract_and_hold(), which should be then
4272  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4273  */
4274 bool
4275 vm_page_wire_mapped(vm_page_t m)
4276 {
4277 	u_int old;
4278 
4279 	old = atomic_load_int(&m->ref_count);
4280 	do {
4281 		KASSERT(old > 0,
4282 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4283 		if ((old & VPRC_BLOCKED) != 0)
4284 			return (false);
4285 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4286 
4287 	if (VPRC_WIRE_COUNT(old) == 0) {
4288 		if ((m->oflags & VPO_UNMANAGED) == 0)
4289 			vm_page_aflag_set(m, PGA_DEQUEUE);
4290 		vm_wire_add(1);
4291 	}
4292 	return (true);
4293 }
4294 
4295 /*
4296  * Release a wiring reference to a managed page.  If the page still belongs to
4297  * an object, update its position in the page queues to reflect the reference.
4298  * If the wiring was the last reference to the page, free the page.
4299  */
4300 static void
4301 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4302 {
4303 	u_int old;
4304 
4305 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4306 	    ("%s: page %p is unmanaged", __func__, m));
4307 
4308 	/*
4309 	 * Update LRU state before releasing the wiring reference.
4310 	 * Use a release store when updating the reference count to
4311 	 * synchronize with vm_page_free_prep().
4312 	 */
4313 	old = atomic_load_int(&m->ref_count);
4314 	do {
4315 		u_int count;
4316 
4317 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4318 		    ("vm_page_unwire: wire count underflow for page %p", m));
4319 
4320 		count = old & ~VPRC_BLOCKED;
4321 		if (count > VPRC_OBJREF + 1) {
4322 			/*
4323 			 * The page has at least one other wiring reference.  An
4324 			 * earlier iteration of this loop may have called
4325 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4326 			 * re-set it if necessary.
4327 			 */
4328 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4329 				vm_page_aflag_set(m, PGA_DEQUEUE);
4330 		} else if (count == VPRC_OBJREF + 1) {
4331 			/*
4332 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4333 			 * update the page's queue state to reflect the
4334 			 * reference.  If the page does not belong to an object
4335 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4336 			 * clear leftover queue state.
4337 			 */
4338 			vm_page_release_toq(m, nqueue, noreuse);
4339 		} else if (count == 1) {
4340 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4341 		}
4342 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4343 
4344 	if (VPRC_WIRE_COUNT(old) == 1) {
4345 		vm_wire_sub(1);
4346 		if (old == 1)
4347 			vm_page_free(m);
4348 	}
4349 }
4350 
4351 /*
4352  * Release one wiring of the specified page, potentially allowing it to be
4353  * paged out.
4354  *
4355  * Only managed pages belonging to an object can be paged out.  If the number
4356  * of wirings transitions to zero and the page is eligible for page out, then
4357  * the page is added to the specified paging queue.  If the released wiring
4358  * represented the last reference to the page, the page is freed.
4359  */
4360 void
4361 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4362 {
4363 
4364 	KASSERT(nqueue < PQ_COUNT,
4365 	    ("vm_page_unwire: invalid queue %u request for page %p",
4366 	    nqueue, m));
4367 
4368 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4369 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4370 			vm_page_free(m);
4371 		return;
4372 	}
4373 	vm_page_unwire_managed(m, nqueue, false);
4374 }
4375 
4376 /*
4377  * Unwire a page without (re-)inserting it into a page queue.  It is up
4378  * to the caller to enqueue, requeue, or free the page as appropriate.
4379  * In most cases involving managed pages, vm_page_unwire() should be used
4380  * instead.
4381  */
4382 bool
4383 vm_page_unwire_noq(vm_page_t m)
4384 {
4385 	u_int old;
4386 
4387 	old = vm_page_drop(m, 1);
4388 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4389 	    ("%s: counter underflow for page %p", __func__,  m));
4390 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4391 	    ("%s: missing ref on fictitious page %p", __func__, m));
4392 
4393 	if (VPRC_WIRE_COUNT(old) > 1)
4394 		return (false);
4395 	if ((m->oflags & VPO_UNMANAGED) == 0)
4396 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4397 	vm_wire_sub(1);
4398 	return (true);
4399 }
4400 
4401 /*
4402  * Ensure that the page ends up in the specified page queue.  If the page is
4403  * active or being moved to the active queue, ensure that its act_count is
4404  * at least ACT_INIT but do not otherwise mess with it.
4405  */
4406 static __always_inline void
4407 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4408 {
4409 	vm_page_astate_t old, new;
4410 
4411 	KASSERT(m->ref_count > 0,
4412 	    ("%s: page %p does not carry any references", __func__, m));
4413 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4414 	    ("%s: invalid flags %x", __func__, nflag));
4415 
4416 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4417 		return;
4418 
4419 	old = vm_page_astate_load(m);
4420 	do {
4421 		if ((old.flags & PGA_DEQUEUE) != 0)
4422 			break;
4423 		new = old;
4424 		new.flags &= ~PGA_QUEUE_OP_MASK;
4425 		if (nqueue == PQ_ACTIVE)
4426 			new.act_count = max(old.act_count, ACT_INIT);
4427 		if (old.queue == nqueue) {
4428 			/*
4429 			 * There is no need to requeue pages already in the
4430 			 * active queue.
4431 			 */
4432 			if (nqueue != PQ_ACTIVE ||
4433 			    (old.flags & PGA_ENQUEUED) == 0)
4434 				new.flags |= nflag;
4435 		} else {
4436 			new.flags |= nflag;
4437 			new.queue = nqueue;
4438 		}
4439 	} while (!vm_page_pqstate_commit(m, &old, new));
4440 }
4441 
4442 /*
4443  * Put the specified page on the active list (if appropriate).
4444  */
4445 void
4446 vm_page_activate(vm_page_t m)
4447 {
4448 
4449 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4450 }
4451 
4452 /*
4453  * Move the specified page to the tail of the inactive queue, or requeue
4454  * the page if it is already in the inactive queue.
4455  */
4456 void
4457 vm_page_deactivate(vm_page_t m)
4458 {
4459 
4460 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4461 }
4462 
4463 void
4464 vm_page_deactivate_noreuse(vm_page_t m)
4465 {
4466 
4467 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4468 }
4469 
4470 /*
4471  * Put a page in the laundry, or requeue it if it is already there.
4472  */
4473 void
4474 vm_page_launder(vm_page_t m)
4475 {
4476 
4477 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4478 }
4479 
4480 /*
4481  * Put a page in the PQ_UNSWAPPABLE holding queue.
4482  */
4483 void
4484 vm_page_unswappable(vm_page_t m)
4485 {
4486 
4487 	VM_OBJECT_ASSERT_LOCKED(m->object);
4488 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4489 	    ("page %p already unswappable", m));
4490 
4491 	vm_page_dequeue(m);
4492 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4493 }
4494 
4495 /*
4496  * Release a page back to the page queues in preparation for unwiring.
4497  */
4498 static void
4499 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4500 {
4501 	vm_page_astate_t old, new;
4502 	uint16_t nflag;
4503 
4504 	/*
4505 	 * Use a check of the valid bits to determine whether we should
4506 	 * accelerate reclamation of the page.  The object lock might not be
4507 	 * held here, in which case the check is racy.  At worst we will either
4508 	 * accelerate reclamation of a valid page and violate LRU, or
4509 	 * unnecessarily defer reclamation of an invalid page.
4510 	 *
4511 	 * If we were asked to not cache the page, place it near the head of the
4512 	 * inactive queue so that is reclaimed sooner.
4513 	 */
4514 	if (noreuse || vm_page_none_valid(m)) {
4515 		nqueue = PQ_INACTIVE;
4516 		nflag = PGA_REQUEUE_HEAD;
4517 	} else {
4518 		nflag = PGA_REQUEUE;
4519 	}
4520 
4521 	old = vm_page_astate_load(m);
4522 	do {
4523 		new = old;
4524 
4525 		/*
4526 		 * If the page is already in the active queue and we are not
4527 		 * trying to accelerate reclamation, simply mark it as
4528 		 * referenced and avoid any queue operations.
4529 		 */
4530 		new.flags &= ~PGA_QUEUE_OP_MASK;
4531 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4532 		    (old.flags & PGA_ENQUEUED) != 0)
4533 			new.flags |= PGA_REFERENCED;
4534 		else {
4535 			new.flags |= nflag;
4536 			new.queue = nqueue;
4537 		}
4538 	} while (!vm_page_pqstate_commit(m, &old, new));
4539 }
4540 
4541 /*
4542  * Unwire a page and either attempt to free it or re-add it to the page queues.
4543  */
4544 void
4545 vm_page_release(vm_page_t m, int flags)
4546 {
4547 	vm_object_t object;
4548 
4549 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4550 	    ("vm_page_release: page %p is unmanaged", m));
4551 
4552 	if ((flags & VPR_TRYFREE) != 0) {
4553 		for (;;) {
4554 			object = atomic_load_ptr(&m->object);
4555 			if (object == NULL)
4556 				break;
4557 			/* Depends on type-stability. */
4558 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4559 				break;
4560 			if (object == m->object) {
4561 				vm_page_release_locked(m, flags);
4562 				VM_OBJECT_WUNLOCK(object);
4563 				return;
4564 			}
4565 			VM_OBJECT_WUNLOCK(object);
4566 		}
4567 	}
4568 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4569 }
4570 
4571 /* See vm_page_release(). */
4572 void
4573 vm_page_release_locked(vm_page_t m, int flags)
4574 {
4575 
4576 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4577 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4578 	    ("vm_page_release_locked: page %p is unmanaged", m));
4579 
4580 	if (vm_page_unwire_noq(m)) {
4581 		if ((flags & VPR_TRYFREE) != 0 &&
4582 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4583 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4584 			/*
4585 			 * An unlocked lookup may have wired the page before the
4586 			 * busy lock was acquired, in which case the page must
4587 			 * not be freed.
4588 			 */
4589 			if (__predict_true(!vm_page_wired(m))) {
4590 				vm_page_free(m);
4591 				return;
4592 			}
4593 			vm_page_xunbusy(m);
4594 		} else {
4595 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4596 		}
4597 	}
4598 }
4599 
4600 static bool
4601 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4602 {
4603 	u_int old;
4604 
4605 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4606 	    ("vm_page_try_blocked_op: page %p has no object", m));
4607 	KASSERT(vm_page_busied(m),
4608 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4609 	VM_OBJECT_ASSERT_LOCKED(m->object);
4610 
4611 	old = atomic_load_int(&m->ref_count);
4612 	do {
4613 		KASSERT(old != 0,
4614 		    ("vm_page_try_blocked_op: page %p has no references", m));
4615 		KASSERT((old & VPRC_BLOCKED) == 0,
4616 		    ("vm_page_try_blocked_op: page %p blocks wirings", m));
4617 		if (VPRC_WIRE_COUNT(old) != 0)
4618 			return (false);
4619 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4620 
4621 	(op)(m);
4622 
4623 	/*
4624 	 * If the object is read-locked, new wirings may be created via an
4625 	 * object lookup.
4626 	 */
4627 	old = vm_page_drop(m, VPRC_BLOCKED);
4628 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4629 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4630 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4631 	    old, m));
4632 	return (true);
4633 }
4634 
4635 /*
4636  * Atomically check for wirings and remove all mappings of the page.
4637  */
4638 bool
4639 vm_page_try_remove_all(vm_page_t m)
4640 {
4641 
4642 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4643 }
4644 
4645 /*
4646  * Atomically check for wirings and remove all writeable mappings of the page.
4647  */
4648 bool
4649 vm_page_try_remove_write(vm_page_t m)
4650 {
4651 
4652 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4653 }
4654 
4655 /*
4656  * vm_page_advise
4657  *
4658  * 	Apply the specified advice to the given page.
4659  */
4660 void
4661 vm_page_advise(vm_page_t m, int advice)
4662 {
4663 
4664 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4665 	vm_page_assert_xbusied(m);
4666 
4667 	if (advice == MADV_FREE)
4668 		/*
4669 		 * Mark the page clean.  This will allow the page to be freed
4670 		 * without first paging it out.  MADV_FREE pages are often
4671 		 * quickly reused by malloc(3), so we do not do anything that
4672 		 * would result in a page fault on a later access.
4673 		 */
4674 		vm_page_undirty(m);
4675 	else if (advice != MADV_DONTNEED) {
4676 		if (advice == MADV_WILLNEED)
4677 			vm_page_activate(m);
4678 		return;
4679 	}
4680 
4681 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4682 		vm_page_dirty(m);
4683 
4684 	/*
4685 	 * Clear any references to the page.  Otherwise, the page daemon will
4686 	 * immediately reactivate the page.
4687 	 */
4688 	vm_page_aflag_clear(m, PGA_REFERENCED);
4689 
4690 	/*
4691 	 * Place clean pages near the head of the inactive queue rather than
4692 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4693 	 * the page will be reused quickly.  Dirty pages not already in the
4694 	 * laundry are moved there.
4695 	 */
4696 	if (m->dirty == 0)
4697 		vm_page_deactivate_noreuse(m);
4698 	else if (!vm_page_in_laundry(m))
4699 		vm_page_launder(m);
4700 }
4701 
4702 /*
4703  *	vm_page_grab_release
4704  *
4705  *	Helper routine for grab functions to release busy on return.
4706  */
4707 static inline void
4708 vm_page_grab_release(vm_page_t m, int allocflags)
4709 {
4710 
4711 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4712 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4713 			vm_page_sunbusy(m);
4714 		else
4715 			vm_page_xunbusy(m);
4716 	}
4717 }
4718 
4719 /*
4720  *	vm_page_grab_sleep
4721  *
4722  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4723  *	if the caller should retry and false otherwise.
4724  *
4725  *	If the object is locked on entry the object will be unlocked with
4726  *	false returns and still locked but possibly having been dropped
4727  *	with true returns.
4728  */
4729 static bool
4730 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4731     const char *wmesg, int allocflags, bool locked)
4732 {
4733 
4734 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4735 		return (false);
4736 
4737 	/*
4738 	 * Reference the page before unlocking and sleeping so that
4739 	 * the page daemon is less likely to reclaim it.
4740 	 */
4741 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4742 		vm_page_reference(m);
4743 
4744 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4745 	    locked)
4746 		VM_OBJECT_WLOCK(object);
4747 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4748 		return (false);
4749 
4750 	return (true);
4751 }
4752 
4753 /*
4754  * Assert that the grab flags are valid.
4755  */
4756 static inline void
4757 vm_page_grab_check(int allocflags)
4758 {
4759 
4760 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4761 	    (allocflags & VM_ALLOC_WIRED) != 0,
4762 	    ("vm_page_grab*: the pages must be busied or wired"));
4763 
4764 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4765 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4766 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4767 }
4768 
4769 /*
4770  * Calculate the page allocation flags for grab.
4771  */
4772 static inline int
4773 vm_page_grab_pflags(int allocflags)
4774 {
4775 	int pflags;
4776 
4777 	pflags = allocflags &
4778 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4779 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4780 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4781 		pflags |= VM_ALLOC_WAITFAIL;
4782 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4783 		pflags |= VM_ALLOC_SBUSY;
4784 
4785 	return (pflags);
4786 }
4787 
4788 /*
4789  * Grab a page, waiting until we are woken up due to the page changing state.
4790  * We keep on waiting, if the page continues to be in the object, unless
4791  * allocflags forbid waiting.
4792  *
4793  * The object must be locked on entry.  This routine may sleep.  The lock will,
4794  * however, be released and reacquired if the routine sleeps.
4795  *
4796  *  Return a grabbed page, or NULL.  Set *found if a page was found, whether or
4797  *  not it was grabbed.
4798  */
4799 static inline vm_page_t
4800 vm_page_grab_lookup(struct pctrie_iter *pages, vm_object_t object,
4801     vm_pindex_t pindex, int allocflags, bool *found)
4802 {
4803 	vm_page_t m;
4804 
4805 	while ((*found = (m = vm_radix_iter_lookup(pages, pindex)) != NULL) &&
4806 	    !vm_page_tryacquire(m, allocflags)) {
4807 		if (!vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4808 		    allocflags, true))
4809 			return (NULL);
4810 		pctrie_iter_reset(pages);
4811 	}
4812 	return (m);
4813 }
4814 
4815 /*
4816  * Grab a page.  Use an iterator parameter. Keep on waiting, as long as the page
4817  * exists in the object.  If the page doesn't exist, first allocate it and then
4818  * conditionally zero it.
4819  *
4820  * The object must be locked on entry.  This routine may sleep.  The lock will,
4821  * however, be released and reacquired if the routine sleeps.
4822  */
4823 vm_page_t
4824 vm_page_grab_iter(vm_object_t object, struct pctrie_iter *pages,
4825     vm_pindex_t pindex, int allocflags)
4826 {
4827 	vm_page_t m, mpred;
4828 	bool found;
4829 
4830 	VM_OBJECT_ASSERT_WLOCKED(object);
4831 	vm_page_grab_check(allocflags);
4832 
4833 	while ((m = vm_page_grab_lookup(
4834 	    pages, object, pindex, allocflags, &found)) == NULL) {
4835 		if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4836 			return (NULL);
4837 		if (found &&
4838 		    (allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4839 			return (NULL);
4840 		mpred = vm_radix_iter_lookup_lt(pages, pindex);
4841 		m = vm_page_alloc_after(object, pages, pindex,
4842 		    vm_page_grab_pflags(allocflags), mpred);
4843 		if (m != NULL) {
4844 			if ((allocflags & VM_ALLOC_ZERO) != 0 &&
4845 			    (m->flags & PG_ZERO) == 0)
4846 				pmap_zero_page(m);
4847 			break;
4848 		}
4849 		if ((allocflags &
4850 		    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4851 			return (NULL);
4852 	}
4853 	vm_page_grab_release(m, allocflags);
4854 
4855 	return (m);
4856 }
4857 
4858 /*
4859  * Grab a page.  Keep on waiting, as long as the page exists in the object.  If
4860  * the page doesn't exist, first allocate it and then conditionally zero it.
4861  *
4862  * The object must be locked on entry.  This routine may sleep.  The lock will,
4863  * however, be released and reacquired if the routine sleeps.
4864  */
4865 vm_page_t
4866 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4867 {
4868 	struct pctrie_iter pages;
4869 
4870 	VM_OBJECT_ASSERT_WLOCKED(object);
4871 	vm_page_iter_init(&pages, object);
4872 	return (vm_page_grab_iter(object, &pages, pindex, allocflags));
4873 }
4874 
4875 /*
4876  * Attempt to validate a page, locklessly acquiring it if necessary, given a
4877  * (object, pindex) tuple and either an invalided page or NULL.  The resulting
4878  * page will be validated against the identity tuple, and busied or wired as
4879  * requested.  A NULL page returned guarantees that the page was not in radix at
4880  * the time of the call but callers must perform higher level synchronization or
4881  * retry the operation under a lock if they require an atomic answer.  This is
4882  * the only lock free validation routine, other routines can depend on the
4883  * resulting page state.
4884  *
4885  * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to
4886  * caller flags.
4887  */
4888 #define PAGE_NOT_ACQUIRED ((vm_page_t)1)
4889 static vm_page_t
4890 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m,
4891     int allocflags)
4892 {
4893 	if (m == NULL)
4894 		m = vm_page_lookup_unlocked(object, pindex);
4895 	for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) {
4896 		if (vm_page_trybusy(m, allocflags)) {
4897 			if (m->object == object && m->pindex == pindex) {
4898 				if ((allocflags & VM_ALLOC_WIRED) != 0)
4899 					vm_page_wire(m);
4900 				vm_page_grab_release(m, allocflags);
4901 				break;
4902 			}
4903 			/* relookup. */
4904 			vm_page_busy_release(m);
4905 			cpu_spinwait();
4906 			continue;
4907 		}
4908 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4909 		    allocflags, false))
4910 			return (PAGE_NOT_ACQUIRED);
4911 	}
4912 	return (m);
4913 }
4914 
4915 /*
4916  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4917  * is not set.
4918  */
4919 vm_page_t
4920 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4921 {
4922 	vm_page_t m;
4923 
4924 	vm_page_grab_check(allocflags);
4925 	m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags);
4926 	if (m == PAGE_NOT_ACQUIRED)
4927 		return (NULL);
4928 	if (m != NULL)
4929 		return (m);
4930 
4931 	/*
4932 	 * The radix lockless lookup should never return a false negative
4933 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4934 	 * was no page present at the instant of the call.  A NOCREAT caller
4935 	 * must handle create races gracefully.
4936 	 */
4937 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4938 		return (NULL);
4939 
4940 	VM_OBJECT_WLOCK(object);
4941 	m = vm_page_grab(object, pindex, allocflags);
4942 	VM_OBJECT_WUNLOCK(object);
4943 
4944 	return (m);
4945 }
4946 
4947 /*
4948  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4949  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4950  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4951  * in simultaneously.  Additional pages will be left on a paging queue but
4952  * will neither be wired nor busy regardless of allocflags.
4953  */
4954 int
4955 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex,
4956     int allocflags)
4957 {
4958 	struct pctrie_iter pages;
4959 	vm_page_t m, mpred;
4960 	vm_page_t ma[VM_INITIAL_PAGEIN];
4961 	int after, i, pflags, rv;
4962 
4963 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4964 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4965 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4966 	KASSERT((allocflags &
4967 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4968 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4969 	VM_OBJECT_ASSERT_WLOCKED(object);
4970 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4971 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4972 	pflags |= VM_ALLOC_WAITFAIL;
4973 	vm_page_iter_init(&pages, object);
4974 
4975 retrylookup:
4976 	if ((m = vm_radix_iter_lookup(&pages, pindex)) != NULL) {
4977 		/*
4978 		 * If the page is fully valid it can only become invalid
4979 		 * with the object lock held.  If it is not valid it can
4980 		 * become valid with the busy lock held.  Therefore, we
4981 		 * may unnecessarily lock the exclusive busy here if we
4982 		 * race with I/O completion not using the object lock.
4983 		 * However, we will not end up with an invalid page and a
4984 		 * shared lock.
4985 		 */
4986 		if (!vm_page_trybusy(m,
4987 		    vm_page_all_valid(m) ? allocflags : 0)) {
4988 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4989 			    allocflags, true);
4990 			pctrie_iter_reset(&pages);
4991 			goto retrylookup;
4992 		}
4993 		if (vm_page_all_valid(m))
4994 			goto out;
4995 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4996 			vm_page_busy_release(m);
4997 			*mp = NULL;
4998 			return (VM_PAGER_FAIL);
4999 		}
5000 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5001 		*mp = NULL;
5002 		return (VM_PAGER_FAIL);
5003 	} else {
5004 		mpred = vm_radix_iter_lookup_lt(&pages, pindex);
5005 		m = vm_page_alloc_after(object, &pages, pindex, pflags, mpred);
5006 		if (m == NULL) {
5007 			if (!vm_pager_can_alloc_page(object, pindex)) {
5008 				*mp = NULL;
5009 				return (VM_PAGER_AGAIN);
5010 			}
5011 			goto retrylookup;
5012 		}
5013 	}
5014 
5015 	vm_page_assert_xbusied(m);
5016 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
5017 		after = MIN(after, VM_INITIAL_PAGEIN);
5018 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
5019 		after = MAX(after, 1);
5020 		ma[0] = mpred = m;
5021 		for (i = 1; i < after; i++) {
5022 			m = vm_radix_iter_lookup(&pages, pindex + i);
5023 			if (m == NULL) {
5024 				m = vm_page_alloc_after(object, &pages,
5025 				    pindex + i, VM_ALLOC_NORMAL, mpred);
5026 				if (m == NULL)
5027 					break;
5028 			} else if (vm_page_any_valid(m) || !vm_page_tryxbusy(m))
5029 				break;
5030 			mpred = ma[i] = m;
5031 		}
5032 		after = i;
5033 		vm_object_pip_add(object, after);
5034 		VM_OBJECT_WUNLOCK(object);
5035 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
5036 		VM_OBJECT_WLOCK(object);
5037 		vm_object_pip_wakeupn(object, after);
5038 		/* Pager may have replaced a page. */
5039 		m = ma[0];
5040 		if (rv != VM_PAGER_OK) {
5041 			for (i = 0; i < after; i++) {
5042 				if (!vm_page_wired(ma[i]))
5043 					vm_page_free(ma[i]);
5044 				else
5045 					vm_page_xunbusy(ma[i]);
5046 			}
5047 			*mp = NULL;
5048 			return (rv);
5049 		}
5050 		for (i = 1; i < after; i++)
5051 			vm_page_readahead_finish(ma[i]);
5052 		MPASS(vm_page_all_valid(m));
5053 	} else {
5054 		vm_page_zero_invalid(m, TRUE);
5055 	}
5056 out:
5057 	if ((allocflags & VM_ALLOC_WIRED) != 0)
5058 		vm_page_wire(m);
5059 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
5060 		vm_page_busy_downgrade(m);
5061 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
5062 		vm_page_busy_release(m);
5063 	*mp = m;
5064 	return (VM_PAGER_OK);
5065 }
5066 
5067 /*
5068  * Grab a page.  Keep on waiting, as long as the page exists in the object.  If
5069  * the page doesn't exist, and the pager has it, allocate it and zero part of
5070  * it.
5071  *
5072  * The object must be locked on entry.  This routine may sleep.  The lock will,
5073  * however, be released and reacquired if the routine sleeps.
5074  */
5075 int
5076 vm_page_grab_zero_partial(vm_object_t object, vm_pindex_t pindex, int base,
5077     int end)
5078 {
5079 	struct pctrie_iter pages;
5080 	vm_page_t m, mpred;
5081 	int allocflags, rv;
5082 	bool found;
5083 
5084 	VM_OBJECT_ASSERT_WLOCKED(object);
5085 	KASSERT(base >= 0, ("%s: base %d", __func__, base));
5086 	KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base,
5087 	    end));
5088 
5089 	allocflags = VM_ALLOC_NOCREAT | VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL;
5090 	vm_page_iter_init(&pages, object);
5091 	while ((m = vm_page_grab_lookup(
5092 	    &pages, object, pindex, allocflags, &found)) == NULL) {
5093 		if (!vm_pager_has_page(object, pindex, NULL, NULL))
5094 			return (0);
5095 		mpred = vm_radix_iter_lookup_lt(&pages, pindex);
5096 		m = vm_page_alloc_after(object, &pages, pindex,
5097 		    vm_page_grab_pflags(allocflags), mpred);
5098 		if (m != NULL) {
5099 			vm_object_pip_add(object, 1);
5100 			VM_OBJECT_WUNLOCK(object);
5101 			rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
5102 			VM_OBJECT_WLOCK(object);
5103 			vm_object_pip_wakeup(object);
5104 			if (rv != VM_PAGER_OK) {
5105 				vm_page_free(m);
5106 				return (EIO);
5107 			}
5108 
5109 			/*
5110 			 * Since the page was not resident, and therefore not
5111 			 * recently accessed, immediately enqueue it for
5112 			 * asynchronous laundering.  The current operation is
5113 			 * not regarded as an access.
5114 			 */
5115 			vm_page_launder(m);
5116 			break;
5117 		}
5118 	}
5119 
5120 	pmap_zero_page_area(m, base, end - base);
5121 	KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m));
5122 	vm_page_set_dirty(m);
5123 	vm_page_xunbusy(m);
5124 	return (0);
5125 }
5126 
5127 /*
5128  * Locklessly grab a valid page.  If the page is not valid or not yet
5129  * allocated this will fall back to the object lock method.
5130  */
5131 int
5132 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
5133     vm_pindex_t pindex, int allocflags)
5134 {
5135 	vm_page_t m;
5136 	int flags;
5137 	int error;
5138 
5139 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
5140 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
5141 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
5142 	    "mismatch"));
5143 	KASSERT((allocflags &
5144 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
5145 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
5146 
5147 	/*
5148 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
5149 	 * before we can inspect the valid field and return a wired page.
5150 	 */
5151 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
5152 	vm_page_grab_check(flags);
5153 	m = vm_page_acquire_unlocked(object, pindex, NULL, flags);
5154 	if (m == PAGE_NOT_ACQUIRED)
5155 		return (VM_PAGER_FAIL);
5156 	if (m != NULL) {
5157 		if (vm_page_all_valid(m)) {
5158 			if ((allocflags & VM_ALLOC_WIRED) != 0)
5159 				vm_page_wire(m);
5160 			vm_page_grab_release(m, allocflags);
5161 			*mp = m;
5162 			return (VM_PAGER_OK);
5163 		}
5164 		vm_page_busy_release(m);
5165 	}
5166 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5167 		*mp = NULL;
5168 		return (VM_PAGER_FAIL);
5169 	}
5170 	VM_OBJECT_WLOCK(object);
5171 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
5172 	VM_OBJECT_WUNLOCK(object);
5173 
5174 	return (error);
5175 }
5176 
5177 /*
5178  * Return the specified range of pages from the given object.  For each
5179  * page offset within the range, if a page already exists within the object
5180  * at that offset and it is busy, then wait for it to change state.  If,
5181  * instead, the page doesn't exist, then allocate it.
5182  *
5183  * The caller must always specify an allocation class.
5184  *
5185  * allocation classes:
5186  *	VM_ALLOC_NORMAL		normal process request
5187  *	VM_ALLOC_SYSTEM		system *really* needs the pages
5188  *
5189  * The caller must always specify that the pages are to be busied and/or
5190  * wired.
5191  *
5192  * optional allocation flags:
5193  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5194  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5195  *	VM_ALLOC_NOWAIT		do not sleep
5196  *	VM_ALLOC_SBUSY		set page to sbusy state
5197  *	VM_ALLOC_WIRED		wire the pages
5198  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5199  *
5200  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5201  * may return a partial prefix of the requested range.
5202  */
5203 int
5204 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5205     vm_page_t *ma, int count)
5206 {
5207 	struct pctrie_iter pages;
5208 	vm_page_t m, mpred;
5209 	int pflags;
5210 	int i;
5211 
5212 	VM_OBJECT_ASSERT_WLOCKED(object);
5213 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5214 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5215 	KASSERT(count > 0,
5216 	    ("vm_page_grab_pages: invalid page count %d", count));
5217 	vm_page_grab_check(allocflags);
5218 
5219 	pflags = vm_page_grab_pflags(allocflags);
5220 	i = 0;
5221 	vm_page_iter_init(&pages, object);
5222 retrylookup:
5223 	mpred = vm_radix_iter_lookup_lt(&pages, pindex + i);
5224 	for (; i < count; i++) {
5225 		m = vm_radix_iter_lookup(&pages, pindex + i);
5226 		if (m != NULL) {
5227 			if (!vm_page_tryacquire(m, allocflags)) {
5228 				if (vm_page_grab_sleep(object, m, pindex + i,
5229 				    "grbmaw", allocflags, true)) {
5230 					pctrie_iter_reset(&pages);
5231 					goto retrylookup;
5232 				}
5233 				break;
5234 			}
5235 		} else {
5236 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5237 				break;
5238 			m = vm_page_alloc_after(object, &pages, pindex + i,
5239 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5240 			if (m == NULL) {
5241 				if ((allocflags & (VM_ALLOC_NOWAIT |
5242 				    VM_ALLOC_WAITFAIL)) != 0)
5243 					break;
5244 				goto retrylookup;
5245 			}
5246 		}
5247 		if (vm_page_none_valid(m) &&
5248 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5249 			if ((m->flags & PG_ZERO) == 0)
5250 				pmap_zero_page(m);
5251 			vm_page_valid(m);
5252 		}
5253 		vm_page_grab_release(m, allocflags);
5254 		ma[i] = mpred = m;
5255 	}
5256 	return (i);
5257 }
5258 
5259 /*
5260  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5261  * and will fall back to the locked variant to handle allocation.
5262  */
5263 int
5264 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5265     int allocflags, vm_page_t *ma, int count)
5266 {
5267 	vm_page_t m;
5268 	int flags;
5269 	int i;
5270 
5271 	KASSERT(count > 0,
5272 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5273 	vm_page_grab_check(allocflags);
5274 
5275 	/*
5276 	 * Modify flags for lockless acquire to hold the page until we
5277 	 * set it valid if necessary.
5278 	 */
5279 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5280 	vm_page_grab_check(flags);
5281 	m = NULL;
5282 	for (i = 0; i < count; i++, pindex++) {
5283 		/*
5284 		 * We may see a false NULL here because the previous page has
5285 		 * been removed or just inserted and the list is loaded without
5286 		 * barriers.  Switch to radix to verify.
5287 		 */
5288 		if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex ||
5289 		    atomic_load_ptr(&m->object) != object) {
5290 			/*
5291 			 * This guarantees the result is instantaneously
5292 			 * correct.
5293 			 */
5294 			m = NULL;
5295 		}
5296 		m = vm_page_acquire_unlocked(object, pindex, m, flags);
5297 		if (m == PAGE_NOT_ACQUIRED)
5298 			return (i);
5299 		if (m == NULL)
5300 			break;
5301 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5302 			if ((m->flags & PG_ZERO) == 0)
5303 				pmap_zero_page(m);
5304 			vm_page_valid(m);
5305 		}
5306 		/* m will still be wired or busy according to flags. */
5307 		vm_page_grab_release(m, allocflags);
5308 		ma[i] = m;
5309 		m = TAILQ_NEXT(m, listq);
5310 	}
5311 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5312 		return (i);
5313 	count -= i;
5314 	VM_OBJECT_WLOCK(object);
5315 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5316 	VM_OBJECT_WUNLOCK(object);
5317 
5318 	return (i);
5319 }
5320 
5321 /*
5322  * Mapping function for valid or dirty bits in a page.
5323  *
5324  * Inputs are required to range within a page.
5325  */
5326 vm_page_bits_t
5327 vm_page_bits(int base, int size)
5328 {
5329 	int first_bit;
5330 	int last_bit;
5331 
5332 	KASSERT(
5333 	    base + size <= PAGE_SIZE,
5334 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5335 	);
5336 
5337 	if (size == 0)		/* handle degenerate case */
5338 		return (0);
5339 
5340 	first_bit = base >> DEV_BSHIFT;
5341 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5342 
5343 	return (((vm_page_bits_t)2 << last_bit) -
5344 	    ((vm_page_bits_t)1 << first_bit));
5345 }
5346 
5347 void
5348 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5349 {
5350 
5351 #if PAGE_SIZE == 32768
5352 	atomic_set_64((uint64_t *)bits, set);
5353 #elif PAGE_SIZE == 16384
5354 	atomic_set_32((uint32_t *)bits, set);
5355 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5356 	atomic_set_16((uint16_t *)bits, set);
5357 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5358 	atomic_set_8((uint8_t *)bits, set);
5359 #else		/* PAGE_SIZE <= 8192 */
5360 	uintptr_t addr;
5361 	int shift;
5362 
5363 	addr = (uintptr_t)bits;
5364 	/*
5365 	 * Use a trick to perform a 32-bit atomic on the
5366 	 * containing aligned word, to not depend on the existence
5367 	 * of atomic_{set, clear}_{8, 16}.
5368 	 */
5369 	shift = addr & (sizeof(uint32_t) - 1);
5370 #if BYTE_ORDER == BIG_ENDIAN
5371 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5372 #else
5373 	shift *= NBBY;
5374 #endif
5375 	addr &= ~(sizeof(uint32_t) - 1);
5376 	atomic_set_32((uint32_t *)addr, set << shift);
5377 #endif		/* PAGE_SIZE */
5378 }
5379 
5380 static inline void
5381 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5382 {
5383 
5384 #if PAGE_SIZE == 32768
5385 	atomic_clear_64((uint64_t *)bits, clear);
5386 #elif PAGE_SIZE == 16384
5387 	atomic_clear_32((uint32_t *)bits, clear);
5388 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5389 	atomic_clear_16((uint16_t *)bits, clear);
5390 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5391 	atomic_clear_8((uint8_t *)bits, clear);
5392 #else		/* PAGE_SIZE <= 8192 */
5393 	uintptr_t addr;
5394 	int shift;
5395 
5396 	addr = (uintptr_t)bits;
5397 	/*
5398 	 * Use a trick to perform a 32-bit atomic on the
5399 	 * containing aligned word, to not depend on the existence
5400 	 * of atomic_{set, clear}_{8, 16}.
5401 	 */
5402 	shift = addr & (sizeof(uint32_t) - 1);
5403 #if BYTE_ORDER == BIG_ENDIAN
5404 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5405 #else
5406 	shift *= NBBY;
5407 #endif
5408 	addr &= ~(sizeof(uint32_t) - 1);
5409 	atomic_clear_32((uint32_t *)addr, clear << shift);
5410 #endif		/* PAGE_SIZE */
5411 }
5412 
5413 static inline vm_page_bits_t
5414 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5415 {
5416 #if PAGE_SIZE == 32768
5417 	uint64_t old;
5418 
5419 	old = *bits;
5420 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5421 	return (old);
5422 #elif PAGE_SIZE == 16384
5423 	uint32_t old;
5424 
5425 	old = *bits;
5426 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5427 	return (old);
5428 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5429 	uint16_t old;
5430 
5431 	old = *bits;
5432 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5433 	return (old);
5434 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5435 	uint8_t old;
5436 
5437 	old = *bits;
5438 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5439 	return (old);
5440 #else		/* PAGE_SIZE <= 4096*/
5441 	uintptr_t addr;
5442 	uint32_t old, new, mask;
5443 	int shift;
5444 
5445 	addr = (uintptr_t)bits;
5446 	/*
5447 	 * Use a trick to perform a 32-bit atomic on the
5448 	 * containing aligned word, to not depend on the existence
5449 	 * of atomic_{set, swap, clear}_{8, 16}.
5450 	 */
5451 	shift = addr & (sizeof(uint32_t) - 1);
5452 #if BYTE_ORDER == BIG_ENDIAN
5453 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5454 #else
5455 	shift *= NBBY;
5456 #endif
5457 	addr &= ~(sizeof(uint32_t) - 1);
5458 	mask = VM_PAGE_BITS_ALL << shift;
5459 
5460 	old = *bits;
5461 	do {
5462 		new = old & ~mask;
5463 		new |= newbits << shift;
5464 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5465 	return (old >> shift);
5466 #endif		/* PAGE_SIZE */
5467 }
5468 
5469 /*
5470  *	vm_page_set_valid_range:
5471  *
5472  *	Sets portions of a page valid.  The arguments are expected
5473  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5474  *	of any partial chunks touched by the range.  The invalid portion of
5475  *	such chunks will be zeroed.
5476  *
5477  *	(base + size) must be less then or equal to PAGE_SIZE.
5478  */
5479 void
5480 vm_page_set_valid_range(vm_page_t m, int base, int size)
5481 {
5482 	int endoff, frag;
5483 	vm_page_bits_t pagebits;
5484 
5485 	vm_page_assert_busied(m);
5486 	if (size == 0)	/* handle degenerate case */
5487 		return;
5488 
5489 	/*
5490 	 * If the base is not DEV_BSIZE aligned and the valid
5491 	 * bit is clear, we have to zero out a portion of the
5492 	 * first block.
5493 	 */
5494 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5495 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5496 		pmap_zero_page_area(m, frag, base - frag);
5497 
5498 	/*
5499 	 * If the ending offset is not DEV_BSIZE aligned and the
5500 	 * valid bit is clear, we have to zero out a portion of
5501 	 * the last block.
5502 	 */
5503 	endoff = base + size;
5504 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5505 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5506 		pmap_zero_page_area(m, endoff,
5507 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5508 
5509 	/*
5510 	 * Assert that no previously invalid block that is now being validated
5511 	 * is already dirty.
5512 	 */
5513 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5514 	    ("vm_page_set_valid_range: page %p is dirty", m));
5515 
5516 	/*
5517 	 * Set valid bits inclusive of any overlap.
5518 	 */
5519 	pagebits = vm_page_bits(base, size);
5520 	if (vm_page_xbusied(m))
5521 		m->valid |= pagebits;
5522 	else
5523 		vm_page_bits_set(m, &m->valid, pagebits);
5524 }
5525 
5526 /*
5527  * Set the page dirty bits and free the invalid swap space if
5528  * present.  Returns the previous dirty bits.
5529  */
5530 vm_page_bits_t
5531 vm_page_set_dirty(vm_page_t m)
5532 {
5533 	vm_page_bits_t old;
5534 
5535 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5536 
5537 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5538 		old = m->dirty;
5539 		m->dirty = VM_PAGE_BITS_ALL;
5540 	} else
5541 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5542 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5543 		vm_pager_page_unswapped(m);
5544 
5545 	return (old);
5546 }
5547 
5548 /*
5549  * Clear the given bits from the specified page's dirty field.
5550  */
5551 static __inline void
5552 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5553 {
5554 
5555 	vm_page_assert_busied(m);
5556 
5557 	/*
5558 	 * If the page is xbusied and not write mapped we are the
5559 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5560 	 * layer can call vm_page_dirty() without holding a distinguished
5561 	 * lock.  The combination of page busy and atomic operations
5562 	 * suffice to guarantee consistency of the page dirty field.
5563 	 */
5564 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5565 		m->dirty &= ~pagebits;
5566 	else
5567 		vm_page_bits_clear(m, &m->dirty, pagebits);
5568 }
5569 
5570 /*
5571  *	vm_page_set_validclean:
5572  *
5573  *	Sets portions of a page valid and clean.  The arguments are expected
5574  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5575  *	of any partial chunks touched by the range.  The invalid portion of
5576  *	such chunks will be zero'd.
5577  *
5578  *	(base + size) must be less then or equal to PAGE_SIZE.
5579  */
5580 void
5581 vm_page_set_validclean(vm_page_t m, int base, int size)
5582 {
5583 	vm_page_bits_t oldvalid, pagebits;
5584 	int endoff, frag;
5585 
5586 	vm_page_assert_busied(m);
5587 	if (size == 0)	/* handle degenerate case */
5588 		return;
5589 
5590 	/*
5591 	 * If the base is not DEV_BSIZE aligned and the valid
5592 	 * bit is clear, we have to zero out a portion of the
5593 	 * first block.
5594 	 */
5595 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5596 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5597 		pmap_zero_page_area(m, frag, base - frag);
5598 
5599 	/*
5600 	 * If the ending offset is not DEV_BSIZE aligned and the
5601 	 * valid bit is clear, we have to zero out a portion of
5602 	 * the last block.
5603 	 */
5604 	endoff = base + size;
5605 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5606 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5607 		pmap_zero_page_area(m, endoff,
5608 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5609 
5610 	/*
5611 	 * Set valid, clear dirty bits.  If validating the entire
5612 	 * page we can safely clear the pmap modify bit.  We also
5613 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5614 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5615 	 * be set again.
5616 	 *
5617 	 * We set valid bits inclusive of any overlap, but we can only
5618 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5619 	 * the range.
5620 	 */
5621 	oldvalid = m->valid;
5622 	pagebits = vm_page_bits(base, size);
5623 	if (vm_page_xbusied(m))
5624 		m->valid |= pagebits;
5625 	else
5626 		vm_page_bits_set(m, &m->valid, pagebits);
5627 #if 0	/* NOT YET */
5628 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5629 		frag = DEV_BSIZE - frag;
5630 		base += frag;
5631 		size -= frag;
5632 		if (size < 0)
5633 			size = 0;
5634 	}
5635 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5636 #endif
5637 	if (base == 0 && size == PAGE_SIZE) {
5638 		/*
5639 		 * The page can only be modified within the pmap if it is
5640 		 * mapped, and it can only be mapped if it was previously
5641 		 * fully valid.
5642 		 */
5643 		if (oldvalid == VM_PAGE_BITS_ALL)
5644 			/*
5645 			 * Perform the pmap_clear_modify() first.  Otherwise,
5646 			 * a concurrent pmap operation, such as
5647 			 * pmap_protect(), could clear a modification in the
5648 			 * pmap and set the dirty field on the page before
5649 			 * pmap_clear_modify() had begun and after the dirty
5650 			 * field was cleared here.
5651 			 */
5652 			pmap_clear_modify(m);
5653 		m->dirty = 0;
5654 		vm_page_aflag_clear(m, PGA_NOSYNC);
5655 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5656 		m->dirty &= ~pagebits;
5657 	else
5658 		vm_page_clear_dirty_mask(m, pagebits);
5659 }
5660 
5661 void
5662 vm_page_clear_dirty(vm_page_t m, int base, int size)
5663 {
5664 
5665 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5666 }
5667 
5668 /*
5669  *	vm_page_set_invalid:
5670  *
5671  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5672  *	valid and dirty bits for the effected areas are cleared.
5673  */
5674 void
5675 vm_page_set_invalid(vm_page_t m, int base, int size)
5676 {
5677 	vm_page_bits_t bits;
5678 	vm_object_t object;
5679 
5680 	/*
5681 	 * The object lock is required so that pages can't be mapped
5682 	 * read-only while we're in the process of invalidating them.
5683 	 */
5684 	object = m->object;
5685 	VM_OBJECT_ASSERT_WLOCKED(object);
5686 	vm_page_assert_busied(m);
5687 
5688 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5689 	    size >= object->un_pager.vnp.vnp_size)
5690 		bits = VM_PAGE_BITS_ALL;
5691 	else
5692 		bits = vm_page_bits(base, size);
5693 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5694 		pmap_remove_all(m);
5695 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5696 	    !pmap_page_is_mapped(m),
5697 	    ("vm_page_set_invalid: page %p is mapped", m));
5698 	if (vm_page_xbusied(m)) {
5699 		m->valid &= ~bits;
5700 		m->dirty &= ~bits;
5701 	} else {
5702 		vm_page_bits_clear(m, &m->valid, bits);
5703 		vm_page_bits_clear(m, &m->dirty, bits);
5704 	}
5705 }
5706 
5707 /*
5708  *	vm_page_invalid:
5709  *
5710  *	Invalidates the entire page.  The page must be busy, unmapped, and
5711  *	the enclosing object must be locked.  The object locks protects
5712  *	against concurrent read-only pmap enter which is done without
5713  *	busy.
5714  */
5715 void
5716 vm_page_invalid(vm_page_t m)
5717 {
5718 
5719 	vm_page_assert_busied(m);
5720 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5721 	MPASS(!pmap_page_is_mapped(m));
5722 
5723 	if (vm_page_xbusied(m))
5724 		m->valid = 0;
5725 	else
5726 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5727 }
5728 
5729 /*
5730  * vm_page_zero_invalid()
5731  *
5732  *	The kernel assumes that the invalid portions of a page contain
5733  *	garbage, but such pages can be mapped into memory by user code.
5734  *	When this occurs, we must zero out the non-valid portions of the
5735  *	page so user code sees what it expects.
5736  *
5737  *	Pages are most often semi-valid when the end of a file is mapped
5738  *	into memory and the file's size is not page aligned.
5739  */
5740 void
5741 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5742 {
5743 	int b;
5744 	int i;
5745 
5746 	/*
5747 	 * Scan the valid bits looking for invalid sections that
5748 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5749 	 * valid bit may be set ) have already been zeroed by
5750 	 * vm_page_set_validclean().
5751 	 */
5752 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5753 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5754 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5755 			if (i > b) {
5756 				pmap_zero_page_area(m,
5757 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5758 			}
5759 			b = i + 1;
5760 		}
5761 	}
5762 
5763 	/*
5764 	 * setvalid is TRUE when we can safely set the zero'd areas
5765 	 * as being valid.  We can do this if there are no cache consistency
5766 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5767 	 */
5768 	if (setvalid)
5769 		vm_page_valid(m);
5770 }
5771 
5772 /*
5773  *	vm_page_is_valid:
5774  *
5775  *	Is (partial) page valid?  Note that the case where size == 0
5776  *	will return FALSE in the degenerate case where the page is
5777  *	entirely invalid, and TRUE otherwise.
5778  *
5779  *	Some callers envoke this routine without the busy lock held and
5780  *	handle races via higher level locks.  Typical callers should
5781  *	hold a busy lock to prevent invalidation.
5782  */
5783 int
5784 vm_page_is_valid(vm_page_t m, int base, int size)
5785 {
5786 	vm_page_bits_t bits;
5787 
5788 	bits = vm_page_bits(base, size);
5789 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5790 }
5791 
5792 /*
5793  * Returns true if all of the specified predicates are true for the entire
5794  * (super)page and false otherwise.
5795  */
5796 bool
5797 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5798 {
5799 	vm_object_t object;
5800 	int i, npages;
5801 
5802 	object = m->object;
5803 	if (skip_m != NULL && skip_m->object != object)
5804 		return (false);
5805 	VM_OBJECT_ASSERT_LOCKED(object);
5806 	KASSERT(psind <= m->psind,
5807 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5808 	npages = atop(pagesizes[psind]);
5809 
5810 	/*
5811 	 * The physically contiguous pages that make up a superpage, i.e., a
5812 	 * page with a page size index ("psind") greater than zero, will
5813 	 * occupy adjacent entries in vm_page_array[].
5814 	 */
5815 	for (i = 0; i < npages; i++) {
5816 		/* Always test object consistency, including "skip_m". */
5817 		if (m[i].object != object)
5818 			return (false);
5819 		if (&m[i] == skip_m)
5820 			continue;
5821 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5822 			return (false);
5823 		if ((flags & PS_ALL_DIRTY) != 0) {
5824 			/*
5825 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5826 			 * might stop this case from spuriously returning
5827 			 * "false".  However, that would require a write lock
5828 			 * on the object containing "m[i]".
5829 			 */
5830 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5831 				return (false);
5832 		}
5833 		if ((flags & PS_ALL_VALID) != 0 &&
5834 		    m[i].valid != VM_PAGE_BITS_ALL)
5835 			return (false);
5836 	}
5837 	return (true);
5838 }
5839 
5840 /*
5841  * Set the page's dirty bits if the page is modified.
5842  */
5843 void
5844 vm_page_test_dirty(vm_page_t m)
5845 {
5846 
5847 	vm_page_assert_busied(m);
5848 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5849 		vm_page_dirty(m);
5850 }
5851 
5852 void
5853 vm_page_valid(vm_page_t m)
5854 {
5855 
5856 	vm_page_assert_busied(m);
5857 	if (vm_page_xbusied(m))
5858 		m->valid = VM_PAGE_BITS_ALL;
5859 	else
5860 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5861 }
5862 
5863 void
5864 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5865 {
5866 
5867 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5868 }
5869 
5870 void
5871 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5872 {
5873 
5874 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5875 }
5876 
5877 int
5878 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5879 {
5880 
5881 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5882 }
5883 
5884 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5885 void
5886 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5887 {
5888 
5889 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5890 }
5891 
5892 void
5893 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5894 {
5895 
5896 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5897 }
5898 #endif
5899 
5900 #ifdef INVARIANTS
5901 void
5902 vm_page_object_busy_assert(vm_page_t m)
5903 {
5904 
5905 	/*
5906 	 * Certain of the page's fields may only be modified by the
5907 	 * holder of a page or object busy.
5908 	 */
5909 	if (m->object != NULL && !vm_page_busied(m))
5910 		VM_OBJECT_ASSERT_BUSY(m->object);
5911 }
5912 
5913 void
5914 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5915 {
5916 
5917 	if ((bits & PGA_WRITEABLE) == 0)
5918 		return;
5919 
5920 	/*
5921 	 * The PGA_WRITEABLE flag can only be set if the page is
5922 	 * managed, is exclusively busied or the object is locked.
5923 	 * Currently, this flag is only set by pmap_enter().
5924 	 */
5925 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5926 	    ("PGA_WRITEABLE on unmanaged page"));
5927 	if (!vm_page_xbusied(m))
5928 		VM_OBJECT_ASSERT_BUSY(m->object);
5929 }
5930 #endif
5931 
5932 #include "opt_ddb.h"
5933 #ifdef DDB
5934 #include <sys/kernel.h>
5935 
5936 #include <ddb/ddb.h>
5937 
5938 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5939 {
5940 
5941 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5942 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5943 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5944 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5945 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5946 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5947 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5948 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5949 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5950 }
5951 
5952 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5953 {
5954 	int dom;
5955 
5956 	db_printf("pq_free %d\n", vm_free_count());
5957 	for (dom = 0; dom < vm_ndomains; dom++) {
5958 		db_printf(
5959     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5960 		    dom,
5961 		    vm_dom[dom].vmd_page_count,
5962 		    vm_dom[dom].vmd_free_count,
5963 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5964 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5965 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5966 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5967 	}
5968 }
5969 
5970 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5971 {
5972 	vm_page_t m;
5973 	boolean_t phys, virt;
5974 
5975 	if (!have_addr) {
5976 		db_printf("show pginfo addr\n");
5977 		return;
5978 	}
5979 
5980 	phys = strchr(modif, 'p') != NULL;
5981 	virt = strchr(modif, 'v') != NULL;
5982 	if (virt)
5983 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5984 	else if (phys)
5985 		m = PHYS_TO_VM_PAGE(addr);
5986 	else
5987 		m = (vm_page_t)addr;
5988 	db_printf(
5989     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5990     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5991 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5992 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5993 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5994 }
5995 #endif /* DDB */
5996