xref: /freebsd/sys/vm/vm_page.c (revision 8b51df0d4fadd2cec40b88b43b7aa5aaf95c7510)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Resident memory management module.
65  */
66 
67 #include <sys/cdefs.h>
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/counter.h>
73 #include <sys/domainset.h>
74 #include <sys/kernel.h>
75 #include <sys/limits.h>
76 #include <sys/linker.h>
77 #include <sys/lock.h>
78 #include <sys/malloc.h>
79 #include <sys/mman.h>
80 #include <sys/msgbuf.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/rwlock.h>
84 #include <sys/sleepqueue.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/smp.h>
88 #include <sys/sysctl.h>
89 #include <sys/vmmeter.h>
90 #include <sys/vnode.h>
91 
92 #include <vm/vm.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_param.h>
95 #include <vm/vm_domainset.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_dumpset.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 struct vm_domain vm_dom[MAXMEMDOM];
114 
115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
116 
117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
118 
119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
120 /* The following fields are protected by the domainset lock. */
121 domainset_t __exclusive_cache_line vm_min_domains;
122 domainset_t __exclusive_cache_line vm_severe_domains;
123 static int vm_min_waiters;
124 static int vm_severe_waiters;
125 static int vm_pageproc_waiters;
126 
127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
128     "VM page statistics");
129 
130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
132     CTLFLAG_RD, &pqstate_commit_retries,
133     "Number of failed per-page atomic queue state updates");
134 
135 static COUNTER_U64_DEFINE_EARLY(queue_ops);
136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
137     CTLFLAG_RD, &queue_ops,
138     "Number of batched queue operations");
139 
140 static COUNTER_U64_DEFINE_EARLY(queue_nops);
141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
142     CTLFLAG_RD, &queue_nops,
143     "Number of batched queue operations with no effects");
144 
145 /*
146  * bogus page -- for I/O to/from partially complete buffers,
147  * or for paging into sparsely invalid regions.
148  */
149 vm_page_t bogus_page;
150 
151 vm_page_t vm_page_array;
152 long vm_page_array_size;
153 long first_page;
154 
155 struct bitset *vm_page_dump;
156 long vm_page_dump_pages;
157 
158 static TAILQ_HEAD(, vm_page) blacklist_head;
159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
161     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
162 
163 static uma_zone_t fakepg_zone;
164 
165 static void vm_page_alloc_check(vm_page_t m);
166 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req);
167 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
168     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
169 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
170 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
171 static bool vm_page_free_prep(vm_page_t m);
172 static void vm_page_free_toq(vm_page_t m);
173 static void vm_page_init(void *dummy);
174 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
175     vm_pindex_t pindex, vm_page_t mpred);
176 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
177     vm_page_t mpred);
178 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
179     const uint16_t nflag);
180 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
181     vm_page_t m_run, vm_paddr_t high);
182 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
183 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
184     int req);
185 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
186     int flags);
187 static void vm_page_zone_release(void *arg, void **store, int cnt);
188 
189 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
190 
191 static void
192 vm_page_init(void *dummy)
193 {
194 
195 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
196 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
197 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE);
198 }
199 
200 static int pgcache_zone_max_pcpu;
201 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
202     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
203     "Per-CPU page cache size");
204 
205 /*
206  * The cache page zone is initialized later since we need to be able to allocate
207  * pages before UMA is fully initialized.
208  */
209 static void
210 vm_page_init_cache_zones(void *dummy __unused)
211 {
212 	struct vm_domain *vmd;
213 	struct vm_pgcache *pgcache;
214 	int cache, domain, maxcache, pool;
215 
216 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
217 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
218 	for (domain = 0; domain < vm_ndomains; domain++) {
219 		vmd = VM_DOMAIN(domain);
220 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
221 			pgcache = &vmd->vmd_pgcache[pool];
222 			pgcache->domain = domain;
223 			pgcache->pool = pool;
224 			pgcache->zone = uma_zcache_create("vm pgcache",
225 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
226 			    vm_page_zone_import, vm_page_zone_release, pgcache,
227 			    UMA_ZONE_VM);
228 
229 			/*
230 			 * Limit each pool's zone to 0.1% of the pages in the
231 			 * domain.
232 			 */
233 			cache = maxcache != 0 ? maxcache :
234 			    vmd->vmd_page_count / 1000;
235 			uma_zone_set_maxcache(pgcache->zone, cache);
236 		}
237 	}
238 }
239 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
240 
241 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
242 #if PAGE_SIZE == 32768
243 #ifdef CTASSERT
244 CTASSERT(sizeof(u_long) >= 8);
245 #endif
246 #endif
247 
248 /*
249  *	vm_set_page_size:
250  *
251  *	Sets the page size, perhaps based upon the memory
252  *	size.  Must be called before any use of page-size
253  *	dependent functions.
254  */
255 void
256 vm_set_page_size(void)
257 {
258 	if (vm_cnt.v_page_size == 0)
259 		vm_cnt.v_page_size = PAGE_SIZE;
260 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
261 		panic("vm_set_page_size: page size not a power of two");
262 }
263 
264 /*
265  *	vm_page_blacklist_next:
266  *
267  *	Find the next entry in the provided string of blacklist
268  *	addresses.  Entries are separated by space, comma, or newline.
269  *	If an invalid integer is encountered then the rest of the
270  *	string is skipped.  Updates the list pointer to the next
271  *	character, or NULL if the string is exhausted or invalid.
272  */
273 static vm_paddr_t
274 vm_page_blacklist_next(char **list, char *end)
275 {
276 	vm_paddr_t bad;
277 	char *cp, *pos;
278 
279 	if (list == NULL || *list == NULL)
280 		return (0);
281 	if (**list =='\0') {
282 		*list = NULL;
283 		return (0);
284 	}
285 
286 	/*
287 	 * If there's no end pointer then the buffer is coming from
288 	 * the kenv and we know it's null-terminated.
289 	 */
290 	if (end == NULL)
291 		end = *list + strlen(*list);
292 
293 	/* Ensure that strtoq() won't walk off the end */
294 	if (*end != '\0') {
295 		if (*end == '\n' || *end == ' ' || *end  == ',')
296 			*end = '\0';
297 		else {
298 			printf("Blacklist not terminated, skipping\n");
299 			*list = NULL;
300 			return (0);
301 		}
302 	}
303 
304 	for (pos = *list; *pos != '\0'; pos = cp) {
305 		bad = strtoq(pos, &cp, 0);
306 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
307 			if (bad == 0) {
308 				if (++cp < end)
309 					continue;
310 				else
311 					break;
312 			}
313 		} else
314 			break;
315 		if (*cp == '\0' || ++cp >= end)
316 			*list = NULL;
317 		else
318 			*list = cp;
319 		return (trunc_page(bad));
320 	}
321 	printf("Garbage in RAM blacklist, skipping\n");
322 	*list = NULL;
323 	return (0);
324 }
325 
326 bool
327 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
328 {
329 	struct vm_domain *vmd;
330 	vm_page_t m;
331 	bool found;
332 
333 	m = vm_phys_paddr_to_vm_page(pa);
334 	if (m == NULL)
335 		return (true); /* page does not exist, no failure */
336 
337 	vmd = VM_DOMAIN(vm_phys_domain(pa));
338 	vm_domain_free_lock(vmd);
339 	found = vm_phys_unfree_page(pa);
340 	vm_domain_free_unlock(vmd);
341 	if (found) {
342 		vm_domain_freecnt_inc(vmd, -1);
343 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
344 		if (verbose)
345 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
346 	}
347 	return (found);
348 }
349 
350 /*
351  *	vm_page_blacklist_check:
352  *
353  *	Iterate through the provided string of blacklist addresses, pulling
354  *	each entry out of the physical allocator free list and putting it
355  *	onto a list for reporting via the vm.page_blacklist sysctl.
356  */
357 static void
358 vm_page_blacklist_check(char *list, char *end)
359 {
360 	vm_paddr_t pa;
361 	char *next;
362 
363 	next = list;
364 	while (next != NULL) {
365 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
366 			continue;
367 		vm_page_blacklist_add(pa, bootverbose);
368 	}
369 }
370 
371 /*
372  *	vm_page_blacklist_load:
373  *
374  *	Search for a special module named "ram_blacklist".  It'll be a
375  *	plain text file provided by the user via the loader directive
376  *	of the same name.
377  */
378 static void
379 vm_page_blacklist_load(char **list, char **end)
380 {
381 	void *mod;
382 	u_char *ptr;
383 	u_int len;
384 
385 	mod = NULL;
386 	ptr = NULL;
387 
388 	mod = preload_search_by_type("ram_blacklist");
389 	if (mod != NULL) {
390 		ptr = preload_fetch_addr(mod);
391 		len = preload_fetch_size(mod);
392         }
393 	*list = ptr;
394 	if (ptr != NULL)
395 		*end = ptr + len;
396 	else
397 		*end = NULL;
398 	return;
399 }
400 
401 static int
402 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
403 {
404 	vm_page_t m;
405 	struct sbuf sbuf;
406 	int error, first;
407 
408 	first = 1;
409 	error = sysctl_wire_old_buffer(req, 0);
410 	if (error != 0)
411 		return (error);
412 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
413 	TAILQ_FOREACH(m, &blacklist_head, listq) {
414 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
415 		    (uintmax_t)m->phys_addr);
416 		first = 0;
417 	}
418 	error = sbuf_finish(&sbuf);
419 	sbuf_delete(&sbuf);
420 	return (error);
421 }
422 
423 /*
424  * Initialize a dummy page for use in scans of the specified paging queue.
425  * In principle, this function only needs to set the flag PG_MARKER.
426  * Nonetheless, it write busies the page as a safety precaution.
427  */
428 void
429 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
430 {
431 
432 	bzero(marker, sizeof(*marker));
433 	marker->flags = PG_MARKER;
434 	marker->a.flags = aflags;
435 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
436 	marker->a.queue = queue;
437 }
438 
439 static void
440 vm_page_domain_init(int domain)
441 {
442 	struct vm_domain *vmd;
443 	struct vm_pagequeue *pq;
444 	int i;
445 
446 	vmd = VM_DOMAIN(domain);
447 	bzero(vmd, sizeof(*vmd));
448 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
449 	    "vm inactive pagequeue";
450 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
451 	    "vm active pagequeue";
452 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
453 	    "vm laundry pagequeue";
454 	*__DECONST(const char **,
455 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
456 	    "vm unswappable pagequeue";
457 	vmd->vmd_domain = domain;
458 	vmd->vmd_page_count = 0;
459 	vmd->vmd_free_count = 0;
460 	vmd->vmd_segs = 0;
461 	vmd->vmd_oom = FALSE;
462 	for (i = 0; i < PQ_COUNT; i++) {
463 		pq = &vmd->vmd_pagequeues[i];
464 		TAILQ_INIT(&pq->pq_pl);
465 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
466 		    MTX_DEF | MTX_DUPOK);
467 		pq->pq_pdpages = 0;
468 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
469 	}
470 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
471 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
472 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
473 
474 	/*
475 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
476 	 * insertions.
477 	 */
478 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
479 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
480 	    &vmd->vmd_inacthead, plinks.q);
481 
482 	/*
483 	 * The clock pages are used to implement active queue scanning without
484 	 * requeues.  Scans start at clock[0], which is advanced after the scan
485 	 * ends.  When the two clock hands meet, they are reset and scanning
486 	 * resumes from the head of the queue.
487 	 */
488 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
489 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
490 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
491 	    &vmd->vmd_clock[0], plinks.q);
492 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
493 	    &vmd->vmd_clock[1], plinks.q);
494 }
495 
496 /*
497  * Initialize a physical page in preparation for adding it to the free
498  * lists.
499  */
500 void
501 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool)
502 {
503 	m->object = NULL;
504 	m->ref_count = 0;
505 	m->busy_lock = VPB_FREED;
506 	m->flags = m->a.flags = 0;
507 	m->phys_addr = pa;
508 	m->a.queue = PQ_NONE;
509 	m->psind = 0;
510 	m->segind = segind;
511 	m->order = VM_NFREEORDER;
512 	m->pool = pool;
513 	m->valid = m->dirty = 0;
514 	pmap_page_init(m);
515 }
516 
517 #ifndef PMAP_HAS_PAGE_ARRAY
518 static vm_paddr_t
519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
520 {
521 	vm_paddr_t new_end;
522 
523 	/*
524 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
525 	 * However, because this page is allocated from KVM, out-of-bounds
526 	 * accesses using the direct map will not be trapped.
527 	 */
528 	*vaddr += PAGE_SIZE;
529 
530 	/*
531 	 * Allocate physical memory for the page structures, and map it.
532 	 */
533 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
534 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
535 	    VM_PROT_READ | VM_PROT_WRITE);
536 	vm_page_array_size = page_range;
537 
538 	return (new_end);
539 }
540 #endif
541 
542 /*
543  *	vm_page_startup:
544  *
545  *	Initializes the resident memory module.  Allocates physical memory for
546  *	bootstrapping UMA and some data structures that are used to manage
547  *	physical pages.  Initializes these structures, and populates the free
548  *	page queues.
549  */
550 vm_offset_t
551 vm_page_startup(vm_offset_t vaddr)
552 {
553 	struct vm_phys_seg *seg;
554 	struct vm_domain *vmd;
555 	vm_page_t m;
556 	char *list, *listend;
557 	vm_paddr_t end, high_avail, low_avail, new_end, size;
558 	vm_paddr_t page_range __unused;
559 	vm_paddr_t last_pa, pa, startp, endp;
560 	u_long pagecount;
561 #if MINIDUMP_PAGE_TRACKING
562 	u_long vm_page_dump_size;
563 #endif
564 	int biggestone, i, segind;
565 #ifdef WITNESS
566 	vm_offset_t mapped;
567 	int witness_size;
568 #endif
569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
570 	long ii;
571 #endif
572 #ifdef VM_FREEPOOL_LAZYINIT
573 	int lazyinit;
574 #endif
575 
576 	vaddr = round_page(vaddr);
577 
578 	vm_phys_early_startup();
579 	biggestone = vm_phys_avail_largest();
580 	end = phys_avail[biggestone+1];
581 
582 	/*
583 	 * Initialize the page and queue locks.
584 	 */
585 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
586 	for (i = 0; i < PA_LOCK_COUNT; i++)
587 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
588 	for (i = 0; i < vm_ndomains; i++)
589 		vm_page_domain_init(i);
590 
591 	new_end = end;
592 #ifdef WITNESS
593 	witness_size = round_page(witness_startup_count());
594 	new_end -= witness_size;
595 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
596 	    VM_PROT_READ | VM_PROT_WRITE);
597 	bzero((void *)mapped, witness_size);
598 	witness_startup((void *)mapped);
599 #endif
600 
601 #if MINIDUMP_PAGE_TRACKING
602 	/*
603 	 * Allocate a bitmap to indicate that a random physical page
604 	 * needs to be included in a minidump.
605 	 *
606 	 * The amd64 port needs this to indicate which direct map pages
607 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
608 	 *
609 	 * However, i386 still needs this workspace internally within the
610 	 * minidump code.  In theory, they are not needed on i386, but are
611 	 * included should the sf_buf code decide to use them.
612 	 */
613 	last_pa = 0;
614 	vm_page_dump_pages = 0;
615 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
616 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
617 		    dump_avail[i] / PAGE_SIZE;
618 		if (dump_avail[i + 1] > last_pa)
619 			last_pa = dump_avail[i + 1];
620 	}
621 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
622 	new_end -= vm_page_dump_size;
623 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
624 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
625 	bzero((void *)vm_page_dump, vm_page_dump_size);
626 #if MINIDUMP_STARTUP_PAGE_TRACKING
627 	/*
628 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
629 	 * in a crash dump.  When pmap_map() uses the direct map, they are
630 	 * not automatically included.
631 	 */
632 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
633 		dump_add_page(pa);
634 #endif
635 #else
636 	(void)last_pa;
637 #endif
638 	phys_avail[biggestone + 1] = new_end;
639 #ifdef __amd64__
640 	/*
641 	 * Request that the physical pages underlying the message buffer be
642 	 * included in a crash dump.  Since the message buffer is accessed
643 	 * through the direct map, they are not automatically included.
644 	 */
645 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
646 	last_pa = pa + round_page(msgbufsize);
647 	while (pa < last_pa) {
648 		dump_add_page(pa);
649 		pa += PAGE_SIZE;
650 	}
651 #endif
652 	/*
653 	 * Compute the number of pages of memory that will be available for
654 	 * use, taking into account the overhead of a page structure per page.
655 	 * In other words, solve
656 	 *	"available physical memory" - round_page(page_range *
657 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
658 	 * for page_range.
659 	 */
660 	low_avail = phys_avail[0];
661 	high_avail = phys_avail[1];
662 	for (i = 0; i < vm_phys_nsegs; i++) {
663 		if (vm_phys_segs[i].start < low_avail)
664 			low_avail = vm_phys_segs[i].start;
665 		if (vm_phys_segs[i].end > high_avail)
666 			high_avail = vm_phys_segs[i].end;
667 	}
668 	/* Skip the first chunk.  It is already accounted for. */
669 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
670 		if (phys_avail[i] < low_avail)
671 			low_avail = phys_avail[i];
672 		if (phys_avail[i + 1] > high_avail)
673 			high_avail = phys_avail[i + 1];
674 	}
675 	first_page = low_avail / PAGE_SIZE;
676 #ifdef VM_PHYSSEG_SPARSE
677 	size = 0;
678 	for (i = 0; i < vm_phys_nsegs; i++)
679 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
680 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
681 		size += phys_avail[i + 1] - phys_avail[i];
682 #elif defined(VM_PHYSSEG_DENSE)
683 	size = high_avail - low_avail;
684 #else
685 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
686 #endif
687 
688 #ifdef PMAP_HAS_PAGE_ARRAY
689 	pmap_page_array_startup(size / PAGE_SIZE);
690 	biggestone = vm_phys_avail_largest();
691 	end = new_end = phys_avail[biggestone + 1];
692 #else
693 #ifdef VM_PHYSSEG_DENSE
694 	/*
695 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
696 	 * the overhead of a page structure per page only if vm_page_array is
697 	 * allocated from the last physical memory chunk.  Otherwise, we must
698 	 * allocate page structures representing the physical memory
699 	 * underlying vm_page_array, even though they will not be used.
700 	 */
701 	if (new_end != high_avail)
702 		page_range = size / PAGE_SIZE;
703 	else
704 #endif
705 	{
706 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
707 
708 		/*
709 		 * If the partial bytes remaining are large enough for
710 		 * a page (PAGE_SIZE) without a corresponding
711 		 * 'struct vm_page', then new_end will contain an
712 		 * extra page after subtracting the length of the VM
713 		 * page array.  Compensate by subtracting an extra
714 		 * page from new_end.
715 		 */
716 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
717 			if (new_end == high_avail)
718 				high_avail -= PAGE_SIZE;
719 			new_end -= PAGE_SIZE;
720 		}
721 	}
722 	end = new_end;
723 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
724 #endif
725 
726 #if VM_NRESERVLEVEL > 0
727 	/*
728 	 * Allocate physical memory for the reservation management system's
729 	 * data structures, and map it.
730 	 */
731 	new_end = vm_reserv_startup(&vaddr, new_end);
732 #endif
733 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
734 	/*
735 	 * Include vm_page_array and vm_reserv_array in a crash dump.
736 	 */
737 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
738 		dump_add_page(pa);
739 #endif
740 	phys_avail[biggestone + 1] = new_end;
741 
742 	/*
743 	 * Add physical memory segments corresponding to the available
744 	 * physical pages.
745 	 */
746 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
747 		if (vm_phys_avail_size(i) != 0)
748 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
749 
750 	/*
751 	 * Initialize the physical memory allocator.
752 	 */
753 	vm_phys_init();
754 
755 #ifdef VM_FREEPOOL_LAZYINIT
756 	lazyinit = 1;
757 	TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit);
758 #endif
759 
760 	/*
761 	 * Initialize the page structures and add every available page to the
762 	 * physical memory allocator's free lists.
763 	 */
764 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
765 	for (ii = 0; ii < vm_page_array_size; ii++) {
766 		m = &vm_page_array[ii];
767 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0,
768 		    VM_FREEPOOL_DEFAULT);
769 		m->flags = PG_FICTITIOUS;
770 	}
771 #endif
772 	vm_cnt.v_page_count = 0;
773 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
774 		seg = &vm_phys_segs[segind];
775 
776 		/*
777 		 * If lazy vm_page initialization is not enabled, simply
778 		 * initialize all of the pages in the segment.  Otherwise, we
779 		 * only initialize:
780 		 * 1. Pages not covered by phys_avail[], since they might be
781 		 *    freed to the allocator at some future point, e.g., by
782 		 *    kmem_bootstrap_free().
783 		 * 2. The first page of each run of free pages handed to the
784 		 *    vm_phys allocator, which in turn defers initialization
785 		 *    of pages until they are needed.
786 		 * This avoids blocking the boot process for long periods, which
787 		 * may be relevant for VMs (which ought to boot as quickly as
788 		 * possible) and/or systems with large amounts of physical
789 		 * memory.
790 		 */
791 #ifdef VM_FREEPOOL_LAZYINIT
792 		if (lazyinit) {
793 			startp = seg->start;
794 			for (i = 0; phys_avail[i + 1] != 0; i += 2) {
795 				if (startp >= seg->end)
796 					break;
797 
798 				if (phys_avail[i + 1] < startp)
799 					continue;
800 				if (phys_avail[i] <= startp) {
801 					startp = phys_avail[i + 1];
802 					continue;
803 				}
804 
805 				m = vm_phys_seg_paddr_to_vm_page(seg, startp);
806 				for (endp = MIN(phys_avail[i], seg->end);
807 				    startp < endp; startp += PAGE_SIZE, m++) {
808 					vm_page_init_page(m, startp, segind,
809 					    VM_FREEPOOL_DEFAULT);
810 				}
811 			}
812 		} else
813 #endif
814 			for (m = seg->first_page, pa = seg->start;
815 			    pa < seg->end; m++, pa += PAGE_SIZE) {
816 				vm_page_init_page(m, pa, segind,
817 				    VM_FREEPOOL_DEFAULT);
818 			}
819 
820 		/*
821 		 * Add the segment's pages that are covered by one of
822 		 * phys_avail's ranges to the free lists.
823 		 */
824 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
825 			if (seg->end <= phys_avail[i] ||
826 			    seg->start >= phys_avail[i + 1])
827 				continue;
828 
829 			startp = MAX(seg->start, phys_avail[i]);
830 			endp = MIN(seg->end, phys_avail[i + 1]);
831 			pagecount = (u_long)atop(endp - startp);
832 			if (pagecount == 0)
833 				continue;
834 
835 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
836 #ifdef VM_FREEPOOL_LAZYINIT
837 			if (lazyinit) {
838 				vm_page_init_page(m, startp, segind,
839 				    VM_FREEPOOL_LAZYINIT);
840 			}
841 #endif
842 			vmd = VM_DOMAIN(seg->domain);
843 			vm_domain_free_lock(vmd);
844 			vm_phys_enqueue_contig(m, pagecount);
845 			vm_domain_free_unlock(vmd);
846 			vm_domain_freecnt_inc(vmd, pagecount);
847 			vm_cnt.v_page_count += (u_int)pagecount;
848 			vmd->vmd_page_count += (u_int)pagecount;
849 			vmd->vmd_segs |= 1UL << segind;
850 		}
851 	}
852 
853 	/*
854 	 * Remove blacklisted pages from the physical memory allocator.
855 	 */
856 	TAILQ_INIT(&blacklist_head);
857 	vm_page_blacklist_load(&list, &listend);
858 	vm_page_blacklist_check(list, listend);
859 
860 	list = kern_getenv("vm.blacklist");
861 	vm_page_blacklist_check(list, NULL);
862 
863 	freeenv(list);
864 #if VM_NRESERVLEVEL > 0
865 	/*
866 	 * Initialize the reservation management system.
867 	 */
868 	vm_reserv_init();
869 #endif
870 
871 	return (vaddr);
872 }
873 
874 void
875 vm_page_reference(vm_page_t m)
876 {
877 
878 	vm_page_aflag_set(m, PGA_REFERENCED);
879 }
880 
881 /*
882  *	vm_page_trybusy
883  *
884  *	Helper routine for grab functions to trylock busy.
885  *
886  *	Returns true on success and false on failure.
887  */
888 static bool
889 vm_page_trybusy(vm_page_t m, int allocflags)
890 {
891 
892 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
893 		return (vm_page_trysbusy(m));
894 	else
895 		return (vm_page_tryxbusy(m));
896 }
897 
898 /*
899  *	vm_page_tryacquire
900  *
901  *	Helper routine for grab functions to trylock busy and wire.
902  *
903  *	Returns true on success and false on failure.
904  */
905 static inline bool
906 vm_page_tryacquire(vm_page_t m, int allocflags)
907 {
908 	bool locked;
909 
910 	locked = vm_page_trybusy(m, allocflags);
911 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
912 		vm_page_wire(m);
913 	return (locked);
914 }
915 
916 /*
917  *	vm_page_busy_acquire:
918  *
919  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
920  *	and drop the object lock if necessary.
921  */
922 bool
923 vm_page_busy_acquire(vm_page_t m, int allocflags)
924 {
925 	vm_object_t obj;
926 	bool locked;
927 
928 	/*
929 	 * The page-specific object must be cached because page
930 	 * identity can change during the sleep, causing the
931 	 * re-lock of a different object.
932 	 * It is assumed that a reference to the object is already
933 	 * held by the callers.
934 	 */
935 	obj = atomic_load_ptr(&m->object);
936 	for (;;) {
937 		if (vm_page_tryacquire(m, allocflags))
938 			return (true);
939 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
940 			return (false);
941 		if (obj != NULL)
942 			locked = VM_OBJECT_WOWNED(obj);
943 		else
944 			locked = false;
945 		MPASS(locked || vm_page_wired(m));
946 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
947 		    locked) && locked)
948 			VM_OBJECT_WLOCK(obj);
949 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
950 			return (false);
951 		KASSERT(m->object == obj || m->object == NULL,
952 		    ("vm_page_busy_acquire: page %p does not belong to %p",
953 		    m, obj));
954 	}
955 }
956 
957 /*
958  *	vm_page_busy_downgrade:
959  *
960  *	Downgrade an exclusive busy page into a single shared busy page.
961  */
962 void
963 vm_page_busy_downgrade(vm_page_t m)
964 {
965 	u_int x;
966 
967 	vm_page_assert_xbusied(m);
968 
969 	x = vm_page_busy_fetch(m);
970 	for (;;) {
971 		if (atomic_fcmpset_rel_int(&m->busy_lock,
972 		    &x, VPB_SHARERS_WORD(1)))
973 			break;
974 	}
975 	if ((x & VPB_BIT_WAITERS) != 0)
976 		wakeup(m);
977 }
978 
979 /*
980  *
981  *	vm_page_busy_tryupgrade:
982  *
983  *	Attempt to upgrade a single shared busy into an exclusive busy.
984  */
985 int
986 vm_page_busy_tryupgrade(vm_page_t m)
987 {
988 	u_int ce, x;
989 
990 	vm_page_assert_sbusied(m);
991 
992 	x = vm_page_busy_fetch(m);
993 	ce = VPB_CURTHREAD_EXCLUSIVE;
994 	for (;;) {
995 		if (VPB_SHARERS(x) > 1)
996 			return (0);
997 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
998 		    ("vm_page_busy_tryupgrade: invalid lock state"));
999 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1000 		    ce | (x & VPB_BIT_WAITERS)))
1001 			continue;
1002 		return (1);
1003 	}
1004 }
1005 
1006 /*
1007  *	vm_page_sbusied:
1008  *
1009  *	Return a positive value if the page is shared busied, 0 otherwise.
1010  */
1011 int
1012 vm_page_sbusied(vm_page_t m)
1013 {
1014 	u_int x;
1015 
1016 	x = vm_page_busy_fetch(m);
1017 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1018 }
1019 
1020 /*
1021  *	vm_page_sunbusy:
1022  *
1023  *	Shared unbusy a page.
1024  */
1025 void
1026 vm_page_sunbusy(vm_page_t m)
1027 {
1028 	u_int x;
1029 
1030 	vm_page_assert_sbusied(m);
1031 
1032 	x = vm_page_busy_fetch(m);
1033 	for (;;) {
1034 		KASSERT(x != VPB_FREED,
1035 		    ("vm_page_sunbusy: Unlocking freed page."));
1036 		if (VPB_SHARERS(x) > 1) {
1037 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1038 			    x - VPB_ONE_SHARER))
1039 				break;
1040 			continue;
1041 		}
1042 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1043 		    ("vm_page_sunbusy: invalid lock state"));
1044 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1045 			continue;
1046 		if ((x & VPB_BIT_WAITERS) == 0)
1047 			break;
1048 		wakeup(m);
1049 		break;
1050 	}
1051 }
1052 
1053 /*
1054  *	vm_page_busy_sleep:
1055  *
1056  *	Sleep if the page is busy, using the page pointer as wchan.
1057  *	This is used to implement the hard-path of the busying mechanism.
1058  *
1059  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1060  *	will not sleep if the page is shared-busy.
1061  *
1062  *	The object lock must be held on entry.
1063  *
1064  *	Returns true if it slept and dropped the object lock, or false
1065  *	if there was no sleep and the lock is still held.
1066  */
1067 bool
1068 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1069 {
1070 	vm_object_t obj;
1071 
1072 	obj = m->object;
1073 	VM_OBJECT_ASSERT_LOCKED(obj);
1074 
1075 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1076 	    true));
1077 }
1078 
1079 /*
1080  *	vm_page_busy_sleep_unlocked:
1081  *
1082  *	Sleep if the page is busy, using the page pointer as wchan.
1083  *	This is used to implement the hard-path of busying mechanism.
1084  *
1085  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1086  *	will not sleep if the page is shared-busy.
1087  *
1088  *	The object lock must not be held on entry.  The operation will
1089  *	return if the page changes identity.
1090  */
1091 void
1092 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1093     const char *wmesg, int allocflags)
1094 {
1095 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1096 
1097 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1098 }
1099 
1100 /*
1101  *	_vm_page_busy_sleep:
1102  *
1103  *	Internal busy sleep function.  Verifies the page identity and
1104  *	lockstate against parameters.  Returns true if it sleeps and
1105  *	false otherwise.
1106  *
1107  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1108  *
1109  *	If locked is true the lock will be dropped for any true returns
1110  *	and held for any false returns.
1111  */
1112 static bool
1113 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1114     const char *wmesg, int allocflags, bool locked)
1115 {
1116 	bool xsleep;
1117 	u_int x;
1118 
1119 	/*
1120 	 * If the object is busy we must wait for that to drain to zero
1121 	 * before trying the page again.
1122 	 */
1123 	if (obj != NULL && vm_object_busied(obj)) {
1124 		if (locked)
1125 			VM_OBJECT_DROP(obj);
1126 		vm_object_busy_wait(obj, wmesg);
1127 		return (true);
1128 	}
1129 
1130 	if (!vm_page_busied(m))
1131 		return (false);
1132 
1133 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1134 	sleepq_lock(m);
1135 	x = vm_page_busy_fetch(m);
1136 	do {
1137 		/*
1138 		 * If the page changes objects or becomes unlocked we can
1139 		 * simply return.
1140 		 */
1141 		if (x == VPB_UNBUSIED ||
1142 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1143 		    m->object != obj || m->pindex != pindex) {
1144 			sleepq_release(m);
1145 			return (false);
1146 		}
1147 		if ((x & VPB_BIT_WAITERS) != 0)
1148 			break;
1149 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1150 	if (locked)
1151 		VM_OBJECT_DROP(obj);
1152 	DROP_GIANT();
1153 	sleepq_add(m, NULL, wmesg, 0, 0);
1154 	sleepq_wait(m, PVM);
1155 	PICKUP_GIANT();
1156 	return (true);
1157 }
1158 
1159 /*
1160  *	vm_page_trysbusy:
1161  *
1162  *	Try to shared busy a page.
1163  *	If the operation succeeds 1 is returned otherwise 0.
1164  *	The operation never sleeps.
1165  */
1166 int
1167 vm_page_trysbusy(vm_page_t m)
1168 {
1169 	vm_object_t obj;
1170 	u_int x;
1171 
1172 	obj = m->object;
1173 	x = vm_page_busy_fetch(m);
1174 	for (;;) {
1175 		if ((x & VPB_BIT_SHARED) == 0)
1176 			return (0);
1177 		/*
1178 		 * Reduce the window for transient busies that will trigger
1179 		 * false negatives in vm_page_ps_test().
1180 		 */
1181 		if (obj != NULL && vm_object_busied(obj))
1182 			return (0);
1183 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1184 		    x + VPB_ONE_SHARER))
1185 			break;
1186 	}
1187 
1188 	/* Refetch the object now that we're guaranteed that it is stable. */
1189 	obj = m->object;
1190 	if (obj != NULL && vm_object_busied(obj)) {
1191 		vm_page_sunbusy(m);
1192 		return (0);
1193 	}
1194 	return (1);
1195 }
1196 
1197 /*
1198  *	vm_page_tryxbusy:
1199  *
1200  *	Try to exclusive busy a page.
1201  *	If the operation succeeds 1 is returned otherwise 0.
1202  *	The operation never sleeps.
1203  */
1204 int
1205 vm_page_tryxbusy(vm_page_t m)
1206 {
1207 	vm_object_t obj;
1208 
1209         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1210             VPB_CURTHREAD_EXCLUSIVE) == 0)
1211 		return (0);
1212 
1213 	obj = m->object;
1214 	if (obj != NULL && vm_object_busied(obj)) {
1215 		vm_page_xunbusy(m);
1216 		return (0);
1217 	}
1218 	return (1);
1219 }
1220 
1221 static void
1222 vm_page_xunbusy_hard_tail(vm_page_t m)
1223 {
1224 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1225 	/* Wake the waiter. */
1226 	wakeup(m);
1227 }
1228 
1229 /*
1230  *	vm_page_xunbusy_hard:
1231  *
1232  *	Called when unbusy has failed because there is a waiter.
1233  */
1234 void
1235 vm_page_xunbusy_hard(vm_page_t m)
1236 {
1237 	vm_page_assert_xbusied(m);
1238 	vm_page_xunbusy_hard_tail(m);
1239 }
1240 
1241 void
1242 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1243 {
1244 	vm_page_assert_xbusied_unchecked(m);
1245 	vm_page_xunbusy_hard_tail(m);
1246 }
1247 
1248 static void
1249 vm_page_busy_free(vm_page_t m)
1250 {
1251 	u_int x;
1252 
1253 	atomic_thread_fence_rel();
1254 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1255 	if ((x & VPB_BIT_WAITERS) != 0)
1256 		wakeup(m);
1257 }
1258 
1259 /*
1260  *	vm_page_unhold_pages:
1261  *
1262  *	Unhold each of the pages that is referenced by the given array.
1263  */
1264 void
1265 vm_page_unhold_pages(vm_page_t *ma, int count)
1266 {
1267 
1268 	for (; count != 0; count--) {
1269 		vm_page_unwire(*ma, PQ_ACTIVE);
1270 		ma++;
1271 	}
1272 }
1273 
1274 vm_page_t
1275 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1276 {
1277 	vm_page_t m;
1278 
1279 #ifdef VM_PHYSSEG_SPARSE
1280 	m = vm_phys_paddr_to_vm_page(pa);
1281 	if (m == NULL)
1282 		m = vm_phys_fictitious_to_vm_page(pa);
1283 	return (m);
1284 #elif defined(VM_PHYSSEG_DENSE)
1285 	long pi;
1286 
1287 	pi = atop(pa);
1288 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1289 		m = &vm_page_array[pi - first_page];
1290 		return (m);
1291 	}
1292 	return (vm_phys_fictitious_to_vm_page(pa));
1293 #else
1294 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1295 #endif
1296 }
1297 
1298 /*
1299  *	vm_page_getfake:
1300  *
1301  *	Create a fictitious page with the specified physical address and
1302  *	memory attribute.  The memory attribute is the only the machine-
1303  *	dependent aspect of a fictitious page that must be initialized.
1304  */
1305 vm_page_t
1306 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1307 {
1308 	vm_page_t m;
1309 
1310 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1311 	vm_page_initfake(m, paddr, memattr);
1312 	return (m);
1313 }
1314 
1315 void
1316 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1317 {
1318 
1319 	if ((m->flags & PG_FICTITIOUS) != 0) {
1320 		/*
1321 		 * The page's memattr might have changed since the
1322 		 * previous initialization.  Update the pmap to the
1323 		 * new memattr.
1324 		 */
1325 		goto memattr;
1326 	}
1327 	m->phys_addr = paddr;
1328 	m->a.queue = PQ_NONE;
1329 	/* Fictitious pages don't use "segind". */
1330 	m->flags = PG_FICTITIOUS;
1331 	/* Fictitious pages don't use "order" or "pool". */
1332 	m->oflags = VPO_UNMANAGED;
1333 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1334 	/* Fictitious pages are unevictable. */
1335 	m->ref_count = 1;
1336 	pmap_page_init(m);
1337 memattr:
1338 	pmap_page_set_memattr(m, memattr);
1339 }
1340 
1341 /*
1342  *	vm_page_putfake:
1343  *
1344  *	Release a fictitious page.
1345  */
1346 void
1347 vm_page_putfake(vm_page_t m)
1348 {
1349 
1350 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1351 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1352 	    ("vm_page_putfake: bad page %p", m));
1353 	vm_page_assert_xbusied(m);
1354 	vm_page_busy_free(m);
1355 	uma_zfree(fakepg_zone, m);
1356 }
1357 
1358 /*
1359  *	vm_page_updatefake:
1360  *
1361  *	Update the given fictitious page to the specified physical address and
1362  *	memory attribute.
1363  */
1364 void
1365 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1366 {
1367 
1368 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1369 	    ("vm_page_updatefake: bad page %p", m));
1370 	m->phys_addr = paddr;
1371 	pmap_page_set_memattr(m, memattr);
1372 }
1373 
1374 /*
1375  *	vm_page_free:
1376  *
1377  *	Free a page.
1378  */
1379 void
1380 vm_page_free(vm_page_t m)
1381 {
1382 
1383 	m->flags &= ~PG_ZERO;
1384 	vm_page_free_toq(m);
1385 }
1386 
1387 /*
1388  *	vm_page_free_zero:
1389  *
1390  *	Free a page to the zerod-pages queue
1391  */
1392 void
1393 vm_page_free_zero(vm_page_t m)
1394 {
1395 
1396 	m->flags |= PG_ZERO;
1397 	vm_page_free_toq(m);
1398 }
1399 
1400 /*
1401  * Unbusy and handle the page queueing for a page from a getpages request that
1402  * was optionally read ahead or behind.
1403  */
1404 void
1405 vm_page_readahead_finish(vm_page_t m)
1406 {
1407 
1408 	/* We shouldn't put invalid pages on queues. */
1409 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1410 
1411 	/*
1412 	 * Since the page is not the actually needed one, whether it should
1413 	 * be activated or deactivated is not obvious.  Empirical results
1414 	 * have shown that deactivating the page is usually the best choice,
1415 	 * unless the page is wanted by another thread.
1416 	 */
1417 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1418 		vm_page_activate(m);
1419 	else
1420 		vm_page_deactivate(m);
1421 	vm_page_xunbusy_unchecked(m);
1422 }
1423 
1424 /*
1425  * Destroy the identity of an invalid page and free it if possible.
1426  * This is intended to be used when reading a page from backing store fails.
1427  */
1428 void
1429 vm_page_free_invalid(vm_page_t m)
1430 {
1431 
1432 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1433 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1434 	KASSERT(m->object != NULL, ("page %p has no object", m));
1435 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1436 
1437 	/*
1438 	 * We may be attempting to free the page as part of the handling for an
1439 	 * I/O error, in which case the page was xbusied by a different thread.
1440 	 */
1441 	vm_page_xbusy_claim(m);
1442 
1443 	/*
1444 	 * If someone has wired this page while the object lock
1445 	 * was not held, then the thread that unwires is responsible
1446 	 * for freeing the page.  Otherwise just free the page now.
1447 	 * The wire count of this unmapped page cannot change while
1448 	 * we have the page xbusy and the page's object wlocked.
1449 	 */
1450 	if (vm_page_remove(m))
1451 		vm_page_free(m);
1452 }
1453 
1454 /*
1455  *	vm_page_dirty_KBI:		[ internal use only ]
1456  *
1457  *	Set all bits in the page's dirty field.
1458  *
1459  *	The object containing the specified page must be locked if the
1460  *	call is made from the machine-independent layer.
1461  *
1462  *	See vm_page_clear_dirty_mask().
1463  *
1464  *	This function should only be called by vm_page_dirty().
1465  */
1466 void
1467 vm_page_dirty_KBI(vm_page_t m)
1468 {
1469 
1470 	/* Refer to this operation by its public name. */
1471 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1472 	m->dirty = VM_PAGE_BITS_ALL;
1473 }
1474 
1475 /*
1476  * Insert the given page into the given object at the given pindex.  mpred is
1477  * used for memq linkage.  From vm_page_insert, lookup is true, mpred is
1478  * initially NULL, and this procedure looks it up.  From vm_page_insert_after
1479  * and vm_page_iter_insert, lookup is false and mpred is known to the caller
1480  * to be valid, and may be NULL if this will be the page with the lowest
1481  * pindex.
1482  *
1483  * The procedure is marked __always_inline to suggest to the compiler to
1484  * eliminate the lookup parameter and the associated alternate branch.
1485  */
1486 static __always_inline int
1487 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1488     struct pctrie_iter *pages, bool iter, vm_page_t mpred, bool lookup)
1489 {
1490 	int error;
1491 
1492 	VM_OBJECT_ASSERT_WLOCKED(object);
1493 	KASSERT(m->object == NULL,
1494 	    ("vm_page_insert: page %p already inserted", m));
1495 
1496 	/*
1497 	 * Record the object/offset pair in this page.
1498 	 */
1499 	m->object = object;
1500 	m->pindex = pindex;
1501 	m->ref_count |= VPRC_OBJREF;
1502 
1503 	/*
1504 	 * Add this page to the object's radix tree, and look up mpred if
1505 	 * needed.
1506 	 */
1507 	if (iter) {
1508 		KASSERT(!lookup, ("%s: cannot lookup mpred", __func__));
1509 		error = vm_radix_iter_insert(pages, m);
1510 	} else if (lookup)
1511 		error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred);
1512 	else
1513 		error = vm_radix_insert(&object->rtree, m);
1514 	if (__predict_false(error != 0)) {
1515 		m->object = NULL;
1516 		m->pindex = 0;
1517 		m->ref_count &= ~VPRC_OBJREF;
1518 		return (1);
1519 	}
1520 
1521 	/*
1522 	 * Now link into the object's ordered list of backed pages.
1523 	 */
1524 	vm_page_insert_radixdone(m, object, mpred);
1525 	vm_pager_page_inserted(object, m);
1526 	return (0);
1527 }
1528 
1529 /*
1530  *	vm_page_insert:		[ internal use only ]
1531  *
1532  *	Inserts the given mem entry into the object and object list.
1533  *
1534  *	The object must be locked.
1535  */
1536 int
1537 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1538 {
1539 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL,
1540 	    true));
1541 }
1542 
1543 /*
1544  *	vm_page_insert_after:
1545  *
1546  *	Inserts the page "m" into the specified object at offset "pindex".
1547  *
1548  *	The page "mpred" must immediately precede the offset "pindex" within
1549  *	the specified object.
1550  *
1551  *	The object must be locked.
1552  */
1553 static int
1554 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1555     vm_page_t mpred)
1556 {
1557 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, mpred,
1558 	    false));
1559 }
1560 
1561 /*
1562  *	vm_page_iter_insert:
1563  *
1564  *	Tries to insert the page "m" into the specified object at offset
1565  *	"pindex" using the iterator "pages".  Returns 0 if the insertion was
1566  *	successful.
1567  *
1568  *	The page "mpred" must immediately precede the offset "pindex" within
1569  *	the specified object.
1570  *
1571  *	The object must be locked.
1572  */
1573 static int
1574 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object,
1575     vm_pindex_t pindex, vm_page_t mpred)
1576 {
1577 	return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred,
1578 	    false));
1579 }
1580 
1581 /*
1582  *	vm_page_insert_radixdone:
1583  *
1584  *	Complete page "m" insertion into the specified object after the
1585  *	radix trie hooking.
1586  *
1587  *	The page "mpred" must precede the offset "m->pindex" within the
1588  *	specified object.
1589  *
1590  *	The object must be locked.
1591  */
1592 static void
1593 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1594 {
1595 
1596 	VM_OBJECT_ASSERT_WLOCKED(object);
1597 	KASSERT(object != NULL && m->object == object,
1598 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1599 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1600 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1601 	if (mpred != NULL) {
1602 		KASSERT(mpred->object == object,
1603 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1604 		KASSERT(mpred->pindex < m->pindex,
1605 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1606 		KASSERT(TAILQ_NEXT(mpred, listq) == NULL ||
1607 		    m->pindex < TAILQ_NEXT(mpred, listq)->pindex,
1608 		    ("vm_page_insert_radixdone: pindex doesn't precede msucc"));
1609 	} else {
1610 		KASSERT(TAILQ_EMPTY(&object->memq) ||
1611 		    m->pindex < TAILQ_FIRST(&object->memq)->pindex,
1612 		    ("vm_page_insert_radixdone: no mpred but not first page"));
1613 	}
1614 
1615 	if (mpred != NULL)
1616 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1617 	else
1618 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1619 
1620 	/*
1621 	 * Show that the object has one more resident page.
1622 	 */
1623 	object->resident_page_count++;
1624 
1625 	/*
1626 	 * Hold the vnode until the last page is released.
1627 	 */
1628 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1629 		vhold(object->handle);
1630 
1631 	/*
1632 	 * Since we are inserting a new and possibly dirty page,
1633 	 * update the object's generation count.
1634 	 */
1635 	if (pmap_page_is_write_mapped(m))
1636 		vm_object_set_writeable_dirty(object);
1637 }
1638 
1639 /*
1640  * Do the work to remove a page from its object.  The caller is responsible for
1641  * updating the page's fields to reflect this removal.
1642  */
1643 static void
1644 vm_page_object_remove(vm_page_t m)
1645 {
1646 	vm_object_t object;
1647 	vm_page_t mrem __diagused;
1648 
1649 	vm_page_assert_xbusied(m);
1650 	object = m->object;
1651 	VM_OBJECT_ASSERT_WLOCKED(object);
1652 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1653 	    ("page %p is missing its object ref", m));
1654 
1655 	/* Deferred free of swap space. */
1656 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1657 		vm_pager_page_unswapped(m);
1658 
1659 	vm_pager_page_removed(object, m);
1660 
1661 	m->object = NULL;
1662 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1663 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1664 
1665 	/*
1666 	 * Now remove from the object's list of backed pages.
1667 	 */
1668 	TAILQ_REMOVE(&object->memq, m, listq);
1669 
1670 	/*
1671 	 * And show that the object has one fewer resident page.
1672 	 */
1673 	object->resident_page_count--;
1674 
1675 	/*
1676 	 * The vnode may now be recycled.
1677 	 */
1678 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1679 		vdrop(object->handle);
1680 }
1681 
1682 /*
1683  *	vm_page_remove:
1684  *
1685  *	Removes the specified page from its containing object, but does not
1686  *	invalidate any backing storage.  Returns true if the object's reference
1687  *	was the last reference to the page, and false otherwise.
1688  *
1689  *	The object must be locked and the page must be exclusively busied.
1690  *	The exclusive busy will be released on return.  If this is not the
1691  *	final ref and the caller does not hold a wire reference it may not
1692  *	continue to access the page.
1693  */
1694 bool
1695 vm_page_remove(vm_page_t m)
1696 {
1697 	bool dropped;
1698 
1699 	dropped = vm_page_remove_xbusy(m);
1700 	vm_page_xunbusy(m);
1701 
1702 	return (dropped);
1703 }
1704 
1705 /*
1706  *	vm_page_remove_xbusy
1707  *
1708  *	Removes the page but leaves the xbusy held.  Returns true if this
1709  *	removed the final ref and false otherwise.
1710  */
1711 bool
1712 vm_page_remove_xbusy(vm_page_t m)
1713 {
1714 
1715 	vm_page_object_remove(m);
1716 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1717 }
1718 
1719 /*
1720  *	vm_page_lookup:
1721  *
1722  *	Returns the page associated with the object/offset
1723  *	pair specified; if none is found, NULL is returned.
1724  *
1725  *	The object must be locked.
1726  */
1727 vm_page_t
1728 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1729 {
1730 
1731 	VM_OBJECT_ASSERT_LOCKED(object);
1732 	return (vm_radix_lookup(&object->rtree, pindex));
1733 }
1734 
1735 /*
1736  *	vm_page_iter_init:
1737  *
1738  *	Initialize iterator for vm pages.
1739  */
1740 void
1741 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object)
1742 {
1743 
1744 	vm_radix_iter_init(pages, &object->rtree);
1745 }
1746 
1747 /*
1748  *	vm_page_iter_init:
1749  *
1750  *	Initialize iterator for vm pages.
1751  */
1752 void
1753 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object,
1754     vm_pindex_t limit)
1755 {
1756 
1757 	vm_radix_iter_limit_init(pages, &object->rtree, limit);
1758 }
1759 
1760 /*
1761  *	vm_page_iter_lookup:
1762  *
1763  *	Returns the page associated with the object/offset pair specified, and
1764  *	stores the path to its position; if none is found, NULL is returned.
1765  *
1766  *	The iter pctrie must be locked.
1767  */
1768 vm_page_t
1769 vm_page_iter_lookup(struct pctrie_iter *pages, vm_pindex_t pindex)
1770 {
1771 
1772 	return (vm_radix_iter_lookup(pages, pindex));
1773 }
1774 
1775 /*
1776  *	vm_page_lookup_unlocked:
1777  *
1778  *	Returns the page associated with the object/offset pair specified;
1779  *	if none is found, NULL is returned.  The page may be no longer be
1780  *	present in the object at the time that this function returns.  Only
1781  *	useful for opportunistic checks such as inmem().
1782  */
1783 vm_page_t
1784 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1785 {
1786 
1787 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1788 }
1789 
1790 /*
1791  *	vm_page_relookup:
1792  *
1793  *	Returns a page that must already have been busied by
1794  *	the caller.  Used for bogus page replacement.
1795  */
1796 vm_page_t
1797 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1798 {
1799 	vm_page_t m;
1800 
1801 	m = vm_page_lookup_unlocked(object, pindex);
1802 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1803 	    m->object == object && m->pindex == pindex,
1804 	    ("vm_page_relookup: Invalid page %p", m));
1805 	return (m);
1806 }
1807 
1808 /*
1809  * This should only be used by lockless functions for releasing transient
1810  * incorrect acquires.  The page may have been freed after we acquired a
1811  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1812  * further to do.
1813  */
1814 static void
1815 vm_page_busy_release(vm_page_t m)
1816 {
1817 	u_int x;
1818 
1819 	x = vm_page_busy_fetch(m);
1820 	for (;;) {
1821 		if (x == VPB_FREED)
1822 			break;
1823 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1824 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1825 			    x - VPB_ONE_SHARER))
1826 				break;
1827 			continue;
1828 		}
1829 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1830 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1831 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1832 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1833 			continue;
1834 		if ((x & VPB_BIT_WAITERS) != 0)
1835 			wakeup(m);
1836 		break;
1837 	}
1838 }
1839 
1840 /*
1841  *	vm_page_find_least:
1842  *
1843  *	Returns the page associated with the object with least pindex
1844  *	greater than or equal to the parameter pindex, or NULL.
1845  *
1846  *	The object must be locked.
1847  */
1848 vm_page_t
1849 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1850 {
1851 	vm_page_t m;
1852 
1853 	VM_OBJECT_ASSERT_LOCKED(object);
1854 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1855 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1856 	return (m);
1857 }
1858 
1859 /*
1860  *	vm_page_iter_lookup_ge:
1861  *
1862  *	Returns the page associated with the object with least pindex
1863  *	greater than or equal to the parameter pindex, or NULL.  Initializes the
1864  *	iterator to point to that page.
1865  *
1866  *	The iter pctrie must be locked.
1867  */
1868 vm_page_t
1869 vm_page_iter_lookup_ge(struct pctrie_iter *pages, vm_pindex_t pindex)
1870 {
1871 
1872 	return (vm_radix_iter_lookup_ge(pages, pindex));
1873 }
1874 
1875 /*
1876  * Returns the given page's successor (by pindex) within the object if it is
1877  * resident; if none is found, NULL is returned.
1878  *
1879  * The object must be locked.
1880  */
1881 vm_page_t
1882 vm_page_next(vm_page_t m)
1883 {
1884 	vm_page_t next;
1885 
1886 	VM_OBJECT_ASSERT_LOCKED(m->object);
1887 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1888 		MPASS(next->object == m->object);
1889 		if (next->pindex != m->pindex + 1)
1890 			next = NULL;
1891 	}
1892 	return (next);
1893 }
1894 
1895 /*
1896  * Returns the given page's predecessor (by pindex) within the object if it is
1897  * resident; if none is found, NULL is returned.
1898  *
1899  * The object must be locked.
1900  */
1901 vm_page_t
1902 vm_page_prev(vm_page_t m)
1903 {
1904 	vm_page_t prev;
1905 
1906 	VM_OBJECT_ASSERT_LOCKED(m->object);
1907 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1908 		MPASS(prev->object == m->object);
1909 		if (prev->pindex != m->pindex - 1)
1910 			prev = NULL;
1911 	}
1912 	return (prev);
1913 }
1914 
1915 /*
1916  * Uses the page mnew as a replacement for an existing page at index
1917  * pindex which must be already present in the object.
1918  *
1919  * Both pages must be exclusively busied on enter.  The old page is
1920  * unbusied on exit.
1921  *
1922  * A return value of true means mold is now free.  If this is not the
1923  * final ref and the caller does not hold a wire reference it may not
1924  * continue to access the page.
1925  */
1926 static bool
1927 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1928     vm_page_t mold)
1929 {
1930 	vm_page_t mret __diagused;
1931 	bool dropped;
1932 
1933 	VM_OBJECT_ASSERT_WLOCKED(object);
1934 	vm_page_assert_xbusied(mold);
1935 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1936 	    ("vm_page_replace: page %p already in object", mnew));
1937 
1938 	/*
1939 	 * This function mostly follows vm_page_insert() and
1940 	 * vm_page_remove() without the radix, object count and vnode
1941 	 * dance.  Double check such functions for more comments.
1942 	 */
1943 
1944 	mnew->object = object;
1945 	mnew->pindex = pindex;
1946 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1947 	mret = vm_radix_replace(&object->rtree, mnew);
1948 	KASSERT(mret == mold,
1949 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1950 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1951 	    (mnew->oflags & VPO_UNMANAGED),
1952 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1953 
1954 	/* Keep the resident page list in sorted order. */
1955 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1956 	TAILQ_REMOVE(&object->memq, mold, listq);
1957 	mold->object = NULL;
1958 
1959 	/*
1960 	 * The object's resident_page_count does not change because we have
1961 	 * swapped one page for another, but the generation count should
1962 	 * change if the page is dirty.
1963 	 */
1964 	if (pmap_page_is_write_mapped(mnew))
1965 		vm_object_set_writeable_dirty(object);
1966 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1967 	vm_page_xunbusy(mold);
1968 
1969 	return (dropped);
1970 }
1971 
1972 void
1973 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1974     vm_page_t mold)
1975 {
1976 
1977 	vm_page_assert_xbusied(mnew);
1978 
1979 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1980 		vm_page_free(mold);
1981 }
1982 
1983 /*
1984  *	vm_page_rename:
1985  *
1986  *	Move the given memory entry from its
1987  *	current object to the specified target object/offset.
1988  *
1989  *	Note: swap associated with the page must be invalidated by the move.  We
1990  *	      have to do this for several reasons:  (1) we aren't freeing the
1991  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1992  *	      moving the page from object A to B, and will then later move
1993  *	      the backing store from A to B and we can't have a conflict.
1994  *
1995  *	Note: we *always* dirty the page.  It is necessary both for the
1996  *	      fact that we moved it, and because we may be invalidating
1997  *	      swap.
1998  *
1999  *	The objects must be locked.
2000  */
2001 int
2002 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
2003 {
2004 	vm_page_t mpred;
2005 	vm_pindex_t opidx;
2006 
2007 	VM_OBJECT_ASSERT_WLOCKED(new_object);
2008 
2009 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
2010 
2011 	/*
2012 	 * Create a custom version of vm_page_insert() which does not depend
2013 	 * by m_prev and can cheat on the implementation aspects of the
2014 	 * function.
2015 	 */
2016 	opidx = m->pindex;
2017 	m->pindex = new_pindex;
2018 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
2019 		m->pindex = opidx;
2020 		return (1);
2021 	}
2022 
2023 	/*
2024 	 * The operation cannot fail anymore.  The removal must happen before
2025 	 * the listq iterator is tainted.
2026 	 */
2027 	m->pindex = opidx;
2028 	vm_page_object_remove(m);
2029 
2030 	/* Return back to the new pindex to complete vm_page_insert(). */
2031 	m->pindex = new_pindex;
2032 	m->object = new_object;
2033 
2034 	vm_page_insert_radixdone(m, new_object, mpred);
2035 	vm_page_dirty(m);
2036 	vm_pager_page_inserted(new_object, m);
2037 	return (0);
2038 }
2039 
2040 /*
2041  *	vm_page_mpred:
2042  *
2043  *	Return the greatest page of the object with index <= pindex,
2044  *	or NULL, if there is none.  Assumes object lock is held.
2045  */
2046 vm_page_t
2047 vm_page_mpred(vm_object_t object, vm_pindex_t pindex)
2048 {
2049 	return (vm_radix_lookup_le(&object->rtree, pindex));
2050 }
2051 
2052 /*
2053  *	vm_page_alloc:
2054  *
2055  *	Allocate and return a page that is associated with the specified
2056  *	object and offset pair.  By default, this page is exclusive busied.
2057  *
2058  *	The caller must always specify an allocation class.
2059  *
2060  *	allocation classes:
2061  *	VM_ALLOC_NORMAL		normal process request
2062  *	VM_ALLOC_SYSTEM		system *really* needs a page
2063  *	VM_ALLOC_INTERRUPT	interrupt time request
2064  *
2065  *	optional allocation flags:
2066  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2067  *				intends to allocate
2068  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2069  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2070  *	VM_ALLOC_SBUSY		shared busy the allocated page
2071  *	VM_ALLOC_WIRED		wire the allocated page
2072  *	VM_ALLOC_ZERO		prefer a zeroed page
2073  */
2074 vm_page_t
2075 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
2076 {
2077 
2078 	return (vm_page_alloc_after(object, pindex, req,
2079 	    vm_page_mpred(object, pindex)));
2080 }
2081 
2082 /*
2083  * Allocate a page in the specified object with the given page index.  To
2084  * optimize insertion of the page into the object, the caller must also specify
2085  * the resident page in the object with largest index smaller than the given
2086  * page index, or NULL if no such page exists.
2087  */
2088 vm_page_t
2089 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
2090     int req, vm_page_t mpred)
2091 {
2092 	struct vm_domainset_iter di;
2093 	vm_page_t m;
2094 	int domain;
2095 
2096 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2097 	do {
2098 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2099 		    mpred);
2100 		if (m != NULL)
2101 			break;
2102 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2103 
2104 	return (m);
2105 }
2106 
2107 /*
2108  * Returns true if the number of free pages exceeds the minimum
2109  * for the request class and false otherwise.
2110  */
2111 static int
2112 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2113 {
2114 	u_int limit, old, new;
2115 
2116 	if (req_class == VM_ALLOC_INTERRUPT)
2117 		limit = 0;
2118 	else if (req_class == VM_ALLOC_SYSTEM)
2119 		limit = vmd->vmd_interrupt_free_min;
2120 	else
2121 		limit = vmd->vmd_free_reserved;
2122 
2123 	/*
2124 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2125 	 */
2126 	limit += npages;
2127 	old = atomic_load_int(&vmd->vmd_free_count);
2128 	do {
2129 		if (old < limit)
2130 			return (0);
2131 		new = old - npages;
2132 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2133 
2134 	/* Wake the page daemon if we've crossed the threshold. */
2135 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2136 		pagedaemon_wakeup(vmd->vmd_domain);
2137 
2138 	/* Only update bitsets on transitions. */
2139 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2140 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2141 		vm_domain_set(vmd);
2142 
2143 	return (1);
2144 }
2145 
2146 int
2147 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2148 {
2149 	int req_class;
2150 
2151 	/*
2152 	 * The page daemon is allowed to dig deeper into the free page list.
2153 	 */
2154 	req_class = req & VM_ALLOC_CLASS_MASK;
2155 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2156 		req_class = VM_ALLOC_SYSTEM;
2157 	return (_vm_domain_allocate(vmd, req_class, npages));
2158 }
2159 
2160 vm_page_t
2161 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2162     int req, vm_page_t mpred)
2163 {
2164 	struct vm_domain *vmd;
2165 	vm_page_t m;
2166 	int flags;
2167 
2168 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2169 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2170 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2171 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2172 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2173 	KASSERT((req & ~VPA_FLAGS) == 0,
2174 	    ("invalid request %#x", req));
2175 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2176 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2177 	    ("invalid request %#x", req));
2178 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2179 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2180 	    (uintmax_t)pindex));
2181 	VM_OBJECT_ASSERT_WLOCKED(object);
2182 
2183 	flags = 0;
2184 	m = NULL;
2185 	if (!vm_pager_can_alloc_page(object, pindex))
2186 		return (NULL);
2187 again:
2188 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2189 		m = vm_page_alloc_nofree_domain(domain, req);
2190 		if (m != NULL)
2191 			goto found;
2192 	}
2193 #if VM_NRESERVLEVEL > 0
2194 	/*
2195 	 * Can we allocate the page from a reservation?
2196 	 */
2197 	if (vm_object_reserv(object) &&
2198 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2199 	    NULL) {
2200 		goto found;
2201 	}
2202 #endif
2203 	vmd = VM_DOMAIN(domain);
2204 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2205 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2206 		    M_NOWAIT | M_NOVM);
2207 		if (m != NULL) {
2208 			flags |= PG_PCPU_CACHE;
2209 			goto found;
2210 		}
2211 	}
2212 	if (vm_domain_allocate(vmd, req, 1)) {
2213 		/*
2214 		 * If not, allocate it from the free page queues.
2215 		 */
2216 		vm_domain_free_lock(vmd);
2217 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2218 		vm_domain_free_unlock(vmd);
2219 		if (m == NULL) {
2220 			vm_domain_freecnt_inc(vmd, 1);
2221 #if VM_NRESERVLEVEL > 0
2222 			if (vm_reserv_reclaim_inactive(domain))
2223 				goto again;
2224 #endif
2225 		}
2226 	}
2227 	if (m == NULL) {
2228 		/*
2229 		 * Not allocatable, give up.
2230 		 */
2231 		if (vm_domain_alloc_fail(vmd, object, req))
2232 			goto again;
2233 		return (NULL);
2234 	}
2235 
2236 	/*
2237 	 * At this point we had better have found a good page.
2238 	 */
2239 found:
2240 	vm_page_dequeue(m);
2241 	vm_page_alloc_check(m);
2242 
2243 	/*
2244 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2245 	 */
2246 	flags |= m->flags & PG_ZERO;
2247 	if ((req & VM_ALLOC_NODUMP) != 0)
2248 		flags |= PG_NODUMP;
2249 	if ((req & VM_ALLOC_NOFREE) != 0)
2250 		flags |= PG_NOFREE;
2251 	m->flags = flags;
2252 	m->a.flags = 0;
2253 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2254 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2255 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2256 	else if ((req & VM_ALLOC_SBUSY) != 0)
2257 		m->busy_lock = VPB_SHARERS_WORD(1);
2258 	else
2259 		m->busy_lock = VPB_UNBUSIED;
2260 	if (req & VM_ALLOC_WIRED) {
2261 		vm_wire_add(1);
2262 		m->ref_count = 1;
2263 	}
2264 	m->a.act_count = 0;
2265 
2266 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2267 		if (req & VM_ALLOC_WIRED) {
2268 			vm_wire_sub(1);
2269 			m->ref_count = 0;
2270 		}
2271 		KASSERT(m->object == NULL, ("page %p has object", m));
2272 		m->oflags = VPO_UNMANAGED;
2273 		m->busy_lock = VPB_UNBUSIED;
2274 		/* Don't change PG_ZERO. */
2275 		vm_page_free_toq(m);
2276 		if (req & VM_ALLOC_WAITFAIL) {
2277 			VM_OBJECT_WUNLOCK(object);
2278 			vm_radix_wait();
2279 			VM_OBJECT_WLOCK(object);
2280 		}
2281 		return (NULL);
2282 	}
2283 
2284 	/* Ignore device objects; the pager sets "memattr" for them. */
2285 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2286 	    (object->flags & OBJ_FICTITIOUS) == 0)
2287 		pmap_page_set_memattr(m, object->memattr);
2288 
2289 	return (m);
2290 }
2291 
2292 /*
2293  *	vm_page_alloc_contig:
2294  *
2295  *	Allocate a contiguous set of physical pages of the given size "npages"
2296  *	from the free lists.  All of the physical pages must be at or above
2297  *	the given physical address "low" and below the given physical address
2298  *	"high".  The given value "alignment" determines the alignment of the
2299  *	first physical page in the set.  If the given value "boundary" is
2300  *	non-zero, then the set of physical pages cannot cross any physical
2301  *	address boundary that is a multiple of that value.  Both "alignment"
2302  *	and "boundary" must be a power of two.
2303  *
2304  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2305  *	then the memory attribute setting for the physical pages is configured
2306  *	to the object's memory attribute setting.  Otherwise, the memory
2307  *	attribute setting for the physical pages is configured to "memattr",
2308  *	overriding the object's memory attribute setting.  However, if the
2309  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2310  *	memory attribute setting for the physical pages cannot be configured
2311  *	to VM_MEMATTR_DEFAULT.
2312  *
2313  *	The specified object may not contain fictitious pages.
2314  *
2315  *	The caller must always specify an allocation class.
2316  *
2317  *	allocation classes:
2318  *	VM_ALLOC_NORMAL		normal process request
2319  *	VM_ALLOC_SYSTEM		system *really* needs a page
2320  *	VM_ALLOC_INTERRUPT	interrupt time request
2321  *
2322  *	optional allocation flags:
2323  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2324  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2325  *	VM_ALLOC_SBUSY		shared busy the allocated page
2326  *	VM_ALLOC_WIRED		wire the allocated page
2327  *	VM_ALLOC_ZERO		prefer a zeroed page
2328  */
2329 vm_page_t
2330 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2331     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2332     vm_paddr_t boundary, vm_memattr_t memattr)
2333 {
2334 	struct vm_domainset_iter di;
2335 	vm_page_t bounds[2];
2336 	vm_page_t m;
2337 	int domain;
2338 	int start_segind;
2339 
2340 	start_segind = -1;
2341 
2342 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2343 	do {
2344 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2345 		    npages, low, high, alignment, boundary, memattr);
2346 		if (m != NULL)
2347 			break;
2348 		if (start_segind == -1)
2349 			start_segind = vm_phys_lookup_segind(low);
2350 		if (vm_phys_find_range(bounds, start_segind, domain,
2351 		    npages, low, high) == -1) {
2352 			vm_domainset_iter_ignore(&di, domain);
2353 		}
2354 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2355 
2356 	return (m);
2357 }
2358 
2359 static vm_page_t
2360 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2361     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2362 {
2363 	struct vm_domain *vmd;
2364 	vm_page_t m_ret;
2365 
2366 	/*
2367 	 * Can we allocate the pages without the number of free pages falling
2368 	 * below the lower bound for the allocation class?
2369 	 */
2370 	vmd = VM_DOMAIN(domain);
2371 	if (!vm_domain_allocate(vmd, req, npages))
2372 		return (NULL);
2373 	/*
2374 	 * Try to allocate the pages from the free page queues.
2375 	 */
2376 	vm_domain_free_lock(vmd);
2377 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2378 	    alignment, boundary);
2379 	vm_domain_free_unlock(vmd);
2380 	if (m_ret != NULL)
2381 		return (m_ret);
2382 #if VM_NRESERVLEVEL > 0
2383 	/*
2384 	 * Try to break a reservation to allocate the pages.
2385 	 */
2386 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2387 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2388 	            high, alignment, boundary);
2389 		if (m_ret != NULL)
2390 			return (m_ret);
2391 	}
2392 #endif
2393 	vm_domain_freecnt_inc(vmd, npages);
2394 	return (NULL);
2395 }
2396 
2397 vm_page_t
2398 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2399     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2400     vm_paddr_t boundary, vm_memattr_t memattr)
2401 {
2402 	struct pctrie_iter pages;
2403 	vm_page_t m, m_ret, mpred;
2404 	u_int busy_lock, flags, oflags;
2405 
2406 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2407 	KASSERT((req & ~VPAC_FLAGS) == 0,
2408 	    ("invalid request %#x", req));
2409 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2410 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2411 	    ("invalid request %#x", req));
2412 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2413 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2414 	    ("invalid request %#x", req));
2415 	VM_OBJECT_ASSERT_WLOCKED(object);
2416 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2417 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2418 	    object));
2419 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2420 
2421 	vm_page_iter_init(&pages, object);
2422 	mpred = vm_radix_iter_lookup_le(&pages, pindex);
2423 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2424 	    ("vm_page_alloc_contig: pindex already allocated"));
2425 	for (;;) {
2426 #if VM_NRESERVLEVEL > 0
2427 		/*
2428 		 * Can we allocate the pages from a reservation?
2429 		 */
2430 		if (vm_object_reserv(object) &&
2431 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2432 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2433 			break;
2434 		}
2435 #endif
2436 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2437 		    low, high, alignment, boundary)) != NULL)
2438 			break;
2439 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2440 			return (NULL);
2441 	}
2442 
2443 	/*
2444 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2445 	 */
2446 	flags = PG_ZERO;
2447 	if ((req & VM_ALLOC_NODUMP) != 0)
2448 		flags |= PG_NODUMP;
2449 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2450 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2451 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2452 	else if ((req & VM_ALLOC_SBUSY) != 0)
2453 		busy_lock = VPB_SHARERS_WORD(1);
2454 	else
2455 		busy_lock = VPB_UNBUSIED;
2456 	if ((req & VM_ALLOC_WIRED) != 0)
2457 		vm_wire_add(npages);
2458 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2459 	    memattr == VM_MEMATTR_DEFAULT)
2460 		memattr = object->memattr;
2461 	for (m = m_ret; m < &m_ret[npages]; m++) {
2462 		vm_page_dequeue(m);
2463 		vm_page_alloc_check(m);
2464 		m->a.flags = 0;
2465 		m->flags = (m->flags | PG_NODUMP) & flags;
2466 		m->busy_lock = busy_lock;
2467 		if ((req & VM_ALLOC_WIRED) != 0)
2468 			m->ref_count = 1;
2469 		m->a.act_count = 0;
2470 		m->oflags = oflags;
2471 		if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) {
2472 			if ((req & VM_ALLOC_WIRED) != 0)
2473 				vm_wire_sub(npages);
2474 			KASSERT(m->object == NULL,
2475 			    ("page %p has object", m));
2476 			mpred = m;
2477 			for (m = m_ret; m < &m_ret[npages]; m++) {
2478 				if (m <= mpred &&
2479 				    (req & VM_ALLOC_WIRED) != 0)
2480 					m->ref_count = 0;
2481 				m->oflags = VPO_UNMANAGED;
2482 				m->busy_lock = VPB_UNBUSIED;
2483 				/* Don't change PG_ZERO. */
2484 				vm_page_free_toq(m);
2485 			}
2486 			if (req & VM_ALLOC_WAITFAIL) {
2487 				VM_OBJECT_WUNLOCK(object);
2488 				vm_radix_wait();
2489 				VM_OBJECT_WLOCK(object);
2490 			}
2491 			return (NULL);
2492 		}
2493 		mpred = m;
2494 		if (memattr != VM_MEMATTR_DEFAULT)
2495 			pmap_page_set_memattr(m, memattr);
2496 		pindex++;
2497 	}
2498 	return (m_ret);
2499 }
2500 
2501 /*
2502  * Allocate a physical page that is not intended to be inserted into a VM
2503  * object.
2504  */
2505 vm_page_t
2506 vm_page_alloc_noobj_domain(int domain, int req)
2507 {
2508 	struct vm_domain *vmd;
2509 	vm_page_t m;
2510 	int flags;
2511 
2512 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2513 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2514 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2515 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2516 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2517 	KASSERT((req & ~VPAN_FLAGS) == 0,
2518 	    ("invalid request %#x", req));
2519 
2520 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2521 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2522 	vmd = VM_DOMAIN(domain);
2523 again:
2524 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2525 		m = vm_page_alloc_nofree_domain(domain, req);
2526 		if (m != NULL)
2527 			goto found;
2528 	}
2529 
2530 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2531 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2532 		    M_NOWAIT | M_NOVM);
2533 		if (m != NULL) {
2534 			flags |= PG_PCPU_CACHE;
2535 			goto found;
2536 		}
2537 	}
2538 
2539 	if (vm_domain_allocate(vmd, req, 1)) {
2540 		vm_domain_free_lock(vmd);
2541 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2542 		vm_domain_free_unlock(vmd);
2543 		if (m == NULL) {
2544 			vm_domain_freecnt_inc(vmd, 1);
2545 #if VM_NRESERVLEVEL > 0
2546 			if (vm_reserv_reclaim_inactive(domain))
2547 				goto again;
2548 #endif
2549 		}
2550 	}
2551 	if (m == NULL) {
2552 		if (vm_domain_alloc_fail(vmd, NULL, req))
2553 			goto again;
2554 		return (NULL);
2555 	}
2556 
2557 found:
2558 	vm_page_dequeue(m);
2559 	vm_page_alloc_check(m);
2560 
2561 	/*
2562 	 * Consumers should not rely on a useful default pindex value.
2563 	 */
2564 	m->pindex = 0xdeadc0dedeadc0de;
2565 	m->flags = (m->flags & PG_ZERO) | flags;
2566 	m->a.flags = 0;
2567 	m->oflags = VPO_UNMANAGED;
2568 	m->busy_lock = VPB_UNBUSIED;
2569 	if ((req & VM_ALLOC_WIRED) != 0) {
2570 		vm_wire_add(1);
2571 		m->ref_count = 1;
2572 	}
2573 
2574 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2575 		pmap_zero_page(m);
2576 
2577 	return (m);
2578 }
2579 
2580 #if VM_NRESERVLEVEL > 1
2581 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2582 #elif VM_NRESERVLEVEL > 0
2583 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2584 #else
2585 #define	VM_NOFREE_IMPORT_ORDER	8
2586 #endif
2587 
2588 /*
2589  * Allocate a single NOFREE page.
2590  *
2591  * This routine hands out NOFREE pages from higher-order
2592  * physical memory blocks in order to reduce memory fragmentation.
2593  * When a NOFREE for a given domain chunk is used up,
2594  * the routine will try to fetch a new one from the freelists
2595  * and discard the old one.
2596  */
2597 static vm_page_t
2598 vm_page_alloc_nofree_domain(int domain, int req)
2599 {
2600 	vm_page_t m;
2601 	struct vm_domain *vmd;
2602 	struct vm_nofreeq *nqp;
2603 
2604 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2605 
2606 	vmd = VM_DOMAIN(domain);
2607 	nqp = &vmd->vmd_nofreeq;
2608 	vm_domain_free_lock(vmd);
2609 	if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) {
2610 		if (!vm_domain_allocate(vmd, req,
2611 		    1 << VM_NOFREE_IMPORT_ORDER)) {
2612 			vm_domain_free_unlock(vmd);
2613 			return (NULL);
2614 		}
2615 		nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2616 		    VM_NOFREE_IMPORT_ORDER);
2617 		if (nqp->ma == NULL) {
2618 			vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER);
2619 			vm_domain_free_unlock(vmd);
2620 			return (NULL);
2621 		}
2622 		nqp->offs = 0;
2623 	}
2624 	m = &nqp->ma[nqp->offs++];
2625 	vm_domain_free_unlock(vmd);
2626 	VM_CNT_ADD(v_nofree_count, 1);
2627 
2628 	return (m);
2629 }
2630 
2631 vm_page_t
2632 vm_page_alloc_noobj(int req)
2633 {
2634 	struct vm_domainset_iter di;
2635 	vm_page_t m;
2636 	int domain;
2637 
2638 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2639 	do {
2640 		m = vm_page_alloc_noobj_domain(domain, req);
2641 		if (m != NULL)
2642 			break;
2643 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2644 
2645 	return (m);
2646 }
2647 
2648 vm_page_t
2649 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2650     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2651     vm_memattr_t memattr)
2652 {
2653 	struct vm_domainset_iter di;
2654 	vm_page_t m;
2655 	int domain;
2656 
2657 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2658 	do {
2659 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2660 		    high, alignment, boundary, memattr);
2661 		if (m != NULL)
2662 			break;
2663 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2664 
2665 	return (m);
2666 }
2667 
2668 vm_page_t
2669 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2670     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2671     vm_memattr_t memattr)
2672 {
2673 	vm_page_t m, m_ret;
2674 	u_int flags;
2675 
2676 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2677 	KASSERT((req & ~VPANC_FLAGS) == 0,
2678 	    ("invalid request %#x", req));
2679 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2680 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2681 	    ("invalid request %#x", req));
2682 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2683 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2684 	    ("invalid request %#x", req));
2685 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2686 
2687 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2688 	    low, high, alignment, boundary)) == NULL) {
2689 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2690 			return (NULL);
2691 	}
2692 
2693 	/*
2694 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2695 	 */
2696 	flags = PG_ZERO;
2697 	if ((req & VM_ALLOC_NODUMP) != 0)
2698 		flags |= PG_NODUMP;
2699 	if ((req & VM_ALLOC_WIRED) != 0)
2700 		vm_wire_add(npages);
2701 	for (m = m_ret; m < &m_ret[npages]; m++) {
2702 		vm_page_dequeue(m);
2703 		vm_page_alloc_check(m);
2704 
2705 		/*
2706 		 * Consumers should not rely on a useful default pindex value.
2707 		 */
2708 		m->pindex = 0xdeadc0dedeadc0de;
2709 		m->a.flags = 0;
2710 		m->flags = (m->flags | PG_NODUMP) & flags;
2711 		m->busy_lock = VPB_UNBUSIED;
2712 		if ((req & VM_ALLOC_WIRED) != 0)
2713 			m->ref_count = 1;
2714 		m->a.act_count = 0;
2715 		m->oflags = VPO_UNMANAGED;
2716 
2717 		/*
2718 		 * Zero the page before updating any mappings since the page is
2719 		 * not yet shared with any devices which might require the
2720 		 * non-default memory attribute.  pmap_page_set_memattr()
2721 		 * flushes data caches before returning.
2722 		 */
2723 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2724 			pmap_zero_page(m);
2725 		if (memattr != VM_MEMATTR_DEFAULT)
2726 			pmap_page_set_memattr(m, memattr);
2727 	}
2728 	return (m_ret);
2729 }
2730 
2731 /*
2732  * Check a page that has been freshly dequeued from a freelist.
2733  */
2734 static void
2735 vm_page_alloc_check(vm_page_t m)
2736 {
2737 
2738 	KASSERT(m->object == NULL, ("page %p has object", m));
2739 	KASSERT(m->a.queue == PQ_NONE &&
2740 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2741 	    ("page %p has unexpected queue %d, flags %#x",
2742 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2743 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2744 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2745 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2746 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2747 	    ("page %p has unexpected memattr %d",
2748 	    m, pmap_page_get_memattr(m)));
2749 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2750 	pmap_vm_page_alloc_check(m);
2751 }
2752 
2753 static int
2754 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2755 {
2756 	struct vm_domain *vmd;
2757 	struct vm_pgcache *pgcache;
2758 	int i;
2759 
2760 	pgcache = arg;
2761 	vmd = VM_DOMAIN(pgcache->domain);
2762 
2763 	/*
2764 	 * The page daemon should avoid creating extra memory pressure since its
2765 	 * main purpose is to replenish the store of free pages.
2766 	 */
2767 	if (vmd->vmd_severeset || curproc == pageproc ||
2768 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2769 		return (0);
2770 	domain = vmd->vmd_domain;
2771 	vm_domain_free_lock(vmd);
2772 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2773 	    (vm_page_t *)store);
2774 	vm_domain_free_unlock(vmd);
2775 	if (cnt != i)
2776 		vm_domain_freecnt_inc(vmd, cnt - i);
2777 
2778 	return (i);
2779 }
2780 
2781 static void
2782 vm_page_zone_release(void *arg, void **store, int cnt)
2783 {
2784 	struct vm_domain *vmd;
2785 	struct vm_pgcache *pgcache;
2786 	vm_page_t m;
2787 	int i;
2788 
2789 	pgcache = arg;
2790 	vmd = VM_DOMAIN(pgcache->domain);
2791 	vm_domain_free_lock(vmd);
2792 	for (i = 0; i < cnt; i++) {
2793 		m = (vm_page_t)store[i];
2794 		vm_phys_free_pages(m, 0);
2795 	}
2796 	vm_domain_free_unlock(vmd);
2797 	vm_domain_freecnt_inc(vmd, cnt);
2798 }
2799 
2800 #define	VPSC_ANY	0	/* No restrictions. */
2801 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2802 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2803 
2804 /*
2805  *	vm_page_scan_contig:
2806  *
2807  *	Scan vm_page_array[] between the specified entries "m_start" and
2808  *	"m_end" for a run of contiguous physical pages that satisfy the
2809  *	specified conditions, and return the lowest page in the run.  The
2810  *	specified "alignment" determines the alignment of the lowest physical
2811  *	page in the run.  If the specified "boundary" is non-zero, then the
2812  *	run of physical pages cannot span a physical address that is a
2813  *	multiple of "boundary".
2814  *
2815  *	"m_end" is never dereferenced, so it need not point to a vm_page
2816  *	structure within vm_page_array[].
2817  *
2818  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2819  *	span a hole (or discontiguity) in the physical address space.  Both
2820  *	"alignment" and "boundary" must be a power of two.
2821  */
2822 static vm_page_t
2823 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2824     u_long alignment, vm_paddr_t boundary, int options)
2825 {
2826 	vm_object_t object;
2827 	vm_paddr_t pa;
2828 	vm_page_t m, m_run;
2829 #if VM_NRESERVLEVEL > 0
2830 	int level;
2831 #endif
2832 	int m_inc, order, run_ext, run_len;
2833 
2834 	KASSERT(npages > 0, ("npages is 0"));
2835 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2836 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2837 	m_run = NULL;
2838 	run_len = 0;
2839 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2840 		KASSERT((m->flags & PG_MARKER) == 0,
2841 		    ("page %p is PG_MARKER", m));
2842 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2843 		    ("fictitious page %p has invalid ref count", m));
2844 
2845 		/*
2846 		 * If the current page would be the start of a run, check its
2847 		 * physical address against the end, alignment, and boundary
2848 		 * conditions.  If it doesn't satisfy these conditions, either
2849 		 * terminate the scan or advance to the next page that
2850 		 * satisfies the failed condition.
2851 		 */
2852 		if (run_len == 0) {
2853 			KASSERT(m_run == NULL, ("m_run != NULL"));
2854 			if (m + npages > m_end)
2855 				break;
2856 			pa = VM_PAGE_TO_PHYS(m);
2857 			if (!vm_addr_align_ok(pa, alignment)) {
2858 				m_inc = atop(roundup2(pa, alignment) - pa);
2859 				continue;
2860 			}
2861 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2862 				m_inc = atop(roundup2(pa, boundary) - pa);
2863 				continue;
2864 			}
2865 		} else
2866 			KASSERT(m_run != NULL, ("m_run == NULL"));
2867 
2868 retry:
2869 		m_inc = 1;
2870 		if (vm_page_wired(m))
2871 			run_ext = 0;
2872 #if VM_NRESERVLEVEL > 0
2873 		else if ((level = vm_reserv_level(m)) >= 0 &&
2874 		    (options & VPSC_NORESERV) != 0) {
2875 			run_ext = 0;
2876 			/* Advance to the end of the reservation. */
2877 			pa = VM_PAGE_TO_PHYS(m);
2878 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2879 			    pa);
2880 		}
2881 #endif
2882 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2883 			/*
2884 			 * The page is considered eligible for relocation if
2885 			 * and only if it could be laundered or reclaimed by
2886 			 * the page daemon.
2887 			 */
2888 			VM_OBJECT_RLOCK(object);
2889 			if (object != m->object) {
2890 				VM_OBJECT_RUNLOCK(object);
2891 				goto retry;
2892 			}
2893 			/* Don't care: PG_NODUMP, PG_ZERO. */
2894 			if ((object->flags & OBJ_SWAP) == 0 &&
2895 			    object->type != OBJT_VNODE) {
2896 				run_ext = 0;
2897 #if VM_NRESERVLEVEL > 0
2898 			} else if ((options & VPSC_NOSUPER) != 0 &&
2899 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2900 				run_ext = 0;
2901 				/* Advance to the end of the superpage. */
2902 				pa = VM_PAGE_TO_PHYS(m);
2903 				m_inc = atop(roundup2(pa + 1,
2904 				    vm_reserv_size(level)) - pa);
2905 #endif
2906 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2907 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2908 				/*
2909 				 * The page is allocated but eligible for
2910 				 * relocation.  Extend the current run by one
2911 				 * page.
2912 				 */
2913 				KASSERT(pmap_page_get_memattr(m) ==
2914 				    VM_MEMATTR_DEFAULT,
2915 				    ("page %p has an unexpected memattr", m));
2916 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2917 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2918 				    ("page %p has unexpected oflags", m));
2919 				/* Don't care: PGA_NOSYNC. */
2920 				run_ext = 1;
2921 			} else
2922 				run_ext = 0;
2923 			VM_OBJECT_RUNLOCK(object);
2924 #if VM_NRESERVLEVEL > 0
2925 		} else if (level >= 0) {
2926 			/*
2927 			 * The page is reserved but not yet allocated.  In
2928 			 * other words, it is still free.  Extend the current
2929 			 * run by one page.
2930 			 */
2931 			run_ext = 1;
2932 #endif
2933 		} else if ((order = m->order) < VM_NFREEORDER) {
2934 			/*
2935 			 * The page is enqueued in the physical memory
2936 			 * allocator's free page queues.  Moreover, it is the
2937 			 * first page in a power-of-two-sized run of
2938 			 * contiguous free pages.  Add these pages to the end
2939 			 * of the current run, and jump ahead.
2940 			 */
2941 			run_ext = 1 << order;
2942 			m_inc = 1 << order;
2943 		} else {
2944 			/*
2945 			 * Skip the page for one of the following reasons: (1)
2946 			 * It is enqueued in the physical memory allocator's
2947 			 * free page queues.  However, it is not the first
2948 			 * page in a run of contiguous free pages.  (This case
2949 			 * rarely occurs because the scan is performed in
2950 			 * ascending order.) (2) It is not reserved, and it is
2951 			 * transitioning from free to allocated.  (Conversely,
2952 			 * the transition from allocated to free for managed
2953 			 * pages is blocked by the page busy lock.) (3) It is
2954 			 * allocated but not contained by an object and not
2955 			 * wired, e.g., allocated by Xen's balloon driver.
2956 			 */
2957 			run_ext = 0;
2958 		}
2959 
2960 		/*
2961 		 * Extend or reset the current run of pages.
2962 		 */
2963 		if (run_ext > 0) {
2964 			if (run_len == 0)
2965 				m_run = m;
2966 			run_len += run_ext;
2967 		} else {
2968 			if (run_len > 0) {
2969 				m_run = NULL;
2970 				run_len = 0;
2971 			}
2972 		}
2973 	}
2974 	if (run_len >= npages)
2975 		return (m_run);
2976 	return (NULL);
2977 }
2978 
2979 /*
2980  *	vm_page_reclaim_run:
2981  *
2982  *	Try to relocate each of the allocated virtual pages within the
2983  *	specified run of physical pages to a new physical address.  Free the
2984  *	physical pages underlying the relocated virtual pages.  A virtual page
2985  *	is relocatable if and only if it could be laundered or reclaimed by
2986  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2987  *	physical address above "high".
2988  *
2989  *	Returns 0 if every physical page within the run was already free or
2990  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2991  *	value indicating why the last attempt to relocate a virtual page was
2992  *	unsuccessful.
2993  *
2994  *	"req_class" must be an allocation class.
2995  */
2996 static int
2997 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2998     vm_paddr_t high)
2999 {
3000 	struct vm_domain *vmd;
3001 	struct spglist free;
3002 	vm_object_t object;
3003 	vm_paddr_t pa;
3004 	vm_page_t m, m_end, m_new;
3005 	int error, order, req;
3006 
3007 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
3008 	    ("req_class is not an allocation class"));
3009 	SLIST_INIT(&free);
3010 	error = 0;
3011 	m = m_run;
3012 	m_end = m_run + npages;
3013 	for (; error == 0 && m < m_end; m++) {
3014 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
3015 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
3016 
3017 		/*
3018 		 * Racily check for wirings.  Races are handled once the object
3019 		 * lock is held and the page is unmapped.
3020 		 */
3021 		if (vm_page_wired(m))
3022 			error = EBUSY;
3023 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
3024 			/*
3025 			 * The page is relocated if and only if it could be
3026 			 * laundered or reclaimed by the page daemon.
3027 			 */
3028 			VM_OBJECT_WLOCK(object);
3029 			/* Don't care: PG_NODUMP, PG_ZERO. */
3030 			if (m->object != object ||
3031 			    ((object->flags & OBJ_SWAP) == 0 &&
3032 			    object->type != OBJT_VNODE))
3033 				error = EINVAL;
3034 			else if (object->memattr != VM_MEMATTR_DEFAULT)
3035 				error = EINVAL;
3036 			else if (vm_page_queue(m) != PQ_NONE &&
3037 			    vm_page_tryxbusy(m) != 0) {
3038 				if (vm_page_wired(m)) {
3039 					vm_page_xunbusy(m);
3040 					error = EBUSY;
3041 					goto unlock;
3042 				}
3043 				KASSERT(pmap_page_get_memattr(m) ==
3044 				    VM_MEMATTR_DEFAULT,
3045 				    ("page %p has an unexpected memattr", m));
3046 				KASSERT(m->oflags == 0,
3047 				    ("page %p has unexpected oflags", m));
3048 				/* Don't care: PGA_NOSYNC. */
3049 				if (!vm_page_none_valid(m)) {
3050 					/*
3051 					 * First, try to allocate a new page
3052 					 * that is above "high".  Failing
3053 					 * that, try to allocate a new page
3054 					 * that is below "m_run".  Allocate
3055 					 * the new page between the end of
3056 					 * "m_run" and "high" only as a last
3057 					 * resort.
3058 					 */
3059 					req = req_class;
3060 					if ((m->flags & PG_NODUMP) != 0)
3061 						req |= VM_ALLOC_NODUMP;
3062 					if (trunc_page(high) !=
3063 					    ~(vm_paddr_t)PAGE_MASK) {
3064 						m_new =
3065 						    vm_page_alloc_noobj_contig(
3066 						    req, 1, round_page(high),
3067 						    ~(vm_paddr_t)0, PAGE_SIZE,
3068 						    0, VM_MEMATTR_DEFAULT);
3069 					} else
3070 						m_new = NULL;
3071 					if (m_new == NULL) {
3072 						pa = VM_PAGE_TO_PHYS(m_run);
3073 						m_new =
3074 						    vm_page_alloc_noobj_contig(
3075 						    req, 1, 0, pa - 1,
3076 						    PAGE_SIZE, 0,
3077 						    VM_MEMATTR_DEFAULT);
3078 					}
3079 					if (m_new == NULL) {
3080 						pa += ptoa(npages);
3081 						m_new =
3082 						    vm_page_alloc_noobj_contig(
3083 						    req, 1, pa, high, PAGE_SIZE,
3084 						    0, VM_MEMATTR_DEFAULT);
3085 					}
3086 					if (m_new == NULL) {
3087 						vm_page_xunbusy(m);
3088 						error = ENOMEM;
3089 						goto unlock;
3090 					}
3091 
3092 					/*
3093 					 * Unmap the page and check for new
3094 					 * wirings that may have been acquired
3095 					 * through a pmap lookup.
3096 					 */
3097 					if (object->ref_count != 0 &&
3098 					    !vm_page_try_remove_all(m)) {
3099 						vm_page_xunbusy(m);
3100 						vm_page_free(m_new);
3101 						error = EBUSY;
3102 						goto unlock;
3103 					}
3104 
3105 					/*
3106 					 * Replace "m" with the new page.  For
3107 					 * vm_page_replace(), "m" must be busy
3108 					 * and dequeued.  Finally, change "m"
3109 					 * as if vm_page_free() was called.
3110 					 */
3111 					m_new->a.flags = m->a.flags &
3112 					    ~PGA_QUEUE_STATE_MASK;
3113 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3114 					    ("page %p is managed", m_new));
3115 					m_new->oflags = 0;
3116 					pmap_copy_page(m, m_new);
3117 					m_new->valid = m->valid;
3118 					m_new->dirty = m->dirty;
3119 					m->flags &= ~PG_ZERO;
3120 					vm_page_dequeue(m);
3121 					if (vm_page_replace_hold(m_new, object,
3122 					    m->pindex, m) &&
3123 					    vm_page_free_prep(m))
3124 						SLIST_INSERT_HEAD(&free, m,
3125 						    plinks.s.ss);
3126 
3127 					/*
3128 					 * The new page must be deactivated
3129 					 * before the object is unlocked.
3130 					 */
3131 					vm_page_deactivate(m_new);
3132 				} else {
3133 					m->flags &= ~PG_ZERO;
3134 					vm_page_dequeue(m);
3135 					if (vm_page_free_prep(m))
3136 						SLIST_INSERT_HEAD(&free, m,
3137 						    plinks.s.ss);
3138 					KASSERT(m->dirty == 0,
3139 					    ("page %p is dirty", m));
3140 				}
3141 			} else
3142 				error = EBUSY;
3143 unlock:
3144 			VM_OBJECT_WUNLOCK(object);
3145 		} else {
3146 			MPASS(vm_page_domain(m) == domain);
3147 			vmd = VM_DOMAIN(domain);
3148 			vm_domain_free_lock(vmd);
3149 			order = m->order;
3150 			if (order < VM_NFREEORDER) {
3151 				/*
3152 				 * The page is enqueued in the physical memory
3153 				 * allocator's free page queues.  Moreover, it
3154 				 * is the first page in a power-of-two-sized
3155 				 * run of contiguous free pages.  Jump ahead
3156 				 * to the last page within that run, and
3157 				 * continue from there.
3158 				 */
3159 				m += (1 << order) - 1;
3160 			}
3161 #if VM_NRESERVLEVEL > 0
3162 			else if (vm_reserv_is_page_free(m))
3163 				order = 0;
3164 #endif
3165 			vm_domain_free_unlock(vmd);
3166 			if (order == VM_NFREEORDER)
3167 				error = EINVAL;
3168 		}
3169 	}
3170 	if ((m = SLIST_FIRST(&free)) != NULL) {
3171 		int cnt;
3172 
3173 		vmd = VM_DOMAIN(domain);
3174 		cnt = 0;
3175 		vm_domain_free_lock(vmd);
3176 		do {
3177 			MPASS(vm_page_domain(m) == domain);
3178 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3179 			vm_phys_free_pages(m, 0);
3180 			cnt++;
3181 		} while ((m = SLIST_FIRST(&free)) != NULL);
3182 		vm_domain_free_unlock(vmd);
3183 		vm_domain_freecnt_inc(vmd, cnt);
3184 	}
3185 	return (error);
3186 }
3187 
3188 #define	NRUNS	16
3189 
3190 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3191 
3192 #define	MIN_RECLAIM	8
3193 
3194 /*
3195  *	vm_page_reclaim_contig:
3196  *
3197  *	Reclaim allocated, contiguous physical memory satisfying the specified
3198  *	conditions by relocating the virtual pages using that physical memory.
3199  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3200  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3201  *	currently able to reclaim the requested number of pages.  Since
3202  *	relocation requires the allocation of physical pages, reclamation may
3203  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3204  *	fails in this manner, callers are expected to perform vm_wait() before
3205  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3206  *
3207  *	The caller must always specify an allocation class through "req".
3208  *
3209  *	allocation classes:
3210  *	VM_ALLOC_NORMAL		normal process request
3211  *	VM_ALLOC_SYSTEM		system *really* needs a page
3212  *	VM_ALLOC_INTERRUPT	interrupt time request
3213  *
3214  *	The optional allocation flags are ignored.
3215  *
3216  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3217  *	must be a power of two.
3218  */
3219 int
3220 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3221     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3222     int desired_runs)
3223 {
3224 	struct vm_domain *vmd;
3225 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3226 	u_long count, minalign, reclaimed;
3227 	int error, i, min_reclaim, nruns, options, req_class;
3228 	int segind, start_segind;
3229 	int ret;
3230 
3231 	KASSERT(npages > 0, ("npages is 0"));
3232 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3233 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3234 
3235 	ret = ENOMEM;
3236 
3237 	/*
3238 	 * If the caller wants to reclaim multiple runs, try to allocate
3239 	 * space to store the runs.  If that fails, fall back to the old
3240 	 * behavior of just reclaiming MIN_RECLAIM pages.
3241 	 */
3242 	if (desired_runs > 1)
3243 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3244 		    M_TEMP, M_NOWAIT);
3245 	else
3246 		m_runs = NULL;
3247 
3248 	if (m_runs == NULL) {
3249 		m_runs = _m_runs;
3250 		nruns = NRUNS;
3251 	} else {
3252 		nruns = NRUNS + desired_runs - 1;
3253 	}
3254 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3255 
3256 	/*
3257 	 * The caller will attempt an allocation after some runs have been
3258 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3259 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3260 	 * power of two or maximum chunk size, and ensure that each run is
3261 	 * suitably aligned.
3262 	 */
3263 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3264 	npages = roundup2(npages, minalign);
3265 	if (alignment < ptoa(minalign))
3266 		alignment = ptoa(minalign);
3267 
3268 	/*
3269 	 * The page daemon is allowed to dig deeper into the free page list.
3270 	 */
3271 	req_class = req & VM_ALLOC_CLASS_MASK;
3272 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3273 		req_class = VM_ALLOC_SYSTEM;
3274 
3275 	start_segind = vm_phys_lookup_segind(low);
3276 
3277 	/*
3278 	 * Return if the number of free pages cannot satisfy the requested
3279 	 * allocation.
3280 	 */
3281 	vmd = VM_DOMAIN(domain);
3282 	count = vmd->vmd_free_count;
3283 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3284 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3285 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3286 		goto done;
3287 
3288 	/*
3289 	 * Scan up to three times, relaxing the restrictions ("options") on
3290 	 * the reclamation of reservations and superpages each time.
3291 	 */
3292 	for (options = VPSC_NORESERV;;) {
3293 		bool phys_range_exists = false;
3294 
3295 		/*
3296 		 * Find the highest runs that satisfy the given constraints
3297 		 * and restrictions, and record them in "m_runs".
3298 		 */
3299 		count = 0;
3300 		segind = start_segind;
3301 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3302 		    npages, low, high)) != -1) {
3303 			phys_range_exists = true;
3304 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3305 			    bounds[1], alignment, boundary, options))) {
3306 				bounds[0] = m_run + npages;
3307 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3308 				count++;
3309 			}
3310 			segind++;
3311 		}
3312 
3313 		if (!phys_range_exists) {
3314 			ret = ERANGE;
3315 			goto done;
3316 		}
3317 
3318 		/*
3319 		 * Reclaim the highest runs in LIFO (descending) order until
3320 		 * the number of reclaimed pages, "reclaimed", is at least
3321 		 * "min_reclaim".  Reset "reclaimed" each time because each
3322 		 * reclamation is idempotent, and runs will (likely) recur
3323 		 * from one scan to the next as restrictions are relaxed.
3324 		 */
3325 		reclaimed = 0;
3326 		for (i = 0; count > 0 && i < nruns; i++) {
3327 			count--;
3328 			m_run = m_runs[RUN_INDEX(count, nruns)];
3329 			error = vm_page_reclaim_run(req_class, domain, npages,
3330 			    m_run, high);
3331 			if (error == 0) {
3332 				reclaimed += npages;
3333 				if (reclaimed >= min_reclaim) {
3334 					ret = 0;
3335 					goto done;
3336 				}
3337 			}
3338 		}
3339 
3340 		/*
3341 		 * Either relax the restrictions on the next scan or return if
3342 		 * the last scan had no restrictions.
3343 		 */
3344 		if (options == VPSC_NORESERV)
3345 			options = VPSC_NOSUPER;
3346 		else if (options == VPSC_NOSUPER)
3347 			options = VPSC_ANY;
3348 		else if (options == VPSC_ANY) {
3349 			if (reclaimed != 0)
3350 				ret = 0;
3351 			goto done;
3352 		}
3353 	}
3354 done:
3355 	if (m_runs != _m_runs)
3356 		free(m_runs, M_TEMP);
3357 	return (ret);
3358 }
3359 
3360 int
3361 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3362     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3363 {
3364 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3365 	    alignment, boundary, 1));
3366 }
3367 
3368 int
3369 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3370     u_long alignment, vm_paddr_t boundary)
3371 {
3372 	struct vm_domainset_iter di;
3373 	int domain, ret, status;
3374 
3375 	ret = ERANGE;
3376 
3377 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3378 	do {
3379 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3380 		    high, alignment, boundary);
3381 		if (status == 0)
3382 			return (0);
3383 		else if (status == ERANGE)
3384 			vm_domainset_iter_ignore(&di, domain);
3385 		else {
3386 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3387 			    "from vm_page_reclaim_contig_domain()", status));
3388 			ret = ENOMEM;
3389 		}
3390 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3391 
3392 	return (ret);
3393 }
3394 
3395 /*
3396  * Set the domain in the appropriate page level domainset.
3397  */
3398 void
3399 vm_domain_set(struct vm_domain *vmd)
3400 {
3401 
3402 	mtx_lock(&vm_domainset_lock);
3403 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3404 		vmd->vmd_minset = 1;
3405 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3406 	}
3407 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3408 		vmd->vmd_severeset = 1;
3409 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3410 	}
3411 	mtx_unlock(&vm_domainset_lock);
3412 }
3413 
3414 /*
3415  * Clear the domain from the appropriate page level domainset.
3416  */
3417 void
3418 vm_domain_clear(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 = 0;
3424 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3425 		if (vm_min_waiters != 0) {
3426 			vm_min_waiters = 0;
3427 			wakeup(&vm_min_domains);
3428 		}
3429 	}
3430 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3431 		vmd->vmd_severeset = 0;
3432 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3433 		if (vm_severe_waiters != 0) {
3434 			vm_severe_waiters = 0;
3435 			wakeup(&vm_severe_domains);
3436 		}
3437 	}
3438 
3439 	/*
3440 	 * If pageout daemon needs pages, then tell it that there are
3441 	 * some free.
3442 	 */
3443 	if (vmd->vmd_pageout_pages_needed &&
3444 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3445 		wakeup(&vmd->vmd_pageout_pages_needed);
3446 		vmd->vmd_pageout_pages_needed = 0;
3447 	}
3448 
3449 	/* See comments in vm_wait_doms(). */
3450 	if (vm_pageproc_waiters) {
3451 		vm_pageproc_waiters = 0;
3452 		wakeup(&vm_pageproc_waiters);
3453 	}
3454 	mtx_unlock(&vm_domainset_lock);
3455 }
3456 
3457 /*
3458  * Wait for free pages to exceed the min threshold globally.
3459  */
3460 void
3461 vm_wait_min(void)
3462 {
3463 
3464 	mtx_lock(&vm_domainset_lock);
3465 	while (vm_page_count_min()) {
3466 		vm_min_waiters++;
3467 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3468 	}
3469 	mtx_unlock(&vm_domainset_lock);
3470 }
3471 
3472 /*
3473  * Wait for free pages to exceed the severe threshold globally.
3474  */
3475 void
3476 vm_wait_severe(void)
3477 {
3478 
3479 	mtx_lock(&vm_domainset_lock);
3480 	while (vm_page_count_severe()) {
3481 		vm_severe_waiters++;
3482 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3483 		    "vmwait", 0);
3484 	}
3485 	mtx_unlock(&vm_domainset_lock);
3486 }
3487 
3488 u_int
3489 vm_wait_count(void)
3490 {
3491 
3492 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3493 }
3494 
3495 int
3496 vm_wait_doms(const domainset_t *wdoms, int mflags)
3497 {
3498 	int error;
3499 
3500 	error = 0;
3501 
3502 	/*
3503 	 * We use racey wakeup synchronization to avoid expensive global
3504 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3505 	 * To handle this, we only sleep for one tick in this instance.  It
3506 	 * is expected that most allocations for the pageproc will come from
3507 	 * kmem or vm_page_grab* which will use the more specific and
3508 	 * race-free vm_wait_domain().
3509 	 */
3510 	if (curproc == pageproc) {
3511 		mtx_lock(&vm_domainset_lock);
3512 		vm_pageproc_waiters++;
3513 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3514 		    PVM | PDROP | mflags, "pageprocwait", 1);
3515 	} else {
3516 		/*
3517 		 * XXX Ideally we would wait only until the allocation could
3518 		 * be satisfied.  This condition can cause new allocators to
3519 		 * consume all freed pages while old allocators wait.
3520 		 */
3521 		mtx_lock(&vm_domainset_lock);
3522 		if (vm_page_count_min_set(wdoms)) {
3523 			if (pageproc == NULL)
3524 				panic("vm_wait in early boot");
3525 			vm_min_waiters++;
3526 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3527 			    PVM | PDROP | mflags, "vmwait", 0);
3528 		} else
3529 			mtx_unlock(&vm_domainset_lock);
3530 	}
3531 	return (error);
3532 }
3533 
3534 /*
3535  *	vm_wait_domain:
3536  *
3537  *	Sleep until free pages are available for allocation.
3538  *	- Called in various places after failed memory allocations.
3539  */
3540 void
3541 vm_wait_domain(int domain)
3542 {
3543 	struct vm_domain *vmd;
3544 	domainset_t wdom;
3545 
3546 	vmd = VM_DOMAIN(domain);
3547 	vm_domain_free_assert_unlocked(vmd);
3548 
3549 	if (curproc == pageproc) {
3550 		mtx_lock(&vm_domainset_lock);
3551 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3552 			vmd->vmd_pageout_pages_needed = 1;
3553 			msleep(&vmd->vmd_pageout_pages_needed,
3554 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3555 		} else
3556 			mtx_unlock(&vm_domainset_lock);
3557 	} else {
3558 		DOMAINSET_ZERO(&wdom);
3559 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3560 		vm_wait_doms(&wdom, 0);
3561 	}
3562 }
3563 
3564 static int
3565 vm_wait_flags(vm_object_t obj, int mflags)
3566 {
3567 	struct domainset *d;
3568 
3569 	d = NULL;
3570 
3571 	/*
3572 	 * Carefully fetch pointers only once: the struct domainset
3573 	 * itself is ummutable but the pointer might change.
3574 	 */
3575 	if (obj != NULL)
3576 		d = obj->domain.dr_policy;
3577 	if (d == NULL)
3578 		d = curthread->td_domain.dr_policy;
3579 
3580 	return (vm_wait_doms(&d->ds_mask, mflags));
3581 }
3582 
3583 /*
3584  *	vm_wait:
3585  *
3586  *	Sleep until free pages are available for allocation in the
3587  *	affinity domains of the obj.  If obj is NULL, the domain set
3588  *	for the calling thread is used.
3589  *	Called in various places after failed memory allocations.
3590  */
3591 void
3592 vm_wait(vm_object_t obj)
3593 {
3594 	(void)vm_wait_flags(obj, 0);
3595 }
3596 
3597 int
3598 vm_wait_intr(vm_object_t obj)
3599 {
3600 	return (vm_wait_flags(obj, PCATCH));
3601 }
3602 
3603 /*
3604  *	vm_domain_alloc_fail:
3605  *
3606  *	Called when a page allocation function fails.  Informs the
3607  *	pagedaemon and performs the requested wait.  Requires the
3608  *	domain_free and object lock on entry.  Returns with the
3609  *	object lock held and free lock released.  Returns an error when
3610  *	retry is necessary.
3611  *
3612  */
3613 static int
3614 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3615 {
3616 
3617 	vm_domain_free_assert_unlocked(vmd);
3618 
3619 	atomic_add_int(&vmd->vmd_pageout_deficit,
3620 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3621 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3622 		if (object != NULL)
3623 			VM_OBJECT_WUNLOCK(object);
3624 		vm_wait_domain(vmd->vmd_domain);
3625 		if (object != NULL)
3626 			VM_OBJECT_WLOCK(object);
3627 		if (req & VM_ALLOC_WAITOK)
3628 			return (EAGAIN);
3629 	}
3630 
3631 	return (0);
3632 }
3633 
3634 /*
3635  *	vm_waitpfault:
3636  *
3637  *	Sleep until free pages are available for allocation.
3638  *	- Called only in vm_fault so that processes page faulting
3639  *	  can be easily tracked.
3640  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3641  *	  processes will be able to grab memory first.  Do not change
3642  *	  this balance without careful testing first.
3643  */
3644 void
3645 vm_waitpfault(struct domainset *dset, int timo)
3646 {
3647 
3648 	/*
3649 	 * XXX Ideally we would wait only until the allocation could
3650 	 * be satisfied.  This condition can cause new allocators to
3651 	 * consume all freed pages while old allocators wait.
3652 	 */
3653 	mtx_lock(&vm_domainset_lock);
3654 	if (vm_page_count_min_set(&dset->ds_mask)) {
3655 		vm_min_waiters++;
3656 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3657 		    "pfault", timo);
3658 	} else
3659 		mtx_unlock(&vm_domainset_lock);
3660 }
3661 
3662 static struct vm_pagequeue *
3663 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3664 {
3665 
3666 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3667 }
3668 
3669 #ifdef INVARIANTS
3670 static struct vm_pagequeue *
3671 vm_page_pagequeue(vm_page_t m)
3672 {
3673 
3674 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3675 }
3676 #endif
3677 
3678 static __always_inline bool
3679 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3680 {
3681 	vm_page_astate_t tmp;
3682 
3683 	tmp = *old;
3684 	do {
3685 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3686 			return (true);
3687 		counter_u64_add(pqstate_commit_retries, 1);
3688 	} while (old->_bits == tmp._bits);
3689 
3690 	return (false);
3691 }
3692 
3693 /*
3694  * Do the work of committing a queue state update that moves the page out of
3695  * its current queue.
3696  */
3697 static bool
3698 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3699     vm_page_astate_t *old, vm_page_astate_t new)
3700 {
3701 	vm_page_t next;
3702 
3703 	vm_pagequeue_assert_locked(pq);
3704 	KASSERT(vm_page_pagequeue(m) == pq,
3705 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3706 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3707 	    ("%s: invalid queue indices %d %d",
3708 	    __func__, old->queue, new.queue));
3709 
3710 	/*
3711 	 * Once the queue index of the page changes there is nothing
3712 	 * synchronizing with further updates to the page's physical
3713 	 * queue state.  Therefore we must speculatively remove the page
3714 	 * from the queue now and be prepared to roll back if the queue
3715 	 * state update fails.  If the page is not physically enqueued then
3716 	 * we just update its queue index.
3717 	 */
3718 	if ((old->flags & PGA_ENQUEUED) != 0) {
3719 		new.flags &= ~PGA_ENQUEUED;
3720 		next = TAILQ_NEXT(m, plinks.q);
3721 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3722 		vm_pagequeue_cnt_dec(pq);
3723 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3724 			if (next == NULL)
3725 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3726 			else
3727 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3728 			vm_pagequeue_cnt_inc(pq);
3729 			return (false);
3730 		} else {
3731 			return (true);
3732 		}
3733 	} else {
3734 		return (vm_page_pqstate_fcmpset(m, old, new));
3735 	}
3736 }
3737 
3738 static bool
3739 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3740     vm_page_astate_t new)
3741 {
3742 	struct vm_pagequeue *pq;
3743 	vm_page_astate_t as;
3744 	bool ret;
3745 
3746 	pq = _vm_page_pagequeue(m, old->queue);
3747 
3748 	/*
3749 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3750 	 * corresponding page queue lock is held.
3751 	 */
3752 	vm_pagequeue_lock(pq);
3753 	as = vm_page_astate_load(m);
3754 	if (__predict_false(as._bits != old->_bits)) {
3755 		*old = as;
3756 		ret = false;
3757 	} else {
3758 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3759 	}
3760 	vm_pagequeue_unlock(pq);
3761 	return (ret);
3762 }
3763 
3764 /*
3765  * Commit a queue state update that enqueues or requeues a page.
3766  */
3767 static bool
3768 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3769     vm_page_astate_t *old, vm_page_astate_t new)
3770 {
3771 	struct vm_domain *vmd;
3772 
3773 	vm_pagequeue_assert_locked(pq);
3774 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3775 	    ("%s: invalid queue indices %d %d",
3776 	    __func__, old->queue, new.queue));
3777 
3778 	new.flags |= PGA_ENQUEUED;
3779 	if (!vm_page_pqstate_fcmpset(m, old, new))
3780 		return (false);
3781 
3782 	if ((old->flags & PGA_ENQUEUED) != 0)
3783 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3784 	else
3785 		vm_pagequeue_cnt_inc(pq);
3786 
3787 	/*
3788 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3789 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3790 	 * applied, even if it was set first.
3791 	 */
3792 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3793 		vmd = vm_pagequeue_domain(m);
3794 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3795 		    ("%s: invalid page queue for page %p", __func__, m));
3796 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3797 	} else {
3798 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3799 	}
3800 	return (true);
3801 }
3802 
3803 /*
3804  * Commit a queue state update that encodes a request for a deferred queue
3805  * operation.
3806  */
3807 static bool
3808 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3809     vm_page_astate_t new)
3810 {
3811 
3812 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3813 	    ("%s: invalid state, queue %d flags %x",
3814 	    __func__, new.queue, new.flags));
3815 
3816 	if (old->_bits != new._bits &&
3817 	    !vm_page_pqstate_fcmpset(m, old, new))
3818 		return (false);
3819 	vm_page_pqbatch_submit(m, new.queue);
3820 	return (true);
3821 }
3822 
3823 /*
3824  * A generic queue state update function.  This handles more cases than the
3825  * specialized functions above.
3826  */
3827 bool
3828 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3829 {
3830 
3831 	if (old->_bits == new._bits)
3832 		return (true);
3833 
3834 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3835 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3836 			return (false);
3837 		if (new.queue != PQ_NONE)
3838 			vm_page_pqbatch_submit(m, new.queue);
3839 	} else {
3840 		if (!vm_page_pqstate_fcmpset(m, old, new))
3841 			return (false);
3842 		if (new.queue != PQ_NONE &&
3843 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3844 			vm_page_pqbatch_submit(m, new.queue);
3845 	}
3846 	return (true);
3847 }
3848 
3849 /*
3850  * Apply deferred queue state updates to a page.
3851  */
3852 static inline void
3853 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3854 {
3855 	vm_page_astate_t new, old;
3856 
3857 	CRITICAL_ASSERT(curthread);
3858 	vm_pagequeue_assert_locked(pq);
3859 	KASSERT(queue < PQ_COUNT,
3860 	    ("%s: invalid queue index %d", __func__, queue));
3861 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3862 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3863 
3864 	for (old = vm_page_astate_load(m);;) {
3865 		if (__predict_false(old.queue != queue ||
3866 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3867 			counter_u64_add(queue_nops, 1);
3868 			break;
3869 		}
3870 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3871 		    ("%s: page %p is unmanaged", __func__, m));
3872 
3873 		new = old;
3874 		if ((old.flags & PGA_DEQUEUE) != 0) {
3875 			new.flags &= ~PGA_QUEUE_OP_MASK;
3876 			new.queue = PQ_NONE;
3877 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3878 			    m, &old, new))) {
3879 				counter_u64_add(queue_ops, 1);
3880 				break;
3881 			}
3882 		} else {
3883 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3884 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3885 			    m, &old, new))) {
3886 				counter_u64_add(queue_ops, 1);
3887 				break;
3888 			}
3889 		}
3890 	}
3891 }
3892 
3893 static void
3894 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3895     uint8_t queue)
3896 {
3897 	int i;
3898 
3899 	for (i = 0; i < bq->bq_cnt; i++)
3900 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3901 	vm_batchqueue_init(bq);
3902 }
3903 
3904 /*
3905  *	vm_page_pqbatch_submit:		[ internal use only ]
3906  *
3907  *	Enqueue a page in the specified page queue's batched work queue.
3908  *	The caller must have encoded the requested operation in the page
3909  *	structure's a.flags field.
3910  */
3911 void
3912 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3913 {
3914 	struct vm_batchqueue *bq;
3915 	struct vm_pagequeue *pq;
3916 	int domain, slots_remaining;
3917 
3918 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3919 
3920 	domain = vm_page_domain(m);
3921 	critical_enter();
3922 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3923 	slots_remaining = vm_batchqueue_insert(bq, m);
3924 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3925 		/* keep building the bq */
3926 		critical_exit();
3927 		return;
3928 	} else if (slots_remaining > 0 ) {
3929 		/* Try to process the bq if we can get the lock */
3930 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3931 		if (vm_pagequeue_trylock(pq)) {
3932 			vm_pqbatch_process(pq, bq, queue);
3933 			vm_pagequeue_unlock(pq);
3934 		}
3935 		critical_exit();
3936 		return;
3937 	}
3938 	critical_exit();
3939 
3940 	/* if we make it here, the bq is full so wait for the lock */
3941 
3942 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3943 	vm_pagequeue_lock(pq);
3944 	critical_enter();
3945 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3946 	vm_pqbatch_process(pq, bq, queue);
3947 	vm_pqbatch_process_page(pq, m, queue);
3948 	vm_pagequeue_unlock(pq);
3949 	critical_exit();
3950 }
3951 
3952 /*
3953  *	vm_page_pqbatch_drain:		[ internal use only ]
3954  *
3955  *	Force all per-CPU page queue batch queues to be drained.  This is
3956  *	intended for use in severe memory shortages, to ensure that pages
3957  *	do not remain stuck in the batch queues.
3958  */
3959 void
3960 vm_page_pqbatch_drain(void)
3961 {
3962 	struct thread *td;
3963 	struct vm_domain *vmd;
3964 	struct vm_pagequeue *pq;
3965 	int cpu, domain, queue;
3966 
3967 	td = curthread;
3968 	CPU_FOREACH(cpu) {
3969 		thread_lock(td);
3970 		sched_bind(td, cpu);
3971 		thread_unlock(td);
3972 
3973 		for (domain = 0; domain < vm_ndomains; domain++) {
3974 			vmd = VM_DOMAIN(domain);
3975 			for (queue = 0; queue < PQ_COUNT; queue++) {
3976 				pq = &vmd->vmd_pagequeues[queue];
3977 				vm_pagequeue_lock(pq);
3978 				critical_enter();
3979 				vm_pqbatch_process(pq,
3980 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3981 				critical_exit();
3982 				vm_pagequeue_unlock(pq);
3983 			}
3984 		}
3985 	}
3986 	thread_lock(td);
3987 	sched_unbind(td);
3988 	thread_unlock(td);
3989 }
3990 
3991 /*
3992  *	vm_page_dequeue_deferred:	[ internal use only ]
3993  *
3994  *	Request removal of the given page from its current page
3995  *	queue.  Physical removal from the queue may be deferred
3996  *	indefinitely.
3997  */
3998 void
3999 vm_page_dequeue_deferred(vm_page_t m)
4000 {
4001 	vm_page_astate_t new, old;
4002 
4003 	old = vm_page_astate_load(m);
4004 	do {
4005 		if (old.queue == PQ_NONE) {
4006 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4007 			    ("%s: page %p has unexpected queue state",
4008 			    __func__, m));
4009 			break;
4010 		}
4011 		new = old;
4012 		new.flags |= PGA_DEQUEUE;
4013 	} while (!vm_page_pqstate_commit_request(m, &old, new));
4014 }
4015 
4016 /*
4017  *	vm_page_dequeue:
4018  *
4019  *	Remove the page from whichever page queue it's in, if any, before
4020  *	returning.
4021  */
4022 void
4023 vm_page_dequeue(vm_page_t m)
4024 {
4025 	vm_page_astate_t new, old;
4026 
4027 	old = vm_page_astate_load(m);
4028 	do {
4029 		if (old.queue == PQ_NONE) {
4030 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4031 			    ("%s: page %p has unexpected queue state",
4032 			    __func__, m));
4033 			break;
4034 		}
4035 		new = old;
4036 		new.flags &= ~PGA_QUEUE_OP_MASK;
4037 		new.queue = PQ_NONE;
4038 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
4039 
4040 }
4041 
4042 /*
4043  * Schedule the given page for insertion into the specified page queue.
4044  * Physical insertion of the page may be deferred indefinitely.
4045  */
4046 static void
4047 vm_page_enqueue(vm_page_t m, uint8_t queue)
4048 {
4049 
4050 	KASSERT(m->a.queue == PQ_NONE &&
4051 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
4052 	    ("%s: page %p is already enqueued", __func__, m));
4053 	KASSERT(m->ref_count > 0,
4054 	    ("%s: page %p does not carry any references", __func__, m));
4055 
4056 	m->a.queue = queue;
4057 	if ((m->a.flags & PGA_REQUEUE) == 0)
4058 		vm_page_aflag_set(m, PGA_REQUEUE);
4059 	vm_page_pqbatch_submit(m, queue);
4060 }
4061 
4062 /*
4063  *	vm_page_free_prep:
4064  *
4065  *	Prepares the given page to be put on the free list,
4066  *	disassociating it from any VM object. The caller may return
4067  *	the page to the free list only if this function returns true.
4068  *
4069  *	The object, if it exists, must be locked, and then the page must
4070  *	be xbusy.  Otherwise the page must be not busied.  A managed
4071  *	page must be unmapped.
4072  */
4073 static bool
4074 vm_page_free_prep(vm_page_t m)
4075 {
4076 
4077 	/*
4078 	 * Synchronize with threads that have dropped a reference to this
4079 	 * page.
4080 	 */
4081 	atomic_thread_fence_acq();
4082 
4083 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
4084 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
4085 		uint64_t *p;
4086 		int i;
4087 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4088 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4089 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4090 			    m, i, (uintmax_t)*p));
4091 	}
4092 #endif
4093 	KASSERT((m->flags & PG_NOFREE) == 0,
4094 	    ("%s: attempting to free a PG_NOFREE page", __func__));
4095 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4096 		KASSERT(!pmap_page_is_mapped(m),
4097 		    ("vm_page_free_prep: freeing mapped page %p", m));
4098 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4099 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4100 	} else {
4101 		KASSERT(m->a.queue == PQ_NONE,
4102 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4103 	}
4104 	VM_CNT_INC(v_tfree);
4105 
4106 	if (m->object != NULL) {
4107 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
4108 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
4109 		    ("vm_page_free_prep: managed flag mismatch for page %p",
4110 		    m));
4111 		vm_page_assert_xbusied(m);
4112 
4113 		/*
4114 		 * The object reference can be released without an atomic
4115 		 * operation.
4116 		 */
4117 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4118 		    m->ref_count == VPRC_OBJREF,
4119 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
4120 		    m, m->ref_count));
4121 		vm_page_object_remove(m);
4122 		m->ref_count -= VPRC_OBJREF;
4123 	} else
4124 		vm_page_assert_unbusied(m);
4125 
4126 	vm_page_busy_free(m);
4127 
4128 	/*
4129 	 * If fictitious remove object association and
4130 	 * return.
4131 	 */
4132 	if ((m->flags & PG_FICTITIOUS) != 0) {
4133 		KASSERT(m->ref_count == 1,
4134 		    ("fictitious page %p is referenced", m));
4135 		KASSERT(m->a.queue == PQ_NONE,
4136 		    ("fictitious page %p is queued", m));
4137 		return (false);
4138 	}
4139 
4140 	/*
4141 	 * Pages need not be dequeued before they are returned to the physical
4142 	 * memory allocator, but they must at least be marked for a deferred
4143 	 * dequeue.
4144 	 */
4145 	if ((m->oflags & VPO_UNMANAGED) == 0)
4146 		vm_page_dequeue_deferred(m);
4147 
4148 	m->valid = 0;
4149 	vm_page_undirty(m);
4150 
4151 	if (m->ref_count != 0)
4152 		panic("vm_page_free_prep: page %p has references", m);
4153 
4154 	/*
4155 	 * Restore the default memory attribute to the page.
4156 	 */
4157 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4158 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4159 
4160 #if VM_NRESERVLEVEL > 0
4161 	/*
4162 	 * Determine whether the page belongs to a reservation.  If the page was
4163 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4164 	 * as an optimization, we avoid the check in that case.
4165 	 */
4166 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4167 		return (false);
4168 #endif
4169 
4170 	return (true);
4171 }
4172 
4173 /*
4174  *	vm_page_free_toq:
4175  *
4176  *	Returns the given page to the free list, disassociating it
4177  *	from any VM object.
4178  *
4179  *	The object must be locked.  The page must be exclusively busied if it
4180  *	belongs to an object.
4181  */
4182 static void
4183 vm_page_free_toq(vm_page_t m)
4184 {
4185 	struct vm_domain *vmd;
4186 	uma_zone_t zone;
4187 
4188 	if (!vm_page_free_prep(m))
4189 		return;
4190 
4191 	vmd = vm_pagequeue_domain(m);
4192 	zone = vmd->vmd_pgcache[m->pool].zone;
4193 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4194 		uma_zfree(zone, m);
4195 		return;
4196 	}
4197 	vm_domain_free_lock(vmd);
4198 	vm_phys_free_pages(m, 0);
4199 	vm_domain_free_unlock(vmd);
4200 	vm_domain_freecnt_inc(vmd, 1);
4201 }
4202 
4203 /*
4204  *	vm_page_free_pages_toq:
4205  *
4206  *	Returns a list of pages to the free list, disassociating it
4207  *	from any VM object.  In other words, this is equivalent to
4208  *	calling vm_page_free_toq() for each page of a list of VM objects.
4209  */
4210 int
4211 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4212 {
4213 	vm_page_t m;
4214 	int count;
4215 
4216 	if (SLIST_EMPTY(free))
4217 		return (0);
4218 
4219 	count = 0;
4220 	while ((m = SLIST_FIRST(free)) != NULL) {
4221 		count++;
4222 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4223 		vm_page_free_toq(m);
4224 	}
4225 
4226 	if (update_wire_count)
4227 		vm_wire_sub(count);
4228 	return (count);
4229 }
4230 
4231 /*
4232  * Mark this page as wired down.  For managed pages, this prevents reclamation
4233  * by the page daemon, or when the containing object, if any, is destroyed.
4234  */
4235 void
4236 vm_page_wire(vm_page_t m)
4237 {
4238 	u_int old;
4239 
4240 #ifdef INVARIANTS
4241 	if (m->object != NULL && !vm_page_busied(m) &&
4242 	    !vm_object_busied(m->object))
4243 		VM_OBJECT_ASSERT_LOCKED(m->object);
4244 #endif
4245 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4246 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4247 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4248 
4249 	old = atomic_fetchadd_int(&m->ref_count, 1);
4250 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4251 	    ("vm_page_wire: counter overflow for page %p", m));
4252 	if (VPRC_WIRE_COUNT(old) == 0) {
4253 		if ((m->oflags & VPO_UNMANAGED) == 0)
4254 			vm_page_aflag_set(m, PGA_DEQUEUE);
4255 		vm_wire_add(1);
4256 	}
4257 }
4258 
4259 /*
4260  * Attempt to wire a mapped page following a pmap lookup of that page.
4261  * This may fail if a thread is concurrently tearing down mappings of the page.
4262  * The transient failure is acceptable because it translates to the
4263  * failure of the caller pmap_extract_and_hold(), which should be then
4264  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4265  */
4266 bool
4267 vm_page_wire_mapped(vm_page_t m)
4268 {
4269 	u_int old;
4270 
4271 	old = atomic_load_int(&m->ref_count);
4272 	do {
4273 		KASSERT(old > 0,
4274 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4275 		if ((old & VPRC_BLOCKED) != 0)
4276 			return (false);
4277 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4278 
4279 	if (VPRC_WIRE_COUNT(old) == 0) {
4280 		if ((m->oflags & VPO_UNMANAGED) == 0)
4281 			vm_page_aflag_set(m, PGA_DEQUEUE);
4282 		vm_wire_add(1);
4283 	}
4284 	return (true);
4285 }
4286 
4287 /*
4288  * Release a wiring reference to a managed page.  If the page still belongs to
4289  * an object, update its position in the page queues to reflect the reference.
4290  * If the wiring was the last reference to the page, free the page.
4291  */
4292 static void
4293 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4294 {
4295 	u_int old;
4296 
4297 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4298 	    ("%s: page %p is unmanaged", __func__, m));
4299 
4300 	/*
4301 	 * Update LRU state before releasing the wiring reference.
4302 	 * Use a release store when updating the reference count to
4303 	 * synchronize with vm_page_free_prep().
4304 	 */
4305 	old = atomic_load_int(&m->ref_count);
4306 	do {
4307 		u_int count;
4308 
4309 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4310 		    ("vm_page_unwire: wire count underflow for page %p", m));
4311 
4312 		count = old & ~VPRC_BLOCKED;
4313 		if (count > VPRC_OBJREF + 1) {
4314 			/*
4315 			 * The page has at least one other wiring reference.  An
4316 			 * earlier iteration of this loop may have called
4317 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4318 			 * re-set it if necessary.
4319 			 */
4320 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4321 				vm_page_aflag_set(m, PGA_DEQUEUE);
4322 		} else if (count == VPRC_OBJREF + 1) {
4323 			/*
4324 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4325 			 * update the page's queue state to reflect the
4326 			 * reference.  If the page does not belong to an object
4327 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4328 			 * clear leftover queue state.
4329 			 */
4330 			vm_page_release_toq(m, nqueue, noreuse);
4331 		} else if (count == 1) {
4332 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4333 		}
4334 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4335 
4336 	if (VPRC_WIRE_COUNT(old) == 1) {
4337 		vm_wire_sub(1);
4338 		if (old == 1)
4339 			vm_page_free(m);
4340 	}
4341 }
4342 
4343 /*
4344  * Release one wiring of the specified page, potentially allowing it to be
4345  * paged out.
4346  *
4347  * Only managed pages belonging to an object can be paged out.  If the number
4348  * of wirings transitions to zero and the page is eligible for page out, then
4349  * the page is added to the specified paging queue.  If the released wiring
4350  * represented the last reference to the page, the page is freed.
4351  */
4352 void
4353 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4354 {
4355 
4356 	KASSERT(nqueue < PQ_COUNT,
4357 	    ("vm_page_unwire: invalid queue %u request for page %p",
4358 	    nqueue, m));
4359 
4360 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4361 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4362 			vm_page_free(m);
4363 		return;
4364 	}
4365 	vm_page_unwire_managed(m, nqueue, false);
4366 }
4367 
4368 /*
4369  * Unwire a page without (re-)inserting it into a page queue.  It is up
4370  * to the caller to enqueue, requeue, or free the page as appropriate.
4371  * In most cases involving managed pages, vm_page_unwire() should be used
4372  * instead.
4373  */
4374 bool
4375 vm_page_unwire_noq(vm_page_t m)
4376 {
4377 	u_int old;
4378 
4379 	old = vm_page_drop(m, 1);
4380 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4381 	    ("%s: counter underflow for page %p", __func__,  m));
4382 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4383 	    ("%s: missing ref on fictitious page %p", __func__, m));
4384 
4385 	if (VPRC_WIRE_COUNT(old) > 1)
4386 		return (false);
4387 	if ((m->oflags & VPO_UNMANAGED) == 0)
4388 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4389 	vm_wire_sub(1);
4390 	return (true);
4391 }
4392 
4393 /*
4394  * Ensure that the page ends up in the specified page queue.  If the page is
4395  * active or being moved to the active queue, ensure that its act_count is
4396  * at least ACT_INIT but do not otherwise mess with it.
4397  */
4398 static __always_inline void
4399 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4400 {
4401 	vm_page_astate_t old, new;
4402 
4403 	KASSERT(m->ref_count > 0,
4404 	    ("%s: page %p does not carry any references", __func__, m));
4405 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4406 	    ("%s: invalid flags %x", __func__, nflag));
4407 
4408 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4409 		return;
4410 
4411 	old = vm_page_astate_load(m);
4412 	do {
4413 		if ((old.flags & PGA_DEQUEUE) != 0)
4414 			break;
4415 		new = old;
4416 		new.flags &= ~PGA_QUEUE_OP_MASK;
4417 		if (nqueue == PQ_ACTIVE)
4418 			new.act_count = max(old.act_count, ACT_INIT);
4419 		if (old.queue == nqueue) {
4420 			/*
4421 			 * There is no need to requeue pages already in the
4422 			 * active queue.
4423 			 */
4424 			if (nqueue != PQ_ACTIVE ||
4425 			    (old.flags & PGA_ENQUEUED) == 0)
4426 				new.flags |= nflag;
4427 		} else {
4428 			new.flags |= nflag;
4429 			new.queue = nqueue;
4430 		}
4431 	} while (!vm_page_pqstate_commit(m, &old, new));
4432 }
4433 
4434 /*
4435  * Put the specified page on the active list (if appropriate).
4436  */
4437 void
4438 vm_page_activate(vm_page_t m)
4439 {
4440 
4441 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4442 }
4443 
4444 /*
4445  * Move the specified page to the tail of the inactive queue, or requeue
4446  * the page if it is already in the inactive queue.
4447  */
4448 void
4449 vm_page_deactivate(vm_page_t m)
4450 {
4451 
4452 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4453 }
4454 
4455 void
4456 vm_page_deactivate_noreuse(vm_page_t m)
4457 {
4458 
4459 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4460 }
4461 
4462 /*
4463  * Put a page in the laundry, or requeue it if it is already there.
4464  */
4465 void
4466 vm_page_launder(vm_page_t m)
4467 {
4468 
4469 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4470 }
4471 
4472 /*
4473  * Put a page in the PQ_UNSWAPPABLE holding queue.
4474  */
4475 void
4476 vm_page_unswappable(vm_page_t m)
4477 {
4478 
4479 	VM_OBJECT_ASSERT_LOCKED(m->object);
4480 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4481 	    ("page %p already unswappable", m));
4482 
4483 	vm_page_dequeue(m);
4484 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4485 }
4486 
4487 /*
4488  * Release a page back to the page queues in preparation for unwiring.
4489  */
4490 static void
4491 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4492 {
4493 	vm_page_astate_t old, new;
4494 	uint16_t nflag;
4495 
4496 	/*
4497 	 * Use a check of the valid bits to determine whether we should
4498 	 * accelerate reclamation of the page.  The object lock might not be
4499 	 * held here, in which case the check is racy.  At worst we will either
4500 	 * accelerate reclamation of a valid page and violate LRU, or
4501 	 * unnecessarily defer reclamation of an invalid page.
4502 	 *
4503 	 * If we were asked to not cache the page, place it near the head of the
4504 	 * inactive queue so that is reclaimed sooner.
4505 	 */
4506 	if (noreuse || vm_page_none_valid(m)) {
4507 		nqueue = PQ_INACTIVE;
4508 		nflag = PGA_REQUEUE_HEAD;
4509 	} else {
4510 		nflag = PGA_REQUEUE;
4511 	}
4512 
4513 	old = vm_page_astate_load(m);
4514 	do {
4515 		new = old;
4516 
4517 		/*
4518 		 * If the page is already in the active queue and we are not
4519 		 * trying to accelerate reclamation, simply mark it as
4520 		 * referenced and avoid any queue operations.
4521 		 */
4522 		new.flags &= ~PGA_QUEUE_OP_MASK;
4523 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4524 		    (old.flags & PGA_ENQUEUED) != 0)
4525 			new.flags |= PGA_REFERENCED;
4526 		else {
4527 			new.flags |= nflag;
4528 			new.queue = nqueue;
4529 		}
4530 	} while (!vm_page_pqstate_commit(m, &old, new));
4531 }
4532 
4533 /*
4534  * Unwire a page and either attempt to free it or re-add it to the page queues.
4535  */
4536 void
4537 vm_page_release(vm_page_t m, int flags)
4538 {
4539 	vm_object_t object;
4540 
4541 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4542 	    ("vm_page_release: page %p is unmanaged", m));
4543 
4544 	if ((flags & VPR_TRYFREE) != 0) {
4545 		for (;;) {
4546 			object = atomic_load_ptr(&m->object);
4547 			if (object == NULL)
4548 				break;
4549 			/* Depends on type-stability. */
4550 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4551 				break;
4552 			if (object == m->object) {
4553 				vm_page_release_locked(m, flags);
4554 				VM_OBJECT_WUNLOCK(object);
4555 				return;
4556 			}
4557 			VM_OBJECT_WUNLOCK(object);
4558 		}
4559 	}
4560 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4561 }
4562 
4563 /* See vm_page_release(). */
4564 void
4565 vm_page_release_locked(vm_page_t m, int flags)
4566 {
4567 
4568 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4569 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4570 	    ("vm_page_release_locked: page %p is unmanaged", m));
4571 
4572 	if (vm_page_unwire_noq(m)) {
4573 		if ((flags & VPR_TRYFREE) != 0 &&
4574 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4575 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4576 			/*
4577 			 * An unlocked lookup may have wired the page before the
4578 			 * busy lock was acquired, in which case the page must
4579 			 * not be freed.
4580 			 */
4581 			if (__predict_true(!vm_page_wired(m))) {
4582 				vm_page_free(m);
4583 				return;
4584 			}
4585 			vm_page_xunbusy(m);
4586 		} else {
4587 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4588 		}
4589 	}
4590 }
4591 
4592 static bool
4593 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4594 {
4595 	u_int old;
4596 
4597 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4598 	    ("vm_page_try_blocked_op: page %p has no object", m));
4599 	KASSERT(vm_page_busied(m),
4600 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4601 	VM_OBJECT_ASSERT_LOCKED(m->object);
4602 
4603 	old = atomic_load_int(&m->ref_count);
4604 	do {
4605 		KASSERT(old != 0,
4606 		    ("vm_page_try_blocked_op: page %p has no references", m));
4607 		KASSERT((old & VPRC_BLOCKED) == 0,
4608 		    ("vm_page_try_blocked_op: page %p blocks wirings", m));
4609 		if (VPRC_WIRE_COUNT(old) != 0)
4610 			return (false);
4611 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4612 
4613 	(op)(m);
4614 
4615 	/*
4616 	 * If the object is read-locked, new wirings may be created via an
4617 	 * object lookup.
4618 	 */
4619 	old = vm_page_drop(m, VPRC_BLOCKED);
4620 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4621 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4622 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4623 	    old, m));
4624 	return (true);
4625 }
4626 
4627 /*
4628  * Atomically check for wirings and remove all mappings of the page.
4629  */
4630 bool
4631 vm_page_try_remove_all(vm_page_t m)
4632 {
4633 
4634 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4635 }
4636 
4637 /*
4638  * Atomically check for wirings and remove all writeable mappings of the page.
4639  */
4640 bool
4641 vm_page_try_remove_write(vm_page_t m)
4642 {
4643 
4644 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4645 }
4646 
4647 /*
4648  * vm_page_advise
4649  *
4650  * 	Apply the specified advice to the given page.
4651  */
4652 void
4653 vm_page_advise(vm_page_t m, int advice)
4654 {
4655 
4656 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4657 	vm_page_assert_xbusied(m);
4658 
4659 	if (advice == MADV_FREE)
4660 		/*
4661 		 * Mark the page clean.  This will allow the page to be freed
4662 		 * without first paging it out.  MADV_FREE pages are often
4663 		 * quickly reused by malloc(3), so we do not do anything that
4664 		 * would result in a page fault on a later access.
4665 		 */
4666 		vm_page_undirty(m);
4667 	else if (advice != MADV_DONTNEED) {
4668 		if (advice == MADV_WILLNEED)
4669 			vm_page_activate(m);
4670 		return;
4671 	}
4672 
4673 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4674 		vm_page_dirty(m);
4675 
4676 	/*
4677 	 * Clear any references to the page.  Otherwise, the page daemon will
4678 	 * immediately reactivate the page.
4679 	 */
4680 	vm_page_aflag_clear(m, PGA_REFERENCED);
4681 
4682 	/*
4683 	 * Place clean pages near the head of the inactive queue rather than
4684 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4685 	 * the page will be reused quickly.  Dirty pages not already in the
4686 	 * laundry are moved there.
4687 	 */
4688 	if (m->dirty == 0)
4689 		vm_page_deactivate_noreuse(m);
4690 	else if (!vm_page_in_laundry(m))
4691 		vm_page_launder(m);
4692 }
4693 
4694 /*
4695  *	vm_page_grab_release
4696  *
4697  *	Helper routine for grab functions to release busy on return.
4698  */
4699 static inline void
4700 vm_page_grab_release(vm_page_t m, int allocflags)
4701 {
4702 
4703 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4704 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4705 			vm_page_sunbusy(m);
4706 		else
4707 			vm_page_xunbusy(m);
4708 	}
4709 }
4710 
4711 /*
4712  *	vm_page_grab_sleep
4713  *
4714  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4715  *	if the caller should retry and false otherwise.
4716  *
4717  *	If the object is locked on entry the object will be unlocked with
4718  *	false returns and still locked but possibly having been dropped
4719  *	with true returns.
4720  */
4721 static bool
4722 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4723     const char *wmesg, int allocflags, bool locked)
4724 {
4725 
4726 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4727 		return (false);
4728 
4729 	/*
4730 	 * Reference the page before unlocking and sleeping so that
4731 	 * the page daemon is less likely to reclaim it.
4732 	 */
4733 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4734 		vm_page_reference(m);
4735 
4736 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4737 	    locked)
4738 		VM_OBJECT_WLOCK(object);
4739 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4740 		return (false);
4741 
4742 	return (true);
4743 }
4744 
4745 /*
4746  * Assert that the grab flags are valid.
4747  */
4748 static inline void
4749 vm_page_grab_check(int allocflags)
4750 {
4751 
4752 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4753 	    (allocflags & VM_ALLOC_WIRED) != 0,
4754 	    ("vm_page_grab*: the pages must be busied or wired"));
4755 
4756 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4757 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4758 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4759 }
4760 
4761 /*
4762  * Calculate the page allocation flags for grab.
4763  */
4764 static inline int
4765 vm_page_grab_pflags(int allocflags)
4766 {
4767 	int pflags;
4768 
4769 	pflags = allocflags &
4770 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4771 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4772 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4773 		pflags |= VM_ALLOC_WAITFAIL;
4774 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4775 		pflags |= VM_ALLOC_SBUSY;
4776 
4777 	return (pflags);
4778 }
4779 
4780 /*
4781  * Grab a page, waiting until we are waken up due to the page
4782  * changing state.  We keep on waiting, if the page continues
4783  * to be in the object.  If the page doesn't exist, first allocate it
4784  * and then conditionally zero it.
4785  *
4786  * This routine may sleep.
4787  *
4788  * The object must be locked on entry.  The lock will, however, be released
4789  * and reacquired if the routine sleeps.
4790  */
4791 vm_page_t
4792 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4793 {
4794 	vm_page_t m;
4795 
4796 	VM_OBJECT_ASSERT_WLOCKED(object);
4797 	vm_page_grab_check(allocflags);
4798 
4799 retrylookup:
4800 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4801 		if (!vm_page_tryacquire(m, allocflags)) {
4802 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4803 			    allocflags, true))
4804 				goto retrylookup;
4805 			return (NULL);
4806 		}
4807 		goto out;
4808 	}
4809 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4810 		return (NULL);
4811 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4812 	if (m == NULL) {
4813 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4814 			return (NULL);
4815 		goto retrylookup;
4816 	}
4817 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4818 		pmap_zero_page(m);
4819 
4820 out:
4821 	vm_page_grab_release(m, allocflags);
4822 
4823 	return (m);
4824 }
4825 
4826 /*
4827  * Attempt to validate a page, locklessly acquiring it if necessary, given a
4828  * (object, pindex) tuple and either an invalided page or NULL.  The resulting
4829  * page will be validated against the identity tuple, and busied or wired as
4830  * requested.  A NULL page returned guarantees that the page was not in radix at
4831  * the time of the call but callers must perform higher level synchronization or
4832  * retry the operation under a lock if they require an atomic answer.  This is
4833  * the only lock free validation routine, other routines can depend on the
4834  * resulting page state.
4835  *
4836  * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to
4837  * caller flags.
4838  */
4839 #define PAGE_NOT_ACQUIRED ((vm_page_t)1)
4840 static vm_page_t
4841 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m,
4842     int allocflags)
4843 {
4844 	if (m == NULL)
4845 		m = vm_page_lookup_unlocked(object, pindex);
4846 	for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) {
4847 		if (vm_page_trybusy(m, allocflags)) {
4848 			if (m->object == object && m->pindex == pindex) {
4849 				if ((allocflags & VM_ALLOC_WIRED) != 0)
4850 					vm_page_wire(m);
4851 				vm_page_grab_release(m, allocflags);
4852 				break;
4853 			}
4854 			/* relookup. */
4855 			vm_page_busy_release(m);
4856 			cpu_spinwait();
4857 			continue;
4858 		}
4859 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4860 		    allocflags, false))
4861 			return (PAGE_NOT_ACQUIRED);
4862 	}
4863 	return (m);
4864 }
4865 
4866 /*
4867  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4868  * is not set.
4869  */
4870 vm_page_t
4871 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4872 {
4873 	vm_page_t m;
4874 
4875 	vm_page_grab_check(allocflags);
4876 	m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags);
4877 	if (m == PAGE_NOT_ACQUIRED)
4878 		return (NULL);
4879 	if (m != NULL)
4880 		return (m);
4881 
4882 	/*
4883 	 * The radix lockless lookup should never return a false negative
4884 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4885 	 * was no page present at the instant of the call.  A NOCREAT caller
4886 	 * must handle create races gracefully.
4887 	 */
4888 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4889 		return (NULL);
4890 
4891 	VM_OBJECT_WLOCK(object);
4892 	m = vm_page_grab(object, pindex, allocflags);
4893 	VM_OBJECT_WUNLOCK(object);
4894 
4895 	return (m);
4896 }
4897 
4898 /*
4899  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4900  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4901  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4902  * in simultaneously.  Additional pages will be left on a paging queue but
4903  * will neither be wired nor busy regardless of allocflags.
4904  */
4905 int
4906 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4907 {
4908 	vm_page_t m;
4909 	vm_page_t ma[VM_INITIAL_PAGEIN];
4910 	int after, i, pflags, rv;
4911 
4912 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4913 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4914 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4915 	KASSERT((allocflags &
4916 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4917 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4918 	VM_OBJECT_ASSERT_WLOCKED(object);
4919 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4920 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4921 	pflags |= VM_ALLOC_WAITFAIL;
4922 
4923 retrylookup:
4924 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4925 		/*
4926 		 * If the page is fully valid it can only become invalid
4927 		 * with the object lock held.  If it is not valid it can
4928 		 * become valid with the busy lock held.  Therefore, we
4929 		 * may unnecessarily lock the exclusive busy here if we
4930 		 * race with I/O completion not using the object lock.
4931 		 * However, we will not end up with an invalid page and a
4932 		 * shared lock.
4933 		 */
4934 		if (!vm_page_trybusy(m,
4935 		    vm_page_all_valid(m) ? allocflags : 0)) {
4936 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4937 			    allocflags, true);
4938 			goto retrylookup;
4939 		}
4940 		if (vm_page_all_valid(m))
4941 			goto out;
4942 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4943 			vm_page_busy_release(m);
4944 			*mp = NULL;
4945 			return (VM_PAGER_FAIL);
4946 		}
4947 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4948 		*mp = NULL;
4949 		return (VM_PAGER_FAIL);
4950 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4951 		if (!vm_pager_can_alloc_page(object, pindex)) {
4952 			*mp = NULL;
4953 			return (VM_PAGER_AGAIN);
4954 		}
4955 		goto retrylookup;
4956 	}
4957 
4958 	vm_page_assert_xbusied(m);
4959 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4960 		after = MIN(after, VM_INITIAL_PAGEIN);
4961 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4962 		after = MAX(after, 1);
4963 		ma[0] = m;
4964 		for (i = 1; i < after; i++) {
4965 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4966 				if (vm_page_any_valid(ma[i]) ||
4967 				    !vm_page_tryxbusy(ma[i]))
4968 					break;
4969 			} else {
4970 				ma[i] = vm_page_alloc(object, m->pindex + i,
4971 				    VM_ALLOC_NORMAL);
4972 				if (ma[i] == NULL)
4973 					break;
4974 			}
4975 		}
4976 		after = i;
4977 		vm_object_pip_add(object, after);
4978 		VM_OBJECT_WUNLOCK(object);
4979 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4980 		VM_OBJECT_WLOCK(object);
4981 		vm_object_pip_wakeupn(object, after);
4982 		/* Pager may have replaced a page. */
4983 		m = ma[0];
4984 		if (rv != VM_PAGER_OK) {
4985 			for (i = 0; i < after; i++) {
4986 				if (!vm_page_wired(ma[i]))
4987 					vm_page_free(ma[i]);
4988 				else
4989 					vm_page_xunbusy(ma[i]);
4990 			}
4991 			*mp = NULL;
4992 			return (rv);
4993 		}
4994 		for (i = 1; i < after; i++)
4995 			vm_page_readahead_finish(ma[i]);
4996 		MPASS(vm_page_all_valid(m));
4997 	} else {
4998 		vm_page_zero_invalid(m, TRUE);
4999 	}
5000 out:
5001 	if ((allocflags & VM_ALLOC_WIRED) != 0)
5002 		vm_page_wire(m);
5003 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
5004 		vm_page_busy_downgrade(m);
5005 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
5006 		vm_page_busy_release(m);
5007 	*mp = m;
5008 	return (VM_PAGER_OK);
5009 }
5010 
5011 /*
5012  * Locklessly grab a valid page.  If the page is not valid or not yet
5013  * allocated this will fall back to the object lock method.
5014  */
5015 int
5016 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
5017     vm_pindex_t pindex, int allocflags)
5018 {
5019 	vm_page_t m;
5020 	int flags;
5021 	int error;
5022 
5023 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
5024 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
5025 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
5026 	    "mismatch"));
5027 	KASSERT((allocflags &
5028 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
5029 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
5030 
5031 	/*
5032 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
5033 	 * before we can inspect the valid field and return a wired page.
5034 	 */
5035 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
5036 	vm_page_grab_check(flags);
5037 	m = vm_page_acquire_unlocked(object, pindex, NULL, flags);
5038 	if (m == PAGE_NOT_ACQUIRED)
5039 		return (VM_PAGER_FAIL);
5040 	if (m != NULL) {
5041 		if (vm_page_all_valid(m)) {
5042 			if ((allocflags & VM_ALLOC_WIRED) != 0)
5043 				vm_page_wire(m);
5044 			vm_page_grab_release(m, allocflags);
5045 			*mp = m;
5046 			return (VM_PAGER_OK);
5047 		}
5048 		vm_page_busy_release(m);
5049 	}
5050 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5051 		*mp = NULL;
5052 		return (VM_PAGER_FAIL);
5053 	}
5054 	VM_OBJECT_WLOCK(object);
5055 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
5056 	VM_OBJECT_WUNLOCK(object);
5057 
5058 	return (error);
5059 }
5060 
5061 /*
5062  * Return the specified range of pages from the given object.  For each
5063  * page offset within the range, if a page already exists within the object
5064  * at that offset and it is busy, then wait for it to change state.  If,
5065  * instead, the page doesn't exist, then allocate it.
5066  *
5067  * The caller must always specify an allocation class.
5068  *
5069  * allocation classes:
5070  *	VM_ALLOC_NORMAL		normal process request
5071  *	VM_ALLOC_SYSTEM		system *really* needs the pages
5072  *
5073  * The caller must always specify that the pages are to be busied and/or
5074  * wired.
5075  *
5076  * optional allocation flags:
5077  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5078  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5079  *	VM_ALLOC_NOWAIT		do not sleep
5080  *	VM_ALLOC_SBUSY		set page to sbusy state
5081  *	VM_ALLOC_WIRED		wire the pages
5082  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5083  *
5084  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5085  * may return a partial prefix of the requested range.
5086  */
5087 int
5088 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5089     vm_page_t *ma, int count)
5090 {
5091 	vm_page_t m, mpred;
5092 	int pflags;
5093 	int i;
5094 
5095 	VM_OBJECT_ASSERT_WLOCKED(object);
5096 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5097 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5098 	KASSERT(count > 0,
5099 	    ("vm_page_grab_pages: invalid page count %d", count));
5100 	vm_page_grab_check(allocflags);
5101 
5102 	pflags = vm_page_grab_pflags(allocflags);
5103 	i = 0;
5104 retrylookup:
5105 	m = vm_page_mpred(object, pindex + i);
5106 	if (m == NULL || m->pindex != pindex + i) {
5107 		mpred = m;
5108 		m = NULL;
5109 	} else
5110 		mpred = TAILQ_PREV(m, pglist, listq);
5111 	for (; i < count; i++) {
5112 		if (m != NULL) {
5113 			if (!vm_page_tryacquire(m, allocflags)) {
5114 				if (vm_page_grab_sleep(object, m, pindex + i,
5115 				    "grbmaw", allocflags, true))
5116 					goto retrylookup;
5117 				break;
5118 			}
5119 		} else {
5120 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5121 				break;
5122 			m = vm_page_alloc_after(object, pindex + i,
5123 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5124 			if (m == NULL) {
5125 				if ((allocflags & (VM_ALLOC_NOWAIT |
5126 				    VM_ALLOC_WAITFAIL)) != 0)
5127 					break;
5128 				goto retrylookup;
5129 			}
5130 		}
5131 		if (vm_page_none_valid(m) &&
5132 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5133 			if ((m->flags & PG_ZERO) == 0)
5134 				pmap_zero_page(m);
5135 			vm_page_valid(m);
5136 		}
5137 		vm_page_grab_release(m, allocflags);
5138 		ma[i] = mpred = m;
5139 		m = vm_page_next(m);
5140 	}
5141 	return (i);
5142 }
5143 
5144 /*
5145  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5146  * and will fall back to the locked variant to handle allocation.
5147  */
5148 int
5149 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5150     int allocflags, vm_page_t *ma, int count)
5151 {
5152 	vm_page_t m;
5153 	int flags;
5154 	int i;
5155 
5156 	KASSERT(count > 0,
5157 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5158 	vm_page_grab_check(allocflags);
5159 
5160 	/*
5161 	 * Modify flags for lockless acquire to hold the page until we
5162 	 * set it valid if necessary.
5163 	 */
5164 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5165 	vm_page_grab_check(flags);
5166 	m = NULL;
5167 	for (i = 0; i < count; i++, pindex++) {
5168 		/*
5169 		 * We may see a false NULL here because the previous page has
5170 		 * been removed or just inserted and the list is loaded without
5171 		 * barriers.  Switch to radix to verify.
5172 		 */
5173 		if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex ||
5174 		    atomic_load_ptr(&m->object) != object) {
5175 			/*
5176 			 * This guarantees the result is instantaneously
5177 			 * correct.
5178 			 */
5179 			m = NULL;
5180 		}
5181 		m = vm_page_acquire_unlocked(object, pindex, m, flags);
5182 		if (m == PAGE_NOT_ACQUIRED)
5183 			return (i);
5184 		if (m == NULL)
5185 			break;
5186 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5187 			if ((m->flags & PG_ZERO) == 0)
5188 				pmap_zero_page(m);
5189 			vm_page_valid(m);
5190 		}
5191 		/* m will still be wired or busy according to flags. */
5192 		vm_page_grab_release(m, allocflags);
5193 		ma[i] = m;
5194 		m = TAILQ_NEXT(m, listq);
5195 	}
5196 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5197 		return (i);
5198 	count -= i;
5199 	VM_OBJECT_WLOCK(object);
5200 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5201 	VM_OBJECT_WUNLOCK(object);
5202 
5203 	return (i);
5204 }
5205 
5206 /*
5207  * Mapping function for valid or dirty bits in a page.
5208  *
5209  * Inputs are required to range within a page.
5210  */
5211 vm_page_bits_t
5212 vm_page_bits(int base, int size)
5213 {
5214 	int first_bit;
5215 	int last_bit;
5216 
5217 	KASSERT(
5218 	    base + size <= PAGE_SIZE,
5219 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5220 	);
5221 
5222 	if (size == 0)		/* handle degenerate case */
5223 		return (0);
5224 
5225 	first_bit = base >> DEV_BSHIFT;
5226 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5227 
5228 	return (((vm_page_bits_t)2 << last_bit) -
5229 	    ((vm_page_bits_t)1 << first_bit));
5230 }
5231 
5232 void
5233 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5234 {
5235 
5236 #if PAGE_SIZE == 32768
5237 	atomic_set_64((uint64_t *)bits, set);
5238 #elif PAGE_SIZE == 16384
5239 	atomic_set_32((uint32_t *)bits, set);
5240 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5241 	atomic_set_16((uint16_t *)bits, set);
5242 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5243 	atomic_set_8((uint8_t *)bits, set);
5244 #else		/* PAGE_SIZE <= 8192 */
5245 	uintptr_t addr;
5246 	int shift;
5247 
5248 	addr = (uintptr_t)bits;
5249 	/*
5250 	 * Use a trick to perform a 32-bit atomic on the
5251 	 * containing aligned word, to not depend on the existence
5252 	 * of atomic_{set, clear}_{8, 16}.
5253 	 */
5254 	shift = addr & (sizeof(uint32_t) - 1);
5255 #if BYTE_ORDER == BIG_ENDIAN
5256 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5257 #else
5258 	shift *= NBBY;
5259 #endif
5260 	addr &= ~(sizeof(uint32_t) - 1);
5261 	atomic_set_32((uint32_t *)addr, set << shift);
5262 #endif		/* PAGE_SIZE */
5263 }
5264 
5265 static inline void
5266 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5267 {
5268 
5269 #if PAGE_SIZE == 32768
5270 	atomic_clear_64((uint64_t *)bits, clear);
5271 #elif PAGE_SIZE == 16384
5272 	atomic_clear_32((uint32_t *)bits, clear);
5273 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5274 	atomic_clear_16((uint16_t *)bits, clear);
5275 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5276 	atomic_clear_8((uint8_t *)bits, clear);
5277 #else		/* PAGE_SIZE <= 8192 */
5278 	uintptr_t addr;
5279 	int shift;
5280 
5281 	addr = (uintptr_t)bits;
5282 	/*
5283 	 * Use a trick to perform a 32-bit atomic on the
5284 	 * containing aligned word, to not depend on the existence
5285 	 * of atomic_{set, clear}_{8, 16}.
5286 	 */
5287 	shift = addr & (sizeof(uint32_t) - 1);
5288 #if BYTE_ORDER == BIG_ENDIAN
5289 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5290 #else
5291 	shift *= NBBY;
5292 #endif
5293 	addr &= ~(sizeof(uint32_t) - 1);
5294 	atomic_clear_32((uint32_t *)addr, clear << shift);
5295 #endif		/* PAGE_SIZE */
5296 }
5297 
5298 static inline vm_page_bits_t
5299 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5300 {
5301 #if PAGE_SIZE == 32768
5302 	uint64_t old;
5303 
5304 	old = *bits;
5305 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5306 	return (old);
5307 #elif PAGE_SIZE == 16384
5308 	uint32_t old;
5309 
5310 	old = *bits;
5311 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5312 	return (old);
5313 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5314 	uint16_t old;
5315 
5316 	old = *bits;
5317 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5318 	return (old);
5319 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5320 	uint8_t old;
5321 
5322 	old = *bits;
5323 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5324 	return (old);
5325 #else		/* PAGE_SIZE <= 4096*/
5326 	uintptr_t addr;
5327 	uint32_t old, new, mask;
5328 	int shift;
5329 
5330 	addr = (uintptr_t)bits;
5331 	/*
5332 	 * Use a trick to perform a 32-bit atomic on the
5333 	 * containing aligned word, to not depend on the existence
5334 	 * of atomic_{set, swap, clear}_{8, 16}.
5335 	 */
5336 	shift = addr & (sizeof(uint32_t) - 1);
5337 #if BYTE_ORDER == BIG_ENDIAN
5338 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5339 #else
5340 	shift *= NBBY;
5341 #endif
5342 	addr &= ~(sizeof(uint32_t) - 1);
5343 	mask = VM_PAGE_BITS_ALL << shift;
5344 
5345 	old = *bits;
5346 	do {
5347 		new = old & ~mask;
5348 		new |= newbits << shift;
5349 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5350 	return (old >> shift);
5351 #endif		/* PAGE_SIZE */
5352 }
5353 
5354 /*
5355  *	vm_page_set_valid_range:
5356  *
5357  *	Sets portions of a page valid.  The arguments are expected
5358  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5359  *	of any partial chunks touched by the range.  The invalid portion of
5360  *	such chunks will be zeroed.
5361  *
5362  *	(base + size) must be less then or equal to PAGE_SIZE.
5363  */
5364 void
5365 vm_page_set_valid_range(vm_page_t m, int base, int size)
5366 {
5367 	int endoff, frag;
5368 	vm_page_bits_t pagebits;
5369 
5370 	vm_page_assert_busied(m);
5371 	if (size == 0)	/* handle degenerate case */
5372 		return;
5373 
5374 	/*
5375 	 * If the base is not DEV_BSIZE aligned and the valid
5376 	 * bit is clear, we have to zero out a portion of the
5377 	 * first block.
5378 	 */
5379 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5380 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5381 		pmap_zero_page_area(m, frag, base - frag);
5382 
5383 	/*
5384 	 * If the ending offset is not DEV_BSIZE aligned and the
5385 	 * valid bit is clear, we have to zero out a portion of
5386 	 * the last block.
5387 	 */
5388 	endoff = base + size;
5389 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5390 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5391 		pmap_zero_page_area(m, endoff,
5392 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5393 
5394 	/*
5395 	 * Assert that no previously invalid block that is now being validated
5396 	 * is already dirty.
5397 	 */
5398 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5399 	    ("vm_page_set_valid_range: page %p is dirty", m));
5400 
5401 	/*
5402 	 * Set valid bits inclusive of any overlap.
5403 	 */
5404 	pagebits = vm_page_bits(base, size);
5405 	if (vm_page_xbusied(m))
5406 		m->valid |= pagebits;
5407 	else
5408 		vm_page_bits_set(m, &m->valid, pagebits);
5409 }
5410 
5411 /*
5412  * Set the page dirty bits and free the invalid swap space if
5413  * present.  Returns the previous dirty bits.
5414  */
5415 vm_page_bits_t
5416 vm_page_set_dirty(vm_page_t m)
5417 {
5418 	vm_page_bits_t old;
5419 
5420 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5421 
5422 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5423 		old = m->dirty;
5424 		m->dirty = VM_PAGE_BITS_ALL;
5425 	} else
5426 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5427 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5428 		vm_pager_page_unswapped(m);
5429 
5430 	return (old);
5431 }
5432 
5433 /*
5434  * Clear the given bits from the specified page's dirty field.
5435  */
5436 static __inline void
5437 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5438 {
5439 
5440 	vm_page_assert_busied(m);
5441 
5442 	/*
5443 	 * If the page is xbusied and not write mapped we are the
5444 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5445 	 * layer can call vm_page_dirty() without holding a distinguished
5446 	 * lock.  The combination of page busy and atomic operations
5447 	 * suffice to guarantee consistency of the page dirty field.
5448 	 */
5449 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5450 		m->dirty &= ~pagebits;
5451 	else
5452 		vm_page_bits_clear(m, &m->dirty, pagebits);
5453 }
5454 
5455 /*
5456  *	vm_page_set_validclean:
5457  *
5458  *	Sets portions of a page valid and clean.  The arguments are expected
5459  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5460  *	of any partial chunks touched by the range.  The invalid portion of
5461  *	such chunks will be zero'd.
5462  *
5463  *	(base + size) must be less then or equal to PAGE_SIZE.
5464  */
5465 void
5466 vm_page_set_validclean(vm_page_t m, int base, int size)
5467 {
5468 	vm_page_bits_t oldvalid, pagebits;
5469 	int endoff, frag;
5470 
5471 	vm_page_assert_busied(m);
5472 	if (size == 0)	/* handle degenerate case */
5473 		return;
5474 
5475 	/*
5476 	 * If the base is not DEV_BSIZE aligned and the valid
5477 	 * bit is clear, we have to zero out a portion of the
5478 	 * first block.
5479 	 */
5480 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5481 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5482 		pmap_zero_page_area(m, frag, base - frag);
5483 
5484 	/*
5485 	 * If the ending offset is not DEV_BSIZE aligned and the
5486 	 * valid bit is clear, we have to zero out a portion of
5487 	 * the last block.
5488 	 */
5489 	endoff = base + size;
5490 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5491 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5492 		pmap_zero_page_area(m, endoff,
5493 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5494 
5495 	/*
5496 	 * Set valid, clear dirty bits.  If validating the entire
5497 	 * page we can safely clear the pmap modify bit.  We also
5498 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5499 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5500 	 * be set again.
5501 	 *
5502 	 * We set valid bits inclusive of any overlap, but we can only
5503 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5504 	 * the range.
5505 	 */
5506 	oldvalid = m->valid;
5507 	pagebits = vm_page_bits(base, size);
5508 	if (vm_page_xbusied(m))
5509 		m->valid |= pagebits;
5510 	else
5511 		vm_page_bits_set(m, &m->valid, pagebits);
5512 #if 0	/* NOT YET */
5513 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5514 		frag = DEV_BSIZE - frag;
5515 		base += frag;
5516 		size -= frag;
5517 		if (size < 0)
5518 			size = 0;
5519 	}
5520 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5521 #endif
5522 	if (base == 0 && size == PAGE_SIZE) {
5523 		/*
5524 		 * The page can only be modified within the pmap if it is
5525 		 * mapped, and it can only be mapped if it was previously
5526 		 * fully valid.
5527 		 */
5528 		if (oldvalid == VM_PAGE_BITS_ALL)
5529 			/*
5530 			 * Perform the pmap_clear_modify() first.  Otherwise,
5531 			 * a concurrent pmap operation, such as
5532 			 * pmap_protect(), could clear a modification in the
5533 			 * pmap and set the dirty field on the page before
5534 			 * pmap_clear_modify() had begun and after the dirty
5535 			 * field was cleared here.
5536 			 */
5537 			pmap_clear_modify(m);
5538 		m->dirty = 0;
5539 		vm_page_aflag_clear(m, PGA_NOSYNC);
5540 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5541 		m->dirty &= ~pagebits;
5542 	else
5543 		vm_page_clear_dirty_mask(m, pagebits);
5544 }
5545 
5546 void
5547 vm_page_clear_dirty(vm_page_t m, int base, int size)
5548 {
5549 
5550 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5551 }
5552 
5553 /*
5554  *	vm_page_set_invalid:
5555  *
5556  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5557  *	valid and dirty bits for the effected areas are cleared.
5558  */
5559 void
5560 vm_page_set_invalid(vm_page_t m, int base, int size)
5561 {
5562 	vm_page_bits_t bits;
5563 	vm_object_t object;
5564 
5565 	/*
5566 	 * The object lock is required so that pages can't be mapped
5567 	 * read-only while we're in the process of invalidating them.
5568 	 */
5569 	object = m->object;
5570 	VM_OBJECT_ASSERT_WLOCKED(object);
5571 	vm_page_assert_busied(m);
5572 
5573 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5574 	    size >= object->un_pager.vnp.vnp_size)
5575 		bits = VM_PAGE_BITS_ALL;
5576 	else
5577 		bits = vm_page_bits(base, size);
5578 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5579 		pmap_remove_all(m);
5580 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5581 	    !pmap_page_is_mapped(m),
5582 	    ("vm_page_set_invalid: page %p is mapped", m));
5583 	if (vm_page_xbusied(m)) {
5584 		m->valid &= ~bits;
5585 		m->dirty &= ~bits;
5586 	} else {
5587 		vm_page_bits_clear(m, &m->valid, bits);
5588 		vm_page_bits_clear(m, &m->dirty, bits);
5589 	}
5590 }
5591 
5592 /*
5593  *	vm_page_invalid:
5594  *
5595  *	Invalidates the entire page.  The page must be busy, unmapped, and
5596  *	the enclosing object must be locked.  The object locks protects
5597  *	against concurrent read-only pmap enter which is done without
5598  *	busy.
5599  */
5600 void
5601 vm_page_invalid(vm_page_t m)
5602 {
5603 
5604 	vm_page_assert_busied(m);
5605 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5606 	MPASS(!pmap_page_is_mapped(m));
5607 
5608 	if (vm_page_xbusied(m))
5609 		m->valid = 0;
5610 	else
5611 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5612 }
5613 
5614 /*
5615  * vm_page_zero_invalid()
5616  *
5617  *	The kernel assumes that the invalid portions of a page contain
5618  *	garbage, but such pages can be mapped into memory by user code.
5619  *	When this occurs, we must zero out the non-valid portions of the
5620  *	page so user code sees what it expects.
5621  *
5622  *	Pages are most often semi-valid when the end of a file is mapped
5623  *	into memory and the file's size is not page aligned.
5624  */
5625 void
5626 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5627 {
5628 	int b;
5629 	int i;
5630 
5631 	/*
5632 	 * Scan the valid bits looking for invalid sections that
5633 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5634 	 * valid bit may be set ) have already been zeroed by
5635 	 * vm_page_set_validclean().
5636 	 */
5637 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5638 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5639 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5640 			if (i > b) {
5641 				pmap_zero_page_area(m,
5642 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5643 			}
5644 			b = i + 1;
5645 		}
5646 	}
5647 
5648 	/*
5649 	 * setvalid is TRUE when we can safely set the zero'd areas
5650 	 * as being valid.  We can do this if there are no cache consistency
5651 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5652 	 */
5653 	if (setvalid)
5654 		vm_page_valid(m);
5655 }
5656 
5657 /*
5658  *	vm_page_is_valid:
5659  *
5660  *	Is (partial) page valid?  Note that the case where size == 0
5661  *	will return FALSE in the degenerate case where the page is
5662  *	entirely invalid, and TRUE otherwise.
5663  *
5664  *	Some callers envoke this routine without the busy lock held and
5665  *	handle races via higher level locks.  Typical callers should
5666  *	hold a busy lock to prevent invalidation.
5667  */
5668 int
5669 vm_page_is_valid(vm_page_t m, int base, int size)
5670 {
5671 	vm_page_bits_t bits;
5672 
5673 	bits = vm_page_bits(base, size);
5674 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5675 }
5676 
5677 /*
5678  * Returns true if all of the specified predicates are true for the entire
5679  * (super)page and false otherwise.
5680  */
5681 bool
5682 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5683 {
5684 	vm_object_t object;
5685 	int i, npages;
5686 
5687 	object = m->object;
5688 	if (skip_m != NULL && skip_m->object != object)
5689 		return (false);
5690 	VM_OBJECT_ASSERT_LOCKED(object);
5691 	KASSERT(psind <= m->psind,
5692 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5693 	npages = atop(pagesizes[psind]);
5694 
5695 	/*
5696 	 * The physically contiguous pages that make up a superpage, i.e., a
5697 	 * page with a page size index ("psind") greater than zero, will
5698 	 * occupy adjacent entries in vm_page_array[].
5699 	 */
5700 	for (i = 0; i < npages; i++) {
5701 		/* Always test object consistency, including "skip_m". */
5702 		if (m[i].object != object)
5703 			return (false);
5704 		if (&m[i] == skip_m)
5705 			continue;
5706 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5707 			return (false);
5708 		if ((flags & PS_ALL_DIRTY) != 0) {
5709 			/*
5710 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5711 			 * might stop this case from spuriously returning
5712 			 * "false".  However, that would require a write lock
5713 			 * on the object containing "m[i]".
5714 			 */
5715 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5716 				return (false);
5717 		}
5718 		if ((flags & PS_ALL_VALID) != 0 &&
5719 		    m[i].valid != VM_PAGE_BITS_ALL)
5720 			return (false);
5721 	}
5722 	return (true);
5723 }
5724 
5725 /*
5726  * Set the page's dirty bits if the page is modified.
5727  */
5728 void
5729 vm_page_test_dirty(vm_page_t m)
5730 {
5731 
5732 	vm_page_assert_busied(m);
5733 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5734 		vm_page_dirty(m);
5735 }
5736 
5737 void
5738 vm_page_valid(vm_page_t m)
5739 {
5740 
5741 	vm_page_assert_busied(m);
5742 	if (vm_page_xbusied(m))
5743 		m->valid = VM_PAGE_BITS_ALL;
5744 	else
5745 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5746 }
5747 
5748 void
5749 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5750 {
5751 
5752 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5753 }
5754 
5755 void
5756 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5757 {
5758 
5759 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5760 }
5761 
5762 int
5763 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5764 {
5765 
5766 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5767 }
5768 
5769 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5770 void
5771 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5772 {
5773 
5774 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5775 }
5776 
5777 void
5778 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5779 {
5780 
5781 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5782 }
5783 #endif
5784 
5785 #ifdef INVARIANTS
5786 void
5787 vm_page_object_busy_assert(vm_page_t m)
5788 {
5789 
5790 	/*
5791 	 * Certain of the page's fields may only be modified by the
5792 	 * holder of a page or object busy.
5793 	 */
5794 	if (m->object != NULL && !vm_page_busied(m))
5795 		VM_OBJECT_ASSERT_BUSY(m->object);
5796 }
5797 
5798 void
5799 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5800 {
5801 
5802 	if ((bits & PGA_WRITEABLE) == 0)
5803 		return;
5804 
5805 	/*
5806 	 * The PGA_WRITEABLE flag can only be set if the page is
5807 	 * managed, is exclusively busied or the object is locked.
5808 	 * Currently, this flag is only set by pmap_enter().
5809 	 */
5810 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5811 	    ("PGA_WRITEABLE on unmanaged page"));
5812 	if (!vm_page_xbusied(m))
5813 		VM_OBJECT_ASSERT_BUSY(m->object);
5814 }
5815 #endif
5816 
5817 #include "opt_ddb.h"
5818 #ifdef DDB
5819 #include <sys/kernel.h>
5820 
5821 #include <ddb/ddb.h>
5822 
5823 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5824 {
5825 
5826 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5827 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5828 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5829 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5830 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5831 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5832 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5833 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5834 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5835 }
5836 
5837 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5838 {
5839 	int dom;
5840 
5841 	db_printf("pq_free %d\n", vm_free_count());
5842 	for (dom = 0; dom < vm_ndomains; dom++) {
5843 		db_printf(
5844     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5845 		    dom,
5846 		    vm_dom[dom].vmd_page_count,
5847 		    vm_dom[dom].vmd_free_count,
5848 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5849 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5850 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5851 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5852 	}
5853 }
5854 
5855 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5856 {
5857 	vm_page_t m;
5858 	boolean_t phys, virt;
5859 
5860 	if (!have_addr) {
5861 		db_printf("show pginfo addr\n");
5862 		return;
5863 	}
5864 
5865 	phys = strchr(modif, 'p') != NULL;
5866 	virt = strchr(modif, 'v') != NULL;
5867 	if (virt)
5868 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5869 	else if (phys)
5870 		m = PHYS_TO_VM_PAGE(addr);
5871 	else
5872 		m = (vm_page_t)addr;
5873 	db_printf(
5874     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5875     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5876 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5877 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5878 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5879 }
5880 #endif /* DDB */
5881