xref: /freebsd/sys/vm/vm_page.c (revision 70f51f0e474ffe1fb74cb427423a2fba3637544d)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *	Resident memory management module.
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_vm.h"
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/lock.h>
82 #include <sys/malloc.h>
83 #include <sys/mman.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
86 #include <sys/proc.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
89 #include <sys/sbuf.h>
90 #include <sys/sched.h>
91 #include <sys/smp.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
95 
96 #include <vm/vm.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
111 #include <vm/vm_dumpset.h>
112 #include <vm/uma.h>
113 #include <vm/uma_int.h>
114 
115 #include <machine/md_var.h>
116 
117 struct vm_domain vm_dom[MAXMEMDOM];
118 
119 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
120 
121 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
122 
123 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
124 /* The following fields are protected by the domainset lock. */
125 domainset_t __exclusive_cache_line vm_min_domains;
126 domainset_t __exclusive_cache_line vm_severe_domains;
127 static int vm_min_waiters;
128 static int vm_severe_waiters;
129 static int vm_pageproc_waiters;
130 
131 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
132     "VM page statistics");
133 
134 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
135 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
136     CTLFLAG_RD, &pqstate_commit_retries,
137     "Number of failed per-page atomic queue state updates");
138 
139 static COUNTER_U64_DEFINE_EARLY(queue_ops);
140 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
141     CTLFLAG_RD, &queue_ops,
142     "Number of batched queue operations");
143 
144 static COUNTER_U64_DEFINE_EARLY(queue_nops);
145 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
146     CTLFLAG_RD, &queue_nops,
147     "Number of batched queue operations with no effects");
148 
149 /*
150  * bogus page -- for I/O to/from partially complete buffers,
151  * or for paging into sparsely invalid regions.
152  */
153 vm_page_t bogus_page;
154 
155 vm_page_t vm_page_array;
156 long vm_page_array_size;
157 long first_page;
158 
159 struct bitset *vm_page_dump;
160 long vm_page_dump_pages;
161 
162 static TAILQ_HEAD(, vm_page) blacklist_head;
163 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
164 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
165     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
166 
167 static uma_zone_t fakepg_zone;
168 
169 static void vm_page_alloc_check(vm_page_t m);
170 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
171     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
172 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
173 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
174 static bool vm_page_free_prep(vm_page_t m);
175 static void vm_page_free_toq(vm_page_t m);
176 static void vm_page_init(void *dummy);
177 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
178     vm_pindex_t pindex, vm_page_t mpred);
179 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
180     vm_page_t mpred);
181 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
182     const uint16_t nflag);
183 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
184     vm_page_t m_run, vm_paddr_t high);
185 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
187     int req);
188 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
189     int flags);
190 static void vm_page_zone_release(void *arg, void **store, int cnt);
191 
192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
193 
194 static void
195 vm_page_init(void *dummy)
196 {
197 
198 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
199 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
200 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED);
201 }
202 
203 /*
204  * The cache page zone is initialized later since we need to be able to allocate
205  * pages before UMA is fully initialized.
206  */
207 static void
208 vm_page_init_cache_zones(void *dummy __unused)
209 {
210 	struct vm_domain *vmd;
211 	struct vm_pgcache *pgcache;
212 	int cache, domain, maxcache, pool;
213 
214 	maxcache = 0;
215 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
216 	maxcache *= mp_ncpus;
217 	for (domain = 0; domain < vm_ndomains; domain++) {
218 		vmd = VM_DOMAIN(domain);
219 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
220 			pgcache = &vmd->vmd_pgcache[pool];
221 			pgcache->domain = domain;
222 			pgcache->pool = pool;
223 			pgcache->zone = uma_zcache_create("vm pgcache",
224 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
225 			    vm_page_zone_import, vm_page_zone_release, pgcache,
226 			    UMA_ZONE_VM);
227 
228 			/*
229 			 * Limit each pool's zone to 0.1% of the pages in the
230 			 * domain.
231 			 */
232 			cache = maxcache != 0 ? maxcache :
233 			    vmd->vmd_page_count / 1000;
234 			uma_zone_set_maxcache(pgcache->zone, cache);
235 		}
236 	}
237 }
238 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
239 
240 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
241 #if PAGE_SIZE == 32768
242 #ifdef CTASSERT
243 CTASSERT(sizeof(u_long) >= 8);
244 #endif
245 #endif
246 
247 /*
248  *	vm_set_page_size:
249  *
250  *	Sets the page size, perhaps based upon the memory
251  *	size.  Must be called before any use of page-size
252  *	dependent functions.
253  */
254 void
255 vm_set_page_size(void)
256 {
257 	if (vm_cnt.v_page_size == 0)
258 		vm_cnt.v_page_size = PAGE_SIZE;
259 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
260 		panic("vm_set_page_size: page size not a power of two");
261 }
262 
263 /*
264  *	vm_page_blacklist_next:
265  *
266  *	Find the next entry in the provided string of blacklist
267  *	addresses.  Entries are separated by space, comma, or newline.
268  *	If an invalid integer is encountered then the rest of the
269  *	string is skipped.  Updates the list pointer to the next
270  *	character, or NULL if the string is exhausted or invalid.
271  */
272 static vm_paddr_t
273 vm_page_blacklist_next(char **list, char *end)
274 {
275 	vm_paddr_t bad;
276 	char *cp, *pos;
277 
278 	if (list == NULL || *list == NULL)
279 		return (0);
280 	if (**list =='\0') {
281 		*list = NULL;
282 		return (0);
283 	}
284 
285 	/*
286 	 * If there's no end pointer then the buffer is coming from
287 	 * the kenv and we know it's null-terminated.
288 	 */
289 	if (end == NULL)
290 		end = *list + strlen(*list);
291 
292 	/* Ensure that strtoq() won't walk off the end */
293 	if (*end != '\0') {
294 		if (*end == '\n' || *end == ' ' || *end  == ',')
295 			*end = '\0';
296 		else {
297 			printf("Blacklist not terminated, skipping\n");
298 			*list = NULL;
299 			return (0);
300 		}
301 	}
302 
303 	for (pos = *list; *pos != '\0'; pos = cp) {
304 		bad = strtoq(pos, &cp, 0);
305 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
306 			if (bad == 0) {
307 				if (++cp < end)
308 					continue;
309 				else
310 					break;
311 			}
312 		} else
313 			break;
314 		if (*cp == '\0' || ++cp >= end)
315 			*list = NULL;
316 		else
317 			*list = cp;
318 		return (trunc_page(bad));
319 	}
320 	printf("Garbage in RAM blacklist, skipping\n");
321 	*list = NULL;
322 	return (0);
323 }
324 
325 bool
326 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
327 {
328 	struct vm_domain *vmd;
329 	vm_page_t m;
330 	int ret;
331 
332 	m = vm_phys_paddr_to_vm_page(pa);
333 	if (m == NULL)
334 		return (true); /* page does not exist, no failure */
335 
336 	vmd = vm_pagequeue_domain(m);
337 	vm_domain_free_lock(vmd);
338 	ret = vm_phys_unfree_page(m);
339 	vm_domain_free_unlock(vmd);
340 	if (ret != 0) {
341 		vm_domain_freecnt_inc(vmd, -1);
342 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
343 		if (verbose)
344 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
345 	}
346 	return (ret);
347 }
348 
349 /*
350  *	vm_page_blacklist_check:
351  *
352  *	Iterate through the provided string of blacklist addresses, pulling
353  *	each entry out of the physical allocator free list and putting it
354  *	onto a list for reporting via the vm.page_blacklist sysctl.
355  */
356 static void
357 vm_page_blacklist_check(char *list, char *end)
358 {
359 	vm_paddr_t pa;
360 	char *next;
361 
362 	next = list;
363 	while (next != NULL) {
364 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
365 			continue;
366 		vm_page_blacklist_add(pa, bootverbose);
367 	}
368 }
369 
370 /*
371  *	vm_page_blacklist_load:
372  *
373  *	Search for a special module named "ram_blacklist".  It'll be a
374  *	plain text file provided by the user via the loader directive
375  *	of the same name.
376  */
377 static void
378 vm_page_blacklist_load(char **list, char **end)
379 {
380 	void *mod;
381 	u_char *ptr;
382 	u_int len;
383 
384 	mod = NULL;
385 	ptr = NULL;
386 
387 	mod = preload_search_by_type("ram_blacklist");
388 	if (mod != NULL) {
389 		ptr = preload_fetch_addr(mod);
390 		len = preload_fetch_size(mod);
391         }
392 	*list = ptr;
393 	if (ptr != NULL)
394 		*end = ptr + len;
395 	else
396 		*end = NULL;
397 	return;
398 }
399 
400 static int
401 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
402 {
403 	vm_page_t m;
404 	struct sbuf sbuf;
405 	int error, first;
406 
407 	first = 1;
408 	error = sysctl_wire_old_buffer(req, 0);
409 	if (error != 0)
410 		return (error);
411 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
412 	TAILQ_FOREACH(m, &blacklist_head, listq) {
413 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
414 		    (uintmax_t)m->phys_addr);
415 		first = 0;
416 	}
417 	error = sbuf_finish(&sbuf);
418 	sbuf_delete(&sbuf);
419 	return (error);
420 }
421 
422 /*
423  * Initialize a dummy page for use in scans of the specified paging queue.
424  * In principle, this function only needs to set the flag PG_MARKER.
425  * Nonetheless, it write busies the page as a safety precaution.
426  */
427 void
428 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
429 {
430 
431 	bzero(marker, sizeof(*marker));
432 	marker->flags = PG_MARKER;
433 	marker->a.flags = aflags;
434 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
435 	marker->a.queue = queue;
436 }
437 
438 static void
439 vm_page_domain_init(int domain)
440 {
441 	struct vm_domain *vmd;
442 	struct vm_pagequeue *pq;
443 	int i;
444 
445 	vmd = VM_DOMAIN(domain);
446 	bzero(vmd, sizeof(*vmd));
447 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
448 	    "vm inactive pagequeue";
449 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
450 	    "vm active pagequeue";
451 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
452 	    "vm laundry pagequeue";
453 	*__DECONST(const char **,
454 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
455 	    "vm unswappable pagequeue";
456 	vmd->vmd_domain = domain;
457 	vmd->vmd_page_count = 0;
458 	vmd->vmd_free_count = 0;
459 	vmd->vmd_segs = 0;
460 	vmd->vmd_oom = FALSE;
461 	for (i = 0; i < PQ_COUNT; i++) {
462 		pq = &vmd->vmd_pagequeues[i];
463 		TAILQ_INIT(&pq->pq_pl);
464 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
465 		    MTX_DEF | MTX_DUPOK);
466 		pq->pq_pdpages = 0;
467 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
468 	}
469 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
470 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
471 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
472 
473 	/*
474 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
475 	 * insertions.
476 	 */
477 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
478 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
479 	    &vmd->vmd_inacthead, plinks.q);
480 
481 	/*
482 	 * The clock pages are used to implement active queue scanning without
483 	 * requeues.  Scans start at clock[0], which is advanced after the scan
484 	 * ends.  When the two clock hands meet, they are reset and scanning
485 	 * resumes from the head of the queue.
486 	 */
487 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
488 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
489 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
490 	    &vmd->vmd_clock[0], plinks.q);
491 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
492 	    &vmd->vmd_clock[1], plinks.q);
493 }
494 
495 /*
496  * Initialize a physical page in preparation for adding it to the free
497  * lists.
498  */
499 void
500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
501 {
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 = VM_FREEPOOL_DEFAULT;
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 
573 	vaddr = round_page(vaddr);
574 
575 	vm_phys_early_startup();
576 	biggestone = vm_phys_avail_largest();
577 	end = phys_avail[biggestone+1];
578 
579 	/*
580 	 * Initialize the page and queue locks.
581 	 */
582 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
583 	for (i = 0; i < PA_LOCK_COUNT; i++)
584 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
585 	for (i = 0; i < vm_ndomains; i++)
586 		vm_page_domain_init(i);
587 
588 	new_end = end;
589 #ifdef WITNESS
590 	witness_size = round_page(witness_startup_count());
591 	new_end -= witness_size;
592 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
593 	    VM_PROT_READ | VM_PROT_WRITE);
594 	bzero((void *)mapped, witness_size);
595 	witness_startup((void *)mapped);
596 #endif
597 
598 #if MINIDUMP_PAGE_TRACKING
599 	/*
600 	 * Allocate a bitmap to indicate that a random physical page
601 	 * needs to be included in a minidump.
602 	 *
603 	 * The amd64 port needs this to indicate which direct map pages
604 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
605 	 *
606 	 * However, i386 still needs this workspace internally within the
607 	 * minidump code.  In theory, they are not needed on i386, but are
608 	 * included should the sf_buf code decide to use them.
609 	 */
610 	last_pa = 0;
611 	vm_page_dump_pages = 0;
612 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
613 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
614 		    dump_avail[i] / PAGE_SIZE;
615 		if (dump_avail[i + 1] > last_pa)
616 			last_pa = dump_avail[i + 1];
617 	}
618 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
619 	new_end -= vm_page_dump_size;
620 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
621 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
622 	bzero((void *)vm_page_dump, vm_page_dump_size);
623 #else
624 	(void)last_pa;
625 #endif
626 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
627     defined(__riscv) || defined(__powerpc64__)
628 	/*
629 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
630 	 * in a crash dump.  When pmap_map() uses the direct map, they are
631 	 * not automatically included.
632 	 */
633 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
634 		dump_add_page(pa);
635 #endif
636 	phys_avail[biggestone + 1] = new_end;
637 #ifdef __amd64__
638 	/*
639 	 * Request that the physical pages underlying the message buffer be
640 	 * included in a crash dump.  Since the message buffer is accessed
641 	 * through the direct map, they are not automatically included.
642 	 */
643 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
644 	last_pa = pa + round_page(msgbufsize);
645 	while (pa < last_pa) {
646 		dump_add_page(pa);
647 		pa += PAGE_SIZE;
648 	}
649 #endif
650 	/*
651 	 * Compute the number of pages of memory that will be available for
652 	 * use, taking into account the overhead of a page structure per page.
653 	 * In other words, solve
654 	 *	"available physical memory" - round_page(page_range *
655 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
656 	 * for page_range.
657 	 */
658 	low_avail = phys_avail[0];
659 	high_avail = phys_avail[1];
660 	for (i = 0; i < vm_phys_nsegs; i++) {
661 		if (vm_phys_segs[i].start < low_avail)
662 			low_avail = vm_phys_segs[i].start;
663 		if (vm_phys_segs[i].end > high_avail)
664 			high_avail = vm_phys_segs[i].end;
665 	}
666 	/* Skip the first chunk.  It is already accounted for. */
667 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
668 		if (phys_avail[i] < low_avail)
669 			low_avail = phys_avail[i];
670 		if (phys_avail[i + 1] > high_avail)
671 			high_avail = phys_avail[i + 1];
672 	}
673 	first_page = low_avail / PAGE_SIZE;
674 #ifdef VM_PHYSSEG_SPARSE
675 	size = 0;
676 	for (i = 0; i < vm_phys_nsegs; i++)
677 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
678 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
679 		size += phys_avail[i + 1] - phys_avail[i];
680 #elif defined(VM_PHYSSEG_DENSE)
681 	size = high_avail - low_avail;
682 #else
683 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
684 #endif
685 
686 #ifdef PMAP_HAS_PAGE_ARRAY
687 	pmap_page_array_startup(size / PAGE_SIZE);
688 	biggestone = vm_phys_avail_largest();
689 	end = new_end = phys_avail[biggestone + 1];
690 #else
691 #ifdef VM_PHYSSEG_DENSE
692 	/*
693 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
694 	 * the overhead of a page structure per page only if vm_page_array is
695 	 * allocated from the last physical memory chunk.  Otherwise, we must
696 	 * allocate page structures representing the physical memory
697 	 * underlying vm_page_array, even though they will not be used.
698 	 */
699 	if (new_end != high_avail)
700 		page_range = size / PAGE_SIZE;
701 	else
702 #endif
703 	{
704 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
705 
706 		/*
707 		 * If the partial bytes remaining are large enough for
708 		 * a page (PAGE_SIZE) without a corresponding
709 		 * 'struct vm_page', then new_end will contain an
710 		 * extra page after subtracting the length of the VM
711 		 * page array.  Compensate by subtracting an extra
712 		 * page from new_end.
713 		 */
714 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
715 			if (new_end == high_avail)
716 				high_avail -= PAGE_SIZE;
717 			new_end -= PAGE_SIZE;
718 		}
719 	}
720 	end = new_end;
721 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
722 #endif
723 
724 #if VM_NRESERVLEVEL > 0
725 	/*
726 	 * Allocate physical memory for the reservation management system's
727 	 * data structures, and map it.
728 	 */
729 	new_end = vm_reserv_startup(&vaddr, new_end);
730 #endif
731 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
732     defined(__riscv) || defined(__powerpc64__)
733 	/*
734 	 * Include vm_page_array and vm_reserv_array in a crash dump.
735 	 */
736 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
737 		dump_add_page(pa);
738 #endif
739 	phys_avail[biggestone + 1] = new_end;
740 
741 	/*
742 	 * Add physical memory segments corresponding to the available
743 	 * physical pages.
744 	 */
745 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
746 		if (vm_phys_avail_size(i) != 0)
747 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
748 
749 	/*
750 	 * Initialize the physical memory allocator.
751 	 */
752 	vm_phys_init();
753 
754 	/*
755 	 * Initialize the page structures and add every available page to the
756 	 * physical memory allocator's free lists.
757 	 */
758 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
759 	for (ii = 0; ii < vm_page_array_size; ii++) {
760 		m = &vm_page_array[ii];
761 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
762 		m->flags = PG_FICTITIOUS;
763 	}
764 #endif
765 	vm_cnt.v_page_count = 0;
766 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
767 		seg = &vm_phys_segs[segind];
768 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
769 		    m++, pa += PAGE_SIZE)
770 			vm_page_init_page(m, pa, segind);
771 
772 		/*
773 		 * Add the segment's pages that are covered by one of
774 		 * phys_avail's ranges to the free lists.
775 		 */
776 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
777 			if (seg->end <= phys_avail[i] ||
778 			    seg->start >= phys_avail[i + 1])
779 				continue;
780 
781 			startp = MAX(seg->start, phys_avail[i]);
782 			endp = MIN(seg->end, phys_avail[i + 1]);
783 			pagecount = (u_long)atop(endp - startp);
784 			if (pagecount == 0)
785 				continue;
786 
787 			m = seg->first_page + atop(startp - seg->start);
788 			vmd = VM_DOMAIN(seg->domain);
789 			vm_domain_free_lock(vmd);
790 			vm_phys_enqueue_contig(m, pagecount);
791 			vm_domain_free_unlock(vmd);
792 			vm_domain_freecnt_inc(vmd, pagecount);
793 			vm_cnt.v_page_count += (u_int)pagecount;
794 			vmd->vmd_page_count += (u_int)pagecount;
795 			vmd->vmd_segs |= 1UL << segind;
796 		}
797 	}
798 
799 	/*
800 	 * Remove blacklisted pages from the physical memory allocator.
801 	 */
802 	TAILQ_INIT(&blacklist_head);
803 	vm_page_blacklist_load(&list, &listend);
804 	vm_page_blacklist_check(list, listend);
805 
806 	list = kern_getenv("vm.blacklist");
807 	vm_page_blacklist_check(list, NULL);
808 
809 	freeenv(list);
810 #if VM_NRESERVLEVEL > 0
811 	/*
812 	 * Initialize the reservation management system.
813 	 */
814 	vm_reserv_init();
815 #endif
816 
817 	return (vaddr);
818 }
819 
820 void
821 vm_page_reference(vm_page_t m)
822 {
823 
824 	vm_page_aflag_set(m, PGA_REFERENCED);
825 }
826 
827 /*
828  *	vm_page_trybusy
829  *
830  *	Helper routine for grab functions to trylock busy.
831  *
832  *	Returns true on success and false on failure.
833  */
834 static bool
835 vm_page_trybusy(vm_page_t m, int allocflags)
836 {
837 
838 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
839 		return (vm_page_trysbusy(m));
840 	else
841 		return (vm_page_tryxbusy(m));
842 }
843 
844 /*
845  *	vm_page_tryacquire
846  *
847  *	Helper routine for grab functions to trylock busy and wire.
848  *
849  *	Returns true on success and false on failure.
850  */
851 static inline bool
852 vm_page_tryacquire(vm_page_t m, int allocflags)
853 {
854 	bool locked;
855 
856 	locked = vm_page_trybusy(m, allocflags);
857 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
858 		vm_page_wire(m);
859 	return (locked);
860 }
861 
862 /*
863  *	vm_page_busy_acquire:
864  *
865  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
866  *	and drop the object lock if necessary.
867  */
868 bool
869 vm_page_busy_acquire(vm_page_t m, int allocflags)
870 {
871 	vm_object_t obj;
872 	bool locked;
873 
874 	/*
875 	 * The page-specific object must be cached because page
876 	 * identity can change during the sleep, causing the
877 	 * re-lock of a different object.
878 	 * It is assumed that a reference to the object is already
879 	 * held by the callers.
880 	 */
881 	obj = atomic_load_ptr(&m->object);
882 	for (;;) {
883 		if (vm_page_tryacquire(m, allocflags))
884 			return (true);
885 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
886 			return (false);
887 		if (obj != NULL)
888 			locked = VM_OBJECT_WOWNED(obj);
889 		else
890 			locked = false;
891 		MPASS(locked || vm_page_wired(m));
892 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
893 		    locked) && locked)
894 			VM_OBJECT_WLOCK(obj);
895 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
896 			return (false);
897 		KASSERT(m->object == obj || m->object == NULL,
898 		    ("vm_page_busy_acquire: page %p does not belong to %p",
899 		    m, obj));
900 	}
901 }
902 
903 /*
904  *	vm_page_busy_downgrade:
905  *
906  *	Downgrade an exclusive busy page into a single shared busy page.
907  */
908 void
909 vm_page_busy_downgrade(vm_page_t m)
910 {
911 	u_int x;
912 
913 	vm_page_assert_xbusied(m);
914 
915 	x = vm_page_busy_fetch(m);
916 	for (;;) {
917 		if (atomic_fcmpset_rel_int(&m->busy_lock,
918 		    &x, VPB_SHARERS_WORD(1)))
919 			break;
920 	}
921 	if ((x & VPB_BIT_WAITERS) != 0)
922 		wakeup(m);
923 }
924 
925 /*
926  *
927  *	vm_page_busy_tryupgrade:
928  *
929  *	Attempt to upgrade a single shared busy into an exclusive busy.
930  */
931 int
932 vm_page_busy_tryupgrade(vm_page_t m)
933 {
934 	u_int ce, x;
935 
936 	vm_page_assert_sbusied(m);
937 
938 	x = vm_page_busy_fetch(m);
939 	ce = VPB_CURTHREAD_EXCLUSIVE;
940 	for (;;) {
941 		if (VPB_SHARERS(x) > 1)
942 			return (0);
943 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
944 		    ("vm_page_busy_tryupgrade: invalid lock state"));
945 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
946 		    ce | (x & VPB_BIT_WAITERS)))
947 			continue;
948 		return (1);
949 	}
950 }
951 
952 /*
953  *	vm_page_sbusied:
954  *
955  *	Return a positive value if the page is shared busied, 0 otherwise.
956  */
957 int
958 vm_page_sbusied(vm_page_t m)
959 {
960 	u_int x;
961 
962 	x = vm_page_busy_fetch(m);
963 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
964 }
965 
966 /*
967  *	vm_page_sunbusy:
968  *
969  *	Shared unbusy a page.
970  */
971 void
972 vm_page_sunbusy(vm_page_t m)
973 {
974 	u_int x;
975 
976 	vm_page_assert_sbusied(m);
977 
978 	x = vm_page_busy_fetch(m);
979 	for (;;) {
980 		KASSERT(x != VPB_FREED,
981 		    ("vm_page_sunbusy: Unlocking freed page."));
982 		if (VPB_SHARERS(x) > 1) {
983 			if (atomic_fcmpset_int(&m->busy_lock, &x,
984 			    x - VPB_ONE_SHARER))
985 				break;
986 			continue;
987 		}
988 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
989 		    ("vm_page_sunbusy: invalid lock state"));
990 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
991 			continue;
992 		if ((x & VPB_BIT_WAITERS) == 0)
993 			break;
994 		wakeup(m);
995 		break;
996 	}
997 }
998 
999 /*
1000  *	vm_page_busy_sleep:
1001  *
1002  *	Sleep if the page is busy, using the page pointer as wchan.
1003  *	This is used to implement the hard-path of busying mechanism.
1004  *
1005  *	If nonshared is true, sleep only if the page is xbusy.
1006  *
1007  *	The object lock must be held on entry and will be released on exit.
1008  */
1009 void
1010 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1011 {
1012 	vm_object_t obj;
1013 
1014 	obj = m->object;
1015 	VM_OBJECT_ASSERT_LOCKED(obj);
1016 	vm_page_lock_assert(m, MA_NOTOWNED);
1017 
1018 	if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1019 	    nonshared ? VM_ALLOC_SBUSY : 0 , true))
1020 		VM_OBJECT_DROP(obj);
1021 }
1022 
1023 /*
1024  *	vm_page_busy_sleep_unlocked:
1025  *
1026  *	Sleep if the page is busy, using the page pointer as wchan.
1027  *	This is used to implement the hard-path of busying mechanism.
1028  *
1029  *	If nonshared is true, sleep only if the page is xbusy.
1030  *
1031  *	The object lock must not be held on entry.  The operation will
1032  *	return if the page changes identity.
1033  */
1034 void
1035 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1036     const char *wmesg, bool nonshared)
1037 {
1038 
1039 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1040 	vm_page_lock_assert(m, MA_NOTOWNED);
1041 
1042 	_vm_page_busy_sleep(obj, m, pindex, wmesg,
1043 	    nonshared ? VM_ALLOC_SBUSY : 0, false);
1044 }
1045 
1046 /*
1047  *	_vm_page_busy_sleep:
1048  *
1049  *	Internal busy sleep function.  Verifies the page identity and
1050  *	lockstate against parameters.  Returns true if it sleeps and
1051  *	false otherwise.
1052  *
1053  *	If locked is true the lock will be dropped for any true returns
1054  *	and held for any false returns.
1055  */
1056 static bool
1057 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1058     const char *wmesg, int allocflags, bool locked)
1059 {
1060 	bool xsleep;
1061 	u_int x;
1062 
1063 	/*
1064 	 * If the object is busy we must wait for that to drain to zero
1065 	 * before trying the page again.
1066 	 */
1067 	if (obj != NULL && vm_object_busied(obj)) {
1068 		if (locked)
1069 			VM_OBJECT_DROP(obj);
1070 		vm_object_busy_wait(obj, wmesg);
1071 		return (true);
1072 	}
1073 
1074 	if (!vm_page_busied(m))
1075 		return (false);
1076 
1077 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1078 	sleepq_lock(m);
1079 	x = vm_page_busy_fetch(m);
1080 	do {
1081 		/*
1082 		 * If the page changes objects or becomes unlocked we can
1083 		 * simply return.
1084 		 */
1085 		if (x == VPB_UNBUSIED ||
1086 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1087 		    m->object != obj || m->pindex != pindex) {
1088 			sleepq_release(m);
1089 			return (false);
1090 		}
1091 		if ((x & VPB_BIT_WAITERS) != 0)
1092 			break;
1093 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1094 	if (locked)
1095 		VM_OBJECT_DROP(obj);
1096 	DROP_GIANT();
1097 	sleepq_add(m, NULL, wmesg, 0, 0);
1098 	sleepq_wait(m, PVM);
1099 	PICKUP_GIANT();
1100 	return (true);
1101 }
1102 
1103 /*
1104  *	vm_page_trysbusy:
1105  *
1106  *	Try to shared busy a page.
1107  *	If the operation succeeds 1 is returned otherwise 0.
1108  *	The operation never sleeps.
1109  */
1110 int
1111 vm_page_trysbusy(vm_page_t m)
1112 {
1113 	vm_object_t obj;
1114 	u_int x;
1115 
1116 	obj = m->object;
1117 	x = vm_page_busy_fetch(m);
1118 	for (;;) {
1119 		if ((x & VPB_BIT_SHARED) == 0)
1120 			return (0);
1121 		/*
1122 		 * Reduce the window for transient busies that will trigger
1123 		 * false negatives in vm_page_ps_test().
1124 		 */
1125 		if (obj != NULL && vm_object_busied(obj))
1126 			return (0);
1127 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1128 		    x + VPB_ONE_SHARER))
1129 			break;
1130 	}
1131 
1132 	/* Refetch the object now that we're guaranteed that it is stable. */
1133 	obj = m->object;
1134 	if (obj != NULL && vm_object_busied(obj)) {
1135 		vm_page_sunbusy(m);
1136 		return (0);
1137 	}
1138 	return (1);
1139 }
1140 
1141 /*
1142  *	vm_page_tryxbusy:
1143  *
1144  *	Try to exclusive busy a page.
1145  *	If the operation succeeds 1 is returned otherwise 0.
1146  *	The operation never sleeps.
1147  */
1148 int
1149 vm_page_tryxbusy(vm_page_t m)
1150 {
1151 	vm_object_t obj;
1152 
1153         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1154             VPB_CURTHREAD_EXCLUSIVE) == 0)
1155 		return (0);
1156 
1157 	obj = m->object;
1158 	if (obj != NULL && vm_object_busied(obj)) {
1159 		vm_page_xunbusy(m);
1160 		return (0);
1161 	}
1162 	return (1);
1163 }
1164 
1165 static void
1166 vm_page_xunbusy_hard_tail(vm_page_t m)
1167 {
1168 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1169 	/* Wake the waiter. */
1170 	wakeup(m);
1171 }
1172 
1173 /*
1174  *	vm_page_xunbusy_hard:
1175  *
1176  *	Called when unbusy has failed because there is a waiter.
1177  */
1178 void
1179 vm_page_xunbusy_hard(vm_page_t m)
1180 {
1181 	vm_page_assert_xbusied(m);
1182 	vm_page_xunbusy_hard_tail(m);
1183 }
1184 
1185 void
1186 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1187 {
1188 	vm_page_assert_xbusied_unchecked(m);
1189 	vm_page_xunbusy_hard_tail(m);
1190 }
1191 
1192 static void
1193 vm_page_busy_free(vm_page_t m)
1194 {
1195 	u_int x;
1196 
1197 	atomic_thread_fence_rel();
1198 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1199 	if ((x & VPB_BIT_WAITERS) != 0)
1200 		wakeup(m);
1201 }
1202 
1203 /*
1204  *	vm_page_unhold_pages:
1205  *
1206  *	Unhold each of the pages that is referenced by the given array.
1207  */
1208 void
1209 vm_page_unhold_pages(vm_page_t *ma, int count)
1210 {
1211 
1212 	for (; count != 0; count--) {
1213 		vm_page_unwire(*ma, PQ_ACTIVE);
1214 		ma++;
1215 	}
1216 }
1217 
1218 vm_page_t
1219 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1220 {
1221 	vm_page_t m;
1222 
1223 #ifdef VM_PHYSSEG_SPARSE
1224 	m = vm_phys_paddr_to_vm_page(pa);
1225 	if (m == NULL)
1226 		m = vm_phys_fictitious_to_vm_page(pa);
1227 	return (m);
1228 #elif defined(VM_PHYSSEG_DENSE)
1229 	long pi;
1230 
1231 	pi = atop(pa);
1232 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1233 		m = &vm_page_array[pi - first_page];
1234 		return (m);
1235 	}
1236 	return (vm_phys_fictitious_to_vm_page(pa));
1237 #else
1238 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1239 #endif
1240 }
1241 
1242 /*
1243  *	vm_page_getfake:
1244  *
1245  *	Create a fictitious page with the specified physical address and
1246  *	memory attribute.  The memory attribute is the only the machine-
1247  *	dependent aspect of a fictitious page that must be initialized.
1248  */
1249 vm_page_t
1250 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1251 {
1252 	vm_page_t m;
1253 
1254 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1255 	vm_page_initfake(m, paddr, memattr);
1256 	return (m);
1257 }
1258 
1259 void
1260 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1261 {
1262 
1263 	if ((m->flags & PG_FICTITIOUS) != 0) {
1264 		/*
1265 		 * The page's memattr might have changed since the
1266 		 * previous initialization.  Update the pmap to the
1267 		 * new memattr.
1268 		 */
1269 		goto memattr;
1270 	}
1271 	m->phys_addr = paddr;
1272 	m->a.queue = PQ_NONE;
1273 	/* Fictitious pages don't use "segind". */
1274 	m->flags = PG_FICTITIOUS;
1275 	/* Fictitious pages don't use "order" or "pool". */
1276 	m->oflags = VPO_UNMANAGED;
1277 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1278 	/* Fictitious pages are unevictable. */
1279 	m->ref_count = 1;
1280 	pmap_page_init(m);
1281 memattr:
1282 	pmap_page_set_memattr(m, memattr);
1283 }
1284 
1285 /*
1286  *	vm_page_putfake:
1287  *
1288  *	Release a fictitious page.
1289  */
1290 void
1291 vm_page_putfake(vm_page_t m)
1292 {
1293 
1294 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1295 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1296 	    ("vm_page_putfake: bad page %p", m));
1297 	vm_page_assert_xbusied(m);
1298 	vm_page_busy_free(m);
1299 	uma_zfree(fakepg_zone, m);
1300 }
1301 
1302 /*
1303  *	vm_page_updatefake:
1304  *
1305  *	Update the given fictitious page to the specified physical address and
1306  *	memory attribute.
1307  */
1308 void
1309 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1310 {
1311 
1312 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1313 	    ("vm_page_updatefake: bad page %p", m));
1314 	m->phys_addr = paddr;
1315 	pmap_page_set_memattr(m, memattr);
1316 }
1317 
1318 /*
1319  *	vm_page_free:
1320  *
1321  *	Free a page.
1322  */
1323 void
1324 vm_page_free(vm_page_t m)
1325 {
1326 
1327 	m->flags &= ~PG_ZERO;
1328 	vm_page_free_toq(m);
1329 }
1330 
1331 /*
1332  *	vm_page_free_zero:
1333  *
1334  *	Free a page to the zerod-pages queue
1335  */
1336 void
1337 vm_page_free_zero(vm_page_t m)
1338 {
1339 
1340 	m->flags |= PG_ZERO;
1341 	vm_page_free_toq(m);
1342 }
1343 
1344 /*
1345  * Unbusy and handle the page queueing for a page from a getpages request that
1346  * was optionally read ahead or behind.
1347  */
1348 void
1349 vm_page_readahead_finish(vm_page_t m)
1350 {
1351 
1352 	/* We shouldn't put invalid pages on queues. */
1353 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1354 
1355 	/*
1356 	 * Since the page is not the actually needed one, whether it should
1357 	 * be activated or deactivated is not obvious.  Empirical results
1358 	 * have shown that deactivating the page is usually the best choice,
1359 	 * unless the page is wanted by another thread.
1360 	 */
1361 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1362 		vm_page_activate(m);
1363 	else
1364 		vm_page_deactivate(m);
1365 	vm_page_xunbusy_unchecked(m);
1366 }
1367 
1368 /*
1369  * Destroy the identity of an invalid page and free it if possible.
1370  * This is intended to be used when reading a page from backing store fails.
1371  */
1372 void
1373 vm_page_free_invalid(vm_page_t m)
1374 {
1375 
1376 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1377 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1378 	KASSERT(m->object != NULL, ("page %p has no object", m));
1379 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1380 
1381 	/*
1382 	 * We may be attempting to free the page as part of the handling for an
1383 	 * I/O error, in which case the page was xbusied by a different thread.
1384 	 */
1385 	vm_page_xbusy_claim(m);
1386 
1387 	/*
1388 	 * If someone has wired this page while the object lock
1389 	 * was not held, then the thread that unwires is responsible
1390 	 * for freeing the page.  Otherwise just free the page now.
1391 	 * The wire count of this unmapped page cannot change while
1392 	 * we have the page xbusy and the page's object wlocked.
1393 	 */
1394 	if (vm_page_remove(m))
1395 		vm_page_free(m);
1396 }
1397 
1398 /*
1399  *	vm_page_sleep_if_busy:
1400  *
1401  *	Sleep and release the object lock if the page is busied.
1402  *	Returns TRUE if the thread slept.
1403  *
1404  *	The given page must be unlocked and object containing it must
1405  *	be locked.
1406  */
1407 int
1408 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1409 {
1410 	vm_object_t obj;
1411 
1412 	vm_page_lock_assert(m, MA_NOTOWNED);
1413 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1414 
1415 	/*
1416 	 * The page-specific object must be cached because page
1417 	 * identity can change during the sleep, causing the
1418 	 * re-lock of a different object.
1419 	 * It is assumed that a reference to the object is already
1420 	 * held by the callers.
1421 	 */
1422 	obj = m->object;
1423 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1424 		VM_OBJECT_WLOCK(obj);
1425 		return (TRUE);
1426 	}
1427 	return (FALSE);
1428 }
1429 
1430 /*
1431  *	vm_page_sleep_if_xbusy:
1432  *
1433  *	Sleep and release the object lock if the page is xbusied.
1434  *	Returns TRUE if the thread slept.
1435  *
1436  *	The given page must be unlocked and object containing it must
1437  *	be locked.
1438  */
1439 int
1440 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1441 {
1442 	vm_object_t obj;
1443 
1444 	vm_page_lock_assert(m, MA_NOTOWNED);
1445 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1446 
1447 	/*
1448 	 * The page-specific object must be cached because page
1449 	 * identity can change during the sleep, causing the
1450 	 * re-lock of a different object.
1451 	 * It is assumed that a reference to the object is already
1452 	 * held by the callers.
1453 	 */
1454 	obj = m->object;
1455 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1456 	    true)) {
1457 		VM_OBJECT_WLOCK(obj);
1458 		return (TRUE);
1459 	}
1460 	return (FALSE);
1461 }
1462 
1463 /*
1464  *	vm_page_dirty_KBI:		[ internal use only ]
1465  *
1466  *	Set all bits in the page's dirty field.
1467  *
1468  *	The object containing the specified page must be locked if the
1469  *	call is made from the machine-independent layer.
1470  *
1471  *	See vm_page_clear_dirty_mask().
1472  *
1473  *	This function should only be called by vm_page_dirty().
1474  */
1475 void
1476 vm_page_dirty_KBI(vm_page_t m)
1477 {
1478 
1479 	/* Refer to this operation by its public name. */
1480 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1481 	m->dirty = VM_PAGE_BITS_ALL;
1482 }
1483 
1484 /*
1485  *	vm_page_insert:		[ internal use only ]
1486  *
1487  *	Inserts the given mem entry into the object and object list.
1488  *
1489  *	The object must be locked.
1490  */
1491 int
1492 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1493 {
1494 	vm_page_t mpred;
1495 
1496 	VM_OBJECT_ASSERT_WLOCKED(object);
1497 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1498 	return (vm_page_insert_after(m, object, pindex, mpred));
1499 }
1500 
1501 /*
1502  *	vm_page_insert_after:
1503  *
1504  *	Inserts the page "m" into the specified object at offset "pindex".
1505  *
1506  *	The page "mpred" must immediately precede the offset "pindex" within
1507  *	the specified object.
1508  *
1509  *	The object must be locked.
1510  */
1511 static int
1512 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1513     vm_page_t mpred)
1514 {
1515 	vm_page_t msucc;
1516 
1517 	VM_OBJECT_ASSERT_WLOCKED(object);
1518 	KASSERT(m->object == NULL,
1519 	    ("vm_page_insert_after: page already inserted"));
1520 	if (mpred != NULL) {
1521 		KASSERT(mpred->object == object,
1522 		    ("vm_page_insert_after: object doesn't contain mpred"));
1523 		KASSERT(mpred->pindex < pindex,
1524 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1525 		msucc = TAILQ_NEXT(mpred, listq);
1526 	} else
1527 		msucc = TAILQ_FIRST(&object->memq);
1528 	if (msucc != NULL)
1529 		KASSERT(msucc->pindex > pindex,
1530 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1531 
1532 	/*
1533 	 * Record the object/offset pair in this page.
1534 	 */
1535 	m->object = object;
1536 	m->pindex = pindex;
1537 	m->ref_count |= VPRC_OBJREF;
1538 
1539 	/*
1540 	 * Now link into the object's ordered list of backed pages.
1541 	 */
1542 	if (vm_radix_insert(&object->rtree, m)) {
1543 		m->object = NULL;
1544 		m->pindex = 0;
1545 		m->ref_count &= ~VPRC_OBJREF;
1546 		return (1);
1547 	}
1548 	vm_page_insert_radixdone(m, object, mpred);
1549 	return (0);
1550 }
1551 
1552 /*
1553  *	vm_page_insert_radixdone:
1554  *
1555  *	Complete page "m" insertion into the specified object after the
1556  *	radix trie hooking.
1557  *
1558  *	The page "mpred" must precede the offset "m->pindex" within the
1559  *	specified object.
1560  *
1561  *	The object must be locked.
1562  */
1563 static void
1564 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1565 {
1566 
1567 	VM_OBJECT_ASSERT_WLOCKED(object);
1568 	KASSERT(object != NULL && m->object == object,
1569 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1570 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1571 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1572 	if (mpred != NULL) {
1573 		KASSERT(mpred->object == object,
1574 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1575 		KASSERT(mpred->pindex < m->pindex,
1576 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1577 	}
1578 
1579 	if (mpred != NULL)
1580 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1581 	else
1582 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1583 
1584 	/*
1585 	 * Show that the object has one more resident page.
1586 	 */
1587 	object->resident_page_count++;
1588 
1589 	/*
1590 	 * Hold the vnode until the last page is released.
1591 	 */
1592 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1593 		vhold(object->handle);
1594 
1595 	/*
1596 	 * Since we are inserting a new and possibly dirty page,
1597 	 * update the object's generation count.
1598 	 */
1599 	if (pmap_page_is_write_mapped(m))
1600 		vm_object_set_writeable_dirty(object);
1601 }
1602 
1603 /*
1604  * Do the work to remove a page from its object.  The caller is responsible for
1605  * updating the page's fields to reflect this removal.
1606  */
1607 static void
1608 vm_page_object_remove(vm_page_t m)
1609 {
1610 	vm_object_t object;
1611 	vm_page_t mrem;
1612 
1613 	vm_page_assert_xbusied(m);
1614 	object = m->object;
1615 	VM_OBJECT_ASSERT_WLOCKED(object);
1616 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1617 	    ("page %p is missing its object ref", m));
1618 
1619 	/* Deferred free of swap space. */
1620 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1621 		vm_pager_page_unswapped(m);
1622 
1623 	m->object = NULL;
1624 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1625 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1626 
1627 	/*
1628 	 * Now remove from the object's list of backed pages.
1629 	 */
1630 	TAILQ_REMOVE(&object->memq, m, listq);
1631 
1632 	/*
1633 	 * And show that the object has one fewer resident page.
1634 	 */
1635 	object->resident_page_count--;
1636 
1637 	/*
1638 	 * The vnode may now be recycled.
1639 	 */
1640 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1641 		vdrop(object->handle);
1642 }
1643 
1644 /*
1645  *	vm_page_remove:
1646  *
1647  *	Removes the specified page from its containing object, but does not
1648  *	invalidate any backing storage.  Returns true if the object's reference
1649  *	was the last reference to the page, and false otherwise.
1650  *
1651  *	The object must be locked and the page must be exclusively busied.
1652  *	The exclusive busy will be released on return.  If this is not the
1653  *	final ref and the caller does not hold a wire reference it may not
1654  *	continue to access the page.
1655  */
1656 bool
1657 vm_page_remove(vm_page_t m)
1658 {
1659 	bool dropped;
1660 
1661 	dropped = vm_page_remove_xbusy(m);
1662 	vm_page_xunbusy(m);
1663 
1664 	return (dropped);
1665 }
1666 
1667 /*
1668  *	vm_page_remove_xbusy
1669  *
1670  *	Removes the page but leaves the xbusy held.  Returns true if this
1671  *	removed the final ref and false otherwise.
1672  */
1673 bool
1674 vm_page_remove_xbusy(vm_page_t m)
1675 {
1676 
1677 	vm_page_object_remove(m);
1678 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1679 }
1680 
1681 /*
1682  *	vm_page_lookup:
1683  *
1684  *	Returns the page associated with the object/offset
1685  *	pair specified; if none is found, NULL is returned.
1686  *
1687  *	The object must be locked.
1688  */
1689 vm_page_t
1690 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1691 {
1692 
1693 	VM_OBJECT_ASSERT_LOCKED(object);
1694 	return (vm_radix_lookup(&object->rtree, pindex));
1695 }
1696 
1697 /*
1698  *	vm_page_lookup_unlocked:
1699  *
1700  *	Returns the page associated with the object/offset pair specified;
1701  *	if none is found, NULL is returned.  The page may be no longer be
1702  *	present in the object at the time that this function returns.  Only
1703  *	useful for opportunistic checks such as inmem().
1704  */
1705 vm_page_t
1706 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1707 {
1708 
1709 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1710 }
1711 
1712 /*
1713  *	vm_page_relookup:
1714  *
1715  *	Returns a page that must already have been busied by
1716  *	the caller.  Used for bogus page replacement.
1717  */
1718 vm_page_t
1719 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1720 {
1721 	vm_page_t m;
1722 
1723 	m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1724 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1725 	    m->object == object && m->pindex == pindex,
1726 	    ("vm_page_relookup: Invalid page %p", m));
1727 	return (m);
1728 }
1729 
1730 /*
1731  * This should only be used by lockless functions for releasing transient
1732  * incorrect acquires.  The page may have been freed after we acquired a
1733  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1734  * further to do.
1735  */
1736 static void
1737 vm_page_busy_release(vm_page_t m)
1738 {
1739 	u_int x;
1740 
1741 	x = vm_page_busy_fetch(m);
1742 	for (;;) {
1743 		if (x == VPB_FREED)
1744 			break;
1745 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1746 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1747 			    x - VPB_ONE_SHARER))
1748 				break;
1749 			continue;
1750 		}
1751 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1752 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1753 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1754 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1755 			continue;
1756 		if ((x & VPB_BIT_WAITERS) != 0)
1757 			wakeup(m);
1758 		break;
1759 	}
1760 }
1761 
1762 /*
1763  *	vm_page_find_least:
1764  *
1765  *	Returns the page associated with the object with least pindex
1766  *	greater than or equal to the parameter pindex, or NULL.
1767  *
1768  *	The object must be locked.
1769  */
1770 vm_page_t
1771 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1772 {
1773 	vm_page_t m;
1774 
1775 	VM_OBJECT_ASSERT_LOCKED(object);
1776 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1777 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1778 	return (m);
1779 }
1780 
1781 /*
1782  * Returns the given page's successor (by pindex) within the object if it is
1783  * resident; if none is found, NULL is returned.
1784  *
1785  * The object must be locked.
1786  */
1787 vm_page_t
1788 vm_page_next(vm_page_t m)
1789 {
1790 	vm_page_t next;
1791 
1792 	VM_OBJECT_ASSERT_LOCKED(m->object);
1793 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1794 		MPASS(next->object == m->object);
1795 		if (next->pindex != m->pindex + 1)
1796 			next = NULL;
1797 	}
1798 	return (next);
1799 }
1800 
1801 /*
1802  * Returns the given page's predecessor (by pindex) within the object if it is
1803  * resident; if none is found, NULL is returned.
1804  *
1805  * The object must be locked.
1806  */
1807 vm_page_t
1808 vm_page_prev(vm_page_t m)
1809 {
1810 	vm_page_t prev;
1811 
1812 	VM_OBJECT_ASSERT_LOCKED(m->object);
1813 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1814 		MPASS(prev->object == m->object);
1815 		if (prev->pindex != m->pindex - 1)
1816 			prev = NULL;
1817 	}
1818 	return (prev);
1819 }
1820 
1821 /*
1822  * Uses the page mnew as a replacement for an existing page at index
1823  * pindex which must be already present in the object.
1824  *
1825  * Both pages must be exclusively busied on enter.  The old page is
1826  * unbusied on exit.
1827  *
1828  * A return value of true means mold is now free.  If this is not the
1829  * final ref and the caller does not hold a wire reference it may not
1830  * continue to access the page.
1831  */
1832 static bool
1833 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1834     vm_page_t mold)
1835 {
1836 	vm_page_t mret;
1837 	bool dropped;
1838 
1839 	VM_OBJECT_ASSERT_WLOCKED(object);
1840 	vm_page_assert_xbusied(mold);
1841 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1842 	    ("vm_page_replace: page %p already in object", mnew));
1843 
1844 	/*
1845 	 * This function mostly follows vm_page_insert() and
1846 	 * vm_page_remove() without the radix, object count and vnode
1847 	 * dance.  Double check such functions for more comments.
1848 	 */
1849 
1850 	mnew->object = object;
1851 	mnew->pindex = pindex;
1852 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1853 	mret = vm_radix_replace(&object->rtree, mnew);
1854 	KASSERT(mret == mold,
1855 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1856 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1857 	    (mnew->oflags & VPO_UNMANAGED),
1858 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1859 
1860 	/* Keep the resident page list in sorted order. */
1861 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1862 	TAILQ_REMOVE(&object->memq, mold, listq);
1863 	mold->object = NULL;
1864 
1865 	/*
1866 	 * The object's resident_page_count does not change because we have
1867 	 * swapped one page for another, but the generation count should
1868 	 * change if the page is dirty.
1869 	 */
1870 	if (pmap_page_is_write_mapped(mnew))
1871 		vm_object_set_writeable_dirty(object);
1872 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1873 	vm_page_xunbusy(mold);
1874 
1875 	return (dropped);
1876 }
1877 
1878 void
1879 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1880     vm_page_t mold)
1881 {
1882 
1883 	vm_page_assert_xbusied(mnew);
1884 
1885 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1886 		vm_page_free(mold);
1887 }
1888 
1889 /*
1890  *	vm_page_rename:
1891  *
1892  *	Move the given memory entry from its
1893  *	current object to the specified target object/offset.
1894  *
1895  *	Note: swap associated with the page must be invalidated by the move.  We
1896  *	      have to do this for several reasons:  (1) we aren't freeing the
1897  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1898  *	      moving the page from object A to B, and will then later move
1899  *	      the backing store from A to B and we can't have a conflict.
1900  *
1901  *	Note: we *always* dirty the page.  It is necessary both for the
1902  *	      fact that we moved it, and because we may be invalidating
1903  *	      swap.
1904  *
1905  *	The objects must be locked.
1906  */
1907 int
1908 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1909 {
1910 	vm_page_t mpred;
1911 	vm_pindex_t opidx;
1912 
1913 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1914 
1915 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1916 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1917 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1918 	    ("vm_page_rename: pindex already renamed"));
1919 
1920 	/*
1921 	 * Create a custom version of vm_page_insert() which does not depend
1922 	 * by m_prev and can cheat on the implementation aspects of the
1923 	 * function.
1924 	 */
1925 	opidx = m->pindex;
1926 	m->pindex = new_pindex;
1927 	if (vm_radix_insert(&new_object->rtree, m)) {
1928 		m->pindex = opidx;
1929 		return (1);
1930 	}
1931 
1932 	/*
1933 	 * The operation cannot fail anymore.  The removal must happen before
1934 	 * the listq iterator is tainted.
1935 	 */
1936 	m->pindex = opidx;
1937 	vm_page_object_remove(m);
1938 
1939 	/* Return back to the new pindex to complete vm_page_insert(). */
1940 	m->pindex = new_pindex;
1941 	m->object = new_object;
1942 
1943 	vm_page_insert_radixdone(m, new_object, mpred);
1944 	vm_page_dirty(m);
1945 	return (0);
1946 }
1947 
1948 /*
1949  *	vm_page_alloc:
1950  *
1951  *	Allocate and return a page that is associated with the specified
1952  *	object and offset pair.  By default, this page is exclusive busied.
1953  *
1954  *	The caller must always specify an allocation class.
1955  *
1956  *	allocation classes:
1957  *	VM_ALLOC_NORMAL		normal process request
1958  *	VM_ALLOC_SYSTEM		system *really* needs a page
1959  *	VM_ALLOC_INTERRUPT	interrupt time request
1960  *
1961  *	optional allocation flags:
1962  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1963  *				intends to allocate
1964  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1965  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1966  *	VM_ALLOC_SBUSY		shared busy the allocated page
1967  *	VM_ALLOC_WIRED		wire the allocated page
1968  *	VM_ALLOC_ZERO		prefer a zeroed page
1969  */
1970 vm_page_t
1971 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1972 {
1973 
1974 	return (vm_page_alloc_after(object, pindex, req,
1975 	    vm_radix_lookup_le(&object->rtree, pindex)));
1976 }
1977 
1978 vm_page_t
1979 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1980     int req)
1981 {
1982 
1983 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1984 	    vm_radix_lookup_le(&object->rtree, pindex)));
1985 }
1986 
1987 /*
1988  * Allocate a page in the specified object with the given page index.  To
1989  * optimize insertion of the page into the object, the caller must also specifiy
1990  * the resident page in the object with largest index smaller than the given
1991  * page index, or NULL if no such page exists.
1992  */
1993 vm_page_t
1994 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1995     int req, vm_page_t mpred)
1996 {
1997 	struct vm_domainset_iter di;
1998 	vm_page_t m;
1999 	int domain;
2000 
2001 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2002 	do {
2003 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2004 		    mpred);
2005 		if (m != NULL)
2006 			break;
2007 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2008 
2009 	return (m);
2010 }
2011 
2012 /*
2013  * Returns true if the number of free pages exceeds the minimum
2014  * for the request class and false otherwise.
2015  */
2016 static int
2017 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2018 {
2019 	u_int limit, old, new;
2020 
2021 	if (req_class == VM_ALLOC_INTERRUPT)
2022 		limit = 0;
2023 	else if (req_class == VM_ALLOC_SYSTEM)
2024 		limit = vmd->vmd_interrupt_free_min;
2025 	else
2026 		limit = vmd->vmd_free_reserved;
2027 
2028 	/*
2029 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2030 	 */
2031 	limit += npages;
2032 	old = vmd->vmd_free_count;
2033 	do {
2034 		if (old < limit)
2035 			return (0);
2036 		new = old - npages;
2037 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2038 
2039 	/* Wake the page daemon if we've crossed the threshold. */
2040 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2041 		pagedaemon_wakeup(vmd->vmd_domain);
2042 
2043 	/* Only update bitsets on transitions. */
2044 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2045 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2046 		vm_domain_set(vmd);
2047 
2048 	return (1);
2049 }
2050 
2051 int
2052 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2053 {
2054 	int req_class;
2055 
2056 	/*
2057 	 * The page daemon is allowed to dig deeper into the free page list.
2058 	 */
2059 	req_class = req & VM_ALLOC_CLASS_MASK;
2060 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2061 		req_class = VM_ALLOC_SYSTEM;
2062 	return (_vm_domain_allocate(vmd, req_class, npages));
2063 }
2064 
2065 vm_page_t
2066 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2067     int req, vm_page_t mpred)
2068 {
2069 	struct vm_domain *vmd;
2070 	vm_page_t m;
2071 	int flags;
2072 
2073 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2074 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2075 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2076 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2077 	KASSERT((req & ~VPA_FLAGS) == 0,
2078 	    ("invalid request %#x", req));
2079 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2080 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2081 	    ("invalid request %#x", req));
2082 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2083 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2084 	    (uintmax_t)pindex));
2085 	VM_OBJECT_ASSERT_WLOCKED(object);
2086 
2087 	flags = 0;
2088 	m = NULL;
2089 again:
2090 #if VM_NRESERVLEVEL > 0
2091 	/*
2092 	 * Can we allocate the page from a reservation?
2093 	 */
2094 	if (vm_object_reserv(object) &&
2095 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2096 	    NULL) {
2097 		goto found;
2098 	}
2099 #endif
2100 	vmd = VM_DOMAIN(domain);
2101 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2102 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2103 		    M_NOWAIT | M_NOVM);
2104 		if (m != NULL) {
2105 			flags |= PG_PCPU_CACHE;
2106 			goto found;
2107 		}
2108 	}
2109 	if (vm_domain_allocate(vmd, req, 1)) {
2110 		/*
2111 		 * If not, allocate it from the free page queues.
2112 		 */
2113 		vm_domain_free_lock(vmd);
2114 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2115 		vm_domain_free_unlock(vmd);
2116 		if (m == NULL) {
2117 			vm_domain_freecnt_inc(vmd, 1);
2118 #if VM_NRESERVLEVEL > 0
2119 			if (vm_reserv_reclaim_inactive(domain))
2120 				goto again;
2121 #endif
2122 		}
2123 	}
2124 	if (m == NULL) {
2125 		/*
2126 		 * Not allocatable, give up.
2127 		 */
2128 		if (vm_domain_alloc_fail(vmd, object, req))
2129 			goto again;
2130 		return (NULL);
2131 	}
2132 
2133 	/*
2134 	 * At this point we had better have found a good page.
2135 	 */
2136 found:
2137 	vm_page_dequeue(m);
2138 	vm_page_alloc_check(m);
2139 
2140 	/*
2141 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2142 	 */
2143 	flags |= m->flags & PG_ZERO;
2144 	if ((req & VM_ALLOC_NODUMP) != 0)
2145 		flags |= PG_NODUMP;
2146 	m->flags = flags;
2147 	m->a.flags = 0;
2148 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2149 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2150 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2151 	else if ((req & VM_ALLOC_SBUSY) != 0)
2152 		m->busy_lock = VPB_SHARERS_WORD(1);
2153 	else
2154 		m->busy_lock = VPB_UNBUSIED;
2155 	if (req & VM_ALLOC_WIRED) {
2156 		vm_wire_add(1);
2157 		m->ref_count = 1;
2158 	}
2159 	m->a.act_count = 0;
2160 
2161 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2162 		if (req & VM_ALLOC_WIRED) {
2163 			vm_wire_sub(1);
2164 			m->ref_count = 0;
2165 		}
2166 		KASSERT(m->object == NULL, ("page %p has object", m));
2167 		m->oflags = VPO_UNMANAGED;
2168 		m->busy_lock = VPB_UNBUSIED;
2169 		/* Don't change PG_ZERO. */
2170 		vm_page_free_toq(m);
2171 		if (req & VM_ALLOC_WAITFAIL) {
2172 			VM_OBJECT_WUNLOCK(object);
2173 			vm_radix_wait();
2174 			VM_OBJECT_WLOCK(object);
2175 		}
2176 		return (NULL);
2177 	}
2178 
2179 	/* Ignore device objects; the pager sets "memattr" for them. */
2180 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2181 	    (object->flags & OBJ_FICTITIOUS) == 0)
2182 		pmap_page_set_memattr(m, object->memattr);
2183 
2184 	return (m);
2185 }
2186 
2187 /*
2188  *	vm_page_alloc_contig:
2189  *
2190  *	Allocate a contiguous set of physical pages of the given size "npages"
2191  *	from the free lists.  All of the physical pages must be at or above
2192  *	the given physical address "low" and below the given physical address
2193  *	"high".  The given value "alignment" determines the alignment of the
2194  *	first physical page in the set.  If the given value "boundary" is
2195  *	non-zero, then the set of physical pages cannot cross any physical
2196  *	address boundary that is a multiple of that value.  Both "alignment"
2197  *	and "boundary" must be a power of two.
2198  *
2199  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2200  *	then the memory attribute setting for the physical pages is configured
2201  *	to the object's memory attribute setting.  Otherwise, the memory
2202  *	attribute setting for the physical pages is configured to "memattr",
2203  *	overriding the object's memory attribute setting.  However, if the
2204  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2205  *	memory attribute setting for the physical pages cannot be configured
2206  *	to VM_MEMATTR_DEFAULT.
2207  *
2208  *	The specified object may not contain fictitious pages.
2209  *
2210  *	The caller must always specify an allocation class.
2211  *
2212  *	allocation classes:
2213  *	VM_ALLOC_NORMAL		normal process request
2214  *	VM_ALLOC_SYSTEM		system *really* needs a page
2215  *	VM_ALLOC_INTERRUPT	interrupt time request
2216  *
2217  *	optional allocation flags:
2218  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2219  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2220  *	VM_ALLOC_SBUSY		shared busy the allocated page
2221  *	VM_ALLOC_WIRED		wire the allocated page
2222  *	VM_ALLOC_ZERO		prefer a zeroed page
2223  */
2224 vm_page_t
2225 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2226     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2227     vm_paddr_t boundary, vm_memattr_t memattr)
2228 {
2229 	struct vm_domainset_iter di;
2230 	vm_page_t m;
2231 	int domain;
2232 
2233 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2234 	do {
2235 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2236 		    npages, low, high, alignment, boundary, memattr);
2237 		if (m != NULL)
2238 			break;
2239 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2240 
2241 	return (m);
2242 }
2243 
2244 vm_page_t
2245 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2246     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2247     vm_paddr_t boundary, vm_memattr_t memattr)
2248 {
2249 	struct vm_domain *vmd;
2250 	vm_page_t m, m_ret, mpred;
2251 	u_int busy_lock, flags, oflags;
2252 
2253 #define	VPAC_FLAGS	VPA_FLAGS
2254 	KASSERT((req & ~VPAC_FLAGS) == 0,
2255 	    ("invalid request %#x", req));
2256 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2257 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2258 	    ("invalid request %#x", req));
2259 	VM_OBJECT_ASSERT_WLOCKED(object);
2260 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2261 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2262 	    object));
2263 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2264 
2265 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
2266 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2267 	    ("vm_page_alloc_contig: pindex already allocated"));
2268 
2269 	/*
2270 	 * Can we allocate the pages without the number of free pages falling
2271 	 * below the lower bound for the allocation class?
2272 	 */
2273 	m_ret = NULL;
2274 again:
2275 #if VM_NRESERVLEVEL > 0
2276 	/*
2277 	 * Can we allocate the pages from a reservation?
2278 	 */
2279 	if (vm_object_reserv(object) &&
2280 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2281 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2282 		goto found;
2283 	}
2284 #endif
2285 	vmd = VM_DOMAIN(domain);
2286 	if (vm_domain_allocate(vmd, req, npages)) {
2287 		/*
2288 		 * allocate them from the free page queues.
2289 		 */
2290 		vm_domain_free_lock(vmd);
2291 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2292 		    alignment, boundary);
2293 		vm_domain_free_unlock(vmd);
2294 		if (m_ret == NULL) {
2295 			vm_domain_freecnt_inc(vmd, npages);
2296 #if VM_NRESERVLEVEL > 0
2297 			if (vm_reserv_reclaim_contig(domain, npages, low,
2298 			    high, alignment, boundary))
2299 				goto again;
2300 #endif
2301 		}
2302 	}
2303 	if (m_ret == NULL) {
2304 		if (vm_domain_alloc_fail(vmd, object, req))
2305 			goto again;
2306 		return (NULL);
2307 	}
2308 #if VM_NRESERVLEVEL > 0
2309 found:
2310 #endif
2311 	for (m = m_ret; m < &m_ret[npages]; m++) {
2312 		vm_page_dequeue(m);
2313 		vm_page_alloc_check(m);
2314 	}
2315 
2316 	/*
2317 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2318 	 */
2319 	flags = PG_ZERO;
2320 	if ((req & VM_ALLOC_NODUMP) != 0)
2321 		flags |= PG_NODUMP;
2322 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2323 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2324 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2325 	else if ((req & VM_ALLOC_SBUSY) != 0)
2326 		busy_lock = VPB_SHARERS_WORD(1);
2327 	else
2328 		busy_lock = VPB_UNBUSIED;
2329 	if ((req & VM_ALLOC_WIRED) != 0)
2330 		vm_wire_add(npages);
2331 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2332 	    memattr == VM_MEMATTR_DEFAULT)
2333 		memattr = object->memattr;
2334 	for (m = m_ret; m < &m_ret[npages]; m++) {
2335 		m->a.flags = 0;
2336 		m->flags = (m->flags | PG_NODUMP) & flags;
2337 		m->busy_lock = busy_lock;
2338 		if ((req & VM_ALLOC_WIRED) != 0)
2339 			m->ref_count = 1;
2340 		m->a.act_count = 0;
2341 		m->oflags = oflags;
2342 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2343 			if ((req & VM_ALLOC_WIRED) != 0)
2344 				vm_wire_sub(npages);
2345 			KASSERT(m->object == NULL,
2346 			    ("page %p has object", m));
2347 			mpred = m;
2348 			for (m = m_ret; m < &m_ret[npages]; m++) {
2349 				if (m <= mpred &&
2350 				    (req & VM_ALLOC_WIRED) != 0)
2351 					m->ref_count = 0;
2352 				m->oflags = VPO_UNMANAGED;
2353 				m->busy_lock = VPB_UNBUSIED;
2354 				/* Don't change PG_ZERO. */
2355 				vm_page_free_toq(m);
2356 			}
2357 			if (req & VM_ALLOC_WAITFAIL) {
2358 				VM_OBJECT_WUNLOCK(object);
2359 				vm_radix_wait();
2360 				VM_OBJECT_WLOCK(object);
2361 			}
2362 			return (NULL);
2363 		}
2364 		mpred = m;
2365 		if (memattr != VM_MEMATTR_DEFAULT)
2366 			pmap_page_set_memattr(m, memattr);
2367 		pindex++;
2368 	}
2369 	return (m_ret);
2370 }
2371 
2372 /*
2373  * Allocate a physical page that is not intended to be inserted into a VM
2374  * object.  If the "freelist" parameter is not equal to VM_NFREELIST, then only
2375  * pages from the specified vm_phys freelist will be returned.
2376  */
2377 static __always_inline vm_page_t
2378 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req)
2379 {
2380 	struct vm_domain *vmd;
2381 	vm_page_t m;
2382 	int flags;
2383 
2384 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2385 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2386 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2387 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2388 	KASSERT((req & ~VPAN_FLAGS) == 0,
2389 	    ("invalid request %#x", req));
2390 
2391 	flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0;
2392 	vmd = VM_DOMAIN(domain);
2393 again:
2394 	if (freelist == VM_NFREELIST &&
2395 	    vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2396 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2397 		    M_NOWAIT | M_NOVM);
2398 		if (m != NULL) {
2399 			flags |= PG_PCPU_CACHE;
2400 			goto found;
2401 		}
2402 	}
2403 
2404 	if (vm_domain_allocate(vmd, req, 1)) {
2405 		vm_domain_free_lock(vmd);
2406 		if (freelist == VM_NFREELIST)
2407 			m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2408 		else
2409 			m = vm_phys_alloc_freelist_pages(domain, freelist,
2410 			    VM_FREEPOOL_DIRECT, 0);
2411 		vm_domain_free_unlock(vmd);
2412 		if (m == NULL) {
2413 			vm_domain_freecnt_inc(vmd, 1);
2414 #if VM_NRESERVLEVEL > 0
2415 			if (freelist == VM_NFREELIST &&
2416 			    vm_reserv_reclaim_inactive(domain))
2417 				goto again;
2418 #endif
2419 		}
2420 	}
2421 	if (m == NULL) {
2422 		if (vm_domain_alloc_fail(vmd, NULL, req))
2423 			goto again;
2424 		return (NULL);
2425 	}
2426 
2427 found:
2428 	vm_page_dequeue(m);
2429 	vm_page_alloc_check(m);
2430 
2431 	/*
2432 	 * Consumers should not rely on a useful default pindex value.
2433 	 */
2434 	m->pindex = 0xdeadc0dedeadc0de;
2435 	m->flags = (m->flags & PG_ZERO) | flags;
2436 	m->a.flags = 0;
2437 	m->oflags = VPO_UNMANAGED;
2438 	m->busy_lock = VPB_UNBUSIED;
2439 	if ((req & VM_ALLOC_WIRED) != 0) {
2440 		vm_wire_add(1);
2441 		m->ref_count = 1;
2442 	}
2443 
2444 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2445 		pmap_zero_page(m);
2446 
2447 	return (m);
2448 }
2449 
2450 vm_page_t
2451 vm_page_alloc_freelist(int freelist, int req)
2452 {
2453 	struct vm_domainset_iter di;
2454 	vm_page_t m;
2455 	int domain;
2456 
2457 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2458 	do {
2459 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2460 		if (m != NULL)
2461 			break;
2462 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2463 
2464 	return (m);
2465 }
2466 
2467 vm_page_t
2468 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2469 {
2470 	KASSERT(freelist >= 0 && freelist < VM_NFREELIST,
2471 	    ("%s: invalid freelist %d", __func__, freelist));
2472 
2473 	return (_vm_page_alloc_noobj_domain(domain, freelist, req));
2474 }
2475 
2476 vm_page_t
2477 vm_page_alloc_noobj(int req)
2478 {
2479 	struct vm_domainset_iter di;
2480 	vm_page_t m;
2481 	int domain;
2482 
2483 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2484 	do {
2485 		m = vm_page_alloc_noobj_domain(domain, req);
2486 		if (m != NULL)
2487 			break;
2488 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2489 
2490 	return (m);
2491 }
2492 
2493 vm_page_t
2494 vm_page_alloc_noobj_domain(int domain, int req)
2495 {
2496 	return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req));
2497 }
2498 
2499 vm_page_t
2500 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2501     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2502     vm_memattr_t memattr)
2503 {
2504 	struct vm_domainset_iter di;
2505 	vm_page_t m;
2506 	int domain;
2507 
2508 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2509 	do {
2510 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2511 		    high, alignment, boundary, memattr);
2512 		if (m != NULL)
2513 			break;
2514 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2515 
2516 	return (m);
2517 }
2518 
2519 vm_page_t
2520 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2521     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2522     vm_memattr_t memattr)
2523 {
2524 	struct vm_domain *vmd;
2525 	vm_page_t m, m_ret;
2526 	u_int flags;
2527 
2528 #define	VPANC_FLAGS	VPAN_FLAGS
2529 	KASSERT((req & ~VPANC_FLAGS) == 0,
2530 	    ("invalid request %#x", req));
2531 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2532 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2533 	    ("invalid request %#x", req));
2534 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2535 
2536 	m_ret = NULL;
2537 again:
2538 	vmd = VM_DOMAIN(domain);
2539 	if (vm_domain_allocate(vmd, req, npages)) {
2540 		/*
2541 		 * allocate them from the free page queues.
2542 		 */
2543 		vm_domain_free_lock(vmd);
2544 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2545 		    alignment, boundary);
2546 		vm_domain_free_unlock(vmd);
2547 		if (m_ret == NULL) {
2548 			vm_domain_freecnt_inc(vmd, npages);
2549 #if VM_NRESERVLEVEL > 0
2550 			if (vm_reserv_reclaim_contig(domain, npages, low,
2551 			    high, alignment, boundary))
2552 				goto again;
2553 #endif
2554 		}
2555 	}
2556 	if (m_ret == NULL) {
2557 		if (vm_domain_alloc_fail(vmd, NULL, req))
2558 			goto again;
2559 		return (NULL);
2560 	}
2561 
2562 	/*
2563 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2564 	 */
2565 	flags = PG_ZERO;
2566 	if ((req & VM_ALLOC_NODUMP) != 0)
2567 		flags |= PG_NODUMP;
2568 	if ((req & VM_ALLOC_WIRED) != 0)
2569 		vm_wire_add(npages);
2570 	for (m = m_ret; m < &m_ret[npages]; m++) {
2571 		vm_page_dequeue(m);
2572 		vm_page_alloc_check(m);
2573 
2574 		/*
2575 		 * Consumers should not rely on a useful default pindex value.
2576 		 */
2577 		m->pindex = 0xdeadc0dedeadc0de;
2578 		m->a.flags = 0;
2579 		m->flags = (m->flags | PG_NODUMP) & flags;
2580 		m->busy_lock = VPB_UNBUSIED;
2581 		if ((req & VM_ALLOC_WIRED) != 0)
2582 			m->ref_count = 1;
2583 		m->a.act_count = 0;
2584 		m->oflags = VPO_UNMANAGED;
2585 
2586 		/*
2587 		 * Zero the page before updating any mappings since the page is
2588 		 * not yet shared with any devices which might require the
2589 		 * non-default memory attribute.  pmap_page_set_memattr()
2590 		 * flushes data caches before returning.
2591 		 */
2592 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2593 			pmap_zero_page(m);
2594 		if (memattr != VM_MEMATTR_DEFAULT)
2595 			pmap_page_set_memattr(m, memattr);
2596 	}
2597 	return (m_ret);
2598 }
2599 
2600 /*
2601  * Check a page that has been freshly dequeued from a freelist.
2602  */
2603 static void
2604 vm_page_alloc_check(vm_page_t m)
2605 {
2606 
2607 	KASSERT(m->object == NULL, ("page %p has object", m));
2608 	KASSERT(m->a.queue == PQ_NONE &&
2609 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2610 	    ("page %p has unexpected queue %d, flags %#x",
2611 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2612 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2613 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2614 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2615 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2616 	    ("page %p has unexpected memattr %d",
2617 	    m, pmap_page_get_memattr(m)));
2618 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2619 	pmap_vm_page_alloc_check(m);
2620 }
2621 
2622 static int
2623 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2624 {
2625 	struct vm_domain *vmd;
2626 	struct vm_pgcache *pgcache;
2627 	int i;
2628 
2629 	pgcache = arg;
2630 	vmd = VM_DOMAIN(pgcache->domain);
2631 
2632 	/*
2633 	 * The page daemon should avoid creating extra memory pressure since its
2634 	 * main purpose is to replenish the store of free pages.
2635 	 */
2636 	if (vmd->vmd_severeset || curproc == pageproc ||
2637 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2638 		return (0);
2639 	domain = vmd->vmd_domain;
2640 	vm_domain_free_lock(vmd);
2641 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2642 	    (vm_page_t *)store);
2643 	vm_domain_free_unlock(vmd);
2644 	if (cnt != i)
2645 		vm_domain_freecnt_inc(vmd, cnt - i);
2646 
2647 	return (i);
2648 }
2649 
2650 static void
2651 vm_page_zone_release(void *arg, void **store, int cnt)
2652 {
2653 	struct vm_domain *vmd;
2654 	struct vm_pgcache *pgcache;
2655 	vm_page_t m;
2656 	int i;
2657 
2658 	pgcache = arg;
2659 	vmd = VM_DOMAIN(pgcache->domain);
2660 	vm_domain_free_lock(vmd);
2661 	for (i = 0; i < cnt; i++) {
2662 		m = (vm_page_t)store[i];
2663 		vm_phys_free_pages(m, 0);
2664 	}
2665 	vm_domain_free_unlock(vmd);
2666 	vm_domain_freecnt_inc(vmd, cnt);
2667 }
2668 
2669 #define	VPSC_ANY	0	/* No restrictions. */
2670 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2671 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2672 
2673 /*
2674  *	vm_page_scan_contig:
2675  *
2676  *	Scan vm_page_array[] between the specified entries "m_start" and
2677  *	"m_end" for a run of contiguous physical pages that satisfy the
2678  *	specified conditions, and return the lowest page in the run.  The
2679  *	specified "alignment" determines the alignment of the lowest physical
2680  *	page in the run.  If the specified "boundary" is non-zero, then the
2681  *	run of physical pages cannot span a physical address that is a
2682  *	multiple of "boundary".
2683  *
2684  *	"m_end" is never dereferenced, so it need not point to a vm_page
2685  *	structure within vm_page_array[].
2686  *
2687  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2688  *	span a hole (or discontiguity) in the physical address space.  Both
2689  *	"alignment" and "boundary" must be a power of two.
2690  */
2691 vm_page_t
2692 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2693     u_long alignment, vm_paddr_t boundary, int options)
2694 {
2695 	vm_object_t object;
2696 	vm_paddr_t pa;
2697 	vm_page_t m, m_run;
2698 #if VM_NRESERVLEVEL > 0
2699 	int level;
2700 #endif
2701 	int m_inc, order, run_ext, run_len;
2702 
2703 	KASSERT(npages > 0, ("npages is 0"));
2704 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2705 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2706 	m_run = NULL;
2707 	run_len = 0;
2708 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2709 		KASSERT((m->flags & PG_MARKER) == 0,
2710 		    ("page %p is PG_MARKER", m));
2711 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2712 		    ("fictitious page %p has invalid ref count", m));
2713 
2714 		/*
2715 		 * If the current page would be the start of a run, check its
2716 		 * physical address against the end, alignment, and boundary
2717 		 * conditions.  If it doesn't satisfy these conditions, either
2718 		 * terminate the scan or advance to the next page that
2719 		 * satisfies the failed condition.
2720 		 */
2721 		if (run_len == 0) {
2722 			KASSERT(m_run == NULL, ("m_run != NULL"));
2723 			if (m + npages > m_end)
2724 				break;
2725 			pa = VM_PAGE_TO_PHYS(m);
2726 			if ((pa & (alignment - 1)) != 0) {
2727 				m_inc = atop(roundup2(pa, alignment) - pa);
2728 				continue;
2729 			}
2730 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2731 			    boundary) != 0) {
2732 				m_inc = atop(roundup2(pa, boundary) - pa);
2733 				continue;
2734 			}
2735 		} else
2736 			KASSERT(m_run != NULL, ("m_run == NULL"));
2737 
2738 retry:
2739 		m_inc = 1;
2740 		if (vm_page_wired(m))
2741 			run_ext = 0;
2742 #if VM_NRESERVLEVEL > 0
2743 		else if ((level = vm_reserv_level(m)) >= 0 &&
2744 		    (options & VPSC_NORESERV) != 0) {
2745 			run_ext = 0;
2746 			/* Advance to the end of the reservation. */
2747 			pa = VM_PAGE_TO_PHYS(m);
2748 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2749 			    pa);
2750 		}
2751 #endif
2752 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2753 			/*
2754 			 * The page is considered eligible for relocation if
2755 			 * and only if it could be laundered or reclaimed by
2756 			 * the page daemon.
2757 			 */
2758 			VM_OBJECT_RLOCK(object);
2759 			if (object != m->object) {
2760 				VM_OBJECT_RUNLOCK(object);
2761 				goto retry;
2762 			}
2763 			/* Don't care: PG_NODUMP, PG_ZERO. */
2764 			if (object->type != OBJT_DEFAULT &&
2765 			    (object->flags & OBJ_SWAP) == 0 &&
2766 			    object->type != OBJT_VNODE) {
2767 				run_ext = 0;
2768 #if VM_NRESERVLEVEL > 0
2769 			} else if ((options & VPSC_NOSUPER) != 0 &&
2770 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2771 				run_ext = 0;
2772 				/* Advance to the end of the superpage. */
2773 				pa = VM_PAGE_TO_PHYS(m);
2774 				m_inc = atop(roundup2(pa + 1,
2775 				    vm_reserv_size(level)) - pa);
2776 #endif
2777 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2778 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2779 				/*
2780 				 * The page is allocated but eligible for
2781 				 * relocation.  Extend the current run by one
2782 				 * page.
2783 				 */
2784 				KASSERT(pmap_page_get_memattr(m) ==
2785 				    VM_MEMATTR_DEFAULT,
2786 				    ("page %p has an unexpected memattr", m));
2787 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2788 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2789 				    ("page %p has unexpected oflags", m));
2790 				/* Don't care: PGA_NOSYNC. */
2791 				run_ext = 1;
2792 			} else
2793 				run_ext = 0;
2794 			VM_OBJECT_RUNLOCK(object);
2795 #if VM_NRESERVLEVEL > 0
2796 		} else if (level >= 0) {
2797 			/*
2798 			 * The page is reserved but not yet allocated.  In
2799 			 * other words, it is still free.  Extend the current
2800 			 * run by one page.
2801 			 */
2802 			run_ext = 1;
2803 #endif
2804 		} else if ((order = m->order) < VM_NFREEORDER) {
2805 			/*
2806 			 * The page is enqueued in the physical memory
2807 			 * allocator's free page queues.  Moreover, it is the
2808 			 * first page in a power-of-two-sized run of
2809 			 * contiguous free pages.  Add these pages to the end
2810 			 * of the current run, and jump ahead.
2811 			 */
2812 			run_ext = 1 << order;
2813 			m_inc = 1 << order;
2814 		} else {
2815 			/*
2816 			 * Skip the page for one of the following reasons: (1)
2817 			 * It is enqueued in the physical memory allocator's
2818 			 * free page queues.  However, it is not the first
2819 			 * page in a run of contiguous free pages.  (This case
2820 			 * rarely occurs because the scan is performed in
2821 			 * ascending order.) (2) It is not reserved, and it is
2822 			 * transitioning from free to allocated.  (Conversely,
2823 			 * the transition from allocated to free for managed
2824 			 * pages is blocked by the page busy lock.) (3) It is
2825 			 * allocated but not contained by an object and not
2826 			 * wired, e.g., allocated by Xen's balloon driver.
2827 			 */
2828 			run_ext = 0;
2829 		}
2830 
2831 		/*
2832 		 * Extend or reset the current run of pages.
2833 		 */
2834 		if (run_ext > 0) {
2835 			if (run_len == 0)
2836 				m_run = m;
2837 			run_len += run_ext;
2838 		} else {
2839 			if (run_len > 0) {
2840 				m_run = NULL;
2841 				run_len = 0;
2842 			}
2843 		}
2844 	}
2845 	if (run_len >= npages)
2846 		return (m_run);
2847 	return (NULL);
2848 }
2849 
2850 /*
2851  *	vm_page_reclaim_run:
2852  *
2853  *	Try to relocate each of the allocated virtual pages within the
2854  *	specified run of physical pages to a new physical address.  Free the
2855  *	physical pages underlying the relocated virtual pages.  A virtual page
2856  *	is relocatable if and only if it could be laundered or reclaimed by
2857  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2858  *	physical address above "high".
2859  *
2860  *	Returns 0 if every physical page within the run was already free or
2861  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2862  *	value indicating why the last attempt to relocate a virtual page was
2863  *	unsuccessful.
2864  *
2865  *	"req_class" must be an allocation class.
2866  */
2867 static int
2868 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2869     vm_paddr_t high)
2870 {
2871 	struct vm_domain *vmd;
2872 	struct spglist free;
2873 	vm_object_t object;
2874 	vm_paddr_t pa;
2875 	vm_page_t m, m_end, m_new;
2876 	int error, order, req;
2877 
2878 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2879 	    ("req_class is not an allocation class"));
2880 	SLIST_INIT(&free);
2881 	error = 0;
2882 	m = m_run;
2883 	m_end = m_run + npages;
2884 	for (; error == 0 && m < m_end; m++) {
2885 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2886 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2887 
2888 		/*
2889 		 * Racily check for wirings.  Races are handled once the object
2890 		 * lock is held and the page is unmapped.
2891 		 */
2892 		if (vm_page_wired(m))
2893 			error = EBUSY;
2894 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2895 			/*
2896 			 * The page is relocated if and only if it could be
2897 			 * laundered or reclaimed by the page daemon.
2898 			 */
2899 			VM_OBJECT_WLOCK(object);
2900 			/* Don't care: PG_NODUMP, PG_ZERO. */
2901 			if (m->object != object ||
2902 			    (object->type != OBJT_DEFAULT &&
2903 			    (object->flags & OBJ_SWAP) == 0 &&
2904 			    object->type != OBJT_VNODE))
2905 				error = EINVAL;
2906 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2907 				error = EINVAL;
2908 			else if (vm_page_queue(m) != PQ_NONE &&
2909 			    vm_page_tryxbusy(m) != 0) {
2910 				if (vm_page_wired(m)) {
2911 					vm_page_xunbusy(m);
2912 					error = EBUSY;
2913 					goto unlock;
2914 				}
2915 				KASSERT(pmap_page_get_memattr(m) ==
2916 				    VM_MEMATTR_DEFAULT,
2917 				    ("page %p has an unexpected memattr", m));
2918 				KASSERT(m->oflags == 0,
2919 				    ("page %p has unexpected oflags", m));
2920 				/* Don't care: PGA_NOSYNC. */
2921 				if (!vm_page_none_valid(m)) {
2922 					/*
2923 					 * First, try to allocate a new page
2924 					 * that is above "high".  Failing
2925 					 * that, try to allocate a new page
2926 					 * that is below "m_run".  Allocate
2927 					 * the new page between the end of
2928 					 * "m_run" and "high" only as a last
2929 					 * resort.
2930 					 */
2931 					req = req_class;
2932 					if ((m->flags & PG_NODUMP) != 0)
2933 						req |= VM_ALLOC_NODUMP;
2934 					if (trunc_page(high) !=
2935 					    ~(vm_paddr_t)PAGE_MASK) {
2936 						m_new =
2937 						    vm_page_alloc_noobj_contig(
2938 						    req, 1, round_page(high),
2939 						    ~(vm_paddr_t)0, PAGE_SIZE,
2940 						    0, VM_MEMATTR_DEFAULT);
2941 					} else
2942 						m_new = NULL;
2943 					if (m_new == NULL) {
2944 						pa = VM_PAGE_TO_PHYS(m_run);
2945 						m_new =
2946 						    vm_page_alloc_noobj_contig(
2947 						    req, 1, 0, pa - 1,
2948 						    PAGE_SIZE, 0,
2949 						    VM_MEMATTR_DEFAULT);
2950 					}
2951 					if (m_new == NULL) {
2952 						pa += ptoa(npages);
2953 						m_new =
2954 						    vm_page_alloc_noobj_contig(
2955 						    req, 1, pa, high, PAGE_SIZE,
2956 						    0, VM_MEMATTR_DEFAULT);
2957 					}
2958 					if (m_new == NULL) {
2959 						vm_page_xunbusy(m);
2960 						error = ENOMEM;
2961 						goto unlock;
2962 					}
2963 
2964 					/*
2965 					 * Unmap the page and check for new
2966 					 * wirings that may have been acquired
2967 					 * through a pmap lookup.
2968 					 */
2969 					if (object->ref_count != 0 &&
2970 					    !vm_page_try_remove_all(m)) {
2971 						vm_page_xunbusy(m);
2972 						vm_page_free(m_new);
2973 						error = EBUSY;
2974 						goto unlock;
2975 					}
2976 
2977 					/*
2978 					 * Replace "m" with the new page.  For
2979 					 * vm_page_replace(), "m" must be busy
2980 					 * and dequeued.  Finally, change "m"
2981 					 * as if vm_page_free() was called.
2982 					 */
2983 					m_new->a.flags = m->a.flags &
2984 					    ~PGA_QUEUE_STATE_MASK;
2985 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2986 					    ("page %p is managed", m_new));
2987 					m_new->oflags = 0;
2988 					pmap_copy_page(m, m_new);
2989 					m_new->valid = m->valid;
2990 					m_new->dirty = m->dirty;
2991 					m->flags &= ~PG_ZERO;
2992 					vm_page_dequeue(m);
2993 					if (vm_page_replace_hold(m_new, object,
2994 					    m->pindex, m) &&
2995 					    vm_page_free_prep(m))
2996 						SLIST_INSERT_HEAD(&free, m,
2997 						    plinks.s.ss);
2998 
2999 					/*
3000 					 * The new page must be deactivated
3001 					 * before the object is unlocked.
3002 					 */
3003 					vm_page_deactivate(m_new);
3004 				} else {
3005 					m->flags &= ~PG_ZERO;
3006 					vm_page_dequeue(m);
3007 					if (vm_page_free_prep(m))
3008 						SLIST_INSERT_HEAD(&free, m,
3009 						    plinks.s.ss);
3010 					KASSERT(m->dirty == 0,
3011 					    ("page %p is dirty", m));
3012 				}
3013 			} else
3014 				error = EBUSY;
3015 unlock:
3016 			VM_OBJECT_WUNLOCK(object);
3017 		} else {
3018 			MPASS(vm_page_domain(m) == domain);
3019 			vmd = VM_DOMAIN(domain);
3020 			vm_domain_free_lock(vmd);
3021 			order = m->order;
3022 			if (order < VM_NFREEORDER) {
3023 				/*
3024 				 * The page is enqueued in the physical memory
3025 				 * allocator's free page queues.  Moreover, it
3026 				 * is the first page in a power-of-two-sized
3027 				 * run of contiguous free pages.  Jump ahead
3028 				 * to the last page within that run, and
3029 				 * continue from there.
3030 				 */
3031 				m += (1 << order) - 1;
3032 			}
3033 #if VM_NRESERVLEVEL > 0
3034 			else if (vm_reserv_is_page_free(m))
3035 				order = 0;
3036 #endif
3037 			vm_domain_free_unlock(vmd);
3038 			if (order == VM_NFREEORDER)
3039 				error = EINVAL;
3040 		}
3041 	}
3042 	if ((m = SLIST_FIRST(&free)) != NULL) {
3043 		int cnt;
3044 
3045 		vmd = VM_DOMAIN(domain);
3046 		cnt = 0;
3047 		vm_domain_free_lock(vmd);
3048 		do {
3049 			MPASS(vm_page_domain(m) == domain);
3050 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3051 			vm_phys_free_pages(m, 0);
3052 			cnt++;
3053 		} while ((m = SLIST_FIRST(&free)) != NULL);
3054 		vm_domain_free_unlock(vmd);
3055 		vm_domain_freecnt_inc(vmd, cnt);
3056 	}
3057 	return (error);
3058 }
3059 
3060 #define	NRUNS	16
3061 
3062 CTASSERT(powerof2(NRUNS));
3063 
3064 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
3065 
3066 #define	MIN_RECLAIM	8
3067 
3068 /*
3069  *	vm_page_reclaim_contig:
3070  *
3071  *	Reclaim allocated, contiguous physical memory satisfying the specified
3072  *	conditions by relocating the virtual pages using that physical memory.
3073  *	Returns true if reclamation is successful and false otherwise.  Since
3074  *	relocation requires the allocation of physical pages, reclamation may
3075  *	fail due to a shortage of free pages.  When reclamation fails, callers
3076  *	are expected to perform vm_wait() before retrying a failed allocation
3077  *	operation, e.g., vm_page_alloc_contig().
3078  *
3079  *	The caller must always specify an allocation class through "req".
3080  *
3081  *	allocation classes:
3082  *	VM_ALLOC_NORMAL		normal process request
3083  *	VM_ALLOC_SYSTEM		system *really* needs a page
3084  *	VM_ALLOC_INTERRUPT	interrupt time request
3085  *
3086  *	The optional allocation flags are ignored.
3087  *
3088  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3089  *	must be a power of two.
3090  */
3091 bool
3092 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3093     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3094 {
3095 	struct vm_domain *vmd;
3096 	vm_paddr_t curr_low;
3097 	vm_page_t m_run, m_runs[NRUNS];
3098 	u_long count, minalign, reclaimed;
3099 	int error, i, options, req_class;
3100 
3101 	KASSERT(npages > 0, ("npages is 0"));
3102 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3103 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3104 
3105 	/*
3106 	 * The caller will attempt an allocation after some runs have been
3107 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3108 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3109 	 * power of two or maximum chunk size, and ensure that each run is
3110 	 * suitably aligned.
3111 	 */
3112 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3113 	npages = roundup2(npages, minalign);
3114 	if (alignment < ptoa(minalign))
3115 		alignment = ptoa(minalign);
3116 
3117 	/*
3118 	 * The page daemon is allowed to dig deeper into the free page list.
3119 	 */
3120 	req_class = req & VM_ALLOC_CLASS_MASK;
3121 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3122 		req_class = VM_ALLOC_SYSTEM;
3123 
3124 	/*
3125 	 * Return if the number of free pages cannot satisfy the requested
3126 	 * allocation.
3127 	 */
3128 	vmd = VM_DOMAIN(domain);
3129 	count = vmd->vmd_free_count;
3130 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3131 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3132 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3133 		return (false);
3134 
3135 	/*
3136 	 * Scan up to three times, relaxing the restrictions ("options") on
3137 	 * the reclamation of reservations and superpages each time.
3138 	 */
3139 	for (options = VPSC_NORESERV;;) {
3140 		/*
3141 		 * Find the highest runs that satisfy the given constraints
3142 		 * and restrictions, and record them in "m_runs".
3143 		 */
3144 		curr_low = low;
3145 		count = 0;
3146 		for (;;) {
3147 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
3148 			    high, alignment, boundary, options);
3149 			if (m_run == NULL)
3150 				break;
3151 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
3152 			m_runs[RUN_INDEX(count)] = m_run;
3153 			count++;
3154 		}
3155 
3156 		/*
3157 		 * Reclaim the highest runs in LIFO (descending) order until
3158 		 * the number of reclaimed pages, "reclaimed", is at least
3159 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
3160 		 * reclamation is idempotent, and runs will (likely) recur
3161 		 * from one scan to the next as restrictions are relaxed.
3162 		 */
3163 		reclaimed = 0;
3164 		for (i = 0; count > 0 && i < NRUNS; i++) {
3165 			count--;
3166 			m_run = m_runs[RUN_INDEX(count)];
3167 			error = vm_page_reclaim_run(req_class, domain, npages,
3168 			    m_run, high);
3169 			if (error == 0) {
3170 				reclaimed += npages;
3171 				if (reclaimed >= MIN_RECLAIM)
3172 					return (true);
3173 			}
3174 		}
3175 
3176 		/*
3177 		 * Either relax the restrictions on the next scan or return if
3178 		 * the last scan had no restrictions.
3179 		 */
3180 		if (options == VPSC_NORESERV)
3181 			options = VPSC_NOSUPER;
3182 		else if (options == VPSC_NOSUPER)
3183 			options = VPSC_ANY;
3184 		else if (options == VPSC_ANY)
3185 			return (reclaimed != 0);
3186 	}
3187 }
3188 
3189 bool
3190 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3191     u_long alignment, vm_paddr_t boundary)
3192 {
3193 	struct vm_domainset_iter di;
3194 	int domain;
3195 	bool ret;
3196 
3197 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3198 	do {
3199 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3200 		    high, alignment, boundary);
3201 		if (ret)
3202 			break;
3203 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3204 
3205 	return (ret);
3206 }
3207 
3208 /*
3209  * Set the domain in the appropriate page level domainset.
3210  */
3211 void
3212 vm_domain_set(struct vm_domain *vmd)
3213 {
3214 
3215 	mtx_lock(&vm_domainset_lock);
3216 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3217 		vmd->vmd_minset = 1;
3218 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3219 	}
3220 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3221 		vmd->vmd_severeset = 1;
3222 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3223 	}
3224 	mtx_unlock(&vm_domainset_lock);
3225 }
3226 
3227 /*
3228  * Clear the domain from the appropriate page level domainset.
3229  */
3230 void
3231 vm_domain_clear(struct vm_domain *vmd)
3232 {
3233 
3234 	mtx_lock(&vm_domainset_lock);
3235 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3236 		vmd->vmd_minset = 0;
3237 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3238 		if (vm_min_waiters != 0) {
3239 			vm_min_waiters = 0;
3240 			wakeup(&vm_min_domains);
3241 		}
3242 	}
3243 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3244 		vmd->vmd_severeset = 0;
3245 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3246 		if (vm_severe_waiters != 0) {
3247 			vm_severe_waiters = 0;
3248 			wakeup(&vm_severe_domains);
3249 		}
3250 	}
3251 
3252 	/*
3253 	 * If pageout daemon needs pages, then tell it that there are
3254 	 * some free.
3255 	 */
3256 	if (vmd->vmd_pageout_pages_needed &&
3257 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3258 		wakeup(&vmd->vmd_pageout_pages_needed);
3259 		vmd->vmd_pageout_pages_needed = 0;
3260 	}
3261 
3262 	/* See comments in vm_wait_doms(). */
3263 	if (vm_pageproc_waiters) {
3264 		vm_pageproc_waiters = 0;
3265 		wakeup(&vm_pageproc_waiters);
3266 	}
3267 	mtx_unlock(&vm_domainset_lock);
3268 }
3269 
3270 /*
3271  * Wait for free pages to exceed the min threshold globally.
3272  */
3273 void
3274 vm_wait_min(void)
3275 {
3276 
3277 	mtx_lock(&vm_domainset_lock);
3278 	while (vm_page_count_min()) {
3279 		vm_min_waiters++;
3280 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3281 	}
3282 	mtx_unlock(&vm_domainset_lock);
3283 }
3284 
3285 /*
3286  * Wait for free pages to exceed the severe threshold globally.
3287  */
3288 void
3289 vm_wait_severe(void)
3290 {
3291 
3292 	mtx_lock(&vm_domainset_lock);
3293 	while (vm_page_count_severe()) {
3294 		vm_severe_waiters++;
3295 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3296 		    "vmwait", 0);
3297 	}
3298 	mtx_unlock(&vm_domainset_lock);
3299 }
3300 
3301 u_int
3302 vm_wait_count(void)
3303 {
3304 
3305 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3306 }
3307 
3308 int
3309 vm_wait_doms(const domainset_t *wdoms, int mflags)
3310 {
3311 	int error;
3312 
3313 	error = 0;
3314 
3315 	/*
3316 	 * We use racey wakeup synchronization to avoid expensive global
3317 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3318 	 * To handle this, we only sleep for one tick in this instance.  It
3319 	 * is expected that most allocations for the pageproc will come from
3320 	 * kmem or vm_page_grab* which will use the more specific and
3321 	 * race-free vm_wait_domain().
3322 	 */
3323 	if (curproc == pageproc) {
3324 		mtx_lock(&vm_domainset_lock);
3325 		vm_pageproc_waiters++;
3326 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3327 		    PVM | PDROP | mflags, "pageprocwait", 1);
3328 	} else {
3329 		/*
3330 		 * XXX Ideally we would wait only until the allocation could
3331 		 * be satisfied.  This condition can cause new allocators to
3332 		 * consume all freed pages while old allocators wait.
3333 		 */
3334 		mtx_lock(&vm_domainset_lock);
3335 		if (vm_page_count_min_set(wdoms)) {
3336 			vm_min_waiters++;
3337 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3338 			    PVM | PDROP | mflags, "vmwait", 0);
3339 		} else
3340 			mtx_unlock(&vm_domainset_lock);
3341 	}
3342 	return (error);
3343 }
3344 
3345 /*
3346  *	vm_wait_domain:
3347  *
3348  *	Sleep until free pages are available for allocation.
3349  *	- Called in various places after failed memory allocations.
3350  */
3351 void
3352 vm_wait_domain(int domain)
3353 {
3354 	struct vm_domain *vmd;
3355 	domainset_t wdom;
3356 
3357 	vmd = VM_DOMAIN(domain);
3358 	vm_domain_free_assert_unlocked(vmd);
3359 
3360 	if (curproc == pageproc) {
3361 		mtx_lock(&vm_domainset_lock);
3362 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3363 			vmd->vmd_pageout_pages_needed = 1;
3364 			msleep(&vmd->vmd_pageout_pages_needed,
3365 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3366 		} else
3367 			mtx_unlock(&vm_domainset_lock);
3368 	} else {
3369 		if (pageproc == NULL)
3370 			panic("vm_wait in early boot");
3371 		DOMAINSET_ZERO(&wdom);
3372 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3373 		vm_wait_doms(&wdom, 0);
3374 	}
3375 }
3376 
3377 static int
3378 vm_wait_flags(vm_object_t obj, int mflags)
3379 {
3380 	struct domainset *d;
3381 
3382 	d = NULL;
3383 
3384 	/*
3385 	 * Carefully fetch pointers only once: the struct domainset
3386 	 * itself is ummutable but the pointer might change.
3387 	 */
3388 	if (obj != NULL)
3389 		d = obj->domain.dr_policy;
3390 	if (d == NULL)
3391 		d = curthread->td_domain.dr_policy;
3392 
3393 	return (vm_wait_doms(&d->ds_mask, mflags));
3394 }
3395 
3396 /*
3397  *	vm_wait:
3398  *
3399  *	Sleep until free pages are available for allocation in the
3400  *	affinity domains of the obj.  If obj is NULL, the domain set
3401  *	for the calling thread is used.
3402  *	Called in various places after failed memory allocations.
3403  */
3404 void
3405 vm_wait(vm_object_t obj)
3406 {
3407 	(void)vm_wait_flags(obj, 0);
3408 }
3409 
3410 int
3411 vm_wait_intr(vm_object_t obj)
3412 {
3413 	return (vm_wait_flags(obj, PCATCH));
3414 }
3415 
3416 /*
3417  *	vm_domain_alloc_fail:
3418  *
3419  *	Called when a page allocation function fails.  Informs the
3420  *	pagedaemon and performs the requested wait.  Requires the
3421  *	domain_free and object lock on entry.  Returns with the
3422  *	object lock held and free lock released.  Returns an error when
3423  *	retry is necessary.
3424  *
3425  */
3426 static int
3427 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3428 {
3429 
3430 	vm_domain_free_assert_unlocked(vmd);
3431 
3432 	atomic_add_int(&vmd->vmd_pageout_deficit,
3433 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3434 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3435 		if (object != NULL)
3436 			VM_OBJECT_WUNLOCK(object);
3437 		vm_wait_domain(vmd->vmd_domain);
3438 		if (object != NULL)
3439 			VM_OBJECT_WLOCK(object);
3440 		if (req & VM_ALLOC_WAITOK)
3441 			return (EAGAIN);
3442 	}
3443 
3444 	return (0);
3445 }
3446 
3447 /*
3448  *	vm_waitpfault:
3449  *
3450  *	Sleep until free pages are available for allocation.
3451  *	- Called only in vm_fault so that processes page faulting
3452  *	  can be easily tracked.
3453  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3454  *	  processes will be able to grab memory first.  Do not change
3455  *	  this balance without careful testing first.
3456  */
3457 void
3458 vm_waitpfault(struct domainset *dset, int timo)
3459 {
3460 
3461 	/*
3462 	 * XXX Ideally we would wait only until the allocation could
3463 	 * be satisfied.  This condition can cause new allocators to
3464 	 * consume all freed pages while old allocators wait.
3465 	 */
3466 	mtx_lock(&vm_domainset_lock);
3467 	if (vm_page_count_min_set(&dset->ds_mask)) {
3468 		vm_min_waiters++;
3469 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3470 		    "pfault", timo);
3471 	} else
3472 		mtx_unlock(&vm_domainset_lock);
3473 }
3474 
3475 static struct vm_pagequeue *
3476 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3477 {
3478 
3479 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3480 }
3481 
3482 #ifdef INVARIANTS
3483 static struct vm_pagequeue *
3484 vm_page_pagequeue(vm_page_t m)
3485 {
3486 
3487 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3488 }
3489 #endif
3490 
3491 static __always_inline bool
3492 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3493 {
3494 	vm_page_astate_t tmp;
3495 
3496 	tmp = *old;
3497 	do {
3498 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3499 			return (true);
3500 		counter_u64_add(pqstate_commit_retries, 1);
3501 	} while (old->_bits == tmp._bits);
3502 
3503 	return (false);
3504 }
3505 
3506 /*
3507  * Do the work of committing a queue state update that moves the page out of
3508  * its current queue.
3509  */
3510 static bool
3511 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3512     vm_page_astate_t *old, vm_page_astate_t new)
3513 {
3514 	vm_page_t next;
3515 
3516 	vm_pagequeue_assert_locked(pq);
3517 	KASSERT(vm_page_pagequeue(m) == pq,
3518 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3519 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3520 	    ("%s: invalid queue indices %d %d",
3521 	    __func__, old->queue, new.queue));
3522 
3523 	/*
3524 	 * Once the queue index of the page changes there is nothing
3525 	 * synchronizing with further updates to the page's physical
3526 	 * queue state.  Therefore we must speculatively remove the page
3527 	 * from the queue now and be prepared to roll back if the queue
3528 	 * state update fails.  If the page is not physically enqueued then
3529 	 * we just update its queue index.
3530 	 */
3531 	if ((old->flags & PGA_ENQUEUED) != 0) {
3532 		new.flags &= ~PGA_ENQUEUED;
3533 		next = TAILQ_NEXT(m, plinks.q);
3534 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3535 		vm_pagequeue_cnt_dec(pq);
3536 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3537 			if (next == NULL)
3538 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3539 			else
3540 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3541 			vm_pagequeue_cnt_inc(pq);
3542 			return (false);
3543 		} else {
3544 			return (true);
3545 		}
3546 	} else {
3547 		return (vm_page_pqstate_fcmpset(m, old, new));
3548 	}
3549 }
3550 
3551 static bool
3552 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3553     vm_page_astate_t new)
3554 {
3555 	struct vm_pagequeue *pq;
3556 	vm_page_astate_t as;
3557 	bool ret;
3558 
3559 	pq = _vm_page_pagequeue(m, old->queue);
3560 
3561 	/*
3562 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3563 	 * corresponding page queue lock is held.
3564 	 */
3565 	vm_pagequeue_lock(pq);
3566 	as = vm_page_astate_load(m);
3567 	if (__predict_false(as._bits != old->_bits)) {
3568 		*old = as;
3569 		ret = false;
3570 	} else {
3571 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3572 	}
3573 	vm_pagequeue_unlock(pq);
3574 	return (ret);
3575 }
3576 
3577 /*
3578  * Commit a queue state update that enqueues or requeues a page.
3579  */
3580 static bool
3581 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3582     vm_page_astate_t *old, vm_page_astate_t new)
3583 {
3584 	struct vm_domain *vmd;
3585 
3586 	vm_pagequeue_assert_locked(pq);
3587 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3588 	    ("%s: invalid queue indices %d %d",
3589 	    __func__, old->queue, new.queue));
3590 
3591 	new.flags |= PGA_ENQUEUED;
3592 	if (!vm_page_pqstate_fcmpset(m, old, new))
3593 		return (false);
3594 
3595 	if ((old->flags & PGA_ENQUEUED) != 0)
3596 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3597 	else
3598 		vm_pagequeue_cnt_inc(pq);
3599 
3600 	/*
3601 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3602 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3603 	 * applied, even if it was set first.
3604 	 */
3605 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3606 		vmd = vm_pagequeue_domain(m);
3607 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3608 		    ("%s: invalid page queue for page %p", __func__, m));
3609 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3610 	} else {
3611 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3612 	}
3613 	return (true);
3614 }
3615 
3616 /*
3617  * Commit a queue state update that encodes a request for a deferred queue
3618  * operation.
3619  */
3620 static bool
3621 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3622     vm_page_astate_t new)
3623 {
3624 
3625 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3626 	    ("%s: invalid state, queue %d flags %x",
3627 	    __func__, new.queue, new.flags));
3628 
3629 	if (old->_bits != new._bits &&
3630 	    !vm_page_pqstate_fcmpset(m, old, new))
3631 		return (false);
3632 	vm_page_pqbatch_submit(m, new.queue);
3633 	return (true);
3634 }
3635 
3636 /*
3637  * A generic queue state update function.  This handles more cases than the
3638  * specialized functions above.
3639  */
3640 bool
3641 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3642 {
3643 
3644 	if (old->_bits == new._bits)
3645 		return (true);
3646 
3647 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3648 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3649 			return (false);
3650 		if (new.queue != PQ_NONE)
3651 			vm_page_pqbatch_submit(m, new.queue);
3652 	} else {
3653 		if (!vm_page_pqstate_fcmpset(m, old, new))
3654 			return (false);
3655 		if (new.queue != PQ_NONE &&
3656 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3657 			vm_page_pqbatch_submit(m, new.queue);
3658 	}
3659 	return (true);
3660 }
3661 
3662 /*
3663  * Apply deferred queue state updates to a page.
3664  */
3665 static inline void
3666 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3667 {
3668 	vm_page_astate_t new, old;
3669 
3670 	CRITICAL_ASSERT(curthread);
3671 	vm_pagequeue_assert_locked(pq);
3672 	KASSERT(queue < PQ_COUNT,
3673 	    ("%s: invalid queue index %d", __func__, queue));
3674 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3675 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3676 
3677 	for (old = vm_page_astate_load(m);;) {
3678 		if (__predict_false(old.queue != queue ||
3679 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3680 			counter_u64_add(queue_nops, 1);
3681 			break;
3682 		}
3683 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3684 		    ("%s: page %p is unmanaged", __func__, m));
3685 
3686 		new = old;
3687 		if ((old.flags & PGA_DEQUEUE) != 0) {
3688 			new.flags &= ~PGA_QUEUE_OP_MASK;
3689 			new.queue = PQ_NONE;
3690 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3691 			    m, &old, new))) {
3692 				counter_u64_add(queue_ops, 1);
3693 				break;
3694 			}
3695 		} else {
3696 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3697 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3698 			    m, &old, new))) {
3699 				counter_u64_add(queue_ops, 1);
3700 				break;
3701 			}
3702 		}
3703 	}
3704 }
3705 
3706 static void
3707 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3708     uint8_t queue)
3709 {
3710 	int i;
3711 
3712 	for (i = 0; i < bq->bq_cnt; i++)
3713 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3714 	vm_batchqueue_init(bq);
3715 }
3716 
3717 /*
3718  *	vm_page_pqbatch_submit:		[ internal use only ]
3719  *
3720  *	Enqueue a page in the specified page queue's batched work queue.
3721  *	The caller must have encoded the requested operation in the page
3722  *	structure's a.flags field.
3723  */
3724 void
3725 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3726 {
3727 	struct vm_batchqueue *bq;
3728 	struct vm_pagequeue *pq;
3729 	int domain;
3730 
3731 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3732 
3733 	domain = vm_page_domain(m);
3734 	critical_enter();
3735 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3736 	if (vm_batchqueue_insert(bq, m)) {
3737 		critical_exit();
3738 		return;
3739 	}
3740 	critical_exit();
3741 
3742 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3743 	vm_pagequeue_lock(pq);
3744 	critical_enter();
3745 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3746 	vm_pqbatch_process(pq, bq, queue);
3747 	vm_pqbatch_process_page(pq, m, queue);
3748 	vm_pagequeue_unlock(pq);
3749 	critical_exit();
3750 }
3751 
3752 /*
3753  *	vm_page_pqbatch_drain:		[ internal use only ]
3754  *
3755  *	Force all per-CPU page queue batch queues to be drained.  This is
3756  *	intended for use in severe memory shortages, to ensure that pages
3757  *	do not remain stuck in the batch queues.
3758  */
3759 void
3760 vm_page_pqbatch_drain(void)
3761 {
3762 	struct thread *td;
3763 	struct vm_domain *vmd;
3764 	struct vm_pagequeue *pq;
3765 	int cpu, domain, queue;
3766 
3767 	td = curthread;
3768 	CPU_FOREACH(cpu) {
3769 		thread_lock(td);
3770 		sched_bind(td, cpu);
3771 		thread_unlock(td);
3772 
3773 		for (domain = 0; domain < vm_ndomains; domain++) {
3774 			vmd = VM_DOMAIN(domain);
3775 			for (queue = 0; queue < PQ_COUNT; queue++) {
3776 				pq = &vmd->vmd_pagequeues[queue];
3777 				vm_pagequeue_lock(pq);
3778 				critical_enter();
3779 				vm_pqbatch_process(pq,
3780 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3781 				critical_exit();
3782 				vm_pagequeue_unlock(pq);
3783 			}
3784 		}
3785 	}
3786 	thread_lock(td);
3787 	sched_unbind(td);
3788 	thread_unlock(td);
3789 }
3790 
3791 /*
3792  *	vm_page_dequeue_deferred:	[ internal use only ]
3793  *
3794  *	Request removal of the given page from its current page
3795  *	queue.  Physical removal from the queue may be deferred
3796  *	indefinitely.
3797  */
3798 void
3799 vm_page_dequeue_deferred(vm_page_t m)
3800 {
3801 	vm_page_astate_t new, old;
3802 
3803 	old = vm_page_astate_load(m);
3804 	do {
3805 		if (old.queue == PQ_NONE) {
3806 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3807 			    ("%s: page %p has unexpected queue state",
3808 			    __func__, m));
3809 			break;
3810 		}
3811 		new = old;
3812 		new.flags |= PGA_DEQUEUE;
3813 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3814 }
3815 
3816 /*
3817  *	vm_page_dequeue:
3818  *
3819  *	Remove the page from whichever page queue it's in, if any, before
3820  *	returning.
3821  */
3822 void
3823 vm_page_dequeue(vm_page_t m)
3824 {
3825 	vm_page_astate_t new, old;
3826 
3827 	old = vm_page_astate_load(m);
3828 	do {
3829 		if (old.queue == PQ_NONE) {
3830 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3831 			    ("%s: page %p has unexpected queue state",
3832 			    __func__, m));
3833 			break;
3834 		}
3835 		new = old;
3836 		new.flags &= ~PGA_QUEUE_OP_MASK;
3837 		new.queue = PQ_NONE;
3838 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3839 
3840 }
3841 
3842 /*
3843  * Schedule the given page for insertion into the specified page queue.
3844  * Physical insertion of the page may be deferred indefinitely.
3845  */
3846 static void
3847 vm_page_enqueue(vm_page_t m, uint8_t queue)
3848 {
3849 
3850 	KASSERT(m->a.queue == PQ_NONE &&
3851 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3852 	    ("%s: page %p is already enqueued", __func__, m));
3853 	KASSERT(m->ref_count > 0,
3854 	    ("%s: page %p does not carry any references", __func__, m));
3855 
3856 	m->a.queue = queue;
3857 	if ((m->a.flags & PGA_REQUEUE) == 0)
3858 		vm_page_aflag_set(m, PGA_REQUEUE);
3859 	vm_page_pqbatch_submit(m, queue);
3860 }
3861 
3862 /*
3863  *	vm_page_free_prep:
3864  *
3865  *	Prepares the given page to be put on the free list,
3866  *	disassociating it from any VM object. The caller may return
3867  *	the page to the free list only if this function returns true.
3868  *
3869  *	The object, if it exists, must be locked, and then the page must
3870  *	be xbusy.  Otherwise the page must be not busied.  A managed
3871  *	page must be unmapped.
3872  */
3873 static bool
3874 vm_page_free_prep(vm_page_t m)
3875 {
3876 
3877 	/*
3878 	 * Synchronize with threads that have dropped a reference to this
3879 	 * page.
3880 	 */
3881 	atomic_thread_fence_acq();
3882 
3883 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3884 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3885 		uint64_t *p;
3886 		int i;
3887 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3888 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3889 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3890 			    m, i, (uintmax_t)*p));
3891 	}
3892 #endif
3893 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3894 		KASSERT(!pmap_page_is_mapped(m),
3895 		    ("vm_page_free_prep: freeing mapped page %p", m));
3896 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3897 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3898 	} else {
3899 		KASSERT(m->a.queue == PQ_NONE,
3900 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3901 	}
3902 	VM_CNT_INC(v_tfree);
3903 
3904 	if (m->object != NULL) {
3905 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3906 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3907 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3908 		    m));
3909 		vm_page_assert_xbusied(m);
3910 
3911 		/*
3912 		 * The object reference can be released without an atomic
3913 		 * operation.
3914 		 */
3915 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3916 		    m->ref_count == VPRC_OBJREF,
3917 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3918 		    m, m->ref_count));
3919 		vm_page_object_remove(m);
3920 		m->ref_count -= VPRC_OBJREF;
3921 	} else
3922 		vm_page_assert_unbusied(m);
3923 
3924 	vm_page_busy_free(m);
3925 
3926 	/*
3927 	 * If fictitious remove object association and
3928 	 * return.
3929 	 */
3930 	if ((m->flags & PG_FICTITIOUS) != 0) {
3931 		KASSERT(m->ref_count == 1,
3932 		    ("fictitious page %p is referenced", m));
3933 		KASSERT(m->a.queue == PQ_NONE,
3934 		    ("fictitious page %p is queued", m));
3935 		return (false);
3936 	}
3937 
3938 	/*
3939 	 * Pages need not be dequeued before they are returned to the physical
3940 	 * memory allocator, but they must at least be marked for a deferred
3941 	 * dequeue.
3942 	 */
3943 	if ((m->oflags & VPO_UNMANAGED) == 0)
3944 		vm_page_dequeue_deferred(m);
3945 
3946 	m->valid = 0;
3947 	vm_page_undirty(m);
3948 
3949 	if (m->ref_count != 0)
3950 		panic("vm_page_free_prep: page %p has references", m);
3951 
3952 	/*
3953 	 * Restore the default memory attribute to the page.
3954 	 */
3955 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3956 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3957 
3958 #if VM_NRESERVLEVEL > 0
3959 	/*
3960 	 * Determine whether the page belongs to a reservation.  If the page was
3961 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3962 	 * as an optimization, we avoid the check in that case.
3963 	 */
3964 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3965 		return (false);
3966 #endif
3967 
3968 	return (true);
3969 }
3970 
3971 /*
3972  *	vm_page_free_toq:
3973  *
3974  *	Returns the given page to the free list, disassociating it
3975  *	from any VM object.
3976  *
3977  *	The object must be locked.  The page must be exclusively busied if it
3978  *	belongs to an object.
3979  */
3980 static void
3981 vm_page_free_toq(vm_page_t m)
3982 {
3983 	struct vm_domain *vmd;
3984 	uma_zone_t zone;
3985 
3986 	if (!vm_page_free_prep(m))
3987 		return;
3988 
3989 	vmd = vm_pagequeue_domain(m);
3990 	zone = vmd->vmd_pgcache[m->pool].zone;
3991 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3992 		uma_zfree(zone, m);
3993 		return;
3994 	}
3995 	vm_domain_free_lock(vmd);
3996 	vm_phys_free_pages(m, 0);
3997 	vm_domain_free_unlock(vmd);
3998 	vm_domain_freecnt_inc(vmd, 1);
3999 }
4000 
4001 /*
4002  *	vm_page_free_pages_toq:
4003  *
4004  *	Returns a list of pages to the free list, disassociating it
4005  *	from any VM object.  In other words, this is equivalent to
4006  *	calling vm_page_free_toq() for each page of a list of VM objects.
4007  */
4008 void
4009 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4010 {
4011 	vm_page_t m;
4012 	int count;
4013 
4014 	if (SLIST_EMPTY(free))
4015 		return;
4016 
4017 	count = 0;
4018 	while ((m = SLIST_FIRST(free)) != NULL) {
4019 		count++;
4020 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4021 		vm_page_free_toq(m);
4022 	}
4023 
4024 	if (update_wire_count)
4025 		vm_wire_sub(count);
4026 }
4027 
4028 /*
4029  * Mark this page as wired down.  For managed pages, this prevents reclamation
4030  * by the page daemon, or when the containing object, if any, is destroyed.
4031  */
4032 void
4033 vm_page_wire(vm_page_t m)
4034 {
4035 	u_int old;
4036 
4037 #ifdef INVARIANTS
4038 	if (m->object != NULL && !vm_page_busied(m) &&
4039 	    !vm_object_busied(m->object))
4040 		VM_OBJECT_ASSERT_LOCKED(m->object);
4041 #endif
4042 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4043 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4044 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4045 
4046 	old = atomic_fetchadd_int(&m->ref_count, 1);
4047 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4048 	    ("vm_page_wire: counter overflow for page %p", m));
4049 	if (VPRC_WIRE_COUNT(old) == 0) {
4050 		if ((m->oflags & VPO_UNMANAGED) == 0)
4051 			vm_page_aflag_set(m, PGA_DEQUEUE);
4052 		vm_wire_add(1);
4053 	}
4054 }
4055 
4056 /*
4057  * Attempt to wire a mapped page following a pmap lookup of that page.
4058  * This may fail if a thread is concurrently tearing down mappings of the page.
4059  * The transient failure is acceptable because it translates to the
4060  * failure of the caller pmap_extract_and_hold(), which should be then
4061  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4062  */
4063 bool
4064 vm_page_wire_mapped(vm_page_t m)
4065 {
4066 	u_int old;
4067 
4068 	old = m->ref_count;
4069 	do {
4070 		KASSERT(old > 0,
4071 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4072 		if ((old & VPRC_BLOCKED) != 0)
4073 			return (false);
4074 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4075 
4076 	if (VPRC_WIRE_COUNT(old) == 0) {
4077 		if ((m->oflags & VPO_UNMANAGED) == 0)
4078 			vm_page_aflag_set(m, PGA_DEQUEUE);
4079 		vm_wire_add(1);
4080 	}
4081 	return (true);
4082 }
4083 
4084 /*
4085  * Release a wiring reference to a managed page.  If the page still belongs to
4086  * an object, update its position in the page queues to reflect the reference.
4087  * If the wiring was the last reference to the page, free the page.
4088  */
4089 static void
4090 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4091 {
4092 	u_int old;
4093 
4094 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4095 	    ("%s: page %p is unmanaged", __func__, m));
4096 
4097 	/*
4098 	 * Update LRU state before releasing the wiring reference.
4099 	 * Use a release store when updating the reference count to
4100 	 * synchronize with vm_page_free_prep().
4101 	 */
4102 	old = m->ref_count;
4103 	do {
4104 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4105 		    ("vm_page_unwire: wire count underflow for page %p", m));
4106 
4107 		if (old > VPRC_OBJREF + 1) {
4108 			/*
4109 			 * The page has at least one other wiring reference.  An
4110 			 * earlier iteration of this loop may have called
4111 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4112 			 * re-set it if necessary.
4113 			 */
4114 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4115 				vm_page_aflag_set(m, PGA_DEQUEUE);
4116 		} else if (old == VPRC_OBJREF + 1) {
4117 			/*
4118 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4119 			 * update the page's queue state to reflect the
4120 			 * reference.  If the page does not belong to an object
4121 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4122 			 * clear leftover queue state.
4123 			 */
4124 			vm_page_release_toq(m, nqueue, noreuse);
4125 		} else if (old == 1) {
4126 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4127 		}
4128 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4129 
4130 	if (VPRC_WIRE_COUNT(old) == 1) {
4131 		vm_wire_sub(1);
4132 		if (old == 1)
4133 			vm_page_free(m);
4134 	}
4135 }
4136 
4137 /*
4138  * Release one wiring of the specified page, potentially allowing it to be
4139  * paged out.
4140  *
4141  * Only managed pages belonging to an object can be paged out.  If the number
4142  * of wirings transitions to zero and the page is eligible for page out, then
4143  * the page is added to the specified paging queue.  If the released wiring
4144  * represented the last reference to the page, the page is freed.
4145  */
4146 void
4147 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4148 {
4149 
4150 	KASSERT(nqueue < PQ_COUNT,
4151 	    ("vm_page_unwire: invalid queue %u request for page %p",
4152 	    nqueue, m));
4153 
4154 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4155 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4156 			vm_page_free(m);
4157 		return;
4158 	}
4159 	vm_page_unwire_managed(m, nqueue, false);
4160 }
4161 
4162 /*
4163  * Unwire a page without (re-)inserting it into a page queue.  It is up
4164  * to the caller to enqueue, requeue, or free the page as appropriate.
4165  * In most cases involving managed pages, vm_page_unwire() should be used
4166  * instead.
4167  */
4168 bool
4169 vm_page_unwire_noq(vm_page_t m)
4170 {
4171 	u_int old;
4172 
4173 	old = vm_page_drop(m, 1);
4174 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4175 	    ("%s: counter underflow for page %p", __func__,  m));
4176 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4177 	    ("%s: missing ref on fictitious page %p", __func__, m));
4178 
4179 	if (VPRC_WIRE_COUNT(old) > 1)
4180 		return (false);
4181 	if ((m->oflags & VPO_UNMANAGED) == 0)
4182 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4183 	vm_wire_sub(1);
4184 	return (true);
4185 }
4186 
4187 /*
4188  * Ensure that the page ends up in the specified page queue.  If the page is
4189  * active or being moved to the active queue, ensure that its act_count is
4190  * at least ACT_INIT but do not otherwise mess with it.
4191  */
4192 static __always_inline void
4193 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4194 {
4195 	vm_page_astate_t old, new;
4196 
4197 	KASSERT(m->ref_count > 0,
4198 	    ("%s: page %p does not carry any references", __func__, m));
4199 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4200 	    ("%s: invalid flags %x", __func__, nflag));
4201 
4202 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4203 		return;
4204 
4205 	old = vm_page_astate_load(m);
4206 	do {
4207 		if ((old.flags & PGA_DEQUEUE) != 0)
4208 			break;
4209 		new = old;
4210 		new.flags &= ~PGA_QUEUE_OP_MASK;
4211 		if (nqueue == PQ_ACTIVE)
4212 			new.act_count = max(old.act_count, ACT_INIT);
4213 		if (old.queue == nqueue) {
4214 			if (nqueue != PQ_ACTIVE)
4215 				new.flags |= nflag;
4216 		} else {
4217 			new.flags |= nflag;
4218 			new.queue = nqueue;
4219 		}
4220 	} while (!vm_page_pqstate_commit(m, &old, new));
4221 }
4222 
4223 /*
4224  * Put the specified page on the active list (if appropriate).
4225  */
4226 void
4227 vm_page_activate(vm_page_t m)
4228 {
4229 
4230 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4231 }
4232 
4233 /*
4234  * Move the specified page to the tail of the inactive queue, or requeue
4235  * the page if it is already in the inactive queue.
4236  */
4237 void
4238 vm_page_deactivate(vm_page_t m)
4239 {
4240 
4241 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4242 }
4243 
4244 void
4245 vm_page_deactivate_noreuse(vm_page_t m)
4246 {
4247 
4248 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4249 }
4250 
4251 /*
4252  * Put a page in the laundry, or requeue it if it is already there.
4253  */
4254 void
4255 vm_page_launder(vm_page_t m)
4256 {
4257 
4258 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4259 }
4260 
4261 /*
4262  * Put a page in the PQ_UNSWAPPABLE holding queue.
4263  */
4264 void
4265 vm_page_unswappable(vm_page_t m)
4266 {
4267 
4268 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4269 	    ("page %p already unswappable", m));
4270 
4271 	vm_page_dequeue(m);
4272 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4273 }
4274 
4275 /*
4276  * Release a page back to the page queues in preparation for unwiring.
4277  */
4278 static void
4279 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4280 {
4281 	vm_page_astate_t old, new;
4282 	uint16_t nflag;
4283 
4284 	/*
4285 	 * Use a check of the valid bits to determine whether we should
4286 	 * accelerate reclamation of the page.  The object lock might not be
4287 	 * held here, in which case the check is racy.  At worst we will either
4288 	 * accelerate reclamation of a valid page and violate LRU, or
4289 	 * unnecessarily defer reclamation of an invalid page.
4290 	 *
4291 	 * If we were asked to not cache the page, place it near the head of the
4292 	 * inactive queue so that is reclaimed sooner.
4293 	 */
4294 	if (noreuse || m->valid == 0) {
4295 		nqueue = PQ_INACTIVE;
4296 		nflag = PGA_REQUEUE_HEAD;
4297 	} else {
4298 		nflag = PGA_REQUEUE;
4299 	}
4300 
4301 	old = vm_page_astate_load(m);
4302 	do {
4303 		new = old;
4304 
4305 		/*
4306 		 * If the page is already in the active queue and we are not
4307 		 * trying to accelerate reclamation, simply mark it as
4308 		 * referenced and avoid any queue operations.
4309 		 */
4310 		new.flags &= ~PGA_QUEUE_OP_MASK;
4311 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4312 			new.flags |= PGA_REFERENCED;
4313 		else {
4314 			new.flags |= nflag;
4315 			new.queue = nqueue;
4316 		}
4317 	} while (!vm_page_pqstate_commit(m, &old, new));
4318 }
4319 
4320 /*
4321  * Unwire a page and either attempt to free it or re-add it to the page queues.
4322  */
4323 void
4324 vm_page_release(vm_page_t m, int flags)
4325 {
4326 	vm_object_t object;
4327 
4328 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4329 	    ("vm_page_release: page %p is unmanaged", m));
4330 
4331 	if ((flags & VPR_TRYFREE) != 0) {
4332 		for (;;) {
4333 			object = atomic_load_ptr(&m->object);
4334 			if (object == NULL)
4335 				break;
4336 			/* Depends on type-stability. */
4337 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4338 				break;
4339 			if (object == m->object) {
4340 				vm_page_release_locked(m, flags);
4341 				VM_OBJECT_WUNLOCK(object);
4342 				return;
4343 			}
4344 			VM_OBJECT_WUNLOCK(object);
4345 		}
4346 	}
4347 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4348 }
4349 
4350 /* See vm_page_release(). */
4351 void
4352 vm_page_release_locked(vm_page_t m, int flags)
4353 {
4354 
4355 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4356 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4357 	    ("vm_page_release_locked: page %p is unmanaged", m));
4358 
4359 	if (vm_page_unwire_noq(m)) {
4360 		if ((flags & VPR_TRYFREE) != 0 &&
4361 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4362 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4363 			/*
4364 			 * An unlocked lookup may have wired the page before the
4365 			 * busy lock was acquired, in which case the page must
4366 			 * not be freed.
4367 			 */
4368 			if (__predict_true(!vm_page_wired(m))) {
4369 				vm_page_free(m);
4370 				return;
4371 			}
4372 			vm_page_xunbusy(m);
4373 		} else {
4374 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4375 		}
4376 	}
4377 }
4378 
4379 static bool
4380 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4381 {
4382 	u_int old;
4383 
4384 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4385 	    ("vm_page_try_blocked_op: page %p has no object", m));
4386 	KASSERT(vm_page_busied(m),
4387 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4388 	VM_OBJECT_ASSERT_LOCKED(m->object);
4389 
4390 	old = m->ref_count;
4391 	do {
4392 		KASSERT(old != 0,
4393 		    ("vm_page_try_blocked_op: page %p has no references", m));
4394 		if (VPRC_WIRE_COUNT(old) != 0)
4395 			return (false);
4396 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4397 
4398 	(op)(m);
4399 
4400 	/*
4401 	 * If the object is read-locked, new wirings may be created via an
4402 	 * object lookup.
4403 	 */
4404 	old = vm_page_drop(m, VPRC_BLOCKED);
4405 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4406 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4407 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4408 	    old, m));
4409 	return (true);
4410 }
4411 
4412 /*
4413  * Atomically check for wirings and remove all mappings of the page.
4414  */
4415 bool
4416 vm_page_try_remove_all(vm_page_t m)
4417 {
4418 
4419 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4420 }
4421 
4422 /*
4423  * Atomically check for wirings and remove all writeable mappings of the page.
4424  */
4425 bool
4426 vm_page_try_remove_write(vm_page_t m)
4427 {
4428 
4429 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4430 }
4431 
4432 /*
4433  * vm_page_advise
4434  *
4435  * 	Apply the specified advice to the given page.
4436  */
4437 void
4438 vm_page_advise(vm_page_t m, int advice)
4439 {
4440 
4441 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4442 	vm_page_assert_xbusied(m);
4443 
4444 	if (advice == MADV_FREE)
4445 		/*
4446 		 * Mark the page clean.  This will allow the page to be freed
4447 		 * without first paging it out.  MADV_FREE pages are often
4448 		 * quickly reused by malloc(3), so we do not do anything that
4449 		 * would result in a page fault on a later access.
4450 		 */
4451 		vm_page_undirty(m);
4452 	else if (advice != MADV_DONTNEED) {
4453 		if (advice == MADV_WILLNEED)
4454 			vm_page_activate(m);
4455 		return;
4456 	}
4457 
4458 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4459 		vm_page_dirty(m);
4460 
4461 	/*
4462 	 * Clear any references to the page.  Otherwise, the page daemon will
4463 	 * immediately reactivate the page.
4464 	 */
4465 	vm_page_aflag_clear(m, PGA_REFERENCED);
4466 
4467 	/*
4468 	 * Place clean pages near the head of the inactive queue rather than
4469 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4470 	 * the page will be reused quickly.  Dirty pages not already in the
4471 	 * laundry are moved there.
4472 	 */
4473 	if (m->dirty == 0)
4474 		vm_page_deactivate_noreuse(m);
4475 	else if (!vm_page_in_laundry(m))
4476 		vm_page_launder(m);
4477 }
4478 
4479 /*
4480  *	vm_page_grab_release
4481  *
4482  *	Helper routine for grab functions to release busy on return.
4483  */
4484 static inline void
4485 vm_page_grab_release(vm_page_t m, int allocflags)
4486 {
4487 
4488 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4489 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4490 			vm_page_sunbusy(m);
4491 		else
4492 			vm_page_xunbusy(m);
4493 	}
4494 }
4495 
4496 /*
4497  *	vm_page_grab_sleep
4498  *
4499  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4500  *	if the caller should retry and false otherwise.
4501  *
4502  *	If the object is locked on entry the object will be unlocked with
4503  *	false returns and still locked but possibly having been dropped
4504  *	with true returns.
4505  */
4506 static bool
4507 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4508     const char *wmesg, int allocflags, bool locked)
4509 {
4510 
4511 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4512 		return (false);
4513 
4514 	/*
4515 	 * Reference the page before unlocking and sleeping so that
4516 	 * the page daemon is less likely to reclaim it.
4517 	 */
4518 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4519 		vm_page_reference(m);
4520 
4521 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4522 	    locked)
4523 		VM_OBJECT_WLOCK(object);
4524 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4525 		return (false);
4526 
4527 	return (true);
4528 }
4529 
4530 /*
4531  * Assert that the grab flags are valid.
4532  */
4533 static inline void
4534 vm_page_grab_check(int allocflags)
4535 {
4536 
4537 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4538 	    (allocflags & VM_ALLOC_WIRED) != 0,
4539 	    ("vm_page_grab*: the pages must be busied or wired"));
4540 
4541 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4542 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4543 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4544 }
4545 
4546 /*
4547  * Calculate the page allocation flags for grab.
4548  */
4549 static inline int
4550 vm_page_grab_pflags(int allocflags)
4551 {
4552 	int pflags;
4553 
4554 	pflags = allocflags &
4555 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4556 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4557 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4558 		pflags |= VM_ALLOC_WAITFAIL;
4559 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4560 		pflags |= VM_ALLOC_SBUSY;
4561 
4562 	return (pflags);
4563 }
4564 
4565 /*
4566  * Grab a page, waiting until we are waken up due to the page
4567  * changing state.  We keep on waiting, if the page continues
4568  * to be in the object.  If the page doesn't exist, first allocate it
4569  * and then conditionally zero it.
4570  *
4571  * This routine may sleep.
4572  *
4573  * The object must be locked on entry.  The lock will, however, be released
4574  * and reacquired if the routine sleeps.
4575  */
4576 vm_page_t
4577 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4578 {
4579 	vm_page_t m;
4580 
4581 	VM_OBJECT_ASSERT_WLOCKED(object);
4582 	vm_page_grab_check(allocflags);
4583 
4584 retrylookup:
4585 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4586 		if (!vm_page_tryacquire(m, allocflags)) {
4587 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4588 			    allocflags, true))
4589 				goto retrylookup;
4590 			return (NULL);
4591 		}
4592 		goto out;
4593 	}
4594 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4595 		return (NULL);
4596 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4597 	if (m == NULL) {
4598 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4599 			return (NULL);
4600 		goto retrylookup;
4601 	}
4602 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4603 		pmap_zero_page(m);
4604 
4605 out:
4606 	vm_page_grab_release(m, allocflags);
4607 
4608 	return (m);
4609 }
4610 
4611 /*
4612  * Locklessly attempt to acquire a page given a (object, pindex) tuple
4613  * and an optional previous page to avoid the radix lookup.  The resulting
4614  * page will be validated against the identity tuple and busied or wired
4615  * as requested.  A NULL *mp return guarantees that the page was not in
4616  * radix at the time of the call but callers must perform higher level
4617  * synchronization or retry the operation under a lock if they require
4618  * an atomic answer.  This is the only lock free validation routine,
4619  * other routines can depend on the resulting page state.
4620  *
4621  * The return value indicates whether the operation failed due to caller
4622  * flags.  The return is tri-state with mp:
4623  *
4624  * (true, *mp != NULL) - The operation was successful.
4625  * (true, *mp == NULL) - The page was not found in tree.
4626  * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4627  */
4628 static bool
4629 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4630     vm_page_t prev, vm_page_t *mp, int allocflags)
4631 {
4632 	vm_page_t m;
4633 
4634 	vm_page_grab_check(allocflags);
4635 	MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4636 
4637 	*mp = NULL;
4638 	for (;;) {
4639 		/*
4640 		 * We may see a false NULL here because the previous page
4641 		 * has been removed or just inserted and the list is loaded
4642 		 * without barriers.  Switch to radix to verify.
4643 		 */
4644 		if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4645 		    QMD_IS_TRASHED(m) || m->pindex != pindex ||
4646 		    atomic_load_ptr(&m->object) != object) {
4647 			prev = NULL;
4648 			/*
4649 			 * This guarantees the result is instantaneously
4650 			 * correct.
4651 			 */
4652 			m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4653 		}
4654 		if (m == NULL)
4655 			return (true);
4656 		if (vm_page_trybusy(m, allocflags)) {
4657 			if (m->object == object && m->pindex == pindex)
4658 				break;
4659 			/* relookup. */
4660 			vm_page_busy_release(m);
4661 			cpu_spinwait();
4662 			continue;
4663 		}
4664 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4665 		    allocflags, false))
4666 			return (false);
4667 	}
4668 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4669 		vm_page_wire(m);
4670 	vm_page_grab_release(m, allocflags);
4671 	*mp = m;
4672 	return (true);
4673 }
4674 
4675 /*
4676  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4677  * is not set.
4678  */
4679 vm_page_t
4680 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4681 {
4682 	vm_page_t m;
4683 
4684 	vm_page_grab_check(allocflags);
4685 
4686 	if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4687 		return (NULL);
4688 	if (m != NULL)
4689 		return (m);
4690 
4691 	/*
4692 	 * The radix lockless lookup should never return a false negative
4693 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4694 	 * was no page present at the instant of the call.  A NOCREAT caller
4695 	 * must handle create races gracefully.
4696 	 */
4697 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4698 		return (NULL);
4699 
4700 	VM_OBJECT_WLOCK(object);
4701 	m = vm_page_grab(object, pindex, allocflags);
4702 	VM_OBJECT_WUNLOCK(object);
4703 
4704 	return (m);
4705 }
4706 
4707 /*
4708  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4709  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4710  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4711  * in simultaneously.  Additional pages will be left on a paging queue but
4712  * will neither be wired nor busy regardless of allocflags.
4713  */
4714 int
4715 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4716 {
4717 	vm_page_t m;
4718 	vm_page_t ma[VM_INITIAL_PAGEIN];
4719 	int after, i, pflags, rv;
4720 
4721 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4722 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4723 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4724 	KASSERT((allocflags &
4725 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4726 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4727 	VM_OBJECT_ASSERT_WLOCKED(object);
4728 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4729 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4730 	pflags |= VM_ALLOC_WAITFAIL;
4731 
4732 retrylookup:
4733 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4734 		/*
4735 		 * If the page is fully valid it can only become invalid
4736 		 * with the object lock held.  If it is not valid it can
4737 		 * become valid with the busy lock held.  Therefore, we
4738 		 * may unnecessarily lock the exclusive busy here if we
4739 		 * race with I/O completion not using the object lock.
4740 		 * However, we will not end up with an invalid page and a
4741 		 * shared lock.
4742 		 */
4743 		if (!vm_page_trybusy(m,
4744 		    vm_page_all_valid(m) ? allocflags : 0)) {
4745 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4746 			    allocflags, true);
4747 			goto retrylookup;
4748 		}
4749 		if (vm_page_all_valid(m))
4750 			goto out;
4751 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4752 			vm_page_busy_release(m);
4753 			*mp = NULL;
4754 			return (VM_PAGER_FAIL);
4755 		}
4756 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4757 		*mp = NULL;
4758 		return (VM_PAGER_FAIL);
4759 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4760 		goto retrylookup;
4761 	}
4762 
4763 	vm_page_assert_xbusied(m);
4764 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4765 		after = MIN(after, VM_INITIAL_PAGEIN);
4766 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4767 		after = MAX(after, 1);
4768 		ma[0] = m;
4769 		for (i = 1; i < after; i++) {
4770 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4771 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4772 					break;
4773 			} else {
4774 				ma[i] = vm_page_alloc(object, m->pindex + i,
4775 				    VM_ALLOC_NORMAL);
4776 				if (ma[i] == NULL)
4777 					break;
4778 			}
4779 		}
4780 		after = i;
4781 		vm_object_pip_add(object, after);
4782 		VM_OBJECT_WUNLOCK(object);
4783 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4784 		VM_OBJECT_WLOCK(object);
4785 		vm_object_pip_wakeupn(object, after);
4786 		/* Pager may have replaced a page. */
4787 		m = ma[0];
4788 		if (rv != VM_PAGER_OK) {
4789 			for (i = 0; i < after; i++) {
4790 				if (!vm_page_wired(ma[i]))
4791 					vm_page_free(ma[i]);
4792 				else
4793 					vm_page_xunbusy(ma[i]);
4794 			}
4795 			*mp = NULL;
4796 			return (rv);
4797 		}
4798 		for (i = 1; i < after; i++)
4799 			vm_page_readahead_finish(ma[i]);
4800 		MPASS(vm_page_all_valid(m));
4801 	} else {
4802 		vm_page_zero_invalid(m, TRUE);
4803 	}
4804 out:
4805 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4806 		vm_page_wire(m);
4807 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4808 		vm_page_busy_downgrade(m);
4809 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4810 		vm_page_busy_release(m);
4811 	*mp = m;
4812 	return (VM_PAGER_OK);
4813 }
4814 
4815 /*
4816  * Locklessly grab a valid page.  If the page is not valid or not yet
4817  * allocated this will fall back to the object lock method.
4818  */
4819 int
4820 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4821     vm_pindex_t pindex, int allocflags)
4822 {
4823 	vm_page_t m;
4824 	int flags;
4825 	int error;
4826 
4827 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4828 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4829 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4830 	    "mismatch"));
4831 	KASSERT((allocflags &
4832 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4833 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4834 
4835 	/*
4836 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
4837 	 * before we can inspect the valid field and return a wired page.
4838 	 */
4839 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4840 	if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4841 		return (VM_PAGER_FAIL);
4842 	if ((m = *mp) != NULL) {
4843 		if (vm_page_all_valid(m)) {
4844 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4845 				vm_page_wire(m);
4846 			vm_page_grab_release(m, allocflags);
4847 			return (VM_PAGER_OK);
4848 		}
4849 		vm_page_busy_release(m);
4850 	}
4851 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4852 		*mp = NULL;
4853 		return (VM_PAGER_FAIL);
4854 	}
4855 	VM_OBJECT_WLOCK(object);
4856 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
4857 	VM_OBJECT_WUNLOCK(object);
4858 
4859 	return (error);
4860 }
4861 
4862 /*
4863  * Return the specified range of pages from the given object.  For each
4864  * page offset within the range, if a page already exists within the object
4865  * at that offset and it is busy, then wait for it to change state.  If,
4866  * instead, the page doesn't exist, then allocate it.
4867  *
4868  * The caller must always specify an allocation class.
4869  *
4870  * allocation classes:
4871  *	VM_ALLOC_NORMAL		normal process request
4872  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4873  *
4874  * The caller must always specify that the pages are to be busied and/or
4875  * wired.
4876  *
4877  * optional allocation flags:
4878  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4879  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4880  *	VM_ALLOC_NOWAIT		do not sleep
4881  *	VM_ALLOC_SBUSY		set page to sbusy state
4882  *	VM_ALLOC_WIRED		wire the pages
4883  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4884  *
4885  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4886  * may return a partial prefix of the requested range.
4887  */
4888 int
4889 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4890     vm_page_t *ma, int count)
4891 {
4892 	vm_page_t m, mpred;
4893 	int pflags;
4894 	int i;
4895 
4896 	VM_OBJECT_ASSERT_WLOCKED(object);
4897 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4898 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4899 	KASSERT(count > 0,
4900 	    ("vm_page_grab_pages: invalid page count %d", count));
4901 	vm_page_grab_check(allocflags);
4902 
4903 	pflags = vm_page_grab_pflags(allocflags);
4904 	i = 0;
4905 retrylookup:
4906 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4907 	if (m == NULL || m->pindex != pindex + i) {
4908 		mpred = m;
4909 		m = NULL;
4910 	} else
4911 		mpred = TAILQ_PREV(m, pglist, listq);
4912 	for (; i < count; i++) {
4913 		if (m != NULL) {
4914 			if (!vm_page_tryacquire(m, allocflags)) {
4915 				if (vm_page_grab_sleep(object, m, pindex + i,
4916 				    "grbmaw", allocflags, true))
4917 					goto retrylookup;
4918 				break;
4919 			}
4920 		} else {
4921 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4922 				break;
4923 			m = vm_page_alloc_after(object, pindex + i,
4924 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4925 			if (m == NULL) {
4926 				if ((allocflags & (VM_ALLOC_NOWAIT |
4927 				    VM_ALLOC_WAITFAIL)) != 0)
4928 					break;
4929 				goto retrylookup;
4930 			}
4931 		}
4932 		if (vm_page_none_valid(m) &&
4933 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4934 			if ((m->flags & PG_ZERO) == 0)
4935 				pmap_zero_page(m);
4936 			vm_page_valid(m);
4937 		}
4938 		vm_page_grab_release(m, allocflags);
4939 		ma[i] = mpred = m;
4940 		m = vm_page_next(m);
4941 	}
4942 	return (i);
4943 }
4944 
4945 /*
4946  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
4947  * and will fall back to the locked variant to handle allocation.
4948  */
4949 int
4950 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4951     int allocflags, vm_page_t *ma, int count)
4952 {
4953 	vm_page_t m, pred;
4954 	int flags;
4955 	int i;
4956 
4957 	KASSERT(count > 0,
4958 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4959 	vm_page_grab_check(allocflags);
4960 
4961 	/*
4962 	 * Modify flags for lockless acquire to hold the page until we
4963 	 * set it valid if necessary.
4964 	 */
4965 	flags = allocflags & ~VM_ALLOC_NOBUSY;
4966 	pred = NULL;
4967 	for (i = 0; i < count; i++, pindex++) {
4968 		if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4969 			return (i);
4970 		if (m == NULL)
4971 			break;
4972 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4973 			if ((m->flags & PG_ZERO) == 0)
4974 				pmap_zero_page(m);
4975 			vm_page_valid(m);
4976 		}
4977 		/* m will still be wired or busy according to flags. */
4978 		vm_page_grab_release(m, allocflags);
4979 		pred = ma[i] = m;
4980 	}
4981 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
4982 		return (i);
4983 	count -= i;
4984 	VM_OBJECT_WLOCK(object);
4985 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4986 	VM_OBJECT_WUNLOCK(object);
4987 
4988 	return (i);
4989 }
4990 
4991 /*
4992  * Mapping function for valid or dirty bits in a page.
4993  *
4994  * Inputs are required to range within a page.
4995  */
4996 vm_page_bits_t
4997 vm_page_bits(int base, int size)
4998 {
4999 	int first_bit;
5000 	int last_bit;
5001 
5002 	KASSERT(
5003 	    base + size <= PAGE_SIZE,
5004 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5005 	);
5006 
5007 	if (size == 0)		/* handle degenerate case */
5008 		return (0);
5009 
5010 	first_bit = base >> DEV_BSHIFT;
5011 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5012 
5013 	return (((vm_page_bits_t)2 << last_bit) -
5014 	    ((vm_page_bits_t)1 << first_bit));
5015 }
5016 
5017 void
5018 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5019 {
5020 
5021 #if PAGE_SIZE == 32768
5022 	atomic_set_64((uint64_t *)bits, set);
5023 #elif PAGE_SIZE == 16384
5024 	atomic_set_32((uint32_t *)bits, set);
5025 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5026 	atomic_set_16((uint16_t *)bits, set);
5027 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5028 	atomic_set_8((uint8_t *)bits, set);
5029 #else		/* PAGE_SIZE <= 8192 */
5030 	uintptr_t addr;
5031 	int shift;
5032 
5033 	addr = (uintptr_t)bits;
5034 	/*
5035 	 * Use a trick to perform a 32-bit atomic on the
5036 	 * containing aligned word, to not depend on the existence
5037 	 * of atomic_{set, clear}_{8, 16}.
5038 	 */
5039 	shift = addr & (sizeof(uint32_t) - 1);
5040 #if BYTE_ORDER == BIG_ENDIAN
5041 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5042 #else
5043 	shift *= NBBY;
5044 #endif
5045 	addr &= ~(sizeof(uint32_t) - 1);
5046 	atomic_set_32((uint32_t *)addr, set << shift);
5047 #endif		/* PAGE_SIZE */
5048 }
5049 
5050 static inline void
5051 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5052 {
5053 
5054 #if PAGE_SIZE == 32768
5055 	atomic_clear_64((uint64_t *)bits, clear);
5056 #elif PAGE_SIZE == 16384
5057 	atomic_clear_32((uint32_t *)bits, clear);
5058 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5059 	atomic_clear_16((uint16_t *)bits, clear);
5060 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5061 	atomic_clear_8((uint8_t *)bits, clear);
5062 #else		/* PAGE_SIZE <= 8192 */
5063 	uintptr_t addr;
5064 	int shift;
5065 
5066 	addr = (uintptr_t)bits;
5067 	/*
5068 	 * Use a trick to perform a 32-bit atomic on the
5069 	 * containing aligned word, to not depend on the existence
5070 	 * of atomic_{set, clear}_{8, 16}.
5071 	 */
5072 	shift = addr & (sizeof(uint32_t) - 1);
5073 #if BYTE_ORDER == BIG_ENDIAN
5074 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5075 #else
5076 	shift *= NBBY;
5077 #endif
5078 	addr &= ~(sizeof(uint32_t) - 1);
5079 	atomic_clear_32((uint32_t *)addr, clear << shift);
5080 #endif		/* PAGE_SIZE */
5081 }
5082 
5083 static inline vm_page_bits_t
5084 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5085 {
5086 #if PAGE_SIZE == 32768
5087 	uint64_t old;
5088 
5089 	old = *bits;
5090 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5091 	return (old);
5092 #elif PAGE_SIZE == 16384
5093 	uint32_t old;
5094 
5095 	old = *bits;
5096 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5097 	return (old);
5098 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5099 	uint16_t old;
5100 
5101 	old = *bits;
5102 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5103 	return (old);
5104 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5105 	uint8_t old;
5106 
5107 	old = *bits;
5108 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5109 	return (old);
5110 #else		/* PAGE_SIZE <= 4096*/
5111 	uintptr_t addr;
5112 	uint32_t old, new, mask;
5113 	int shift;
5114 
5115 	addr = (uintptr_t)bits;
5116 	/*
5117 	 * Use a trick to perform a 32-bit atomic on the
5118 	 * containing aligned word, to not depend on the existence
5119 	 * of atomic_{set, swap, clear}_{8, 16}.
5120 	 */
5121 	shift = addr & (sizeof(uint32_t) - 1);
5122 #if BYTE_ORDER == BIG_ENDIAN
5123 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5124 #else
5125 	shift *= NBBY;
5126 #endif
5127 	addr &= ~(sizeof(uint32_t) - 1);
5128 	mask = VM_PAGE_BITS_ALL << shift;
5129 
5130 	old = *bits;
5131 	do {
5132 		new = old & ~mask;
5133 		new |= newbits << shift;
5134 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5135 	return (old >> shift);
5136 #endif		/* PAGE_SIZE */
5137 }
5138 
5139 /*
5140  *	vm_page_set_valid_range:
5141  *
5142  *	Sets portions of a page valid.  The arguments are expected
5143  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5144  *	of any partial chunks touched by the range.  The invalid portion of
5145  *	such chunks will be zeroed.
5146  *
5147  *	(base + size) must be less then or equal to PAGE_SIZE.
5148  */
5149 void
5150 vm_page_set_valid_range(vm_page_t m, int base, int size)
5151 {
5152 	int endoff, frag;
5153 	vm_page_bits_t pagebits;
5154 
5155 	vm_page_assert_busied(m);
5156 	if (size == 0)	/* handle degenerate case */
5157 		return;
5158 
5159 	/*
5160 	 * If the base is not DEV_BSIZE aligned and the valid
5161 	 * bit is clear, we have to zero out a portion of the
5162 	 * first block.
5163 	 */
5164 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5165 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5166 		pmap_zero_page_area(m, frag, base - frag);
5167 
5168 	/*
5169 	 * If the ending offset is not DEV_BSIZE aligned and the
5170 	 * valid bit is clear, we have to zero out a portion of
5171 	 * the last block.
5172 	 */
5173 	endoff = base + size;
5174 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5175 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5176 		pmap_zero_page_area(m, endoff,
5177 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5178 
5179 	/*
5180 	 * Assert that no previously invalid block that is now being validated
5181 	 * is already dirty.
5182 	 */
5183 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5184 	    ("vm_page_set_valid_range: page %p is dirty", m));
5185 
5186 	/*
5187 	 * Set valid bits inclusive of any overlap.
5188 	 */
5189 	pagebits = vm_page_bits(base, size);
5190 	if (vm_page_xbusied(m))
5191 		m->valid |= pagebits;
5192 	else
5193 		vm_page_bits_set(m, &m->valid, pagebits);
5194 }
5195 
5196 /*
5197  * Set the page dirty bits and free the invalid swap space if
5198  * present.  Returns the previous dirty bits.
5199  */
5200 vm_page_bits_t
5201 vm_page_set_dirty(vm_page_t m)
5202 {
5203 	vm_page_bits_t old;
5204 
5205 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5206 
5207 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5208 		old = m->dirty;
5209 		m->dirty = VM_PAGE_BITS_ALL;
5210 	} else
5211 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5212 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5213 		vm_pager_page_unswapped(m);
5214 
5215 	return (old);
5216 }
5217 
5218 /*
5219  * Clear the given bits from the specified page's dirty field.
5220  */
5221 static __inline void
5222 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5223 {
5224 
5225 	vm_page_assert_busied(m);
5226 
5227 	/*
5228 	 * If the page is xbusied and not write mapped we are the
5229 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5230 	 * layer can call vm_page_dirty() without holding a distinguished
5231 	 * lock.  The combination of page busy and atomic operations
5232 	 * suffice to guarantee consistency of the page dirty field.
5233 	 */
5234 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5235 		m->dirty &= ~pagebits;
5236 	else
5237 		vm_page_bits_clear(m, &m->dirty, pagebits);
5238 }
5239 
5240 /*
5241  *	vm_page_set_validclean:
5242  *
5243  *	Sets portions of a page valid and clean.  The arguments are expected
5244  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5245  *	of any partial chunks touched by the range.  The invalid portion of
5246  *	such chunks will be zero'd.
5247  *
5248  *	(base + size) must be less then or equal to PAGE_SIZE.
5249  */
5250 void
5251 vm_page_set_validclean(vm_page_t m, int base, int size)
5252 {
5253 	vm_page_bits_t oldvalid, pagebits;
5254 	int endoff, frag;
5255 
5256 	vm_page_assert_busied(m);
5257 	if (size == 0)	/* handle degenerate case */
5258 		return;
5259 
5260 	/*
5261 	 * If the base is not DEV_BSIZE aligned and the valid
5262 	 * bit is clear, we have to zero out a portion of the
5263 	 * first block.
5264 	 */
5265 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5266 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5267 		pmap_zero_page_area(m, frag, base - frag);
5268 
5269 	/*
5270 	 * If the ending offset is not DEV_BSIZE aligned and the
5271 	 * valid bit is clear, we have to zero out a portion of
5272 	 * the last block.
5273 	 */
5274 	endoff = base + size;
5275 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5276 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5277 		pmap_zero_page_area(m, endoff,
5278 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5279 
5280 	/*
5281 	 * Set valid, clear dirty bits.  If validating the entire
5282 	 * page we can safely clear the pmap modify bit.  We also
5283 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5284 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5285 	 * be set again.
5286 	 *
5287 	 * We set valid bits inclusive of any overlap, but we can only
5288 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5289 	 * the range.
5290 	 */
5291 	oldvalid = m->valid;
5292 	pagebits = vm_page_bits(base, size);
5293 	if (vm_page_xbusied(m))
5294 		m->valid |= pagebits;
5295 	else
5296 		vm_page_bits_set(m, &m->valid, pagebits);
5297 #if 0	/* NOT YET */
5298 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5299 		frag = DEV_BSIZE - frag;
5300 		base += frag;
5301 		size -= frag;
5302 		if (size < 0)
5303 			size = 0;
5304 	}
5305 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5306 #endif
5307 	if (base == 0 && size == PAGE_SIZE) {
5308 		/*
5309 		 * The page can only be modified within the pmap if it is
5310 		 * mapped, and it can only be mapped if it was previously
5311 		 * fully valid.
5312 		 */
5313 		if (oldvalid == VM_PAGE_BITS_ALL)
5314 			/*
5315 			 * Perform the pmap_clear_modify() first.  Otherwise,
5316 			 * a concurrent pmap operation, such as
5317 			 * pmap_protect(), could clear a modification in the
5318 			 * pmap and set the dirty field on the page before
5319 			 * pmap_clear_modify() had begun and after the dirty
5320 			 * field was cleared here.
5321 			 */
5322 			pmap_clear_modify(m);
5323 		m->dirty = 0;
5324 		vm_page_aflag_clear(m, PGA_NOSYNC);
5325 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5326 		m->dirty &= ~pagebits;
5327 	else
5328 		vm_page_clear_dirty_mask(m, pagebits);
5329 }
5330 
5331 void
5332 vm_page_clear_dirty(vm_page_t m, int base, int size)
5333 {
5334 
5335 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5336 }
5337 
5338 /*
5339  *	vm_page_set_invalid:
5340  *
5341  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5342  *	valid and dirty bits for the effected areas are cleared.
5343  */
5344 void
5345 vm_page_set_invalid(vm_page_t m, int base, int size)
5346 {
5347 	vm_page_bits_t bits;
5348 	vm_object_t object;
5349 
5350 	/*
5351 	 * The object lock is required so that pages can't be mapped
5352 	 * read-only while we're in the process of invalidating them.
5353 	 */
5354 	object = m->object;
5355 	VM_OBJECT_ASSERT_WLOCKED(object);
5356 	vm_page_assert_busied(m);
5357 
5358 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5359 	    size >= object->un_pager.vnp.vnp_size)
5360 		bits = VM_PAGE_BITS_ALL;
5361 	else
5362 		bits = vm_page_bits(base, size);
5363 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5364 		pmap_remove_all(m);
5365 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5366 	    !pmap_page_is_mapped(m),
5367 	    ("vm_page_set_invalid: page %p is mapped", m));
5368 	if (vm_page_xbusied(m)) {
5369 		m->valid &= ~bits;
5370 		m->dirty &= ~bits;
5371 	} else {
5372 		vm_page_bits_clear(m, &m->valid, bits);
5373 		vm_page_bits_clear(m, &m->dirty, bits);
5374 	}
5375 }
5376 
5377 /*
5378  *	vm_page_invalid:
5379  *
5380  *	Invalidates the entire page.  The page must be busy, unmapped, and
5381  *	the enclosing object must be locked.  The object locks protects
5382  *	against concurrent read-only pmap enter which is done without
5383  *	busy.
5384  */
5385 void
5386 vm_page_invalid(vm_page_t m)
5387 {
5388 
5389 	vm_page_assert_busied(m);
5390 	VM_OBJECT_ASSERT_LOCKED(m->object);
5391 	MPASS(!pmap_page_is_mapped(m));
5392 
5393 	if (vm_page_xbusied(m))
5394 		m->valid = 0;
5395 	else
5396 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5397 }
5398 
5399 /*
5400  * vm_page_zero_invalid()
5401  *
5402  *	The kernel assumes that the invalid portions of a page contain
5403  *	garbage, but such pages can be mapped into memory by user code.
5404  *	When this occurs, we must zero out the non-valid portions of the
5405  *	page so user code sees what it expects.
5406  *
5407  *	Pages are most often semi-valid when the end of a file is mapped
5408  *	into memory and the file's size is not page aligned.
5409  */
5410 void
5411 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5412 {
5413 	int b;
5414 	int i;
5415 
5416 	/*
5417 	 * Scan the valid bits looking for invalid sections that
5418 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5419 	 * valid bit may be set ) have already been zeroed by
5420 	 * vm_page_set_validclean().
5421 	 */
5422 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5423 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5424 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5425 			if (i > b) {
5426 				pmap_zero_page_area(m,
5427 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5428 			}
5429 			b = i + 1;
5430 		}
5431 	}
5432 
5433 	/*
5434 	 * setvalid is TRUE when we can safely set the zero'd areas
5435 	 * as being valid.  We can do this if there are no cache consistancy
5436 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5437 	 */
5438 	if (setvalid)
5439 		vm_page_valid(m);
5440 }
5441 
5442 /*
5443  *	vm_page_is_valid:
5444  *
5445  *	Is (partial) page valid?  Note that the case where size == 0
5446  *	will return FALSE in the degenerate case where the page is
5447  *	entirely invalid, and TRUE otherwise.
5448  *
5449  *	Some callers envoke this routine without the busy lock held and
5450  *	handle races via higher level locks.  Typical callers should
5451  *	hold a busy lock to prevent invalidation.
5452  */
5453 int
5454 vm_page_is_valid(vm_page_t m, int base, int size)
5455 {
5456 	vm_page_bits_t bits;
5457 
5458 	bits = vm_page_bits(base, size);
5459 	return (m->valid != 0 && (m->valid & bits) == bits);
5460 }
5461 
5462 /*
5463  * Returns true if all of the specified predicates are true for the entire
5464  * (super)page and false otherwise.
5465  */
5466 bool
5467 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5468 {
5469 	vm_object_t object;
5470 	int i, npages;
5471 
5472 	object = m->object;
5473 	if (skip_m != NULL && skip_m->object != object)
5474 		return (false);
5475 	VM_OBJECT_ASSERT_LOCKED(object);
5476 	npages = atop(pagesizes[m->psind]);
5477 
5478 	/*
5479 	 * The physically contiguous pages that make up a superpage, i.e., a
5480 	 * page with a page size index ("psind") greater than zero, will
5481 	 * occupy adjacent entries in vm_page_array[].
5482 	 */
5483 	for (i = 0; i < npages; i++) {
5484 		/* Always test object consistency, including "skip_m". */
5485 		if (m[i].object != object)
5486 			return (false);
5487 		if (&m[i] == skip_m)
5488 			continue;
5489 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5490 			return (false);
5491 		if ((flags & PS_ALL_DIRTY) != 0) {
5492 			/*
5493 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5494 			 * might stop this case from spuriously returning
5495 			 * "false".  However, that would require a write lock
5496 			 * on the object containing "m[i]".
5497 			 */
5498 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5499 				return (false);
5500 		}
5501 		if ((flags & PS_ALL_VALID) != 0 &&
5502 		    m[i].valid != VM_PAGE_BITS_ALL)
5503 			return (false);
5504 	}
5505 	return (true);
5506 }
5507 
5508 /*
5509  * Set the page's dirty bits if the page is modified.
5510  */
5511 void
5512 vm_page_test_dirty(vm_page_t m)
5513 {
5514 
5515 	vm_page_assert_busied(m);
5516 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5517 		vm_page_dirty(m);
5518 }
5519 
5520 void
5521 vm_page_valid(vm_page_t m)
5522 {
5523 
5524 	vm_page_assert_busied(m);
5525 	if (vm_page_xbusied(m))
5526 		m->valid = VM_PAGE_BITS_ALL;
5527 	else
5528 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5529 }
5530 
5531 void
5532 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5533 {
5534 
5535 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5536 }
5537 
5538 void
5539 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5540 {
5541 
5542 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5543 }
5544 
5545 int
5546 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5547 {
5548 
5549 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5550 }
5551 
5552 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5553 void
5554 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5555 {
5556 
5557 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5558 }
5559 
5560 void
5561 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5562 {
5563 
5564 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5565 }
5566 #endif
5567 
5568 #ifdef INVARIANTS
5569 void
5570 vm_page_object_busy_assert(vm_page_t m)
5571 {
5572 
5573 	/*
5574 	 * Certain of the page's fields may only be modified by the
5575 	 * holder of a page or object busy.
5576 	 */
5577 	if (m->object != NULL && !vm_page_busied(m))
5578 		VM_OBJECT_ASSERT_BUSY(m->object);
5579 }
5580 
5581 void
5582 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5583 {
5584 
5585 	if ((bits & PGA_WRITEABLE) == 0)
5586 		return;
5587 
5588 	/*
5589 	 * The PGA_WRITEABLE flag can only be set if the page is
5590 	 * managed, is exclusively busied or the object is locked.
5591 	 * Currently, this flag is only set by pmap_enter().
5592 	 */
5593 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5594 	    ("PGA_WRITEABLE on unmanaged page"));
5595 	if (!vm_page_xbusied(m))
5596 		VM_OBJECT_ASSERT_BUSY(m->object);
5597 }
5598 #endif
5599 
5600 #include "opt_ddb.h"
5601 #ifdef DDB
5602 #include <sys/kernel.h>
5603 
5604 #include <ddb/ddb.h>
5605 
5606 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5607 {
5608 
5609 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5610 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5611 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5612 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5613 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5614 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5615 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5616 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5617 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5618 }
5619 
5620 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5621 {
5622 	int dom;
5623 
5624 	db_printf("pq_free %d\n", vm_free_count());
5625 	for (dom = 0; dom < vm_ndomains; dom++) {
5626 		db_printf(
5627     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5628 		    dom,
5629 		    vm_dom[dom].vmd_page_count,
5630 		    vm_dom[dom].vmd_free_count,
5631 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5632 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5633 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5634 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5635 	}
5636 }
5637 
5638 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5639 {
5640 	vm_page_t m;
5641 	boolean_t phys, virt;
5642 
5643 	if (!have_addr) {
5644 		db_printf("show pginfo addr\n");
5645 		return;
5646 	}
5647 
5648 	phys = strchr(modif, 'p') != NULL;
5649 	virt = strchr(modif, 'v') != NULL;
5650 	if (virt)
5651 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5652 	else if (phys)
5653 		m = PHYS_TO_VM_PAGE(addr);
5654 	else
5655 		m = (vm_page_t)addr;
5656 	db_printf(
5657     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5658     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5659 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5660 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5661 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5662 }
5663 #endif /* DDB */
5664