xref: /freebsd/sys/vm/vm_page.c (revision fa4b5aa51f1b3971e92f831b8e67e83aaed3c212)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Resident memory management module.
65  */
66 
67 #include <sys/cdefs.h>
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/counter.h>
73 #include <sys/domainset.h>
74 #include <sys/kernel.h>
75 #include <sys/limits.h>
76 #include <sys/linker.h>
77 #include <sys/lock.h>
78 #include <sys/malloc.h>
79 #include <sys/mman.h>
80 #include <sys/msgbuf.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/rwlock.h>
84 #include <sys/sleepqueue.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/smp.h>
88 #include <sys/sysctl.h>
89 #include <sys/vmmeter.h>
90 #include <sys/vnode.h>
91 
92 #include <vm/vm.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_param.h>
95 #include <vm/vm_domainset.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_dumpset.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 struct vm_domain vm_dom[MAXMEMDOM];
114 
115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
116 
117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
118 
119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
120 /* The following fields are protected by the domainset lock. */
121 domainset_t __exclusive_cache_line vm_min_domains;
122 domainset_t __exclusive_cache_line vm_severe_domains;
123 static int vm_min_waiters;
124 static int vm_severe_waiters;
125 static int vm_pageproc_waiters;
126 
127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
128     "VM page statistics");
129 
130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
132     CTLFLAG_RD, &pqstate_commit_retries,
133     "Number of failed per-page atomic queue state updates");
134 
135 static COUNTER_U64_DEFINE_EARLY(queue_ops);
136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
137     CTLFLAG_RD, &queue_ops,
138     "Number of batched queue operations");
139 
140 static COUNTER_U64_DEFINE_EARLY(queue_nops);
141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
142     CTLFLAG_RD, &queue_nops,
143     "Number of batched queue operations with no effects");
144 
145 /*
146  * bogus page -- for I/O to/from partially complete buffers,
147  * or for paging into sparsely invalid regions.
148  */
149 vm_page_t bogus_page;
150 
151 vm_page_t vm_page_array;
152 long vm_page_array_size;
153 long first_page;
154 
155 struct bitset *vm_page_dump;
156 long vm_page_dump_pages;
157 
158 static TAILQ_HEAD(, vm_page) blacklist_head;
159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
161     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
162 
163 static uma_zone_t fakepg_zone;
164 
165 static void vm_page_alloc_check(vm_page_t m);
166 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req);
167 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
168     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
169 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
170 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
171 static bool vm_page_free_prep(vm_page_t m);
172 static void vm_page_free_toq(vm_page_t m);
173 static void vm_page_init(void *dummy);
174 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
175     vm_pindex_t pindex, vm_page_t mpred);
176 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
177     vm_page_t mpred);
178 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
179     const uint16_t nflag);
180 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
181     vm_page_t m_run, vm_paddr_t high);
182 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
183 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
184     int req);
185 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
186     int flags);
187 static void vm_page_zone_release(void *arg, void **store, int cnt);
188 
189 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
190 
191 static void
192 vm_page_init(void *dummy)
193 {
194 
195 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
196 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
197 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED);
198 }
199 
200 static int pgcache_zone_max_pcpu;
201 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
202     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
203     "Per-CPU page cache size");
204 
205 /*
206  * The cache page zone is initialized later since we need to be able to allocate
207  * pages before UMA is fully initialized.
208  */
209 static void
210 vm_page_init_cache_zones(void *dummy __unused)
211 {
212 	struct vm_domain *vmd;
213 	struct vm_pgcache *pgcache;
214 	int cache, domain, maxcache, pool;
215 
216 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
217 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
218 	for (domain = 0; domain < vm_ndomains; domain++) {
219 		vmd = VM_DOMAIN(domain);
220 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
221 			pgcache = &vmd->vmd_pgcache[pool];
222 			pgcache->domain = domain;
223 			pgcache->pool = pool;
224 			pgcache->zone = uma_zcache_create("vm pgcache",
225 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
226 			    vm_page_zone_import, vm_page_zone_release, pgcache,
227 			    UMA_ZONE_VM);
228 
229 			/*
230 			 * Limit each pool's zone to 0.1% of the pages in the
231 			 * domain.
232 			 */
233 			cache = maxcache != 0 ? maxcache :
234 			    vmd->vmd_page_count / 1000;
235 			uma_zone_set_maxcache(pgcache->zone, cache);
236 		}
237 	}
238 }
239 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
240 
241 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
242 #if PAGE_SIZE == 32768
243 #ifdef CTASSERT
244 CTASSERT(sizeof(u_long) >= 8);
245 #endif
246 #endif
247 
248 /*
249  *	vm_set_page_size:
250  *
251  *	Sets the page size, perhaps based upon the memory
252  *	size.  Must be called before any use of page-size
253  *	dependent functions.
254  */
255 void
256 vm_set_page_size(void)
257 {
258 	if (vm_cnt.v_page_size == 0)
259 		vm_cnt.v_page_size = PAGE_SIZE;
260 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
261 		panic("vm_set_page_size: page size not a power of two");
262 }
263 
264 /*
265  *	vm_page_blacklist_next:
266  *
267  *	Find the next entry in the provided string of blacklist
268  *	addresses.  Entries are separated by space, comma, or newline.
269  *	If an invalid integer is encountered then the rest of the
270  *	string is skipped.  Updates the list pointer to the next
271  *	character, or NULL if the string is exhausted or invalid.
272  */
273 static vm_paddr_t
274 vm_page_blacklist_next(char **list, char *end)
275 {
276 	vm_paddr_t bad;
277 	char *cp, *pos;
278 
279 	if (list == NULL || *list == NULL)
280 		return (0);
281 	if (**list =='\0') {
282 		*list = NULL;
283 		return (0);
284 	}
285 
286 	/*
287 	 * If there's no end pointer then the buffer is coming from
288 	 * the kenv and we know it's null-terminated.
289 	 */
290 	if (end == NULL)
291 		end = *list + strlen(*list);
292 
293 	/* Ensure that strtoq() won't walk off the end */
294 	if (*end != '\0') {
295 		if (*end == '\n' || *end == ' ' || *end  == ',')
296 			*end = '\0';
297 		else {
298 			printf("Blacklist not terminated, skipping\n");
299 			*list = NULL;
300 			return (0);
301 		}
302 	}
303 
304 	for (pos = *list; *pos != '\0'; pos = cp) {
305 		bad = strtoq(pos, &cp, 0);
306 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
307 			if (bad == 0) {
308 				if (++cp < end)
309 					continue;
310 				else
311 					break;
312 			}
313 		} else
314 			break;
315 		if (*cp == '\0' || ++cp >= end)
316 			*list = NULL;
317 		else
318 			*list = cp;
319 		return (trunc_page(bad));
320 	}
321 	printf("Garbage in RAM blacklist, skipping\n");
322 	*list = NULL;
323 	return (0);
324 }
325 
326 bool
327 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
328 {
329 	struct vm_domain *vmd;
330 	vm_page_t m;
331 	bool found;
332 
333 	m = vm_phys_paddr_to_vm_page(pa);
334 	if (m == NULL)
335 		return (true); /* page does not exist, no failure */
336 
337 	vmd = VM_DOMAIN(vm_phys_domain(pa));
338 	vm_domain_free_lock(vmd);
339 	found = vm_phys_unfree_page(pa);
340 	vm_domain_free_unlock(vmd);
341 	if (found) {
342 		vm_domain_freecnt_inc(vmd, -1);
343 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
344 		if (verbose)
345 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
346 	}
347 	return (found);
348 }
349 
350 /*
351  *	vm_page_blacklist_check:
352  *
353  *	Iterate through the provided string of blacklist addresses, pulling
354  *	each entry out of the physical allocator free list and putting it
355  *	onto a list for reporting via the vm.page_blacklist sysctl.
356  */
357 static void
358 vm_page_blacklist_check(char *list, char *end)
359 {
360 	vm_paddr_t pa;
361 	char *next;
362 
363 	next = list;
364 	while (next != NULL) {
365 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
366 			continue;
367 		vm_page_blacklist_add(pa, bootverbose);
368 	}
369 }
370 
371 /*
372  *	vm_page_blacklist_load:
373  *
374  *	Search for a special module named "ram_blacklist".  It'll be a
375  *	plain text file provided by the user via the loader directive
376  *	of the same name.
377  */
378 static void
379 vm_page_blacklist_load(char **list, char **end)
380 {
381 	void *mod;
382 	u_char *ptr;
383 	u_int len;
384 
385 	mod = NULL;
386 	ptr = NULL;
387 
388 	mod = preload_search_by_type("ram_blacklist");
389 	if (mod != NULL) {
390 		ptr = preload_fetch_addr(mod);
391 		len = preload_fetch_size(mod);
392         }
393 	*list = ptr;
394 	if (ptr != NULL)
395 		*end = ptr + len;
396 	else
397 		*end = NULL;
398 	return;
399 }
400 
401 static int
402 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
403 {
404 	vm_page_t m;
405 	struct sbuf sbuf;
406 	int error, first;
407 
408 	first = 1;
409 	error = sysctl_wire_old_buffer(req, 0);
410 	if (error != 0)
411 		return (error);
412 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
413 	TAILQ_FOREACH(m, &blacklist_head, listq) {
414 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
415 		    (uintmax_t)m->phys_addr);
416 		first = 0;
417 	}
418 	error = sbuf_finish(&sbuf);
419 	sbuf_delete(&sbuf);
420 	return (error);
421 }
422 
423 /*
424  * Initialize a dummy page for use in scans of the specified paging queue.
425  * In principle, this function only needs to set the flag PG_MARKER.
426  * Nonetheless, it write busies the page as a safety precaution.
427  */
428 void
429 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
430 {
431 
432 	bzero(marker, sizeof(*marker));
433 	marker->flags = PG_MARKER;
434 	marker->a.flags = aflags;
435 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
436 	marker->a.queue = queue;
437 }
438 
439 static void
440 vm_page_domain_init(int domain)
441 {
442 	struct vm_domain *vmd;
443 	struct vm_pagequeue *pq;
444 	int i;
445 
446 	vmd = VM_DOMAIN(domain);
447 	bzero(vmd, sizeof(*vmd));
448 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
449 	    "vm inactive pagequeue";
450 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
451 	    "vm active pagequeue";
452 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
453 	    "vm laundry pagequeue";
454 	*__DECONST(const char **,
455 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
456 	    "vm unswappable pagequeue";
457 	vmd->vmd_domain = domain;
458 	vmd->vmd_page_count = 0;
459 	vmd->vmd_free_count = 0;
460 	vmd->vmd_segs = 0;
461 	vmd->vmd_oom = FALSE;
462 	for (i = 0; i < PQ_COUNT; i++) {
463 		pq = &vmd->vmd_pagequeues[i];
464 		TAILQ_INIT(&pq->pq_pl);
465 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
466 		    MTX_DEF | MTX_DUPOK);
467 		pq->pq_pdpages = 0;
468 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
469 	}
470 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
471 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
472 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
473 
474 	/*
475 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
476 	 * insertions.
477 	 */
478 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
479 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
480 	    &vmd->vmd_inacthead, plinks.q);
481 
482 	/*
483 	 * The clock pages are used to implement active queue scanning without
484 	 * requeues.  Scans start at clock[0], which is advanced after the scan
485 	 * ends.  When the two clock hands meet, they are reset and scanning
486 	 * resumes from the head of the queue.
487 	 */
488 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
489 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
490 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
491 	    &vmd->vmd_clock[0], plinks.q);
492 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
493 	    &vmd->vmd_clock[1], plinks.q);
494 }
495 
496 /*
497  * Initialize a physical page in preparation for adding it to the free
498  * lists.
499  */
500 void
501 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool)
502 {
503 	m->object = NULL;
504 	m->ref_count = 0;
505 	m->busy_lock = VPB_FREED;
506 	m->flags = m->a.flags = 0;
507 	m->phys_addr = pa;
508 	m->a.queue = PQ_NONE;
509 	m->psind = 0;
510 	m->segind = segind;
511 	m->order = VM_NFREEORDER;
512 	m->pool = pool;
513 	m->valid = m->dirty = 0;
514 	pmap_page_init(m);
515 }
516 
517 #ifndef PMAP_HAS_PAGE_ARRAY
518 static vm_paddr_t
519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
520 {
521 	vm_paddr_t new_end;
522 
523 	/*
524 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
525 	 * However, because this page is allocated from KVM, out-of-bounds
526 	 * accesses using the direct map will not be trapped.
527 	 */
528 	*vaddr += PAGE_SIZE;
529 
530 	/*
531 	 * Allocate physical memory for the page structures, and map it.
532 	 */
533 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
534 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
535 	    VM_PROT_READ | VM_PROT_WRITE);
536 	vm_page_array_size = page_range;
537 
538 	return (new_end);
539 }
540 #endif
541 
542 /*
543  *	vm_page_startup:
544  *
545  *	Initializes the resident memory module.  Allocates physical memory for
546  *	bootstrapping UMA and some data structures that are used to manage
547  *	physical pages.  Initializes these structures, and populates the free
548  *	page queues.
549  */
550 vm_offset_t
551 vm_page_startup(vm_offset_t vaddr)
552 {
553 	struct vm_phys_seg *seg;
554 	struct vm_domain *vmd;
555 	vm_page_t m;
556 	char *list, *listend;
557 	vm_paddr_t end, high_avail, low_avail, new_end, size;
558 	vm_paddr_t page_range __unused;
559 	vm_paddr_t last_pa, pa, startp, endp;
560 	u_long pagecount;
561 #if MINIDUMP_PAGE_TRACKING
562 	u_long vm_page_dump_size;
563 #endif
564 	int biggestone, i, segind;
565 #ifdef WITNESS
566 	vm_offset_t mapped;
567 	int witness_size;
568 #endif
569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
570 	long ii;
571 #endif
572 #ifdef VM_FREEPOOL_LAZYINIT
573 	int lazyinit;
574 #endif
575 
576 	vaddr = round_page(vaddr);
577 
578 	vm_phys_early_startup();
579 	biggestone = vm_phys_avail_largest();
580 	end = phys_avail[biggestone+1];
581 
582 	/*
583 	 * Initialize the page and queue locks.
584 	 */
585 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
586 	for (i = 0; i < PA_LOCK_COUNT; i++)
587 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
588 	for (i = 0; i < vm_ndomains; i++)
589 		vm_page_domain_init(i);
590 
591 	new_end = end;
592 #ifdef WITNESS
593 	witness_size = round_page(witness_startup_count());
594 	new_end -= witness_size;
595 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
596 	    VM_PROT_READ | VM_PROT_WRITE);
597 	bzero((void *)mapped, witness_size);
598 	witness_startup((void *)mapped);
599 #endif
600 
601 #if MINIDUMP_PAGE_TRACKING
602 	/*
603 	 * Allocate a bitmap to indicate that a random physical page
604 	 * needs to be included in a minidump.
605 	 *
606 	 * The amd64 port needs this to indicate which direct map pages
607 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
608 	 *
609 	 * However, i386 still needs this workspace internally within the
610 	 * minidump code.  In theory, they are not needed on i386, but are
611 	 * included should the sf_buf code decide to use them.
612 	 */
613 	last_pa = 0;
614 	vm_page_dump_pages = 0;
615 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
616 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
617 		    dump_avail[i] / PAGE_SIZE;
618 		if (dump_avail[i + 1] > last_pa)
619 			last_pa = dump_avail[i + 1];
620 	}
621 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
622 	new_end -= vm_page_dump_size;
623 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
624 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
625 	bzero((void *)vm_page_dump, vm_page_dump_size);
626 #if MINIDUMP_STARTUP_PAGE_TRACKING
627 	/*
628 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
629 	 * in a crash dump.  When pmap_map() uses the direct map, they are
630 	 * not automatically included.
631 	 */
632 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
633 		dump_add_page(pa);
634 #endif
635 #else
636 	(void)last_pa;
637 #endif
638 	phys_avail[biggestone + 1] = new_end;
639 #ifdef __amd64__
640 	/*
641 	 * Request that the physical pages underlying the message buffer be
642 	 * included in a crash dump.  Since the message buffer is accessed
643 	 * through the direct map, they are not automatically included.
644 	 */
645 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
646 	last_pa = pa + round_page(msgbufsize);
647 	while (pa < last_pa) {
648 		dump_add_page(pa);
649 		pa += PAGE_SIZE;
650 	}
651 #endif
652 	/*
653 	 * Compute the number of pages of memory that will be available for
654 	 * use, taking into account the overhead of a page structure per page.
655 	 * In other words, solve
656 	 *	"available physical memory" - round_page(page_range *
657 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
658 	 * for page_range.
659 	 */
660 	low_avail = phys_avail[0];
661 	high_avail = phys_avail[1];
662 	for (i = 0; i < vm_phys_nsegs; i++) {
663 		if (vm_phys_segs[i].start < low_avail)
664 			low_avail = vm_phys_segs[i].start;
665 		if (vm_phys_segs[i].end > high_avail)
666 			high_avail = vm_phys_segs[i].end;
667 	}
668 	/* Skip the first chunk.  It is already accounted for. */
669 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
670 		if (phys_avail[i] < low_avail)
671 			low_avail = phys_avail[i];
672 		if (phys_avail[i + 1] > high_avail)
673 			high_avail = phys_avail[i + 1];
674 	}
675 	first_page = low_avail / PAGE_SIZE;
676 #ifdef VM_PHYSSEG_SPARSE
677 	size = 0;
678 	for (i = 0; i < vm_phys_nsegs; i++)
679 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
680 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
681 		size += phys_avail[i + 1] - phys_avail[i];
682 #elif defined(VM_PHYSSEG_DENSE)
683 	size = high_avail - low_avail;
684 #else
685 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
686 #endif
687 
688 #ifdef PMAP_HAS_PAGE_ARRAY
689 	pmap_page_array_startup(size / PAGE_SIZE);
690 	biggestone = vm_phys_avail_largest();
691 	end = new_end = phys_avail[biggestone + 1];
692 #else
693 #ifdef VM_PHYSSEG_DENSE
694 	/*
695 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
696 	 * the overhead of a page structure per page only if vm_page_array is
697 	 * allocated from the last physical memory chunk.  Otherwise, we must
698 	 * allocate page structures representing the physical memory
699 	 * underlying vm_page_array, even though they will not be used.
700 	 */
701 	if (new_end != high_avail)
702 		page_range = size / PAGE_SIZE;
703 	else
704 #endif
705 	{
706 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
707 
708 		/*
709 		 * If the partial bytes remaining are large enough for
710 		 * a page (PAGE_SIZE) without a corresponding
711 		 * 'struct vm_page', then new_end will contain an
712 		 * extra page after subtracting the length of the VM
713 		 * page array.  Compensate by subtracting an extra
714 		 * page from new_end.
715 		 */
716 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
717 			if (new_end == high_avail)
718 				high_avail -= PAGE_SIZE;
719 			new_end -= PAGE_SIZE;
720 		}
721 	}
722 	end = new_end;
723 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
724 #endif
725 
726 #if VM_NRESERVLEVEL > 0
727 	/*
728 	 * Allocate physical memory for the reservation management system's
729 	 * data structures, and map it.
730 	 */
731 	new_end = vm_reserv_startup(&vaddr, new_end);
732 #endif
733 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
734 	/*
735 	 * Include vm_page_array and vm_reserv_array in a crash dump.
736 	 */
737 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
738 		dump_add_page(pa);
739 #endif
740 	phys_avail[biggestone + 1] = new_end;
741 
742 	/*
743 	 * Add physical memory segments corresponding to the available
744 	 * physical pages.
745 	 */
746 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
747 		if (vm_phys_avail_size(i) != 0)
748 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
749 
750 	/*
751 	 * Initialize the physical memory allocator.
752 	 */
753 	vm_phys_init();
754 
755 #ifdef VM_FREEPOOL_LAZYINIT
756 	lazyinit = 1;
757 	TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit);
758 #endif
759 
760 	/*
761 	 * Initialize the page structures and add every available page to the
762 	 * physical memory allocator's free lists.
763 	 */
764 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
765 	for (ii = 0; ii < vm_page_array_size; ii++) {
766 		m = &vm_page_array[ii];
767 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0,
768 		    VM_FREEPOOL_DEFAULT);
769 		m->flags = PG_FICTITIOUS;
770 	}
771 #endif
772 	vm_cnt.v_page_count = 0;
773 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
774 		seg = &vm_phys_segs[segind];
775 
776 		/*
777 		 * If lazy vm_page initialization is not enabled, simply
778 		 * initialize all of the pages in the segment.  Otherwise, we
779 		 * only initialize:
780 		 * 1. Pages not covered by phys_avail[], since they might be
781 		 *    freed to the allocator at some future point, e.g., by
782 		 *    kmem_bootstrap_free().
783 		 * 2. The first page of each run of free pages handed to the
784 		 *    vm_phys allocator, which in turn defers initialization
785 		 *    of pages until they are needed.
786 		 * This avoids blocking the boot process for long periods, which
787 		 * may be relevant for VMs (which ought to boot as quickly as
788 		 * possible) and/or systems with large amounts of physical
789 		 * memory.
790 		 */
791 #ifdef VM_FREEPOOL_LAZYINIT
792 		if (lazyinit) {
793 			startp = seg->start;
794 			for (i = 0; phys_avail[i + 1] != 0; i += 2) {
795 				if (startp >= seg->end)
796 					break;
797 
798 				if (phys_avail[i + 1] < startp)
799 					continue;
800 				if (phys_avail[i] <= startp) {
801 					startp = phys_avail[i + 1];
802 					continue;
803 				}
804 
805 				m = vm_phys_seg_paddr_to_vm_page(seg, startp);
806 				for (endp = MIN(phys_avail[i], seg->end);
807 				    startp < endp; startp += PAGE_SIZE, m++) {
808 					vm_page_init_page(m, startp, segind,
809 					    VM_FREEPOOL_DEFAULT);
810 				}
811 			}
812 		} else
813 #endif
814 			for (m = seg->first_page, pa = seg->start;
815 			    pa < seg->end; m++, pa += PAGE_SIZE) {
816 				vm_page_init_page(m, pa, segind,
817 				    VM_FREEPOOL_DEFAULT);
818 			}
819 
820 		/*
821 		 * Add the segment's pages that are covered by one of
822 		 * phys_avail's ranges to the free lists.
823 		 */
824 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
825 			if (seg->end <= phys_avail[i] ||
826 			    seg->start >= phys_avail[i + 1])
827 				continue;
828 
829 			startp = MAX(seg->start, phys_avail[i]);
830 			endp = MIN(seg->end, phys_avail[i + 1]);
831 			pagecount = (u_long)atop(endp - startp);
832 			if (pagecount == 0)
833 				continue;
834 
835 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
836 #ifdef VM_FREEPOOL_LAZYINIT
837 			if (lazyinit) {
838 				vm_page_init_page(m, startp, segind,
839 				    VM_FREEPOOL_LAZYINIT);
840 			}
841 #endif
842 			vmd = VM_DOMAIN(seg->domain);
843 			vm_domain_free_lock(vmd);
844 			vm_phys_enqueue_contig(m, pagecount);
845 			vm_domain_free_unlock(vmd);
846 			vm_domain_freecnt_inc(vmd, pagecount);
847 			vm_cnt.v_page_count += (u_int)pagecount;
848 			vmd->vmd_page_count += (u_int)pagecount;
849 			vmd->vmd_segs |= 1UL << segind;
850 		}
851 	}
852 
853 	/*
854 	 * Remove blacklisted pages from the physical memory allocator.
855 	 */
856 	TAILQ_INIT(&blacklist_head);
857 	vm_page_blacklist_load(&list, &listend);
858 	vm_page_blacklist_check(list, listend);
859 
860 	list = kern_getenv("vm.blacklist");
861 	vm_page_blacklist_check(list, NULL);
862 
863 	freeenv(list);
864 #if VM_NRESERVLEVEL > 0
865 	/*
866 	 * Initialize the reservation management system.
867 	 */
868 	vm_reserv_init();
869 #endif
870 
871 	return (vaddr);
872 }
873 
874 void
875 vm_page_reference(vm_page_t m)
876 {
877 
878 	vm_page_aflag_set(m, PGA_REFERENCED);
879 }
880 
881 /*
882  *	vm_page_trybusy
883  *
884  *	Helper routine for grab functions to trylock busy.
885  *
886  *	Returns true on success and false on failure.
887  */
888 static bool
889 vm_page_trybusy(vm_page_t m, int allocflags)
890 {
891 
892 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
893 		return (vm_page_trysbusy(m));
894 	else
895 		return (vm_page_tryxbusy(m));
896 }
897 
898 /*
899  *	vm_page_tryacquire
900  *
901  *	Helper routine for grab functions to trylock busy and wire.
902  *
903  *	Returns true on success and false on failure.
904  */
905 static inline bool
906 vm_page_tryacquire(vm_page_t m, int allocflags)
907 {
908 	bool locked;
909 
910 	locked = vm_page_trybusy(m, allocflags);
911 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
912 		vm_page_wire(m);
913 	return (locked);
914 }
915 
916 /*
917  *	vm_page_busy_acquire:
918  *
919  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
920  *	and drop the object lock if necessary.
921  */
922 bool
923 vm_page_busy_acquire(vm_page_t m, int allocflags)
924 {
925 	vm_object_t obj;
926 	bool locked;
927 
928 	/*
929 	 * The page-specific object must be cached because page
930 	 * identity can change during the sleep, causing the
931 	 * re-lock of a different object.
932 	 * It is assumed that a reference to the object is already
933 	 * held by the callers.
934 	 */
935 	obj = atomic_load_ptr(&m->object);
936 	for (;;) {
937 		if (vm_page_tryacquire(m, allocflags))
938 			return (true);
939 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
940 			return (false);
941 		if (obj != NULL)
942 			locked = VM_OBJECT_WOWNED(obj);
943 		else
944 			locked = false;
945 		MPASS(locked || vm_page_wired(m));
946 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
947 		    locked) && locked)
948 			VM_OBJECT_WLOCK(obj);
949 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
950 			return (false);
951 		KASSERT(m->object == obj || m->object == NULL,
952 		    ("vm_page_busy_acquire: page %p does not belong to %p",
953 		    m, obj));
954 	}
955 }
956 
957 /*
958  *	vm_page_busy_downgrade:
959  *
960  *	Downgrade an exclusive busy page into a single shared busy page.
961  */
962 void
963 vm_page_busy_downgrade(vm_page_t m)
964 {
965 	u_int x;
966 
967 	vm_page_assert_xbusied(m);
968 
969 	x = vm_page_busy_fetch(m);
970 	for (;;) {
971 		if (atomic_fcmpset_rel_int(&m->busy_lock,
972 		    &x, VPB_SHARERS_WORD(1)))
973 			break;
974 	}
975 	if ((x & VPB_BIT_WAITERS) != 0)
976 		wakeup(m);
977 }
978 
979 /*
980  *
981  *	vm_page_busy_tryupgrade:
982  *
983  *	Attempt to upgrade a single shared busy into an exclusive busy.
984  */
985 int
986 vm_page_busy_tryupgrade(vm_page_t m)
987 {
988 	u_int ce, x;
989 
990 	vm_page_assert_sbusied(m);
991 
992 	x = vm_page_busy_fetch(m);
993 	ce = VPB_CURTHREAD_EXCLUSIVE;
994 	for (;;) {
995 		if (VPB_SHARERS(x) > 1)
996 			return (0);
997 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
998 		    ("vm_page_busy_tryupgrade: invalid lock state"));
999 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1000 		    ce | (x & VPB_BIT_WAITERS)))
1001 			continue;
1002 		return (1);
1003 	}
1004 }
1005 
1006 /*
1007  *	vm_page_sbusied:
1008  *
1009  *	Return a positive value if the page is shared busied, 0 otherwise.
1010  */
1011 int
1012 vm_page_sbusied(vm_page_t m)
1013 {
1014 	u_int x;
1015 
1016 	x = vm_page_busy_fetch(m);
1017 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1018 }
1019 
1020 /*
1021  *	vm_page_sunbusy:
1022  *
1023  *	Shared unbusy a page.
1024  */
1025 void
1026 vm_page_sunbusy(vm_page_t m)
1027 {
1028 	u_int x;
1029 
1030 	vm_page_assert_sbusied(m);
1031 
1032 	x = vm_page_busy_fetch(m);
1033 	for (;;) {
1034 		KASSERT(x != VPB_FREED,
1035 		    ("vm_page_sunbusy: Unlocking freed page."));
1036 		if (VPB_SHARERS(x) > 1) {
1037 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1038 			    x - VPB_ONE_SHARER))
1039 				break;
1040 			continue;
1041 		}
1042 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1043 		    ("vm_page_sunbusy: invalid lock state"));
1044 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1045 			continue;
1046 		if ((x & VPB_BIT_WAITERS) == 0)
1047 			break;
1048 		wakeup(m);
1049 		break;
1050 	}
1051 }
1052 
1053 /*
1054  *	vm_page_busy_sleep:
1055  *
1056  *	Sleep if the page is busy, using the page pointer as wchan.
1057  *	This is used to implement the hard-path of the busying mechanism.
1058  *
1059  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1060  *	will not sleep if the page is shared-busy.
1061  *
1062  *	The object lock must be held on entry.
1063  *
1064  *	Returns true if it slept and dropped the object lock, or false
1065  *	if there was no sleep and the lock is still held.
1066  */
1067 bool
1068 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1069 {
1070 	vm_object_t obj;
1071 
1072 	obj = m->object;
1073 	VM_OBJECT_ASSERT_LOCKED(obj);
1074 
1075 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1076 	    true));
1077 }
1078 
1079 /*
1080  *	vm_page_busy_sleep_unlocked:
1081  *
1082  *	Sleep if the page is busy, using the page pointer as wchan.
1083  *	This is used to implement the hard-path of busying mechanism.
1084  *
1085  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1086  *	will not sleep if the page is shared-busy.
1087  *
1088  *	The object lock must not be held on entry.  The operation will
1089  *	return if the page changes identity.
1090  */
1091 void
1092 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1093     const char *wmesg, int allocflags)
1094 {
1095 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1096 
1097 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1098 }
1099 
1100 /*
1101  *	_vm_page_busy_sleep:
1102  *
1103  *	Internal busy sleep function.  Verifies the page identity and
1104  *	lockstate against parameters.  Returns true if it sleeps and
1105  *	false otherwise.
1106  *
1107  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1108  *
1109  *	If locked is true the lock will be dropped for any true returns
1110  *	and held for any false returns.
1111  */
1112 static bool
1113 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1114     const char *wmesg, int allocflags, bool locked)
1115 {
1116 	bool xsleep;
1117 	u_int x;
1118 
1119 	/*
1120 	 * If the object is busy we must wait for that to drain to zero
1121 	 * before trying the page again.
1122 	 */
1123 	if (obj != NULL && vm_object_busied(obj)) {
1124 		if (locked)
1125 			VM_OBJECT_DROP(obj);
1126 		vm_object_busy_wait(obj, wmesg);
1127 		return (true);
1128 	}
1129 
1130 	if (!vm_page_busied(m))
1131 		return (false);
1132 
1133 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1134 	sleepq_lock(m);
1135 	x = vm_page_busy_fetch(m);
1136 	do {
1137 		/*
1138 		 * If the page changes objects or becomes unlocked we can
1139 		 * simply return.
1140 		 */
1141 		if (x == VPB_UNBUSIED ||
1142 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1143 		    m->object != obj || m->pindex != pindex) {
1144 			sleepq_release(m);
1145 			return (false);
1146 		}
1147 		if ((x & VPB_BIT_WAITERS) != 0)
1148 			break;
1149 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1150 	if (locked)
1151 		VM_OBJECT_DROP(obj);
1152 	DROP_GIANT();
1153 	sleepq_add(m, NULL, wmesg, 0, 0);
1154 	sleepq_wait(m, PVM);
1155 	PICKUP_GIANT();
1156 	return (true);
1157 }
1158 
1159 /*
1160  *	vm_page_trysbusy:
1161  *
1162  *	Try to shared busy a page.
1163  *	If the operation succeeds 1 is returned otherwise 0.
1164  *	The operation never sleeps.
1165  */
1166 int
1167 vm_page_trysbusy(vm_page_t m)
1168 {
1169 	vm_object_t obj;
1170 	u_int x;
1171 
1172 	obj = m->object;
1173 	x = vm_page_busy_fetch(m);
1174 	for (;;) {
1175 		if ((x & VPB_BIT_SHARED) == 0)
1176 			return (0);
1177 		/*
1178 		 * Reduce the window for transient busies that will trigger
1179 		 * false negatives in vm_page_ps_test().
1180 		 */
1181 		if (obj != NULL && vm_object_busied(obj))
1182 			return (0);
1183 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1184 		    x + VPB_ONE_SHARER))
1185 			break;
1186 	}
1187 
1188 	/* Refetch the object now that we're guaranteed that it is stable. */
1189 	obj = m->object;
1190 	if (obj != NULL && vm_object_busied(obj)) {
1191 		vm_page_sunbusy(m);
1192 		return (0);
1193 	}
1194 	return (1);
1195 }
1196 
1197 /*
1198  *	vm_page_tryxbusy:
1199  *
1200  *	Try to exclusive busy a page.
1201  *	If the operation succeeds 1 is returned otherwise 0.
1202  *	The operation never sleeps.
1203  */
1204 int
1205 vm_page_tryxbusy(vm_page_t m)
1206 {
1207 	vm_object_t obj;
1208 
1209         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1210             VPB_CURTHREAD_EXCLUSIVE) == 0)
1211 		return (0);
1212 
1213 	obj = m->object;
1214 	if (obj != NULL && vm_object_busied(obj)) {
1215 		vm_page_xunbusy(m);
1216 		return (0);
1217 	}
1218 	return (1);
1219 }
1220 
1221 static void
1222 vm_page_xunbusy_hard_tail(vm_page_t m)
1223 {
1224 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1225 	/* Wake the waiter. */
1226 	wakeup(m);
1227 }
1228 
1229 /*
1230  *	vm_page_xunbusy_hard:
1231  *
1232  *	Called when unbusy has failed because there is a waiter.
1233  */
1234 void
1235 vm_page_xunbusy_hard(vm_page_t m)
1236 {
1237 	vm_page_assert_xbusied(m);
1238 	vm_page_xunbusy_hard_tail(m);
1239 }
1240 
1241 void
1242 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1243 {
1244 	vm_page_assert_xbusied_unchecked(m);
1245 	vm_page_xunbusy_hard_tail(m);
1246 }
1247 
1248 static void
1249 vm_page_busy_free(vm_page_t m)
1250 {
1251 	u_int x;
1252 
1253 	atomic_thread_fence_rel();
1254 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1255 	if ((x & VPB_BIT_WAITERS) != 0)
1256 		wakeup(m);
1257 }
1258 
1259 /*
1260  *	vm_page_unhold_pages:
1261  *
1262  *	Unhold each of the pages that is referenced by the given array.
1263  */
1264 void
1265 vm_page_unhold_pages(vm_page_t *ma, int count)
1266 {
1267 
1268 	for (; count != 0; count--) {
1269 		vm_page_unwire(*ma, PQ_ACTIVE);
1270 		ma++;
1271 	}
1272 }
1273 
1274 vm_page_t
1275 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1276 {
1277 	vm_page_t m;
1278 
1279 #ifdef VM_PHYSSEG_SPARSE
1280 	m = vm_phys_paddr_to_vm_page(pa);
1281 	if (m == NULL)
1282 		m = vm_phys_fictitious_to_vm_page(pa);
1283 	return (m);
1284 #elif defined(VM_PHYSSEG_DENSE)
1285 	long pi;
1286 
1287 	pi = atop(pa);
1288 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1289 		m = &vm_page_array[pi - first_page];
1290 		return (m);
1291 	}
1292 	return (vm_phys_fictitious_to_vm_page(pa));
1293 #else
1294 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1295 #endif
1296 }
1297 
1298 /*
1299  *	vm_page_getfake:
1300  *
1301  *	Create a fictitious page with the specified physical address and
1302  *	memory attribute.  The memory attribute is the only the machine-
1303  *	dependent aspect of a fictitious page that must be initialized.
1304  */
1305 vm_page_t
1306 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1307 {
1308 	vm_page_t m;
1309 
1310 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1311 	vm_page_initfake(m, paddr, memattr);
1312 	return (m);
1313 }
1314 
1315 void
1316 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1317 {
1318 
1319 	if ((m->flags & PG_FICTITIOUS) != 0) {
1320 		/*
1321 		 * The page's memattr might have changed since the
1322 		 * previous initialization.  Update the pmap to the
1323 		 * new memattr.
1324 		 */
1325 		goto memattr;
1326 	}
1327 	m->phys_addr = paddr;
1328 	m->a.queue = PQ_NONE;
1329 	/* Fictitious pages don't use "segind". */
1330 	m->flags = PG_FICTITIOUS;
1331 	/* Fictitious pages don't use "order" or "pool". */
1332 	m->oflags = VPO_UNMANAGED;
1333 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1334 	/* Fictitious pages are unevictable. */
1335 	m->ref_count = 1;
1336 	pmap_page_init(m);
1337 memattr:
1338 	pmap_page_set_memattr(m, memattr);
1339 }
1340 
1341 /*
1342  *	vm_page_putfake:
1343  *
1344  *	Release a fictitious page.
1345  */
1346 void
1347 vm_page_putfake(vm_page_t m)
1348 {
1349 
1350 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1351 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1352 	    ("vm_page_putfake: bad page %p", m));
1353 	vm_page_assert_xbusied(m);
1354 	vm_page_busy_free(m);
1355 	uma_zfree(fakepg_zone, m);
1356 }
1357 
1358 /*
1359  *	vm_page_updatefake:
1360  *
1361  *	Update the given fictitious page to the specified physical address and
1362  *	memory attribute.
1363  */
1364 void
1365 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1366 {
1367 
1368 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1369 	    ("vm_page_updatefake: bad page %p", m));
1370 	m->phys_addr = paddr;
1371 	pmap_page_set_memattr(m, memattr);
1372 }
1373 
1374 /*
1375  *	vm_page_free:
1376  *
1377  *	Free a page.
1378  */
1379 void
1380 vm_page_free(vm_page_t m)
1381 {
1382 
1383 	m->flags &= ~PG_ZERO;
1384 	vm_page_free_toq(m);
1385 }
1386 
1387 /*
1388  *	vm_page_free_zero:
1389  *
1390  *	Free a page to the zerod-pages queue
1391  */
1392 void
1393 vm_page_free_zero(vm_page_t m)
1394 {
1395 
1396 	m->flags |= PG_ZERO;
1397 	vm_page_free_toq(m);
1398 }
1399 
1400 /*
1401  * Unbusy and handle the page queueing for a page from a getpages request that
1402  * was optionally read ahead or behind.
1403  */
1404 void
1405 vm_page_readahead_finish(vm_page_t m)
1406 {
1407 
1408 	/* We shouldn't put invalid pages on queues. */
1409 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1410 
1411 	/*
1412 	 * Since the page is not the actually needed one, whether it should
1413 	 * be activated or deactivated is not obvious.  Empirical results
1414 	 * have shown that deactivating the page is usually the best choice,
1415 	 * unless the page is wanted by another thread.
1416 	 */
1417 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1418 		vm_page_activate(m);
1419 	else
1420 		vm_page_deactivate(m);
1421 	vm_page_xunbusy_unchecked(m);
1422 }
1423 
1424 /*
1425  * Destroy the identity of an invalid page and free it if possible.
1426  * This is intended to be used when reading a page from backing store fails.
1427  */
1428 void
1429 vm_page_free_invalid(vm_page_t m)
1430 {
1431 
1432 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1433 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1434 	KASSERT(m->object != NULL, ("page %p has no object", m));
1435 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1436 
1437 	/*
1438 	 * We may be attempting to free the page as part of the handling for an
1439 	 * I/O error, in which case the page was xbusied by a different thread.
1440 	 */
1441 	vm_page_xbusy_claim(m);
1442 
1443 	/*
1444 	 * If someone has wired this page while the object lock
1445 	 * was not held, then the thread that unwires is responsible
1446 	 * for freeing the page.  Otherwise just free the page now.
1447 	 * The wire count of this unmapped page cannot change while
1448 	 * we have the page xbusy and the page's object wlocked.
1449 	 */
1450 	if (vm_page_remove(m))
1451 		vm_page_free(m);
1452 }
1453 
1454 /*
1455  *	vm_page_dirty_KBI:		[ internal use only ]
1456  *
1457  *	Set all bits in the page's dirty field.
1458  *
1459  *	The object containing the specified page must be locked if the
1460  *	call is made from the machine-independent layer.
1461  *
1462  *	See vm_page_clear_dirty_mask().
1463  *
1464  *	This function should only be called by vm_page_dirty().
1465  */
1466 void
1467 vm_page_dirty_KBI(vm_page_t m)
1468 {
1469 
1470 	/* Refer to this operation by its public name. */
1471 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1472 	m->dirty = VM_PAGE_BITS_ALL;
1473 }
1474 
1475 /*
1476  * Insert the given page into the given object at the given pindex.  mpred is
1477  * used for memq linkage.  From vm_page_insert, lookup is true, mpred is
1478  * initially NULL, and this procedure looks it up.  From vm_page_insert_after,
1479  * lookup is false and mpred is known to the caller to be valid, and may be
1480  * NULL if this will be the page with the lowest pindex.
1481  *
1482  * The procedure is marked __always_inline to suggest to the compiler to
1483  * eliminate the lookup parameter and the associated alternate branch.
1484  */
1485 static __always_inline int
1486 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1487     vm_page_t mpred, bool lookup)
1488 {
1489 	int error;
1490 
1491 	VM_OBJECT_ASSERT_WLOCKED(object);
1492 	KASSERT(m->object == NULL,
1493 	    ("vm_page_insert: page %p already inserted", m));
1494 
1495 	/*
1496 	 * Record the object/offset pair in this page.
1497 	 */
1498 	m->object = object;
1499 	m->pindex = pindex;
1500 	m->ref_count |= VPRC_OBJREF;
1501 
1502 	/*
1503 	 * Add this page to the object's radix tree, and look up mpred if
1504 	 * needed.
1505 	 */
1506 	if (lookup)
1507 		error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred);
1508 	else
1509 		error = vm_radix_insert(&object->rtree, m);
1510 	if (__predict_false(error != 0)) {
1511 		m->object = NULL;
1512 		m->pindex = 0;
1513 		m->ref_count &= ~VPRC_OBJREF;
1514 		return (1);
1515 	}
1516 
1517 	/*
1518 	 * Now link into the object's ordered list of backed pages.
1519 	 */
1520 	vm_page_insert_radixdone(m, object, mpred);
1521 	vm_pager_page_inserted(object, m);
1522 	return (0);
1523 }
1524 
1525 /*
1526  *	vm_page_insert:		[ internal use only ]
1527  *
1528  *	Inserts the given mem entry into the object and object list.
1529  *
1530  *	The object must be locked.
1531  */
1532 int
1533 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1534 {
1535 	return (vm_page_insert_lookup(m, object, pindex, NULL, true));
1536 }
1537 
1538 /*
1539  *	vm_page_insert_after:
1540  *
1541  *	Inserts the page "m" into the specified object at offset "pindex".
1542  *
1543  *	The page "mpred" must immediately precede the offset "pindex" within
1544  *	the specified object.
1545  *
1546  *	The object must be locked.
1547  */
1548 static int
1549 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1550     vm_page_t mpred)
1551 {
1552 	return (vm_page_insert_lookup(m, object, pindex, mpred, false));
1553 }
1554 
1555 /*
1556  *	vm_page_insert_radixdone:
1557  *
1558  *	Complete page "m" insertion into the specified object after the
1559  *	radix trie hooking.
1560  *
1561  *	The page "mpred" must precede the offset "m->pindex" within the
1562  *	specified object.
1563  *
1564  *	The object must be locked.
1565  */
1566 static void
1567 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1568 {
1569 
1570 	VM_OBJECT_ASSERT_WLOCKED(object);
1571 	KASSERT(object != NULL && m->object == object,
1572 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1573 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1574 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1575 	if (mpred != NULL) {
1576 		KASSERT(mpred->object == object,
1577 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1578 		KASSERT(mpred->pindex < m->pindex,
1579 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1580 		KASSERT(TAILQ_NEXT(mpred, listq) == NULL ||
1581 		    m->pindex < TAILQ_NEXT(mpred, listq)->pindex,
1582 		    ("vm_page_insert_radixdone: pindex doesn't precede msucc"));
1583 	} else {
1584 		KASSERT(TAILQ_EMPTY(&object->memq) ||
1585 		    m->pindex < TAILQ_FIRST(&object->memq)->pindex,
1586 		    ("vm_page_insert_radixdone: no mpred but not first page"));
1587 	}
1588 
1589 	if (mpred != NULL)
1590 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1591 	else
1592 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1593 
1594 	/*
1595 	 * Show that the object has one more resident page.
1596 	 */
1597 	object->resident_page_count++;
1598 
1599 	/*
1600 	 * Hold the vnode until the last page is released.
1601 	 */
1602 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1603 		vhold(object->handle);
1604 
1605 	/*
1606 	 * Since we are inserting a new and possibly dirty page,
1607 	 * update the object's generation count.
1608 	 */
1609 	if (pmap_page_is_write_mapped(m))
1610 		vm_object_set_writeable_dirty(object);
1611 }
1612 
1613 /*
1614  * Do the work to remove a page from its object.  The caller is responsible for
1615  * updating the page's fields to reflect this removal.
1616  */
1617 static void
1618 vm_page_object_remove(vm_page_t m)
1619 {
1620 	vm_object_t object;
1621 	vm_page_t mrem __diagused;
1622 
1623 	vm_page_assert_xbusied(m);
1624 	object = m->object;
1625 	VM_OBJECT_ASSERT_WLOCKED(object);
1626 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1627 	    ("page %p is missing its object ref", m));
1628 
1629 	/* Deferred free of swap space. */
1630 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1631 		vm_pager_page_unswapped(m);
1632 
1633 	vm_pager_page_removed(object, m);
1634 
1635 	m->object = NULL;
1636 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1637 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1638 
1639 	/*
1640 	 * Now remove from the object's list of backed pages.
1641 	 */
1642 	TAILQ_REMOVE(&object->memq, m, listq);
1643 
1644 	/*
1645 	 * And show that the object has one fewer resident page.
1646 	 */
1647 	object->resident_page_count--;
1648 
1649 	/*
1650 	 * The vnode may now be recycled.
1651 	 */
1652 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1653 		vdrop(object->handle);
1654 }
1655 
1656 /*
1657  *	vm_page_remove:
1658  *
1659  *	Removes the specified page from its containing object, but does not
1660  *	invalidate any backing storage.  Returns true if the object's reference
1661  *	was the last reference to the page, and false otherwise.
1662  *
1663  *	The object must be locked and the page must be exclusively busied.
1664  *	The exclusive busy will be released on return.  If this is not the
1665  *	final ref and the caller does not hold a wire reference it may not
1666  *	continue to access the page.
1667  */
1668 bool
1669 vm_page_remove(vm_page_t m)
1670 {
1671 	bool dropped;
1672 
1673 	dropped = vm_page_remove_xbusy(m);
1674 	vm_page_xunbusy(m);
1675 
1676 	return (dropped);
1677 }
1678 
1679 /*
1680  *	vm_page_remove_xbusy
1681  *
1682  *	Removes the page but leaves the xbusy held.  Returns true if this
1683  *	removed the final ref and false otherwise.
1684  */
1685 bool
1686 vm_page_remove_xbusy(vm_page_t m)
1687 {
1688 
1689 	vm_page_object_remove(m);
1690 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1691 }
1692 
1693 /*
1694  *	vm_page_lookup:
1695  *
1696  *	Returns the page associated with the object/offset
1697  *	pair specified; if none is found, NULL is returned.
1698  *
1699  *	The object must be locked.
1700  */
1701 vm_page_t
1702 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1703 {
1704 
1705 	VM_OBJECT_ASSERT_LOCKED(object);
1706 	return (vm_radix_lookup(&object->rtree, pindex));
1707 }
1708 
1709 /*
1710  *	vm_page_lookup_unlocked:
1711  *
1712  *	Returns the page associated with the object/offset pair specified;
1713  *	if none is found, NULL is returned.  The page may be no longer be
1714  *	present in the object at the time that this function returns.  Only
1715  *	useful for opportunistic checks such as inmem().
1716  */
1717 vm_page_t
1718 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1719 {
1720 
1721 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1722 }
1723 
1724 /*
1725  *	vm_page_relookup:
1726  *
1727  *	Returns a page that must already have been busied by
1728  *	the caller.  Used for bogus page replacement.
1729  */
1730 vm_page_t
1731 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1732 {
1733 	vm_page_t m;
1734 
1735 	m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1736 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1737 	    m->object == object && m->pindex == pindex,
1738 	    ("vm_page_relookup: Invalid page %p", m));
1739 	return (m);
1740 }
1741 
1742 /*
1743  * This should only be used by lockless functions for releasing transient
1744  * incorrect acquires.  The page may have been freed after we acquired a
1745  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1746  * further to do.
1747  */
1748 static void
1749 vm_page_busy_release(vm_page_t m)
1750 {
1751 	u_int x;
1752 
1753 	x = vm_page_busy_fetch(m);
1754 	for (;;) {
1755 		if (x == VPB_FREED)
1756 			break;
1757 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1758 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1759 			    x - VPB_ONE_SHARER))
1760 				break;
1761 			continue;
1762 		}
1763 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1764 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1765 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1766 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1767 			continue;
1768 		if ((x & VPB_BIT_WAITERS) != 0)
1769 			wakeup(m);
1770 		break;
1771 	}
1772 }
1773 
1774 /*
1775  *	vm_page_find_least:
1776  *
1777  *	Returns the page associated with the object with least pindex
1778  *	greater than or equal to the parameter pindex, or NULL.
1779  *
1780  *	The object must be locked.
1781  */
1782 vm_page_t
1783 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1784 {
1785 	vm_page_t m;
1786 
1787 	VM_OBJECT_ASSERT_LOCKED(object);
1788 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1789 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1790 	return (m);
1791 }
1792 
1793 /*
1794  * Returns the given page's successor (by pindex) within the object if it is
1795  * resident; if none is found, NULL is returned.
1796  *
1797  * The object must be locked.
1798  */
1799 vm_page_t
1800 vm_page_next(vm_page_t m)
1801 {
1802 	vm_page_t next;
1803 
1804 	VM_OBJECT_ASSERT_LOCKED(m->object);
1805 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1806 		MPASS(next->object == m->object);
1807 		if (next->pindex != m->pindex + 1)
1808 			next = NULL;
1809 	}
1810 	return (next);
1811 }
1812 
1813 /*
1814  * Returns the given page's predecessor (by pindex) within the object if it is
1815  * resident; if none is found, NULL is returned.
1816  *
1817  * The object must be locked.
1818  */
1819 vm_page_t
1820 vm_page_prev(vm_page_t m)
1821 {
1822 	vm_page_t prev;
1823 
1824 	VM_OBJECT_ASSERT_LOCKED(m->object);
1825 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1826 		MPASS(prev->object == m->object);
1827 		if (prev->pindex != m->pindex - 1)
1828 			prev = NULL;
1829 	}
1830 	return (prev);
1831 }
1832 
1833 /*
1834  * Uses the page mnew as a replacement for an existing page at index
1835  * pindex which must be already present in the object.
1836  *
1837  * Both pages must be exclusively busied on enter.  The old page is
1838  * unbusied on exit.
1839  *
1840  * A return value of true means mold is now free.  If this is not the
1841  * final ref and the caller does not hold a wire reference it may not
1842  * continue to access the page.
1843  */
1844 static bool
1845 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1846     vm_page_t mold)
1847 {
1848 	vm_page_t mret __diagused;
1849 	bool dropped;
1850 
1851 	VM_OBJECT_ASSERT_WLOCKED(object);
1852 	vm_page_assert_xbusied(mold);
1853 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1854 	    ("vm_page_replace: page %p already in object", mnew));
1855 
1856 	/*
1857 	 * This function mostly follows vm_page_insert() and
1858 	 * vm_page_remove() without the radix, object count and vnode
1859 	 * dance.  Double check such functions for more comments.
1860 	 */
1861 
1862 	mnew->object = object;
1863 	mnew->pindex = pindex;
1864 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1865 	mret = vm_radix_replace(&object->rtree, mnew);
1866 	KASSERT(mret == mold,
1867 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1868 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1869 	    (mnew->oflags & VPO_UNMANAGED),
1870 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1871 
1872 	/* Keep the resident page list in sorted order. */
1873 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1874 	TAILQ_REMOVE(&object->memq, mold, listq);
1875 	mold->object = NULL;
1876 
1877 	/*
1878 	 * The object's resident_page_count does not change because we have
1879 	 * swapped one page for another, but the generation count should
1880 	 * change if the page is dirty.
1881 	 */
1882 	if (pmap_page_is_write_mapped(mnew))
1883 		vm_object_set_writeable_dirty(object);
1884 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1885 	vm_page_xunbusy(mold);
1886 
1887 	return (dropped);
1888 }
1889 
1890 void
1891 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1892     vm_page_t mold)
1893 {
1894 
1895 	vm_page_assert_xbusied(mnew);
1896 
1897 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1898 		vm_page_free(mold);
1899 }
1900 
1901 /*
1902  *	vm_page_rename:
1903  *
1904  *	Move the given memory entry from its
1905  *	current object to the specified target object/offset.
1906  *
1907  *	Note: swap associated with the page must be invalidated by the move.  We
1908  *	      have to do this for several reasons:  (1) we aren't freeing the
1909  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1910  *	      moving the page from object A to B, and will then later move
1911  *	      the backing store from A to B and we can't have a conflict.
1912  *
1913  *	Note: we *always* dirty the page.  It is necessary both for the
1914  *	      fact that we moved it, and because we may be invalidating
1915  *	      swap.
1916  *
1917  *	The objects must be locked.
1918  */
1919 int
1920 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1921 {
1922 	vm_page_t mpred;
1923 	vm_pindex_t opidx;
1924 
1925 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1926 
1927 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1928 
1929 	/*
1930 	 * Create a custom version of vm_page_insert() which does not depend
1931 	 * by m_prev and can cheat on the implementation aspects of the
1932 	 * function.
1933 	 */
1934 	opidx = m->pindex;
1935 	m->pindex = new_pindex;
1936 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
1937 		m->pindex = opidx;
1938 		return (1);
1939 	}
1940 
1941 	/*
1942 	 * The operation cannot fail anymore.  The removal must happen before
1943 	 * the listq iterator is tainted.
1944 	 */
1945 	m->pindex = opidx;
1946 	vm_page_object_remove(m);
1947 
1948 	/* Return back to the new pindex to complete vm_page_insert(). */
1949 	m->pindex = new_pindex;
1950 	m->object = new_object;
1951 
1952 	vm_page_insert_radixdone(m, new_object, mpred);
1953 	vm_page_dirty(m);
1954 	vm_pager_page_inserted(new_object, m);
1955 	return (0);
1956 }
1957 
1958 /*
1959  *	vm_page_alloc:
1960  *
1961  *	Allocate and return a page that is associated with the specified
1962  *	object and offset pair.  By default, this page is exclusive busied.
1963  *
1964  *	The caller must always specify an allocation class.
1965  *
1966  *	allocation classes:
1967  *	VM_ALLOC_NORMAL		normal process request
1968  *	VM_ALLOC_SYSTEM		system *really* needs a page
1969  *	VM_ALLOC_INTERRUPT	interrupt time request
1970  *
1971  *	optional allocation flags:
1972  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1973  *				intends to allocate
1974  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1975  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1976  *	VM_ALLOC_SBUSY		shared busy the allocated page
1977  *	VM_ALLOC_WIRED		wire the allocated page
1978  *	VM_ALLOC_ZERO		prefer a zeroed page
1979  */
1980 vm_page_t
1981 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1982 {
1983 
1984 	return (vm_page_alloc_after(object, pindex, req,
1985 	    vm_radix_lookup_le(&object->rtree, pindex)));
1986 }
1987 
1988 vm_page_t
1989 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1990     int req)
1991 {
1992 
1993 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1994 	    vm_radix_lookup_le(&object->rtree, pindex)));
1995 }
1996 
1997 /*
1998  * Allocate a page in the specified object with the given page index.  To
1999  * optimize insertion of the page into the object, the caller must also specify
2000  * the resident page in the object with largest index smaller than the given
2001  * page index, or NULL if no such page exists.
2002  */
2003 vm_page_t
2004 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
2005     int req, vm_page_t mpred)
2006 {
2007 	struct vm_domainset_iter di;
2008 	vm_page_t m;
2009 	int domain;
2010 
2011 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2012 	do {
2013 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2014 		    mpred);
2015 		if (m != NULL)
2016 			break;
2017 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2018 
2019 	return (m);
2020 }
2021 
2022 /*
2023  * Returns true if the number of free pages exceeds the minimum
2024  * for the request class and false otherwise.
2025  */
2026 static int
2027 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2028 {
2029 	u_int limit, old, new;
2030 
2031 	if (req_class == VM_ALLOC_INTERRUPT)
2032 		limit = 0;
2033 	else if (req_class == VM_ALLOC_SYSTEM)
2034 		limit = vmd->vmd_interrupt_free_min;
2035 	else
2036 		limit = vmd->vmd_free_reserved;
2037 
2038 	/*
2039 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2040 	 */
2041 	limit += npages;
2042 	old = vmd->vmd_free_count;
2043 	do {
2044 		if (old < limit)
2045 			return (0);
2046 		new = old - npages;
2047 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2048 
2049 	/* Wake the page daemon if we've crossed the threshold. */
2050 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2051 		pagedaemon_wakeup(vmd->vmd_domain);
2052 
2053 	/* Only update bitsets on transitions. */
2054 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2055 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2056 		vm_domain_set(vmd);
2057 
2058 	return (1);
2059 }
2060 
2061 int
2062 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2063 {
2064 	int req_class;
2065 
2066 	/*
2067 	 * The page daemon is allowed to dig deeper into the free page list.
2068 	 */
2069 	req_class = req & VM_ALLOC_CLASS_MASK;
2070 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2071 		req_class = VM_ALLOC_SYSTEM;
2072 	return (_vm_domain_allocate(vmd, req_class, npages));
2073 }
2074 
2075 vm_page_t
2076 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2077     int req, vm_page_t mpred)
2078 {
2079 	struct vm_domain *vmd;
2080 	vm_page_t m;
2081 	int flags;
2082 
2083 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2084 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2085 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2086 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2087 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2088 	KASSERT((req & ~VPA_FLAGS) == 0,
2089 	    ("invalid request %#x", req));
2090 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2091 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2092 	    ("invalid request %#x", req));
2093 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2094 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2095 	    (uintmax_t)pindex));
2096 	VM_OBJECT_ASSERT_WLOCKED(object);
2097 
2098 	flags = 0;
2099 	m = NULL;
2100 	if (!vm_pager_can_alloc_page(object, pindex))
2101 		return (NULL);
2102 again:
2103 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2104 		m = vm_page_alloc_nofree_domain(domain, req);
2105 		if (m != NULL)
2106 			goto found;
2107 	}
2108 #if VM_NRESERVLEVEL > 0
2109 	/*
2110 	 * Can we allocate the page from a reservation?
2111 	 */
2112 	if (vm_object_reserv(object) &&
2113 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2114 	    NULL) {
2115 		goto found;
2116 	}
2117 #endif
2118 	vmd = VM_DOMAIN(domain);
2119 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2120 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2121 		    M_NOWAIT | M_NOVM);
2122 		if (m != NULL) {
2123 			flags |= PG_PCPU_CACHE;
2124 			goto found;
2125 		}
2126 	}
2127 	if (vm_domain_allocate(vmd, req, 1)) {
2128 		/*
2129 		 * If not, allocate it from the free page queues.
2130 		 */
2131 		vm_domain_free_lock(vmd);
2132 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2133 		vm_domain_free_unlock(vmd);
2134 		if (m == NULL) {
2135 			vm_domain_freecnt_inc(vmd, 1);
2136 #if VM_NRESERVLEVEL > 0
2137 			if (vm_reserv_reclaim_inactive(domain))
2138 				goto again;
2139 #endif
2140 		}
2141 	}
2142 	if (m == NULL) {
2143 		/*
2144 		 * Not allocatable, give up.
2145 		 */
2146 		if (vm_domain_alloc_fail(vmd, object, req))
2147 			goto again;
2148 		return (NULL);
2149 	}
2150 
2151 	/*
2152 	 * At this point we had better have found a good page.
2153 	 */
2154 found:
2155 	vm_page_dequeue(m);
2156 	vm_page_alloc_check(m);
2157 
2158 	/*
2159 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2160 	 */
2161 	flags |= m->flags & PG_ZERO;
2162 	if ((req & VM_ALLOC_NODUMP) != 0)
2163 		flags |= PG_NODUMP;
2164 	if ((req & VM_ALLOC_NOFREE) != 0)
2165 		flags |= PG_NOFREE;
2166 	m->flags = flags;
2167 	m->a.flags = 0;
2168 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2169 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2170 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2171 	else if ((req & VM_ALLOC_SBUSY) != 0)
2172 		m->busy_lock = VPB_SHARERS_WORD(1);
2173 	else
2174 		m->busy_lock = VPB_UNBUSIED;
2175 	if (req & VM_ALLOC_WIRED) {
2176 		vm_wire_add(1);
2177 		m->ref_count = 1;
2178 	}
2179 	m->a.act_count = 0;
2180 
2181 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2182 		if (req & VM_ALLOC_WIRED) {
2183 			vm_wire_sub(1);
2184 			m->ref_count = 0;
2185 		}
2186 		KASSERT(m->object == NULL, ("page %p has object", m));
2187 		m->oflags = VPO_UNMANAGED;
2188 		m->busy_lock = VPB_UNBUSIED;
2189 		/* Don't change PG_ZERO. */
2190 		vm_page_free_toq(m);
2191 		if (req & VM_ALLOC_WAITFAIL) {
2192 			VM_OBJECT_WUNLOCK(object);
2193 			vm_radix_wait();
2194 			VM_OBJECT_WLOCK(object);
2195 		}
2196 		return (NULL);
2197 	}
2198 
2199 	/* Ignore device objects; the pager sets "memattr" for them. */
2200 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2201 	    (object->flags & OBJ_FICTITIOUS) == 0)
2202 		pmap_page_set_memattr(m, object->memattr);
2203 
2204 	return (m);
2205 }
2206 
2207 /*
2208  *	vm_page_alloc_contig:
2209  *
2210  *	Allocate a contiguous set of physical pages of the given size "npages"
2211  *	from the free lists.  All of the physical pages must be at or above
2212  *	the given physical address "low" and below the given physical address
2213  *	"high".  The given value "alignment" determines the alignment of the
2214  *	first physical page in the set.  If the given value "boundary" is
2215  *	non-zero, then the set of physical pages cannot cross any physical
2216  *	address boundary that is a multiple of that value.  Both "alignment"
2217  *	and "boundary" must be a power of two.
2218  *
2219  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2220  *	then the memory attribute setting for the physical pages is configured
2221  *	to the object's memory attribute setting.  Otherwise, the memory
2222  *	attribute setting for the physical pages is configured to "memattr",
2223  *	overriding the object's memory attribute setting.  However, if the
2224  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2225  *	memory attribute setting for the physical pages cannot be configured
2226  *	to VM_MEMATTR_DEFAULT.
2227  *
2228  *	The specified object may not contain fictitious pages.
2229  *
2230  *	The caller must always specify an allocation class.
2231  *
2232  *	allocation classes:
2233  *	VM_ALLOC_NORMAL		normal process request
2234  *	VM_ALLOC_SYSTEM		system *really* needs a page
2235  *	VM_ALLOC_INTERRUPT	interrupt time request
2236  *
2237  *	optional allocation flags:
2238  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2239  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2240  *	VM_ALLOC_SBUSY		shared busy the allocated page
2241  *	VM_ALLOC_WIRED		wire the allocated page
2242  *	VM_ALLOC_ZERO		prefer a zeroed page
2243  */
2244 vm_page_t
2245 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2246     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_domainset_iter di;
2250 	vm_page_t bounds[2];
2251 	vm_page_t m;
2252 	int domain;
2253 	int start_segind;
2254 
2255 	start_segind = -1;
2256 
2257 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2258 	do {
2259 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2260 		    npages, low, high, alignment, boundary, memattr);
2261 		if (m != NULL)
2262 			break;
2263 		if (start_segind == -1)
2264 			start_segind = vm_phys_lookup_segind(low);
2265 		if (vm_phys_find_range(bounds, start_segind, domain,
2266 		    npages, low, high) == -1) {
2267 			vm_domainset_iter_ignore(&di, domain);
2268 		}
2269 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2270 
2271 	return (m);
2272 }
2273 
2274 static vm_page_t
2275 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2276     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2277 {
2278 	struct vm_domain *vmd;
2279 	vm_page_t m_ret;
2280 
2281 	/*
2282 	 * Can we allocate the pages without the number of free pages falling
2283 	 * below the lower bound for the allocation class?
2284 	 */
2285 	vmd = VM_DOMAIN(domain);
2286 	if (!vm_domain_allocate(vmd, req, npages))
2287 		return (NULL);
2288 	/*
2289 	 * Try to allocate the pages from the free page queues.
2290 	 */
2291 	vm_domain_free_lock(vmd);
2292 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2293 	    alignment, boundary);
2294 	vm_domain_free_unlock(vmd);
2295 	if (m_ret != NULL)
2296 		return (m_ret);
2297 #if VM_NRESERVLEVEL > 0
2298 	/*
2299 	 * Try to break a reservation to allocate the pages.
2300 	 */
2301 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2302 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2303 	            high, alignment, boundary);
2304 		if (m_ret != NULL)
2305 			return (m_ret);
2306 	}
2307 #endif
2308 	vm_domain_freecnt_inc(vmd, npages);
2309 	return (NULL);
2310 }
2311 
2312 vm_page_t
2313 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2314     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2315     vm_paddr_t boundary, vm_memattr_t memattr)
2316 {
2317 	vm_page_t m, m_ret, mpred;
2318 	u_int busy_lock, flags, oflags;
2319 
2320 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2321 	KASSERT((req & ~VPAC_FLAGS) == 0,
2322 	    ("invalid request %#x", req));
2323 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2324 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2325 	    ("invalid request %#x", req));
2326 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2327 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2328 	    ("invalid request %#x", req));
2329 	VM_OBJECT_ASSERT_WLOCKED(object);
2330 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2331 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2332 	    object));
2333 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2334 
2335 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
2336 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2337 	    ("vm_page_alloc_contig: pindex already allocated"));
2338 	for (;;) {
2339 #if VM_NRESERVLEVEL > 0
2340 		/*
2341 		 * Can we allocate the pages from a reservation?
2342 		 */
2343 		if (vm_object_reserv(object) &&
2344 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2345 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2346 			break;
2347 		}
2348 #endif
2349 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2350 		    low, high, alignment, boundary)) != NULL)
2351 			break;
2352 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2353 			return (NULL);
2354 	}
2355 
2356 	/*
2357 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2358 	 */
2359 	flags = PG_ZERO;
2360 	if ((req & VM_ALLOC_NODUMP) != 0)
2361 		flags |= PG_NODUMP;
2362 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2363 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2364 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2365 	else if ((req & VM_ALLOC_SBUSY) != 0)
2366 		busy_lock = VPB_SHARERS_WORD(1);
2367 	else
2368 		busy_lock = VPB_UNBUSIED;
2369 	if ((req & VM_ALLOC_WIRED) != 0)
2370 		vm_wire_add(npages);
2371 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2372 	    memattr == VM_MEMATTR_DEFAULT)
2373 		memattr = object->memattr;
2374 	for (m = m_ret; m < &m_ret[npages]; m++) {
2375 		vm_page_dequeue(m);
2376 		vm_page_alloc_check(m);
2377 		m->a.flags = 0;
2378 		m->flags = (m->flags | PG_NODUMP) & flags;
2379 		m->busy_lock = busy_lock;
2380 		if ((req & VM_ALLOC_WIRED) != 0)
2381 			m->ref_count = 1;
2382 		m->a.act_count = 0;
2383 		m->oflags = oflags;
2384 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2385 			if ((req & VM_ALLOC_WIRED) != 0)
2386 				vm_wire_sub(npages);
2387 			KASSERT(m->object == NULL,
2388 			    ("page %p has object", m));
2389 			mpred = m;
2390 			for (m = m_ret; m < &m_ret[npages]; m++) {
2391 				if (m <= mpred &&
2392 				    (req & VM_ALLOC_WIRED) != 0)
2393 					m->ref_count = 0;
2394 				m->oflags = VPO_UNMANAGED;
2395 				m->busy_lock = VPB_UNBUSIED;
2396 				/* Don't change PG_ZERO. */
2397 				vm_page_free_toq(m);
2398 			}
2399 			if (req & VM_ALLOC_WAITFAIL) {
2400 				VM_OBJECT_WUNLOCK(object);
2401 				vm_radix_wait();
2402 				VM_OBJECT_WLOCK(object);
2403 			}
2404 			return (NULL);
2405 		}
2406 		mpred = m;
2407 		if (memattr != VM_MEMATTR_DEFAULT)
2408 			pmap_page_set_memattr(m, memattr);
2409 		pindex++;
2410 	}
2411 	return (m_ret);
2412 }
2413 
2414 /*
2415  * Allocate a physical page that is not intended to be inserted into a VM
2416  * object.
2417  */
2418 vm_page_t
2419 vm_page_alloc_noobj_domain(int domain, int req)
2420 {
2421 	struct vm_domain *vmd;
2422 	vm_page_t m;
2423 	int flags;
2424 
2425 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2426 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2427 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2428 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2429 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2430 	KASSERT((req & ~VPAN_FLAGS) == 0,
2431 	    ("invalid request %#x", req));
2432 
2433 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2434 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2435 	vmd = VM_DOMAIN(domain);
2436 again:
2437 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2438 		m = vm_page_alloc_nofree_domain(domain, req);
2439 		if (m != NULL)
2440 			goto found;
2441 	}
2442 
2443 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2444 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2445 		    M_NOWAIT | M_NOVM);
2446 		if (m != NULL) {
2447 			flags |= PG_PCPU_CACHE;
2448 			goto found;
2449 		}
2450 	}
2451 
2452 	if (vm_domain_allocate(vmd, req, 1)) {
2453 		vm_domain_free_lock(vmd);
2454 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2455 		vm_domain_free_unlock(vmd);
2456 		if (m == NULL) {
2457 			vm_domain_freecnt_inc(vmd, 1);
2458 #if VM_NRESERVLEVEL > 0
2459 			if (vm_reserv_reclaim_inactive(domain))
2460 				goto again;
2461 #endif
2462 		}
2463 	}
2464 	if (m == NULL) {
2465 		if (vm_domain_alloc_fail(vmd, NULL, req))
2466 			goto again;
2467 		return (NULL);
2468 	}
2469 
2470 found:
2471 	vm_page_dequeue(m);
2472 	vm_page_alloc_check(m);
2473 
2474 	/*
2475 	 * Consumers should not rely on a useful default pindex value.
2476 	 */
2477 	m->pindex = 0xdeadc0dedeadc0de;
2478 	m->flags = (m->flags & PG_ZERO) | flags;
2479 	m->a.flags = 0;
2480 	m->oflags = VPO_UNMANAGED;
2481 	m->busy_lock = VPB_UNBUSIED;
2482 	if ((req & VM_ALLOC_WIRED) != 0) {
2483 		vm_wire_add(1);
2484 		m->ref_count = 1;
2485 	}
2486 
2487 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2488 		pmap_zero_page(m);
2489 
2490 	return (m);
2491 }
2492 
2493 #if VM_NRESERVLEVEL > 1
2494 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2495 #elif VM_NRESERVLEVEL > 0
2496 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2497 #else
2498 #define	VM_NOFREE_IMPORT_ORDER	8
2499 #endif
2500 
2501 /*
2502  * Allocate a single NOFREE page.
2503  *
2504  * This routine hands out NOFREE pages from higher-order
2505  * physical memory blocks in order to reduce memory fragmentation.
2506  * When a NOFREE for a given domain chunk is used up,
2507  * the routine will try to fetch a new one from the freelists
2508  * and discard the old one.
2509  */
2510 static vm_page_t
2511 vm_page_alloc_nofree_domain(int domain, int req)
2512 {
2513 	vm_page_t m;
2514 	struct vm_domain *vmd;
2515 	struct vm_nofreeq *nqp;
2516 
2517 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2518 
2519 	vmd = VM_DOMAIN(domain);
2520 	nqp = &vmd->vmd_nofreeq;
2521 	vm_domain_free_lock(vmd);
2522 	if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) {
2523 		if (!vm_domain_allocate(vmd, req,
2524 		    1 << VM_NOFREE_IMPORT_ORDER)) {
2525 			vm_domain_free_unlock(vmd);
2526 			return (NULL);
2527 		}
2528 		nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2529 		    VM_NOFREE_IMPORT_ORDER);
2530 		if (nqp->ma == NULL) {
2531 			vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER);
2532 			vm_domain_free_unlock(vmd);
2533 			return (NULL);
2534 		}
2535 		nqp->offs = 0;
2536 	}
2537 	m = &nqp->ma[nqp->offs++];
2538 	vm_domain_free_unlock(vmd);
2539 
2540 	return (m);
2541 }
2542 
2543 vm_page_t
2544 vm_page_alloc_noobj(int req)
2545 {
2546 	struct vm_domainset_iter di;
2547 	vm_page_t m;
2548 	int domain;
2549 
2550 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2551 	do {
2552 		m = vm_page_alloc_noobj_domain(domain, req);
2553 		if (m != NULL)
2554 			break;
2555 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2556 
2557 	return (m);
2558 }
2559 
2560 vm_page_t
2561 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2562     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2563     vm_memattr_t memattr)
2564 {
2565 	struct vm_domainset_iter di;
2566 	vm_page_t m;
2567 	int domain;
2568 
2569 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2570 	do {
2571 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2572 		    high, alignment, boundary, memattr);
2573 		if (m != NULL)
2574 			break;
2575 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2576 
2577 	return (m);
2578 }
2579 
2580 vm_page_t
2581 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2582     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2583     vm_memattr_t memattr)
2584 {
2585 	vm_page_t m, m_ret;
2586 	u_int flags;
2587 
2588 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2589 	KASSERT((req & ~VPANC_FLAGS) == 0,
2590 	    ("invalid request %#x", req));
2591 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2592 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2593 	    ("invalid request %#x", req));
2594 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2595 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2596 	    ("invalid request %#x", req));
2597 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2598 
2599 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2600 	    low, high, alignment, boundary)) == NULL) {
2601 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2602 			return (NULL);
2603 	}
2604 
2605 	/*
2606 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2607 	 */
2608 	flags = PG_ZERO;
2609 	if ((req & VM_ALLOC_NODUMP) != 0)
2610 		flags |= PG_NODUMP;
2611 	if ((req & VM_ALLOC_WIRED) != 0)
2612 		vm_wire_add(npages);
2613 	for (m = m_ret; m < &m_ret[npages]; m++) {
2614 		vm_page_dequeue(m);
2615 		vm_page_alloc_check(m);
2616 
2617 		/*
2618 		 * Consumers should not rely on a useful default pindex value.
2619 		 */
2620 		m->pindex = 0xdeadc0dedeadc0de;
2621 		m->a.flags = 0;
2622 		m->flags = (m->flags | PG_NODUMP) & flags;
2623 		m->busy_lock = VPB_UNBUSIED;
2624 		if ((req & VM_ALLOC_WIRED) != 0)
2625 			m->ref_count = 1;
2626 		m->a.act_count = 0;
2627 		m->oflags = VPO_UNMANAGED;
2628 
2629 		/*
2630 		 * Zero the page before updating any mappings since the page is
2631 		 * not yet shared with any devices which might require the
2632 		 * non-default memory attribute.  pmap_page_set_memattr()
2633 		 * flushes data caches before returning.
2634 		 */
2635 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2636 			pmap_zero_page(m);
2637 		if (memattr != VM_MEMATTR_DEFAULT)
2638 			pmap_page_set_memattr(m, memattr);
2639 	}
2640 	return (m_ret);
2641 }
2642 
2643 /*
2644  * Check a page that has been freshly dequeued from a freelist.
2645  */
2646 static void
2647 vm_page_alloc_check(vm_page_t m)
2648 {
2649 
2650 	KASSERT(m->object == NULL, ("page %p has object", m));
2651 	KASSERT(m->a.queue == PQ_NONE &&
2652 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2653 	    ("page %p has unexpected queue %d, flags %#x",
2654 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2655 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2656 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2657 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2658 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2659 	    ("page %p has unexpected memattr %d",
2660 	    m, pmap_page_get_memattr(m)));
2661 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2662 	pmap_vm_page_alloc_check(m);
2663 }
2664 
2665 static int
2666 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2667 {
2668 	struct vm_domain *vmd;
2669 	struct vm_pgcache *pgcache;
2670 	int i;
2671 
2672 	pgcache = arg;
2673 	vmd = VM_DOMAIN(pgcache->domain);
2674 
2675 	/*
2676 	 * The page daemon should avoid creating extra memory pressure since its
2677 	 * main purpose is to replenish the store of free pages.
2678 	 */
2679 	if (vmd->vmd_severeset || curproc == pageproc ||
2680 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2681 		return (0);
2682 	domain = vmd->vmd_domain;
2683 	vm_domain_free_lock(vmd);
2684 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2685 	    (vm_page_t *)store);
2686 	vm_domain_free_unlock(vmd);
2687 	if (cnt != i)
2688 		vm_domain_freecnt_inc(vmd, cnt - i);
2689 
2690 	return (i);
2691 }
2692 
2693 static void
2694 vm_page_zone_release(void *arg, void **store, int cnt)
2695 {
2696 	struct vm_domain *vmd;
2697 	struct vm_pgcache *pgcache;
2698 	vm_page_t m;
2699 	int i;
2700 
2701 	pgcache = arg;
2702 	vmd = VM_DOMAIN(pgcache->domain);
2703 	vm_domain_free_lock(vmd);
2704 	for (i = 0; i < cnt; i++) {
2705 		m = (vm_page_t)store[i];
2706 		vm_phys_free_pages(m, 0);
2707 	}
2708 	vm_domain_free_unlock(vmd);
2709 	vm_domain_freecnt_inc(vmd, cnt);
2710 }
2711 
2712 #define	VPSC_ANY	0	/* No restrictions. */
2713 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2714 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2715 
2716 /*
2717  *	vm_page_scan_contig:
2718  *
2719  *	Scan vm_page_array[] between the specified entries "m_start" and
2720  *	"m_end" for a run of contiguous physical pages that satisfy the
2721  *	specified conditions, and return the lowest page in the run.  The
2722  *	specified "alignment" determines the alignment of the lowest physical
2723  *	page in the run.  If the specified "boundary" is non-zero, then the
2724  *	run of physical pages cannot span a physical address that is a
2725  *	multiple of "boundary".
2726  *
2727  *	"m_end" is never dereferenced, so it need not point to a vm_page
2728  *	structure within vm_page_array[].
2729  *
2730  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2731  *	span a hole (or discontiguity) in the physical address space.  Both
2732  *	"alignment" and "boundary" must be a power of two.
2733  */
2734 static vm_page_t
2735 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2736     u_long alignment, vm_paddr_t boundary, int options)
2737 {
2738 	vm_object_t object;
2739 	vm_paddr_t pa;
2740 	vm_page_t m, m_run;
2741 #if VM_NRESERVLEVEL > 0
2742 	int level;
2743 #endif
2744 	int m_inc, order, run_ext, run_len;
2745 
2746 	KASSERT(npages > 0, ("npages is 0"));
2747 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2748 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2749 	m_run = NULL;
2750 	run_len = 0;
2751 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2752 		KASSERT((m->flags & PG_MARKER) == 0,
2753 		    ("page %p is PG_MARKER", m));
2754 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2755 		    ("fictitious page %p has invalid ref count", m));
2756 
2757 		/*
2758 		 * If the current page would be the start of a run, check its
2759 		 * physical address against the end, alignment, and boundary
2760 		 * conditions.  If it doesn't satisfy these conditions, either
2761 		 * terminate the scan or advance to the next page that
2762 		 * satisfies the failed condition.
2763 		 */
2764 		if (run_len == 0) {
2765 			KASSERT(m_run == NULL, ("m_run != NULL"));
2766 			if (m + npages > m_end)
2767 				break;
2768 			pa = VM_PAGE_TO_PHYS(m);
2769 			if (!vm_addr_align_ok(pa, alignment)) {
2770 				m_inc = atop(roundup2(pa, alignment) - pa);
2771 				continue;
2772 			}
2773 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2774 				m_inc = atop(roundup2(pa, boundary) - pa);
2775 				continue;
2776 			}
2777 		} else
2778 			KASSERT(m_run != NULL, ("m_run == NULL"));
2779 
2780 retry:
2781 		m_inc = 1;
2782 		if (vm_page_wired(m))
2783 			run_ext = 0;
2784 #if VM_NRESERVLEVEL > 0
2785 		else if ((level = vm_reserv_level(m)) >= 0 &&
2786 		    (options & VPSC_NORESERV) != 0) {
2787 			run_ext = 0;
2788 			/* Advance to the end of the reservation. */
2789 			pa = VM_PAGE_TO_PHYS(m);
2790 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2791 			    pa);
2792 		}
2793 #endif
2794 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2795 			/*
2796 			 * The page is considered eligible for relocation if
2797 			 * and only if it could be laundered or reclaimed by
2798 			 * the page daemon.
2799 			 */
2800 			VM_OBJECT_RLOCK(object);
2801 			if (object != m->object) {
2802 				VM_OBJECT_RUNLOCK(object);
2803 				goto retry;
2804 			}
2805 			/* Don't care: PG_NODUMP, PG_ZERO. */
2806 			if ((object->flags & OBJ_SWAP) == 0 &&
2807 			    object->type != OBJT_VNODE) {
2808 				run_ext = 0;
2809 #if VM_NRESERVLEVEL > 0
2810 			} else if ((options & VPSC_NOSUPER) != 0 &&
2811 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2812 				run_ext = 0;
2813 				/* Advance to the end of the superpage. */
2814 				pa = VM_PAGE_TO_PHYS(m);
2815 				m_inc = atop(roundup2(pa + 1,
2816 				    vm_reserv_size(level)) - pa);
2817 #endif
2818 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2819 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2820 				/*
2821 				 * The page is allocated but eligible for
2822 				 * relocation.  Extend the current run by one
2823 				 * page.
2824 				 */
2825 				KASSERT(pmap_page_get_memattr(m) ==
2826 				    VM_MEMATTR_DEFAULT,
2827 				    ("page %p has an unexpected memattr", m));
2828 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2829 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2830 				    ("page %p has unexpected oflags", m));
2831 				/* Don't care: PGA_NOSYNC. */
2832 				run_ext = 1;
2833 			} else
2834 				run_ext = 0;
2835 			VM_OBJECT_RUNLOCK(object);
2836 #if VM_NRESERVLEVEL > 0
2837 		} else if (level >= 0) {
2838 			/*
2839 			 * The page is reserved but not yet allocated.  In
2840 			 * other words, it is still free.  Extend the current
2841 			 * run by one page.
2842 			 */
2843 			run_ext = 1;
2844 #endif
2845 		} else if ((order = m->order) < VM_NFREEORDER) {
2846 			/*
2847 			 * The page is enqueued in the physical memory
2848 			 * allocator's free page queues.  Moreover, it is the
2849 			 * first page in a power-of-two-sized run of
2850 			 * contiguous free pages.  Add these pages to the end
2851 			 * of the current run, and jump ahead.
2852 			 */
2853 			run_ext = 1 << order;
2854 			m_inc = 1 << order;
2855 		} else {
2856 			/*
2857 			 * Skip the page for one of the following reasons: (1)
2858 			 * It is enqueued in the physical memory allocator's
2859 			 * free page queues.  However, it is not the first
2860 			 * page in a run of contiguous free pages.  (This case
2861 			 * rarely occurs because the scan is performed in
2862 			 * ascending order.) (2) It is not reserved, and it is
2863 			 * transitioning from free to allocated.  (Conversely,
2864 			 * the transition from allocated to free for managed
2865 			 * pages is blocked by the page busy lock.) (3) It is
2866 			 * allocated but not contained by an object and not
2867 			 * wired, e.g., allocated by Xen's balloon driver.
2868 			 */
2869 			run_ext = 0;
2870 		}
2871 
2872 		/*
2873 		 * Extend or reset the current run of pages.
2874 		 */
2875 		if (run_ext > 0) {
2876 			if (run_len == 0)
2877 				m_run = m;
2878 			run_len += run_ext;
2879 		} else {
2880 			if (run_len > 0) {
2881 				m_run = NULL;
2882 				run_len = 0;
2883 			}
2884 		}
2885 	}
2886 	if (run_len >= npages)
2887 		return (m_run);
2888 	return (NULL);
2889 }
2890 
2891 /*
2892  *	vm_page_reclaim_run:
2893  *
2894  *	Try to relocate each of the allocated virtual pages within the
2895  *	specified run of physical pages to a new physical address.  Free the
2896  *	physical pages underlying the relocated virtual pages.  A virtual page
2897  *	is relocatable if and only if it could be laundered or reclaimed by
2898  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2899  *	physical address above "high".
2900  *
2901  *	Returns 0 if every physical page within the run was already free or
2902  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2903  *	value indicating why the last attempt to relocate a virtual page was
2904  *	unsuccessful.
2905  *
2906  *	"req_class" must be an allocation class.
2907  */
2908 static int
2909 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2910     vm_paddr_t high)
2911 {
2912 	struct vm_domain *vmd;
2913 	struct spglist free;
2914 	vm_object_t object;
2915 	vm_paddr_t pa;
2916 	vm_page_t m, m_end, m_new;
2917 	int error, order, req;
2918 
2919 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2920 	    ("req_class is not an allocation class"));
2921 	SLIST_INIT(&free);
2922 	error = 0;
2923 	m = m_run;
2924 	m_end = m_run + npages;
2925 	for (; error == 0 && m < m_end; m++) {
2926 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2927 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2928 
2929 		/*
2930 		 * Racily check for wirings.  Races are handled once the object
2931 		 * lock is held and the page is unmapped.
2932 		 */
2933 		if (vm_page_wired(m))
2934 			error = EBUSY;
2935 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2936 			/*
2937 			 * The page is relocated if and only if it could be
2938 			 * laundered or reclaimed by the page daemon.
2939 			 */
2940 			VM_OBJECT_WLOCK(object);
2941 			/* Don't care: PG_NODUMP, PG_ZERO. */
2942 			if (m->object != object ||
2943 			    ((object->flags & OBJ_SWAP) == 0 &&
2944 			    object->type != OBJT_VNODE))
2945 				error = EINVAL;
2946 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2947 				error = EINVAL;
2948 			else if (vm_page_queue(m) != PQ_NONE &&
2949 			    vm_page_tryxbusy(m) != 0) {
2950 				if (vm_page_wired(m)) {
2951 					vm_page_xunbusy(m);
2952 					error = EBUSY;
2953 					goto unlock;
2954 				}
2955 				KASSERT(pmap_page_get_memattr(m) ==
2956 				    VM_MEMATTR_DEFAULT,
2957 				    ("page %p has an unexpected memattr", m));
2958 				KASSERT(m->oflags == 0,
2959 				    ("page %p has unexpected oflags", m));
2960 				/* Don't care: PGA_NOSYNC. */
2961 				if (!vm_page_none_valid(m)) {
2962 					/*
2963 					 * First, try to allocate a new page
2964 					 * that is above "high".  Failing
2965 					 * that, try to allocate a new page
2966 					 * that is below "m_run".  Allocate
2967 					 * the new page between the end of
2968 					 * "m_run" and "high" only as a last
2969 					 * resort.
2970 					 */
2971 					req = req_class;
2972 					if ((m->flags & PG_NODUMP) != 0)
2973 						req |= VM_ALLOC_NODUMP;
2974 					if (trunc_page(high) !=
2975 					    ~(vm_paddr_t)PAGE_MASK) {
2976 						m_new =
2977 						    vm_page_alloc_noobj_contig(
2978 						    req, 1, round_page(high),
2979 						    ~(vm_paddr_t)0, PAGE_SIZE,
2980 						    0, VM_MEMATTR_DEFAULT);
2981 					} else
2982 						m_new = NULL;
2983 					if (m_new == NULL) {
2984 						pa = VM_PAGE_TO_PHYS(m_run);
2985 						m_new =
2986 						    vm_page_alloc_noobj_contig(
2987 						    req, 1, 0, pa - 1,
2988 						    PAGE_SIZE, 0,
2989 						    VM_MEMATTR_DEFAULT);
2990 					}
2991 					if (m_new == NULL) {
2992 						pa += ptoa(npages);
2993 						m_new =
2994 						    vm_page_alloc_noobj_contig(
2995 						    req, 1, pa, high, PAGE_SIZE,
2996 						    0, VM_MEMATTR_DEFAULT);
2997 					}
2998 					if (m_new == NULL) {
2999 						vm_page_xunbusy(m);
3000 						error = ENOMEM;
3001 						goto unlock;
3002 					}
3003 
3004 					/*
3005 					 * Unmap the page and check for new
3006 					 * wirings that may have been acquired
3007 					 * through a pmap lookup.
3008 					 */
3009 					if (object->ref_count != 0 &&
3010 					    !vm_page_try_remove_all(m)) {
3011 						vm_page_xunbusy(m);
3012 						vm_page_free(m_new);
3013 						error = EBUSY;
3014 						goto unlock;
3015 					}
3016 
3017 					/*
3018 					 * Replace "m" with the new page.  For
3019 					 * vm_page_replace(), "m" must be busy
3020 					 * and dequeued.  Finally, change "m"
3021 					 * as if vm_page_free() was called.
3022 					 */
3023 					m_new->a.flags = m->a.flags &
3024 					    ~PGA_QUEUE_STATE_MASK;
3025 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3026 					    ("page %p is managed", m_new));
3027 					m_new->oflags = 0;
3028 					pmap_copy_page(m, m_new);
3029 					m_new->valid = m->valid;
3030 					m_new->dirty = m->dirty;
3031 					m->flags &= ~PG_ZERO;
3032 					vm_page_dequeue(m);
3033 					if (vm_page_replace_hold(m_new, object,
3034 					    m->pindex, m) &&
3035 					    vm_page_free_prep(m))
3036 						SLIST_INSERT_HEAD(&free, m,
3037 						    plinks.s.ss);
3038 
3039 					/*
3040 					 * The new page must be deactivated
3041 					 * before the object is unlocked.
3042 					 */
3043 					vm_page_deactivate(m_new);
3044 				} else {
3045 					m->flags &= ~PG_ZERO;
3046 					vm_page_dequeue(m);
3047 					if (vm_page_free_prep(m))
3048 						SLIST_INSERT_HEAD(&free, m,
3049 						    plinks.s.ss);
3050 					KASSERT(m->dirty == 0,
3051 					    ("page %p is dirty", m));
3052 				}
3053 			} else
3054 				error = EBUSY;
3055 unlock:
3056 			VM_OBJECT_WUNLOCK(object);
3057 		} else {
3058 			MPASS(vm_page_domain(m) == domain);
3059 			vmd = VM_DOMAIN(domain);
3060 			vm_domain_free_lock(vmd);
3061 			order = m->order;
3062 			if (order < VM_NFREEORDER) {
3063 				/*
3064 				 * The page is enqueued in the physical memory
3065 				 * allocator's free page queues.  Moreover, it
3066 				 * is the first page in a power-of-two-sized
3067 				 * run of contiguous free pages.  Jump ahead
3068 				 * to the last page within that run, and
3069 				 * continue from there.
3070 				 */
3071 				m += (1 << order) - 1;
3072 			}
3073 #if VM_NRESERVLEVEL > 0
3074 			else if (vm_reserv_is_page_free(m))
3075 				order = 0;
3076 #endif
3077 			vm_domain_free_unlock(vmd);
3078 			if (order == VM_NFREEORDER)
3079 				error = EINVAL;
3080 		}
3081 	}
3082 	if ((m = SLIST_FIRST(&free)) != NULL) {
3083 		int cnt;
3084 
3085 		vmd = VM_DOMAIN(domain);
3086 		cnt = 0;
3087 		vm_domain_free_lock(vmd);
3088 		do {
3089 			MPASS(vm_page_domain(m) == domain);
3090 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3091 			vm_phys_free_pages(m, 0);
3092 			cnt++;
3093 		} while ((m = SLIST_FIRST(&free)) != NULL);
3094 		vm_domain_free_unlock(vmd);
3095 		vm_domain_freecnt_inc(vmd, cnt);
3096 	}
3097 	return (error);
3098 }
3099 
3100 #define	NRUNS	16
3101 
3102 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3103 
3104 #define	MIN_RECLAIM	8
3105 
3106 /*
3107  *	vm_page_reclaim_contig:
3108  *
3109  *	Reclaim allocated, contiguous physical memory satisfying the specified
3110  *	conditions by relocating the virtual pages using that physical memory.
3111  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3112  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3113  *	currently able to reclaim the requested number of pages.  Since
3114  *	relocation requires the allocation of physical pages, reclamation may
3115  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3116  *	fails in this manner, callers are expected to perform vm_wait() before
3117  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3118  *
3119  *	The caller must always specify an allocation class through "req".
3120  *
3121  *	allocation classes:
3122  *	VM_ALLOC_NORMAL		normal process request
3123  *	VM_ALLOC_SYSTEM		system *really* needs a page
3124  *	VM_ALLOC_INTERRUPT	interrupt time request
3125  *
3126  *	The optional allocation flags are ignored.
3127  *
3128  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3129  *	must be a power of two.
3130  */
3131 int
3132 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3133     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3134     int desired_runs)
3135 {
3136 	struct vm_domain *vmd;
3137 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3138 	u_long count, minalign, reclaimed;
3139 	int error, i, min_reclaim, nruns, options, req_class;
3140 	int segind, start_segind;
3141 	int ret;
3142 
3143 	KASSERT(npages > 0, ("npages is 0"));
3144 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3145 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3146 
3147 	ret = ENOMEM;
3148 
3149 	/*
3150 	 * If the caller wants to reclaim multiple runs, try to allocate
3151 	 * space to store the runs.  If that fails, fall back to the old
3152 	 * behavior of just reclaiming MIN_RECLAIM pages.
3153 	 */
3154 	if (desired_runs > 1)
3155 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3156 		    M_TEMP, M_NOWAIT);
3157 	else
3158 		m_runs = NULL;
3159 
3160 	if (m_runs == NULL) {
3161 		m_runs = _m_runs;
3162 		nruns = NRUNS;
3163 	} else {
3164 		nruns = NRUNS + desired_runs - 1;
3165 	}
3166 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3167 
3168 	/*
3169 	 * The caller will attempt an allocation after some runs have been
3170 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3171 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3172 	 * power of two or maximum chunk size, and ensure that each run is
3173 	 * suitably aligned.
3174 	 */
3175 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3176 	npages = roundup2(npages, minalign);
3177 	if (alignment < ptoa(minalign))
3178 		alignment = ptoa(minalign);
3179 
3180 	/*
3181 	 * The page daemon is allowed to dig deeper into the free page list.
3182 	 */
3183 	req_class = req & VM_ALLOC_CLASS_MASK;
3184 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3185 		req_class = VM_ALLOC_SYSTEM;
3186 
3187 	start_segind = vm_phys_lookup_segind(low);
3188 
3189 	/*
3190 	 * Return if the number of free pages cannot satisfy the requested
3191 	 * allocation.
3192 	 */
3193 	vmd = VM_DOMAIN(domain);
3194 	count = vmd->vmd_free_count;
3195 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3196 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3197 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3198 		goto done;
3199 
3200 	/*
3201 	 * Scan up to three times, relaxing the restrictions ("options") on
3202 	 * the reclamation of reservations and superpages each time.
3203 	 */
3204 	for (options = VPSC_NORESERV;;) {
3205 		bool phys_range_exists = false;
3206 
3207 		/*
3208 		 * Find the highest runs that satisfy the given constraints
3209 		 * and restrictions, and record them in "m_runs".
3210 		 */
3211 		count = 0;
3212 		segind = start_segind;
3213 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3214 		    npages, low, high)) != -1) {
3215 			phys_range_exists = true;
3216 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3217 			    bounds[1], alignment, boundary, options))) {
3218 				bounds[0] = m_run + npages;
3219 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3220 				count++;
3221 			}
3222 			segind++;
3223 		}
3224 
3225 		if (!phys_range_exists) {
3226 			ret = ERANGE;
3227 			goto done;
3228 		}
3229 
3230 		/*
3231 		 * Reclaim the highest runs in LIFO (descending) order until
3232 		 * the number of reclaimed pages, "reclaimed", is at least
3233 		 * "min_reclaim".  Reset "reclaimed" each time because each
3234 		 * reclamation is idempotent, and runs will (likely) recur
3235 		 * from one scan to the next as restrictions are relaxed.
3236 		 */
3237 		reclaimed = 0;
3238 		for (i = 0; count > 0 && i < nruns; i++) {
3239 			count--;
3240 			m_run = m_runs[RUN_INDEX(count, nruns)];
3241 			error = vm_page_reclaim_run(req_class, domain, npages,
3242 			    m_run, high);
3243 			if (error == 0) {
3244 				reclaimed += npages;
3245 				if (reclaimed >= min_reclaim) {
3246 					ret = 0;
3247 					goto done;
3248 				}
3249 			}
3250 		}
3251 
3252 		/*
3253 		 * Either relax the restrictions on the next scan or return if
3254 		 * the last scan had no restrictions.
3255 		 */
3256 		if (options == VPSC_NORESERV)
3257 			options = VPSC_NOSUPER;
3258 		else if (options == VPSC_NOSUPER)
3259 			options = VPSC_ANY;
3260 		else if (options == VPSC_ANY) {
3261 			if (reclaimed != 0)
3262 				ret = 0;
3263 			goto done;
3264 		}
3265 	}
3266 done:
3267 	if (m_runs != _m_runs)
3268 		free(m_runs, M_TEMP);
3269 	return (ret);
3270 }
3271 
3272 int
3273 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3274     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3275 {
3276 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3277 	    alignment, boundary, 1));
3278 }
3279 
3280 int
3281 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3282     u_long alignment, vm_paddr_t boundary)
3283 {
3284 	struct vm_domainset_iter di;
3285 	int domain, ret, status;
3286 
3287 	ret = ERANGE;
3288 
3289 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3290 	do {
3291 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3292 		    high, alignment, boundary);
3293 		if (status == 0)
3294 			return (0);
3295 		else if (status == ERANGE)
3296 			vm_domainset_iter_ignore(&di, domain);
3297 		else {
3298 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3299 			    "from vm_page_reclaim_contig_domain()", status));
3300 			ret = ENOMEM;
3301 		}
3302 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3303 
3304 	return (ret);
3305 }
3306 
3307 /*
3308  * Set the domain in the appropriate page level domainset.
3309  */
3310 void
3311 vm_domain_set(struct vm_domain *vmd)
3312 {
3313 
3314 	mtx_lock(&vm_domainset_lock);
3315 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3316 		vmd->vmd_minset = 1;
3317 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3318 	}
3319 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3320 		vmd->vmd_severeset = 1;
3321 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3322 	}
3323 	mtx_unlock(&vm_domainset_lock);
3324 }
3325 
3326 /*
3327  * Clear the domain from the appropriate page level domainset.
3328  */
3329 void
3330 vm_domain_clear(struct vm_domain *vmd)
3331 {
3332 
3333 	mtx_lock(&vm_domainset_lock);
3334 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3335 		vmd->vmd_minset = 0;
3336 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3337 		if (vm_min_waiters != 0) {
3338 			vm_min_waiters = 0;
3339 			wakeup(&vm_min_domains);
3340 		}
3341 	}
3342 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3343 		vmd->vmd_severeset = 0;
3344 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3345 		if (vm_severe_waiters != 0) {
3346 			vm_severe_waiters = 0;
3347 			wakeup(&vm_severe_domains);
3348 		}
3349 	}
3350 
3351 	/*
3352 	 * If pageout daemon needs pages, then tell it that there are
3353 	 * some free.
3354 	 */
3355 	if (vmd->vmd_pageout_pages_needed &&
3356 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3357 		wakeup(&vmd->vmd_pageout_pages_needed);
3358 		vmd->vmd_pageout_pages_needed = 0;
3359 	}
3360 
3361 	/* See comments in vm_wait_doms(). */
3362 	if (vm_pageproc_waiters) {
3363 		vm_pageproc_waiters = 0;
3364 		wakeup(&vm_pageproc_waiters);
3365 	}
3366 	mtx_unlock(&vm_domainset_lock);
3367 }
3368 
3369 /*
3370  * Wait for free pages to exceed the min threshold globally.
3371  */
3372 void
3373 vm_wait_min(void)
3374 {
3375 
3376 	mtx_lock(&vm_domainset_lock);
3377 	while (vm_page_count_min()) {
3378 		vm_min_waiters++;
3379 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3380 	}
3381 	mtx_unlock(&vm_domainset_lock);
3382 }
3383 
3384 /*
3385  * Wait for free pages to exceed the severe threshold globally.
3386  */
3387 void
3388 vm_wait_severe(void)
3389 {
3390 
3391 	mtx_lock(&vm_domainset_lock);
3392 	while (vm_page_count_severe()) {
3393 		vm_severe_waiters++;
3394 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3395 		    "vmwait", 0);
3396 	}
3397 	mtx_unlock(&vm_domainset_lock);
3398 }
3399 
3400 u_int
3401 vm_wait_count(void)
3402 {
3403 
3404 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3405 }
3406 
3407 int
3408 vm_wait_doms(const domainset_t *wdoms, int mflags)
3409 {
3410 	int error;
3411 
3412 	error = 0;
3413 
3414 	/*
3415 	 * We use racey wakeup synchronization to avoid expensive global
3416 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3417 	 * To handle this, we only sleep for one tick in this instance.  It
3418 	 * is expected that most allocations for the pageproc will come from
3419 	 * kmem or vm_page_grab* which will use the more specific and
3420 	 * race-free vm_wait_domain().
3421 	 */
3422 	if (curproc == pageproc) {
3423 		mtx_lock(&vm_domainset_lock);
3424 		vm_pageproc_waiters++;
3425 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3426 		    PVM | PDROP | mflags, "pageprocwait", 1);
3427 	} else {
3428 		/*
3429 		 * XXX Ideally we would wait only until the allocation could
3430 		 * be satisfied.  This condition can cause new allocators to
3431 		 * consume all freed pages while old allocators wait.
3432 		 */
3433 		mtx_lock(&vm_domainset_lock);
3434 		if (vm_page_count_min_set(wdoms)) {
3435 			if (pageproc == NULL)
3436 				panic("vm_wait in early boot");
3437 			vm_min_waiters++;
3438 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3439 			    PVM | PDROP | mflags, "vmwait", 0);
3440 		} else
3441 			mtx_unlock(&vm_domainset_lock);
3442 	}
3443 	return (error);
3444 }
3445 
3446 /*
3447  *	vm_wait_domain:
3448  *
3449  *	Sleep until free pages are available for allocation.
3450  *	- Called in various places after failed memory allocations.
3451  */
3452 void
3453 vm_wait_domain(int domain)
3454 {
3455 	struct vm_domain *vmd;
3456 	domainset_t wdom;
3457 
3458 	vmd = VM_DOMAIN(domain);
3459 	vm_domain_free_assert_unlocked(vmd);
3460 
3461 	if (curproc == pageproc) {
3462 		mtx_lock(&vm_domainset_lock);
3463 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3464 			vmd->vmd_pageout_pages_needed = 1;
3465 			msleep(&vmd->vmd_pageout_pages_needed,
3466 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3467 		} else
3468 			mtx_unlock(&vm_domainset_lock);
3469 	} else {
3470 		DOMAINSET_ZERO(&wdom);
3471 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3472 		vm_wait_doms(&wdom, 0);
3473 	}
3474 }
3475 
3476 static int
3477 vm_wait_flags(vm_object_t obj, int mflags)
3478 {
3479 	struct domainset *d;
3480 
3481 	d = NULL;
3482 
3483 	/*
3484 	 * Carefully fetch pointers only once: the struct domainset
3485 	 * itself is ummutable but the pointer might change.
3486 	 */
3487 	if (obj != NULL)
3488 		d = obj->domain.dr_policy;
3489 	if (d == NULL)
3490 		d = curthread->td_domain.dr_policy;
3491 
3492 	return (vm_wait_doms(&d->ds_mask, mflags));
3493 }
3494 
3495 /*
3496  *	vm_wait:
3497  *
3498  *	Sleep until free pages are available for allocation in the
3499  *	affinity domains of the obj.  If obj is NULL, the domain set
3500  *	for the calling thread is used.
3501  *	Called in various places after failed memory allocations.
3502  */
3503 void
3504 vm_wait(vm_object_t obj)
3505 {
3506 	(void)vm_wait_flags(obj, 0);
3507 }
3508 
3509 int
3510 vm_wait_intr(vm_object_t obj)
3511 {
3512 	return (vm_wait_flags(obj, PCATCH));
3513 }
3514 
3515 /*
3516  *	vm_domain_alloc_fail:
3517  *
3518  *	Called when a page allocation function fails.  Informs the
3519  *	pagedaemon and performs the requested wait.  Requires the
3520  *	domain_free and object lock on entry.  Returns with the
3521  *	object lock held and free lock released.  Returns an error when
3522  *	retry is necessary.
3523  *
3524  */
3525 static int
3526 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3527 {
3528 
3529 	vm_domain_free_assert_unlocked(vmd);
3530 
3531 	atomic_add_int(&vmd->vmd_pageout_deficit,
3532 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3533 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3534 		if (object != NULL)
3535 			VM_OBJECT_WUNLOCK(object);
3536 		vm_wait_domain(vmd->vmd_domain);
3537 		if (object != NULL)
3538 			VM_OBJECT_WLOCK(object);
3539 		if (req & VM_ALLOC_WAITOK)
3540 			return (EAGAIN);
3541 	}
3542 
3543 	return (0);
3544 }
3545 
3546 /*
3547  *	vm_waitpfault:
3548  *
3549  *	Sleep until free pages are available for allocation.
3550  *	- Called only in vm_fault so that processes page faulting
3551  *	  can be easily tracked.
3552  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3553  *	  processes will be able to grab memory first.  Do not change
3554  *	  this balance without careful testing first.
3555  */
3556 void
3557 vm_waitpfault(struct domainset *dset, int timo)
3558 {
3559 
3560 	/*
3561 	 * XXX Ideally we would wait only until the allocation could
3562 	 * be satisfied.  This condition can cause new allocators to
3563 	 * consume all freed pages while old allocators wait.
3564 	 */
3565 	mtx_lock(&vm_domainset_lock);
3566 	if (vm_page_count_min_set(&dset->ds_mask)) {
3567 		vm_min_waiters++;
3568 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3569 		    "pfault", timo);
3570 	} else
3571 		mtx_unlock(&vm_domainset_lock);
3572 }
3573 
3574 static struct vm_pagequeue *
3575 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3576 {
3577 
3578 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3579 }
3580 
3581 #ifdef INVARIANTS
3582 static struct vm_pagequeue *
3583 vm_page_pagequeue(vm_page_t m)
3584 {
3585 
3586 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3587 }
3588 #endif
3589 
3590 static __always_inline bool
3591 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3592 {
3593 	vm_page_astate_t tmp;
3594 
3595 	tmp = *old;
3596 	do {
3597 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3598 			return (true);
3599 		counter_u64_add(pqstate_commit_retries, 1);
3600 	} while (old->_bits == tmp._bits);
3601 
3602 	return (false);
3603 }
3604 
3605 /*
3606  * Do the work of committing a queue state update that moves the page out of
3607  * its current queue.
3608  */
3609 static bool
3610 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3611     vm_page_astate_t *old, vm_page_astate_t new)
3612 {
3613 	vm_page_t next;
3614 
3615 	vm_pagequeue_assert_locked(pq);
3616 	KASSERT(vm_page_pagequeue(m) == pq,
3617 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3618 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3619 	    ("%s: invalid queue indices %d %d",
3620 	    __func__, old->queue, new.queue));
3621 
3622 	/*
3623 	 * Once the queue index of the page changes there is nothing
3624 	 * synchronizing with further updates to the page's physical
3625 	 * queue state.  Therefore we must speculatively remove the page
3626 	 * from the queue now and be prepared to roll back if the queue
3627 	 * state update fails.  If the page is not physically enqueued then
3628 	 * we just update its queue index.
3629 	 */
3630 	if ((old->flags & PGA_ENQUEUED) != 0) {
3631 		new.flags &= ~PGA_ENQUEUED;
3632 		next = TAILQ_NEXT(m, plinks.q);
3633 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3634 		vm_pagequeue_cnt_dec(pq);
3635 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3636 			if (next == NULL)
3637 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3638 			else
3639 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3640 			vm_pagequeue_cnt_inc(pq);
3641 			return (false);
3642 		} else {
3643 			return (true);
3644 		}
3645 	} else {
3646 		return (vm_page_pqstate_fcmpset(m, old, new));
3647 	}
3648 }
3649 
3650 static bool
3651 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3652     vm_page_astate_t new)
3653 {
3654 	struct vm_pagequeue *pq;
3655 	vm_page_astate_t as;
3656 	bool ret;
3657 
3658 	pq = _vm_page_pagequeue(m, old->queue);
3659 
3660 	/*
3661 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3662 	 * corresponding page queue lock is held.
3663 	 */
3664 	vm_pagequeue_lock(pq);
3665 	as = vm_page_astate_load(m);
3666 	if (__predict_false(as._bits != old->_bits)) {
3667 		*old = as;
3668 		ret = false;
3669 	} else {
3670 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3671 	}
3672 	vm_pagequeue_unlock(pq);
3673 	return (ret);
3674 }
3675 
3676 /*
3677  * Commit a queue state update that enqueues or requeues a page.
3678  */
3679 static bool
3680 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3681     vm_page_astate_t *old, vm_page_astate_t new)
3682 {
3683 	struct vm_domain *vmd;
3684 
3685 	vm_pagequeue_assert_locked(pq);
3686 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3687 	    ("%s: invalid queue indices %d %d",
3688 	    __func__, old->queue, new.queue));
3689 
3690 	new.flags |= PGA_ENQUEUED;
3691 	if (!vm_page_pqstate_fcmpset(m, old, new))
3692 		return (false);
3693 
3694 	if ((old->flags & PGA_ENQUEUED) != 0)
3695 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3696 	else
3697 		vm_pagequeue_cnt_inc(pq);
3698 
3699 	/*
3700 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3701 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3702 	 * applied, even if it was set first.
3703 	 */
3704 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3705 		vmd = vm_pagequeue_domain(m);
3706 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3707 		    ("%s: invalid page queue for page %p", __func__, m));
3708 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3709 	} else {
3710 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3711 	}
3712 	return (true);
3713 }
3714 
3715 /*
3716  * Commit a queue state update that encodes a request for a deferred queue
3717  * operation.
3718  */
3719 static bool
3720 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3721     vm_page_astate_t new)
3722 {
3723 
3724 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3725 	    ("%s: invalid state, queue %d flags %x",
3726 	    __func__, new.queue, new.flags));
3727 
3728 	if (old->_bits != new._bits &&
3729 	    !vm_page_pqstate_fcmpset(m, old, new))
3730 		return (false);
3731 	vm_page_pqbatch_submit(m, new.queue);
3732 	return (true);
3733 }
3734 
3735 /*
3736  * A generic queue state update function.  This handles more cases than the
3737  * specialized functions above.
3738  */
3739 bool
3740 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3741 {
3742 
3743 	if (old->_bits == new._bits)
3744 		return (true);
3745 
3746 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3747 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3748 			return (false);
3749 		if (new.queue != PQ_NONE)
3750 			vm_page_pqbatch_submit(m, new.queue);
3751 	} else {
3752 		if (!vm_page_pqstate_fcmpset(m, old, new))
3753 			return (false);
3754 		if (new.queue != PQ_NONE &&
3755 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3756 			vm_page_pqbatch_submit(m, new.queue);
3757 	}
3758 	return (true);
3759 }
3760 
3761 /*
3762  * Apply deferred queue state updates to a page.
3763  */
3764 static inline void
3765 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3766 {
3767 	vm_page_astate_t new, old;
3768 
3769 	CRITICAL_ASSERT(curthread);
3770 	vm_pagequeue_assert_locked(pq);
3771 	KASSERT(queue < PQ_COUNT,
3772 	    ("%s: invalid queue index %d", __func__, queue));
3773 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3774 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3775 
3776 	for (old = vm_page_astate_load(m);;) {
3777 		if (__predict_false(old.queue != queue ||
3778 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3779 			counter_u64_add(queue_nops, 1);
3780 			break;
3781 		}
3782 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3783 		    ("%s: page %p is unmanaged", __func__, m));
3784 
3785 		new = old;
3786 		if ((old.flags & PGA_DEQUEUE) != 0) {
3787 			new.flags &= ~PGA_QUEUE_OP_MASK;
3788 			new.queue = PQ_NONE;
3789 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3790 			    m, &old, new))) {
3791 				counter_u64_add(queue_ops, 1);
3792 				break;
3793 			}
3794 		} else {
3795 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3796 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3797 			    m, &old, new))) {
3798 				counter_u64_add(queue_ops, 1);
3799 				break;
3800 			}
3801 		}
3802 	}
3803 }
3804 
3805 static void
3806 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3807     uint8_t queue)
3808 {
3809 	int i;
3810 
3811 	for (i = 0; i < bq->bq_cnt; i++)
3812 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3813 	vm_batchqueue_init(bq);
3814 }
3815 
3816 /*
3817  *	vm_page_pqbatch_submit:		[ internal use only ]
3818  *
3819  *	Enqueue a page in the specified page queue's batched work queue.
3820  *	The caller must have encoded the requested operation in the page
3821  *	structure's a.flags field.
3822  */
3823 void
3824 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3825 {
3826 	struct vm_batchqueue *bq;
3827 	struct vm_pagequeue *pq;
3828 	int domain, slots_remaining;
3829 
3830 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3831 
3832 	domain = vm_page_domain(m);
3833 	critical_enter();
3834 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3835 	slots_remaining = vm_batchqueue_insert(bq, m);
3836 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3837 		/* keep building the bq */
3838 		critical_exit();
3839 		return;
3840 	} else if (slots_remaining > 0 ) {
3841 		/* Try to process the bq if we can get the lock */
3842 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3843 		if (vm_pagequeue_trylock(pq)) {
3844 			vm_pqbatch_process(pq, bq, queue);
3845 			vm_pagequeue_unlock(pq);
3846 		}
3847 		critical_exit();
3848 		return;
3849 	}
3850 	critical_exit();
3851 
3852 	/* if we make it here, the bq is full so wait for the lock */
3853 
3854 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3855 	vm_pagequeue_lock(pq);
3856 	critical_enter();
3857 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3858 	vm_pqbatch_process(pq, bq, queue);
3859 	vm_pqbatch_process_page(pq, m, queue);
3860 	vm_pagequeue_unlock(pq);
3861 	critical_exit();
3862 }
3863 
3864 /*
3865  *	vm_page_pqbatch_drain:		[ internal use only ]
3866  *
3867  *	Force all per-CPU page queue batch queues to be drained.  This is
3868  *	intended for use in severe memory shortages, to ensure that pages
3869  *	do not remain stuck in the batch queues.
3870  */
3871 void
3872 vm_page_pqbatch_drain(void)
3873 {
3874 	struct thread *td;
3875 	struct vm_domain *vmd;
3876 	struct vm_pagequeue *pq;
3877 	int cpu, domain, queue;
3878 
3879 	td = curthread;
3880 	CPU_FOREACH(cpu) {
3881 		thread_lock(td);
3882 		sched_bind(td, cpu);
3883 		thread_unlock(td);
3884 
3885 		for (domain = 0; domain < vm_ndomains; domain++) {
3886 			vmd = VM_DOMAIN(domain);
3887 			for (queue = 0; queue < PQ_COUNT; queue++) {
3888 				pq = &vmd->vmd_pagequeues[queue];
3889 				vm_pagequeue_lock(pq);
3890 				critical_enter();
3891 				vm_pqbatch_process(pq,
3892 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3893 				critical_exit();
3894 				vm_pagequeue_unlock(pq);
3895 			}
3896 		}
3897 	}
3898 	thread_lock(td);
3899 	sched_unbind(td);
3900 	thread_unlock(td);
3901 }
3902 
3903 /*
3904  *	vm_page_dequeue_deferred:	[ internal use only ]
3905  *
3906  *	Request removal of the given page from its current page
3907  *	queue.  Physical removal from the queue may be deferred
3908  *	indefinitely.
3909  */
3910 void
3911 vm_page_dequeue_deferred(vm_page_t m)
3912 {
3913 	vm_page_astate_t new, old;
3914 
3915 	old = vm_page_astate_load(m);
3916 	do {
3917 		if (old.queue == PQ_NONE) {
3918 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3919 			    ("%s: page %p has unexpected queue state",
3920 			    __func__, m));
3921 			break;
3922 		}
3923 		new = old;
3924 		new.flags |= PGA_DEQUEUE;
3925 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3926 }
3927 
3928 /*
3929  *	vm_page_dequeue:
3930  *
3931  *	Remove the page from whichever page queue it's in, if any, before
3932  *	returning.
3933  */
3934 void
3935 vm_page_dequeue(vm_page_t m)
3936 {
3937 	vm_page_astate_t new, old;
3938 
3939 	old = vm_page_astate_load(m);
3940 	do {
3941 		if (old.queue == PQ_NONE) {
3942 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3943 			    ("%s: page %p has unexpected queue state",
3944 			    __func__, m));
3945 			break;
3946 		}
3947 		new = old;
3948 		new.flags &= ~PGA_QUEUE_OP_MASK;
3949 		new.queue = PQ_NONE;
3950 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3951 
3952 }
3953 
3954 /*
3955  * Schedule the given page for insertion into the specified page queue.
3956  * Physical insertion of the page may be deferred indefinitely.
3957  */
3958 static void
3959 vm_page_enqueue(vm_page_t m, uint8_t queue)
3960 {
3961 
3962 	KASSERT(m->a.queue == PQ_NONE &&
3963 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3964 	    ("%s: page %p is already enqueued", __func__, m));
3965 	KASSERT(m->ref_count > 0,
3966 	    ("%s: page %p does not carry any references", __func__, m));
3967 
3968 	m->a.queue = queue;
3969 	if ((m->a.flags & PGA_REQUEUE) == 0)
3970 		vm_page_aflag_set(m, PGA_REQUEUE);
3971 	vm_page_pqbatch_submit(m, queue);
3972 }
3973 
3974 /*
3975  *	vm_page_free_prep:
3976  *
3977  *	Prepares the given page to be put on the free list,
3978  *	disassociating it from any VM object. The caller may return
3979  *	the page to the free list only if this function returns true.
3980  *
3981  *	The object, if it exists, must be locked, and then the page must
3982  *	be xbusy.  Otherwise the page must be not busied.  A managed
3983  *	page must be unmapped.
3984  */
3985 static bool
3986 vm_page_free_prep(vm_page_t m)
3987 {
3988 
3989 	/*
3990 	 * Synchronize with threads that have dropped a reference to this
3991 	 * page.
3992 	 */
3993 	atomic_thread_fence_acq();
3994 
3995 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3996 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3997 		uint64_t *p;
3998 		int i;
3999 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4000 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4001 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4002 			    m, i, (uintmax_t)*p));
4003 	}
4004 #endif
4005 	KASSERT((m->flags & PG_NOFREE) == 0,
4006 	    ("%s: attempting to free a PG_NOFREE page", __func__));
4007 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4008 		KASSERT(!pmap_page_is_mapped(m),
4009 		    ("vm_page_free_prep: freeing mapped page %p", m));
4010 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4011 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4012 	} else {
4013 		KASSERT(m->a.queue == PQ_NONE,
4014 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4015 	}
4016 	VM_CNT_INC(v_tfree);
4017 
4018 	if (m->object != NULL) {
4019 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
4020 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
4021 		    ("vm_page_free_prep: managed flag mismatch for page %p",
4022 		    m));
4023 		vm_page_assert_xbusied(m);
4024 
4025 		/*
4026 		 * The object reference can be released without an atomic
4027 		 * operation.
4028 		 */
4029 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4030 		    m->ref_count == VPRC_OBJREF,
4031 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
4032 		    m, m->ref_count));
4033 		vm_page_object_remove(m);
4034 		m->ref_count -= VPRC_OBJREF;
4035 	} else
4036 		vm_page_assert_unbusied(m);
4037 
4038 	vm_page_busy_free(m);
4039 
4040 	/*
4041 	 * If fictitious remove object association and
4042 	 * return.
4043 	 */
4044 	if ((m->flags & PG_FICTITIOUS) != 0) {
4045 		KASSERT(m->ref_count == 1,
4046 		    ("fictitious page %p is referenced", m));
4047 		KASSERT(m->a.queue == PQ_NONE,
4048 		    ("fictitious page %p is queued", m));
4049 		return (false);
4050 	}
4051 
4052 	/*
4053 	 * Pages need not be dequeued before they are returned to the physical
4054 	 * memory allocator, but they must at least be marked for a deferred
4055 	 * dequeue.
4056 	 */
4057 	if ((m->oflags & VPO_UNMANAGED) == 0)
4058 		vm_page_dequeue_deferred(m);
4059 
4060 	m->valid = 0;
4061 	vm_page_undirty(m);
4062 
4063 	if (m->ref_count != 0)
4064 		panic("vm_page_free_prep: page %p has references", m);
4065 
4066 	/*
4067 	 * Restore the default memory attribute to the page.
4068 	 */
4069 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4070 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4071 
4072 #if VM_NRESERVLEVEL > 0
4073 	/*
4074 	 * Determine whether the page belongs to a reservation.  If the page was
4075 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4076 	 * as an optimization, we avoid the check in that case.
4077 	 */
4078 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4079 		return (false);
4080 #endif
4081 
4082 	return (true);
4083 }
4084 
4085 /*
4086  *	vm_page_free_toq:
4087  *
4088  *	Returns the given page to the free list, disassociating it
4089  *	from any VM object.
4090  *
4091  *	The object must be locked.  The page must be exclusively busied if it
4092  *	belongs to an object.
4093  */
4094 static void
4095 vm_page_free_toq(vm_page_t m)
4096 {
4097 	struct vm_domain *vmd;
4098 	uma_zone_t zone;
4099 
4100 	if (!vm_page_free_prep(m))
4101 		return;
4102 
4103 	vmd = vm_pagequeue_domain(m);
4104 	zone = vmd->vmd_pgcache[m->pool].zone;
4105 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4106 		uma_zfree(zone, m);
4107 		return;
4108 	}
4109 	vm_domain_free_lock(vmd);
4110 	vm_phys_free_pages(m, 0);
4111 	vm_domain_free_unlock(vmd);
4112 	vm_domain_freecnt_inc(vmd, 1);
4113 }
4114 
4115 /*
4116  *	vm_page_free_pages_toq:
4117  *
4118  *	Returns a list of pages to the free list, disassociating it
4119  *	from any VM object.  In other words, this is equivalent to
4120  *	calling vm_page_free_toq() for each page of a list of VM objects.
4121  */
4122 void
4123 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4124 {
4125 	vm_page_t m;
4126 	int count;
4127 
4128 	if (SLIST_EMPTY(free))
4129 		return;
4130 
4131 	count = 0;
4132 	while ((m = SLIST_FIRST(free)) != NULL) {
4133 		count++;
4134 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4135 		vm_page_free_toq(m);
4136 	}
4137 
4138 	if (update_wire_count)
4139 		vm_wire_sub(count);
4140 }
4141 
4142 /*
4143  * Mark this page as wired down.  For managed pages, this prevents reclamation
4144  * by the page daemon, or when the containing object, if any, is destroyed.
4145  */
4146 void
4147 vm_page_wire(vm_page_t m)
4148 {
4149 	u_int old;
4150 
4151 #ifdef INVARIANTS
4152 	if (m->object != NULL && !vm_page_busied(m) &&
4153 	    !vm_object_busied(m->object))
4154 		VM_OBJECT_ASSERT_LOCKED(m->object);
4155 #endif
4156 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4157 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4158 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4159 
4160 	old = atomic_fetchadd_int(&m->ref_count, 1);
4161 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4162 	    ("vm_page_wire: counter overflow for page %p", m));
4163 	if (VPRC_WIRE_COUNT(old) == 0) {
4164 		if ((m->oflags & VPO_UNMANAGED) == 0)
4165 			vm_page_aflag_set(m, PGA_DEQUEUE);
4166 		vm_wire_add(1);
4167 	}
4168 }
4169 
4170 /*
4171  * Attempt to wire a mapped page following a pmap lookup of that page.
4172  * This may fail if a thread is concurrently tearing down mappings of the page.
4173  * The transient failure is acceptable because it translates to the
4174  * failure of the caller pmap_extract_and_hold(), which should be then
4175  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4176  */
4177 bool
4178 vm_page_wire_mapped(vm_page_t m)
4179 {
4180 	u_int old;
4181 
4182 	old = m->ref_count;
4183 	do {
4184 		KASSERT(old > 0,
4185 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4186 		if ((old & VPRC_BLOCKED) != 0)
4187 			return (false);
4188 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4189 
4190 	if (VPRC_WIRE_COUNT(old) == 0) {
4191 		if ((m->oflags & VPO_UNMANAGED) == 0)
4192 			vm_page_aflag_set(m, PGA_DEQUEUE);
4193 		vm_wire_add(1);
4194 	}
4195 	return (true);
4196 }
4197 
4198 /*
4199  * Release a wiring reference to a managed page.  If the page still belongs to
4200  * an object, update its position in the page queues to reflect the reference.
4201  * If the wiring was the last reference to the page, free the page.
4202  */
4203 static void
4204 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4205 {
4206 	u_int old;
4207 
4208 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4209 	    ("%s: page %p is unmanaged", __func__, m));
4210 
4211 	/*
4212 	 * Update LRU state before releasing the wiring reference.
4213 	 * Use a release store when updating the reference count to
4214 	 * synchronize with vm_page_free_prep().
4215 	 */
4216 	old = m->ref_count;
4217 	do {
4218 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4219 		    ("vm_page_unwire: wire count underflow for page %p", m));
4220 
4221 		if (old > VPRC_OBJREF + 1) {
4222 			/*
4223 			 * The page has at least one other wiring reference.  An
4224 			 * earlier iteration of this loop may have called
4225 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4226 			 * re-set it if necessary.
4227 			 */
4228 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4229 				vm_page_aflag_set(m, PGA_DEQUEUE);
4230 		} else if (old == VPRC_OBJREF + 1) {
4231 			/*
4232 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4233 			 * update the page's queue state to reflect the
4234 			 * reference.  If the page does not belong to an object
4235 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4236 			 * clear leftover queue state.
4237 			 */
4238 			vm_page_release_toq(m, nqueue, noreuse);
4239 		} else if (old == 1) {
4240 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4241 		}
4242 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4243 
4244 	if (VPRC_WIRE_COUNT(old) == 1) {
4245 		vm_wire_sub(1);
4246 		if (old == 1)
4247 			vm_page_free(m);
4248 	}
4249 }
4250 
4251 /*
4252  * Release one wiring of the specified page, potentially allowing it to be
4253  * paged out.
4254  *
4255  * Only managed pages belonging to an object can be paged out.  If the number
4256  * of wirings transitions to zero and the page is eligible for page out, then
4257  * the page is added to the specified paging queue.  If the released wiring
4258  * represented the last reference to the page, the page is freed.
4259  */
4260 void
4261 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4262 {
4263 
4264 	KASSERT(nqueue < PQ_COUNT,
4265 	    ("vm_page_unwire: invalid queue %u request for page %p",
4266 	    nqueue, m));
4267 
4268 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4269 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4270 			vm_page_free(m);
4271 		return;
4272 	}
4273 	vm_page_unwire_managed(m, nqueue, false);
4274 }
4275 
4276 /*
4277  * Unwire a page without (re-)inserting it into a page queue.  It is up
4278  * to the caller to enqueue, requeue, or free the page as appropriate.
4279  * In most cases involving managed pages, vm_page_unwire() should be used
4280  * instead.
4281  */
4282 bool
4283 vm_page_unwire_noq(vm_page_t m)
4284 {
4285 	u_int old;
4286 
4287 	old = vm_page_drop(m, 1);
4288 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4289 	    ("%s: counter underflow for page %p", __func__,  m));
4290 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4291 	    ("%s: missing ref on fictitious page %p", __func__, m));
4292 
4293 	if (VPRC_WIRE_COUNT(old) > 1)
4294 		return (false);
4295 	if ((m->oflags & VPO_UNMANAGED) == 0)
4296 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4297 	vm_wire_sub(1);
4298 	return (true);
4299 }
4300 
4301 /*
4302  * Ensure that the page ends up in the specified page queue.  If the page is
4303  * active or being moved to the active queue, ensure that its act_count is
4304  * at least ACT_INIT but do not otherwise mess with it.
4305  */
4306 static __always_inline void
4307 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4308 {
4309 	vm_page_astate_t old, new;
4310 
4311 	KASSERT(m->ref_count > 0,
4312 	    ("%s: page %p does not carry any references", __func__, m));
4313 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4314 	    ("%s: invalid flags %x", __func__, nflag));
4315 
4316 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4317 		return;
4318 
4319 	old = vm_page_astate_load(m);
4320 	do {
4321 		if ((old.flags & PGA_DEQUEUE) != 0)
4322 			break;
4323 		new = old;
4324 		new.flags &= ~PGA_QUEUE_OP_MASK;
4325 		if (nqueue == PQ_ACTIVE)
4326 			new.act_count = max(old.act_count, ACT_INIT);
4327 		if (old.queue == nqueue) {
4328 			/*
4329 			 * There is no need to requeue pages already in the
4330 			 * active queue.
4331 			 */
4332 			if (nqueue != PQ_ACTIVE ||
4333 			    (old.flags & PGA_ENQUEUED) == 0)
4334 				new.flags |= nflag;
4335 		} else {
4336 			new.flags |= nflag;
4337 			new.queue = nqueue;
4338 		}
4339 	} while (!vm_page_pqstate_commit(m, &old, new));
4340 }
4341 
4342 /*
4343  * Put the specified page on the active list (if appropriate).
4344  */
4345 void
4346 vm_page_activate(vm_page_t m)
4347 {
4348 
4349 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4350 }
4351 
4352 /*
4353  * Move the specified page to the tail of the inactive queue, or requeue
4354  * the page if it is already in the inactive queue.
4355  */
4356 void
4357 vm_page_deactivate(vm_page_t m)
4358 {
4359 
4360 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4361 }
4362 
4363 void
4364 vm_page_deactivate_noreuse(vm_page_t m)
4365 {
4366 
4367 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4368 }
4369 
4370 /*
4371  * Put a page in the laundry, or requeue it if it is already there.
4372  */
4373 void
4374 vm_page_launder(vm_page_t m)
4375 {
4376 
4377 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4378 }
4379 
4380 /*
4381  * Put a page in the PQ_UNSWAPPABLE holding queue.
4382  */
4383 void
4384 vm_page_unswappable(vm_page_t m)
4385 {
4386 
4387 	VM_OBJECT_ASSERT_LOCKED(m->object);
4388 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4389 	    ("page %p already unswappable", m));
4390 
4391 	vm_page_dequeue(m);
4392 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4393 }
4394 
4395 /*
4396  * Release a page back to the page queues in preparation for unwiring.
4397  */
4398 static void
4399 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4400 {
4401 	vm_page_astate_t old, new;
4402 	uint16_t nflag;
4403 
4404 	/*
4405 	 * Use a check of the valid bits to determine whether we should
4406 	 * accelerate reclamation of the page.  The object lock might not be
4407 	 * held here, in which case the check is racy.  At worst we will either
4408 	 * accelerate reclamation of a valid page and violate LRU, or
4409 	 * unnecessarily defer reclamation of an invalid page.
4410 	 *
4411 	 * If we were asked to not cache the page, place it near the head of the
4412 	 * inactive queue so that is reclaimed sooner.
4413 	 */
4414 	if (noreuse || vm_page_none_valid(m)) {
4415 		nqueue = PQ_INACTIVE;
4416 		nflag = PGA_REQUEUE_HEAD;
4417 	} else {
4418 		nflag = PGA_REQUEUE;
4419 	}
4420 
4421 	old = vm_page_astate_load(m);
4422 	do {
4423 		new = old;
4424 
4425 		/*
4426 		 * If the page is already in the active queue and we are not
4427 		 * trying to accelerate reclamation, simply mark it as
4428 		 * referenced and avoid any queue operations.
4429 		 */
4430 		new.flags &= ~PGA_QUEUE_OP_MASK;
4431 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4432 		    (old.flags & PGA_ENQUEUED) != 0)
4433 			new.flags |= PGA_REFERENCED;
4434 		else {
4435 			new.flags |= nflag;
4436 			new.queue = nqueue;
4437 		}
4438 	} while (!vm_page_pqstate_commit(m, &old, new));
4439 }
4440 
4441 /*
4442  * Unwire a page and either attempt to free it or re-add it to the page queues.
4443  */
4444 void
4445 vm_page_release(vm_page_t m, int flags)
4446 {
4447 	vm_object_t object;
4448 
4449 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4450 	    ("vm_page_release: page %p is unmanaged", m));
4451 
4452 	if ((flags & VPR_TRYFREE) != 0) {
4453 		for (;;) {
4454 			object = atomic_load_ptr(&m->object);
4455 			if (object == NULL)
4456 				break;
4457 			/* Depends on type-stability. */
4458 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4459 				break;
4460 			if (object == m->object) {
4461 				vm_page_release_locked(m, flags);
4462 				VM_OBJECT_WUNLOCK(object);
4463 				return;
4464 			}
4465 			VM_OBJECT_WUNLOCK(object);
4466 		}
4467 	}
4468 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4469 }
4470 
4471 /* See vm_page_release(). */
4472 void
4473 vm_page_release_locked(vm_page_t m, int flags)
4474 {
4475 
4476 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4477 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4478 	    ("vm_page_release_locked: page %p is unmanaged", m));
4479 
4480 	if (vm_page_unwire_noq(m)) {
4481 		if ((flags & VPR_TRYFREE) != 0 &&
4482 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4483 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4484 			/*
4485 			 * An unlocked lookup may have wired the page before the
4486 			 * busy lock was acquired, in which case the page must
4487 			 * not be freed.
4488 			 */
4489 			if (__predict_true(!vm_page_wired(m))) {
4490 				vm_page_free(m);
4491 				return;
4492 			}
4493 			vm_page_xunbusy(m);
4494 		} else {
4495 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4496 		}
4497 	}
4498 }
4499 
4500 static bool
4501 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4502 {
4503 	u_int old;
4504 
4505 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4506 	    ("vm_page_try_blocked_op: page %p has no object", m));
4507 	KASSERT(vm_page_busied(m),
4508 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4509 	VM_OBJECT_ASSERT_LOCKED(m->object);
4510 
4511 	old = m->ref_count;
4512 	do {
4513 		KASSERT(old != 0,
4514 		    ("vm_page_try_blocked_op: page %p has no references", m));
4515 		if (VPRC_WIRE_COUNT(old) != 0)
4516 			return (false);
4517 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4518 
4519 	(op)(m);
4520 
4521 	/*
4522 	 * If the object is read-locked, new wirings may be created via an
4523 	 * object lookup.
4524 	 */
4525 	old = vm_page_drop(m, VPRC_BLOCKED);
4526 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4527 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4528 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4529 	    old, m));
4530 	return (true);
4531 }
4532 
4533 /*
4534  * Atomically check for wirings and remove all mappings of the page.
4535  */
4536 bool
4537 vm_page_try_remove_all(vm_page_t m)
4538 {
4539 
4540 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4541 }
4542 
4543 /*
4544  * Atomically check for wirings and remove all writeable mappings of the page.
4545  */
4546 bool
4547 vm_page_try_remove_write(vm_page_t m)
4548 {
4549 
4550 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4551 }
4552 
4553 /*
4554  * vm_page_advise
4555  *
4556  * 	Apply the specified advice to the given page.
4557  */
4558 void
4559 vm_page_advise(vm_page_t m, int advice)
4560 {
4561 
4562 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4563 	vm_page_assert_xbusied(m);
4564 
4565 	if (advice == MADV_FREE)
4566 		/*
4567 		 * Mark the page clean.  This will allow the page to be freed
4568 		 * without first paging it out.  MADV_FREE pages are often
4569 		 * quickly reused by malloc(3), so we do not do anything that
4570 		 * would result in a page fault on a later access.
4571 		 */
4572 		vm_page_undirty(m);
4573 	else if (advice != MADV_DONTNEED) {
4574 		if (advice == MADV_WILLNEED)
4575 			vm_page_activate(m);
4576 		return;
4577 	}
4578 
4579 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4580 		vm_page_dirty(m);
4581 
4582 	/*
4583 	 * Clear any references to the page.  Otherwise, the page daemon will
4584 	 * immediately reactivate the page.
4585 	 */
4586 	vm_page_aflag_clear(m, PGA_REFERENCED);
4587 
4588 	/*
4589 	 * Place clean pages near the head of the inactive queue rather than
4590 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4591 	 * the page will be reused quickly.  Dirty pages not already in the
4592 	 * laundry are moved there.
4593 	 */
4594 	if (m->dirty == 0)
4595 		vm_page_deactivate_noreuse(m);
4596 	else if (!vm_page_in_laundry(m))
4597 		vm_page_launder(m);
4598 }
4599 
4600 /*
4601  *	vm_page_grab_release
4602  *
4603  *	Helper routine for grab functions to release busy on return.
4604  */
4605 static inline void
4606 vm_page_grab_release(vm_page_t m, int allocflags)
4607 {
4608 
4609 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4610 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4611 			vm_page_sunbusy(m);
4612 		else
4613 			vm_page_xunbusy(m);
4614 	}
4615 }
4616 
4617 /*
4618  *	vm_page_grab_sleep
4619  *
4620  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4621  *	if the caller should retry and false otherwise.
4622  *
4623  *	If the object is locked on entry the object will be unlocked with
4624  *	false returns and still locked but possibly having been dropped
4625  *	with true returns.
4626  */
4627 static bool
4628 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4629     const char *wmesg, int allocflags, bool locked)
4630 {
4631 
4632 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4633 		return (false);
4634 
4635 	/*
4636 	 * Reference the page before unlocking and sleeping so that
4637 	 * the page daemon is less likely to reclaim it.
4638 	 */
4639 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4640 		vm_page_reference(m);
4641 
4642 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4643 	    locked)
4644 		VM_OBJECT_WLOCK(object);
4645 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4646 		return (false);
4647 
4648 	return (true);
4649 }
4650 
4651 /*
4652  * Assert that the grab flags are valid.
4653  */
4654 static inline void
4655 vm_page_grab_check(int allocflags)
4656 {
4657 
4658 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4659 	    (allocflags & VM_ALLOC_WIRED) != 0,
4660 	    ("vm_page_grab*: the pages must be busied or wired"));
4661 
4662 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4663 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4664 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4665 }
4666 
4667 /*
4668  * Calculate the page allocation flags for grab.
4669  */
4670 static inline int
4671 vm_page_grab_pflags(int allocflags)
4672 {
4673 	int pflags;
4674 
4675 	pflags = allocflags &
4676 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4677 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4678 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4679 		pflags |= VM_ALLOC_WAITFAIL;
4680 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4681 		pflags |= VM_ALLOC_SBUSY;
4682 
4683 	return (pflags);
4684 }
4685 
4686 /*
4687  * Grab a page, waiting until we are waken up due to the page
4688  * changing state.  We keep on waiting, if the page continues
4689  * to be in the object.  If the page doesn't exist, first allocate it
4690  * and then conditionally zero it.
4691  *
4692  * This routine may sleep.
4693  *
4694  * The object must be locked on entry.  The lock will, however, be released
4695  * and reacquired if the routine sleeps.
4696  */
4697 vm_page_t
4698 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4699 {
4700 	vm_page_t m;
4701 
4702 	VM_OBJECT_ASSERT_WLOCKED(object);
4703 	vm_page_grab_check(allocflags);
4704 
4705 retrylookup:
4706 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4707 		if (!vm_page_tryacquire(m, allocflags)) {
4708 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4709 			    allocflags, true))
4710 				goto retrylookup;
4711 			return (NULL);
4712 		}
4713 		goto out;
4714 	}
4715 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4716 		return (NULL);
4717 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4718 	if (m == NULL) {
4719 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4720 			return (NULL);
4721 		goto retrylookup;
4722 	}
4723 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4724 		pmap_zero_page(m);
4725 
4726 out:
4727 	vm_page_grab_release(m, allocflags);
4728 
4729 	return (m);
4730 }
4731 
4732 /*
4733  * Locklessly attempt to acquire a page given a (object, pindex) tuple
4734  * and an optional previous page to avoid the radix lookup.  The resulting
4735  * page will be validated against the identity tuple and busied or wired
4736  * as requested.  A NULL *mp return guarantees that the page was not in
4737  * radix at the time of the call but callers must perform higher level
4738  * synchronization or retry the operation under a lock if they require
4739  * an atomic answer.  This is the only lock free validation routine,
4740  * other routines can depend on the resulting page state.
4741  *
4742  * The return value indicates whether the operation failed due to caller
4743  * flags.  The return is tri-state with mp:
4744  *
4745  * (true, *mp != NULL) - The operation was successful.
4746  * (true, *mp == NULL) - The page was not found in tree.
4747  * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4748  */
4749 static bool
4750 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4751     vm_page_t prev, vm_page_t *mp, int allocflags)
4752 {
4753 	vm_page_t m;
4754 
4755 	vm_page_grab_check(allocflags);
4756 	MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4757 
4758 	*mp = NULL;
4759 	for (;;) {
4760 		/*
4761 		 * We may see a false NULL here because the previous page
4762 		 * has been removed or just inserted and the list is loaded
4763 		 * without barriers.  Switch to radix to verify.
4764 		 */
4765 		if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4766 		    QMD_IS_TRASHED(m) || m->pindex != pindex ||
4767 		    atomic_load_ptr(&m->object) != object) {
4768 			prev = NULL;
4769 			/*
4770 			 * This guarantees the result is instantaneously
4771 			 * correct.
4772 			 */
4773 			m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4774 		}
4775 		if (m == NULL)
4776 			return (true);
4777 		if (vm_page_trybusy(m, allocflags)) {
4778 			if (m->object == object && m->pindex == pindex)
4779 				break;
4780 			/* relookup. */
4781 			vm_page_busy_release(m);
4782 			cpu_spinwait();
4783 			continue;
4784 		}
4785 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4786 		    allocflags, false))
4787 			return (false);
4788 	}
4789 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4790 		vm_page_wire(m);
4791 	vm_page_grab_release(m, allocflags);
4792 	*mp = m;
4793 	return (true);
4794 }
4795 
4796 /*
4797  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4798  * is not set.
4799  */
4800 vm_page_t
4801 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4802 {
4803 	vm_page_t m;
4804 
4805 	vm_page_grab_check(allocflags);
4806 
4807 	if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4808 		return (NULL);
4809 	if (m != NULL)
4810 		return (m);
4811 
4812 	/*
4813 	 * The radix lockless lookup should never return a false negative
4814 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4815 	 * was no page present at the instant of the call.  A NOCREAT caller
4816 	 * must handle create races gracefully.
4817 	 */
4818 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4819 		return (NULL);
4820 
4821 	VM_OBJECT_WLOCK(object);
4822 	m = vm_page_grab(object, pindex, allocflags);
4823 	VM_OBJECT_WUNLOCK(object);
4824 
4825 	return (m);
4826 }
4827 
4828 /*
4829  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4830  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4831  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4832  * in simultaneously.  Additional pages will be left on a paging queue but
4833  * will neither be wired nor busy regardless of allocflags.
4834  */
4835 int
4836 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4837 {
4838 	vm_page_t m;
4839 	vm_page_t ma[VM_INITIAL_PAGEIN];
4840 	int after, i, pflags, rv;
4841 
4842 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4843 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4844 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4845 	KASSERT((allocflags &
4846 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4847 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4848 	VM_OBJECT_ASSERT_WLOCKED(object);
4849 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4850 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4851 	pflags |= VM_ALLOC_WAITFAIL;
4852 
4853 retrylookup:
4854 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4855 		/*
4856 		 * If the page is fully valid it can only become invalid
4857 		 * with the object lock held.  If it is not valid it can
4858 		 * become valid with the busy lock held.  Therefore, we
4859 		 * may unnecessarily lock the exclusive busy here if we
4860 		 * race with I/O completion not using the object lock.
4861 		 * However, we will not end up with an invalid page and a
4862 		 * shared lock.
4863 		 */
4864 		if (!vm_page_trybusy(m,
4865 		    vm_page_all_valid(m) ? allocflags : 0)) {
4866 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4867 			    allocflags, true);
4868 			goto retrylookup;
4869 		}
4870 		if (vm_page_all_valid(m))
4871 			goto out;
4872 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4873 			vm_page_busy_release(m);
4874 			*mp = NULL;
4875 			return (VM_PAGER_FAIL);
4876 		}
4877 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4878 		*mp = NULL;
4879 		return (VM_PAGER_FAIL);
4880 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4881 		if (!vm_pager_can_alloc_page(object, pindex)) {
4882 			*mp = NULL;
4883 			return (VM_PAGER_AGAIN);
4884 		}
4885 		goto retrylookup;
4886 	}
4887 
4888 	vm_page_assert_xbusied(m);
4889 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4890 		after = MIN(after, VM_INITIAL_PAGEIN);
4891 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4892 		after = MAX(after, 1);
4893 		ma[0] = m;
4894 		for (i = 1; i < after; i++) {
4895 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4896 				if (vm_page_any_valid(ma[i]) ||
4897 				    !vm_page_tryxbusy(ma[i]))
4898 					break;
4899 			} else {
4900 				ma[i] = vm_page_alloc(object, m->pindex + i,
4901 				    VM_ALLOC_NORMAL);
4902 				if (ma[i] == NULL)
4903 					break;
4904 			}
4905 		}
4906 		after = i;
4907 		vm_object_pip_add(object, after);
4908 		VM_OBJECT_WUNLOCK(object);
4909 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4910 		VM_OBJECT_WLOCK(object);
4911 		vm_object_pip_wakeupn(object, after);
4912 		/* Pager may have replaced a page. */
4913 		m = ma[0];
4914 		if (rv != VM_PAGER_OK) {
4915 			for (i = 0; i < after; i++) {
4916 				if (!vm_page_wired(ma[i]))
4917 					vm_page_free(ma[i]);
4918 				else
4919 					vm_page_xunbusy(ma[i]);
4920 			}
4921 			*mp = NULL;
4922 			return (rv);
4923 		}
4924 		for (i = 1; i < after; i++)
4925 			vm_page_readahead_finish(ma[i]);
4926 		MPASS(vm_page_all_valid(m));
4927 	} else {
4928 		vm_page_zero_invalid(m, TRUE);
4929 	}
4930 out:
4931 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4932 		vm_page_wire(m);
4933 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4934 		vm_page_busy_downgrade(m);
4935 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4936 		vm_page_busy_release(m);
4937 	*mp = m;
4938 	return (VM_PAGER_OK);
4939 }
4940 
4941 /*
4942  * Locklessly grab a valid page.  If the page is not valid or not yet
4943  * allocated this will fall back to the object lock method.
4944  */
4945 int
4946 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4947     vm_pindex_t pindex, int allocflags)
4948 {
4949 	vm_page_t m;
4950 	int flags;
4951 	int error;
4952 
4953 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4954 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4955 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4956 	    "mismatch"));
4957 	KASSERT((allocflags &
4958 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4959 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4960 
4961 	/*
4962 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
4963 	 * before we can inspect the valid field and return a wired page.
4964 	 */
4965 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4966 	if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4967 		return (VM_PAGER_FAIL);
4968 	if ((m = *mp) != NULL) {
4969 		if (vm_page_all_valid(m)) {
4970 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4971 				vm_page_wire(m);
4972 			vm_page_grab_release(m, allocflags);
4973 			return (VM_PAGER_OK);
4974 		}
4975 		vm_page_busy_release(m);
4976 	}
4977 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4978 		*mp = NULL;
4979 		return (VM_PAGER_FAIL);
4980 	}
4981 	VM_OBJECT_WLOCK(object);
4982 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
4983 	VM_OBJECT_WUNLOCK(object);
4984 
4985 	return (error);
4986 }
4987 
4988 /*
4989  * Return the specified range of pages from the given object.  For each
4990  * page offset within the range, if a page already exists within the object
4991  * at that offset and it is busy, then wait for it to change state.  If,
4992  * instead, the page doesn't exist, then allocate it.
4993  *
4994  * The caller must always specify an allocation class.
4995  *
4996  * allocation classes:
4997  *	VM_ALLOC_NORMAL		normal process request
4998  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4999  *
5000  * The caller must always specify that the pages are to be busied and/or
5001  * wired.
5002  *
5003  * optional allocation flags:
5004  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5005  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5006  *	VM_ALLOC_NOWAIT		do not sleep
5007  *	VM_ALLOC_SBUSY		set page to sbusy state
5008  *	VM_ALLOC_WIRED		wire the pages
5009  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5010  *
5011  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5012  * may return a partial prefix of the requested range.
5013  */
5014 int
5015 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5016     vm_page_t *ma, int count)
5017 {
5018 	vm_page_t m, mpred;
5019 	int pflags;
5020 	int i;
5021 
5022 	VM_OBJECT_ASSERT_WLOCKED(object);
5023 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5024 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5025 	KASSERT(count > 0,
5026 	    ("vm_page_grab_pages: invalid page count %d", count));
5027 	vm_page_grab_check(allocflags);
5028 
5029 	pflags = vm_page_grab_pflags(allocflags);
5030 	i = 0;
5031 retrylookup:
5032 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
5033 	if (m == NULL || m->pindex != pindex + i) {
5034 		mpred = m;
5035 		m = NULL;
5036 	} else
5037 		mpred = TAILQ_PREV(m, pglist, listq);
5038 	for (; i < count; i++) {
5039 		if (m != NULL) {
5040 			if (!vm_page_tryacquire(m, allocflags)) {
5041 				if (vm_page_grab_sleep(object, m, pindex + i,
5042 				    "grbmaw", allocflags, true))
5043 					goto retrylookup;
5044 				break;
5045 			}
5046 		} else {
5047 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5048 				break;
5049 			m = vm_page_alloc_after(object, pindex + i,
5050 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5051 			if (m == NULL) {
5052 				if ((allocflags & (VM_ALLOC_NOWAIT |
5053 				    VM_ALLOC_WAITFAIL)) != 0)
5054 					break;
5055 				goto retrylookup;
5056 			}
5057 		}
5058 		if (vm_page_none_valid(m) &&
5059 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5060 			if ((m->flags & PG_ZERO) == 0)
5061 				pmap_zero_page(m);
5062 			vm_page_valid(m);
5063 		}
5064 		vm_page_grab_release(m, allocflags);
5065 		ma[i] = mpred = m;
5066 		m = vm_page_next(m);
5067 	}
5068 	return (i);
5069 }
5070 
5071 /*
5072  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5073  * and will fall back to the locked variant to handle allocation.
5074  */
5075 int
5076 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5077     int allocflags, vm_page_t *ma, int count)
5078 {
5079 	vm_page_t m, pred;
5080 	int flags;
5081 	int i;
5082 
5083 	KASSERT(count > 0,
5084 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5085 	vm_page_grab_check(allocflags);
5086 
5087 	/*
5088 	 * Modify flags for lockless acquire to hold the page until we
5089 	 * set it valid if necessary.
5090 	 */
5091 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5092 	pred = NULL;
5093 	for (i = 0; i < count; i++, pindex++) {
5094 		if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
5095 			return (i);
5096 		if (m == NULL)
5097 			break;
5098 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5099 			if ((m->flags & PG_ZERO) == 0)
5100 				pmap_zero_page(m);
5101 			vm_page_valid(m);
5102 		}
5103 		/* m will still be wired or busy according to flags. */
5104 		vm_page_grab_release(m, allocflags);
5105 		pred = ma[i] = m;
5106 	}
5107 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5108 		return (i);
5109 	count -= i;
5110 	VM_OBJECT_WLOCK(object);
5111 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5112 	VM_OBJECT_WUNLOCK(object);
5113 
5114 	return (i);
5115 }
5116 
5117 /*
5118  * Mapping function for valid or dirty bits in a page.
5119  *
5120  * Inputs are required to range within a page.
5121  */
5122 vm_page_bits_t
5123 vm_page_bits(int base, int size)
5124 {
5125 	int first_bit;
5126 	int last_bit;
5127 
5128 	KASSERT(
5129 	    base + size <= PAGE_SIZE,
5130 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5131 	);
5132 
5133 	if (size == 0)		/* handle degenerate case */
5134 		return (0);
5135 
5136 	first_bit = base >> DEV_BSHIFT;
5137 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5138 
5139 	return (((vm_page_bits_t)2 << last_bit) -
5140 	    ((vm_page_bits_t)1 << first_bit));
5141 }
5142 
5143 void
5144 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5145 {
5146 
5147 #if PAGE_SIZE == 32768
5148 	atomic_set_64((uint64_t *)bits, set);
5149 #elif PAGE_SIZE == 16384
5150 	atomic_set_32((uint32_t *)bits, set);
5151 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5152 	atomic_set_16((uint16_t *)bits, set);
5153 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5154 	atomic_set_8((uint8_t *)bits, set);
5155 #else		/* PAGE_SIZE <= 8192 */
5156 	uintptr_t addr;
5157 	int shift;
5158 
5159 	addr = (uintptr_t)bits;
5160 	/*
5161 	 * Use a trick to perform a 32-bit atomic on the
5162 	 * containing aligned word, to not depend on the existence
5163 	 * of atomic_{set, clear}_{8, 16}.
5164 	 */
5165 	shift = addr & (sizeof(uint32_t) - 1);
5166 #if BYTE_ORDER == BIG_ENDIAN
5167 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5168 #else
5169 	shift *= NBBY;
5170 #endif
5171 	addr &= ~(sizeof(uint32_t) - 1);
5172 	atomic_set_32((uint32_t *)addr, set << shift);
5173 #endif		/* PAGE_SIZE */
5174 }
5175 
5176 static inline void
5177 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5178 {
5179 
5180 #if PAGE_SIZE == 32768
5181 	atomic_clear_64((uint64_t *)bits, clear);
5182 #elif PAGE_SIZE == 16384
5183 	atomic_clear_32((uint32_t *)bits, clear);
5184 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5185 	atomic_clear_16((uint16_t *)bits, clear);
5186 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5187 	atomic_clear_8((uint8_t *)bits, clear);
5188 #else		/* PAGE_SIZE <= 8192 */
5189 	uintptr_t addr;
5190 	int shift;
5191 
5192 	addr = (uintptr_t)bits;
5193 	/*
5194 	 * Use a trick to perform a 32-bit atomic on the
5195 	 * containing aligned word, to not depend on the existence
5196 	 * of atomic_{set, clear}_{8, 16}.
5197 	 */
5198 	shift = addr & (sizeof(uint32_t) - 1);
5199 #if BYTE_ORDER == BIG_ENDIAN
5200 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5201 #else
5202 	shift *= NBBY;
5203 #endif
5204 	addr &= ~(sizeof(uint32_t) - 1);
5205 	atomic_clear_32((uint32_t *)addr, clear << shift);
5206 #endif		/* PAGE_SIZE */
5207 }
5208 
5209 static inline vm_page_bits_t
5210 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5211 {
5212 #if PAGE_SIZE == 32768
5213 	uint64_t old;
5214 
5215 	old = *bits;
5216 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5217 	return (old);
5218 #elif PAGE_SIZE == 16384
5219 	uint32_t old;
5220 
5221 	old = *bits;
5222 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5223 	return (old);
5224 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5225 	uint16_t old;
5226 
5227 	old = *bits;
5228 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5229 	return (old);
5230 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5231 	uint8_t old;
5232 
5233 	old = *bits;
5234 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5235 	return (old);
5236 #else		/* PAGE_SIZE <= 4096*/
5237 	uintptr_t addr;
5238 	uint32_t old, new, mask;
5239 	int shift;
5240 
5241 	addr = (uintptr_t)bits;
5242 	/*
5243 	 * Use a trick to perform a 32-bit atomic on the
5244 	 * containing aligned word, to not depend on the existence
5245 	 * of atomic_{set, swap, clear}_{8, 16}.
5246 	 */
5247 	shift = addr & (sizeof(uint32_t) - 1);
5248 #if BYTE_ORDER == BIG_ENDIAN
5249 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5250 #else
5251 	shift *= NBBY;
5252 #endif
5253 	addr &= ~(sizeof(uint32_t) - 1);
5254 	mask = VM_PAGE_BITS_ALL << shift;
5255 
5256 	old = *bits;
5257 	do {
5258 		new = old & ~mask;
5259 		new |= newbits << shift;
5260 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5261 	return (old >> shift);
5262 #endif		/* PAGE_SIZE */
5263 }
5264 
5265 /*
5266  *	vm_page_set_valid_range:
5267  *
5268  *	Sets portions of a page valid.  The arguments are expected
5269  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5270  *	of any partial chunks touched by the range.  The invalid portion of
5271  *	such chunks will be zeroed.
5272  *
5273  *	(base + size) must be less then or equal to PAGE_SIZE.
5274  */
5275 void
5276 vm_page_set_valid_range(vm_page_t m, int base, int size)
5277 {
5278 	int endoff, frag;
5279 	vm_page_bits_t pagebits;
5280 
5281 	vm_page_assert_busied(m);
5282 	if (size == 0)	/* handle degenerate case */
5283 		return;
5284 
5285 	/*
5286 	 * If the base is not DEV_BSIZE aligned and the valid
5287 	 * bit is clear, we have to zero out a portion of the
5288 	 * first block.
5289 	 */
5290 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5291 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5292 		pmap_zero_page_area(m, frag, base - frag);
5293 
5294 	/*
5295 	 * If the ending offset is not DEV_BSIZE aligned and the
5296 	 * valid bit is clear, we have to zero out a portion of
5297 	 * the last block.
5298 	 */
5299 	endoff = base + size;
5300 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5301 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5302 		pmap_zero_page_area(m, endoff,
5303 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5304 
5305 	/*
5306 	 * Assert that no previously invalid block that is now being validated
5307 	 * is already dirty.
5308 	 */
5309 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5310 	    ("vm_page_set_valid_range: page %p is dirty", m));
5311 
5312 	/*
5313 	 * Set valid bits inclusive of any overlap.
5314 	 */
5315 	pagebits = vm_page_bits(base, size);
5316 	if (vm_page_xbusied(m))
5317 		m->valid |= pagebits;
5318 	else
5319 		vm_page_bits_set(m, &m->valid, pagebits);
5320 }
5321 
5322 /*
5323  * Set the page dirty bits and free the invalid swap space if
5324  * present.  Returns the previous dirty bits.
5325  */
5326 vm_page_bits_t
5327 vm_page_set_dirty(vm_page_t m)
5328 {
5329 	vm_page_bits_t old;
5330 
5331 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5332 
5333 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5334 		old = m->dirty;
5335 		m->dirty = VM_PAGE_BITS_ALL;
5336 	} else
5337 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5338 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5339 		vm_pager_page_unswapped(m);
5340 
5341 	return (old);
5342 }
5343 
5344 /*
5345  * Clear the given bits from the specified page's dirty field.
5346  */
5347 static __inline void
5348 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5349 {
5350 
5351 	vm_page_assert_busied(m);
5352 
5353 	/*
5354 	 * If the page is xbusied and not write mapped we are the
5355 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5356 	 * layer can call vm_page_dirty() without holding a distinguished
5357 	 * lock.  The combination of page busy and atomic operations
5358 	 * suffice to guarantee consistency of the page dirty field.
5359 	 */
5360 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5361 		m->dirty &= ~pagebits;
5362 	else
5363 		vm_page_bits_clear(m, &m->dirty, pagebits);
5364 }
5365 
5366 /*
5367  *	vm_page_set_validclean:
5368  *
5369  *	Sets portions of a page valid and clean.  The arguments are expected
5370  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5371  *	of any partial chunks touched by the range.  The invalid portion of
5372  *	such chunks will be zero'd.
5373  *
5374  *	(base + size) must be less then or equal to PAGE_SIZE.
5375  */
5376 void
5377 vm_page_set_validclean(vm_page_t m, int base, int size)
5378 {
5379 	vm_page_bits_t oldvalid, pagebits;
5380 	int endoff, frag;
5381 
5382 	vm_page_assert_busied(m);
5383 	if (size == 0)	/* handle degenerate case */
5384 		return;
5385 
5386 	/*
5387 	 * If the base is not DEV_BSIZE aligned and the valid
5388 	 * bit is clear, we have to zero out a portion of the
5389 	 * first block.
5390 	 */
5391 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5392 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5393 		pmap_zero_page_area(m, frag, base - frag);
5394 
5395 	/*
5396 	 * If the ending offset is not DEV_BSIZE aligned and the
5397 	 * valid bit is clear, we have to zero out a portion of
5398 	 * the last block.
5399 	 */
5400 	endoff = base + size;
5401 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5402 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5403 		pmap_zero_page_area(m, endoff,
5404 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5405 
5406 	/*
5407 	 * Set valid, clear dirty bits.  If validating the entire
5408 	 * page we can safely clear the pmap modify bit.  We also
5409 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5410 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5411 	 * be set again.
5412 	 *
5413 	 * We set valid bits inclusive of any overlap, but we can only
5414 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5415 	 * the range.
5416 	 */
5417 	oldvalid = m->valid;
5418 	pagebits = vm_page_bits(base, size);
5419 	if (vm_page_xbusied(m))
5420 		m->valid |= pagebits;
5421 	else
5422 		vm_page_bits_set(m, &m->valid, pagebits);
5423 #if 0	/* NOT YET */
5424 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5425 		frag = DEV_BSIZE - frag;
5426 		base += frag;
5427 		size -= frag;
5428 		if (size < 0)
5429 			size = 0;
5430 	}
5431 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5432 #endif
5433 	if (base == 0 && size == PAGE_SIZE) {
5434 		/*
5435 		 * The page can only be modified within the pmap if it is
5436 		 * mapped, and it can only be mapped if it was previously
5437 		 * fully valid.
5438 		 */
5439 		if (oldvalid == VM_PAGE_BITS_ALL)
5440 			/*
5441 			 * Perform the pmap_clear_modify() first.  Otherwise,
5442 			 * a concurrent pmap operation, such as
5443 			 * pmap_protect(), could clear a modification in the
5444 			 * pmap and set the dirty field on the page before
5445 			 * pmap_clear_modify() had begun and after the dirty
5446 			 * field was cleared here.
5447 			 */
5448 			pmap_clear_modify(m);
5449 		m->dirty = 0;
5450 		vm_page_aflag_clear(m, PGA_NOSYNC);
5451 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5452 		m->dirty &= ~pagebits;
5453 	else
5454 		vm_page_clear_dirty_mask(m, pagebits);
5455 }
5456 
5457 void
5458 vm_page_clear_dirty(vm_page_t m, int base, int size)
5459 {
5460 
5461 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5462 }
5463 
5464 /*
5465  *	vm_page_set_invalid:
5466  *
5467  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5468  *	valid and dirty bits for the effected areas are cleared.
5469  */
5470 void
5471 vm_page_set_invalid(vm_page_t m, int base, int size)
5472 {
5473 	vm_page_bits_t bits;
5474 	vm_object_t object;
5475 
5476 	/*
5477 	 * The object lock is required so that pages can't be mapped
5478 	 * read-only while we're in the process of invalidating them.
5479 	 */
5480 	object = m->object;
5481 	VM_OBJECT_ASSERT_WLOCKED(object);
5482 	vm_page_assert_busied(m);
5483 
5484 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5485 	    size >= object->un_pager.vnp.vnp_size)
5486 		bits = VM_PAGE_BITS_ALL;
5487 	else
5488 		bits = vm_page_bits(base, size);
5489 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5490 		pmap_remove_all(m);
5491 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5492 	    !pmap_page_is_mapped(m),
5493 	    ("vm_page_set_invalid: page %p is mapped", m));
5494 	if (vm_page_xbusied(m)) {
5495 		m->valid &= ~bits;
5496 		m->dirty &= ~bits;
5497 	} else {
5498 		vm_page_bits_clear(m, &m->valid, bits);
5499 		vm_page_bits_clear(m, &m->dirty, bits);
5500 	}
5501 }
5502 
5503 /*
5504  *	vm_page_invalid:
5505  *
5506  *	Invalidates the entire page.  The page must be busy, unmapped, and
5507  *	the enclosing object must be locked.  The object locks protects
5508  *	against concurrent read-only pmap enter which is done without
5509  *	busy.
5510  */
5511 void
5512 vm_page_invalid(vm_page_t m)
5513 {
5514 
5515 	vm_page_assert_busied(m);
5516 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5517 	MPASS(!pmap_page_is_mapped(m));
5518 
5519 	if (vm_page_xbusied(m))
5520 		m->valid = 0;
5521 	else
5522 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5523 }
5524 
5525 /*
5526  * vm_page_zero_invalid()
5527  *
5528  *	The kernel assumes that the invalid portions of a page contain
5529  *	garbage, but such pages can be mapped into memory by user code.
5530  *	When this occurs, we must zero out the non-valid portions of the
5531  *	page so user code sees what it expects.
5532  *
5533  *	Pages are most often semi-valid when the end of a file is mapped
5534  *	into memory and the file's size is not page aligned.
5535  */
5536 void
5537 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5538 {
5539 	int b;
5540 	int i;
5541 
5542 	/*
5543 	 * Scan the valid bits looking for invalid sections that
5544 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5545 	 * valid bit may be set ) have already been zeroed by
5546 	 * vm_page_set_validclean().
5547 	 */
5548 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5549 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5550 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5551 			if (i > b) {
5552 				pmap_zero_page_area(m,
5553 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5554 			}
5555 			b = i + 1;
5556 		}
5557 	}
5558 
5559 	/*
5560 	 * setvalid is TRUE when we can safely set the zero'd areas
5561 	 * as being valid.  We can do this if there are no cache consistency
5562 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5563 	 */
5564 	if (setvalid)
5565 		vm_page_valid(m);
5566 }
5567 
5568 /*
5569  *	vm_page_is_valid:
5570  *
5571  *	Is (partial) page valid?  Note that the case where size == 0
5572  *	will return FALSE in the degenerate case where the page is
5573  *	entirely invalid, and TRUE otherwise.
5574  *
5575  *	Some callers envoke this routine without the busy lock held and
5576  *	handle races via higher level locks.  Typical callers should
5577  *	hold a busy lock to prevent invalidation.
5578  */
5579 int
5580 vm_page_is_valid(vm_page_t m, int base, int size)
5581 {
5582 	vm_page_bits_t bits;
5583 
5584 	bits = vm_page_bits(base, size);
5585 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5586 }
5587 
5588 /*
5589  * Returns true if all of the specified predicates are true for the entire
5590  * (super)page and false otherwise.
5591  */
5592 bool
5593 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5594 {
5595 	vm_object_t object;
5596 	int i, npages;
5597 
5598 	object = m->object;
5599 	if (skip_m != NULL && skip_m->object != object)
5600 		return (false);
5601 	VM_OBJECT_ASSERT_LOCKED(object);
5602 	KASSERT(psind <= m->psind,
5603 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5604 	npages = atop(pagesizes[psind]);
5605 
5606 	/*
5607 	 * The physically contiguous pages that make up a superpage, i.e., a
5608 	 * page with a page size index ("psind") greater than zero, will
5609 	 * occupy adjacent entries in vm_page_array[].
5610 	 */
5611 	for (i = 0; i < npages; i++) {
5612 		/* Always test object consistency, including "skip_m". */
5613 		if (m[i].object != object)
5614 			return (false);
5615 		if (&m[i] == skip_m)
5616 			continue;
5617 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5618 			return (false);
5619 		if ((flags & PS_ALL_DIRTY) != 0) {
5620 			/*
5621 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5622 			 * might stop this case from spuriously returning
5623 			 * "false".  However, that would require a write lock
5624 			 * on the object containing "m[i]".
5625 			 */
5626 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5627 				return (false);
5628 		}
5629 		if ((flags & PS_ALL_VALID) != 0 &&
5630 		    m[i].valid != VM_PAGE_BITS_ALL)
5631 			return (false);
5632 	}
5633 	return (true);
5634 }
5635 
5636 /*
5637  * Set the page's dirty bits if the page is modified.
5638  */
5639 void
5640 vm_page_test_dirty(vm_page_t m)
5641 {
5642 
5643 	vm_page_assert_busied(m);
5644 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5645 		vm_page_dirty(m);
5646 }
5647 
5648 void
5649 vm_page_valid(vm_page_t m)
5650 {
5651 
5652 	vm_page_assert_busied(m);
5653 	if (vm_page_xbusied(m))
5654 		m->valid = VM_PAGE_BITS_ALL;
5655 	else
5656 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5657 }
5658 
5659 void
5660 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5661 {
5662 
5663 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5664 }
5665 
5666 void
5667 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5668 {
5669 
5670 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5671 }
5672 
5673 int
5674 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5675 {
5676 
5677 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5678 }
5679 
5680 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5681 void
5682 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5683 {
5684 
5685 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5686 }
5687 
5688 void
5689 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5690 {
5691 
5692 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5693 }
5694 #endif
5695 
5696 #ifdef INVARIANTS
5697 void
5698 vm_page_object_busy_assert(vm_page_t m)
5699 {
5700 
5701 	/*
5702 	 * Certain of the page's fields may only be modified by the
5703 	 * holder of a page or object busy.
5704 	 */
5705 	if (m->object != NULL && !vm_page_busied(m))
5706 		VM_OBJECT_ASSERT_BUSY(m->object);
5707 }
5708 
5709 void
5710 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5711 {
5712 
5713 	if ((bits & PGA_WRITEABLE) == 0)
5714 		return;
5715 
5716 	/*
5717 	 * The PGA_WRITEABLE flag can only be set if the page is
5718 	 * managed, is exclusively busied or the object is locked.
5719 	 * Currently, this flag is only set by pmap_enter().
5720 	 */
5721 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5722 	    ("PGA_WRITEABLE on unmanaged page"));
5723 	if (!vm_page_xbusied(m))
5724 		VM_OBJECT_ASSERT_BUSY(m->object);
5725 }
5726 #endif
5727 
5728 #include "opt_ddb.h"
5729 #ifdef DDB
5730 #include <sys/kernel.h>
5731 
5732 #include <ddb/ddb.h>
5733 
5734 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5735 {
5736 
5737 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5738 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5739 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5740 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5741 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5742 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5743 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5744 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5745 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5746 }
5747 
5748 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5749 {
5750 	int dom;
5751 
5752 	db_printf("pq_free %d\n", vm_free_count());
5753 	for (dom = 0; dom < vm_ndomains; dom++) {
5754 		db_printf(
5755     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5756 		    dom,
5757 		    vm_dom[dom].vmd_page_count,
5758 		    vm_dom[dom].vmd_free_count,
5759 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5760 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5761 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5762 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5763 	}
5764 }
5765 
5766 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5767 {
5768 	vm_page_t m;
5769 	boolean_t phys, virt;
5770 
5771 	if (!have_addr) {
5772 		db_printf("show pginfo addr\n");
5773 		return;
5774 	}
5775 
5776 	phys = strchr(modif, 'p') != NULL;
5777 	virt = strchr(modif, 'v') != NULL;
5778 	if (virt)
5779 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5780 	else if (phys)
5781 		m = PHYS_TO_VM_PAGE(addr);
5782 	else
5783 		m = (vm_page_t)addr;
5784 	db_printf(
5785     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5786     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5787 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5788 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5789 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5790 }
5791 #endif /* DDB */
5792