xref: /freebsd/sys/vm/vm_page.c (revision 55aa53fbf2bbeb7c032eae1a3767e043390a28da)
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 vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
166     int req, vm_page_t mpred);
167 static void vm_page_alloc_check(vm_page_t m);
168 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req);
169 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
170     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
171 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
172 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
173 static bool vm_page_free_prep(vm_page_t m);
174 static void vm_page_free_toq(vm_page_t m);
175 static void vm_page_init(void *dummy);
176 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
177     vm_pindex_t pindex, vm_page_t mpred);
178 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
179     vm_page_t mpred);
180 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
181     const uint16_t nflag);
182 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
183     vm_page_t m_run, vm_paddr_t high);
184 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
185 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
186     int req);
187 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
188     int flags);
189 static void vm_page_zone_release(void *arg, void **store, int cnt);
190 
191 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
192 
193 static void
194 vm_page_init(void *dummy)
195 {
196 
197 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
198 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
199 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE);
200 }
201 
202 static int pgcache_zone_max_pcpu;
203 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
204     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
205     "Per-CPU page cache size");
206 
207 /*
208  * The cache page zone is initialized later since we need to be able to allocate
209  * pages before UMA is fully initialized.
210  */
211 static void
212 vm_page_init_cache_zones(void *dummy __unused)
213 {
214 	struct vm_domain *vmd;
215 	struct vm_pgcache *pgcache;
216 	int cache, domain, maxcache, pool;
217 
218 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
219 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
220 	for (domain = 0; domain < vm_ndomains; domain++) {
221 		vmd = VM_DOMAIN(domain);
222 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
223 			pgcache = &vmd->vmd_pgcache[pool];
224 			pgcache->domain = domain;
225 			pgcache->pool = pool;
226 			pgcache->zone = uma_zcache_create("vm pgcache",
227 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
228 			    vm_page_zone_import, vm_page_zone_release, pgcache,
229 			    UMA_ZONE_VM);
230 
231 			/*
232 			 * Limit each pool's zone to 0.1% of the pages in the
233 			 * domain.
234 			 */
235 			cache = maxcache != 0 ? maxcache :
236 			    vmd->vmd_page_count / 1000;
237 			uma_zone_set_maxcache(pgcache->zone, cache);
238 		}
239 	}
240 }
241 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
242 
243 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
244 #if PAGE_SIZE == 32768
245 #ifdef CTASSERT
246 CTASSERT(sizeof(u_long) >= 8);
247 #endif
248 #endif
249 
250 /*
251  *	vm_set_page_size:
252  *
253  *	Sets the page size, perhaps based upon the memory
254  *	size.  Must be called before any use of page-size
255  *	dependent functions.
256  */
257 void
258 vm_set_page_size(void)
259 {
260 	if (vm_cnt.v_page_size == 0)
261 		vm_cnt.v_page_size = PAGE_SIZE;
262 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
263 		panic("vm_set_page_size: page size not a power of two");
264 }
265 
266 /*
267  *	vm_page_blacklist_next:
268  *
269  *	Find the next entry in the provided string of blacklist
270  *	addresses.  Entries are separated by space, comma, or newline.
271  *	If an invalid integer is encountered then the rest of the
272  *	string is skipped.  Updates the list pointer to the next
273  *	character, or NULL if the string is exhausted or invalid.
274  */
275 static vm_paddr_t
276 vm_page_blacklist_next(char **list, char *end)
277 {
278 	vm_paddr_t bad;
279 	char *cp, *pos;
280 
281 	if (list == NULL || *list == NULL)
282 		return (0);
283 	if (**list =='\0') {
284 		*list = NULL;
285 		return (0);
286 	}
287 
288 	/*
289 	 * If there's no end pointer then the buffer is coming from
290 	 * the kenv and we know it's null-terminated.
291 	 */
292 	if (end == NULL)
293 		end = *list + strlen(*list);
294 
295 	/* Ensure that strtoq() won't walk off the end */
296 	if (*end != '\0') {
297 		if (*end == '\n' || *end == ' ' || *end  == ',')
298 			*end = '\0';
299 		else {
300 			printf("Blacklist not terminated, skipping\n");
301 			*list = NULL;
302 			return (0);
303 		}
304 	}
305 
306 	for (pos = *list; *pos != '\0'; pos = cp) {
307 		bad = strtoq(pos, &cp, 0);
308 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
309 			if (bad == 0) {
310 				if (++cp < end)
311 					continue;
312 				else
313 					break;
314 			}
315 		} else
316 			break;
317 		if (*cp == '\0' || ++cp >= end)
318 			*list = NULL;
319 		else
320 			*list = cp;
321 		return (trunc_page(bad));
322 	}
323 	printf("Garbage in RAM blacklist, skipping\n");
324 	*list = NULL;
325 	return (0);
326 }
327 
328 bool
329 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
330 {
331 	struct vm_domain *vmd;
332 	vm_page_t m;
333 	bool found;
334 
335 	m = vm_phys_paddr_to_vm_page(pa);
336 	if (m == NULL)
337 		return (true); /* page does not exist, no failure */
338 
339 	vmd = VM_DOMAIN(vm_phys_domain(pa));
340 	vm_domain_free_lock(vmd);
341 	found = vm_phys_unfree_page(pa);
342 	vm_domain_free_unlock(vmd);
343 	if (found) {
344 		vm_domain_freecnt_inc(vmd, -1);
345 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
346 		if (verbose)
347 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
348 	}
349 	return (found);
350 }
351 
352 /*
353  *	vm_page_blacklist_check:
354  *
355  *	Iterate through the provided string of blacklist addresses, pulling
356  *	each entry out of the physical allocator free list and putting it
357  *	onto a list for reporting via the vm.page_blacklist sysctl.
358  */
359 static void
360 vm_page_blacklist_check(char *list, char *end)
361 {
362 	vm_paddr_t pa;
363 	char *next;
364 
365 	next = list;
366 	while (next != NULL) {
367 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
368 			continue;
369 		vm_page_blacklist_add(pa, bootverbose);
370 	}
371 }
372 
373 /*
374  *	vm_page_blacklist_load:
375  *
376  *	Search for a special module named "ram_blacklist".  It'll be a
377  *	plain text file provided by the user via the loader directive
378  *	of the same name.
379  */
380 static void
381 vm_page_blacklist_load(char **list, char **end)
382 {
383 	void *mod;
384 	u_char *ptr;
385 	u_int len;
386 
387 	mod = NULL;
388 	ptr = NULL;
389 
390 	mod = preload_search_by_type("ram_blacklist");
391 	if (mod != NULL) {
392 		ptr = preload_fetch_addr(mod);
393 		len = preload_fetch_size(mod);
394         }
395 	*list = ptr;
396 	if (ptr != NULL)
397 		*end = ptr + len;
398 	else
399 		*end = NULL;
400 	return;
401 }
402 
403 static int
404 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
405 {
406 	vm_page_t m;
407 	struct sbuf sbuf;
408 	int error, first;
409 
410 	first = 1;
411 	error = sysctl_wire_old_buffer(req, 0);
412 	if (error != 0)
413 		return (error);
414 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
415 	TAILQ_FOREACH(m, &blacklist_head, listq) {
416 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
417 		    (uintmax_t)m->phys_addr);
418 		first = 0;
419 	}
420 	error = sbuf_finish(&sbuf);
421 	sbuf_delete(&sbuf);
422 	return (error);
423 }
424 
425 /*
426  * Initialize a dummy page for use in scans of the specified paging queue.
427  * In principle, this function only needs to set the flag PG_MARKER.
428  * Nonetheless, it write busies the page as a safety precaution.
429  */
430 void
431 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
432 {
433 
434 	bzero(marker, sizeof(*marker));
435 	marker->flags = PG_MARKER;
436 	marker->a.flags = aflags;
437 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
438 	marker->a.queue = queue;
439 }
440 
441 static void
442 vm_page_domain_init(int domain)
443 {
444 	struct vm_domain *vmd;
445 	struct vm_pagequeue *pq;
446 	int i;
447 
448 	vmd = VM_DOMAIN(domain);
449 	bzero(vmd, sizeof(*vmd));
450 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
451 	    "vm inactive pagequeue";
452 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
453 	    "vm active pagequeue";
454 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
455 	    "vm laundry pagequeue";
456 	*__DECONST(const char **,
457 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
458 	    "vm unswappable pagequeue";
459 	vmd->vmd_domain = domain;
460 	vmd->vmd_page_count = 0;
461 	vmd->vmd_free_count = 0;
462 	vmd->vmd_segs = 0;
463 	vmd->vmd_oom = FALSE;
464 	for (i = 0; i < PQ_COUNT; i++) {
465 		pq = &vmd->vmd_pagequeues[i];
466 		TAILQ_INIT(&pq->pq_pl);
467 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
468 		    MTX_DEF | MTX_DUPOK);
469 		pq->pq_pdpages = 0;
470 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
471 	}
472 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
473 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
474 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
475 
476 	/*
477 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
478 	 * insertions.
479 	 */
480 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
481 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
482 	    &vmd->vmd_inacthead, plinks.q);
483 
484 	/*
485 	 * The clock pages are used to implement active queue scanning without
486 	 * requeues.  Scans start at clock[0], which is advanced after the scan
487 	 * ends.  When the two clock hands meet, they are reset and scanning
488 	 * resumes from the head of the queue.
489 	 */
490 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
491 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
492 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
493 	    &vmd->vmd_clock[0], plinks.q);
494 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
495 	    &vmd->vmd_clock[1], plinks.q);
496 }
497 
498 /*
499  * Initialize a physical page in preparation for adding it to the free
500  * lists.
501  */
502 void
503 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool)
504 {
505 	m->object = NULL;
506 	m->ref_count = 0;
507 	m->busy_lock = VPB_FREED;
508 	m->flags = m->a.flags = 0;
509 	m->phys_addr = pa;
510 	m->a.queue = PQ_NONE;
511 	m->psind = 0;
512 	m->segind = segind;
513 	m->order = VM_NFREEORDER;
514 	m->pool = pool;
515 	m->valid = m->dirty = 0;
516 	pmap_page_init(m);
517 }
518 
519 #ifndef PMAP_HAS_PAGE_ARRAY
520 static vm_paddr_t
521 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
522 {
523 	vm_paddr_t new_end;
524 
525 	/*
526 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
527 	 * However, because this page is allocated from KVM, out-of-bounds
528 	 * accesses using the direct map will not be trapped.
529 	 */
530 	*vaddr += PAGE_SIZE;
531 
532 	/*
533 	 * Allocate physical memory for the page structures, and map it.
534 	 */
535 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
536 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
537 	    VM_PROT_READ | VM_PROT_WRITE);
538 	vm_page_array_size = page_range;
539 
540 	return (new_end);
541 }
542 #endif
543 
544 /*
545  *	vm_page_startup:
546  *
547  *	Initializes the resident memory module.  Allocates physical memory for
548  *	bootstrapping UMA and some data structures that are used to manage
549  *	physical pages.  Initializes these structures, and populates the free
550  *	page queues.
551  */
552 vm_offset_t
553 vm_page_startup(vm_offset_t vaddr)
554 {
555 	struct vm_phys_seg *seg;
556 	struct vm_domain *vmd;
557 	vm_page_t m;
558 	char *list, *listend;
559 	vm_paddr_t end, high_avail, low_avail, new_end, size;
560 	vm_paddr_t page_range __unused;
561 	vm_paddr_t last_pa, pa, startp, endp;
562 	u_long pagecount;
563 #if MINIDUMP_PAGE_TRACKING
564 	u_long vm_page_dump_size;
565 #endif
566 	int biggestone, i, segind;
567 #ifdef WITNESS
568 	vm_offset_t mapped;
569 	int witness_size;
570 #endif
571 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
572 	long ii;
573 #endif
574 #ifdef VM_FREEPOOL_LAZYINIT
575 	int lazyinit;
576 #endif
577 
578 	vaddr = round_page(vaddr);
579 
580 	vm_phys_early_startup();
581 	biggestone = vm_phys_avail_largest();
582 	end = phys_avail[biggestone+1];
583 
584 	/*
585 	 * Initialize the page and queue locks.
586 	 */
587 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
588 	for (i = 0; i < PA_LOCK_COUNT; i++)
589 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
590 	for (i = 0; i < vm_ndomains; i++)
591 		vm_page_domain_init(i);
592 
593 	new_end = end;
594 #ifdef WITNESS
595 	witness_size = round_page(witness_startup_count());
596 	new_end -= witness_size;
597 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
598 	    VM_PROT_READ | VM_PROT_WRITE);
599 	bzero((void *)mapped, witness_size);
600 	witness_startup((void *)mapped);
601 #endif
602 
603 #if MINIDUMP_PAGE_TRACKING
604 	/*
605 	 * Allocate a bitmap to indicate that a random physical page
606 	 * needs to be included in a minidump.
607 	 *
608 	 * The amd64 port needs this to indicate which direct map pages
609 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
610 	 *
611 	 * However, i386 still needs this workspace internally within the
612 	 * minidump code.  In theory, they are not needed on i386, but are
613 	 * included should the sf_buf code decide to use them.
614 	 */
615 	last_pa = 0;
616 	vm_page_dump_pages = 0;
617 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
618 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
619 		    dump_avail[i] / PAGE_SIZE;
620 		if (dump_avail[i + 1] > last_pa)
621 			last_pa = dump_avail[i + 1];
622 	}
623 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
624 	new_end -= vm_page_dump_size;
625 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
626 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
627 	bzero((void *)vm_page_dump, vm_page_dump_size);
628 #if MINIDUMP_STARTUP_PAGE_TRACKING
629 	/*
630 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
631 	 * in a crash dump.  When pmap_map() uses the direct map, they are
632 	 * not automatically included.
633 	 */
634 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
635 		dump_add_page(pa);
636 #endif
637 #else
638 	(void)last_pa;
639 #endif
640 	phys_avail[biggestone + 1] = new_end;
641 #ifdef __amd64__
642 	/*
643 	 * Request that the physical pages underlying the message buffer be
644 	 * included in a crash dump.  Since the message buffer is accessed
645 	 * through the direct map, they are not automatically included.
646 	 */
647 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
648 	last_pa = pa + round_page(msgbufsize);
649 	while (pa < last_pa) {
650 		dump_add_page(pa);
651 		pa += PAGE_SIZE;
652 	}
653 #endif
654 	/*
655 	 * Compute the number of pages of memory that will be available for
656 	 * use, taking into account the overhead of a page structure per page.
657 	 * In other words, solve
658 	 *	"available physical memory" - round_page(page_range *
659 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
660 	 * for page_range.
661 	 */
662 	low_avail = phys_avail[0];
663 	high_avail = phys_avail[1];
664 	for (i = 0; i < vm_phys_nsegs; i++) {
665 		if (vm_phys_segs[i].start < low_avail)
666 			low_avail = vm_phys_segs[i].start;
667 		if (vm_phys_segs[i].end > high_avail)
668 			high_avail = vm_phys_segs[i].end;
669 	}
670 	/* Skip the first chunk.  It is already accounted for. */
671 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
672 		if (phys_avail[i] < low_avail)
673 			low_avail = phys_avail[i];
674 		if (phys_avail[i + 1] > high_avail)
675 			high_avail = phys_avail[i + 1];
676 	}
677 	first_page = low_avail / PAGE_SIZE;
678 #ifdef VM_PHYSSEG_SPARSE
679 	size = 0;
680 	for (i = 0; i < vm_phys_nsegs; i++)
681 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
682 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
683 		size += phys_avail[i + 1] - phys_avail[i];
684 #elif defined(VM_PHYSSEG_DENSE)
685 	size = high_avail - low_avail;
686 #else
687 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
688 #endif
689 
690 #ifdef PMAP_HAS_PAGE_ARRAY
691 	pmap_page_array_startup(size / PAGE_SIZE);
692 	biggestone = vm_phys_avail_largest();
693 	end = new_end = phys_avail[biggestone + 1];
694 #else
695 #ifdef VM_PHYSSEG_DENSE
696 	/*
697 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
698 	 * the overhead of a page structure per page only if vm_page_array is
699 	 * allocated from the last physical memory chunk.  Otherwise, we must
700 	 * allocate page structures representing the physical memory
701 	 * underlying vm_page_array, even though they will not be used.
702 	 */
703 	if (new_end != high_avail)
704 		page_range = size / PAGE_SIZE;
705 	else
706 #endif
707 	{
708 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
709 
710 		/*
711 		 * If the partial bytes remaining are large enough for
712 		 * a page (PAGE_SIZE) without a corresponding
713 		 * 'struct vm_page', then new_end will contain an
714 		 * extra page after subtracting the length of the VM
715 		 * page array.  Compensate by subtracting an extra
716 		 * page from new_end.
717 		 */
718 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
719 			if (new_end == high_avail)
720 				high_avail -= PAGE_SIZE;
721 			new_end -= PAGE_SIZE;
722 		}
723 	}
724 	end = new_end;
725 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
726 #endif
727 
728 #if VM_NRESERVLEVEL > 0
729 	/*
730 	 * Allocate physical memory for the reservation management system's
731 	 * data structures, and map it.
732 	 */
733 	new_end = vm_reserv_startup(&vaddr, new_end);
734 #endif
735 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
736 	/*
737 	 * Include vm_page_array and vm_reserv_array in a crash dump.
738 	 */
739 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
740 		dump_add_page(pa);
741 #endif
742 	phys_avail[biggestone + 1] = new_end;
743 
744 	/*
745 	 * Add physical memory segments corresponding to the available
746 	 * physical pages.
747 	 */
748 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
749 		if (vm_phys_avail_size(i) != 0)
750 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
751 
752 	/*
753 	 * Initialize the physical memory allocator.
754 	 */
755 	vm_phys_init();
756 
757 #ifdef VM_FREEPOOL_LAZYINIT
758 	lazyinit = 1;
759 	TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit);
760 #endif
761 
762 	/*
763 	 * Initialize the page structures and add every available page to the
764 	 * physical memory allocator's free lists.
765 	 */
766 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
767 	for (ii = 0; ii < vm_page_array_size; ii++) {
768 		m = &vm_page_array[ii];
769 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0,
770 		    VM_FREEPOOL_DEFAULT);
771 		m->flags = PG_FICTITIOUS;
772 	}
773 #endif
774 	vm_cnt.v_page_count = 0;
775 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
776 		seg = &vm_phys_segs[segind];
777 
778 		/*
779 		 * If lazy vm_page initialization is not enabled, simply
780 		 * initialize all of the pages in the segment.  Otherwise, we
781 		 * only initialize:
782 		 * 1. Pages not covered by phys_avail[], since they might be
783 		 *    freed to the allocator at some future point, e.g., by
784 		 *    kmem_bootstrap_free().
785 		 * 2. The first page of each run of free pages handed to the
786 		 *    vm_phys allocator, which in turn defers initialization
787 		 *    of pages until they are needed.
788 		 * This avoids blocking the boot process for long periods, which
789 		 * may be relevant for VMs (which ought to boot as quickly as
790 		 * possible) and/or systems with large amounts of physical
791 		 * memory.
792 		 */
793 #ifdef VM_FREEPOOL_LAZYINIT
794 		if (lazyinit) {
795 			startp = seg->start;
796 			for (i = 0; phys_avail[i + 1] != 0; i += 2) {
797 				if (startp >= seg->end)
798 					break;
799 
800 				if (phys_avail[i + 1] < startp)
801 					continue;
802 				if (phys_avail[i] <= startp) {
803 					startp = phys_avail[i + 1];
804 					continue;
805 				}
806 
807 				m = vm_phys_seg_paddr_to_vm_page(seg, startp);
808 				for (endp = MIN(phys_avail[i], seg->end);
809 				    startp < endp; startp += PAGE_SIZE, m++) {
810 					vm_page_init_page(m, startp, segind,
811 					    VM_FREEPOOL_DEFAULT);
812 				}
813 			}
814 		} else
815 #endif
816 			for (m = seg->first_page, pa = seg->start;
817 			    pa < seg->end; m++, pa += PAGE_SIZE) {
818 				vm_page_init_page(m, pa, segind,
819 				    VM_FREEPOOL_DEFAULT);
820 			}
821 
822 		/*
823 		 * Add the segment's pages that are covered by one of
824 		 * phys_avail's ranges to the free lists.
825 		 */
826 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
827 			if (seg->end <= phys_avail[i] ||
828 			    seg->start >= phys_avail[i + 1])
829 				continue;
830 
831 			startp = MAX(seg->start, phys_avail[i]);
832 			endp = MIN(seg->end, phys_avail[i + 1]);
833 			pagecount = (u_long)atop(endp - startp);
834 			if (pagecount == 0)
835 				continue;
836 
837 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
838 #ifdef VM_FREEPOOL_LAZYINIT
839 			if (lazyinit) {
840 				vm_page_init_page(m, startp, segind,
841 				    VM_FREEPOOL_LAZYINIT);
842 			}
843 #endif
844 			vmd = VM_DOMAIN(seg->domain);
845 			vm_domain_free_lock(vmd);
846 			vm_phys_enqueue_contig(m, pagecount);
847 			vm_domain_free_unlock(vmd);
848 			vm_domain_freecnt_inc(vmd, pagecount);
849 			vm_cnt.v_page_count += (u_int)pagecount;
850 			vmd->vmd_page_count += (u_int)pagecount;
851 			vmd->vmd_segs |= 1UL << segind;
852 		}
853 	}
854 
855 	/*
856 	 * Remove blacklisted pages from the physical memory allocator.
857 	 */
858 	TAILQ_INIT(&blacklist_head);
859 	vm_page_blacklist_load(&list, &listend);
860 	vm_page_blacklist_check(list, listend);
861 
862 	list = kern_getenv("vm.blacklist");
863 	vm_page_blacklist_check(list, NULL);
864 
865 	freeenv(list);
866 #if VM_NRESERVLEVEL > 0
867 	/*
868 	 * Initialize the reservation management system.
869 	 */
870 	vm_reserv_init();
871 #endif
872 
873 	return (vaddr);
874 }
875 
876 void
877 vm_page_reference(vm_page_t m)
878 {
879 
880 	vm_page_aflag_set(m, PGA_REFERENCED);
881 }
882 
883 /*
884  *	vm_page_trybusy
885  *
886  *	Helper routine for grab functions to trylock busy.
887  *
888  *	Returns true on success and false on failure.
889  */
890 static bool
891 vm_page_trybusy(vm_page_t m, int allocflags)
892 {
893 
894 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
895 		return (vm_page_trysbusy(m));
896 	else
897 		return (vm_page_tryxbusy(m));
898 }
899 
900 /*
901  *	vm_page_tryacquire
902  *
903  *	Helper routine for grab functions to trylock busy and wire.
904  *
905  *	Returns true on success and false on failure.
906  */
907 static inline bool
908 vm_page_tryacquire(vm_page_t m, int allocflags)
909 {
910 	bool locked;
911 
912 	locked = vm_page_trybusy(m, allocflags);
913 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
914 		vm_page_wire(m);
915 	return (locked);
916 }
917 
918 /*
919  *	vm_page_busy_acquire:
920  *
921  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
922  *	and drop the object lock if necessary.
923  */
924 bool
925 vm_page_busy_acquire(vm_page_t m, int allocflags)
926 {
927 	vm_object_t obj;
928 	bool locked;
929 
930 	/*
931 	 * The page-specific object must be cached because page
932 	 * identity can change during the sleep, causing the
933 	 * re-lock of a different object.
934 	 * It is assumed that a reference to the object is already
935 	 * held by the callers.
936 	 */
937 	obj = atomic_load_ptr(&m->object);
938 	for (;;) {
939 		if (vm_page_tryacquire(m, allocflags))
940 			return (true);
941 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
942 			return (false);
943 		if (obj != NULL)
944 			locked = VM_OBJECT_WOWNED(obj);
945 		else
946 			locked = false;
947 		MPASS(locked || vm_page_wired(m));
948 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
949 		    locked) && locked)
950 			VM_OBJECT_WLOCK(obj);
951 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
952 			return (false);
953 		KASSERT(m->object == obj || m->object == NULL,
954 		    ("vm_page_busy_acquire: page %p does not belong to %p",
955 		    m, obj));
956 	}
957 }
958 
959 /*
960  *	vm_page_busy_downgrade:
961  *
962  *	Downgrade an exclusive busy page into a single shared busy page.
963  */
964 void
965 vm_page_busy_downgrade(vm_page_t m)
966 {
967 	u_int x;
968 
969 	vm_page_assert_xbusied(m);
970 
971 	x = vm_page_busy_fetch(m);
972 	for (;;) {
973 		if (atomic_fcmpset_rel_int(&m->busy_lock,
974 		    &x, VPB_SHARERS_WORD(1)))
975 			break;
976 	}
977 	if ((x & VPB_BIT_WAITERS) != 0)
978 		wakeup(m);
979 }
980 
981 /*
982  *
983  *	vm_page_busy_tryupgrade:
984  *
985  *	Attempt to upgrade a single shared busy into an exclusive busy.
986  */
987 int
988 vm_page_busy_tryupgrade(vm_page_t m)
989 {
990 	u_int ce, x;
991 
992 	vm_page_assert_sbusied(m);
993 
994 	x = vm_page_busy_fetch(m);
995 	ce = VPB_CURTHREAD_EXCLUSIVE;
996 	for (;;) {
997 		if (VPB_SHARERS(x) > 1)
998 			return (0);
999 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1000 		    ("vm_page_busy_tryupgrade: invalid lock state"));
1001 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1002 		    ce | (x & VPB_BIT_WAITERS)))
1003 			continue;
1004 		return (1);
1005 	}
1006 }
1007 
1008 /*
1009  *	vm_page_sbusied:
1010  *
1011  *	Return a positive value if the page is shared busied, 0 otherwise.
1012  */
1013 int
1014 vm_page_sbusied(vm_page_t m)
1015 {
1016 	u_int x;
1017 
1018 	x = vm_page_busy_fetch(m);
1019 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1020 }
1021 
1022 /*
1023  *	vm_page_sunbusy:
1024  *
1025  *	Shared unbusy a page.
1026  */
1027 void
1028 vm_page_sunbusy(vm_page_t m)
1029 {
1030 	u_int x;
1031 
1032 	vm_page_assert_sbusied(m);
1033 
1034 	x = vm_page_busy_fetch(m);
1035 	for (;;) {
1036 		KASSERT(x != VPB_FREED,
1037 		    ("vm_page_sunbusy: Unlocking freed page."));
1038 		if (VPB_SHARERS(x) > 1) {
1039 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1040 			    x - VPB_ONE_SHARER))
1041 				break;
1042 			continue;
1043 		}
1044 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1045 		    ("vm_page_sunbusy: invalid lock state"));
1046 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1047 			continue;
1048 		if ((x & VPB_BIT_WAITERS) == 0)
1049 			break;
1050 		wakeup(m);
1051 		break;
1052 	}
1053 }
1054 
1055 /*
1056  *	vm_page_busy_sleep:
1057  *
1058  *	Sleep if the page is busy, using the page pointer as wchan.
1059  *	This is used to implement the hard-path of the busying mechanism.
1060  *
1061  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1062  *	will not sleep if the page is shared-busy.
1063  *
1064  *	The object lock must be held on entry.
1065  *
1066  *	Returns true if it slept and dropped the object lock, or false
1067  *	if there was no sleep and the lock is still held.
1068  */
1069 bool
1070 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1071 {
1072 	vm_object_t obj;
1073 
1074 	obj = m->object;
1075 	VM_OBJECT_ASSERT_LOCKED(obj);
1076 
1077 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1078 	    true));
1079 }
1080 
1081 /*
1082  *	vm_page_busy_sleep_unlocked:
1083  *
1084  *	Sleep if the page is busy, using the page pointer as wchan.
1085  *	This is used to implement the hard-path of busying mechanism.
1086  *
1087  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1088  *	will not sleep if the page is shared-busy.
1089  *
1090  *	The object lock must not be held on entry.  The operation will
1091  *	return if the page changes identity.
1092  */
1093 void
1094 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1095     const char *wmesg, int allocflags)
1096 {
1097 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1098 
1099 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1100 }
1101 
1102 /*
1103  *	_vm_page_busy_sleep:
1104  *
1105  *	Internal busy sleep function.  Verifies the page identity and
1106  *	lockstate against parameters.  Returns true if it sleeps and
1107  *	false otherwise.
1108  *
1109  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1110  *
1111  *	If locked is true the lock will be dropped for any true returns
1112  *	and held for any false returns.
1113  */
1114 static bool
1115 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1116     const char *wmesg, int allocflags, bool locked)
1117 {
1118 	bool xsleep;
1119 	u_int x;
1120 
1121 	/*
1122 	 * If the object is busy we must wait for that to drain to zero
1123 	 * before trying the page again.
1124 	 */
1125 	if (obj != NULL && vm_object_busied(obj)) {
1126 		if (locked)
1127 			VM_OBJECT_DROP(obj);
1128 		vm_object_busy_wait(obj, wmesg);
1129 		return (true);
1130 	}
1131 
1132 	if (!vm_page_busied(m))
1133 		return (false);
1134 
1135 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1136 	sleepq_lock(m);
1137 	x = vm_page_busy_fetch(m);
1138 	do {
1139 		/*
1140 		 * If the page changes objects or becomes unlocked we can
1141 		 * simply return.
1142 		 */
1143 		if (x == VPB_UNBUSIED ||
1144 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1145 		    m->object != obj || m->pindex != pindex) {
1146 			sleepq_release(m);
1147 			return (false);
1148 		}
1149 		if ((x & VPB_BIT_WAITERS) != 0)
1150 			break;
1151 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1152 	if (locked)
1153 		VM_OBJECT_DROP(obj);
1154 	DROP_GIANT();
1155 	sleepq_add(m, NULL, wmesg, 0, 0);
1156 	sleepq_wait(m, PVM);
1157 	PICKUP_GIANT();
1158 	return (true);
1159 }
1160 
1161 /*
1162  *	vm_page_trysbusy:
1163  *
1164  *	Try to shared busy a page.
1165  *	If the operation succeeds 1 is returned otherwise 0.
1166  *	The operation never sleeps.
1167  */
1168 int
1169 vm_page_trysbusy(vm_page_t m)
1170 {
1171 	vm_object_t obj;
1172 	u_int x;
1173 
1174 	obj = m->object;
1175 	x = vm_page_busy_fetch(m);
1176 	for (;;) {
1177 		if ((x & VPB_BIT_SHARED) == 0)
1178 			return (0);
1179 		/*
1180 		 * Reduce the window for transient busies that will trigger
1181 		 * false negatives in vm_page_ps_test().
1182 		 */
1183 		if (obj != NULL && vm_object_busied(obj))
1184 			return (0);
1185 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1186 		    x + VPB_ONE_SHARER))
1187 			break;
1188 	}
1189 
1190 	/* Refetch the object now that we're guaranteed that it is stable. */
1191 	obj = m->object;
1192 	if (obj != NULL && vm_object_busied(obj)) {
1193 		vm_page_sunbusy(m);
1194 		return (0);
1195 	}
1196 	return (1);
1197 }
1198 
1199 /*
1200  *	vm_page_tryxbusy:
1201  *
1202  *	Try to exclusive busy a page.
1203  *	If the operation succeeds 1 is returned otherwise 0.
1204  *	The operation never sleeps.
1205  */
1206 int
1207 vm_page_tryxbusy(vm_page_t m)
1208 {
1209 	vm_object_t obj;
1210 
1211         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1212             VPB_CURTHREAD_EXCLUSIVE) == 0)
1213 		return (0);
1214 
1215 	obj = m->object;
1216 	if (obj != NULL && vm_object_busied(obj)) {
1217 		vm_page_xunbusy(m);
1218 		return (0);
1219 	}
1220 	return (1);
1221 }
1222 
1223 static void
1224 vm_page_xunbusy_hard_tail(vm_page_t m)
1225 {
1226 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1227 	/* Wake the waiter. */
1228 	wakeup(m);
1229 }
1230 
1231 /*
1232  *	vm_page_xunbusy_hard:
1233  *
1234  *	Called when unbusy has failed because there is a waiter.
1235  */
1236 void
1237 vm_page_xunbusy_hard(vm_page_t m)
1238 {
1239 	vm_page_assert_xbusied(m);
1240 	vm_page_xunbusy_hard_tail(m);
1241 }
1242 
1243 void
1244 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1245 {
1246 	vm_page_assert_xbusied_unchecked(m);
1247 	vm_page_xunbusy_hard_tail(m);
1248 }
1249 
1250 static void
1251 vm_page_busy_free(vm_page_t m)
1252 {
1253 	u_int x;
1254 
1255 	atomic_thread_fence_rel();
1256 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1257 	if ((x & VPB_BIT_WAITERS) != 0)
1258 		wakeup(m);
1259 }
1260 
1261 /*
1262  *	vm_page_unhold_pages:
1263  *
1264  *	Unhold each of the pages that is referenced by the given array.
1265  */
1266 void
1267 vm_page_unhold_pages(vm_page_t *ma, int count)
1268 {
1269 
1270 	for (; count != 0; count--) {
1271 		vm_page_unwire(*ma, PQ_ACTIVE);
1272 		ma++;
1273 	}
1274 }
1275 
1276 vm_page_t
1277 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1278 {
1279 	vm_page_t m;
1280 
1281 #ifdef VM_PHYSSEG_SPARSE
1282 	m = vm_phys_paddr_to_vm_page(pa);
1283 	if (m == NULL)
1284 		m = vm_phys_fictitious_to_vm_page(pa);
1285 	return (m);
1286 #elif defined(VM_PHYSSEG_DENSE)
1287 	long pi;
1288 
1289 	pi = atop(pa);
1290 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1291 		m = &vm_page_array[pi - first_page];
1292 		return (m);
1293 	}
1294 	return (vm_phys_fictitious_to_vm_page(pa));
1295 #else
1296 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1297 #endif
1298 }
1299 
1300 /*
1301  *	vm_page_getfake:
1302  *
1303  *	Create a fictitious page with the specified physical address and
1304  *	memory attribute.  The memory attribute is the only the machine-
1305  *	dependent aspect of a fictitious page that must be initialized.
1306  */
1307 vm_page_t
1308 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1309 {
1310 	vm_page_t m;
1311 
1312 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1313 	vm_page_initfake(m, paddr, memattr);
1314 	return (m);
1315 }
1316 
1317 void
1318 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1319 {
1320 
1321 	if ((m->flags & PG_FICTITIOUS) != 0) {
1322 		/*
1323 		 * The page's memattr might have changed since the
1324 		 * previous initialization.  Update the pmap to the
1325 		 * new memattr.
1326 		 */
1327 		goto memattr;
1328 	}
1329 	m->phys_addr = paddr;
1330 	m->a.queue = PQ_NONE;
1331 	/* Fictitious pages don't use "segind". */
1332 	m->flags = PG_FICTITIOUS;
1333 	/* Fictitious pages don't use "order" or "pool". */
1334 	m->oflags = VPO_UNMANAGED;
1335 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1336 	/* Fictitious pages are unevictable. */
1337 	m->ref_count = 1;
1338 	pmap_page_init(m);
1339 memattr:
1340 	pmap_page_set_memattr(m, memattr);
1341 }
1342 
1343 /*
1344  *	vm_page_putfake:
1345  *
1346  *	Release a fictitious page.
1347  */
1348 void
1349 vm_page_putfake(vm_page_t m)
1350 {
1351 
1352 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1353 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1354 	    ("vm_page_putfake: bad page %p", m));
1355 	vm_page_assert_xbusied(m);
1356 	vm_page_busy_free(m);
1357 	uma_zfree(fakepg_zone, m);
1358 }
1359 
1360 /*
1361  *	vm_page_updatefake:
1362  *
1363  *	Update the given fictitious page to the specified physical address and
1364  *	memory attribute.
1365  */
1366 void
1367 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1368 {
1369 
1370 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1371 	    ("vm_page_updatefake: bad page %p", m));
1372 	m->phys_addr = paddr;
1373 	pmap_page_set_memattr(m, memattr);
1374 }
1375 
1376 /*
1377  *	vm_page_free:
1378  *
1379  *	Free a page.
1380  */
1381 void
1382 vm_page_free(vm_page_t m)
1383 {
1384 
1385 	m->flags &= ~PG_ZERO;
1386 	vm_page_free_toq(m);
1387 }
1388 
1389 /*
1390  *	vm_page_free_zero:
1391  *
1392  *	Free a page to the zerod-pages queue
1393  */
1394 void
1395 vm_page_free_zero(vm_page_t m)
1396 {
1397 
1398 	m->flags |= PG_ZERO;
1399 	vm_page_free_toq(m);
1400 }
1401 
1402 /*
1403  * Unbusy and handle the page queueing for a page from a getpages request that
1404  * was optionally read ahead or behind.
1405  */
1406 void
1407 vm_page_readahead_finish(vm_page_t m)
1408 {
1409 
1410 	/* We shouldn't put invalid pages on queues. */
1411 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1412 
1413 	/*
1414 	 * Since the page is not the actually needed one, whether it should
1415 	 * be activated or deactivated is not obvious.  Empirical results
1416 	 * have shown that deactivating the page is usually the best choice,
1417 	 * unless the page is wanted by another thread.
1418 	 */
1419 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1420 		vm_page_activate(m);
1421 	else
1422 		vm_page_deactivate(m);
1423 	vm_page_xunbusy_unchecked(m);
1424 }
1425 
1426 /*
1427  * Destroy the identity of an invalid page and free it if possible.
1428  * This is intended to be used when reading a page from backing store fails.
1429  */
1430 void
1431 vm_page_free_invalid(vm_page_t m)
1432 {
1433 
1434 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1435 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1436 	KASSERT(m->object != NULL, ("page %p has no object", m));
1437 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1438 
1439 	/*
1440 	 * We may be attempting to free the page as part of the handling for an
1441 	 * I/O error, in which case the page was xbusied by a different thread.
1442 	 */
1443 	vm_page_xbusy_claim(m);
1444 
1445 	/*
1446 	 * If someone has wired this page while the object lock
1447 	 * was not held, then the thread that unwires is responsible
1448 	 * for freeing the page.  Otherwise just free the page now.
1449 	 * The wire count of this unmapped page cannot change while
1450 	 * we have the page xbusy and the page's object wlocked.
1451 	 */
1452 	if (vm_page_remove(m))
1453 		vm_page_free(m);
1454 }
1455 
1456 /*
1457  *	vm_page_dirty_KBI:		[ internal use only ]
1458  *
1459  *	Set all bits in the page's dirty field.
1460  *
1461  *	The object containing the specified page must be locked if the
1462  *	call is made from the machine-independent layer.
1463  *
1464  *	See vm_page_clear_dirty_mask().
1465  *
1466  *	This function should only be called by vm_page_dirty().
1467  */
1468 void
1469 vm_page_dirty_KBI(vm_page_t m)
1470 {
1471 
1472 	/* Refer to this operation by its public name. */
1473 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1474 	m->dirty = VM_PAGE_BITS_ALL;
1475 }
1476 
1477 /*
1478  * Insert the given page into the given object at the given pindex.  mpred is
1479  * used for memq linkage.  From vm_page_insert, lookup is true, mpred is
1480  * initially NULL, and this procedure looks it up.  From vm_page_insert_after
1481  * and vm_page_iter_insert, lookup is false and mpred is known to the caller
1482  * to be valid, and may be NULL if this will be the page with the lowest
1483  * pindex.
1484  *
1485  * The procedure is marked __always_inline to suggest to the compiler to
1486  * eliminate the lookup parameter and the associated alternate branch.
1487  */
1488 static __always_inline int
1489 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1490     struct pctrie_iter *pages, bool iter, vm_page_t mpred, bool lookup)
1491 {
1492 	int error;
1493 
1494 	VM_OBJECT_ASSERT_WLOCKED(object);
1495 	KASSERT(m->object == NULL,
1496 	    ("vm_page_insert: page %p already inserted", m));
1497 
1498 	/*
1499 	 * Record the object/offset pair in this page.
1500 	 */
1501 	m->object = object;
1502 	m->pindex = pindex;
1503 	m->ref_count |= VPRC_OBJREF;
1504 
1505 	/*
1506 	 * Add this page to the object's radix tree, and look up mpred if
1507 	 * needed.
1508 	 */
1509 	if (iter) {
1510 		KASSERT(!lookup, ("%s: cannot lookup mpred", __func__));
1511 		error = vm_radix_iter_insert(pages, m);
1512 	} else if (lookup)
1513 		error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred);
1514 	else
1515 		error = vm_radix_insert(&object->rtree, m);
1516 	if (__predict_false(error != 0)) {
1517 		m->object = NULL;
1518 		m->pindex = 0;
1519 		m->ref_count &= ~VPRC_OBJREF;
1520 		return (1);
1521 	}
1522 
1523 	/*
1524 	 * Now link into the object's ordered list of backed pages.
1525 	 */
1526 	vm_page_insert_radixdone(m, object, mpred);
1527 	vm_pager_page_inserted(object, m);
1528 	return (0);
1529 }
1530 
1531 /*
1532  *	vm_page_insert:		[ internal use only ]
1533  *
1534  *	Inserts the given mem entry into the object and object list.
1535  *
1536  *	The object must be locked.
1537  */
1538 int
1539 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1540 {
1541 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL,
1542 	    true));
1543 }
1544 
1545 /*
1546  *	vm_page_insert_after:
1547  *
1548  *	Inserts the page "m" into the specified object at offset "pindex".
1549  *
1550  *	The page "mpred" must immediately precede the offset "pindex" within
1551  *	the specified object.
1552  *
1553  *	The object must be locked.
1554  */
1555 static int
1556 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1557     vm_page_t mpred)
1558 {
1559 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, mpred,
1560 	    false));
1561 }
1562 
1563 /*
1564  *	vm_page_iter_insert:
1565  *
1566  *	Tries to insert the page "m" into the specified object at offset
1567  *	"pindex" using the iterator "pages".  Returns 0 if the insertion was
1568  *	successful.
1569  *
1570  *	The page "mpred" must immediately precede the offset "pindex" within
1571  *	the specified object.
1572  *
1573  *	The object must be locked.
1574  */
1575 static int
1576 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object,
1577     vm_pindex_t pindex, vm_page_t mpred)
1578 {
1579 	return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred,
1580 	    false));
1581 }
1582 
1583 /*
1584  *	vm_page_insert_radixdone:
1585  *
1586  *	Complete page "m" insertion into the specified object after the
1587  *	radix trie hooking.
1588  *
1589  *	The page "mpred" must precede the offset "m->pindex" within the
1590  *	specified object.
1591  *
1592  *	The object must be locked.
1593  */
1594 static void
1595 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1596 {
1597 
1598 	VM_OBJECT_ASSERT_WLOCKED(object);
1599 	KASSERT(object != NULL && m->object == object,
1600 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1601 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1602 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1603 	if (mpred != NULL) {
1604 		KASSERT(mpred->object == object,
1605 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1606 		KASSERT(mpred->pindex < m->pindex,
1607 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1608 		KASSERT(TAILQ_NEXT(mpred, listq) == NULL ||
1609 		    m->pindex < TAILQ_NEXT(mpred, listq)->pindex,
1610 		    ("vm_page_insert_radixdone: pindex doesn't precede msucc"));
1611 	} else {
1612 		KASSERT(TAILQ_EMPTY(&object->memq) ||
1613 		    m->pindex < TAILQ_FIRST(&object->memq)->pindex,
1614 		    ("vm_page_insert_radixdone: no mpred but not first page"));
1615 	}
1616 
1617 	if (mpred != NULL)
1618 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1619 	else
1620 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1621 
1622 	/*
1623 	 * Show that the object has one more resident page.
1624 	 */
1625 	object->resident_page_count++;
1626 
1627 	/*
1628 	 * Hold the vnode until the last page is released.
1629 	 */
1630 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1631 		vhold(object->handle);
1632 
1633 	/*
1634 	 * Since we are inserting a new and possibly dirty page,
1635 	 * update the object's generation count.
1636 	 */
1637 	if (pmap_page_is_write_mapped(m))
1638 		vm_object_set_writeable_dirty(object);
1639 }
1640 
1641 /*
1642  * Do the work to remove a page from its object.  The caller is responsible for
1643  * updating the page's fields to reflect this removal.
1644  */
1645 static void
1646 vm_page_object_remove(vm_page_t m)
1647 {
1648 	vm_object_t object;
1649 	vm_page_t mrem __diagused;
1650 
1651 	vm_page_assert_xbusied(m);
1652 	object = m->object;
1653 	VM_OBJECT_ASSERT_WLOCKED(object);
1654 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1655 	    ("page %p is missing its object ref", m));
1656 
1657 	/* Deferred free of swap space. */
1658 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1659 		vm_pager_page_unswapped(m);
1660 
1661 	vm_pager_page_removed(object, m);
1662 
1663 	m->object = NULL;
1664 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1665 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1666 
1667 	/*
1668 	 * Now remove from the object's list of backed pages.
1669 	 */
1670 	TAILQ_REMOVE(&object->memq, m, listq);
1671 
1672 	/*
1673 	 * And show that the object has one fewer resident page.
1674 	 */
1675 	object->resident_page_count--;
1676 
1677 	/*
1678 	 * The vnode may now be recycled.
1679 	 */
1680 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1681 		vdrop(object->handle);
1682 }
1683 
1684 /*
1685  *	vm_page_remove:
1686  *
1687  *	Removes the specified page from its containing object, but does not
1688  *	invalidate any backing storage.  Returns true if the object's reference
1689  *	was the last reference to the page, and false otherwise.
1690  *
1691  *	The object must be locked and the page must be exclusively busied.
1692  *	The exclusive busy will be released on return.  If this is not the
1693  *	final ref and the caller does not hold a wire reference it may not
1694  *	continue to access the page.
1695  */
1696 bool
1697 vm_page_remove(vm_page_t m)
1698 {
1699 	bool dropped;
1700 
1701 	dropped = vm_page_remove_xbusy(m);
1702 	vm_page_xunbusy(m);
1703 
1704 	return (dropped);
1705 }
1706 
1707 /*
1708  *	vm_page_remove_xbusy
1709  *
1710  *	Removes the page but leaves the xbusy held.  Returns true if this
1711  *	removed the final ref and false otherwise.
1712  */
1713 bool
1714 vm_page_remove_xbusy(vm_page_t m)
1715 {
1716 
1717 	vm_page_object_remove(m);
1718 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1719 }
1720 
1721 /*
1722  *	vm_page_lookup:
1723  *
1724  *	Returns the page associated with the object/offset
1725  *	pair specified; if none is found, NULL is returned.
1726  *
1727  *	The object must be locked.
1728  */
1729 vm_page_t
1730 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1731 {
1732 
1733 	VM_OBJECT_ASSERT_LOCKED(object);
1734 	return (vm_radix_lookup(&object->rtree, pindex));
1735 }
1736 
1737 /*
1738  *	vm_page_iter_init:
1739  *
1740  *	Initialize iterator for vm pages.
1741  */
1742 void
1743 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object)
1744 {
1745 
1746 	vm_radix_iter_init(pages, &object->rtree);
1747 }
1748 
1749 /*
1750  *	vm_page_iter_init:
1751  *
1752  *	Initialize iterator for vm pages.
1753  */
1754 void
1755 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object,
1756     vm_pindex_t limit)
1757 {
1758 
1759 	vm_radix_iter_limit_init(pages, &object->rtree, limit);
1760 }
1761 
1762 /*
1763  *	vm_page_iter_lookup:
1764  *
1765  *	Returns the page associated with the object/offset pair specified, and
1766  *	stores the path to its position; if none is found, NULL is returned.
1767  *
1768  *	The iter pctrie must be locked.
1769  */
1770 vm_page_t
1771 vm_page_iter_lookup(struct pctrie_iter *pages, vm_pindex_t pindex)
1772 {
1773 
1774 	return (vm_radix_iter_lookup(pages, pindex));
1775 }
1776 
1777 /*
1778  *	vm_page_lookup_unlocked:
1779  *
1780  *	Returns the page associated with the object/offset pair specified;
1781  *	if none is found, NULL is returned.  The page may be no longer be
1782  *	present in the object at the time that this function returns.  Only
1783  *	useful for opportunistic checks such as inmem().
1784  */
1785 vm_page_t
1786 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1787 {
1788 
1789 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1790 }
1791 
1792 /*
1793  *	vm_page_relookup:
1794  *
1795  *	Returns a page that must already have been busied by
1796  *	the caller.  Used for bogus page replacement.
1797  */
1798 vm_page_t
1799 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1800 {
1801 	vm_page_t m;
1802 
1803 	m = vm_page_lookup_unlocked(object, pindex);
1804 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1805 	    m->object == object && m->pindex == pindex,
1806 	    ("vm_page_relookup: Invalid page %p", m));
1807 	return (m);
1808 }
1809 
1810 /*
1811  * This should only be used by lockless functions for releasing transient
1812  * incorrect acquires.  The page may have been freed after we acquired a
1813  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1814  * further to do.
1815  */
1816 static void
1817 vm_page_busy_release(vm_page_t m)
1818 {
1819 	u_int x;
1820 
1821 	x = vm_page_busy_fetch(m);
1822 	for (;;) {
1823 		if (x == VPB_FREED)
1824 			break;
1825 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1826 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1827 			    x - VPB_ONE_SHARER))
1828 				break;
1829 			continue;
1830 		}
1831 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1832 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1833 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1834 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1835 			continue;
1836 		if ((x & VPB_BIT_WAITERS) != 0)
1837 			wakeup(m);
1838 		break;
1839 	}
1840 }
1841 
1842 /*
1843  *	vm_page_find_least:
1844  *
1845  *	Returns the page associated with the object with least pindex
1846  *	greater than or equal to the parameter pindex, or NULL.
1847  *
1848  *	The object must be locked.
1849  */
1850 vm_page_t
1851 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1852 {
1853 	vm_page_t m;
1854 
1855 	VM_OBJECT_ASSERT_LOCKED(object);
1856 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1857 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1858 	return (m);
1859 }
1860 
1861 /*
1862  *	vm_page_iter_lookup_ge:
1863  *
1864  *	Returns the page associated with the object with least pindex
1865  *	greater than or equal to the parameter pindex, or NULL.  Initializes the
1866  *	iterator to point to that page.
1867  *
1868  *	The iter pctrie must be locked.
1869  */
1870 vm_page_t
1871 vm_page_iter_lookup_ge(struct pctrie_iter *pages, vm_pindex_t pindex)
1872 {
1873 
1874 	return (vm_radix_iter_lookup_ge(pages, pindex));
1875 }
1876 
1877 /*
1878  * Returns the given page's successor (by pindex) within the object if it is
1879  * resident; if none is found, NULL is returned.
1880  *
1881  * The object must be locked.
1882  */
1883 vm_page_t
1884 vm_page_next(vm_page_t m)
1885 {
1886 	vm_page_t next;
1887 
1888 	VM_OBJECT_ASSERT_LOCKED(m->object);
1889 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1890 		MPASS(next->object == m->object);
1891 		if (next->pindex != m->pindex + 1)
1892 			next = NULL;
1893 	}
1894 	return (next);
1895 }
1896 
1897 /*
1898  * Returns the given page's predecessor (by pindex) within the object if it is
1899  * resident; if none is found, NULL is returned.
1900  *
1901  * The object must be locked.
1902  */
1903 vm_page_t
1904 vm_page_prev(vm_page_t m)
1905 {
1906 	vm_page_t prev;
1907 
1908 	VM_OBJECT_ASSERT_LOCKED(m->object);
1909 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1910 		MPASS(prev->object == m->object);
1911 		if (prev->pindex != m->pindex - 1)
1912 			prev = NULL;
1913 	}
1914 	return (prev);
1915 }
1916 
1917 /*
1918  * Uses the page mnew as a replacement for an existing page at index
1919  * pindex which must be already present in the object.
1920  *
1921  * Both pages must be exclusively busied on enter.  The old page is
1922  * unbusied on exit.
1923  *
1924  * A return value of true means mold is now free.  If this is not the
1925  * final ref and the caller does not hold a wire reference it may not
1926  * continue to access the page.
1927  */
1928 static bool
1929 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1930     vm_page_t mold)
1931 {
1932 	vm_page_t mret __diagused;
1933 	bool dropped;
1934 
1935 	VM_OBJECT_ASSERT_WLOCKED(object);
1936 	vm_page_assert_xbusied(mold);
1937 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1938 	    ("vm_page_replace: page %p already in object", mnew));
1939 
1940 	/*
1941 	 * This function mostly follows vm_page_insert() and
1942 	 * vm_page_remove() without the radix, object count and vnode
1943 	 * dance.  Double check such functions for more comments.
1944 	 */
1945 
1946 	mnew->object = object;
1947 	mnew->pindex = pindex;
1948 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1949 	mret = vm_radix_replace(&object->rtree, mnew);
1950 	KASSERT(mret == mold,
1951 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1952 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1953 	    (mnew->oflags & VPO_UNMANAGED),
1954 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1955 
1956 	/* Keep the resident page list in sorted order. */
1957 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1958 	TAILQ_REMOVE(&object->memq, mold, listq);
1959 	mold->object = NULL;
1960 
1961 	/*
1962 	 * The object's resident_page_count does not change because we have
1963 	 * swapped one page for another, but the generation count should
1964 	 * change if the page is dirty.
1965 	 */
1966 	if (pmap_page_is_write_mapped(mnew))
1967 		vm_object_set_writeable_dirty(object);
1968 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1969 	vm_page_xunbusy(mold);
1970 
1971 	return (dropped);
1972 }
1973 
1974 void
1975 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1976     vm_page_t mold)
1977 {
1978 
1979 	vm_page_assert_xbusied(mnew);
1980 
1981 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1982 		vm_page_free(mold);
1983 }
1984 
1985 /*
1986  *	vm_page_rename:
1987  *
1988  *	Move the given memory entry from its
1989  *	current object to the specified target object/offset.
1990  *
1991  *	Note: swap associated with the page must be invalidated by the move.  We
1992  *	      have to do this for several reasons:  (1) we aren't freeing the
1993  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1994  *	      moving the page from object A to B, and will then later move
1995  *	      the backing store from A to B and we can't have a conflict.
1996  *
1997  *	Note: we *always* dirty the page.  It is necessary both for the
1998  *	      fact that we moved it, and because we may be invalidating
1999  *	      swap.
2000  *
2001  *	The objects must be locked.
2002  */
2003 int
2004 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
2005 {
2006 	vm_page_t mpred;
2007 	vm_pindex_t opidx;
2008 
2009 	VM_OBJECT_ASSERT_WLOCKED(new_object);
2010 
2011 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
2012 
2013 	/*
2014 	 * Create a custom version of vm_page_insert() which does not depend
2015 	 * by m_prev and can cheat on the implementation aspects of the
2016 	 * function.
2017 	 */
2018 	opidx = m->pindex;
2019 	m->pindex = new_pindex;
2020 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
2021 		m->pindex = opidx;
2022 		return (1);
2023 	}
2024 
2025 	/*
2026 	 * The operation cannot fail anymore.  The removal must happen before
2027 	 * the listq iterator is tainted.
2028 	 */
2029 	m->pindex = opidx;
2030 	vm_page_object_remove(m);
2031 
2032 	/* Return back to the new pindex to complete vm_page_insert(). */
2033 	m->pindex = new_pindex;
2034 	m->object = new_object;
2035 
2036 	vm_page_insert_radixdone(m, new_object, mpred);
2037 	vm_page_dirty(m);
2038 	vm_pager_page_inserted(new_object, m);
2039 	return (0);
2040 }
2041 
2042 /*
2043  *	vm_page_mpred:
2044  *
2045  *	Return the greatest page of the object with index <= pindex,
2046  *	or NULL, if there is none.  Assumes object lock is held.
2047  */
2048 vm_page_t
2049 vm_page_mpred(vm_object_t object, vm_pindex_t pindex)
2050 {
2051 	return (vm_radix_lookup_le(&object->rtree, pindex));
2052 }
2053 
2054 /*
2055  *	vm_page_alloc:
2056  *
2057  *	Allocate and return a page that is associated with the specified
2058  *	object and offset pair.  By default, this page is exclusive busied.
2059  *
2060  *	The caller must always specify an allocation class.
2061  *
2062  *	allocation classes:
2063  *	VM_ALLOC_NORMAL		normal process request
2064  *	VM_ALLOC_SYSTEM		system *really* needs a page
2065  *	VM_ALLOC_INTERRUPT	interrupt time request
2066  *
2067  *	optional allocation flags:
2068  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2069  *				intends to allocate
2070  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2071  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2072  *	VM_ALLOC_SBUSY		shared busy the allocated page
2073  *	VM_ALLOC_WIRED		wire the allocated page
2074  *	VM_ALLOC_ZERO		prefer a zeroed page
2075  */
2076 vm_page_t
2077 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
2078 {
2079 
2080 	return (vm_page_alloc_after(object, pindex, req,
2081 	    vm_page_mpred(object, pindex)));
2082 }
2083 
2084 /*
2085  * Allocate a page in the specified object with the given page index.  To
2086  * optimize insertion of the page into the object, the caller must also specify
2087  * the resident page in the object with largest index smaller than the given
2088  * page index, or NULL if no such page exists.
2089  */
2090 static vm_page_t
2091 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
2092     int req, vm_page_t mpred)
2093 {
2094 	struct vm_domainset_iter di;
2095 	vm_page_t m;
2096 	int domain;
2097 
2098 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2099 	do {
2100 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2101 		    mpred);
2102 		if (m != NULL)
2103 			break;
2104 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2105 
2106 	return (m);
2107 }
2108 
2109 /*
2110  * Returns true if the number of free pages exceeds the minimum
2111  * for the request class and false otherwise.
2112  */
2113 static int
2114 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2115 {
2116 	u_int limit, old, new;
2117 
2118 	if (req_class == VM_ALLOC_INTERRUPT)
2119 		limit = 0;
2120 	else if (req_class == VM_ALLOC_SYSTEM)
2121 		limit = vmd->vmd_interrupt_free_min;
2122 	else
2123 		limit = vmd->vmd_free_reserved;
2124 
2125 	/*
2126 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2127 	 */
2128 	limit += npages;
2129 	old = atomic_load_int(&vmd->vmd_free_count);
2130 	do {
2131 		if (old < limit)
2132 			return (0);
2133 		new = old - npages;
2134 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2135 
2136 	/* Wake the page daemon if we've crossed the threshold. */
2137 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2138 		pagedaemon_wakeup(vmd->vmd_domain);
2139 
2140 	/* Only update bitsets on transitions. */
2141 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2142 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2143 		vm_domain_set(vmd);
2144 
2145 	return (1);
2146 }
2147 
2148 int
2149 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2150 {
2151 	int req_class;
2152 
2153 	/*
2154 	 * The page daemon is allowed to dig deeper into the free page list.
2155 	 */
2156 	req_class = req & VM_ALLOC_CLASS_MASK;
2157 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2158 		req_class = VM_ALLOC_SYSTEM;
2159 	return (_vm_domain_allocate(vmd, req_class, npages));
2160 }
2161 
2162 vm_page_t
2163 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2164     int req, vm_page_t mpred)
2165 {
2166 	struct vm_domain *vmd;
2167 	vm_page_t m;
2168 	int flags;
2169 
2170 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2171 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2172 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2173 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2174 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2175 	KASSERT((req & ~VPA_FLAGS) == 0,
2176 	    ("invalid request %#x", req));
2177 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2178 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2179 	    ("invalid request %#x", req));
2180 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2181 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2182 	    (uintmax_t)pindex));
2183 	VM_OBJECT_ASSERT_WLOCKED(object);
2184 
2185 	flags = 0;
2186 	m = NULL;
2187 	if (!vm_pager_can_alloc_page(object, pindex))
2188 		return (NULL);
2189 again:
2190 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2191 		m = vm_page_alloc_nofree_domain(domain, req);
2192 		if (m != NULL)
2193 			goto found;
2194 	}
2195 #if VM_NRESERVLEVEL > 0
2196 	/*
2197 	 * Can we allocate the page from a reservation?
2198 	 */
2199 	if (vm_object_reserv(object) &&
2200 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2201 	    NULL) {
2202 		goto found;
2203 	}
2204 #endif
2205 	vmd = VM_DOMAIN(domain);
2206 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2207 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2208 		    M_NOWAIT | M_NOVM);
2209 		if (m != NULL) {
2210 			flags |= PG_PCPU_CACHE;
2211 			goto found;
2212 		}
2213 	}
2214 	if (vm_domain_allocate(vmd, req, 1)) {
2215 		/*
2216 		 * If not, allocate it from the free page queues.
2217 		 */
2218 		vm_domain_free_lock(vmd);
2219 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2220 		vm_domain_free_unlock(vmd);
2221 		if (m == NULL) {
2222 			vm_domain_freecnt_inc(vmd, 1);
2223 #if VM_NRESERVLEVEL > 0
2224 			if (vm_reserv_reclaim_inactive(domain))
2225 				goto again;
2226 #endif
2227 		}
2228 	}
2229 	if (m == NULL) {
2230 		/*
2231 		 * Not allocatable, give up.
2232 		 */
2233 		if (vm_domain_alloc_fail(vmd, object, req))
2234 			goto again;
2235 		return (NULL);
2236 	}
2237 
2238 	/*
2239 	 * At this point we had better have found a good page.
2240 	 */
2241 found:
2242 	vm_page_dequeue(m);
2243 	vm_page_alloc_check(m);
2244 
2245 	/*
2246 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2247 	 */
2248 	flags |= m->flags & PG_ZERO;
2249 	if ((req & VM_ALLOC_NODUMP) != 0)
2250 		flags |= PG_NODUMP;
2251 	if ((req & VM_ALLOC_NOFREE) != 0)
2252 		flags |= PG_NOFREE;
2253 	m->flags = flags;
2254 	m->a.flags = 0;
2255 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2256 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2257 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2258 	else if ((req & VM_ALLOC_SBUSY) != 0)
2259 		m->busy_lock = VPB_SHARERS_WORD(1);
2260 	else
2261 		m->busy_lock = VPB_UNBUSIED;
2262 	if (req & VM_ALLOC_WIRED) {
2263 		vm_wire_add(1);
2264 		m->ref_count = 1;
2265 	}
2266 	m->a.act_count = 0;
2267 
2268 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2269 		if (req & VM_ALLOC_WIRED) {
2270 			vm_wire_sub(1);
2271 			m->ref_count = 0;
2272 		}
2273 		KASSERT(m->object == NULL, ("page %p has object", m));
2274 		m->oflags = VPO_UNMANAGED;
2275 		m->busy_lock = VPB_UNBUSIED;
2276 		/* Don't change PG_ZERO. */
2277 		vm_page_free_toq(m);
2278 		if (req & VM_ALLOC_WAITFAIL) {
2279 			VM_OBJECT_WUNLOCK(object);
2280 			vm_radix_wait();
2281 			VM_OBJECT_WLOCK(object);
2282 		}
2283 		return (NULL);
2284 	}
2285 
2286 	/* Ignore device objects; the pager sets "memattr" for them. */
2287 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2288 	    (object->flags & OBJ_FICTITIOUS) == 0)
2289 		pmap_page_set_memattr(m, object->memattr);
2290 
2291 	return (m);
2292 }
2293 
2294 /*
2295  *	vm_page_alloc_contig:
2296  *
2297  *	Allocate a contiguous set of physical pages of the given size "npages"
2298  *	from the free lists.  All of the physical pages must be at or above
2299  *	the given physical address "low" and below the given physical address
2300  *	"high".  The given value "alignment" determines the alignment of the
2301  *	first physical page in the set.  If the given value "boundary" is
2302  *	non-zero, then the set of physical pages cannot cross any physical
2303  *	address boundary that is a multiple of that value.  Both "alignment"
2304  *	and "boundary" must be a power of two.
2305  *
2306  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2307  *	then the memory attribute setting for the physical pages is configured
2308  *	to the object's memory attribute setting.  Otherwise, the memory
2309  *	attribute setting for the physical pages is configured to "memattr",
2310  *	overriding the object's memory attribute setting.  However, if the
2311  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2312  *	memory attribute setting for the physical pages cannot be configured
2313  *	to VM_MEMATTR_DEFAULT.
2314  *
2315  *	The specified object may not contain fictitious pages.
2316  *
2317  *	The caller must always specify an allocation class.
2318  *
2319  *	allocation classes:
2320  *	VM_ALLOC_NORMAL		normal process request
2321  *	VM_ALLOC_SYSTEM		system *really* needs a page
2322  *	VM_ALLOC_INTERRUPT	interrupt time request
2323  *
2324  *	optional allocation flags:
2325  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2326  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2327  *	VM_ALLOC_SBUSY		shared busy the allocated page
2328  *	VM_ALLOC_WIRED		wire the allocated page
2329  *	VM_ALLOC_ZERO		prefer a zeroed page
2330  */
2331 vm_page_t
2332 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2333     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2334     vm_paddr_t boundary, vm_memattr_t memattr)
2335 {
2336 	struct vm_domainset_iter di;
2337 	vm_page_t bounds[2];
2338 	vm_page_t m;
2339 	int domain;
2340 	int start_segind;
2341 
2342 	start_segind = -1;
2343 
2344 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2345 	do {
2346 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2347 		    npages, low, high, alignment, boundary, memattr);
2348 		if (m != NULL)
2349 			break;
2350 		if (start_segind == -1)
2351 			start_segind = vm_phys_lookup_segind(low);
2352 		if (vm_phys_find_range(bounds, start_segind, domain,
2353 		    npages, low, high) == -1) {
2354 			vm_domainset_iter_ignore(&di, domain);
2355 		}
2356 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2357 
2358 	return (m);
2359 }
2360 
2361 static vm_page_t
2362 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2363     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2364 {
2365 	struct vm_domain *vmd;
2366 	vm_page_t m_ret;
2367 
2368 	/*
2369 	 * Can we allocate the pages without the number of free pages falling
2370 	 * below the lower bound for the allocation class?
2371 	 */
2372 	vmd = VM_DOMAIN(domain);
2373 	if (!vm_domain_allocate(vmd, req, npages))
2374 		return (NULL);
2375 	/*
2376 	 * Try to allocate the pages from the free page queues.
2377 	 */
2378 	vm_domain_free_lock(vmd);
2379 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2380 	    alignment, boundary);
2381 	vm_domain_free_unlock(vmd);
2382 	if (m_ret != NULL)
2383 		return (m_ret);
2384 #if VM_NRESERVLEVEL > 0
2385 	/*
2386 	 * Try to break a reservation to allocate the pages.
2387 	 */
2388 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2389 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2390 	            high, alignment, boundary);
2391 		if (m_ret != NULL)
2392 			return (m_ret);
2393 	}
2394 #endif
2395 	vm_domain_freecnt_inc(vmd, npages);
2396 	return (NULL);
2397 }
2398 
2399 vm_page_t
2400 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2401     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2402     vm_paddr_t boundary, vm_memattr_t memattr)
2403 {
2404 	struct pctrie_iter pages;
2405 	vm_page_t m, m_ret, mpred;
2406 	u_int busy_lock, flags, oflags;
2407 
2408 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2409 	KASSERT((req & ~VPAC_FLAGS) == 0,
2410 	    ("invalid request %#x", req));
2411 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2412 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2413 	    ("invalid request %#x", req));
2414 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2415 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2416 	    ("invalid request %#x", req));
2417 	VM_OBJECT_ASSERT_WLOCKED(object);
2418 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2419 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2420 	    object));
2421 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2422 
2423 	vm_page_iter_init(&pages, object);
2424 	mpred = vm_radix_iter_lookup_le(&pages, pindex);
2425 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2426 	    ("vm_page_alloc_contig: pindex already allocated"));
2427 	for (;;) {
2428 #if VM_NRESERVLEVEL > 0
2429 		/*
2430 		 * Can we allocate the pages from a reservation?
2431 		 */
2432 		if (vm_object_reserv(object) &&
2433 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2434 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2435 			break;
2436 		}
2437 #endif
2438 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2439 		    low, high, alignment, boundary)) != NULL)
2440 			break;
2441 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2442 			return (NULL);
2443 	}
2444 
2445 	/*
2446 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2447 	 */
2448 	flags = PG_ZERO;
2449 	if ((req & VM_ALLOC_NODUMP) != 0)
2450 		flags |= PG_NODUMP;
2451 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2452 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2453 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2454 	else if ((req & VM_ALLOC_SBUSY) != 0)
2455 		busy_lock = VPB_SHARERS_WORD(1);
2456 	else
2457 		busy_lock = VPB_UNBUSIED;
2458 	if ((req & VM_ALLOC_WIRED) != 0)
2459 		vm_wire_add(npages);
2460 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2461 	    memattr == VM_MEMATTR_DEFAULT)
2462 		memattr = object->memattr;
2463 	for (m = m_ret; m < &m_ret[npages]; m++) {
2464 		vm_page_dequeue(m);
2465 		vm_page_alloc_check(m);
2466 		m->a.flags = 0;
2467 		m->flags = (m->flags | PG_NODUMP) & flags;
2468 		m->busy_lock = busy_lock;
2469 		if ((req & VM_ALLOC_WIRED) != 0)
2470 			m->ref_count = 1;
2471 		m->a.act_count = 0;
2472 		m->oflags = oflags;
2473 		if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) {
2474 			if ((req & VM_ALLOC_WIRED) != 0)
2475 				vm_wire_sub(npages);
2476 			KASSERT(m->object == NULL,
2477 			    ("page %p has object", m));
2478 			mpred = m;
2479 			for (m = m_ret; m < &m_ret[npages]; m++) {
2480 				if (m <= mpred &&
2481 				    (req & VM_ALLOC_WIRED) != 0)
2482 					m->ref_count = 0;
2483 				m->oflags = VPO_UNMANAGED;
2484 				m->busy_lock = VPB_UNBUSIED;
2485 				/* Don't change PG_ZERO. */
2486 				vm_page_free_toq(m);
2487 			}
2488 			if (req & VM_ALLOC_WAITFAIL) {
2489 				VM_OBJECT_WUNLOCK(object);
2490 				vm_radix_wait();
2491 				VM_OBJECT_WLOCK(object);
2492 			}
2493 			return (NULL);
2494 		}
2495 		mpred = m;
2496 		if (memattr != VM_MEMATTR_DEFAULT)
2497 			pmap_page_set_memattr(m, memattr);
2498 		pindex++;
2499 	}
2500 	return (m_ret);
2501 }
2502 
2503 /*
2504  * Allocate a physical page that is not intended to be inserted into a VM
2505  * object.
2506  */
2507 vm_page_t
2508 vm_page_alloc_noobj_domain(int domain, int req)
2509 {
2510 	struct vm_domain *vmd;
2511 	vm_page_t m;
2512 	int flags;
2513 
2514 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2515 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2516 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2517 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2518 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2519 	KASSERT((req & ~VPAN_FLAGS) == 0,
2520 	    ("invalid request %#x", req));
2521 
2522 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2523 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2524 	vmd = VM_DOMAIN(domain);
2525 again:
2526 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2527 		m = vm_page_alloc_nofree_domain(domain, req);
2528 		if (m != NULL)
2529 			goto found;
2530 	}
2531 
2532 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2533 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2534 		    M_NOWAIT | M_NOVM);
2535 		if (m != NULL) {
2536 			flags |= PG_PCPU_CACHE;
2537 			goto found;
2538 		}
2539 	}
2540 
2541 	if (vm_domain_allocate(vmd, req, 1)) {
2542 		vm_domain_free_lock(vmd);
2543 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2544 		vm_domain_free_unlock(vmd);
2545 		if (m == NULL) {
2546 			vm_domain_freecnt_inc(vmd, 1);
2547 #if VM_NRESERVLEVEL > 0
2548 			if (vm_reserv_reclaim_inactive(domain))
2549 				goto again;
2550 #endif
2551 		}
2552 	}
2553 	if (m == NULL) {
2554 		if (vm_domain_alloc_fail(vmd, NULL, req))
2555 			goto again;
2556 		return (NULL);
2557 	}
2558 
2559 found:
2560 	vm_page_dequeue(m);
2561 	vm_page_alloc_check(m);
2562 
2563 	/*
2564 	 * Consumers should not rely on a useful default pindex value.
2565 	 */
2566 	m->pindex = 0xdeadc0dedeadc0de;
2567 	m->flags = (m->flags & PG_ZERO) | flags;
2568 	m->a.flags = 0;
2569 	m->oflags = VPO_UNMANAGED;
2570 	m->busy_lock = VPB_UNBUSIED;
2571 	if ((req & VM_ALLOC_WIRED) != 0) {
2572 		vm_wire_add(1);
2573 		m->ref_count = 1;
2574 	}
2575 
2576 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2577 		pmap_zero_page(m);
2578 
2579 	return (m);
2580 }
2581 
2582 #if VM_NRESERVLEVEL > 1
2583 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2584 #elif VM_NRESERVLEVEL > 0
2585 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2586 #else
2587 #define	VM_NOFREE_IMPORT_ORDER	8
2588 #endif
2589 
2590 /*
2591  * Allocate a single NOFREE page.
2592  *
2593  * This routine hands out NOFREE pages from higher-order
2594  * physical memory blocks in order to reduce memory fragmentation.
2595  * When a NOFREE for a given domain chunk is used up,
2596  * the routine will try to fetch a new one from the freelists
2597  * and discard the old one.
2598  */
2599 static vm_page_t
2600 vm_page_alloc_nofree_domain(int domain, int req)
2601 {
2602 	vm_page_t m;
2603 	struct vm_domain *vmd;
2604 	struct vm_nofreeq *nqp;
2605 
2606 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2607 
2608 	vmd = VM_DOMAIN(domain);
2609 	nqp = &vmd->vmd_nofreeq;
2610 	vm_domain_free_lock(vmd);
2611 	if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) {
2612 		if (!vm_domain_allocate(vmd, req,
2613 		    1 << VM_NOFREE_IMPORT_ORDER)) {
2614 			vm_domain_free_unlock(vmd);
2615 			return (NULL);
2616 		}
2617 		nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2618 		    VM_NOFREE_IMPORT_ORDER);
2619 		if (nqp->ma == NULL) {
2620 			vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER);
2621 			vm_domain_free_unlock(vmd);
2622 			return (NULL);
2623 		}
2624 		nqp->offs = 0;
2625 	}
2626 	m = &nqp->ma[nqp->offs++];
2627 	vm_domain_free_unlock(vmd);
2628 	VM_CNT_ADD(v_nofree_count, 1);
2629 
2630 	return (m);
2631 }
2632 
2633 vm_page_t
2634 vm_page_alloc_noobj(int req)
2635 {
2636 	struct vm_domainset_iter di;
2637 	vm_page_t m;
2638 	int domain;
2639 
2640 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2641 	do {
2642 		m = vm_page_alloc_noobj_domain(domain, req);
2643 		if (m != NULL)
2644 			break;
2645 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2646 
2647 	return (m);
2648 }
2649 
2650 vm_page_t
2651 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2652     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2653     vm_memattr_t memattr)
2654 {
2655 	struct vm_domainset_iter di;
2656 	vm_page_t m;
2657 	int domain;
2658 
2659 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2660 	do {
2661 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2662 		    high, alignment, boundary, memattr);
2663 		if (m != NULL)
2664 			break;
2665 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2666 
2667 	return (m);
2668 }
2669 
2670 vm_page_t
2671 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2672     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2673     vm_memattr_t memattr)
2674 {
2675 	vm_page_t m, m_ret;
2676 	u_int flags;
2677 
2678 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2679 	KASSERT((req & ~VPANC_FLAGS) == 0,
2680 	    ("invalid request %#x", req));
2681 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2682 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2683 	    ("invalid request %#x", req));
2684 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2685 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2686 	    ("invalid request %#x", req));
2687 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2688 
2689 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2690 	    low, high, alignment, boundary)) == NULL) {
2691 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2692 			return (NULL);
2693 	}
2694 
2695 	/*
2696 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2697 	 */
2698 	flags = PG_ZERO;
2699 	if ((req & VM_ALLOC_NODUMP) != 0)
2700 		flags |= PG_NODUMP;
2701 	if ((req & VM_ALLOC_WIRED) != 0)
2702 		vm_wire_add(npages);
2703 	for (m = m_ret; m < &m_ret[npages]; m++) {
2704 		vm_page_dequeue(m);
2705 		vm_page_alloc_check(m);
2706 
2707 		/*
2708 		 * Consumers should not rely on a useful default pindex value.
2709 		 */
2710 		m->pindex = 0xdeadc0dedeadc0de;
2711 		m->a.flags = 0;
2712 		m->flags = (m->flags | PG_NODUMP) & flags;
2713 		m->busy_lock = VPB_UNBUSIED;
2714 		if ((req & VM_ALLOC_WIRED) != 0)
2715 			m->ref_count = 1;
2716 		m->a.act_count = 0;
2717 		m->oflags = VPO_UNMANAGED;
2718 
2719 		/*
2720 		 * Zero the page before updating any mappings since the page is
2721 		 * not yet shared with any devices which might require the
2722 		 * non-default memory attribute.  pmap_page_set_memattr()
2723 		 * flushes data caches before returning.
2724 		 */
2725 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2726 			pmap_zero_page(m);
2727 		if (memattr != VM_MEMATTR_DEFAULT)
2728 			pmap_page_set_memattr(m, memattr);
2729 	}
2730 	return (m_ret);
2731 }
2732 
2733 /*
2734  * Check a page that has been freshly dequeued from a freelist.
2735  */
2736 static void
2737 vm_page_alloc_check(vm_page_t m)
2738 {
2739 
2740 	KASSERT(m->object == NULL, ("page %p has object", m));
2741 	KASSERT(m->a.queue == PQ_NONE &&
2742 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2743 	    ("page %p has unexpected queue %d, flags %#x",
2744 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2745 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2746 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2747 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2748 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2749 	    ("page %p has unexpected memattr %d",
2750 	    m, pmap_page_get_memattr(m)));
2751 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2752 	pmap_vm_page_alloc_check(m);
2753 }
2754 
2755 static int
2756 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2757 {
2758 	struct vm_domain *vmd;
2759 	struct vm_pgcache *pgcache;
2760 	int i;
2761 
2762 	pgcache = arg;
2763 	vmd = VM_DOMAIN(pgcache->domain);
2764 
2765 	/*
2766 	 * The page daemon should avoid creating extra memory pressure since its
2767 	 * main purpose is to replenish the store of free pages.
2768 	 */
2769 	if (vmd->vmd_severeset || curproc == pageproc ||
2770 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2771 		return (0);
2772 	domain = vmd->vmd_domain;
2773 	vm_domain_free_lock(vmd);
2774 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2775 	    (vm_page_t *)store);
2776 	vm_domain_free_unlock(vmd);
2777 	if (cnt != i)
2778 		vm_domain_freecnt_inc(vmd, cnt - i);
2779 
2780 	return (i);
2781 }
2782 
2783 static void
2784 vm_page_zone_release(void *arg, void **store, int cnt)
2785 {
2786 	struct vm_domain *vmd;
2787 	struct vm_pgcache *pgcache;
2788 	vm_page_t m;
2789 	int i;
2790 
2791 	pgcache = arg;
2792 	vmd = VM_DOMAIN(pgcache->domain);
2793 	vm_domain_free_lock(vmd);
2794 	for (i = 0; i < cnt; i++) {
2795 		m = (vm_page_t)store[i];
2796 		vm_phys_free_pages(m, 0);
2797 	}
2798 	vm_domain_free_unlock(vmd);
2799 	vm_domain_freecnt_inc(vmd, cnt);
2800 }
2801 
2802 #define	VPSC_ANY	0	/* No restrictions. */
2803 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2804 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2805 
2806 /*
2807  *	vm_page_scan_contig:
2808  *
2809  *	Scan vm_page_array[] between the specified entries "m_start" and
2810  *	"m_end" for a run of contiguous physical pages that satisfy the
2811  *	specified conditions, and return the lowest page in the run.  The
2812  *	specified "alignment" determines the alignment of the lowest physical
2813  *	page in the run.  If the specified "boundary" is non-zero, then the
2814  *	run of physical pages cannot span a physical address that is a
2815  *	multiple of "boundary".
2816  *
2817  *	"m_end" is never dereferenced, so it need not point to a vm_page
2818  *	structure within vm_page_array[].
2819  *
2820  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2821  *	span a hole (or discontiguity) in the physical address space.  Both
2822  *	"alignment" and "boundary" must be a power of two.
2823  */
2824 static vm_page_t
2825 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2826     u_long alignment, vm_paddr_t boundary, int options)
2827 {
2828 	vm_object_t object;
2829 	vm_paddr_t pa;
2830 	vm_page_t m, m_run;
2831 #if VM_NRESERVLEVEL > 0
2832 	int level;
2833 #endif
2834 	int m_inc, order, run_ext, run_len;
2835 
2836 	KASSERT(npages > 0, ("npages is 0"));
2837 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2838 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2839 	m_run = NULL;
2840 	run_len = 0;
2841 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2842 		KASSERT((m->flags & PG_MARKER) == 0,
2843 		    ("page %p is PG_MARKER", m));
2844 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2845 		    ("fictitious page %p has invalid ref count", m));
2846 
2847 		/*
2848 		 * If the current page would be the start of a run, check its
2849 		 * physical address against the end, alignment, and boundary
2850 		 * conditions.  If it doesn't satisfy these conditions, either
2851 		 * terminate the scan or advance to the next page that
2852 		 * satisfies the failed condition.
2853 		 */
2854 		if (run_len == 0) {
2855 			KASSERT(m_run == NULL, ("m_run != NULL"));
2856 			if (m + npages > m_end)
2857 				break;
2858 			pa = VM_PAGE_TO_PHYS(m);
2859 			if (!vm_addr_align_ok(pa, alignment)) {
2860 				m_inc = atop(roundup2(pa, alignment) - pa);
2861 				continue;
2862 			}
2863 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2864 				m_inc = atop(roundup2(pa, boundary) - pa);
2865 				continue;
2866 			}
2867 		} else
2868 			KASSERT(m_run != NULL, ("m_run == NULL"));
2869 
2870 retry:
2871 		m_inc = 1;
2872 		if (vm_page_wired(m))
2873 			run_ext = 0;
2874 #if VM_NRESERVLEVEL > 0
2875 		else if ((level = vm_reserv_level(m)) >= 0 &&
2876 		    (options & VPSC_NORESERV) != 0) {
2877 			run_ext = 0;
2878 			/* Advance to the end of the reservation. */
2879 			pa = VM_PAGE_TO_PHYS(m);
2880 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2881 			    pa);
2882 		}
2883 #endif
2884 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2885 			/*
2886 			 * The page is considered eligible for relocation if
2887 			 * and only if it could be laundered or reclaimed by
2888 			 * the page daemon.
2889 			 */
2890 			VM_OBJECT_RLOCK(object);
2891 			if (object != m->object) {
2892 				VM_OBJECT_RUNLOCK(object);
2893 				goto retry;
2894 			}
2895 			/* Don't care: PG_NODUMP, PG_ZERO. */
2896 			if ((object->flags & OBJ_SWAP) == 0 &&
2897 			    object->type != OBJT_VNODE) {
2898 				run_ext = 0;
2899 #if VM_NRESERVLEVEL > 0
2900 			} else if ((options & VPSC_NOSUPER) != 0 &&
2901 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2902 				run_ext = 0;
2903 				/* Advance to the end of the superpage. */
2904 				pa = VM_PAGE_TO_PHYS(m);
2905 				m_inc = atop(roundup2(pa + 1,
2906 				    vm_reserv_size(level)) - pa);
2907 #endif
2908 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2909 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2910 				/*
2911 				 * The page is allocated but eligible for
2912 				 * relocation.  Extend the current run by one
2913 				 * page.
2914 				 */
2915 				KASSERT(pmap_page_get_memattr(m) ==
2916 				    VM_MEMATTR_DEFAULT,
2917 				    ("page %p has an unexpected memattr", m));
2918 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2919 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2920 				    ("page %p has unexpected oflags", m));
2921 				/* Don't care: PGA_NOSYNC. */
2922 				run_ext = 1;
2923 			} else
2924 				run_ext = 0;
2925 			VM_OBJECT_RUNLOCK(object);
2926 #if VM_NRESERVLEVEL > 0
2927 		} else if (level >= 0) {
2928 			/*
2929 			 * The page is reserved but not yet allocated.  In
2930 			 * other words, it is still free.  Extend the current
2931 			 * run by one page.
2932 			 */
2933 			run_ext = 1;
2934 #endif
2935 		} else if ((order = m->order) < VM_NFREEORDER) {
2936 			/*
2937 			 * The page is enqueued in the physical memory
2938 			 * allocator's free page queues.  Moreover, it is the
2939 			 * first page in a power-of-two-sized run of
2940 			 * contiguous free pages.  Add these pages to the end
2941 			 * of the current run, and jump ahead.
2942 			 */
2943 			run_ext = 1 << order;
2944 			m_inc = 1 << order;
2945 		} else {
2946 			/*
2947 			 * Skip the page for one of the following reasons: (1)
2948 			 * It is enqueued in the physical memory allocator's
2949 			 * free page queues.  However, it is not the first
2950 			 * page in a run of contiguous free pages.  (This case
2951 			 * rarely occurs because the scan is performed in
2952 			 * ascending order.) (2) It is not reserved, and it is
2953 			 * transitioning from free to allocated.  (Conversely,
2954 			 * the transition from allocated to free for managed
2955 			 * pages is blocked by the page busy lock.) (3) It is
2956 			 * allocated but not contained by an object and not
2957 			 * wired, e.g., allocated by Xen's balloon driver.
2958 			 */
2959 			run_ext = 0;
2960 		}
2961 
2962 		/*
2963 		 * Extend or reset the current run of pages.
2964 		 */
2965 		if (run_ext > 0) {
2966 			if (run_len == 0)
2967 				m_run = m;
2968 			run_len += run_ext;
2969 		} else {
2970 			if (run_len > 0) {
2971 				m_run = NULL;
2972 				run_len = 0;
2973 			}
2974 		}
2975 	}
2976 	if (run_len >= npages)
2977 		return (m_run);
2978 	return (NULL);
2979 }
2980 
2981 /*
2982  *	vm_page_reclaim_run:
2983  *
2984  *	Try to relocate each of the allocated virtual pages within the
2985  *	specified run of physical pages to a new physical address.  Free the
2986  *	physical pages underlying the relocated virtual pages.  A virtual page
2987  *	is relocatable if and only if it could be laundered or reclaimed by
2988  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2989  *	physical address above "high".
2990  *
2991  *	Returns 0 if every physical page within the run was already free or
2992  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2993  *	value indicating why the last attempt to relocate a virtual page was
2994  *	unsuccessful.
2995  *
2996  *	"req_class" must be an allocation class.
2997  */
2998 static int
2999 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
3000     vm_paddr_t high)
3001 {
3002 	struct vm_domain *vmd;
3003 	struct spglist free;
3004 	vm_object_t object;
3005 	vm_paddr_t pa;
3006 	vm_page_t m, m_end, m_new;
3007 	int error, order, req;
3008 
3009 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
3010 	    ("req_class is not an allocation class"));
3011 	SLIST_INIT(&free);
3012 	error = 0;
3013 	m = m_run;
3014 	m_end = m_run + npages;
3015 	for (; error == 0 && m < m_end; m++) {
3016 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
3017 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
3018 
3019 		/*
3020 		 * Racily check for wirings.  Races are handled once the object
3021 		 * lock is held and the page is unmapped.
3022 		 */
3023 		if (vm_page_wired(m))
3024 			error = EBUSY;
3025 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
3026 			/*
3027 			 * The page is relocated if and only if it could be
3028 			 * laundered or reclaimed by the page daemon.
3029 			 */
3030 			VM_OBJECT_WLOCK(object);
3031 			/* Don't care: PG_NODUMP, PG_ZERO. */
3032 			if (m->object != object ||
3033 			    ((object->flags & OBJ_SWAP) == 0 &&
3034 			    object->type != OBJT_VNODE))
3035 				error = EINVAL;
3036 			else if (object->memattr != VM_MEMATTR_DEFAULT)
3037 				error = EINVAL;
3038 			else if (vm_page_queue(m) != PQ_NONE &&
3039 			    vm_page_tryxbusy(m) != 0) {
3040 				if (vm_page_wired(m)) {
3041 					vm_page_xunbusy(m);
3042 					error = EBUSY;
3043 					goto unlock;
3044 				}
3045 				KASSERT(pmap_page_get_memattr(m) ==
3046 				    VM_MEMATTR_DEFAULT,
3047 				    ("page %p has an unexpected memattr", m));
3048 				KASSERT(m->oflags == 0,
3049 				    ("page %p has unexpected oflags", m));
3050 				/* Don't care: PGA_NOSYNC. */
3051 				if (!vm_page_none_valid(m)) {
3052 					/*
3053 					 * First, try to allocate a new page
3054 					 * that is above "high".  Failing
3055 					 * that, try to allocate a new page
3056 					 * that is below "m_run".  Allocate
3057 					 * the new page between the end of
3058 					 * "m_run" and "high" only as a last
3059 					 * resort.
3060 					 */
3061 					req = req_class;
3062 					if ((m->flags & PG_NODUMP) != 0)
3063 						req |= VM_ALLOC_NODUMP;
3064 					if (trunc_page(high) !=
3065 					    ~(vm_paddr_t)PAGE_MASK) {
3066 						m_new =
3067 						    vm_page_alloc_noobj_contig(
3068 						    req, 1, round_page(high),
3069 						    ~(vm_paddr_t)0, PAGE_SIZE,
3070 						    0, VM_MEMATTR_DEFAULT);
3071 					} else
3072 						m_new = NULL;
3073 					if (m_new == NULL) {
3074 						pa = VM_PAGE_TO_PHYS(m_run);
3075 						m_new =
3076 						    vm_page_alloc_noobj_contig(
3077 						    req, 1, 0, pa - 1,
3078 						    PAGE_SIZE, 0,
3079 						    VM_MEMATTR_DEFAULT);
3080 					}
3081 					if (m_new == NULL) {
3082 						pa += ptoa(npages);
3083 						m_new =
3084 						    vm_page_alloc_noobj_contig(
3085 						    req, 1, pa, high, PAGE_SIZE,
3086 						    0, VM_MEMATTR_DEFAULT);
3087 					}
3088 					if (m_new == NULL) {
3089 						vm_page_xunbusy(m);
3090 						error = ENOMEM;
3091 						goto unlock;
3092 					}
3093 
3094 					/*
3095 					 * Unmap the page and check for new
3096 					 * wirings that may have been acquired
3097 					 * through a pmap lookup.
3098 					 */
3099 					if (object->ref_count != 0 &&
3100 					    !vm_page_try_remove_all(m)) {
3101 						vm_page_xunbusy(m);
3102 						vm_page_free(m_new);
3103 						error = EBUSY;
3104 						goto unlock;
3105 					}
3106 
3107 					/*
3108 					 * Replace "m" with the new page.  For
3109 					 * vm_page_replace(), "m" must be busy
3110 					 * and dequeued.  Finally, change "m"
3111 					 * as if vm_page_free() was called.
3112 					 */
3113 					m_new->a.flags = m->a.flags &
3114 					    ~PGA_QUEUE_STATE_MASK;
3115 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3116 					    ("page %p is managed", m_new));
3117 					m_new->oflags = 0;
3118 					pmap_copy_page(m, m_new);
3119 					m_new->valid = m->valid;
3120 					m_new->dirty = m->dirty;
3121 					m->flags &= ~PG_ZERO;
3122 					vm_page_dequeue(m);
3123 					if (vm_page_replace_hold(m_new, object,
3124 					    m->pindex, m) &&
3125 					    vm_page_free_prep(m))
3126 						SLIST_INSERT_HEAD(&free, m,
3127 						    plinks.s.ss);
3128 
3129 					/*
3130 					 * The new page must be deactivated
3131 					 * before the object is unlocked.
3132 					 */
3133 					vm_page_deactivate(m_new);
3134 				} else {
3135 					m->flags &= ~PG_ZERO;
3136 					vm_page_dequeue(m);
3137 					if (vm_page_free_prep(m))
3138 						SLIST_INSERT_HEAD(&free, m,
3139 						    plinks.s.ss);
3140 					KASSERT(m->dirty == 0,
3141 					    ("page %p is dirty", m));
3142 				}
3143 			} else
3144 				error = EBUSY;
3145 unlock:
3146 			VM_OBJECT_WUNLOCK(object);
3147 		} else {
3148 			MPASS(vm_page_domain(m) == domain);
3149 			vmd = VM_DOMAIN(domain);
3150 			vm_domain_free_lock(vmd);
3151 			order = m->order;
3152 			if (order < VM_NFREEORDER) {
3153 				/*
3154 				 * The page is enqueued in the physical memory
3155 				 * allocator's free page queues.  Moreover, it
3156 				 * is the first page in a power-of-two-sized
3157 				 * run of contiguous free pages.  Jump ahead
3158 				 * to the last page within that run, and
3159 				 * continue from there.
3160 				 */
3161 				m += (1 << order) - 1;
3162 			}
3163 #if VM_NRESERVLEVEL > 0
3164 			else if (vm_reserv_is_page_free(m))
3165 				order = 0;
3166 #endif
3167 			vm_domain_free_unlock(vmd);
3168 			if (order == VM_NFREEORDER)
3169 				error = EINVAL;
3170 		}
3171 	}
3172 	if ((m = SLIST_FIRST(&free)) != NULL) {
3173 		int cnt;
3174 
3175 		vmd = VM_DOMAIN(domain);
3176 		cnt = 0;
3177 		vm_domain_free_lock(vmd);
3178 		do {
3179 			MPASS(vm_page_domain(m) == domain);
3180 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3181 			vm_phys_free_pages(m, 0);
3182 			cnt++;
3183 		} while ((m = SLIST_FIRST(&free)) != NULL);
3184 		vm_domain_free_unlock(vmd);
3185 		vm_domain_freecnt_inc(vmd, cnt);
3186 	}
3187 	return (error);
3188 }
3189 
3190 #define	NRUNS	16
3191 
3192 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3193 
3194 #define	MIN_RECLAIM	8
3195 
3196 /*
3197  *	vm_page_reclaim_contig:
3198  *
3199  *	Reclaim allocated, contiguous physical memory satisfying the specified
3200  *	conditions by relocating the virtual pages using that physical memory.
3201  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3202  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3203  *	currently able to reclaim the requested number of pages.  Since
3204  *	relocation requires the allocation of physical pages, reclamation may
3205  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3206  *	fails in this manner, callers are expected to perform vm_wait() before
3207  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3208  *
3209  *	The caller must always specify an allocation class through "req".
3210  *
3211  *	allocation classes:
3212  *	VM_ALLOC_NORMAL		normal process request
3213  *	VM_ALLOC_SYSTEM		system *really* needs a page
3214  *	VM_ALLOC_INTERRUPT	interrupt time request
3215  *
3216  *	The optional allocation flags are ignored.
3217  *
3218  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3219  *	must be a power of two.
3220  */
3221 int
3222 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3223     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3224     int desired_runs)
3225 {
3226 	struct vm_domain *vmd;
3227 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3228 	u_long count, minalign, reclaimed;
3229 	int error, i, min_reclaim, nruns, options, req_class;
3230 	int segind, start_segind;
3231 	int ret;
3232 
3233 	KASSERT(npages > 0, ("npages is 0"));
3234 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3235 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3236 
3237 	ret = ENOMEM;
3238 
3239 	/*
3240 	 * If the caller wants to reclaim multiple runs, try to allocate
3241 	 * space to store the runs.  If that fails, fall back to the old
3242 	 * behavior of just reclaiming MIN_RECLAIM pages.
3243 	 */
3244 	if (desired_runs > 1)
3245 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3246 		    M_TEMP, M_NOWAIT);
3247 	else
3248 		m_runs = NULL;
3249 
3250 	if (m_runs == NULL) {
3251 		m_runs = _m_runs;
3252 		nruns = NRUNS;
3253 	} else {
3254 		nruns = NRUNS + desired_runs - 1;
3255 	}
3256 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3257 
3258 	/*
3259 	 * The caller will attempt an allocation after some runs have been
3260 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3261 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3262 	 * power of two or maximum chunk size, and ensure that each run is
3263 	 * suitably aligned.
3264 	 */
3265 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3266 	npages = roundup2(npages, minalign);
3267 	if (alignment < ptoa(minalign))
3268 		alignment = ptoa(minalign);
3269 
3270 	/*
3271 	 * The page daemon is allowed to dig deeper into the free page list.
3272 	 */
3273 	req_class = req & VM_ALLOC_CLASS_MASK;
3274 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3275 		req_class = VM_ALLOC_SYSTEM;
3276 
3277 	start_segind = vm_phys_lookup_segind(low);
3278 
3279 	/*
3280 	 * Return if the number of free pages cannot satisfy the requested
3281 	 * allocation.
3282 	 */
3283 	vmd = VM_DOMAIN(domain);
3284 	count = vmd->vmd_free_count;
3285 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3286 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3287 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3288 		goto done;
3289 
3290 	/*
3291 	 * Scan up to three times, relaxing the restrictions ("options") on
3292 	 * the reclamation of reservations and superpages each time.
3293 	 */
3294 	for (options = VPSC_NORESERV;;) {
3295 		bool phys_range_exists = false;
3296 
3297 		/*
3298 		 * Find the highest runs that satisfy the given constraints
3299 		 * and restrictions, and record them in "m_runs".
3300 		 */
3301 		count = 0;
3302 		segind = start_segind;
3303 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3304 		    npages, low, high)) != -1) {
3305 			phys_range_exists = true;
3306 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3307 			    bounds[1], alignment, boundary, options))) {
3308 				bounds[0] = m_run + npages;
3309 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3310 				count++;
3311 			}
3312 			segind++;
3313 		}
3314 
3315 		if (!phys_range_exists) {
3316 			ret = ERANGE;
3317 			goto done;
3318 		}
3319 
3320 		/*
3321 		 * Reclaim the highest runs in LIFO (descending) order until
3322 		 * the number of reclaimed pages, "reclaimed", is at least
3323 		 * "min_reclaim".  Reset "reclaimed" each time because each
3324 		 * reclamation is idempotent, and runs will (likely) recur
3325 		 * from one scan to the next as restrictions are relaxed.
3326 		 */
3327 		reclaimed = 0;
3328 		for (i = 0; count > 0 && i < nruns; i++) {
3329 			count--;
3330 			m_run = m_runs[RUN_INDEX(count, nruns)];
3331 			error = vm_page_reclaim_run(req_class, domain, npages,
3332 			    m_run, high);
3333 			if (error == 0) {
3334 				reclaimed += npages;
3335 				if (reclaimed >= min_reclaim) {
3336 					ret = 0;
3337 					goto done;
3338 				}
3339 			}
3340 		}
3341 
3342 		/*
3343 		 * Either relax the restrictions on the next scan or return if
3344 		 * the last scan had no restrictions.
3345 		 */
3346 		if (options == VPSC_NORESERV)
3347 			options = VPSC_NOSUPER;
3348 		else if (options == VPSC_NOSUPER)
3349 			options = VPSC_ANY;
3350 		else if (options == VPSC_ANY) {
3351 			if (reclaimed != 0)
3352 				ret = 0;
3353 			goto done;
3354 		}
3355 	}
3356 done:
3357 	if (m_runs != _m_runs)
3358 		free(m_runs, M_TEMP);
3359 	return (ret);
3360 }
3361 
3362 int
3363 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3364     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3365 {
3366 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3367 	    alignment, boundary, 1));
3368 }
3369 
3370 int
3371 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3372     u_long alignment, vm_paddr_t boundary)
3373 {
3374 	struct vm_domainset_iter di;
3375 	int domain, ret, status;
3376 
3377 	ret = ERANGE;
3378 
3379 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3380 	do {
3381 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3382 		    high, alignment, boundary);
3383 		if (status == 0)
3384 			return (0);
3385 		else if (status == ERANGE)
3386 			vm_domainset_iter_ignore(&di, domain);
3387 		else {
3388 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3389 			    "from vm_page_reclaim_contig_domain()", status));
3390 			ret = ENOMEM;
3391 		}
3392 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3393 
3394 	return (ret);
3395 }
3396 
3397 /*
3398  * Set the domain in the appropriate page level domainset.
3399  */
3400 void
3401 vm_domain_set(struct vm_domain *vmd)
3402 {
3403 
3404 	mtx_lock(&vm_domainset_lock);
3405 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3406 		vmd->vmd_minset = 1;
3407 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3408 	}
3409 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3410 		vmd->vmd_severeset = 1;
3411 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3412 	}
3413 	mtx_unlock(&vm_domainset_lock);
3414 }
3415 
3416 /*
3417  * Clear the domain from the appropriate page level domainset.
3418  */
3419 void
3420 vm_domain_clear(struct vm_domain *vmd)
3421 {
3422 
3423 	mtx_lock(&vm_domainset_lock);
3424 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3425 		vmd->vmd_minset = 0;
3426 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3427 		if (vm_min_waiters != 0) {
3428 			vm_min_waiters = 0;
3429 			wakeup(&vm_min_domains);
3430 		}
3431 	}
3432 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3433 		vmd->vmd_severeset = 0;
3434 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3435 		if (vm_severe_waiters != 0) {
3436 			vm_severe_waiters = 0;
3437 			wakeup(&vm_severe_domains);
3438 		}
3439 	}
3440 
3441 	/*
3442 	 * If pageout daemon needs pages, then tell it that there are
3443 	 * some free.
3444 	 */
3445 	if (vmd->vmd_pageout_pages_needed &&
3446 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3447 		wakeup(&vmd->vmd_pageout_pages_needed);
3448 		vmd->vmd_pageout_pages_needed = 0;
3449 	}
3450 
3451 	/* See comments in vm_wait_doms(). */
3452 	if (vm_pageproc_waiters) {
3453 		vm_pageproc_waiters = 0;
3454 		wakeup(&vm_pageproc_waiters);
3455 	}
3456 	mtx_unlock(&vm_domainset_lock);
3457 }
3458 
3459 /*
3460  * Wait for free pages to exceed the min threshold globally.
3461  */
3462 void
3463 vm_wait_min(void)
3464 {
3465 
3466 	mtx_lock(&vm_domainset_lock);
3467 	while (vm_page_count_min()) {
3468 		vm_min_waiters++;
3469 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3470 	}
3471 	mtx_unlock(&vm_domainset_lock);
3472 }
3473 
3474 /*
3475  * Wait for free pages to exceed the severe threshold globally.
3476  */
3477 void
3478 vm_wait_severe(void)
3479 {
3480 
3481 	mtx_lock(&vm_domainset_lock);
3482 	while (vm_page_count_severe()) {
3483 		vm_severe_waiters++;
3484 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3485 		    "vmwait", 0);
3486 	}
3487 	mtx_unlock(&vm_domainset_lock);
3488 }
3489 
3490 u_int
3491 vm_wait_count(void)
3492 {
3493 
3494 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3495 }
3496 
3497 int
3498 vm_wait_doms(const domainset_t *wdoms, int mflags)
3499 {
3500 	int error;
3501 
3502 	error = 0;
3503 
3504 	/*
3505 	 * We use racey wakeup synchronization to avoid expensive global
3506 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3507 	 * To handle this, we only sleep for one tick in this instance.  It
3508 	 * is expected that most allocations for the pageproc will come from
3509 	 * kmem or vm_page_grab* which will use the more specific and
3510 	 * race-free vm_wait_domain().
3511 	 */
3512 	if (curproc == pageproc) {
3513 		mtx_lock(&vm_domainset_lock);
3514 		vm_pageproc_waiters++;
3515 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3516 		    PVM | PDROP | mflags, "pageprocwait", 1);
3517 	} else {
3518 		/*
3519 		 * XXX Ideally we would wait only until the allocation could
3520 		 * be satisfied.  This condition can cause new allocators to
3521 		 * consume all freed pages while old allocators wait.
3522 		 */
3523 		mtx_lock(&vm_domainset_lock);
3524 		if (vm_page_count_min_set(wdoms)) {
3525 			if (pageproc == NULL)
3526 				panic("vm_wait in early boot");
3527 			vm_min_waiters++;
3528 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3529 			    PVM | PDROP | mflags, "vmwait", 0);
3530 		} else
3531 			mtx_unlock(&vm_domainset_lock);
3532 	}
3533 	return (error);
3534 }
3535 
3536 /*
3537  *	vm_wait_domain:
3538  *
3539  *	Sleep until free pages are available for allocation.
3540  *	- Called in various places after failed memory allocations.
3541  */
3542 void
3543 vm_wait_domain(int domain)
3544 {
3545 	struct vm_domain *vmd;
3546 	domainset_t wdom;
3547 
3548 	vmd = VM_DOMAIN(domain);
3549 	vm_domain_free_assert_unlocked(vmd);
3550 
3551 	if (curproc == pageproc) {
3552 		mtx_lock(&vm_domainset_lock);
3553 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3554 			vmd->vmd_pageout_pages_needed = 1;
3555 			msleep(&vmd->vmd_pageout_pages_needed,
3556 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3557 		} else
3558 			mtx_unlock(&vm_domainset_lock);
3559 	} else {
3560 		DOMAINSET_ZERO(&wdom);
3561 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3562 		vm_wait_doms(&wdom, 0);
3563 	}
3564 }
3565 
3566 static int
3567 vm_wait_flags(vm_object_t obj, int mflags)
3568 {
3569 	struct domainset *d;
3570 
3571 	d = NULL;
3572 
3573 	/*
3574 	 * Carefully fetch pointers only once: the struct domainset
3575 	 * itself is ummutable but the pointer might change.
3576 	 */
3577 	if (obj != NULL)
3578 		d = obj->domain.dr_policy;
3579 	if (d == NULL)
3580 		d = curthread->td_domain.dr_policy;
3581 
3582 	return (vm_wait_doms(&d->ds_mask, mflags));
3583 }
3584 
3585 /*
3586  *	vm_wait:
3587  *
3588  *	Sleep until free pages are available for allocation in the
3589  *	affinity domains of the obj.  If obj is NULL, the domain set
3590  *	for the calling thread is used.
3591  *	Called in various places after failed memory allocations.
3592  */
3593 void
3594 vm_wait(vm_object_t obj)
3595 {
3596 	(void)vm_wait_flags(obj, 0);
3597 }
3598 
3599 int
3600 vm_wait_intr(vm_object_t obj)
3601 {
3602 	return (vm_wait_flags(obj, PCATCH));
3603 }
3604 
3605 /*
3606  *	vm_domain_alloc_fail:
3607  *
3608  *	Called when a page allocation function fails.  Informs the
3609  *	pagedaemon and performs the requested wait.  Requires the
3610  *	domain_free and object lock on entry.  Returns with the
3611  *	object lock held and free lock released.  Returns an error when
3612  *	retry is necessary.
3613  *
3614  */
3615 static int
3616 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3617 {
3618 
3619 	vm_domain_free_assert_unlocked(vmd);
3620 
3621 	atomic_add_int(&vmd->vmd_pageout_deficit,
3622 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3623 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3624 		if (object != NULL)
3625 			VM_OBJECT_WUNLOCK(object);
3626 		vm_wait_domain(vmd->vmd_domain);
3627 		if (object != NULL)
3628 			VM_OBJECT_WLOCK(object);
3629 		if (req & VM_ALLOC_WAITOK)
3630 			return (EAGAIN);
3631 	}
3632 
3633 	return (0);
3634 }
3635 
3636 /*
3637  *	vm_waitpfault:
3638  *
3639  *	Sleep until free pages are available for allocation.
3640  *	- Called only in vm_fault so that processes page faulting
3641  *	  can be easily tracked.
3642  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3643  *	  processes will be able to grab memory first.  Do not change
3644  *	  this balance without careful testing first.
3645  */
3646 void
3647 vm_waitpfault(struct domainset *dset, int timo)
3648 {
3649 
3650 	/*
3651 	 * XXX Ideally we would wait only until the allocation could
3652 	 * be satisfied.  This condition can cause new allocators to
3653 	 * consume all freed pages while old allocators wait.
3654 	 */
3655 	mtx_lock(&vm_domainset_lock);
3656 	if (vm_page_count_min_set(&dset->ds_mask)) {
3657 		vm_min_waiters++;
3658 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3659 		    "pfault", timo);
3660 	} else
3661 		mtx_unlock(&vm_domainset_lock);
3662 }
3663 
3664 static struct vm_pagequeue *
3665 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3666 {
3667 
3668 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3669 }
3670 
3671 #ifdef INVARIANTS
3672 static struct vm_pagequeue *
3673 vm_page_pagequeue(vm_page_t m)
3674 {
3675 
3676 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3677 }
3678 #endif
3679 
3680 static __always_inline bool
3681 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3682 {
3683 	vm_page_astate_t tmp;
3684 
3685 	tmp = *old;
3686 	do {
3687 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3688 			return (true);
3689 		counter_u64_add(pqstate_commit_retries, 1);
3690 	} while (old->_bits == tmp._bits);
3691 
3692 	return (false);
3693 }
3694 
3695 /*
3696  * Do the work of committing a queue state update that moves the page out of
3697  * its current queue.
3698  */
3699 static bool
3700 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3701     vm_page_astate_t *old, vm_page_astate_t new)
3702 {
3703 	vm_page_t next;
3704 
3705 	vm_pagequeue_assert_locked(pq);
3706 	KASSERT(vm_page_pagequeue(m) == pq,
3707 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3708 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3709 	    ("%s: invalid queue indices %d %d",
3710 	    __func__, old->queue, new.queue));
3711 
3712 	/*
3713 	 * Once the queue index of the page changes there is nothing
3714 	 * synchronizing with further updates to the page's physical
3715 	 * queue state.  Therefore we must speculatively remove the page
3716 	 * from the queue now and be prepared to roll back if the queue
3717 	 * state update fails.  If the page is not physically enqueued then
3718 	 * we just update its queue index.
3719 	 */
3720 	if ((old->flags & PGA_ENQUEUED) != 0) {
3721 		new.flags &= ~PGA_ENQUEUED;
3722 		next = TAILQ_NEXT(m, plinks.q);
3723 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3724 		vm_pagequeue_cnt_dec(pq);
3725 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3726 			if (next == NULL)
3727 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3728 			else
3729 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3730 			vm_pagequeue_cnt_inc(pq);
3731 			return (false);
3732 		} else {
3733 			return (true);
3734 		}
3735 	} else {
3736 		return (vm_page_pqstate_fcmpset(m, old, new));
3737 	}
3738 }
3739 
3740 static bool
3741 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3742     vm_page_astate_t new)
3743 {
3744 	struct vm_pagequeue *pq;
3745 	vm_page_astate_t as;
3746 	bool ret;
3747 
3748 	pq = _vm_page_pagequeue(m, old->queue);
3749 
3750 	/*
3751 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3752 	 * corresponding page queue lock is held.
3753 	 */
3754 	vm_pagequeue_lock(pq);
3755 	as = vm_page_astate_load(m);
3756 	if (__predict_false(as._bits != old->_bits)) {
3757 		*old = as;
3758 		ret = false;
3759 	} else {
3760 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3761 	}
3762 	vm_pagequeue_unlock(pq);
3763 	return (ret);
3764 }
3765 
3766 /*
3767  * Commit a queue state update that enqueues or requeues a page.
3768  */
3769 static bool
3770 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3771     vm_page_astate_t *old, vm_page_astate_t new)
3772 {
3773 	struct vm_domain *vmd;
3774 
3775 	vm_pagequeue_assert_locked(pq);
3776 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3777 	    ("%s: invalid queue indices %d %d",
3778 	    __func__, old->queue, new.queue));
3779 
3780 	new.flags |= PGA_ENQUEUED;
3781 	if (!vm_page_pqstate_fcmpset(m, old, new))
3782 		return (false);
3783 
3784 	if ((old->flags & PGA_ENQUEUED) != 0)
3785 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3786 	else
3787 		vm_pagequeue_cnt_inc(pq);
3788 
3789 	/*
3790 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3791 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3792 	 * applied, even if it was set first.
3793 	 */
3794 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3795 		vmd = vm_pagequeue_domain(m);
3796 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3797 		    ("%s: invalid page queue for page %p", __func__, m));
3798 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3799 	} else {
3800 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3801 	}
3802 	return (true);
3803 }
3804 
3805 /*
3806  * Commit a queue state update that encodes a request for a deferred queue
3807  * operation.
3808  */
3809 static bool
3810 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3811     vm_page_astate_t new)
3812 {
3813 
3814 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3815 	    ("%s: invalid state, queue %d flags %x",
3816 	    __func__, new.queue, new.flags));
3817 
3818 	if (old->_bits != new._bits &&
3819 	    !vm_page_pqstate_fcmpset(m, old, new))
3820 		return (false);
3821 	vm_page_pqbatch_submit(m, new.queue);
3822 	return (true);
3823 }
3824 
3825 /*
3826  * A generic queue state update function.  This handles more cases than the
3827  * specialized functions above.
3828  */
3829 bool
3830 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3831 {
3832 
3833 	if (old->_bits == new._bits)
3834 		return (true);
3835 
3836 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3837 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3838 			return (false);
3839 		if (new.queue != PQ_NONE)
3840 			vm_page_pqbatch_submit(m, new.queue);
3841 	} else {
3842 		if (!vm_page_pqstate_fcmpset(m, old, new))
3843 			return (false);
3844 		if (new.queue != PQ_NONE &&
3845 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3846 			vm_page_pqbatch_submit(m, new.queue);
3847 	}
3848 	return (true);
3849 }
3850 
3851 /*
3852  * Apply deferred queue state updates to a page.
3853  */
3854 static inline void
3855 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3856 {
3857 	vm_page_astate_t new, old;
3858 
3859 	CRITICAL_ASSERT(curthread);
3860 	vm_pagequeue_assert_locked(pq);
3861 	KASSERT(queue < PQ_COUNT,
3862 	    ("%s: invalid queue index %d", __func__, queue));
3863 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3864 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3865 
3866 	for (old = vm_page_astate_load(m);;) {
3867 		if (__predict_false(old.queue != queue ||
3868 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3869 			counter_u64_add(queue_nops, 1);
3870 			break;
3871 		}
3872 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3873 		    ("%s: page %p is unmanaged", __func__, m));
3874 
3875 		new = old;
3876 		if ((old.flags & PGA_DEQUEUE) != 0) {
3877 			new.flags &= ~PGA_QUEUE_OP_MASK;
3878 			new.queue = PQ_NONE;
3879 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3880 			    m, &old, new))) {
3881 				counter_u64_add(queue_ops, 1);
3882 				break;
3883 			}
3884 		} else {
3885 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3886 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3887 			    m, &old, new))) {
3888 				counter_u64_add(queue_ops, 1);
3889 				break;
3890 			}
3891 		}
3892 	}
3893 }
3894 
3895 static void
3896 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3897     uint8_t queue)
3898 {
3899 	int i;
3900 
3901 	for (i = 0; i < bq->bq_cnt; i++)
3902 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3903 	vm_batchqueue_init(bq);
3904 }
3905 
3906 /*
3907  *	vm_page_pqbatch_submit:		[ internal use only ]
3908  *
3909  *	Enqueue a page in the specified page queue's batched work queue.
3910  *	The caller must have encoded the requested operation in the page
3911  *	structure's a.flags field.
3912  */
3913 void
3914 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3915 {
3916 	struct vm_batchqueue *bq;
3917 	struct vm_pagequeue *pq;
3918 	int domain, slots_remaining;
3919 
3920 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3921 
3922 	domain = vm_page_domain(m);
3923 	critical_enter();
3924 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3925 	slots_remaining = vm_batchqueue_insert(bq, m);
3926 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3927 		/* keep building the bq */
3928 		critical_exit();
3929 		return;
3930 	} else if (slots_remaining > 0 ) {
3931 		/* Try to process the bq if we can get the lock */
3932 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3933 		if (vm_pagequeue_trylock(pq)) {
3934 			vm_pqbatch_process(pq, bq, queue);
3935 			vm_pagequeue_unlock(pq);
3936 		}
3937 		critical_exit();
3938 		return;
3939 	}
3940 	critical_exit();
3941 
3942 	/* if we make it here, the bq is full so wait for the lock */
3943 
3944 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3945 	vm_pagequeue_lock(pq);
3946 	critical_enter();
3947 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3948 	vm_pqbatch_process(pq, bq, queue);
3949 	vm_pqbatch_process_page(pq, m, queue);
3950 	vm_pagequeue_unlock(pq);
3951 	critical_exit();
3952 }
3953 
3954 /*
3955  *	vm_page_pqbatch_drain:		[ internal use only ]
3956  *
3957  *	Force all per-CPU page queue batch queues to be drained.  This is
3958  *	intended for use in severe memory shortages, to ensure that pages
3959  *	do not remain stuck in the batch queues.
3960  */
3961 void
3962 vm_page_pqbatch_drain(void)
3963 {
3964 	struct thread *td;
3965 	struct vm_domain *vmd;
3966 	struct vm_pagequeue *pq;
3967 	int cpu, domain, queue;
3968 
3969 	td = curthread;
3970 	CPU_FOREACH(cpu) {
3971 		thread_lock(td);
3972 		sched_bind(td, cpu);
3973 		thread_unlock(td);
3974 
3975 		for (domain = 0; domain < vm_ndomains; domain++) {
3976 			vmd = VM_DOMAIN(domain);
3977 			for (queue = 0; queue < PQ_COUNT; queue++) {
3978 				pq = &vmd->vmd_pagequeues[queue];
3979 				vm_pagequeue_lock(pq);
3980 				critical_enter();
3981 				vm_pqbatch_process(pq,
3982 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3983 				critical_exit();
3984 				vm_pagequeue_unlock(pq);
3985 			}
3986 		}
3987 	}
3988 	thread_lock(td);
3989 	sched_unbind(td);
3990 	thread_unlock(td);
3991 }
3992 
3993 /*
3994  *	vm_page_dequeue_deferred:	[ internal use only ]
3995  *
3996  *	Request removal of the given page from its current page
3997  *	queue.  Physical removal from the queue may be deferred
3998  *	indefinitely.
3999  */
4000 void
4001 vm_page_dequeue_deferred(vm_page_t m)
4002 {
4003 	vm_page_astate_t new, old;
4004 
4005 	old = vm_page_astate_load(m);
4006 	do {
4007 		if (old.queue == PQ_NONE) {
4008 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4009 			    ("%s: page %p has unexpected queue state",
4010 			    __func__, m));
4011 			break;
4012 		}
4013 		new = old;
4014 		new.flags |= PGA_DEQUEUE;
4015 	} while (!vm_page_pqstate_commit_request(m, &old, new));
4016 }
4017 
4018 /*
4019  *	vm_page_dequeue:
4020  *
4021  *	Remove the page from whichever page queue it's in, if any, before
4022  *	returning.
4023  */
4024 void
4025 vm_page_dequeue(vm_page_t m)
4026 {
4027 	vm_page_astate_t new, old;
4028 
4029 	old = vm_page_astate_load(m);
4030 	do {
4031 		if (old.queue == PQ_NONE) {
4032 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4033 			    ("%s: page %p has unexpected queue state",
4034 			    __func__, m));
4035 			break;
4036 		}
4037 		new = old;
4038 		new.flags &= ~PGA_QUEUE_OP_MASK;
4039 		new.queue = PQ_NONE;
4040 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
4041 
4042 }
4043 
4044 /*
4045  * Schedule the given page for insertion into the specified page queue.
4046  * Physical insertion of the page may be deferred indefinitely.
4047  */
4048 static void
4049 vm_page_enqueue(vm_page_t m, uint8_t queue)
4050 {
4051 
4052 	KASSERT(m->a.queue == PQ_NONE &&
4053 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
4054 	    ("%s: page %p is already enqueued", __func__, m));
4055 	KASSERT(m->ref_count > 0,
4056 	    ("%s: page %p does not carry any references", __func__, m));
4057 
4058 	m->a.queue = queue;
4059 	if ((m->a.flags & PGA_REQUEUE) == 0)
4060 		vm_page_aflag_set(m, PGA_REQUEUE);
4061 	vm_page_pqbatch_submit(m, queue);
4062 }
4063 
4064 /*
4065  *	vm_page_free_prep:
4066  *
4067  *	Prepares the given page to be put on the free list,
4068  *	disassociating it from any VM object. The caller may return
4069  *	the page to the free list only if this function returns true.
4070  *
4071  *	The object, if it exists, must be locked, and then the page must
4072  *	be xbusy.  Otherwise the page must be not busied.  A managed
4073  *	page must be unmapped.
4074  */
4075 static bool
4076 vm_page_free_prep(vm_page_t m)
4077 {
4078 
4079 	/*
4080 	 * Synchronize with threads that have dropped a reference to this
4081 	 * page.
4082 	 */
4083 	atomic_thread_fence_acq();
4084 
4085 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
4086 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
4087 		uint64_t *p;
4088 		int i;
4089 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4090 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4091 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4092 			    m, i, (uintmax_t)*p));
4093 	}
4094 #endif
4095 	KASSERT((m->flags & PG_NOFREE) == 0,
4096 	    ("%s: attempting to free a PG_NOFREE page", __func__));
4097 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4098 		KASSERT(!pmap_page_is_mapped(m),
4099 		    ("vm_page_free_prep: freeing mapped page %p", m));
4100 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4101 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4102 	} else {
4103 		KASSERT(m->a.queue == PQ_NONE,
4104 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4105 	}
4106 	VM_CNT_INC(v_tfree);
4107 
4108 	if (m->object != NULL) {
4109 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
4110 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
4111 		    ("vm_page_free_prep: managed flag mismatch for page %p",
4112 		    m));
4113 		vm_page_assert_xbusied(m);
4114 
4115 		/*
4116 		 * The object reference can be released without an atomic
4117 		 * operation.
4118 		 */
4119 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4120 		    m->ref_count == VPRC_OBJREF,
4121 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
4122 		    m, m->ref_count));
4123 		vm_page_object_remove(m);
4124 		m->ref_count -= VPRC_OBJREF;
4125 	} else
4126 		vm_page_assert_unbusied(m);
4127 
4128 	vm_page_busy_free(m);
4129 
4130 	/*
4131 	 * If fictitious remove object association and
4132 	 * return.
4133 	 */
4134 	if ((m->flags & PG_FICTITIOUS) != 0) {
4135 		KASSERT(m->ref_count == 1,
4136 		    ("fictitious page %p is referenced", m));
4137 		KASSERT(m->a.queue == PQ_NONE,
4138 		    ("fictitious page %p is queued", m));
4139 		return (false);
4140 	}
4141 
4142 	/*
4143 	 * Pages need not be dequeued before they are returned to the physical
4144 	 * memory allocator, but they must at least be marked for a deferred
4145 	 * dequeue.
4146 	 */
4147 	if ((m->oflags & VPO_UNMANAGED) == 0)
4148 		vm_page_dequeue_deferred(m);
4149 
4150 	m->valid = 0;
4151 	vm_page_undirty(m);
4152 
4153 	if (m->ref_count != 0)
4154 		panic("vm_page_free_prep: page %p has references", m);
4155 
4156 	/*
4157 	 * Restore the default memory attribute to the page.
4158 	 */
4159 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4160 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4161 
4162 #if VM_NRESERVLEVEL > 0
4163 	/*
4164 	 * Determine whether the page belongs to a reservation.  If the page was
4165 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4166 	 * as an optimization, we avoid the check in that case.
4167 	 */
4168 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4169 		return (false);
4170 #endif
4171 
4172 	return (true);
4173 }
4174 
4175 /*
4176  *	vm_page_free_toq:
4177  *
4178  *	Returns the given page to the free list, disassociating it
4179  *	from any VM object.
4180  *
4181  *	The object must be locked.  The page must be exclusively busied if it
4182  *	belongs to an object.
4183  */
4184 static void
4185 vm_page_free_toq(vm_page_t m)
4186 {
4187 	struct vm_domain *vmd;
4188 	uma_zone_t zone;
4189 
4190 	if (!vm_page_free_prep(m))
4191 		return;
4192 
4193 	vmd = vm_pagequeue_domain(m);
4194 	zone = vmd->vmd_pgcache[m->pool].zone;
4195 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4196 		uma_zfree(zone, m);
4197 		return;
4198 	}
4199 	vm_domain_free_lock(vmd);
4200 	vm_phys_free_pages(m, 0);
4201 	vm_domain_free_unlock(vmd);
4202 	vm_domain_freecnt_inc(vmd, 1);
4203 }
4204 
4205 /*
4206  *	vm_page_free_pages_toq:
4207  *
4208  *	Returns a list of pages to the free list, disassociating it
4209  *	from any VM object.  In other words, this is equivalent to
4210  *	calling vm_page_free_toq() for each page of a list of VM objects.
4211  */
4212 int
4213 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4214 {
4215 	vm_page_t m;
4216 	int count;
4217 
4218 	if (SLIST_EMPTY(free))
4219 		return (0);
4220 
4221 	count = 0;
4222 	while ((m = SLIST_FIRST(free)) != NULL) {
4223 		count++;
4224 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4225 		vm_page_free_toq(m);
4226 	}
4227 
4228 	if (update_wire_count)
4229 		vm_wire_sub(count);
4230 	return (count);
4231 }
4232 
4233 /*
4234  * Mark this page as wired down.  For managed pages, this prevents reclamation
4235  * by the page daemon, or when the containing object, if any, is destroyed.
4236  */
4237 void
4238 vm_page_wire(vm_page_t m)
4239 {
4240 	u_int old;
4241 
4242 #ifdef INVARIANTS
4243 	if (m->object != NULL && !vm_page_busied(m) &&
4244 	    !vm_object_busied(m->object))
4245 		VM_OBJECT_ASSERT_LOCKED(m->object);
4246 #endif
4247 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4248 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4249 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4250 
4251 	old = atomic_fetchadd_int(&m->ref_count, 1);
4252 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4253 	    ("vm_page_wire: counter overflow for page %p", m));
4254 	if (VPRC_WIRE_COUNT(old) == 0) {
4255 		if ((m->oflags & VPO_UNMANAGED) == 0)
4256 			vm_page_aflag_set(m, PGA_DEQUEUE);
4257 		vm_wire_add(1);
4258 	}
4259 }
4260 
4261 /*
4262  * Attempt to wire a mapped page following a pmap lookup of that page.
4263  * This may fail if a thread is concurrently tearing down mappings of the page.
4264  * The transient failure is acceptable because it translates to the
4265  * failure of the caller pmap_extract_and_hold(), which should be then
4266  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4267  */
4268 bool
4269 vm_page_wire_mapped(vm_page_t m)
4270 {
4271 	u_int old;
4272 
4273 	old = atomic_load_int(&m->ref_count);
4274 	do {
4275 		KASSERT(old > 0,
4276 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4277 		if ((old & VPRC_BLOCKED) != 0)
4278 			return (false);
4279 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4280 
4281 	if (VPRC_WIRE_COUNT(old) == 0) {
4282 		if ((m->oflags & VPO_UNMANAGED) == 0)
4283 			vm_page_aflag_set(m, PGA_DEQUEUE);
4284 		vm_wire_add(1);
4285 	}
4286 	return (true);
4287 }
4288 
4289 /*
4290  * Release a wiring reference to a managed page.  If the page still belongs to
4291  * an object, update its position in the page queues to reflect the reference.
4292  * If the wiring was the last reference to the page, free the page.
4293  */
4294 static void
4295 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4296 {
4297 	u_int old;
4298 
4299 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4300 	    ("%s: page %p is unmanaged", __func__, m));
4301 
4302 	/*
4303 	 * Update LRU state before releasing the wiring reference.
4304 	 * Use a release store when updating the reference count to
4305 	 * synchronize with vm_page_free_prep().
4306 	 */
4307 	old = atomic_load_int(&m->ref_count);
4308 	do {
4309 		u_int count;
4310 
4311 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4312 		    ("vm_page_unwire: wire count underflow for page %p", m));
4313 
4314 		count = old & ~VPRC_BLOCKED;
4315 		if (count > VPRC_OBJREF + 1) {
4316 			/*
4317 			 * The page has at least one other wiring reference.  An
4318 			 * earlier iteration of this loop may have called
4319 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4320 			 * re-set it if necessary.
4321 			 */
4322 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4323 				vm_page_aflag_set(m, PGA_DEQUEUE);
4324 		} else if (count == VPRC_OBJREF + 1) {
4325 			/*
4326 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4327 			 * update the page's queue state to reflect the
4328 			 * reference.  If the page does not belong to an object
4329 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4330 			 * clear leftover queue state.
4331 			 */
4332 			vm_page_release_toq(m, nqueue, noreuse);
4333 		} else if (count == 1) {
4334 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4335 		}
4336 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4337 
4338 	if (VPRC_WIRE_COUNT(old) == 1) {
4339 		vm_wire_sub(1);
4340 		if (old == 1)
4341 			vm_page_free(m);
4342 	}
4343 }
4344 
4345 /*
4346  * Release one wiring of the specified page, potentially allowing it to be
4347  * paged out.
4348  *
4349  * Only managed pages belonging to an object can be paged out.  If the number
4350  * of wirings transitions to zero and the page is eligible for page out, then
4351  * the page is added to the specified paging queue.  If the released wiring
4352  * represented the last reference to the page, the page is freed.
4353  */
4354 void
4355 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4356 {
4357 
4358 	KASSERT(nqueue < PQ_COUNT,
4359 	    ("vm_page_unwire: invalid queue %u request for page %p",
4360 	    nqueue, m));
4361 
4362 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4363 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4364 			vm_page_free(m);
4365 		return;
4366 	}
4367 	vm_page_unwire_managed(m, nqueue, false);
4368 }
4369 
4370 /*
4371  * Unwire a page without (re-)inserting it into a page queue.  It is up
4372  * to the caller to enqueue, requeue, or free the page as appropriate.
4373  * In most cases involving managed pages, vm_page_unwire() should be used
4374  * instead.
4375  */
4376 bool
4377 vm_page_unwire_noq(vm_page_t m)
4378 {
4379 	u_int old;
4380 
4381 	old = vm_page_drop(m, 1);
4382 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4383 	    ("%s: counter underflow for page %p", __func__,  m));
4384 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4385 	    ("%s: missing ref on fictitious page %p", __func__, m));
4386 
4387 	if (VPRC_WIRE_COUNT(old) > 1)
4388 		return (false);
4389 	if ((m->oflags & VPO_UNMANAGED) == 0)
4390 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4391 	vm_wire_sub(1);
4392 	return (true);
4393 }
4394 
4395 /*
4396  * Ensure that the page ends up in the specified page queue.  If the page is
4397  * active or being moved to the active queue, ensure that its act_count is
4398  * at least ACT_INIT but do not otherwise mess with it.
4399  */
4400 static __always_inline void
4401 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4402 {
4403 	vm_page_astate_t old, new;
4404 
4405 	KASSERT(m->ref_count > 0,
4406 	    ("%s: page %p does not carry any references", __func__, m));
4407 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4408 	    ("%s: invalid flags %x", __func__, nflag));
4409 
4410 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4411 		return;
4412 
4413 	old = vm_page_astate_load(m);
4414 	do {
4415 		if ((old.flags & PGA_DEQUEUE) != 0)
4416 			break;
4417 		new = old;
4418 		new.flags &= ~PGA_QUEUE_OP_MASK;
4419 		if (nqueue == PQ_ACTIVE)
4420 			new.act_count = max(old.act_count, ACT_INIT);
4421 		if (old.queue == nqueue) {
4422 			/*
4423 			 * There is no need to requeue pages already in the
4424 			 * active queue.
4425 			 */
4426 			if (nqueue != PQ_ACTIVE ||
4427 			    (old.flags & PGA_ENQUEUED) == 0)
4428 				new.flags |= nflag;
4429 		} else {
4430 			new.flags |= nflag;
4431 			new.queue = nqueue;
4432 		}
4433 	} while (!vm_page_pqstate_commit(m, &old, new));
4434 }
4435 
4436 /*
4437  * Put the specified page on the active list (if appropriate).
4438  */
4439 void
4440 vm_page_activate(vm_page_t m)
4441 {
4442 
4443 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4444 }
4445 
4446 /*
4447  * Move the specified page to the tail of the inactive queue, or requeue
4448  * the page if it is already in the inactive queue.
4449  */
4450 void
4451 vm_page_deactivate(vm_page_t m)
4452 {
4453 
4454 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4455 }
4456 
4457 void
4458 vm_page_deactivate_noreuse(vm_page_t m)
4459 {
4460 
4461 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4462 }
4463 
4464 /*
4465  * Put a page in the laundry, or requeue it if it is already there.
4466  */
4467 void
4468 vm_page_launder(vm_page_t m)
4469 {
4470 
4471 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4472 }
4473 
4474 /*
4475  * Put a page in the PQ_UNSWAPPABLE holding queue.
4476  */
4477 void
4478 vm_page_unswappable(vm_page_t m)
4479 {
4480 
4481 	VM_OBJECT_ASSERT_LOCKED(m->object);
4482 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4483 	    ("page %p already unswappable", m));
4484 
4485 	vm_page_dequeue(m);
4486 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4487 }
4488 
4489 /*
4490  * Release a page back to the page queues in preparation for unwiring.
4491  */
4492 static void
4493 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4494 {
4495 	vm_page_astate_t old, new;
4496 	uint16_t nflag;
4497 
4498 	/*
4499 	 * Use a check of the valid bits to determine whether we should
4500 	 * accelerate reclamation of the page.  The object lock might not be
4501 	 * held here, in which case the check is racy.  At worst we will either
4502 	 * accelerate reclamation of a valid page and violate LRU, or
4503 	 * unnecessarily defer reclamation of an invalid page.
4504 	 *
4505 	 * If we were asked to not cache the page, place it near the head of the
4506 	 * inactive queue so that is reclaimed sooner.
4507 	 */
4508 	if (noreuse || vm_page_none_valid(m)) {
4509 		nqueue = PQ_INACTIVE;
4510 		nflag = PGA_REQUEUE_HEAD;
4511 	} else {
4512 		nflag = PGA_REQUEUE;
4513 	}
4514 
4515 	old = vm_page_astate_load(m);
4516 	do {
4517 		new = old;
4518 
4519 		/*
4520 		 * If the page is already in the active queue and we are not
4521 		 * trying to accelerate reclamation, simply mark it as
4522 		 * referenced and avoid any queue operations.
4523 		 */
4524 		new.flags &= ~PGA_QUEUE_OP_MASK;
4525 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4526 		    (old.flags & PGA_ENQUEUED) != 0)
4527 			new.flags |= PGA_REFERENCED;
4528 		else {
4529 			new.flags |= nflag;
4530 			new.queue = nqueue;
4531 		}
4532 	} while (!vm_page_pqstate_commit(m, &old, new));
4533 }
4534 
4535 /*
4536  * Unwire a page and either attempt to free it or re-add it to the page queues.
4537  */
4538 void
4539 vm_page_release(vm_page_t m, int flags)
4540 {
4541 	vm_object_t object;
4542 
4543 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4544 	    ("vm_page_release: page %p is unmanaged", m));
4545 
4546 	if ((flags & VPR_TRYFREE) != 0) {
4547 		for (;;) {
4548 			object = atomic_load_ptr(&m->object);
4549 			if (object == NULL)
4550 				break;
4551 			/* Depends on type-stability. */
4552 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4553 				break;
4554 			if (object == m->object) {
4555 				vm_page_release_locked(m, flags);
4556 				VM_OBJECT_WUNLOCK(object);
4557 				return;
4558 			}
4559 			VM_OBJECT_WUNLOCK(object);
4560 		}
4561 	}
4562 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4563 }
4564 
4565 /* See vm_page_release(). */
4566 void
4567 vm_page_release_locked(vm_page_t m, int flags)
4568 {
4569 
4570 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4571 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4572 	    ("vm_page_release_locked: page %p is unmanaged", m));
4573 
4574 	if (vm_page_unwire_noq(m)) {
4575 		if ((flags & VPR_TRYFREE) != 0 &&
4576 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4577 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4578 			/*
4579 			 * An unlocked lookup may have wired the page before the
4580 			 * busy lock was acquired, in which case the page must
4581 			 * not be freed.
4582 			 */
4583 			if (__predict_true(!vm_page_wired(m))) {
4584 				vm_page_free(m);
4585 				return;
4586 			}
4587 			vm_page_xunbusy(m);
4588 		} else {
4589 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4590 		}
4591 	}
4592 }
4593 
4594 static bool
4595 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4596 {
4597 	u_int old;
4598 
4599 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4600 	    ("vm_page_try_blocked_op: page %p has no object", m));
4601 	KASSERT(vm_page_busied(m),
4602 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4603 	VM_OBJECT_ASSERT_LOCKED(m->object);
4604 
4605 	old = atomic_load_int(&m->ref_count);
4606 	do {
4607 		KASSERT(old != 0,
4608 		    ("vm_page_try_blocked_op: page %p has no references", m));
4609 		KASSERT((old & VPRC_BLOCKED) == 0,
4610 		    ("vm_page_try_blocked_op: page %p blocks wirings", m));
4611 		if (VPRC_WIRE_COUNT(old) != 0)
4612 			return (false);
4613 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4614 
4615 	(op)(m);
4616 
4617 	/*
4618 	 * If the object is read-locked, new wirings may be created via an
4619 	 * object lookup.
4620 	 */
4621 	old = vm_page_drop(m, VPRC_BLOCKED);
4622 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4623 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4624 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4625 	    old, m));
4626 	return (true);
4627 }
4628 
4629 /*
4630  * Atomically check for wirings and remove all mappings of the page.
4631  */
4632 bool
4633 vm_page_try_remove_all(vm_page_t m)
4634 {
4635 
4636 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4637 }
4638 
4639 /*
4640  * Atomically check for wirings and remove all writeable mappings of the page.
4641  */
4642 bool
4643 vm_page_try_remove_write(vm_page_t m)
4644 {
4645 
4646 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4647 }
4648 
4649 /*
4650  * vm_page_advise
4651  *
4652  * 	Apply the specified advice to the given page.
4653  */
4654 void
4655 vm_page_advise(vm_page_t m, int advice)
4656 {
4657 
4658 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4659 	vm_page_assert_xbusied(m);
4660 
4661 	if (advice == MADV_FREE)
4662 		/*
4663 		 * Mark the page clean.  This will allow the page to be freed
4664 		 * without first paging it out.  MADV_FREE pages are often
4665 		 * quickly reused by malloc(3), so we do not do anything that
4666 		 * would result in a page fault on a later access.
4667 		 */
4668 		vm_page_undirty(m);
4669 	else if (advice != MADV_DONTNEED) {
4670 		if (advice == MADV_WILLNEED)
4671 			vm_page_activate(m);
4672 		return;
4673 	}
4674 
4675 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4676 		vm_page_dirty(m);
4677 
4678 	/*
4679 	 * Clear any references to the page.  Otherwise, the page daemon will
4680 	 * immediately reactivate the page.
4681 	 */
4682 	vm_page_aflag_clear(m, PGA_REFERENCED);
4683 
4684 	/*
4685 	 * Place clean pages near the head of the inactive queue rather than
4686 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4687 	 * the page will be reused quickly.  Dirty pages not already in the
4688 	 * laundry are moved there.
4689 	 */
4690 	if (m->dirty == 0)
4691 		vm_page_deactivate_noreuse(m);
4692 	else if (!vm_page_in_laundry(m))
4693 		vm_page_launder(m);
4694 }
4695 
4696 /*
4697  *	vm_page_grab_release
4698  *
4699  *	Helper routine for grab functions to release busy on return.
4700  */
4701 static inline void
4702 vm_page_grab_release(vm_page_t m, int allocflags)
4703 {
4704 
4705 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4706 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4707 			vm_page_sunbusy(m);
4708 		else
4709 			vm_page_xunbusy(m);
4710 	}
4711 }
4712 
4713 /*
4714  *	vm_page_grab_sleep
4715  *
4716  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4717  *	if the caller should retry and false otherwise.
4718  *
4719  *	If the object is locked on entry the object will be unlocked with
4720  *	false returns and still locked but possibly having been dropped
4721  *	with true returns.
4722  */
4723 static bool
4724 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4725     const char *wmesg, int allocflags, bool locked)
4726 {
4727 
4728 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4729 		return (false);
4730 
4731 	/*
4732 	 * Reference the page before unlocking and sleeping so that
4733 	 * the page daemon is less likely to reclaim it.
4734 	 */
4735 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4736 		vm_page_reference(m);
4737 
4738 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4739 	    locked)
4740 		VM_OBJECT_WLOCK(object);
4741 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4742 		return (false);
4743 
4744 	return (true);
4745 }
4746 
4747 /*
4748  * Assert that the grab flags are valid.
4749  */
4750 static inline void
4751 vm_page_grab_check(int allocflags)
4752 {
4753 
4754 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4755 	    (allocflags & VM_ALLOC_WIRED) != 0,
4756 	    ("vm_page_grab*: the pages must be busied or wired"));
4757 
4758 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4759 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4760 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4761 }
4762 
4763 /*
4764  * Calculate the page allocation flags for grab.
4765  */
4766 static inline int
4767 vm_page_grab_pflags(int allocflags)
4768 {
4769 	int pflags;
4770 
4771 	pflags = allocflags &
4772 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4773 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4774 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4775 		pflags |= VM_ALLOC_WAITFAIL;
4776 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4777 		pflags |= VM_ALLOC_SBUSY;
4778 
4779 	return (pflags);
4780 }
4781 
4782 /*
4783  * Grab a page, waiting until we are waken up due to the page
4784  * changing state.  We keep on waiting, if the page continues
4785  * to be in the object.  If the page doesn't exist, first allocate it
4786  * and then conditionally zero it.
4787  *
4788  * This routine may sleep.
4789  *
4790  * The object must be locked on entry.  The lock will, however, be released
4791  * and reacquired if the routine sleeps.
4792  */
4793 vm_page_t
4794 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4795 {
4796 	vm_page_t m;
4797 
4798 	VM_OBJECT_ASSERT_WLOCKED(object);
4799 	vm_page_grab_check(allocflags);
4800 
4801 retrylookup:
4802 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4803 		if (!vm_page_tryacquire(m, allocflags)) {
4804 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4805 			    allocflags, true))
4806 				goto retrylookup;
4807 			return (NULL);
4808 		}
4809 		goto out;
4810 	}
4811 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4812 		return (NULL);
4813 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4814 	if (m == NULL) {
4815 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4816 			return (NULL);
4817 		goto retrylookup;
4818 	}
4819 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4820 		pmap_zero_page(m);
4821 
4822 out:
4823 	vm_page_grab_release(m, allocflags);
4824 
4825 	return (m);
4826 }
4827 
4828 /*
4829  * Attempt to validate a page, locklessly acquiring it if necessary, given a
4830  * (object, pindex) tuple and either an invalided page or NULL.  The resulting
4831  * page will be validated against the identity tuple, and busied or wired as
4832  * requested.  A NULL page returned guarantees that the page was not in radix at
4833  * the time of the call but callers must perform higher level synchronization or
4834  * retry the operation under a lock if they require an atomic answer.  This is
4835  * the only lock free validation routine, other routines can depend on the
4836  * resulting page state.
4837  *
4838  * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to
4839  * caller flags.
4840  */
4841 #define PAGE_NOT_ACQUIRED ((vm_page_t)1)
4842 static vm_page_t
4843 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m,
4844     int allocflags)
4845 {
4846 	if (m == NULL)
4847 		m = vm_page_lookup_unlocked(object, pindex);
4848 	for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) {
4849 		if (vm_page_trybusy(m, allocflags)) {
4850 			if (m->object == object && m->pindex == pindex) {
4851 				if ((allocflags & VM_ALLOC_WIRED) != 0)
4852 					vm_page_wire(m);
4853 				vm_page_grab_release(m, allocflags);
4854 				break;
4855 			}
4856 			/* relookup. */
4857 			vm_page_busy_release(m);
4858 			cpu_spinwait();
4859 			continue;
4860 		}
4861 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4862 		    allocflags, false))
4863 			return (PAGE_NOT_ACQUIRED);
4864 	}
4865 	return (m);
4866 }
4867 
4868 /*
4869  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4870  * is not set.
4871  */
4872 vm_page_t
4873 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4874 {
4875 	vm_page_t m;
4876 
4877 	vm_page_grab_check(allocflags);
4878 	m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags);
4879 	if (m == PAGE_NOT_ACQUIRED)
4880 		return (NULL);
4881 	if (m != NULL)
4882 		return (m);
4883 
4884 	/*
4885 	 * The radix lockless lookup should never return a false negative
4886 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4887 	 * was no page present at the instant of the call.  A NOCREAT caller
4888 	 * must handle create races gracefully.
4889 	 */
4890 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4891 		return (NULL);
4892 
4893 	VM_OBJECT_WLOCK(object);
4894 	m = vm_page_grab(object, pindex, allocflags);
4895 	VM_OBJECT_WUNLOCK(object);
4896 
4897 	return (m);
4898 }
4899 
4900 /*
4901  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4902  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4903  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4904  * in simultaneously.  Additional pages will be left on a paging queue but
4905  * will neither be wired nor busy regardless of allocflags.
4906  */
4907 int
4908 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4909 {
4910 	vm_page_t m;
4911 	vm_page_t ma[VM_INITIAL_PAGEIN];
4912 	int after, i, pflags, rv;
4913 
4914 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4915 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4916 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4917 	KASSERT((allocflags &
4918 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4919 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4920 	VM_OBJECT_ASSERT_WLOCKED(object);
4921 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4922 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4923 	pflags |= VM_ALLOC_WAITFAIL;
4924 
4925 retrylookup:
4926 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4927 		/*
4928 		 * If the page is fully valid it can only become invalid
4929 		 * with the object lock held.  If it is not valid it can
4930 		 * become valid with the busy lock held.  Therefore, we
4931 		 * may unnecessarily lock the exclusive busy here if we
4932 		 * race with I/O completion not using the object lock.
4933 		 * However, we will not end up with an invalid page and a
4934 		 * shared lock.
4935 		 */
4936 		if (!vm_page_trybusy(m,
4937 		    vm_page_all_valid(m) ? allocflags : 0)) {
4938 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4939 			    allocflags, true);
4940 			goto retrylookup;
4941 		}
4942 		if (vm_page_all_valid(m))
4943 			goto out;
4944 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4945 			vm_page_busy_release(m);
4946 			*mp = NULL;
4947 			return (VM_PAGER_FAIL);
4948 		}
4949 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4950 		*mp = NULL;
4951 		return (VM_PAGER_FAIL);
4952 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4953 		if (!vm_pager_can_alloc_page(object, pindex)) {
4954 			*mp = NULL;
4955 			return (VM_PAGER_AGAIN);
4956 		}
4957 		goto retrylookup;
4958 	}
4959 
4960 	vm_page_assert_xbusied(m);
4961 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4962 		after = MIN(after, VM_INITIAL_PAGEIN);
4963 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4964 		after = MAX(after, 1);
4965 		ma[0] = m;
4966 		for (i = 1; i < after; i++) {
4967 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4968 				if (vm_page_any_valid(ma[i]) ||
4969 				    !vm_page_tryxbusy(ma[i]))
4970 					break;
4971 			} else {
4972 				ma[i] = vm_page_alloc(object, m->pindex + i,
4973 				    VM_ALLOC_NORMAL);
4974 				if (ma[i] == NULL)
4975 					break;
4976 			}
4977 		}
4978 		after = i;
4979 		vm_object_pip_add(object, after);
4980 		VM_OBJECT_WUNLOCK(object);
4981 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4982 		VM_OBJECT_WLOCK(object);
4983 		vm_object_pip_wakeupn(object, after);
4984 		/* Pager may have replaced a page. */
4985 		m = ma[0];
4986 		if (rv != VM_PAGER_OK) {
4987 			for (i = 0; i < after; i++) {
4988 				if (!vm_page_wired(ma[i]))
4989 					vm_page_free(ma[i]);
4990 				else
4991 					vm_page_xunbusy(ma[i]);
4992 			}
4993 			*mp = NULL;
4994 			return (rv);
4995 		}
4996 		for (i = 1; i < after; i++)
4997 			vm_page_readahead_finish(ma[i]);
4998 		MPASS(vm_page_all_valid(m));
4999 	} else {
5000 		vm_page_zero_invalid(m, TRUE);
5001 	}
5002 out:
5003 	if ((allocflags & VM_ALLOC_WIRED) != 0)
5004 		vm_page_wire(m);
5005 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
5006 		vm_page_busy_downgrade(m);
5007 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
5008 		vm_page_busy_release(m);
5009 	*mp = m;
5010 	return (VM_PAGER_OK);
5011 }
5012 
5013 /*
5014  * Locklessly grab a valid page.  If the page is not valid or not yet
5015  * allocated this will fall back to the object lock method.
5016  */
5017 int
5018 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
5019     vm_pindex_t pindex, int allocflags)
5020 {
5021 	vm_page_t m;
5022 	int flags;
5023 	int error;
5024 
5025 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
5026 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
5027 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
5028 	    "mismatch"));
5029 	KASSERT((allocflags &
5030 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
5031 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
5032 
5033 	/*
5034 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
5035 	 * before we can inspect the valid field and return a wired page.
5036 	 */
5037 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
5038 	vm_page_grab_check(flags);
5039 	m = vm_page_acquire_unlocked(object, pindex, NULL, flags);
5040 	if (m == PAGE_NOT_ACQUIRED)
5041 		return (VM_PAGER_FAIL);
5042 	if (m != NULL) {
5043 		if (vm_page_all_valid(m)) {
5044 			if ((allocflags & VM_ALLOC_WIRED) != 0)
5045 				vm_page_wire(m);
5046 			vm_page_grab_release(m, allocflags);
5047 			*mp = m;
5048 			return (VM_PAGER_OK);
5049 		}
5050 		vm_page_busy_release(m);
5051 	}
5052 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5053 		*mp = NULL;
5054 		return (VM_PAGER_FAIL);
5055 	}
5056 	VM_OBJECT_WLOCK(object);
5057 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
5058 	VM_OBJECT_WUNLOCK(object);
5059 
5060 	return (error);
5061 }
5062 
5063 /*
5064  * Return the specified range of pages from the given object.  For each
5065  * page offset within the range, if a page already exists within the object
5066  * at that offset and it is busy, then wait for it to change state.  If,
5067  * instead, the page doesn't exist, then allocate it.
5068  *
5069  * The caller must always specify an allocation class.
5070  *
5071  * allocation classes:
5072  *	VM_ALLOC_NORMAL		normal process request
5073  *	VM_ALLOC_SYSTEM		system *really* needs the pages
5074  *
5075  * The caller must always specify that the pages are to be busied and/or
5076  * wired.
5077  *
5078  * optional allocation flags:
5079  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5080  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5081  *	VM_ALLOC_NOWAIT		do not sleep
5082  *	VM_ALLOC_SBUSY		set page to sbusy state
5083  *	VM_ALLOC_WIRED		wire the pages
5084  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5085  *
5086  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5087  * may return a partial prefix of the requested range.
5088  */
5089 int
5090 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5091     vm_page_t *ma, int count)
5092 {
5093 	vm_page_t m, mpred;
5094 	int pflags;
5095 	int i;
5096 
5097 	VM_OBJECT_ASSERT_WLOCKED(object);
5098 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5099 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5100 	KASSERT(count > 0,
5101 	    ("vm_page_grab_pages: invalid page count %d", count));
5102 	vm_page_grab_check(allocflags);
5103 
5104 	pflags = vm_page_grab_pflags(allocflags);
5105 	i = 0;
5106 retrylookup:
5107 	m = vm_page_mpred(object, pindex + i);
5108 	if (m == NULL || m->pindex != pindex + i) {
5109 		mpred = m;
5110 		m = NULL;
5111 	} else
5112 		mpred = TAILQ_PREV(m, pglist, listq);
5113 	for (; i < count; i++) {
5114 		if (m != NULL) {
5115 			if (!vm_page_tryacquire(m, allocflags)) {
5116 				if (vm_page_grab_sleep(object, m, pindex + i,
5117 				    "grbmaw", allocflags, true))
5118 					goto retrylookup;
5119 				break;
5120 			}
5121 		} else {
5122 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5123 				break;
5124 			m = vm_page_alloc_after(object, pindex + i,
5125 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5126 			if (m == NULL) {
5127 				if ((allocflags & (VM_ALLOC_NOWAIT |
5128 				    VM_ALLOC_WAITFAIL)) != 0)
5129 					break;
5130 				goto retrylookup;
5131 			}
5132 		}
5133 		if (vm_page_none_valid(m) &&
5134 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5135 			if ((m->flags & PG_ZERO) == 0)
5136 				pmap_zero_page(m);
5137 			vm_page_valid(m);
5138 		}
5139 		vm_page_grab_release(m, allocflags);
5140 		ma[i] = mpred = m;
5141 		m = vm_page_next(m);
5142 	}
5143 	return (i);
5144 }
5145 
5146 /*
5147  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5148  * and will fall back to the locked variant to handle allocation.
5149  */
5150 int
5151 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5152     int allocflags, vm_page_t *ma, int count)
5153 {
5154 	vm_page_t m;
5155 	int flags;
5156 	int i;
5157 
5158 	KASSERT(count > 0,
5159 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5160 	vm_page_grab_check(allocflags);
5161 
5162 	/*
5163 	 * Modify flags for lockless acquire to hold the page until we
5164 	 * set it valid if necessary.
5165 	 */
5166 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5167 	vm_page_grab_check(flags);
5168 	m = NULL;
5169 	for (i = 0; i < count; i++, pindex++) {
5170 		/*
5171 		 * We may see a false NULL here because the previous page has
5172 		 * been removed or just inserted and the list is loaded without
5173 		 * barriers.  Switch to radix to verify.
5174 		 */
5175 		if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex ||
5176 		    atomic_load_ptr(&m->object) != object) {
5177 			/*
5178 			 * This guarantees the result is instantaneously
5179 			 * correct.
5180 			 */
5181 			m = NULL;
5182 		}
5183 		m = vm_page_acquire_unlocked(object, pindex, m, flags);
5184 		if (m == PAGE_NOT_ACQUIRED)
5185 			return (i);
5186 		if (m == NULL)
5187 			break;
5188 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5189 			if ((m->flags & PG_ZERO) == 0)
5190 				pmap_zero_page(m);
5191 			vm_page_valid(m);
5192 		}
5193 		/* m will still be wired or busy according to flags. */
5194 		vm_page_grab_release(m, allocflags);
5195 		ma[i] = m;
5196 		m = TAILQ_NEXT(m, listq);
5197 	}
5198 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5199 		return (i);
5200 	count -= i;
5201 	VM_OBJECT_WLOCK(object);
5202 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5203 	VM_OBJECT_WUNLOCK(object);
5204 
5205 	return (i);
5206 }
5207 
5208 /*
5209  * Mapping function for valid or dirty bits in a page.
5210  *
5211  * Inputs are required to range within a page.
5212  */
5213 vm_page_bits_t
5214 vm_page_bits(int base, int size)
5215 {
5216 	int first_bit;
5217 	int last_bit;
5218 
5219 	KASSERT(
5220 	    base + size <= PAGE_SIZE,
5221 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5222 	);
5223 
5224 	if (size == 0)		/* handle degenerate case */
5225 		return (0);
5226 
5227 	first_bit = base >> DEV_BSHIFT;
5228 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5229 
5230 	return (((vm_page_bits_t)2 << last_bit) -
5231 	    ((vm_page_bits_t)1 << first_bit));
5232 }
5233 
5234 void
5235 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5236 {
5237 
5238 #if PAGE_SIZE == 32768
5239 	atomic_set_64((uint64_t *)bits, set);
5240 #elif PAGE_SIZE == 16384
5241 	atomic_set_32((uint32_t *)bits, set);
5242 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5243 	atomic_set_16((uint16_t *)bits, set);
5244 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5245 	atomic_set_8((uint8_t *)bits, set);
5246 #else		/* PAGE_SIZE <= 8192 */
5247 	uintptr_t addr;
5248 	int shift;
5249 
5250 	addr = (uintptr_t)bits;
5251 	/*
5252 	 * Use a trick to perform a 32-bit atomic on the
5253 	 * containing aligned word, to not depend on the existence
5254 	 * of atomic_{set, clear}_{8, 16}.
5255 	 */
5256 	shift = addr & (sizeof(uint32_t) - 1);
5257 #if BYTE_ORDER == BIG_ENDIAN
5258 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5259 #else
5260 	shift *= NBBY;
5261 #endif
5262 	addr &= ~(sizeof(uint32_t) - 1);
5263 	atomic_set_32((uint32_t *)addr, set << shift);
5264 #endif		/* PAGE_SIZE */
5265 }
5266 
5267 static inline void
5268 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5269 {
5270 
5271 #if PAGE_SIZE == 32768
5272 	atomic_clear_64((uint64_t *)bits, clear);
5273 #elif PAGE_SIZE == 16384
5274 	atomic_clear_32((uint32_t *)bits, clear);
5275 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5276 	atomic_clear_16((uint16_t *)bits, clear);
5277 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5278 	atomic_clear_8((uint8_t *)bits, clear);
5279 #else		/* PAGE_SIZE <= 8192 */
5280 	uintptr_t addr;
5281 	int shift;
5282 
5283 	addr = (uintptr_t)bits;
5284 	/*
5285 	 * Use a trick to perform a 32-bit atomic on the
5286 	 * containing aligned word, to not depend on the existence
5287 	 * of atomic_{set, clear}_{8, 16}.
5288 	 */
5289 	shift = addr & (sizeof(uint32_t) - 1);
5290 #if BYTE_ORDER == BIG_ENDIAN
5291 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5292 #else
5293 	shift *= NBBY;
5294 #endif
5295 	addr &= ~(sizeof(uint32_t) - 1);
5296 	atomic_clear_32((uint32_t *)addr, clear << shift);
5297 #endif		/* PAGE_SIZE */
5298 }
5299 
5300 static inline vm_page_bits_t
5301 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5302 {
5303 #if PAGE_SIZE == 32768
5304 	uint64_t old;
5305 
5306 	old = *bits;
5307 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5308 	return (old);
5309 #elif PAGE_SIZE == 16384
5310 	uint32_t old;
5311 
5312 	old = *bits;
5313 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5314 	return (old);
5315 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5316 	uint16_t old;
5317 
5318 	old = *bits;
5319 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5320 	return (old);
5321 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5322 	uint8_t old;
5323 
5324 	old = *bits;
5325 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5326 	return (old);
5327 #else		/* PAGE_SIZE <= 4096*/
5328 	uintptr_t addr;
5329 	uint32_t old, new, mask;
5330 	int shift;
5331 
5332 	addr = (uintptr_t)bits;
5333 	/*
5334 	 * Use a trick to perform a 32-bit atomic on the
5335 	 * containing aligned word, to not depend on the existence
5336 	 * of atomic_{set, swap, clear}_{8, 16}.
5337 	 */
5338 	shift = addr & (sizeof(uint32_t) - 1);
5339 #if BYTE_ORDER == BIG_ENDIAN
5340 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5341 #else
5342 	shift *= NBBY;
5343 #endif
5344 	addr &= ~(sizeof(uint32_t) - 1);
5345 	mask = VM_PAGE_BITS_ALL << shift;
5346 
5347 	old = *bits;
5348 	do {
5349 		new = old & ~mask;
5350 		new |= newbits << shift;
5351 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5352 	return (old >> shift);
5353 #endif		/* PAGE_SIZE */
5354 }
5355 
5356 /*
5357  *	vm_page_set_valid_range:
5358  *
5359  *	Sets portions of a page valid.  The arguments are expected
5360  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5361  *	of any partial chunks touched by the range.  The invalid portion of
5362  *	such chunks will be zeroed.
5363  *
5364  *	(base + size) must be less then or equal to PAGE_SIZE.
5365  */
5366 void
5367 vm_page_set_valid_range(vm_page_t m, int base, int size)
5368 {
5369 	int endoff, frag;
5370 	vm_page_bits_t pagebits;
5371 
5372 	vm_page_assert_busied(m);
5373 	if (size == 0)	/* handle degenerate case */
5374 		return;
5375 
5376 	/*
5377 	 * If the base is not DEV_BSIZE aligned and the valid
5378 	 * bit is clear, we have to zero out a portion of the
5379 	 * first block.
5380 	 */
5381 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5382 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5383 		pmap_zero_page_area(m, frag, base - frag);
5384 
5385 	/*
5386 	 * If the ending offset is not DEV_BSIZE aligned and the
5387 	 * valid bit is clear, we have to zero out a portion of
5388 	 * the last block.
5389 	 */
5390 	endoff = base + size;
5391 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5392 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5393 		pmap_zero_page_area(m, endoff,
5394 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5395 
5396 	/*
5397 	 * Assert that no previously invalid block that is now being validated
5398 	 * is already dirty.
5399 	 */
5400 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5401 	    ("vm_page_set_valid_range: page %p is dirty", m));
5402 
5403 	/*
5404 	 * Set valid bits inclusive of any overlap.
5405 	 */
5406 	pagebits = vm_page_bits(base, size);
5407 	if (vm_page_xbusied(m))
5408 		m->valid |= pagebits;
5409 	else
5410 		vm_page_bits_set(m, &m->valid, pagebits);
5411 }
5412 
5413 /*
5414  * Set the page dirty bits and free the invalid swap space if
5415  * present.  Returns the previous dirty bits.
5416  */
5417 vm_page_bits_t
5418 vm_page_set_dirty(vm_page_t m)
5419 {
5420 	vm_page_bits_t old;
5421 
5422 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5423 
5424 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5425 		old = m->dirty;
5426 		m->dirty = VM_PAGE_BITS_ALL;
5427 	} else
5428 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5429 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5430 		vm_pager_page_unswapped(m);
5431 
5432 	return (old);
5433 }
5434 
5435 /*
5436  * Clear the given bits from the specified page's dirty field.
5437  */
5438 static __inline void
5439 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5440 {
5441 
5442 	vm_page_assert_busied(m);
5443 
5444 	/*
5445 	 * If the page is xbusied and not write mapped we are the
5446 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5447 	 * layer can call vm_page_dirty() without holding a distinguished
5448 	 * lock.  The combination of page busy and atomic operations
5449 	 * suffice to guarantee consistency of the page dirty field.
5450 	 */
5451 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5452 		m->dirty &= ~pagebits;
5453 	else
5454 		vm_page_bits_clear(m, &m->dirty, pagebits);
5455 }
5456 
5457 /*
5458  *	vm_page_set_validclean:
5459  *
5460  *	Sets portions of a page valid and clean.  The arguments are expected
5461  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5462  *	of any partial chunks touched by the range.  The invalid portion of
5463  *	such chunks will be zero'd.
5464  *
5465  *	(base + size) must be less then or equal to PAGE_SIZE.
5466  */
5467 void
5468 vm_page_set_validclean(vm_page_t m, int base, int size)
5469 {
5470 	vm_page_bits_t oldvalid, pagebits;
5471 	int endoff, frag;
5472 
5473 	vm_page_assert_busied(m);
5474 	if (size == 0)	/* handle degenerate case */
5475 		return;
5476 
5477 	/*
5478 	 * If the base is not DEV_BSIZE aligned and the valid
5479 	 * bit is clear, we have to zero out a portion of the
5480 	 * first block.
5481 	 */
5482 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5483 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5484 		pmap_zero_page_area(m, frag, base - frag);
5485 
5486 	/*
5487 	 * If the ending offset is not DEV_BSIZE aligned and the
5488 	 * valid bit is clear, we have to zero out a portion of
5489 	 * the last block.
5490 	 */
5491 	endoff = base + size;
5492 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5493 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5494 		pmap_zero_page_area(m, endoff,
5495 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5496 
5497 	/*
5498 	 * Set valid, clear dirty bits.  If validating the entire
5499 	 * page we can safely clear the pmap modify bit.  We also
5500 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5501 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5502 	 * be set again.
5503 	 *
5504 	 * We set valid bits inclusive of any overlap, but we can only
5505 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5506 	 * the range.
5507 	 */
5508 	oldvalid = m->valid;
5509 	pagebits = vm_page_bits(base, size);
5510 	if (vm_page_xbusied(m))
5511 		m->valid |= pagebits;
5512 	else
5513 		vm_page_bits_set(m, &m->valid, pagebits);
5514 #if 0	/* NOT YET */
5515 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5516 		frag = DEV_BSIZE - frag;
5517 		base += frag;
5518 		size -= frag;
5519 		if (size < 0)
5520 			size = 0;
5521 	}
5522 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5523 #endif
5524 	if (base == 0 && size == PAGE_SIZE) {
5525 		/*
5526 		 * The page can only be modified within the pmap if it is
5527 		 * mapped, and it can only be mapped if it was previously
5528 		 * fully valid.
5529 		 */
5530 		if (oldvalid == VM_PAGE_BITS_ALL)
5531 			/*
5532 			 * Perform the pmap_clear_modify() first.  Otherwise,
5533 			 * a concurrent pmap operation, such as
5534 			 * pmap_protect(), could clear a modification in the
5535 			 * pmap and set the dirty field on the page before
5536 			 * pmap_clear_modify() had begun and after the dirty
5537 			 * field was cleared here.
5538 			 */
5539 			pmap_clear_modify(m);
5540 		m->dirty = 0;
5541 		vm_page_aflag_clear(m, PGA_NOSYNC);
5542 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5543 		m->dirty &= ~pagebits;
5544 	else
5545 		vm_page_clear_dirty_mask(m, pagebits);
5546 }
5547 
5548 void
5549 vm_page_clear_dirty(vm_page_t m, int base, int size)
5550 {
5551 
5552 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5553 }
5554 
5555 /*
5556  *	vm_page_set_invalid:
5557  *
5558  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5559  *	valid and dirty bits for the effected areas are cleared.
5560  */
5561 void
5562 vm_page_set_invalid(vm_page_t m, int base, int size)
5563 {
5564 	vm_page_bits_t bits;
5565 	vm_object_t object;
5566 
5567 	/*
5568 	 * The object lock is required so that pages can't be mapped
5569 	 * read-only while we're in the process of invalidating them.
5570 	 */
5571 	object = m->object;
5572 	VM_OBJECT_ASSERT_WLOCKED(object);
5573 	vm_page_assert_busied(m);
5574 
5575 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5576 	    size >= object->un_pager.vnp.vnp_size)
5577 		bits = VM_PAGE_BITS_ALL;
5578 	else
5579 		bits = vm_page_bits(base, size);
5580 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5581 		pmap_remove_all(m);
5582 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5583 	    !pmap_page_is_mapped(m),
5584 	    ("vm_page_set_invalid: page %p is mapped", m));
5585 	if (vm_page_xbusied(m)) {
5586 		m->valid &= ~bits;
5587 		m->dirty &= ~bits;
5588 	} else {
5589 		vm_page_bits_clear(m, &m->valid, bits);
5590 		vm_page_bits_clear(m, &m->dirty, bits);
5591 	}
5592 }
5593 
5594 /*
5595  *	vm_page_invalid:
5596  *
5597  *	Invalidates the entire page.  The page must be busy, unmapped, and
5598  *	the enclosing object must be locked.  The object locks protects
5599  *	against concurrent read-only pmap enter which is done without
5600  *	busy.
5601  */
5602 void
5603 vm_page_invalid(vm_page_t m)
5604 {
5605 
5606 	vm_page_assert_busied(m);
5607 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5608 	MPASS(!pmap_page_is_mapped(m));
5609 
5610 	if (vm_page_xbusied(m))
5611 		m->valid = 0;
5612 	else
5613 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5614 }
5615 
5616 /*
5617  * vm_page_zero_invalid()
5618  *
5619  *	The kernel assumes that the invalid portions of a page contain
5620  *	garbage, but such pages can be mapped into memory by user code.
5621  *	When this occurs, we must zero out the non-valid portions of the
5622  *	page so user code sees what it expects.
5623  *
5624  *	Pages are most often semi-valid when the end of a file is mapped
5625  *	into memory and the file's size is not page aligned.
5626  */
5627 void
5628 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5629 {
5630 	int b;
5631 	int i;
5632 
5633 	/*
5634 	 * Scan the valid bits looking for invalid sections that
5635 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5636 	 * valid bit may be set ) have already been zeroed by
5637 	 * vm_page_set_validclean().
5638 	 */
5639 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5640 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5641 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5642 			if (i > b) {
5643 				pmap_zero_page_area(m,
5644 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5645 			}
5646 			b = i + 1;
5647 		}
5648 	}
5649 
5650 	/*
5651 	 * setvalid is TRUE when we can safely set the zero'd areas
5652 	 * as being valid.  We can do this if there are no cache consistency
5653 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5654 	 */
5655 	if (setvalid)
5656 		vm_page_valid(m);
5657 }
5658 
5659 /*
5660  *	vm_page_is_valid:
5661  *
5662  *	Is (partial) page valid?  Note that the case where size == 0
5663  *	will return FALSE in the degenerate case where the page is
5664  *	entirely invalid, and TRUE otherwise.
5665  *
5666  *	Some callers envoke this routine without the busy lock held and
5667  *	handle races via higher level locks.  Typical callers should
5668  *	hold a busy lock to prevent invalidation.
5669  */
5670 int
5671 vm_page_is_valid(vm_page_t m, int base, int size)
5672 {
5673 	vm_page_bits_t bits;
5674 
5675 	bits = vm_page_bits(base, size);
5676 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5677 }
5678 
5679 /*
5680  * Returns true if all of the specified predicates are true for the entire
5681  * (super)page and false otherwise.
5682  */
5683 bool
5684 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5685 {
5686 	vm_object_t object;
5687 	int i, npages;
5688 
5689 	object = m->object;
5690 	if (skip_m != NULL && skip_m->object != object)
5691 		return (false);
5692 	VM_OBJECT_ASSERT_LOCKED(object);
5693 	KASSERT(psind <= m->psind,
5694 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5695 	npages = atop(pagesizes[psind]);
5696 
5697 	/*
5698 	 * The physically contiguous pages that make up a superpage, i.e., a
5699 	 * page with a page size index ("psind") greater than zero, will
5700 	 * occupy adjacent entries in vm_page_array[].
5701 	 */
5702 	for (i = 0; i < npages; i++) {
5703 		/* Always test object consistency, including "skip_m". */
5704 		if (m[i].object != object)
5705 			return (false);
5706 		if (&m[i] == skip_m)
5707 			continue;
5708 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5709 			return (false);
5710 		if ((flags & PS_ALL_DIRTY) != 0) {
5711 			/*
5712 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5713 			 * might stop this case from spuriously returning
5714 			 * "false".  However, that would require a write lock
5715 			 * on the object containing "m[i]".
5716 			 */
5717 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5718 				return (false);
5719 		}
5720 		if ((flags & PS_ALL_VALID) != 0 &&
5721 		    m[i].valid != VM_PAGE_BITS_ALL)
5722 			return (false);
5723 	}
5724 	return (true);
5725 }
5726 
5727 /*
5728  * Set the page's dirty bits if the page is modified.
5729  */
5730 void
5731 vm_page_test_dirty(vm_page_t m)
5732 {
5733 
5734 	vm_page_assert_busied(m);
5735 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5736 		vm_page_dirty(m);
5737 }
5738 
5739 void
5740 vm_page_valid(vm_page_t m)
5741 {
5742 
5743 	vm_page_assert_busied(m);
5744 	if (vm_page_xbusied(m))
5745 		m->valid = VM_PAGE_BITS_ALL;
5746 	else
5747 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5748 }
5749 
5750 void
5751 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5752 {
5753 
5754 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5755 }
5756 
5757 void
5758 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5759 {
5760 
5761 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5762 }
5763 
5764 int
5765 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5766 {
5767 
5768 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5769 }
5770 
5771 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5772 void
5773 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5774 {
5775 
5776 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5777 }
5778 
5779 void
5780 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5781 {
5782 
5783 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5784 }
5785 #endif
5786 
5787 #ifdef INVARIANTS
5788 void
5789 vm_page_object_busy_assert(vm_page_t m)
5790 {
5791 
5792 	/*
5793 	 * Certain of the page's fields may only be modified by the
5794 	 * holder of a page or object busy.
5795 	 */
5796 	if (m->object != NULL && !vm_page_busied(m))
5797 		VM_OBJECT_ASSERT_BUSY(m->object);
5798 }
5799 
5800 void
5801 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5802 {
5803 
5804 	if ((bits & PGA_WRITEABLE) == 0)
5805 		return;
5806 
5807 	/*
5808 	 * The PGA_WRITEABLE flag can only be set if the page is
5809 	 * managed, is exclusively busied or the object is locked.
5810 	 * Currently, this flag is only set by pmap_enter().
5811 	 */
5812 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5813 	    ("PGA_WRITEABLE on unmanaged page"));
5814 	if (!vm_page_xbusied(m))
5815 		VM_OBJECT_ASSERT_BUSY(m->object);
5816 }
5817 #endif
5818 
5819 #include "opt_ddb.h"
5820 #ifdef DDB
5821 #include <sys/kernel.h>
5822 
5823 #include <ddb/ddb.h>
5824 
5825 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5826 {
5827 
5828 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5829 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5830 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5831 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5832 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5833 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5834 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5835 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5836 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5837 }
5838 
5839 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5840 {
5841 	int dom;
5842 
5843 	db_printf("pq_free %d\n", vm_free_count());
5844 	for (dom = 0; dom < vm_ndomains; dom++) {
5845 		db_printf(
5846     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5847 		    dom,
5848 		    vm_dom[dom].vmd_page_count,
5849 		    vm_dom[dom].vmd_free_count,
5850 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5851 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5852 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5853 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5854 	}
5855 }
5856 
5857 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5858 {
5859 	vm_page_t m;
5860 	boolean_t phys, virt;
5861 
5862 	if (!have_addr) {
5863 		db_printf("show pginfo addr\n");
5864 		return;
5865 	}
5866 
5867 	phys = strchr(modif, 'p') != NULL;
5868 	virt = strchr(modif, 'v') != NULL;
5869 	if (virt)
5870 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5871 	else if (phys)
5872 		m = PHYS_TO_VM_PAGE(addr);
5873 	else
5874 		m = (vm_page_t)addr;
5875 	db_printf(
5876     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5877     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5878 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5879 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5880 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5881 }
5882 #endif /* DDB */
5883