xref: /freebsd/sys/vm/vm_page.c (revision 4015ff43cbbe819316104258da6a79bef76e52c2)
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 	vmd->vmd_helper_threads_enabled = true;
465 	for (i = 0; i < PQ_COUNT; i++) {
466 		pq = &vmd->vmd_pagequeues[i];
467 		TAILQ_INIT(&pq->pq_pl);
468 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
469 		    MTX_DEF | MTX_DUPOK);
470 		pq->pq_pdpages = 0;
471 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
472 	}
473 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
474 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
475 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
476 
477 	/*
478 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
479 	 * insertions.
480 	 */
481 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
482 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
483 	    &vmd->vmd_inacthead, plinks.q);
484 
485 	/*
486 	 * The clock pages are used to implement active queue scanning without
487 	 * requeues.  Scans start at clock[0], which is advanced after the scan
488 	 * ends.  When the two clock hands meet, they are reset and scanning
489 	 * resumes from the head of the queue.
490 	 */
491 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
492 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
493 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
494 	    &vmd->vmd_clock[0], plinks.q);
495 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
496 	    &vmd->vmd_clock[1], plinks.q);
497 }
498 
499 /*
500  * Initialize a physical page in preparation for adding it to the free
501  * lists.
502  */
503 void
504 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool)
505 {
506 	m->object = NULL;
507 	m->ref_count = 0;
508 	m->busy_lock = VPB_FREED;
509 	m->flags = m->a.flags = 0;
510 	m->phys_addr = pa;
511 	m->a.queue = PQ_NONE;
512 	m->psind = 0;
513 	m->segind = segind;
514 	m->order = VM_NFREEORDER;
515 	m->pool = pool;
516 	m->valid = m->dirty = 0;
517 	pmap_page_init(m);
518 }
519 
520 #ifndef PMAP_HAS_PAGE_ARRAY
521 static vm_paddr_t
522 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
523 {
524 	vm_paddr_t new_end;
525 
526 	/*
527 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
528 	 * However, because this page is allocated from KVM, out-of-bounds
529 	 * accesses using the direct map will not be trapped.
530 	 */
531 	*vaddr += PAGE_SIZE;
532 
533 	/*
534 	 * Allocate physical memory for the page structures, and map it.
535 	 */
536 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
537 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
538 	    VM_PROT_READ | VM_PROT_WRITE);
539 	vm_page_array_size = page_range;
540 
541 	return (new_end);
542 }
543 #endif
544 
545 /*
546  *	vm_page_startup:
547  *
548  *	Initializes the resident memory module.  Allocates physical memory for
549  *	bootstrapping UMA and some data structures that are used to manage
550  *	physical pages.  Initializes these structures, and populates the free
551  *	page queues.
552  */
553 vm_offset_t
554 vm_page_startup(vm_offset_t vaddr)
555 {
556 	struct vm_phys_seg *seg;
557 	struct vm_domain *vmd;
558 	vm_page_t m;
559 	char *list, *listend;
560 	vm_paddr_t end, high_avail, low_avail, new_end, size;
561 	vm_paddr_t page_range __unused;
562 	vm_paddr_t last_pa, pa, startp, endp;
563 	u_long pagecount;
564 #if MINIDUMP_PAGE_TRACKING
565 	u_long vm_page_dump_size;
566 #endif
567 	int biggestone, i, segind;
568 #ifdef WITNESS
569 	vm_offset_t mapped;
570 	int witness_size;
571 #endif
572 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
573 	long ii;
574 #endif
575 	int pool;
576 #ifdef VM_FREEPOOL_LAZYINIT
577 	int lazyinit;
578 #endif
579 
580 	vaddr = round_page(vaddr);
581 
582 	vm_phys_early_startup();
583 	biggestone = vm_phys_avail_largest();
584 	end = phys_avail[biggestone+1];
585 
586 	/*
587 	 * Initialize the page and queue locks.
588 	 */
589 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
590 	for (i = 0; i < PA_LOCK_COUNT; i++)
591 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
592 	for (i = 0; i < vm_ndomains; i++)
593 		vm_page_domain_init(i);
594 
595 	new_end = end;
596 #ifdef WITNESS
597 	witness_size = round_page(witness_startup_count());
598 	new_end -= witness_size;
599 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
600 	    VM_PROT_READ | VM_PROT_WRITE);
601 	bzero((void *)mapped, witness_size);
602 	witness_startup((void *)mapped);
603 #endif
604 
605 #if MINIDUMP_PAGE_TRACKING
606 	/*
607 	 * Allocate a bitmap to indicate that a random physical page
608 	 * needs to be included in a minidump.
609 	 *
610 	 * The amd64 port needs this to indicate which direct map pages
611 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
612 	 *
613 	 * However, i386 still needs this workspace internally within the
614 	 * minidump code.  In theory, they are not needed on i386, but are
615 	 * included should the sf_buf code decide to use them.
616 	 */
617 	last_pa = 0;
618 	vm_page_dump_pages = 0;
619 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
620 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
621 		    dump_avail[i] / PAGE_SIZE;
622 		if (dump_avail[i + 1] > last_pa)
623 			last_pa = dump_avail[i + 1];
624 	}
625 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
626 	new_end -= vm_page_dump_size;
627 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
628 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
629 	bzero((void *)vm_page_dump, vm_page_dump_size);
630 #if MINIDUMP_STARTUP_PAGE_TRACKING
631 	/*
632 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
633 	 * in a crash dump.  When pmap_map() uses the direct map, they are
634 	 * not automatically included.
635 	 */
636 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
637 		dump_add_page(pa);
638 #endif
639 #else
640 	(void)last_pa;
641 #endif
642 	phys_avail[biggestone + 1] = new_end;
643 #ifdef __amd64__
644 	/*
645 	 * Request that the physical pages underlying the message buffer be
646 	 * included in a crash dump.  Since the message buffer is accessed
647 	 * through the direct map, they are not automatically included.
648 	 */
649 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
650 	last_pa = pa + round_page(msgbufsize);
651 	while (pa < last_pa) {
652 		dump_add_page(pa);
653 		pa += PAGE_SIZE;
654 	}
655 #else
656 	(void)pa;
657 #endif
658 	/*
659 	 * Compute the number of pages of memory that will be available for
660 	 * use, taking into account the overhead of a page structure per page.
661 	 * In other words, solve
662 	 *	"available physical memory" - round_page(page_range *
663 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
664 	 * for page_range.
665 	 */
666 	low_avail = phys_avail[0];
667 	high_avail = phys_avail[1];
668 	for (i = 0; i < vm_phys_nsegs; i++) {
669 		if (vm_phys_segs[i].start < low_avail)
670 			low_avail = vm_phys_segs[i].start;
671 		if (vm_phys_segs[i].end > high_avail)
672 			high_avail = vm_phys_segs[i].end;
673 	}
674 	/* Skip the first chunk.  It is already accounted for. */
675 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
676 		if (phys_avail[i] < low_avail)
677 			low_avail = phys_avail[i];
678 		if (phys_avail[i + 1] > high_avail)
679 			high_avail = phys_avail[i + 1];
680 	}
681 	first_page = low_avail / PAGE_SIZE;
682 #ifdef VM_PHYSSEG_SPARSE
683 	size = 0;
684 	for (i = 0; i < vm_phys_nsegs; i++)
685 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
686 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
687 		size += phys_avail[i + 1] - phys_avail[i];
688 #elif defined(VM_PHYSSEG_DENSE)
689 	size = high_avail - low_avail;
690 #else
691 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
692 #endif
693 
694 #ifdef PMAP_HAS_PAGE_ARRAY
695 	pmap_page_array_startup(size / PAGE_SIZE);
696 	biggestone = vm_phys_avail_largest();
697 	end = new_end = phys_avail[biggestone + 1];
698 #else
699 #ifdef VM_PHYSSEG_DENSE
700 	/*
701 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
702 	 * the overhead of a page structure per page only if vm_page_array is
703 	 * allocated from the last physical memory chunk.  Otherwise, we must
704 	 * allocate page structures representing the physical memory
705 	 * underlying vm_page_array, even though they will not be used.
706 	 */
707 	if (new_end != high_avail)
708 		page_range = size / PAGE_SIZE;
709 	else
710 #endif
711 	{
712 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
713 
714 		/*
715 		 * If the partial bytes remaining are large enough for
716 		 * a page (PAGE_SIZE) without a corresponding
717 		 * 'struct vm_page', then new_end will contain an
718 		 * extra page after subtracting the length of the VM
719 		 * page array.  Compensate by subtracting an extra
720 		 * page from new_end.
721 		 */
722 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
723 			if (new_end == high_avail)
724 				high_avail -= PAGE_SIZE;
725 			new_end -= PAGE_SIZE;
726 		}
727 	}
728 	end = new_end;
729 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
730 #endif
731 
732 #if VM_NRESERVLEVEL > 0
733 	/*
734 	 * Allocate physical memory for the reservation management system's
735 	 * data structures, and map it.
736 	 */
737 	new_end = vm_reserv_startup(&vaddr, new_end);
738 #endif
739 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
740 	/*
741 	 * Include vm_page_array and vm_reserv_array in a crash dump.
742 	 */
743 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
744 		dump_add_page(pa);
745 #endif
746 	phys_avail[biggestone + 1] = new_end;
747 
748 	/*
749 	 * Add physical memory segments corresponding to the available
750 	 * physical pages.
751 	 */
752 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
753 		if (vm_phys_avail_size(i) != 0)
754 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
755 
756 	/*
757 	 * Initialize the physical memory allocator.
758 	 */
759 	vm_phys_init();
760 
761 	pool = VM_FREEPOOL_DEFAULT;
762 #ifdef VM_FREEPOOL_LAZYINIT
763 	lazyinit = 1;
764 	TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit);
765 	if (lazyinit)
766 		pool = VM_FREEPOOL_LAZYINIT;
767 #endif
768 
769 	/*
770 	 * Initialize the page structures and add every available page to the
771 	 * physical memory allocator's free lists.
772 	 */
773 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
774 	for (ii = 0; ii < vm_page_array_size; ii++) {
775 		m = &vm_page_array[ii];
776 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0,
777 		    VM_FREEPOOL_DEFAULT);
778 		m->flags = PG_FICTITIOUS;
779 	}
780 #endif
781 	vm_cnt.v_page_count = 0;
782 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
783 		seg = &vm_phys_segs[segind];
784 
785 		/*
786 		 * Initialize pages not covered by phys_avail[], since they
787 		 * might be freed to the allocator at some future point, e.g.,
788 		 * by kmem_bootstrap_free().
789 		 */
790 		startp = seg->start;
791 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
792 			if (startp >= seg->end)
793 				break;
794 			if (phys_avail[i + 1] < startp)
795 				continue;
796 			if (phys_avail[i] <= startp) {
797 				startp = phys_avail[i + 1];
798 				continue;
799 			}
800 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
801 			for (endp = MIN(phys_avail[i], seg->end);
802 			    startp < endp; startp += PAGE_SIZE, m++) {
803 				vm_page_init_page(m, startp, segind,
804 				    VM_FREEPOOL_DEFAULT);
805 			}
806 		}
807 
808 		/*
809 		 * Add the segment's pages that are covered by one of
810 		 * phys_avail's ranges to the free lists.
811 		 */
812 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
813 			if (seg->end <= phys_avail[i] ||
814 			    seg->start >= phys_avail[i + 1])
815 				continue;
816 
817 			startp = MAX(seg->start, phys_avail[i]);
818 			endp = MIN(seg->end, phys_avail[i + 1]);
819 			pagecount = (u_long)atop(endp - startp);
820 			if (pagecount == 0)
821 				continue;
822 
823 			/*
824 			 * If lazy vm_page initialization is not enabled, simply
825 			 * initialize all of the pages in the segment covered by
826 			 * phys_avail.  Otherwise, initialize only the first
827 			 * page of each run of free pages handed to the vm_phys
828 			 * allocator, which in turn defers initialization of
829 			 * pages until they are needed.
830 			 *
831 			 * This avoids blocking the boot process for long
832 			 * periods, which may be relevant for VMs (which ought
833 			 * to boot as quickly as possible) and/or systems with
834 			 * large amounts of physical memory.
835 			 */
836 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
837 			vm_page_init_page(m, startp, segind, pool);
838 			if (pool == VM_FREEPOOL_DEFAULT) {
839 				for (u_long j = 1; j < pagecount; j++) {
840 					vm_page_init_page(&m[j],
841 					    startp + ptoa((vm_paddr_t)j),
842 					    segind, pool);
843 				}
844 			}
845 			vmd = VM_DOMAIN(seg->domain);
846 			vm_domain_free_lock(vmd);
847 			vm_phys_enqueue_contig(m, pool, pagecount);
848 			vm_domain_free_unlock(vmd);
849 			vm_domain_freecnt_inc(vmd, pagecount);
850 			vm_cnt.v_page_count += (u_int)pagecount;
851 			vmd->vmd_page_count += (u_int)pagecount;
852 			vmd->vmd_segs |= 1UL << segind;
853 		}
854 	}
855 
856 	/*
857 	 * Remove blacklisted pages from the physical memory allocator.
858 	 */
859 	TAILQ_INIT(&blacklist_head);
860 	vm_page_blacklist_load(&list, &listend);
861 	vm_page_blacklist_check(list, listend);
862 
863 	list = kern_getenv("vm.blacklist");
864 	vm_page_blacklist_check(list, NULL);
865 
866 	freeenv(list);
867 #if VM_NRESERVLEVEL > 0
868 	/*
869 	 * Initialize the reservation management system.
870 	 */
871 	vm_reserv_init();
872 #endif
873 
874 	return (vaddr);
875 }
876 
877 void
878 vm_page_reference(vm_page_t m)
879 {
880 
881 	vm_page_aflag_set(m, PGA_REFERENCED);
882 }
883 
884 /*
885  *	vm_page_trybusy
886  *
887  *	Helper routine for grab functions to trylock busy.
888  *
889  *	Returns true on success and false on failure.
890  */
891 static bool
892 vm_page_trybusy(vm_page_t m, int allocflags)
893 {
894 
895 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
896 		return (vm_page_trysbusy(m));
897 	else
898 		return (vm_page_tryxbusy(m));
899 }
900 
901 /*
902  *	vm_page_tryacquire
903  *
904  *	Helper routine for grab functions to trylock busy and wire.
905  *
906  *	Returns true on success and false on failure.
907  */
908 static inline bool
909 vm_page_tryacquire(vm_page_t m, int allocflags)
910 {
911 	bool locked;
912 
913 	locked = vm_page_trybusy(m, allocflags);
914 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
915 		vm_page_wire(m);
916 	return (locked);
917 }
918 
919 /*
920  *	vm_page_busy_acquire:
921  *
922  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
923  *	and drop the object lock if necessary.
924  */
925 bool
926 vm_page_busy_acquire(vm_page_t m, int allocflags)
927 {
928 	vm_object_t obj;
929 	bool locked;
930 
931 	/*
932 	 * The page-specific object must be cached because page
933 	 * identity can change during the sleep, causing the
934 	 * re-lock of a different object.
935 	 * It is assumed that a reference to the object is already
936 	 * held by the callers.
937 	 */
938 	obj = atomic_load_ptr(&m->object);
939 	for (;;) {
940 		if (vm_page_tryacquire(m, allocflags))
941 			return (true);
942 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
943 			return (false);
944 		if (obj != NULL)
945 			locked = VM_OBJECT_WOWNED(obj);
946 		else
947 			locked = false;
948 		MPASS(locked || vm_page_wired(m));
949 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
950 		    locked) && locked)
951 			VM_OBJECT_WLOCK(obj);
952 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
953 			return (false);
954 		KASSERT(m->object == obj || m->object == NULL,
955 		    ("vm_page_busy_acquire: page %p does not belong to %p",
956 		    m, obj));
957 	}
958 }
959 
960 /*
961  *	vm_page_busy_downgrade:
962  *
963  *	Downgrade an exclusive busy page into a single shared busy page.
964  */
965 void
966 vm_page_busy_downgrade(vm_page_t m)
967 {
968 	u_int x;
969 
970 	vm_page_assert_xbusied(m);
971 
972 	x = vm_page_busy_fetch(m);
973 	for (;;) {
974 		if (atomic_fcmpset_rel_int(&m->busy_lock,
975 		    &x, VPB_SHARERS_WORD(1)))
976 			break;
977 	}
978 	if ((x & VPB_BIT_WAITERS) != 0)
979 		wakeup(m);
980 }
981 
982 /*
983  *
984  *	vm_page_busy_tryupgrade:
985  *
986  *	Attempt to upgrade a single shared busy into an exclusive busy.
987  */
988 int
989 vm_page_busy_tryupgrade(vm_page_t m)
990 {
991 	u_int ce, x;
992 
993 	vm_page_assert_sbusied(m);
994 
995 	x = vm_page_busy_fetch(m);
996 	ce = VPB_CURTHREAD_EXCLUSIVE;
997 	for (;;) {
998 		if (VPB_SHARERS(x) > 1)
999 			return (0);
1000 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1001 		    ("vm_page_busy_tryupgrade: invalid lock state"));
1002 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1003 		    ce | (x & VPB_BIT_WAITERS)))
1004 			continue;
1005 		return (1);
1006 	}
1007 }
1008 
1009 /*
1010  *	vm_page_sbusied:
1011  *
1012  *	Return a positive value if the page is shared busied, 0 otherwise.
1013  */
1014 int
1015 vm_page_sbusied(vm_page_t m)
1016 {
1017 	u_int x;
1018 
1019 	x = vm_page_busy_fetch(m);
1020 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1021 }
1022 
1023 /*
1024  *	vm_page_sunbusy:
1025  *
1026  *	Shared unbusy a page.
1027  */
1028 void
1029 vm_page_sunbusy(vm_page_t m)
1030 {
1031 	u_int x;
1032 
1033 	vm_page_assert_sbusied(m);
1034 
1035 	x = vm_page_busy_fetch(m);
1036 	for (;;) {
1037 		KASSERT(x != VPB_FREED,
1038 		    ("vm_page_sunbusy: Unlocking freed page."));
1039 		if (VPB_SHARERS(x) > 1) {
1040 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1041 			    x - VPB_ONE_SHARER))
1042 				break;
1043 			continue;
1044 		}
1045 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1046 		    ("vm_page_sunbusy: invalid lock state"));
1047 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1048 			continue;
1049 		if ((x & VPB_BIT_WAITERS) == 0)
1050 			break;
1051 		wakeup(m);
1052 		break;
1053 	}
1054 }
1055 
1056 /*
1057  *	vm_page_busy_sleep:
1058  *
1059  *	Sleep if the page is busy, using the page pointer as wchan.
1060  *	This is used to implement the hard-path of the busying mechanism.
1061  *
1062  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1063  *	will not sleep if the page is shared-busy.
1064  *
1065  *	The object lock must be held on entry.
1066  *
1067  *	Returns true if it slept and dropped the object lock, or false
1068  *	if there was no sleep and the lock is still held.
1069  */
1070 bool
1071 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1072 {
1073 	vm_object_t obj;
1074 
1075 	obj = m->object;
1076 	VM_OBJECT_ASSERT_LOCKED(obj);
1077 
1078 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1079 	    true));
1080 }
1081 
1082 /*
1083  *	vm_page_busy_sleep_unlocked:
1084  *
1085  *	Sleep if the page is busy, using the page pointer as wchan.
1086  *	This is used to implement the hard-path of busying mechanism.
1087  *
1088  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1089  *	will not sleep if the page is shared-busy.
1090  *
1091  *	The object lock must not be held on entry.  The operation will
1092  *	return if the page changes identity.
1093  */
1094 void
1095 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1096     const char *wmesg, int allocflags)
1097 {
1098 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1099 
1100 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1101 }
1102 
1103 /*
1104  *	_vm_page_busy_sleep:
1105  *
1106  *	Internal busy sleep function.  Verifies the page identity and
1107  *	lockstate against parameters.  Returns true if it sleeps and
1108  *	false otherwise.
1109  *
1110  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1111  *
1112  *	If locked is true the lock will be dropped for any true returns
1113  *	and held for any false returns.
1114  */
1115 static bool
1116 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1117     const char *wmesg, int allocflags, bool locked)
1118 {
1119 	bool xsleep;
1120 	u_int x;
1121 
1122 	/*
1123 	 * If the object is busy we must wait for that to drain to zero
1124 	 * before trying the page again.
1125 	 */
1126 	if (obj != NULL && vm_object_busied(obj)) {
1127 		if (locked)
1128 			VM_OBJECT_DROP(obj);
1129 		vm_object_busy_wait(obj, wmesg);
1130 		return (true);
1131 	}
1132 
1133 	if (!vm_page_busied(m))
1134 		return (false);
1135 
1136 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1137 	sleepq_lock(m);
1138 	x = vm_page_busy_fetch(m);
1139 	do {
1140 		/*
1141 		 * If the page changes objects or becomes unlocked we can
1142 		 * simply return.
1143 		 */
1144 		if (x == VPB_UNBUSIED ||
1145 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1146 		    m->object != obj || m->pindex != pindex) {
1147 			sleepq_release(m);
1148 			return (false);
1149 		}
1150 		if ((x & VPB_BIT_WAITERS) != 0)
1151 			break;
1152 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1153 	if (locked)
1154 		VM_OBJECT_DROP(obj);
1155 	DROP_GIANT();
1156 	sleepq_add(m, NULL, wmesg, 0, 0);
1157 	sleepq_wait(m, PVM);
1158 	PICKUP_GIANT();
1159 	return (true);
1160 }
1161 
1162 /*
1163  *	vm_page_trysbusy:
1164  *
1165  *	Try to shared busy a page.
1166  *	If the operation succeeds 1 is returned otherwise 0.
1167  *	The operation never sleeps.
1168  */
1169 int
1170 vm_page_trysbusy(vm_page_t m)
1171 {
1172 	vm_object_t obj;
1173 	u_int x;
1174 
1175 	obj = m->object;
1176 	x = vm_page_busy_fetch(m);
1177 	for (;;) {
1178 		if ((x & VPB_BIT_SHARED) == 0)
1179 			return (0);
1180 		/*
1181 		 * Reduce the window for transient busies that will trigger
1182 		 * false negatives in vm_page_ps_test().
1183 		 */
1184 		if (obj != NULL && vm_object_busied(obj))
1185 			return (0);
1186 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1187 		    x + VPB_ONE_SHARER))
1188 			break;
1189 	}
1190 
1191 	/* Refetch the object now that we're guaranteed that it is stable. */
1192 	obj = m->object;
1193 	if (obj != NULL && vm_object_busied(obj)) {
1194 		vm_page_sunbusy(m);
1195 		return (0);
1196 	}
1197 	return (1);
1198 }
1199 
1200 /*
1201  *	vm_page_tryxbusy:
1202  *
1203  *	Try to exclusive busy a page.
1204  *	If the operation succeeds 1 is returned otherwise 0.
1205  *	The operation never sleeps.
1206  */
1207 int
1208 vm_page_tryxbusy(vm_page_t m)
1209 {
1210 	vm_object_t obj;
1211 
1212         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1213             VPB_CURTHREAD_EXCLUSIVE) == 0)
1214 		return (0);
1215 
1216 	obj = m->object;
1217 	if (obj != NULL && vm_object_busied(obj)) {
1218 		vm_page_xunbusy(m);
1219 		return (0);
1220 	}
1221 	return (1);
1222 }
1223 
1224 static void
1225 vm_page_xunbusy_hard_tail(vm_page_t m)
1226 {
1227 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1228 	/* Wake the waiter. */
1229 	wakeup(m);
1230 }
1231 
1232 /*
1233  *	vm_page_xunbusy_hard:
1234  *
1235  *	Called when unbusy has failed because there is a waiter.
1236  */
1237 void
1238 vm_page_xunbusy_hard(vm_page_t m)
1239 {
1240 	vm_page_assert_xbusied(m);
1241 	vm_page_xunbusy_hard_tail(m);
1242 }
1243 
1244 void
1245 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1246 {
1247 	vm_page_assert_xbusied_unchecked(m);
1248 	vm_page_xunbusy_hard_tail(m);
1249 }
1250 
1251 static void
1252 vm_page_busy_free(vm_page_t m)
1253 {
1254 	u_int x;
1255 
1256 	atomic_thread_fence_rel();
1257 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1258 	if ((x & VPB_BIT_WAITERS) != 0)
1259 		wakeup(m);
1260 }
1261 
1262 /*
1263  *	vm_page_unhold_pages:
1264  *
1265  *	Unhold each of the pages that is referenced by the given array.
1266  */
1267 void
1268 vm_page_unhold_pages(vm_page_t *ma, int count)
1269 {
1270 
1271 	for (; count != 0; count--) {
1272 		vm_page_unwire(*ma, PQ_ACTIVE);
1273 		ma++;
1274 	}
1275 }
1276 
1277 vm_page_t
1278 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1279 {
1280 	vm_page_t m;
1281 
1282 #ifdef VM_PHYSSEG_SPARSE
1283 	m = vm_phys_paddr_to_vm_page(pa);
1284 	if (m == NULL)
1285 		m = vm_phys_fictitious_to_vm_page(pa);
1286 	return (m);
1287 #elif defined(VM_PHYSSEG_DENSE)
1288 	long pi;
1289 
1290 	pi = atop(pa);
1291 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1292 		m = &vm_page_array[pi - first_page];
1293 		return (m);
1294 	}
1295 	return (vm_phys_fictitious_to_vm_page(pa));
1296 #else
1297 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1298 #endif
1299 }
1300 
1301 /*
1302  *	vm_page_getfake:
1303  *
1304  *	Create a fictitious page with the specified physical address and
1305  *	memory attribute.  The memory attribute is the only the machine-
1306  *	dependent aspect of a fictitious page that must be initialized.
1307  */
1308 vm_page_t
1309 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1310 {
1311 	vm_page_t m;
1312 
1313 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1314 	vm_page_initfake(m, paddr, memattr);
1315 	return (m);
1316 }
1317 
1318 void
1319 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1320 {
1321 
1322 	if ((m->flags & PG_FICTITIOUS) != 0) {
1323 		/*
1324 		 * The page's memattr might have changed since the
1325 		 * previous initialization.  Update the pmap to the
1326 		 * new memattr.
1327 		 */
1328 		goto memattr;
1329 	}
1330 	m->phys_addr = paddr;
1331 	m->a.queue = PQ_NONE;
1332 	/* Fictitious pages don't use "segind". */
1333 	m->flags = PG_FICTITIOUS;
1334 	/* Fictitious pages don't use "order" or "pool". */
1335 	m->oflags = VPO_UNMANAGED;
1336 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1337 	/* Fictitious pages are unevictable. */
1338 	m->ref_count = 1;
1339 	pmap_page_init(m);
1340 memattr:
1341 	pmap_page_set_memattr(m, memattr);
1342 }
1343 
1344 /*
1345  *	vm_page_putfake:
1346  *
1347  *	Release a fictitious page.
1348  */
1349 void
1350 vm_page_putfake(vm_page_t m)
1351 {
1352 
1353 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1354 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1355 	    ("vm_page_putfake: bad page %p", m));
1356 	vm_page_assert_xbusied(m);
1357 	vm_page_busy_free(m);
1358 	uma_zfree(fakepg_zone, m);
1359 }
1360 
1361 /*
1362  *	vm_page_updatefake:
1363  *
1364  *	Update the given fictitious page to the specified physical address and
1365  *	memory attribute.
1366  */
1367 void
1368 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1369 {
1370 
1371 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1372 	    ("vm_page_updatefake: bad page %p", m));
1373 	m->phys_addr = paddr;
1374 	pmap_page_set_memattr(m, memattr);
1375 }
1376 
1377 /*
1378  *	vm_page_free:
1379  *
1380  *	Free a page.
1381  */
1382 void
1383 vm_page_free(vm_page_t m)
1384 {
1385 
1386 	m->flags &= ~PG_ZERO;
1387 	vm_page_free_toq(m);
1388 }
1389 
1390 /*
1391  *	vm_page_free_zero:
1392  *
1393  *	Free a page to the zerod-pages queue
1394  */
1395 void
1396 vm_page_free_zero(vm_page_t m)
1397 {
1398 
1399 	m->flags |= PG_ZERO;
1400 	vm_page_free_toq(m);
1401 }
1402 
1403 /*
1404  * Unbusy and handle the page queueing for a page from a getpages request that
1405  * was optionally read ahead or behind.
1406  */
1407 void
1408 vm_page_readahead_finish(vm_page_t m)
1409 {
1410 
1411 	/* We shouldn't put invalid pages on queues. */
1412 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1413 
1414 	/*
1415 	 * Since the page is not the actually needed one, whether it should
1416 	 * be activated or deactivated is not obvious.  Empirical results
1417 	 * have shown that deactivating the page is usually the best choice,
1418 	 * unless the page is wanted by another thread.
1419 	 */
1420 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1421 		vm_page_activate(m);
1422 	else
1423 		vm_page_deactivate(m);
1424 	vm_page_xunbusy_unchecked(m);
1425 }
1426 
1427 /*
1428  * Destroy the identity of an invalid page and free it if possible.
1429  * This is intended to be used when reading a page from backing store fails.
1430  */
1431 void
1432 vm_page_free_invalid(vm_page_t m)
1433 {
1434 
1435 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1436 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1437 	KASSERT(m->object != NULL, ("page %p has no object", m));
1438 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1439 
1440 	/*
1441 	 * We may be attempting to free the page as part of the handling for an
1442 	 * I/O error, in which case the page was xbusied by a different thread.
1443 	 */
1444 	vm_page_xbusy_claim(m);
1445 
1446 	/*
1447 	 * If someone has wired this page while the object lock
1448 	 * was not held, then the thread that unwires is responsible
1449 	 * for freeing the page.  Otherwise just free the page now.
1450 	 * The wire count of this unmapped page cannot change while
1451 	 * we have the page xbusy and the page's object wlocked.
1452 	 */
1453 	if (vm_page_remove(m))
1454 		vm_page_free(m);
1455 }
1456 
1457 /*
1458  *	vm_page_dirty_KBI:		[ internal use only ]
1459  *
1460  *	Set all bits in the page's dirty field.
1461  *
1462  *	The object containing the specified page must be locked if the
1463  *	call is made from the machine-independent layer.
1464  *
1465  *	See vm_page_clear_dirty_mask().
1466  *
1467  *	This function should only be called by vm_page_dirty().
1468  */
1469 void
1470 vm_page_dirty_KBI(vm_page_t m)
1471 {
1472 
1473 	/* Refer to this operation by its public name. */
1474 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1475 	m->dirty = VM_PAGE_BITS_ALL;
1476 }
1477 
1478 /*
1479  * Insert the given page into the given object at the given pindex.  mpred is
1480  * used for memq linkage.  From vm_page_insert, lookup is true, mpred is
1481  * initially NULL, and this procedure looks it up.  From vm_page_insert_after
1482  * and vm_page_iter_insert, lookup is false and mpred is known to the caller
1483  * to be valid, and may be NULL if this will be the page with the lowest
1484  * pindex.
1485  *
1486  * The procedure is marked __always_inline to suggest to the compiler to
1487  * eliminate the lookup parameter and the associated alternate branch.
1488  */
1489 static __always_inline int
1490 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1491     struct pctrie_iter *pages, bool iter, vm_page_t mpred, bool lookup)
1492 {
1493 	int error;
1494 
1495 	VM_OBJECT_ASSERT_WLOCKED(object);
1496 	KASSERT(m->object == NULL,
1497 	    ("vm_page_insert: page %p already inserted", m));
1498 
1499 	/*
1500 	 * Record the object/offset pair in this page.
1501 	 */
1502 	m->object = object;
1503 	m->pindex = pindex;
1504 	m->ref_count |= VPRC_OBJREF;
1505 
1506 	/*
1507 	 * Add this page to the object's radix tree, and look up mpred if
1508 	 * needed.
1509 	 */
1510 	if (iter) {
1511 		KASSERT(!lookup, ("%s: cannot lookup mpred", __func__));
1512 		error = vm_radix_iter_insert(pages, m);
1513 	} else if (lookup)
1514 		error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred);
1515 	else
1516 		error = vm_radix_insert(&object->rtree, m);
1517 	if (__predict_false(error != 0)) {
1518 		m->object = NULL;
1519 		m->pindex = 0;
1520 		m->ref_count &= ~VPRC_OBJREF;
1521 		return (1);
1522 	}
1523 
1524 	/*
1525 	 * Now link into the object's ordered list of backed pages.
1526 	 */
1527 	vm_page_insert_radixdone(m, object, mpred);
1528 	vm_pager_page_inserted(object, m);
1529 	return (0);
1530 }
1531 
1532 /*
1533  *	vm_page_insert:		[ internal use only ]
1534  *
1535  *	Inserts the given mem entry into the object and object list.
1536  *
1537  *	The object must be locked.
1538  */
1539 int
1540 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1541 {
1542 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL,
1543 	    true));
1544 }
1545 
1546 /*
1547  *	vm_page_insert_after:
1548  *
1549  *	Inserts the page "m" into the specified object at offset "pindex".
1550  *
1551  *	The page "mpred" must immediately precede the offset "pindex" within
1552  *	the specified object.
1553  *
1554  *	The object must be locked.
1555  */
1556 static int
1557 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1558     vm_page_t mpred)
1559 {
1560 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, mpred,
1561 	    false));
1562 }
1563 
1564 /*
1565  *	vm_page_iter_insert:
1566  *
1567  *	Tries to insert the page "m" into the specified object at offset
1568  *	"pindex" using the iterator "pages".  Returns 0 if the insertion was
1569  *	successful.
1570  *
1571  *	The page "mpred" must immediately precede the offset "pindex" within
1572  *	the specified object.
1573  *
1574  *	The object must be locked.
1575  */
1576 static int
1577 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object,
1578     vm_pindex_t pindex, vm_page_t mpred)
1579 {
1580 	return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred,
1581 	    false));
1582 }
1583 
1584 /*
1585  *	vm_page_insert_radixdone:
1586  *
1587  *	Complete page "m" insertion into the specified object after the
1588  *	radix trie hooking.
1589  *
1590  *	The page "mpred" must precede the offset "m->pindex" within the
1591  *	specified object.
1592  *
1593  *	The object must be locked.
1594  */
1595 static void
1596 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1597 {
1598 
1599 	VM_OBJECT_ASSERT_WLOCKED(object);
1600 	KASSERT(object != NULL && m->object == object,
1601 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1602 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1603 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1604 	if (mpred != NULL) {
1605 		KASSERT(mpred->object == object,
1606 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1607 		KASSERT(mpred->pindex < m->pindex,
1608 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1609 		KASSERT(TAILQ_NEXT(mpred, listq) == NULL ||
1610 		    m->pindex < TAILQ_NEXT(mpred, listq)->pindex,
1611 		    ("vm_page_insert_radixdone: pindex doesn't precede msucc"));
1612 	} else {
1613 		KASSERT(TAILQ_EMPTY(&object->memq) ||
1614 		    m->pindex < TAILQ_FIRST(&object->memq)->pindex,
1615 		    ("vm_page_insert_radixdone: no mpred but not first page"));
1616 	}
1617 
1618 	if (mpred != NULL)
1619 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1620 	else
1621 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1622 
1623 	/*
1624 	 * Show that the object has one more resident page.
1625 	 */
1626 	object->resident_page_count++;
1627 
1628 	/*
1629 	 * Hold the vnode until the last page is released.
1630 	 */
1631 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1632 		vhold(object->handle);
1633 
1634 	/*
1635 	 * Since we are inserting a new and possibly dirty page,
1636 	 * update the object's generation count.
1637 	 */
1638 	if (pmap_page_is_write_mapped(m))
1639 		vm_object_set_writeable_dirty(object);
1640 }
1641 
1642 /*
1643  *	vm_page_remove_radixdone
1644  *
1645  *	Complete page "m" removal from the specified object after the radix trie
1646  *	unhooking.
1647  *
1648  *	The caller is responsible for updating the page's fields to reflect this
1649  *	removal.
1650  */
1651 static void
1652 vm_page_remove_radixdone(vm_page_t m)
1653 {
1654 	vm_object_t object;
1655 
1656 	vm_page_assert_xbusied(m);
1657 	object = m->object;
1658 	VM_OBJECT_ASSERT_WLOCKED(object);
1659 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1660 	    ("page %p is missing its object ref", m));
1661 
1662 	/* Deferred free of swap space. */
1663 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1664 		vm_pager_page_unswapped(m);
1665 
1666 	vm_pager_page_removed(object, m);
1667 	m->object = NULL;
1668 
1669 	/*
1670 	 * Now remove from the object's list of backed pages.
1671 	 */
1672 	TAILQ_REMOVE(&object->memq, m, listq);
1673 
1674 	/*
1675 	 * And show that the object has one fewer resident page.
1676 	 */
1677 	object->resident_page_count--;
1678 
1679 	/*
1680 	 * The vnode may now be recycled.
1681 	 */
1682 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1683 		vdrop(object->handle);
1684 }
1685 
1686 /*
1687  *	vm_page_free_object_prep:
1688  *
1689  *	Disassociates the given page from its VM object.
1690  *
1691  *	The object must be locked, and the page must be xbusy.
1692  */
1693 static void
1694 vm_page_free_object_prep(vm_page_t m)
1695 {
1696 	KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
1697 	    ((m->object->flags & OBJ_UNMANAGED) != 0),
1698 	    ("%s: managed flag mismatch for page %p",
1699 	     __func__, m));
1700 	vm_page_assert_xbusied(m);
1701 
1702 	/*
1703 	 * The object reference can be released without an atomic
1704 	 * operation.
1705 	 */
1706 	KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
1707 	    m->ref_count == VPRC_OBJREF,
1708 	    ("%s: page %p has unexpected ref_count %u",
1709 	    __func__, m, m->ref_count));
1710 	vm_page_remove_radixdone(m);
1711 	m->ref_count -= VPRC_OBJREF;
1712 }
1713 
1714 /*
1715  *	vm_page_iter_free:
1716  *
1717  *	Free the given page, and use the iterator to remove it from the radix
1718  *	tree.
1719  */
1720 void
1721 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m)
1722 {
1723 	vm_radix_iter_remove(pages);
1724 	vm_page_free_object_prep(m);
1725 	vm_page_xunbusy(m);
1726 	m->flags &= ~PG_ZERO;
1727 	vm_page_free_toq(m);
1728 }
1729 
1730 /*
1731  *	vm_page_remove:
1732  *
1733  *	Removes the specified page from its containing object, but does not
1734  *	invalidate any backing storage.  Returns true if the object's reference
1735  *	was the last reference to the page, and false otherwise.
1736  *
1737  *	The object must be locked and the page must be exclusively busied.
1738  *	The exclusive busy will be released on return.  If this is not the
1739  *	final ref and the caller does not hold a wire reference it may not
1740  *	continue to access the page.
1741  */
1742 bool
1743 vm_page_remove(vm_page_t m)
1744 {
1745 	bool dropped;
1746 
1747 	dropped = vm_page_remove_xbusy(m);
1748 	vm_page_xunbusy(m);
1749 
1750 	return (dropped);
1751 }
1752 
1753 /*
1754  *	vm_page_iter_remove:
1755  *
1756  *	Remove the current page, and use the iterator to remove it from the
1757  *	radix tree.
1758  */
1759 bool
1760 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m)
1761 {
1762 	bool dropped;
1763 
1764 	vm_radix_iter_remove(pages);
1765 	vm_page_remove_radixdone(m);
1766 	dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1767 	vm_page_xunbusy(m);
1768 
1769 	return (dropped);
1770 }
1771 
1772 /*
1773  *	vm_page_radix_remove
1774  *
1775  *	Removes the specified page from the radix tree.
1776  */
1777 static void
1778 vm_page_radix_remove(vm_page_t m)
1779 {
1780 	vm_page_t mrem __diagused;
1781 
1782 	mrem = vm_radix_remove(&m->object->rtree, m->pindex);
1783 	KASSERT(mrem == m,
1784 	    ("removed page %p, expected page %p", mrem, m));
1785 }
1786 
1787 /*
1788  *	vm_page_remove_xbusy
1789  *
1790  *	Removes the page but leaves the xbusy held.  Returns true if this
1791  *	removed the final ref and false otherwise.
1792  */
1793 bool
1794 vm_page_remove_xbusy(vm_page_t m)
1795 {
1796 
1797 	vm_page_radix_remove(m);
1798 	vm_page_remove_radixdone(m);
1799 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1800 }
1801 
1802 /*
1803  *	vm_page_lookup:
1804  *
1805  *	Returns the page associated with the object/offset
1806  *	pair specified; if none is found, NULL is returned.
1807  *
1808  *	The object must be locked.
1809  */
1810 vm_page_t
1811 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1812 {
1813 
1814 	VM_OBJECT_ASSERT_LOCKED(object);
1815 	return (vm_radix_lookup(&object->rtree, pindex));
1816 }
1817 
1818 /*
1819  *	vm_page_iter_init:
1820  *
1821  *	Initialize iterator for vm pages.
1822  */
1823 void
1824 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object)
1825 {
1826 
1827 	vm_radix_iter_init(pages, &object->rtree);
1828 }
1829 
1830 /*
1831  *	vm_page_iter_init:
1832  *
1833  *	Initialize iterator for vm pages.
1834  */
1835 void
1836 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object,
1837     vm_pindex_t limit)
1838 {
1839 
1840 	vm_radix_iter_limit_init(pages, &object->rtree, limit);
1841 }
1842 
1843 /*
1844  *	vm_page_iter_lookup:
1845  *
1846  *	Returns the page associated with the object/offset pair specified, and
1847  *	stores the path to its position; if none is found, NULL is returned.
1848  *
1849  *	The iter pctrie must be locked.
1850  */
1851 vm_page_t
1852 vm_page_iter_lookup(struct pctrie_iter *pages, vm_pindex_t pindex)
1853 {
1854 
1855 	return (vm_radix_iter_lookup(pages, pindex));
1856 }
1857 
1858 /*
1859  *	vm_page_lookup_unlocked:
1860  *
1861  *	Returns the page associated with the object/offset pair specified;
1862  *	if none is found, NULL is returned.  The page may be no longer be
1863  *	present in the object at the time that this function returns.  Only
1864  *	useful for opportunistic checks such as inmem().
1865  */
1866 vm_page_t
1867 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1868 {
1869 
1870 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1871 }
1872 
1873 /*
1874  *	vm_page_relookup:
1875  *
1876  *	Returns a page that must already have been busied by
1877  *	the caller.  Used for bogus page replacement.
1878  */
1879 vm_page_t
1880 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1881 {
1882 	vm_page_t m;
1883 
1884 	m = vm_page_lookup_unlocked(object, pindex);
1885 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1886 	    m->object == object && m->pindex == pindex,
1887 	    ("vm_page_relookup: Invalid page %p", m));
1888 	return (m);
1889 }
1890 
1891 /*
1892  * This should only be used by lockless functions for releasing transient
1893  * incorrect acquires.  The page may have been freed after we acquired a
1894  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1895  * further to do.
1896  */
1897 static void
1898 vm_page_busy_release(vm_page_t m)
1899 {
1900 	u_int x;
1901 
1902 	x = vm_page_busy_fetch(m);
1903 	for (;;) {
1904 		if (x == VPB_FREED)
1905 			break;
1906 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1907 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1908 			    x - VPB_ONE_SHARER))
1909 				break;
1910 			continue;
1911 		}
1912 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1913 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1914 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1915 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1916 			continue;
1917 		if ((x & VPB_BIT_WAITERS) != 0)
1918 			wakeup(m);
1919 		break;
1920 	}
1921 }
1922 
1923 /*
1924  *	vm_page_find_least:
1925  *
1926  *	Returns the page associated with the object with least pindex
1927  *	greater than or equal to the parameter pindex, or NULL.
1928  *
1929  *	The object must be locked.
1930  */
1931 vm_page_t
1932 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1933 {
1934 	vm_page_t m;
1935 
1936 	VM_OBJECT_ASSERT_LOCKED(object);
1937 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1938 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1939 	return (m);
1940 }
1941 
1942 /*
1943  *	vm_page_iter_lookup_ge:
1944  *
1945  *	Returns the page associated with the object with least pindex
1946  *	greater than or equal to the parameter pindex, or NULL.  Initializes the
1947  *	iterator to point to that page.
1948  *
1949  *	The iter pctrie must be locked.
1950  */
1951 vm_page_t
1952 vm_page_iter_lookup_ge(struct pctrie_iter *pages, vm_pindex_t pindex)
1953 {
1954 
1955 	return (vm_radix_iter_lookup_ge(pages, pindex));
1956 }
1957 
1958 /*
1959  * Returns the given page's successor (by pindex) within the object if it is
1960  * resident; if none is found, NULL is returned.
1961  *
1962  * The object must be locked.
1963  */
1964 vm_page_t
1965 vm_page_next(vm_page_t m)
1966 {
1967 	vm_page_t next;
1968 
1969 	VM_OBJECT_ASSERT_LOCKED(m->object);
1970 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1971 		MPASS(next->object == m->object);
1972 		if (next->pindex != m->pindex + 1)
1973 			next = NULL;
1974 	}
1975 	return (next);
1976 }
1977 
1978 /*
1979  * Returns the given page's predecessor (by pindex) within the object if it is
1980  * resident; if none is found, NULL is returned.
1981  *
1982  * The object must be locked.
1983  */
1984 vm_page_t
1985 vm_page_prev(vm_page_t m)
1986 {
1987 	vm_page_t prev;
1988 
1989 	VM_OBJECT_ASSERT_LOCKED(m->object);
1990 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1991 		MPASS(prev->object == m->object);
1992 		if (prev->pindex != m->pindex - 1)
1993 			prev = NULL;
1994 	}
1995 	return (prev);
1996 }
1997 
1998 /*
1999  * Uses the page mnew as a replacement for an existing page at index
2000  * pindex which must be already present in the object.
2001  *
2002  * Both pages must be exclusively busied on enter.  The old page is
2003  * unbusied on exit.
2004  *
2005  * A return value of true means mold is now free.  If this is not the
2006  * final ref and the caller does not hold a wire reference it may not
2007  * continue to access the page.
2008  */
2009 static bool
2010 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
2011     vm_page_t mold)
2012 {
2013 	vm_page_t mret __diagused;
2014 	bool dropped;
2015 
2016 	VM_OBJECT_ASSERT_WLOCKED(object);
2017 	vm_page_assert_xbusied(mold);
2018 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
2019 	    ("vm_page_replace: page %p already in object", mnew));
2020 
2021 	/*
2022 	 * This function mostly follows vm_page_insert() and
2023 	 * vm_page_remove() without the radix, object count and vnode
2024 	 * dance.  Double check such functions for more comments.
2025 	 */
2026 
2027 	mnew->object = object;
2028 	mnew->pindex = pindex;
2029 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
2030 	mret = vm_radix_replace(&object->rtree, mnew);
2031 	KASSERT(mret == mold,
2032 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
2033 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
2034 	    (mnew->oflags & VPO_UNMANAGED),
2035 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
2036 
2037 	/* Keep the resident page list in sorted order. */
2038 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
2039 	TAILQ_REMOVE(&object->memq, mold, listq);
2040 	mold->object = NULL;
2041 
2042 	/*
2043 	 * The object's resident_page_count does not change because we have
2044 	 * swapped one page for another, but the generation count should
2045 	 * change if the page is dirty.
2046 	 */
2047 	if (pmap_page_is_write_mapped(mnew))
2048 		vm_object_set_writeable_dirty(object);
2049 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
2050 	vm_page_xunbusy(mold);
2051 
2052 	return (dropped);
2053 }
2054 
2055 void
2056 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
2057     vm_page_t mold)
2058 {
2059 
2060 	vm_page_assert_xbusied(mnew);
2061 
2062 	if (vm_page_replace_hold(mnew, object, pindex, mold))
2063 		vm_page_free(mold);
2064 }
2065 
2066 /*
2067  *	vm_page_iter_rename:
2068  *
2069  *	Tries to move the specified page from its current object to a new object
2070  *	and pindex, using the given iterator to remove the page from its current
2071  *	object.  Returns true if the move was successful, and false if the move
2072  *	was aborted due to a failed memory allocation.
2073  *
2074  *	Panics if a page already resides in the new object at the new pindex.
2075  *
2076  *	Note: swap associated with the page must be invalidated by the move.  We
2077  *	      have to do this for several reasons:  (1) we aren't freeing the
2078  *	      page, (2) we are dirtying the page, (3) the VM system is probably
2079  *	      moving the page from object A to B, and will then later move
2080  *	      the backing store from A to B and we can't have a conflict.
2081  *
2082  *	Note: we *always* dirty the page.  It is necessary both for the
2083  *	      fact that we moved it, and because we may be invalidating
2084  *	      swap.
2085  *
2086  *	The objects must be locked.
2087  */
2088 bool
2089 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m,
2090     vm_object_t new_object, vm_pindex_t new_pindex)
2091 {
2092 	vm_page_t mpred;
2093 	vm_pindex_t opidx;
2094 
2095 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
2096 	    ("%s: page %p is missing object ref", __func__, m));
2097 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2098 	VM_OBJECT_ASSERT_WLOCKED(new_object);
2099 
2100 	/*
2101 	 * Create a custom version of vm_page_insert() which does not depend
2102 	 * by m_prev and can cheat on the implementation aspects of the
2103 	 * function.
2104 	 */
2105 	opidx = m->pindex;
2106 	m->pindex = new_pindex;
2107 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
2108 		m->pindex = opidx;
2109 		return (false);
2110 	}
2111 
2112 	/*
2113 	 * The operation cannot fail anymore.  The removal must happen before
2114 	 * the listq iterator is tainted.
2115 	 */
2116 	m->pindex = opidx;
2117 	vm_radix_iter_remove(old_pages);
2118 	vm_page_remove_radixdone(m);
2119 
2120 	/* Return back to the new pindex to complete vm_page_insert(). */
2121 	m->pindex = new_pindex;
2122 	m->object = new_object;
2123 
2124 	vm_page_insert_radixdone(m, new_object, mpred);
2125 	vm_page_dirty(m);
2126 	vm_pager_page_inserted(new_object, m);
2127 	return (true);
2128 }
2129 
2130 /*
2131  *	vm_page_mpred:
2132  *
2133  *	Return the greatest page of the object with index <= pindex,
2134  *	or NULL, if there is none.  Assumes object lock is held.
2135  */
2136 vm_page_t
2137 vm_page_mpred(vm_object_t object, vm_pindex_t pindex)
2138 {
2139 	return (vm_radix_lookup_le(&object->rtree, pindex));
2140 }
2141 
2142 /*
2143  *	vm_page_alloc:
2144  *
2145  *	Allocate and return a page that is associated with the specified
2146  *	object and offset pair.  By default, this page is exclusive busied.
2147  *
2148  *	The caller must always specify an allocation class.
2149  *
2150  *	allocation classes:
2151  *	VM_ALLOC_NORMAL		normal process request
2152  *	VM_ALLOC_SYSTEM		system *really* needs a page
2153  *	VM_ALLOC_INTERRUPT	interrupt time request
2154  *
2155  *	optional allocation flags:
2156  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2157  *				intends to allocate
2158  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2159  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2160  *	VM_ALLOC_SBUSY		shared busy the allocated page
2161  *	VM_ALLOC_WIRED		wire the allocated page
2162  *	VM_ALLOC_ZERO		prefer a zeroed page
2163  */
2164 vm_page_t
2165 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
2166 {
2167 
2168 	return (vm_page_alloc_after(object, pindex, req,
2169 	    vm_page_mpred(object, pindex)));
2170 }
2171 
2172 /*
2173  * Allocate a page in the specified object with the given page index.  To
2174  * optimize insertion of the page into the object, the caller must also specify
2175  * the resident page in the object with largest index smaller than the given
2176  * page index, or NULL if no such page exists.
2177  */
2178 static vm_page_t
2179 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
2180     int req, vm_page_t mpred)
2181 {
2182 	struct vm_domainset_iter di;
2183 	vm_page_t m;
2184 	int domain;
2185 
2186 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2187 	do {
2188 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2189 		    mpred);
2190 		if (m != NULL)
2191 			break;
2192 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2193 
2194 	return (m);
2195 }
2196 
2197 /*
2198  * Returns true if the number of free pages exceeds the minimum
2199  * for the request class and false otherwise.
2200  */
2201 static int
2202 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2203 {
2204 	u_int limit, old, new;
2205 
2206 	if (req_class == VM_ALLOC_INTERRUPT)
2207 		limit = 0;
2208 	else if (req_class == VM_ALLOC_SYSTEM)
2209 		limit = vmd->vmd_interrupt_free_min;
2210 	else
2211 		limit = vmd->vmd_free_reserved;
2212 
2213 	/*
2214 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2215 	 */
2216 	limit += npages;
2217 	old = atomic_load_int(&vmd->vmd_free_count);
2218 	do {
2219 		if (old < limit)
2220 			return (0);
2221 		new = old - npages;
2222 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2223 
2224 	/* Wake the page daemon if we've crossed the threshold. */
2225 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2226 		pagedaemon_wakeup(vmd->vmd_domain);
2227 
2228 	/* Only update bitsets on transitions. */
2229 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2230 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2231 		vm_domain_set(vmd);
2232 
2233 	return (1);
2234 }
2235 
2236 int
2237 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2238 {
2239 	int req_class;
2240 
2241 	/*
2242 	 * The page daemon is allowed to dig deeper into the free page list.
2243 	 */
2244 	req_class = req & VM_ALLOC_CLASS_MASK;
2245 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2246 		req_class = VM_ALLOC_SYSTEM;
2247 	return (_vm_domain_allocate(vmd, req_class, npages));
2248 }
2249 
2250 vm_page_t
2251 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2252     int req, vm_page_t mpred)
2253 {
2254 	struct vm_domain *vmd;
2255 	vm_page_t m;
2256 	int flags;
2257 
2258 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2259 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2260 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2261 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2262 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2263 	KASSERT((req & ~VPA_FLAGS) == 0,
2264 	    ("invalid request %#x", req));
2265 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2266 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2267 	    ("invalid request %#x", req));
2268 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2269 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2270 	    (uintmax_t)pindex));
2271 	VM_OBJECT_ASSERT_WLOCKED(object);
2272 
2273 	flags = 0;
2274 	m = NULL;
2275 	if (!vm_pager_can_alloc_page(object, pindex))
2276 		return (NULL);
2277 again:
2278 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2279 		m = vm_page_alloc_nofree_domain(domain, req);
2280 		if (m != NULL)
2281 			goto found;
2282 	}
2283 #if VM_NRESERVLEVEL > 0
2284 	/*
2285 	 * Can we allocate the page from a reservation?
2286 	 */
2287 	if (vm_object_reserv(object) &&
2288 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2289 	    NULL) {
2290 		goto found;
2291 	}
2292 #endif
2293 	vmd = VM_DOMAIN(domain);
2294 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2295 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2296 		    M_NOWAIT | M_NOVM);
2297 		if (m != NULL) {
2298 			flags |= PG_PCPU_CACHE;
2299 			goto found;
2300 		}
2301 	}
2302 	if (vm_domain_allocate(vmd, req, 1)) {
2303 		/*
2304 		 * If not, allocate it from the free page queues.
2305 		 */
2306 		vm_domain_free_lock(vmd);
2307 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2308 		vm_domain_free_unlock(vmd);
2309 		if (m == NULL) {
2310 			vm_domain_freecnt_inc(vmd, 1);
2311 #if VM_NRESERVLEVEL > 0
2312 			if (vm_reserv_reclaim_inactive(domain))
2313 				goto again;
2314 #endif
2315 		}
2316 	}
2317 	if (m == NULL) {
2318 		/*
2319 		 * Not allocatable, give up.
2320 		 */
2321 		if (vm_domain_alloc_fail(vmd, object, req))
2322 			goto again;
2323 		return (NULL);
2324 	}
2325 
2326 	/*
2327 	 * At this point we had better have found a good page.
2328 	 */
2329 found:
2330 	vm_page_dequeue(m);
2331 	vm_page_alloc_check(m);
2332 
2333 	/*
2334 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2335 	 */
2336 	flags |= m->flags & PG_ZERO;
2337 	if ((req & VM_ALLOC_NODUMP) != 0)
2338 		flags |= PG_NODUMP;
2339 	if ((req & VM_ALLOC_NOFREE) != 0)
2340 		flags |= PG_NOFREE;
2341 	m->flags = flags;
2342 	m->a.flags = 0;
2343 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2344 	m->pool = VM_FREEPOOL_DEFAULT;
2345 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2346 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2347 	else if ((req & VM_ALLOC_SBUSY) != 0)
2348 		m->busy_lock = VPB_SHARERS_WORD(1);
2349 	else
2350 		m->busy_lock = VPB_UNBUSIED;
2351 	if (req & VM_ALLOC_WIRED) {
2352 		vm_wire_add(1);
2353 		m->ref_count = 1;
2354 	}
2355 	m->a.act_count = 0;
2356 
2357 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2358 		if (req & VM_ALLOC_WIRED) {
2359 			vm_wire_sub(1);
2360 			m->ref_count = 0;
2361 		}
2362 		KASSERT(m->object == NULL, ("page %p has object", m));
2363 		m->oflags = VPO_UNMANAGED;
2364 		m->busy_lock = VPB_UNBUSIED;
2365 		/* Don't change PG_ZERO. */
2366 		vm_page_free_toq(m);
2367 		if (req & VM_ALLOC_WAITFAIL) {
2368 			VM_OBJECT_WUNLOCK(object);
2369 			vm_radix_wait();
2370 			VM_OBJECT_WLOCK(object);
2371 		}
2372 		return (NULL);
2373 	}
2374 
2375 	/* Ignore device objects; the pager sets "memattr" for them. */
2376 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2377 	    (object->flags & OBJ_FICTITIOUS) == 0)
2378 		pmap_page_set_memattr(m, object->memattr);
2379 
2380 	return (m);
2381 }
2382 
2383 /*
2384  *	vm_page_alloc_contig:
2385  *
2386  *	Allocate a contiguous set of physical pages of the given size "npages"
2387  *	from the free lists.  All of the physical pages must be at or above
2388  *	the given physical address "low" and below the given physical address
2389  *	"high".  The given value "alignment" determines the alignment of the
2390  *	first physical page in the set.  If the given value "boundary" is
2391  *	non-zero, then the set of physical pages cannot cross any physical
2392  *	address boundary that is a multiple of that value.  Both "alignment"
2393  *	and "boundary" must be a power of two.
2394  *
2395  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2396  *	then the memory attribute setting for the physical pages is configured
2397  *	to the object's memory attribute setting.  Otherwise, the memory
2398  *	attribute setting for the physical pages is configured to "memattr",
2399  *	overriding the object's memory attribute setting.  However, if the
2400  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2401  *	memory attribute setting for the physical pages cannot be configured
2402  *	to VM_MEMATTR_DEFAULT.
2403  *
2404  *	The specified object may not contain fictitious pages.
2405  *
2406  *	The caller must always specify an allocation class.
2407  *
2408  *	allocation classes:
2409  *	VM_ALLOC_NORMAL		normal process request
2410  *	VM_ALLOC_SYSTEM		system *really* needs a page
2411  *	VM_ALLOC_INTERRUPT	interrupt time request
2412  *
2413  *	optional allocation flags:
2414  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2415  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2416  *	VM_ALLOC_SBUSY		shared busy the allocated page
2417  *	VM_ALLOC_WIRED		wire the allocated page
2418  *	VM_ALLOC_ZERO		prefer a zeroed page
2419  */
2420 vm_page_t
2421 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2422     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2423     vm_paddr_t boundary, vm_memattr_t memattr)
2424 {
2425 	struct vm_domainset_iter di;
2426 	vm_page_t bounds[2];
2427 	vm_page_t m;
2428 	int domain;
2429 	int start_segind;
2430 
2431 	start_segind = -1;
2432 
2433 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2434 	do {
2435 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2436 		    npages, low, high, alignment, boundary, memattr);
2437 		if (m != NULL)
2438 			break;
2439 		if (start_segind == -1)
2440 			start_segind = vm_phys_lookup_segind(low);
2441 		if (vm_phys_find_range(bounds, start_segind, domain,
2442 		    npages, low, high) == -1) {
2443 			vm_domainset_iter_ignore(&di, domain);
2444 		}
2445 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2446 
2447 	return (m);
2448 }
2449 
2450 static vm_page_t
2451 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2452     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2453 {
2454 	struct vm_domain *vmd;
2455 	vm_page_t m_ret;
2456 
2457 	/*
2458 	 * Can we allocate the pages without the number of free pages falling
2459 	 * below the lower bound for the allocation class?
2460 	 */
2461 	vmd = VM_DOMAIN(domain);
2462 	if (!vm_domain_allocate(vmd, req, npages))
2463 		return (NULL);
2464 	/*
2465 	 * Try to allocate the pages from the free page queues.
2466 	 */
2467 	vm_domain_free_lock(vmd);
2468 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2469 	    alignment, boundary);
2470 	vm_domain_free_unlock(vmd);
2471 	if (m_ret != NULL)
2472 		return (m_ret);
2473 #if VM_NRESERVLEVEL > 0
2474 	/*
2475 	 * Try to break a reservation to allocate the pages.
2476 	 */
2477 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2478 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2479 	            high, alignment, boundary);
2480 		if (m_ret != NULL)
2481 			return (m_ret);
2482 	}
2483 #endif
2484 	vm_domain_freecnt_inc(vmd, npages);
2485 	return (NULL);
2486 }
2487 
2488 vm_page_t
2489 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2490     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2491     vm_paddr_t boundary, vm_memattr_t memattr)
2492 {
2493 	struct pctrie_iter pages;
2494 	vm_page_t m, m_ret, mpred;
2495 	u_int busy_lock, flags, oflags;
2496 
2497 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2498 	KASSERT((req & ~VPAC_FLAGS) == 0,
2499 	    ("invalid request %#x", req));
2500 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2501 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2502 	    ("invalid request %#x", req));
2503 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2504 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2505 	    ("invalid request %#x", req));
2506 	VM_OBJECT_ASSERT_WLOCKED(object);
2507 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2508 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2509 	    object));
2510 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2511 
2512 	vm_page_iter_init(&pages, object);
2513 	mpred = vm_radix_iter_lookup_le(&pages, pindex);
2514 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2515 	    ("vm_page_alloc_contig: pindex already allocated"));
2516 	for (;;) {
2517 #if VM_NRESERVLEVEL > 0
2518 		/*
2519 		 * Can we allocate the pages from a reservation?
2520 		 */
2521 		if (vm_object_reserv(object) &&
2522 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2523 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2524 			break;
2525 		}
2526 #endif
2527 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2528 		    low, high, alignment, boundary)) != NULL)
2529 			break;
2530 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2531 			return (NULL);
2532 	}
2533 
2534 	/*
2535 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2536 	 */
2537 	flags = PG_ZERO;
2538 	if ((req & VM_ALLOC_NODUMP) != 0)
2539 		flags |= PG_NODUMP;
2540 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2541 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2542 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2543 	else if ((req & VM_ALLOC_SBUSY) != 0)
2544 		busy_lock = VPB_SHARERS_WORD(1);
2545 	else
2546 		busy_lock = VPB_UNBUSIED;
2547 	if ((req & VM_ALLOC_WIRED) != 0)
2548 		vm_wire_add(npages);
2549 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2550 	    memattr == VM_MEMATTR_DEFAULT)
2551 		memattr = object->memattr;
2552 	for (m = m_ret; m < &m_ret[npages]; m++) {
2553 		vm_page_dequeue(m);
2554 		vm_page_alloc_check(m);
2555 		m->a.flags = 0;
2556 		m->flags = (m->flags | PG_NODUMP) & flags;
2557 		m->busy_lock = busy_lock;
2558 		if ((req & VM_ALLOC_WIRED) != 0)
2559 			m->ref_count = 1;
2560 		m->a.act_count = 0;
2561 		m->oflags = oflags;
2562 		m->pool = VM_FREEPOOL_DEFAULT;
2563 		if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) {
2564 			if ((req & VM_ALLOC_WIRED) != 0)
2565 				vm_wire_sub(npages);
2566 			KASSERT(m->object == NULL,
2567 			    ("page %p has object", m));
2568 			mpred = m;
2569 			for (m = m_ret; m < &m_ret[npages]; m++) {
2570 				if (m <= mpred &&
2571 				    (req & VM_ALLOC_WIRED) != 0)
2572 					m->ref_count = 0;
2573 				m->oflags = VPO_UNMANAGED;
2574 				m->busy_lock = VPB_UNBUSIED;
2575 				/* Don't change PG_ZERO. */
2576 				vm_page_free_toq(m);
2577 			}
2578 			if (req & VM_ALLOC_WAITFAIL) {
2579 				VM_OBJECT_WUNLOCK(object);
2580 				vm_radix_wait();
2581 				VM_OBJECT_WLOCK(object);
2582 			}
2583 			return (NULL);
2584 		}
2585 		mpred = m;
2586 		if (memattr != VM_MEMATTR_DEFAULT)
2587 			pmap_page_set_memattr(m, memattr);
2588 		pindex++;
2589 	}
2590 	return (m_ret);
2591 }
2592 
2593 /*
2594  * Allocate a physical page that is not intended to be inserted into a VM
2595  * object.
2596  */
2597 vm_page_t
2598 vm_page_alloc_noobj_domain(int domain, int req)
2599 {
2600 	struct vm_domain *vmd;
2601 	vm_page_t m;
2602 	int flags;
2603 
2604 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2605 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2606 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2607 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2608 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2609 	KASSERT((req & ~VPAN_FLAGS) == 0,
2610 	    ("invalid request %#x", req));
2611 
2612 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2613 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2614 	vmd = VM_DOMAIN(domain);
2615 again:
2616 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2617 		m = vm_page_alloc_nofree_domain(domain, req);
2618 		if (m != NULL)
2619 			goto found;
2620 	}
2621 
2622 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2623 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2624 		    M_NOWAIT | M_NOVM);
2625 		if (m != NULL) {
2626 			flags |= PG_PCPU_CACHE;
2627 			goto found;
2628 		}
2629 	}
2630 
2631 	if (vm_domain_allocate(vmd, req, 1)) {
2632 		vm_domain_free_lock(vmd);
2633 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2634 		vm_domain_free_unlock(vmd);
2635 		if (m == NULL) {
2636 			vm_domain_freecnt_inc(vmd, 1);
2637 #if VM_NRESERVLEVEL > 0
2638 			if (vm_reserv_reclaim_inactive(domain))
2639 				goto again;
2640 #endif
2641 		}
2642 	}
2643 	if (m == NULL) {
2644 		if (vm_domain_alloc_fail(vmd, NULL, req))
2645 			goto again;
2646 		return (NULL);
2647 	}
2648 
2649 found:
2650 	vm_page_dequeue(m);
2651 	vm_page_alloc_check(m);
2652 
2653 	/*
2654 	 * Consumers should not rely on a useful default pindex value.
2655 	 */
2656 	m->pindex = 0xdeadc0dedeadc0de;
2657 	m->flags = (m->flags & PG_ZERO) | flags;
2658 	m->a.flags = 0;
2659 	m->oflags = VPO_UNMANAGED;
2660 	m->pool = VM_FREEPOOL_DIRECT;
2661 	m->busy_lock = VPB_UNBUSIED;
2662 	if ((req & VM_ALLOC_WIRED) != 0) {
2663 		vm_wire_add(1);
2664 		m->ref_count = 1;
2665 	}
2666 
2667 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2668 		pmap_zero_page(m);
2669 
2670 	return (m);
2671 }
2672 
2673 #if VM_NRESERVLEVEL > 1
2674 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2675 #elif VM_NRESERVLEVEL > 0
2676 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2677 #else
2678 #define	VM_NOFREE_IMPORT_ORDER	8
2679 #endif
2680 
2681 /*
2682  * Allocate a single NOFREE page.
2683  *
2684  * This routine hands out NOFREE pages from higher-order
2685  * physical memory blocks in order to reduce memory fragmentation.
2686  * When a NOFREE for a given domain chunk is used up,
2687  * the routine will try to fetch a new one from the freelists
2688  * and discard the old one.
2689  */
2690 static vm_page_t
2691 vm_page_alloc_nofree_domain(int domain, int req)
2692 {
2693 	vm_page_t m;
2694 	struct vm_domain *vmd;
2695 	struct vm_nofreeq *nqp;
2696 
2697 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2698 
2699 	vmd = VM_DOMAIN(domain);
2700 	nqp = &vmd->vmd_nofreeq;
2701 	vm_domain_free_lock(vmd);
2702 	if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) {
2703 		if (!vm_domain_allocate(vmd, req,
2704 		    1 << VM_NOFREE_IMPORT_ORDER)) {
2705 			vm_domain_free_unlock(vmd);
2706 			return (NULL);
2707 		}
2708 		nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2709 		    VM_NOFREE_IMPORT_ORDER);
2710 		if (nqp->ma == NULL) {
2711 			vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER);
2712 			vm_domain_free_unlock(vmd);
2713 			return (NULL);
2714 		}
2715 		nqp->offs = 0;
2716 	}
2717 	m = &nqp->ma[nqp->offs++];
2718 	vm_domain_free_unlock(vmd);
2719 	VM_CNT_ADD(v_nofree_count, 1);
2720 
2721 	return (m);
2722 }
2723 
2724 vm_page_t
2725 vm_page_alloc_noobj(int req)
2726 {
2727 	struct vm_domainset_iter di;
2728 	vm_page_t m;
2729 	int domain;
2730 
2731 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2732 	do {
2733 		m = vm_page_alloc_noobj_domain(domain, req);
2734 		if (m != NULL)
2735 			break;
2736 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2737 
2738 	return (m);
2739 }
2740 
2741 vm_page_t
2742 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2743     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2744     vm_memattr_t memattr)
2745 {
2746 	struct vm_domainset_iter di;
2747 	vm_page_t m;
2748 	int domain;
2749 
2750 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2751 	do {
2752 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2753 		    high, alignment, boundary, memattr);
2754 		if (m != NULL)
2755 			break;
2756 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2757 
2758 	return (m);
2759 }
2760 
2761 vm_page_t
2762 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2763     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2764     vm_memattr_t memattr)
2765 {
2766 	vm_page_t m, m_ret;
2767 	u_int flags;
2768 
2769 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2770 	KASSERT((req & ~VPANC_FLAGS) == 0,
2771 	    ("invalid request %#x", req));
2772 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2773 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2774 	    ("invalid request %#x", req));
2775 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2776 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2777 	    ("invalid request %#x", req));
2778 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2779 
2780 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2781 	    low, high, alignment, boundary)) == NULL) {
2782 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2783 			return (NULL);
2784 	}
2785 
2786 	/*
2787 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2788 	 */
2789 	flags = PG_ZERO;
2790 	if ((req & VM_ALLOC_NODUMP) != 0)
2791 		flags |= PG_NODUMP;
2792 	if ((req & VM_ALLOC_WIRED) != 0)
2793 		vm_wire_add(npages);
2794 	for (m = m_ret; m < &m_ret[npages]; m++) {
2795 		vm_page_dequeue(m);
2796 		vm_page_alloc_check(m);
2797 
2798 		/*
2799 		 * Consumers should not rely on a useful default pindex value.
2800 		 */
2801 		m->pindex = 0xdeadc0dedeadc0de;
2802 		m->a.flags = 0;
2803 		m->flags = (m->flags | PG_NODUMP) & flags;
2804 		m->busy_lock = VPB_UNBUSIED;
2805 		if ((req & VM_ALLOC_WIRED) != 0)
2806 			m->ref_count = 1;
2807 		m->a.act_count = 0;
2808 		m->oflags = VPO_UNMANAGED;
2809 		m->pool = VM_FREEPOOL_DIRECT;
2810 
2811 		/*
2812 		 * Zero the page before updating any mappings since the page is
2813 		 * not yet shared with any devices which might require the
2814 		 * non-default memory attribute.  pmap_page_set_memattr()
2815 		 * flushes data caches before returning.
2816 		 */
2817 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2818 			pmap_zero_page(m);
2819 		if (memattr != VM_MEMATTR_DEFAULT)
2820 			pmap_page_set_memattr(m, memattr);
2821 	}
2822 	return (m_ret);
2823 }
2824 
2825 /*
2826  * Check a page that has been freshly dequeued from a freelist.
2827  */
2828 static void
2829 vm_page_alloc_check(vm_page_t m)
2830 {
2831 
2832 	KASSERT(m->object == NULL, ("page %p has object", m));
2833 	KASSERT(m->a.queue == PQ_NONE &&
2834 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2835 	    ("page %p has unexpected queue %d, flags %#x",
2836 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2837 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2838 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2839 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2840 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2841 	    ("page %p has unexpected memattr %d",
2842 	    m, pmap_page_get_memattr(m)));
2843 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2844 	pmap_vm_page_alloc_check(m);
2845 }
2846 
2847 static int
2848 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2849 {
2850 	struct vm_domain *vmd;
2851 	struct vm_pgcache *pgcache;
2852 	int i;
2853 
2854 	pgcache = arg;
2855 	vmd = VM_DOMAIN(pgcache->domain);
2856 
2857 	/*
2858 	 * The page daemon should avoid creating extra memory pressure since its
2859 	 * main purpose is to replenish the store of free pages.
2860 	 */
2861 	if (vmd->vmd_severeset || curproc == pageproc ||
2862 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2863 		return (0);
2864 	domain = vmd->vmd_domain;
2865 	vm_domain_free_lock(vmd);
2866 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2867 	    (vm_page_t *)store);
2868 	vm_domain_free_unlock(vmd);
2869 	if (cnt != i)
2870 		vm_domain_freecnt_inc(vmd, cnt - i);
2871 
2872 	return (i);
2873 }
2874 
2875 static void
2876 vm_page_zone_release(void *arg, void **store, int cnt)
2877 {
2878 	struct vm_domain *vmd;
2879 	struct vm_pgcache *pgcache;
2880 	vm_page_t m;
2881 	int i;
2882 
2883 	pgcache = arg;
2884 	vmd = VM_DOMAIN(pgcache->domain);
2885 	vm_domain_free_lock(vmd);
2886 	for (i = 0; i < cnt; i++) {
2887 		m = (vm_page_t)store[i];
2888 		vm_phys_free_pages(m, pgcache->pool, 0);
2889 	}
2890 	vm_domain_free_unlock(vmd);
2891 	vm_domain_freecnt_inc(vmd, cnt);
2892 }
2893 
2894 #define	VPSC_ANY	0	/* No restrictions. */
2895 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2896 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2897 
2898 /*
2899  *	vm_page_scan_contig:
2900  *
2901  *	Scan vm_page_array[] between the specified entries "m_start" and
2902  *	"m_end" for a run of contiguous physical pages that satisfy the
2903  *	specified conditions, and return the lowest page in the run.  The
2904  *	specified "alignment" determines the alignment of the lowest physical
2905  *	page in the run.  If the specified "boundary" is non-zero, then the
2906  *	run of physical pages cannot span a physical address that is a
2907  *	multiple of "boundary".
2908  *
2909  *	"m_end" is never dereferenced, so it need not point to a vm_page
2910  *	structure within vm_page_array[].
2911  *
2912  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2913  *	span a hole (or discontiguity) in the physical address space.  Both
2914  *	"alignment" and "boundary" must be a power of two.
2915  */
2916 static vm_page_t
2917 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2918     u_long alignment, vm_paddr_t boundary, int options)
2919 {
2920 	vm_object_t object;
2921 	vm_paddr_t pa;
2922 	vm_page_t m, m_run;
2923 #if VM_NRESERVLEVEL > 0
2924 	int level;
2925 #endif
2926 	int m_inc, order, run_ext, run_len;
2927 
2928 	KASSERT(npages > 0, ("npages is 0"));
2929 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2930 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2931 	m_run = NULL;
2932 	run_len = 0;
2933 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2934 		KASSERT((m->flags & PG_MARKER) == 0,
2935 		    ("page %p is PG_MARKER", m));
2936 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2937 		    ("fictitious page %p has invalid ref count", m));
2938 
2939 		/*
2940 		 * If the current page would be the start of a run, check its
2941 		 * physical address against the end, alignment, and boundary
2942 		 * conditions.  If it doesn't satisfy these conditions, either
2943 		 * terminate the scan or advance to the next page that
2944 		 * satisfies the failed condition.
2945 		 */
2946 		if (run_len == 0) {
2947 			KASSERT(m_run == NULL, ("m_run != NULL"));
2948 			if (m + npages > m_end)
2949 				break;
2950 			pa = VM_PAGE_TO_PHYS(m);
2951 			if (!vm_addr_align_ok(pa, alignment)) {
2952 				m_inc = atop(roundup2(pa, alignment) - pa);
2953 				continue;
2954 			}
2955 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2956 				m_inc = atop(roundup2(pa, boundary) - pa);
2957 				continue;
2958 			}
2959 		} else
2960 			KASSERT(m_run != NULL, ("m_run == NULL"));
2961 
2962 retry:
2963 		m_inc = 1;
2964 		if (vm_page_wired(m))
2965 			run_ext = 0;
2966 #if VM_NRESERVLEVEL > 0
2967 		else if ((level = vm_reserv_level(m)) >= 0 &&
2968 		    (options & VPSC_NORESERV) != 0) {
2969 			run_ext = 0;
2970 			/* Advance to the end of the reservation. */
2971 			pa = VM_PAGE_TO_PHYS(m);
2972 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2973 			    pa);
2974 		}
2975 #endif
2976 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2977 			/*
2978 			 * The page is considered eligible for relocation if
2979 			 * and only if it could be laundered or reclaimed by
2980 			 * the page daemon.
2981 			 */
2982 			VM_OBJECT_RLOCK(object);
2983 			if (object != m->object) {
2984 				VM_OBJECT_RUNLOCK(object);
2985 				goto retry;
2986 			}
2987 			/* Don't care: PG_NODUMP, PG_ZERO. */
2988 			if ((object->flags & OBJ_SWAP) == 0 &&
2989 			    object->type != OBJT_VNODE) {
2990 				run_ext = 0;
2991 #if VM_NRESERVLEVEL > 0
2992 			} else if ((options & VPSC_NOSUPER) != 0 &&
2993 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2994 				run_ext = 0;
2995 				/* Advance to the end of the superpage. */
2996 				pa = VM_PAGE_TO_PHYS(m);
2997 				m_inc = atop(roundup2(pa + 1,
2998 				    vm_reserv_size(level)) - pa);
2999 #endif
3000 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
3001 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
3002 				/*
3003 				 * The page is allocated but eligible for
3004 				 * relocation.  Extend the current run by one
3005 				 * page.
3006 				 */
3007 				KASSERT(pmap_page_get_memattr(m) ==
3008 				    VM_MEMATTR_DEFAULT,
3009 				    ("page %p has an unexpected memattr", m));
3010 				KASSERT((m->oflags & (VPO_SWAPINPROG |
3011 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
3012 				    ("page %p has unexpected oflags", m));
3013 				/* Don't care: PGA_NOSYNC. */
3014 				run_ext = 1;
3015 			} else
3016 				run_ext = 0;
3017 			VM_OBJECT_RUNLOCK(object);
3018 #if VM_NRESERVLEVEL > 0
3019 		} else if (level >= 0) {
3020 			/*
3021 			 * The page is reserved but not yet allocated.  In
3022 			 * other words, it is still free.  Extend the current
3023 			 * run by one page.
3024 			 */
3025 			run_ext = 1;
3026 #endif
3027 		} else if ((order = m->order) < VM_NFREEORDER) {
3028 			/*
3029 			 * The page is enqueued in the physical memory
3030 			 * allocator's free page queues.  Moreover, it is the
3031 			 * first page in a power-of-two-sized run of
3032 			 * contiguous free pages.  Add these pages to the end
3033 			 * of the current run, and jump ahead.
3034 			 */
3035 			run_ext = 1 << order;
3036 			m_inc = 1 << order;
3037 		} else {
3038 			/*
3039 			 * Skip the page for one of the following reasons: (1)
3040 			 * It is enqueued in the physical memory allocator's
3041 			 * free page queues.  However, it is not the first
3042 			 * page in a run of contiguous free pages.  (This case
3043 			 * rarely occurs because the scan is performed in
3044 			 * ascending order.) (2) It is not reserved, and it is
3045 			 * transitioning from free to allocated.  (Conversely,
3046 			 * the transition from allocated to free for managed
3047 			 * pages is blocked by the page busy lock.) (3) It is
3048 			 * allocated but not contained by an object and not
3049 			 * wired, e.g., allocated by Xen's balloon driver.
3050 			 */
3051 			run_ext = 0;
3052 		}
3053 
3054 		/*
3055 		 * Extend or reset the current run of pages.
3056 		 */
3057 		if (run_ext > 0) {
3058 			if (run_len == 0)
3059 				m_run = m;
3060 			run_len += run_ext;
3061 		} else {
3062 			if (run_len > 0) {
3063 				m_run = NULL;
3064 				run_len = 0;
3065 			}
3066 		}
3067 	}
3068 	if (run_len >= npages)
3069 		return (m_run);
3070 	return (NULL);
3071 }
3072 
3073 /*
3074  *	vm_page_reclaim_run:
3075  *
3076  *	Try to relocate each of the allocated virtual pages within the
3077  *	specified run of physical pages to a new physical address.  Free the
3078  *	physical pages underlying the relocated virtual pages.  A virtual page
3079  *	is relocatable if and only if it could be laundered or reclaimed by
3080  *	the page daemon.  Whenever possible, a virtual page is relocated to a
3081  *	physical address above "high".
3082  *
3083  *	Returns 0 if every physical page within the run was already free or
3084  *	just freed by a successful relocation.  Otherwise, returns a non-zero
3085  *	value indicating why the last attempt to relocate a virtual page was
3086  *	unsuccessful.
3087  *
3088  *	"req_class" must be an allocation class.
3089  */
3090 static int
3091 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
3092     vm_paddr_t high)
3093 {
3094 	struct vm_domain *vmd;
3095 	struct spglist free;
3096 	vm_object_t object;
3097 	vm_paddr_t pa;
3098 	vm_page_t m, m_end, m_new;
3099 	int error, order, req;
3100 
3101 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
3102 	    ("req_class is not an allocation class"));
3103 	SLIST_INIT(&free);
3104 	error = 0;
3105 	m = m_run;
3106 	m_end = m_run + npages;
3107 	for (; error == 0 && m < m_end; m++) {
3108 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
3109 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
3110 
3111 		/*
3112 		 * Racily check for wirings.  Races are handled once the object
3113 		 * lock is held and the page is unmapped.
3114 		 */
3115 		if (vm_page_wired(m))
3116 			error = EBUSY;
3117 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
3118 			/*
3119 			 * The page is relocated if and only if it could be
3120 			 * laundered or reclaimed by the page daemon.
3121 			 */
3122 			VM_OBJECT_WLOCK(object);
3123 			/* Don't care: PG_NODUMP, PG_ZERO. */
3124 			if (m->object != object ||
3125 			    ((object->flags & OBJ_SWAP) == 0 &&
3126 			    object->type != OBJT_VNODE))
3127 				error = EINVAL;
3128 			else if (object->memattr != VM_MEMATTR_DEFAULT)
3129 				error = EINVAL;
3130 			else if (vm_page_queue(m) != PQ_NONE &&
3131 			    vm_page_tryxbusy(m) != 0) {
3132 				if (vm_page_wired(m)) {
3133 					vm_page_xunbusy(m);
3134 					error = EBUSY;
3135 					goto unlock;
3136 				}
3137 				KASSERT(pmap_page_get_memattr(m) ==
3138 				    VM_MEMATTR_DEFAULT,
3139 				    ("page %p has an unexpected memattr", m));
3140 				KASSERT(m->oflags == 0,
3141 				    ("page %p has unexpected oflags", m));
3142 				/* Don't care: PGA_NOSYNC. */
3143 				if (!vm_page_none_valid(m)) {
3144 					/*
3145 					 * First, try to allocate a new page
3146 					 * that is above "high".  Failing
3147 					 * that, try to allocate a new page
3148 					 * that is below "m_run".  Allocate
3149 					 * the new page between the end of
3150 					 * "m_run" and "high" only as a last
3151 					 * resort.
3152 					 */
3153 					req = req_class;
3154 					if ((m->flags & PG_NODUMP) != 0)
3155 						req |= VM_ALLOC_NODUMP;
3156 					if (trunc_page(high) !=
3157 					    ~(vm_paddr_t)PAGE_MASK) {
3158 						m_new =
3159 						    vm_page_alloc_noobj_contig(
3160 						    req, 1, round_page(high),
3161 						    ~(vm_paddr_t)0, PAGE_SIZE,
3162 						    0, VM_MEMATTR_DEFAULT);
3163 					} else
3164 						m_new = NULL;
3165 					if (m_new == NULL) {
3166 						pa = VM_PAGE_TO_PHYS(m_run);
3167 						m_new =
3168 						    vm_page_alloc_noobj_contig(
3169 						    req, 1, 0, pa - 1,
3170 						    PAGE_SIZE, 0,
3171 						    VM_MEMATTR_DEFAULT);
3172 					}
3173 					if (m_new == NULL) {
3174 						pa += ptoa(npages);
3175 						m_new =
3176 						    vm_page_alloc_noobj_contig(
3177 						    req, 1, pa, high, PAGE_SIZE,
3178 						    0, VM_MEMATTR_DEFAULT);
3179 					}
3180 					if (m_new == NULL) {
3181 						vm_page_xunbusy(m);
3182 						error = ENOMEM;
3183 						goto unlock;
3184 					}
3185 
3186 					/*
3187 					 * Unmap the page and check for new
3188 					 * wirings that may have been acquired
3189 					 * through a pmap lookup.
3190 					 */
3191 					if (object->ref_count != 0 &&
3192 					    !vm_page_try_remove_all(m)) {
3193 						vm_page_xunbusy(m);
3194 						vm_page_free(m_new);
3195 						error = EBUSY;
3196 						goto unlock;
3197 					}
3198 
3199 					/*
3200 					 * Replace "m" with the new page.  For
3201 					 * vm_page_replace(), "m" must be busy
3202 					 * and dequeued.  Finally, change "m"
3203 					 * as if vm_page_free() was called.
3204 					 */
3205 					m_new->a.flags = m->a.flags &
3206 					    ~PGA_QUEUE_STATE_MASK;
3207 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3208 					    ("page %p is managed", m_new));
3209 					m_new->oflags = 0;
3210 					pmap_copy_page(m, m_new);
3211 					m_new->valid = m->valid;
3212 					m_new->dirty = m->dirty;
3213 					m->flags &= ~PG_ZERO;
3214 					vm_page_dequeue(m);
3215 					if (vm_page_replace_hold(m_new, object,
3216 					    m->pindex, m) &&
3217 					    vm_page_free_prep(m))
3218 						SLIST_INSERT_HEAD(&free, m,
3219 						    plinks.s.ss);
3220 
3221 					/*
3222 					 * The new page must be deactivated
3223 					 * before the object is unlocked.
3224 					 */
3225 					vm_page_deactivate(m_new);
3226 				} else {
3227 					m->flags &= ~PG_ZERO;
3228 					vm_page_dequeue(m);
3229 					if (vm_page_free_prep(m))
3230 						SLIST_INSERT_HEAD(&free, m,
3231 						    plinks.s.ss);
3232 					KASSERT(m->dirty == 0,
3233 					    ("page %p is dirty", m));
3234 				}
3235 			} else
3236 				error = EBUSY;
3237 unlock:
3238 			VM_OBJECT_WUNLOCK(object);
3239 		} else {
3240 			MPASS(vm_page_domain(m) == domain);
3241 			vmd = VM_DOMAIN(domain);
3242 			vm_domain_free_lock(vmd);
3243 			order = m->order;
3244 			if (order < VM_NFREEORDER) {
3245 				/*
3246 				 * The page is enqueued in the physical memory
3247 				 * allocator's free page queues.  Moreover, it
3248 				 * is the first page in a power-of-two-sized
3249 				 * run of contiguous free pages.  Jump ahead
3250 				 * to the last page within that run, and
3251 				 * continue from there.
3252 				 */
3253 				m += (1 << order) - 1;
3254 			}
3255 #if VM_NRESERVLEVEL > 0
3256 			else if (vm_reserv_is_page_free(m))
3257 				order = 0;
3258 #endif
3259 			vm_domain_free_unlock(vmd);
3260 			if (order == VM_NFREEORDER)
3261 				error = EINVAL;
3262 		}
3263 	}
3264 	if ((m = SLIST_FIRST(&free)) != NULL) {
3265 		int cnt;
3266 
3267 		vmd = VM_DOMAIN(domain);
3268 		cnt = 0;
3269 		vm_domain_free_lock(vmd);
3270 		do {
3271 			MPASS(vm_page_domain(m) == domain);
3272 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3273 			vm_phys_free_pages(m, m->pool, 0);
3274 			cnt++;
3275 		} while ((m = SLIST_FIRST(&free)) != NULL);
3276 		vm_domain_free_unlock(vmd);
3277 		vm_domain_freecnt_inc(vmd, cnt);
3278 	}
3279 	return (error);
3280 }
3281 
3282 #define	NRUNS	16
3283 
3284 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3285 
3286 #define	MIN_RECLAIM	8
3287 
3288 /*
3289  *	vm_page_reclaim_contig:
3290  *
3291  *	Reclaim allocated, contiguous physical memory satisfying the specified
3292  *	conditions by relocating the virtual pages using that physical memory.
3293  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3294  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3295  *	currently able to reclaim the requested number of pages.  Since
3296  *	relocation requires the allocation of physical pages, reclamation may
3297  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3298  *	fails in this manner, callers are expected to perform vm_wait() before
3299  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3300  *
3301  *	The caller must always specify an allocation class through "req".
3302  *
3303  *	allocation classes:
3304  *	VM_ALLOC_NORMAL		normal process request
3305  *	VM_ALLOC_SYSTEM		system *really* needs a page
3306  *	VM_ALLOC_INTERRUPT	interrupt time request
3307  *
3308  *	The optional allocation flags are ignored.
3309  *
3310  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3311  *	must be a power of two.
3312  */
3313 int
3314 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3315     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3316     int desired_runs)
3317 {
3318 	struct vm_domain *vmd;
3319 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3320 	u_long count, minalign, reclaimed;
3321 	int error, i, min_reclaim, nruns, options, req_class;
3322 	int segind, start_segind;
3323 	int ret;
3324 
3325 	KASSERT(npages > 0, ("npages is 0"));
3326 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3327 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3328 
3329 	ret = ENOMEM;
3330 
3331 	/*
3332 	 * If the caller wants to reclaim multiple runs, try to allocate
3333 	 * space to store the runs.  If that fails, fall back to the old
3334 	 * behavior of just reclaiming MIN_RECLAIM pages.
3335 	 */
3336 	if (desired_runs > 1)
3337 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3338 		    M_TEMP, M_NOWAIT);
3339 	else
3340 		m_runs = NULL;
3341 
3342 	if (m_runs == NULL) {
3343 		m_runs = _m_runs;
3344 		nruns = NRUNS;
3345 	} else {
3346 		nruns = NRUNS + desired_runs - 1;
3347 	}
3348 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3349 
3350 	/*
3351 	 * The caller will attempt an allocation after some runs have been
3352 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3353 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3354 	 * power of two or maximum chunk size, and ensure that each run is
3355 	 * suitably aligned.
3356 	 */
3357 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3358 	npages = roundup2(npages, minalign);
3359 	if (alignment < ptoa(minalign))
3360 		alignment = ptoa(minalign);
3361 
3362 	/*
3363 	 * The page daemon is allowed to dig deeper into the free page list.
3364 	 */
3365 	req_class = req & VM_ALLOC_CLASS_MASK;
3366 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3367 		req_class = VM_ALLOC_SYSTEM;
3368 
3369 	start_segind = vm_phys_lookup_segind(low);
3370 
3371 	/*
3372 	 * Return if the number of free pages cannot satisfy the requested
3373 	 * allocation.
3374 	 */
3375 	vmd = VM_DOMAIN(domain);
3376 	count = vmd->vmd_free_count;
3377 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3378 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3379 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3380 		goto done;
3381 
3382 	/*
3383 	 * Scan up to three times, relaxing the restrictions ("options") on
3384 	 * the reclamation of reservations and superpages each time.
3385 	 */
3386 	for (options = VPSC_NORESERV;;) {
3387 		bool phys_range_exists = false;
3388 
3389 		/*
3390 		 * Find the highest runs that satisfy the given constraints
3391 		 * and restrictions, and record them in "m_runs".
3392 		 */
3393 		count = 0;
3394 		segind = start_segind;
3395 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3396 		    npages, low, high)) != -1) {
3397 			phys_range_exists = true;
3398 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3399 			    bounds[1], alignment, boundary, options))) {
3400 				bounds[0] = m_run + npages;
3401 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3402 				count++;
3403 			}
3404 			segind++;
3405 		}
3406 
3407 		if (!phys_range_exists) {
3408 			ret = ERANGE;
3409 			goto done;
3410 		}
3411 
3412 		/*
3413 		 * Reclaim the highest runs in LIFO (descending) order until
3414 		 * the number of reclaimed pages, "reclaimed", is at least
3415 		 * "min_reclaim".  Reset "reclaimed" each time because each
3416 		 * reclamation is idempotent, and runs will (likely) recur
3417 		 * from one scan to the next as restrictions are relaxed.
3418 		 */
3419 		reclaimed = 0;
3420 		for (i = 0; count > 0 && i < nruns; i++) {
3421 			count--;
3422 			m_run = m_runs[RUN_INDEX(count, nruns)];
3423 			error = vm_page_reclaim_run(req_class, domain, npages,
3424 			    m_run, high);
3425 			if (error == 0) {
3426 				reclaimed += npages;
3427 				if (reclaimed >= min_reclaim) {
3428 					ret = 0;
3429 					goto done;
3430 				}
3431 			}
3432 		}
3433 
3434 		/*
3435 		 * Either relax the restrictions on the next scan or return if
3436 		 * the last scan had no restrictions.
3437 		 */
3438 		if (options == VPSC_NORESERV)
3439 			options = VPSC_NOSUPER;
3440 		else if (options == VPSC_NOSUPER)
3441 			options = VPSC_ANY;
3442 		else if (options == VPSC_ANY) {
3443 			if (reclaimed != 0)
3444 				ret = 0;
3445 			goto done;
3446 		}
3447 	}
3448 done:
3449 	if (m_runs != _m_runs)
3450 		free(m_runs, M_TEMP);
3451 	return (ret);
3452 }
3453 
3454 int
3455 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3456     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3457 {
3458 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3459 	    alignment, boundary, 1));
3460 }
3461 
3462 int
3463 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3464     u_long alignment, vm_paddr_t boundary)
3465 {
3466 	struct vm_domainset_iter di;
3467 	int domain, ret, status;
3468 
3469 	ret = ERANGE;
3470 
3471 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3472 	do {
3473 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3474 		    high, alignment, boundary);
3475 		if (status == 0)
3476 			return (0);
3477 		else if (status == ERANGE)
3478 			vm_domainset_iter_ignore(&di, domain);
3479 		else {
3480 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3481 			    "from vm_page_reclaim_contig_domain()", status));
3482 			ret = ENOMEM;
3483 		}
3484 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3485 
3486 	return (ret);
3487 }
3488 
3489 /*
3490  * Set the domain in the appropriate page level domainset.
3491  */
3492 void
3493 vm_domain_set(struct vm_domain *vmd)
3494 {
3495 
3496 	mtx_lock(&vm_domainset_lock);
3497 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3498 		vmd->vmd_minset = 1;
3499 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3500 	}
3501 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3502 		vmd->vmd_severeset = 1;
3503 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3504 	}
3505 	mtx_unlock(&vm_domainset_lock);
3506 }
3507 
3508 /*
3509  * Clear the domain from the appropriate page level domainset.
3510  */
3511 void
3512 vm_domain_clear(struct vm_domain *vmd)
3513 {
3514 
3515 	mtx_lock(&vm_domainset_lock);
3516 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3517 		vmd->vmd_minset = 0;
3518 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3519 		if (vm_min_waiters != 0) {
3520 			vm_min_waiters = 0;
3521 			wakeup(&vm_min_domains);
3522 		}
3523 	}
3524 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3525 		vmd->vmd_severeset = 0;
3526 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3527 		if (vm_severe_waiters != 0) {
3528 			vm_severe_waiters = 0;
3529 			wakeup(&vm_severe_domains);
3530 		}
3531 	}
3532 
3533 	/*
3534 	 * If pageout daemon needs pages, then tell it that there are
3535 	 * some free.
3536 	 */
3537 	if (vmd->vmd_pageout_pages_needed &&
3538 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3539 		wakeup(&vmd->vmd_pageout_pages_needed);
3540 		vmd->vmd_pageout_pages_needed = 0;
3541 	}
3542 
3543 	/* See comments in vm_wait_doms(). */
3544 	if (vm_pageproc_waiters) {
3545 		vm_pageproc_waiters = 0;
3546 		wakeup(&vm_pageproc_waiters);
3547 	}
3548 	mtx_unlock(&vm_domainset_lock);
3549 }
3550 
3551 /*
3552  * Wait for free pages to exceed the min threshold globally.
3553  */
3554 void
3555 vm_wait_min(void)
3556 {
3557 
3558 	mtx_lock(&vm_domainset_lock);
3559 	while (vm_page_count_min()) {
3560 		vm_min_waiters++;
3561 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3562 	}
3563 	mtx_unlock(&vm_domainset_lock);
3564 }
3565 
3566 /*
3567  * Wait for free pages to exceed the severe threshold globally.
3568  */
3569 void
3570 vm_wait_severe(void)
3571 {
3572 
3573 	mtx_lock(&vm_domainset_lock);
3574 	while (vm_page_count_severe()) {
3575 		vm_severe_waiters++;
3576 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3577 		    "vmwait", 0);
3578 	}
3579 	mtx_unlock(&vm_domainset_lock);
3580 }
3581 
3582 u_int
3583 vm_wait_count(void)
3584 {
3585 
3586 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3587 }
3588 
3589 int
3590 vm_wait_doms(const domainset_t *wdoms, int mflags)
3591 {
3592 	int error;
3593 
3594 	error = 0;
3595 
3596 	/*
3597 	 * We use racey wakeup synchronization to avoid expensive global
3598 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3599 	 * To handle this, we only sleep for one tick in this instance.  It
3600 	 * is expected that most allocations for the pageproc will come from
3601 	 * kmem or vm_page_grab* which will use the more specific and
3602 	 * race-free vm_wait_domain().
3603 	 */
3604 	if (curproc == pageproc) {
3605 		mtx_lock(&vm_domainset_lock);
3606 		vm_pageproc_waiters++;
3607 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3608 		    PVM | PDROP | mflags, "pageprocwait", 1);
3609 	} else {
3610 		/*
3611 		 * XXX Ideally we would wait only until the allocation could
3612 		 * be satisfied.  This condition can cause new allocators to
3613 		 * consume all freed pages while old allocators wait.
3614 		 */
3615 		mtx_lock(&vm_domainset_lock);
3616 		if (vm_page_count_min_set(wdoms)) {
3617 			if (pageproc == NULL)
3618 				panic("vm_wait in early boot");
3619 			vm_min_waiters++;
3620 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3621 			    PVM | PDROP | mflags, "vmwait", 0);
3622 		} else
3623 			mtx_unlock(&vm_domainset_lock);
3624 	}
3625 	return (error);
3626 }
3627 
3628 /*
3629  *	vm_wait_domain:
3630  *
3631  *	Sleep until free pages are available for allocation.
3632  *	- Called in various places after failed memory allocations.
3633  */
3634 void
3635 vm_wait_domain(int domain)
3636 {
3637 	struct vm_domain *vmd;
3638 	domainset_t wdom;
3639 
3640 	vmd = VM_DOMAIN(domain);
3641 	vm_domain_free_assert_unlocked(vmd);
3642 
3643 	if (curproc == pageproc) {
3644 		mtx_lock(&vm_domainset_lock);
3645 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3646 			vmd->vmd_pageout_pages_needed = 1;
3647 			msleep(&vmd->vmd_pageout_pages_needed,
3648 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3649 		} else
3650 			mtx_unlock(&vm_domainset_lock);
3651 	} else {
3652 		DOMAINSET_ZERO(&wdom);
3653 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3654 		vm_wait_doms(&wdom, 0);
3655 	}
3656 }
3657 
3658 static int
3659 vm_wait_flags(vm_object_t obj, int mflags)
3660 {
3661 	struct domainset *d;
3662 
3663 	d = NULL;
3664 
3665 	/*
3666 	 * Carefully fetch pointers only once: the struct domainset
3667 	 * itself is ummutable but the pointer might change.
3668 	 */
3669 	if (obj != NULL)
3670 		d = obj->domain.dr_policy;
3671 	if (d == NULL)
3672 		d = curthread->td_domain.dr_policy;
3673 
3674 	return (vm_wait_doms(&d->ds_mask, mflags));
3675 }
3676 
3677 /*
3678  *	vm_wait:
3679  *
3680  *	Sleep until free pages are available for allocation in the
3681  *	affinity domains of the obj.  If obj is NULL, the domain set
3682  *	for the calling thread is used.
3683  *	Called in various places after failed memory allocations.
3684  */
3685 void
3686 vm_wait(vm_object_t obj)
3687 {
3688 	(void)vm_wait_flags(obj, 0);
3689 }
3690 
3691 int
3692 vm_wait_intr(vm_object_t obj)
3693 {
3694 	return (vm_wait_flags(obj, PCATCH));
3695 }
3696 
3697 /*
3698  *	vm_domain_alloc_fail:
3699  *
3700  *	Called when a page allocation function fails.  Informs the
3701  *	pagedaemon and performs the requested wait.  Requires the
3702  *	domain_free and object lock on entry.  Returns with the
3703  *	object lock held and free lock released.  Returns an error when
3704  *	retry is necessary.
3705  *
3706  */
3707 static int
3708 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3709 {
3710 
3711 	vm_domain_free_assert_unlocked(vmd);
3712 
3713 	atomic_add_int(&vmd->vmd_pageout_deficit,
3714 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3715 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3716 		if (object != NULL)
3717 			VM_OBJECT_WUNLOCK(object);
3718 		vm_wait_domain(vmd->vmd_domain);
3719 		if (object != NULL)
3720 			VM_OBJECT_WLOCK(object);
3721 		if (req & VM_ALLOC_WAITOK)
3722 			return (EAGAIN);
3723 	}
3724 
3725 	return (0);
3726 }
3727 
3728 /*
3729  *	vm_waitpfault:
3730  *
3731  *	Sleep until free pages are available for allocation.
3732  *	- Called only in vm_fault so that processes page faulting
3733  *	  can be easily tracked.
3734  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3735  *	  processes will be able to grab memory first.  Do not change
3736  *	  this balance without careful testing first.
3737  */
3738 void
3739 vm_waitpfault(struct domainset *dset, int timo)
3740 {
3741 
3742 	/*
3743 	 * XXX Ideally we would wait only until the allocation could
3744 	 * be satisfied.  This condition can cause new allocators to
3745 	 * consume all freed pages while old allocators wait.
3746 	 */
3747 	mtx_lock(&vm_domainset_lock);
3748 	if (vm_page_count_min_set(&dset->ds_mask)) {
3749 		vm_min_waiters++;
3750 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3751 		    "pfault", timo);
3752 	} else
3753 		mtx_unlock(&vm_domainset_lock);
3754 }
3755 
3756 static struct vm_pagequeue *
3757 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3758 {
3759 
3760 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3761 }
3762 
3763 #ifdef INVARIANTS
3764 static struct vm_pagequeue *
3765 vm_page_pagequeue(vm_page_t m)
3766 {
3767 
3768 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3769 }
3770 #endif
3771 
3772 static __always_inline bool
3773 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3774 {
3775 	vm_page_astate_t tmp;
3776 
3777 	tmp = *old;
3778 	do {
3779 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3780 			return (true);
3781 		counter_u64_add(pqstate_commit_retries, 1);
3782 	} while (old->_bits == tmp._bits);
3783 
3784 	return (false);
3785 }
3786 
3787 /*
3788  * Do the work of committing a queue state update that moves the page out of
3789  * its current queue.
3790  */
3791 static bool
3792 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3793     vm_page_astate_t *old, vm_page_astate_t new)
3794 {
3795 	vm_page_t next;
3796 
3797 	vm_pagequeue_assert_locked(pq);
3798 	KASSERT(vm_page_pagequeue(m) == pq,
3799 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3800 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3801 	    ("%s: invalid queue indices %d %d",
3802 	    __func__, old->queue, new.queue));
3803 
3804 	/*
3805 	 * Once the queue index of the page changes there is nothing
3806 	 * synchronizing with further updates to the page's physical
3807 	 * queue state.  Therefore we must speculatively remove the page
3808 	 * from the queue now and be prepared to roll back if the queue
3809 	 * state update fails.  If the page is not physically enqueued then
3810 	 * we just update its queue index.
3811 	 */
3812 	if ((old->flags & PGA_ENQUEUED) != 0) {
3813 		new.flags &= ~PGA_ENQUEUED;
3814 		next = TAILQ_NEXT(m, plinks.q);
3815 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3816 		vm_pagequeue_cnt_dec(pq);
3817 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3818 			if (next == NULL)
3819 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3820 			else
3821 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3822 			vm_pagequeue_cnt_inc(pq);
3823 			return (false);
3824 		} else {
3825 			return (true);
3826 		}
3827 	} else {
3828 		return (vm_page_pqstate_fcmpset(m, old, new));
3829 	}
3830 }
3831 
3832 static bool
3833 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3834     vm_page_astate_t new)
3835 {
3836 	struct vm_pagequeue *pq;
3837 	vm_page_astate_t as;
3838 	bool ret;
3839 
3840 	pq = _vm_page_pagequeue(m, old->queue);
3841 
3842 	/*
3843 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3844 	 * corresponding page queue lock is held.
3845 	 */
3846 	vm_pagequeue_lock(pq);
3847 	as = vm_page_astate_load(m);
3848 	if (__predict_false(as._bits != old->_bits)) {
3849 		*old = as;
3850 		ret = false;
3851 	} else {
3852 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3853 	}
3854 	vm_pagequeue_unlock(pq);
3855 	return (ret);
3856 }
3857 
3858 /*
3859  * Commit a queue state update that enqueues or requeues a page.
3860  */
3861 static bool
3862 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3863     vm_page_astate_t *old, vm_page_astate_t new)
3864 {
3865 	struct vm_domain *vmd;
3866 
3867 	vm_pagequeue_assert_locked(pq);
3868 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3869 	    ("%s: invalid queue indices %d %d",
3870 	    __func__, old->queue, new.queue));
3871 
3872 	new.flags |= PGA_ENQUEUED;
3873 	if (!vm_page_pqstate_fcmpset(m, old, new))
3874 		return (false);
3875 
3876 	if ((old->flags & PGA_ENQUEUED) != 0)
3877 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3878 	else
3879 		vm_pagequeue_cnt_inc(pq);
3880 
3881 	/*
3882 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3883 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3884 	 * applied, even if it was set first.
3885 	 */
3886 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3887 		vmd = vm_pagequeue_domain(m);
3888 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3889 		    ("%s: invalid page queue for page %p", __func__, m));
3890 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3891 	} else {
3892 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3893 	}
3894 	return (true);
3895 }
3896 
3897 /*
3898  * Commit a queue state update that encodes a request for a deferred queue
3899  * operation.
3900  */
3901 static bool
3902 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3903     vm_page_astate_t new)
3904 {
3905 
3906 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3907 	    ("%s: invalid state, queue %d flags %x",
3908 	    __func__, new.queue, new.flags));
3909 
3910 	if (old->_bits != new._bits &&
3911 	    !vm_page_pqstate_fcmpset(m, old, new))
3912 		return (false);
3913 	vm_page_pqbatch_submit(m, new.queue);
3914 	return (true);
3915 }
3916 
3917 /*
3918  * A generic queue state update function.  This handles more cases than the
3919  * specialized functions above.
3920  */
3921 bool
3922 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3923 {
3924 
3925 	if (old->_bits == new._bits)
3926 		return (true);
3927 
3928 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3929 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3930 			return (false);
3931 		if (new.queue != PQ_NONE)
3932 			vm_page_pqbatch_submit(m, new.queue);
3933 	} else {
3934 		if (!vm_page_pqstate_fcmpset(m, old, new))
3935 			return (false);
3936 		if (new.queue != PQ_NONE &&
3937 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3938 			vm_page_pqbatch_submit(m, new.queue);
3939 	}
3940 	return (true);
3941 }
3942 
3943 /*
3944  * Apply deferred queue state updates to a page.
3945  */
3946 static inline void
3947 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3948 {
3949 	vm_page_astate_t new, old;
3950 
3951 	CRITICAL_ASSERT(curthread);
3952 	vm_pagequeue_assert_locked(pq);
3953 	KASSERT(queue < PQ_COUNT,
3954 	    ("%s: invalid queue index %d", __func__, queue));
3955 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3956 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3957 
3958 	for (old = vm_page_astate_load(m);;) {
3959 		if (__predict_false(old.queue != queue ||
3960 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3961 			counter_u64_add(queue_nops, 1);
3962 			break;
3963 		}
3964 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3965 		    ("%s: page %p is unmanaged", __func__, m));
3966 
3967 		new = old;
3968 		if ((old.flags & PGA_DEQUEUE) != 0) {
3969 			new.flags &= ~PGA_QUEUE_OP_MASK;
3970 			new.queue = PQ_NONE;
3971 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3972 			    m, &old, new))) {
3973 				counter_u64_add(queue_ops, 1);
3974 				break;
3975 			}
3976 		} else {
3977 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3978 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3979 			    m, &old, new))) {
3980 				counter_u64_add(queue_ops, 1);
3981 				break;
3982 			}
3983 		}
3984 	}
3985 }
3986 
3987 static void
3988 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3989     uint8_t queue)
3990 {
3991 	int i;
3992 
3993 	for (i = 0; i < bq->bq_cnt; i++)
3994 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3995 	vm_batchqueue_init(bq);
3996 }
3997 
3998 /*
3999  *	vm_page_pqbatch_submit:		[ internal use only ]
4000  *
4001  *	Enqueue a page in the specified page queue's batched work queue.
4002  *	The caller must have encoded the requested operation in the page
4003  *	structure's a.flags field.
4004  */
4005 void
4006 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
4007 {
4008 	struct vm_batchqueue *bq;
4009 	struct vm_pagequeue *pq;
4010 	int domain, slots_remaining;
4011 
4012 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
4013 
4014 	domain = vm_page_domain(m);
4015 	critical_enter();
4016 	bq = DPCPU_PTR(pqbatch[domain][queue]);
4017 	slots_remaining = vm_batchqueue_insert(bq, m);
4018 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
4019 		/* keep building the bq */
4020 		critical_exit();
4021 		return;
4022 	} else if (slots_remaining > 0 ) {
4023 		/* Try to process the bq if we can get the lock */
4024 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
4025 		if (vm_pagequeue_trylock(pq)) {
4026 			vm_pqbatch_process(pq, bq, queue);
4027 			vm_pagequeue_unlock(pq);
4028 		}
4029 		critical_exit();
4030 		return;
4031 	}
4032 	critical_exit();
4033 
4034 	/* if we make it here, the bq is full so wait for the lock */
4035 
4036 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
4037 	vm_pagequeue_lock(pq);
4038 	critical_enter();
4039 	bq = DPCPU_PTR(pqbatch[domain][queue]);
4040 	vm_pqbatch_process(pq, bq, queue);
4041 	vm_pqbatch_process_page(pq, m, queue);
4042 	vm_pagequeue_unlock(pq);
4043 	critical_exit();
4044 }
4045 
4046 /*
4047  *	vm_page_pqbatch_drain:		[ internal use only ]
4048  *
4049  *	Force all per-CPU page queue batch queues to be drained.  This is
4050  *	intended for use in severe memory shortages, to ensure that pages
4051  *	do not remain stuck in the batch queues.
4052  */
4053 void
4054 vm_page_pqbatch_drain(void)
4055 {
4056 	struct thread *td;
4057 	struct vm_domain *vmd;
4058 	struct vm_pagequeue *pq;
4059 	int cpu, domain, queue;
4060 
4061 	td = curthread;
4062 	CPU_FOREACH(cpu) {
4063 		thread_lock(td);
4064 		sched_bind(td, cpu);
4065 		thread_unlock(td);
4066 
4067 		for (domain = 0; domain < vm_ndomains; domain++) {
4068 			vmd = VM_DOMAIN(domain);
4069 			for (queue = 0; queue < PQ_COUNT; queue++) {
4070 				pq = &vmd->vmd_pagequeues[queue];
4071 				vm_pagequeue_lock(pq);
4072 				critical_enter();
4073 				vm_pqbatch_process(pq,
4074 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
4075 				critical_exit();
4076 				vm_pagequeue_unlock(pq);
4077 			}
4078 		}
4079 	}
4080 	thread_lock(td);
4081 	sched_unbind(td);
4082 	thread_unlock(td);
4083 }
4084 
4085 /*
4086  *	vm_page_dequeue_deferred:	[ internal use only ]
4087  *
4088  *	Request removal of the given page from its current page
4089  *	queue.  Physical removal from the queue may be deferred
4090  *	indefinitely.
4091  */
4092 void
4093 vm_page_dequeue_deferred(vm_page_t m)
4094 {
4095 	vm_page_astate_t new, old;
4096 
4097 	old = vm_page_astate_load(m);
4098 	do {
4099 		if (old.queue == PQ_NONE) {
4100 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4101 			    ("%s: page %p has unexpected queue state",
4102 			    __func__, m));
4103 			break;
4104 		}
4105 		new = old;
4106 		new.flags |= PGA_DEQUEUE;
4107 	} while (!vm_page_pqstate_commit_request(m, &old, new));
4108 }
4109 
4110 /*
4111  *	vm_page_dequeue:
4112  *
4113  *	Remove the page from whichever page queue it's in, if any, before
4114  *	returning.
4115  */
4116 void
4117 vm_page_dequeue(vm_page_t m)
4118 {
4119 	vm_page_astate_t new, old;
4120 
4121 	old = vm_page_astate_load(m);
4122 	do {
4123 		if (old.queue == PQ_NONE) {
4124 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4125 			    ("%s: page %p has unexpected queue state",
4126 			    __func__, m));
4127 			break;
4128 		}
4129 		new = old;
4130 		new.flags &= ~PGA_QUEUE_OP_MASK;
4131 		new.queue = PQ_NONE;
4132 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
4133 
4134 }
4135 
4136 /*
4137  * Schedule the given page for insertion into the specified page queue.
4138  * Physical insertion of the page may be deferred indefinitely.
4139  */
4140 static void
4141 vm_page_enqueue(vm_page_t m, uint8_t queue)
4142 {
4143 
4144 	KASSERT(m->a.queue == PQ_NONE &&
4145 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
4146 	    ("%s: page %p is already enqueued", __func__, m));
4147 	KASSERT(m->ref_count > 0,
4148 	    ("%s: page %p does not carry any references", __func__, m));
4149 
4150 	m->a.queue = queue;
4151 	if ((m->a.flags & PGA_REQUEUE) == 0)
4152 		vm_page_aflag_set(m, PGA_REQUEUE);
4153 	vm_page_pqbatch_submit(m, queue);
4154 }
4155 
4156 /*
4157  *	vm_page_free_prep:
4158  *
4159  *	Prepares the given page to be put on the free list,
4160  *	disassociating it from any VM object. The caller may return
4161  *	the page to the free list only if this function returns true.
4162  *
4163  *	The object, if it exists, must be locked, and then the page must
4164  *	be xbusy.  Otherwise the page must be not busied.  A managed
4165  *	page must be unmapped.
4166  */
4167 static bool
4168 vm_page_free_prep(vm_page_t m)
4169 {
4170 
4171 	/*
4172 	 * Synchronize with threads that have dropped a reference to this
4173 	 * page.
4174 	 */
4175 	atomic_thread_fence_acq();
4176 
4177 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
4178 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
4179 		uint64_t *p;
4180 		int i;
4181 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4182 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4183 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4184 			    m, i, (uintmax_t)*p));
4185 	}
4186 #endif
4187 	KASSERT((m->flags & PG_NOFREE) == 0,
4188 	    ("%s: attempting to free a PG_NOFREE page", __func__));
4189 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4190 		KASSERT(!pmap_page_is_mapped(m),
4191 		    ("vm_page_free_prep: freeing mapped page %p", m));
4192 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4193 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4194 	} else {
4195 		KASSERT(m->a.queue == PQ_NONE,
4196 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4197 	}
4198 	VM_CNT_INC(v_tfree);
4199 
4200 	if (m->object != NULL) {
4201 		vm_page_radix_remove(m);
4202 		vm_page_free_object_prep(m);
4203 	} else
4204 		vm_page_assert_unbusied(m);
4205 
4206 	vm_page_busy_free(m);
4207 
4208 	/*
4209 	 * If fictitious remove object association and
4210 	 * return.
4211 	 */
4212 	if ((m->flags & PG_FICTITIOUS) != 0) {
4213 		KASSERT(m->ref_count == 1,
4214 		    ("fictitious page %p is referenced", m));
4215 		KASSERT(m->a.queue == PQ_NONE,
4216 		    ("fictitious page %p is queued", m));
4217 		return (false);
4218 	}
4219 
4220 	/*
4221 	 * Pages need not be dequeued before they are returned to the physical
4222 	 * memory allocator, but they must at least be marked for a deferred
4223 	 * dequeue.
4224 	 */
4225 	if ((m->oflags & VPO_UNMANAGED) == 0)
4226 		vm_page_dequeue_deferred(m);
4227 
4228 	m->valid = 0;
4229 	vm_page_undirty(m);
4230 
4231 	if (m->ref_count != 0)
4232 		panic("vm_page_free_prep: page %p has references", m);
4233 
4234 	/*
4235 	 * Restore the default memory attribute to the page.
4236 	 */
4237 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4238 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4239 
4240 #if VM_NRESERVLEVEL > 0
4241 	/*
4242 	 * Determine whether the page belongs to a reservation.  If the page was
4243 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4244 	 * as an optimization, we avoid the check in that case.
4245 	 */
4246 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4247 		return (false);
4248 #endif
4249 
4250 	return (true);
4251 }
4252 
4253 /*
4254  *	vm_page_free_toq:
4255  *
4256  *	Returns the given page to the free list, disassociating it
4257  *	from any VM object.
4258  *
4259  *	The object must be locked.  The page must be exclusively busied if it
4260  *	belongs to an object.
4261  */
4262 static void
4263 vm_page_free_toq(vm_page_t m)
4264 {
4265 	struct vm_domain *vmd;
4266 	uma_zone_t zone;
4267 
4268 	if (!vm_page_free_prep(m))
4269 		return;
4270 
4271 	vmd = vm_pagequeue_domain(m);
4272 	zone = vmd->vmd_pgcache[m->pool].zone;
4273 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4274 		uma_zfree(zone, m);
4275 		return;
4276 	}
4277 	vm_domain_free_lock(vmd);
4278 	vm_phys_free_pages(m, m->pool, 0);
4279 	vm_domain_free_unlock(vmd);
4280 	vm_domain_freecnt_inc(vmd, 1);
4281 }
4282 
4283 /*
4284  *	vm_page_free_pages_toq:
4285  *
4286  *	Returns a list of pages to the free list, disassociating it
4287  *	from any VM object.  In other words, this is equivalent to
4288  *	calling vm_page_free_toq() for each page of a list of VM objects.
4289  */
4290 int
4291 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4292 {
4293 	vm_page_t m;
4294 	int count;
4295 
4296 	if (SLIST_EMPTY(free))
4297 		return (0);
4298 
4299 	count = 0;
4300 	while ((m = SLIST_FIRST(free)) != NULL) {
4301 		count++;
4302 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4303 		vm_page_free_toq(m);
4304 	}
4305 
4306 	if (update_wire_count)
4307 		vm_wire_sub(count);
4308 	return (count);
4309 }
4310 
4311 /*
4312  * Mark this page as wired down.  For managed pages, this prevents reclamation
4313  * by the page daemon, or when the containing object, if any, is destroyed.
4314  */
4315 void
4316 vm_page_wire(vm_page_t m)
4317 {
4318 	u_int old;
4319 
4320 #ifdef INVARIANTS
4321 	if (m->object != NULL && !vm_page_busied(m) &&
4322 	    !vm_object_busied(m->object))
4323 		VM_OBJECT_ASSERT_LOCKED(m->object);
4324 #endif
4325 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4326 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4327 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4328 
4329 	old = atomic_fetchadd_int(&m->ref_count, 1);
4330 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4331 	    ("vm_page_wire: counter overflow for page %p", m));
4332 	if (VPRC_WIRE_COUNT(old) == 0) {
4333 		if ((m->oflags & VPO_UNMANAGED) == 0)
4334 			vm_page_aflag_set(m, PGA_DEQUEUE);
4335 		vm_wire_add(1);
4336 	}
4337 }
4338 
4339 /*
4340  * Attempt to wire a mapped page following a pmap lookup of that page.
4341  * This may fail if a thread is concurrently tearing down mappings of the page.
4342  * The transient failure is acceptable because it translates to the
4343  * failure of the caller pmap_extract_and_hold(), which should be then
4344  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4345  */
4346 bool
4347 vm_page_wire_mapped(vm_page_t m)
4348 {
4349 	u_int old;
4350 
4351 	old = atomic_load_int(&m->ref_count);
4352 	do {
4353 		KASSERT(old > 0,
4354 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4355 		if ((old & VPRC_BLOCKED) != 0)
4356 			return (false);
4357 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4358 
4359 	if (VPRC_WIRE_COUNT(old) == 0) {
4360 		if ((m->oflags & VPO_UNMANAGED) == 0)
4361 			vm_page_aflag_set(m, PGA_DEQUEUE);
4362 		vm_wire_add(1);
4363 	}
4364 	return (true);
4365 }
4366 
4367 /*
4368  * Release a wiring reference to a managed page.  If the page still belongs to
4369  * an object, update its position in the page queues to reflect the reference.
4370  * If the wiring was the last reference to the page, free the page.
4371  */
4372 static void
4373 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4374 {
4375 	u_int old;
4376 
4377 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4378 	    ("%s: page %p is unmanaged", __func__, m));
4379 
4380 	/*
4381 	 * Update LRU state before releasing the wiring reference.
4382 	 * Use a release store when updating the reference count to
4383 	 * synchronize with vm_page_free_prep().
4384 	 */
4385 	old = atomic_load_int(&m->ref_count);
4386 	do {
4387 		u_int count;
4388 
4389 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4390 		    ("vm_page_unwire: wire count underflow for page %p", m));
4391 
4392 		count = old & ~VPRC_BLOCKED;
4393 		if (count > VPRC_OBJREF + 1) {
4394 			/*
4395 			 * The page has at least one other wiring reference.  An
4396 			 * earlier iteration of this loop may have called
4397 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4398 			 * re-set it if necessary.
4399 			 */
4400 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4401 				vm_page_aflag_set(m, PGA_DEQUEUE);
4402 		} else if (count == VPRC_OBJREF + 1) {
4403 			/*
4404 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4405 			 * update the page's queue state to reflect the
4406 			 * reference.  If the page does not belong to an object
4407 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4408 			 * clear leftover queue state.
4409 			 */
4410 			vm_page_release_toq(m, nqueue, noreuse);
4411 		} else if (count == 1) {
4412 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4413 		}
4414 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4415 
4416 	if (VPRC_WIRE_COUNT(old) == 1) {
4417 		vm_wire_sub(1);
4418 		if (old == 1)
4419 			vm_page_free(m);
4420 	}
4421 }
4422 
4423 /*
4424  * Release one wiring of the specified page, potentially allowing it to be
4425  * paged out.
4426  *
4427  * Only managed pages belonging to an object can be paged out.  If the number
4428  * of wirings transitions to zero and the page is eligible for page out, then
4429  * the page is added to the specified paging queue.  If the released wiring
4430  * represented the last reference to the page, the page is freed.
4431  */
4432 void
4433 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4434 {
4435 
4436 	KASSERT(nqueue < PQ_COUNT,
4437 	    ("vm_page_unwire: invalid queue %u request for page %p",
4438 	    nqueue, m));
4439 
4440 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4441 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4442 			vm_page_free(m);
4443 		return;
4444 	}
4445 	vm_page_unwire_managed(m, nqueue, false);
4446 }
4447 
4448 /*
4449  * Unwire a page without (re-)inserting it into a page queue.  It is up
4450  * to the caller to enqueue, requeue, or free the page as appropriate.
4451  * In most cases involving managed pages, vm_page_unwire() should be used
4452  * instead.
4453  */
4454 bool
4455 vm_page_unwire_noq(vm_page_t m)
4456 {
4457 	u_int old;
4458 
4459 	old = vm_page_drop(m, 1);
4460 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4461 	    ("%s: counter underflow for page %p", __func__,  m));
4462 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4463 	    ("%s: missing ref on fictitious page %p", __func__, m));
4464 
4465 	if (VPRC_WIRE_COUNT(old) > 1)
4466 		return (false);
4467 	if ((m->oflags & VPO_UNMANAGED) == 0)
4468 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4469 	vm_wire_sub(1);
4470 	return (true);
4471 }
4472 
4473 /*
4474  * Ensure that the page ends up in the specified page queue.  If the page is
4475  * active or being moved to the active queue, ensure that its act_count is
4476  * at least ACT_INIT but do not otherwise mess with it.
4477  */
4478 static __always_inline void
4479 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4480 {
4481 	vm_page_astate_t old, new;
4482 
4483 	KASSERT(m->ref_count > 0,
4484 	    ("%s: page %p does not carry any references", __func__, m));
4485 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4486 	    ("%s: invalid flags %x", __func__, nflag));
4487 
4488 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4489 		return;
4490 
4491 	old = vm_page_astate_load(m);
4492 	do {
4493 		if ((old.flags & PGA_DEQUEUE) != 0)
4494 			break;
4495 		new = old;
4496 		new.flags &= ~PGA_QUEUE_OP_MASK;
4497 		if (nqueue == PQ_ACTIVE)
4498 			new.act_count = max(old.act_count, ACT_INIT);
4499 		if (old.queue == nqueue) {
4500 			/*
4501 			 * There is no need to requeue pages already in the
4502 			 * active queue.
4503 			 */
4504 			if (nqueue != PQ_ACTIVE ||
4505 			    (old.flags & PGA_ENQUEUED) == 0)
4506 				new.flags |= nflag;
4507 		} else {
4508 			new.flags |= nflag;
4509 			new.queue = nqueue;
4510 		}
4511 	} while (!vm_page_pqstate_commit(m, &old, new));
4512 }
4513 
4514 /*
4515  * Put the specified page on the active list (if appropriate).
4516  */
4517 void
4518 vm_page_activate(vm_page_t m)
4519 {
4520 
4521 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4522 }
4523 
4524 /*
4525  * Move the specified page to the tail of the inactive queue, or requeue
4526  * the page if it is already in the inactive queue.
4527  */
4528 void
4529 vm_page_deactivate(vm_page_t m)
4530 {
4531 
4532 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4533 }
4534 
4535 void
4536 vm_page_deactivate_noreuse(vm_page_t m)
4537 {
4538 
4539 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4540 }
4541 
4542 /*
4543  * Put a page in the laundry, or requeue it if it is already there.
4544  */
4545 void
4546 vm_page_launder(vm_page_t m)
4547 {
4548 
4549 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4550 }
4551 
4552 /*
4553  * Put a page in the PQ_UNSWAPPABLE holding queue.
4554  */
4555 void
4556 vm_page_unswappable(vm_page_t m)
4557 {
4558 
4559 	VM_OBJECT_ASSERT_LOCKED(m->object);
4560 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4561 	    ("page %p already unswappable", m));
4562 
4563 	vm_page_dequeue(m);
4564 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4565 }
4566 
4567 /*
4568  * Release a page back to the page queues in preparation for unwiring.
4569  */
4570 static void
4571 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4572 {
4573 	vm_page_astate_t old, new;
4574 	uint16_t nflag;
4575 
4576 	/*
4577 	 * Use a check of the valid bits to determine whether we should
4578 	 * accelerate reclamation of the page.  The object lock might not be
4579 	 * held here, in which case the check is racy.  At worst we will either
4580 	 * accelerate reclamation of a valid page and violate LRU, or
4581 	 * unnecessarily defer reclamation of an invalid page.
4582 	 *
4583 	 * If we were asked to not cache the page, place it near the head of the
4584 	 * inactive queue so that is reclaimed sooner.
4585 	 */
4586 	if (noreuse || vm_page_none_valid(m)) {
4587 		nqueue = PQ_INACTIVE;
4588 		nflag = PGA_REQUEUE_HEAD;
4589 	} else {
4590 		nflag = PGA_REQUEUE;
4591 	}
4592 
4593 	old = vm_page_astate_load(m);
4594 	do {
4595 		new = old;
4596 
4597 		/*
4598 		 * If the page is already in the active queue and we are not
4599 		 * trying to accelerate reclamation, simply mark it as
4600 		 * referenced and avoid any queue operations.
4601 		 */
4602 		new.flags &= ~PGA_QUEUE_OP_MASK;
4603 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4604 		    (old.flags & PGA_ENQUEUED) != 0)
4605 			new.flags |= PGA_REFERENCED;
4606 		else {
4607 			new.flags |= nflag;
4608 			new.queue = nqueue;
4609 		}
4610 	} while (!vm_page_pqstate_commit(m, &old, new));
4611 }
4612 
4613 /*
4614  * Unwire a page and either attempt to free it or re-add it to the page queues.
4615  */
4616 void
4617 vm_page_release(vm_page_t m, int flags)
4618 {
4619 	vm_object_t object;
4620 
4621 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4622 	    ("vm_page_release: page %p is unmanaged", m));
4623 
4624 	if ((flags & VPR_TRYFREE) != 0) {
4625 		for (;;) {
4626 			object = atomic_load_ptr(&m->object);
4627 			if (object == NULL)
4628 				break;
4629 			/* Depends on type-stability. */
4630 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4631 				break;
4632 			if (object == m->object) {
4633 				vm_page_release_locked(m, flags);
4634 				VM_OBJECT_WUNLOCK(object);
4635 				return;
4636 			}
4637 			VM_OBJECT_WUNLOCK(object);
4638 		}
4639 	}
4640 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4641 }
4642 
4643 /* See vm_page_release(). */
4644 void
4645 vm_page_release_locked(vm_page_t m, int flags)
4646 {
4647 
4648 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4649 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4650 	    ("vm_page_release_locked: page %p is unmanaged", m));
4651 
4652 	if (vm_page_unwire_noq(m)) {
4653 		if ((flags & VPR_TRYFREE) != 0 &&
4654 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4655 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4656 			/*
4657 			 * An unlocked lookup may have wired the page before the
4658 			 * busy lock was acquired, in which case the page must
4659 			 * not be freed.
4660 			 */
4661 			if (__predict_true(!vm_page_wired(m))) {
4662 				vm_page_free(m);
4663 				return;
4664 			}
4665 			vm_page_xunbusy(m);
4666 		} else {
4667 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4668 		}
4669 	}
4670 }
4671 
4672 static bool
4673 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4674 {
4675 	u_int old;
4676 
4677 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4678 	    ("vm_page_try_blocked_op: page %p has no object", m));
4679 	KASSERT(vm_page_busied(m),
4680 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4681 	VM_OBJECT_ASSERT_LOCKED(m->object);
4682 
4683 	old = atomic_load_int(&m->ref_count);
4684 	do {
4685 		KASSERT(old != 0,
4686 		    ("vm_page_try_blocked_op: page %p has no references", m));
4687 		KASSERT((old & VPRC_BLOCKED) == 0,
4688 		    ("vm_page_try_blocked_op: page %p blocks wirings", m));
4689 		if (VPRC_WIRE_COUNT(old) != 0)
4690 			return (false);
4691 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4692 
4693 	(op)(m);
4694 
4695 	/*
4696 	 * If the object is read-locked, new wirings may be created via an
4697 	 * object lookup.
4698 	 */
4699 	old = vm_page_drop(m, VPRC_BLOCKED);
4700 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4701 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4702 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4703 	    old, m));
4704 	return (true);
4705 }
4706 
4707 /*
4708  * Atomically check for wirings and remove all mappings of the page.
4709  */
4710 bool
4711 vm_page_try_remove_all(vm_page_t m)
4712 {
4713 
4714 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4715 }
4716 
4717 /*
4718  * Atomically check for wirings and remove all writeable mappings of the page.
4719  */
4720 bool
4721 vm_page_try_remove_write(vm_page_t m)
4722 {
4723 
4724 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4725 }
4726 
4727 /*
4728  * vm_page_advise
4729  *
4730  * 	Apply the specified advice to the given page.
4731  */
4732 void
4733 vm_page_advise(vm_page_t m, int advice)
4734 {
4735 
4736 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4737 	vm_page_assert_xbusied(m);
4738 
4739 	if (advice == MADV_FREE)
4740 		/*
4741 		 * Mark the page clean.  This will allow the page to be freed
4742 		 * without first paging it out.  MADV_FREE pages are often
4743 		 * quickly reused by malloc(3), so we do not do anything that
4744 		 * would result in a page fault on a later access.
4745 		 */
4746 		vm_page_undirty(m);
4747 	else if (advice != MADV_DONTNEED) {
4748 		if (advice == MADV_WILLNEED)
4749 			vm_page_activate(m);
4750 		return;
4751 	}
4752 
4753 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4754 		vm_page_dirty(m);
4755 
4756 	/*
4757 	 * Clear any references to the page.  Otherwise, the page daemon will
4758 	 * immediately reactivate the page.
4759 	 */
4760 	vm_page_aflag_clear(m, PGA_REFERENCED);
4761 
4762 	/*
4763 	 * Place clean pages near the head of the inactive queue rather than
4764 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4765 	 * the page will be reused quickly.  Dirty pages not already in the
4766 	 * laundry are moved there.
4767 	 */
4768 	if (m->dirty == 0)
4769 		vm_page_deactivate_noreuse(m);
4770 	else if (!vm_page_in_laundry(m))
4771 		vm_page_launder(m);
4772 }
4773 
4774 /*
4775  *	vm_page_grab_release
4776  *
4777  *	Helper routine for grab functions to release busy on return.
4778  */
4779 static inline void
4780 vm_page_grab_release(vm_page_t m, int allocflags)
4781 {
4782 
4783 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4784 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4785 			vm_page_sunbusy(m);
4786 		else
4787 			vm_page_xunbusy(m);
4788 	}
4789 }
4790 
4791 /*
4792  *	vm_page_grab_sleep
4793  *
4794  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4795  *	if the caller should retry and false otherwise.
4796  *
4797  *	If the object is locked on entry the object will be unlocked with
4798  *	false returns and still locked but possibly having been dropped
4799  *	with true returns.
4800  */
4801 static bool
4802 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4803     const char *wmesg, int allocflags, bool locked)
4804 {
4805 
4806 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4807 		return (false);
4808 
4809 	/*
4810 	 * Reference the page before unlocking and sleeping so that
4811 	 * the page daemon is less likely to reclaim it.
4812 	 */
4813 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4814 		vm_page_reference(m);
4815 
4816 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4817 	    locked)
4818 		VM_OBJECT_WLOCK(object);
4819 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4820 		return (false);
4821 
4822 	return (true);
4823 }
4824 
4825 /*
4826  * Assert that the grab flags are valid.
4827  */
4828 static inline void
4829 vm_page_grab_check(int allocflags)
4830 {
4831 
4832 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4833 	    (allocflags & VM_ALLOC_WIRED) != 0,
4834 	    ("vm_page_grab*: the pages must be busied or wired"));
4835 
4836 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4837 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4838 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4839 }
4840 
4841 /*
4842  * Calculate the page allocation flags for grab.
4843  */
4844 static inline int
4845 vm_page_grab_pflags(int allocflags)
4846 {
4847 	int pflags;
4848 
4849 	pflags = allocflags &
4850 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4851 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4852 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4853 		pflags |= VM_ALLOC_WAITFAIL;
4854 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4855 		pflags |= VM_ALLOC_SBUSY;
4856 
4857 	return (pflags);
4858 }
4859 
4860 /*
4861  * Grab a page, waiting until we are waken up due to the page
4862  * changing state.  We keep on waiting, if the page continues
4863  * to be in the object.  If the page doesn't exist, first allocate it
4864  * and then conditionally zero it.
4865  *
4866  * This routine may sleep.
4867  *
4868  * The object must be locked on entry.  The lock will, however, be released
4869  * and reacquired if the routine sleeps.
4870  */
4871 vm_page_t
4872 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4873 {
4874 	vm_page_t m;
4875 
4876 	VM_OBJECT_ASSERT_WLOCKED(object);
4877 	vm_page_grab_check(allocflags);
4878 
4879 retrylookup:
4880 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4881 		if (!vm_page_tryacquire(m, allocflags)) {
4882 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4883 			    allocflags, true))
4884 				goto retrylookup;
4885 			return (NULL);
4886 		}
4887 		goto out;
4888 	}
4889 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4890 		return (NULL);
4891 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4892 	if (m == NULL) {
4893 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4894 			return (NULL);
4895 		goto retrylookup;
4896 	}
4897 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4898 		pmap_zero_page(m);
4899 
4900 out:
4901 	vm_page_grab_release(m, allocflags);
4902 
4903 	return (m);
4904 }
4905 
4906 /*
4907  * Attempt to validate a page, locklessly acquiring it if necessary, given a
4908  * (object, pindex) tuple and either an invalided page or NULL.  The resulting
4909  * page will be validated against the identity tuple, and busied or wired as
4910  * requested.  A NULL page returned guarantees that the page was not in radix at
4911  * the time of the call but callers must perform higher level synchronization or
4912  * retry the operation under a lock if they require an atomic answer.  This is
4913  * the only lock free validation routine, other routines can depend on the
4914  * resulting page state.
4915  *
4916  * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to
4917  * caller flags.
4918  */
4919 #define PAGE_NOT_ACQUIRED ((vm_page_t)1)
4920 static vm_page_t
4921 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m,
4922     int allocflags)
4923 {
4924 	if (m == NULL)
4925 		m = vm_page_lookup_unlocked(object, pindex);
4926 	for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) {
4927 		if (vm_page_trybusy(m, allocflags)) {
4928 			if (m->object == object && m->pindex == pindex) {
4929 				if ((allocflags & VM_ALLOC_WIRED) != 0)
4930 					vm_page_wire(m);
4931 				vm_page_grab_release(m, allocflags);
4932 				break;
4933 			}
4934 			/* relookup. */
4935 			vm_page_busy_release(m);
4936 			cpu_spinwait();
4937 			continue;
4938 		}
4939 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4940 		    allocflags, false))
4941 			return (PAGE_NOT_ACQUIRED);
4942 	}
4943 	return (m);
4944 }
4945 
4946 /*
4947  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4948  * is not set.
4949  */
4950 vm_page_t
4951 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4952 {
4953 	vm_page_t m;
4954 
4955 	vm_page_grab_check(allocflags);
4956 	m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags);
4957 	if (m == PAGE_NOT_ACQUIRED)
4958 		return (NULL);
4959 	if (m != NULL)
4960 		return (m);
4961 
4962 	/*
4963 	 * The radix lockless lookup should never return a false negative
4964 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4965 	 * was no page present at the instant of the call.  A NOCREAT caller
4966 	 * must handle create races gracefully.
4967 	 */
4968 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4969 		return (NULL);
4970 
4971 	VM_OBJECT_WLOCK(object);
4972 	m = vm_page_grab(object, pindex, allocflags);
4973 	VM_OBJECT_WUNLOCK(object);
4974 
4975 	return (m);
4976 }
4977 
4978 /*
4979  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4980  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4981  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4982  * in simultaneously.  Additional pages will be left on a paging queue but
4983  * will neither be wired nor busy regardless of allocflags.
4984  */
4985 int
4986 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4987 {
4988 	vm_page_t m;
4989 	vm_page_t ma[VM_INITIAL_PAGEIN];
4990 	int after, i, pflags, rv;
4991 
4992 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4993 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4994 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4995 	KASSERT((allocflags &
4996 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4997 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4998 	VM_OBJECT_ASSERT_WLOCKED(object);
4999 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
5000 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
5001 	pflags |= VM_ALLOC_WAITFAIL;
5002 
5003 retrylookup:
5004 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
5005 		/*
5006 		 * If the page is fully valid it can only become invalid
5007 		 * with the object lock held.  If it is not valid it can
5008 		 * become valid with the busy lock held.  Therefore, we
5009 		 * may unnecessarily lock the exclusive busy here if we
5010 		 * race with I/O completion not using the object lock.
5011 		 * However, we will not end up with an invalid page and a
5012 		 * shared lock.
5013 		 */
5014 		if (!vm_page_trybusy(m,
5015 		    vm_page_all_valid(m) ? allocflags : 0)) {
5016 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
5017 			    allocflags, true);
5018 			goto retrylookup;
5019 		}
5020 		if (vm_page_all_valid(m))
5021 			goto out;
5022 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5023 			vm_page_busy_release(m);
5024 			*mp = NULL;
5025 			return (VM_PAGER_FAIL);
5026 		}
5027 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5028 		*mp = NULL;
5029 		return (VM_PAGER_FAIL);
5030 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
5031 		if (!vm_pager_can_alloc_page(object, pindex)) {
5032 			*mp = NULL;
5033 			return (VM_PAGER_AGAIN);
5034 		}
5035 		goto retrylookup;
5036 	}
5037 
5038 	vm_page_assert_xbusied(m);
5039 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
5040 		after = MIN(after, VM_INITIAL_PAGEIN);
5041 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
5042 		after = MAX(after, 1);
5043 		ma[0] = m;
5044 		for (i = 1; i < after; i++) {
5045 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
5046 				if (vm_page_any_valid(ma[i]) ||
5047 				    !vm_page_tryxbusy(ma[i]))
5048 					break;
5049 			} else {
5050 				ma[i] = vm_page_alloc(object, m->pindex + i,
5051 				    VM_ALLOC_NORMAL);
5052 				if (ma[i] == NULL)
5053 					break;
5054 			}
5055 		}
5056 		after = i;
5057 		vm_object_pip_add(object, after);
5058 		VM_OBJECT_WUNLOCK(object);
5059 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
5060 		VM_OBJECT_WLOCK(object);
5061 		vm_object_pip_wakeupn(object, after);
5062 		/* Pager may have replaced a page. */
5063 		m = ma[0];
5064 		if (rv != VM_PAGER_OK) {
5065 			for (i = 0; i < after; i++) {
5066 				if (!vm_page_wired(ma[i]))
5067 					vm_page_free(ma[i]);
5068 				else
5069 					vm_page_xunbusy(ma[i]);
5070 			}
5071 			*mp = NULL;
5072 			return (rv);
5073 		}
5074 		for (i = 1; i < after; i++)
5075 			vm_page_readahead_finish(ma[i]);
5076 		MPASS(vm_page_all_valid(m));
5077 	} else {
5078 		vm_page_zero_invalid(m, TRUE);
5079 	}
5080 out:
5081 	if ((allocflags & VM_ALLOC_WIRED) != 0)
5082 		vm_page_wire(m);
5083 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
5084 		vm_page_busy_downgrade(m);
5085 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
5086 		vm_page_busy_release(m);
5087 	*mp = m;
5088 	return (VM_PAGER_OK);
5089 }
5090 
5091 /*
5092  * Locklessly grab a valid page.  If the page is not valid or not yet
5093  * allocated this will fall back to the object lock method.
5094  */
5095 int
5096 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
5097     vm_pindex_t pindex, int allocflags)
5098 {
5099 	vm_page_t m;
5100 	int flags;
5101 	int error;
5102 
5103 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
5104 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
5105 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
5106 	    "mismatch"));
5107 	KASSERT((allocflags &
5108 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
5109 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
5110 
5111 	/*
5112 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
5113 	 * before we can inspect the valid field and return a wired page.
5114 	 */
5115 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
5116 	vm_page_grab_check(flags);
5117 	m = vm_page_acquire_unlocked(object, pindex, NULL, flags);
5118 	if (m == PAGE_NOT_ACQUIRED)
5119 		return (VM_PAGER_FAIL);
5120 	if (m != NULL) {
5121 		if (vm_page_all_valid(m)) {
5122 			if ((allocflags & VM_ALLOC_WIRED) != 0)
5123 				vm_page_wire(m);
5124 			vm_page_grab_release(m, allocflags);
5125 			*mp = m;
5126 			return (VM_PAGER_OK);
5127 		}
5128 		vm_page_busy_release(m);
5129 	}
5130 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5131 		*mp = NULL;
5132 		return (VM_PAGER_FAIL);
5133 	}
5134 	VM_OBJECT_WLOCK(object);
5135 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
5136 	VM_OBJECT_WUNLOCK(object);
5137 
5138 	return (error);
5139 }
5140 
5141 /*
5142  * Return the specified range of pages from the given object.  For each
5143  * page offset within the range, if a page already exists within the object
5144  * at that offset and it is busy, then wait for it to change state.  If,
5145  * instead, the page doesn't exist, then allocate it.
5146  *
5147  * The caller must always specify an allocation class.
5148  *
5149  * allocation classes:
5150  *	VM_ALLOC_NORMAL		normal process request
5151  *	VM_ALLOC_SYSTEM		system *really* needs the pages
5152  *
5153  * The caller must always specify that the pages are to be busied and/or
5154  * wired.
5155  *
5156  * optional allocation flags:
5157  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5158  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5159  *	VM_ALLOC_NOWAIT		do not sleep
5160  *	VM_ALLOC_SBUSY		set page to sbusy state
5161  *	VM_ALLOC_WIRED		wire the pages
5162  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5163  *
5164  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5165  * may return a partial prefix of the requested range.
5166  */
5167 int
5168 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5169     vm_page_t *ma, int count)
5170 {
5171 	vm_page_t m, mpred;
5172 	int pflags;
5173 	int i;
5174 
5175 	VM_OBJECT_ASSERT_WLOCKED(object);
5176 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5177 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5178 	KASSERT(count > 0,
5179 	    ("vm_page_grab_pages: invalid page count %d", count));
5180 	vm_page_grab_check(allocflags);
5181 
5182 	pflags = vm_page_grab_pflags(allocflags);
5183 	i = 0;
5184 retrylookup:
5185 	m = vm_page_mpred(object, pindex + i);
5186 	if (m == NULL || m->pindex != pindex + i) {
5187 		mpred = m;
5188 		m = NULL;
5189 	} else
5190 		mpred = TAILQ_PREV(m, pglist, listq);
5191 	for (; i < count; i++) {
5192 		if (m != NULL) {
5193 			if (!vm_page_tryacquire(m, allocflags)) {
5194 				if (vm_page_grab_sleep(object, m, pindex + i,
5195 				    "grbmaw", allocflags, true))
5196 					goto retrylookup;
5197 				break;
5198 			}
5199 		} else {
5200 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5201 				break;
5202 			m = vm_page_alloc_after(object, pindex + i,
5203 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5204 			if (m == NULL) {
5205 				if ((allocflags & (VM_ALLOC_NOWAIT |
5206 				    VM_ALLOC_WAITFAIL)) != 0)
5207 					break;
5208 				goto retrylookup;
5209 			}
5210 		}
5211 		if (vm_page_none_valid(m) &&
5212 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5213 			if ((m->flags & PG_ZERO) == 0)
5214 				pmap_zero_page(m);
5215 			vm_page_valid(m);
5216 		}
5217 		vm_page_grab_release(m, allocflags);
5218 		ma[i] = mpred = m;
5219 		m = vm_page_next(m);
5220 	}
5221 	return (i);
5222 }
5223 
5224 /*
5225  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5226  * and will fall back to the locked variant to handle allocation.
5227  */
5228 int
5229 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5230     int allocflags, vm_page_t *ma, int count)
5231 {
5232 	vm_page_t m;
5233 	int flags;
5234 	int i;
5235 
5236 	KASSERT(count > 0,
5237 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5238 	vm_page_grab_check(allocflags);
5239 
5240 	/*
5241 	 * Modify flags for lockless acquire to hold the page until we
5242 	 * set it valid if necessary.
5243 	 */
5244 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5245 	vm_page_grab_check(flags);
5246 	m = NULL;
5247 	for (i = 0; i < count; i++, pindex++) {
5248 		/*
5249 		 * We may see a false NULL here because the previous page has
5250 		 * been removed or just inserted and the list is loaded without
5251 		 * barriers.  Switch to radix to verify.
5252 		 */
5253 		if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex ||
5254 		    atomic_load_ptr(&m->object) != object) {
5255 			/*
5256 			 * This guarantees the result is instantaneously
5257 			 * correct.
5258 			 */
5259 			m = NULL;
5260 		}
5261 		m = vm_page_acquire_unlocked(object, pindex, m, flags);
5262 		if (m == PAGE_NOT_ACQUIRED)
5263 			return (i);
5264 		if (m == NULL)
5265 			break;
5266 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5267 			if ((m->flags & PG_ZERO) == 0)
5268 				pmap_zero_page(m);
5269 			vm_page_valid(m);
5270 		}
5271 		/* m will still be wired or busy according to flags. */
5272 		vm_page_grab_release(m, allocflags);
5273 		ma[i] = m;
5274 		m = TAILQ_NEXT(m, listq);
5275 	}
5276 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5277 		return (i);
5278 	count -= i;
5279 	VM_OBJECT_WLOCK(object);
5280 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5281 	VM_OBJECT_WUNLOCK(object);
5282 
5283 	return (i);
5284 }
5285 
5286 /*
5287  * Mapping function for valid or dirty bits in a page.
5288  *
5289  * Inputs are required to range within a page.
5290  */
5291 vm_page_bits_t
5292 vm_page_bits(int base, int size)
5293 {
5294 	int first_bit;
5295 	int last_bit;
5296 
5297 	KASSERT(
5298 	    base + size <= PAGE_SIZE,
5299 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5300 	);
5301 
5302 	if (size == 0)		/* handle degenerate case */
5303 		return (0);
5304 
5305 	first_bit = base >> DEV_BSHIFT;
5306 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5307 
5308 	return (((vm_page_bits_t)2 << last_bit) -
5309 	    ((vm_page_bits_t)1 << first_bit));
5310 }
5311 
5312 void
5313 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5314 {
5315 
5316 #if PAGE_SIZE == 32768
5317 	atomic_set_64((uint64_t *)bits, set);
5318 #elif PAGE_SIZE == 16384
5319 	atomic_set_32((uint32_t *)bits, set);
5320 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5321 	atomic_set_16((uint16_t *)bits, set);
5322 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5323 	atomic_set_8((uint8_t *)bits, set);
5324 #else		/* PAGE_SIZE <= 8192 */
5325 	uintptr_t addr;
5326 	int shift;
5327 
5328 	addr = (uintptr_t)bits;
5329 	/*
5330 	 * Use a trick to perform a 32-bit atomic on the
5331 	 * containing aligned word, to not depend on the existence
5332 	 * of atomic_{set, clear}_{8, 16}.
5333 	 */
5334 	shift = addr & (sizeof(uint32_t) - 1);
5335 #if BYTE_ORDER == BIG_ENDIAN
5336 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5337 #else
5338 	shift *= NBBY;
5339 #endif
5340 	addr &= ~(sizeof(uint32_t) - 1);
5341 	atomic_set_32((uint32_t *)addr, set << shift);
5342 #endif		/* PAGE_SIZE */
5343 }
5344 
5345 static inline void
5346 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5347 {
5348 
5349 #if PAGE_SIZE == 32768
5350 	atomic_clear_64((uint64_t *)bits, clear);
5351 #elif PAGE_SIZE == 16384
5352 	atomic_clear_32((uint32_t *)bits, clear);
5353 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5354 	atomic_clear_16((uint16_t *)bits, clear);
5355 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5356 	atomic_clear_8((uint8_t *)bits, clear);
5357 #else		/* PAGE_SIZE <= 8192 */
5358 	uintptr_t addr;
5359 	int shift;
5360 
5361 	addr = (uintptr_t)bits;
5362 	/*
5363 	 * Use a trick to perform a 32-bit atomic on the
5364 	 * containing aligned word, to not depend on the existence
5365 	 * of atomic_{set, clear}_{8, 16}.
5366 	 */
5367 	shift = addr & (sizeof(uint32_t) - 1);
5368 #if BYTE_ORDER == BIG_ENDIAN
5369 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5370 #else
5371 	shift *= NBBY;
5372 #endif
5373 	addr &= ~(sizeof(uint32_t) - 1);
5374 	atomic_clear_32((uint32_t *)addr, clear << shift);
5375 #endif		/* PAGE_SIZE */
5376 }
5377 
5378 static inline vm_page_bits_t
5379 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5380 {
5381 #if PAGE_SIZE == 32768
5382 	uint64_t old;
5383 
5384 	old = *bits;
5385 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5386 	return (old);
5387 #elif PAGE_SIZE == 16384
5388 	uint32_t old;
5389 
5390 	old = *bits;
5391 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5392 	return (old);
5393 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5394 	uint16_t old;
5395 
5396 	old = *bits;
5397 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5398 	return (old);
5399 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5400 	uint8_t old;
5401 
5402 	old = *bits;
5403 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5404 	return (old);
5405 #else		/* PAGE_SIZE <= 4096*/
5406 	uintptr_t addr;
5407 	uint32_t old, new, mask;
5408 	int shift;
5409 
5410 	addr = (uintptr_t)bits;
5411 	/*
5412 	 * Use a trick to perform a 32-bit atomic on the
5413 	 * containing aligned word, to not depend on the existence
5414 	 * of atomic_{set, swap, clear}_{8, 16}.
5415 	 */
5416 	shift = addr & (sizeof(uint32_t) - 1);
5417 #if BYTE_ORDER == BIG_ENDIAN
5418 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5419 #else
5420 	shift *= NBBY;
5421 #endif
5422 	addr &= ~(sizeof(uint32_t) - 1);
5423 	mask = VM_PAGE_BITS_ALL << shift;
5424 
5425 	old = *bits;
5426 	do {
5427 		new = old & ~mask;
5428 		new |= newbits << shift;
5429 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5430 	return (old >> shift);
5431 #endif		/* PAGE_SIZE */
5432 }
5433 
5434 /*
5435  *	vm_page_set_valid_range:
5436  *
5437  *	Sets portions of a page valid.  The arguments are expected
5438  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5439  *	of any partial chunks touched by the range.  The invalid portion of
5440  *	such chunks will be zeroed.
5441  *
5442  *	(base + size) must be less then or equal to PAGE_SIZE.
5443  */
5444 void
5445 vm_page_set_valid_range(vm_page_t m, int base, int size)
5446 {
5447 	int endoff, frag;
5448 	vm_page_bits_t pagebits;
5449 
5450 	vm_page_assert_busied(m);
5451 	if (size == 0)	/* handle degenerate case */
5452 		return;
5453 
5454 	/*
5455 	 * If the base is not DEV_BSIZE aligned and the valid
5456 	 * bit is clear, we have to zero out a portion of the
5457 	 * first block.
5458 	 */
5459 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5460 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5461 		pmap_zero_page_area(m, frag, base - frag);
5462 
5463 	/*
5464 	 * If the ending offset is not DEV_BSIZE aligned and the
5465 	 * valid bit is clear, we have to zero out a portion of
5466 	 * the last block.
5467 	 */
5468 	endoff = base + size;
5469 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5470 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5471 		pmap_zero_page_area(m, endoff,
5472 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5473 
5474 	/*
5475 	 * Assert that no previously invalid block that is now being validated
5476 	 * is already dirty.
5477 	 */
5478 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5479 	    ("vm_page_set_valid_range: page %p is dirty", m));
5480 
5481 	/*
5482 	 * Set valid bits inclusive of any overlap.
5483 	 */
5484 	pagebits = vm_page_bits(base, size);
5485 	if (vm_page_xbusied(m))
5486 		m->valid |= pagebits;
5487 	else
5488 		vm_page_bits_set(m, &m->valid, pagebits);
5489 }
5490 
5491 /*
5492  * Set the page dirty bits and free the invalid swap space if
5493  * present.  Returns the previous dirty bits.
5494  */
5495 vm_page_bits_t
5496 vm_page_set_dirty(vm_page_t m)
5497 {
5498 	vm_page_bits_t old;
5499 
5500 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5501 
5502 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5503 		old = m->dirty;
5504 		m->dirty = VM_PAGE_BITS_ALL;
5505 	} else
5506 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5507 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5508 		vm_pager_page_unswapped(m);
5509 
5510 	return (old);
5511 }
5512 
5513 /*
5514  * Clear the given bits from the specified page's dirty field.
5515  */
5516 static __inline void
5517 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5518 {
5519 
5520 	vm_page_assert_busied(m);
5521 
5522 	/*
5523 	 * If the page is xbusied and not write mapped we are the
5524 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5525 	 * layer can call vm_page_dirty() without holding a distinguished
5526 	 * lock.  The combination of page busy and atomic operations
5527 	 * suffice to guarantee consistency of the page dirty field.
5528 	 */
5529 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5530 		m->dirty &= ~pagebits;
5531 	else
5532 		vm_page_bits_clear(m, &m->dirty, pagebits);
5533 }
5534 
5535 /*
5536  *	vm_page_set_validclean:
5537  *
5538  *	Sets portions of a page valid and clean.  The arguments are expected
5539  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5540  *	of any partial chunks touched by the range.  The invalid portion of
5541  *	such chunks will be zero'd.
5542  *
5543  *	(base + size) must be less then or equal to PAGE_SIZE.
5544  */
5545 void
5546 vm_page_set_validclean(vm_page_t m, int base, int size)
5547 {
5548 	vm_page_bits_t oldvalid, pagebits;
5549 	int endoff, frag;
5550 
5551 	vm_page_assert_busied(m);
5552 	if (size == 0)	/* handle degenerate case */
5553 		return;
5554 
5555 	/*
5556 	 * If the base is not DEV_BSIZE aligned and the valid
5557 	 * bit is clear, we have to zero out a portion of the
5558 	 * first block.
5559 	 */
5560 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5561 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5562 		pmap_zero_page_area(m, frag, base - frag);
5563 
5564 	/*
5565 	 * If the ending offset is not DEV_BSIZE aligned and the
5566 	 * valid bit is clear, we have to zero out a portion of
5567 	 * the last block.
5568 	 */
5569 	endoff = base + size;
5570 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5571 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5572 		pmap_zero_page_area(m, endoff,
5573 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5574 
5575 	/*
5576 	 * Set valid, clear dirty bits.  If validating the entire
5577 	 * page we can safely clear the pmap modify bit.  We also
5578 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5579 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5580 	 * be set again.
5581 	 *
5582 	 * We set valid bits inclusive of any overlap, but we can only
5583 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5584 	 * the range.
5585 	 */
5586 	oldvalid = m->valid;
5587 	pagebits = vm_page_bits(base, size);
5588 	if (vm_page_xbusied(m))
5589 		m->valid |= pagebits;
5590 	else
5591 		vm_page_bits_set(m, &m->valid, pagebits);
5592 #if 0	/* NOT YET */
5593 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5594 		frag = DEV_BSIZE - frag;
5595 		base += frag;
5596 		size -= frag;
5597 		if (size < 0)
5598 			size = 0;
5599 	}
5600 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5601 #endif
5602 	if (base == 0 && size == PAGE_SIZE) {
5603 		/*
5604 		 * The page can only be modified within the pmap if it is
5605 		 * mapped, and it can only be mapped if it was previously
5606 		 * fully valid.
5607 		 */
5608 		if (oldvalid == VM_PAGE_BITS_ALL)
5609 			/*
5610 			 * Perform the pmap_clear_modify() first.  Otherwise,
5611 			 * a concurrent pmap operation, such as
5612 			 * pmap_protect(), could clear a modification in the
5613 			 * pmap and set the dirty field on the page before
5614 			 * pmap_clear_modify() had begun and after the dirty
5615 			 * field was cleared here.
5616 			 */
5617 			pmap_clear_modify(m);
5618 		m->dirty = 0;
5619 		vm_page_aflag_clear(m, PGA_NOSYNC);
5620 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5621 		m->dirty &= ~pagebits;
5622 	else
5623 		vm_page_clear_dirty_mask(m, pagebits);
5624 }
5625 
5626 void
5627 vm_page_clear_dirty(vm_page_t m, int base, int size)
5628 {
5629 
5630 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5631 }
5632 
5633 /*
5634  *	vm_page_set_invalid:
5635  *
5636  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5637  *	valid and dirty bits for the effected areas are cleared.
5638  */
5639 void
5640 vm_page_set_invalid(vm_page_t m, int base, int size)
5641 {
5642 	vm_page_bits_t bits;
5643 	vm_object_t object;
5644 
5645 	/*
5646 	 * The object lock is required so that pages can't be mapped
5647 	 * read-only while we're in the process of invalidating them.
5648 	 */
5649 	object = m->object;
5650 	VM_OBJECT_ASSERT_WLOCKED(object);
5651 	vm_page_assert_busied(m);
5652 
5653 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5654 	    size >= object->un_pager.vnp.vnp_size)
5655 		bits = VM_PAGE_BITS_ALL;
5656 	else
5657 		bits = vm_page_bits(base, size);
5658 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5659 		pmap_remove_all(m);
5660 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5661 	    !pmap_page_is_mapped(m),
5662 	    ("vm_page_set_invalid: page %p is mapped", m));
5663 	if (vm_page_xbusied(m)) {
5664 		m->valid &= ~bits;
5665 		m->dirty &= ~bits;
5666 	} else {
5667 		vm_page_bits_clear(m, &m->valid, bits);
5668 		vm_page_bits_clear(m, &m->dirty, bits);
5669 	}
5670 }
5671 
5672 /*
5673  *	vm_page_invalid:
5674  *
5675  *	Invalidates the entire page.  The page must be busy, unmapped, and
5676  *	the enclosing object must be locked.  The object locks protects
5677  *	against concurrent read-only pmap enter which is done without
5678  *	busy.
5679  */
5680 void
5681 vm_page_invalid(vm_page_t m)
5682 {
5683 
5684 	vm_page_assert_busied(m);
5685 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5686 	MPASS(!pmap_page_is_mapped(m));
5687 
5688 	if (vm_page_xbusied(m))
5689 		m->valid = 0;
5690 	else
5691 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5692 }
5693 
5694 /*
5695  * vm_page_zero_invalid()
5696  *
5697  *	The kernel assumes that the invalid portions of a page contain
5698  *	garbage, but such pages can be mapped into memory by user code.
5699  *	When this occurs, we must zero out the non-valid portions of the
5700  *	page so user code sees what it expects.
5701  *
5702  *	Pages are most often semi-valid when the end of a file is mapped
5703  *	into memory and the file's size is not page aligned.
5704  */
5705 void
5706 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5707 {
5708 	int b;
5709 	int i;
5710 
5711 	/*
5712 	 * Scan the valid bits looking for invalid sections that
5713 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5714 	 * valid bit may be set ) have already been zeroed by
5715 	 * vm_page_set_validclean().
5716 	 */
5717 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5718 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5719 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5720 			if (i > b) {
5721 				pmap_zero_page_area(m,
5722 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5723 			}
5724 			b = i + 1;
5725 		}
5726 	}
5727 
5728 	/*
5729 	 * setvalid is TRUE when we can safely set the zero'd areas
5730 	 * as being valid.  We can do this if there are no cache consistency
5731 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5732 	 */
5733 	if (setvalid)
5734 		vm_page_valid(m);
5735 }
5736 
5737 /*
5738  *	vm_page_is_valid:
5739  *
5740  *	Is (partial) page valid?  Note that the case where size == 0
5741  *	will return FALSE in the degenerate case where the page is
5742  *	entirely invalid, and TRUE otherwise.
5743  *
5744  *	Some callers envoke this routine without the busy lock held and
5745  *	handle races via higher level locks.  Typical callers should
5746  *	hold a busy lock to prevent invalidation.
5747  */
5748 int
5749 vm_page_is_valid(vm_page_t m, int base, int size)
5750 {
5751 	vm_page_bits_t bits;
5752 
5753 	bits = vm_page_bits(base, size);
5754 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5755 }
5756 
5757 /*
5758  * Returns true if all of the specified predicates are true for the entire
5759  * (super)page and false otherwise.
5760  */
5761 bool
5762 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5763 {
5764 	vm_object_t object;
5765 	int i, npages;
5766 
5767 	object = m->object;
5768 	if (skip_m != NULL && skip_m->object != object)
5769 		return (false);
5770 	VM_OBJECT_ASSERT_LOCKED(object);
5771 	KASSERT(psind <= m->psind,
5772 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5773 	npages = atop(pagesizes[psind]);
5774 
5775 	/*
5776 	 * The physically contiguous pages that make up a superpage, i.e., a
5777 	 * page with a page size index ("psind") greater than zero, will
5778 	 * occupy adjacent entries in vm_page_array[].
5779 	 */
5780 	for (i = 0; i < npages; i++) {
5781 		/* Always test object consistency, including "skip_m". */
5782 		if (m[i].object != object)
5783 			return (false);
5784 		if (&m[i] == skip_m)
5785 			continue;
5786 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5787 			return (false);
5788 		if ((flags & PS_ALL_DIRTY) != 0) {
5789 			/*
5790 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5791 			 * might stop this case from spuriously returning
5792 			 * "false".  However, that would require a write lock
5793 			 * on the object containing "m[i]".
5794 			 */
5795 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5796 				return (false);
5797 		}
5798 		if ((flags & PS_ALL_VALID) != 0 &&
5799 		    m[i].valid != VM_PAGE_BITS_ALL)
5800 			return (false);
5801 	}
5802 	return (true);
5803 }
5804 
5805 /*
5806  * Set the page's dirty bits if the page is modified.
5807  */
5808 void
5809 vm_page_test_dirty(vm_page_t m)
5810 {
5811 
5812 	vm_page_assert_busied(m);
5813 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5814 		vm_page_dirty(m);
5815 }
5816 
5817 void
5818 vm_page_valid(vm_page_t m)
5819 {
5820 
5821 	vm_page_assert_busied(m);
5822 	if (vm_page_xbusied(m))
5823 		m->valid = VM_PAGE_BITS_ALL;
5824 	else
5825 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5826 }
5827 
5828 void
5829 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5830 {
5831 
5832 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5833 }
5834 
5835 void
5836 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5837 {
5838 
5839 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5840 }
5841 
5842 int
5843 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5844 {
5845 
5846 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5847 }
5848 
5849 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5850 void
5851 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5852 {
5853 
5854 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5855 }
5856 
5857 void
5858 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5859 {
5860 
5861 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5862 }
5863 #endif
5864 
5865 #ifdef INVARIANTS
5866 void
5867 vm_page_object_busy_assert(vm_page_t m)
5868 {
5869 
5870 	/*
5871 	 * Certain of the page's fields may only be modified by the
5872 	 * holder of a page or object busy.
5873 	 */
5874 	if (m->object != NULL && !vm_page_busied(m))
5875 		VM_OBJECT_ASSERT_BUSY(m->object);
5876 }
5877 
5878 void
5879 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5880 {
5881 
5882 	if ((bits & PGA_WRITEABLE) == 0)
5883 		return;
5884 
5885 	/*
5886 	 * The PGA_WRITEABLE flag can only be set if the page is
5887 	 * managed, is exclusively busied or the object is locked.
5888 	 * Currently, this flag is only set by pmap_enter().
5889 	 */
5890 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5891 	    ("PGA_WRITEABLE on unmanaged page"));
5892 	if (!vm_page_xbusied(m))
5893 		VM_OBJECT_ASSERT_BUSY(m->object);
5894 }
5895 #endif
5896 
5897 #include "opt_ddb.h"
5898 #ifdef DDB
5899 #include <sys/kernel.h>
5900 
5901 #include <ddb/ddb.h>
5902 
5903 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5904 {
5905 
5906 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5907 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5908 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5909 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5910 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5911 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5912 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5913 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5914 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5915 }
5916 
5917 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5918 {
5919 	int dom;
5920 
5921 	db_printf("pq_free %d\n", vm_free_count());
5922 	for (dom = 0; dom < vm_ndomains; dom++) {
5923 		db_printf(
5924     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5925 		    dom,
5926 		    vm_dom[dom].vmd_page_count,
5927 		    vm_dom[dom].vmd_free_count,
5928 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5929 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5930 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5931 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5932 	}
5933 }
5934 
5935 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5936 {
5937 	vm_page_t m;
5938 	boolean_t phys, virt;
5939 
5940 	if (!have_addr) {
5941 		db_printf("show pginfo addr\n");
5942 		return;
5943 	}
5944 
5945 	phys = strchr(modif, 'p') != NULL;
5946 	virt = strchr(modif, 'v') != NULL;
5947 	if (virt)
5948 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5949 	else if (phys)
5950 		m = PHYS_TO_VM_PAGE(addr);
5951 	else
5952 		m = (vm_page_t)addr;
5953 	db_printf(
5954     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5955     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5956 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5957 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5958 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5959 }
5960 #endif /* DDB */
5961