xref: /freebsd/sys/vm/vm_page.c (revision a0ca4af9455b844c5e094fc1b09b1390ffa979fc)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Resident memory management module.
65  */
66 
67 #include <sys/cdefs.h>
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/counter.h>
73 #include <sys/domainset.h>
74 #include <sys/kernel.h>
75 #include <sys/limits.h>
76 #include <sys/linker.h>
77 #include <sys/lock.h>
78 #include <sys/malloc.h>
79 #include <sys/mman.h>
80 #include <sys/msgbuf.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/rwlock.h>
84 #include <sys/sleepqueue.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/smp.h>
88 #include <sys/sysctl.h>
89 #include <sys/vmmeter.h>
90 #include <sys/vnode.h>
91 
92 #include <vm/vm.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_param.h>
95 #include <vm/vm_domainset.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_dumpset.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 struct vm_domain vm_dom[MAXMEMDOM];
114 
115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
116 
117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
118 
119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
120 /* The following fields are protected by the domainset lock. */
121 domainset_t __exclusive_cache_line vm_min_domains;
122 domainset_t __exclusive_cache_line vm_severe_domains;
123 static int vm_min_waiters;
124 static int vm_severe_waiters;
125 static int vm_pageproc_waiters;
126 
127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
128     "VM page statistics");
129 
130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
132     CTLFLAG_RD, &pqstate_commit_retries,
133     "Number of failed per-page atomic queue state updates");
134 
135 static COUNTER_U64_DEFINE_EARLY(queue_ops);
136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
137     CTLFLAG_RD, &queue_ops,
138     "Number of batched queue operations");
139 
140 static COUNTER_U64_DEFINE_EARLY(queue_nops);
141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
142     CTLFLAG_RD, &queue_nops,
143     "Number of batched queue operations with no effects");
144 
145 /*
146  * bogus page -- for I/O to/from partially complete buffers,
147  * or for paging into sparsely invalid regions.
148  */
149 vm_page_t bogus_page;
150 
151 vm_page_t vm_page_array;
152 long vm_page_array_size;
153 long first_page;
154 
155 struct bitset *vm_page_dump;
156 long vm_page_dump_pages;
157 
158 static TAILQ_HEAD(, vm_page) blacklist_head;
159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
161     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
162 
163 static uma_zone_t fakepg_zone;
164 
165 static void vm_page_alloc_check(vm_page_t m);
166 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
167     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
168 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
169 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
170 static bool vm_page_free_prep(vm_page_t m);
171 static void vm_page_free_toq(vm_page_t m);
172 static void vm_page_init(void *dummy);
173 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
174     vm_pindex_t pindex, vm_page_t mpred);
175 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
176     vm_page_t mpred);
177 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
178     const uint16_t nflag);
179 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
180     vm_page_t m_run, vm_paddr_t high);
181 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
182 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
183     int req);
184 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
185     int flags);
186 static void vm_page_zone_release(void *arg, void **store, int cnt);
187 
188 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
189 
190 static void
191 vm_page_init(void *dummy)
192 {
193 
194 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
195 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
196 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED);
197 }
198 
199 static int pgcache_zone_max_pcpu;
200 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
201     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
202     "Per-CPU page cache size");
203 
204 /*
205  * The cache page zone is initialized later since we need to be able to allocate
206  * pages before UMA is fully initialized.
207  */
208 static void
209 vm_page_init_cache_zones(void *dummy __unused)
210 {
211 	struct vm_domain *vmd;
212 	struct vm_pgcache *pgcache;
213 	int cache, domain, maxcache, pool;
214 
215 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
216 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
217 	for (domain = 0; domain < vm_ndomains; domain++) {
218 		vmd = VM_DOMAIN(domain);
219 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
220 			pgcache = &vmd->vmd_pgcache[pool];
221 			pgcache->domain = domain;
222 			pgcache->pool = pool;
223 			pgcache->zone = uma_zcache_create("vm pgcache",
224 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
225 			    vm_page_zone_import, vm_page_zone_release, pgcache,
226 			    UMA_ZONE_VM);
227 
228 			/*
229 			 * Limit each pool's zone to 0.1% of the pages in the
230 			 * domain.
231 			 */
232 			cache = maxcache != 0 ? maxcache :
233 			    vmd->vmd_page_count / 1000;
234 			uma_zone_set_maxcache(pgcache->zone, cache);
235 		}
236 	}
237 }
238 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
239 
240 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
241 #if PAGE_SIZE == 32768
242 #ifdef CTASSERT
243 CTASSERT(sizeof(u_long) >= 8);
244 #endif
245 #endif
246 
247 /*
248  *	vm_set_page_size:
249  *
250  *	Sets the page size, perhaps based upon the memory
251  *	size.  Must be called before any use of page-size
252  *	dependent functions.
253  */
254 void
255 vm_set_page_size(void)
256 {
257 	if (vm_cnt.v_page_size == 0)
258 		vm_cnt.v_page_size = PAGE_SIZE;
259 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
260 		panic("vm_set_page_size: page size not a power of two");
261 }
262 
263 /*
264  *	vm_page_blacklist_next:
265  *
266  *	Find the next entry in the provided string of blacklist
267  *	addresses.  Entries are separated by space, comma, or newline.
268  *	If an invalid integer is encountered then the rest of the
269  *	string is skipped.  Updates the list pointer to the next
270  *	character, or NULL if the string is exhausted or invalid.
271  */
272 static vm_paddr_t
273 vm_page_blacklist_next(char **list, char *end)
274 {
275 	vm_paddr_t bad;
276 	char *cp, *pos;
277 
278 	if (list == NULL || *list == NULL)
279 		return (0);
280 	if (**list =='\0') {
281 		*list = NULL;
282 		return (0);
283 	}
284 
285 	/*
286 	 * If there's no end pointer then the buffer is coming from
287 	 * the kenv and we know it's null-terminated.
288 	 */
289 	if (end == NULL)
290 		end = *list + strlen(*list);
291 
292 	/* Ensure that strtoq() won't walk off the end */
293 	if (*end != '\0') {
294 		if (*end == '\n' || *end == ' ' || *end  == ',')
295 			*end = '\0';
296 		else {
297 			printf("Blacklist not terminated, skipping\n");
298 			*list = NULL;
299 			return (0);
300 		}
301 	}
302 
303 	for (pos = *list; *pos != '\0'; pos = cp) {
304 		bad = strtoq(pos, &cp, 0);
305 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
306 			if (bad == 0) {
307 				if (++cp < end)
308 					continue;
309 				else
310 					break;
311 			}
312 		} else
313 			break;
314 		if (*cp == '\0' || ++cp >= end)
315 			*list = NULL;
316 		else
317 			*list = cp;
318 		return (trunc_page(bad));
319 	}
320 	printf("Garbage in RAM blacklist, skipping\n");
321 	*list = NULL;
322 	return (0);
323 }
324 
325 bool
326 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
327 {
328 	struct vm_domain *vmd;
329 	vm_page_t m;
330 	bool found;
331 
332 	m = vm_phys_paddr_to_vm_page(pa);
333 	if (m == NULL)
334 		return (true); /* page does not exist, no failure */
335 
336 	vmd = vm_pagequeue_domain(m);
337 	vm_domain_free_lock(vmd);
338 	found = vm_phys_unfree_page(m);
339 	vm_domain_free_unlock(vmd);
340 	if (found) {
341 		vm_domain_freecnt_inc(vmd, -1);
342 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
343 		if (verbose)
344 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
345 	}
346 	return (found);
347 }
348 
349 /*
350  *	vm_page_blacklist_check:
351  *
352  *	Iterate through the provided string of blacklist addresses, pulling
353  *	each entry out of the physical allocator free list and putting it
354  *	onto a list for reporting via the vm.page_blacklist sysctl.
355  */
356 static void
357 vm_page_blacklist_check(char *list, char *end)
358 {
359 	vm_paddr_t pa;
360 	char *next;
361 
362 	next = list;
363 	while (next != NULL) {
364 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
365 			continue;
366 		vm_page_blacklist_add(pa, bootverbose);
367 	}
368 }
369 
370 /*
371  *	vm_page_blacklist_load:
372  *
373  *	Search for a special module named "ram_blacklist".  It'll be a
374  *	plain text file provided by the user via the loader directive
375  *	of the same name.
376  */
377 static void
378 vm_page_blacklist_load(char **list, char **end)
379 {
380 	void *mod;
381 	u_char *ptr;
382 	u_int len;
383 
384 	mod = NULL;
385 	ptr = NULL;
386 
387 	mod = preload_search_by_type("ram_blacklist");
388 	if (mod != NULL) {
389 		ptr = preload_fetch_addr(mod);
390 		len = preload_fetch_size(mod);
391         }
392 	*list = ptr;
393 	if (ptr != NULL)
394 		*end = ptr + len;
395 	else
396 		*end = NULL;
397 	return;
398 }
399 
400 static int
401 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
402 {
403 	vm_page_t m;
404 	struct sbuf sbuf;
405 	int error, first;
406 
407 	first = 1;
408 	error = sysctl_wire_old_buffer(req, 0);
409 	if (error != 0)
410 		return (error);
411 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
412 	TAILQ_FOREACH(m, &blacklist_head, listq) {
413 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
414 		    (uintmax_t)m->phys_addr);
415 		first = 0;
416 	}
417 	error = sbuf_finish(&sbuf);
418 	sbuf_delete(&sbuf);
419 	return (error);
420 }
421 
422 /*
423  * Initialize a dummy page for use in scans of the specified paging queue.
424  * In principle, this function only needs to set the flag PG_MARKER.
425  * Nonetheless, it write busies the page as a safety precaution.
426  */
427 void
428 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
429 {
430 
431 	bzero(marker, sizeof(*marker));
432 	marker->flags = PG_MARKER;
433 	marker->a.flags = aflags;
434 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
435 	marker->a.queue = queue;
436 }
437 
438 static void
439 vm_page_domain_init(int domain)
440 {
441 	struct vm_domain *vmd;
442 	struct vm_pagequeue *pq;
443 	int i;
444 
445 	vmd = VM_DOMAIN(domain);
446 	bzero(vmd, sizeof(*vmd));
447 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
448 	    "vm inactive pagequeue";
449 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
450 	    "vm active pagequeue";
451 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
452 	    "vm laundry pagequeue";
453 	*__DECONST(const char **,
454 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
455 	    "vm unswappable pagequeue";
456 	vmd->vmd_domain = domain;
457 	vmd->vmd_page_count = 0;
458 	vmd->vmd_free_count = 0;
459 	vmd->vmd_segs = 0;
460 	vmd->vmd_oom = FALSE;
461 	for (i = 0; i < PQ_COUNT; i++) {
462 		pq = &vmd->vmd_pagequeues[i];
463 		TAILQ_INIT(&pq->pq_pl);
464 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
465 		    MTX_DEF | MTX_DUPOK);
466 		pq->pq_pdpages = 0;
467 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
468 	}
469 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
470 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
471 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
472 
473 	/*
474 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
475 	 * insertions.
476 	 */
477 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
478 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
479 	    &vmd->vmd_inacthead, plinks.q);
480 
481 	/*
482 	 * The clock pages are used to implement active queue scanning without
483 	 * requeues.  Scans start at clock[0], which is advanced after the scan
484 	 * ends.  When the two clock hands meet, they are reset and scanning
485 	 * resumes from the head of the queue.
486 	 */
487 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
488 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
489 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
490 	    &vmd->vmd_clock[0], plinks.q);
491 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
492 	    &vmd->vmd_clock[1], plinks.q);
493 }
494 
495 /*
496  * Initialize a physical page in preparation for adding it to the free
497  * lists.
498  */
499 void
500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
501 {
502 
503 	m->object = NULL;
504 	m->ref_count = 0;
505 	m->busy_lock = VPB_FREED;
506 	m->flags = m->a.flags = 0;
507 	m->phys_addr = pa;
508 	m->a.queue = PQ_NONE;
509 	m->psind = 0;
510 	m->segind = segind;
511 	m->order = VM_NFREEORDER;
512 	m->pool = VM_FREEPOOL_DEFAULT;
513 	m->valid = m->dirty = 0;
514 	pmap_page_init(m);
515 }
516 
517 #ifndef PMAP_HAS_PAGE_ARRAY
518 static vm_paddr_t
519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
520 {
521 	vm_paddr_t new_end;
522 
523 	/*
524 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
525 	 * However, because this page is allocated from KVM, out-of-bounds
526 	 * accesses using the direct map will not be trapped.
527 	 */
528 	*vaddr += PAGE_SIZE;
529 
530 	/*
531 	 * Allocate physical memory for the page structures, and map it.
532 	 */
533 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
534 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
535 	    VM_PROT_READ | VM_PROT_WRITE);
536 	vm_page_array_size = page_range;
537 
538 	return (new_end);
539 }
540 #endif
541 
542 /*
543  *	vm_page_startup:
544  *
545  *	Initializes the resident memory module.  Allocates physical memory for
546  *	bootstrapping UMA and some data structures that are used to manage
547  *	physical pages.  Initializes these structures, and populates the free
548  *	page queues.
549  */
550 vm_offset_t
551 vm_page_startup(vm_offset_t vaddr)
552 {
553 	struct vm_phys_seg *seg;
554 	struct vm_domain *vmd;
555 	vm_page_t m;
556 	char *list, *listend;
557 	vm_paddr_t end, high_avail, low_avail, new_end, size;
558 	vm_paddr_t page_range __unused;
559 	vm_paddr_t last_pa, pa, startp, endp;
560 	u_long pagecount;
561 #if MINIDUMP_PAGE_TRACKING
562 	u_long vm_page_dump_size;
563 #endif
564 	int biggestone, i, segind;
565 #ifdef WITNESS
566 	vm_offset_t mapped;
567 	int witness_size;
568 #endif
569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
570 	long ii;
571 #endif
572 
573 	vaddr = round_page(vaddr);
574 
575 	vm_phys_early_startup();
576 	biggestone = vm_phys_avail_largest();
577 	end = phys_avail[biggestone+1];
578 
579 	/*
580 	 * Initialize the page and queue locks.
581 	 */
582 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
583 	for (i = 0; i < PA_LOCK_COUNT; i++)
584 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
585 	for (i = 0; i < vm_ndomains; i++)
586 		vm_page_domain_init(i);
587 
588 	new_end = end;
589 #ifdef WITNESS
590 	witness_size = round_page(witness_startup_count());
591 	new_end -= witness_size;
592 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
593 	    VM_PROT_READ | VM_PROT_WRITE);
594 	bzero((void *)mapped, witness_size);
595 	witness_startup((void *)mapped);
596 #endif
597 
598 #if MINIDUMP_PAGE_TRACKING
599 	/*
600 	 * Allocate a bitmap to indicate that a random physical page
601 	 * needs to be included in a minidump.
602 	 *
603 	 * The amd64 port needs this to indicate which direct map pages
604 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
605 	 *
606 	 * However, i386 still needs this workspace internally within the
607 	 * minidump code.  In theory, they are not needed on i386, but are
608 	 * included should the sf_buf code decide to use them.
609 	 */
610 	last_pa = 0;
611 	vm_page_dump_pages = 0;
612 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
613 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
614 		    dump_avail[i] / PAGE_SIZE;
615 		if (dump_avail[i + 1] > last_pa)
616 			last_pa = dump_avail[i + 1];
617 	}
618 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
619 	new_end -= vm_page_dump_size;
620 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
621 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
622 	bzero((void *)vm_page_dump, vm_page_dump_size);
623 #if MINIDUMP_STARTUP_PAGE_TRACKING
624 	/*
625 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
626 	 * in a crash dump.  When pmap_map() uses the direct map, they are
627 	 * not automatically included.
628 	 */
629 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
630 		dump_add_page(pa);
631 #endif
632 #else
633 	(void)last_pa;
634 #endif
635 	phys_avail[biggestone + 1] = new_end;
636 #ifdef __amd64__
637 	/*
638 	 * Request that the physical pages underlying the message buffer be
639 	 * included in a crash dump.  Since the message buffer is accessed
640 	 * through the direct map, they are not automatically included.
641 	 */
642 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
643 	last_pa = pa + round_page(msgbufsize);
644 	while (pa < last_pa) {
645 		dump_add_page(pa);
646 		pa += PAGE_SIZE;
647 	}
648 #endif
649 	/*
650 	 * Compute the number of pages of memory that will be available for
651 	 * use, taking into account the overhead of a page structure per page.
652 	 * In other words, solve
653 	 *	"available physical memory" - round_page(page_range *
654 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
655 	 * for page_range.
656 	 */
657 	low_avail = phys_avail[0];
658 	high_avail = phys_avail[1];
659 	for (i = 0; i < vm_phys_nsegs; i++) {
660 		if (vm_phys_segs[i].start < low_avail)
661 			low_avail = vm_phys_segs[i].start;
662 		if (vm_phys_segs[i].end > high_avail)
663 			high_avail = vm_phys_segs[i].end;
664 	}
665 	/* Skip the first chunk.  It is already accounted for. */
666 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
667 		if (phys_avail[i] < low_avail)
668 			low_avail = phys_avail[i];
669 		if (phys_avail[i + 1] > high_avail)
670 			high_avail = phys_avail[i + 1];
671 	}
672 	first_page = low_avail / PAGE_SIZE;
673 #ifdef VM_PHYSSEG_SPARSE
674 	size = 0;
675 	for (i = 0; i < vm_phys_nsegs; i++)
676 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
677 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
678 		size += phys_avail[i + 1] - phys_avail[i];
679 #elif defined(VM_PHYSSEG_DENSE)
680 	size = high_avail - low_avail;
681 #else
682 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
683 #endif
684 
685 #ifdef PMAP_HAS_PAGE_ARRAY
686 	pmap_page_array_startup(size / PAGE_SIZE);
687 	biggestone = vm_phys_avail_largest();
688 	end = new_end = phys_avail[biggestone + 1];
689 #else
690 #ifdef VM_PHYSSEG_DENSE
691 	/*
692 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
693 	 * the overhead of a page structure per page only if vm_page_array is
694 	 * allocated from the last physical memory chunk.  Otherwise, we must
695 	 * allocate page structures representing the physical memory
696 	 * underlying vm_page_array, even though they will not be used.
697 	 */
698 	if (new_end != high_avail)
699 		page_range = size / PAGE_SIZE;
700 	else
701 #endif
702 	{
703 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
704 
705 		/*
706 		 * If the partial bytes remaining are large enough for
707 		 * a page (PAGE_SIZE) without a corresponding
708 		 * 'struct vm_page', then new_end will contain an
709 		 * extra page after subtracting the length of the VM
710 		 * page array.  Compensate by subtracting an extra
711 		 * page from new_end.
712 		 */
713 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
714 			if (new_end == high_avail)
715 				high_avail -= PAGE_SIZE;
716 			new_end -= PAGE_SIZE;
717 		}
718 	}
719 	end = new_end;
720 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
721 #endif
722 
723 #if VM_NRESERVLEVEL > 0
724 	/*
725 	 * Allocate physical memory for the reservation management system's
726 	 * data structures, and map it.
727 	 */
728 	new_end = vm_reserv_startup(&vaddr, new_end);
729 #endif
730 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
731 	/*
732 	 * Include vm_page_array and vm_reserv_array in a crash dump.
733 	 */
734 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
735 		dump_add_page(pa);
736 #endif
737 	phys_avail[biggestone + 1] = new_end;
738 
739 	/*
740 	 * Add physical memory segments corresponding to the available
741 	 * physical pages.
742 	 */
743 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
744 		if (vm_phys_avail_size(i) != 0)
745 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
746 
747 	/*
748 	 * Initialize the physical memory allocator.
749 	 */
750 	vm_phys_init();
751 
752 	/*
753 	 * Initialize the page structures and add every available page to the
754 	 * physical memory allocator's free lists.
755 	 */
756 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
757 	for (ii = 0; ii < vm_page_array_size; ii++) {
758 		m = &vm_page_array[ii];
759 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
760 		m->flags = PG_FICTITIOUS;
761 	}
762 #endif
763 	vm_cnt.v_page_count = 0;
764 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
765 		seg = &vm_phys_segs[segind];
766 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
767 		    m++, pa += PAGE_SIZE)
768 			vm_page_init_page(m, pa, segind);
769 
770 		/*
771 		 * Add the segment's pages that are covered by one of
772 		 * phys_avail's ranges to the free lists.
773 		 */
774 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
775 			if (seg->end <= phys_avail[i] ||
776 			    seg->start >= phys_avail[i + 1])
777 				continue;
778 
779 			startp = MAX(seg->start, phys_avail[i]);
780 			endp = MIN(seg->end, phys_avail[i + 1]);
781 			pagecount = (u_long)atop(endp - startp);
782 			if (pagecount == 0)
783 				continue;
784 
785 			m = seg->first_page + atop(startp - seg->start);
786 			vmd = VM_DOMAIN(seg->domain);
787 			vm_domain_free_lock(vmd);
788 			vm_phys_enqueue_contig(m, pagecount);
789 			vm_domain_free_unlock(vmd);
790 			vm_domain_freecnt_inc(vmd, pagecount);
791 			vm_cnt.v_page_count += (u_int)pagecount;
792 			vmd->vmd_page_count += (u_int)pagecount;
793 			vmd->vmd_segs |= 1UL << segind;
794 		}
795 	}
796 
797 	/*
798 	 * Remove blacklisted pages from the physical memory allocator.
799 	 */
800 	TAILQ_INIT(&blacklist_head);
801 	vm_page_blacklist_load(&list, &listend);
802 	vm_page_blacklist_check(list, listend);
803 
804 	list = kern_getenv("vm.blacklist");
805 	vm_page_blacklist_check(list, NULL);
806 
807 	freeenv(list);
808 #if VM_NRESERVLEVEL > 0
809 	/*
810 	 * Initialize the reservation management system.
811 	 */
812 	vm_reserv_init();
813 #endif
814 
815 	return (vaddr);
816 }
817 
818 void
819 vm_page_reference(vm_page_t m)
820 {
821 
822 	vm_page_aflag_set(m, PGA_REFERENCED);
823 }
824 
825 /*
826  *	vm_page_trybusy
827  *
828  *	Helper routine for grab functions to trylock busy.
829  *
830  *	Returns true on success and false on failure.
831  */
832 static bool
833 vm_page_trybusy(vm_page_t m, int allocflags)
834 {
835 
836 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
837 		return (vm_page_trysbusy(m));
838 	else
839 		return (vm_page_tryxbusy(m));
840 }
841 
842 /*
843  *	vm_page_tryacquire
844  *
845  *	Helper routine for grab functions to trylock busy and wire.
846  *
847  *	Returns true on success and false on failure.
848  */
849 static inline bool
850 vm_page_tryacquire(vm_page_t m, int allocflags)
851 {
852 	bool locked;
853 
854 	locked = vm_page_trybusy(m, allocflags);
855 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
856 		vm_page_wire(m);
857 	return (locked);
858 }
859 
860 /*
861  *	vm_page_busy_acquire:
862  *
863  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
864  *	and drop the object lock if necessary.
865  */
866 bool
867 vm_page_busy_acquire(vm_page_t m, int allocflags)
868 {
869 	vm_object_t obj;
870 	bool locked;
871 
872 	/*
873 	 * The page-specific object must be cached because page
874 	 * identity can change during the sleep, causing the
875 	 * re-lock of a different object.
876 	 * It is assumed that a reference to the object is already
877 	 * held by the callers.
878 	 */
879 	obj = atomic_load_ptr(&m->object);
880 	for (;;) {
881 		if (vm_page_tryacquire(m, allocflags))
882 			return (true);
883 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
884 			return (false);
885 		if (obj != NULL)
886 			locked = VM_OBJECT_WOWNED(obj);
887 		else
888 			locked = false;
889 		MPASS(locked || vm_page_wired(m));
890 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
891 		    locked) && locked)
892 			VM_OBJECT_WLOCK(obj);
893 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
894 			return (false);
895 		KASSERT(m->object == obj || m->object == NULL,
896 		    ("vm_page_busy_acquire: page %p does not belong to %p",
897 		    m, obj));
898 	}
899 }
900 
901 /*
902  *	vm_page_busy_downgrade:
903  *
904  *	Downgrade an exclusive busy page into a single shared busy page.
905  */
906 void
907 vm_page_busy_downgrade(vm_page_t m)
908 {
909 	u_int x;
910 
911 	vm_page_assert_xbusied(m);
912 
913 	x = vm_page_busy_fetch(m);
914 	for (;;) {
915 		if (atomic_fcmpset_rel_int(&m->busy_lock,
916 		    &x, VPB_SHARERS_WORD(1)))
917 			break;
918 	}
919 	if ((x & VPB_BIT_WAITERS) != 0)
920 		wakeup(m);
921 }
922 
923 /*
924  *
925  *	vm_page_busy_tryupgrade:
926  *
927  *	Attempt to upgrade a single shared busy into an exclusive busy.
928  */
929 int
930 vm_page_busy_tryupgrade(vm_page_t m)
931 {
932 	u_int ce, x;
933 
934 	vm_page_assert_sbusied(m);
935 
936 	x = vm_page_busy_fetch(m);
937 	ce = VPB_CURTHREAD_EXCLUSIVE;
938 	for (;;) {
939 		if (VPB_SHARERS(x) > 1)
940 			return (0);
941 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
942 		    ("vm_page_busy_tryupgrade: invalid lock state"));
943 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
944 		    ce | (x & VPB_BIT_WAITERS)))
945 			continue;
946 		return (1);
947 	}
948 }
949 
950 /*
951  *	vm_page_sbusied:
952  *
953  *	Return a positive value if the page is shared busied, 0 otherwise.
954  */
955 int
956 vm_page_sbusied(vm_page_t m)
957 {
958 	u_int x;
959 
960 	x = vm_page_busy_fetch(m);
961 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
962 }
963 
964 /*
965  *	vm_page_sunbusy:
966  *
967  *	Shared unbusy a page.
968  */
969 void
970 vm_page_sunbusy(vm_page_t m)
971 {
972 	u_int x;
973 
974 	vm_page_assert_sbusied(m);
975 
976 	x = vm_page_busy_fetch(m);
977 	for (;;) {
978 		KASSERT(x != VPB_FREED,
979 		    ("vm_page_sunbusy: Unlocking freed page."));
980 		if (VPB_SHARERS(x) > 1) {
981 			if (atomic_fcmpset_int(&m->busy_lock, &x,
982 			    x - VPB_ONE_SHARER))
983 				break;
984 			continue;
985 		}
986 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
987 		    ("vm_page_sunbusy: invalid lock state"));
988 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
989 			continue;
990 		if ((x & VPB_BIT_WAITERS) == 0)
991 			break;
992 		wakeup(m);
993 		break;
994 	}
995 }
996 
997 /*
998  *	vm_page_busy_sleep:
999  *
1000  *	Sleep if the page is busy, using the page pointer as wchan.
1001  *	This is used to implement the hard-path of the busying mechanism.
1002  *
1003  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1004  *	will not sleep if the page is shared-busy.
1005  *
1006  *	The object lock must be held on entry.
1007  *
1008  *	Returns true if it slept and dropped the object lock, or false
1009  *	if there was no sleep and the lock is still held.
1010  */
1011 bool
1012 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1013 {
1014 	vm_object_t obj;
1015 
1016 	obj = m->object;
1017 	VM_OBJECT_ASSERT_LOCKED(obj);
1018 
1019 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1020 	    true));
1021 }
1022 
1023 /*
1024  *	vm_page_busy_sleep_unlocked:
1025  *
1026  *	Sleep if the page is busy, using the page pointer as wchan.
1027  *	This is used to implement the hard-path of busying mechanism.
1028  *
1029  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1030  *	will not sleep if the page is shared-busy.
1031  *
1032  *	The object lock must not be held on entry.  The operation will
1033  *	return if the page changes identity.
1034  */
1035 void
1036 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1037     const char *wmesg, int allocflags)
1038 {
1039 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1040 
1041 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1042 }
1043 
1044 /*
1045  *	_vm_page_busy_sleep:
1046  *
1047  *	Internal busy sleep function.  Verifies the page identity and
1048  *	lockstate against parameters.  Returns true if it sleeps and
1049  *	false otherwise.
1050  *
1051  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1052  *
1053  *	If locked is true the lock will be dropped for any true returns
1054  *	and held for any false returns.
1055  */
1056 static bool
1057 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1058     const char *wmesg, int allocflags, bool locked)
1059 {
1060 	bool xsleep;
1061 	u_int x;
1062 
1063 	/*
1064 	 * If the object is busy we must wait for that to drain to zero
1065 	 * before trying the page again.
1066 	 */
1067 	if (obj != NULL && vm_object_busied(obj)) {
1068 		if (locked)
1069 			VM_OBJECT_DROP(obj);
1070 		vm_object_busy_wait(obj, wmesg);
1071 		return (true);
1072 	}
1073 
1074 	if (!vm_page_busied(m))
1075 		return (false);
1076 
1077 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1078 	sleepq_lock(m);
1079 	x = vm_page_busy_fetch(m);
1080 	do {
1081 		/*
1082 		 * If the page changes objects or becomes unlocked we can
1083 		 * simply return.
1084 		 */
1085 		if (x == VPB_UNBUSIED ||
1086 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1087 		    m->object != obj || m->pindex != pindex) {
1088 			sleepq_release(m);
1089 			return (false);
1090 		}
1091 		if ((x & VPB_BIT_WAITERS) != 0)
1092 			break;
1093 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1094 	if (locked)
1095 		VM_OBJECT_DROP(obj);
1096 	DROP_GIANT();
1097 	sleepq_add(m, NULL, wmesg, 0, 0);
1098 	sleepq_wait(m, PVM);
1099 	PICKUP_GIANT();
1100 	return (true);
1101 }
1102 
1103 /*
1104  *	vm_page_trysbusy:
1105  *
1106  *	Try to shared busy a page.
1107  *	If the operation succeeds 1 is returned otherwise 0.
1108  *	The operation never sleeps.
1109  */
1110 int
1111 vm_page_trysbusy(vm_page_t m)
1112 {
1113 	vm_object_t obj;
1114 	u_int x;
1115 
1116 	obj = m->object;
1117 	x = vm_page_busy_fetch(m);
1118 	for (;;) {
1119 		if ((x & VPB_BIT_SHARED) == 0)
1120 			return (0);
1121 		/*
1122 		 * Reduce the window for transient busies that will trigger
1123 		 * false negatives in vm_page_ps_test().
1124 		 */
1125 		if (obj != NULL && vm_object_busied(obj))
1126 			return (0);
1127 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1128 		    x + VPB_ONE_SHARER))
1129 			break;
1130 	}
1131 
1132 	/* Refetch the object now that we're guaranteed that it is stable. */
1133 	obj = m->object;
1134 	if (obj != NULL && vm_object_busied(obj)) {
1135 		vm_page_sunbusy(m);
1136 		return (0);
1137 	}
1138 	return (1);
1139 }
1140 
1141 /*
1142  *	vm_page_tryxbusy:
1143  *
1144  *	Try to exclusive busy a page.
1145  *	If the operation succeeds 1 is returned otherwise 0.
1146  *	The operation never sleeps.
1147  */
1148 int
1149 vm_page_tryxbusy(vm_page_t m)
1150 {
1151 	vm_object_t obj;
1152 
1153         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1154             VPB_CURTHREAD_EXCLUSIVE) == 0)
1155 		return (0);
1156 
1157 	obj = m->object;
1158 	if (obj != NULL && vm_object_busied(obj)) {
1159 		vm_page_xunbusy(m);
1160 		return (0);
1161 	}
1162 	return (1);
1163 }
1164 
1165 static void
1166 vm_page_xunbusy_hard_tail(vm_page_t m)
1167 {
1168 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1169 	/* Wake the waiter. */
1170 	wakeup(m);
1171 }
1172 
1173 /*
1174  *	vm_page_xunbusy_hard:
1175  *
1176  *	Called when unbusy has failed because there is a waiter.
1177  */
1178 void
1179 vm_page_xunbusy_hard(vm_page_t m)
1180 {
1181 	vm_page_assert_xbusied(m);
1182 	vm_page_xunbusy_hard_tail(m);
1183 }
1184 
1185 void
1186 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1187 {
1188 	vm_page_assert_xbusied_unchecked(m);
1189 	vm_page_xunbusy_hard_tail(m);
1190 }
1191 
1192 static void
1193 vm_page_busy_free(vm_page_t m)
1194 {
1195 	u_int x;
1196 
1197 	atomic_thread_fence_rel();
1198 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1199 	if ((x & VPB_BIT_WAITERS) != 0)
1200 		wakeup(m);
1201 }
1202 
1203 /*
1204  *	vm_page_unhold_pages:
1205  *
1206  *	Unhold each of the pages that is referenced by the given array.
1207  */
1208 void
1209 vm_page_unhold_pages(vm_page_t *ma, int count)
1210 {
1211 
1212 	for (; count != 0; count--) {
1213 		vm_page_unwire(*ma, PQ_ACTIVE);
1214 		ma++;
1215 	}
1216 }
1217 
1218 vm_page_t
1219 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1220 {
1221 	vm_page_t m;
1222 
1223 #ifdef VM_PHYSSEG_SPARSE
1224 	m = vm_phys_paddr_to_vm_page(pa);
1225 	if (m == NULL)
1226 		m = vm_phys_fictitious_to_vm_page(pa);
1227 	return (m);
1228 #elif defined(VM_PHYSSEG_DENSE)
1229 	long pi;
1230 
1231 	pi = atop(pa);
1232 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1233 		m = &vm_page_array[pi - first_page];
1234 		return (m);
1235 	}
1236 	return (vm_phys_fictitious_to_vm_page(pa));
1237 #else
1238 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1239 #endif
1240 }
1241 
1242 /*
1243  *	vm_page_getfake:
1244  *
1245  *	Create a fictitious page with the specified physical address and
1246  *	memory attribute.  The memory attribute is the only the machine-
1247  *	dependent aspect of a fictitious page that must be initialized.
1248  */
1249 vm_page_t
1250 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1251 {
1252 	vm_page_t m;
1253 
1254 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1255 	vm_page_initfake(m, paddr, memattr);
1256 	return (m);
1257 }
1258 
1259 void
1260 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1261 {
1262 
1263 	if ((m->flags & PG_FICTITIOUS) != 0) {
1264 		/*
1265 		 * The page's memattr might have changed since the
1266 		 * previous initialization.  Update the pmap to the
1267 		 * new memattr.
1268 		 */
1269 		goto memattr;
1270 	}
1271 	m->phys_addr = paddr;
1272 	m->a.queue = PQ_NONE;
1273 	/* Fictitious pages don't use "segind". */
1274 	m->flags = PG_FICTITIOUS;
1275 	/* Fictitious pages don't use "order" or "pool". */
1276 	m->oflags = VPO_UNMANAGED;
1277 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1278 	/* Fictitious pages are unevictable. */
1279 	m->ref_count = 1;
1280 	pmap_page_init(m);
1281 memattr:
1282 	pmap_page_set_memattr(m, memattr);
1283 }
1284 
1285 /*
1286  *	vm_page_putfake:
1287  *
1288  *	Release a fictitious page.
1289  */
1290 void
1291 vm_page_putfake(vm_page_t m)
1292 {
1293 
1294 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1295 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1296 	    ("vm_page_putfake: bad page %p", m));
1297 	vm_page_assert_xbusied(m);
1298 	vm_page_busy_free(m);
1299 	uma_zfree(fakepg_zone, m);
1300 }
1301 
1302 /*
1303  *	vm_page_updatefake:
1304  *
1305  *	Update the given fictitious page to the specified physical address and
1306  *	memory attribute.
1307  */
1308 void
1309 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1310 {
1311 
1312 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1313 	    ("vm_page_updatefake: bad page %p", m));
1314 	m->phys_addr = paddr;
1315 	pmap_page_set_memattr(m, memattr);
1316 }
1317 
1318 /*
1319  *	vm_page_free:
1320  *
1321  *	Free a page.
1322  */
1323 void
1324 vm_page_free(vm_page_t m)
1325 {
1326 
1327 	m->flags &= ~PG_ZERO;
1328 	vm_page_free_toq(m);
1329 }
1330 
1331 /*
1332  *	vm_page_free_zero:
1333  *
1334  *	Free a page to the zerod-pages queue
1335  */
1336 void
1337 vm_page_free_zero(vm_page_t m)
1338 {
1339 
1340 	m->flags |= PG_ZERO;
1341 	vm_page_free_toq(m);
1342 }
1343 
1344 /*
1345  * Unbusy and handle the page queueing for a page from a getpages request that
1346  * was optionally read ahead or behind.
1347  */
1348 void
1349 vm_page_readahead_finish(vm_page_t m)
1350 {
1351 
1352 	/* We shouldn't put invalid pages on queues. */
1353 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1354 
1355 	/*
1356 	 * Since the page is not the actually needed one, whether it should
1357 	 * be activated or deactivated is not obvious.  Empirical results
1358 	 * have shown that deactivating the page is usually the best choice,
1359 	 * unless the page is wanted by another thread.
1360 	 */
1361 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1362 		vm_page_activate(m);
1363 	else
1364 		vm_page_deactivate(m);
1365 	vm_page_xunbusy_unchecked(m);
1366 }
1367 
1368 /*
1369  * Destroy the identity of an invalid page and free it if possible.
1370  * This is intended to be used when reading a page from backing store fails.
1371  */
1372 void
1373 vm_page_free_invalid(vm_page_t m)
1374 {
1375 
1376 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1377 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1378 	KASSERT(m->object != NULL, ("page %p has no object", m));
1379 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1380 
1381 	/*
1382 	 * We may be attempting to free the page as part of the handling for an
1383 	 * I/O error, in which case the page was xbusied by a different thread.
1384 	 */
1385 	vm_page_xbusy_claim(m);
1386 
1387 	/*
1388 	 * If someone has wired this page while the object lock
1389 	 * was not held, then the thread that unwires is responsible
1390 	 * for freeing the page.  Otherwise just free the page now.
1391 	 * The wire count of this unmapped page cannot change while
1392 	 * we have the page xbusy and the page's object wlocked.
1393 	 */
1394 	if (vm_page_remove(m))
1395 		vm_page_free(m);
1396 }
1397 
1398 /*
1399  *	vm_page_dirty_KBI:		[ internal use only ]
1400  *
1401  *	Set all bits in the page's dirty field.
1402  *
1403  *	The object containing the specified page must be locked if the
1404  *	call is made from the machine-independent layer.
1405  *
1406  *	See vm_page_clear_dirty_mask().
1407  *
1408  *	This function should only be called by vm_page_dirty().
1409  */
1410 void
1411 vm_page_dirty_KBI(vm_page_t m)
1412 {
1413 
1414 	/* Refer to this operation by its public name. */
1415 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1416 	m->dirty = VM_PAGE_BITS_ALL;
1417 }
1418 
1419 /*
1420  *	vm_page_insert:		[ internal use only ]
1421  *
1422  *	Inserts the given mem entry into the object and object list.
1423  *
1424  *	The object must be locked.
1425  */
1426 int
1427 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1428 {
1429 	vm_page_t mpred;
1430 
1431 	VM_OBJECT_ASSERT_WLOCKED(object);
1432 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1433 	return (vm_page_insert_after(m, object, pindex, mpred));
1434 }
1435 
1436 /*
1437  *	vm_page_insert_after:
1438  *
1439  *	Inserts the page "m" into the specified object at offset "pindex".
1440  *
1441  *	The page "mpred" must immediately precede the offset "pindex" within
1442  *	the specified object.
1443  *
1444  *	The object must be locked.
1445  */
1446 static int
1447 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1448     vm_page_t mpred)
1449 {
1450 	vm_page_t msucc;
1451 
1452 	VM_OBJECT_ASSERT_WLOCKED(object);
1453 	KASSERT(m->object == NULL,
1454 	    ("vm_page_insert_after: page already inserted"));
1455 	if (mpred != NULL) {
1456 		KASSERT(mpred->object == object,
1457 		    ("vm_page_insert_after: object doesn't contain mpred"));
1458 		KASSERT(mpred->pindex < pindex,
1459 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1460 		msucc = TAILQ_NEXT(mpred, listq);
1461 	} else
1462 		msucc = TAILQ_FIRST(&object->memq);
1463 	if (msucc != NULL)
1464 		KASSERT(msucc->pindex > pindex,
1465 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1466 
1467 	/*
1468 	 * Record the object/offset pair in this page.
1469 	 */
1470 	m->object = object;
1471 	m->pindex = pindex;
1472 	m->ref_count |= VPRC_OBJREF;
1473 
1474 	/*
1475 	 * Now link into the object's ordered list of backed pages.
1476 	 */
1477 	if (vm_radix_insert(&object->rtree, m)) {
1478 		m->object = NULL;
1479 		m->pindex = 0;
1480 		m->ref_count &= ~VPRC_OBJREF;
1481 		return (1);
1482 	}
1483 	vm_page_insert_radixdone(m, object, mpred);
1484 	vm_pager_page_inserted(object, m);
1485 	return (0);
1486 }
1487 
1488 /*
1489  *	vm_page_insert_radixdone:
1490  *
1491  *	Complete page "m" insertion into the specified object after the
1492  *	radix trie hooking.
1493  *
1494  *	The page "mpred" must precede the offset "m->pindex" within the
1495  *	specified object.
1496  *
1497  *	The object must be locked.
1498  */
1499 static void
1500 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1501 {
1502 
1503 	VM_OBJECT_ASSERT_WLOCKED(object);
1504 	KASSERT(object != NULL && m->object == object,
1505 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1506 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1507 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1508 	if (mpred != NULL) {
1509 		KASSERT(mpred->object == object,
1510 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1511 		KASSERT(mpred->pindex < m->pindex,
1512 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1513 	}
1514 
1515 	if (mpred != NULL)
1516 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1517 	else
1518 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1519 
1520 	/*
1521 	 * Show that the object has one more resident page.
1522 	 */
1523 	object->resident_page_count++;
1524 
1525 	/*
1526 	 * Hold the vnode until the last page is released.
1527 	 */
1528 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1529 		vhold(object->handle);
1530 
1531 	/*
1532 	 * Since we are inserting a new and possibly dirty page,
1533 	 * update the object's generation count.
1534 	 */
1535 	if (pmap_page_is_write_mapped(m))
1536 		vm_object_set_writeable_dirty(object);
1537 }
1538 
1539 /*
1540  * Do the work to remove a page from its object.  The caller is responsible for
1541  * updating the page's fields to reflect this removal.
1542  */
1543 static void
1544 vm_page_object_remove(vm_page_t m)
1545 {
1546 	vm_object_t object;
1547 	vm_page_t mrem __diagused;
1548 
1549 	vm_page_assert_xbusied(m);
1550 	object = m->object;
1551 	VM_OBJECT_ASSERT_WLOCKED(object);
1552 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1553 	    ("page %p is missing its object ref", m));
1554 
1555 	/* Deferred free of swap space. */
1556 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1557 		vm_pager_page_unswapped(m);
1558 
1559 	vm_pager_page_removed(object, m);
1560 
1561 	m->object = NULL;
1562 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1563 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1564 
1565 	/*
1566 	 * Now remove from the object's list of backed pages.
1567 	 */
1568 	TAILQ_REMOVE(&object->memq, m, listq);
1569 
1570 	/*
1571 	 * And show that the object has one fewer resident page.
1572 	 */
1573 	object->resident_page_count--;
1574 
1575 	/*
1576 	 * The vnode may now be recycled.
1577 	 */
1578 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1579 		vdrop(object->handle);
1580 }
1581 
1582 /*
1583  *	vm_page_remove:
1584  *
1585  *	Removes the specified page from its containing object, but does not
1586  *	invalidate any backing storage.  Returns true if the object's reference
1587  *	was the last reference to the page, and false otherwise.
1588  *
1589  *	The object must be locked and the page must be exclusively busied.
1590  *	The exclusive busy will be released on return.  If this is not the
1591  *	final ref and the caller does not hold a wire reference it may not
1592  *	continue to access the page.
1593  */
1594 bool
1595 vm_page_remove(vm_page_t m)
1596 {
1597 	bool dropped;
1598 
1599 	dropped = vm_page_remove_xbusy(m);
1600 	vm_page_xunbusy(m);
1601 
1602 	return (dropped);
1603 }
1604 
1605 /*
1606  *	vm_page_remove_xbusy
1607  *
1608  *	Removes the page but leaves the xbusy held.  Returns true if this
1609  *	removed the final ref and false otherwise.
1610  */
1611 bool
1612 vm_page_remove_xbusy(vm_page_t m)
1613 {
1614 
1615 	vm_page_object_remove(m);
1616 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1617 }
1618 
1619 /*
1620  *	vm_page_lookup:
1621  *
1622  *	Returns the page associated with the object/offset
1623  *	pair specified; if none is found, NULL is returned.
1624  *
1625  *	The object must be locked.
1626  */
1627 vm_page_t
1628 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1629 {
1630 
1631 	VM_OBJECT_ASSERT_LOCKED(object);
1632 	return (vm_radix_lookup(&object->rtree, pindex));
1633 }
1634 
1635 /*
1636  *	vm_page_lookup_unlocked:
1637  *
1638  *	Returns the page associated with the object/offset pair specified;
1639  *	if none is found, NULL is returned.  The page may be no longer be
1640  *	present in the object at the time that this function returns.  Only
1641  *	useful for opportunistic checks such as inmem().
1642  */
1643 vm_page_t
1644 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1645 {
1646 
1647 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1648 }
1649 
1650 /*
1651  *	vm_page_relookup:
1652  *
1653  *	Returns a page that must already have been busied by
1654  *	the caller.  Used for bogus page replacement.
1655  */
1656 vm_page_t
1657 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1658 {
1659 	vm_page_t m;
1660 
1661 	m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1662 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1663 	    m->object == object && m->pindex == pindex,
1664 	    ("vm_page_relookup: Invalid page %p", m));
1665 	return (m);
1666 }
1667 
1668 /*
1669  * This should only be used by lockless functions for releasing transient
1670  * incorrect acquires.  The page may have been freed after we acquired a
1671  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1672  * further to do.
1673  */
1674 static void
1675 vm_page_busy_release(vm_page_t m)
1676 {
1677 	u_int x;
1678 
1679 	x = vm_page_busy_fetch(m);
1680 	for (;;) {
1681 		if (x == VPB_FREED)
1682 			break;
1683 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1684 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1685 			    x - VPB_ONE_SHARER))
1686 				break;
1687 			continue;
1688 		}
1689 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1690 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1691 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1692 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1693 			continue;
1694 		if ((x & VPB_BIT_WAITERS) != 0)
1695 			wakeup(m);
1696 		break;
1697 	}
1698 }
1699 
1700 /*
1701  *	vm_page_find_least:
1702  *
1703  *	Returns the page associated with the object with least pindex
1704  *	greater than or equal to the parameter pindex, or NULL.
1705  *
1706  *	The object must be locked.
1707  */
1708 vm_page_t
1709 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1710 {
1711 	vm_page_t m;
1712 
1713 	VM_OBJECT_ASSERT_LOCKED(object);
1714 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1715 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1716 	return (m);
1717 }
1718 
1719 /*
1720  * Returns the given page's successor (by pindex) within the object if it is
1721  * resident; if none is found, NULL is returned.
1722  *
1723  * The object must be locked.
1724  */
1725 vm_page_t
1726 vm_page_next(vm_page_t m)
1727 {
1728 	vm_page_t next;
1729 
1730 	VM_OBJECT_ASSERT_LOCKED(m->object);
1731 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1732 		MPASS(next->object == m->object);
1733 		if (next->pindex != m->pindex + 1)
1734 			next = NULL;
1735 	}
1736 	return (next);
1737 }
1738 
1739 /*
1740  * Returns the given page's predecessor (by pindex) within the object if it is
1741  * resident; if none is found, NULL is returned.
1742  *
1743  * The object must be locked.
1744  */
1745 vm_page_t
1746 vm_page_prev(vm_page_t m)
1747 {
1748 	vm_page_t prev;
1749 
1750 	VM_OBJECT_ASSERT_LOCKED(m->object);
1751 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1752 		MPASS(prev->object == m->object);
1753 		if (prev->pindex != m->pindex - 1)
1754 			prev = NULL;
1755 	}
1756 	return (prev);
1757 }
1758 
1759 /*
1760  * Uses the page mnew as a replacement for an existing page at index
1761  * pindex which must be already present in the object.
1762  *
1763  * Both pages must be exclusively busied on enter.  The old page is
1764  * unbusied on exit.
1765  *
1766  * A return value of true means mold is now free.  If this is not the
1767  * final ref and the caller does not hold a wire reference it may not
1768  * continue to access the page.
1769  */
1770 static bool
1771 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1772     vm_page_t mold)
1773 {
1774 	vm_page_t mret __diagused;
1775 	bool dropped;
1776 
1777 	VM_OBJECT_ASSERT_WLOCKED(object);
1778 	vm_page_assert_xbusied(mold);
1779 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1780 	    ("vm_page_replace: page %p already in object", mnew));
1781 
1782 	/*
1783 	 * This function mostly follows vm_page_insert() and
1784 	 * vm_page_remove() without the radix, object count and vnode
1785 	 * dance.  Double check such functions for more comments.
1786 	 */
1787 
1788 	mnew->object = object;
1789 	mnew->pindex = pindex;
1790 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1791 	mret = vm_radix_replace(&object->rtree, mnew);
1792 	KASSERT(mret == mold,
1793 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1794 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1795 	    (mnew->oflags & VPO_UNMANAGED),
1796 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1797 
1798 	/* Keep the resident page list in sorted order. */
1799 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1800 	TAILQ_REMOVE(&object->memq, mold, listq);
1801 	mold->object = NULL;
1802 
1803 	/*
1804 	 * The object's resident_page_count does not change because we have
1805 	 * swapped one page for another, but the generation count should
1806 	 * change if the page is dirty.
1807 	 */
1808 	if (pmap_page_is_write_mapped(mnew))
1809 		vm_object_set_writeable_dirty(object);
1810 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1811 	vm_page_xunbusy(mold);
1812 
1813 	return (dropped);
1814 }
1815 
1816 void
1817 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1818     vm_page_t mold)
1819 {
1820 
1821 	vm_page_assert_xbusied(mnew);
1822 
1823 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1824 		vm_page_free(mold);
1825 }
1826 
1827 /*
1828  *	vm_page_rename:
1829  *
1830  *	Move the given memory entry from its
1831  *	current object to the specified target object/offset.
1832  *
1833  *	Note: swap associated with the page must be invalidated by the move.  We
1834  *	      have to do this for several reasons:  (1) we aren't freeing the
1835  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1836  *	      moving the page from object A to B, and will then later move
1837  *	      the backing store from A to B and we can't have a conflict.
1838  *
1839  *	Note: we *always* dirty the page.  It is necessary both for the
1840  *	      fact that we moved it, and because we may be invalidating
1841  *	      swap.
1842  *
1843  *	The objects must be locked.
1844  */
1845 int
1846 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1847 {
1848 	vm_page_t mpred;
1849 	vm_pindex_t opidx;
1850 
1851 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1852 
1853 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1854 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1855 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1856 	    ("vm_page_rename: pindex already renamed"));
1857 
1858 	/*
1859 	 * Create a custom version of vm_page_insert() which does not depend
1860 	 * by m_prev and can cheat on the implementation aspects of the
1861 	 * function.
1862 	 */
1863 	opidx = m->pindex;
1864 	m->pindex = new_pindex;
1865 	if (vm_radix_insert(&new_object->rtree, m)) {
1866 		m->pindex = opidx;
1867 		return (1);
1868 	}
1869 
1870 	/*
1871 	 * The operation cannot fail anymore.  The removal must happen before
1872 	 * the listq iterator is tainted.
1873 	 */
1874 	m->pindex = opidx;
1875 	vm_page_object_remove(m);
1876 
1877 	/* Return back to the new pindex to complete vm_page_insert(). */
1878 	m->pindex = new_pindex;
1879 	m->object = new_object;
1880 
1881 	vm_page_insert_radixdone(m, new_object, mpred);
1882 	vm_page_dirty(m);
1883 	vm_pager_page_inserted(new_object, m);
1884 	return (0);
1885 }
1886 
1887 /*
1888  *	vm_page_alloc:
1889  *
1890  *	Allocate and return a page that is associated with the specified
1891  *	object and offset pair.  By default, this page is exclusive busied.
1892  *
1893  *	The caller must always specify an allocation class.
1894  *
1895  *	allocation classes:
1896  *	VM_ALLOC_NORMAL		normal process request
1897  *	VM_ALLOC_SYSTEM		system *really* needs a page
1898  *	VM_ALLOC_INTERRUPT	interrupt time request
1899  *
1900  *	optional allocation flags:
1901  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1902  *				intends to allocate
1903  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1904  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1905  *	VM_ALLOC_SBUSY		shared busy the allocated page
1906  *	VM_ALLOC_WIRED		wire the allocated page
1907  *	VM_ALLOC_ZERO		prefer a zeroed page
1908  */
1909 vm_page_t
1910 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1911 {
1912 
1913 	return (vm_page_alloc_after(object, pindex, req,
1914 	    vm_radix_lookup_le(&object->rtree, pindex)));
1915 }
1916 
1917 vm_page_t
1918 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1919     int req)
1920 {
1921 
1922 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1923 	    vm_radix_lookup_le(&object->rtree, pindex)));
1924 }
1925 
1926 /*
1927  * Allocate a page in the specified object with the given page index.  To
1928  * optimize insertion of the page into the object, the caller must also specifiy
1929  * the resident page in the object with largest index smaller than the given
1930  * page index, or NULL if no such page exists.
1931  */
1932 vm_page_t
1933 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1934     int req, vm_page_t mpred)
1935 {
1936 	struct vm_domainset_iter di;
1937 	vm_page_t m;
1938 	int domain;
1939 
1940 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1941 	do {
1942 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1943 		    mpred);
1944 		if (m != NULL)
1945 			break;
1946 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1947 
1948 	return (m);
1949 }
1950 
1951 /*
1952  * Returns true if the number of free pages exceeds the minimum
1953  * for the request class and false otherwise.
1954  */
1955 static int
1956 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1957 {
1958 	u_int limit, old, new;
1959 
1960 	if (req_class == VM_ALLOC_INTERRUPT)
1961 		limit = 0;
1962 	else if (req_class == VM_ALLOC_SYSTEM)
1963 		limit = vmd->vmd_interrupt_free_min;
1964 	else
1965 		limit = vmd->vmd_free_reserved;
1966 
1967 	/*
1968 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1969 	 */
1970 	limit += npages;
1971 	old = vmd->vmd_free_count;
1972 	do {
1973 		if (old < limit)
1974 			return (0);
1975 		new = old - npages;
1976 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1977 
1978 	/* Wake the page daemon if we've crossed the threshold. */
1979 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1980 		pagedaemon_wakeup(vmd->vmd_domain);
1981 
1982 	/* Only update bitsets on transitions. */
1983 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1984 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1985 		vm_domain_set(vmd);
1986 
1987 	return (1);
1988 }
1989 
1990 int
1991 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1992 {
1993 	int req_class;
1994 
1995 	/*
1996 	 * The page daemon is allowed to dig deeper into the free page list.
1997 	 */
1998 	req_class = req & VM_ALLOC_CLASS_MASK;
1999 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2000 		req_class = VM_ALLOC_SYSTEM;
2001 	return (_vm_domain_allocate(vmd, req_class, npages));
2002 }
2003 
2004 vm_page_t
2005 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2006     int req, vm_page_t mpred)
2007 {
2008 	struct vm_domain *vmd;
2009 	vm_page_t m;
2010 	int flags;
2011 
2012 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2013 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2014 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2015 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2016 	KASSERT((req & ~VPA_FLAGS) == 0,
2017 	    ("invalid request %#x", req));
2018 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2019 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2020 	    ("invalid request %#x", req));
2021 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2022 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2023 	    (uintmax_t)pindex));
2024 	VM_OBJECT_ASSERT_WLOCKED(object);
2025 
2026 	flags = 0;
2027 	m = NULL;
2028 	if (!vm_pager_can_alloc_page(object, pindex))
2029 		return (NULL);
2030 again:
2031 #if VM_NRESERVLEVEL > 0
2032 	/*
2033 	 * Can we allocate the page from a reservation?
2034 	 */
2035 	if (vm_object_reserv(object) &&
2036 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2037 	    NULL) {
2038 		goto found;
2039 	}
2040 #endif
2041 	vmd = VM_DOMAIN(domain);
2042 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2043 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2044 		    M_NOWAIT | M_NOVM);
2045 		if (m != NULL) {
2046 			flags |= PG_PCPU_CACHE;
2047 			goto found;
2048 		}
2049 	}
2050 	if (vm_domain_allocate(vmd, req, 1)) {
2051 		/*
2052 		 * If not, allocate it from the free page queues.
2053 		 */
2054 		vm_domain_free_lock(vmd);
2055 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2056 		vm_domain_free_unlock(vmd);
2057 		if (m == NULL) {
2058 			vm_domain_freecnt_inc(vmd, 1);
2059 #if VM_NRESERVLEVEL > 0
2060 			if (vm_reserv_reclaim_inactive(domain))
2061 				goto again;
2062 #endif
2063 		}
2064 	}
2065 	if (m == NULL) {
2066 		/*
2067 		 * Not allocatable, give up.
2068 		 */
2069 		if (vm_domain_alloc_fail(vmd, object, req))
2070 			goto again;
2071 		return (NULL);
2072 	}
2073 
2074 	/*
2075 	 * At this point we had better have found a good page.
2076 	 */
2077 found:
2078 	vm_page_dequeue(m);
2079 	vm_page_alloc_check(m);
2080 
2081 	/*
2082 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2083 	 */
2084 	flags |= m->flags & PG_ZERO;
2085 	if ((req & VM_ALLOC_NODUMP) != 0)
2086 		flags |= PG_NODUMP;
2087 	m->flags = flags;
2088 	m->a.flags = 0;
2089 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2090 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2091 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2092 	else if ((req & VM_ALLOC_SBUSY) != 0)
2093 		m->busy_lock = VPB_SHARERS_WORD(1);
2094 	else
2095 		m->busy_lock = VPB_UNBUSIED;
2096 	if (req & VM_ALLOC_WIRED) {
2097 		vm_wire_add(1);
2098 		m->ref_count = 1;
2099 	}
2100 	m->a.act_count = 0;
2101 
2102 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2103 		if (req & VM_ALLOC_WIRED) {
2104 			vm_wire_sub(1);
2105 			m->ref_count = 0;
2106 		}
2107 		KASSERT(m->object == NULL, ("page %p has object", m));
2108 		m->oflags = VPO_UNMANAGED;
2109 		m->busy_lock = VPB_UNBUSIED;
2110 		/* Don't change PG_ZERO. */
2111 		vm_page_free_toq(m);
2112 		if (req & VM_ALLOC_WAITFAIL) {
2113 			VM_OBJECT_WUNLOCK(object);
2114 			vm_radix_wait();
2115 			VM_OBJECT_WLOCK(object);
2116 		}
2117 		return (NULL);
2118 	}
2119 
2120 	/* Ignore device objects; the pager sets "memattr" for them. */
2121 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2122 	    (object->flags & OBJ_FICTITIOUS) == 0)
2123 		pmap_page_set_memattr(m, object->memattr);
2124 
2125 	return (m);
2126 }
2127 
2128 /*
2129  *	vm_page_alloc_contig:
2130  *
2131  *	Allocate a contiguous set of physical pages of the given size "npages"
2132  *	from the free lists.  All of the physical pages must be at or above
2133  *	the given physical address "low" and below the given physical address
2134  *	"high".  The given value "alignment" determines the alignment of the
2135  *	first physical page in the set.  If the given value "boundary" is
2136  *	non-zero, then the set of physical pages cannot cross any physical
2137  *	address boundary that is a multiple of that value.  Both "alignment"
2138  *	and "boundary" must be a power of two.
2139  *
2140  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2141  *	then the memory attribute setting for the physical pages is configured
2142  *	to the object's memory attribute setting.  Otherwise, the memory
2143  *	attribute setting for the physical pages is configured to "memattr",
2144  *	overriding the object's memory attribute setting.  However, if the
2145  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2146  *	memory attribute setting for the physical pages cannot be configured
2147  *	to VM_MEMATTR_DEFAULT.
2148  *
2149  *	The specified object may not contain fictitious pages.
2150  *
2151  *	The caller must always specify an allocation class.
2152  *
2153  *	allocation classes:
2154  *	VM_ALLOC_NORMAL		normal process request
2155  *	VM_ALLOC_SYSTEM		system *really* needs a page
2156  *	VM_ALLOC_INTERRUPT	interrupt time request
2157  *
2158  *	optional allocation flags:
2159  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2160  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2161  *	VM_ALLOC_SBUSY		shared busy the allocated page
2162  *	VM_ALLOC_WIRED		wire the allocated page
2163  *	VM_ALLOC_ZERO		prefer a zeroed page
2164  */
2165 vm_page_t
2166 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2167     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2168     vm_paddr_t boundary, vm_memattr_t memattr)
2169 {
2170 	struct vm_domainset_iter di;
2171 	vm_page_t bounds[2];
2172 	vm_page_t m;
2173 	int domain;
2174 	int start_segind;
2175 
2176 	start_segind = -1;
2177 
2178 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2179 	do {
2180 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2181 		    npages, low, high, alignment, boundary, memattr);
2182 		if (m != NULL)
2183 			break;
2184 		if (start_segind == -1)
2185 			start_segind = vm_phys_lookup_segind(low);
2186 		if (vm_phys_find_range(bounds, start_segind, domain,
2187 		    npages, low, high) == -1) {
2188 			vm_domainset_iter_ignore(&di, domain);
2189 		}
2190 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2191 
2192 	return (m);
2193 }
2194 
2195 static vm_page_t
2196 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2197     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2198 {
2199 	struct vm_domain *vmd;
2200 	vm_page_t m_ret;
2201 
2202 	/*
2203 	 * Can we allocate the pages without the number of free pages falling
2204 	 * below the lower bound for the allocation class?
2205 	 */
2206 	vmd = VM_DOMAIN(domain);
2207 	if (!vm_domain_allocate(vmd, req, npages))
2208 		return (NULL);
2209 	/*
2210 	 * Try to allocate the pages from the free page queues.
2211 	 */
2212 	vm_domain_free_lock(vmd);
2213 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2214 	    alignment, boundary);
2215 	vm_domain_free_unlock(vmd);
2216 	if (m_ret != NULL)
2217 		return (m_ret);
2218 #if VM_NRESERVLEVEL > 0
2219 	/*
2220 	 * Try to break a reservation to allocate the pages.
2221 	 */
2222 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2223 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2224 	            high, alignment, boundary);
2225 		if (m_ret != NULL)
2226 			return (m_ret);
2227 	}
2228 #endif
2229 	vm_domain_freecnt_inc(vmd, npages);
2230 	return (NULL);
2231 }
2232 
2233 vm_page_t
2234 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2235     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2236     vm_paddr_t boundary, vm_memattr_t memattr)
2237 {
2238 	vm_page_t m, m_ret, mpred;
2239 	u_int busy_lock, flags, oflags;
2240 
2241 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2242 	KASSERT((req & ~VPAC_FLAGS) == 0,
2243 	    ("invalid request %#x", req));
2244 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2245 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2246 	    ("invalid request %#x", req));
2247 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2248 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2249 	    ("invalid request %#x", req));
2250 	VM_OBJECT_ASSERT_WLOCKED(object);
2251 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2252 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2253 	    object));
2254 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2255 
2256 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
2257 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2258 	    ("vm_page_alloc_contig: pindex already allocated"));
2259 	for (;;) {
2260 #if VM_NRESERVLEVEL > 0
2261 		/*
2262 		 * Can we allocate the pages from a reservation?
2263 		 */
2264 		if (vm_object_reserv(object) &&
2265 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2266 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2267 			break;
2268 		}
2269 #endif
2270 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2271 		    low, high, alignment, boundary)) != NULL)
2272 			break;
2273 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2274 			return (NULL);
2275 	}
2276 	for (m = m_ret; m < &m_ret[npages]; m++) {
2277 		vm_page_dequeue(m);
2278 		vm_page_alloc_check(m);
2279 	}
2280 
2281 	/*
2282 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2283 	 */
2284 	flags = PG_ZERO;
2285 	if ((req & VM_ALLOC_NODUMP) != 0)
2286 		flags |= PG_NODUMP;
2287 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2288 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2289 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2290 	else if ((req & VM_ALLOC_SBUSY) != 0)
2291 		busy_lock = VPB_SHARERS_WORD(1);
2292 	else
2293 		busy_lock = VPB_UNBUSIED;
2294 	if ((req & VM_ALLOC_WIRED) != 0)
2295 		vm_wire_add(npages);
2296 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2297 	    memattr == VM_MEMATTR_DEFAULT)
2298 		memattr = object->memattr;
2299 	for (m = m_ret; m < &m_ret[npages]; m++) {
2300 		m->a.flags = 0;
2301 		m->flags = (m->flags | PG_NODUMP) & flags;
2302 		m->busy_lock = busy_lock;
2303 		if ((req & VM_ALLOC_WIRED) != 0)
2304 			m->ref_count = 1;
2305 		m->a.act_count = 0;
2306 		m->oflags = oflags;
2307 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2308 			if ((req & VM_ALLOC_WIRED) != 0)
2309 				vm_wire_sub(npages);
2310 			KASSERT(m->object == NULL,
2311 			    ("page %p has object", m));
2312 			mpred = m;
2313 			for (m = m_ret; m < &m_ret[npages]; m++) {
2314 				if (m <= mpred &&
2315 				    (req & VM_ALLOC_WIRED) != 0)
2316 					m->ref_count = 0;
2317 				m->oflags = VPO_UNMANAGED;
2318 				m->busy_lock = VPB_UNBUSIED;
2319 				/* Don't change PG_ZERO. */
2320 				vm_page_free_toq(m);
2321 			}
2322 			if (req & VM_ALLOC_WAITFAIL) {
2323 				VM_OBJECT_WUNLOCK(object);
2324 				vm_radix_wait();
2325 				VM_OBJECT_WLOCK(object);
2326 			}
2327 			return (NULL);
2328 		}
2329 		mpred = m;
2330 		if (memattr != VM_MEMATTR_DEFAULT)
2331 			pmap_page_set_memattr(m, memattr);
2332 		pindex++;
2333 	}
2334 	return (m_ret);
2335 }
2336 
2337 /*
2338  * Allocate a physical page that is not intended to be inserted into a VM
2339  * object.  If the "freelist" parameter is not equal to VM_NFREELIST, then only
2340  * pages from the specified vm_phys freelist will be returned.
2341  */
2342 static __always_inline vm_page_t
2343 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req)
2344 {
2345 	struct vm_domain *vmd;
2346 	vm_page_t m;
2347 	int flags;
2348 
2349 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2350 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2351 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2352 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2353 	KASSERT((req & ~VPAN_FLAGS) == 0,
2354 	    ("invalid request %#x", req));
2355 
2356 	flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0;
2357 	vmd = VM_DOMAIN(domain);
2358 again:
2359 	if (freelist == VM_NFREELIST &&
2360 	    vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2361 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2362 		    M_NOWAIT | M_NOVM);
2363 		if (m != NULL) {
2364 			flags |= PG_PCPU_CACHE;
2365 			goto found;
2366 		}
2367 	}
2368 
2369 	if (vm_domain_allocate(vmd, req, 1)) {
2370 		vm_domain_free_lock(vmd);
2371 		if (freelist == VM_NFREELIST)
2372 			m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2373 		else
2374 			m = vm_phys_alloc_freelist_pages(domain, freelist,
2375 			    VM_FREEPOOL_DIRECT, 0);
2376 		vm_domain_free_unlock(vmd);
2377 		if (m == NULL) {
2378 			vm_domain_freecnt_inc(vmd, 1);
2379 #if VM_NRESERVLEVEL > 0
2380 			if (freelist == VM_NFREELIST &&
2381 			    vm_reserv_reclaim_inactive(domain))
2382 				goto again;
2383 #endif
2384 		}
2385 	}
2386 	if (m == NULL) {
2387 		if (vm_domain_alloc_fail(vmd, NULL, req))
2388 			goto again;
2389 		return (NULL);
2390 	}
2391 
2392 found:
2393 	vm_page_dequeue(m);
2394 	vm_page_alloc_check(m);
2395 
2396 	/*
2397 	 * Consumers should not rely on a useful default pindex value.
2398 	 */
2399 	m->pindex = 0xdeadc0dedeadc0de;
2400 	m->flags = (m->flags & PG_ZERO) | flags;
2401 	m->a.flags = 0;
2402 	m->oflags = VPO_UNMANAGED;
2403 	m->busy_lock = VPB_UNBUSIED;
2404 	if ((req & VM_ALLOC_WIRED) != 0) {
2405 		vm_wire_add(1);
2406 		m->ref_count = 1;
2407 	}
2408 
2409 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2410 		pmap_zero_page(m);
2411 
2412 	return (m);
2413 }
2414 
2415 vm_page_t
2416 vm_page_alloc_freelist(int freelist, int req)
2417 {
2418 	struct vm_domainset_iter di;
2419 	vm_page_t m;
2420 	int domain;
2421 
2422 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2423 	do {
2424 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2425 		if (m != NULL)
2426 			break;
2427 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2428 
2429 	return (m);
2430 }
2431 
2432 vm_page_t
2433 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2434 {
2435 	KASSERT(freelist >= 0 && freelist < VM_NFREELIST,
2436 	    ("%s: invalid freelist %d", __func__, freelist));
2437 
2438 	return (_vm_page_alloc_noobj_domain(domain, freelist, req));
2439 }
2440 
2441 vm_page_t
2442 vm_page_alloc_noobj(int req)
2443 {
2444 	struct vm_domainset_iter di;
2445 	vm_page_t m;
2446 	int domain;
2447 
2448 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2449 	do {
2450 		m = vm_page_alloc_noobj_domain(domain, req);
2451 		if (m != NULL)
2452 			break;
2453 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2454 
2455 	return (m);
2456 }
2457 
2458 vm_page_t
2459 vm_page_alloc_noobj_domain(int domain, int req)
2460 {
2461 	return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req));
2462 }
2463 
2464 vm_page_t
2465 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2466     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2467     vm_memattr_t memattr)
2468 {
2469 	struct vm_domainset_iter di;
2470 	vm_page_t m;
2471 	int domain;
2472 
2473 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2474 	do {
2475 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2476 		    high, alignment, boundary, memattr);
2477 		if (m != NULL)
2478 			break;
2479 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2480 
2481 	return (m);
2482 }
2483 
2484 vm_page_t
2485 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2486     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2487     vm_memattr_t memattr)
2488 {
2489 	vm_page_t m, m_ret;
2490 	u_int flags;
2491 
2492 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2493 	KASSERT((req & ~VPANC_FLAGS) == 0,
2494 	    ("invalid request %#x", req));
2495 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2496 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2497 	    ("invalid request %#x", req));
2498 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2499 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2500 	    ("invalid request %#x", req));
2501 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2502 
2503 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2504 	    low, high, alignment, boundary)) == NULL) {
2505 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2506 			return (NULL);
2507 	}
2508 
2509 	/*
2510 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2511 	 */
2512 	flags = PG_ZERO;
2513 	if ((req & VM_ALLOC_NODUMP) != 0)
2514 		flags |= PG_NODUMP;
2515 	if ((req & VM_ALLOC_WIRED) != 0)
2516 		vm_wire_add(npages);
2517 	for (m = m_ret; m < &m_ret[npages]; m++) {
2518 		vm_page_dequeue(m);
2519 		vm_page_alloc_check(m);
2520 
2521 		/*
2522 		 * Consumers should not rely on a useful default pindex value.
2523 		 */
2524 		m->pindex = 0xdeadc0dedeadc0de;
2525 		m->a.flags = 0;
2526 		m->flags = (m->flags | PG_NODUMP) & flags;
2527 		m->busy_lock = VPB_UNBUSIED;
2528 		if ((req & VM_ALLOC_WIRED) != 0)
2529 			m->ref_count = 1;
2530 		m->a.act_count = 0;
2531 		m->oflags = VPO_UNMANAGED;
2532 
2533 		/*
2534 		 * Zero the page before updating any mappings since the page is
2535 		 * not yet shared with any devices which might require the
2536 		 * non-default memory attribute.  pmap_page_set_memattr()
2537 		 * flushes data caches before returning.
2538 		 */
2539 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2540 			pmap_zero_page(m);
2541 		if (memattr != VM_MEMATTR_DEFAULT)
2542 			pmap_page_set_memattr(m, memattr);
2543 	}
2544 	return (m_ret);
2545 }
2546 
2547 /*
2548  * Check a page that has been freshly dequeued from a freelist.
2549  */
2550 static void
2551 vm_page_alloc_check(vm_page_t m)
2552 {
2553 
2554 	KASSERT(m->object == NULL, ("page %p has object", m));
2555 	KASSERT(m->a.queue == PQ_NONE &&
2556 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2557 	    ("page %p has unexpected queue %d, flags %#x",
2558 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2559 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2560 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2561 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2562 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2563 	    ("page %p has unexpected memattr %d",
2564 	    m, pmap_page_get_memattr(m)));
2565 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2566 	pmap_vm_page_alloc_check(m);
2567 }
2568 
2569 static int
2570 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2571 {
2572 	struct vm_domain *vmd;
2573 	struct vm_pgcache *pgcache;
2574 	int i;
2575 
2576 	pgcache = arg;
2577 	vmd = VM_DOMAIN(pgcache->domain);
2578 
2579 	/*
2580 	 * The page daemon should avoid creating extra memory pressure since its
2581 	 * main purpose is to replenish the store of free pages.
2582 	 */
2583 	if (vmd->vmd_severeset || curproc == pageproc ||
2584 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2585 		return (0);
2586 	domain = vmd->vmd_domain;
2587 	vm_domain_free_lock(vmd);
2588 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2589 	    (vm_page_t *)store);
2590 	vm_domain_free_unlock(vmd);
2591 	if (cnt != i)
2592 		vm_domain_freecnt_inc(vmd, cnt - i);
2593 
2594 	return (i);
2595 }
2596 
2597 static void
2598 vm_page_zone_release(void *arg, void **store, int cnt)
2599 {
2600 	struct vm_domain *vmd;
2601 	struct vm_pgcache *pgcache;
2602 	vm_page_t m;
2603 	int i;
2604 
2605 	pgcache = arg;
2606 	vmd = VM_DOMAIN(pgcache->domain);
2607 	vm_domain_free_lock(vmd);
2608 	for (i = 0; i < cnt; i++) {
2609 		m = (vm_page_t)store[i];
2610 		vm_phys_free_pages(m, 0);
2611 	}
2612 	vm_domain_free_unlock(vmd);
2613 	vm_domain_freecnt_inc(vmd, cnt);
2614 }
2615 
2616 #define	VPSC_ANY	0	/* No restrictions. */
2617 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2618 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2619 
2620 /*
2621  *	vm_page_scan_contig:
2622  *
2623  *	Scan vm_page_array[] between the specified entries "m_start" and
2624  *	"m_end" for a run of contiguous physical pages that satisfy the
2625  *	specified conditions, and return the lowest page in the run.  The
2626  *	specified "alignment" determines the alignment of the lowest physical
2627  *	page in the run.  If the specified "boundary" is non-zero, then the
2628  *	run of physical pages cannot span a physical address that is a
2629  *	multiple of "boundary".
2630  *
2631  *	"m_end" is never dereferenced, so it need not point to a vm_page
2632  *	structure within vm_page_array[].
2633  *
2634  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2635  *	span a hole (or discontiguity) in the physical address space.  Both
2636  *	"alignment" and "boundary" must be a power of two.
2637  */
2638 static vm_page_t
2639 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2640     u_long alignment, vm_paddr_t boundary, int options)
2641 {
2642 	vm_object_t object;
2643 	vm_paddr_t pa;
2644 	vm_page_t m, m_run;
2645 #if VM_NRESERVLEVEL > 0
2646 	int level;
2647 #endif
2648 	int m_inc, order, run_ext, run_len;
2649 
2650 	KASSERT(npages > 0, ("npages is 0"));
2651 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2652 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2653 	m_run = NULL;
2654 	run_len = 0;
2655 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2656 		KASSERT((m->flags & PG_MARKER) == 0,
2657 		    ("page %p is PG_MARKER", m));
2658 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2659 		    ("fictitious page %p has invalid ref count", m));
2660 
2661 		/*
2662 		 * If the current page would be the start of a run, check its
2663 		 * physical address against the end, alignment, and boundary
2664 		 * conditions.  If it doesn't satisfy these conditions, either
2665 		 * terminate the scan or advance to the next page that
2666 		 * satisfies the failed condition.
2667 		 */
2668 		if (run_len == 0) {
2669 			KASSERT(m_run == NULL, ("m_run != NULL"));
2670 			if (m + npages > m_end)
2671 				break;
2672 			pa = VM_PAGE_TO_PHYS(m);
2673 			if (!vm_addr_align_ok(pa, alignment)) {
2674 				m_inc = atop(roundup2(pa, alignment) - pa);
2675 				continue;
2676 			}
2677 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2678 				m_inc = atop(roundup2(pa, boundary) - pa);
2679 				continue;
2680 			}
2681 		} else
2682 			KASSERT(m_run != NULL, ("m_run == NULL"));
2683 
2684 retry:
2685 		m_inc = 1;
2686 		if (vm_page_wired(m))
2687 			run_ext = 0;
2688 #if VM_NRESERVLEVEL > 0
2689 		else if ((level = vm_reserv_level(m)) >= 0 &&
2690 		    (options & VPSC_NORESERV) != 0) {
2691 			run_ext = 0;
2692 			/* Advance to the end of the reservation. */
2693 			pa = VM_PAGE_TO_PHYS(m);
2694 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2695 			    pa);
2696 		}
2697 #endif
2698 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2699 			/*
2700 			 * The page is considered eligible for relocation if
2701 			 * and only if it could be laundered or reclaimed by
2702 			 * the page daemon.
2703 			 */
2704 			VM_OBJECT_RLOCK(object);
2705 			if (object != m->object) {
2706 				VM_OBJECT_RUNLOCK(object);
2707 				goto retry;
2708 			}
2709 			/* Don't care: PG_NODUMP, PG_ZERO. */
2710 			if ((object->flags & OBJ_SWAP) == 0 &&
2711 			    object->type != OBJT_VNODE) {
2712 				run_ext = 0;
2713 #if VM_NRESERVLEVEL > 0
2714 			} else if ((options & VPSC_NOSUPER) != 0 &&
2715 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2716 				run_ext = 0;
2717 				/* Advance to the end of the superpage. */
2718 				pa = VM_PAGE_TO_PHYS(m);
2719 				m_inc = atop(roundup2(pa + 1,
2720 				    vm_reserv_size(level)) - pa);
2721 #endif
2722 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2723 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2724 				/*
2725 				 * The page is allocated but eligible for
2726 				 * relocation.  Extend the current run by one
2727 				 * page.
2728 				 */
2729 				KASSERT(pmap_page_get_memattr(m) ==
2730 				    VM_MEMATTR_DEFAULT,
2731 				    ("page %p has an unexpected memattr", m));
2732 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2733 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2734 				    ("page %p has unexpected oflags", m));
2735 				/* Don't care: PGA_NOSYNC. */
2736 				run_ext = 1;
2737 			} else
2738 				run_ext = 0;
2739 			VM_OBJECT_RUNLOCK(object);
2740 #if VM_NRESERVLEVEL > 0
2741 		} else if (level >= 0) {
2742 			/*
2743 			 * The page is reserved but not yet allocated.  In
2744 			 * other words, it is still free.  Extend the current
2745 			 * run by one page.
2746 			 */
2747 			run_ext = 1;
2748 #endif
2749 		} else if ((order = m->order) < VM_NFREEORDER) {
2750 			/*
2751 			 * The page is enqueued in the physical memory
2752 			 * allocator's free page queues.  Moreover, it is the
2753 			 * first page in a power-of-two-sized run of
2754 			 * contiguous free pages.  Add these pages to the end
2755 			 * of the current run, and jump ahead.
2756 			 */
2757 			run_ext = 1 << order;
2758 			m_inc = 1 << order;
2759 		} else {
2760 			/*
2761 			 * Skip the page for one of the following reasons: (1)
2762 			 * It is enqueued in the physical memory allocator's
2763 			 * free page queues.  However, it is not the first
2764 			 * page in a run of contiguous free pages.  (This case
2765 			 * rarely occurs because the scan is performed in
2766 			 * ascending order.) (2) It is not reserved, and it is
2767 			 * transitioning from free to allocated.  (Conversely,
2768 			 * the transition from allocated to free for managed
2769 			 * pages is blocked by the page busy lock.) (3) It is
2770 			 * allocated but not contained by an object and not
2771 			 * wired, e.g., allocated by Xen's balloon driver.
2772 			 */
2773 			run_ext = 0;
2774 		}
2775 
2776 		/*
2777 		 * Extend or reset the current run of pages.
2778 		 */
2779 		if (run_ext > 0) {
2780 			if (run_len == 0)
2781 				m_run = m;
2782 			run_len += run_ext;
2783 		} else {
2784 			if (run_len > 0) {
2785 				m_run = NULL;
2786 				run_len = 0;
2787 			}
2788 		}
2789 	}
2790 	if (run_len >= npages)
2791 		return (m_run);
2792 	return (NULL);
2793 }
2794 
2795 /*
2796  *	vm_page_reclaim_run:
2797  *
2798  *	Try to relocate each of the allocated virtual pages within the
2799  *	specified run of physical pages to a new physical address.  Free the
2800  *	physical pages underlying the relocated virtual pages.  A virtual page
2801  *	is relocatable if and only if it could be laundered or reclaimed by
2802  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2803  *	physical address above "high".
2804  *
2805  *	Returns 0 if every physical page within the run was already free or
2806  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2807  *	value indicating why the last attempt to relocate a virtual page was
2808  *	unsuccessful.
2809  *
2810  *	"req_class" must be an allocation class.
2811  */
2812 static int
2813 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2814     vm_paddr_t high)
2815 {
2816 	struct vm_domain *vmd;
2817 	struct spglist free;
2818 	vm_object_t object;
2819 	vm_paddr_t pa;
2820 	vm_page_t m, m_end, m_new;
2821 	int error, order, req;
2822 
2823 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2824 	    ("req_class is not an allocation class"));
2825 	SLIST_INIT(&free);
2826 	error = 0;
2827 	m = m_run;
2828 	m_end = m_run + npages;
2829 	for (; error == 0 && m < m_end; m++) {
2830 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2831 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2832 
2833 		/*
2834 		 * Racily check for wirings.  Races are handled once the object
2835 		 * lock is held and the page is unmapped.
2836 		 */
2837 		if (vm_page_wired(m))
2838 			error = EBUSY;
2839 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2840 			/*
2841 			 * The page is relocated if and only if it could be
2842 			 * laundered or reclaimed by the page daemon.
2843 			 */
2844 			VM_OBJECT_WLOCK(object);
2845 			/* Don't care: PG_NODUMP, PG_ZERO. */
2846 			if (m->object != object ||
2847 			    ((object->flags & OBJ_SWAP) == 0 &&
2848 			    object->type != OBJT_VNODE))
2849 				error = EINVAL;
2850 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2851 				error = EINVAL;
2852 			else if (vm_page_queue(m) != PQ_NONE &&
2853 			    vm_page_tryxbusy(m) != 0) {
2854 				if (vm_page_wired(m)) {
2855 					vm_page_xunbusy(m);
2856 					error = EBUSY;
2857 					goto unlock;
2858 				}
2859 				KASSERT(pmap_page_get_memattr(m) ==
2860 				    VM_MEMATTR_DEFAULT,
2861 				    ("page %p has an unexpected memattr", m));
2862 				KASSERT(m->oflags == 0,
2863 				    ("page %p has unexpected oflags", m));
2864 				/* Don't care: PGA_NOSYNC. */
2865 				if (!vm_page_none_valid(m)) {
2866 					/*
2867 					 * First, try to allocate a new page
2868 					 * that is above "high".  Failing
2869 					 * that, try to allocate a new page
2870 					 * that is below "m_run".  Allocate
2871 					 * the new page between the end of
2872 					 * "m_run" and "high" only as a last
2873 					 * resort.
2874 					 */
2875 					req = req_class;
2876 					if ((m->flags & PG_NODUMP) != 0)
2877 						req |= VM_ALLOC_NODUMP;
2878 					if (trunc_page(high) !=
2879 					    ~(vm_paddr_t)PAGE_MASK) {
2880 						m_new =
2881 						    vm_page_alloc_noobj_contig(
2882 						    req, 1, round_page(high),
2883 						    ~(vm_paddr_t)0, PAGE_SIZE,
2884 						    0, VM_MEMATTR_DEFAULT);
2885 					} else
2886 						m_new = NULL;
2887 					if (m_new == NULL) {
2888 						pa = VM_PAGE_TO_PHYS(m_run);
2889 						m_new =
2890 						    vm_page_alloc_noobj_contig(
2891 						    req, 1, 0, pa - 1,
2892 						    PAGE_SIZE, 0,
2893 						    VM_MEMATTR_DEFAULT);
2894 					}
2895 					if (m_new == NULL) {
2896 						pa += ptoa(npages);
2897 						m_new =
2898 						    vm_page_alloc_noobj_contig(
2899 						    req, 1, pa, high, PAGE_SIZE,
2900 						    0, VM_MEMATTR_DEFAULT);
2901 					}
2902 					if (m_new == NULL) {
2903 						vm_page_xunbusy(m);
2904 						error = ENOMEM;
2905 						goto unlock;
2906 					}
2907 
2908 					/*
2909 					 * Unmap the page and check for new
2910 					 * wirings that may have been acquired
2911 					 * through a pmap lookup.
2912 					 */
2913 					if (object->ref_count != 0 &&
2914 					    !vm_page_try_remove_all(m)) {
2915 						vm_page_xunbusy(m);
2916 						vm_page_free(m_new);
2917 						error = EBUSY;
2918 						goto unlock;
2919 					}
2920 
2921 					/*
2922 					 * Replace "m" with the new page.  For
2923 					 * vm_page_replace(), "m" must be busy
2924 					 * and dequeued.  Finally, change "m"
2925 					 * as if vm_page_free() was called.
2926 					 */
2927 					m_new->a.flags = m->a.flags &
2928 					    ~PGA_QUEUE_STATE_MASK;
2929 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2930 					    ("page %p is managed", m_new));
2931 					m_new->oflags = 0;
2932 					pmap_copy_page(m, m_new);
2933 					m_new->valid = m->valid;
2934 					m_new->dirty = m->dirty;
2935 					m->flags &= ~PG_ZERO;
2936 					vm_page_dequeue(m);
2937 					if (vm_page_replace_hold(m_new, object,
2938 					    m->pindex, m) &&
2939 					    vm_page_free_prep(m))
2940 						SLIST_INSERT_HEAD(&free, m,
2941 						    plinks.s.ss);
2942 
2943 					/*
2944 					 * The new page must be deactivated
2945 					 * before the object is unlocked.
2946 					 */
2947 					vm_page_deactivate(m_new);
2948 				} else {
2949 					m->flags &= ~PG_ZERO;
2950 					vm_page_dequeue(m);
2951 					if (vm_page_free_prep(m))
2952 						SLIST_INSERT_HEAD(&free, m,
2953 						    plinks.s.ss);
2954 					KASSERT(m->dirty == 0,
2955 					    ("page %p is dirty", m));
2956 				}
2957 			} else
2958 				error = EBUSY;
2959 unlock:
2960 			VM_OBJECT_WUNLOCK(object);
2961 		} else {
2962 			MPASS(vm_page_domain(m) == domain);
2963 			vmd = VM_DOMAIN(domain);
2964 			vm_domain_free_lock(vmd);
2965 			order = m->order;
2966 			if (order < VM_NFREEORDER) {
2967 				/*
2968 				 * The page is enqueued in the physical memory
2969 				 * allocator's free page queues.  Moreover, it
2970 				 * is the first page in a power-of-two-sized
2971 				 * run of contiguous free pages.  Jump ahead
2972 				 * to the last page within that run, and
2973 				 * continue from there.
2974 				 */
2975 				m += (1 << order) - 1;
2976 			}
2977 #if VM_NRESERVLEVEL > 0
2978 			else if (vm_reserv_is_page_free(m))
2979 				order = 0;
2980 #endif
2981 			vm_domain_free_unlock(vmd);
2982 			if (order == VM_NFREEORDER)
2983 				error = EINVAL;
2984 		}
2985 	}
2986 	if ((m = SLIST_FIRST(&free)) != NULL) {
2987 		int cnt;
2988 
2989 		vmd = VM_DOMAIN(domain);
2990 		cnt = 0;
2991 		vm_domain_free_lock(vmd);
2992 		do {
2993 			MPASS(vm_page_domain(m) == domain);
2994 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2995 			vm_phys_free_pages(m, 0);
2996 			cnt++;
2997 		} while ((m = SLIST_FIRST(&free)) != NULL);
2998 		vm_domain_free_unlock(vmd);
2999 		vm_domain_freecnt_inc(vmd, cnt);
3000 	}
3001 	return (error);
3002 }
3003 
3004 #define	NRUNS	16
3005 
3006 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3007 
3008 #define	MIN_RECLAIM	8
3009 
3010 /*
3011  *	vm_page_reclaim_contig:
3012  *
3013  *	Reclaim allocated, contiguous physical memory satisfying the specified
3014  *	conditions by relocating the virtual pages using that physical memory.
3015  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3016  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3017  *	currently able to reclaim the requested number of pages.  Since
3018  *	relocation requires the allocation of physical pages, reclamation may
3019  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3020  *	fails in this manner, callers are expected to perform vm_wait() before
3021  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3022  *
3023  *	The caller must always specify an allocation class through "req".
3024  *
3025  *	allocation classes:
3026  *	VM_ALLOC_NORMAL		normal process request
3027  *	VM_ALLOC_SYSTEM		system *really* needs a page
3028  *	VM_ALLOC_INTERRUPT	interrupt time request
3029  *
3030  *	The optional allocation flags are ignored.
3031  *
3032  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3033  *	must be a power of two.
3034  */
3035 int
3036 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3037     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3038     int desired_runs)
3039 {
3040 	struct vm_domain *vmd;
3041 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3042 	u_long count, minalign, reclaimed;
3043 	int error, i, min_reclaim, nruns, options, req_class;
3044 	int segind, start_segind;
3045 	int ret;
3046 
3047 	KASSERT(npages > 0, ("npages is 0"));
3048 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3049 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3050 
3051 	ret = ENOMEM;
3052 
3053 	/*
3054 	 * If the caller wants to reclaim multiple runs, try to allocate
3055 	 * space to store the runs.  If that fails, fall back to the old
3056 	 * behavior of just reclaiming MIN_RECLAIM pages.
3057 	 */
3058 	if (desired_runs > 1)
3059 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3060 		    M_TEMP, M_NOWAIT);
3061 	else
3062 		m_runs = NULL;
3063 
3064 	if (m_runs == NULL) {
3065 		m_runs = _m_runs;
3066 		nruns = NRUNS;
3067 	} else {
3068 		nruns = NRUNS + desired_runs - 1;
3069 	}
3070 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3071 
3072 	/*
3073 	 * The caller will attempt an allocation after some runs have been
3074 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3075 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3076 	 * power of two or maximum chunk size, and ensure that each run is
3077 	 * suitably aligned.
3078 	 */
3079 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3080 	npages = roundup2(npages, minalign);
3081 	if (alignment < ptoa(minalign))
3082 		alignment = ptoa(minalign);
3083 
3084 	/*
3085 	 * The page daemon is allowed to dig deeper into the free page list.
3086 	 */
3087 	req_class = req & VM_ALLOC_CLASS_MASK;
3088 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3089 		req_class = VM_ALLOC_SYSTEM;
3090 
3091 	start_segind = vm_phys_lookup_segind(low);
3092 
3093 	/*
3094 	 * Return if the number of free pages cannot satisfy the requested
3095 	 * allocation.
3096 	 */
3097 	vmd = VM_DOMAIN(domain);
3098 	count = vmd->vmd_free_count;
3099 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3100 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3101 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3102 		goto done;
3103 
3104 	/*
3105 	 * Scan up to three times, relaxing the restrictions ("options") on
3106 	 * the reclamation of reservations and superpages each time.
3107 	 */
3108 	for (options = VPSC_NORESERV;;) {
3109 		bool phys_range_exists = false;
3110 
3111 		/*
3112 		 * Find the highest runs that satisfy the given constraints
3113 		 * and restrictions, and record them in "m_runs".
3114 		 */
3115 		count = 0;
3116 		segind = start_segind;
3117 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3118 		    npages, low, high)) != -1) {
3119 			phys_range_exists = true;
3120 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3121 			    bounds[1], alignment, boundary, options))) {
3122 				bounds[0] = m_run + npages;
3123 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3124 				count++;
3125 			}
3126 			segind++;
3127 		}
3128 
3129 		if (!phys_range_exists) {
3130 			ret = ERANGE;
3131 			goto done;
3132 		}
3133 
3134 		/*
3135 		 * Reclaim the highest runs in LIFO (descending) order until
3136 		 * the number of reclaimed pages, "reclaimed", is at least
3137 		 * "min_reclaim".  Reset "reclaimed" each time because each
3138 		 * reclamation is idempotent, and runs will (likely) recur
3139 		 * from one scan to the next as restrictions are relaxed.
3140 		 */
3141 		reclaimed = 0;
3142 		for (i = 0; count > 0 && i < nruns; i++) {
3143 			count--;
3144 			m_run = m_runs[RUN_INDEX(count, nruns)];
3145 			error = vm_page_reclaim_run(req_class, domain, npages,
3146 			    m_run, high);
3147 			if (error == 0) {
3148 				reclaimed += npages;
3149 				if (reclaimed >= min_reclaim) {
3150 					ret = 0;
3151 					goto done;
3152 				}
3153 			}
3154 		}
3155 
3156 		/*
3157 		 * Either relax the restrictions on the next scan or return if
3158 		 * the last scan had no restrictions.
3159 		 */
3160 		if (options == VPSC_NORESERV)
3161 			options = VPSC_NOSUPER;
3162 		else if (options == VPSC_NOSUPER)
3163 			options = VPSC_ANY;
3164 		else if (options == VPSC_ANY) {
3165 			if (reclaimed != 0)
3166 				ret = 0;
3167 			goto done;
3168 		}
3169 	}
3170 done:
3171 	if (m_runs != _m_runs)
3172 		free(m_runs, M_TEMP);
3173 	return (ret);
3174 }
3175 
3176 int
3177 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3178     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3179 {
3180 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3181 	    alignment, boundary, 1));
3182 }
3183 
3184 int
3185 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3186     u_long alignment, vm_paddr_t boundary)
3187 {
3188 	struct vm_domainset_iter di;
3189 	int domain, ret, status;
3190 
3191 	ret = ERANGE;
3192 
3193 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3194 	do {
3195 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3196 		    high, alignment, boundary);
3197 		if (status == 0)
3198 			return (0);
3199 		else if (status == ERANGE)
3200 			vm_domainset_iter_ignore(&di, domain);
3201 		else {
3202 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3203 			    "from vm_page_reclaim_contig_domain()", status));
3204 			ret = ENOMEM;
3205 		}
3206 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3207 
3208 	return (ret);
3209 }
3210 
3211 /*
3212  * Set the domain in the appropriate page level domainset.
3213  */
3214 void
3215 vm_domain_set(struct vm_domain *vmd)
3216 {
3217 
3218 	mtx_lock(&vm_domainset_lock);
3219 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3220 		vmd->vmd_minset = 1;
3221 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3222 	}
3223 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3224 		vmd->vmd_severeset = 1;
3225 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3226 	}
3227 	mtx_unlock(&vm_domainset_lock);
3228 }
3229 
3230 /*
3231  * Clear the domain from the appropriate page level domainset.
3232  */
3233 void
3234 vm_domain_clear(struct vm_domain *vmd)
3235 {
3236 
3237 	mtx_lock(&vm_domainset_lock);
3238 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3239 		vmd->vmd_minset = 0;
3240 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3241 		if (vm_min_waiters != 0) {
3242 			vm_min_waiters = 0;
3243 			wakeup(&vm_min_domains);
3244 		}
3245 	}
3246 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3247 		vmd->vmd_severeset = 0;
3248 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3249 		if (vm_severe_waiters != 0) {
3250 			vm_severe_waiters = 0;
3251 			wakeup(&vm_severe_domains);
3252 		}
3253 	}
3254 
3255 	/*
3256 	 * If pageout daemon needs pages, then tell it that there are
3257 	 * some free.
3258 	 */
3259 	if (vmd->vmd_pageout_pages_needed &&
3260 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3261 		wakeup(&vmd->vmd_pageout_pages_needed);
3262 		vmd->vmd_pageout_pages_needed = 0;
3263 	}
3264 
3265 	/* See comments in vm_wait_doms(). */
3266 	if (vm_pageproc_waiters) {
3267 		vm_pageproc_waiters = 0;
3268 		wakeup(&vm_pageproc_waiters);
3269 	}
3270 	mtx_unlock(&vm_domainset_lock);
3271 }
3272 
3273 /*
3274  * Wait for free pages to exceed the min threshold globally.
3275  */
3276 void
3277 vm_wait_min(void)
3278 {
3279 
3280 	mtx_lock(&vm_domainset_lock);
3281 	while (vm_page_count_min()) {
3282 		vm_min_waiters++;
3283 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3284 	}
3285 	mtx_unlock(&vm_domainset_lock);
3286 }
3287 
3288 /*
3289  * Wait for free pages to exceed the severe threshold globally.
3290  */
3291 void
3292 vm_wait_severe(void)
3293 {
3294 
3295 	mtx_lock(&vm_domainset_lock);
3296 	while (vm_page_count_severe()) {
3297 		vm_severe_waiters++;
3298 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3299 		    "vmwait", 0);
3300 	}
3301 	mtx_unlock(&vm_domainset_lock);
3302 }
3303 
3304 u_int
3305 vm_wait_count(void)
3306 {
3307 
3308 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3309 }
3310 
3311 int
3312 vm_wait_doms(const domainset_t *wdoms, int mflags)
3313 {
3314 	int error;
3315 
3316 	error = 0;
3317 
3318 	/*
3319 	 * We use racey wakeup synchronization to avoid expensive global
3320 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3321 	 * To handle this, we only sleep for one tick in this instance.  It
3322 	 * is expected that most allocations for the pageproc will come from
3323 	 * kmem or vm_page_grab* which will use the more specific and
3324 	 * race-free vm_wait_domain().
3325 	 */
3326 	if (curproc == pageproc) {
3327 		mtx_lock(&vm_domainset_lock);
3328 		vm_pageproc_waiters++;
3329 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3330 		    PVM | PDROP | mflags, "pageprocwait", 1);
3331 	} else {
3332 		/*
3333 		 * XXX Ideally we would wait only until the allocation could
3334 		 * be satisfied.  This condition can cause new allocators to
3335 		 * consume all freed pages while old allocators wait.
3336 		 */
3337 		mtx_lock(&vm_domainset_lock);
3338 		if (vm_page_count_min_set(wdoms)) {
3339 			if (pageproc == NULL)
3340 				panic("vm_wait in early boot");
3341 			vm_min_waiters++;
3342 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3343 			    PVM | PDROP | mflags, "vmwait", 0);
3344 		} else
3345 			mtx_unlock(&vm_domainset_lock);
3346 	}
3347 	return (error);
3348 }
3349 
3350 /*
3351  *	vm_wait_domain:
3352  *
3353  *	Sleep until free pages are available for allocation.
3354  *	- Called in various places after failed memory allocations.
3355  */
3356 void
3357 vm_wait_domain(int domain)
3358 {
3359 	struct vm_domain *vmd;
3360 	domainset_t wdom;
3361 
3362 	vmd = VM_DOMAIN(domain);
3363 	vm_domain_free_assert_unlocked(vmd);
3364 
3365 	if (curproc == pageproc) {
3366 		mtx_lock(&vm_domainset_lock);
3367 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3368 			vmd->vmd_pageout_pages_needed = 1;
3369 			msleep(&vmd->vmd_pageout_pages_needed,
3370 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3371 		} else
3372 			mtx_unlock(&vm_domainset_lock);
3373 	} else {
3374 		DOMAINSET_ZERO(&wdom);
3375 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3376 		vm_wait_doms(&wdom, 0);
3377 	}
3378 }
3379 
3380 static int
3381 vm_wait_flags(vm_object_t obj, int mflags)
3382 {
3383 	struct domainset *d;
3384 
3385 	d = NULL;
3386 
3387 	/*
3388 	 * Carefully fetch pointers only once: the struct domainset
3389 	 * itself is ummutable but the pointer might change.
3390 	 */
3391 	if (obj != NULL)
3392 		d = obj->domain.dr_policy;
3393 	if (d == NULL)
3394 		d = curthread->td_domain.dr_policy;
3395 
3396 	return (vm_wait_doms(&d->ds_mask, mflags));
3397 }
3398 
3399 /*
3400  *	vm_wait:
3401  *
3402  *	Sleep until free pages are available for allocation in the
3403  *	affinity domains of the obj.  If obj is NULL, the domain set
3404  *	for the calling thread is used.
3405  *	Called in various places after failed memory allocations.
3406  */
3407 void
3408 vm_wait(vm_object_t obj)
3409 {
3410 	(void)vm_wait_flags(obj, 0);
3411 }
3412 
3413 int
3414 vm_wait_intr(vm_object_t obj)
3415 {
3416 	return (vm_wait_flags(obj, PCATCH));
3417 }
3418 
3419 /*
3420  *	vm_domain_alloc_fail:
3421  *
3422  *	Called when a page allocation function fails.  Informs the
3423  *	pagedaemon and performs the requested wait.  Requires the
3424  *	domain_free and object lock on entry.  Returns with the
3425  *	object lock held and free lock released.  Returns an error when
3426  *	retry is necessary.
3427  *
3428  */
3429 static int
3430 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3431 {
3432 
3433 	vm_domain_free_assert_unlocked(vmd);
3434 
3435 	atomic_add_int(&vmd->vmd_pageout_deficit,
3436 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3437 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3438 		if (object != NULL)
3439 			VM_OBJECT_WUNLOCK(object);
3440 		vm_wait_domain(vmd->vmd_domain);
3441 		if (object != NULL)
3442 			VM_OBJECT_WLOCK(object);
3443 		if (req & VM_ALLOC_WAITOK)
3444 			return (EAGAIN);
3445 	}
3446 
3447 	return (0);
3448 }
3449 
3450 /*
3451  *	vm_waitpfault:
3452  *
3453  *	Sleep until free pages are available for allocation.
3454  *	- Called only in vm_fault so that processes page faulting
3455  *	  can be easily tracked.
3456  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3457  *	  processes will be able to grab memory first.  Do not change
3458  *	  this balance without careful testing first.
3459  */
3460 void
3461 vm_waitpfault(struct domainset *dset, int timo)
3462 {
3463 
3464 	/*
3465 	 * XXX Ideally we would wait only until the allocation could
3466 	 * be satisfied.  This condition can cause new allocators to
3467 	 * consume all freed pages while old allocators wait.
3468 	 */
3469 	mtx_lock(&vm_domainset_lock);
3470 	if (vm_page_count_min_set(&dset->ds_mask)) {
3471 		vm_min_waiters++;
3472 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3473 		    "pfault", timo);
3474 	} else
3475 		mtx_unlock(&vm_domainset_lock);
3476 }
3477 
3478 static struct vm_pagequeue *
3479 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3480 {
3481 
3482 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3483 }
3484 
3485 #ifdef INVARIANTS
3486 static struct vm_pagequeue *
3487 vm_page_pagequeue(vm_page_t m)
3488 {
3489 
3490 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3491 }
3492 #endif
3493 
3494 static __always_inline bool
3495 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3496 {
3497 	vm_page_astate_t tmp;
3498 
3499 	tmp = *old;
3500 	do {
3501 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3502 			return (true);
3503 		counter_u64_add(pqstate_commit_retries, 1);
3504 	} while (old->_bits == tmp._bits);
3505 
3506 	return (false);
3507 }
3508 
3509 /*
3510  * Do the work of committing a queue state update that moves the page out of
3511  * its current queue.
3512  */
3513 static bool
3514 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3515     vm_page_astate_t *old, vm_page_astate_t new)
3516 {
3517 	vm_page_t next;
3518 
3519 	vm_pagequeue_assert_locked(pq);
3520 	KASSERT(vm_page_pagequeue(m) == pq,
3521 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3522 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3523 	    ("%s: invalid queue indices %d %d",
3524 	    __func__, old->queue, new.queue));
3525 
3526 	/*
3527 	 * Once the queue index of the page changes there is nothing
3528 	 * synchronizing with further updates to the page's physical
3529 	 * queue state.  Therefore we must speculatively remove the page
3530 	 * from the queue now and be prepared to roll back if the queue
3531 	 * state update fails.  If the page is not physically enqueued then
3532 	 * we just update its queue index.
3533 	 */
3534 	if ((old->flags & PGA_ENQUEUED) != 0) {
3535 		new.flags &= ~PGA_ENQUEUED;
3536 		next = TAILQ_NEXT(m, plinks.q);
3537 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3538 		vm_pagequeue_cnt_dec(pq);
3539 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3540 			if (next == NULL)
3541 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3542 			else
3543 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3544 			vm_pagequeue_cnt_inc(pq);
3545 			return (false);
3546 		} else {
3547 			return (true);
3548 		}
3549 	} else {
3550 		return (vm_page_pqstate_fcmpset(m, old, new));
3551 	}
3552 }
3553 
3554 static bool
3555 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3556     vm_page_astate_t new)
3557 {
3558 	struct vm_pagequeue *pq;
3559 	vm_page_astate_t as;
3560 	bool ret;
3561 
3562 	pq = _vm_page_pagequeue(m, old->queue);
3563 
3564 	/*
3565 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3566 	 * corresponding page queue lock is held.
3567 	 */
3568 	vm_pagequeue_lock(pq);
3569 	as = vm_page_astate_load(m);
3570 	if (__predict_false(as._bits != old->_bits)) {
3571 		*old = as;
3572 		ret = false;
3573 	} else {
3574 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3575 	}
3576 	vm_pagequeue_unlock(pq);
3577 	return (ret);
3578 }
3579 
3580 /*
3581  * Commit a queue state update that enqueues or requeues a page.
3582  */
3583 static bool
3584 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3585     vm_page_astate_t *old, vm_page_astate_t new)
3586 {
3587 	struct vm_domain *vmd;
3588 
3589 	vm_pagequeue_assert_locked(pq);
3590 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3591 	    ("%s: invalid queue indices %d %d",
3592 	    __func__, old->queue, new.queue));
3593 
3594 	new.flags |= PGA_ENQUEUED;
3595 	if (!vm_page_pqstate_fcmpset(m, old, new))
3596 		return (false);
3597 
3598 	if ((old->flags & PGA_ENQUEUED) != 0)
3599 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3600 	else
3601 		vm_pagequeue_cnt_inc(pq);
3602 
3603 	/*
3604 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3605 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3606 	 * applied, even if it was set first.
3607 	 */
3608 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3609 		vmd = vm_pagequeue_domain(m);
3610 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3611 		    ("%s: invalid page queue for page %p", __func__, m));
3612 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3613 	} else {
3614 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3615 	}
3616 	return (true);
3617 }
3618 
3619 /*
3620  * Commit a queue state update that encodes a request for a deferred queue
3621  * operation.
3622  */
3623 static bool
3624 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3625     vm_page_astate_t new)
3626 {
3627 
3628 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3629 	    ("%s: invalid state, queue %d flags %x",
3630 	    __func__, new.queue, new.flags));
3631 
3632 	if (old->_bits != new._bits &&
3633 	    !vm_page_pqstate_fcmpset(m, old, new))
3634 		return (false);
3635 	vm_page_pqbatch_submit(m, new.queue);
3636 	return (true);
3637 }
3638 
3639 /*
3640  * A generic queue state update function.  This handles more cases than the
3641  * specialized functions above.
3642  */
3643 bool
3644 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3645 {
3646 
3647 	if (old->_bits == new._bits)
3648 		return (true);
3649 
3650 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3651 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3652 			return (false);
3653 		if (new.queue != PQ_NONE)
3654 			vm_page_pqbatch_submit(m, new.queue);
3655 	} else {
3656 		if (!vm_page_pqstate_fcmpset(m, old, new))
3657 			return (false);
3658 		if (new.queue != PQ_NONE &&
3659 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3660 			vm_page_pqbatch_submit(m, new.queue);
3661 	}
3662 	return (true);
3663 }
3664 
3665 /*
3666  * Apply deferred queue state updates to a page.
3667  */
3668 static inline void
3669 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3670 {
3671 	vm_page_astate_t new, old;
3672 
3673 	CRITICAL_ASSERT(curthread);
3674 	vm_pagequeue_assert_locked(pq);
3675 	KASSERT(queue < PQ_COUNT,
3676 	    ("%s: invalid queue index %d", __func__, queue));
3677 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3678 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3679 
3680 	for (old = vm_page_astate_load(m);;) {
3681 		if (__predict_false(old.queue != queue ||
3682 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3683 			counter_u64_add(queue_nops, 1);
3684 			break;
3685 		}
3686 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3687 		    ("%s: page %p is unmanaged", __func__, m));
3688 
3689 		new = old;
3690 		if ((old.flags & PGA_DEQUEUE) != 0) {
3691 			new.flags &= ~PGA_QUEUE_OP_MASK;
3692 			new.queue = PQ_NONE;
3693 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3694 			    m, &old, new))) {
3695 				counter_u64_add(queue_ops, 1);
3696 				break;
3697 			}
3698 		} else {
3699 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3700 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3701 			    m, &old, new))) {
3702 				counter_u64_add(queue_ops, 1);
3703 				break;
3704 			}
3705 		}
3706 	}
3707 }
3708 
3709 static void
3710 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3711     uint8_t queue)
3712 {
3713 	int i;
3714 
3715 	for (i = 0; i < bq->bq_cnt; i++)
3716 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3717 	vm_batchqueue_init(bq);
3718 }
3719 
3720 /*
3721  *	vm_page_pqbatch_submit:		[ internal use only ]
3722  *
3723  *	Enqueue a page in the specified page queue's batched work queue.
3724  *	The caller must have encoded the requested operation in the page
3725  *	structure's a.flags field.
3726  */
3727 void
3728 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3729 {
3730 	struct vm_batchqueue *bq;
3731 	struct vm_pagequeue *pq;
3732 	int domain, slots_remaining;
3733 
3734 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3735 
3736 	domain = vm_page_domain(m);
3737 	critical_enter();
3738 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3739 	slots_remaining = vm_batchqueue_insert(bq, m);
3740 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3741 		/* keep building the bq */
3742 		critical_exit();
3743 		return;
3744 	} else if (slots_remaining > 0 ) {
3745 		/* Try to process the bq if we can get the lock */
3746 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3747 		if (vm_pagequeue_trylock(pq)) {
3748 			vm_pqbatch_process(pq, bq, queue);
3749 			vm_pagequeue_unlock(pq);
3750 		}
3751 		critical_exit();
3752 		return;
3753 	}
3754 	critical_exit();
3755 
3756 	/* if we make it here, the bq is full so wait for the lock */
3757 
3758 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3759 	vm_pagequeue_lock(pq);
3760 	critical_enter();
3761 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3762 	vm_pqbatch_process(pq, bq, queue);
3763 	vm_pqbatch_process_page(pq, m, queue);
3764 	vm_pagequeue_unlock(pq);
3765 	critical_exit();
3766 }
3767 
3768 /*
3769  *	vm_page_pqbatch_drain:		[ internal use only ]
3770  *
3771  *	Force all per-CPU page queue batch queues to be drained.  This is
3772  *	intended for use in severe memory shortages, to ensure that pages
3773  *	do not remain stuck in the batch queues.
3774  */
3775 void
3776 vm_page_pqbatch_drain(void)
3777 {
3778 	struct thread *td;
3779 	struct vm_domain *vmd;
3780 	struct vm_pagequeue *pq;
3781 	int cpu, domain, queue;
3782 
3783 	td = curthread;
3784 	CPU_FOREACH(cpu) {
3785 		thread_lock(td);
3786 		sched_bind(td, cpu);
3787 		thread_unlock(td);
3788 
3789 		for (domain = 0; domain < vm_ndomains; domain++) {
3790 			vmd = VM_DOMAIN(domain);
3791 			for (queue = 0; queue < PQ_COUNT; queue++) {
3792 				pq = &vmd->vmd_pagequeues[queue];
3793 				vm_pagequeue_lock(pq);
3794 				critical_enter();
3795 				vm_pqbatch_process(pq,
3796 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3797 				critical_exit();
3798 				vm_pagequeue_unlock(pq);
3799 			}
3800 		}
3801 	}
3802 	thread_lock(td);
3803 	sched_unbind(td);
3804 	thread_unlock(td);
3805 }
3806 
3807 /*
3808  *	vm_page_dequeue_deferred:	[ internal use only ]
3809  *
3810  *	Request removal of the given page from its current page
3811  *	queue.  Physical removal from the queue may be deferred
3812  *	indefinitely.
3813  */
3814 void
3815 vm_page_dequeue_deferred(vm_page_t m)
3816 {
3817 	vm_page_astate_t new, old;
3818 
3819 	old = vm_page_astate_load(m);
3820 	do {
3821 		if (old.queue == PQ_NONE) {
3822 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3823 			    ("%s: page %p has unexpected queue state",
3824 			    __func__, m));
3825 			break;
3826 		}
3827 		new = old;
3828 		new.flags |= PGA_DEQUEUE;
3829 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3830 }
3831 
3832 /*
3833  *	vm_page_dequeue:
3834  *
3835  *	Remove the page from whichever page queue it's in, if any, before
3836  *	returning.
3837  */
3838 void
3839 vm_page_dequeue(vm_page_t m)
3840 {
3841 	vm_page_astate_t new, old;
3842 
3843 	old = vm_page_astate_load(m);
3844 	do {
3845 		if (old.queue == PQ_NONE) {
3846 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3847 			    ("%s: page %p has unexpected queue state",
3848 			    __func__, m));
3849 			break;
3850 		}
3851 		new = old;
3852 		new.flags &= ~PGA_QUEUE_OP_MASK;
3853 		new.queue = PQ_NONE;
3854 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3855 
3856 }
3857 
3858 /*
3859  * Schedule the given page for insertion into the specified page queue.
3860  * Physical insertion of the page may be deferred indefinitely.
3861  */
3862 static void
3863 vm_page_enqueue(vm_page_t m, uint8_t queue)
3864 {
3865 
3866 	KASSERT(m->a.queue == PQ_NONE &&
3867 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3868 	    ("%s: page %p is already enqueued", __func__, m));
3869 	KASSERT(m->ref_count > 0,
3870 	    ("%s: page %p does not carry any references", __func__, m));
3871 
3872 	m->a.queue = queue;
3873 	if ((m->a.flags & PGA_REQUEUE) == 0)
3874 		vm_page_aflag_set(m, PGA_REQUEUE);
3875 	vm_page_pqbatch_submit(m, queue);
3876 }
3877 
3878 /*
3879  *	vm_page_free_prep:
3880  *
3881  *	Prepares the given page to be put on the free list,
3882  *	disassociating it from any VM object. The caller may return
3883  *	the page to the free list only if this function returns true.
3884  *
3885  *	The object, if it exists, must be locked, and then the page must
3886  *	be xbusy.  Otherwise the page must be not busied.  A managed
3887  *	page must be unmapped.
3888  */
3889 static bool
3890 vm_page_free_prep(vm_page_t m)
3891 {
3892 
3893 	/*
3894 	 * Synchronize with threads that have dropped a reference to this
3895 	 * page.
3896 	 */
3897 	atomic_thread_fence_acq();
3898 
3899 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3900 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3901 		uint64_t *p;
3902 		int i;
3903 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3904 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3905 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3906 			    m, i, (uintmax_t)*p));
3907 	}
3908 #endif
3909 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3910 		KASSERT(!pmap_page_is_mapped(m),
3911 		    ("vm_page_free_prep: freeing mapped page %p", m));
3912 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3913 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3914 	} else {
3915 		KASSERT(m->a.queue == PQ_NONE,
3916 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3917 	}
3918 	VM_CNT_INC(v_tfree);
3919 
3920 	if (m->object != NULL) {
3921 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3922 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3923 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3924 		    m));
3925 		vm_page_assert_xbusied(m);
3926 
3927 		/*
3928 		 * The object reference can be released without an atomic
3929 		 * operation.
3930 		 */
3931 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3932 		    m->ref_count == VPRC_OBJREF,
3933 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3934 		    m, m->ref_count));
3935 		vm_page_object_remove(m);
3936 		m->ref_count -= VPRC_OBJREF;
3937 	} else
3938 		vm_page_assert_unbusied(m);
3939 
3940 	vm_page_busy_free(m);
3941 
3942 	/*
3943 	 * If fictitious remove object association and
3944 	 * return.
3945 	 */
3946 	if ((m->flags & PG_FICTITIOUS) != 0) {
3947 		KASSERT(m->ref_count == 1,
3948 		    ("fictitious page %p is referenced", m));
3949 		KASSERT(m->a.queue == PQ_NONE,
3950 		    ("fictitious page %p is queued", m));
3951 		return (false);
3952 	}
3953 
3954 	/*
3955 	 * Pages need not be dequeued before they are returned to the physical
3956 	 * memory allocator, but they must at least be marked for a deferred
3957 	 * dequeue.
3958 	 */
3959 	if ((m->oflags & VPO_UNMANAGED) == 0)
3960 		vm_page_dequeue_deferred(m);
3961 
3962 	m->valid = 0;
3963 	vm_page_undirty(m);
3964 
3965 	if (m->ref_count != 0)
3966 		panic("vm_page_free_prep: page %p has references", m);
3967 
3968 	/*
3969 	 * Restore the default memory attribute to the page.
3970 	 */
3971 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3972 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3973 
3974 #if VM_NRESERVLEVEL > 0
3975 	/*
3976 	 * Determine whether the page belongs to a reservation.  If the page was
3977 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3978 	 * as an optimization, we avoid the check in that case.
3979 	 */
3980 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3981 		return (false);
3982 #endif
3983 
3984 	return (true);
3985 }
3986 
3987 /*
3988  *	vm_page_free_toq:
3989  *
3990  *	Returns the given page to the free list, disassociating it
3991  *	from any VM object.
3992  *
3993  *	The object must be locked.  The page must be exclusively busied if it
3994  *	belongs to an object.
3995  */
3996 static void
3997 vm_page_free_toq(vm_page_t m)
3998 {
3999 	struct vm_domain *vmd;
4000 	uma_zone_t zone;
4001 
4002 	if (!vm_page_free_prep(m))
4003 		return;
4004 
4005 	vmd = vm_pagequeue_domain(m);
4006 	zone = vmd->vmd_pgcache[m->pool].zone;
4007 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4008 		uma_zfree(zone, m);
4009 		return;
4010 	}
4011 	vm_domain_free_lock(vmd);
4012 	vm_phys_free_pages(m, 0);
4013 	vm_domain_free_unlock(vmd);
4014 	vm_domain_freecnt_inc(vmd, 1);
4015 }
4016 
4017 /*
4018  *	vm_page_free_pages_toq:
4019  *
4020  *	Returns a list of pages to the free list, disassociating it
4021  *	from any VM object.  In other words, this is equivalent to
4022  *	calling vm_page_free_toq() for each page of a list of VM objects.
4023  */
4024 void
4025 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4026 {
4027 	vm_page_t m;
4028 	int count;
4029 
4030 	if (SLIST_EMPTY(free))
4031 		return;
4032 
4033 	count = 0;
4034 	while ((m = SLIST_FIRST(free)) != NULL) {
4035 		count++;
4036 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4037 		vm_page_free_toq(m);
4038 	}
4039 
4040 	if (update_wire_count)
4041 		vm_wire_sub(count);
4042 }
4043 
4044 /*
4045  * Mark this page as wired down.  For managed pages, this prevents reclamation
4046  * by the page daemon, or when the containing object, if any, is destroyed.
4047  */
4048 void
4049 vm_page_wire(vm_page_t m)
4050 {
4051 	u_int old;
4052 
4053 #ifdef INVARIANTS
4054 	if (m->object != NULL && !vm_page_busied(m) &&
4055 	    !vm_object_busied(m->object))
4056 		VM_OBJECT_ASSERT_LOCKED(m->object);
4057 #endif
4058 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4059 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4060 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4061 
4062 	old = atomic_fetchadd_int(&m->ref_count, 1);
4063 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4064 	    ("vm_page_wire: counter overflow for page %p", m));
4065 	if (VPRC_WIRE_COUNT(old) == 0) {
4066 		if ((m->oflags & VPO_UNMANAGED) == 0)
4067 			vm_page_aflag_set(m, PGA_DEQUEUE);
4068 		vm_wire_add(1);
4069 	}
4070 }
4071 
4072 /*
4073  * Attempt to wire a mapped page following a pmap lookup of that page.
4074  * This may fail if a thread is concurrently tearing down mappings of the page.
4075  * The transient failure is acceptable because it translates to the
4076  * failure of the caller pmap_extract_and_hold(), which should be then
4077  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4078  */
4079 bool
4080 vm_page_wire_mapped(vm_page_t m)
4081 {
4082 	u_int old;
4083 
4084 	old = m->ref_count;
4085 	do {
4086 		KASSERT(old > 0,
4087 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4088 		if ((old & VPRC_BLOCKED) != 0)
4089 			return (false);
4090 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4091 
4092 	if (VPRC_WIRE_COUNT(old) == 0) {
4093 		if ((m->oflags & VPO_UNMANAGED) == 0)
4094 			vm_page_aflag_set(m, PGA_DEQUEUE);
4095 		vm_wire_add(1);
4096 	}
4097 	return (true);
4098 }
4099 
4100 /*
4101  * Release a wiring reference to a managed page.  If the page still belongs to
4102  * an object, update its position in the page queues to reflect the reference.
4103  * If the wiring was the last reference to the page, free the page.
4104  */
4105 static void
4106 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4107 {
4108 	u_int old;
4109 
4110 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4111 	    ("%s: page %p is unmanaged", __func__, m));
4112 
4113 	/*
4114 	 * Update LRU state before releasing the wiring reference.
4115 	 * Use a release store when updating the reference count to
4116 	 * synchronize with vm_page_free_prep().
4117 	 */
4118 	old = m->ref_count;
4119 	do {
4120 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4121 		    ("vm_page_unwire: wire count underflow for page %p", m));
4122 
4123 		if (old > VPRC_OBJREF + 1) {
4124 			/*
4125 			 * The page has at least one other wiring reference.  An
4126 			 * earlier iteration of this loop may have called
4127 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4128 			 * re-set it if necessary.
4129 			 */
4130 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4131 				vm_page_aflag_set(m, PGA_DEQUEUE);
4132 		} else if (old == VPRC_OBJREF + 1) {
4133 			/*
4134 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4135 			 * update the page's queue state to reflect the
4136 			 * reference.  If the page does not belong to an object
4137 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4138 			 * clear leftover queue state.
4139 			 */
4140 			vm_page_release_toq(m, nqueue, noreuse);
4141 		} else if (old == 1) {
4142 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4143 		}
4144 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4145 
4146 	if (VPRC_WIRE_COUNT(old) == 1) {
4147 		vm_wire_sub(1);
4148 		if (old == 1)
4149 			vm_page_free(m);
4150 	}
4151 }
4152 
4153 /*
4154  * Release one wiring of the specified page, potentially allowing it to be
4155  * paged out.
4156  *
4157  * Only managed pages belonging to an object can be paged out.  If the number
4158  * of wirings transitions to zero and the page is eligible for page out, then
4159  * the page is added to the specified paging queue.  If the released wiring
4160  * represented the last reference to the page, the page is freed.
4161  */
4162 void
4163 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4164 {
4165 
4166 	KASSERT(nqueue < PQ_COUNT,
4167 	    ("vm_page_unwire: invalid queue %u request for page %p",
4168 	    nqueue, m));
4169 
4170 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4171 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4172 			vm_page_free(m);
4173 		return;
4174 	}
4175 	vm_page_unwire_managed(m, nqueue, false);
4176 }
4177 
4178 /*
4179  * Unwire a page without (re-)inserting it into a page queue.  It is up
4180  * to the caller to enqueue, requeue, or free the page as appropriate.
4181  * In most cases involving managed pages, vm_page_unwire() should be used
4182  * instead.
4183  */
4184 bool
4185 vm_page_unwire_noq(vm_page_t m)
4186 {
4187 	u_int old;
4188 
4189 	old = vm_page_drop(m, 1);
4190 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4191 	    ("%s: counter underflow for page %p", __func__,  m));
4192 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4193 	    ("%s: missing ref on fictitious page %p", __func__, m));
4194 
4195 	if (VPRC_WIRE_COUNT(old) > 1)
4196 		return (false);
4197 	if ((m->oflags & VPO_UNMANAGED) == 0)
4198 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4199 	vm_wire_sub(1);
4200 	return (true);
4201 }
4202 
4203 /*
4204  * Ensure that the page ends up in the specified page queue.  If the page is
4205  * active or being moved to the active queue, ensure that its act_count is
4206  * at least ACT_INIT but do not otherwise mess with it.
4207  */
4208 static __always_inline void
4209 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4210 {
4211 	vm_page_astate_t old, new;
4212 
4213 	KASSERT(m->ref_count > 0,
4214 	    ("%s: page %p does not carry any references", __func__, m));
4215 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4216 	    ("%s: invalid flags %x", __func__, nflag));
4217 
4218 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4219 		return;
4220 
4221 	old = vm_page_astate_load(m);
4222 	do {
4223 		if ((old.flags & PGA_DEQUEUE) != 0)
4224 			break;
4225 		new = old;
4226 		new.flags &= ~PGA_QUEUE_OP_MASK;
4227 		if (nqueue == PQ_ACTIVE)
4228 			new.act_count = max(old.act_count, ACT_INIT);
4229 		if (old.queue == nqueue) {
4230 			/*
4231 			 * There is no need to requeue pages already in the
4232 			 * active queue.
4233 			 */
4234 			if (nqueue != PQ_ACTIVE ||
4235 			    (old.flags & PGA_ENQUEUED) == 0)
4236 				new.flags |= nflag;
4237 		} else {
4238 			new.flags |= nflag;
4239 			new.queue = nqueue;
4240 		}
4241 	} while (!vm_page_pqstate_commit(m, &old, new));
4242 }
4243 
4244 /*
4245  * Put the specified page on the active list (if appropriate).
4246  */
4247 void
4248 vm_page_activate(vm_page_t m)
4249 {
4250 
4251 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4252 }
4253 
4254 /*
4255  * Move the specified page to the tail of the inactive queue, or requeue
4256  * the page if it is already in the inactive queue.
4257  */
4258 void
4259 vm_page_deactivate(vm_page_t m)
4260 {
4261 
4262 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4263 }
4264 
4265 void
4266 vm_page_deactivate_noreuse(vm_page_t m)
4267 {
4268 
4269 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4270 }
4271 
4272 /*
4273  * Put a page in the laundry, or requeue it if it is already there.
4274  */
4275 void
4276 vm_page_launder(vm_page_t m)
4277 {
4278 
4279 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4280 }
4281 
4282 /*
4283  * Put a page in the PQ_UNSWAPPABLE holding queue.
4284  */
4285 void
4286 vm_page_unswappable(vm_page_t m)
4287 {
4288 
4289 	VM_OBJECT_ASSERT_LOCKED(m->object);
4290 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4291 	    ("page %p already unswappable", m));
4292 
4293 	vm_page_dequeue(m);
4294 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4295 }
4296 
4297 /*
4298  * Release a page back to the page queues in preparation for unwiring.
4299  */
4300 static void
4301 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4302 {
4303 	vm_page_astate_t old, new;
4304 	uint16_t nflag;
4305 
4306 	/*
4307 	 * Use a check of the valid bits to determine whether we should
4308 	 * accelerate reclamation of the page.  The object lock might not be
4309 	 * held here, in which case the check is racy.  At worst we will either
4310 	 * accelerate reclamation of a valid page and violate LRU, or
4311 	 * unnecessarily defer reclamation of an invalid page.
4312 	 *
4313 	 * If we were asked to not cache the page, place it near the head of the
4314 	 * inactive queue so that is reclaimed sooner.
4315 	 */
4316 	if (noreuse || vm_page_none_valid(m)) {
4317 		nqueue = PQ_INACTIVE;
4318 		nflag = PGA_REQUEUE_HEAD;
4319 	} else {
4320 		nflag = PGA_REQUEUE;
4321 	}
4322 
4323 	old = vm_page_astate_load(m);
4324 	do {
4325 		new = old;
4326 
4327 		/*
4328 		 * If the page is already in the active queue and we are not
4329 		 * trying to accelerate reclamation, simply mark it as
4330 		 * referenced and avoid any queue operations.
4331 		 */
4332 		new.flags &= ~PGA_QUEUE_OP_MASK;
4333 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4334 		    (old.flags & PGA_ENQUEUED) != 0)
4335 			new.flags |= PGA_REFERENCED;
4336 		else {
4337 			new.flags |= nflag;
4338 			new.queue = nqueue;
4339 		}
4340 	} while (!vm_page_pqstate_commit(m, &old, new));
4341 }
4342 
4343 /*
4344  * Unwire a page and either attempt to free it or re-add it to the page queues.
4345  */
4346 void
4347 vm_page_release(vm_page_t m, int flags)
4348 {
4349 	vm_object_t object;
4350 
4351 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4352 	    ("vm_page_release: page %p is unmanaged", m));
4353 
4354 	if ((flags & VPR_TRYFREE) != 0) {
4355 		for (;;) {
4356 			object = atomic_load_ptr(&m->object);
4357 			if (object == NULL)
4358 				break;
4359 			/* Depends on type-stability. */
4360 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4361 				break;
4362 			if (object == m->object) {
4363 				vm_page_release_locked(m, flags);
4364 				VM_OBJECT_WUNLOCK(object);
4365 				return;
4366 			}
4367 			VM_OBJECT_WUNLOCK(object);
4368 		}
4369 	}
4370 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4371 }
4372 
4373 /* See vm_page_release(). */
4374 void
4375 vm_page_release_locked(vm_page_t m, int flags)
4376 {
4377 
4378 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4379 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4380 	    ("vm_page_release_locked: page %p is unmanaged", m));
4381 
4382 	if (vm_page_unwire_noq(m)) {
4383 		if ((flags & VPR_TRYFREE) != 0 &&
4384 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4385 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4386 			/*
4387 			 * An unlocked lookup may have wired the page before the
4388 			 * busy lock was acquired, in which case the page must
4389 			 * not be freed.
4390 			 */
4391 			if (__predict_true(!vm_page_wired(m))) {
4392 				vm_page_free(m);
4393 				return;
4394 			}
4395 			vm_page_xunbusy(m);
4396 		} else {
4397 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4398 		}
4399 	}
4400 }
4401 
4402 static bool
4403 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4404 {
4405 	u_int old;
4406 
4407 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4408 	    ("vm_page_try_blocked_op: page %p has no object", m));
4409 	KASSERT(vm_page_busied(m),
4410 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4411 	VM_OBJECT_ASSERT_LOCKED(m->object);
4412 
4413 	old = m->ref_count;
4414 	do {
4415 		KASSERT(old != 0,
4416 		    ("vm_page_try_blocked_op: page %p has no references", m));
4417 		if (VPRC_WIRE_COUNT(old) != 0)
4418 			return (false);
4419 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4420 
4421 	(op)(m);
4422 
4423 	/*
4424 	 * If the object is read-locked, new wirings may be created via an
4425 	 * object lookup.
4426 	 */
4427 	old = vm_page_drop(m, VPRC_BLOCKED);
4428 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4429 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4430 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4431 	    old, m));
4432 	return (true);
4433 }
4434 
4435 /*
4436  * Atomically check for wirings and remove all mappings of the page.
4437  */
4438 bool
4439 vm_page_try_remove_all(vm_page_t m)
4440 {
4441 
4442 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4443 }
4444 
4445 /*
4446  * Atomically check for wirings and remove all writeable mappings of the page.
4447  */
4448 bool
4449 vm_page_try_remove_write(vm_page_t m)
4450 {
4451 
4452 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4453 }
4454 
4455 /*
4456  * vm_page_advise
4457  *
4458  * 	Apply the specified advice to the given page.
4459  */
4460 void
4461 vm_page_advise(vm_page_t m, int advice)
4462 {
4463 
4464 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4465 	vm_page_assert_xbusied(m);
4466 
4467 	if (advice == MADV_FREE)
4468 		/*
4469 		 * Mark the page clean.  This will allow the page to be freed
4470 		 * without first paging it out.  MADV_FREE pages are often
4471 		 * quickly reused by malloc(3), so we do not do anything that
4472 		 * would result in a page fault on a later access.
4473 		 */
4474 		vm_page_undirty(m);
4475 	else if (advice != MADV_DONTNEED) {
4476 		if (advice == MADV_WILLNEED)
4477 			vm_page_activate(m);
4478 		return;
4479 	}
4480 
4481 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4482 		vm_page_dirty(m);
4483 
4484 	/*
4485 	 * Clear any references to the page.  Otherwise, the page daemon will
4486 	 * immediately reactivate the page.
4487 	 */
4488 	vm_page_aflag_clear(m, PGA_REFERENCED);
4489 
4490 	/*
4491 	 * Place clean pages near the head of the inactive queue rather than
4492 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4493 	 * the page will be reused quickly.  Dirty pages not already in the
4494 	 * laundry are moved there.
4495 	 */
4496 	if (m->dirty == 0)
4497 		vm_page_deactivate_noreuse(m);
4498 	else if (!vm_page_in_laundry(m))
4499 		vm_page_launder(m);
4500 }
4501 
4502 /*
4503  *	vm_page_grab_release
4504  *
4505  *	Helper routine for grab functions to release busy on return.
4506  */
4507 static inline void
4508 vm_page_grab_release(vm_page_t m, int allocflags)
4509 {
4510 
4511 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4512 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4513 			vm_page_sunbusy(m);
4514 		else
4515 			vm_page_xunbusy(m);
4516 	}
4517 }
4518 
4519 /*
4520  *	vm_page_grab_sleep
4521  *
4522  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4523  *	if the caller should retry and false otherwise.
4524  *
4525  *	If the object is locked on entry the object will be unlocked with
4526  *	false returns and still locked but possibly having been dropped
4527  *	with true returns.
4528  */
4529 static bool
4530 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4531     const char *wmesg, int allocflags, bool locked)
4532 {
4533 
4534 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4535 		return (false);
4536 
4537 	/*
4538 	 * Reference the page before unlocking and sleeping so that
4539 	 * the page daemon is less likely to reclaim it.
4540 	 */
4541 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4542 		vm_page_reference(m);
4543 
4544 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4545 	    locked)
4546 		VM_OBJECT_WLOCK(object);
4547 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4548 		return (false);
4549 
4550 	return (true);
4551 }
4552 
4553 /*
4554  * Assert that the grab flags are valid.
4555  */
4556 static inline void
4557 vm_page_grab_check(int allocflags)
4558 {
4559 
4560 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4561 	    (allocflags & VM_ALLOC_WIRED) != 0,
4562 	    ("vm_page_grab*: the pages must be busied or wired"));
4563 
4564 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4565 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4566 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4567 }
4568 
4569 /*
4570  * Calculate the page allocation flags for grab.
4571  */
4572 static inline int
4573 vm_page_grab_pflags(int allocflags)
4574 {
4575 	int pflags;
4576 
4577 	pflags = allocflags &
4578 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4579 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4580 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4581 		pflags |= VM_ALLOC_WAITFAIL;
4582 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4583 		pflags |= VM_ALLOC_SBUSY;
4584 
4585 	return (pflags);
4586 }
4587 
4588 /*
4589  * Grab a page, waiting until we are waken up due to the page
4590  * changing state.  We keep on waiting, if the page continues
4591  * to be in the object.  If the page doesn't exist, first allocate it
4592  * and then conditionally zero it.
4593  *
4594  * This routine may sleep.
4595  *
4596  * The object must be locked on entry.  The lock will, however, be released
4597  * and reacquired if the routine sleeps.
4598  */
4599 vm_page_t
4600 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4601 {
4602 	vm_page_t m;
4603 
4604 	VM_OBJECT_ASSERT_WLOCKED(object);
4605 	vm_page_grab_check(allocflags);
4606 
4607 retrylookup:
4608 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4609 		if (!vm_page_tryacquire(m, allocflags)) {
4610 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4611 			    allocflags, true))
4612 				goto retrylookup;
4613 			return (NULL);
4614 		}
4615 		goto out;
4616 	}
4617 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4618 		return (NULL);
4619 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4620 	if (m == NULL) {
4621 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4622 			return (NULL);
4623 		goto retrylookup;
4624 	}
4625 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4626 		pmap_zero_page(m);
4627 
4628 out:
4629 	vm_page_grab_release(m, allocflags);
4630 
4631 	return (m);
4632 }
4633 
4634 /*
4635  * Locklessly attempt to acquire a page given a (object, pindex) tuple
4636  * and an optional previous page to avoid the radix lookup.  The resulting
4637  * page will be validated against the identity tuple and busied or wired
4638  * as requested.  A NULL *mp return guarantees that the page was not in
4639  * radix at the time of the call but callers must perform higher level
4640  * synchronization or retry the operation under a lock if they require
4641  * an atomic answer.  This is the only lock free validation routine,
4642  * other routines can depend on the resulting page state.
4643  *
4644  * The return value indicates whether the operation failed due to caller
4645  * flags.  The return is tri-state with mp:
4646  *
4647  * (true, *mp != NULL) - The operation was successful.
4648  * (true, *mp == NULL) - The page was not found in tree.
4649  * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4650  */
4651 static bool
4652 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4653     vm_page_t prev, vm_page_t *mp, int allocflags)
4654 {
4655 	vm_page_t m;
4656 
4657 	vm_page_grab_check(allocflags);
4658 	MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4659 
4660 	*mp = NULL;
4661 	for (;;) {
4662 		/*
4663 		 * We may see a false NULL here because the previous page
4664 		 * has been removed or just inserted and the list is loaded
4665 		 * without barriers.  Switch to radix to verify.
4666 		 */
4667 		if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4668 		    QMD_IS_TRASHED(m) || m->pindex != pindex ||
4669 		    atomic_load_ptr(&m->object) != object) {
4670 			prev = NULL;
4671 			/*
4672 			 * This guarantees the result is instantaneously
4673 			 * correct.
4674 			 */
4675 			m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4676 		}
4677 		if (m == NULL)
4678 			return (true);
4679 		if (vm_page_trybusy(m, allocflags)) {
4680 			if (m->object == object && m->pindex == pindex)
4681 				break;
4682 			/* relookup. */
4683 			vm_page_busy_release(m);
4684 			cpu_spinwait();
4685 			continue;
4686 		}
4687 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4688 		    allocflags, false))
4689 			return (false);
4690 	}
4691 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4692 		vm_page_wire(m);
4693 	vm_page_grab_release(m, allocflags);
4694 	*mp = m;
4695 	return (true);
4696 }
4697 
4698 /*
4699  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4700  * is not set.
4701  */
4702 vm_page_t
4703 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4704 {
4705 	vm_page_t m;
4706 
4707 	vm_page_grab_check(allocflags);
4708 
4709 	if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4710 		return (NULL);
4711 	if (m != NULL)
4712 		return (m);
4713 
4714 	/*
4715 	 * The radix lockless lookup should never return a false negative
4716 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4717 	 * was no page present at the instant of the call.  A NOCREAT caller
4718 	 * must handle create races gracefully.
4719 	 */
4720 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4721 		return (NULL);
4722 
4723 	VM_OBJECT_WLOCK(object);
4724 	m = vm_page_grab(object, pindex, allocflags);
4725 	VM_OBJECT_WUNLOCK(object);
4726 
4727 	return (m);
4728 }
4729 
4730 /*
4731  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4732  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4733  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4734  * in simultaneously.  Additional pages will be left on a paging queue but
4735  * will neither be wired nor busy regardless of allocflags.
4736  */
4737 int
4738 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4739 {
4740 	vm_page_t m;
4741 	vm_page_t ma[VM_INITIAL_PAGEIN];
4742 	int after, i, pflags, rv;
4743 
4744 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4745 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4746 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4747 	KASSERT((allocflags &
4748 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4749 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4750 	VM_OBJECT_ASSERT_WLOCKED(object);
4751 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4752 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4753 	pflags |= VM_ALLOC_WAITFAIL;
4754 
4755 retrylookup:
4756 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4757 		/*
4758 		 * If the page is fully valid it can only become invalid
4759 		 * with the object lock held.  If it is not valid it can
4760 		 * become valid with the busy lock held.  Therefore, we
4761 		 * may unnecessarily lock the exclusive busy here if we
4762 		 * race with I/O completion not using the object lock.
4763 		 * However, we will not end up with an invalid page and a
4764 		 * shared lock.
4765 		 */
4766 		if (!vm_page_trybusy(m,
4767 		    vm_page_all_valid(m) ? allocflags : 0)) {
4768 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4769 			    allocflags, true);
4770 			goto retrylookup;
4771 		}
4772 		if (vm_page_all_valid(m))
4773 			goto out;
4774 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4775 			vm_page_busy_release(m);
4776 			*mp = NULL;
4777 			return (VM_PAGER_FAIL);
4778 		}
4779 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4780 		*mp = NULL;
4781 		return (VM_PAGER_FAIL);
4782 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4783 		if (!vm_pager_can_alloc_page(object, pindex)) {
4784 			*mp = NULL;
4785 			return (VM_PAGER_AGAIN);
4786 		}
4787 		goto retrylookup;
4788 	}
4789 
4790 	vm_page_assert_xbusied(m);
4791 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4792 		after = MIN(after, VM_INITIAL_PAGEIN);
4793 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4794 		after = MAX(after, 1);
4795 		ma[0] = m;
4796 		for (i = 1; i < after; i++) {
4797 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4798 				if (vm_page_any_valid(ma[i]) ||
4799 				    !vm_page_tryxbusy(ma[i]))
4800 					break;
4801 			} else {
4802 				ma[i] = vm_page_alloc(object, m->pindex + i,
4803 				    VM_ALLOC_NORMAL);
4804 				if (ma[i] == NULL)
4805 					break;
4806 			}
4807 		}
4808 		after = i;
4809 		vm_object_pip_add(object, after);
4810 		VM_OBJECT_WUNLOCK(object);
4811 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4812 		VM_OBJECT_WLOCK(object);
4813 		vm_object_pip_wakeupn(object, after);
4814 		/* Pager may have replaced a page. */
4815 		m = ma[0];
4816 		if (rv != VM_PAGER_OK) {
4817 			for (i = 0; i < after; i++) {
4818 				if (!vm_page_wired(ma[i]))
4819 					vm_page_free(ma[i]);
4820 				else
4821 					vm_page_xunbusy(ma[i]);
4822 			}
4823 			*mp = NULL;
4824 			return (rv);
4825 		}
4826 		for (i = 1; i < after; i++)
4827 			vm_page_readahead_finish(ma[i]);
4828 		MPASS(vm_page_all_valid(m));
4829 	} else {
4830 		vm_page_zero_invalid(m, TRUE);
4831 	}
4832 out:
4833 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4834 		vm_page_wire(m);
4835 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4836 		vm_page_busy_downgrade(m);
4837 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4838 		vm_page_busy_release(m);
4839 	*mp = m;
4840 	return (VM_PAGER_OK);
4841 }
4842 
4843 /*
4844  * Locklessly grab a valid page.  If the page is not valid or not yet
4845  * allocated this will fall back to the object lock method.
4846  */
4847 int
4848 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4849     vm_pindex_t pindex, int allocflags)
4850 {
4851 	vm_page_t m;
4852 	int flags;
4853 	int error;
4854 
4855 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4856 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4857 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4858 	    "mismatch"));
4859 	KASSERT((allocflags &
4860 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4861 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4862 
4863 	/*
4864 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
4865 	 * before we can inspect the valid field and return a wired page.
4866 	 */
4867 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4868 	if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4869 		return (VM_PAGER_FAIL);
4870 	if ((m = *mp) != NULL) {
4871 		if (vm_page_all_valid(m)) {
4872 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4873 				vm_page_wire(m);
4874 			vm_page_grab_release(m, allocflags);
4875 			return (VM_PAGER_OK);
4876 		}
4877 		vm_page_busy_release(m);
4878 	}
4879 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4880 		*mp = NULL;
4881 		return (VM_PAGER_FAIL);
4882 	}
4883 	VM_OBJECT_WLOCK(object);
4884 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
4885 	VM_OBJECT_WUNLOCK(object);
4886 
4887 	return (error);
4888 }
4889 
4890 /*
4891  * Return the specified range of pages from the given object.  For each
4892  * page offset within the range, if a page already exists within the object
4893  * at that offset and it is busy, then wait for it to change state.  If,
4894  * instead, the page doesn't exist, then allocate it.
4895  *
4896  * The caller must always specify an allocation class.
4897  *
4898  * allocation classes:
4899  *	VM_ALLOC_NORMAL		normal process request
4900  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4901  *
4902  * The caller must always specify that the pages are to be busied and/or
4903  * wired.
4904  *
4905  * optional allocation flags:
4906  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4907  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4908  *	VM_ALLOC_NOWAIT		do not sleep
4909  *	VM_ALLOC_SBUSY		set page to sbusy state
4910  *	VM_ALLOC_WIRED		wire the pages
4911  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4912  *
4913  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4914  * may return a partial prefix of the requested range.
4915  */
4916 int
4917 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4918     vm_page_t *ma, int count)
4919 {
4920 	vm_page_t m, mpred;
4921 	int pflags;
4922 	int i;
4923 
4924 	VM_OBJECT_ASSERT_WLOCKED(object);
4925 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4926 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4927 	KASSERT(count > 0,
4928 	    ("vm_page_grab_pages: invalid page count %d", count));
4929 	vm_page_grab_check(allocflags);
4930 
4931 	pflags = vm_page_grab_pflags(allocflags);
4932 	i = 0;
4933 retrylookup:
4934 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4935 	if (m == NULL || m->pindex != pindex + i) {
4936 		mpred = m;
4937 		m = NULL;
4938 	} else
4939 		mpred = TAILQ_PREV(m, pglist, listq);
4940 	for (; i < count; i++) {
4941 		if (m != NULL) {
4942 			if (!vm_page_tryacquire(m, allocflags)) {
4943 				if (vm_page_grab_sleep(object, m, pindex + i,
4944 				    "grbmaw", allocflags, true))
4945 					goto retrylookup;
4946 				break;
4947 			}
4948 		} else {
4949 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4950 				break;
4951 			m = vm_page_alloc_after(object, pindex + i,
4952 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4953 			if (m == NULL) {
4954 				if ((allocflags & (VM_ALLOC_NOWAIT |
4955 				    VM_ALLOC_WAITFAIL)) != 0)
4956 					break;
4957 				goto retrylookup;
4958 			}
4959 		}
4960 		if (vm_page_none_valid(m) &&
4961 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4962 			if ((m->flags & PG_ZERO) == 0)
4963 				pmap_zero_page(m);
4964 			vm_page_valid(m);
4965 		}
4966 		vm_page_grab_release(m, allocflags);
4967 		ma[i] = mpred = m;
4968 		m = vm_page_next(m);
4969 	}
4970 	return (i);
4971 }
4972 
4973 /*
4974  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
4975  * and will fall back to the locked variant to handle allocation.
4976  */
4977 int
4978 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4979     int allocflags, vm_page_t *ma, int count)
4980 {
4981 	vm_page_t m, pred;
4982 	int flags;
4983 	int i;
4984 
4985 	KASSERT(count > 0,
4986 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4987 	vm_page_grab_check(allocflags);
4988 
4989 	/*
4990 	 * Modify flags for lockless acquire to hold the page until we
4991 	 * set it valid if necessary.
4992 	 */
4993 	flags = allocflags & ~VM_ALLOC_NOBUSY;
4994 	pred = NULL;
4995 	for (i = 0; i < count; i++, pindex++) {
4996 		if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4997 			return (i);
4998 		if (m == NULL)
4999 			break;
5000 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5001 			if ((m->flags & PG_ZERO) == 0)
5002 				pmap_zero_page(m);
5003 			vm_page_valid(m);
5004 		}
5005 		/* m will still be wired or busy according to flags. */
5006 		vm_page_grab_release(m, allocflags);
5007 		pred = ma[i] = m;
5008 	}
5009 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5010 		return (i);
5011 	count -= i;
5012 	VM_OBJECT_WLOCK(object);
5013 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5014 	VM_OBJECT_WUNLOCK(object);
5015 
5016 	return (i);
5017 }
5018 
5019 /*
5020  * Mapping function for valid or dirty bits in a page.
5021  *
5022  * Inputs are required to range within a page.
5023  */
5024 vm_page_bits_t
5025 vm_page_bits(int base, int size)
5026 {
5027 	int first_bit;
5028 	int last_bit;
5029 
5030 	KASSERT(
5031 	    base + size <= PAGE_SIZE,
5032 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5033 	);
5034 
5035 	if (size == 0)		/* handle degenerate case */
5036 		return (0);
5037 
5038 	first_bit = base >> DEV_BSHIFT;
5039 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5040 
5041 	return (((vm_page_bits_t)2 << last_bit) -
5042 	    ((vm_page_bits_t)1 << first_bit));
5043 }
5044 
5045 void
5046 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5047 {
5048 
5049 #if PAGE_SIZE == 32768
5050 	atomic_set_64((uint64_t *)bits, set);
5051 #elif PAGE_SIZE == 16384
5052 	atomic_set_32((uint32_t *)bits, set);
5053 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5054 	atomic_set_16((uint16_t *)bits, set);
5055 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5056 	atomic_set_8((uint8_t *)bits, set);
5057 #else		/* PAGE_SIZE <= 8192 */
5058 	uintptr_t addr;
5059 	int shift;
5060 
5061 	addr = (uintptr_t)bits;
5062 	/*
5063 	 * Use a trick to perform a 32-bit atomic on the
5064 	 * containing aligned word, to not depend on the existence
5065 	 * of atomic_{set, clear}_{8, 16}.
5066 	 */
5067 	shift = addr & (sizeof(uint32_t) - 1);
5068 #if BYTE_ORDER == BIG_ENDIAN
5069 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5070 #else
5071 	shift *= NBBY;
5072 #endif
5073 	addr &= ~(sizeof(uint32_t) - 1);
5074 	atomic_set_32((uint32_t *)addr, set << shift);
5075 #endif		/* PAGE_SIZE */
5076 }
5077 
5078 static inline void
5079 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5080 {
5081 
5082 #if PAGE_SIZE == 32768
5083 	atomic_clear_64((uint64_t *)bits, clear);
5084 #elif PAGE_SIZE == 16384
5085 	atomic_clear_32((uint32_t *)bits, clear);
5086 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5087 	atomic_clear_16((uint16_t *)bits, clear);
5088 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5089 	atomic_clear_8((uint8_t *)bits, clear);
5090 #else		/* PAGE_SIZE <= 8192 */
5091 	uintptr_t addr;
5092 	int shift;
5093 
5094 	addr = (uintptr_t)bits;
5095 	/*
5096 	 * Use a trick to perform a 32-bit atomic on the
5097 	 * containing aligned word, to not depend on the existence
5098 	 * of atomic_{set, clear}_{8, 16}.
5099 	 */
5100 	shift = addr & (sizeof(uint32_t) - 1);
5101 #if BYTE_ORDER == BIG_ENDIAN
5102 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5103 #else
5104 	shift *= NBBY;
5105 #endif
5106 	addr &= ~(sizeof(uint32_t) - 1);
5107 	atomic_clear_32((uint32_t *)addr, clear << shift);
5108 #endif		/* PAGE_SIZE */
5109 }
5110 
5111 static inline vm_page_bits_t
5112 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5113 {
5114 #if PAGE_SIZE == 32768
5115 	uint64_t old;
5116 
5117 	old = *bits;
5118 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5119 	return (old);
5120 #elif PAGE_SIZE == 16384
5121 	uint32_t old;
5122 
5123 	old = *bits;
5124 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5125 	return (old);
5126 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5127 	uint16_t old;
5128 
5129 	old = *bits;
5130 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5131 	return (old);
5132 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5133 	uint8_t old;
5134 
5135 	old = *bits;
5136 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5137 	return (old);
5138 #else		/* PAGE_SIZE <= 4096*/
5139 	uintptr_t addr;
5140 	uint32_t old, new, mask;
5141 	int shift;
5142 
5143 	addr = (uintptr_t)bits;
5144 	/*
5145 	 * Use a trick to perform a 32-bit atomic on the
5146 	 * containing aligned word, to not depend on the existence
5147 	 * of atomic_{set, swap, clear}_{8, 16}.
5148 	 */
5149 	shift = addr & (sizeof(uint32_t) - 1);
5150 #if BYTE_ORDER == BIG_ENDIAN
5151 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5152 #else
5153 	shift *= NBBY;
5154 #endif
5155 	addr &= ~(sizeof(uint32_t) - 1);
5156 	mask = VM_PAGE_BITS_ALL << shift;
5157 
5158 	old = *bits;
5159 	do {
5160 		new = old & ~mask;
5161 		new |= newbits << shift;
5162 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5163 	return (old >> shift);
5164 #endif		/* PAGE_SIZE */
5165 }
5166 
5167 /*
5168  *	vm_page_set_valid_range:
5169  *
5170  *	Sets portions of a page valid.  The arguments are expected
5171  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5172  *	of any partial chunks touched by the range.  The invalid portion of
5173  *	such chunks will be zeroed.
5174  *
5175  *	(base + size) must be less then or equal to PAGE_SIZE.
5176  */
5177 void
5178 vm_page_set_valid_range(vm_page_t m, int base, int size)
5179 {
5180 	int endoff, frag;
5181 	vm_page_bits_t pagebits;
5182 
5183 	vm_page_assert_busied(m);
5184 	if (size == 0)	/* handle degenerate case */
5185 		return;
5186 
5187 	/*
5188 	 * If the base is not DEV_BSIZE aligned and the valid
5189 	 * bit is clear, we have to zero out a portion of the
5190 	 * first block.
5191 	 */
5192 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5193 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5194 		pmap_zero_page_area(m, frag, base - frag);
5195 
5196 	/*
5197 	 * If the ending offset is not DEV_BSIZE aligned and the
5198 	 * valid bit is clear, we have to zero out a portion of
5199 	 * the last block.
5200 	 */
5201 	endoff = base + size;
5202 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5203 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5204 		pmap_zero_page_area(m, endoff,
5205 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5206 
5207 	/*
5208 	 * Assert that no previously invalid block that is now being validated
5209 	 * is already dirty.
5210 	 */
5211 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5212 	    ("vm_page_set_valid_range: page %p is dirty", m));
5213 
5214 	/*
5215 	 * Set valid bits inclusive of any overlap.
5216 	 */
5217 	pagebits = vm_page_bits(base, size);
5218 	if (vm_page_xbusied(m))
5219 		m->valid |= pagebits;
5220 	else
5221 		vm_page_bits_set(m, &m->valid, pagebits);
5222 }
5223 
5224 /*
5225  * Set the page dirty bits and free the invalid swap space if
5226  * present.  Returns the previous dirty bits.
5227  */
5228 vm_page_bits_t
5229 vm_page_set_dirty(vm_page_t m)
5230 {
5231 	vm_page_bits_t old;
5232 
5233 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5234 
5235 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5236 		old = m->dirty;
5237 		m->dirty = VM_PAGE_BITS_ALL;
5238 	} else
5239 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5240 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5241 		vm_pager_page_unswapped(m);
5242 
5243 	return (old);
5244 }
5245 
5246 /*
5247  * Clear the given bits from the specified page's dirty field.
5248  */
5249 static __inline void
5250 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5251 {
5252 
5253 	vm_page_assert_busied(m);
5254 
5255 	/*
5256 	 * If the page is xbusied and not write mapped we are the
5257 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5258 	 * layer can call vm_page_dirty() without holding a distinguished
5259 	 * lock.  The combination of page busy and atomic operations
5260 	 * suffice to guarantee consistency of the page dirty field.
5261 	 */
5262 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5263 		m->dirty &= ~pagebits;
5264 	else
5265 		vm_page_bits_clear(m, &m->dirty, pagebits);
5266 }
5267 
5268 /*
5269  *	vm_page_set_validclean:
5270  *
5271  *	Sets portions of a page valid and clean.  The arguments are expected
5272  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5273  *	of any partial chunks touched by the range.  The invalid portion of
5274  *	such chunks will be zero'd.
5275  *
5276  *	(base + size) must be less then or equal to PAGE_SIZE.
5277  */
5278 void
5279 vm_page_set_validclean(vm_page_t m, int base, int size)
5280 {
5281 	vm_page_bits_t oldvalid, pagebits;
5282 	int endoff, frag;
5283 
5284 	vm_page_assert_busied(m);
5285 	if (size == 0)	/* handle degenerate case */
5286 		return;
5287 
5288 	/*
5289 	 * If the base is not DEV_BSIZE aligned and the valid
5290 	 * bit is clear, we have to zero out a portion of the
5291 	 * first block.
5292 	 */
5293 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5294 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5295 		pmap_zero_page_area(m, frag, base - frag);
5296 
5297 	/*
5298 	 * If the ending offset is not DEV_BSIZE aligned and the
5299 	 * valid bit is clear, we have to zero out a portion of
5300 	 * the last block.
5301 	 */
5302 	endoff = base + size;
5303 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5304 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5305 		pmap_zero_page_area(m, endoff,
5306 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5307 
5308 	/*
5309 	 * Set valid, clear dirty bits.  If validating the entire
5310 	 * page we can safely clear the pmap modify bit.  We also
5311 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5312 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5313 	 * be set again.
5314 	 *
5315 	 * We set valid bits inclusive of any overlap, but we can only
5316 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5317 	 * the range.
5318 	 */
5319 	oldvalid = m->valid;
5320 	pagebits = vm_page_bits(base, size);
5321 	if (vm_page_xbusied(m))
5322 		m->valid |= pagebits;
5323 	else
5324 		vm_page_bits_set(m, &m->valid, pagebits);
5325 #if 0	/* NOT YET */
5326 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5327 		frag = DEV_BSIZE - frag;
5328 		base += frag;
5329 		size -= frag;
5330 		if (size < 0)
5331 			size = 0;
5332 	}
5333 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5334 #endif
5335 	if (base == 0 && size == PAGE_SIZE) {
5336 		/*
5337 		 * The page can only be modified within the pmap if it is
5338 		 * mapped, and it can only be mapped if it was previously
5339 		 * fully valid.
5340 		 */
5341 		if (oldvalid == VM_PAGE_BITS_ALL)
5342 			/*
5343 			 * Perform the pmap_clear_modify() first.  Otherwise,
5344 			 * a concurrent pmap operation, such as
5345 			 * pmap_protect(), could clear a modification in the
5346 			 * pmap and set the dirty field on the page before
5347 			 * pmap_clear_modify() had begun and after the dirty
5348 			 * field was cleared here.
5349 			 */
5350 			pmap_clear_modify(m);
5351 		m->dirty = 0;
5352 		vm_page_aflag_clear(m, PGA_NOSYNC);
5353 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5354 		m->dirty &= ~pagebits;
5355 	else
5356 		vm_page_clear_dirty_mask(m, pagebits);
5357 }
5358 
5359 void
5360 vm_page_clear_dirty(vm_page_t m, int base, int size)
5361 {
5362 
5363 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5364 }
5365 
5366 /*
5367  *	vm_page_set_invalid:
5368  *
5369  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5370  *	valid and dirty bits for the effected areas are cleared.
5371  */
5372 void
5373 vm_page_set_invalid(vm_page_t m, int base, int size)
5374 {
5375 	vm_page_bits_t bits;
5376 	vm_object_t object;
5377 
5378 	/*
5379 	 * The object lock is required so that pages can't be mapped
5380 	 * read-only while we're in the process of invalidating them.
5381 	 */
5382 	object = m->object;
5383 	VM_OBJECT_ASSERT_WLOCKED(object);
5384 	vm_page_assert_busied(m);
5385 
5386 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5387 	    size >= object->un_pager.vnp.vnp_size)
5388 		bits = VM_PAGE_BITS_ALL;
5389 	else
5390 		bits = vm_page_bits(base, size);
5391 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5392 		pmap_remove_all(m);
5393 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5394 	    !pmap_page_is_mapped(m),
5395 	    ("vm_page_set_invalid: page %p is mapped", m));
5396 	if (vm_page_xbusied(m)) {
5397 		m->valid &= ~bits;
5398 		m->dirty &= ~bits;
5399 	} else {
5400 		vm_page_bits_clear(m, &m->valid, bits);
5401 		vm_page_bits_clear(m, &m->dirty, bits);
5402 	}
5403 }
5404 
5405 /*
5406  *	vm_page_invalid:
5407  *
5408  *	Invalidates the entire page.  The page must be busy, unmapped, and
5409  *	the enclosing object must be locked.  The object locks protects
5410  *	against concurrent read-only pmap enter which is done without
5411  *	busy.
5412  */
5413 void
5414 vm_page_invalid(vm_page_t m)
5415 {
5416 
5417 	vm_page_assert_busied(m);
5418 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5419 	MPASS(!pmap_page_is_mapped(m));
5420 
5421 	if (vm_page_xbusied(m))
5422 		m->valid = 0;
5423 	else
5424 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5425 }
5426 
5427 /*
5428  * vm_page_zero_invalid()
5429  *
5430  *	The kernel assumes that the invalid portions of a page contain
5431  *	garbage, but such pages can be mapped into memory by user code.
5432  *	When this occurs, we must zero out the non-valid portions of the
5433  *	page so user code sees what it expects.
5434  *
5435  *	Pages are most often semi-valid when the end of a file is mapped
5436  *	into memory and the file's size is not page aligned.
5437  */
5438 void
5439 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5440 {
5441 	int b;
5442 	int i;
5443 
5444 	/*
5445 	 * Scan the valid bits looking for invalid sections that
5446 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5447 	 * valid bit may be set ) have already been zeroed by
5448 	 * vm_page_set_validclean().
5449 	 */
5450 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5451 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5452 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5453 			if (i > b) {
5454 				pmap_zero_page_area(m,
5455 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5456 			}
5457 			b = i + 1;
5458 		}
5459 	}
5460 
5461 	/*
5462 	 * setvalid is TRUE when we can safely set the zero'd areas
5463 	 * as being valid.  We can do this if there are no cache consistency
5464 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5465 	 */
5466 	if (setvalid)
5467 		vm_page_valid(m);
5468 }
5469 
5470 /*
5471  *	vm_page_is_valid:
5472  *
5473  *	Is (partial) page valid?  Note that the case where size == 0
5474  *	will return FALSE in the degenerate case where the page is
5475  *	entirely invalid, and TRUE otherwise.
5476  *
5477  *	Some callers envoke this routine without the busy lock held and
5478  *	handle races via higher level locks.  Typical callers should
5479  *	hold a busy lock to prevent invalidation.
5480  */
5481 int
5482 vm_page_is_valid(vm_page_t m, int base, int size)
5483 {
5484 	vm_page_bits_t bits;
5485 
5486 	bits = vm_page_bits(base, size);
5487 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5488 }
5489 
5490 /*
5491  * Returns true if all of the specified predicates are true for the entire
5492  * (super)page and false otherwise.
5493  */
5494 bool
5495 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5496 {
5497 	vm_object_t object;
5498 	int i, npages;
5499 
5500 	object = m->object;
5501 	if (skip_m != NULL && skip_m->object != object)
5502 		return (false);
5503 	VM_OBJECT_ASSERT_LOCKED(object);
5504 	npages = atop(pagesizes[m->psind]);
5505 
5506 	/*
5507 	 * The physically contiguous pages that make up a superpage, i.e., a
5508 	 * page with a page size index ("psind") greater than zero, will
5509 	 * occupy adjacent entries in vm_page_array[].
5510 	 */
5511 	for (i = 0; i < npages; i++) {
5512 		/* Always test object consistency, including "skip_m". */
5513 		if (m[i].object != object)
5514 			return (false);
5515 		if (&m[i] == skip_m)
5516 			continue;
5517 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5518 			return (false);
5519 		if ((flags & PS_ALL_DIRTY) != 0) {
5520 			/*
5521 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5522 			 * might stop this case from spuriously returning
5523 			 * "false".  However, that would require a write lock
5524 			 * on the object containing "m[i]".
5525 			 */
5526 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5527 				return (false);
5528 		}
5529 		if ((flags & PS_ALL_VALID) != 0 &&
5530 		    m[i].valid != VM_PAGE_BITS_ALL)
5531 			return (false);
5532 	}
5533 	return (true);
5534 }
5535 
5536 /*
5537  * Set the page's dirty bits if the page is modified.
5538  */
5539 void
5540 vm_page_test_dirty(vm_page_t m)
5541 {
5542 
5543 	vm_page_assert_busied(m);
5544 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5545 		vm_page_dirty(m);
5546 }
5547 
5548 void
5549 vm_page_valid(vm_page_t m)
5550 {
5551 
5552 	vm_page_assert_busied(m);
5553 	if (vm_page_xbusied(m))
5554 		m->valid = VM_PAGE_BITS_ALL;
5555 	else
5556 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5557 }
5558 
5559 void
5560 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5561 {
5562 
5563 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5564 }
5565 
5566 void
5567 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5568 {
5569 
5570 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5571 }
5572 
5573 int
5574 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5575 {
5576 
5577 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5578 }
5579 
5580 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5581 void
5582 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5583 {
5584 
5585 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5586 }
5587 
5588 void
5589 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5590 {
5591 
5592 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5593 }
5594 #endif
5595 
5596 #ifdef INVARIANTS
5597 void
5598 vm_page_object_busy_assert(vm_page_t m)
5599 {
5600 
5601 	/*
5602 	 * Certain of the page's fields may only be modified by the
5603 	 * holder of a page or object busy.
5604 	 */
5605 	if (m->object != NULL && !vm_page_busied(m))
5606 		VM_OBJECT_ASSERT_BUSY(m->object);
5607 }
5608 
5609 void
5610 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5611 {
5612 
5613 	if ((bits & PGA_WRITEABLE) == 0)
5614 		return;
5615 
5616 	/*
5617 	 * The PGA_WRITEABLE flag can only be set if the page is
5618 	 * managed, is exclusively busied or the object is locked.
5619 	 * Currently, this flag is only set by pmap_enter().
5620 	 */
5621 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5622 	    ("PGA_WRITEABLE on unmanaged page"));
5623 	if (!vm_page_xbusied(m))
5624 		VM_OBJECT_ASSERT_BUSY(m->object);
5625 }
5626 #endif
5627 
5628 #include "opt_ddb.h"
5629 #ifdef DDB
5630 #include <sys/kernel.h>
5631 
5632 #include <ddb/ddb.h>
5633 
5634 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5635 {
5636 
5637 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5638 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5639 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5640 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5641 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5642 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5643 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5644 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5645 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5646 }
5647 
5648 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5649 {
5650 	int dom;
5651 
5652 	db_printf("pq_free %d\n", vm_free_count());
5653 	for (dom = 0; dom < vm_ndomains; dom++) {
5654 		db_printf(
5655     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5656 		    dom,
5657 		    vm_dom[dom].vmd_page_count,
5658 		    vm_dom[dom].vmd_free_count,
5659 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5660 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5661 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5662 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5663 	}
5664 }
5665 
5666 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5667 {
5668 	vm_page_t m;
5669 	boolean_t phys, virt;
5670 
5671 	if (!have_addr) {
5672 		db_printf("show pginfo addr\n");
5673 		return;
5674 	}
5675 
5676 	phys = strchr(modif, 'p') != NULL;
5677 	virt = strchr(modif, 'v') != NULL;
5678 	if (virt)
5679 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5680 	else if (phys)
5681 		m = PHYS_TO_VM_PAGE(addr);
5682 	else
5683 		m = (vm_page_t)addr;
5684 	db_printf(
5685     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5686     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5687 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5688 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5689 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5690 }
5691 #endif /* DDB */
5692