xref: /freebsd/sys/vm/vm_page.c (revision 473957941922d17be72089e385e2e2a995fd0e1c)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *	Resident memory management module.
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_vm.h"
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/lock.h>
82 #include <sys/malloc.h>
83 #include <sys/mman.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
86 #include <sys/proc.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
89 #include <sys/sbuf.h>
90 #include <sys/sched.h>
91 #include <sys/smp.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
95 
96 #include <vm/vm.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
111 #include <vm/uma.h>
112 #include <vm/uma_int.h>
113 
114 #include <machine/md_var.h>
115 
116 struct vm_domain vm_dom[MAXMEMDOM];
117 
118 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
119 
120 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
121 
122 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
123 /* The following fields are protected by the domainset lock. */
124 domainset_t __exclusive_cache_line vm_min_domains;
125 domainset_t __exclusive_cache_line vm_severe_domains;
126 static int vm_min_waiters;
127 static int vm_severe_waiters;
128 static int vm_pageproc_waiters;
129 
130 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0,
131     "VM page statistics");
132 
133 static counter_u64_t pqstate_commit_retries = EARLY_COUNTER;
134 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
135     CTLFLAG_RD, &pqstate_commit_retries,
136     "Number of failed per-page atomic queue state updates");
137 
138 static counter_u64_t queue_ops = EARLY_COUNTER;
139 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
140     CTLFLAG_RD, &queue_ops,
141     "Number of batched queue operations");
142 
143 static counter_u64_t queue_nops = EARLY_COUNTER;
144 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
145     CTLFLAG_RD, &queue_nops,
146     "Number of batched queue operations with no effects");
147 
148 static void
149 counter_startup(void)
150 {
151 
152 	pqstate_commit_retries = counter_u64_alloc(M_WAITOK);
153 	queue_ops = counter_u64_alloc(M_WAITOK);
154 	queue_nops = counter_u64_alloc(M_WAITOK);
155 }
156 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
157 
158 /*
159  * bogus page -- for I/O to/from partially complete buffers,
160  * or for paging into sparsely invalid regions.
161  */
162 vm_page_t bogus_page;
163 
164 vm_page_t vm_page_array;
165 long vm_page_array_size;
166 long first_page;
167 
168 static TAILQ_HEAD(, vm_page) blacklist_head;
169 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
170 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
171     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
172 
173 static uma_zone_t fakepg_zone;
174 
175 static void vm_page_alloc_check(vm_page_t m);
176 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
177     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
178 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
179 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
180 static bool vm_page_free_prep(vm_page_t m);
181 static void vm_page_free_toq(vm_page_t m);
182 static void vm_page_init(void *dummy);
183 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
184     vm_pindex_t pindex, vm_page_t mpred);
185 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
186     vm_page_t mpred);
187 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
188     const uint16_t nflag);
189 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
190     vm_page_t m_run, vm_paddr_t high);
191 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
192 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
193     int req);
194 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
195     int flags);
196 static void vm_page_zone_release(void *arg, void **store, int cnt);
197 
198 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
199 
200 static void
201 vm_page_init(void *dummy)
202 {
203 
204 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
205 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
206 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
207 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
208 }
209 
210 /*
211  * The cache page zone is initialized later since we need to be able to allocate
212  * pages before UMA is fully initialized.
213  */
214 static void
215 vm_page_init_cache_zones(void *dummy __unused)
216 {
217 	struct vm_domain *vmd;
218 	struct vm_pgcache *pgcache;
219 	int cache, domain, maxcache, pool;
220 
221 	maxcache = 0;
222 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
223 	maxcache *= mp_ncpus;
224 	for (domain = 0; domain < vm_ndomains; domain++) {
225 		vmd = VM_DOMAIN(domain);
226 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
227 			pgcache = &vmd->vmd_pgcache[pool];
228 			pgcache->domain = domain;
229 			pgcache->pool = pool;
230 			pgcache->zone = uma_zcache_create("vm pgcache",
231 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
232 			    vm_page_zone_import, vm_page_zone_release, pgcache,
233 			    UMA_ZONE_VM);
234 
235 			/*
236 			 * Limit each pool's zone to 0.1% of the pages in the
237 			 * domain.
238 			 */
239 			cache = maxcache != 0 ? maxcache :
240 			    vmd->vmd_page_count / 1000;
241 			uma_zone_set_maxcache(pgcache->zone, cache);
242 		}
243 	}
244 }
245 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
246 
247 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
248 #if PAGE_SIZE == 32768
249 #ifdef CTASSERT
250 CTASSERT(sizeof(u_long) >= 8);
251 #endif
252 #endif
253 
254 /*
255  *	vm_set_page_size:
256  *
257  *	Sets the page size, perhaps based upon the memory
258  *	size.  Must be called before any use of page-size
259  *	dependent functions.
260  */
261 void
262 vm_set_page_size(void)
263 {
264 	if (vm_cnt.v_page_size == 0)
265 		vm_cnt.v_page_size = PAGE_SIZE;
266 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
267 		panic("vm_set_page_size: page size not a power of two");
268 }
269 
270 /*
271  *	vm_page_blacklist_next:
272  *
273  *	Find the next entry in the provided string of blacklist
274  *	addresses.  Entries are separated by space, comma, or newline.
275  *	If an invalid integer is encountered then the rest of the
276  *	string is skipped.  Updates the list pointer to the next
277  *	character, or NULL if the string is exhausted or invalid.
278  */
279 static vm_paddr_t
280 vm_page_blacklist_next(char **list, char *end)
281 {
282 	vm_paddr_t bad;
283 	char *cp, *pos;
284 
285 	if (list == NULL || *list == NULL)
286 		return (0);
287 	if (**list =='\0') {
288 		*list = NULL;
289 		return (0);
290 	}
291 
292 	/*
293 	 * If there's no end pointer then the buffer is coming from
294 	 * the kenv and we know it's null-terminated.
295 	 */
296 	if (end == NULL)
297 		end = *list + strlen(*list);
298 
299 	/* Ensure that strtoq() won't walk off the end */
300 	if (*end != '\0') {
301 		if (*end == '\n' || *end == ' ' || *end  == ',')
302 			*end = '\0';
303 		else {
304 			printf("Blacklist not terminated, skipping\n");
305 			*list = NULL;
306 			return (0);
307 		}
308 	}
309 
310 	for (pos = *list; *pos != '\0'; pos = cp) {
311 		bad = strtoq(pos, &cp, 0);
312 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
313 			if (bad == 0) {
314 				if (++cp < end)
315 					continue;
316 				else
317 					break;
318 			}
319 		} else
320 			break;
321 		if (*cp == '\0' || ++cp >= end)
322 			*list = NULL;
323 		else
324 			*list = cp;
325 		return (trunc_page(bad));
326 	}
327 	printf("Garbage in RAM blacklist, skipping\n");
328 	*list = NULL;
329 	return (0);
330 }
331 
332 bool
333 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
334 {
335 	struct vm_domain *vmd;
336 	vm_page_t m;
337 	int ret;
338 
339 	m = vm_phys_paddr_to_vm_page(pa);
340 	if (m == NULL)
341 		return (true); /* page does not exist, no failure */
342 
343 	vmd = vm_pagequeue_domain(m);
344 	vm_domain_free_lock(vmd);
345 	ret = vm_phys_unfree_page(m);
346 	vm_domain_free_unlock(vmd);
347 	if (ret != 0) {
348 		vm_domain_freecnt_inc(vmd, -1);
349 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
350 		if (verbose)
351 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
352 	}
353 	return (ret);
354 }
355 
356 /*
357  *	vm_page_blacklist_check:
358  *
359  *	Iterate through the provided string of blacklist addresses, pulling
360  *	each entry out of the physical allocator free list and putting it
361  *	onto a list for reporting via the vm.page_blacklist sysctl.
362  */
363 static void
364 vm_page_blacklist_check(char *list, char *end)
365 {
366 	vm_paddr_t pa;
367 	char *next;
368 
369 	next = list;
370 	while (next != NULL) {
371 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
372 			continue;
373 		vm_page_blacklist_add(pa, bootverbose);
374 	}
375 }
376 
377 /*
378  *	vm_page_blacklist_load:
379  *
380  *	Search for a special module named "ram_blacklist".  It'll be a
381  *	plain text file provided by the user via the loader directive
382  *	of the same name.
383  */
384 static void
385 vm_page_blacklist_load(char **list, char **end)
386 {
387 	void *mod;
388 	u_char *ptr;
389 	u_int len;
390 
391 	mod = NULL;
392 	ptr = NULL;
393 
394 	mod = preload_search_by_type("ram_blacklist");
395 	if (mod != NULL) {
396 		ptr = preload_fetch_addr(mod);
397 		len = preload_fetch_size(mod);
398         }
399 	*list = ptr;
400 	if (ptr != NULL)
401 		*end = ptr + len;
402 	else
403 		*end = NULL;
404 	return;
405 }
406 
407 static int
408 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
409 {
410 	vm_page_t m;
411 	struct sbuf sbuf;
412 	int error, first;
413 
414 	first = 1;
415 	error = sysctl_wire_old_buffer(req, 0);
416 	if (error != 0)
417 		return (error);
418 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
419 	TAILQ_FOREACH(m, &blacklist_head, listq) {
420 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
421 		    (uintmax_t)m->phys_addr);
422 		first = 0;
423 	}
424 	error = sbuf_finish(&sbuf);
425 	sbuf_delete(&sbuf);
426 	return (error);
427 }
428 
429 /*
430  * Initialize a dummy page for use in scans of the specified paging queue.
431  * In principle, this function only needs to set the flag PG_MARKER.
432  * Nonetheless, it write busies the page as a safety precaution.
433  */
434 static void
435 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
436 {
437 
438 	bzero(marker, sizeof(*marker));
439 	marker->flags = PG_MARKER;
440 	marker->a.flags = aflags;
441 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
442 	marker->a.queue = queue;
443 }
444 
445 static void
446 vm_page_domain_init(int domain)
447 {
448 	struct vm_domain *vmd;
449 	struct vm_pagequeue *pq;
450 	int i;
451 
452 	vmd = VM_DOMAIN(domain);
453 	bzero(vmd, sizeof(*vmd));
454 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
455 	    "vm inactive pagequeue";
456 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
457 	    "vm active pagequeue";
458 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
459 	    "vm laundry pagequeue";
460 	*__DECONST(const char **,
461 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
462 	    "vm unswappable pagequeue";
463 	vmd->vmd_domain = domain;
464 	vmd->vmd_page_count = 0;
465 	vmd->vmd_free_count = 0;
466 	vmd->vmd_segs = 0;
467 	vmd->vmd_oom = FALSE;
468 	for (i = 0; i < PQ_COUNT; i++) {
469 		pq = &vmd->vmd_pagequeues[i];
470 		TAILQ_INIT(&pq->pq_pl);
471 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
472 		    MTX_DEF | MTX_DUPOK);
473 		pq->pq_pdpages = 0;
474 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
475 	}
476 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
477 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
478 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
479 
480 	/*
481 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
482 	 * insertions.
483 	 */
484 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
485 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
486 	    &vmd->vmd_inacthead, plinks.q);
487 
488 	/*
489 	 * The clock pages are used to implement active queue scanning without
490 	 * requeues.  Scans start at clock[0], which is advanced after the scan
491 	 * ends.  When the two clock hands meet, they are reset and scanning
492 	 * resumes from the head of the queue.
493 	 */
494 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
495 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
496 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
497 	    &vmd->vmd_clock[0], plinks.q);
498 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
499 	    &vmd->vmd_clock[1], plinks.q);
500 }
501 
502 /*
503  * Initialize a physical page in preparation for adding it to the free
504  * lists.
505  */
506 static void
507 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
508 {
509 
510 	m->object = NULL;
511 	m->ref_count = 0;
512 	m->busy_lock = VPB_FREED;
513 	m->flags = m->a.flags = 0;
514 	m->phys_addr = pa;
515 	m->a.queue = PQ_NONE;
516 	m->psind = 0;
517 	m->segind = segind;
518 	m->order = VM_NFREEORDER;
519 	m->pool = VM_FREEPOOL_DEFAULT;
520 	m->valid = m->dirty = 0;
521 	pmap_page_init(m);
522 }
523 
524 #ifndef PMAP_HAS_PAGE_ARRAY
525 static vm_paddr_t
526 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
527 {
528 	vm_paddr_t new_end;
529 
530 	/*
531 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
532 	 * However, because this page is allocated from KVM, out-of-bounds
533 	 * accesses using the direct map will not be trapped.
534 	 */
535 	*vaddr += PAGE_SIZE;
536 
537 	/*
538 	 * Allocate physical memory for the page structures, and map it.
539 	 */
540 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
541 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
542 	    VM_PROT_READ | VM_PROT_WRITE);
543 	vm_page_array_size = page_range;
544 
545 	return (new_end);
546 }
547 #endif
548 
549 /*
550  *	vm_page_startup:
551  *
552  *	Initializes the resident memory module.  Allocates physical memory for
553  *	bootstrapping UMA and some data structures that are used to manage
554  *	physical pages.  Initializes these structures, and populates the free
555  *	page queues.
556  */
557 vm_offset_t
558 vm_page_startup(vm_offset_t vaddr)
559 {
560 	struct vm_phys_seg *seg;
561 	vm_page_t m;
562 	char *list, *listend;
563 	vm_paddr_t end, high_avail, low_avail, new_end, size;
564 	vm_paddr_t page_range __unused;
565 	vm_paddr_t last_pa, pa;
566 	u_long pagecount;
567 	int biggestone, i, segind;
568 #ifdef WITNESS
569 	vm_offset_t mapped;
570 	int witness_size;
571 #endif
572 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
573 	long ii;
574 #endif
575 
576 	vaddr = round_page(vaddr);
577 
578 	vm_phys_early_startup();
579 	biggestone = vm_phys_avail_largest();
580 	end = phys_avail[biggestone+1];
581 
582 	/*
583 	 * Initialize the page and queue locks.
584 	 */
585 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
586 	for (i = 0; i < PA_LOCK_COUNT; i++)
587 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
588 	for (i = 0; i < vm_ndomains; i++)
589 		vm_page_domain_init(i);
590 
591 	new_end = end;
592 #ifdef WITNESS
593 	witness_size = round_page(witness_startup_count());
594 	new_end -= witness_size;
595 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
596 	    VM_PROT_READ | VM_PROT_WRITE);
597 	bzero((void *)mapped, witness_size);
598 	witness_startup((void *)mapped);
599 #endif
600 
601 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
602     defined(__i386__) || defined(__mips__) || defined(__riscv) || \
603     defined(__powerpc64__)
604 	/*
605 	 * Allocate a bitmap to indicate that a random physical page
606 	 * needs to be included in a minidump.
607 	 *
608 	 * The amd64 port needs this to indicate which direct map pages
609 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
610 	 *
611 	 * However, i386 still needs this workspace internally within the
612 	 * minidump code.  In theory, they are not needed on i386, but are
613 	 * included should the sf_buf code decide to use them.
614 	 */
615 	last_pa = 0;
616 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
617 		if (dump_avail[i + 1] > last_pa)
618 			last_pa = dump_avail[i + 1];
619 	page_range = last_pa / PAGE_SIZE;
620 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
621 	new_end -= vm_page_dump_size;
622 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
623 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
624 	bzero((void *)vm_page_dump, vm_page_dump_size);
625 #else
626 	(void)last_pa;
627 #endif
628 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
629     defined(__riscv) || defined(__powerpc64__)
630 	/*
631 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
632 	 * in a crash dump.  When pmap_map() uses the direct map, they are
633 	 * not automatically included.
634 	 */
635 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
636 		dump_add_page(pa);
637 #endif
638 	phys_avail[biggestone + 1] = new_end;
639 #ifdef __amd64__
640 	/*
641 	 * Request that the physical pages underlying the message buffer be
642 	 * included in a crash dump.  Since the message buffer is accessed
643 	 * through the direct map, they are not automatically included.
644 	 */
645 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
646 	last_pa = pa + round_page(msgbufsize);
647 	while (pa < last_pa) {
648 		dump_add_page(pa);
649 		pa += PAGE_SIZE;
650 	}
651 #endif
652 	/*
653 	 * Compute the number of pages of memory that will be available for
654 	 * use, taking into account the overhead of a page structure per page.
655 	 * In other words, solve
656 	 *	"available physical memory" - round_page(page_range *
657 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
658 	 * for page_range.
659 	 */
660 	low_avail = phys_avail[0];
661 	high_avail = phys_avail[1];
662 	for (i = 0; i < vm_phys_nsegs; i++) {
663 		if (vm_phys_segs[i].start < low_avail)
664 			low_avail = vm_phys_segs[i].start;
665 		if (vm_phys_segs[i].end > high_avail)
666 			high_avail = vm_phys_segs[i].end;
667 	}
668 	/* Skip the first chunk.  It is already accounted for. */
669 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
670 		if (phys_avail[i] < low_avail)
671 			low_avail = phys_avail[i];
672 		if (phys_avail[i + 1] > high_avail)
673 			high_avail = phys_avail[i + 1];
674 	}
675 	first_page = low_avail / PAGE_SIZE;
676 #ifdef VM_PHYSSEG_SPARSE
677 	size = 0;
678 	for (i = 0; i < vm_phys_nsegs; i++)
679 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
680 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
681 		size += phys_avail[i + 1] - phys_avail[i];
682 #elif defined(VM_PHYSSEG_DENSE)
683 	size = high_avail - low_avail;
684 #else
685 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
686 #endif
687 
688 #ifdef PMAP_HAS_PAGE_ARRAY
689 	pmap_page_array_startup(size / PAGE_SIZE);
690 	biggestone = vm_phys_avail_largest();
691 	end = new_end = phys_avail[biggestone + 1];
692 #else
693 #ifdef VM_PHYSSEG_DENSE
694 	/*
695 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
696 	 * the overhead of a page structure per page only if vm_page_array is
697 	 * allocated from the last physical memory chunk.  Otherwise, we must
698 	 * allocate page structures representing the physical memory
699 	 * underlying vm_page_array, even though they will not be used.
700 	 */
701 	if (new_end != high_avail)
702 		page_range = size / PAGE_SIZE;
703 	else
704 #endif
705 	{
706 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
707 
708 		/*
709 		 * If the partial bytes remaining are large enough for
710 		 * a page (PAGE_SIZE) without a corresponding
711 		 * 'struct vm_page', then new_end will contain an
712 		 * extra page after subtracting the length of the VM
713 		 * page array.  Compensate by subtracting an extra
714 		 * page from new_end.
715 		 */
716 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
717 			if (new_end == high_avail)
718 				high_avail -= PAGE_SIZE;
719 			new_end -= PAGE_SIZE;
720 		}
721 	}
722 	end = new_end;
723 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
724 #endif
725 
726 #if VM_NRESERVLEVEL > 0
727 	/*
728 	 * Allocate physical memory for the reservation management system's
729 	 * data structures, and map it.
730 	 */
731 	new_end = vm_reserv_startup(&vaddr, new_end);
732 #endif
733 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
734     defined(__riscv) || defined(__powerpc64__)
735 	/*
736 	 * Include vm_page_array and vm_reserv_array in a crash dump.
737 	 */
738 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
739 		dump_add_page(pa);
740 #endif
741 	phys_avail[biggestone + 1] = new_end;
742 
743 	/*
744 	 * Add physical memory segments corresponding to the available
745 	 * physical pages.
746 	 */
747 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
748 		if (vm_phys_avail_size(i) != 0)
749 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
750 
751 	/*
752 	 * Initialize the physical memory allocator.
753 	 */
754 	vm_phys_init();
755 
756 	/*
757 	 * Initialize the page structures and add every available page to the
758 	 * physical memory allocator's free lists.
759 	 */
760 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
761 	for (ii = 0; ii < vm_page_array_size; ii++) {
762 		m = &vm_page_array[ii];
763 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
764 		m->flags = PG_FICTITIOUS;
765 	}
766 #endif
767 	vm_cnt.v_page_count = 0;
768 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
769 		seg = &vm_phys_segs[segind];
770 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
771 		    m++, pa += PAGE_SIZE)
772 			vm_page_init_page(m, pa, segind);
773 
774 		/*
775 		 * Add the segment to the free lists only if it is covered by
776 		 * one of the ranges in phys_avail.  Because we've added the
777 		 * ranges to the vm_phys_segs array, we can assume that each
778 		 * segment is either entirely contained in one of the ranges,
779 		 * or doesn't overlap any of them.
780 		 */
781 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
782 			struct vm_domain *vmd;
783 
784 			if (seg->start < phys_avail[i] ||
785 			    seg->end > phys_avail[i + 1])
786 				continue;
787 
788 			m = seg->first_page;
789 			pagecount = (u_long)atop(seg->end - seg->start);
790 
791 			vmd = VM_DOMAIN(seg->domain);
792 			vm_domain_free_lock(vmd);
793 			vm_phys_enqueue_contig(m, pagecount);
794 			vm_domain_free_unlock(vmd);
795 			vm_domain_freecnt_inc(vmd, pagecount);
796 			vm_cnt.v_page_count += (u_int)pagecount;
797 
798 			vmd = VM_DOMAIN(seg->domain);
799 			vmd->vmd_page_count += (u_int)pagecount;
800 			vmd->vmd_segs |= 1UL << m->segind;
801 			break;
802 		}
803 	}
804 
805 	/*
806 	 * Remove blacklisted pages from the physical memory allocator.
807 	 */
808 	TAILQ_INIT(&blacklist_head);
809 	vm_page_blacklist_load(&list, &listend);
810 	vm_page_blacklist_check(list, listend);
811 
812 	list = kern_getenv("vm.blacklist");
813 	vm_page_blacklist_check(list, NULL);
814 
815 	freeenv(list);
816 #if VM_NRESERVLEVEL > 0
817 	/*
818 	 * Initialize the reservation management system.
819 	 */
820 	vm_reserv_init();
821 #endif
822 
823 	return (vaddr);
824 }
825 
826 void
827 vm_page_reference(vm_page_t m)
828 {
829 
830 	vm_page_aflag_set(m, PGA_REFERENCED);
831 }
832 
833 static bool
834 vm_page_acquire_flags(vm_page_t m, int allocflags)
835 {
836 	bool locked;
837 
838 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
839 		locked = vm_page_trysbusy(m);
840 	else
841 		locked = vm_page_tryxbusy(m);
842 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
843 		vm_page_wire(m);
844 	return (locked);
845 }
846 
847 /*
848  *	vm_page_busy_sleep_flags
849  *
850  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
851  *	if the caller should retry and false otherwise.
852  */
853 static bool
854 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg,
855     int allocflags)
856 {
857 
858 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
859 		return (false);
860 
861 	/*
862 	 * Reference the page before unlocking and sleeping so that
863 	 * the page daemon is less likely to reclaim it.
864 	 */
865 	if ((allocflags & VM_ALLOC_NOCREAT) == 0)
866 		vm_page_reference(m);
867 
868 	if (_vm_page_busy_sleep(object, m, m->pindex, wmesg, allocflags, true))
869 		VM_OBJECT_WLOCK(object);
870 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
871 		return (false);
872 
873 	return (true);
874 }
875 
876 /*
877  *	vm_page_busy_acquire:
878  *
879  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
880  *	and drop the object lock if necessary.
881  */
882 bool
883 vm_page_busy_acquire(vm_page_t m, int allocflags)
884 {
885 	vm_object_t obj;
886 	bool locked;
887 
888 	/*
889 	 * The page-specific object must be cached because page
890 	 * identity can change during the sleep, causing the
891 	 * re-lock of a different object.
892 	 * It is assumed that a reference to the object is already
893 	 * held by the callers.
894 	 */
895 	obj = m->object;
896 	for (;;) {
897 		if (vm_page_acquire_flags(m, allocflags))
898 			return (true);
899 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
900 			return (false);
901 		if (obj != NULL)
902 			locked = VM_OBJECT_WOWNED(obj);
903 		else
904 			locked = false;
905 		MPASS(locked || vm_page_wired(m));
906 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
907 		    locked) && locked)
908 			VM_OBJECT_WLOCK(obj);
909 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
910 			return (false);
911 		KASSERT(m->object == obj || m->object == NULL,
912 		    ("vm_page_busy_acquire: page %p does not belong to %p",
913 		    m, obj));
914 	}
915 }
916 
917 /*
918  *	vm_page_busy_downgrade:
919  *
920  *	Downgrade an exclusive busy page into a single shared busy page.
921  */
922 void
923 vm_page_busy_downgrade(vm_page_t m)
924 {
925 	u_int x;
926 
927 	vm_page_assert_xbusied(m);
928 
929 	x = m->busy_lock;
930 	for (;;) {
931 		if (atomic_fcmpset_rel_int(&m->busy_lock,
932 		    &x, VPB_SHARERS_WORD(1)))
933 			break;
934 	}
935 	if ((x & VPB_BIT_WAITERS) != 0)
936 		wakeup(m);
937 }
938 
939 /*
940  *
941  *	vm_page_busy_tryupgrade:
942  *
943  *	Attempt to upgrade a single shared busy into an exclusive busy.
944  */
945 int
946 vm_page_busy_tryupgrade(vm_page_t m)
947 {
948 	u_int ce, x;
949 
950 	vm_page_assert_sbusied(m);
951 
952 	x = m->busy_lock;
953 	ce = VPB_CURTHREAD_EXCLUSIVE;
954 	for (;;) {
955 		if (VPB_SHARERS(x) > 1)
956 			return (0);
957 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
958 		    ("vm_page_busy_tryupgrade: invalid lock state"));
959 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
960 		    ce | (x & VPB_BIT_WAITERS)))
961 			continue;
962 		return (1);
963 	}
964 }
965 
966 /*
967  *	vm_page_sbusied:
968  *
969  *	Return a positive value if the page is shared busied, 0 otherwise.
970  */
971 int
972 vm_page_sbusied(vm_page_t m)
973 {
974 	u_int x;
975 
976 	x = m->busy_lock;
977 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
978 }
979 
980 /*
981  *	vm_page_sunbusy:
982  *
983  *	Shared unbusy a page.
984  */
985 void
986 vm_page_sunbusy(vm_page_t m)
987 {
988 	u_int x;
989 
990 	vm_page_assert_sbusied(m);
991 
992 	x = m->busy_lock;
993 	for (;;) {
994 		KASSERT(x != VPB_FREED,
995 		    ("vm_page_sunbusy: Unlocking freed page."));
996 		if (VPB_SHARERS(x) > 1) {
997 			if (atomic_fcmpset_int(&m->busy_lock, &x,
998 			    x - VPB_ONE_SHARER))
999 				break;
1000 			continue;
1001 		}
1002 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1003 		    ("vm_page_sunbusy: invalid lock state"));
1004 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1005 			continue;
1006 		if ((x & VPB_BIT_WAITERS) == 0)
1007 			break;
1008 		wakeup(m);
1009 		break;
1010 	}
1011 }
1012 
1013 /*
1014  *	vm_page_busy_sleep:
1015  *
1016  *	Sleep if the page is busy, using the page pointer as wchan.
1017  *	This is used to implement the hard-path of busying mechanism.
1018  *
1019  *	If nonshared is true, sleep only if the page is xbusy.
1020  *
1021  *	The object lock must be held on entry and will be released on exit.
1022  */
1023 void
1024 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1025 {
1026 	vm_object_t obj;
1027 
1028 	obj = m->object;
1029 	VM_OBJECT_ASSERT_LOCKED(obj);
1030 	vm_page_lock_assert(m, MA_NOTOWNED);
1031 
1032 	if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1033 	    nonshared ? VM_ALLOC_SBUSY : 0 , true))
1034 		VM_OBJECT_DROP(obj);
1035 }
1036 
1037 /*
1038  *	vm_page_busy_sleep_unlocked:
1039  *
1040  *	Sleep if the page is busy, using the page pointer as wchan.
1041  *	This is used to implement the hard-path of busying mechanism.
1042  *
1043  *	If nonshared is true, sleep only if the page is xbusy.
1044  *
1045  *	The object lock must not be held on entry.  The operation will
1046  *	return if the page changes identity.
1047  */
1048 void
1049 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1050     const char *wmesg, bool nonshared)
1051 {
1052 
1053 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1054 	vm_page_lock_assert(m, MA_NOTOWNED);
1055 
1056 	_vm_page_busy_sleep(obj, m, pindex, wmesg,
1057 	    nonshared ? VM_ALLOC_SBUSY : 0, false);
1058 }
1059 
1060 /*
1061  *	_vm_page_busy_sleep:
1062  *
1063  *	Internal busy sleep function.  Verifies the page identity and
1064  *	lockstate against parameters.  Returns true if it sleeps and
1065  *	false otherwise.
1066  *
1067  *	If locked is true the lock will be dropped for any true returns
1068  *	and held for any false returns.
1069  */
1070 static bool
1071 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1072     const char *wmesg, int allocflags, bool locked)
1073 {
1074 	bool xsleep;
1075 	u_int x;
1076 
1077 	/*
1078 	 * If the object is busy we must wait for that to drain to zero
1079 	 * before trying the page again.
1080 	 */
1081 	if (obj != NULL && vm_object_busied(obj)) {
1082 		if (locked)
1083 			VM_OBJECT_DROP(obj);
1084 		vm_object_busy_wait(obj, wmesg);
1085 		return (true);
1086 	}
1087 
1088 	if (!vm_page_busied(m))
1089 		return (false);
1090 
1091 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1092 	sleepq_lock(m);
1093 	x = atomic_load_int(&m->busy_lock);
1094 	do {
1095 		/*
1096 		 * If the page changes objects or becomes unlocked we can
1097 		 * simply return.
1098 		 */
1099 		if (x == VPB_UNBUSIED ||
1100 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1101 		    m->object != obj || m->pindex != pindex) {
1102 			sleepq_release(m);
1103 			return (false);
1104 		}
1105 		if ((x & VPB_BIT_WAITERS) != 0)
1106 			break;
1107 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1108 	if (locked)
1109 		VM_OBJECT_DROP(obj);
1110 	DROP_GIANT();
1111 	sleepq_add(m, NULL, wmesg, 0, 0);
1112 	sleepq_wait(m, PVM);
1113 	PICKUP_GIANT();
1114 	return (true);
1115 }
1116 
1117 /*
1118  *	vm_page_trysbusy:
1119  *
1120  *	Try to shared busy a page.
1121  *	If the operation succeeds 1 is returned otherwise 0.
1122  *	The operation never sleeps.
1123  */
1124 int
1125 vm_page_trysbusy(vm_page_t m)
1126 {
1127 	vm_object_t obj;
1128 	u_int x;
1129 
1130 	obj = m->object;
1131 	x = m->busy_lock;
1132 	for (;;) {
1133 		if ((x & VPB_BIT_SHARED) == 0)
1134 			return (0);
1135 		/*
1136 		 * Reduce the window for transient busies that will trigger
1137 		 * false negatives in vm_page_ps_test().
1138 		 */
1139 		if (obj != NULL && vm_object_busied(obj))
1140 			return (0);
1141 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1142 		    x + VPB_ONE_SHARER))
1143 			break;
1144 	}
1145 
1146 	/* Refetch the object now that we're guaranteed that it is stable. */
1147 	obj = m->object;
1148 	if (obj != NULL && vm_object_busied(obj)) {
1149 		vm_page_sunbusy(m);
1150 		return (0);
1151 	}
1152 	return (1);
1153 }
1154 
1155 /*
1156  *	vm_page_tryxbusy:
1157  *
1158  *	Try to exclusive busy a page.
1159  *	If the operation succeeds 1 is returned otherwise 0.
1160  *	The operation never sleeps.
1161  */
1162 int
1163 vm_page_tryxbusy(vm_page_t m)
1164 {
1165 	vm_object_t obj;
1166 
1167         if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1168             VPB_CURTHREAD_EXCLUSIVE) == 0)
1169 		return (0);
1170 
1171 	obj = m->object;
1172 	if (obj != NULL && vm_object_busied(obj)) {
1173 		vm_page_xunbusy(m);
1174 		return (0);
1175 	}
1176 	return (1);
1177 }
1178 
1179 static void
1180 vm_page_xunbusy_hard_tail(vm_page_t m)
1181 {
1182 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1183 	/* Wake the waiter. */
1184 	wakeup(m);
1185 }
1186 
1187 /*
1188  *	vm_page_xunbusy_hard:
1189  *
1190  *	Called when unbusy has failed because there is a waiter.
1191  */
1192 void
1193 vm_page_xunbusy_hard(vm_page_t m)
1194 {
1195 	vm_page_assert_xbusied(m);
1196 	vm_page_xunbusy_hard_tail(m);
1197 }
1198 
1199 void
1200 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1201 {
1202 	vm_page_assert_xbusied_unchecked(m);
1203 	vm_page_xunbusy_hard_tail(m);
1204 }
1205 
1206 static void
1207 vm_page_busy_free(vm_page_t m)
1208 {
1209 	u_int x;
1210 
1211 	atomic_thread_fence_rel();
1212 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1213 	if ((x & VPB_BIT_WAITERS) != 0)
1214 		wakeup(m);
1215 }
1216 
1217 /*
1218  *	vm_page_unhold_pages:
1219  *
1220  *	Unhold each of the pages that is referenced by the given array.
1221  */
1222 void
1223 vm_page_unhold_pages(vm_page_t *ma, int count)
1224 {
1225 
1226 	for (; count != 0; count--) {
1227 		vm_page_unwire(*ma, PQ_ACTIVE);
1228 		ma++;
1229 	}
1230 }
1231 
1232 vm_page_t
1233 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1234 {
1235 	vm_page_t m;
1236 
1237 #ifdef VM_PHYSSEG_SPARSE
1238 	m = vm_phys_paddr_to_vm_page(pa);
1239 	if (m == NULL)
1240 		m = vm_phys_fictitious_to_vm_page(pa);
1241 	return (m);
1242 #elif defined(VM_PHYSSEG_DENSE)
1243 	long pi;
1244 
1245 	pi = atop(pa);
1246 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1247 		m = &vm_page_array[pi - first_page];
1248 		return (m);
1249 	}
1250 	return (vm_phys_fictitious_to_vm_page(pa));
1251 #else
1252 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1253 #endif
1254 }
1255 
1256 /*
1257  *	vm_page_getfake:
1258  *
1259  *	Create a fictitious page with the specified physical address and
1260  *	memory attribute.  The memory attribute is the only the machine-
1261  *	dependent aspect of a fictitious page that must be initialized.
1262  */
1263 vm_page_t
1264 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1265 {
1266 	vm_page_t m;
1267 
1268 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1269 	vm_page_initfake(m, paddr, memattr);
1270 	return (m);
1271 }
1272 
1273 void
1274 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1275 {
1276 
1277 	if ((m->flags & PG_FICTITIOUS) != 0) {
1278 		/*
1279 		 * The page's memattr might have changed since the
1280 		 * previous initialization.  Update the pmap to the
1281 		 * new memattr.
1282 		 */
1283 		goto memattr;
1284 	}
1285 	m->phys_addr = paddr;
1286 	m->a.queue = PQ_NONE;
1287 	/* Fictitious pages don't use "segind". */
1288 	m->flags = PG_FICTITIOUS;
1289 	/* Fictitious pages don't use "order" or "pool". */
1290 	m->oflags = VPO_UNMANAGED;
1291 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1292 	/* Fictitious pages are unevictable. */
1293 	m->ref_count = 1;
1294 	pmap_page_init(m);
1295 memattr:
1296 	pmap_page_set_memattr(m, memattr);
1297 }
1298 
1299 /*
1300  *	vm_page_putfake:
1301  *
1302  *	Release a fictitious page.
1303  */
1304 void
1305 vm_page_putfake(vm_page_t m)
1306 {
1307 
1308 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1309 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1310 	    ("vm_page_putfake: bad page %p", m));
1311 	vm_page_assert_xbusied(m);
1312 	vm_page_busy_free(m);
1313 	uma_zfree(fakepg_zone, m);
1314 }
1315 
1316 /*
1317  *	vm_page_updatefake:
1318  *
1319  *	Update the given fictitious page to the specified physical address and
1320  *	memory attribute.
1321  */
1322 void
1323 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1324 {
1325 
1326 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1327 	    ("vm_page_updatefake: bad page %p", m));
1328 	m->phys_addr = paddr;
1329 	pmap_page_set_memattr(m, memattr);
1330 }
1331 
1332 /*
1333  *	vm_page_free:
1334  *
1335  *	Free a page.
1336  */
1337 void
1338 vm_page_free(vm_page_t m)
1339 {
1340 
1341 	m->flags &= ~PG_ZERO;
1342 	vm_page_free_toq(m);
1343 }
1344 
1345 /*
1346  *	vm_page_free_zero:
1347  *
1348  *	Free a page to the zerod-pages queue
1349  */
1350 void
1351 vm_page_free_zero(vm_page_t m)
1352 {
1353 
1354 	m->flags |= PG_ZERO;
1355 	vm_page_free_toq(m);
1356 }
1357 
1358 /*
1359  * Unbusy and handle the page queueing for a page from a getpages request that
1360  * was optionally read ahead or behind.
1361  */
1362 void
1363 vm_page_readahead_finish(vm_page_t m)
1364 {
1365 
1366 	/* We shouldn't put invalid pages on queues. */
1367 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1368 
1369 	/*
1370 	 * Since the page is not the actually needed one, whether it should
1371 	 * be activated or deactivated is not obvious.  Empirical results
1372 	 * have shown that deactivating the page is usually the best choice,
1373 	 * unless the page is wanted by another thread.
1374 	 */
1375 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1376 		vm_page_activate(m);
1377 	else
1378 		vm_page_deactivate(m);
1379 	vm_page_xunbusy_unchecked(m);
1380 }
1381 
1382 /*
1383  *	vm_page_sleep_if_busy:
1384  *
1385  *	Sleep and release the object lock if the page is busied.
1386  *	Returns TRUE if the thread slept.
1387  *
1388  *	The given page must be unlocked and object containing it must
1389  *	be locked.
1390  */
1391 int
1392 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1393 {
1394 	vm_object_t obj;
1395 
1396 	vm_page_lock_assert(m, MA_NOTOWNED);
1397 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1398 
1399 	/*
1400 	 * The page-specific object must be cached because page
1401 	 * identity can change during the sleep, causing the
1402 	 * re-lock of a different object.
1403 	 * It is assumed that a reference to the object is already
1404 	 * held by the callers.
1405 	 */
1406 	obj = m->object;
1407 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1408 		VM_OBJECT_WLOCK(obj);
1409 		return (TRUE);
1410 	}
1411 	return (FALSE);
1412 }
1413 
1414 /*
1415  *	vm_page_sleep_if_xbusy:
1416  *
1417  *	Sleep and release the object lock if the page is xbusied.
1418  *	Returns TRUE if the thread slept.
1419  *
1420  *	The given page must be unlocked and object containing it must
1421  *	be locked.
1422  */
1423 int
1424 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1425 {
1426 	vm_object_t obj;
1427 
1428 	vm_page_lock_assert(m, MA_NOTOWNED);
1429 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1430 
1431 	/*
1432 	 * The page-specific object must be cached because page
1433 	 * identity can change during the sleep, causing the
1434 	 * re-lock of a different object.
1435 	 * It is assumed that a reference to the object is already
1436 	 * held by the callers.
1437 	 */
1438 	obj = m->object;
1439 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1440 	    true)) {
1441 		VM_OBJECT_WLOCK(obj);
1442 		return (TRUE);
1443 	}
1444 	return (FALSE);
1445 }
1446 
1447 /*
1448  *	vm_page_dirty_KBI:		[ internal use only ]
1449  *
1450  *	Set all bits in the page's dirty field.
1451  *
1452  *	The object containing the specified page must be locked if the
1453  *	call is made from the machine-independent layer.
1454  *
1455  *	See vm_page_clear_dirty_mask().
1456  *
1457  *	This function should only be called by vm_page_dirty().
1458  */
1459 void
1460 vm_page_dirty_KBI(vm_page_t m)
1461 {
1462 
1463 	/* Refer to this operation by its public name. */
1464 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1465 	m->dirty = VM_PAGE_BITS_ALL;
1466 }
1467 
1468 /*
1469  *	vm_page_insert:		[ internal use only ]
1470  *
1471  *	Inserts the given mem entry into the object and object list.
1472  *
1473  *	The object must be locked.
1474  */
1475 int
1476 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1477 {
1478 	vm_page_t mpred;
1479 
1480 	VM_OBJECT_ASSERT_WLOCKED(object);
1481 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1482 	return (vm_page_insert_after(m, object, pindex, mpred));
1483 }
1484 
1485 /*
1486  *	vm_page_insert_after:
1487  *
1488  *	Inserts the page "m" into the specified object at offset "pindex".
1489  *
1490  *	The page "mpred" must immediately precede the offset "pindex" within
1491  *	the specified object.
1492  *
1493  *	The object must be locked.
1494  */
1495 static int
1496 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1497     vm_page_t mpred)
1498 {
1499 	vm_page_t msucc;
1500 
1501 	VM_OBJECT_ASSERT_WLOCKED(object);
1502 	KASSERT(m->object == NULL,
1503 	    ("vm_page_insert_after: page already inserted"));
1504 	if (mpred != NULL) {
1505 		KASSERT(mpred->object == object,
1506 		    ("vm_page_insert_after: object doesn't contain mpred"));
1507 		KASSERT(mpred->pindex < pindex,
1508 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1509 		msucc = TAILQ_NEXT(mpred, listq);
1510 	} else
1511 		msucc = TAILQ_FIRST(&object->memq);
1512 	if (msucc != NULL)
1513 		KASSERT(msucc->pindex > pindex,
1514 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1515 
1516 	/*
1517 	 * Record the object/offset pair in this page.
1518 	 */
1519 	m->object = object;
1520 	m->pindex = pindex;
1521 	m->ref_count |= VPRC_OBJREF;
1522 
1523 	/*
1524 	 * Now link into the object's ordered list of backed pages.
1525 	 */
1526 	if (vm_radix_insert(&object->rtree, m)) {
1527 		m->object = NULL;
1528 		m->pindex = 0;
1529 		m->ref_count &= ~VPRC_OBJREF;
1530 		return (1);
1531 	}
1532 	vm_page_insert_radixdone(m, object, mpred);
1533 	return (0);
1534 }
1535 
1536 /*
1537  *	vm_page_insert_radixdone:
1538  *
1539  *	Complete page "m" insertion into the specified object after the
1540  *	radix trie hooking.
1541  *
1542  *	The page "mpred" must precede the offset "m->pindex" within the
1543  *	specified object.
1544  *
1545  *	The object must be locked.
1546  */
1547 static void
1548 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1549 {
1550 
1551 	VM_OBJECT_ASSERT_WLOCKED(object);
1552 	KASSERT(object != NULL && m->object == object,
1553 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1554 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1555 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1556 	if (mpred != NULL) {
1557 		KASSERT(mpred->object == object,
1558 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1559 		KASSERT(mpred->pindex < m->pindex,
1560 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1561 	}
1562 
1563 	if (mpred != NULL)
1564 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1565 	else
1566 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1567 
1568 	/*
1569 	 * Show that the object has one more resident page.
1570 	 */
1571 	object->resident_page_count++;
1572 
1573 	/*
1574 	 * Hold the vnode until the last page is released.
1575 	 */
1576 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1577 		vhold(object->handle);
1578 
1579 	/*
1580 	 * Since we are inserting a new and possibly dirty page,
1581 	 * update the object's generation count.
1582 	 */
1583 	if (pmap_page_is_write_mapped(m))
1584 		vm_object_set_writeable_dirty(object);
1585 }
1586 
1587 /*
1588  * Do the work to remove a page from its object.  The caller is responsible for
1589  * updating the page's fields to reflect this removal.
1590  */
1591 static void
1592 vm_page_object_remove(vm_page_t m)
1593 {
1594 	vm_object_t object;
1595 	vm_page_t mrem;
1596 
1597 	vm_page_assert_xbusied(m);
1598 	object = m->object;
1599 	VM_OBJECT_ASSERT_WLOCKED(object);
1600 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1601 	    ("page %p is missing its object ref", m));
1602 
1603 	/* Deferred free of swap space. */
1604 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1605 		vm_pager_page_unswapped(m);
1606 
1607 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1608 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1609 
1610 	/*
1611 	 * Now remove from the object's list of backed pages.
1612 	 */
1613 	TAILQ_REMOVE(&object->memq, m, listq);
1614 
1615 	/*
1616 	 * And show that the object has one fewer resident page.
1617 	 */
1618 	object->resident_page_count--;
1619 
1620 	/*
1621 	 * The vnode may now be recycled.
1622 	 */
1623 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1624 		vdrop(object->handle);
1625 }
1626 
1627 /*
1628  *	vm_page_remove:
1629  *
1630  *	Removes the specified page from its containing object, but does not
1631  *	invalidate any backing storage.  Returns true if the object's reference
1632  *	was the last reference to the page, and false otherwise.
1633  *
1634  *	The object must be locked and the page must be exclusively busied.
1635  *	The exclusive busy will be released on return.  If this is not the
1636  *	final ref and the caller does not hold a wire reference it may not
1637  *	continue to access the page.
1638  */
1639 bool
1640 vm_page_remove(vm_page_t m)
1641 {
1642 	bool dropped;
1643 
1644 	dropped = vm_page_remove_xbusy(m);
1645 	vm_page_xunbusy(m);
1646 
1647 	return (dropped);
1648 }
1649 
1650 /*
1651  *	vm_page_remove_xbusy
1652  *
1653  *	Removes the page but leaves the xbusy held.  Returns true if this
1654  *	removed the final ref and false otherwise.
1655  */
1656 bool
1657 vm_page_remove_xbusy(vm_page_t m)
1658 {
1659 
1660 	vm_page_object_remove(m);
1661 	m->object = NULL;
1662 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1663 }
1664 
1665 /*
1666  *	vm_page_lookup:
1667  *
1668  *	Returns the page associated with the object/offset
1669  *	pair specified; if none is found, NULL is returned.
1670  *
1671  *	The object must be locked.
1672  */
1673 vm_page_t
1674 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1675 {
1676 
1677 	VM_OBJECT_ASSERT_LOCKED(object);
1678 	return (vm_radix_lookup(&object->rtree, pindex));
1679 }
1680 
1681 /*
1682  *	vm_page_find_least:
1683  *
1684  *	Returns the page associated with the object with least pindex
1685  *	greater than or equal to the parameter pindex, or NULL.
1686  *
1687  *	The object must be locked.
1688  */
1689 vm_page_t
1690 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1691 {
1692 	vm_page_t m;
1693 
1694 	VM_OBJECT_ASSERT_LOCKED(object);
1695 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1696 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1697 	return (m);
1698 }
1699 
1700 /*
1701  * Returns the given page's successor (by pindex) within the object if it is
1702  * resident; if none is found, NULL is returned.
1703  *
1704  * The object must be locked.
1705  */
1706 vm_page_t
1707 vm_page_next(vm_page_t m)
1708 {
1709 	vm_page_t next;
1710 
1711 	VM_OBJECT_ASSERT_LOCKED(m->object);
1712 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1713 		MPASS(next->object == m->object);
1714 		if (next->pindex != m->pindex + 1)
1715 			next = NULL;
1716 	}
1717 	return (next);
1718 }
1719 
1720 /*
1721  * Returns the given page's predecessor (by pindex) within the object if it is
1722  * resident; if none is found, NULL is returned.
1723  *
1724  * The object must be locked.
1725  */
1726 vm_page_t
1727 vm_page_prev(vm_page_t m)
1728 {
1729 	vm_page_t prev;
1730 
1731 	VM_OBJECT_ASSERT_LOCKED(m->object);
1732 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1733 		MPASS(prev->object == m->object);
1734 		if (prev->pindex != m->pindex - 1)
1735 			prev = NULL;
1736 	}
1737 	return (prev);
1738 }
1739 
1740 /*
1741  * Uses the page mnew as a replacement for an existing page at index
1742  * pindex which must be already present in the object.
1743  *
1744  * Both pages must be exclusively busied on enter.  The old page is
1745  * unbusied on exit.
1746  *
1747  * A return value of true means mold is now free.  If this is not the
1748  * final ref and the caller does not hold a wire reference it may not
1749  * continue to access the page.
1750  */
1751 static bool
1752 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1753     vm_page_t mold)
1754 {
1755 	vm_page_t mret;
1756 	bool dropped;
1757 
1758 	VM_OBJECT_ASSERT_WLOCKED(object);
1759 	vm_page_assert_xbusied(mold);
1760 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1761 	    ("vm_page_replace: page %p already in object", mnew));
1762 
1763 	/*
1764 	 * This function mostly follows vm_page_insert() and
1765 	 * vm_page_remove() without the radix, object count and vnode
1766 	 * dance.  Double check such functions for more comments.
1767 	 */
1768 
1769 	mnew->object = object;
1770 	mnew->pindex = pindex;
1771 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1772 	mret = vm_radix_replace(&object->rtree, mnew);
1773 	KASSERT(mret == mold,
1774 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1775 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1776 	    (mnew->oflags & VPO_UNMANAGED),
1777 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1778 
1779 	/* Keep the resident page list in sorted order. */
1780 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1781 	TAILQ_REMOVE(&object->memq, mold, listq);
1782 	mold->object = NULL;
1783 
1784 	/*
1785 	 * The object's resident_page_count does not change because we have
1786 	 * swapped one page for another, but the generation count should
1787 	 * change if the page is dirty.
1788 	 */
1789 	if (pmap_page_is_write_mapped(mnew))
1790 		vm_object_set_writeable_dirty(object);
1791 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1792 	vm_page_xunbusy(mold);
1793 
1794 	return (dropped);
1795 }
1796 
1797 void
1798 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1799     vm_page_t mold)
1800 {
1801 
1802 	vm_page_assert_xbusied(mnew);
1803 
1804 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1805 		vm_page_free(mold);
1806 }
1807 
1808 /*
1809  *	vm_page_rename:
1810  *
1811  *	Move the given memory entry from its
1812  *	current object to the specified target object/offset.
1813  *
1814  *	Note: swap associated with the page must be invalidated by the move.  We
1815  *	      have to do this for several reasons:  (1) we aren't freeing the
1816  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1817  *	      moving the page from object A to B, and will then later move
1818  *	      the backing store from A to B and we can't have a conflict.
1819  *
1820  *	Note: we *always* dirty the page.  It is necessary both for the
1821  *	      fact that we moved it, and because we may be invalidating
1822  *	      swap.
1823  *
1824  *	The objects must be locked.
1825  */
1826 int
1827 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1828 {
1829 	vm_page_t mpred;
1830 	vm_pindex_t opidx;
1831 
1832 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1833 
1834 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1835 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1836 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1837 	    ("vm_page_rename: pindex already renamed"));
1838 
1839 	/*
1840 	 * Create a custom version of vm_page_insert() which does not depend
1841 	 * by m_prev and can cheat on the implementation aspects of the
1842 	 * function.
1843 	 */
1844 	opidx = m->pindex;
1845 	m->pindex = new_pindex;
1846 	if (vm_radix_insert(&new_object->rtree, m)) {
1847 		m->pindex = opidx;
1848 		return (1);
1849 	}
1850 
1851 	/*
1852 	 * The operation cannot fail anymore.  The removal must happen before
1853 	 * the listq iterator is tainted.
1854 	 */
1855 	m->pindex = opidx;
1856 	vm_page_object_remove(m);
1857 
1858 	/* Return back to the new pindex to complete vm_page_insert(). */
1859 	m->pindex = new_pindex;
1860 	m->object = new_object;
1861 
1862 	vm_page_insert_radixdone(m, new_object, mpred);
1863 	vm_page_dirty(m);
1864 	return (0);
1865 }
1866 
1867 /*
1868  *	vm_page_alloc:
1869  *
1870  *	Allocate and return a page that is associated with the specified
1871  *	object and offset pair.  By default, this page is exclusive busied.
1872  *
1873  *	The caller must always specify an allocation class.
1874  *
1875  *	allocation classes:
1876  *	VM_ALLOC_NORMAL		normal process request
1877  *	VM_ALLOC_SYSTEM		system *really* needs a page
1878  *	VM_ALLOC_INTERRUPT	interrupt time request
1879  *
1880  *	optional allocation flags:
1881  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1882  *				intends to allocate
1883  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1884  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1885  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1886  *				should not be exclusive busy
1887  *	VM_ALLOC_SBUSY		shared busy the allocated page
1888  *	VM_ALLOC_WIRED		wire the allocated page
1889  *	VM_ALLOC_ZERO		prefer a zeroed page
1890  */
1891 vm_page_t
1892 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1893 {
1894 
1895 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1896 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1897 }
1898 
1899 vm_page_t
1900 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1901     int req)
1902 {
1903 
1904 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1905 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1906 	    NULL));
1907 }
1908 
1909 /*
1910  * Allocate a page in the specified object with the given page index.  To
1911  * optimize insertion of the page into the object, the caller must also specifiy
1912  * the resident page in the object with largest index smaller than the given
1913  * page index, or NULL if no such page exists.
1914  */
1915 vm_page_t
1916 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1917     int req, vm_page_t mpred)
1918 {
1919 	struct vm_domainset_iter di;
1920 	vm_page_t m;
1921 	int domain;
1922 
1923 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1924 	do {
1925 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1926 		    mpred);
1927 		if (m != NULL)
1928 			break;
1929 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1930 
1931 	return (m);
1932 }
1933 
1934 /*
1935  * Returns true if the number of free pages exceeds the minimum
1936  * for the request class and false otherwise.
1937  */
1938 static int
1939 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1940 {
1941 	u_int limit, old, new;
1942 
1943 	if (req_class == VM_ALLOC_INTERRUPT)
1944 		limit = 0;
1945 	else if (req_class == VM_ALLOC_SYSTEM)
1946 		limit = vmd->vmd_interrupt_free_min;
1947 	else
1948 		limit = vmd->vmd_free_reserved;
1949 
1950 	/*
1951 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1952 	 */
1953 	limit += npages;
1954 	old = vmd->vmd_free_count;
1955 	do {
1956 		if (old < limit)
1957 			return (0);
1958 		new = old - npages;
1959 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1960 
1961 	/* Wake the page daemon if we've crossed the threshold. */
1962 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1963 		pagedaemon_wakeup(vmd->vmd_domain);
1964 
1965 	/* Only update bitsets on transitions. */
1966 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1967 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1968 		vm_domain_set(vmd);
1969 
1970 	return (1);
1971 }
1972 
1973 int
1974 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1975 {
1976 	int req_class;
1977 
1978 	/*
1979 	 * The page daemon is allowed to dig deeper into the free page list.
1980 	 */
1981 	req_class = req & VM_ALLOC_CLASS_MASK;
1982 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1983 		req_class = VM_ALLOC_SYSTEM;
1984 	return (_vm_domain_allocate(vmd, req_class, npages));
1985 }
1986 
1987 vm_page_t
1988 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1989     int req, vm_page_t mpred)
1990 {
1991 	struct vm_domain *vmd;
1992 	vm_page_t m;
1993 	int flags, pool;
1994 
1995 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1996 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1997 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1998 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1999 	    ("inconsistent object(%p)/req(%x)", object, req));
2000 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2001 	    ("Can't sleep and retry object insertion."));
2002 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2003 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2004 	    (uintmax_t)pindex));
2005 	if (object != NULL)
2006 		VM_OBJECT_ASSERT_WLOCKED(object);
2007 
2008 	flags = 0;
2009 	m = NULL;
2010 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2011 again:
2012 #if VM_NRESERVLEVEL > 0
2013 	/*
2014 	 * Can we allocate the page from a reservation?
2015 	 */
2016 	if (vm_object_reserv(object) &&
2017 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2018 	    NULL) {
2019 		domain = vm_phys_domain(m);
2020 		vmd = VM_DOMAIN(domain);
2021 		goto found;
2022 	}
2023 #endif
2024 	vmd = VM_DOMAIN(domain);
2025 	if (vmd->vmd_pgcache[pool].zone != NULL) {
2026 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM);
2027 		if (m != NULL) {
2028 			flags |= PG_PCPU_CACHE;
2029 			goto found;
2030 		}
2031 	}
2032 	if (vm_domain_allocate(vmd, req, 1)) {
2033 		/*
2034 		 * If not, allocate it from the free page queues.
2035 		 */
2036 		vm_domain_free_lock(vmd);
2037 		m = vm_phys_alloc_pages(domain, pool, 0);
2038 		vm_domain_free_unlock(vmd);
2039 		if (m == NULL) {
2040 			vm_domain_freecnt_inc(vmd, 1);
2041 #if VM_NRESERVLEVEL > 0
2042 			if (vm_reserv_reclaim_inactive(domain))
2043 				goto again;
2044 #endif
2045 		}
2046 	}
2047 	if (m == NULL) {
2048 		/*
2049 		 * Not allocatable, give up.
2050 		 */
2051 		if (vm_domain_alloc_fail(vmd, object, req))
2052 			goto again;
2053 		return (NULL);
2054 	}
2055 
2056 	/*
2057 	 * At this point we had better have found a good page.
2058 	 */
2059 found:
2060 	vm_page_dequeue(m);
2061 	vm_page_alloc_check(m);
2062 
2063 	/*
2064 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2065 	 */
2066 	if ((req & VM_ALLOC_ZERO) != 0)
2067 		flags |= (m->flags & PG_ZERO);
2068 	if ((req & VM_ALLOC_NODUMP) != 0)
2069 		flags |= PG_NODUMP;
2070 	m->flags = flags;
2071 	m->a.flags = 0;
2072 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2073 	    VPO_UNMANAGED : 0;
2074 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2075 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2076 	else if ((req & VM_ALLOC_SBUSY) != 0)
2077 		m->busy_lock = VPB_SHARERS_WORD(1);
2078 	else
2079 		m->busy_lock = VPB_UNBUSIED;
2080 	if (req & VM_ALLOC_WIRED) {
2081 		vm_wire_add(1);
2082 		m->ref_count = 1;
2083 	}
2084 	m->a.act_count = 0;
2085 
2086 	if (object != NULL) {
2087 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2088 			if (req & VM_ALLOC_WIRED) {
2089 				vm_wire_sub(1);
2090 				m->ref_count = 0;
2091 			}
2092 			KASSERT(m->object == NULL, ("page %p has object", m));
2093 			m->oflags = VPO_UNMANAGED;
2094 			m->busy_lock = VPB_UNBUSIED;
2095 			/* Don't change PG_ZERO. */
2096 			vm_page_free_toq(m);
2097 			if (req & VM_ALLOC_WAITFAIL) {
2098 				VM_OBJECT_WUNLOCK(object);
2099 				vm_radix_wait();
2100 				VM_OBJECT_WLOCK(object);
2101 			}
2102 			return (NULL);
2103 		}
2104 
2105 		/* Ignore device objects; the pager sets "memattr" for them. */
2106 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2107 		    (object->flags & OBJ_FICTITIOUS) == 0)
2108 			pmap_page_set_memattr(m, object->memattr);
2109 	} else
2110 		m->pindex = pindex;
2111 
2112 	return (m);
2113 }
2114 
2115 /*
2116  *	vm_page_alloc_contig:
2117  *
2118  *	Allocate a contiguous set of physical pages of the given size "npages"
2119  *	from the free lists.  All of the physical pages must be at or above
2120  *	the given physical address "low" and below the given physical address
2121  *	"high".  The given value "alignment" determines the alignment of the
2122  *	first physical page in the set.  If the given value "boundary" is
2123  *	non-zero, then the set of physical pages cannot cross any physical
2124  *	address boundary that is a multiple of that value.  Both "alignment"
2125  *	and "boundary" must be a power of two.
2126  *
2127  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2128  *	then the memory attribute setting for the physical pages is configured
2129  *	to the object's memory attribute setting.  Otherwise, the memory
2130  *	attribute setting for the physical pages is configured to "memattr",
2131  *	overriding the object's memory attribute setting.  However, if the
2132  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2133  *	memory attribute setting for the physical pages cannot be configured
2134  *	to VM_MEMATTR_DEFAULT.
2135  *
2136  *	The specified object may not contain fictitious pages.
2137  *
2138  *	The caller must always specify an allocation class.
2139  *
2140  *	allocation classes:
2141  *	VM_ALLOC_NORMAL		normal process request
2142  *	VM_ALLOC_SYSTEM		system *really* needs a page
2143  *	VM_ALLOC_INTERRUPT	interrupt time request
2144  *
2145  *	optional allocation flags:
2146  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2147  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2148  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2149  *				should not be exclusive busy
2150  *	VM_ALLOC_SBUSY		shared busy the allocated page
2151  *	VM_ALLOC_WIRED		wire the allocated page
2152  *	VM_ALLOC_ZERO		prefer a zeroed page
2153  */
2154 vm_page_t
2155 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2156     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2157     vm_paddr_t boundary, vm_memattr_t memattr)
2158 {
2159 	struct vm_domainset_iter di;
2160 	vm_page_t m;
2161 	int domain;
2162 
2163 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2164 	do {
2165 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2166 		    npages, low, high, alignment, boundary, memattr);
2167 		if (m != NULL)
2168 			break;
2169 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2170 
2171 	return (m);
2172 }
2173 
2174 vm_page_t
2175 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2176     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2177     vm_paddr_t boundary, vm_memattr_t memattr)
2178 {
2179 	struct vm_domain *vmd;
2180 	vm_page_t m, m_ret, mpred;
2181 	u_int busy_lock, flags, oflags;
2182 
2183 	mpred = NULL;	/* XXX: pacify gcc */
2184 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2185 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2186 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2187 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2188 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2189 	    req));
2190 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2191 	    ("Can't sleep and retry object insertion."));
2192 	if (object != NULL) {
2193 		VM_OBJECT_ASSERT_WLOCKED(object);
2194 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2195 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2196 		    object));
2197 	}
2198 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2199 
2200 	if (object != NULL) {
2201 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2202 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2203 		    ("vm_page_alloc_contig: pindex already allocated"));
2204 	}
2205 
2206 	/*
2207 	 * Can we allocate the pages without the number of free pages falling
2208 	 * below the lower bound for the allocation class?
2209 	 */
2210 	m_ret = NULL;
2211 again:
2212 #if VM_NRESERVLEVEL > 0
2213 	/*
2214 	 * Can we allocate the pages from a reservation?
2215 	 */
2216 	if (vm_object_reserv(object) &&
2217 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2218 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2219 		domain = vm_phys_domain(m_ret);
2220 		vmd = VM_DOMAIN(domain);
2221 		goto found;
2222 	}
2223 #endif
2224 	vmd = VM_DOMAIN(domain);
2225 	if (vm_domain_allocate(vmd, req, npages)) {
2226 		/*
2227 		 * allocate them from the free page queues.
2228 		 */
2229 		vm_domain_free_lock(vmd);
2230 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2231 		    alignment, boundary);
2232 		vm_domain_free_unlock(vmd);
2233 		if (m_ret == NULL) {
2234 			vm_domain_freecnt_inc(vmd, npages);
2235 #if VM_NRESERVLEVEL > 0
2236 			if (vm_reserv_reclaim_contig(domain, npages, low,
2237 			    high, alignment, boundary))
2238 				goto again;
2239 #endif
2240 		}
2241 	}
2242 	if (m_ret == NULL) {
2243 		if (vm_domain_alloc_fail(vmd, object, req))
2244 			goto again;
2245 		return (NULL);
2246 	}
2247 #if VM_NRESERVLEVEL > 0
2248 found:
2249 #endif
2250 	for (m = m_ret; m < &m_ret[npages]; m++) {
2251 		vm_page_dequeue(m);
2252 		vm_page_alloc_check(m);
2253 	}
2254 
2255 	/*
2256 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2257 	 */
2258 	flags = 0;
2259 	if ((req & VM_ALLOC_ZERO) != 0)
2260 		flags = PG_ZERO;
2261 	if ((req & VM_ALLOC_NODUMP) != 0)
2262 		flags |= PG_NODUMP;
2263 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2264 	    VPO_UNMANAGED : 0;
2265 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2266 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2267 	else if ((req & VM_ALLOC_SBUSY) != 0)
2268 		busy_lock = VPB_SHARERS_WORD(1);
2269 	else
2270 		busy_lock = VPB_UNBUSIED;
2271 	if ((req & VM_ALLOC_WIRED) != 0)
2272 		vm_wire_add(npages);
2273 	if (object != NULL) {
2274 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2275 		    memattr == VM_MEMATTR_DEFAULT)
2276 			memattr = object->memattr;
2277 	}
2278 	for (m = m_ret; m < &m_ret[npages]; m++) {
2279 		m->a.flags = 0;
2280 		m->flags = (m->flags | PG_NODUMP) & flags;
2281 		m->busy_lock = busy_lock;
2282 		if ((req & VM_ALLOC_WIRED) != 0)
2283 			m->ref_count = 1;
2284 		m->a.act_count = 0;
2285 		m->oflags = oflags;
2286 		if (object != NULL) {
2287 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2288 				if ((req & VM_ALLOC_WIRED) != 0)
2289 					vm_wire_sub(npages);
2290 				KASSERT(m->object == NULL,
2291 				    ("page %p has object", m));
2292 				mpred = m;
2293 				for (m = m_ret; m < &m_ret[npages]; m++) {
2294 					if (m <= mpred &&
2295 					    (req & VM_ALLOC_WIRED) != 0)
2296 						m->ref_count = 0;
2297 					m->oflags = VPO_UNMANAGED;
2298 					m->busy_lock = VPB_UNBUSIED;
2299 					/* Don't change PG_ZERO. */
2300 					vm_page_free_toq(m);
2301 				}
2302 				if (req & VM_ALLOC_WAITFAIL) {
2303 					VM_OBJECT_WUNLOCK(object);
2304 					vm_radix_wait();
2305 					VM_OBJECT_WLOCK(object);
2306 				}
2307 				return (NULL);
2308 			}
2309 			mpred = m;
2310 		} else
2311 			m->pindex = pindex;
2312 		if (memattr != VM_MEMATTR_DEFAULT)
2313 			pmap_page_set_memattr(m, memattr);
2314 		pindex++;
2315 	}
2316 	return (m_ret);
2317 }
2318 
2319 /*
2320  * Check a page that has been freshly dequeued from a freelist.
2321  */
2322 static void
2323 vm_page_alloc_check(vm_page_t m)
2324 {
2325 
2326 	KASSERT(m->object == NULL, ("page %p has object", m));
2327 	KASSERT(m->a.queue == PQ_NONE &&
2328 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2329 	    ("page %p has unexpected queue %d, flags %#x",
2330 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2331 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2332 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2333 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2334 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2335 	    ("page %p has unexpected memattr %d",
2336 	    m, pmap_page_get_memattr(m)));
2337 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2338 }
2339 
2340 /*
2341  * 	vm_page_alloc_freelist:
2342  *
2343  *	Allocate a physical page from the specified free page list.
2344  *
2345  *	The caller must always specify an allocation class.
2346  *
2347  *	allocation classes:
2348  *	VM_ALLOC_NORMAL		normal process request
2349  *	VM_ALLOC_SYSTEM		system *really* needs a page
2350  *	VM_ALLOC_INTERRUPT	interrupt time request
2351  *
2352  *	optional allocation flags:
2353  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2354  *				intends to allocate
2355  *	VM_ALLOC_WIRED		wire the allocated page
2356  *	VM_ALLOC_ZERO		prefer a zeroed page
2357  */
2358 vm_page_t
2359 vm_page_alloc_freelist(int freelist, int req)
2360 {
2361 	struct vm_domainset_iter di;
2362 	vm_page_t m;
2363 	int domain;
2364 
2365 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2366 	do {
2367 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2368 		if (m != NULL)
2369 			break;
2370 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2371 
2372 	return (m);
2373 }
2374 
2375 vm_page_t
2376 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2377 {
2378 	struct vm_domain *vmd;
2379 	vm_page_t m;
2380 	u_int flags;
2381 
2382 	m = NULL;
2383 	vmd = VM_DOMAIN(domain);
2384 again:
2385 	if (vm_domain_allocate(vmd, req, 1)) {
2386 		vm_domain_free_lock(vmd);
2387 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2388 		    VM_FREEPOOL_DIRECT, 0);
2389 		vm_domain_free_unlock(vmd);
2390 		if (m == NULL)
2391 			vm_domain_freecnt_inc(vmd, 1);
2392 	}
2393 	if (m == NULL) {
2394 		if (vm_domain_alloc_fail(vmd, NULL, req))
2395 			goto again;
2396 		return (NULL);
2397 	}
2398 	vm_page_dequeue(m);
2399 	vm_page_alloc_check(m);
2400 
2401 	/*
2402 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2403 	 */
2404 	m->a.flags = 0;
2405 	flags = 0;
2406 	if ((req & VM_ALLOC_ZERO) != 0)
2407 		flags = PG_ZERO;
2408 	m->flags &= flags;
2409 	if ((req & VM_ALLOC_WIRED) != 0) {
2410 		vm_wire_add(1);
2411 		m->ref_count = 1;
2412 	}
2413 	/* Unmanaged pages don't use "act_count". */
2414 	m->oflags = VPO_UNMANAGED;
2415 	return (m);
2416 }
2417 
2418 static int
2419 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2420 {
2421 	struct vm_domain *vmd;
2422 	struct vm_pgcache *pgcache;
2423 	int i;
2424 
2425 	pgcache = arg;
2426 	vmd = VM_DOMAIN(pgcache->domain);
2427 
2428 	/*
2429 	 * The page daemon should avoid creating extra memory pressure since its
2430 	 * main purpose is to replenish the store of free pages.
2431 	 */
2432 	if (vmd->vmd_severeset || curproc == pageproc ||
2433 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2434 		return (0);
2435 	domain = vmd->vmd_domain;
2436 	vm_domain_free_lock(vmd);
2437 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2438 	    (vm_page_t *)store);
2439 	vm_domain_free_unlock(vmd);
2440 	if (cnt != i)
2441 		vm_domain_freecnt_inc(vmd, cnt - i);
2442 
2443 	return (i);
2444 }
2445 
2446 static void
2447 vm_page_zone_release(void *arg, void **store, int cnt)
2448 {
2449 	struct vm_domain *vmd;
2450 	struct vm_pgcache *pgcache;
2451 	vm_page_t m;
2452 	int i;
2453 
2454 	pgcache = arg;
2455 	vmd = VM_DOMAIN(pgcache->domain);
2456 	vm_domain_free_lock(vmd);
2457 	for (i = 0; i < cnt; i++) {
2458 		m = (vm_page_t)store[i];
2459 		vm_phys_free_pages(m, 0);
2460 	}
2461 	vm_domain_free_unlock(vmd);
2462 	vm_domain_freecnt_inc(vmd, cnt);
2463 }
2464 
2465 #define	VPSC_ANY	0	/* No restrictions. */
2466 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2467 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2468 
2469 /*
2470  *	vm_page_scan_contig:
2471  *
2472  *	Scan vm_page_array[] between the specified entries "m_start" and
2473  *	"m_end" for a run of contiguous physical pages that satisfy the
2474  *	specified conditions, and return the lowest page in the run.  The
2475  *	specified "alignment" determines the alignment of the lowest physical
2476  *	page in the run.  If the specified "boundary" is non-zero, then the
2477  *	run of physical pages cannot span a physical address that is a
2478  *	multiple of "boundary".
2479  *
2480  *	"m_end" is never dereferenced, so it need not point to a vm_page
2481  *	structure within vm_page_array[].
2482  *
2483  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2484  *	span a hole (or discontiguity) in the physical address space.  Both
2485  *	"alignment" and "boundary" must be a power of two.
2486  */
2487 vm_page_t
2488 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2489     u_long alignment, vm_paddr_t boundary, int options)
2490 {
2491 	vm_object_t object;
2492 	vm_paddr_t pa;
2493 	vm_page_t m, m_run;
2494 #if VM_NRESERVLEVEL > 0
2495 	int level;
2496 #endif
2497 	int m_inc, order, run_ext, run_len;
2498 
2499 	KASSERT(npages > 0, ("npages is 0"));
2500 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2501 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2502 	m_run = NULL;
2503 	run_len = 0;
2504 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2505 		KASSERT((m->flags & PG_MARKER) == 0,
2506 		    ("page %p is PG_MARKER", m));
2507 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2508 		    ("fictitious page %p has invalid ref count", m));
2509 
2510 		/*
2511 		 * If the current page would be the start of a run, check its
2512 		 * physical address against the end, alignment, and boundary
2513 		 * conditions.  If it doesn't satisfy these conditions, either
2514 		 * terminate the scan or advance to the next page that
2515 		 * satisfies the failed condition.
2516 		 */
2517 		if (run_len == 0) {
2518 			KASSERT(m_run == NULL, ("m_run != NULL"));
2519 			if (m + npages > m_end)
2520 				break;
2521 			pa = VM_PAGE_TO_PHYS(m);
2522 			if ((pa & (alignment - 1)) != 0) {
2523 				m_inc = atop(roundup2(pa, alignment) - pa);
2524 				continue;
2525 			}
2526 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2527 			    boundary) != 0) {
2528 				m_inc = atop(roundup2(pa, boundary) - pa);
2529 				continue;
2530 			}
2531 		} else
2532 			KASSERT(m_run != NULL, ("m_run == NULL"));
2533 
2534 retry:
2535 		m_inc = 1;
2536 		if (vm_page_wired(m))
2537 			run_ext = 0;
2538 #if VM_NRESERVLEVEL > 0
2539 		else if ((level = vm_reserv_level(m)) >= 0 &&
2540 		    (options & VPSC_NORESERV) != 0) {
2541 			run_ext = 0;
2542 			/* Advance to the end of the reservation. */
2543 			pa = VM_PAGE_TO_PHYS(m);
2544 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2545 			    pa);
2546 		}
2547 #endif
2548 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2549 			/*
2550 			 * The page is considered eligible for relocation if
2551 			 * and only if it could be laundered or reclaimed by
2552 			 * the page daemon.
2553 			 */
2554 			VM_OBJECT_RLOCK(object);
2555 			if (object != m->object) {
2556 				VM_OBJECT_RUNLOCK(object);
2557 				goto retry;
2558 			}
2559 			/* Don't care: PG_NODUMP, PG_ZERO. */
2560 			if (object->type != OBJT_DEFAULT &&
2561 			    object->type != OBJT_SWAP &&
2562 			    object->type != OBJT_VNODE) {
2563 				run_ext = 0;
2564 #if VM_NRESERVLEVEL > 0
2565 			} else if ((options & VPSC_NOSUPER) != 0 &&
2566 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2567 				run_ext = 0;
2568 				/* Advance to the end of the superpage. */
2569 				pa = VM_PAGE_TO_PHYS(m);
2570 				m_inc = atop(roundup2(pa + 1,
2571 				    vm_reserv_size(level)) - pa);
2572 #endif
2573 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2574 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2575 				/*
2576 				 * The page is allocated but eligible for
2577 				 * relocation.  Extend the current run by one
2578 				 * page.
2579 				 */
2580 				KASSERT(pmap_page_get_memattr(m) ==
2581 				    VM_MEMATTR_DEFAULT,
2582 				    ("page %p has an unexpected memattr", m));
2583 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2584 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2585 				    ("page %p has unexpected oflags", m));
2586 				/* Don't care: PGA_NOSYNC. */
2587 				run_ext = 1;
2588 			} else
2589 				run_ext = 0;
2590 			VM_OBJECT_RUNLOCK(object);
2591 #if VM_NRESERVLEVEL > 0
2592 		} else if (level >= 0) {
2593 			/*
2594 			 * The page is reserved but not yet allocated.  In
2595 			 * other words, it is still free.  Extend the current
2596 			 * run by one page.
2597 			 */
2598 			run_ext = 1;
2599 #endif
2600 		} else if ((order = m->order) < VM_NFREEORDER) {
2601 			/*
2602 			 * The page is enqueued in the physical memory
2603 			 * allocator's free page queues.  Moreover, it is the
2604 			 * first page in a power-of-two-sized run of
2605 			 * contiguous free pages.  Add these pages to the end
2606 			 * of the current run, and jump ahead.
2607 			 */
2608 			run_ext = 1 << order;
2609 			m_inc = 1 << order;
2610 		} else {
2611 			/*
2612 			 * Skip the page for one of the following reasons: (1)
2613 			 * It is enqueued in the physical memory allocator's
2614 			 * free page queues.  However, it is not the first
2615 			 * page in a run of contiguous free pages.  (This case
2616 			 * rarely occurs because the scan is performed in
2617 			 * ascending order.) (2) It is not reserved, and it is
2618 			 * transitioning from free to allocated.  (Conversely,
2619 			 * the transition from allocated to free for managed
2620 			 * pages is blocked by the page lock.) (3) It is
2621 			 * allocated but not contained by an object and not
2622 			 * wired, e.g., allocated by Xen's balloon driver.
2623 			 */
2624 			run_ext = 0;
2625 		}
2626 
2627 		/*
2628 		 * Extend or reset the current run of pages.
2629 		 */
2630 		if (run_ext > 0) {
2631 			if (run_len == 0)
2632 				m_run = m;
2633 			run_len += run_ext;
2634 		} else {
2635 			if (run_len > 0) {
2636 				m_run = NULL;
2637 				run_len = 0;
2638 			}
2639 		}
2640 	}
2641 	if (run_len >= npages)
2642 		return (m_run);
2643 	return (NULL);
2644 }
2645 
2646 /*
2647  *	vm_page_reclaim_run:
2648  *
2649  *	Try to relocate each of the allocated virtual pages within the
2650  *	specified run of physical pages to a new physical address.  Free the
2651  *	physical pages underlying the relocated virtual pages.  A virtual page
2652  *	is relocatable if and only if it could be laundered or reclaimed by
2653  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2654  *	physical address above "high".
2655  *
2656  *	Returns 0 if every physical page within the run was already free or
2657  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2658  *	value indicating why the last attempt to relocate a virtual page was
2659  *	unsuccessful.
2660  *
2661  *	"req_class" must be an allocation class.
2662  */
2663 static int
2664 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2665     vm_paddr_t high)
2666 {
2667 	struct vm_domain *vmd;
2668 	struct spglist free;
2669 	vm_object_t object;
2670 	vm_paddr_t pa;
2671 	vm_page_t m, m_end, m_new;
2672 	int error, order, req;
2673 
2674 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2675 	    ("req_class is not an allocation class"));
2676 	SLIST_INIT(&free);
2677 	error = 0;
2678 	m = m_run;
2679 	m_end = m_run + npages;
2680 	for (; error == 0 && m < m_end; m++) {
2681 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2682 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2683 
2684 		/*
2685 		 * Racily check for wirings.  Races are handled once the object
2686 		 * lock is held and the page is unmapped.
2687 		 */
2688 		if (vm_page_wired(m))
2689 			error = EBUSY;
2690 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2691 			/*
2692 			 * The page is relocated if and only if it could be
2693 			 * laundered or reclaimed by the page daemon.
2694 			 */
2695 			VM_OBJECT_WLOCK(object);
2696 			/* Don't care: PG_NODUMP, PG_ZERO. */
2697 			if (m->object != object ||
2698 			    (object->type != OBJT_DEFAULT &&
2699 			    object->type != OBJT_SWAP &&
2700 			    object->type != OBJT_VNODE))
2701 				error = EINVAL;
2702 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2703 				error = EINVAL;
2704 			else if (vm_page_queue(m) != PQ_NONE &&
2705 			    vm_page_tryxbusy(m) != 0) {
2706 				if (vm_page_wired(m)) {
2707 					vm_page_xunbusy(m);
2708 					error = EBUSY;
2709 					goto unlock;
2710 				}
2711 				KASSERT(pmap_page_get_memattr(m) ==
2712 				    VM_MEMATTR_DEFAULT,
2713 				    ("page %p has an unexpected memattr", m));
2714 				KASSERT(m->oflags == 0,
2715 				    ("page %p has unexpected oflags", m));
2716 				/* Don't care: PGA_NOSYNC. */
2717 				if (!vm_page_none_valid(m)) {
2718 					/*
2719 					 * First, try to allocate a new page
2720 					 * that is above "high".  Failing
2721 					 * that, try to allocate a new page
2722 					 * that is below "m_run".  Allocate
2723 					 * the new page between the end of
2724 					 * "m_run" and "high" only as a last
2725 					 * resort.
2726 					 */
2727 					req = req_class | VM_ALLOC_NOOBJ;
2728 					if ((m->flags & PG_NODUMP) != 0)
2729 						req |= VM_ALLOC_NODUMP;
2730 					if (trunc_page(high) !=
2731 					    ~(vm_paddr_t)PAGE_MASK) {
2732 						m_new = vm_page_alloc_contig(
2733 						    NULL, 0, req, 1,
2734 						    round_page(high),
2735 						    ~(vm_paddr_t)0,
2736 						    PAGE_SIZE, 0,
2737 						    VM_MEMATTR_DEFAULT);
2738 					} else
2739 						m_new = NULL;
2740 					if (m_new == NULL) {
2741 						pa = VM_PAGE_TO_PHYS(m_run);
2742 						m_new = vm_page_alloc_contig(
2743 						    NULL, 0, req, 1,
2744 						    0, pa - 1, PAGE_SIZE, 0,
2745 						    VM_MEMATTR_DEFAULT);
2746 					}
2747 					if (m_new == NULL) {
2748 						pa += ptoa(npages);
2749 						m_new = vm_page_alloc_contig(
2750 						    NULL, 0, req, 1,
2751 						    pa, high, PAGE_SIZE, 0,
2752 						    VM_MEMATTR_DEFAULT);
2753 					}
2754 					if (m_new == NULL) {
2755 						vm_page_xunbusy(m);
2756 						error = ENOMEM;
2757 						goto unlock;
2758 					}
2759 
2760 					/*
2761 					 * Unmap the page and check for new
2762 					 * wirings that may have been acquired
2763 					 * through a pmap lookup.
2764 					 */
2765 					if (object->ref_count != 0 &&
2766 					    !vm_page_try_remove_all(m)) {
2767 						vm_page_xunbusy(m);
2768 						vm_page_free(m_new);
2769 						error = EBUSY;
2770 						goto unlock;
2771 					}
2772 
2773 					/*
2774 					 * Replace "m" with the new page.  For
2775 					 * vm_page_replace(), "m" must be busy
2776 					 * and dequeued.  Finally, change "m"
2777 					 * as if vm_page_free() was called.
2778 					 */
2779 					m_new->a.flags = m->a.flags &
2780 					    ~PGA_QUEUE_STATE_MASK;
2781 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2782 					    ("page %p is managed", m_new));
2783 					m_new->oflags = 0;
2784 					pmap_copy_page(m, m_new);
2785 					m_new->valid = m->valid;
2786 					m_new->dirty = m->dirty;
2787 					m->flags &= ~PG_ZERO;
2788 					vm_page_dequeue(m);
2789 					if (vm_page_replace_hold(m_new, object,
2790 					    m->pindex, m) &&
2791 					    vm_page_free_prep(m))
2792 						SLIST_INSERT_HEAD(&free, m,
2793 						    plinks.s.ss);
2794 
2795 					/*
2796 					 * The new page must be deactivated
2797 					 * before the object is unlocked.
2798 					 */
2799 					vm_page_deactivate(m_new);
2800 				} else {
2801 					m->flags &= ~PG_ZERO;
2802 					vm_page_dequeue(m);
2803 					if (vm_page_free_prep(m))
2804 						SLIST_INSERT_HEAD(&free, m,
2805 						    plinks.s.ss);
2806 					KASSERT(m->dirty == 0,
2807 					    ("page %p is dirty", m));
2808 				}
2809 			} else
2810 				error = EBUSY;
2811 unlock:
2812 			VM_OBJECT_WUNLOCK(object);
2813 		} else {
2814 			MPASS(vm_phys_domain(m) == domain);
2815 			vmd = VM_DOMAIN(domain);
2816 			vm_domain_free_lock(vmd);
2817 			order = m->order;
2818 			if (order < VM_NFREEORDER) {
2819 				/*
2820 				 * The page is enqueued in the physical memory
2821 				 * allocator's free page queues.  Moreover, it
2822 				 * is the first page in a power-of-two-sized
2823 				 * run of contiguous free pages.  Jump ahead
2824 				 * to the last page within that run, and
2825 				 * continue from there.
2826 				 */
2827 				m += (1 << order) - 1;
2828 			}
2829 #if VM_NRESERVLEVEL > 0
2830 			else if (vm_reserv_is_page_free(m))
2831 				order = 0;
2832 #endif
2833 			vm_domain_free_unlock(vmd);
2834 			if (order == VM_NFREEORDER)
2835 				error = EINVAL;
2836 		}
2837 	}
2838 	if ((m = SLIST_FIRST(&free)) != NULL) {
2839 		int cnt;
2840 
2841 		vmd = VM_DOMAIN(domain);
2842 		cnt = 0;
2843 		vm_domain_free_lock(vmd);
2844 		do {
2845 			MPASS(vm_phys_domain(m) == domain);
2846 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2847 			vm_phys_free_pages(m, 0);
2848 			cnt++;
2849 		} while ((m = SLIST_FIRST(&free)) != NULL);
2850 		vm_domain_free_unlock(vmd);
2851 		vm_domain_freecnt_inc(vmd, cnt);
2852 	}
2853 	return (error);
2854 }
2855 
2856 #define	NRUNS	16
2857 
2858 CTASSERT(powerof2(NRUNS));
2859 
2860 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2861 
2862 #define	MIN_RECLAIM	8
2863 
2864 /*
2865  *	vm_page_reclaim_contig:
2866  *
2867  *	Reclaim allocated, contiguous physical memory satisfying the specified
2868  *	conditions by relocating the virtual pages using that physical memory.
2869  *	Returns true if reclamation is successful and false otherwise.  Since
2870  *	relocation requires the allocation of physical pages, reclamation may
2871  *	fail due to a shortage of free pages.  When reclamation fails, callers
2872  *	are expected to perform vm_wait() before retrying a failed allocation
2873  *	operation, e.g., vm_page_alloc_contig().
2874  *
2875  *	The caller must always specify an allocation class through "req".
2876  *
2877  *	allocation classes:
2878  *	VM_ALLOC_NORMAL		normal process request
2879  *	VM_ALLOC_SYSTEM		system *really* needs a page
2880  *	VM_ALLOC_INTERRUPT	interrupt time request
2881  *
2882  *	The optional allocation flags are ignored.
2883  *
2884  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2885  *	must be a power of two.
2886  */
2887 bool
2888 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2889     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2890 {
2891 	struct vm_domain *vmd;
2892 	vm_paddr_t curr_low;
2893 	vm_page_t m_run, m_runs[NRUNS];
2894 	u_long count, reclaimed;
2895 	int error, i, options, req_class;
2896 
2897 	KASSERT(npages > 0, ("npages is 0"));
2898 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2899 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2900 	req_class = req & VM_ALLOC_CLASS_MASK;
2901 
2902 	/*
2903 	 * The page daemon is allowed to dig deeper into the free page list.
2904 	 */
2905 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2906 		req_class = VM_ALLOC_SYSTEM;
2907 
2908 	/*
2909 	 * Return if the number of free pages cannot satisfy the requested
2910 	 * allocation.
2911 	 */
2912 	vmd = VM_DOMAIN(domain);
2913 	count = vmd->vmd_free_count;
2914 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2915 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2916 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2917 		return (false);
2918 
2919 	/*
2920 	 * Scan up to three times, relaxing the restrictions ("options") on
2921 	 * the reclamation of reservations and superpages each time.
2922 	 */
2923 	for (options = VPSC_NORESERV;;) {
2924 		/*
2925 		 * Find the highest runs that satisfy the given constraints
2926 		 * and restrictions, and record them in "m_runs".
2927 		 */
2928 		curr_low = low;
2929 		count = 0;
2930 		for (;;) {
2931 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2932 			    high, alignment, boundary, options);
2933 			if (m_run == NULL)
2934 				break;
2935 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2936 			m_runs[RUN_INDEX(count)] = m_run;
2937 			count++;
2938 		}
2939 
2940 		/*
2941 		 * Reclaim the highest runs in LIFO (descending) order until
2942 		 * the number of reclaimed pages, "reclaimed", is at least
2943 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2944 		 * reclamation is idempotent, and runs will (likely) recur
2945 		 * from one scan to the next as restrictions are relaxed.
2946 		 */
2947 		reclaimed = 0;
2948 		for (i = 0; count > 0 && i < NRUNS; i++) {
2949 			count--;
2950 			m_run = m_runs[RUN_INDEX(count)];
2951 			error = vm_page_reclaim_run(req_class, domain, npages,
2952 			    m_run, high);
2953 			if (error == 0) {
2954 				reclaimed += npages;
2955 				if (reclaimed >= MIN_RECLAIM)
2956 					return (true);
2957 			}
2958 		}
2959 
2960 		/*
2961 		 * Either relax the restrictions on the next scan or return if
2962 		 * the last scan had no restrictions.
2963 		 */
2964 		if (options == VPSC_NORESERV)
2965 			options = VPSC_NOSUPER;
2966 		else if (options == VPSC_NOSUPER)
2967 			options = VPSC_ANY;
2968 		else if (options == VPSC_ANY)
2969 			return (reclaimed != 0);
2970 	}
2971 }
2972 
2973 bool
2974 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2975     u_long alignment, vm_paddr_t boundary)
2976 {
2977 	struct vm_domainset_iter di;
2978 	int domain;
2979 	bool ret;
2980 
2981 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2982 	do {
2983 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2984 		    high, alignment, boundary);
2985 		if (ret)
2986 			break;
2987 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2988 
2989 	return (ret);
2990 }
2991 
2992 /*
2993  * Set the domain in the appropriate page level domainset.
2994  */
2995 void
2996 vm_domain_set(struct vm_domain *vmd)
2997 {
2998 
2999 	mtx_lock(&vm_domainset_lock);
3000 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3001 		vmd->vmd_minset = 1;
3002 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3003 	}
3004 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3005 		vmd->vmd_severeset = 1;
3006 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3007 	}
3008 	mtx_unlock(&vm_domainset_lock);
3009 }
3010 
3011 /*
3012  * Clear the domain from the appropriate page level domainset.
3013  */
3014 void
3015 vm_domain_clear(struct vm_domain *vmd)
3016 {
3017 
3018 	mtx_lock(&vm_domainset_lock);
3019 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3020 		vmd->vmd_minset = 0;
3021 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3022 		if (vm_min_waiters != 0) {
3023 			vm_min_waiters = 0;
3024 			wakeup(&vm_min_domains);
3025 		}
3026 	}
3027 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3028 		vmd->vmd_severeset = 0;
3029 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3030 		if (vm_severe_waiters != 0) {
3031 			vm_severe_waiters = 0;
3032 			wakeup(&vm_severe_domains);
3033 		}
3034 	}
3035 
3036 	/*
3037 	 * If pageout daemon needs pages, then tell it that there are
3038 	 * some free.
3039 	 */
3040 	if (vmd->vmd_pageout_pages_needed &&
3041 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3042 		wakeup(&vmd->vmd_pageout_pages_needed);
3043 		vmd->vmd_pageout_pages_needed = 0;
3044 	}
3045 
3046 	/* See comments in vm_wait_doms(). */
3047 	if (vm_pageproc_waiters) {
3048 		vm_pageproc_waiters = 0;
3049 		wakeup(&vm_pageproc_waiters);
3050 	}
3051 	mtx_unlock(&vm_domainset_lock);
3052 }
3053 
3054 /*
3055  * Wait for free pages to exceed the min threshold globally.
3056  */
3057 void
3058 vm_wait_min(void)
3059 {
3060 
3061 	mtx_lock(&vm_domainset_lock);
3062 	while (vm_page_count_min()) {
3063 		vm_min_waiters++;
3064 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3065 	}
3066 	mtx_unlock(&vm_domainset_lock);
3067 }
3068 
3069 /*
3070  * Wait for free pages to exceed the severe threshold globally.
3071  */
3072 void
3073 vm_wait_severe(void)
3074 {
3075 
3076 	mtx_lock(&vm_domainset_lock);
3077 	while (vm_page_count_severe()) {
3078 		vm_severe_waiters++;
3079 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3080 		    "vmwait", 0);
3081 	}
3082 	mtx_unlock(&vm_domainset_lock);
3083 }
3084 
3085 u_int
3086 vm_wait_count(void)
3087 {
3088 
3089 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3090 }
3091 
3092 void
3093 vm_wait_doms(const domainset_t *wdoms)
3094 {
3095 
3096 	/*
3097 	 * We use racey wakeup synchronization to avoid expensive global
3098 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3099 	 * To handle this, we only sleep for one tick in this instance.  It
3100 	 * is expected that most allocations for the pageproc will come from
3101 	 * kmem or vm_page_grab* which will use the more specific and
3102 	 * race-free vm_wait_domain().
3103 	 */
3104 	if (curproc == pageproc) {
3105 		mtx_lock(&vm_domainset_lock);
3106 		vm_pageproc_waiters++;
3107 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3108 		    "pageprocwait", 1);
3109 	} else {
3110 		/*
3111 		 * XXX Ideally we would wait only until the allocation could
3112 		 * be satisfied.  This condition can cause new allocators to
3113 		 * consume all freed pages while old allocators wait.
3114 		 */
3115 		mtx_lock(&vm_domainset_lock);
3116 		if (vm_page_count_min_set(wdoms)) {
3117 			vm_min_waiters++;
3118 			msleep(&vm_min_domains, &vm_domainset_lock,
3119 			    PVM | PDROP, "vmwait", 0);
3120 		} else
3121 			mtx_unlock(&vm_domainset_lock);
3122 	}
3123 }
3124 
3125 /*
3126  *	vm_wait_domain:
3127  *
3128  *	Sleep until free pages are available for allocation.
3129  *	- Called in various places after failed memory allocations.
3130  */
3131 void
3132 vm_wait_domain(int domain)
3133 {
3134 	struct vm_domain *vmd;
3135 	domainset_t wdom;
3136 
3137 	vmd = VM_DOMAIN(domain);
3138 	vm_domain_free_assert_unlocked(vmd);
3139 
3140 	if (curproc == pageproc) {
3141 		mtx_lock(&vm_domainset_lock);
3142 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3143 			vmd->vmd_pageout_pages_needed = 1;
3144 			msleep(&vmd->vmd_pageout_pages_needed,
3145 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3146 		} else
3147 			mtx_unlock(&vm_domainset_lock);
3148 	} else {
3149 		if (pageproc == NULL)
3150 			panic("vm_wait in early boot");
3151 		DOMAINSET_ZERO(&wdom);
3152 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3153 		vm_wait_doms(&wdom);
3154 	}
3155 }
3156 
3157 /*
3158  *	vm_wait:
3159  *
3160  *	Sleep until free pages are available for allocation in the
3161  *	affinity domains of the obj.  If obj is NULL, the domain set
3162  *	for the calling thread is used.
3163  *	Called in various places after failed memory allocations.
3164  */
3165 void
3166 vm_wait(vm_object_t obj)
3167 {
3168 	struct domainset *d;
3169 
3170 	d = NULL;
3171 
3172 	/*
3173 	 * Carefully fetch pointers only once: the struct domainset
3174 	 * itself is ummutable but the pointer might change.
3175 	 */
3176 	if (obj != NULL)
3177 		d = obj->domain.dr_policy;
3178 	if (d == NULL)
3179 		d = curthread->td_domain.dr_policy;
3180 
3181 	vm_wait_doms(&d->ds_mask);
3182 }
3183 
3184 /*
3185  *	vm_domain_alloc_fail:
3186  *
3187  *	Called when a page allocation function fails.  Informs the
3188  *	pagedaemon and performs the requested wait.  Requires the
3189  *	domain_free and object lock on entry.  Returns with the
3190  *	object lock held and free lock released.  Returns an error when
3191  *	retry is necessary.
3192  *
3193  */
3194 static int
3195 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3196 {
3197 
3198 	vm_domain_free_assert_unlocked(vmd);
3199 
3200 	atomic_add_int(&vmd->vmd_pageout_deficit,
3201 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3202 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3203 		if (object != NULL)
3204 			VM_OBJECT_WUNLOCK(object);
3205 		vm_wait_domain(vmd->vmd_domain);
3206 		if (object != NULL)
3207 			VM_OBJECT_WLOCK(object);
3208 		if (req & VM_ALLOC_WAITOK)
3209 			return (EAGAIN);
3210 	}
3211 
3212 	return (0);
3213 }
3214 
3215 /*
3216  *	vm_waitpfault:
3217  *
3218  *	Sleep until free pages are available for allocation.
3219  *	- Called only in vm_fault so that processes page faulting
3220  *	  can be easily tracked.
3221  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3222  *	  processes will be able to grab memory first.  Do not change
3223  *	  this balance without careful testing first.
3224  */
3225 void
3226 vm_waitpfault(struct domainset *dset, int timo)
3227 {
3228 
3229 	/*
3230 	 * XXX Ideally we would wait only until the allocation could
3231 	 * be satisfied.  This condition can cause new allocators to
3232 	 * consume all freed pages while old allocators wait.
3233 	 */
3234 	mtx_lock(&vm_domainset_lock);
3235 	if (vm_page_count_min_set(&dset->ds_mask)) {
3236 		vm_min_waiters++;
3237 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3238 		    "pfault", timo);
3239 	} else
3240 		mtx_unlock(&vm_domainset_lock);
3241 }
3242 
3243 static struct vm_pagequeue *
3244 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3245 {
3246 
3247 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3248 }
3249 
3250 #ifdef INVARIANTS
3251 static struct vm_pagequeue *
3252 vm_page_pagequeue(vm_page_t m)
3253 {
3254 
3255 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3256 }
3257 #endif
3258 
3259 static __always_inline bool
3260 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3261 {
3262 	vm_page_astate_t tmp;
3263 
3264 	tmp = *old;
3265 	do {
3266 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3267 			return (true);
3268 		counter_u64_add(pqstate_commit_retries, 1);
3269 	} while (old->_bits == tmp._bits);
3270 
3271 	return (false);
3272 }
3273 
3274 /*
3275  * Do the work of committing a queue state update that moves the page out of
3276  * its current queue.
3277  */
3278 static bool
3279 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3280     vm_page_astate_t *old, vm_page_astate_t new)
3281 {
3282 	vm_page_t next;
3283 
3284 	vm_pagequeue_assert_locked(pq);
3285 	KASSERT(vm_page_pagequeue(m) == pq,
3286 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3287 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3288 	    ("%s: invalid queue indices %d %d",
3289 	    __func__, old->queue, new.queue));
3290 
3291 	/*
3292 	 * Once the queue index of the page changes there is nothing
3293 	 * synchronizing with further updates to the page's physical
3294 	 * queue state.  Therefore we must speculatively remove the page
3295 	 * from the queue now and be prepared to roll back if the queue
3296 	 * state update fails.  If the page is not physically enqueued then
3297 	 * we just update its queue index.
3298 	 */
3299 	if ((old->flags & PGA_ENQUEUED) != 0) {
3300 		new.flags &= ~PGA_ENQUEUED;
3301 		next = TAILQ_NEXT(m, plinks.q);
3302 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3303 		vm_pagequeue_cnt_dec(pq);
3304 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3305 			if (next == NULL)
3306 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3307 			else
3308 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3309 			vm_pagequeue_cnt_inc(pq);
3310 			return (false);
3311 		} else {
3312 			return (true);
3313 		}
3314 	} else {
3315 		return (vm_page_pqstate_fcmpset(m, old, new));
3316 	}
3317 }
3318 
3319 static bool
3320 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3321     vm_page_astate_t new)
3322 {
3323 	struct vm_pagequeue *pq;
3324 	vm_page_astate_t as;
3325 	bool ret;
3326 
3327 	pq = _vm_page_pagequeue(m, old->queue);
3328 
3329 	/*
3330 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3331 	 * corresponding page queue lock is held.
3332 	 */
3333 	vm_pagequeue_lock(pq);
3334 	as = vm_page_astate_load(m);
3335 	if (__predict_false(as._bits != old->_bits)) {
3336 		*old = as;
3337 		ret = false;
3338 	} else {
3339 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3340 	}
3341 	vm_pagequeue_unlock(pq);
3342 	return (ret);
3343 }
3344 
3345 /*
3346  * Commit a queue state update that enqueues or requeues a page.
3347  */
3348 static bool
3349 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3350     vm_page_astate_t *old, vm_page_astate_t new)
3351 {
3352 	struct vm_domain *vmd;
3353 
3354 	vm_pagequeue_assert_locked(pq);
3355 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3356 	    ("%s: invalid queue indices %d %d",
3357 	    __func__, old->queue, new.queue));
3358 
3359 	new.flags |= PGA_ENQUEUED;
3360 	if (!vm_page_pqstate_fcmpset(m, old, new))
3361 		return (false);
3362 
3363 	if ((old->flags & PGA_ENQUEUED) != 0)
3364 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3365 	else
3366 		vm_pagequeue_cnt_inc(pq);
3367 
3368 	/*
3369 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3370 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3371 	 * applied, even if it was set first.
3372 	 */
3373 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3374 		vmd = vm_pagequeue_domain(m);
3375 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3376 		    ("%s: invalid page queue for page %p", __func__, m));
3377 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3378 	} else {
3379 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3380 	}
3381 	return (true);
3382 }
3383 
3384 /*
3385  * Commit a queue state update that encodes a request for a deferred queue
3386  * operation.
3387  */
3388 static bool
3389 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3390     vm_page_astate_t new)
3391 {
3392 
3393 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3394 	    ("%s: invalid state, queue %d flags %x",
3395 	    __func__, new.queue, new.flags));
3396 
3397 	if (old->_bits != new._bits &&
3398 	    !vm_page_pqstate_fcmpset(m, old, new))
3399 		return (false);
3400 	vm_page_pqbatch_submit(m, new.queue);
3401 	return (true);
3402 }
3403 
3404 /*
3405  * A generic queue state update function.  This handles more cases than the
3406  * specialized functions above.
3407  */
3408 bool
3409 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3410 {
3411 
3412 	if (old->_bits == new._bits)
3413 		return (true);
3414 
3415 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3416 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3417 			return (false);
3418 		if (new.queue != PQ_NONE)
3419 			vm_page_pqbatch_submit(m, new.queue);
3420 	} else {
3421 		if (!vm_page_pqstate_fcmpset(m, old, new))
3422 			return (false);
3423 		if (new.queue != PQ_NONE &&
3424 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3425 			vm_page_pqbatch_submit(m, new.queue);
3426 	}
3427 	return (true);
3428 }
3429 
3430 /*
3431  * Apply deferred queue state updates to a page.
3432  */
3433 static inline void
3434 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3435 {
3436 	vm_page_astate_t new, old;
3437 
3438 	CRITICAL_ASSERT(curthread);
3439 	vm_pagequeue_assert_locked(pq);
3440 	KASSERT(queue < PQ_COUNT,
3441 	    ("%s: invalid queue index %d", __func__, queue));
3442 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3443 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3444 
3445 	for (old = vm_page_astate_load(m);;) {
3446 		if (__predict_false(old.queue != queue ||
3447 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3448 			counter_u64_add(queue_nops, 1);
3449 			break;
3450 		}
3451 		KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3452 		    ("%s: page %p has unexpected queue state", __func__, m));
3453 
3454 		new = old;
3455 		if ((old.flags & PGA_DEQUEUE) != 0) {
3456 			new.flags &= ~PGA_QUEUE_OP_MASK;
3457 			new.queue = PQ_NONE;
3458 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3459 			    m, &old, new))) {
3460 				counter_u64_add(queue_ops, 1);
3461 				break;
3462 			}
3463 		} else {
3464 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3465 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3466 			    m, &old, new))) {
3467 				counter_u64_add(queue_ops, 1);
3468 				break;
3469 			}
3470 		}
3471 	}
3472 }
3473 
3474 static void
3475 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3476     uint8_t queue)
3477 {
3478 	int i;
3479 
3480 	for (i = 0; i < bq->bq_cnt; i++)
3481 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3482 	vm_batchqueue_init(bq);
3483 }
3484 
3485 /*
3486  *	vm_page_pqbatch_submit:		[ internal use only ]
3487  *
3488  *	Enqueue a page in the specified page queue's batched work queue.
3489  *	The caller must have encoded the requested operation in the page
3490  *	structure's a.flags field.
3491  */
3492 void
3493 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3494 {
3495 	struct vm_batchqueue *bq;
3496 	struct vm_pagequeue *pq;
3497 	int domain;
3498 
3499 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3500 	    ("page %p is unmanaged", m));
3501 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3502 
3503 	domain = vm_phys_domain(m);
3504 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3505 
3506 	critical_enter();
3507 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3508 	if (vm_batchqueue_insert(bq, m)) {
3509 		critical_exit();
3510 		return;
3511 	}
3512 	critical_exit();
3513 	vm_pagequeue_lock(pq);
3514 	critical_enter();
3515 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3516 	vm_pqbatch_process(pq, bq, queue);
3517 	vm_pqbatch_process_page(pq, m, queue);
3518 	vm_pagequeue_unlock(pq);
3519 	critical_exit();
3520 }
3521 
3522 /*
3523  *	vm_page_pqbatch_drain:		[ internal use only ]
3524  *
3525  *	Force all per-CPU page queue batch queues to be drained.  This is
3526  *	intended for use in severe memory shortages, to ensure that pages
3527  *	do not remain stuck in the batch queues.
3528  */
3529 void
3530 vm_page_pqbatch_drain(void)
3531 {
3532 	struct thread *td;
3533 	struct vm_domain *vmd;
3534 	struct vm_pagequeue *pq;
3535 	int cpu, domain, queue;
3536 
3537 	td = curthread;
3538 	CPU_FOREACH(cpu) {
3539 		thread_lock(td);
3540 		sched_bind(td, cpu);
3541 		thread_unlock(td);
3542 
3543 		for (domain = 0; domain < vm_ndomains; domain++) {
3544 			vmd = VM_DOMAIN(domain);
3545 			for (queue = 0; queue < PQ_COUNT; queue++) {
3546 				pq = &vmd->vmd_pagequeues[queue];
3547 				vm_pagequeue_lock(pq);
3548 				critical_enter();
3549 				vm_pqbatch_process(pq,
3550 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3551 				critical_exit();
3552 				vm_pagequeue_unlock(pq);
3553 			}
3554 		}
3555 	}
3556 	thread_lock(td);
3557 	sched_unbind(td);
3558 	thread_unlock(td);
3559 }
3560 
3561 /*
3562  *	vm_page_dequeue_deferred:	[ internal use only ]
3563  *
3564  *	Request removal of the given page from its current page
3565  *	queue.  Physical removal from the queue may be deferred
3566  *	indefinitely.
3567  *
3568  *	The page must be locked.
3569  */
3570 void
3571 vm_page_dequeue_deferred(vm_page_t m)
3572 {
3573 	vm_page_astate_t new, old;
3574 
3575 	old = vm_page_astate_load(m);
3576 	do {
3577 		if (old.queue == PQ_NONE) {
3578 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3579 			    ("%s: page %p has unexpected queue state",
3580 			    __func__, m));
3581 			break;
3582 		}
3583 		new = old;
3584 		new.flags |= PGA_DEQUEUE;
3585 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3586 }
3587 
3588 /*
3589  *	vm_page_dequeue:
3590  *
3591  *	Remove the page from whichever page queue it's in, if any, before
3592  *	returning.
3593  */
3594 void
3595 vm_page_dequeue(vm_page_t m)
3596 {
3597 	vm_page_astate_t new, old;
3598 
3599 	old = vm_page_astate_load(m);
3600 	do {
3601 		if (old.queue == PQ_NONE) {
3602 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3603 			    ("%s: page %p has unexpected queue state",
3604 			    __func__, m));
3605 			break;
3606 		}
3607 		new = old;
3608 		new.flags &= ~PGA_QUEUE_OP_MASK;
3609 		new.queue = PQ_NONE;
3610 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3611 
3612 }
3613 
3614 /*
3615  * Schedule the given page for insertion into the specified page queue.
3616  * Physical insertion of the page may be deferred indefinitely.
3617  */
3618 static void
3619 vm_page_enqueue(vm_page_t m, uint8_t queue)
3620 {
3621 
3622 	KASSERT(m->a.queue == PQ_NONE &&
3623 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3624 	    ("%s: page %p is already enqueued", __func__, m));
3625 	KASSERT(m->ref_count > 0,
3626 	    ("%s: page %p does not carry any references", __func__, m));
3627 
3628 	m->a.queue = queue;
3629 	if ((m->a.flags & PGA_REQUEUE) == 0)
3630 		vm_page_aflag_set(m, PGA_REQUEUE);
3631 	vm_page_pqbatch_submit(m, queue);
3632 }
3633 
3634 /*
3635  *	vm_page_free_prep:
3636  *
3637  *	Prepares the given page to be put on the free list,
3638  *	disassociating it from any VM object. The caller may return
3639  *	the page to the free list only if this function returns true.
3640  *
3641  *	The object must be locked.  The page must be locked if it is
3642  *	managed.
3643  */
3644 static bool
3645 vm_page_free_prep(vm_page_t m)
3646 {
3647 
3648 	/*
3649 	 * Synchronize with threads that have dropped a reference to this
3650 	 * page.
3651 	 */
3652 	atomic_thread_fence_acq();
3653 
3654 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3655 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3656 		uint64_t *p;
3657 		int i;
3658 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3659 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3660 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3661 			    m, i, (uintmax_t)*p));
3662 	}
3663 #endif
3664 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3665 		KASSERT(!pmap_page_is_mapped(m),
3666 		    ("vm_page_free_prep: freeing mapped page %p", m));
3667 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3668 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3669 	} else {
3670 		KASSERT(m->a.queue == PQ_NONE,
3671 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3672 	}
3673 	VM_CNT_INC(v_tfree);
3674 
3675 	if (m->object != NULL) {
3676 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3677 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3678 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3679 		    m));
3680 		vm_page_assert_xbusied(m);
3681 
3682 		/*
3683 		 * The object reference can be released without an atomic
3684 		 * operation.
3685 		 */
3686 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3687 		    m->ref_count == VPRC_OBJREF,
3688 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3689 		    m, m->ref_count));
3690 		vm_page_object_remove(m);
3691 		m->object = NULL;
3692 		m->ref_count -= VPRC_OBJREF;
3693 	} else
3694 		vm_page_assert_unbusied(m);
3695 
3696 	vm_page_busy_free(m);
3697 
3698 	/*
3699 	 * If fictitious remove object association and
3700 	 * return.
3701 	 */
3702 	if ((m->flags & PG_FICTITIOUS) != 0) {
3703 		KASSERT(m->ref_count == 1,
3704 		    ("fictitious page %p is referenced", m));
3705 		KASSERT(m->a.queue == PQ_NONE,
3706 		    ("fictitious page %p is queued", m));
3707 		return (false);
3708 	}
3709 
3710 	/*
3711 	 * Pages need not be dequeued before they are returned to the physical
3712 	 * memory allocator, but they must at least be marked for a deferred
3713 	 * dequeue.
3714 	 */
3715 	if ((m->oflags & VPO_UNMANAGED) == 0)
3716 		vm_page_dequeue_deferred(m);
3717 
3718 	m->valid = 0;
3719 	vm_page_undirty(m);
3720 
3721 	if (m->ref_count != 0)
3722 		panic("vm_page_free_prep: page %p has references", m);
3723 
3724 	/*
3725 	 * Restore the default memory attribute to the page.
3726 	 */
3727 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3728 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3729 
3730 #if VM_NRESERVLEVEL > 0
3731 	/*
3732 	 * Determine whether the page belongs to a reservation.  If the page was
3733 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3734 	 * as an optimization, we avoid the check in that case.
3735 	 */
3736 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3737 		return (false);
3738 #endif
3739 
3740 	return (true);
3741 }
3742 
3743 /*
3744  *	vm_page_free_toq:
3745  *
3746  *	Returns the given page to the free list, disassociating it
3747  *	from any VM object.
3748  *
3749  *	The object must be locked.  The page must be locked if it is
3750  *	managed.
3751  */
3752 static void
3753 vm_page_free_toq(vm_page_t m)
3754 {
3755 	struct vm_domain *vmd;
3756 	uma_zone_t zone;
3757 
3758 	if (!vm_page_free_prep(m))
3759 		return;
3760 
3761 	vmd = vm_pagequeue_domain(m);
3762 	zone = vmd->vmd_pgcache[m->pool].zone;
3763 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3764 		uma_zfree(zone, m);
3765 		return;
3766 	}
3767 	vm_domain_free_lock(vmd);
3768 	vm_phys_free_pages(m, 0);
3769 	vm_domain_free_unlock(vmd);
3770 	vm_domain_freecnt_inc(vmd, 1);
3771 }
3772 
3773 /*
3774  *	vm_page_free_pages_toq:
3775  *
3776  *	Returns a list of pages to the free list, disassociating it
3777  *	from any VM object.  In other words, this is equivalent to
3778  *	calling vm_page_free_toq() for each page of a list of VM objects.
3779  *
3780  *	The objects must be locked.  The pages must be locked if it is
3781  *	managed.
3782  */
3783 void
3784 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3785 {
3786 	vm_page_t m;
3787 	int count;
3788 
3789 	if (SLIST_EMPTY(free))
3790 		return;
3791 
3792 	count = 0;
3793 	while ((m = SLIST_FIRST(free)) != NULL) {
3794 		count++;
3795 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3796 		vm_page_free_toq(m);
3797 	}
3798 
3799 	if (update_wire_count)
3800 		vm_wire_sub(count);
3801 }
3802 
3803 /*
3804  * Mark this page as wired down, preventing reclamation by the page daemon
3805  * or when the containing object is destroyed.
3806  */
3807 void
3808 vm_page_wire(vm_page_t m)
3809 {
3810 	u_int old;
3811 
3812 	KASSERT(m->object != NULL,
3813 	    ("vm_page_wire: page %p does not belong to an object", m));
3814 	if (!vm_page_busied(m) && !vm_object_busied(m->object))
3815 		VM_OBJECT_ASSERT_LOCKED(m->object);
3816 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3817 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3818 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3819 
3820 	old = atomic_fetchadd_int(&m->ref_count, 1);
3821 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3822 	    ("vm_page_wire: counter overflow for page %p", m));
3823 	if (VPRC_WIRE_COUNT(old) == 0) {
3824 		if ((m->oflags & VPO_UNMANAGED) == 0)
3825 			vm_page_aflag_set(m, PGA_DEQUEUE);
3826 		vm_wire_add(1);
3827 	}
3828 }
3829 
3830 /*
3831  * Attempt to wire a mapped page following a pmap lookup of that page.
3832  * This may fail if a thread is concurrently tearing down mappings of the page.
3833  * The transient failure is acceptable because it translates to the
3834  * failure of the caller pmap_extract_and_hold(), which should be then
3835  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3836  */
3837 bool
3838 vm_page_wire_mapped(vm_page_t m)
3839 {
3840 	u_int old;
3841 
3842 	old = m->ref_count;
3843 	do {
3844 		KASSERT(old > 0,
3845 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3846 		if ((old & VPRC_BLOCKED) != 0)
3847 			return (false);
3848 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3849 
3850 	if (VPRC_WIRE_COUNT(old) == 0) {
3851 		if ((m->oflags & VPO_UNMANAGED) == 0)
3852 			vm_page_aflag_set(m, PGA_DEQUEUE);
3853 		vm_wire_add(1);
3854 	}
3855 	return (true);
3856 }
3857 
3858 /*
3859  * Release a wiring reference to a managed page.  If the page still belongs to
3860  * an object, update its position in the page queues to reflect the reference.
3861  * If the wiring was the last reference to the page, free the page.
3862  */
3863 static void
3864 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3865 {
3866 	u_int old;
3867 
3868 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3869 	    ("%s: page %p is unmanaged", __func__, m));
3870 
3871 	/*
3872 	 * Update LRU state before releasing the wiring reference.
3873 	 * Use a release store when updating the reference count to
3874 	 * synchronize with vm_page_free_prep().
3875 	 */
3876 	old = m->ref_count;
3877 	do {
3878 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3879 		    ("vm_page_unwire: wire count underflow for page %p", m));
3880 
3881 		if (old > VPRC_OBJREF + 1) {
3882 			/*
3883 			 * The page has at least one other wiring reference.  An
3884 			 * earlier iteration of this loop may have called
3885 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3886 			 * re-set it if necessary.
3887 			 */
3888 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3889 				vm_page_aflag_set(m, PGA_DEQUEUE);
3890 		} else if (old == VPRC_OBJREF + 1) {
3891 			/*
3892 			 * This is the last wiring.  Clear PGA_DEQUEUE and
3893 			 * update the page's queue state to reflect the
3894 			 * reference.  If the page does not belong to an object
3895 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3896 			 * clear leftover queue state.
3897 			 */
3898 			vm_page_release_toq(m, nqueue, false);
3899 		} else if (old == 1) {
3900 			vm_page_aflag_clear(m, PGA_DEQUEUE);
3901 		}
3902 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3903 
3904 	if (VPRC_WIRE_COUNT(old) == 1) {
3905 		vm_wire_sub(1);
3906 		if (old == 1)
3907 			vm_page_free(m);
3908 	}
3909 }
3910 
3911 /*
3912  * Release one wiring of the specified page, potentially allowing it to be
3913  * paged out.
3914  *
3915  * Only managed pages belonging to an object can be paged out.  If the number
3916  * of wirings transitions to zero and the page is eligible for page out, then
3917  * the page is added to the specified paging queue.  If the released wiring
3918  * represented the last reference to the page, the page is freed.
3919  *
3920  * A managed page must be locked.
3921  */
3922 void
3923 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3924 {
3925 
3926 	KASSERT(nqueue < PQ_COUNT,
3927 	    ("vm_page_unwire: invalid queue %u request for page %p",
3928 	    nqueue, m));
3929 
3930 	if ((m->oflags & VPO_UNMANAGED) != 0) {
3931 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
3932 			vm_page_free(m);
3933 		return;
3934 	}
3935 	vm_page_unwire_managed(m, nqueue, false);
3936 }
3937 
3938 /*
3939  * Unwire a page without (re-)inserting it into a page queue.  It is up
3940  * to the caller to enqueue, requeue, or free the page as appropriate.
3941  * In most cases involving managed pages, vm_page_unwire() should be used
3942  * instead.
3943  */
3944 bool
3945 vm_page_unwire_noq(vm_page_t m)
3946 {
3947 	u_int old;
3948 
3949 	old = vm_page_drop(m, 1);
3950 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
3951 	    ("vm_page_unref: counter underflow for page %p", m));
3952 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3953 	    ("vm_page_unref: missing ref on fictitious page %p", m));
3954 
3955 	if (VPRC_WIRE_COUNT(old) > 1)
3956 		return (false);
3957 	if ((m->oflags & VPO_UNMANAGED) == 0)
3958 		vm_page_aflag_clear(m, PGA_DEQUEUE);
3959 	vm_wire_sub(1);
3960 	return (true);
3961 }
3962 
3963 /*
3964  * Ensure that the page ends up in the specified page queue.  If the page is
3965  * active or being moved to the active queue, ensure that its act_count is
3966  * at least ACT_INIT but do not otherwise mess with it.
3967  *
3968  * A managed page must be locked.
3969  */
3970 static __always_inline void
3971 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
3972 {
3973 	vm_page_astate_t old, new;
3974 
3975 	KASSERT(m->ref_count > 0,
3976 	    ("%s: page %p does not carry any references", __func__, m));
3977 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
3978 	    ("%s: invalid flags %x", __func__, nflag));
3979 
3980 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3981 		return;
3982 
3983 	old = vm_page_astate_load(m);
3984 	do {
3985 		if ((old.flags & PGA_DEQUEUE) != 0)
3986 			break;
3987 		new = old;
3988 		new.flags &= ~PGA_QUEUE_OP_MASK;
3989 		if (nqueue == PQ_ACTIVE)
3990 			new.act_count = max(old.act_count, ACT_INIT);
3991 		if (old.queue == nqueue) {
3992 			if (nqueue != PQ_ACTIVE)
3993 				new.flags |= nflag;
3994 		} else {
3995 			new.flags |= nflag;
3996 			new.queue = nqueue;
3997 		}
3998 	} while (!vm_page_pqstate_commit(m, &old, new));
3999 }
4000 
4001 /*
4002  * Put the specified page on the active list (if appropriate).
4003  */
4004 void
4005 vm_page_activate(vm_page_t m)
4006 {
4007 
4008 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4009 }
4010 
4011 /*
4012  * Move the specified page to the tail of the inactive queue, or requeue
4013  * the page if it is already in the inactive queue.
4014  */
4015 void
4016 vm_page_deactivate(vm_page_t m)
4017 {
4018 
4019 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4020 }
4021 
4022 void
4023 vm_page_deactivate_noreuse(vm_page_t m)
4024 {
4025 
4026 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4027 }
4028 
4029 /*
4030  * Put a page in the laundry, or requeue it if it is already there.
4031  */
4032 void
4033 vm_page_launder(vm_page_t m)
4034 {
4035 
4036 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4037 }
4038 
4039 /*
4040  * Put a page in the PQ_UNSWAPPABLE holding queue.
4041  */
4042 void
4043 vm_page_unswappable(vm_page_t m)
4044 {
4045 
4046 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4047 	    ("page %p already unswappable", m));
4048 
4049 	vm_page_dequeue(m);
4050 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4051 }
4052 
4053 /*
4054  * Release a page back to the page queues in preparation for unwiring.
4055  */
4056 static void
4057 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4058 {
4059 	vm_page_astate_t old, new;
4060 	uint16_t nflag;
4061 
4062 	/*
4063 	 * Use a check of the valid bits to determine whether we should
4064 	 * accelerate reclamation of the page.  The object lock might not be
4065 	 * held here, in which case the check is racy.  At worst we will either
4066 	 * accelerate reclamation of a valid page and violate LRU, or
4067 	 * unnecessarily defer reclamation of an invalid page.
4068 	 *
4069 	 * If we were asked to not cache the page, place it near the head of the
4070 	 * inactive queue so that is reclaimed sooner.
4071 	 */
4072 	if (noreuse || m->valid == 0) {
4073 		nqueue = PQ_INACTIVE;
4074 		nflag = PGA_REQUEUE_HEAD;
4075 	} else {
4076 		nflag = PGA_REQUEUE;
4077 	}
4078 
4079 	old = vm_page_astate_load(m);
4080 	do {
4081 		new = old;
4082 
4083 		/*
4084 		 * If the page is already in the active queue and we are not
4085 		 * trying to accelerate reclamation, simply mark it as
4086 		 * referenced and avoid any queue operations.
4087 		 */
4088 		new.flags &= ~PGA_QUEUE_OP_MASK;
4089 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4090 			new.flags |= PGA_REFERENCED;
4091 		else {
4092 			new.flags |= nflag;
4093 			new.queue = nqueue;
4094 		}
4095 	} while (!vm_page_pqstate_commit(m, &old, new));
4096 }
4097 
4098 /*
4099  * Unwire a page and either attempt to free it or re-add it to the page queues.
4100  */
4101 void
4102 vm_page_release(vm_page_t m, int flags)
4103 {
4104 	vm_object_t object;
4105 
4106 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4107 	    ("vm_page_release: page %p is unmanaged", m));
4108 
4109 	if ((flags & VPR_TRYFREE) != 0) {
4110 		for (;;) {
4111 			object = atomic_load_ptr(&m->object);
4112 			if (object == NULL)
4113 				break;
4114 			/* Depends on type-stability. */
4115 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4116 				break;
4117 			if (object == m->object) {
4118 				vm_page_release_locked(m, flags);
4119 				VM_OBJECT_WUNLOCK(object);
4120 				return;
4121 			}
4122 			VM_OBJECT_WUNLOCK(object);
4123 		}
4124 	}
4125 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4126 }
4127 
4128 /* See vm_page_release(). */
4129 void
4130 vm_page_release_locked(vm_page_t m, int flags)
4131 {
4132 
4133 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4134 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4135 	    ("vm_page_release_locked: page %p is unmanaged", m));
4136 
4137 	if (vm_page_unwire_noq(m)) {
4138 		if ((flags & VPR_TRYFREE) != 0 &&
4139 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4140 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4141 			vm_page_free(m);
4142 		} else {
4143 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4144 		}
4145 	}
4146 }
4147 
4148 static bool
4149 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4150 {
4151 	u_int old;
4152 
4153 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4154 	    ("vm_page_try_blocked_op: page %p has no object", m));
4155 	KASSERT(vm_page_busied(m),
4156 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4157 	VM_OBJECT_ASSERT_LOCKED(m->object);
4158 
4159 	old = m->ref_count;
4160 	do {
4161 		KASSERT(old != 0,
4162 		    ("vm_page_try_blocked_op: page %p has no references", m));
4163 		if (VPRC_WIRE_COUNT(old) != 0)
4164 			return (false);
4165 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4166 
4167 	(op)(m);
4168 
4169 	/*
4170 	 * If the object is read-locked, new wirings may be created via an
4171 	 * object lookup.
4172 	 */
4173 	old = vm_page_drop(m, VPRC_BLOCKED);
4174 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4175 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4176 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4177 	    old, m));
4178 	return (true);
4179 }
4180 
4181 /*
4182  * Atomically check for wirings and remove all mappings of the page.
4183  */
4184 bool
4185 vm_page_try_remove_all(vm_page_t m)
4186 {
4187 
4188 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4189 }
4190 
4191 /*
4192  * Atomically check for wirings and remove all writeable mappings of the page.
4193  */
4194 bool
4195 vm_page_try_remove_write(vm_page_t m)
4196 {
4197 
4198 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4199 }
4200 
4201 /*
4202  * vm_page_advise
4203  *
4204  * 	Apply the specified advice to the given page.
4205  *
4206  *	The object and page must be locked.
4207  */
4208 void
4209 vm_page_advise(vm_page_t m, int advice)
4210 {
4211 
4212 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4213 	if (advice == MADV_FREE)
4214 		/*
4215 		 * Mark the page clean.  This will allow the page to be freed
4216 		 * without first paging it out.  MADV_FREE pages are often
4217 		 * quickly reused by malloc(3), so we do not do anything that
4218 		 * would result in a page fault on a later access.
4219 		 */
4220 		vm_page_undirty(m);
4221 	else if (advice != MADV_DONTNEED) {
4222 		if (advice == MADV_WILLNEED)
4223 			vm_page_activate(m);
4224 		return;
4225 	}
4226 
4227 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4228 		vm_page_dirty(m);
4229 
4230 	/*
4231 	 * Clear any references to the page.  Otherwise, the page daemon will
4232 	 * immediately reactivate the page.
4233 	 */
4234 	vm_page_aflag_clear(m, PGA_REFERENCED);
4235 
4236 	/*
4237 	 * Place clean pages near the head of the inactive queue rather than
4238 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4239 	 * the page will be reused quickly.  Dirty pages not already in the
4240 	 * laundry are moved there.
4241 	 */
4242 	if (m->dirty == 0)
4243 		vm_page_deactivate_noreuse(m);
4244 	else if (!vm_page_in_laundry(m))
4245 		vm_page_launder(m);
4246 }
4247 
4248 static inline int
4249 vm_page_grab_pflags(int allocflags)
4250 {
4251 	int pflags;
4252 
4253 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4254 	    (allocflags & VM_ALLOC_WIRED) != 0,
4255 	    ("vm_page_grab_pflags: the pages must be busied or wired"));
4256 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4257 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4258 	    ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4259 	    "mismatch"));
4260 	pflags = allocflags &
4261 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4262 	    VM_ALLOC_NOBUSY);
4263 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4264 		pflags |= VM_ALLOC_WAITFAIL;
4265 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4266 		pflags |= VM_ALLOC_SBUSY;
4267 
4268 	return (pflags);
4269 }
4270 
4271 /*
4272  * Grab a page, waiting until we are waken up due to the page
4273  * changing state.  We keep on waiting, if the page continues
4274  * to be in the object.  If the page doesn't exist, first allocate it
4275  * and then conditionally zero it.
4276  *
4277  * This routine may sleep.
4278  *
4279  * The object must be locked on entry.  The lock will, however, be released
4280  * and reacquired if the routine sleeps.
4281  */
4282 vm_page_t
4283 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4284 {
4285 	vm_page_t m;
4286 	int pflags;
4287 
4288 	VM_OBJECT_ASSERT_WLOCKED(object);
4289 	pflags = vm_page_grab_pflags(allocflags);
4290 retrylookup:
4291 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4292 		if (!vm_page_acquire_flags(m, allocflags)) {
4293 			if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4294 			    allocflags))
4295 				goto retrylookup;
4296 			return (NULL);
4297 		}
4298 		goto out;
4299 	}
4300 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4301 		return (NULL);
4302 	m = vm_page_alloc(object, pindex, pflags);
4303 	if (m == NULL) {
4304 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4305 			return (NULL);
4306 		goto retrylookup;
4307 	}
4308 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4309 		pmap_zero_page(m);
4310 
4311 out:
4312 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4313 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4314 			vm_page_sunbusy(m);
4315 		else
4316 			vm_page_xunbusy(m);
4317 	}
4318 	return (m);
4319 }
4320 
4321 /*
4322  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4323  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4324  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4325  * in simultaneously.  Additional pages will be left on a paging queue but
4326  * will neither be wired nor busy regardless of allocflags.
4327  */
4328 int
4329 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4330 {
4331 	vm_page_t m;
4332 	vm_page_t ma[VM_INITIAL_PAGEIN];
4333 	bool sleep, xbusy;
4334 	int after, i, pflags, rv;
4335 
4336 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4337 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4338 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4339 	KASSERT((allocflags &
4340 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4341 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4342 	VM_OBJECT_ASSERT_WLOCKED(object);
4343 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4344 	pflags |= VM_ALLOC_WAITFAIL;
4345 
4346 retrylookup:
4347 	xbusy = false;
4348 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4349 		/*
4350 		 * If the page is fully valid it can only become invalid
4351 		 * with the object lock held.  If it is not valid it can
4352 		 * become valid with the busy lock held.  Therefore, we
4353 		 * may unnecessarily lock the exclusive busy here if we
4354 		 * race with I/O completion not using the object lock.
4355 		 * However, we will not end up with an invalid page and a
4356 		 * shared lock.
4357 		 */
4358 		if (!vm_page_all_valid(m) ||
4359 		    (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4360 			sleep = !vm_page_tryxbusy(m);
4361 			xbusy = true;
4362 		} else
4363 			sleep = !vm_page_trysbusy(m);
4364 		if (sleep) {
4365 			(void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4366 			    allocflags);
4367 			goto retrylookup;
4368 		}
4369 		if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4370 		   !vm_page_all_valid(m)) {
4371 			if (xbusy)
4372 				vm_page_xunbusy(m);
4373 			else
4374 				vm_page_sunbusy(m);
4375 			*mp = NULL;
4376 			return (VM_PAGER_FAIL);
4377 		}
4378 		if ((allocflags & VM_ALLOC_WIRED) != 0)
4379 			vm_page_wire(m);
4380 		if (vm_page_all_valid(m))
4381 			goto out;
4382 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4383 		*mp = NULL;
4384 		return (VM_PAGER_FAIL);
4385 	} else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4386 		xbusy = true;
4387 	} else {
4388 		goto retrylookup;
4389 	}
4390 
4391 	vm_page_assert_xbusied(m);
4392 	MPASS(xbusy);
4393 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4394 		after = MIN(after, VM_INITIAL_PAGEIN);
4395 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4396 		after = MAX(after, 1);
4397 		ma[0] = m;
4398 		for (i = 1; i < after; i++) {
4399 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4400 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4401 					break;
4402 			} else {
4403 				ma[i] = vm_page_alloc(object, m->pindex + i,
4404 				    VM_ALLOC_NORMAL);
4405 				if (ma[i] == NULL)
4406 					break;
4407 			}
4408 		}
4409 		after = i;
4410 		vm_object_pip_add(object, after);
4411 		VM_OBJECT_WUNLOCK(object);
4412 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4413 		VM_OBJECT_WLOCK(object);
4414 		vm_object_pip_wakeupn(object, after);
4415 		/* Pager may have replaced a page. */
4416 		m = ma[0];
4417 		if (rv != VM_PAGER_OK) {
4418 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4419 				vm_page_unwire_noq(m);
4420 			for (i = 0; i < after; i++) {
4421 				if (!vm_page_wired(ma[i]))
4422 					vm_page_free(ma[i]);
4423 				else
4424 					vm_page_xunbusy(ma[i]);
4425 			}
4426 			*mp = NULL;
4427 			return (rv);
4428 		}
4429 		for (i = 1; i < after; i++)
4430 			vm_page_readahead_finish(ma[i]);
4431 		MPASS(vm_page_all_valid(m));
4432 	} else {
4433 		vm_page_zero_invalid(m, TRUE);
4434 	}
4435 out:
4436 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4437 		if (xbusy)
4438 			vm_page_xunbusy(m);
4439 		else
4440 			vm_page_sunbusy(m);
4441 	}
4442 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4443 		vm_page_busy_downgrade(m);
4444 	*mp = m;
4445 	return (VM_PAGER_OK);
4446 }
4447 
4448 /*
4449  * Return the specified range of pages from the given object.  For each
4450  * page offset within the range, if a page already exists within the object
4451  * at that offset and it is busy, then wait for it to change state.  If,
4452  * instead, the page doesn't exist, then allocate it.
4453  *
4454  * The caller must always specify an allocation class.
4455  *
4456  * allocation classes:
4457  *	VM_ALLOC_NORMAL		normal process request
4458  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4459  *
4460  * The caller must always specify that the pages are to be busied and/or
4461  * wired.
4462  *
4463  * optional allocation flags:
4464  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4465  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4466  *	VM_ALLOC_NOWAIT		do not sleep
4467  *	VM_ALLOC_SBUSY		set page to sbusy state
4468  *	VM_ALLOC_WIRED		wire the pages
4469  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4470  *
4471  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4472  * may return a partial prefix of the requested range.
4473  */
4474 int
4475 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4476     vm_page_t *ma, int count)
4477 {
4478 	vm_page_t m, mpred;
4479 	int pflags;
4480 	int i;
4481 
4482 	VM_OBJECT_ASSERT_WLOCKED(object);
4483 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4484 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4485 
4486 	pflags = vm_page_grab_pflags(allocflags);
4487 	if (count == 0)
4488 		return (0);
4489 
4490 	i = 0;
4491 retrylookup:
4492 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4493 	if (m == NULL || m->pindex != pindex + i) {
4494 		mpred = m;
4495 		m = NULL;
4496 	} else
4497 		mpred = TAILQ_PREV(m, pglist, listq);
4498 	for (; i < count; i++) {
4499 		if (m != NULL) {
4500 			if (!vm_page_acquire_flags(m, allocflags)) {
4501 				if (vm_page_busy_sleep_flags(object, m,
4502 				    "grbmaw", allocflags))
4503 					goto retrylookup;
4504 				break;
4505 			}
4506 		} else {
4507 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4508 				break;
4509 			m = vm_page_alloc_after(object, pindex + i,
4510 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4511 			if (m == NULL) {
4512 				if ((allocflags & (VM_ALLOC_NOWAIT |
4513 				    VM_ALLOC_WAITFAIL)) != 0)
4514 					break;
4515 				goto retrylookup;
4516 			}
4517 		}
4518 		if (vm_page_none_valid(m) &&
4519 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4520 			if ((m->flags & PG_ZERO) == 0)
4521 				pmap_zero_page(m);
4522 			vm_page_valid(m);
4523 		}
4524 		if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4525 			if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4526 				vm_page_sunbusy(m);
4527 			else
4528 				vm_page_xunbusy(m);
4529 		}
4530 		ma[i] = mpred = m;
4531 		m = vm_page_next(m);
4532 	}
4533 	return (i);
4534 }
4535 
4536 /*
4537  * Mapping function for valid or dirty bits in a page.
4538  *
4539  * Inputs are required to range within a page.
4540  */
4541 vm_page_bits_t
4542 vm_page_bits(int base, int size)
4543 {
4544 	int first_bit;
4545 	int last_bit;
4546 
4547 	KASSERT(
4548 	    base + size <= PAGE_SIZE,
4549 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4550 	);
4551 
4552 	if (size == 0)		/* handle degenerate case */
4553 		return (0);
4554 
4555 	first_bit = base >> DEV_BSHIFT;
4556 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4557 
4558 	return (((vm_page_bits_t)2 << last_bit) -
4559 	    ((vm_page_bits_t)1 << first_bit));
4560 }
4561 
4562 void
4563 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4564 {
4565 
4566 #if PAGE_SIZE == 32768
4567 	atomic_set_64((uint64_t *)bits, set);
4568 #elif PAGE_SIZE == 16384
4569 	atomic_set_32((uint32_t *)bits, set);
4570 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4571 	atomic_set_16((uint16_t *)bits, set);
4572 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4573 	atomic_set_8((uint8_t *)bits, set);
4574 #else		/* PAGE_SIZE <= 8192 */
4575 	uintptr_t addr;
4576 	int shift;
4577 
4578 	addr = (uintptr_t)bits;
4579 	/*
4580 	 * Use a trick to perform a 32-bit atomic on the
4581 	 * containing aligned word, to not depend on the existence
4582 	 * of atomic_{set, clear}_{8, 16}.
4583 	 */
4584 	shift = addr & (sizeof(uint32_t) - 1);
4585 #if BYTE_ORDER == BIG_ENDIAN
4586 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4587 #else
4588 	shift *= NBBY;
4589 #endif
4590 	addr &= ~(sizeof(uint32_t) - 1);
4591 	atomic_set_32((uint32_t *)addr, set << shift);
4592 #endif		/* PAGE_SIZE */
4593 }
4594 
4595 static inline void
4596 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4597 {
4598 
4599 #if PAGE_SIZE == 32768
4600 	atomic_clear_64((uint64_t *)bits, clear);
4601 #elif PAGE_SIZE == 16384
4602 	atomic_clear_32((uint32_t *)bits, clear);
4603 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4604 	atomic_clear_16((uint16_t *)bits, clear);
4605 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4606 	atomic_clear_8((uint8_t *)bits, clear);
4607 #else		/* PAGE_SIZE <= 8192 */
4608 	uintptr_t addr;
4609 	int shift;
4610 
4611 	addr = (uintptr_t)bits;
4612 	/*
4613 	 * Use a trick to perform a 32-bit atomic on the
4614 	 * containing aligned word, to not depend on the existence
4615 	 * of atomic_{set, clear}_{8, 16}.
4616 	 */
4617 	shift = addr & (sizeof(uint32_t) - 1);
4618 #if BYTE_ORDER == BIG_ENDIAN
4619 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4620 #else
4621 	shift *= NBBY;
4622 #endif
4623 	addr &= ~(sizeof(uint32_t) - 1);
4624 	atomic_clear_32((uint32_t *)addr, clear << shift);
4625 #endif		/* PAGE_SIZE */
4626 }
4627 
4628 static inline vm_page_bits_t
4629 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4630 {
4631 #if PAGE_SIZE == 32768
4632 	uint64_t old;
4633 
4634 	old = *bits;
4635 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4636 	return (old);
4637 #elif PAGE_SIZE == 16384
4638 	uint32_t old;
4639 
4640 	old = *bits;
4641 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4642 	return (old);
4643 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4644 	uint16_t old;
4645 
4646 	old = *bits;
4647 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4648 	return (old);
4649 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4650 	uint8_t old;
4651 
4652 	old = *bits;
4653 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4654 	return (old);
4655 #else		/* PAGE_SIZE <= 4096*/
4656 	uintptr_t addr;
4657 	uint32_t old, new, mask;
4658 	int shift;
4659 
4660 	addr = (uintptr_t)bits;
4661 	/*
4662 	 * Use a trick to perform a 32-bit atomic on the
4663 	 * containing aligned word, to not depend on the existence
4664 	 * of atomic_{set, swap, clear}_{8, 16}.
4665 	 */
4666 	shift = addr & (sizeof(uint32_t) - 1);
4667 #if BYTE_ORDER == BIG_ENDIAN
4668 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4669 #else
4670 	shift *= NBBY;
4671 #endif
4672 	addr &= ~(sizeof(uint32_t) - 1);
4673 	mask = VM_PAGE_BITS_ALL << shift;
4674 
4675 	old = *bits;
4676 	do {
4677 		new = old & ~mask;
4678 		new |= newbits << shift;
4679 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4680 	return (old >> shift);
4681 #endif		/* PAGE_SIZE */
4682 }
4683 
4684 /*
4685  *	vm_page_set_valid_range:
4686  *
4687  *	Sets portions of a page valid.  The arguments are expected
4688  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4689  *	of any partial chunks touched by the range.  The invalid portion of
4690  *	such chunks will be zeroed.
4691  *
4692  *	(base + size) must be less then or equal to PAGE_SIZE.
4693  */
4694 void
4695 vm_page_set_valid_range(vm_page_t m, int base, int size)
4696 {
4697 	int endoff, frag;
4698 	vm_page_bits_t pagebits;
4699 
4700 	vm_page_assert_busied(m);
4701 	if (size == 0)	/* handle degenerate case */
4702 		return;
4703 
4704 	/*
4705 	 * If the base is not DEV_BSIZE aligned and the valid
4706 	 * bit is clear, we have to zero out a portion of the
4707 	 * first block.
4708 	 */
4709 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4710 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4711 		pmap_zero_page_area(m, frag, base - frag);
4712 
4713 	/*
4714 	 * If the ending offset is not DEV_BSIZE aligned and the
4715 	 * valid bit is clear, we have to zero out a portion of
4716 	 * the last block.
4717 	 */
4718 	endoff = base + size;
4719 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4720 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4721 		pmap_zero_page_area(m, endoff,
4722 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4723 
4724 	/*
4725 	 * Assert that no previously invalid block that is now being validated
4726 	 * is already dirty.
4727 	 */
4728 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4729 	    ("vm_page_set_valid_range: page %p is dirty", m));
4730 
4731 	/*
4732 	 * Set valid bits inclusive of any overlap.
4733 	 */
4734 	pagebits = vm_page_bits(base, size);
4735 	if (vm_page_xbusied(m))
4736 		m->valid |= pagebits;
4737 	else
4738 		vm_page_bits_set(m, &m->valid, pagebits);
4739 }
4740 
4741 /*
4742  * Set the page dirty bits and free the invalid swap space if
4743  * present.  Returns the previous dirty bits.
4744  */
4745 vm_page_bits_t
4746 vm_page_set_dirty(vm_page_t m)
4747 {
4748 	vm_page_bits_t old;
4749 
4750 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
4751 
4752 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4753 		old = m->dirty;
4754 		m->dirty = VM_PAGE_BITS_ALL;
4755 	} else
4756 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4757 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4758 		vm_pager_page_unswapped(m);
4759 
4760 	return (old);
4761 }
4762 
4763 /*
4764  * Clear the given bits from the specified page's dirty field.
4765  */
4766 static __inline void
4767 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4768 {
4769 
4770 	vm_page_assert_busied(m);
4771 
4772 	/*
4773 	 * If the page is xbusied and not write mapped we are the
4774 	 * only thread that can modify dirty bits.  Otherwise, The pmap
4775 	 * layer can call vm_page_dirty() without holding a distinguished
4776 	 * lock.  The combination of page busy and atomic operations
4777 	 * suffice to guarantee consistency of the page dirty field.
4778 	 */
4779 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4780 		m->dirty &= ~pagebits;
4781 	else
4782 		vm_page_bits_clear(m, &m->dirty, pagebits);
4783 }
4784 
4785 /*
4786  *	vm_page_set_validclean:
4787  *
4788  *	Sets portions of a page valid and clean.  The arguments are expected
4789  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4790  *	of any partial chunks touched by the range.  The invalid portion of
4791  *	such chunks will be zero'd.
4792  *
4793  *	(base + size) must be less then or equal to PAGE_SIZE.
4794  */
4795 void
4796 vm_page_set_validclean(vm_page_t m, int base, int size)
4797 {
4798 	vm_page_bits_t oldvalid, pagebits;
4799 	int endoff, frag;
4800 
4801 	vm_page_assert_busied(m);
4802 	if (size == 0)	/* handle degenerate case */
4803 		return;
4804 
4805 	/*
4806 	 * If the base is not DEV_BSIZE aligned and the valid
4807 	 * bit is clear, we have to zero out a portion of the
4808 	 * first block.
4809 	 */
4810 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4811 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4812 		pmap_zero_page_area(m, frag, base - frag);
4813 
4814 	/*
4815 	 * If the ending offset is not DEV_BSIZE aligned and the
4816 	 * valid bit is clear, we have to zero out a portion of
4817 	 * the last block.
4818 	 */
4819 	endoff = base + size;
4820 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4821 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4822 		pmap_zero_page_area(m, endoff,
4823 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4824 
4825 	/*
4826 	 * Set valid, clear dirty bits.  If validating the entire
4827 	 * page we can safely clear the pmap modify bit.  We also
4828 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
4829 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4830 	 * be set again.
4831 	 *
4832 	 * We set valid bits inclusive of any overlap, but we can only
4833 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4834 	 * the range.
4835 	 */
4836 	oldvalid = m->valid;
4837 	pagebits = vm_page_bits(base, size);
4838 	if (vm_page_xbusied(m))
4839 		m->valid |= pagebits;
4840 	else
4841 		vm_page_bits_set(m, &m->valid, pagebits);
4842 #if 0	/* NOT YET */
4843 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4844 		frag = DEV_BSIZE - frag;
4845 		base += frag;
4846 		size -= frag;
4847 		if (size < 0)
4848 			size = 0;
4849 	}
4850 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4851 #endif
4852 	if (base == 0 && size == PAGE_SIZE) {
4853 		/*
4854 		 * The page can only be modified within the pmap if it is
4855 		 * mapped, and it can only be mapped if it was previously
4856 		 * fully valid.
4857 		 */
4858 		if (oldvalid == VM_PAGE_BITS_ALL)
4859 			/*
4860 			 * Perform the pmap_clear_modify() first.  Otherwise,
4861 			 * a concurrent pmap operation, such as
4862 			 * pmap_protect(), could clear a modification in the
4863 			 * pmap and set the dirty field on the page before
4864 			 * pmap_clear_modify() had begun and after the dirty
4865 			 * field was cleared here.
4866 			 */
4867 			pmap_clear_modify(m);
4868 		m->dirty = 0;
4869 		vm_page_aflag_clear(m, PGA_NOSYNC);
4870 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4871 		m->dirty &= ~pagebits;
4872 	else
4873 		vm_page_clear_dirty_mask(m, pagebits);
4874 }
4875 
4876 void
4877 vm_page_clear_dirty(vm_page_t m, int base, int size)
4878 {
4879 
4880 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4881 }
4882 
4883 /*
4884  *	vm_page_set_invalid:
4885  *
4886  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4887  *	valid and dirty bits for the effected areas are cleared.
4888  */
4889 void
4890 vm_page_set_invalid(vm_page_t m, int base, int size)
4891 {
4892 	vm_page_bits_t bits;
4893 	vm_object_t object;
4894 
4895 	/*
4896 	 * The object lock is required so that pages can't be mapped
4897 	 * read-only while we're in the process of invalidating them.
4898 	 */
4899 	object = m->object;
4900 	VM_OBJECT_ASSERT_WLOCKED(object);
4901 	vm_page_assert_busied(m);
4902 
4903 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4904 	    size >= object->un_pager.vnp.vnp_size)
4905 		bits = VM_PAGE_BITS_ALL;
4906 	else
4907 		bits = vm_page_bits(base, size);
4908 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4909 		pmap_remove_all(m);
4910 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4911 	    !pmap_page_is_mapped(m),
4912 	    ("vm_page_set_invalid: page %p is mapped", m));
4913 	if (vm_page_xbusied(m)) {
4914 		m->valid &= ~bits;
4915 		m->dirty &= ~bits;
4916 	} else {
4917 		vm_page_bits_clear(m, &m->valid, bits);
4918 		vm_page_bits_clear(m, &m->dirty, bits);
4919 	}
4920 }
4921 
4922 /*
4923  *	vm_page_invalid:
4924  *
4925  *	Invalidates the entire page.  The page must be busy, unmapped, and
4926  *	the enclosing object must be locked.  The object locks protects
4927  *	against concurrent read-only pmap enter which is done without
4928  *	busy.
4929  */
4930 void
4931 vm_page_invalid(vm_page_t m)
4932 {
4933 
4934 	vm_page_assert_busied(m);
4935 	VM_OBJECT_ASSERT_LOCKED(m->object);
4936 	MPASS(!pmap_page_is_mapped(m));
4937 
4938 	if (vm_page_xbusied(m))
4939 		m->valid = 0;
4940 	else
4941 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4942 }
4943 
4944 /*
4945  * vm_page_zero_invalid()
4946  *
4947  *	The kernel assumes that the invalid portions of a page contain
4948  *	garbage, but such pages can be mapped into memory by user code.
4949  *	When this occurs, we must zero out the non-valid portions of the
4950  *	page so user code sees what it expects.
4951  *
4952  *	Pages are most often semi-valid when the end of a file is mapped
4953  *	into memory and the file's size is not page aligned.
4954  */
4955 void
4956 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4957 {
4958 	int b;
4959 	int i;
4960 
4961 	/*
4962 	 * Scan the valid bits looking for invalid sections that
4963 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4964 	 * valid bit may be set ) have already been zeroed by
4965 	 * vm_page_set_validclean().
4966 	 */
4967 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4968 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4969 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4970 			if (i > b) {
4971 				pmap_zero_page_area(m,
4972 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4973 			}
4974 			b = i + 1;
4975 		}
4976 	}
4977 
4978 	/*
4979 	 * setvalid is TRUE when we can safely set the zero'd areas
4980 	 * as being valid.  We can do this if there are no cache consistancy
4981 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4982 	 */
4983 	if (setvalid)
4984 		vm_page_valid(m);
4985 }
4986 
4987 /*
4988  *	vm_page_is_valid:
4989  *
4990  *	Is (partial) page valid?  Note that the case where size == 0
4991  *	will return FALSE in the degenerate case where the page is
4992  *	entirely invalid, and TRUE otherwise.
4993  *
4994  *	Some callers envoke this routine without the busy lock held and
4995  *	handle races via higher level locks.  Typical callers should
4996  *	hold a busy lock to prevent invalidation.
4997  */
4998 int
4999 vm_page_is_valid(vm_page_t m, int base, int size)
5000 {
5001 	vm_page_bits_t bits;
5002 
5003 	bits = vm_page_bits(base, size);
5004 	return (m->valid != 0 && (m->valid & bits) == bits);
5005 }
5006 
5007 /*
5008  * Returns true if all of the specified predicates are true for the entire
5009  * (super)page and false otherwise.
5010  */
5011 bool
5012 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5013 {
5014 	vm_object_t object;
5015 	int i, npages;
5016 
5017 	object = m->object;
5018 	if (skip_m != NULL && skip_m->object != object)
5019 		return (false);
5020 	VM_OBJECT_ASSERT_LOCKED(object);
5021 	npages = atop(pagesizes[m->psind]);
5022 
5023 	/*
5024 	 * The physically contiguous pages that make up a superpage, i.e., a
5025 	 * page with a page size index ("psind") greater than zero, will
5026 	 * occupy adjacent entries in vm_page_array[].
5027 	 */
5028 	for (i = 0; i < npages; i++) {
5029 		/* Always test object consistency, including "skip_m". */
5030 		if (m[i].object != object)
5031 			return (false);
5032 		if (&m[i] == skip_m)
5033 			continue;
5034 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5035 			return (false);
5036 		if ((flags & PS_ALL_DIRTY) != 0) {
5037 			/*
5038 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5039 			 * might stop this case from spuriously returning
5040 			 * "false".  However, that would require a write lock
5041 			 * on the object containing "m[i]".
5042 			 */
5043 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5044 				return (false);
5045 		}
5046 		if ((flags & PS_ALL_VALID) != 0 &&
5047 		    m[i].valid != VM_PAGE_BITS_ALL)
5048 			return (false);
5049 	}
5050 	return (true);
5051 }
5052 
5053 /*
5054  * Set the page's dirty bits if the page is modified.
5055  */
5056 void
5057 vm_page_test_dirty(vm_page_t m)
5058 {
5059 
5060 	vm_page_assert_busied(m);
5061 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5062 		vm_page_dirty(m);
5063 }
5064 
5065 void
5066 vm_page_valid(vm_page_t m)
5067 {
5068 
5069 	vm_page_assert_busied(m);
5070 	if (vm_page_xbusied(m))
5071 		m->valid = VM_PAGE_BITS_ALL;
5072 	else
5073 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5074 }
5075 
5076 void
5077 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5078 {
5079 
5080 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5081 }
5082 
5083 void
5084 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5085 {
5086 
5087 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5088 }
5089 
5090 int
5091 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5092 {
5093 
5094 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5095 }
5096 
5097 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5098 void
5099 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5100 {
5101 
5102 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5103 }
5104 
5105 void
5106 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5107 {
5108 
5109 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5110 }
5111 #endif
5112 
5113 #ifdef INVARIANTS
5114 void
5115 vm_page_object_busy_assert(vm_page_t m)
5116 {
5117 
5118 	/*
5119 	 * Certain of the page's fields may only be modified by the
5120 	 * holder of a page or object busy.
5121 	 */
5122 	if (m->object != NULL && !vm_page_busied(m))
5123 		VM_OBJECT_ASSERT_BUSY(m->object);
5124 }
5125 
5126 void
5127 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5128 {
5129 
5130 	if ((bits & PGA_WRITEABLE) == 0)
5131 		return;
5132 
5133 	/*
5134 	 * The PGA_WRITEABLE flag can only be set if the page is
5135 	 * managed, is exclusively busied or the object is locked.
5136 	 * Currently, this flag is only set by pmap_enter().
5137 	 */
5138 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5139 	    ("PGA_WRITEABLE on unmanaged page"));
5140 	if (!vm_page_xbusied(m))
5141 		VM_OBJECT_ASSERT_BUSY(m->object);
5142 }
5143 #endif
5144 
5145 #include "opt_ddb.h"
5146 #ifdef DDB
5147 #include <sys/kernel.h>
5148 
5149 #include <ddb/ddb.h>
5150 
5151 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5152 {
5153 
5154 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5155 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5156 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5157 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5158 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5159 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5160 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5161 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5162 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5163 }
5164 
5165 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5166 {
5167 	int dom;
5168 
5169 	db_printf("pq_free %d\n", vm_free_count());
5170 	for (dom = 0; dom < vm_ndomains; dom++) {
5171 		db_printf(
5172     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5173 		    dom,
5174 		    vm_dom[dom].vmd_page_count,
5175 		    vm_dom[dom].vmd_free_count,
5176 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5177 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5178 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5179 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5180 	}
5181 }
5182 
5183 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5184 {
5185 	vm_page_t m;
5186 	boolean_t phys, virt;
5187 
5188 	if (!have_addr) {
5189 		db_printf("show pginfo addr\n");
5190 		return;
5191 	}
5192 
5193 	phys = strchr(modif, 'p') != NULL;
5194 	virt = strchr(modif, 'v') != NULL;
5195 	if (virt)
5196 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5197 	else if (phys)
5198 		m = PHYS_TO_VM_PAGE(addr);
5199 	else
5200 		m = (vm_page_t)addr;
5201 	db_printf(
5202     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5203     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5204 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5205 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5206 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5207 }
5208 #endif /* DDB */
5209