xref: /freebsd/sys/vm/vm_page.c (revision d34048812292b714a0bf99967270d18fe3097c62)
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  *			GENERAL RULES ON VM_PAGE MANIPULATION
67  *
68  *	- A page queue lock is required when adding or removing a page from a
69  *	  page queue regardless of other locks or the busy state of a page.
70  *
71  *		* In general, no thread besides the page daemon can acquire or
72  *		  hold more than one page queue lock at a time.
73  *
74  *		* The page daemon can acquire and hold any pair of page queue
75  *		  locks in any order.
76  *
77  *	- The object lock is required when inserting or removing
78  *	  pages from an object (vm_page_insert() or vm_page_remove()).
79  *
80  */
81 
82 /*
83  *	Resident memory management module.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 #include "opt_vm.h"
90 
91 #include <sys/param.h>
92 #include <sys/systm.h>
93 #include <sys/lock.h>
94 #include <sys/domainset.h>
95 #include <sys/kernel.h>
96 #include <sys/limits.h>
97 #include <sys/linker.h>
98 #include <sys/malloc.h>
99 #include <sys/mman.h>
100 #include <sys/msgbuf.h>
101 #include <sys/mutex.h>
102 #include <sys/proc.h>
103 #include <sys/rwlock.h>
104 #include <sys/sbuf.h>
105 #include <sys/sched.h>
106 #include <sys/smp.h>
107 #include <sys/sysctl.h>
108 #include <sys/vmmeter.h>
109 #include <sys/vnode.h>
110 
111 #include <vm/vm.h>
112 #include <vm/pmap.h>
113 #include <vm/vm_param.h>
114 #include <vm/vm_domainset.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_map.h>
117 #include <vm/vm_object.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_pageout.h>
120 #include <vm/vm_phys.h>
121 #include <vm/vm_pagequeue.h>
122 #include <vm/vm_pager.h>
123 #include <vm/vm_radix.h>
124 #include <vm/vm_reserv.h>
125 #include <vm/vm_extern.h>
126 #include <vm/uma.h>
127 #include <vm/uma_int.h>
128 
129 #include <machine/md_var.h>
130 
131 extern int	uma_startup_count(int);
132 extern void	uma_startup(void *, int);
133 extern int	vmem_startup_count(void);
134 
135 struct vm_domain vm_dom[MAXMEMDOM];
136 
137 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
138 
139 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
140 
141 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
142 /* The following fields are protected by the domainset lock. */
143 domainset_t __exclusive_cache_line vm_min_domains;
144 domainset_t __exclusive_cache_line vm_severe_domains;
145 static int vm_min_waiters;
146 static int vm_severe_waiters;
147 static int vm_pageproc_waiters;
148 
149 /*
150  * bogus page -- for I/O to/from partially complete buffers,
151  * or for paging into sparsely invalid regions.
152  */
153 vm_page_t bogus_page;
154 
155 vm_page_t vm_page_array;
156 long vm_page_array_size;
157 long first_page;
158 
159 static int boot_pages;
160 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
161     &boot_pages, 0,
162     "number of pages allocated for bootstrapping the VM system");
163 
164 static int pa_tryrelock_restart;
165 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
166     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
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 void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
177 static void vm_page_dequeue_complete(vm_page_t m);
178 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
179 static void vm_page_init(void *dummy);
180 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
181     vm_pindex_t pindex, vm_page_t mpred);
182 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
183     vm_page_t mpred);
184 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
185     vm_page_t m_run, vm_paddr_t high);
186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
187     int req);
188 static int vm_page_import(void *arg, void **store, int cnt, int domain,
189     int flags);
190 static void vm_page_release(void *arg, void **store, int cnt);
191 
192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
193 
194 static void
195 vm_page_init(void *dummy)
196 {
197 
198 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
199 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
200 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
201 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
202 }
203 
204 /*
205  * The cache page zone is initialized later since we need to be able to allocate
206  * pages before UMA is fully initialized.
207  */
208 static void
209 vm_page_init_cache_zones(void *dummy __unused)
210 {
211 	struct vm_domain *vmd;
212 	int i;
213 
214 	for (i = 0; i < vm_ndomains; i++) {
215 		vmd = VM_DOMAIN(i);
216 		/*
217 		 * Don't allow the page cache to take up more than .25% of
218 		 * memory.
219 		 */
220 		if (vmd->vmd_page_count / 400 < 256 * mp_ncpus)
221 			continue;
222 		vmd->vmd_pgcache = uma_zcache_create("vm pgcache",
223 		    sizeof(struct vm_page), NULL, NULL, NULL, NULL,
224 		    vm_page_import, vm_page_release, vmd,
225 		    UMA_ZONE_NOBUCKETCACHE | UMA_ZONE_MAXBUCKET | UMA_ZONE_VM);
226 	}
227 }
228 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
229 
230 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
231 #if PAGE_SIZE == 32768
232 #ifdef CTASSERT
233 CTASSERT(sizeof(u_long) >= 8);
234 #endif
235 #endif
236 
237 /*
238  * Try to acquire a physical address lock while a pmap is locked.  If we
239  * fail to trylock we unlock and lock the pmap directly and cache the
240  * locked pa in *locked.  The caller should then restart their loop in case
241  * the virtual to physical mapping has changed.
242  */
243 int
244 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
245 {
246 	vm_paddr_t lockpa;
247 
248 	lockpa = *locked;
249 	*locked = pa;
250 	if (lockpa) {
251 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
252 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
253 			return (0);
254 		PA_UNLOCK(lockpa);
255 	}
256 	if (PA_TRYLOCK(pa))
257 		return (0);
258 	PMAP_UNLOCK(pmap);
259 	atomic_add_int(&pa_tryrelock_restart, 1);
260 	PA_LOCK(pa);
261 	PMAP_LOCK(pmap);
262 	return (EAGAIN);
263 }
264 
265 /*
266  *	vm_set_page_size:
267  *
268  *	Sets the page size, perhaps based upon the memory
269  *	size.  Must be called before any use of page-size
270  *	dependent functions.
271  */
272 void
273 vm_set_page_size(void)
274 {
275 	if (vm_cnt.v_page_size == 0)
276 		vm_cnt.v_page_size = PAGE_SIZE;
277 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
278 		panic("vm_set_page_size: page size not a power of two");
279 }
280 
281 /*
282  *	vm_page_blacklist_next:
283  *
284  *	Find the next entry in the provided string of blacklist
285  *	addresses.  Entries are separated by space, comma, or newline.
286  *	If an invalid integer is encountered then the rest of the
287  *	string is skipped.  Updates the list pointer to the next
288  *	character, or NULL if the string is exhausted or invalid.
289  */
290 static vm_paddr_t
291 vm_page_blacklist_next(char **list, char *end)
292 {
293 	vm_paddr_t bad;
294 	char *cp, *pos;
295 
296 	if (list == NULL || *list == NULL)
297 		return (0);
298 	if (**list =='\0') {
299 		*list = NULL;
300 		return (0);
301 	}
302 
303 	/*
304 	 * If there's no end pointer then the buffer is coming from
305 	 * the kenv and we know it's null-terminated.
306 	 */
307 	if (end == NULL)
308 		end = *list + strlen(*list);
309 
310 	/* Ensure that strtoq() won't walk off the end */
311 	if (*end != '\0') {
312 		if (*end == '\n' || *end == ' ' || *end  == ',')
313 			*end = '\0';
314 		else {
315 			printf("Blacklist not terminated, skipping\n");
316 			*list = NULL;
317 			return (0);
318 		}
319 	}
320 
321 	for (pos = *list; *pos != '\0'; pos = cp) {
322 		bad = strtoq(pos, &cp, 0);
323 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
324 			if (bad == 0) {
325 				if (++cp < end)
326 					continue;
327 				else
328 					break;
329 			}
330 		} else
331 			break;
332 		if (*cp == '\0' || ++cp >= end)
333 			*list = NULL;
334 		else
335 			*list = cp;
336 		return (trunc_page(bad));
337 	}
338 	printf("Garbage in RAM blacklist, skipping\n");
339 	*list = NULL;
340 	return (0);
341 }
342 
343 bool
344 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
345 {
346 	struct vm_domain *vmd;
347 	vm_page_t m;
348 	int ret;
349 
350 	m = vm_phys_paddr_to_vm_page(pa);
351 	if (m == NULL)
352 		return (true); /* page does not exist, no failure */
353 
354 	vmd = vm_pagequeue_domain(m);
355 	vm_domain_free_lock(vmd);
356 	ret = vm_phys_unfree_page(m);
357 	vm_domain_free_unlock(vmd);
358 	if (ret) {
359 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
360 		if (verbose)
361 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
362 	}
363 	return (ret);
364 }
365 
366 /*
367  *	vm_page_blacklist_check:
368  *
369  *	Iterate through the provided string of blacklist addresses, pulling
370  *	each entry out of the physical allocator free list and putting it
371  *	onto a list for reporting via the vm.page_blacklist sysctl.
372  */
373 static void
374 vm_page_blacklist_check(char *list, char *end)
375 {
376 	vm_paddr_t pa;
377 	char *next;
378 
379 	next = list;
380 	while (next != NULL) {
381 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
382 			continue;
383 		vm_page_blacklist_add(pa, bootverbose);
384 	}
385 }
386 
387 /*
388  *	vm_page_blacklist_load:
389  *
390  *	Search for a special module named "ram_blacklist".  It'll be a
391  *	plain text file provided by the user via the loader directive
392  *	of the same name.
393  */
394 static void
395 vm_page_blacklist_load(char **list, char **end)
396 {
397 	void *mod;
398 	u_char *ptr;
399 	u_int len;
400 
401 	mod = NULL;
402 	ptr = NULL;
403 
404 	mod = preload_search_by_type("ram_blacklist");
405 	if (mod != NULL) {
406 		ptr = preload_fetch_addr(mod);
407 		len = preload_fetch_size(mod);
408         }
409 	*list = ptr;
410 	if (ptr != NULL)
411 		*end = ptr + len;
412 	else
413 		*end = NULL;
414 	return;
415 }
416 
417 static int
418 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
419 {
420 	vm_page_t m;
421 	struct sbuf sbuf;
422 	int error, first;
423 
424 	first = 1;
425 	error = sysctl_wire_old_buffer(req, 0);
426 	if (error != 0)
427 		return (error);
428 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
429 	TAILQ_FOREACH(m, &blacklist_head, listq) {
430 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
431 		    (uintmax_t)m->phys_addr);
432 		first = 0;
433 	}
434 	error = sbuf_finish(&sbuf);
435 	sbuf_delete(&sbuf);
436 	return (error);
437 }
438 
439 /*
440  * Initialize a dummy page for use in scans of the specified paging queue.
441  * In principle, this function only needs to set the flag PG_MARKER.
442  * Nonetheless, it write busies and initializes the hold count to one as
443  * safety precautions.
444  */
445 static void
446 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
447 {
448 
449 	bzero(marker, sizeof(*marker));
450 	marker->flags = PG_MARKER;
451 	marker->aflags = aflags;
452 	marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
453 	marker->queue = queue;
454 	marker->hold_count = 1;
455 }
456 
457 static void
458 vm_page_domain_init(int domain)
459 {
460 	struct vm_domain *vmd;
461 	struct vm_pagequeue *pq;
462 	int i;
463 
464 	vmd = VM_DOMAIN(domain);
465 	bzero(vmd, sizeof(*vmd));
466 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
467 	    "vm inactive pagequeue";
468 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
469 	    "vm active pagequeue";
470 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
471 	    "vm laundry pagequeue";
472 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
473 	    "vm unswappable pagequeue";
474 	vmd->vmd_domain = domain;
475 	vmd->vmd_page_count = 0;
476 	vmd->vmd_free_count = 0;
477 	vmd->vmd_segs = 0;
478 	vmd->vmd_oom = FALSE;
479 	for (i = 0; i < PQ_COUNT; i++) {
480 		pq = &vmd->vmd_pagequeues[i];
481 		TAILQ_INIT(&pq->pq_pl);
482 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
483 		    MTX_DEF | MTX_DUPOK);
484 		pq->pq_pdpages = 0;
485 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
486 	}
487 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
488 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
489 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
490 
491 	/*
492 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
493 	 * insertions.
494 	 */
495 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
496 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
497 	    &vmd->vmd_inacthead, plinks.q);
498 
499 	/*
500 	 * The clock pages are used to implement active queue scanning without
501 	 * requeues.  Scans start at clock[0], which is advanced after the scan
502 	 * ends.  When the two clock hands meet, they are reset and scanning
503 	 * resumes from the head of the queue.
504 	 */
505 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
506 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
507 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
508 	    &vmd->vmd_clock[0], plinks.q);
509 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
510 	    &vmd->vmd_clock[1], plinks.q);
511 }
512 
513 /*
514  * Initialize a physical page in preparation for adding it to the free
515  * lists.
516  */
517 static void
518 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
519 {
520 
521 	m->object = NULL;
522 	m->wire_count = 0;
523 	m->busy_lock = VPB_UNBUSIED;
524 	m->hold_count = 0;
525 	m->flags = m->aflags = 0;
526 	m->phys_addr = pa;
527 	m->queue = PQ_NONE;
528 	m->psind = 0;
529 	m->segind = segind;
530 	m->order = VM_NFREEORDER;
531 	m->pool = VM_FREEPOOL_DEFAULT;
532 	m->valid = m->dirty = 0;
533 	pmap_page_init(m);
534 }
535 
536 /*
537  *	vm_page_startup:
538  *
539  *	Initializes the resident memory module.  Allocates physical memory for
540  *	bootstrapping UMA and some data structures that are used to manage
541  *	physical pages.  Initializes these structures, and populates the free
542  *	page queues.
543  */
544 vm_offset_t
545 vm_page_startup(vm_offset_t vaddr)
546 {
547 	struct vm_phys_seg *seg;
548 	vm_page_t m;
549 	char *list, *listend;
550 	vm_offset_t mapped;
551 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
552 	vm_paddr_t biggestsize, last_pa, pa;
553 	u_long pagecount;
554 	int biggestone, i, segind;
555 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
556 	long ii;
557 #endif
558 
559 	biggestsize = 0;
560 	biggestone = 0;
561 	vaddr = round_page(vaddr);
562 
563 	for (i = 0; phys_avail[i + 1]; i += 2) {
564 		phys_avail[i] = round_page(phys_avail[i]);
565 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
566 	}
567 	for (i = 0; phys_avail[i + 1]; i += 2) {
568 		size = phys_avail[i + 1] - phys_avail[i];
569 		if (size > biggestsize) {
570 			biggestone = i;
571 			biggestsize = size;
572 		}
573 	}
574 
575 	end = phys_avail[biggestone+1];
576 
577 	/*
578 	 * Initialize the page and queue locks.
579 	 */
580 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
581 	for (i = 0; i < PA_LOCK_COUNT; i++)
582 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
583 	for (i = 0; i < vm_ndomains; i++)
584 		vm_page_domain_init(i);
585 
586 	/*
587 	 * Allocate memory for use when boot strapping the kernel memory
588 	 * allocator.  Tell UMA how many zones we are going to create
589 	 * before going fully functional.  UMA will add its zones.
590 	 *
591 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
592 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
593 	 */
594 	boot_pages = uma_startup_count(8);
595 
596 #ifndef UMA_MD_SMALL_ALLOC
597 	/* vmem_startup() calls uma_prealloc(). */
598 	boot_pages += vmem_startup_count();
599 	/* vm_map_startup() calls uma_prealloc(). */
600 	boot_pages += howmany(MAX_KMAP,
601 	    UMA_SLAB_SPACE / sizeof(struct vm_map));
602 
603 	/*
604 	 * Before going fully functional kmem_init() does allocation
605 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
606 	 */
607 	boot_pages += 2;
608 #endif
609 	/*
610 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
611 	 * manually fetch the value.
612 	 */
613 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
614 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
615 	new_end = trunc_page(new_end);
616 	mapped = pmap_map(&vaddr, new_end, end,
617 	    VM_PROT_READ | VM_PROT_WRITE);
618 	bzero((void *)mapped, end - new_end);
619 	uma_startup((void *)mapped, boot_pages);
620 
621 #ifdef WITNESS
622 	end = new_end;
623 	new_end = end - round_page(witness_startup_count());
624 	mapped = pmap_map(&vaddr, new_end, end,
625 	    VM_PROT_READ | VM_PROT_WRITE);
626 	bzero((void *)mapped, end - new_end);
627 	witness_startup((void *)mapped);
628 #endif
629 
630 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
631     defined(__i386__) || defined(__mips__)
632 	/*
633 	 * Allocate a bitmap to indicate that a random physical page
634 	 * needs to be included in a minidump.
635 	 *
636 	 * The amd64 port needs this to indicate which direct map pages
637 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
638 	 *
639 	 * However, i386 still needs this workspace internally within the
640 	 * minidump code.  In theory, they are not needed on i386, but are
641 	 * included should the sf_buf code decide to use them.
642 	 */
643 	last_pa = 0;
644 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
645 		if (dump_avail[i + 1] > last_pa)
646 			last_pa = dump_avail[i + 1];
647 	page_range = last_pa / PAGE_SIZE;
648 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
649 	new_end -= vm_page_dump_size;
650 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
651 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
652 	bzero((void *)vm_page_dump, vm_page_dump_size);
653 #else
654 	(void)last_pa;
655 #endif
656 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
657 	/*
658 	 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
659 	 * When pmap_map() uses the direct map, they are not automatically
660 	 * included.
661 	 */
662 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
663 		dump_add_page(pa);
664 #endif
665 	phys_avail[biggestone + 1] = new_end;
666 #ifdef __amd64__
667 	/*
668 	 * Request that the physical pages underlying the message buffer be
669 	 * included in a crash dump.  Since the message buffer is accessed
670 	 * through the direct map, they are not automatically included.
671 	 */
672 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
673 	last_pa = pa + round_page(msgbufsize);
674 	while (pa < last_pa) {
675 		dump_add_page(pa);
676 		pa += PAGE_SIZE;
677 	}
678 #endif
679 	/*
680 	 * Compute the number of pages of memory that will be available for
681 	 * use, taking into account the overhead of a page structure per page.
682 	 * In other words, solve
683 	 *	"available physical memory" - round_page(page_range *
684 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
685 	 * for page_range.
686 	 */
687 	low_avail = phys_avail[0];
688 	high_avail = phys_avail[1];
689 	for (i = 0; i < vm_phys_nsegs; i++) {
690 		if (vm_phys_segs[i].start < low_avail)
691 			low_avail = vm_phys_segs[i].start;
692 		if (vm_phys_segs[i].end > high_avail)
693 			high_avail = vm_phys_segs[i].end;
694 	}
695 	/* Skip the first chunk.  It is already accounted for. */
696 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
697 		if (phys_avail[i] < low_avail)
698 			low_avail = phys_avail[i];
699 		if (phys_avail[i + 1] > high_avail)
700 			high_avail = phys_avail[i + 1];
701 	}
702 	first_page = low_avail / PAGE_SIZE;
703 #ifdef VM_PHYSSEG_SPARSE
704 	size = 0;
705 	for (i = 0; i < vm_phys_nsegs; i++)
706 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
707 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
708 		size += phys_avail[i + 1] - phys_avail[i];
709 #elif defined(VM_PHYSSEG_DENSE)
710 	size = high_avail - low_avail;
711 #else
712 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
713 #endif
714 
715 #ifdef VM_PHYSSEG_DENSE
716 	/*
717 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
718 	 * the overhead of a page structure per page only if vm_page_array is
719 	 * allocated from the last physical memory chunk.  Otherwise, we must
720 	 * allocate page structures representing the physical memory
721 	 * underlying vm_page_array, even though they will not be used.
722 	 */
723 	if (new_end != high_avail)
724 		page_range = size / PAGE_SIZE;
725 	else
726 #endif
727 	{
728 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
729 
730 		/*
731 		 * If the partial bytes remaining are large enough for
732 		 * a page (PAGE_SIZE) without a corresponding
733 		 * 'struct vm_page', then new_end will contain an
734 		 * extra page after subtracting the length of the VM
735 		 * page array.  Compensate by subtracting an extra
736 		 * page from new_end.
737 		 */
738 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
739 			if (new_end == high_avail)
740 				high_avail -= PAGE_SIZE;
741 			new_end -= PAGE_SIZE;
742 		}
743 	}
744 	end = new_end;
745 
746 	/*
747 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
748 	 * However, because this page is allocated from KVM, out-of-bounds
749 	 * accesses using the direct map will not be trapped.
750 	 */
751 	vaddr += PAGE_SIZE;
752 
753 	/*
754 	 * Allocate physical memory for the page structures, and map it.
755 	 */
756 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
757 	mapped = pmap_map(&vaddr, new_end, end,
758 	    VM_PROT_READ | VM_PROT_WRITE);
759 	vm_page_array = (vm_page_t)mapped;
760 	vm_page_array_size = page_range;
761 
762 #if VM_NRESERVLEVEL > 0
763 	/*
764 	 * Allocate physical memory for the reservation management system's
765 	 * data structures, and map it.
766 	 */
767 	if (high_avail == end)
768 		high_avail = new_end;
769 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
770 #endif
771 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
772 	/*
773 	 * Include vm_page_array and vm_reserv_array in a crash dump.
774 	 */
775 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
776 		dump_add_page(pa);
777 #endif
778 	phys_avail[biggestone + 1] = new_end;
779 
780 	/*
781 	 * Add physical memory segments corresponding to the available
782 	 * physical pages.
783 	 */
784 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
785 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
786 
787 	/*
788 	 * Initialize the physical memory allocator.
789 	 */
790 	vm_phys_init();
791 
792 	/*
793 	 * Initialize the page structures and add every available page to the
794 	 * physical memory allocator's free lists.
795 	 */
796 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
797 	for (ii = 0; ii < vm_page_array_size; ii++) {
798 		m = &vm_page_array[ii];
799 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
800 		m->flags = PG_FICTITIOUS;
801 	}
802 #endif
803 	vm_cnt.v_page_count = 0;
804 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
805 		seg = &vm_phys_segs[segind];
806 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
807 		    m++, pa += PAGE_SIZE)
808 			vm_page_init_page(m, pa, segind);
809 
810 		/*
811 		 * Add the segment to the free lists only if it is covered by
812 		 * one of the ranges in phys_avail.  Because we've added the
813 		 * ranges to the vm_phys_segs array, we can assume that each
814 		 * segment is either entirely contained in one of the ranges,
815 		 * or doesn't overlap any of them.
816 		 */
817 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
818 			struct vm_domain *vmd;
819 
820 			if (seg->start < phys_avail[i] ||
821 			    seg->end > phys_avail[i + 1])
822 				continue;
823 
824 			m = seg->first_page;
825 			pagecount = (u_long)atop(seg->end - seg->start);
826 
827 			vmd = VM_DOMAIN(seg->domain);
828 			vm_domain_free_lock(vmd);
829 			vm_phys_free_contig(m, pagecount);
830 			vm_domain_free_unlock(vmd);
831 			vm_domain_freecnt_inc(vmd, pagecount);
832 			vm_cnt.v_page_count += (u_int)pagecount;
833 
834 			vmd = VM_DOMAIN(seg->domain);
835 			vmd->vmd_page_count += (u_int)pagecount;
836 			vmd->vmd_segs |= 1UL << m->segind;
837 			break;
838 		}
839 	}
840 
841 	/*
842 	 * Remove blacklisted pages from the physical memory allocator.
843 	 */
844 	TAILQ_INIT(&blacklist_head);
845 	vm_page_blacklist_load(&list, &listend);
846 	vm_page_blacklist_check(list, listend);
847 
848 	list = kern_getenv("vm.blacklist");
849 	vm_page_blacklist_check(list, NULL);
850 
851 	freeenv(list);
852 #if VM_NRESERVLEVEL > 0
853 	/*
854 	 * Initialize the reservation management system.
855 	 */
856 	vm_reserv_init();
857 #endif
858 
859 	return (vaddr);
860 }
861 
862 void
863 vm_page_reference(vm_page_t m)
864 {
865 
866 	vm_page_aflag_set(m, PGA_REFERENCED);
867 }
868 
869 /*
870  *	vm_page_busy_downgrade:
871  *
872  *	Downgrade an exclusive busy page into a single shared busy page.
873  */
874 void
875 vm_page_busy_downgrade(vm_page_t m)
876 {
877 	u_int x;
878 	bool locked;
879 
880 	vm_page_assert_xbusied(m);
881 	locked = mtx_owned(vm_page_lockptr(m));
882 
883 	for (;;) {
884 		x = m->busy_lock;
885 		x &= VPB_BIT_WAITERS;
886 		if (x != 0 && !locked)
887 			vm_page_lock(m);
888 		if (atomic_cmpset_rel_int(&m->busy_lock,
889 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
890 			break;
891 		if (x != 0 && !locked)
892 			vm_page_unlock(m);
893 	}
894 	if (x != 0) {
895 		wakeup(m);
896 		if (!locked)
897 			vm_page_unlock(m);
898 	}
899 }
900 
901 /*
902  *	vm_page_sbusied:
903  *
904  *	Return a positive value if the page is shared busied, 0 otherwise.
905  */
906 int
907 vm_page_sbusied(vm_page_t m)
908 {
909 	u_int x;
910 
911 	x = m->busy_lock;
912 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
913 }
914 
915 /*
916  *	vm_page_sunbusy:
917  *
918  *	Shared unbusy a page.
919  */
920 void
921 vm_page_sunbusy(vm_page_t m)
922 {
923 	u_int x;
924 
925 	vm_page_lock_assert(m, MA_NOTOWNED);
926 	vm_page_assert_sbusied(m);
927 
928 	for (;;) {
929 		x = m->busy_lock;
930 		if (VPB_SHARERS(x) > 1) {
931 			if (atomic_cmpset_int(&m->busy_lock, x,
932 			    x - VPB_ONE_SHARER))
933 				break;
934 			continue;
935 		}
936 		if ((x & VPB_BIT_WAITERS) == 0) {
937 			KASSERT(x == VPB_SHARERS_WORD(1),
938 			    ("vm_page_sunbusy: invalid lock state"));
939 			if (atomic_cmpset_int(&m->busy_lock,
940 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
941 				break;
942 			continue;
943 		}
944 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
945 		    ("vm_page_sunbusy: invalid lock state for waiters"));
946 
947 		vm_page_lock(m);
948 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
949 			vm_page_unlock(m);
950 			continue;
951 		}
952 		wakeup(m);
953 		vm_page_unlock(m);
954 		break;
955 	}
956 }
957 
958 /*
959  *	vm_page_busy_sleep:
960  *
961  *	Sleep and release the page lock, using the page pointer as wchan.
962  *	This is used to implement the hard-path of busying mechanism.
963  *
964  *	The given page must be locked.
965  *
966  *	If nonshared is true, sleep only if the page is xbusy.
967  */
968 void
969 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
970 {
971 	u_int x;
972 
973 	vm_page_assert_locked(m);
974 
975 	x = m->busy_lock;
976 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
977 	    ((x & VPB_BIT_WAITERS) == 0 &&
978 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
979 		vm_page_unlock(m);
980 		return;
981 	}
982 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
983 }
984 
985 /*
986  *	vm_page_trysbusy:
987  *
988  *	Try to shared busy a page.
989  *	If the operation succeeds 1 is returned otherwise 0.
990  *	The operation never sleeps.
991  */
992 int
993 vm_page_trysbusy(vm_page_t m)
994 {
995 	u_int x;
996 
997 	for (;;) {
998 		x = m->busy_lock;
999 		if ((x & VPB_BIT_SHARED) == 0)
1000 			return (0);
1001 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
1002 			return (1);
1003 	}
1004 }
1005 
1006 static void
1007 vm_page_xunbusy_locked(vm_page_t m)
1008 {
1009 
1010 	vm_page_assert_xbusied(m);
1011 	vm_page_assert_locked(m);
1012 
1013 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1014 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
1015 	wakeup(m);
1016 }
1017 
1018 void
1019 vm_page_xunbusy_maybelocked(vm_page_t m)
1020 {
1021 	bool lockacq;
1022 
1023 	vm_page_assert_xbusied(m);
1024 
1025 	/*
1026 	 * Fast path for unbusy.  If it succeeds, we know that there
1027 	 * are no waiters, so we do not need a wakeup.
1028 	 */
1029 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
1030 	    VPB_UNBUSIED))
1031 		return;
1032 
1033 	lockacq = !mtx_owned(vm_page_lockptr(m));
1034 	if (lockacq)
1035 		vm_page_lock(m);
1036 	vm_page_xunbusy_locked(m);
1037 	if (lockacq)
1038 		vm_page_unlock(m);
1039 }
1040 
1041 /*
1042  *	vm_page_xunbusy_hard:
1043  *
1044  *	Called after the first try the exclusive unbusy of a page failed.
1045  *	It is assumed that the waiters bit is on.
1046  */
1047 void
1048 vm_page_xunbusy_hard(vm_page_t m)
1049 {
1050 
1051 	vm_page_assert_xbusied(m);
1052 
1053 	vm_page_lock(m);
1054 	vm_page_xunbusy_locked(m);
1055 	vm_page_unlock(m);
1056 }
1057 
1058 /*
1059  *	vm_page_flash:
1060  *
1061  *	Wakeup anyone waiting for the page.
1062  *	The ownership bits do not change.
1063  *
1064  *	The given page must be locked.
1065  */
1066 void
1067 vm_page_flash(vm_page_t m)
1068 {
1069 	u_int x;
1070 
1071 	vm_page_lock_assert(m, MA_OWNED);
1072 
1073 	for (;;) {
1074 		x = m->busy_lock;
1075 		if ((x & VPB_BIT_WAITERS) == 0)
1076 			return;
1077 		if (atomic_cmpset_int(&m->busy_lock, x,
1078 		    x & (~VPB_BIT_WAITERS)))
1079 			break;
1080 	}
1081 	wakeup(m);
1082 }
1083 
1084 /*
1085  * Avoid releasing and reacquiring the same page lock.
1086  */
1087 void
1088 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1089 {
1090 	struct mtx *mtx1;
1091 
1092 	mtx1 = vm_page_lockptr(m);
1093 	if (*mtx == mtx1)
1094 		return;
1095 	if (*mtx != NULL)
1096 		mtx_unlock(*mtx);
1097 	*mtx = mtx1;
1098 	mtx_lock(mtx1);
1099 }
1100 
1101 /*
1102  * Keep page from being freed by the page daemon
1103  * much of the same effect as wiring, except much lower
1104  * overhead and should be used only for *very* temporary
1105  * holding ("wiring").
1106  */
1107 void
1108 vm_page_hold(vm_page_t mem)
1109 {
1110 
1111 	vm_page_lock_assert(mem, MA_OWNED);
1112         mem->hold_count++;
1113 }
1114 
1115 void
1116 vm_page_unhold(vm_page_t mem)
1117 {
1118 
1119 	vm_page_lock_assert(mem, MA_OWNED);
1120 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1121 	--mem->hold_count;
1122 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1123 		vm_page_free_toq(mem);
1124 }
1125 
1126 /*
1127  *	vm_page_unhold_pages:
1128  *
1129  *	Unhold each of the pages that is referenced by the given array.
1130  */
1131 void
1132 vm_page_unhold_pages(vm_page_t *ma, int count)
1133 {
1134 	struct mtx *mtx;
1135 
1136 	mtx = NULL;
1137 	for (; count != 0; count--) {
1138 		vm_page_change_lock(*ma, &mtx);
1139 		vm_page_unhold(*ma);
1140 		ma++;
1141 	}
1142 	if (mtx != NULL)
1143 		mtx_unlock(mtx);
1144 }
1145 
1146 vm_page_t
1147 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1148 {
1149 	vm_page_t m;
1150 
1151 #ifdef VM_PHYSSEG_SPARSE
1152 	m = vm_phys_paddr_to_vm_page(pa);
1153 	if (m == NULL)
1154 		m = vm_phys_fictitious_to_vm_page(pa);
1155 	return (m);
1156 #elif defined(VM_PHYSSEG_DENSE)
1157 	long pi;
1158 
1159 	pi = atop(pa);
1160 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1161 		m = &vm_page_array[pi - first_page];
1162 		return (m);
1163 	}
1164 	return (vm_phys_fictitious_to_vm_page(pa));
1165 #else
1166 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1167 #endif
1168 }
1169 
1170 /*
1171  *	vm_page_getfake:
1172  *
1173  *	Create a fictitious page with the specified physical address and
1174  *	memory attribute.  The memory attribute is the only the machine-
1175  *	dependent aspect of a fictitious page that must be initialized.
1176  */
1177 vm_page_t
1178 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1179 {
1180 	vm_page_t m;
1181 
1182 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1183 	vm_page_initfake(m, paddr, memattr);
1184 	return (m);
1185 }
1186 
1187 void
1188 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1189 {
1190 
1191 	if ((m->flags & PG_FICTITIOUS) != 0) {
1192 		/*
1193 		 * The page's memattr might have changed since the
1194 		 * previous initialization.  Update the pmap to the
1195 		 * new memattr.
1196 		 */
1197 		goto memattr;
1198 	}
1199 	m->phys_addr = paddr;
1200 	m->queue = PQ_NONE;
1201 	/* Fictitious pages don't use "segind". */
1202 	m->flags = PG_FICTITIOUS;
1203 	/* Fictitious pages don't use "order" or "pool". */
1204 	m->oflags = VPO_UNMANAGED;
1205 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1206 	m->wire_count = 1;
1207 	pmap_page_init(m);
1208 memattr:
1209 	pmap_page_set_memattr(m, memattr);
1210 }
1211 
1212 /*
1213  *	vm_page_putfake:
1214  *
1215  *	Release a fictitious page.
1216  */
1217 void
1218 vm_page_putfake(vm_page_t m)
1219 {
1220 
1221 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1222 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1223 	    ("vm_page_putfake: bad page %p", m));
1224 	uma_zfree(fakepg_zone, m);
1225 }
1226 
1227 /*
1228  *	vm_page_updatefake:
1229  *
1230  *	Update the given fictitious page to the specified physical address and
1231  *	memory attribute.
1232  */
1233 void
1234 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1235 {
1236 
1237 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1238 	    ("vm_page_updatefake: bad page %p", m));
1239 	m->phys_addr = paddr;
1240 	pmap_page_set_memattr(m, memattr);
1241 }
1242 
1243 /*
1244  *	vm_page_free:
1245  *
1246  *	Free a page.
1247  */
1248 void
1249 vm_page_free(vm_page_t m)
1250 {
1251 
1252 	m->flags &= ~PG_ZERO;
1253 	vm_page_free_toq(m);
1254 }
1255 
1256 /*
1257  *	vm_page_free_zero:
1258  *
1259  *	Free a page to the zerod-pages queue
1260  */
1261 void
1262 vm_page_free_zero(vm_page_t m)
1263 {
1264 
1265 	m->flags |= PG_ZERO;
1266 	vm_page_free_toq(m);
1267 }
1268 
1269 /*
1270  * Unbusy and handle the page queueing for a page from a getpages request that
1271  * was optionally read ahead or behind.
1272  */
1273 void
1274 vm_page_readahead_finish(vm_page_t m)
1275 {
1276 
1277 	/* We shouldn't put invalid pages on queues. */
1278 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1279 
1280 	/*
1281 	 * Since the page is not the actually needed one, whether it should
1282 	 * be activated or deactivated is not obvious.  Empirical results
1283 	 * have shown that deactivating the page is usually the best choice,
1284 	 * unless the page is wanted by another thread.
1285 	 */
1286 	vm_page_lock(m);
1287 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1288 		vm_page_activate(m);
1289 	else
1290 		vm_page_deactivate(m);
1291 	vm_page_unlock(m);
1292 	vm_page_xunbusy(m);
1293 }
1294 
1295 /*
1296  *	vm_page_sleep_if_busy:
1297  *
1298  *	Sleep and release the page queues lock if the page is busied.
1299  *	Returns TRUE if the thread slept.
1300  *
1301  *	The given page must be unlocked and object containing it must
1302  *	be locked.
1303  */
1304 int
1305 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1306 {
1307 	vm_object_t obj;
1308 
1309 	vm_page_lock_assert(m, MA_NOTOWNED);
1310 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1311 
1312 	if (vm_page_busied(m)) {
1313 		/*
1314 		 * The page-specific object must be cached because page
1315 		 * identity can change during the sleep, causing the
1316 		 * re-lock of a different object.
1317 		 * It is assumed that a reference to the object is already
1318 		 * held by the callers.
1319 		 */
1320 		obj = m->object;
1321 		vm_page_lock(m);
1322 		VM_OBJECT_WUNLOCK(obj);
1323 		vm_page_busy_sleep(m, msg, false);
1324 		VM_OBJECT_WLOCK(obj);
1325 		return (TRUE);
1326 	}
1327 	return (FALSE);
1328 }
1329 
1330 /*
1331  *	vm_page_dirty_KBI:		[ internal use only ]
1332  *
1333  *	Set all bits in the page's dirty field.
1334  *
1335  *	The object containing the specified page must be locked if the
1336  *	call is made from the machine-independent layer.
1337  *
1338  *	See vm_page_clear_dirty_mask().
1339  *
1340  *	This function should only be called by vm_page_dirty().
1341  */
1342 void
1343 vm_page_dirty_KBI(vm_page_t m)
1344 {
1345 
1346 	/* Refer to this operation by its public name. */
1347 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1348 	    ("vm_page_dirty: page is invalid!"));
1349 	m->dirty = VM_PAGE_BITS_ALL;
1350 }
1351 
1352 /*
1353  *	vm_page_insert:		[ internal use only ]
1354  *
1355  *	Inserts the given mem entry into the object and object list.
1356  *
1357  *	The object must be locked.
1358  */
1359 int
1360 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1361 {
1362 	vm_page_t mpred;
1363 
1364 	VM_OBJECT_ASSERT_WLOCKED(object);
1365 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1366 	return (vm_page_insert_after(m, object, pindex, mpred));
1367 }
1368 
1369 /*
1370  *	vm_page_insert_after:
1371  *
1372  *	Inserts the page "m" into the specified object at offset "pindex".
1373  *
1374  *	The page "mpred" must immediately precede the offset "pindex" within
1375  *	the specified object.
1376  *
1377  *	The object must be locked.
1378  */
1379 static int
1380 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1381     vm_page_t mpred)
1382 {
1383 	vm_page_t msucc;
1384 
1385 	VM_OBJECT_ASSERT_WLOCKED(object);
1386 	KASSERT(m->object == NULL,
1387 	    ("vm_page_insert_after: page already inserted"));
1388 	if (mpred != NULL) {
1389 		KASSERT(mpred->object == object,
1390 		    ("vm_page_insert_after: object doesn't contain mpred"));
1391 		KASSERT(mpred->pindex < pindex,
1392 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1393 		msucc = TAILQ_NEXT(mpred, listq);
1394 	} else
1395 		msucc = TAILQ_FIRST(&object->memq);
1396 	if (msucc != NULL)
1397 		KASSERT(msucc->pindex > pindex,
1398 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1399 
1400 	/*
1401 	 * Record the object/offset pair in this page
1402 	 */
1403 	m->object = object;
1404 	m->pindex = pindex;
1405 
1406 	/*
1407 	 * Now link into the object's ordered list of backed pages.
1408 	 */
1409 	if (vm_radix_insert(&object->rtree, m)) {
1410 		m->object = NULL;
1411 		m->pindex = 0;
1412 		return (1);
1413 	}
1414 	vm_page_insert_radixdone(m, object, mpred);
1415 	return (0);
1416 }
1417 
1418 /*
1419  *	vm_page_insert_radixdone:
1420  *
1421  *	Complete page "m" insertion into the specified object after the
1422  *	radix trie hooking.
1423  *
1424  *	The page "mpred" must precede the offset "m->pindex" within the
1425  *	specified object.
1426  *
1427  *	The object must be locked.
1428  */
1429 static void
1430 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1431 {
1432 
1433 	VM_OBJECT_ASSERT_WLOCKED(object);
1434 	KASSERT(object != NULL && m->object == object,
1435 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1436 	if (mpred != NULL) {
1437 		KASSERT(mpred->object == object,
1438 		    ("vm_page_insert_after: object doesn't contain mpred"));
1439 		KASSERT(mpred->pindex < m->pindex,
1440 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1441 	}
1442 
1443 	if (mpred != NULL)
1444 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1445 	else
1446 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1447 
1448 	/*
1449 	 * Show that the object has one more resident page.
1450 	 */
1451 	object->resident_page_count++;
1452 
1453 	/*
1454 	 * Hold the vnode until the last page is released.
1455 	 */
1456 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1457 		vhold(object->handle);
1458 
1459 	/*
1460 	 * Since we are inserting a new and possibly dirty page,
1461 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1462 	 */
1463 	if (pmap_page_is_write_mapped(m))
1464 		vm_object_set_writeable_dirty(object);
1465 }
1466 
1467 /*
1468  *	vm_page_remove:
1469  *
1470  *	Removes the specified page from its containing object, but does not
1471  *	invalidate any backing storage.
1472  *
1473  *	The object must be locked.  The page must be locked if it is managed.
1474  */
1475 void
1476 vm_page_remove(vm_page_t m)
1477 {
1478 	vm_object_t object;
1479 	vm_page_t mrem;
1480 
1481 	if ((m->oflags & VPO_UNMANAGED) == 0)
1482 		vm_page_assert_locked(m);
1483 	if ((object = m->object) == NULL)
1484 		return;
1485 	VM_OBJECT_ASSERT_WLOCKED(object);
1486 	if (vm_page_xbusied(m))
1487 		vm_page_xunbusy_maybelocked(m);
1488 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1489 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1490 
1491 	/*
1492 	 * Now remove from the object's list of backed pages.
1493 	 */
1494 	TAILQ_REMOVE(&object->memq, m, listq);
1495 
1496 	/*
1497 	 * And show that the object has one fewer resident page.
1498 	 */
1499 	object->resident_page_count--;
1500 
1501 	/*
1502 	 * The vnode may now be recycled.
1503 	 */
1504 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1505 		vdrop(object->handle);
1506 
1507 	m->object = NULL;
1508 }
1509 
1510 /*
1511  *	vm_page_lookup:
1512  *
1513  *	Returns the page associated with the object/offset
1514  *	pair specified; if none is found, NULL is returned.
1515  *
1516  *	The object must be locked.
1517  */
1518 vm_page_t
1519 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1520 {
1521 
1522 	VM_OBJECT_ASSERT_LOCKED(object);
1523 	return (vm_radix_lookup(&object->rtree, pindex));
1524 }
1525 
1526 /*
1527  *	vm_page_find_least:
1528  *
1529  *	Returns the page associated with the object with least pindex
1530  *	greater than or equal to the parameter pindex, or NULL.
1531  *
1532  *	The object must be locked.
1533  */
1534 vm_page_t
1535 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1536 {
1537 	vm_page_t m;
1538 
1539 	VM_OBJECT_ASSERT_LOCKED(object);
1540 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1541 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1542 	return (m);
1543 }
1544 
1545 /*
1546  * Returns the given page's successor (by pindex) within the object if it is
1547  * resident; if none is found, NULL is returned.
1548  *
1549  * The object must be locked.
1550  */
1551 vm_page_t
1552 vm_page_next(vm_page_t m)
1553 {
1554 	vm_page_t next;
1555 
1556 	VM_OBJECT_ASSERT_LOCKED(m->object);
1557 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1558 		MPASS(next->object == m->object);
1559 		if (next->pindex != m->pindex + 1)
1560 			next = NULL;
1561 	}
1562 	return (next);
1563 }
1564 
1565 /*
1566  * Returns the given page's predecessor (by pindex) within the object if it is
1567  * resident; if none is found, NULL is returned.
1568  *
1569  * The object must be locked.
1570  */
1571 vm_page_t
1572 vm_page_prev(vm_page_t m)
1573 {
1574 	vm_page_t prev;
1575 
1576 	VM_OBJECT_ASSERT_LOCKED(m->object);
1577 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1578 		MPASS(prev->object == m->object);
1579 		if (prev->pindex != m->pindex - 1)
1580 			prev = NULL;
1581 	}
1582 	return (prev);
1583 }
1584 
1585 /*
1586  * Uses the page mnew as a replacement for an existing page at index
1587  * pindex which must be already present in the object.
1588  *
1589  * The existing page must not be on a paging queue.
1590  */
1591 vm_page_t
1592 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1593 {
1594 	vm_page_t mold;
1595 
1596 	VM_OBJECT_ASSERT_WLOCKED(object);
1597 	KASSERT(mnew->object == NULL,
1598 	    ("vm_page_replace: page %p already in object", mnew));
1599 	KASSERT(mnew->queue == PQ_NONE,
1600 	    ("vm_page_replace: new page %p is on a paging queue", mnew));
1601 
1602 	/*
1603 	 * This function mostly follows vm_page_insert() and
1604 	 * vm_page_remove() without the radix, object count and vnode
1605 	 * dance.  Double check such functions for more comments.
1606 	 */
1607 
1608 	mnew->object = object;
1609 	mnew->pindex = pindex;
1610 	mold = vm_radix_replace(&object->rtree, mnew);
1611 	KASSERT(mold->queue == PQ_NONE,
1612 	    ("vm_page_replace: old page %p is on a paging queue", mold));
1613 
1614 	/* Keep the resident page list in sorted order. */
1615 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1616 	TAILQ_REMOVE(&object->memq, mold, listq);
1617 
1618 	mold->object = NULL;
1619 	vm_page_xunbusy_maybelocked(mold);
1620 
1621 	/*
1622 	 * The object's resident_page_count does not change because we have
1623 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1624 	 */
1625 	if (pmap_page_is_write_mapped(mnew))
1626 		vm_object_set_writeable_dirty(object);
1627 	return (mold);
1628 }
1629 
1630 /*
1631  *	vm_page_rename:
1632  *
1633  *	Move the given memory entry from its
1634  *	current object to the specified target object/offset.
1635  *
1636  *	Note: swap associated with the page must be invalidated by the move.  We
1637  *	      have to do this for several reasons:  (1) we aren't freeing the
1638  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1639  *	      moving the page from object A to B, and will then later move
1640  *	      the backing store from A to B and we can't have a conflict.
1641  *
1642  *	Note: we *always* dirty the page.  It is necessary both for the
1643  *	      fact that we moved it, and because we may be invalidating
1644  *	      swap.
1645  *
1646  *	The objects must be locked.
1647  */
1648 int
1649 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1650 {
1651 	vm_page_t mpred;
1652 	vm_pindex_t opidx;
1653 
1654 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1655 
1656 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1657 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1658 	    ("vm_page_rename: pindex already renamed"));
1659 
1660 	/*
1661 	 * Create a custom version of vm_page_insert() which does not depend
1662 	 * by m_prev and can cheat on the implementation aspects of the
1663 	 * function.
1664 	 */
1665 	opidx = m->pindex;
1666 	m->pindex = new_pindex;
1667 	if (vm_radix_insert(&new_object->rtree, m)) {
1668 		m->pindex = opidx;
1669 		return (1);
1670 	}
1671 
1672 	/*
1673 	 * The operation cannot fail anymore.  The removal must happen before
1674 	 * the listq iterator is tainted.
1675 	 */
1676 	m->pindex = opidx;
1677 	vm_page_lock(m);
1678 	vm_page_remove(m);
1679 
1680 	/* Return back to the new pindex to complete vm_page_insert(). */
1681 	m->pindex = new_pindex;
1682 	m->object = new_object;
1683 	vm_page_unlock(m);
1684 	vm_page_insert_radixdone(m, new_object, mpred);
1685 	vm_page_dirty(m);
1686 	return (0);
1687 }
1688 
1689 /*
1690  *	vm_page_alloc:
1691  *
1692  *	Allocate and return a page that is associated with the specified
1693  *	object and offset pair.  By default, this page is exclusive busied.
1694  *
1695  *	The caller must always specify an allocation class.
1696  *
1697  *	allocation classes:
1698  *	VM_ALLOC_NORMAL		normal process request
1699  *	VM_ALLOC_SYSTEM		system *really* needs a page
1700  *	VM_ALLOC_INTERRUPT	interrupt time request
1701  *
1702  *	optional allocation flags:
1703  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1704  *				intends to allocate
1705  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1706  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1707  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1708  *				should not be exclusive busy
1709  *	VM_ALLOC_SBUSY		shared busy the allocated page
1710  *	VM_ALLOC_WIRED		wire the allocated page
1711  *	VM_ALLOC_ZERO		prefer a zeroed page
1712  */
1713 vm_page_t
1714 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1715 {
1716 
1717 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1718 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1719 }
1720 
1721 vm_page_t
1722 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1723     int req)
1724 {
1725 
1726 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1727 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1728 	    NULL));
1729 }
1730 
1731 /*
1732  * Allocate a page in the specified object with the given page index.  To
1733  * optimize insertion of the page into the object, the caller must also specifiy
1734  * the resident page in the object with largest index smaller than the given
1735  * page index, or NULL if no such page exists.
1736  */
1737 vm_page_t
1738 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1739     int req, vm_page_t mpred)
1740 {
1741 	struct vm_domainset_iter di;
1742 	vm_page_t m;
1743 	int domain;
1744 
1745 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1746 	do {
1747 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1748 		    mpred);
1749 		if (m != NULL)
1750 			break;
1751 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1752 
1753 	return (m);
1754 }
1755 
1756 /*
1757  * Returns true if the number of free pages exceeds the minimum
1758  * for the request class and false otherwise.
1759  */
1760 int
1761 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1762 {
1763 	u_int limit, old, new;
1764 
1765 	req = req & VM_ALLOC_CLASS_MASK;
1766 
1767 	/*
1768 	 * The page daemon is allowed to dig deeper into the free page list.
1769 	 */
1770 	if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1771 		req = VM_ALLOC_SYSTEM;
1772 	if (req == VM_ALLOC_INTERRUPT)
1773 		limit = 0;
1774 	else if (req == VM_ALLOC_SYSTEM)
1775 		limit = vmd->vmd_interrupt_free_min;
1776 	else
1777 		limit = vmd->vmd_free_reserved;
1778 
1779 	/*
1780 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1781 	 */
1782 	limit += npages;
1783 	old = vmd->vmd_free_count;
1784 	do {
1785 		if (old < limit)
1786 			return (0);
1787 		new = old - npages;
1788 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1789 
1790 	/* Wake the page daemon if we've crossed the threshold. */
1791 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1792 		pagedaemon_wakeup(vmd->vmd_domain);
1793 
1794 	/* Only update bitsets on transitions. */
1795 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1796 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1797 		vm_domain_set(vmd);
1798 
1799 	return (1);
1800 }
1801 
1802 vm_page_t
1803 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1804     int req, vm_page_t mpred)
1805 {
1806 	struct vm_domain *vmd;
1807 	vm_page_t m;
1808 	int flags;
1809 
1810 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1811 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1812 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1813 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1814 	    ("inconsistent object(%p)/req(%x)", object, req));
1815 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1816 	    ("Can't sleep and retry object insertion."));
1817 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1818 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1819 	    (uintmax_t)pindex));
1820 	if (object != NULL)
1821 		VM_OBJECT_ASSERT_WLOCKED(object);
1822 
1823 again:
1824 	m = NULL;
1825 #if VM_NRESERVLEVEL > 0
1826 	/*
1827 	 * Can we allocate the page from a reservation?
1828 	 */
1829 	if (vm_object_reserv(object) &&
1830 	    ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL ||
1831 	    (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) {
1832 		domain = vm_phys_domain(m);
1833 		vmd = VM_DOMAIN(domain);
1834 		goto found;
1835 	}
1836 #endif
1837 	vmd = VM_DOMAIN(domain);
1838 	if (object != NULL && vmd->vmd_pgcache != NULL) {
1839 		m = uma_zalloc(vmd->vmd_pgcache, M_NOWAIT);
1840 		if (m != NULL)
1841 			goto found;
1842 	}
1843 	if (vm_domain_allocate(vmd, req, 1)) {
1844 		/*
1845 		 * If not, allocate it from the free page queues.
1846 		 */
1847 		vm_domain_free_lock(vmd);
1848 		m = vm_phys_alloc_pages(domain, object != NULL ?
1849 		    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1850 		vm_domain_free_unlock(vmd);
1851 		if (m == NULL) {
1852 			vm_domain_freecnt_inc(vmd, 1);
1853 #if VM_NRESERVLEVEL > 0
1854 			if (vm_reserv_reclaim_inactive(domain))
1855 				goto again;
1856 #endif
1857 		}
1858 	}
1859 	if (m == NULL) {
1860 		/*
1861 		 * Not allocatable, give up.
1862 		 */
1863 		if (vm_domain_alloc_fail(vmd, object, req))
1864 			goto again;
1865 		return (NULL);
1866 	}
1867 
1868 	/*
1869 	 *  At this point we had better have found a good page.
1870 	 */
1871 	KASSERT(m != NULL, ("missing page"));
1872 
1873 found:
1874 	vm_page_dequeue(m);
1875 	vm_page_alloc_check(m);
1876 
1877 	/*
1878 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1879 	 */
1880 	flags = 0;
1881 	if ((req & VM_ALLOC_ZERO) != 0)
1882 		flags = PG_ZERO;
1883 	flags &= m->flags;
1884 	if ((req & VM_ALLOC_NODUMP) != 0)
1885 		flags |= PG_NODUMP;
1886 	m->flags = flags;
1887 	m->aflags = 0;
1888 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1889 	    VPO_UNMANAGED : 0;
1890 	m->busy_lock = VPB_UNBUSIED;
1891 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1892 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1893 	if ((req & VM_ALLOC_SBUSY) != 0)
1894 		m->busy_lock = VPB_SHARERS_WORD(1);
1895 	if (req & VM_ALLOC_WIRED) {
1896 		/*
1897 		 * The page lock is not required for wiring a page until that
1898 		 * page is inserted into the object.
1899 		 */
1900 		vm_wire_add(1);
1901 		m->wire_count = 1;
1902 	}
1903 	m->act_count = 0;
1904 
1905 	if (object != NULL) {
1906 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1907 			if (req & VM_ALLOC_WIRED) {
1908 				vm_wire_sub(1);
1909 				m->wire_count = 0;
1910 			}
1911 			KASSERT(m->object == NULL, ("page %p has object", m));
1912 			m->oflags = VPO_UNMANAGED;
1913 			m->busy_lock = VPB_UNBUSIED;
1914 			/* Don't change PG_ZERO. */
1915 			vm_page_free_toq(m);
1916 			if (req & VM_ALLOC_WAITFAIL) {
1917 				VM_OBJECT_WUNLOCK(object);
1918 				vm_radix_wait();
1919 				VM_OBJECT_WLOCK(object);
1920 			}
1921 			return (NULL);
1922 		}
1923 
1924 		/* Ignore device objects; the pager sets "memattr" for them. */
1925 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1926 		    (object->flags & OBJ_FICTITIOUS) == 0)
1927 			pmap_page_set_memattr(m, object->memattr);
1928 	} else
1929 		m->pindex = pindex;
1930 
1931 	return (m);
1932 }
1933 
1934 /*
1935  *	vm_page_alloc_contig:
1936  *
1937  *	Allocate a contiguous set of physical pages of the given size "npages"
1938  *	from the free lists.  All of the physical pages must be at or above
1939  *	the given physical address "low" and below the given physical address
1940  *	"high".  The given value "alignment" determines the alignment of the
1941  *	first physical page in the set.  If the given value "boundary" is
1942  *	non-zero, then the set of physical pages cannot cross any physical
1943  *	address boundary that is a multiple of that value.  Both "alignment"
1944  *	and "boundary" must be a power of two.
1945  *
1946  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1947  *	then the memory attribute setting for the physical pages is configured
1948  *	to the object's memory attribute setting.  Otherwise, the memory
1949  *	attribute setting for the physical pages is configured to "memattr",
1950  *	overriding the object's memory attribute setting.  However, if the
1951  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1952  *	memory attribute setting for the physical pages cannot be configured
1953  *	to VM_MEMATTR_DEFAULT.
1954  *
1955  *	The specified object may not contain fictitious pages.
1956  *
1957  *	The caller must always specify an allocation class.
1958  *
1959  *	allocation classes:
1960  *	VM_ALLOC_NORMAL		normal process request
1961  *	VM_ALLOC_SYSTEM		system *really* needs a page
1962  *	VM_ALLOC_INTERRUPT	interrupt time request
1963  *
1964  *	optional allocation flags:
1965  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1966  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1967  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1968  *				should not be exclusive busy
1969  *	VM_ALLOC_SBUSY		shared busy the allocated page
1970  *	VM_ALLOC_WIRED		wire the allocated page
1971  *	VM_ALLOC_ZERO		prefer a zeroed page
1972  */
1973 vm_page_t
1974 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1975     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1976     vm_paddr_t boundary, vm_memattr_t memattr)
1977 {
1978 	struct vm_domainset_iter di;
1979 	vm_page_t m;
1980 	int domain;
1981 
1982 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1983 	do {
1984 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1985 		    npages, low, high, alignment, boundary, memattr);
1986 		if (m != NULL)
1987 			break;
1988 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1989 
1990 	return (m);
1991 }
1992 
1993 vm_page_t
1994 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1995     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1996     vm_paddr_t boundary, vm_memattr_t memattr)
1997 {
1998 	struct vm_domain *vmd;
1999 	vm_page_t m, m_ret, mpred;
2000 	u_int busy_lock, flags, oflags;
2001 
2002 	mpred = NULL;	/* XXX: pacify gcc */
2003 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2004 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2005 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2006 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2007 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2008 	    req));
2009 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2010 	    ("Can't sleep and retry object insertion."));
2011 	if (object != NULL) {
2012 		VM_OBJECT_ASSERT_WLOCKED(object);
2013 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2014 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2015 		    object));
2016 	}
2017 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2018 
2019 	if (object != NULL) {
2020 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2021 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2022 		    ("vm_page_alloc_contig: pindex already allocated"));
2023 	}
2024 
2025 	/*
2026 	 * Can we allocate the pages without the number of free pages falling
2027 	 * below the lower bound for the allocation class?
2028 	 */
2029 again:
2030 #if VM_NRESERVLEVEL > 0
2031 	/*
2032 	 * Can we allocate the pages from a reservation?
2033 	 */
2034 	if (vm_object_reserv(object) &&
2035 	    ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
2036 	    npages, low, high, alignment, boundary, mpred)) != NULL ||
2037 	    (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain,
2038 	    npages, low, high, alignment, boundary, mpred)) != NULL)) {
2039 		domain = vm_phys_domain(m_ret);
2040 		vmd = VM_DOMAIN(domain);
2041 		goto found;
2042 	}
2043 #endif
2044 	m_ret = NULL;
2045 	vmd = VM_DOMAIN(domain);
2046 	if (vm_domain_allocate(vmd, req, npages)) {
2047 		/*
2048 		 * allocate them from the free page queues.
2049 		 */
2050 		vm_domain_free_lock(vmd);
2051 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2052 		    alignment, boundary);
2053 		vm_domain_free_unlock(vmd);
2054 		if (m_ret == NULL) {
2055 			vm_domain_freecnt_inc(vmd, npages);
2056 #if VM_NRESERVLEVEL > 0
2057 			if (vm_reserv_reclaim_contig(domain, npages, low,
2058 			    high, alignment, boundary))
2059 				goto again;
2060 #endif
2061 		}
2062 	}
2063 	if (m_ret == NULL) {
2064 		if (vm_domain_alloc_fail(vmd, object, req))
2065 			goto again;
2066 		return (NULL);
2067 	}
2068 #if VM_NRESERVLEVEL > 0
2069 found:
2070 #endif
2071 	for (m = m_ret; m < &m_ret[npages]; m++) {
2072 		vm_page_dequeue(m);
2073 		vm_page_alloc_check(m);
2074 	}
2075 
2076 	/*
2077 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2078 	 */
2079 	flags = 0;
2080 	if ((req & VM_ALLOC_ZERO) != 0)
2081 		flags = PG_ZERO;
2082 	if ((req & VM_ALLOC_NODUMP) != 0)
2083 		flags |= PG_NODUMP;
2084 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2085 	    VPO_UNMANAGED : 0;
2086 	busy_lock = VPB_UNBUSIED;
2087 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2088 		busy_lock = VPB_SINGLE_EXCLUSIVER;
2089 	if ((req & VM_ALLOC_SBUSY) != 0)
2090 		busy_lock = VPB_SHARERS_WORD(1);
2091 	if ((req & VM_ALLOC_WIRED) != 0)
2092 		vm_wire_add(npages);
2093 	if (object != NULL) {
2094 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2095 		    memattr == VM_MEMATTR_DEFAULT)
2096 			memattr = object->memattr;
2097 	}
2098 	for (m = m_ret; m < &m_ret[npages]; m++) {
2099 		m->aflags = 0;
2100 		m->flags = (m->flags | PG_NODUMP) & flags;
2101 		m->busy_lock = busy_lock;
2102 		if ((req & VM_ALLOC_WIRED) != 0)
2103 			m->wire_count = 1;
2104 		m->act_count = 0;
2105 		m->oflags = oflags;
2106 		if (object != NULL) {
2107 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2108 				if ((req & VM_ALLOC_WIRED) != 0)
2109 					vm_wire_sub(npages);
2110 				KASSERT(m->object == NULL,
2111 				    ("page %p has object", m));
2112 				mpred = m;
2113 				for (m = m_ret; m < &m_ret[npages]; m++) {
2114 					if (m <= mpred &&
2115 					    (req & VM_ALLOC_WIRED) != 0)
2116 						m->wire_count = 0;
2117 					m->oflags = VPO_UNMANAGED;
2118 					m->busy_lock = VPB_UNBUSIED;
2119 					/* Don't change PG_ZERO. */
2120 					vm_page_free_toq(m);
2121 				}
2122 				if (req & VM_ALLOC_WAITFAIL) {
2123 					VM_OBJECT_WUNLOCK(object);
2124 					vm_radix_wait();
2125 					VM_OBJECT_WLOCK(object);
2126 				}
2127 				return (NULL);
2128 			}
2129 			mpred = m;
2130 		} else
2131 			m->pindex = pindex;
2132 		if (memattr != VM_MEMATTR_DEFAULT)
2133 			pmap_page_set_memattr(m, memattr);
2134 		pindex++;
2135 	}
2136 	return (m_ret);
2137 }
2138 
2139 /*
2140  * Check a page that has been freshly dequeued from a freelist.
2141  */
2142 static void
2143 vm_page_alloc_check(vm_page_t m)
2144 {
2145 
2146 	KASSERT(m->object == NULL, ("page %p has object", m));
2147 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2148 	    ("page %p has unexpected queue %d, flags %#x",
2149 	    m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2150 	KASSERT(!vm_page_held(m), ("page %p is held", m));
2151 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2152 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2153 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2154 	    ("page %p has unexpected memattr %d",
2155 	    m, pmap_page_get_memattr(m)));
2156 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2157 }
2158 
2159 /*
2160  * 	vm_page_alloc_freelist:
2161  *
2162  *	Allocate a physical page from the specified free page list.
2163  *
2164  *	The caller must always specify an allocation class.
2165  *
2166  *	allocation classes:
2167  *	VM_ALLOC_NORMAL		normal process request
2168  *	VM_ALLOC_SYSTEM		system *really* needs a page
2169  *	VM_ALLOC_INTERRUPT	interrupt time request
2170  *
2171  *	optional allocation flags:
2172  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2173  *				intends to allocate
2174  *	VM_ALLOC_WIRED		wire the allocated page
2175  *	VM_ALLOC_ZERO		prefer a zeroed page
2176  */
2177 vm_page_t
2178 vm_page_alloc_freelist(int freelist, int req)
2179 {
2180 	struct vm_domainset_iter di;
2181 	vm_page_t m;
2182 	int domain;
2183 
2184 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2185 	do {
2186 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2187 		if (m != NULL)
2188 			break;
2189 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2190 
2191 	return (m);
2192 }
2193 
2194 vm_page_t
2195 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2196 {
2197 	struct vm_domain *vmd;
2198 	vm_page_t m;
2199 	u_int flags;
2200 
2201 	m = NULL;
2202 	vmd = VM_DOMAIN(domain);
2203 again:
2204 	if (vm_domain_allocate(vmd, req, 1)) {
2205 		vm_domain_free_lock(vmd);
2206 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2207 		    VM_FREEPOOL_DIRECT, 0);
2208 		vm_domain_free_unlock(vmd);
2209 		if (m == NULL)
2210 			vm_domain_freecnt_inc(vmd, 1);
2211 	}
2212 	if (m == NULL) {
2213 		if (vm_domain_alloc_fail(vmd, NULL, req))
2214 			goto again;
2215 		return (NULL);
2216 	}
2217 	vm_page_dequeue(m);
2218 	vm_page_alloc_check(m);
2219 
2220 	/*
2221 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2222 	 */
2223 	m->aflags = 0;
2224 	flags = 0;
2225 	if ((req & VM_ALLOC_ZERO) != 0)
2226 		flags = PG_ZERO;
2227 	m->flags &= flags;
2228 	if ((req & VM_ALLOC_WIRED) != 0) {
2229 		/*
2230 		 * The page lock is not required for wiring a page that does
2231 		 * not belong to an object.
2232 		 */
2233 		vm_wire_add(1);
2234 		m->wire_count = 1;
2235 	}
2236 	/* Unmanaged pages don't use "act_count". */
2237 	m->oflags = VPO_UNMANAGED;
2238 	return (m);
2239 }
2240 
2241 static int
2242 vm_page_import(void *arg, void **store, int cnt, int domain, int flags)
2243 {
2244 	struct vm_domain *vmd;
2245 	int i;
2246 
2247 	vmd = arg;
2248 	/* Only import if we can bring in a full bucket. */
2249 	if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2250 		return (0);
2251 	domain = vmd->vmd_domain;
2252 	vm_domain_free_lock(vmd);
2253 	i = vm_phys_alloc_npages(domain, VM_FREEPOOL_DEFAULT, cnt,
2254 	    (vm_page_t *)store);
2255 	vm_domain_free_unlock(vmd);
2256 	if (cnt != i)
2257 		vm_domain_freecnt_inc(vmd, cnt - i);
2258 
2259 	return (i);
2260 }
2261 
2262 static void
2263 vm_page_release(void *arg, void **store, int cnt)
2264 {
2265 	struct vm_domain *vmd;
2266 	vm_page_t m;
2267 	int i;
2268 
2269 	vmd = arg;
2270 	vm_domain_free_lock(vmd);
2271 	for (i = 0; i < cnt; i++) {
2272 		m = (vm_page_t)store[i];
2273 		vm_phys_free_pages(m, 0);
2274 	}
2275 	vm_domain_free_unlock(vmd);
2276 	vm_domain_freecnt_inc(vmd, cnt);
2277 }
2278 
2279 #define	VPSC_ANY	0	/* No restrictions. */
2280 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2281 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2282 
2283 /*
2284  *	vm_page_scan_contig:
2285  *
2286  *	Scan vm_page_array[] between the specified entries "m_start" and
2287  *	"m_end" for a run of contiguous physical pages that satisfy the
2288  *	specified conditions, and return the lowest page in the run.  The
2289  *	specified "alignment" determines the alignment of the lowest physical
2290  *	page in the run.  If the specified "boundary" is non-zero, then the
2291  *	run of physical pages cannot span a physical address that is a
2292  *	multiple of "boundary".
2293  *
2294  *	"m_end" is never dereferenced, so it need not point to a vm_page
2295  *	structure within vm_page_array[].
2296  *
2297  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2298  *	span a hole (or discontiguity) in the physical address space.  Both
2299  *	"alignment" and "boundary" must be a power of two.
2300  */
2301 vm_page_t
2302 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2303     u_long alignment, vm_paddr_t boundary, int options)
2304 {
2305 	struct mtx *m_mtx;
2306 	vm_object_t object;
2307 	vm_paddr_t pa;
2308 	vm_page_t m, m_run;
2309 #if VM_NRESERVLEVEL > 0
2310 	int level;
2311 #endif
2312 	int m_inc, order, run_ext, run_len;
2313 
2314 	KASSERT(npages > 0, ("npages is 0"));
2315 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2316 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2317 	m_run = NULL;
2318 	run_len = 0;
2319 	m_mtx = NULL;
2320 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2321 		KASSERT((m->flags & PG_MARKER) == 0,
2322 		    ("page %p is PG_MARKER", m));
2323 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2324 		    ("fictitious page %p has invalid wire count", m));
2325 
2326 		/*
2327 		 * If the current page would be the start of a run, check its
2328 		 * physical address against the end, alignment, and boundary
2329 		 * conditions.  If it doesn't satisfy these conditions, either
2330 		 * terminate the scan or advance to the next page that
2331 		 * satisfies the failed condition.
2332 		 */
2333 		if (run_len == 0) {
2334 			KASSERT(m_run == NULL, ("m_run != NULL"));
2335 			if (m + npages > m_end)
2336 				break;
2337 			pa = VM_PAGE_TO_PHYS(m);
2338 			if ((pa & (alignment - 1)) != 0) {
2339 				m_inc = atop(roundup2(pa, alignment) - pa);
2340 				continue;
2341 			}
2342 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2343 			    boundary) != 0) {
2344 				m_inc = atop(roundup2(pa, boundary) - pa);
2345 				continue;
2346 			}
2347 		} else
2348 			KASSERT(m_run != NULL, ("m_run == NULL"));
2349 
2350 		vm_page_change_lock(m, &m_mtx);
2351 		m_inc = 1;
2352 retry:
2353 		if (vm_page_held(m))
2354 			run_ext = 0;
2355 #if VM_NRESERVLEVEL > 0
2356 		else if ((level = vm_reserv_level(m)) >= 0 &&
2357 		    (options & VPSC_NORESERV) != 0) {
2358 			run_ext = 0;
2359 			/* Advance to the end of the reservation. */
2360 			pa = VM_PAGE_TO_PHYS(m);
2361 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2362 			    pa);
2363 		}
2364 #endif
2365 		else if ((object = m->object) != NULL) {
2366 			/*
2367 			 * The page is considered eligible for relocation if
2368 			 * and only if it could be laundered or reclaimed by
2369 			 * the page daemon.
2370 			 */
2371 			if (!VM_OBJECT_TRYRLOCK(object)) {
2372 				mtx_unlock(m_mtx);
2373 				VM_OBJECT_RLOCK(object);
2374 				mtx_lock(m_mtx);
2375 				if (m->object != object) {
2376 					/*
2377 					 * The page may have been freed.
2378 					 */
2379 					VM_OBJECT_RUNLOCK(object);
2380 					goto retry;
2381 				} else if (vm_page_held(m)) {
2382 					run_ext = 0;
2383 					goto unlock;
2384 				}
2385 			}
2386 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2387 			    ("page %p is PG_UNHOLDFREE", m));
2388 			/* Don't care: PG_NODUMP, PG_ZERO. */
2389 			if (object->type != OBJT_DEFAULT &&
2390 			    object->type != OBJT_SWAP &&
2391 			    object->type != OBJT_VNODE) {
2392 				run_ext = 0;
2393 #if VM_NRESERVLEVEL > 0
2394 			} else if ((options & VPSC_NOSUPER) != 0 &&
2395 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2396 				run_ext = 0;
2397 				/* Advance to the end of the superpage. */
2398 				pa = VM_PAGE_TO_PHYS(m);
2399 				m_inc = atop(roundup2(pa + 1,
2400 				    vm_reserv_size(level)) - pa);
2401 #endif
2402 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2403 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2404 				/*
2405 				 * The page is allocated but eligible for
2406 				 * relocation.  Extend the current run by one
2407 				 * page.
2408 				 */
2409 				KASSERT(pmap_page_get_memattr(m) ==
2410 				    VM_MEMATTR_DEFAULT,
2411 				    ("page %p has an unexpected memattr", m));
2412 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2413 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2414 				    ("page %p has unexpected oflags", m));
2415 				/* Don't care: VPO_NOSYNC. */
2416 				run_ext = 1;
2417 			} else
2418 				run_ext = 0;
2419 unlock:
2420 			VM_OBJECT_RUNLOCK(object);
2421 #if VM_NRESERVLEVEL > 0
2422 		} else if (level >= 0) {
2423 			/*
2424 			 * The page is reserved but not yet allocated.  In
2425 			 * other words, it is still free.  Extend the current
2426 			 * run by one page.
2427 			 */
2428 			run_ext = 1;
2429 #endif
2430 		} else if ((order = m->order) < VM_NFREEORDER) {
2431 			/*
2432 			 * The page is enqueued in the physical memory
2433 			 * allocator's free page queues.  Moreover, it is the
2434 			 * first page in a power-of-two-sized run of
2435 			 * contiguous free pages.  Add these pages to the end
2436 			 * of the current run, and jump ahead.
2437 			 */
2438 			run_ext = 1 << order;
2439 			m_inc = 1 << order;
2440 		} else {
2441 			/*
2442 			 * Skip the page for one of the following reasons: (1)
2443 			 * It is enqueued in the physical memory allocator's
2444 			 * free page queues.  However, it is not the first
2445 			 * page in a run of contiguous free pages.  (This case
2446 			 * rarely occurs because the scan is performed in
2447 			 * ascending order.) (2) It is not reserved, and it is
2448 			 * transitioning from free to allocated.  (Conversely,
2449 			 * the transition from allocated to free for managed
2450 			 * pages is blocked by the page lock.) (3) It is
2451 			 * allocated but not contained by an object and not
2452 			 * wired, e.g., allocated by Xen's balloon driver.
2453 			 */
2454 			run_ext = 0;
2455 		}
2456 
2457 		/*
2458 		 * Extend or reset the current run of pages.
2459 		 */
2460 		if (run_ext > 0) {
2461 			if (run_len == 0)
2462 				m_run = m;
2463 			run_len += run_ext;
2464 		} else {
2465 			if (run_len > 0) {
2466 				m_run = NULL;
2467 				run_len = 0;
2468 			}
2469 		}
2470 	}
2471 	if (m_mtx != NULL)
2472 		mtx_unlock(m_mtx);
2473 	if (run_len >= npages)
2474 		return (m_run);
2475 	return (NULL);
2476 }
2477 
2478 /*
2479  *	vm_page_reclaim_run:
2480  *
2481  *	Try to relocate each of the allocated virtual pages within the
2482  *	specified run of physical pages to a new physical address.  Free the
2483  *	physical pages underlying the relocated virtual pages.  A virtual page
2484  *	is relocatable if and only if it could be laundered or reclaimed by
2485  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2486  *	physical address above "high".
2487  *
2488  *	Returns 0 if every physical page within the run was already free or
2489  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2490  *	value indicating why the last attempt to relocate a virtual page was
2491  *	unsuccessful.
2492  *
2493  *	"req_class" must be an allocation class.
2494  */
2495 static int
2496 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2497     vm_paddr_t high)
2498 {
2499 	struct vm_domain *vmd;
2500 	struct mtx *m_mtx;
2501 	struct spglist free;
2502 	vm_object_t object;
2503 	vm_paddr_t pa;
2504 	vm_page_t m, m_end, m_new;
2505 	int error, order, req;
2506 
2507 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2508 	    ("req_class is not an allocation class"));
2509 	SLIST_INIT(&free);
2510 	error = 0;
2511 	m = m_run;
2512 	m_end = m_run + npages;
2513 	m_mtx = NULL;
2514 	for (; error == 0 && m < m_end; m++) {
2515 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2516 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2517 
2518 		/*
2519 		 * Avoid releasing and reacquiring the same page lock.
2520 		 */
2521 		vm_page_change_lock(m, &m_mtx);
2522 retry:
2523 		if (vm_page_held(m))
2524 			error = EBUSY;
2525 		else if ((object = m->object) != NULL) {
2526 			/*
2527 			 * The page is relocated if and only if it could be
2528 			 * laundered or reclaimed by the page daemon.
2529 			 */
2530 			if (!VM_OBJECT_TRYWLOCK(object)) {
2531 				mtx_unlock(m_mtx);
2532 				VM_OBJECT_WLOCK(object);
2533 				mtx_lock(m_mtx);
2534 				if (m->object != object) {
2535 					/*
2536 					 * The page may have been freed.
2537 					 */
2538 					VM_OBJECT_WUNLOCK(object);
2539 					goto retry;
2540 				} else if (vm_page_held(m)) {
2541 					error = EBUSY;
2542 					goto unlock;
2543 				}
2544 			}
2545 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2546 			    ("page %p is PG_UNHOLDFREE", m));
2547 			/* Don't care: PG_NODUMP, PG_ZERO. */
2548 			if (object->type != OBJT_DEFAULT &&
2549 			    object->type != OBJT_SWAP &&
2550 			    object->type != OBJT_VNODE)
2551 				error = EINVAL;
2552 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2553 				error = EINVAL;
2554 			else if (vm_page_queue(m) != PQ_NONE &&
2555 			    !vm_page_busied(m)) {
2556 				KASSERT(pmap_page_get_memattr(m) ==
2557 				    VM_MEMATTR_DEFAULT,
2558 				    ("page %p has an unexpected memattr", m));
2559 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2560 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2561 				    ("page %p has unexpected oflags", m));
2562 				/* Don't care: VPO_NOSYNC. */
2563 				if (m->valid != 0) {
2564 					/*
2565 					 * First, try to allocate a new page
2566 					 * that is above "high".  Failing
2567 					 * that, try to allocate a new page
2568 					 * that is below "m_run".  Allocate
2569 					 * the new page between the end of
2570 					 * "m_run" and "high" only as a last
2571 					 * resort.
2572 					 */
2573 					req = req_class | VM_ALLOC_NOOBJ;
2574 					if ((m->flags & PG_NODUMP) != 0)
2575 						req |= VM_ALLOC_NODUMP;
2576 					if (trunc_page(high) !=
2577 					    ~(vm_paddr_t)PAGE_MASK) {
2578 						m_new = vm_page_alloc_contig(
2579 						    NULL, 0, req, 1,
2580 						    round_page(high),
2581 						    ~(vm_paddr_t)0,
2582 						    PAGE_SIZE, 0,
2583 						    VM_MEMATTR_DEFAULT);
2584 					} else
2585 						m_new = NULL;
2586 					if (m_new == NULL) {
2587 						pa = VM_PAGE_TO_PHYS(m_run);
2588 						m_new = vm_page_alloc_contig(
2589 						    NULL, 0, req, 1,
2590 						    0, pa - 1, PAGE_SIZE, 0,
2591 						    VM_MEMATTR_DEFAULT);
2592 					}
2593 					if (m_new == NULL) {
2594 						pa += ptoa(npages);
2595 						m_new = vm_page_alloc_contig(
2596 						    NULL, 0, req, 1,
2597 						    pa, high, PAGE_SIZE, 0,
2598 						    VM_MEMATTR_DEFAULT);
2599 					}
2600 					if (m_new == NULL) {
2601 						error = ENOMEM;
2602 						goto unlock;
2603 					}
2604 					KASSERT(m_new->wire_count == 0,
2605 					    ("page %p is wired", m_new));
2606 
2607 					/*
2608 					 * Replace "m" with the new page.  For
2609 					 * vm_page_replace(), "m" must be busy
2610 					 * and dequeued.  Finally, change "m"
2611 					 * as if vm_page_free() was called.
2612 					 */
2613 					if (object->ref_count != 0)
2614 						pmap_remove_all(m);
2615 					m_new->aflags = m->aflags &
2616 					    ~PGA_QUEUE_STATE_MASK;
2617 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2618 					    ("page %p is managed", m_new));
2619 					m_new->oflags = m->oflags & VPO_NOSYNC;
2620 					pmap_copy_page(m, m_new);
2621 					m_new->valid = m->valid;
2622 					m_new->dirty = m->dirty;
2623 					m->flags &= ~PG_ZERO;
2624 					vm_page_xbusy(m);
2625 					vm_page_dequeue(m);
2626 					vm_page_replace_checked(m_new, object,
2627 					    m->pindex, m);
2628 					if (vm_page_free_prep(m))
2629 						SLIST_INSERT_HEAD(&free, m,
2630 						    plinks.s.ss);
2631 
2632 					/*
2633 					 * The new page must be deactivated
2634 					 * before the object is unlocked.
2635 					 */
2636 					vm_page_change_lock(m_new, &m_mtx);
2637 					vm_page_deactivate(m_new);
2638 				} else {
2639 					m->flags &= ~PG_ZERO;
2640 					vm_page_dequeue(m);
2641 					vm_page_remove(m);
2642 					if (vm_page_free_prep(m))
2643 						SLIST_INSERT_HEAD(&free, m,
2644 						    plinks.s.ss);
2645 					KASSERT(m->dirty == 0,
2646 					    ("page %p is dirty", m));
2647 				}
2648 			} else
2649 				error = EBUSY;
2650 unlock:
2651 			VM_OBJECT_WUNLOCK(object);
2652 		} else {
2653 			MPASS(vm_phys_domain(m) == domain);
2654 			vmd = VM_DOMAIN(domain);
2655 			vm_domain_free_lock(vmd);
2656 			order = m->order;
2657 			if (order < VM_NFREEORDER) {
2658 				/*
2659 				 * The page is enqueued in the physical memory
2660 				 * allocator's free page queues.  Moreover, it
2661 				 * is the first page in a power-of-two-sized
2662 				 * run of contiguous free pages.  Jump ahead
2663 				 * to the last page within that run, and
2664 				 * continue from there.
2665 				 */
2666 				m += (1 << order) - 1;
2667 			}
2668 #if VM_NRESERVLEVEL > 0
2669 			else if (vm_reserv_is_page_free(m))
2670 				order = 0;
2671 #endif
2672 			vm_domain_free_unlock(vmd);
2673 			if (order == VM_NFREEORDER)
2674 				error = EINVAL;
2675 		}
2676 	}
2677 	if (m_mtx != NULL)
2678 		mtx_unlock(m_mtx);
2679 	if ((m = SLIST_FIRST(&free)) != NULL) {
2680 		int cnt;
2681 
2682 		vmd = VM_DOMAIN(domain);
2683 		cnt = 0;
2684 		vm_domain_free_lock(vmd);
2685 		do {
2686 			MPASS(vm_phys_domain(m) == domain);
2687 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2688 			vm_phys_free_pages(m, 0);
2689 			cnt++;
2690 		} while ((m = SLIST_FIRST(&free)) != NULL);
2691 		vm_domain_free_unlock(vmd);
2692 		vm_domain_freecnt_inc(vmd, cnt);
2693 	}
2694 	return (error);
2695 }
2696 
2697 #define	NRUNS	16
2698 
2699 CTASSERT(powerof2(NRUNS));
2700 
2701 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2702 
2703 #define	MIN_RECLAIM	8
2704 
2705 /*
2706  *	vm_page_reclaim_contig:
2707  *
2708  *	Reclaim allocated, contiguous physical memory satisfying the specified
2709  *	conditions by relocating the virtual pages using that physical memory.
2710  *	Returns true if reclamation is successful and false otherwise.  Since
2711  *	relocation requires the allocation of physical pages, reclamation may
2712  *	fail due to a shortage of free pages.  When reclamation fails, callers
2713  *	are expected to perform vm_wait() before retrying a failed allocation
2714  *	operation, e.g., vm_page_alloc_contig().
2715  *
2716  *	The caller must always specify an allocation class through "req".
2717  *
2718  *	allocation classes:
2719  *	VM_ALLOC_NORMAL		normal process request
2720  *	VM_ALLOC_SYSTEM		system *really* needs a page
2721  *	VM_ALLOC_INTERRUPT	interrupt time request
2722  *
2723  *	The optional allocation flags are ignored.
2724  *
2725  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2726  *	must be a power of two.
2727  */
2728 bool
2729 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2730     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2731 {
2732 	struct vm_domain *vmd;
2733 	vm_paddr_t curr_low;
2734 	vm_page_t m_run, m_runs[NRUNS];
2735 	u_long count, reclaimed;
2736 	int error, i, options, req_class;
2737 
2738 	KASSERT(npages > 0, ("npages is 0"));
2739 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2740 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2741 	req_class = req & VM_ALLOC_CLASS_MASK;
2742 
2743 	/*
2744 	 * The page daemon is allowed to dig deeper into the free page list.
2745 	 */
2746 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2747 		req_class = VM_ALLOC_SYSTEM;
2748 
2749 	/*
2750 	 * Return if the number of free pages cannot satisfy the requested
2751 	 * allocation.
2752 	 */
2753 	vmd = VM_DOMAIN(domain);
2754 	count = vmd->vmd_free_count;
2755 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2756 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2757 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2758 		return (false);
2759 
2760 	/*
2761 	 * Scan up to three times, relaxing the restrictions ("options") on
2762 	 * the reclamation of reservations and superpages each time.
2763 	 */
2764 	for (options = VPSC_NORESERV;;) {
2765 		/*
2766 		 * Find the highest runs that satisfy the given constraints
2767 		 * and restrictions, and record them in "m_runs".
2768 		 */
2769 		curr_low = low;
2770 		count = 0;
2771 		for (;;) {
2772 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2773 			    high, alignment, boundary, options);
2774 			if (m_run == NULL)
2775 				break;
2776 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2777 			m_runs[RUN_INDEX(count)] = m_run;
2778 			count++;
2779 		}
2780 
2781 		/*
2782 		 * Reclaim the highest runs in LIFO (descending) order until
2783 		 * the number of reclaimed pages, "reclaimed", is at least
2784 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2785 		 * reclamation is idempotent, and runs will (likely) recur
2786 		 * from one scan to the next as restrictions are relaxed.
2787 		 */
2788 		reclaimed = 0;
2789 		for (i = 0; count > 0 && i < NRUNS; i++) {
2790 			count--;
2791 			m_run = m_runs[RUN_INDEX(count)];
2792 			error = vm_page_reclaim_run(req_class, domain, npages,
2793 			    m_run, high);
2794 			if (error == 0) {
2795 				reclaimed += npages;
2796 				if (reclaimed >= MIN_RECLAIM)
2797 					return (true);
2798 			}
2799 		}
2800 
2801 		/*
2802 		 * Either relax the restrictions on the next scan or return if
2803 		 * the last scan had no restrictions.
2804 		 */
2805 		if (options == VPSC_NORESERV)
2806 			options = VPSC_NOSUPER;
2807 		else if (options == VPSC_NOSUPER)
2808 			options = VPSC_ANY;
2809 		else if (options == VPSC_ANY)
2810 			return (reclaimed != 0);
2811 	}
2812 }
2813 
2814 bool
2815 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2816     u_long alignment, vm_paddr_t boundary)
2817 {
2818 	struct vm_domainset_iter di;
2819 	int domain;
2820 	bool ret;
2821 
2822 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2823 	do {
2824 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2825 		    high, alignment, boundary);
2826 		if (ret)
2827 			break;
2828 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2829 
2830 	return (ret);
2831 }
2832 
2833 /*
2834  * Set the domain in the appropriate page level domainset.
2835  */
2836 void
2837 vm_domain_set(struct vm_domain *vmd)
2838 {
2839 
2840 	mtx_lock(&vm_domainset_lock);
2841 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2842 		vmd->vmd_minset = 1;
2843 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2844 	}
2845 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2846 		vmd->vmd_severeset = 1;
2847 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2848 	}
2849 	mtx_unlock(&vm_domainset_lock);
2850 }
2851 
2852 /*
2853  * Clear the domain from the appropriate page level domainset.
2854  */
2855 void
2856 vm_domain_clear(struct vm_domain *vmd)
2857 {
2858 
2859 	mtx_lock(&vm_domainset_lock);
2860 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2861 		vmd->vmd_minset = 0;
2862 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2863 		if (vm_min_waiters != 0) {
2864 			vm_min_waiters = 0;
2865 			wakeup(&vm_min_domains);
2866 		}
2867 	}
2868 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2869 		vmd->vmd_severeset = 0;
2870 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2871 		if (vm_severe_waiters != 0) {
2872 			vm_severe_waiters = 0;
2873 			wakeup(&vm_severe_domains);
2874 		}
2875 	}
2876 
2877 	/*
2878 	 * If pageout daemon needs pages, then tell it that there are
2879 	 * some free.
2880 	 */
2881 	if (vmd->vmd_pageout_pages_needed &&
2882 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2883 		wakeup(&vmd->vmd_pageout_pages_needed);
2884 		vmd->vmd_pageout_pages_needed = 0;
2885 	}
2886 
2887 	/* See comments in vm_wait_doms(). */
2888 	if (vm_pageproc_waiters) {
2889 		vm_pageproc_waiters = 0;
2890 		wakeup(&vm_pageproc_waiters);
2891 	}
2892 	mtx_unlock(&vm_domainset_lock);
2893 }
2894 
2895 /*
2896  * Wait for free pages to exceed the min threshold globally.
2897  */
2898 void
2899 vm_wait_min(void)
2900 {
2901 
2902 	mtx_lock(&vm_domainset_lock);
2903 	while (vm_page_count_min()) {
2904 		vm_min_waiters++;
2905 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2906 	}
2907 	mtx_unlock(&vm_domainset_lock);
2908 }
2909 
2910 /*
2911  * Wait for free pages to exceed the severe threshold globally.
2912  */
2913 void
2914 vm_wait_severe(void)
2915 {
2916 
2917 	mtx_lock(&vm_domainset_lock);
2918 	while (vm_page_count_severe()) {
2919 		vm_severe_waiters++;
2920 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2921 		    "vmwait", 0);
2922 	}
2923 	mtx_unlock(&vm_domainset_lock);
2924 }
2925 
2926 u_int
2927 vm_wait_count(void)
2928 {
2929 
2930 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2931 }
2932 
2933 void
2934 vm_wait_doms(const domainset_t *wdoms)
2935 {
2936 
2937 	/*
2938 	 * We use racey wakeup synchronization to avoid expensive global
2939 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
2940 	 * To handle this, we only sleep for one tick in this instance.  It
2941 	 * is expected that most allocations for the pageproc will come from
2942 	 * kmem or vm_page_grab* which will use the more specific and
2943 	 * race-free vm_wait_domain().
2944 	 */
2945 	if (curproc == pageproc) {
2946 		mtx_lock(&vm_domainset_lock);
2947 		vm_pageproc_waiters++;
2948 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2949 		    "pageprocwait", 1);
2950 	} else {
2951 		/*
2952 		 * XXX Ideally we would wait only until the allocation could
2953 		 * be satisfied.  This condition can cause new allocators to
2954 		 * consume all freed pages while old allocators wait.
2955 		 */
2956 		mtx_lock(&vm_domainset_lock);
2957 		if (vm_page_count_min_set(wdoms)) {
2958 			vm_min_waiters++;
2959 			msleep(&vm_min_domains, &vm_domainset_lock,
2960 			    PVM | PDROP, "vmwait", 0);
2961 		} else
2962 			mtx_unlock(&vm_domainset_lock);
2963 	}
2964 }
2965 
2966 /*
2967  *	vm_wait_domain:
2968  *
2969  *	Sleep until free pages are available for allocation.
2970  *	- Called in various places after failed memory allocations.
2971  */
2972 void
2973 vm_wait_domain(int domain)
2974 {
2975 	struct vm_domain *vmd;
2976 	domainset_t wdom;
2977 
2978 	vmd = VM_DOMAIN(domain);
2979 	vm_domain_free_assert_unlocked(vmd);
2980 
2981 	if (curproc == pageproc) {
2982 		mtx_lock(&vm_domainset_lock);
2983 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2984 			vmd->vmd_pageout_pages_needed = 1;
2985 			msleep(&vmd->vmd_pageout_pages_needed,
2986 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
2987 		} else
2988 			mtx_unlock(&vm_domainset_lock);
2989 	} else {
2990 		if (pageproc == NULL)
2991 			panic("vm_wait in early boot");
2992 		DOMAINSET_ZERO(&wdom);
2993 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
2994 		vm_wait_doms(&wdom);
2995 	}
2996 }
2997 
2998 /*
2999  *	vm_wait:
3000  *
3001  *	Sleep until free pages are available for allocation in the
3002  *	affinity domains of the obj.  If obj is NULL, the domain set
3003  *	for the calling thread is used.
3004  *	Called in various places after failed memory allocations.
3005  */
3006 void
3007 vm_wait(vm_object_t obj)
3008 {
3009 	struct domainset *d;
3010 
3011 	d = NULL;
3012 
3013 	/*
3014 	 * Carefully fetch pointers only once: the struct domainset
3015 	 * itself is ummutable but the pointer might change.
3016 	 */
3017 	if (obj != NULL)
3018 		d = obj->domain.dr_policy;
3019 	if (d == NULL)
3020 		d = curthread->td_domain.dr_policy;
3021 
3022 	vm_wait_doms(&d->ds_mask);
3023 }
3024 
3025 /*
3026  *	vm_domain_alloc_fail:
3027  *
3028  *	Called when a page allocation function fails.  Informs the
3029  *	pagedaemon and performs the requested wait.  Requires the
3030  *	domain_free and object lock on entry.  Returns with the
3031  *	object lock held and free lock released.  Returns an error when
3032  *	retry is necessary.
3033  *
3034  */
3035 static int
3036 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3037 {
3038 
3039 	vm_domain_free_assert_unlocked(vmd);
3040 
3041 	atomic_add_int(&vmd->vmd_pageout_deficit,
3042 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3043 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3044 		if (object != NULL)
3045 			VM_OBJECT_WUNLOCK(object);
3046 		vm_wait_domain(vmd->vmd_domain);
3047 		if (object != NULL)
3048 			VM_OBJECT_WLOCK(object);
3049 		if (req & VM_ALLOC_WAITOK)
3050 			return (EAGAIN);
3051 	}
3052 
3053 	return (0);
3054 }
3055 
3056 /*
3057  *	vm_waitpfault:
3058  *
3059  *	Sleep until free pages are available for allocation.
3060  *	- Called only in vm_fault so that processes page faulting
3061  *	  can be easily tracked.
3062  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3063  *	  processes will be able to grab memory first.  Do not change
3064  *	  this balance without careful testing first.
3065  */
3066 void
3067 vm_waitpfault(struct domainset *dset)
3068 {
3069 
3070 	/*
3071 	 * XXX Ideally we would wait only until the allocation could
3072 	 * be satisfied.  This condition can cause new allocators to
3073 	 * consume all freed pages while old allocators wait.
3074 	 */
3075 	mtx_lock(&vm_domainset_lock);
3076 	if (vm_page_count_min_set(&dset->ds_mask)) {
3077 		vm_min_waiters++;
3078 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3079 		    "pfault", 0);
3080 	} else
3081 		mtx_unlock(&vm_domainset_lock);
3082 }
3083 
3084 struct vm_pagequeue *
3085 vm_page_pagequeue(vm_page_t m)
3086 {
3087 
3088 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
3089 }
3090 
3091 static struct mtx *
3092 vm_page_pagequeue_lockptr(vm_page_t m)
3093 {
3094 	uint8_t queue;
3095 
3096 	if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3097 		return (NULL);
3098 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex);
3099 }
3100 
3101 static inline void
3102 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3103 {
3104 	struct vm_domain *vmd;
3105 	uint8_t qflags;
3106 
3107 	CRITICAL_ASSERT(curthread);
3108 	vm_pagequeue_assert_locked(pq);
3109 
3110 	/*
3111 	 * The page daemon is allowed to set m->queue = PQ_NONE without
3112 	 * the page queue lock held.  In this case it is about to free the page,
3113 	 * which must not have any queue state.
3114 	 */
3115 	qflags = atomic_load_8(&m->aflags) & PGA_QUEUE_STATE_MASK;
3116 	KASSERT(pq == vm_page_pagequeue(m) || qflags == 0,
3117 	    ("page %p doesn't belong to queue %p but has queue state %#x",
3118 	    m, pq, qflags));
3119 
3120 	if ((qflags & PGA_DEQUEUE) != 0) {
3121 		if (__predict_true((qflags & PGA_ENQUEUED) != 0)) {
3122 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3123 			vm_pagequeue_cnt_dec(pq);
3124 		}
3125 		vm_page_dequeue_complete(m);
3126 	} else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3127 		if ((qflags & PGA_ENQUEUED) != 0)
3128 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3129 		else {
3130 			vm_pagequeue_cnt_inc(pq);
3131 			vm_page_aflag_set(m, PGA_ENQUEUED);
3132 		}
3133 		if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3134 			KASSERT(m->queue == PQ_INACTIVE,
3135 			    ("head enqueue not supported for page %p", m));
3136 			vmd = vm_pagequeue_domain(m);
3137 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3138 		} else
3139 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3140 
3141 		/*
3142 		 * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after
3143 		 * setting PGA_ENQUEUED in order to synchronize with the
3144 		 * page daemon.
3145 		 */
3146 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
3147 	}
3148 }
3149 
3150 static void
3151 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3152     uint8_t queue)
3153 {
3154 	vm_page_t m;
3155 	int i;
3156 
3157 	for (i = 0; i < bq->bq_cnt; i++) {
3158 		m = bq->bq_pa[i];
3159 		if (__predict_false(m->queue != queue))
3160 			continue;
3161 		vm_pqbatch_process_page(pq, m);
3162 	}
3163 	vm_batchqueue_init(bq);
3164 }
3165 
3166 static void
3167 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3168 {
3169 	struct vm_batchqueue *bq;
3170 	struct vm_pagequeue *pq;
3171 	int domain;
3172 
3173 	vm_page_assert_locked(m);
3174 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3175 
3176 	domain = vm_phys_domain(m);
3177 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3178 
3179 	critical_enter();
3180 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3181 	if (vm_batchqueue_insert(bq, m)) {
3182 		critical_exit();
3183 		return;
3184 	}
3185 	if (!vm_pagequeue_trylock(pq)) {
3186 		critical_exit();
3187 		vm_pagequeue_lock(pq);
3188 		critical_enter();
3189 		bq = DPCPU_PTR(pqbatch[domain][queue]);
3190 	}
3191 	vm_pqbatch_process(pq, bq, queue);
3192 
3193 	/*
3194 	 * The page may have been logically dequeued before we acquired the
3195 	 * page queue lock.  In this case, the page lock prevents the page
3196 	 * from being logically enqueued elsewhere.
3197 	 */
3198 	if (__predict_true(m->queue == queue))
3199 		vm_pqbatch_process_page(pq, m);
3200 	else {
3201 		KASSERT(m->queue == PQ_NONE,
3202 		    ("invalid queue transition for page %p", m));
3203 		KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3204 		    ("page %p is enqueued with invalid queue index", m));
3205 		vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3206 	}
3207 	vm_pagequeue_unlock(pq);
3208 	critical_exit();
3209 }
3210 
3211 /*
3212  *	vm_page_drain_pqbatch:		[ internal use only ]
3213  *
3214  *	Force all per-CPU page queue batch queues to be drained.  This is
3215  *	intended for use in severe memory shortages, to ensure that pages
3216  *	do not remain stuck in the batch queues.
3217  */
3218 void
3219 vm_page_drain_pqbatch(void)
3220 {
3221 	struct thread *td;
3222 	struct vm_domain *vmd;
3223 	struct vm_pagequeue *pq;
3224 	int cpu, domain, queue;
3225 
3226 	td = curthread;
3227 	CPU_FOREACH(cpu) {
3228 		thread_lock(td);
3229 		sched_bind(td, cpu);
3230 		thread_unlock(td);
3231 
3232 		for (domain = 0; domain < vm_ndomains; domain++) {
3233 			vmd = VM_DOMAIN(domain);
3234 			for (queue = 0; queue < PQ_COUNT; queue++) {
3235 				pq = &vmd->vmd_pagequeues[queue];
3236 				vm_pagequeue_lock(pq);
3237 				critical_enter();
3238 				vm_pqbatch_process(pq,
3239 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3240 				critical_exit();
3241 				vm_pagequeue_unlock(pq);
3242 			}
3243 		}
3244 	}
3245 	thread_lock(td);
3246 	sched_unbind(td);
3247 	thread_unlock(td);
3248 }
3249 
3250 /*
3251  * Complete the logical removal of a page from a page queue.  We must be
3252  * careful to synchronize with the page daemon, which may be concurrently
3253  * examining the page with only the page lock held.  The page must not be
3254  * in a state where it appears to be logically enqueued.
3255  */
3256 static void
3257 vm_page_dequeue_complete(vm_page_t m)
3258 {
3259 
3260 	m->queue = PQ_NONE;
3261 	atomic_thread_fence_rel();
3262 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3263 }
3264 
3265 /*
3266  *	vm_page_dequeue_deferred:	[ internal use only ]
3267  *
3268  *	Request removal of the given page from its current page
3269  *	queue.  Physical removal from the queue may be deferred
3270  *	indefinitely.
3271  *
3272  *	The page must be locked.
3273  */
3274 void
3275 vm_page_dequeue_deferred(vm_page_t m)
3276 {
3277 	int queue;
3278 
3279 	vm_page_assert_locked(m);
3280 
3281 	queue = atomic_load_8(&m->queue);
3282 	if (queue == PQ_NONE) {
3283 		KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3284 		    ("page %p has queue state", m));
3285 		return;
3286 	}
3287 	if ((m->aflags & PGA_DEQUEUE) == 0)
3288 		vm_page_aflag_set(m, PGA_DEQUEUE);
3289 	vm_pqbatch_submit_page(m, queue);
3290 }
3291 
3292 /*
3293  *	vm_page_dequeue:
3294  *
3295  *	Remove the page from whichever page queue it's in, if any.
3296  *	The page must either be locked or unallocated.  This constraint
3297  *	ensures that the queue state of the page will remain consistent
3298  *	after this function returns.
3299  */
3300 void
3301 vm_page_dequeue(vm_page_t m)
3302 {
3303 	struct mtx *lock, *lock1;
3304 	struct vm_pagequeue *pq;
3305 	uint8_t aflags;
3306 
3307 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->order == VM_NFREEORDER,
3308 	    ("page %p is allocated and unlocked", m));
3309 
3310 	for (;;) {
3311 		lock = vm_page_pagequeue_lockptr(m);
3312 		if (lock == NULL) {
3313 			/*
3314 			 * A thread may be concurrently executing
3315 			 * vm_page_dequeue_complete().  Ensure that all queue
3316 			 * state is cleared before we return.
3317 			 */
3318 			aflags = atomic_load_8(&m->aflags);
3319 			if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3320 				return;
3321 			KASSERT((aflags & PGA_DEQUEUE) != 0,
3322 			    ("page %p has unexpected queue state flags %#x",
3323 			    m, aflags));
3324 
3325 			/*
3326 			 * Busy wait until the thread updating queue state is
3327 			 * finished.  Such a thread must be executing in a
3328 			 * critical section.
3329 			 */
3330 			cpu_spinwait();
3331 			continue;
3332 		}
3333 		mtx_lock(lock);
3334 		if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock)
3335 			break;
3336 		mtx_unlock(lock);
3337 		lock = lock1;
3338 	}
3339 	KASSERT(lock == vm_page_pagequeue_lockptr(m),
3340 	    ("%s: page %p migrated directly between queues", __func__, m));
3341 	KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3342 	    mtx_owned(vm_page_lockptr(m)),
3343 	    ("%s: queued unlocked page %p", __func__, m));
3344 
3345 	if ((m->aflags & PGA_ENQUEUED) != 0) {
3346 		pq = vm_page_pagequeue(m);
3347 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3348 		vm_pagequeue_cnt_dec(pq);
3349 	}
3350 	vm_page_dequeue_complete(m);
3351 	mtx_unlock(lock);
3352 }
3353 
3354 /*
3355  * Schedule the given page for insertion into the specified page queue.
3356  * Physical insertion of the page may be deferred indefinitely.
3357  */
3358 static void
3359 vm_page_enqueue(vm_page_t m, uint8_t queue)
3360 {
3361 
3362 	vm_page_assert_locked(m);
3363 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3364 	    ("%s: page %p is already enqueued", __func__, m));
3365 
3366 	m->queue = queue;
3367 	if ((m->aflags & PGA_REQUEUE) == 0)
3368 		vm_page_aflag_set(m, PGA_REQUEUE);
3369 	vm_pqbatch_submit_page(m, queue);
3370 }
3371 
3372 /*
3373  *	vm_page_requeue:		[ internal use only ]
3374  *
3375  *	Schedule a requeue of the given page.
3376  *
3377  *	The page must be locked.
3378  */
3379 void
3380 vm_page_requeue(vm_page_t m)
3381 {
3382 
3383 	vm_page_assert_locked(m);
3384 	KASSERT(m->queue != PQ_NONE,
3385 	    ("%s: page %p is not logically enqueued", __func__, m));
3386 
3387 	if ((m->aflags & PGA_REQUEUE) == 0)
3388 		vm_page_aflag_set(m, PGA_REQUEUE);
3389 	vm_pqbatch_submit_page(m, atomic_load_8(&m->queue));
3390 }
3391 
3392 /*
3393  *	vm_page_activate:
3394  *
3395  *	Put the specified page on the active list (if appropriate).
3396  *	Ensure that act_count is at least ACT_INIT but do not otherwise
3397  *	mess with it.
3398  *
3399  *	The page must be locked.
3400  */
3401 void
3402 vm_page_activate(vm_page_t m)
3403 {
3404 
3405 	vm_page_assert_locked(m);
3406 
3407 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3408 		return;
3409 	if (vm_page_queue(m) == PQ_ACTIVE) {
3410 		if (m->act_count < ACT_INIT)
3411 			m->act_count = ACT_INIT;
3412 		return;
3413 	}
3414 
3415 	vm_page_dequeue(m);
3416 	if (m->act_count < ACT_INIT)
3417 		m->act_count = ACT_INIT;
3418 	vm_page_enqueue(m, PQ_ACTIVE);
3419 }
3420 
3421 /*
3422  *	vm_page_free_prep:
3423  *
3424  *	Prepares the given page to be put on the free list,
3425  *	disassociating it from any VM object. The caller may return
3426  *	the page to the free list only if this function returns true.
3427  *
3428  *	The object must be locked.  The page must be locked if it is
3429  *	managed.
3430  */
3431 bool
3432 vm_page_free_prep(vm_page_t m)
3433 {
3434 
3435 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3436 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3437 		uint64_t *p;
3438 		int i;
3439 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3440 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3441 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3442 			    m, i, (uintmax_t)*p));
3443 	}
3444 #endif
3445 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3446 		vm_page_lock_assert(m, MA_OWNED);
3447 		KASSERT(!pmap_page_is_mapped(m),
3448 		    ("vm_page_free_prep: freeing mapped page %p", m));
3449 	} else
3450 		KASSERT(m->queue == PQ_NONE,
3451 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3452 	VM_CNT_INC(v_tfree);
3453 
3454 	if (vm_page_sbusied(m))
3455 		panic("vm_page_free_prep: freeing busy page %p", m);
3456 
3457 	vm_page_remove(m);
3458 
3459 	/*
3460 	 * If fictitious remove object association and
3461 	 * return.
3462 	 */
3463 	if ((m->flags & PG_FICTITIOUS) != 0) {
3464 		KASSERT(m->wire_count == 1,
3465 		    ("fictitious page %p is not wired", m));
3466 		KASSERT(m->queue == PQ_NONE,
3467 		    ("fictitious page %p is queued", m));
3468 		return (false);
3469 	}
3470 
3471 	/*
3472 	 * Pages need not be dequeued before they are returned to the physical
3473 	 * memory allocator, but they must at least be marked for a deferred
3474 	 * dequeue.
3475 	 */
3476 	if ((m->oflags & VPO_UNMANAGED) == 0)
3477 		vm_page_dequeue_deferred(m);
3478 
3479 	m->valid = 0;
3480 	vm_page_undirty(m);
3481 
3482 	if (m->wire_count != 0)
3483 		panic("vm_page_free_prep: freeing wired page %p", m);
3484 	if (m->hold_count != 0) {
3485 		m->flags &= ~PG_ZERO;
3486 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3487 		    ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m));
3488 		m->flags |= PG_UNHOLDFREE;
3489 		return (false);
3490 	}
3491 
3492 	/*
3493 	 * Restore the default memory attribute to the page.
3494 	 */
3495 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3496 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3497 
3498 #if VM_NRESERVLEVEL > 0
3499 	if (vm_reserv_free_page(m))
3500 		return (false);
3501 #endif
3502 
3503 	return (true);
3504 }
3505 
3506 /*
3507  *	vm_page_free_toq:
3508  *
3509  *	Returns the given page to the free list, disassociating it
3510  *	from any VM object.
3511  *
3512  *	The object must be locked.  The page must be locked if it is
3513  *	managed.
3514  */
3515 void
3516 vm_page_free_toq(vm_page_t m)
3517 {
3518 	struct vm_domain *vmd;
3519 
3520 	if (!vm_page_free_prep(m))
3521 		return;
3522 
3523 	vmd = vm_pagequeue_domain(m);
3524 	if (m->pool == VM_FREEPOOL_DEFAULT && vmd->vmd_pgcache != NULL) {
3525 		uma_zfree(vmd->vmd_pgcache, m);
3526 		return;
3527 	}
3528 	vm_domain_free_lock(vmd);
3529 	vm_phys_free_pages(m, 0);
3530 	vm_domain_free_unlock(vmd);
3531 	vm_domain_freecnt_inc(vmd, 1);
3532 }
3533 
3534 /*
3535  *	vm_page_free_pages_toq:
3536  *
3537  *	Returns a list of pages to the free list, disassociating it
3538  *	from any VM object.  In other words, this is equivalent to
3539  *	calling vm_page_free_toq() for each page of a list of VM objects.
3540  *
3541  *	The objects must be locked.  The pages must be locked if it is
3542  *	managed.
3543  */
3544 void
3545 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3546 {
3547 	vm_page_t m;
3548 	int count;
3549 
3550 	if (SLIST_EMPTY(free))
3551 		return;
3552 
3553 	count = 0;
3554 	while ((m = SLIST_FIRST(free)) != NULL) {
3555 		count++;
3556 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3557 		vm_page_free_toq(m);
3558 	}
3559 
3560 	if (update_wire_count)
3561 		vm_wire_sub(count);
3562 }
3563 
3564 /*
3565  *	vm_page_wire:
3566  *
3567  * Mark this page as wired down.  If the page is fictitious, then
3568  * its wire count must remain one.
3569  *
3570  * The page must be locked.
3571  */
3572 void
3573 vm_page_wire(vm_page_t m)
3574 {
3575 
3576 	vm_page_assert_locked(m);
3577 	if ((m->flags & PG_FICTITIOUS) != 0) {
3578 		KASSERT(m->wire_count == 1,
3579 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
3580 		    m));
3581 		return;
3582 	}
3583 	if (m->wire_count == 0) {
3584 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3585 		    m->queue == PQ_NONE,
3586 		    ("vm_page_wire: unmanaged page %p is queued", m));
3587 		vm_wire_add(1);
3588 	}
3589 	m->wire_count++;
3590 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3591 }
3592 
3593 /*
3594  * vm_page_unwire:
3595  *
3596  * Release one wiring of the specified page, potentially allowing it to be
3597  * paged out.  Returns TRUE if the number of wirings transitions to zero and
3598  * FALSE otherwise.
3599  *
3600  * Only managed pages belonging to an object can be paged out.  If the number
3601  * of wirings transitions to zero and the page is eligible for page out, then
3602  * the page is added to the specified paging queue (unless PQ_NONE is
3603  * specified, in which case the page is dequeued if it belongs to a paging
3604  * queue).
3605  *
3606  * If a page is fictitious, then its wire count must always be one.
3607  *
3608  * A managed page must be locked.
3609  */
3610 bool
3611 vm_page_unwire(vm_page_t m, uint8_t queue)
3612 {
3613 	bool unwired;
3614 
3615 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3616 	    ("vm_page_unwire: invalid queue %u request for page %p",
3617 	    queue, m));
3618 	if ((m->oflags & VPO_UNMANAGED) == 0)
3619 		vm_page_assert_locked(m);
3620 
3621 	unwired = vm_page_unwire_noq(m);
3622 	if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3623 		return (unwired);
3624 
3625 	if (vm_page_queue(m) == queue) {
3626 		if (queue == PQ_ACTIVE)
3627 			vm_page_reference(m);
3628 		else if (queue != PQ_NONE)
3629 			vm_page_requeue(m);
3630 	} else {
3631 		vm_page_dequeue(m);
3632 		if (queue != PQ_NONE) {
3633 			vm_page_enqueue(m, queue);
3634 			if (queue == PQ_ACTIVE)
3635 				/* Initialize act_count. */
3636 				vm_page_activate(m);
3637 		}
3638 	}
3639 	return (unwired);
3640 }
3641 
3642 /*
3643  *
3644  * vm_page_unwire_noq:
3645  *
3646  * Unwire a page without (re-)inserting it into a page queue.  It is up
3647  * to the caller to enqueue, requeue, or free the page as appropriate.
3648  * In most cases, vm_page_unwire() should be used instead.
3649  */
3650 bool
3651 vm_page_unwire_noq(vm_page_t m)
3652 {
3653 
3654 	if ((m->oflags & VPO_UNMANAGED) == 0)
3655 		vm_page_assert_locked(m);
3656 	if ((m->flags & PG_FICTITIOUS) != 0) {
3657 		KASSERT(m->wire_count == 1,
3658 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3659 		return (false);
3660 	}
3661 	if (m->wire_count == 0)
3662 		panic("vm_page_unwire: page %p's wire count is zero", m);
3663 	m->wire_count--;
3664 	if (m->wire_count == 0) {
3665 		vm_wire_sub(1);
3666 		return (true);
3667 	} else
3668 		return (false);
3669 }
3670 
3671 /*
3672  * Move the specified page to the tail of the inactive queue, or requeue
3673  * the page if it is already in the inactive queue.
3674  *
3675  * The page must be locked.
3676  */
3677 void
3678 vm_page_deactivate(vm_page_t m)
3679 {
3680 
3681 	vm_page_assert_locked(m);
3682 
3683 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3684 		return;
3685 
3686 	if (!vm_page_inactive(m)) {
3687 		vm_page_dequeue(m);
3688 		vm_page_enqueue(m, PQ_INACTIVE);
3689 	} else
3690 		vm_page_requeue(m);
3691 }
3692 
3693 /*
3694  * Move the specified page close to the head of the inactive queue,
3695  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3696  * As with regular enqueues, we use a per-CPU batch queue to reduce
3697  * contention on the page queue lock.
3698  *
3699  * The page must be locked.
3700  */
3701 void
3702 vm_page_deactivate_noreuse(vm_page_t m)
3703 {
3704 
3705 	vm_page_assert_locked(m);
3706 
3707 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3708 		return;
3709 
3710 	if (!vm_page_inactive(m)) {
3711 		vm_page_dequeue(m);
3712 		m->queue = PQ_INACTIVE;
3713 	}
3714 	if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3715 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3716 	vm_pqbatch_submit_page(m, PQ_INACTIVE);
3717 }
3718 
3719 /*
3720  * vm_page_launder
3721  *
3722  * 	Put a page in the laundry, or requeue it if it is already there.
3723  */
3724 void
3725 vm_page_launder(vm_page_t m)
3726 {
3727 
3728 	vm_page_assert_locked(m);
3729 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3730 		return;
3731 
3732 	if (vm_page_in_laundry(m))
3733 		vm_page_requeue(m);
3734 	else {
3735 		vm_page_dequeue(m);
3736 		vm_page_enqueue(m, PQ_LAUNDRY);
3737 	}
3738 }
3739 
3740 /*
3741  * vm_page_unswappable
3742  *
3743  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3744  */
3745 void
3746 vm_page_unswappable(vm_page_t m)
3747 {
3748 
3749 	vm_page_assert_locked(m);
3750 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3751 	    ("page %p already unswappable", m));
3752 
3753 	vm_page_dequeue(m);
3754 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
3755 }
3756 
3757 /*
3758  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3759  * if the page is freed and false otherwise.
3760  *
3761  * The page must be managed.  The page and its containing object must be
3762  * locked.
3763  */
3764 bool
3765 vm_page_try_to_free(vm_page_t m)
3766 {
3767 
3768 	vm_page_assert_locked(m);
3769 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3770 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3771 	if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m))
3772 		return (false);
3773 	if (m->object->ref_count != 0) {
3774 		pmap_remove_all(m);
3775 		if (m->dirty != 0)
3776 			return (false);
3777 	}
3778 	vm_page_free(m);
3779 	return (true);
3780 }
3781 
3782 /*
3783  * vm_page_advise
3784  *
3785  * 	Apply the specified advice to the given page.
3786  *
3787  *	The object and page must be locked.
3788  */
3789 void
3790 vm_page_advise(vm_page_t m, int advice)
3791 {
3792 
3793 	vm_page_assert_locked(m);
3794 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3795 	if (advice == MADV_FREE)
3796 		/*
3797 		 * Mark the page clean.  This will allow the page to be freed
3798 		 * without first paging it out.  MADV_FREE pages are often
3799 		 * quickly reused by malloc(3), so we do not do anything that
3800 		 * would result in a page fault on a later access.
3801 		 */
3802 		vm_page_undirty(m);
3803 	else if (advice != MADV_DONTNEED) {
3804 		if (advice == MADV_WILLNEED)
3805 			vm_page_activate(m);
3806 		return;
3807 	}
3808 
3809 	/*
3810 	 * Clear any references to the page.  Otherwise, the page daemon will
3811 	 * immediately reactivate the page.
3812 	 */
3813 	vm_page_aflag_clear(m, PGA_REFERENCED);
3814 
3815 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3816 		vm_page_dirty(m);
3817 
3818 	/*
3819 	 * Place clean pages near the head of the inactive queue rather than
3820 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3821 	 * the page will be reused quickly.  Dirty pages not already in the
3822 	 * laundry are moved there.
3823 	 */
3824 	if (m->dirty == 0)
3825 		vm_page_deactivate_noreuse(m);
3826 	else if (!vm_page_in_laundry(m))
3827 		vm_page_launder(m);
3828 }
3829 
3830 /*
3831  * Grab a page, waiting until we are waken up due to the page
3832  * changing state.  We keep on waiting, if the page continues
3833  * to be in the object.  If the page doesn't exist, first allocate it
3834  * and then conditionally zero it.
3835  *
3836  * This routine may sleep.
3837  *
3838  * The object must be locked on entry.  The lock will, however, be released
3839  * and reacquired if the routine sleeps.
3840  */
3841 vm_page_t
3842 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3843 {
3844 	vm_page_t m;
3845 	int sleep;
3846 	int pflags;
3847 
3848 	VM_OBJECT_ASSERT_WLOCKED(object);
3849 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3850 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3851 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3852 	pflags = allocflags &
3853 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3854 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3855 		pflags |= VM_ALLOC_WAITFAIL;
3856 retrylookup:
3857 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3858 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3859 		    vm_page_xbusied(m) : vm_page_busied(m);
3860 		if (sleep) {
3861 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3862 				return (NULL);
3863 			/*
3864 			 * Reference the page before unlocking and
3865 			 * sleeping so that the page daemon is less
3866 			 * likely to reclaim it.
3867 			 */
3868 			vm_page_aflag_set(m, PGA_REFERENCED);
3869 			vm_page_lock(m);
3870 			VM_OBJECT_WUNLOCK(object);
3871 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3872 			    VM_ALLOC_IGN_SBUSY) != 0);
3873 			VM_OBJECT_WLOCK(object);
3874 			goto retrylookup;
3875 		} else {
3876 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3877 				vm_page_lock(m);
3878 				vm_page_wire(m);
3879 				vm_page_unlock(m);
3880 			}
3881 			if ((allocflags &
3882 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3883 				vm_page_xbusy(m);
3884 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3885 				vm_page_sbusy(m);
3886 			return (m);
3887 		}
3888 	}
3889 	m = vm_page_alloc(object, pindex, pflags);
3890 	if (m == NULL) {
3891 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3892 			return (NULL);
3893 		goto retrylookup;
3894 	}
3895 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3896 		pmap_zero_page(m);
3897 	return (m);
3898 }
3899 
3900 /*
3901  * Return the specified range of pages from the given object.  For each
3902  * page offset within the range, if a page already exists within the object
3903  * at that offset and it is busy, then wait for it to change state.  If,
3904  * instead, the page doesn't exist, then allocate it.
3905  *
3906  * The caller must always specify an allocation class.
3907  *
3908  * allocation classes:
3909  *	VM_ALLOC_NORMAL		normal process request
3910  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3911  *
3912  * The caller must always specify that the pages are to be busied and/or
3913  * wired.
3914  *
3915  * optional allocation flags:
3916  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3917  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3918  *	VM_ALLOC_NOWAIT		do not sleep
3919  *	VM_ALLOC_SBUSY		set page to sbusy state
3920  *	VM_ALLOC_WIRED		wire the pages
3921  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3922  *
3923  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3924  * may return a partial prefix of the requested range.
3925  */
3926 int
3927 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3928     vm_page_t *ma, int count)
3929 {
3930 	vm_page_t m, mpred;
3931 	int pflags;
3932 	int i;
3933 	bool sleep;
3934 
3935 	VM_OBJECT_ASSERT_WLOCKED(object);
3936 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3937 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3938 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3939 	    (allocflags & VM_ALLOC_WIRED) != 0,
3940 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3941 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3942 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3943 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3944 	if (count == 0)
3945 		return (0);
3946 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3947 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3948 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3949 		pflags |= VM_ALLOC_WAITFAIL;
3950 	i = 0;
3951 retrylookup:
3952 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3953 	if (m == NULL || m->pindex != pindex + i) {
3954 		mpred = m;
3955 		m = NULL;
3956 	} else
3957 		mpred = TAILQ_PREV(m, pglist, listq);
3958 	for (; i < count; i++) {
3959 		if (m != NULL) {
3960 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3961 			    vm_page_xbusied(m) : vm_page_busied(m);
3962 			if (sleep) {
3963 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3964 					break;
3965 				/*
3966 				 * Reference the page before unlocking and
3967 				 * sleeping so that the page daemon is less
3968 				 * likely to reclaim it.
3969 				 */
3970 				vm_page_aflag_set(m, PGA_REFERENCED);
3971 				vm_page_lock(m);
3972 				VM_OBJECT_WUNLOCK(object);
3973 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
3974 				    VM_ALLOC_IGN_SBUSY) != 0);
3975 				VM_OBJECT_WLOCK(object);
3976 				goto retrylookup;
3977 			}
3978 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3979 				vm_page_lock(m);
3980 				vm_page_wire(m);
3981 				vm_page_unlock(m);
3982 			}
3983 			if ((allocflags & (VM_ALLOC_NOBUSY |
3984 			    VM_ALLOC_SBUSY)) == 0)
3985 				vm_page_xbusy(m);
3986 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3987 				vm_page_sbusy(m);
3988 		} else {
3989 			m = vm_page_alloc_after(object, pindex + i,
3990 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
3991 			if (m == NULL) {
3992 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3993 					break;
3994 				goto retrylookup;
3995 			}
3996 		}
3997 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3998 			if ((m->flags & PG_ZERO) == 0)
3999 				pmap_zero_page(m);
4000 			m->valid = VM_PAGE_BITS_ALL;
4001 		}
4002 		ma[i] = mpred = m;
4003 		m = vm_page_next(m);
4004 	}
4005 	return (i);
4006 }
4007 
4008 /*
4009  * Mapping function for valid or dirty bits in a page.
4010  *
4011  * Inputs are required to range within a page.
4012  */
4013 vm_page_bits_t
4014 vm_page_bits(int base, int size)
4015 {
4016 	int first_bit;
4017 	int last_bit;
4018 
4019 	KASSERT(
4020 	    base + size <= PAGE_SIZE,
4021 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4022 	);
4023 
4024 	if (size == 0)		/* handle degenerate case */
4025 		return (0);
4026 
4027 	first_bit = base >> DEV_BSHIFT;
4028 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4029 
4030 	return (((vm_page_bits_t)2 << last_bit) -
4031 	    ((vm_page_bits_t)1 << first_bit));
4032 }
4033 
4034 /*
4035  *	vm_page_set_valid_range:
4036  *
4037  *	Sets portions of a page valid.  The arguments are expected
4038  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4039  *	of any partial chunks touched by the range.  The invalid portion of
4040  *	such chunks will be zeroed.
4041  *
4042  *	(base + size) must be less then or equal to PAGE_SIZE.
4043  */
4044 void
4045 vm_page_set_valid_range(vm_page_t m, int base, int size)
4046 {
4047 	int endoff, frag;
4048 
4049 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4050 	if (size == 0)	/* handle degenerate case */
4051 		return;
4052 
4053 	/*
4054 	 * If the base is not DEV_BSIZE aligned and the valid
4055 	 * bit is clear, we have to zero out a portion of the
4056 	 * first block.
4057 	 */
4058 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4059 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4060 		pmap_zero_page_area(m, frag, base - frag);
4061 
4062 	/*
4063 	 * If the ending offset is not DEV_BSIZE aligned and the
4064 	 * valid bit is clear, we have to zero out a portion of
4065 	 * the last block.
4066 	 */
4067 	endoff = base + size;
4068 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4069 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4070 		pmap_zero_page_area(m, endoff,
4071 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4072 
4073 	/*
4074 	 * Assert that no previously invalid block that is now being validated
4075 	 * is already dirty.
4076 	 */
4077 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4078 	    ("vm_page_set_valid_range: page %p is dirty", m));
4079 
4080 	/*
4081 	 * Set valid bits inclusive of any overlap.
4082 	 */
4083 	m->valid |= vm_page_bits(base, size);
4084 }
4085 
4086 /*
4087  * Clear the given bits from the specified page's dirty field.
4088  */
4089 static __inline void
4090 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4091 {
4092 	uintptr_t addr;
4093 #if PAGE_SIZE < 16384
4094 	int shift;
4095 #endif
4096 
4097 	/*
4098 	 * If the object is locked and the page is neither exclusive busy nor
4099 	 * write mapped, then the page's dirty field cannot possibly be
4100 	 * set by a concurrent pmap operation.
4101 	 */
4102 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4103 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4104 		m->dirty &= ~pagebits;
4105 	else {
4106 		/*
4107 		 * The pmap layer can call vm_page_dirty() without
4108 		 * holding a distinguished lock.  The combination of
4109 		 * the object's lock and an atomic operation suffice
4110 		 * to guarantee consistency of the page dirty field.
4111 		 *
4112 		 * For PAGE_SIZE == 32768 case, compiler already
4113 		 * properly aligns the dirty field, so no forcible
4114 		 * alignment is needed. Only require existence of
4115 		 * atomic_clear_64 when page size is 32768.
4116 		 */
4117 		addr = (uintptr_t)&m->dirty;
4118 #if PAGE_SIZE == 32768
4119 		atomic_clear_64((uint64_t *)addr, pagebits);
4120 #elif PAGE_SIZE == 16384
4121 		atomic_clear_32((uint32_t *)addr, pagebits);
4122 #else		/* PAGE_SIZE <= 8192 */
4123 		/*
4124 		 * Use a trick to perform a 32-bit atomic on the
4125 		 * containing aligned word, to not depend on the existence
4126 		 * of atomic_clear_{8, 16}.
4127 		 */
4128 		shift = addr & (sizeof(uint32_t) - 1);
4129 #if BYTE_ORDER == BIG_ENDIAN
4130 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4131 #else
4132 		shift *= NBBY;
4133 #endif
4134 		addr &= ~(sizeof(uint32_t) - 1);
4135 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
4136 #endif		/* PAGE_SIZE */
4137 	}
4138 }
4139 
4140 /*
4141  *	vm_page_set_validclean:
4142  *
4143  *	Sets portions of a page valid and clean.  The arguments are expected
4144  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4145  *	of any partial chunks touched by the range.  The invalid portion of
4146  *	such chunks will be zero'd.
4147  *
4148  *	(base + size) must be less then or equal to PAGE_SIZE.
4149  */
4150 void
4151 vm_page_set_validclean(vm_page_t m, int base, int size)
4152 {
4153 	vm_page_bits_t oldvalid, pagebits;
4154 	int endoff, frag;
4155 
4156 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4157 	if (size == 0)	/* handle degenerate case */
4158 		return;
4159 
4160 	/*
4161 	 * If the base is not DEV_BSIZE aligned and the valid
4162 	 * bit is clear, we have to zero out a portion of the
4163 	 * first block.
4164 	 */
4165 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4166 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4167 		pmap_zero_page_area(m, frag, base - frag);
4168 
4169 	/*
4170 	 * If the ending offset is not DEV_BSIZE aligned and the
4171 	 * valid bit is clear, we have to zero out a portion of
4172 	 * the last block.
4173 	 */
4174 	endoff = base + size;
4175 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4176 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4177 		pmap_zero_page_area(m, endoff,
4178 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4179 
4180 	/*
4181 	 * Set valid, clear dirty bits.  If validating the entire
4182 	 * page we can safely clear the pmap modify bit.  We also
4183 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
4184 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4185 	 * be set again.
4186 	 *
4187 	 * We set valid bits inclusive of any overlap, but we can only
4188 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4189 	 * the range.
4190 	 */
4191 	oldvalid = m->valid;
4192 	pagebits = vm_page_bits(base, size);
4193 	m->valid |= pagebits;
4194 #if 0	/* NOT YET */
4195 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4196 		frag = DEV_BSIZE - frag;
4197 		base += frag;
4198 		size -= frag;
4199 		if (size < 0)
4200 			size = 0;
4201 	}
4202 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4203 #endif
4204 	if (base == 0 && size == PAGE_SIZE) {
4205 		/*
4206 		 * The page can only be modified within the pmap if it is
4207 		 * mapped, and it can only be mapped if it was previously
4208 		 * fully valid.
4209 		 */
4210 		if (oldvalid == VM_PAGE_BITS_ALL)
4211 			/*
4212 			 * Perform the pmap_clear_modify() first.  Otherwise,
4213 			 * a concurrent pmap operation, such as
4214 			 * pmap_protect(), could clear a modification in the
4215 			 * pmap and set the dirty field on the page before
4216 			 * pmap_clear_modify() had begun and after the dirty
4217 			 * field was cleared here.
4218 			 */
4219 			pmap_clear_modify(m);
4220 		m->dirty = 0;
4221 		m->oflags &= ~VPO_NOSYNC;
4222 	} else if (oldvalid != VM_PAGE_BITS_ALL)
4223 		m->dirty &= ~pagebits;
4224 	else
4225 		vm_page_clear_dirty_mask(m, pagebits);
4226 }
4227 
4228 void
4229 vm_page_clear_dirty(vm_page_t m, int base, int size)
4230 {
4231 
4232 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4233 }
4234 
4235 /*
4236  *	vm_page_set_invalid:
4237  *
4238  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4239  *	valid and dirty bits for the effected areas are cleared.
4240  */
4241 void
4242 vm_page_set_invalid(vm_page_t m, int base, int size)
4243 {
4244 	vm_page_bits_t bits;
4245 	vm_object_t object;
4246 
4247 	object = m->object;
4248 	VM_OBJECT_ASSERT_WLOCKED(object);
4249 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4250 	    size >= object->un_pager.vnp.vnp_size)
4251 		bits = VM_PAGE_BITS_ALL;
4252 	else
4253 		bits = vm_page_bits(base, size);
4254 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4255 	    bits != 0)
4256 		pmap_remove_all(m);
4257 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4258 	    !pmap_page_is_mapped(m),
4259 	    ("vm_page_set_invalid: page %p is mapped", m));
4260 	m->valid &= ~bits;
4261 	m->dirty &= ~bits;
4262 }
4263 
4264 /*
4265  * vm_page_zero_invalid()
4266  *
4267  *	The kernel assumes that the invalid portions of a page contain
4268  *	garbage, but such pages can be mapped into memory by user code.
4269  *	When this occurs, we must zero out the non-valid portions of the
4270  *	page so user code sees what it expects.
4271  *
4272  *	Pages are most often semi-valid when the end of a file is mapped
4273  *	into memory and the file's size is not page aligned.
4274  */
4275 void
4276 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4277 {
4278 	int b;
4279 	int i;
4280 
4281 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4282 	/*
4283 	 * Scan the valid bits looking for invalid sections that
4284 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4285 	 * valid bit may be set ) have already been zeroed by
4286 	 * vm_page_set_validclean().
4287 	 */
4288 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4289 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4290 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4291 			if (i > b) {
4292 				pmap_zero_page_area(m,
4293 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4294 			}
4295 			b = i + 1;
4296 		}
4297 	}
4298 
4299 	/*
4300 	 * setvalid is TRUE when we can safely set the zero'd areas
4301 	 * as being valid.  We can do this if there are no cache consistancy
4302 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4303 	 */
4304 	if (setvalid)
4305 		m->valid = VM_PAGE_BITS_ALL;
4306 }
4307 
4308 /*
4309  *	vm_page_is_valid:
4310  *
4311  *	Is (partial) page valid?  Note that the case where size == 0
4312  *	will return FALSE in the degenerate case where the page is
4313  *	entirely invalid, and TRUE otherwise.
4314  */
4315 int
4316 vm_page_is_valid(vm_page_t m, int base, int size)
4317 {
4318 	vm_page_bits_t bits;
4319 
4320 	VM_OBJECT_ASSERT_LOCKED(m->object);
4321 	bits = vm_page_bits(base, size);
4322 	return (m->valid != 0 && (m->valid & bits) == bits);
4323 }
4324 
4325 /*
4326  * Returns true if all of the specified predicates are true for the entire
4327  * (super)page and false otherwise.
4328  */
4329 bool
4330 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4331 {
4332 	vm_object_t object;
4333 	int i, npages;
4334 
4335 	object = m->object;
4336 	if (skip_m != NULL && skip_m->object != object)
4337 		return (false);
4338 	VM_OBJECT_ASSERT_LOCKED(object);
4339 	npages = atop(pagesizes[m->psind]);
4340 
4341 	/*
4342 	 * The physically contiguous pages that make up a superpage, i.e., a
4343 	 * page with a page size index ("psind") greater than zero, will
4344 	 * occupy adjacent entries in vm_page_array[].
4345 	 */
4346 	for (i = 0; i < npages; i++) {
4347 		/* Always test object consistency, including "skip_m". */
4348 		if (m[i].object != object)
4349 			return (false);
4350 		if (&m[i] == skip_m)
4351 			continue;
4352 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4353 			return (false);
4354 		if ((flags & PS_ALL_DIRTY) != 0) {
4355 			/*
4356 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4357 			 * might stop this case from spuriously returning
4358 			 * "false".  However, that would require a write lock
4359 			 * on the object containing "m[i]".
4360 			 */
4361 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4362 				return (false);
4363 		}
4364 		if ((flags & PS_ALL_VALID) != 0 &&
4365 		    m[i].valid != VM_PAGE_BITS_ALL)
4366 			return (false);
4367 	}
4368 	return (true);
4369 }
4370 
4371 /*
4372  * Set the page's dirty bits if the page is modified.
4373  */
4374 void
4375 vm_page_test_dirty(vm_page_t m)
4376 {
4377 
4378 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4379 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4380 		vm_page_dirty(m);
4381 }
4382 
4383 void
4384 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4385 {
4386 
4387 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4388 }
4389 
4390 void
4391 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4392 {
4393 
4394 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4395 }
4396 
4397 int
4398 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4399 {
4400 
4401 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4402 }
4403 
4404 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4405 void
4406 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4407 {
4408 
4409 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4410 }
4411 
4412 void
4413 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4414 {
4415 
4416 	mtx_assert_(vm_page_lockptr(m), a, file, line);
4417 }
4418 #endif
4419 
4420 #ifdef INVARIANTS
4421 void
4422 vm_page_object_lock_assert(vm_page_t m)
4423 {
4424 
4425 	/*
4426 	 * Certain of the page's fields may only be modified by the
4427 	 * holder of the containing object's lock or the exclusive busy.
4428 	 * holder.  Unfortunately, the holder of the write busy is
4429 	 * not recorded, and thus cannot be checked here.
4430 	 */
4431 	if (m->object != NULL && !vm_page_xbusied(m))
4432 		VM_OBJECT_ASSERT_WLOCKED(m->object);
4433 }
4434 
4435 void
4436 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4437 {
4438 
4439 	if ((bits & PGA_WRITEABLE) == 0)
4440 		return;
4441 
4442 	/*
4443 	 * The PGA_WRITEABLE flag can only be set if the page is
4444 	 * managed, is exclusively busied or the object is locked.
4445 	 * Currently, this flag is only set by pmap_enter().
4446 	 */
4447 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4448 	    ("PGA_WRITEABLE on unmanaged page"));
4449 	if (!vm_page_xbusied(m))
4450 		VM_OBJECT_ASSERT_LOCKED(m->object);
4451 }
4452 #endif
4453 
4454 #include "opt_ddb.h"
4455 #ifdef DDB
4456 #include <sys/kernel.h>
4457 
4458 #include <ddb/ddb.h>
4459 
4460 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4461 {
4462 
4463 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4464 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4465 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4466 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4467 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4468 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4469 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4470 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4471 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4472 }
4473 
4474 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4475 {
4476 	int dom;
4477 
4478 	db_printf("pq_free %d\n", vm_free_count());
4479 	for (dom = 0; dom < vm_ndomains; dom++) {
4480 		db_printf(
4481     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4482 		    dom,
4483 		    vm_dom[dom].vmd_page_count,
4484 		    vm_dom[dom].vmd_free_count,
4485 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4486 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4487 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4488 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4489 	}
4490 }
4491 
4492 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4493 {
4494 	vm_page_t m;
4495 	boolean_t phys;
4496 
4497 	if (!have_addr) {
4498 		db_printf("show pginfo addr\n");
4499 		return;
4500 	}
4501 
4502 	phys = strchr(modif, 'p') != NULL;
4503 	if (phys)
4504 		m = PHYS_TO_VM_PAGE(addr);
4505 	else
4506 		m = (vm_page_t)addr;
4507 	db_printf(
4508     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4509     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4510 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4511 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4512 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4513 }
4514 #endif /* DDB */
4515