xref: /freebsd/sys/vm/vm_page.c (revision d0ba1baed3f6e4936a0c1b89c25f6c59168ef6de)
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 static DPCPU_DEFINE(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 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
485 	}
486 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
487 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
488 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
489 
490 	/*
491 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
492 	 * insertions.
493 	 */
494 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
495 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
496 	    &vmd->vmd_inacthead, plinks.q);
497 
498 	/*
499 	 * The clock pages are used to implement active queue scanning without
500 	 * requeues.  Scans start at clock[0], which is advanced after the scan
501 	 * ends.  When the two clock hands meet, they are reset and scanning
502 	 * resumes from the head of the queue.
503 	 */
504 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
505 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
506 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
507 	    &vmd->vmd_clock[0], plinks.q);
508 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
509 	    &vmd->vmd_clock[1], plinks.q);
510 }
511 
512 /*
513  * Initialize a physical page in preparation for adding it to the free
514  * lists.
515  */
516 static void
517 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
518 {
519 
520 	m->object = NULL;
521 	m->wire_count = 0;
522 	m->busy_lock = VPB_UNBUSIED;
523 	m->hold_count = 0;
524 	m->flags = 0;
525 	m->phys_addr = pa;
526 	m->queue = PQ_NONE;
527 	m->psind = 0;
528 	m->segind = segind;
529 	m->order = VM_NFREEORDER;
530 	m->pool = VM_FREEPOOL_DEFAULT;
531 	m->valid = m->dirty = 0;
532 	pmap_page_init(m);
533 }
534 
535 /*
536  *	vm_page_startup:
537  *
538  *	Initializes the resident memory module.  Allocates physical memory for
539  *	bootstrapping UMA and some data structures that are used to manage
540  *	physical pages.  Initializes these structures, and populates the free
541  *	page queues.
542  */
543 vm_offset_t
544 vm_page_startup(vm_offset_t vaddr)
545 {
546 	struct vm_phys_seg *seg;
547 	vm_page_t m;
548 	char *list, *listend;
549 	vm_offset_t mapped;
550 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
551 	vm_paddr_t biggestsize, last_pa, pa;
552 	u_long pagecount;
553 	int biggestone, i, segind;
554 
555 	biggestsize = 0;
556 	biggestone = 0;
557 	vaddr = round_page(vaddr);
558 
559 	for (i = 0; phys_avail[i + 1]; i += 2) {
560 		phys_avail[i] = round_page(phys_avail[i]);
561 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
562 	}
563 	for (i = 0; phys_avail[i + 1]; i += 2) {
564 		size = phys_avail[i + 1] - phys_avail[i];
565 		if (size > biggestsize) {
566 			biggestone = i;
567 			biggestsize = size;
568 		}
569 	}
570 
571 	end = phys_avail[biggestone+1];
572 
573 	/*
574 	 * Initialize the page and queue locks.
575 	 */
576 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
577 	for (i = 0; i < PA_LOCK_COUNT; i++)
578 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
579 	for (i = 0; i < vm_ndomains; i++)
580 		vm_page_domain_init(i);
581 
582 	/*
583 	 * Allocate memory for use when boot strapping the kernel memory
584 	 * allocator.  Tell UMA how many zones we are going to create
585 	 * before going fully functional.  UMA will add its zones.
586 	 *
587 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
588 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
589 	 */
590 	boot_pages = uma_startup_count(8);
591 
592 #ifndef UMA_MD_SMALL_ALLOC
593 	/* vmem_startup() calls uma_prealloc(). */
594 	boot_pages += vmem_startup_count();
595 	/* vm_map_startup() calls uma_prealloc(). */
596 	boot_pages += howmany(MAX_KMAP,
597 	    UMA_SLAB_SPACE / sizeof(struct vm_map));
598 
599 	/*
600 	 * Before going fully functional kmem_init() does allocation
601 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
602 	 */
603 	boot_pages += 2;
604 #endif
605 	/*
606 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
607 	 * manually fetch the value.
608 	 */
609 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
610 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
611 	new_end = trunc_page(new_end);
612 	mapped = pmap_map(&vaddr, new_end, end,
613 	    VM_PROT_READ | VM_PROT_WRITE);
614 	bzero((void *)mapped, end - new_end);
615 	uma_startup((void *)mapped, boot_pages);
616 
617 #ifdef WITNESS
618 	end = new_end;
619 	new_end = end - round_page(witness_startup_count());
620 	mapped = pmap_map(&vaddr, new_end, end,
621 	    VM_PROT_READ | VM_PROT_WRITE);
622 	bzero((void *)mapped, end - new_end);
623 	witness_startup((void *)mapped);
624 #endif
625 
626 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
627     defined(__i386__) || defined(__mips__)
628 	/*
629 	 * Allocate a bitmap to indicate that a random physical page
630 	 * needs to be included in a minidump.
631 	 *
632 	 * The amd64 port needs this to indicate which direct map pages
633 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
634 	 *
635 	 * However, i386 still needs this workspace internally within the
636 	 * minidump code.  In theory, they are not needed on i386, but are
637 	 * included should the sf_buf code decide to use them.
638 	 */
639 	last_pa = 0;
640 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
641 		if (dump_avail[i + 1] > last_pa)
642 			last_pa = dump_avail[i + 1];
643 	page_range = last_pa / PAGE_SIZE;
644 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
645 	new_end -= vm_page_dump_size;
646 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
647 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
648 	bzero((void *)vm_page_dump, vm_page_dump_size);
649 #else
650 	(void)last_pa;
651 #endif
652 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
653 	/*
654 	 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
655 	 * When pmap_map() uses the direct map, they are not automatically
656 	 * included.
657 	 */
658 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
659 		dump_add_page(pa);
660 #endif
661 	phys_avail[biggestone + 1] = new_end;
662 #ifdef __amd64__
663 	/*
664 	 * Request that the physical pages underlying the message buffer be
665 	 * included in a crash dump.  Since the message buffer is accessed
666 	 * through the direct map, they are not automatically included.
667 	 */
668 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
669 	last_pa = pa + round_page(msgbufsize);
670 	while (pa < last_pa) {
671 		dump_add_page(pa);
672 		pa += PAGE_SIZE;
673 	}
674 #endif
675 	/*
676 	 * Compute the number of pages of memory that will be available for
677 	 * use, taking into account the overhead of a page structure per page.
678 	 * In other words, solve
679 	 *	"available physical memory" - round_page(page_range *
680 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
681 	 * for page_range.
682 	 */
683 	low_avail = phys_avail[0];
684 	high_avail = phys_avail[1];
685 	for (i = 0; i < vm_phys_nsegs; i++) {
686 		if (vm_phys_segs[i].start < low_avail)
687 			low_avail = vm_phys_segs[i].start;
688 		if (vm_phys_segs[i].end > high_avail)
689 			high_avail = vm_phys_segs[i].end;
690 	}
691 	/* Skip the first chunk.  It is already accounted for. */
692 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
693 		if (phys_avail[i] < low_avail)
694 			low_avail = phys_avail[i];
695 		if (phys_avail[i + 1] > high_avail)
696 			high_avail = phys_avail[i + 1];
697 	}
698 	first_page = low_avail / PAGE_SIZE;
699 #ifdef VM_PHYSSEG_SPARSE
700 	size = 0;
701 	for (i = 0; i < vm_phys_nsegs; i++)
702 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
703 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
704 		size += phys_avail[i + 1] - phys_avail[i];
705 #elif defined(VM_PHYSSEG_DENSE)
706 	size = high_avail - low_avail;
707 #else
708 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
709 #endif
710 
711 #ifdef VM_PHYSSEG_DENSE
712 	/*
713 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
714 	 * the overhead of a page structure per page only if vm_page_array is
715 	 * allocated from the last physical memory chunk.  Otherwise, we must
716 	 * allocate page structures representing the physical memory
717 	 * underlying vm_page_array, even though they will not be used.
718 	 */
719 	if (new_end != high_avail)
720 		page_range = size / PAGE_SIZE;
721 	else
722 #endif
723 	{
724 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
725 
726 		/*
727 		 * If the partial bytes remaining are large enough for
728 		 * a page (PAGE_SIZE) without a corresponding
729 		 * 'struct vm_page', then new_end will contain an
730 		 * extra page after subtracting the length of the VM
731 		 * page array.  Compensate by subtracting an extra
732 		 * page from new_end.
733 		 */
734 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
735 			if (new_end == high_avail)
736 				high_avail -= PAGE_SIZE;
737 			new_end -= PAGE_SIZE;
738 		}
739 	}
740 	end = new_end;
741 
742 	/*
743 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
744 	 * However, because this page is allocated from KVM, out-of-bounds
745 	 * accesses using the direct map will not be trapped.
746 	 */
747 	vaddr += PAGE_SIZE;
748 
749 	/*
750 	 * Allocate physical memory for the page structures, and map it.
751 	 */
752 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
753 	mapped = pmap_map(&vaddr, new_end, end,
754 	    VM_PROT_READ | VM_PROT_WRITE);
755 	vm_page_array = (vm_page_t)mapped;
756 	vm_page_array_size = page_range;
757 
758 #if VM_NRESERVLEVEL > 0
759 	/*
760 	 * Allocate physical memory for the reservation management system's
761 	 * data structures, and map it.
762 	 */
763 	if (high_avail == end)
764 		high_avail = new_end;
765 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
766 #endif
767 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
768 	/*
769 	 * Include vm_page_array and vm_reserv_array in a crash dump.
770 	 */
771 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
772 		dump_add_page(pa);
773 #endif
774 	phys_avail[biggestone + 1] = new_end;
775 
776 	/*
777 	 * Add physical memory segments corresponding to the available
778 	 * physical pages.
779 	 */
780 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
781 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
782 
783 	/*
784 	 * Initialize the physical memory allocator.
785 	 */
786 	vm_phys_init();
787 
788 	/*
789 	 * Initialize the page structures and add every available page to the
790 	 * physical memory allocator's free lists.
791 	 */
792 	vm_cnt.v_page_count = 0;
793 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
794 		seg = &vm_phys_segs[segind];
795 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
796 		    m++, pa += PAGE_SIZE)
797 			vm_page_init_page(m, pa, segind);
798 
799 		/*
800 		 * Add the segment to the free lists only if it is covered by
801 		 * one of the ranges in phys_avail.  Because we've added the
802 		 * ranges to the vm_phys_segs array, we can assume that each
803 		 * segment is either entirely contained in one of the ranges,
804 		 * or doesn't overlap any of them.
805 		 */
806 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
807 			struct vm_domain *vmd;
808 
809 			if (seg->start < phys_avail[i] ||
810 			    seg->end > phys_avail[i + 1])
811 				continue;
812 
813 			m = seg->first_page;
814 			pagecount = (u_long)atop(seg->end - seg->start);
815 
816 			vmd = VM_DOMAIN(seg->domain);
817 			vm_domain_free_lock(vmd);
818 			vm_phys_free_contig(m, pagecount);
819 			vm_domain_free_unlock(vmd);
820 			vm_domain_freecnt_inc(vmd, pagecount);
821 			vm_cnt.v_page_count += (u_int)pagecount;
822 
823 			vmd = VM_DOMAIN(seg->domain);
824 			vmd->vmd_page_count += (u_int)pagecount;
825 			vmd->vmd_segs |= 1UL << m->segind;
826 			break;
827 		}
828 	}
829 
830 	/*
831 	 * Remove blacklisted pages from the physical memory allocator.
832 	 */
833 	TAILQ_INIT(&blacklist_head);
834 	vm_page_blacklist_load(&list, &listend);
835 	vm_page_blacklist_check(list, listend);
836 
837 	list = kern_getenv("vm.blacklist");
838 	vm_page_blacklist_check(list, NULL);
839 
840 	freeenv(list);
841 #if VM_NRESERVLEVEL > 0
842 	/*
843 	 * Initialize the reservation management system.
844 	 */
845 	vm_reserv_init();
846 #endif
847 	/*
848 	 * Set an initial domain policy for thread0 so that allocations
849 	 * can work.
850 	 */
851 	domainset_zero();
852 
853 	return (vaddr);
854 }
855 
856 void
857 vm_page_reference(vm_page_t m)
858 {
859 
860 	vm_page_aflag_set(m, PGA_REFERENCED);
861 }
862 
863 /*
864  *	vm_page_busy_downgrade:
865  *
866  *	Downgrade an exclusive busy page into a single shared busy page.
867  */
868 void
869 vm_page_busy_downgrade(vm_page_t m)
870 {
871 	u_int x;
872 	bool locked;
873 
874 	vm_page_assert_xbusied(m);
875 	locked = mtx_owned(vm_page_lockptr(m));
876 
877 	for (;;) {
878 		x = m->busy_lock;
879 		x &= VPB_BIT_WAITERS;
880 		if (x != 0 && !locked)
881 			vm_page_lock(m);
882 		if (atomic_cmpset_rel_int(&m->busy_lock,
883 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
884 			break;
885 		if (x != 0 && !locked)
886 			vm_page_unlock(m);
887 	}
888 	if (x != 0) {
889 		wakeup(m);
890 		if (!locked)
891 			vm_page_unlock(m);
892 	}
893 }
894 
895 /*
896  *	vm_page_sbusied:
897  *
898  *	Return a positive value if the page is shared busied, 0 otherwise.
899  */
900 int
901 vm_page_sbusied(vm_page_t m)
902 {
903 	u_int x;
904 
905 	x = m->busy_lock;
906 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
907 }
908 
909 /*
910  *	vm_page_sunbusy:
911  *
912  *	Shared unbusy a page.
913  */
914 void
915 vm_page_sunbusy(vm_page_t m)
916 {
917 	u_int x;
918 
919 	vm_page_lock_assert(m, MA_NOTOWNED);
920 	vm_page_assert_sbusied(m);
921 
922 	for (;;) {
923 		x = m->busy_lock;
924 		if (VPB_SHARERS(x) > 1) {
925 			if (atomic_cmpset_int(&m->busy_lock, x,
926 			    x - VPB_ONE_SHARER))
927 				break;
928 			continue;
929 		}
930 		if ((x & VPB_BIT_WAITERS) == 0) {
931 			KASSERT(x == VPB_SHARERS_WORD(1),
932 			    ("vm_page_sunbusy: invalid lock state"));
933 			if (atomic_cmpset_int(&m->busy_lock,
934 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
935 				break;
936 			continue;
937 		}
938 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
939 		    ("vm_page_sunbusy: invalid lock state for waiters"));
940 
941 		vm_page_lock(m);
942 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
943 			vm_page_unlock(m);
944 			continue;
945 		}
946 		wakeup(m);
947 		vm_page_unlock(m);
948 		break;
949 	}
950 }
951 
952 /*
953  *	vm_page_busy_sleep:
954  *
955  *	Sleep and release the page lock, using the page pointer as wchan.
956  *	This is used to implement the hard-path of busying mechanism.
957  *
958  *	The given page must be locked.
959  *
960  *	If nonshared is true, sleep only if the page is xbusy.
961  */
962 void
963 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
964 {
965 	u_int x;
966 
967 	vm_page_assert_locked(m);
968 
969 	x = m->busy_lock;
970 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
971 	    ((x & VPB_BIT_WAITERS) == 0 &&
972 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
973 		vm_page_unlock(m);
974 		return;
975 	}
976 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
977 }
978 
979 /*
980  *	vm_page_trysbusy:
981  *
982  *	Try to shared busy a page.
983  *	If the operation succeeds 1 is returned otherwise 0.
984  *	The operation never sleeps.
985  */
986 int
987 vm_page_trysbusy(vm_page_t m)
988 {
989 	u_int x;
990 
991 	for (;;) {
992 		x = m->busy_lock;
993 		if ((x & VPB_BIT_SHARED) == 0)
994 			return (0);
995 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
996 			return (1);
997 	}
998 }
999 
1000 static void
1001 vm_page_xunbusy_locked(vm_page_t m)
1002 {
1003 
1004 	vm_page_assert_xbusied(m);
1005 	vm_page_assert_locked(m);
1006 
1007 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1008 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
1009 	wakeup(m);
1010 }
1011 
1012 void
1013 vm_page_xunbusy_maybelocked(vm_page_t m)
1014 {
1015 	bool lockacq;
1016 
1017 	vm_page_assert_xbusied(m);
1018 
1019 	/*
1020 	 * Fast path for unbusy.  If it succeeds, we know that there
1021 	 * are no waiters, so we do not need a wakeup.
1022 	 */
1023 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
1024 	    VPB_UNBUSIED))
1025 		return;
1026 
1027 	lockacq = !mtx_owned(vm_page_lockptr(m));
1028 	if (lockacq)
1029 		vm_page_lock(m);
1030 	vm_page_xunbusy_locked(m);
1031 	if (lockacq)
1032 		vm_page_unlock(m);
1033 }
1034 
1035 /*
1036  *	vm_page_xunbusy_hard:
1037  *
1038  *	Called after the first try the exclusive unbusy of a page failed.
1039  *	It is assumed that the waiters bit is on.
1040  */
1041 void
1042 vm_page_xunbusy_hard(vm_page_t m)
1043 {
1044 
1045 	vm_page_assert_xbusied(m);
1046 
1047 	vm_page_lock(m);
1048 	vm_page_xunbusy_locked(m);
1049 	vm_page_unlock(m);
1050 }
1051 
1052 /*
1053  *	vm_page_flash:
1054  *
1055  *	Wakeup anyone waiting for the page.
1056  *	The ownership bits do not change.
1057  *
1058  *	The given page must be locked.
1059  */
1060 void
1061 vm_page_flash(vm_page_t m)
1062 {
1063 	u_int x;
1064 
1065 	vm_page_lock_assert(m, MA_OWNED);
1066 
1067 	for (;;) {
1068 		x = m->busy_lock;
1069 		if ((x & VPB_BIT_WAITERS) == 0)
1070 			return;
1071 		if (atomic_cmpset_int(&m->busy_lock, x,
1072 		    x & (~VPB_BIT_WAITERS)))
1073 			break;
1074 	}
1075 	wakeup(m);
1076 }
1077 
1078 /*
1079  * Avoid releasing and reacquiring the same page lock.
1080  */
1081 void
1082 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1083 {
1084 	struct mtx *mtx1;
1085 
1086 	mtx1 = vm_page_lockptr(m);
1087 	if (*mtx == mtx1)
1088 		return;
1089 	if (*mtx != NULL)
1090 		mtx_unlock(*mtx);
1091 	*mtx = mtx1;
1092 	mtx_lock(mtx1);
1093 }
1094 
1095 /*
1096  * Keep page from being freed by the page daemon
1097  * much of the same effect as wiring, except much lower
1098  * overhead and should be used only for *very* temporary
1099  * holding ("wiring").
1100  */
1101 void
1102 vm_page_hold(vm_page_t mem)
1103 {
1104 
1105 	vm_page_lock_assert(mem, MA_OWNED);
1106         mem->hold_count++;
1107 }
1108 
1109 void
1110 vm_page_unhold(vm_page_t mem)
1111 {
1112 
1113 	vm_page_lock_assert(mem, MA_OWNED);
1114 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1115 	--mem->hold_count;
1116 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1117 		vm_page_free_toq(mem);
1118 }
1119 
1120 /*
1121  *	vm_page_unhold_pages:
1122  *
1123  *	Unhold each of the pages that is referenced by the given array.
1124  */
1125 void
1126 vm_page_unhold_pages(vm_page_t *ma, int count)
1127 {
1128 	struct mtx *mtx;
1129 
1130 	mtx = NULL;
1131 	for (; count != 0; count--) {
1132 		vm_page_change_lock(*ma, &mtx);
1133 		vm_page_unhold(*ma);
1134 		ma++;
1135 	}
1136 	if (mtx != NULL)
1137 		mtx_unlock(mtx);
1138 }
1139 
1140 vm_page_t
1141 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1142 {
1143 	vm_page_t m;
1144 
1145 #ifdef VM_PHYSSEG_SPARSE
1146 	m = vm_phys_paddr_to_vm_page(pa);
1147 	if (m == NULL)
1148 		m = vm_phys_fictitious_to_vm_page(pa);
1149 	return (m);
1150 #elif defined(VM_PHYSSEG_DENSE)
1151 	long pi;
1152 
1153 	pi = atop(pa);
1154 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1155 		m = &vm_page_array[pi - first_page];
1156 		return (m);
1157 	}
1158 	return (vm_phys_fictitious_to_vm_page(pa));
1159 #else
1160 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1161 #endif
1162 }
1163 
1164 /*
1165  *	vm_page_getfake:
1166  *
1167  *	Create a fictitious page with the specified physical address and
1168  *	memory attribute.  The memory attribute is the only the machine-
1169  *	dependent aspect of a fictitious page that must be initialized.
1170  */
1171 vm_page_t
1172 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1173 {
1174 	vm_page_t m;
1175 
1176 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1177 	vm_page_initfake(m, paddr, memattr);
1178 	return (m);
1179 }
1180 
1181 void
1182 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1183 {
1184 
1185 	if ((m->flags & PG_FICTITIOUS) != 0) {
1186 		/*
1187 		 * The page's memattr might have changed since the
1188 		 * previous initialization.  Update the pmap to the
1189 		 * new memattr.
1190 		 */
1191 		goto memattr;
1192 	}
1193 	m->phys_addr = paddr;
1194 	m->queue = PQ_NONE;
1195 	/* Fictitious pages don't use "segind". */
1196 	m->flags = PG_FICTITIOUS;
1197 	/* Fictitious pages don't use "order" or "pool". */
1198 	m->oflags = VPO_UNMANAGED;
1199 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1200 	m->wire_count = 1;
1201 	pmap_page_init(m);
1202 memattr:
1203 	pmap_page_set_memattr(m, memattr);
1204 }
1205 
1206 /*
1207  *	vm_page_putfake:
1208  *
1209  *	Release a fictitious page.
1210  */
1211 void
1212 vm_page_putfake(vm_page_t m)
1213 {
1214 
1215 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1216 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1217 	    ("vm_page_putfake: bad page %p", m));
1218 	uma_zfree(fakepg_zone, m);
1219 }
1220 
1221 /*
1222  *	vm_page_updatefake:
1223  *
1224  *	Update the given fictitious page to the specified physical address and
1225  *	memory attribute.
1226  */
1227 void
1228 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1229 {
1230 
1231 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1232 	    ("vm_page_updatefake: bad page %p", m));
1233 	m->phys_addr = paddr;
1234 	pmap_page_set_memattr(m, memattr);
1235 }
1236 
1237 /*
1238  *	vm_page_free:
1239  *
1240  *	Free a page.
1241  */
1242 void
1243 vm_page_free(vm_page_t m)
1244 {
1245 
1246 	m->flags &= ~PG_ZERO;
1247 	vm_page_free_toq(m);
1248 }
1249 
1250 /*
1251  *	vm_page_free_zero:
1252  *
1253  *	Free a page to the zerod-pages queue
1254  */
1255 void
1256 vm_page_free_zero(vm_page_t m)
1257 {
1258 
1259 	m->flags |= PG_ZERO;
1260 	vm_page_free_toq(m);
1261 }
1262 
1263 /*
1264  * Unbusy and handle the page queueing for a page from a getpages request that
1265  * was optionally read ahead or behind.
1266  */
1267 void
1268 vm_page_readahead_finish(vm_page_t m)
1269 {
1270 
1271 	/* We shouldn't put invalid pages on queues. */
1272 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1273 
1274 	/*
1275 	 * Since the page is not the actually needed one, whether it should
1276 	 * be activated or deactivated is not obvious.  Empirical results
1277 	 * have shown that deactivating the page is usually the best choice,
1278 	 * unless the page is wanted by another thread.
1279 	 */
1280 	vm_page_lock(m);
1281 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1282 		vm_page_activate(m);
1283 	else
1284 		vm_page_deactivate(m);
1285 	vm_page_unlock(m);
1286 	vm_page_xunbusy(m);
1287 }
1288 
1289 /*
1290  *	vm_page_sleep_if_busy:
1291  *
1292  *	Sleep and release the page queues lock if the page is busied.
1293  *	Returns TRUE if the thread slept.
1294  *
1295  *	The given page must be unlocked and object containing it must
1296  *	be locked.
1297  */
1298 int
1299 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1300 {
1301 	vm_object_t obj;
1302 
1303 	vm_page_lock_assert(m, MA_NOTOWNED);
1304 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1305 
1306 	if (vm_page_busied(m)) {
1307 		/*
1308 		 * The page-specific object must be cached because page
1309 		 * identity can change during the sleep, causing the
1310 		 * re-lock of a different object.
1311 		 * It is assumed that a reference to the object is already
1312 		 * held by the callers.
1313 		 */
1314 		obj = m->object;
1315 		vm_page_lock(m);
1316 		VM_OBJECT_WUNLOCK(obj);
1317 		vm_page_busy_sleep(m, msg, false);
1318 		VM_OBJECT_WLOCK(obj);
1319 		return (TRUE);
1320 	}
1321 	return (FALSE);
1322 }
1323 
1324 /*
1325  *	vm_page_dirty_KBI:		[ internal use only ]
1326  *
1327  *	Set all bits in the page's dirty field.
1328  *
1329  *	The object containing the specified page must be locked if the
1330  *	call is made from the machine-independent layer.
1331  *
1332  *	See vm_page_clear_dirty_mask().
1333  *
1334  *	This function should only be called by vm_page_dirty().
1335  */
1336 void
1337 vm_page_dirty_KBI(vm_page_t m)
1338 {
1339 
1340 	/* Refer to this operation by its public name. */
1341 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1342 	    ("vm_page_dirty: page is invalid!"));
1343 	m->dirty = VM_PAGE_BITS_ALL;
1344 }
1345 
1346 /*
1347  *	vm_page_insert:		[ internal use only ]
1348  *
1349  *	Inserts the given mem entry into the object and object list.
1350  *
1351  *	The object must be locked.
1352  */
1353 int
1354 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1355 {
1356 	vm_page_t mpred;
1357 
1358 	VM_OBJECT_ASSERT_WLOCKED(object);
1359 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1360 	return (vm_page_insert_after(m, object, pindex, mpred));
1361 }
1362 
1363 /*
1364  *	vm_page_insert_after:
1365  *
1366  *	Inserts the page "m" into the specified object at offset "pindex".
1367  *
1368  *	The page "mpred" must immediately precede the offset "pindex" within
1369  *	the specified object.
1370  *
1371  *	The object must be locked.
1372  */
1373 static int
1374 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1375     vm_page_t mpred)
1376 {
1377 	vm_page_t msucc;
1378 
1379 	VM_OBJECT_ASSERT_WLOCKED(object);
1380 	KASSERT(m->object == NULL,
1381 	    ("vm_page_insert_after: page already inserted"));
1382 	if (mpred != NULL) {
1383 		KASSERT(mpred->object == object,
1384 		    ("vm_page_insert_after: object doesn't contain mpred"));
1385 		KASSERT(mpred->pindex < pindex,
1386 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1387 		msucc = TAILQ_NEXT(mpred, listq);
1388 	} else
1389 		msucc = TAILQ_FIRST(&object->memq);
1390 	if (msucc != NULL)
1391 		KASSERT(msucc->pindex > pindex,
1392 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1393 
1394 	/*
1395 	 * Record the object/offset pair in this page
1396 	 */
1397 	m->object = object;
1398 	m->pindex = pindex;
1399 
1400 	/*
1401 	 * Now link into the object's ordered list of backed pages.
1402 	 */
1403 	if (vm_radix_insert(&object->rtree, m)) {
1404 		m->object = NULL;
1405 		m->pindex = 0;
1406 		return (1);
1407 	}
1408 	vm_page_insert_radixdone(m, object, mpred);
1409 	return (0);
1410 }
1411 
1412 /*
1413  *	vm_page_insert_radixdone:
1414  *
1415  *	Complete page "m" insertion into the specified object after the
1416  *	radix trie hooking.
1417  *
1418  *	The page "mpred" must precede the offset "m->pindex" within the
1419  *	specified object.
1420  *
1421  *	The object must be locked.
1422  */
1423 static void
1424 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1425 {
1426 
1427 	VM_OBJECT_ASSERT_WLOCKED(object);
1428 	KASSERT(object != NULL && m->object == object,
1429 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1430 	if (mpred != NULL) {
1431 		KASSERT(mpred->object == object,
1432 		    ("vm_page_insert_after: object doesn't contain mpred"));
1433 		KASSERT(mpred->pindex < m->pindex,
1434 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1435 	}
1436 
1437 	if (mpred != NULL)
1438 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1439 	else
1440 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1441 
1442 	/*
1443 	 * Show that the object has one more resident page.
1444 	 */
1445 	object->resident_page_count++;
1446 
1447 	/*
1448 	 * Hold the vnode until the last page is released.
1449 	 */
1450 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1451 		vhold(object->handle);
1452 
1453 	/*
1454 	 * Since we are inserting a new and possibly dirty page,
1455 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1456 	 */
1457 	if (pmap_page_is_write_mapped(m))
1458 		vm_object_set_writeable_dirty(object);
1459 }
1460 
1461 /*
1462  *	vm_page_remove:
1463  *
1464  *	Removes the specified page from its containing object, but does not
1465  *	invalidate any backing storage.
1466  *
1467  *	The object must be locked.  The page must be locked if it is managed.
1468  */
1469 void
1470 vm_page_remove(vm_page_t m)
1471 {
1472 	vm_object_t object;
1473 	vm_page_t mrem;
1474 
1475 	if ((m->oflags & VPO_UNMANAGED) == 0)
1476 		vm_page_assert_locked(m);
1477 	if ((object = m->object) == NULL)
1478 		return;
1479 	VM_OBJECT_ASSERT_WLOCKED(object);
1480 	if (vm_page_xbusied(m))
1481 		vm_page_xunbusy_maybelocked(m);
1482 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1483 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1484 
1485 	/*
1486 	 * Now remove from the object's list of backed pages.
1487 	 */
1488 	TAILQ_REMOVE(&object->memq, m, listq);
1489 
1490 	/*
1491 	 * And show that the object has one fewer resident page.
1492 	 */
1493 	object->resident_page_count--;
1494 
1495 	/*
1496 	 * The vnode may now be recycled.
1497 	 */
1498 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1499 		vdrop(object->handle);
1500 
1501 	m->object = NULL;
1502 }
1503 
1504 /*
1505  *	vm_page_lookup:
1506  *
1507  *	Returns the page associated with the object/offset
1508  *	pair specified; if none is found, NULL is returned.
1509  *
1510  *	The object must be locked.
1511  */
1512 vm_page_t
1513 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1514 {
1515 
1516 	VM_OBJECT_ASSERT_LOCKED(object);
1517 	return (vm_radix_lookup(&object->rtree, pindex));
1518 }
1519 
1520 /*
1521  *	vm_page_find_least:
1522  *
1523  *	Returns the page associated with the object with least pindex
1524  *	greater than or equal to the parameter pindex, or NULL.
1525  *
1526  *	The object must be locked.
1527  */
1528 vm_page_t
1529 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1530 {
1531 	vm_page_t m;
1532 
1533 	VM_OBJECT_ASSERT_LOCKED(object);
1534 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1535 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1536 	return (m);
1537 }
1538 
1539 /*
1540  * Returns the given page's successor (by pindex) within the object if it is
1541  * resident; if none is found, NULL is returned.
1542  *
1543  * The object must be locked.
1544  */
1545 vm_page_t
1546 vm_page_next(vm_page_t m)
1547 {
1548 	vm_page_t next;
1549 
1550 	VM_OBJECT_ASSERT_LOCKED(m->object);
1551 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1552 		MPASS(next->object == m->object);
1553 		if (next->pindex != m->pindex + 1)
1554 			next = NULL;
1555 	}
1556 	return (next);
1557 }
1558 
1559 /*
1560  * Returns the given page's predecessor (by pindex) within the object if it is
1561  * resident; if none is found, NULL is returned.
1562  *
1563  * The object must be locked.
1564  */
1565 vm_page_t
1566 vm_page_prev(vm_page_t m)
1567 {
1568 	vm_page_t prev;
1569 
1570 	VM_OBJECT_ASSERT_LOCKED(m->object);
1571 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1572 		MPASS(prev->object == m->object);
1573 		if (prev->pindex != m->pindex - 1)
1574 			prev = NULL;
1575 	}
1576 	return (prev);
1577 }
1578 
1579 /*
1580  * Uses the page mnew as a replacement for an existing page at index
1581  * pindex which must be already present in the object.
1582  *
1583  * The existing page must not be on a paging queue.
1584  */
1585 vm_page_t
1586 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1587 {
1588 	vm_page_t mold;
1589 
1590 	VM_OBJECT_ASSERT_WLOCKED(object);
1591 	KASSERT(mnew->object == NULL,
1592 	    ("vm_page_replace: page %p already in object", mnew));
1593 	KASSERT(mnew->queue == PQ_NONE,
1594 	    ("vm_page_replace: new page %p is on a paging queue", mnew));
1595 
1596 	/*
1597 	 * This function mostly follows vm_page_insert() and
1598 	 * vm_page_remove() without the radix, object count and vnode
1599 	 * dance.  Double check such functions for more comments.
1600 	 */
1601 
1602 	mnew->object = object;
1603 	mnew->pindex = pindex;
1604 	mold = vm_radix_replace(&object->rtree, mnew);
1605 	KASSERT(mold->queue == PQ_NONE,
1606 	    ("vm_page_replace: old page %p is on a paging queue", mold));
1607 
1608 	/* Keep the resident page list in sorted order. */
1609 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1610 	TAILQ_REMOVE(&object->memq, mold, listq);
1611 
1612 	mold->object = NULL;
1613 	vm_page_xunbusy_maybelocked(mold);
1614 
1615 	/*
1616 	 * The object's resident_page_count does not change because we have
1617 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1618 	 */
1619 	if (pmap_page_is_write_mapped(mnew))
1620 		vm_object_set_writeable_dirty(object);
1621 	return (mold);
1622 }
1623 
1624 /*
1625  *	vm_page_rename:
1626  *
1627  *	Move the given memory entry from its
1628  *	current object to the specified target object/offset.
1629  *
1630  *	Note: swap associated with the page must be invalidated by the move.  We
1631  *	      have to do this for several reasons:  (1) we aren't freeing the
1632  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1633  *	      moving the page from object A to B, and will then later move
1634  *	      the backing store from A to B and we can't have a conflict.
1635  *
1636  *	Note: we *always* dirty the page.  It is necessary both for the
1637  *	      fact that we moved it, and because we may be invalidating
1638  *	      swap.
1639  *
1640  *	The objects must be locked.
1641  */
1642 int
1643 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1644 {
1645 	vm_page_t mpred;
1646 	vm_pindex_t opidx;
1647 
1648 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1649 
1650 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1651 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1652 	    ("vm_page_rename: pindex already renamed"));
1653 
1654 	/*
1655 	 * Create a custom version of vm_page_insert() which does not depend
1656 	 * by m_prev and can cheat on the implementation aspects of the
1657 	 * function.
1658 	 */
1659 	opidx = m->pindex;
1660 	m->pindex = new_pindex;
1661 	if (vm_radix_insert(&new_object->rtree, m)) {
1662 		m->pindex = opidx;
1663 		return (1);
1664 	}
1665 
1666 	/*
1667 	 * The operation cannot fail anymore.  The removal must happen before
1668 	 * the listq iterator is tainted.
1669 	 */
1670 	m->pindex = opidx;
1671 	vm_page_lock(m);
1672 	vm_page_remove(m);
1673 
1674 	/* Return back to the new pindex to complete vm_page_insert(). */
1675 	m->pindex = new_pindex;
1676 	m->object = new_object;
1677 	vm_page_unlock(m);
1678 	vm_page_insert_radixdone(m, new_object, mpred);
1679 	vm_page_dirty(m);
1680 	return (0);
1681 }
1682 
1683 /*
1684  *	vm_page_alloc:
1685  *
1686  *	Allocate and return a page that is associated with the specified
1687  *	object and offset pair.  By default, this page is exclusive busied.
1688  *
1689  *	The caller must always specify an allocation class.
1690  *
1691  *	allocation classes:
1692  *	VM_ALLOC_NORMAL		normal process request
1693  *	VM_ALLOC_SYSTEM		system *really* needs a page
1694  *	VM_ALLOC_INTERRUPT	interrupt time request
1695  *
1696  *	optional allocation flags:
1697  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1698  *				intends to allocate
1699  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1700  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1701  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1702  *				should not be exclusive busy
1703  *	VM_ALLOC_SBUSY		shared busy the allocated page
1704  *	VM_ALLOC_WIRED		wire the allocated page
1705  *	VM_ALLOC_ZERO		prefer a zeroed page
1706  */
1707 vm_page_t
1708 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1709 {
1710 
1711 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1712 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1713 }
1714 
1715 vm_page_t
1716 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1717     int req)
1718 {
1719 
1720 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1721 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1722 	    NULL));
1723 }
1724 
1725 /*
1726  * Allocate a page in the specified object with the given page index.  To
1727  * optimize insertion of the page into the object, the caller must also specifiy
1728  * the resident page in the object with largest index smaller than the given
1729  * page index, or NULL if no such page exists.
1730  */
1731 vm_page_t
1732 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1733     int req, vm_page_t mpred)
1734 {
1735 	struct vm_domainset_iter di;
1736 	vm_page_t m;
1737 	int domain;
1738 
1739 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1740 	do {
1741 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1742 		    mpred);
1743 		if (m != NULL)
1744 			break;
1745 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1746 
1747 	return (m);
1748 }
1749 
1750 /*
1751  * Returns true if the number of free pages exceeds the minimum
1752  * for the request class and false otherwise.
1753  */
1754 int
1755 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1756 {
1757 	u_int limit, old, new;
1758 
1759 	req = req & VM_ALLOC_CLASS_MASK;
1760 
1761 	/*
1762 	 * The page daemon is allowed to dig deeper into the free page list.
1763 	 */
1764 	if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1765 		req = VM_ALLOC_SYSTEM;
1766 	if (req == VM_ALLOC_INTERRUPT)
1767 		limit = 0;
1768 	else if (req == VM_ALLOC_SYSTEM)
1769 		limit = vmd->vmd_interrupt_free_min;
1770 	else
1771 		limit = vmd->vmd_free_reserved;
1772 
1773 	/*
1774 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1775 	 */
1776 	limit += npages;
1777 	old = vmd->vmd_free_count;
1778 	do {
1779 		if (old < limit)
1780 			return (0);
1781 		new = old - npages;
1782 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1783 
1784 	/* Wake the page daemon if we've crossed the threshold. */
1785 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1786 		pagedaemon_wakeup(vmd->vmd_domain);
1787 
1788 	/* Only update bitsets on transitions. */
1789 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1790 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1791 		vm_domain_set(vmd);
1792 
1793 	return (1);
1794 }
1795 
1796 vm_page_t
1797 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1798     int req, vm_page_t mpred)
1799 {
1800 	struct vm_domain *vmd;
1801 	vm_page_t m;
1802 	int flags;
1803 
1804 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1805 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1806 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1807 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1808 	    ("inconsistent object(%p)/req(%x)", object, req));
1809 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1810 	    ("Can't sleep and retry object insertion."));
1811 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1812 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1813 	    (uintmax_t)pindex));
1814 	if (object != NULL)
1815 		VM_OBJECT_ASSERT_WLOCKED(object);
1816 
1817 again:
1818 	m = NULL;
1819 #if VM_NRESERVLEVEL > 0
1820 	/*
1821 	 * Can we allocate the page from a reservation?
1822 	 */
1823 	if (vm_object_reserv(object) &&
1824 	    ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL ||
1825 	    (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) {
1826 		domain = vm_phys_domain(m);
1827 		vmd = VM_DOMAIN(domain);
1828 		goto found;
1829 	}
1830 #endif
1831 	vmd = VM_DOMAIN(domain);
1832 	if (object != NULL && vmd->vmd_pgcache != NULL) {
1833 		m = uma_zalloc(vmd->vmd_pgcache, M_NOWAIT);
1834 		if (m != NULL)
1835 			goto found;
1836 	}
1837 	if (vm_domain_allocate(vmd, req, 1)) {
1838 		/*
1839 		 * If not, allocate it from the free page queues.
1840 		 */
1841 		vm_domain_free_lock(vmd);
1842 		m = vm_phys_alloc_pages(domain, object != NULL ?
1843 		    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1844 		vm_domain_free_unlock(vmd);
1845 		if (m == NULL) {
1846 			vm_domain_freecnt_inc(vmd, 1);
1847 #if VM_NRESERVLEVEL > 0
1848 			if (vm_reserv_reclaim_inactive(domain))
1849 				goto again;
1850 #endif
1851 		}
1852 	}
1853 	if (m == NULL) {
1854 		/*
1855 		 * Not allocatable, give up.
1856 		 */
1857 		if (vm_domain_alloc_fail(vmd, object, req))
1858 			goto again;
1859 		return (NULL);
1860 	}
1861 
1862 	/*
1863 	 *  At this point we had better have found a good page.
1864 	 */
1865 	KASSERT(m != NULL, ("missing page"));
1866 
1867 found:
1868 	vm_page_dequeue(m);
1869 	vm_page_alloc_check(m);
1870 
1871 	/*
1872 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1873 	 */
1874 	flags = 0;
1875 	if ((req & VM_ALLOC_ZERO) != 0)
1876 		flags = PG_ZERO;
1877 	flags &= m->flags;
1878 	if ((req & VM_ALLOC_NODUMP) != 0)
1879 		flags |= PG_NODUMP;
1880 	m->flags = flags;
1881 	m->aflags = 0;
1882 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1883 	    VPO_UNMANAGED : 0;
1884 	m->busy_lock = VPB_UNBUSIED;
1885 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1886 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1887 	if ((req & VM_ALLOC_SBUSY) != 0)
1888 		m->busy_lock = VPB_SHARERS_WORD(1);
1889 	if (req & VM_ALLOC_WIRED) {
1890 		/*
1891 		 * The page lock is not required for wiring a page until that
1892 		 * page is inserted into the object.
1893 		 */
1894 		vm_wire_add(1);
1895 		m->wire_count = 1;
1896 	}
1897 	m->act_count = 0;
1898 
1899 	if (object != NULL) {
1900 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1901 			if (req & VM_ALLOC_WIRED) {
1902 				vm_wire_sub(1);
1903 				m->wire_count = 0;
1904 			}
1905 			KASSERT(m->object == NULL, ("page %p has object", m));
1906 			m->oflags = VPO_UNMANAGED;
1907 			m->busy_lock = VPB_UNBUSIED;
1908 			/* Don't change PG_ZERO. */
1909 			vm_page_free_toq(m);
1910 			if (req & VM_ALLOC_WAITFAIL) {
1911 				VM_OBJECT_WUNLOCK(object);
1912 				vm_radix_wait();
1913 				VM_OBJECT_WLOCK(object);
1914 			}
1915 			return (NULL);
1916 		}
1917 
1918 		/* Ignore device objects; the pager sets "memattr" for them. */
1919 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1920 		    (object->flags & OBJ_FICTITIOUS) == 0)
1921 			pmap_page_set_memattr(m, object->memattr);
1922 	} else
1923 		m->pindex = pindex;
1924 
1925 	return (m);
1926 }
1927 
1928 /*
1929  *	vm_page_alloc_contig:
1930  *
1931  *	Allocate a contiguous set of physical pages of the given size "npages"
1932  *	from the free lists.  All of the physical pages must be at or above
1933  *	the given physical address "low" and below the given physical address
1934  *	"high".  The given value "alignment" determines the alignment of the
1935  *	first physical page in the set.  If the given value "boundary" is
1936  *	non-zero, then the set of physical pages cannot cross any physical
1937  *	address boundary that is a multiple of that value.  Both "alignment"
1938  *	and "boundary" must be a power of two.
1939  *
1940  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1941  *	then the memory attribute setting for the physical pages is configured
1942  *	to the object's memory attribute setting.  Otherwise, the memory
1943  *	attribute setting for the physical pages is configured to "memattr",
1944  *	overriding the object's memory attribute setting.  However, if the
1945  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1946  *	memory attribute setting for the physical pages cannot be configured
1947  *	to VM_MEMATTR_DEFAULT.
1948  *
1949  *	The specified object may not contain fictitious pages.
1950  *
1951  *	The caller must always specify an allocation class.
1952  *
1953  *	allocation classes:
1954  *	VM_ALLOC_NORMAL		normal process request
1955  *	VM_ALLOC_SYSTEM		system *really* needs a page
1956  *	VM_ALLOC_INTERRUPT	interrupt time request
1957  *
1958  *	optional allocation flags:
1959  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1960  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1961  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1962  *				should not be exclusive busy
1963  *	VM_ALLOC_SBUSY		shared busy the allocated page
1964  *	VM_ALLOC_WIRED		wire the allocated page
1965  *	VM_ALLOC_ZERO		prefer a zeroed page
1966  */
1967 vm_page_t
1968 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1969     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1970     vm_paddr_t boundary, vm_memattr_t memattr)
1971 {
1972 	struct vm_domainset_iter di;
1973 	vm_page_t m;
1974 	int domain;
1975 
1976 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1977 	do {
1978 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1979 		    npages, low, high, alignment, boundary, memattr);
1980 		if (m != NULL)
1981 			break;
1982 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1983 
1984 	return (m);
1985 }
1986 
1987 vm_page_t
1988 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1989     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1990     vm_paddr_t boundary, vm_memattr_t memattr)
1991 {
1992 	struct vm_domain *vmd;
1993 	vm_page_t m, m_ret, mpred;
1994 	u_int busy_lock, flags, oflags;
1995 
1996 	mpred = NULL;	/* XXX: pacify gcc */
1997 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1998 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1999 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2000 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2001 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2002 	    req));
2003 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2004 	    ("Can't sleep and retry object insertion."));
2005 	if (object != NULL) {
2006 		VM_OBJECT_ASSERT_WLOCKED(object);
2007 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2008 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2009 		    object));
2010 	}
2011 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2012 
2013 	if (object != NULL) {
2014 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2015 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2016 		    ("vm_page_alloc_contig: pindex already allocated"));
2017 	}
2018 
2019 	/*
2020 	 * Can we allocate the pages without the number of free pages falling
2021 	 * below the lower bound for the allocation class?
2022 	 */
2023 again:
2024 #if VM_NRESERVLEVEL > 0
2025 	/*
2026 	 * Can we allocate the pages from a reservation?
2027 	 */
2028 	if (vm_object_reserv(object) &&
2029 	    ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
2030 	    npages, low, high, alignment, boundary, mpred)) != NULL ||
2031 	    (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain,
2032 	    npages, low, high, alignment, boundary, mpred)) != NULL)) {
2033 		domain = vm_phys_domain(m_ret);
2034 		vmd = VM_DOMAIN(domain);
2035 		goto found;
2036 	}
2037 #endif
2038 	m_ret = NULL;
2039 	vmd = VM_DOMAIN(domain);
2040 	if (vm_domain_allocate(vmd, req, npages)) {
2041 		/*
2042 		 * allocate them from the free page queues.
2043 		 */
2044 		vm_domain_free_lock(vmd);
2045 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2046 		    alignment, boundary);
2047 		vm_domain_free_unlock(vmd);
2048 		if (m_ret == NULL) {
2049 			vm_domain_freecnt_inc(vmd, npages);
2050 #if VM_NRESERVLEVEL > 0
2051 			if (vm_reserv_reclaim_contig(domain, npages, low,
2052 			    high, alignment, boundary))
2053 				goto again;
2054 #endif
2055 		}
2056 	}
2057 	if (m_ret == NULL) {
2058 		if (vm_domain_alloc_fail(vmd, object, req))
2059 			goto again;
2060 		return (NULL);
2061 	}
2062 #if VM_NRESERVLEVEL > 0
2063 found:
2064 #endif
2065 	for (m = m_ret; m < &m_ret[npages]; m++) {
2066 		vm_page_dequeue(m);
2067 		vm_page_alloc_check(m);
2068 	}
2069 
2070 	/*
2071 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2072 	 */
2073 	flags = 0;
2074 	if ((req & VM_ALLOC_ZERO) != 0)
2075 		flags = PG_ZERO;
2076 	if ((req & VM_ALLOC_NODUMP) != 0)
2077 		flags |= PG_NODUMP;
2078 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2079 	    VPO_UNMANAGED : 0;
2080 	busy_lock = VPB_UNBUSIED;
2081 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2082 		busy_lock = VPB_SINGLE_EXCLUSIVER;
2083 	if ((req & VM_ALLOC_SBUSY) != 0)
2084 		busy_lock = VPB_SHARERS_WORD(1);
2085 	if ((req & VM_ALLOC_WIRED) != 0)
2086 		vm_wire_add(npages);
2087 	if (object != NULL) {
2088 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2089 		    memattr == VM_MEMATTR_DEFAULT)
2090 			memattr = object->memattr;
2091 	}
2092 	for (m = m_ret; m < &m_ret[npages]; m++) {
2093 		m->aflags = 0;
2094 		m->flags = (m->flags | PG_NODUMP) & flags;
2095 		m->busy_lock = busy_lock;
2096 		if ((req & VM_ALLOC_WIRED) != 0)
2097 			m->wire_count = 1;
2098 		m->act_count = 0;
2099 		m->oflags = oflags;
2100 		if (object != NULL) {
2101 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2102 				if ((req & VM_ALLOC_WIRED) != 0)
2103 					vm_wire_sub(npages);
2104 				KASSERT(m->object == NULL,
2105 				    ("page %p has object", m));
2106 				mpred = m;
2107 				for (m = m_ret; m < &m_ret[npages]; m++) {
2108 					if (m <= mpred &&
2109 					    (req & VM_ALLOC_WIRED) != 0)
2110 						m->wire_count = 0;
2111 					m->oflags = VPO_UNMANAGED;
2112 					m->busy_lock = VPB_UNBUSIED;
2113 					/* Don't change PG_ZERO. */
2114 					vm_page_free_toq(m);
2115 				}
2116 				if (req & VM_ALLOC_WAITFAIL) {
2117 					VM_OBJECT_WUNLOCK(object);
2118 					vm_radix_wait();
2119 					VM_OBJECT_WLOCK(object);
2120 				}
2121 				return (NULL);
2122 			}
2123 			mpred = m;
2124 		} else
2125 			m->pindex = pindex;
2126 		if (memattr != VM_MEMATTR_DEFAULT)
2127 			pmap_page_set_memattr(m, memattr);
2128 		pindex++;
2129 	}
2130 	return (m_ret);
2131 }
2132 
2133 /*
2134  * Check a page that has been freshly dequeued from a freelist.
2135  */
2136 static void
2137 vm_page_alloc_check(vm_page_t m)
2138 {
2139 
2140 	KASSERT(m->object == NULL, ("page %p has object", m));
2141 	KASSERT(m->queue == PQ_NONE,
2142 	    ("page %p has unexpected queue %d", m, m->queue));
2143 	KASSERT(!vm_page_held(m), ("page %p is held", m));
2144 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2145 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2146 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2147 	    ("page %p has unexpected memattr %d",
2148 	    m, pmap_page_get_memattr(m)));
2149 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2150 }
2151 
2152 /*
2153  * 	vm_page_alloc_freelist:
2154  *
2155  *	Allocate a physical page from the specified free page list.
2156  *
2157  *	The caller must always specify an allocation class.
2158  *
2159  *	allocation classes:
2160  *	VM_ALLOC_NORMAL		normal process request
2161  *	VM_ALLOC_SYSTEM		system *really* needs a page
2162  *	VM_ALLOC_INTERRUPT	interrupt time request
2163  *
2164  *	optional allocation flags:
2165  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2166  *				intends to allocate
2167  *	VM_ALLOC_WIRED		wire the allocated page
2168  *	VM_ALLOC_ZERO		prefer a zeroed page
2169  */
2170 vm_page_t
2171 vm_page_alloc_freelist(int freelist, int req)
2172 {
2173 	struct vm_domainset_iter di;
2174 	vm_page_t m;
2175 	int domain;
2176 
2177 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2178 	do {
2179 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2180 		if (m != NULL)
2181 			break;
2182 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2183 
2184 	return (m);
2185 }
2186 
2187 vm_page_t
2188 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2189 {
2190 	struct vm_domain *vmd;
2191 	vm_page_t m;
2192 	u_int flags;
2193 
2194 	/*
2195 	 * Do not allocate reserved pages unless the req has asked for it.
2196 	 */
2197 	vmd = VM_DOMAIN(domain);
2198 again:
2199 	if (vm_domain_allocate(vmd, req, 1)) {
2200 		vm_domain_free_lock(vmd);
2201 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2202 		    VM_FREEPOOL_DIRECT, 0);
2203 		vm_domain_free_unlock(vmd);
2204 		if (m == NULL)
2205 			vm_domain_freecnt_inc(vmd, 1);
2206 	}
2207 	if (m == NULL) {
2208 		if (vm_domain_alloc_fail(vmd, NULL, req))
2209 			goto again;
2210 		return (NULL);
2211 	}
2212 	vm_page_dequeue(m);
2213 	vm_page_alloc_check(m);
2214 
2215 	/*
2216 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2217 	 */
2218 	m->aflags = 0;
2219 	flags = 0;
2220 	if ((req & VM_ALLOC_ZERO) != 0)
2221 		flags = PG_ZERO;
2222 	m->flags &= flags;
2223 	if ((req & VM_ALLOC_WIRED) != 0) {
2224 		/*
2225 		 * The page lock is not required for wiring a page that does
2226 		 * not belong to an object.
2227 		 */
2228 		vm_wire_add(1);
2229 		m->wire_count = 1;
2230 	}
2231 	/* Unmanaged pages don't use "act_count". */
2232 	m->oflags = VPO_UNMANAGED;
2233 	return (m);
2234 }
2235 
2236 static int
2237 vm_page_import(void *arg, void **store, int cnt, int domain, int flags)
2238 {
2239 	struct vm_domain *vmd;
2240 	vm_page_t m;
2241 	int i, j, n;
2242 
2243 	vmd = arg;
2244 	/* Only import if we can bring in a full bucket. */
2245 	if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2246 		return (0);
2247 	domain = vmd->vmd_domain;
2248 	n = 64;	/* Starting stride, arbitrary. */
2249 	vm_domain_free_lock(vmd);
2250 	for (i = 0; i < cnt; i+=n) {
2251 		n = vm_phys_alloc_npages(domain, VM_FREELIST_DEFAULT, &m,
2252 		    MIN(n, cnt-i));
2253 		if (n == 0)
2254 			break;
2255 		for (j = 0; j < n; j++)
2256 			store[i+j] = m++;
2257 	}
2258 	vm_domain_free_unlock(vmd);
2259 	if (cnt != i)
2260 		vm_domain_freecnt_inc(vmd, cnt - i);
2261 
2262 	return (i);
2263 }
2264 
2265 static void
2266 vm_page_release(void *arg, void **store, int cnt)
2267 {
2268 	struct vm_domain *vmd;
2269 	vm_page_t m;
2270 	int i;
2271 
2272 	vmd = arg;
2273 	vm_domain_free_lock(vmd);
2274 	for (i = 0; i < cnt; i++) {
2275 		m = (vm_page_t)store[i];
2276 		vm_phys_free_pages(m, 0);
2277 	}
2278 	vm_domain_free_unlock(vmd);
2279 	vm_domain_freecnt_inc(vmd, cnt);
2280 }
2281 
2282 #define	VPSC_ANY	0	/* No restrictions. */
2283 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2284 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2285 
2286 /*
2287  *	vm_page_scan_contig:
2288  *
2289  *	Scan vm_page_array[] between the specified entries "m_start" and
2290  *	"m_end" for a run of contiguous physical pages that satisfy the
2291  *	specified conditions, and return the lowest page in the run.  The
2292  *	specified "alignment" determines the alignment of the lowest physical
2293  *	page in the run.  If the specified "boundary" is non-zero, then the
2294  *	run of physical pages cannot span a physical address that is a
2295  *	multiple of "boundary".
2296  *
2297  *	"m_end" is never dereferenced, so it need not point to a vm_page
2298  *	structure within vm_page_array[].
2299  *
2300  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2301  *	span a hole (or discontiguity) in the physical address space.  Both
2302  *	"alignment" and "boundary" must be a power of two.
2303  */
2304 vm_page_t
2305 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2306     u_long alignment, vm_paddr_t boundary, int options)
2307 {
2308 	struct mtx *m_mtx;
2309 	vm_object_t object;
2310 	vm_paddr_t pa;
2311 	vm_page_t m, m_run;
2312 #if VM_NRESERVLEVEL > 0
2313 	int level;
2314 #endif
2315 	int m_inc, order, run_ext, run_len;
2316 
2317 	KASSERT(npages > 0, ("npages is 0"));
2318 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2319 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2320 	m_run = NULL;
2321 	run_len = 0;
2322 	m_mtx = NULL;
2323 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2324 		KASSERT((m->flags & PG_MARKER) == 0,
2325 		    ("page %p is PG_MARKER", m));
2326 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2327 		    ("fictitious page %p has invalid wire count", m));
2328 
2329 		/*
2330 		 * If the current page would be the start of a run, check its
2331 		 * physical address against the end, alignment, and boundary
2332 		 * conditions.  If it doesn't satisfy these conditions, either
2333 		 * terminate the scan or advance to the next page that
2334 		 * satisfies the failed condition.
2335 		 */
2336 		if (run_len == 0) {
2337 			KASSERT(m_run == NULL, ("m_run != NULL"));
2338 			if (m + npages > m_end)
2339 				break;
2340 			pa = VM_PAGE_TO_PHYS(m);
2341 			if ((pa & (alignment - 1)) != 0) {
2342 				m_inc = atop(roundup2(pa, alignment) - pa);
2343 				continue;
2344 			}
2345 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2346 			    boundary) != 0) {
2347 				m_inc = atop(roundup2(pa, boundary) - pa);
2348 				continue;
2349 			}
2350 		} else
2351 			KASSERT(m_run != NULL, ("m_run == NULL"));
2352 
2353 		vm_page_change_lock(m, &m_mtx);
2354 		m_inc = 1;
2355 retry:
2356 		if (vm_page_held(m))
2357 			run_ext = 0;
2358 #if VM_NRESERVLEVEL > 0
2359 		else if ((level = vm_reserv_level(m)) >= 0 &&
2360 		    (options & VPSC_NORESERV) != 0) {
2361 			run_ext = 0;
2362 			/* Advance to the end of the reservation. */
2363 			pa = VM_PAGE_TO_PHYS(m);
2364 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2365 			    pa);
2366 		}
2367 #endif
2368 		else if ((object = m->object) != NULL) {
2369 			/*
2370 			 * The page is considered eligible for relocation if
2371 			 * and only if it could be laundered or reclaimed by
2372 			 * the page daemon.
2373 			 */
2374 			if (!VM_OBJECT_TRYRLOCK(object)) {
2375 				mtx_unlock(m_mtx);
2376 				VM_OBJECT_RLOCK(object);
2377 				mtx_lock(m_mtx);
2378 				if (m->object != object) {
2379 					/*
2380 					 * The page may have been freed.
2381 					 */
2382 					VM_OBJECT_RUNLOCK(object);
2383 					goto retry;
2384 				} else if (vm_page_held(m)) {
2385 					run_ext = 0;
2386 					goto unlock;
2387 				}
2388 			}
2389 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2390 			    ("page %p is PG_UNHOLDFREE", m));
2391 			/* Don't care: PG_NODUMP, PG_ZERO. */
2392 			if (object->type != OBJT_DEFAULT &&
2393 			    object->type != OBJT_SWAP &&
2394 			    object->type != OBJT_VNODE) {
2395 				run_ext = 0;
2396 #if VM_NRESERVLEVEL > 0
2397 			} else if ((options & VPSC_NOSUPER) != 0 &&
2398 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2399 				run_ext = 0;
2400 				/* Advance to the end of the superpage. */
2401 				pa = VM_PAGE_TO_PHYS(m);
2402 				m_inc = atop(roundup2(pa + 1,
2403 				    vm_reserv_size(level)) - pa);
2404 #endif
2405 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2406 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2407 				/*
2408 				 * The page is allocated but eligible for
2409 				 * relocation.  Extend the current run by one
2410 				 * page.
2411 				 */
2412 				KASSERT(pmap_page_get_memattr(m) ==
2413 				    VM_MEMATTR_DEFAULT,
2414 				    ("page %p has an unexpected memattr", m));
2415 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2416 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2417 				    ("page %p has unexpected oflags", m));
2418 				/* Don't care: VPO_NOSYNC. */
2419 				run_ext = 1;
2420 			} else
2421 				run_ext = 0;
2422 unlock:
2423 			VM_OBJECT_RUNLOCK(object);
2424 #if VM_NRESERVLEVEL > 0
2425 		} else if (level >= 0) {
2426 			/*
2427 			 * The page is reserved but not yet allocated.  In
2428 			 * other words, it is still free.  Extend the current
2429 			 * run by one page.
2430 			 */
2431 			run_ext = 1;
2432 #endif
2433 		} else if ((order = m->order) < VM_NFREEORDER) {
2434 			/*
2435 			 * The page is enqueued in the physical memory
2436 			 * allocator's free page queues.  Moreover, it is the
2437 			 * first page in a power-of-two-sized run of
2438 			 * contiguous free pages.  Add these pages to the end
2439 			 * of the current run, and jump ahead.
2440 			 */
2441 			run_ext = 1 << order;
2442 			m_inc = 1 << order;
2443 		} else {
2444 			/*
2445 			 * Skip the page for one of the following reasons: (1)
2446 			 * It is enqueued in the physical memory allocator's
2447 			 * free page queues.  However, it is not the first
2448 			 * page in a run of contiguous free pages.  (This case
2449 			 * rarely occurs because the scan is performed in
2450 			 * ascending order.) (2) It is not reserved, and it is
2451 			 * transitioning from free to allocated.  (Conversely,
2452 			 * the transition from allocated to free for managed
2453 			 * pages is blocked by the page lock.) (3) It is
2454 			 * allocated but not contained by an object and not
2455 			 * wired, e.g., allocated by Xen's balloon driver.
2456 			 */
2457 			run_ext = 0;
2458 		}
2459 
2460 		/*
2461 		 * Extend or reset the current run of pages.
2462 		 */
2463 		if (run_ext > 0) {
2464 			if (run_len == 0)
2465 				m_run = m;
2466 			run_len += run_ext;
2467 		} else {
2468 			if (run_len > 0) {
2469 				m_run = NULL;
2470 				run_len = 0;
2471 			}
2472 		}
2473 	}
2474 	if (m_mtx != NULL)
2475 		mtx_unlock(m_mtx);
2476 	if (run_len >= npages)
2477 		return (m_run);
2478 	return (NULL);
2479 }
2480 
2481 /*
2482  *	vm_page_reclaim_run:
2483  *
2484  *	Try to relocate each of the allocated virtual pages within the
2485  *	specified run of physical pages to a new physical address.  Free the
2486  *	physical pages underlying the relocated virtual pages.  A virtual page
2487  *	is relocatable if and only if it could be laundered or reclaimed by
2488  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2489  *	physical address above "high".
2490  *
2491  *	Returns 0 if every physical page within the run was already free or
2492  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2493  *	value indicating why the last attempt to relocate a virtual page was
2494  *	unsuccessful.
2495  *
2496  *	"req_class" must be an allocation class.
2497  */
2498 static int
2499 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2500     vm_paddr_t high)
2501 {
2502 	struct vm_domain *vmd;
2503 	struct mtx *m_mtx;
2504 	struct spglist free;
2505 	vm_object_t object;
2506 	vm_paddr_t pa;
2507 	vm_page_t m, m_end, m_new;
2508 	int error, order, req;
2509 
2510 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2511 	    ("req_class is not an allocation class"));
2512 	SLIST_INIT(&free);
2513 	error = 0;
2514 	m = m_run;
2515 	m_end = m_run + npages;
2516 	m_mtx = NULL;
2517 	for (; error == 0 && m < m_end; m++) {
2518 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2519 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2520 
2521 		/*
2522 		 * Avoid releasing and reacquiring the same page lock.
2523 		 */
2524 		vm_page_change_lock(m, &m_mtx);
2525 retry:
2526 		if (vm_page_held(m))
2527 			error = EBUSY;
2528 		else if ((object = m->object) != NULL) {
2529 			/*
2530 			 * The page is relocated if and only if it could be
2531 			 * laundered or reclaimed by the page daemon.
2532 			 */
2533 			if (!VM_OBJECT_TRYWLOCK(object)) {
2534 				mtx_unlock(m_mtx);
2535 				VM_OBJECT_WLOCK(object);
2536 				mtx_lock(m_mtx);
2537 				if (m->object != object) {
2538 					/*
2539 					 * The page may have been freed.
2540 					 */
2541 					VM_OBJECT_WUNLOCK(object);
2542 					goto retry;
2543 				} else if (vm_page_held(m)) {
2544 					error = EBUSY;
2545 					goto unlock;
2546 				}
2547 			}
2548 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2549 			    ("page %p is PG_UNHOLDFREE", m));
2550 			/* Don't care: PG_NODUMP, PG_ZERO. */
2551 			if (object->type != OBJT_DEFAULT &&
2552 			    object->type != OBJT_SWAP &&
2553 			    object->type != OBJT_VNODE)
2554 				error = EINVAL;
2555 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2556 				error = EINVAL;
2557 			else if (vm_page_queue(m) != PQ_NONE &&
2558 			    !vm_page_busied(m)) {
2559 				KASSERT(pmap_page_get_memattr(m) ==
2560 				    VM_MEMATTR_DEFAULT,
2561 				    ("page %p has an unexpected memattr", m));
2562 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2563 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2564 				    ("page %p has unexpected oflags", m));
2565 				/* Don't care: VPO_NOSYNC. */
2566 				if (m->valid != 0) {
2567 					/*
2568 					 * First, try to allocate a new page
2569 					 * that is above "high".  Failing
2570 					 * that, try to allocate a new page
2571 					 * that is below "m_run".  Allocate
2572 					 * the new page between the end of
2573 					 * "m_run" and "high" only as a last
2574 					 * resort.
2575 					 */
2576 					req = req_class | VM_ALLOC_NOOBJ;
2577 					if ((m->flags & PG_NODUMP) != 0)
2578 						req |= VM_ALLOC_NODUMP;
2579 					if (trunc_page(high) !=
2580 					    ~(vm_paddr_t)PAGE_MASK) {
2581 						m_new = vm_page_alloc_contig(
2582 						    NULL, 0, req, 1,
2583 						    round_page(high),
2584 						    ~(vm_paddr_t)0,
2585 						    PAGE_SIZE, 0,
2586 						    VM_MEMATTR_DEFAULT);
2587 					} else
2588 						m_new = NULL;
2589 					if (m_new == NULL) {
2590 						pa = VM_PAGE_TO_PHYS(m_run);
2591 						m_new = vm_page_alloc_contig(
2592 						    NULL, 0, req, 1,
2593 						    0, pa - 1, PAGE_SIZE, 0,
2594 						    VM_MEMATTR_DEFAULT);
2595 					}
2596 					if (m_new == NULL) {
2597 						pa += ptoa(npages);
2598 						m_new = vm_page_alloc_contig(
2599 						    NULL, 0, req, 1,
2600 						    pa, high, PAGE_SIZE, 0,
2601 						    VM_MEMATTR_DEFAULT);
2602 					}
2603 					if (m_new == NULL) {
2604 						error = ENOMEM;
2605 						goto unlock;
2606 					}
2607 					KASSERT(m_new->wire_count == 0,
2608 					    ("page %p is wired", m_new));
2609 
2610 					/*
2611 					 * Replace "m" with the new page.  For
2612 					 * vm_page_replace(), "m" must be busy
2613 					 * and dequeued.  Finally, change "m"
2614 					 * as if vm_page_free() was called.
2615 					 */
2616 					if (object->ref_count != 0)
2617 						pmap_remove_all(m);
2618 					m_new->aflags = m->aflags &
2619 					    ~PGA_QUEUE_STATE_MASK;
2620 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2621 					    ("page %p is managed", m_new));
2622 					m_new->oflags = m->oflags & VPO_NOSYNC;
2623 					pmap_copy_page(m, m_new);
2624 					m_new->valid = m->valid;
2625 					m_new->dirty = m->dirty;
2626 					m->flags &= ~PG_ZERO;
2627 					vm_page_xbusy(m);
2628 					vm_page_remque(m);
2629 					vm_page_replace_checked(m_new, object,
2630 					    m->pindex, m);
2631 					if (vm_page_free_prep(m))
2632 						SLIST_INSERT_HEAD(&free, m,
2633 						    plinks.s.ss);
2634 
2635 					/*
2636 					 * The new page must be deactivated
2637 					 * before the object is unlocked.
2638 					 */
2639 					vm_page_change_lock(m_new, &m_mtx);
2640 					vm_page_deactivate(m_new);
2641 				} else {
2642 					m->flags &= ~PG_ZERO;
2643 					vm_page_remque(m);
2644 					vm_page_remove(m);
2645 					if (vm_page_free_prep(m))
2646 						SLIST_INSERT_HEAD(&free, m,
2647 						    plinks.s.ss);
2648 					KASSERT(m->dirty == 0,
2649 					    ("page %p is dirty", m));
2650 				}
2651 			} else
2652 				error = EBUSY;
2653 unlock:
2654 			VM_OBJECT_WUNLOCK(object);
2655 		} else {
2656 			MPASS(vm_phys_domain(m) == domain);
2657 			vmd = VM_DOMAIN(domain);
2658 			vm_domain_free_lock(vmd);
2659 			order = m->order;
2660 			if (order < VM_NFREEORDER) {
2661 				/*
2662 				 * The page is enqueued in the physical memory
2663 				 * allocator's free page queues.  Moreover, it
2664 				 * is the first page in a power-of-two-sized
2665 				 * run of contiguous free pages.  Jump ahead
2666 				 * to the last page within that run, and
2667 				 * continue from there.
2668 				 */
2669 				m += (1 << order) - 1;
2670 			}
2671 #if VM_NRESERVLEVEL > 0
2672 			else if (vm_reserv_is_page_free(m))
2673 				order = 0;
2674 #endif
2675 			vm_domain_free_unlock(vmd);
2676 			if (order == VM_NFREEORDER)
2677 				error = EINVAL;
2678 		}
2679 	}
2680 	if (m_mtx != NULL)
2681 		mtx_unlock(m_mtx);
2682 	if ((m = SLIST_FIRST(&free)) != NULL) {
2683 		int cnt;
2684 
2685 		vmd = VM_DOMAIN(domain);
2686 		cnt = 0;
2687 		vm_domain_free_lock(vmd);
2688 		do {
2689 			MPASS(vm_phys_domain(m) == domain);
2690 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2691 			vm_phys_free_pages(m, 0);
2692 			cnt++;
2693 		} while ((m = SLIST_FIRST(&free)) != NULL);
2694 		vm_domain_free_unlock(vmd);
2695 		vm_domain_freecnt_inc(vmd, cnt);
2696 	}
2697 	return (error);
2698 }
2699 
2700 #define	NRUNS	16
2701 
2702 CTASSERT(powerof2(NRUNS));
2703 
2704 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2705 
2706 #define	MIN_RECLAIM	8
2707 
2708 /*
2709  *	vm_page_reclaim_contig:
2710  *
2711  *	Reclaim allocated, contiguous physical memory satisfying the specified
2712  *	conditions by relocating the virtual pages using that physical memory.
2713  *	Returns true if reclamation is successful and false otherwise.  Since
2714  *	relocation requires the allocation of physical pages, reclamation may
2715  *	fail due to a shortage of free pages.  When reclamation fails, callers
2716  *	are expected to perform vm_wait() before retrying a failed allocation
2717  *	operation, e.g., vm_page_alloc_contig().
2718  *
2719  *	The caller must always specify an allocation class through "req".
2720  *
2721  *	allocation classes:
2722  *	VM_ALLOC_NORMAL		normal process request
2723  *	VM_ALLOC_SYSTEM		system *really* needs a page
2724  *	VM_ALLOC_INTERRUPT	interrupt time request
2725  *
2726  *	The optional allocation flags are ignored.
2727  *
2728  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2729  *	must be a power of two.
2730  */
2731 bool
2732 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2733     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2734 {
2735 	struct vm_domain *vmd;
2736 	vm_paddr_t curr_low;
2737 	vm_page_t m_run, m_runs[NRUNS];
2738 	u_long count, reclaimed;
2739 	int error, i, options, req_class;
2740 
2741 	KASSERT(npages > 0, ("npages is 0"));
2742 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2743 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2744 	req_class = req & VM_ALLOC_CLASS_MASK;
2745 
2746 	/*
2747 	 * The page daemon is allowed to dig deeper into the free page list.
2748 	 */
2749 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2750 		req_class = VM_ALLOC_SYSTEM;
2751 
2752 	/*
2753 	 * Return if the number of free pages cannot satisfy the requested
2754 	 * allocation.
2755 	 */
2756 	vmd = VM_DOMAIN(domain);
2757 	count = vmd->vmd_free_count;
2758 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2759 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2760 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2761 		return (false);
2762 
2763 	/*
2764 	 * Scan up to three times, relaxing the restrictions ("options") on
2765 	 * the reclamation of reservations and superpages each time.
2766 	 */
2767 	for (options = VPSC_NORESERV;;) {
2768 		/*
2769 		 * Find the highest runs that satisfy the given constraints
2770 		 * and restrictions, and record them in "m_runs".
2771 		 */
2772 		curr_low = low;
2773 		count = 0;
2774 		for (;;) {
2775 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2776 			    high, alignment, boundary, options);
2777 			if (m_run == NULL)
2778 				break;
2779 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2780 			m_runs[RUN_INDEX(count)] = m_run;
2781 			count++;
2782 		}
2783 
2784 		/*
2785 		 * Reclaim the highest runs in LIFO (descending) order until
2786 		 * the number of reclaimed pages, "reclaimed", is at least
2787 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2788 		 * reclamation is idempotent, and runs will (likely) recur
2789 		 * from one scan to the next as restrictions are relaxed.
2790 		 */
2791 		reclaimed = 0;
2792 		for (i = 0; count > 0 && i < NRUNS; i++) {
2793 			count--;
2794 			m_run = m_runs[RUN_INDEX(count)];
2795 			error = vm_page_reclaim_run(req_class, domain, npages,
2796 			    m_run, high);
2797 			if (error == 0) {
2798 				reclaimed += npages;
2799 				if (reclaimed >= MIN_RECLAIM)
2800 					return (true);
2801 			}
2802 		}
2803 
2804 		/*
2805 		 * Either relax the restrictions on the next scan or return if
2806 		 * the last scan had no restrictions.
2807 		 */
2808 		if (options == VPSC_NORESERV)
2809 			options = VPSC_NOSUPER;
2810 		else if (options == VPSC_NOSUPER)
2811 			options = VPSC_ANY;
2812 		else if (options == VPSC_ANY)
2813 			return (reclaimed != 0);
2814 	}
2815 }
2816 
2817 bool
2818 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2819     u_long alignment, vm_paddr_t boundary)
2820 {
2821 	struct vm_domainset_iter di;
2822 	int domain;
2823 	bool ret;
2824 
2825 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2826 	do {
2827 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2828 		    high, alignment, boundary);
2829 		if (ret)
2830 			break;
2831 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2832 
2833 	return (ret);
2834 }
2835 
2836 /*
2837  * Set the domain in the appropriate page level domainset.
2838  */
2839 void
2840 vm_domain_set(struct vm_domain *vmd)
2841 {
2842 
2843 	mtx_lock(&vm_domainset_lock);
2844 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2845 		vmd->vmd_minset = 1;
2846 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2847 	}
2848 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2849 		vmd->vmd_severeset = 1;
2850 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2851 	}
2852 	mtx_unlock(&vm_domainset_lock);
2853 }
2854 
2855 /*
2856  * Clear the domain from the appropriate page level domainset.
2857  */
2858 void
2859 vm_domain_clear(struct vm_domain *vmd)
2860 {
2861 
2862 	mtx_lock(&vm_domainset_lock);
2863 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2864 		vmd->vmd_minset = 0;
2865 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2866 		if (vm_min_waiters != 0) {
2867 			vm_min_waiters = 0;
2868 			wakeup(&vm_min_domains);
2869 		}
2870 	}
2871 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2872 		vmd->vmd_severeset = 0;
2873 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2874 		if (vm_severe_waiters != 0) {
2875 			vm_severe_waiters = 0;
2876 			wakeup(&vm_severe_domains);
2877 		}
2878 	}
2879 
2880 	/*
2881 	 * If pageout daemon needs pages, then tell it that there are
2882 	 * some free.
2883 	 */
2884 	if (vmd->vmd_pageout_pages_needed &&
2885 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2886 		wakeup(&vmd->vmd_pageout_pages_needed);
2887 		vmd->vmd_pageout_pages_needed = 0;
2888 	}
2889 
2890 	/* See comments in vm_wait_doms(). */
2891 	if (vm_pageproc_waiters) {
2892 		vm_pageproc_waiters = 0;
2893 		wakeup(&vm_pageproc_waiters);
2894 	}
2895 	mtx_unlock(&vm_domainset_lock);
2896 }
2897 
2898 /*
2899  * Wait for free pages to exceed the min threshold globally.
2900  */
2901 void
2902 vm_wait_min(void)
2903 {
2904 
2905 	mtx_lock(&vm_domainset_lock);
2906 	while (vm_page_count_min()) {
2907 		vm_min_waiters++;
2908 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2909 	}
2910 	mtx_unlock(&vm_domainset_lock);
2911 }
2912 
2913 /*
2914  * Wait for free pages to exceed the severe threshold globally.
2915  */
2916 void
2917 vm_wait_severe(void)
2918 {
2919 
2920 	mtx_lock(&vm_domainset_lock);
2921 	while (vm_page_count_severe()) {
2922 		vm_severe_waiters++;
2923 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2924 		    "vmwait", 0);
2925 	}
2926 	mtx_unlock(&vm_domainset_lock);
2927 }
2928 
2929 u_int
2930 vm_wait_count(void)
2931 {
2932 
2933 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2934 }
2935 
2936 static void
2937 vm_wait_doms(const domainset_t *wdoms)
2938 {
2939 
2940 	/*
2941 	 * We use racey wakeup synchronization to avoid expensive global
2942 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
2943 	 * To handle this, we only sleep for one tick in this instance.  It
2944 	 * is expected that most allocations for the pageproc will come from
2945 	 * kmem or vm_page_grab* which will use the more specific and
2946 	 * race-free vm_wait_domain().
2947 	 */
2948 	if (curproc == pageproc) {
2949 		mtx_lock(&vm_domainset_lock);
2950 		vm_pageproc_waiters++;
2951 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2952 		    "pageprocwait", 1);
2953 	} else {
2954 		/*
2955 		 * XXX Ideally we would wait only until the allocation could
2956 		 * be satisfied.  This condition can cause new allocators to
2957 		 * consume all freed pages while old allocators wait.
2958 		 */
2959 		mtx_lock(&vm_domainset_lock);
2960 		if (DOMAINSET_SUBSET(&vm_min_domains, wdoms)) {
2961 			vm_min_waiters++;
2962 			msleep(&vm_min_domains, &vm_domainset_lock, PVM,
2963 			    "vmwait", 0);
2964 		}
2965 		mtx_unlock(&vm_domainset_lock);
2966 	}
2967 }
2968 
2969 /*
2970  *	vm_wait_domain:
2971  *
2972  *	Sleep until free pages are available for allocation.
2973  *	- Called in various places after failed memory allocations.
2974  */
2975 void
2976 vm_wait_domain(int domain)
2977 {
2978 	struct vm_domain *vmd;
2979 	domainset_t wdom;
2980 
2981 	vmd = VM_DOMAIN(domain);
2982 	vm_domain_free_assert_unlocked(vmd);
2983 
2984 	if (curproc == pageproc) {
2985 		mtx_lock(&vm_domainset_lock);
2986 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2987 			vmd->vmd_pageout_pages_needed = 1;
2988 			msleep(&vmd->vmd_pageout_pages_needed,
2989 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
2990 		} else
2991 			mtx_unlock(&vm_domainset_lock);
2992 	} else {
2993 		if (pageproc == NULL)
2994 			panic("vm_wait in early boot");
2995 		DOMAINSET_ZERO(&wdom);
2996 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
2997 		vm_wait_doms(&wdom);
2998 	}
2999 }
3000 
3001 /*
3002  *	vm_wait:
3003  *
3004  *	Sleep until free pages are available for allocation in the
3005  *	affinity domains of the obj.  If obj is NULL, the domain set
3006  *	for the calling thread is used.
3007  *	Called in various places after failed memory allocations.
3008  */
3009 void
3010 vm_wait(vm_object_t obj)
3011 {
3012 	struct domainset *d;
3013 
3014 	d = NULL;
3015 
3016 	/*
3017 	 * Carefully fetch pointers only once: the struct domainset
3018 	 * itself is ummutable but the pointer might change.
3019 	 */
3020 	if (obj != NULL)
3021 		d = obj->domain.dr_policy;
3022 	if (d == NULL)
3023 		d = curthread->td_domain.dr_policy;
3024 
3025 	vm_wait_doms(&d->ds_mask);
3026 }
3027 
3028 /*
3029  *	vm_domain_alloc_fail:
3030  *
3031  *	Called when a page allocation function fails.  Informs the
3032  *	pagedaemon and performs the requested wait.  Requires the
3033  *	domain_free and object lock on entry.  Returns with the
3034  *	object lock held and free lock released.  Returns an error when
3035  *	retry is necessary.
3036  *
3037  */
3038 static int
3039 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3040 {
3041 
3042 	vm_domain_free_assert_unlocked(vmd);
3043 
3044 	atomic_add_int(&vmd->vmd_pageout_deficit,
3045 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3046 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3047 		if (object != NULL)
3048 			VM_OBJECT_WUNLOCK(object);
3049 		vm_wait_domain(vmd->vmd_domain);
3050 		if (object != NULL)
3051 			VM_OBJECT_WLOCK(object);
3052 		if (req & VM_ALLOC_WAITOK)
3053 			return (EAGAIN);
3054 	}
3055 
3056 	return (0);
3057 }
3058 
3059 /*
3060  *	vm_waitpfault:
3061  *
3062  *	Sleep until free pages are available for allocation.
3063  *	- Called only in vm_fault so that processes page faulting
3064  *	  can be easily tracked.
3065  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3066  *	  processes will be able to grab memory first.  Do not change
3067  *	  this balance without careful testing first.
3068  */
3069 void
3070 vm_waitpfault(void)
3071 {
3072 
3073 	mtx_lock(&vm_domainset_lock);
3074 	if (vm_page_count_min()) {
3075 		vm_min_waiters++;
3076 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER, "pfault", 0);
3077 	}
3078 	mtx_unlock(&vm_domainset_lock);
3079 }
3080 
3081 struct vm_pagequeue *
3082 vm_page_pagequeue(vm_page_t m)
3083 {
3084 
3085 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
3086 }
3087 
3088 static struct mtx *
3089 vm_page_pagequeue_lockptr(vm_page_t m)
3090 {
3091 	uint8_t queue;
3092 
3093 	if ((queue = m->queue) == PQ_NONE)
3094 		return (NULL);
3095 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex);
3096 }
3097 
3098 static inline void
3099 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3100 {
3101 	struct vm_domain *vmd;
3102 	uint8_t aflags;
3103 
3104 	vm_pagequeue_assert_locked(pq);
3105 	KASSERT(pq == vm_page_pagequeue(m),
3106 	    ("page %p doesn't belong to %p", m, pq));
3107 
3108 	aflags = m->aflags;
3109 	if ((aflags & PGA_DEQUEUE) != 0) {
3110 		if (__predict_true((aflags & PGA_ENQUEUED) != 0)) {
3111 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3112 			vm_pagequeue_cnt_dec(pq);
3113 		}
3114 		vm_page_dequeue_complete(m);
3115 	} else if ((aflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3116 		if ((aflags & PGA_ENQUEUED) != 0)
3117 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3118 		else {
3119 			vm_pagequeue_cnt_inc(pq);
3120 			vm_page_aflag_set(m, PGA_ENQUEUED);
3121 		}
3122 		if ((aflags & PGA_REQUEUE_HEAD) != 0) {
3123 			KASSERT(m->queue == PQ_INACTIVE,
3124 			    ("head enqueue not supported for page %p", m));
3125 			vmd = vm_pagequeue_domain(m);
3126 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3127 		} else
3128 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3129 
3130 		/*
3131 		 * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after
3132 		 * setting PGA_ENQUEUED in order to synchronize with the
3133 		 * page daemon.
3134 		 */
3135 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
3136 	}
3137 }
3138 
3139 static void
3140 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3141     uint8_t queue)
3142 {
3143 	vm_page_t m;
3144 	int i;
3145 
3146 	for (i = 0; i < bq->bq_cnt; i++) {
3147 		m = bq->bq_pa[i];
3148 		if (__predict_false(m->queue != queue))
3149 			continue;
3150 		vm_pqbatch_process_page(pq, m);
3151 	}
3152 	vm_batchqueue_init(bq);
3153 }
3154 
3155 static void
3156 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3157 {
3158 	struct vm_batchqueue *bq;
3159 	struct vm_pagequeue *pq;
3160 	int domain;
3161 
3162 	vm_page_assert_locked(m);
3163 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3164 
3165 	domain = vm_phys_domain(m);
3166 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3167 
3168 	critical_enter();
3169 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3170 	if (vm_batchqueue_insert(bq, m)) {
3171 		critical_exit();
3172 		return;
3173 	}
3174 	if (!vm_pagequeue_trylock(pq)) {
3175 		critical_exit();
3176 		vm_pagequeue_lock(pq);
3177 		critical_enter();
3178 		bq = DPCPU_PTR(pqbatch[domain][queue]);
3179 	}
3180 	vm_pqbatch_process(pq, bq, queue);
3181 
3182 	/*
3183 	 * The page may have been logically dequeued before we acquired the
3184 	 * page queue lock.  In this case, the page lock prevents the page
3185 	 * from being logically enqueued elsewhere.
3186 	 */
3187 	if (__predict_true(m->queue == queue))
3188 		vm_pqbatch_process_page(pq, m);
3189 	else {
3190 		KASSERT(m->queue == PQ_NONE,
3191 		    ("invalid queue transition for page %p", m));
3192 		KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3193 		    ("page %p is enqueued with invalid queue index", m));
3194 		vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3195 	}
3196 	vm_pagequeue_unlock(pq);
3197 	critical_exit();
3198 }
3199 
3200 /*
3201  *	vm_page_drain_pqbatch:		[ internal use only ]
3202  *
3203  *	Force all per-CPU page queue batch queues to be drained.  This is
3204  *	intended for use in severe memory shortages, to ensure that pages
3205  *	do not remain stuck in the batch queues.
3206  */
3207 void
3208 vm_page_drain_pqbatch(void)
3209 {
3210 	struct thread *td;
3211 	struct vm_domain *vmd;
3212 	struct vm_pagequeue *pq;
3213 	int cpu, domain, queue;
3214 
3215 	td = curthread;
3216 	CPU_FOREACH(cpu) {
3217 		thread_lock(td);
3218 		sched_bind(td, cpu);
3219 		thread_unlock(td);
3220 
3221 		for (domain = 0; domain < vm_ndomains; domain++) {
3222 			vmd = VM_DOMAIN(domain);
3223 			for (queue = 0; queue < PQ_COUNT; queue++) {
3224 				pq = &vmd->vmd_pagequeues[queue];
3225 				vm_pagequeue_lock(pq);
3226 				critical_enter();
3227 				vm_pqbatch_process(pq,
3228 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3229 				critical_exit();
3230 				vm_pagequeue_unlock(pq);
3231 			}
3232 		}
3233 	}
3234 	thread_lock(td);
3235 	sched_unbind(td);
3236 	thread_unlock(td);
3237 }
3238 
3239 /*
3240  * Complete the logical removal of a page from a page queue.  We must be
3241  * careful to synchronize with the page daemon, which may be concurrently
3242  * examining the page with only the page lock held.  The page must not be
3243  * in a state where it appears to be logically enqueued.
3244  */
3245 static void
3246 vm_page_dequeue_complete(vm_page_t m)
3247 {
3248 
3249 	m->queue = PQ_NONE;
3250 	atomic_thread_fence_rel();
3251 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3252 }
3253 
3254 /*
3255  *	vm_page_dequeue_deferred:	[ internal use only ]
3256  *
3257  *	Request removal of the given page from its current page
3258  *	queue.  Physical removal from the queue may be deferred
3259  *	indefinitely.
3260  *
3261  *	The page must be locked.
3262  */
3263 void
3264 vm_page_dequeue_deferred(vm_page_t m)
3265 {
3266 	int queue;
3267 
3268 	vm_page_assert_locked(m);
3269 
3270 	queue = m->queue;
3271 	if (queue == PQ_NONE) {
3272 		KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3273 		    ("page %p has queue state", m));
3274 		return;
3275 	}
3276 	if ((m->aflags & PGA_DEQUEUE) == 0)
3277 		vm_page_aflag_set(m, PGA_DEQUEUE);
3278 	vm_pqbatch_submit_page(m, queue);
3279 }
3280 
3281 /*
3282  *	vm_page_dequeue_locked:
3283  *
3284  *	Remove the page from its page queue, which must be locked.
3285  *	If the page lock is not held, there is no guarantee that the
3286  *	page will not be enqueued by another thread before this function
3287  *	returns.  In this case, it is up to the caller to ensure that
3288  *	no other threads hold a reference to the page.
3289  *
3290  *	The page queue lock must be held.  If the page is not already
3291  *	logically dequeued, the page lock must be held as well.
3292  */
3293 void
3294 vm_page_dequeue_locked(vm_page_t m)
3295 {
3296 	struct vm_pagequeue *pq;
3297 
3298 	pq = vm_page_pagequeue(m);
3299 
3300 	KASSERT(m->queue != PQ_NONE,
3301 	    ("%s: page %p queue field is PQ_NONE", __func__, m));
3302 	vm_pagequeue_assert_locked(pq);
3303 	KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3304 	    mtx_owned(vm_page_lockptr(m)),
3305 	    ("%s: queued unlocked page %p", __func__, m));
3306 
3307 	if ((m->aflags & PGA_ENQUEUED) != 0) {
3308 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3309 		vm_pagequeue_cnt_dec(pq);
3310 	}
3311 	vm_page_dequeue_complete(m);
3312 }
3313 
3314 /*
3315  *	vm_page_dequeue:
3316  *
3317  *	Remove the page from whichever page queue it's in, if any.
3318  *	If the page lock is not held, there is no guarantee that the
3319  *	page will not be enqueued by another thread before this function
3320  *	returns.  In this case, it is up to the caller to ensure that
3321  *	no other threads hold a reference to the page.
3322  */
3323 void
3324 vm_page_dequeue(vm_page_t m)
3325 {
3326 	struct mtx *lock, *lock1;
3327 
3328 	lock = vm_page_pagequeue_lockptr(m);
3329 	for (;;) {
3330 		if (lock == NULL)
3331 			return;
3332 		mtx_lock(lock);
3333 		if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock)
3334 			break;
3335 		mtx_unlock(lock);
3336 		lock = lock1;
3337 	}
3338 	KASSERT(lock == vm_page_pagequeue_lockptr(m),
3339 	    ("%s: page %p migrated directly between queues", __func__, m));
3340 	vm_page_dequeue_locked(m);
3341 	mtx_unlock(lock);
3342 }
3343 
3344 /*
3345  * Schedule the given page for insertion into the specified page queue.
3346  * Physical insertion of the page may be deferred indefinitely.
3347  */
3348 static void
3349 vm_page_enqueue(vm_page_t m, uint8_t queue)
3350 {
3351 
3352 	vm_page_assert_locked(m);
3353 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3354 	    ("%s: page %p is already enqueued", __func__, m));
3355 
3356 	m->queue = queue;
3357 	if ((m->aflags & PGA_REQUEUE) == 0)
3358 		vm_page_aflag_set(m, PGA_REQUEUE);
3359 	vm_pqbatch_submit_page(m, queue);
3360 }
3361 
3362 /*
3363  *	vm_page_requeue:		[ internal use only ]
3364  *
3365  *	Schedule a requeue of the given page.
3366  *
3367  *	The page must be locked.
3368  */
3369 void
3370 vm_page_requeue(vm_page_t m)
3371 {
3372 
3373 	vm_page_assert_locked(m);
3374 	KASSERT(m->queue != PQ_NONE,
3375 	    ("%s: page %p is not logically enqueued", __func__, m));
3376 
3377 	if ((m->aflags & PGA_REQUEUE) == 0)
3378 		vm_page_aflag_set(m, PGA_REQUEUE);
3379 	vm_pqbatch_submit_page(m, m->queue);
3380 }
3381 
3382 /*
3383  *	vm_page_activate:
3384  *
3385  *	Put the specified page on the active list (if appropriate).
3386  *	Ensure that act_count is at least ACT_INIT but do not otherwise
3387  *	mess with it.
3388  *
3389  *	The page must be locked.
3390  */
3391 void
3392 vm_page_activate(vm_page_t m)
3393 {
3394 	int queue;
3395 
3396 	vm_page_assert_locked(m);
3397 
3398 	if ((queue = vm_page_queue(m)) == PQ_ACTIVE || m->wire_count > 0 ||
3399 	    (m->oflags & VPO_UNMANAGED) != 0) {
3400 		if (queue == PQ_ACTIVE && m->act_count < ACT_INIT)
3401 			m->act_count = ACT_INIT;
3402 		return;
3403 	}
3404 
3405 	vm_page_remque(m);
3406 	if (m->act_count < ACT_INIT)
3407 		m->act_count = ACT_INIT;
3408 	vm_page_enqueue(m, PQ_ACTIVE);
3409 }
3410 
3411 /*
3412  *	vm_page_free_prep:
3413  *
3414  *	Prepares the given page to be put on the free list,
3415  *	disassociating it from any VM object. The caller may return
3416  *	the page to the free list only if this function returns true.
3417  *
3418  *	The object must be locked.  The page must be locked if it is
3419  *	managed.
3420  */
3421 bool
3422 vm_page_free_prep(vm_page_t m)
3423 {
3424 
3425 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3426 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3427 		uint64_t *p;
3428 		int i;
3429 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3430 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3431 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3432 			    m, i, (uintmax_t)*p));
3433 	}
3434 #endif
3435 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3436 		vm_page_lock_assert(m, MA_OWNED);
3437 		KASSERT(!pmap_page_is_mapped(m),
3438 		    ("vm_page_free_prep: freeing mapped page %p", m));
3439 	} else
3440 		KASSERT(m->queue == PQ_NONE,
3441 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3442 	VM_CNT_INC(v_tfree);
3443 
3444 	if (vm_page_sbusied(m))
3445 		panic("vm_page_free_prep: freeing busy page %p", m);
3446 
3447 	vm_page_remove(m);
3448 
3449 	/*
3450 	 * If fictitious remove object association and
3451 	 * return.
3452 	 */
3453 	if ((m->flags & PG_FICTITIOUS) != 0) {
3454 		KASSERT(m->wire_count == 1,
3455 		    ("fictitious page %p is not wired", m));
3456 		KASSERT(m->queue == PQ_NONE,
3457 		    ("fictitious page %p is queued", m));
3458 		return (false);
3459 	}
3460 
3461 	/*
3462 	 * Pages need not be dequeued before they are returned to the physical
3463 	 * memory allocator, but they must at least be marked for a deferred
3464 	 * dequeue.
3465 	 */
3466 	if ((m->oflags & VPO_UNMANAGED) == 0)
3467 		vm_page_dequeue_deferred(m);
3468 
3469 	m->valid = 0;
3470 	vm_page_undirty(m);
3471 
3472 	if (m->wire_count != 0)
3473 		panic("vm_page_free_prep: freeing wired page %p", m);
3474 	if (m->hold_count != 0) {
3475 		m->flags &= ~PG_ZERO;
3476 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3477 		    ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m));
3478 		m->flags |= PG_UNHOLDFREE;
3479 		return (false);
3480 	}
3481 
3482 	/*
3483 	 * Restore the default memory attribute to the page.
3484 	 */
3485 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3486 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3487 
3488 #if VM_NRESERVLEVEL > 0
3489 	if (vm_reserv_free_page(m))
3490 		return (false);
3491 #endif
3492 
3493 	return (true);
3494 }
3495 
3496 /*
3497  *	vm_page_free_toq:
3498  *
3499  *	Returns the given page to the free list, disassociating it
3500  *	from any VM object.
3501  *
3502  *	The object must be locked.  The page must be locked if it is
3503  *	managed.
3504  */
3505 void
3506 vm_page_free_toq(vm_page_t m)
3507 {
3508 	struct vm_domain *vmd;
3509 
3510 	if (!vm_page_free_prep(m))
3511 		return;
3512 
3513 	vmd = vm_pagequeue_domain(m);
3514 	if (m->pool == VM_FREEPOOL_DEFAULT && vmd->vmd_pgcache != NULL) {
3515 		uma_zfree(vmd->vmd_pgcache, m);
3516 		return;
3517 	}
3518 	vm_domain_free_lock(vmd);
3519 	vm_phys_free_pages(m, 0);
3520 	vm_domain_free_unlock(vmd);
3521 	vm_domain_freecnt_inc(vmd, 1);
3522 }
3523 
3524 /*
3525  *	vm_page_free_pages_toq:
3526  *
3527  *	Returns a list of pages to the free list, disassociating it
3528  *	from any VM object.  In other words, this is equivalent to
3529  *	calling vm_page_free_toq() for each page of a list of VM objects.
3530  *
3531  *	The objects must be locked.  The pages must be locked if it is
3532  *	managed.
3533  */
3534 void
3535 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3536 {
3537 	vm_page_t m;
3538 	int count;
3539 
3540 	if (SLIST_EMPTY(free))
3541 		return;
3542 
3543 	count = 0;
3544 	while ((m = SLIST_FIRST(free)) != NULL) {
3545 		count++;
3546 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3547 		vm_page_free_toq(m);
3548 	}
3549 
3550 	if (update_wire_count)
3551 		vm_wire_sub(count);
3552 }
3553 
3554 /*
3555  *	vm_page_wire:
3556  *
3557  * Mark this page as wired down.  If the page is fictitious, then
3558  * its wire count must remain one.
3559  *
3560  * The page must be locked.
3561  */
3562 void
3563 vm_page_wire(vm_page_t m)
3564 {
3565 
3566 	vm_page_assert_locked(m);
3567 	if ((m->flags & PG_FICTITIOUS) != 0) {
3568 		KASSERT(m->wire_count == 1,
3569 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
3570 		    m));
3571 		return;
3572 	}
3573 	if (m->wire_count == 0) {
3574 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3575 		    m->queue == PQ_NONE,
3576 		    ("vm_page_wire: unmanaged page %p is queued", m));
3577 		vm_wire_add(1);
3578 	}
3579 	m->wire_count++;
3580 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3581 }
3582 
3583 /*
3584  * vm_page_unwire:
3585  *
3586  * Release one wiring of the specified page, potentially allowing it to be
3587  * paged out.  Returns TRUE if the number of wirings transitions to zero and
3588  * FALSE otherwise.
3589  *
3590  * Only managed pages belonging to an object can be paged out.  If the number
3591  * of wirings transitions to zero and the page is eligible for page out, then
3592  * the page is added to the specified paging queue (unless PQ_NONE is
3593  * specified, in which case the page is dequeued if it belongs to a paging
3594  * queue).
3595  *
3596  * If a page is fictitious, then its wire count must always be one.
3597  *
3598  * A managed page must be locked.
3599  */
3600 bool
3601 vm_page_unwire(vm_page_t m, uint8_t queue)
3602 {
3603 	bool unwired;
3604 
3605 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3606 	    ("vm_page_unwire: invalid queue %u request for page %p",
3607 	    queue, m));
3608 	if ((m->oflags & VPO_UNMANAGED) == 0)
3609 		vm_page_assert_locked(m);
3610 
3611 	unwired = vm_page_unwire_noq(m);
3612 	if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3613 		return (unwired);
3614 
3615 	if (vm_page_queue(m) == queue) {
3616 		if (queue == PQ_ACTIVE)
3617 			vm_page_reference(m);
3618 		else if (queue != PQ_NONE)
3619 			vm_page_requeue(m);
3620 	} else {
3621 		vm_page_dequeue(m);
3622 		if (queue != PQ_NONE) {
3623 			vm_page_enqueue(m, queue);
3624 			if (queue == PQ_ACTIVE)
3625 				/* Initialize act_count. */
3626 				vm_page_activate(m);
3627 		}
3628 	}
3629 	return (unwired);
3630 }
3631 
3632 /*
3633  *
3634  * vm_page_unwire_noq:
3635  *
3636  * Unwire a page without (re-)inserting it into a page queue.  It is up
3637  * to the caller to enqueue, requeue, or free the page as appropriate.
3638  * In most cases, vm_page_unwire() should be used instead.
3639  */
3640 bool
3641 vm_page_unwire_noq(vm_page_t m)
3642 {
3643 
3644 	if ((m->oflags & VPO_UNMANAGED) == 0)
3645 		vm_page_assert_locked(m);
3646 	if ((m->flags & PG_FICTITIOUS) != 0) {
3647 		KASSERT(m->wire_count == 1,
3648 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3649 		return (false);
3650 	}
3651 	if (m->wire_count == 0)
3652 		panic("vm_page_unwire: page %p's wire count is zero", m);
3653 	m->wire_count--;
3654 	if (m->wire_count == 0) {
3655 		vm_wire_sub(1);
3656 		return (true);
3657 	} else
3658 		return (false);
3659 }
3660 
3661 /*
3662  * Move the specified page to the tail of the inactive queue, or requeue
3663  * the page if it is already in the inactive queue.
3664  *
3665  * The page must be locked.
3666  */
3667 void
3668 vm_page_deactivate(vm_page_t m)
3669 {
3670 
3671 	vm_page_assert_locked(m);
3672 
3673 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3674 		return;
3675 
3676 	if (!vm_page_inactive(m)) {
3677 		vm_page_remque(m);
3678 		vm_page_enqueue(m, PQ_INACTIVE);
3679 	} else
3680 		vm_page_requeue(m);
3681 }
3682 
3683 /*
3684  * Move the specified page close to the head of the inactive queue,
3685  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3686  * As with regular enqueues, we use a per-CPU batch queue to reduce
3687  * contention on the page queue lock.
3688  *
3689  * The page must be locked.
3690  */
3691 void
3692 vm_page_deactivate_noreuse(vm_page_t m)
3693 {
3694 
3695 	vm_page_assert_locked(m);
3696 
3697 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3698 		return;
3699 
3700 	if (!vm_page_inactive(m))
3701 		vm_page_remque(m);
3702 	m->queue = PQ_INACTIVE;
3703 	if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3704 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3705 	vm_pqbatch_submit_page(m, PQ_INACTIVE);
3706 }
3707 
3708 /*
3709  * vm_page_launder
3710  *
3711  * 	Put a page in the laundry, or requeue it if it is already there.
3712  */
3713 void
3714 vm_page_launder(vm_page_t m)
3715 {
3716 
3717 	vm_page_assert_locked(m);
3718 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3719 		return;
3720 
3721 	if (vm_page_in_laundry(m))
3722 		vm_page_requeue(m);
3723 	else {
3724 		vm_page_remque(m);
3725 		vm_page_enqueue(m, PQ_LAUNDRY);
3726 	}
3727 }
3728 
3729 /*
3730  * vm_page_unswappable
3731  *
3732  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3733  */
3734 void
3735 vm_page_unswappable(vm_page_t m)
3736 {
3737 
3738 	vm_page_assert_locked(m);
3739 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3740 	    ("page %p already unswappable", m));
3741 
3742 	vm_page_remque(m);
3743 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
3744 }
3745 
3746 /*
3747  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3748  * if the page is freed and false otherwise.
3749  *
3750  * The page must be managed.  The page and its containing object must be
3751  * locked.
3752  */
3753 bool
3754 vm_page_try_to_free(vm_page_t m)
3755 {
3756 
3757 	vm_page_assert_locked(m);
3758 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3759 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3760 	if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m))
3761 		return (false);
3762 	if (m->object->ref_count != 0) {
3763 		pmap_remove_all(m);
3764 		if (m->dirty != 0)
3765 			return (false);
3766 	}
3767 	vm_page_free(m);
3768 	return (true);
3769 }
3770 
3771 /*
3772  * vm_page_advise
3773  *
3774  * 	Apply the specified advice to the given page.
3775  *
3776  *	The object and page must be locked.
3777  */
3778 void
3779 vm_page_advise(vm_page_t m, int advice)
3780 {
3781 
3782 	vm_page_assert_locked(m);
3783 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3784 	if (advice == MADV_FREE)
3785 		/*
3786 		 * Mark the page clean.  This will allow the page to be freed
3787 		 * without first paging it out.  MADV_FREE pages are often
3788 		 * quickly reused by malloc(3), so we do not do anything that
3789 		 * would result in a page fault on a later access.
3790 		 */
3791 		vm_page_undirty(m);
3792 	else if (advice != MADV_DONTNEED) {
3793 		if (advice == MADV_WILLNEED)
3794 			vm_page_activate(m);
3795 		return;
3796 	}
3797 
3798 	/*
3799 	 * Clear any references to the page.  Otherwise, the page daemon will
3800 	 * immediately reactivate the page.
3801 	 */
3802 	vm_page_aflag_clear(m, PGA_REFERENCED);
3803 
3804 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3805 		vm_page_dirty(m);
3806 
3807 	/*
3808 	 * Place clean pages near the head of the inactive queue rather than
3809 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3810 	 * the page will be reused quickly.  Dirty pages not already in the
3811 	 * laundry are moved there.
3812 	 */
3813 	if (m->dirty == 0)
3814 		vm_page_deactivate_noreuse(m);
3815 	else if (!vm_page_in_laundry(m))
3816 		vm_page_launder(m);
3817 }
3818 
3819 /*
3820  * Grab a page, waiting until we are waken up due to the page
3821  * changing state.  We keep on waiting, if the page continues
3822  * to be in the object.  If the page doesn't exist, first allocate it
3823  * and then conditionally zero it.
3824  *
3825  * This routine may sleep.
3826  *
3827  * The object must be locked on entry.  The lock will, however, be released
3828  * and reacquired if the routine sleeps.
3829  */
3830 vm_page_t
3831 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3832 {
3833 	vm_page_t m;
3834 	int sleep;
3835 	int pflags;
3836 
3837 	VM_OBJECT_ASSERT_WLOCKED(object);
3838 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3839 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3840 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3841 	pflags = allocflags &
3842 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3843 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3844 		pflags |= VM_ALLOC_WAITFAIL;
3845 retrylookup:
3846 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3847 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3848 		    vm_page_xbusied(m) : vm_page_busied(m);
3849 		if (sleep) {
3850 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3851 				return (NULL);
3852 			/*
3853 			 * Reference the page before unlocking and
3854 			 * sleeping so that the page daemon is less
3855 			 * likely to reclaim it.
3856 			 */
3857 			vm_page_aflag_set(m, PGA_REFERENCED);
3858 			vm_page_lock(m);
3859 			VM_OBJECT_WUNLOCK(object);
3860 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3861 			    VM_ALLOC_IGN_SBUSY) != 0);
3862 			VM_OBJECT_WLOCK(object);
3863 			goto retrylookup;
3864 		} else {
3865 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3866 				vm_page_lock(m);
3867 				vm_page_wire(m);
3868 				vm_page_unlock(m);
3869 			}
3870 			if ((allocflags &
3871 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3872 				vm_page_xbusy(m);
3873 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3874 				vm_page_sbusy(m);
3875 			return (m);
3876 		}
3877 	}
3878 	m = vm_page_alloc(object, pindex, pflags);
3879 	if (m == NULL) {
3880 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3881 			return (NULL);
3882 		goto retrylookup;
3883 	}
3884 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3885 		pmap_zero_page(m);
3886 	return (m);
3887 }
3888 
3889 /*
3890  * Return the specified range of pages from the given object.  For each
3891  * page offset within the range, if a page already exists within the object
3892  * at that offset and it is busy, then wait for it to change state.  If,
3893  * instead, the page doesn't exist, then allocate it.
3894  *
3895  * The caller must always specify an allocation class.
3896  *
3897  * allocation classes:
3898  *	VM_ALLOC_NORMAL		normal process request
3899  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3900  *
3901  * The caller must always specify that the pages are to be busied and/or
3902  * wired.
3903  *
3904  * optional allocation flags:
3905  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3906  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3907  *	VM_ALLOC_NOWAIT		do not sleep
3908  *	VM_ALLOC_SBUSY		set page to sbusy state
3909  *	VM_ALLOC_WIRED		wire the pages
3910  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3911  *
3912  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3913  * may return a partial prefix of the requested range.
3914  */
3915 int
3916 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3917     vm_page_t *ma, int count)
3918 {
3919 	vm_page_t m, mpred;
3920 	int pflags;
3921 	int i;
3922 	bool sleep;
3923 
3924 	VM_OBJECT_ASSERT_WLOCKED(object);
3925 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3926 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3927 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3928 	    (allocflags & VM_ALLOC_WIRED) != 0,
3929 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3930 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3931 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3932 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3933 	if (count == 0)
3934 		return (0);
3935 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3936 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3937 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3938 		pflags |= VM_ALLOC_WAITFAIL;
3939 	i = 0;
3940 retrylookup:
3941 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3942 	if (m == NULL || m->pindex != pindex + i) {
3943 		mpred = m;
3944 		m = NULL;
3945 	} else
3946 		mpred = TAILQ_PREV(m, pglist, listq);
3947 	for (; i < count; i++) {
3948 		if (m != NULL) {
3949 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3950 			    vm_page_xbusied(m) : vm_page_busied(m);
3951 			if (sleep) {
3952 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3953 					break;
3954 				/*
3955 				 * Reference the page before unlocking and
3956 				 * sleeping so that the page daemon is less
3957 				 * likely to reclaim it.
3958 				 */
3959 				vm_page_aflag_set(m, PGA_REFERENCED);
3960 				vm_page_lock(m);
3961 				VM_OBJECT_WUNLOCK(object);
3962 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
3963 				    VM_ALLOC_IGN_SBUSY) != 0);
3964 				VM_OBJECT_WLOCK(object);
3965 				goto retrylookup;
3966 			}
3967 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3968 				vm_page_lock(m);
3969 				vm_page_wire(m);
3970 				vm_page_unlock(m);
3971 			}
3972 			if ((allocflags & (VM_ALLOC_NOBUSY |
3973 			    VM_ALLOC_SBUSY)) == 0)
3974 				vm_page_xbusy(m);
3975 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3976 				vm_page_sbusy(m);
3977 		} else {
3978 			m = vm_page_alloc_after(object, pindex + i,
3979 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
3980 			if (m == NULL) {
3981 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3982 					break;
3983 				goto retrylookup;
3984 			}
3985 		}
3986 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3987 			if ((m->flags & PG_ZERO) == 0)
3988 				pmap_zero_page(m);
3989 			m->valid = VM_PAGE_BITS_ALL;
3990 		}
3991 		ma[i] = mpred = m;
3992 		m = vm_page_next(m);
3993 	}
3994 	return (i);
3995 }
3996 
3997 /*
3998  * Mapping function for valid or dirty bits in a page.
3999  *
4000  * Inputs are required to range within a page.
4001  */
4002 vm_page_bits_t
4003 vm_page_bits(int base, int size)
4004 {
4005 	int first_bit;
4006 	int last_bit;
4007 
4008 	KASSERT(
4009 	    base + size <= PAGE_SIZE,
4010 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4011 	);
4012 
4013 	if (size == 0)		/* handle degenerate case */
4014 		return (0);
4015 
4016 	first_bit = base >> DEV_BSHIFT;
4017 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4018 
4019 	return (((vm_page_bits_t)2 << last_bit) -
4020 	    ((vm_page_bits_t)1 << first_bit));
4021 }
4022 
4023 /*
4024  *	vm_page_set_valid_range:
4025  *
4026  *	Sets portions of a page valid.  The arguments are expected
4027  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4028  *	of any partial chunks touched by the range.  The invalid portion of
4029  *	such chunks will be zeroed.
4030  *
4031  *	(base + size) must be less then or equal to PAGE_SIZE.
4032  */
4033 void
4034 vm_page_set_valid_range(vm_page_t m, int base, int size)
4035 {
4036 	int endoff, frag;
4037 
4038 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4039 	if (size == 0)	/* handle degenerate case */
4040 		return;
4041 
4042 	/*
4043 	 * If the base is not DEV_BSIZE aligned and the valid
4044 	 * bit is clear, we have to zero out a portion of the
4045 	 * first block.
4046 	 */
4047 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4048 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4049 		pmap_zero_page_area(m, frag, base - frag);
4050 
4051 	/*
4052 	 * If the ending offset is not DEV_BSIZE aligned and the
4053 	 * valid bit is clear, we have to zero out a portion of
4054 	 * the last block.
4055 	 */
4056 	endoff = base + size;
4057 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4058 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4059 		pmap_zero_page_area(m, endoff,
4060 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4061 
4062 	/*
4063 	 * Assert that no previously invalid block that is now being validated
4064 	 * is already dirty.
4065 	 */
4066 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4067 	    ("vm_page_set_valid_range: page %p is dirty", m));
4068 
4069 	/*
4070 	 * Set valid bits inclusive of any overlap.
4071 	 */
4072 	m->valid |= vm_page_bits(base, size);
4073 }
4074 
4075 /*
4076  * Clear the given bits from the specified page's dirty field.
4077  */
4078 static __inline void
4079 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4080 {
4081 	uintptr_t addr;
4082 #if PAGE_SIZE < 16384
4083 	int shift;
4084 #endif
4085 
4086 	/*
4087 	 * If the object is locked and the page is neither exclusive busy nor
4088 	 * write mapped, then the page's dirty field cannot possibly be
4089 	 * set by a concurrent pmap operation.
4090 	 */
4091 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4092 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4093 		m->dirty &= ~pagebits;
4094 	else {
4095 		/*
4096 		 * The pmap layer can call vm_page_dirty() without
4097 		 * holding a distinguished lock.  The combination of
4098 		 * the object's lock and an atomic operation suffice
4099 		 * to guarantee consistency of the page dirty field.
4100 		 *
4101 		 * For PAGE_SIZE == 32768 case, compiler already
4102 		 * properly aligns the dirty field, so no forcible
4103 		 * alignment is needed. Only require existence of
4104 		 * atomic_clear_64 when page size is 32768.
4105 		 */
4106 		addr = (uintptr_t)&m->dirty;
4107 #if PAGE_SIZE == 32768
4108 		atomic_clear_64((uint64_t *)addr, pagebits);
4109 #elif PAGE_SIZE == 16384
4110 		atomic_clear_32((uint32_t *)addr, pagebits);
4111 #else		/* PAGE_SIZE <= 8192 */
4112 		/*
4113 		 * Use a trick to perform a 32-bit atomic on the
4114 		 * containing aligned word, to not depend on the existence
4115 		 * of atomic_clear_{8, 16}.
4116 		 */
4117 		shift = addr & (sizeof(uint32_t) - 1);
4118 #if BYTE_ORDER == BIG_ENDIAN
4119 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4120 #else
4121 		shift *= NBBY;
4122 #endif
4123 		addr &= ~(sizeof(uint32_t) - 1);
4124 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
4125 #endif		/* PAGE_SIZE */
4126 	}
4127 }
4128 
4129 /*
4130  *	vm_page_set_validclean:
4131  *
4132  *	Sets portions of a page valid and clean.  The arguments are expected
4133  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4134  *	of any partial chunks touched by the range.  The invalid portion of
4135  *	such chunks will be zero'd.
4136  *
4137  *	(base + size) must be less then or equal to PAGE_SIZE.
4138  */
4139 void
4140 vm_page_set_validclean(vm_page_t m, int base, int size)
4141 {
4142 	vm_page_bits_t oldvalid, pagebits;
4143 	int endoff, frag;
4144 
4145 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4146 	if (size == 0)	/* handle degenerate case */
4147 		return;
4148 
4149 	/*
4150 	 * If the base is not DEV_BSIZE aligned and the valid
4151 	 * bit is clear, we have to zero out a portion of the
4152 	 * first block.
4153 	 */
4154 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4155 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4156 		pmap_zero_page_area(m, frag, base - frag);
4157 
4158 	/*
4159 	 * If the ending offset is not DEV_BSIZE aligned and the
4160 	 * valid bit is clear, we have to zero out a portion of
4161 	 * the last block.
4162 	 */
4163 	endoff = base + size;
4164 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4165 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4166 		pmap_zero_page_area(m, endoff,
4167 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4168 
4169 	/*
4170 	 * Set valid, clear dirty bits.  If validating the entire
4171 	 * page we can safely clear the pmap modify bit.  We also
4172 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
4173 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4174 	 * be set again.
4175 	 *
4176 	 * We set valid bits inclusive of any overlap, but we can only
4177 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4178 	 * the range.
4179 	 */
4180 	oldvalid = m->valid;
4181 	pagebits = vm_page_bits(base, size);
4182 	m->valid |= pagebits;
4183 #if 0	/* NOT YET */
4184 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4185 		frag = DEV_BSIZE - frag;
4186 		base += frag;
4187 		size -= frag;
4188 		if (size < 0)
4189 			size = 0;
4190 	}
4191 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4192 #endif
4193 	if (base == 0 && size == PAGE_SIZE) {
4194 		/*
4195 		 * The page can only be modified within the pmap if it is
4196 		 * mapped, and it can only be mapped if it was previously
4197 		 * fully valid.
4198 		 */
4199 		if (oldvalid == VM_PAGE_BITS_ALL)
4200 			/*
4201 			 * Perform the pmap_clear_modify() first.  Otherwise,
4202 			 * a concurrent pmap operation, such as
4203 			 * pmap_protect(), could clear a modification in the
4204 			 * pmap and set the dirty field on the page before
4205 			 * pmap_clear_modify() had begun and after the dirty
4206 			 * field was cleared here.
4207 			 */
4208 			pmap_clear_modify(m);
4209 		m->dirty = 0;
4210 		m->oflags &= ~VPO_NOSYNC;
4211 	} else if (oldvalid != VM_PAGE_BITS_ALL)
4212 		m->dirty &= ~pagebits;
4213 	else
4214 		vm_page_clear_dirty_mask(m, pagebits);
4215 }
4216 
4217 void
4218 vm_page_clear_dirty(vm_page_t m, int base, int size)
4219 {
4220 
4221 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4222 }
4223 
4224 /*
4225  *	vm_page_set_invalid:
4226  *
4227  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4228  *	valid and dirty bits for the effected areas are cleared.
4229  */
4230 void
4231 vm_page_set_invalid(vm_page_t m, int base, int size)
4232 {
4233 	vm_page_bits_t bits;
4234 	vm_object_t object;
4235 
4236 	object = m->object;
4237 	VM_OBJECT_ASSERT_WLOCKED(object);
4238 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4239 	    size >= object->un_pager.vnp.vnp_size)
4240 		bits = VM_PAGE_BITS_ALL;
4241 	else
4242 		bits = vm_page_bits(base, size);
4243 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4244 	    bits != 0)
4245 		pmap_remove_all(m);
4246 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4247 	    !pmap_page_is_mapped(m),
4248 	    ("vm_page_set_invalid: page %p is mapped", m));
4249 	m->valid &= ~bits;
4250 	m->dirty &= ~bits;
4251 }
4252 
4253 /*
4254  * vm_page_zero_invalid()
4255  *
4256  *	The kernel assumes that the invalid portions of a page contain
4257  *	garbage, but such pages can be mapped into memory by user code.
4258  *	When this occurs, we must zero out the non-valid portions of the
4259  *	page so user code sees what it expects.
4260  *
4261  *	Pages are most often semi-valid when the end of a file is mapped
4262  *	into memory and the file's size is not page aligned.
4263  */
4264 void
4265 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4266 {
4267 	int b;
4268 	int i;
4269 
4270 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4271 	/*
4272 	 * Scan the valid bits looking for invalid sections that
4273 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4274 	 * valid bit may be set ) have already been zeroed by
4275 	 * vm_page_set_validclean().
4276 	 */
4277 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4278 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4279 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4280 			if (i > b) {
4281 				pmap_zero_page_area(m,
4282 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4283 			}
4284 			b = i + 1;
4285 		}
4286 	}
4287 
4288 	/*
4289 	 * setvalid is TRUE when we can safely set the zero'd areas
4290 	 * as being valid.  We can do this if there are no cache consistancy
4291 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4292 	 */
4293 	if (setvalid)
4294 		m->valid = VM_PAGE_BITS_ALL;
4295 }
4296 
4297 /*
4298  *	vm_page_is_valid:
4299  *
4300  *	Is (partial) page valid?  Note that the case where size == 0
4301  *	will return FALSE in the degenerate case where the page is
4302  *	entirely invalid, and TRUE otherwise.
4303  */
4304 int
4305 vm_page_is_valid(vm_page_t m, int base, int size)
4306 {
4307 	vm_page_bits_t bits;
4308 
4309 	VM_OBJECT_ASSERT_LOCKED(m->object);
4310 	bits = vm_page_bits(base, size);
4311 	return (m->valid != 0 && (m->valid & bits) == bits);
4312 }
4313 
4314 /*
4315  * Returns true if all of the specified predicates are true for the entire
4316  * (super)page and false otherwise.
4317  */
4318 bool
4319 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4320 {
4321 	vm_object_t object;
4322 	int i, npages;
4323 
4324 	object = m->object;
4325 	if (skip_m != NULL && skip_m->object != object)
4326 		return (false);
4327 	VM_OBJECT_ASSERT_LOCKED(object);
4328 	npages = atop(pagesizes[m->psind]);
4329 
4330 	/*
4331 	 * The physically contiguous pages that make up a superpage, i.e., a
4332 	 * page with a page size index ("psind") greater than zero, will
4333 	 * occupy adjacent entries in vm_page_array[].
4334 	 */
4335 	for (i = 0; i < npages; i++) {
4336 		/* Always test object consistency, including "skip_m". */
4337 		if (m[i].object != object)
4338 			return (false);
4339 		if (&m[i] == skip_m)
4340 			continue;
4341 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4342 			return (false);
4343 		if ((flags & PS_ALL_DIRTY) != 0) {
4344 			/*
4345 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4346 			 * might stop this case from spuriously returning
4347 			 * "false".  However, that would require a write lock
4348 			 * on the object containing "m[i]".
4349 			 */
4350 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4351 				return (false);
4352 		}
4353 		if ((flags & PS_ALL_VALID) != 0 &&
4354 		    m[i].valid != VM_PAGE_BITS_ALL)
4355 			return (false);
4356 	}
4357 	return (true);
4358 }
4359 
4360 /*
4361  * Set the page's dirty bits if the page is modified.
4362  */
4363 void
4364 vm_page_test_dirty(vm_page_t m)
4365 {
4366 
4367 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4368 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4369 		vm_page_dirty(m);
4370 }
4371 
4372 void
4373 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4374 {
4375 
4376 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4377 }
4378 
4379 void
4380 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4381 {
4382 
4383 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4384 }
4385 
4386 int
4387 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4388 {
4389 
4390 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4391 }
4392 
4393 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4394 void
4395 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4396 {
4397 
4398 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4399 }
4400 
4401 void
4402 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4403 {
4404 
4405 	mtx_assert_(vm_page_lockptr(m), a, file, line);
4406 }
4407 #endif
4408 
4409 #ifdef INVARIANTS
4410 void
4411 vm_page_object_lock_assert(vm_page_t m)
4412 {
4413 
4414 	/*
4415 	 * Certain of the page's fields may only be modified by the
4416 	 * holder of the containing object's lock or the exclusive busy.
4417 	 * holder.  Unfortunately, the holder of the write busy is
4418 	 * not recorded, and thus cannot be checked here.
4419 	 */
4420 	if (m->object != NULL && !vm_page_xbusied(m))
4421 		VM_OBJECT_ASSERT_WLOCKED(m->object);
4422 }
4423 
4424 void
4425 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4426 {
4427 
4428 	if ((bits & PGA_WRITEABLE) == 0)
4429 		return;
4430 
4431 	/*
4432 	 * The PGA_WRITEABLE flag can only be set if the page is
4433 	 * managed, is exclusively busied or the object is locked.
4434 	 * Currently, this flag is only set by pmap_enter().
4435 	 */
4436 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4437 	    ("PGA_WRITEABLE on unmanaged page"));
4438 	if (!vm_page_xbusied(m))
4439 		VM_OBJECT_ASSERT_LOCKED(m->object);
4440 }
4441 #endif
4442 
4443 #include "opt_ddb.h"
4444 #ifdef DDB
4445 #include <sys/kernel.h>
4446 
4447 #include <ddb/ddb.h>
4448 
4449 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4450 {
4451 
4452 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4453 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4454 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4455 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4456 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4457 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4458 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4459 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4460 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4461 }
4462 
4463 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4464 {
4465 	int dom;
4466 
4467 	db_printf("pq_free %d\n", vm_free_count());
4468 	for (dom = 0; dom < vm_ndomains; dom++) {
4469 		db_printf(
4470     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4471 		    dom,
4472 		    vm_dom[dom].vmd_page_count,
4473 		    vm_dom[dom].vmd_free_count,
4474 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4475 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4476 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4477 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4478 	}
4479 }
4480 
4481 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4482 {
4483 	vm_page_t m;
4484 	boolean_t phys;
4485 
4486 	if (!have_addr) {
4487 		db_printf("show pginfo addr\n");
4488 		return;
4489 	}
4490 
4491 	phys = strchr(modif, 'p') != NULL;
4492 	if (phys)
4493 		m = PHYS_TO_VM_PAGE(addr);
4494 	else
4495 		m = (vm_page_t)addr;
4496 	db_printf(
4497     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4498     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4499 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4500 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4501 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4502 }
4503 #endif /* DDB */
4504