xref: /freebsd/sys/vm/vm_page.c (revision ae500c1ff8974130f7f2692772cf288b90349e0d)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *			GENERAL RULES ON VM_PAGE MANIPULATION
67  *
68  *	- A page queue lock is required when adding or removing a page from a
69  *	  page queue regardless of other locks or the busy state of a page.
70  *
71  *		* In general, no thread besides the page daemon can acquire or
72  *		  hold more than one page queue lock at a time.
73  *
74  *		* The page daemon can acquire and hold any pair of page queue
75  *		  locks in any order.
76  *
77  *	- The object lock is required when inserting or removing
78  *	  pages from an object (vm_page_insert() or vm_page_remove()).
79  *
80  */
81 
82 /*
83  *	Resident memory management module.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 #include "opt_vm.h"
90 
91 #include <sys/param.h>
92 #include <sys/systm.h>
93 #include <sys/lock.h>
94 #include <sys/domainset.h>
95 #include <sys/kernel.h>
96 #include <sys/limits.h>
97 #include <sys/linker.h>
98 #include <sys/malloc.h>
99 #include <sys/mman.h>
100 #include <sys/msgbuf.h>
101 #include <sys/mutex.h>
102 #include <sys/proc.h>
103 #include <sys/rwlock.h>
104 #include <sys/sbuf.h>
105 #include <sys/sched.h>
106 #include <sys/smp.h>
107 #include <sys/sysctl.h>
108 #include <sys/vmmeter.h>
109 #include <sys/vnode.h>
110 
111 #include <vm/vm.h>
112 #include <vm/pmap.h>
113 #include <vm/vm_param.h>
114 #include <vm/vm_domainset.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_map.h>
117 #include <vm/vm_object.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_pageout.h>
120 #include <vm/vm_phys.h>
121 #include <vm/vm_pagequeue.h>
122 #include <vm/vm_pager.h>
123 #include <vm/vm_radix.h>
124 #include <vm/vm_reserv.h>
125 #include <vm/vm_extern.h>
126 #include <vm/uma.h>
127 #include <vm/uma_int.h>
128 
129 #include <machine/md_var.h>
130 
131 extern int	uma_startup_count(int);
132 extern void	uma_startup(void *, int);
133 extern int	vmem_startup_count(void);
134 
135 struct vm_domain vm_dom[MAXMEMDOM];
136 
137 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
138 
139 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
140 
141 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
142 /* The following fields are protected by the domainset lock. */
143 domainset_t __exclusive_cache_line vm_min_domains;
144 domainset_t __exclusive_cache_line vm_severe_domains;
145 static int vm_min_waiters;
146 static int vm_severe_waiters;
147 static int vm_pageproc_waiters;
148 
149 /*
150  * bogus page -- for I/O to/from partially complete buffers,
151  * or for paging into sparsely invalid regions.
152  */
153 vm_page_t bogus_page;
154 
155 vm_page_t vm_page_array;
156 long vm_page_array_size;
157 long first_page;
158 
159 static int boot_pages;
160 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
161     &boot_pages, 0,
162     "number of pages allocated for bootstrapping the VM system");
163 
164 static int pa_tryrelock_restart;
165 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
166     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
167 
168 static TAILQ_HEAD(, vm_page) blacklist_head;
169 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
170 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
171     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
172 
173 static uma_zone_t fakepg_zone;
174 
175 static void vm_page_alloc_check(vm_page_t m);
176 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
177 static void vm_page_dequeue_complete(vm_page_t m);
178 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
179 static void vm_page_init(void *dummy);
180 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
181     vm_pindex_t pindex, vm_page_t mpred);
182 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
183     vm_page_t mpred);
184 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
185     vm_page_t m_run, vm_paddr_t high);
186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
187     int req);
188 static int vm_page_import(void *arg, void **store, int cnt, int domain,
189     int flags);
190 static void vm_page_release(void *arg, void **store, int cnt);
191 
192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
193 
194 static void
195 vm_page_init(void *dummy)
196 {
197 
198 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
199 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
200 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
201 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
202 }
203 
204 /*
205  * The cache page zone is initialized later since we need to be able to allocate
206  * pages before UMA is fully initialized.
207  */
208 static void
209 vm_page_init_cache_zones(void *dummy __unused)
210 {
211 	struct vm_domain *vmd;
212 	int i;
213 
214 	for (i = 0; i < vm_ndomains; i++) {
215 		vmd = VM_DOMAIN(i);
216 		/*
217 		 * Don't allow the page cache to take up more than .25% of
218 		 * memory.
219 		 */
220 		if (vmd->vmd_page_count / 400 < 256 * mp_ncpus)
221 			continue;
222 		vmd->vmd_pgcache = uma_zcache_create("vm pgcache",
223 		    sizeof(struct vm_page), NULL, NULL, NULL, NULL,
224 		    vm_page_import, vm_page_release, vmd,
225 		    UMA_ZONE_NOBUCKETCACHE | UMA_ZONE_MAXBUCKET | UMA_ZONE_VM);
226 	}
227 }
228 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
229 
230 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
231 #if PAGE_SIZE == 32768
232 #ifdef CTASSERT
233 CTASSERT(sizeof(u_long) >= 8);
234 #endif
235 #endif
236 
237 /*
238  * Try to acquire a physical address lock while a pmap is locked.  If we
239  * fail to trylock we unlock and lock the pmap directly and cache the
240  * locked pa in *locked.  The caller should then restart their loop in case
241  * the virtual to physical mapping has changed.
242  */
243 int
244 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
245 {
246 	vm_paddr_t lockpa;
247 
248 	lockpa = *locked;
249 	*locked = pa;
250 	if (lockpa) {
251 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
252 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
253 			return (0);
254 		PA_UNLOCK(lockpa);
255 	}
256 	if (PA_TRYLOCK(pa))
257 		return (0);
258 	PMAP_UNLOCK(pmap);
259 	atomic_add_int(&pa_tryrelock_restart, 1);
260 	PA_LOCK(pa);
261 	PMAP_LOCK(pmap);
262 	return (EAGAIN);
263 }
264 
265 /*
266  *	vm_set_page_size:
267  *
268  *	Sets the page size, perhaps based upon the memory
269  *	size.  Must be called before any use of page-size
270  *	dependent functions.
271  */
272 void
273 vm_set_page_size(void)
274 {
275 	if (vm_cnt.v_page_size == 0)
276 		vm_cnt.v_page_size = PAGE_SIZE;
277 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
278 		panic("vm_set_page_size: page size not a power of two");
279 }
280 
281 /*
282  *	vm_page_blacklist_next:
283  *
284  *	Find the next entry in the provided string of blacklist
285  *	addresses.  Entries are separated by space, comma, or newline.
286  *	If an invalid integer is encountered then the rest of the
287  *	string is skipped.  Updates the list pointer to the next
288  *	character, or NULL if the string is exhausted or invalid.
289  */
290 static vm_paddr_t
291 vm_page_blacklist_next(char **list, char *end)
292 {
293 	vm_paddr_t bad;
294 	char *cp, *pos;
295 
296 	if (list == NULL || *list == NULL)
297 		return (0);
298 	if (**list =='\0') {
299 		*list = NULL;
300 		return (0);
301 	}
302 
303 	/*
304 	 * If there's no end pointer then the buffer is coming from
305 	 * the kenv and we know it's null-terminated.
306 	 */
307 	if (end == NULL)
308 		end = *list + strlen(*list);
309 
310 	/* Ensure that strtoq() won't walk off the end */
311 	if (*end != '\0') {
312 		if (*end == '\n' || *end == ' ' || *end  == ',')
313 			*end = '\0';
314 		else {
315 			printf("Blacklist not terminated, skipping\n");
316 			*list = NULL;
317 			return (0);
318 		}
319 	}
320 
321 	for (pos = *list; *pos != '\0'; pos = cp) {
322 		bad = strtoq(pos, &cp, 0);
323 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
324 			if (bad == 0) {
325 				if (++cp < end)
326 					continue;
327 				else
328 					break;
329 			}
330 		} else
331 			break;
332 		if (*cp == '\0' || ++cp >= end)
333 			*list = NULL;
334 		else
335 			*list = cp;
336 		return (trunc_page(bad));
337 	}
338 	printf("Garbage in RAM blacklist, skipping\n");
339 	*list = NULL;
340 	return (0);
341 }
342 
343 bool
344 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
345 {
346 	struct vm_domain *vmd;
347 	vm_page_t m;
348 	int ret;
349 
350 	m = vm_phys_paddr_to_vm_page(pa);
351 	if (m == NULL)
352 		return (true); /* page does not exist, no failure */
353 
354 	vmd = vm_pagequeue_domain(m);
355 	vm_domain_free_lock(vmd);
356 	ret = vm_phys_unfree_page(m);
357 	vm_domain_free_unlock(vmd);
358 	if (ret != 0) {
359 		vm_domain_freecnt_inc(vmd, -1);
360 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
361 		if (verbose)
362 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
363 	}
364 	return (ret);
365 }
366 
367 /*
368  *	vm_page_blacklist_check:
369  *
370  *	Iterate through the provided string of blacklist addresses, pulling
371  *	each entry out of the physical allocator free list and putting it
372  *	onto a list for reporting via the vm.page_blacklist sysctl.
373  */
374 static void
375 vm_page_blacklist_check(char *list, char *end)
376 {
377 	vm_paddr_t pa;
378 	char *next;
379 
380 	next = list;
381 	while (next != NULL) {
382 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
383 			continue;
384 		vm_page_blacklist_add(pa, bootverbose);
385 	}
386 }
387 
388 /*
389  *	vm_page_blacklist_load:
390  *
391  *	Search for a special module named "ram_blacklist".  It'll be a
392  *	plain text file provided by the user via the loader directive
393  *	of the same name.
394  */
395 static void
396 vm_page_blacklist_load(char **list, char **end)
397 {
398 	void *mod;
399 	u_char *ptr;
400 	u_int len;
401 
402 	mod = NULL;
403 	ptr = NULL;
404 
405 	mod = preload_search_by_type("ram_blacklist");
406 	if (mod != NULL) {
407 		ptr = preload_fetch_addr(mod);
408 		len = preload_fetch_size(mod);
409         }
410 	*list = ptr;
411 	if (ptr != NULL)
412 		*end = ptr + len;
413 	else
414 		*end = NULL;
415 	return;
416 }
417 
418 static int
419 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
420 {
421 	vm_page_t m;
422 	struct sbuf sbuf;
423 	int error, first;
424 
425 	first = 1;
426 	error = sysctl_wire_old_buffer(req, 0);
427 	if (error != 0)
428 		return (error);
429 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
430 	TAILQ_FOREACH(m, &blacklist_head, listq) {
431 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
432 		    (uintmax_t)m->phys_addr);
433 		first = 0;
434 	}
435 	error = sbuf_finish(&sbuf);
436 	sbuf_delete(&sbuf);
437 	return (error);
438 }
439 
440 /*
441  * Initialize a dummy page for use in scans of the specified paging queue.
442  * In principle, this function only needs to set the flag PG_MARKER.
443  * Nonetheless, it write busies and initializes the hold count to one as
444  * safety precautions.
445  */
446 static void
447 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
448 {
449 
450 	bzero(marker, sizeof(*marker));
451 	marker->flags = PG_MARKER;
452 	marker->aflags = aflags;
453 	marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
454 	marker->queue = queue;
455 	marker->hold_count = 1;
456 }
457 
458 static void
459 vm_page_domain_init(int domain)
460 {
461 	struct vm_domain *vmd;
462 	struct vm_pagequeue *pq;
463 	int i;
464 
465 	vmd = VM_DOMAIN(domain);
466 	bzero(vmd, sizeof(*vmd));
467 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
468 	    "vm inactive pagequeue";
469 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
470 	    "vm active pagequeue";
471 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
472 	    "vm laundry pagequeue";
473 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
474 	    "vm unswappable pagequeue";
475 	vmd->vmd_domain = domain;
476 	vmd->vmd_page_count = 0;
477 	vmd->vmd_free_count = 0;
478 	vmd->vmd_segs = 0;
479 	vmd->vmd_oom = FALSE;
480 	for (i = 0; i < PQ_COUNT; i++) {
481 		pq = &vmd->vmd_pagequeues[i];
482 		TAILQ_INIT(&pq->pq_pl);
483 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
484 		    MTX_DEF | MTX_DUPOK);
485 		pq->pq_pdpages = 0;
486 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
487 	}
488 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
489 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
490 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
491 
492 	/*
493 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
494 	 * insertions.
495 	 */
496 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
497 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
498 	    &vmd->vmd_inacthead, plinks.q);
499 
500 	/*
501 	 * The clock pages are used to implement active queue scanning without
502 	 * requeues.  Scans start at clock[0], which is advanced after the scan
503 	 * ends.  When the two clock hands meet, they are reset and scanning
504 	 * resumes from the head of the queue.
505 	 */
506 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
507 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
508 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
509 	    &vmd->vmd_clock[0], plinks.q);
510 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
511 	    &vmd->vmd_clock[1], plinks.q);
512 }
513 
514 /*
515  * Initialize a physical page in preparation for adding it to the free
516  * lists.
517  */
518 static void
519 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
520 {
521 
522 	m->object = NULL;
523 	m->wire_count = 0;
524 	m->busy_lock = VPB_UNBUSIED;
525 	m->hold_count = 0;
526 	m->flags = m->aflags = 0;
527 	m->phys_addr = pa;
528 	m->queue = PQ_NONE;
529 	m->psind = 0;
530 	m->segind = segind;
531 	m->order = VM_NFREEORDER;
532 	m->pool = VM_FREEPOOL_DEFAULT;
533 	m->valid = m->dirty = 0;
534 	pmap_page_init(m);
535 }
536 
537 /*
538  *	vm_page_startup:
539  *
540  *	Initializes the resident memory module.  Allocates physical memory for
541  *	bootstrapping UMA and some data structures that are used to manage
542  *	physical pages.  Initializes these structures, and populates the free
543  *	page queues.
544  */
545 vm_offset_t
546 vm_page_startup(vm_offset_t vaddr)
547 {
548 	struct vm_phys_seg *seg;
549 	vm_page_t m;
550 	char *list, *listend;
551 	vm_offset_t mapped;
552 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
553 	vm_paddr_t biggestsize, last_pa, pa;
554 	u_long pagecount;
555 	int biggestone, i, segind;
556 #ifdef WITNESS
557 	int witness_size;
558 #endif
559 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
560 	long ii;
561 #endif
562 
563 	biggestsize = 0;
564 	biggestone = 0;
565 	vaddr = round_page(vaddr);
566 
567 	for (i = 0; phys_avail[i + 1]; i += 2) {
568 		phys_avail[i] = round_page(phys_avail[i]);
569 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
570 	}
571 	for (i = 0; phys_avail[i + 1]; i += 2) {
572 		size = phys_avail[i + 1] - phys_avail[i];
573 		if (size > biggestsize) {
574 			biggestone = i;
575 			biggestsize = size;
576 		}
577 	}
578 
579 	end = phys_avail[biggestone+1];
580 
581 	/*
582 	 * Initialize the page and queue locks.
583 	 */
584 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
585 	for (i = 0; i < PA_LOCK_COUNT; i++)
586 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
587 	for (i = 0; i < vm_ndomains; i++)
588 		vm_page_domain_init(i);
589 
590 	/*
591 	 * Allocate memory for use when boot strapping the kernel memory
592 	 * allocator.  Tell UMA how many zones we are going to create
593 	 * before going fully functional.  UMA will add its zones.
594 	 *
595 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
596 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
597 	 */
598 	boot_pages = uma_startup_count(8);
599 
600 #ifndef UMA_MD_SMALL_ALLOC
601 	/* vmem_startup() calls uma_prealloc(). */
602 	boot_pages += vmem_startup_count();
603 	/* vm_map_startup() calls uma_prealloc(). */
604 	boot_pages += howmany(MAX_KMAP,
605 	    UMA_SLAB_SPACE / sizeof(struct vm_map));
606 
607 	/*
608 	 * Before going fully functional kmem_init() does allocation
609 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
610 	 */
611 	boot_pages += 2;
612 #endif
613 	/*
614 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
615 	 * manually fetch the value.
616 	 */
617 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
618 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
619 	new_end = trunc_page(new_end);
620 	mapped = pmap_map(&vaddr, new_end, end,
621 	    VM_PROT_READ | VM_PROT_WRITE);
622 	bzero((void *)mapped, end - new_end);
623 	uma_startup((void *)mapped, boot_pages);
624 
625 #ifdef WITNESS
626 	witness_size = round_page(witness_startup_count());
627 	new_end -= witness_size;
628 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
629 	    VM_PROT_READ | VM_PROT_WRITE);
630 	bzero((void *)mapped, witness_size);
631 	witness_startup((void *)mapped);
632 #endif
633 
634 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
635     defined(__i386__) || defined(__mips__)
636 	/*
637 	 * Allocate a bitmap to indicate that a random physical page
638 	 * needs to be included in a minidump.
639 	 *
640 	 * The amd64 port needs this to indicate which direct map pages
641 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
642 	 *
643 	 * However, i386 still needs this workspace internally within the
644 	 * minidump code.  In theory, they are not needed on i386, but are
645 	 * included should the sf_buf code decide to use them.
646 	 */
647 	last_pa = 0;
648 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
649 		if (dump_avail[i + 1] > last_pa)
650 			last_pa = dump_avail[i + 1];
651 	page_range = last_pa / PAGE_SIZE;
652 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
653 	new_end -= vm_page_dump_size;
654 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
655 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
656 	bzero((void *)vm_page_dump, vm_page_dump_size);
657 #else
658 	(void)last_pa;
659 #endif
660 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
661 	/*
662 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
663 	 * in a crash dump.  When pmap_map() uses the direct map, they are
664 	 * not automatically included.
665 	 */
666 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
667 		dump_add_page(pa);
668 #endif
669 	phys_avail[biggestone + 1] = new_end;
670 #ifdef __amd64__
671 	/*
672 	 * Request that the physical pages underlying the message buffer be
673 	 * included in a crash dump.  Since the message buffer is accessed
674 	 * through the direct map, they are not automatically included.
675 	 */
676 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
677 	last_pa = pa + round_page(msgbufsize);
678 	while (pa < last_pa) {
679 		dump_add_page(pa);
680 		pa += PAGE_SIZE;
681 	}
682 #endif
683 	/*
684 	 * Compute the number of pages of memory that will be available for
685 	 * use, taking into account the overhead of a page structure per page.
686 	 * In other words, solve
687 	 *	"available physical memory" - round_page(page_range *
688 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
689 	 * for page_range.
690 	 */
691 	low_avail = phys_avail[0];
692 	high_avail = phys_avail[1];
693 	for (i = 0; i < vm_phys_nsegs; i++) {
694 		if (vm_phys_segs[i].start < low_avail)
695 			low_avail = vm_phys_segs[i].start;
696 		if (vm_phys_segs[i].end > high_avail)
697 			high_avail = vm_phys_segs[i].end;
698 	}
699 	/* Skip the first chunk.  It is already accounted for. */
700 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
701 		if (phys_avail[i] < low_avail)
702 			low_avail = phys_avail[i];
703 		if (phys_avail[i + 1] > high_avail)
704 			high_avail = phys_avail[i + 1];
705 	}
706 	first_page = low_avail / PAGE_SIZE;
707 #ifdef VM_PHYSSEG_SPARSE
708 	size = 0;
709 	for (i = 0; i < vm_phys_nsegs; i++)
710 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
711 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
712 		size += phys_avail[i + 1] - phys_avail[i];
713 #elif defined(VM_PHYSSEG_DENSE)
714 	size = high_avail - low_avail;
715 #else
716 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
717 #endif
718 
719 #ifdef VM_PHYSSEG_DENSE
720 	/*
721 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
722 	 * the overhead of a page structure per page only if vm_page_array is
723 	 * allocated from the last physical memory chunk.  Otherwise, we must
724 	 * allocate page structures representing the physical memory
725 	 * underlying vm_page_array, even though they will not be used.
726 	 */
727 	if (new_end != high_avail)
728 		page_range = size / PAGE_SIZE;
729 	else
730 #endif
731 	{
732 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
733 
734 		/*
735 		 * If the partial bytes remaining are large enough for
736 		 * a page (PAGE_SIZE) without a corresponding
737 		 * 'struct vm_page', then new_end will contain an
738 		 * extra page after subtracting the length of the VM
739 		 * page array.  Compensate by subtracting an extra
740 		 * page from new_end.
741 		 */
742 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
743 			if (new_end == high_avail)
744 				high_avail -= PAGE_SIZE;
745 			new_end -= PAGE_SIZE;
746 		}
747 	}
748 	end = new_end;
749 
750 	/*
751 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
752 	 * However, because this page is allocated from KVM, out-of-bounds
753 	 * accesses using the direct map will not be trapped.
754 	 */
755 	vaddr += PAGE_SIZE;
756 
757 	/*
758 	 * Allocate physical memory for the page structures, and map it.
759 	 */
760 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
761 	mapped = pmap_map(&vaddr, new_end, end,
762 	    VM_PROT_READ | VM_PROT_WRITE);
763 	vm_page_array = (vm_page_t)mapped;
764 	vm_page_array_size = page_range;
765 
766 #if VM_NRESERVLEVEL > 0
767 	/*
768 	 * Allocate physical memory for the reservation management system's
769 	 * data structures, and map it.
770 	 */
771 	if (high_avail == end)
772 		high_avail = new_end;
773 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
774 #endif
775 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
776 	/*
777 	 * Include vm_page_array and vm_reserv_array in a crash dump.
778 	 */
779 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
780 		dump_add_page(pa);
781 #endif
782 	phys_avail[biggestone + 1] = new_end;
783 
784 	/*
785 	 * Add physical memory segments corresponding to the available
786 	 * physical pages.
787 	 */
788 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
789 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
790 
791 	/*
792 	 * Initialize the physical memory allocator.
793 	 */
794 	vm_phys_init();
795 
796 	/*
797 	 * Initialize the page structures and add every available page to the
798 	 * physical memory allocator's free lists.
799 	 */
800 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
801 	for (ii = 0; ii < vm_page_array_size; ii++) {
802 		m = &vm_page_array[ii];
803 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
804 		m->flags = PG_FICTITIOUS;
805 	}
806 #endif
807 	vm_cnt.v_page_count = 0;
808 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
809 		seg = &vm_phys_segs[segind];
810 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
811 		    m++, pa += PAGE_SIZE)
812 			vm_page_init_page(m, pa, segind);
813 
814 		/*
815 		 * Add the segment to the free lists only if it is covered by
816 		 * one of the ranges in phys_avail.  Because we've added the
817 		 * ranges to the vm_phys_segs array, we can assume that each
818 		 * segment is either entirely contained in one of the ranges,
819 		 * or doesn't overlap any of them.
820 		 */
821 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
822 			struct vm_domain *vmd;
823 
824 			if (seg->start < phys_avail[i] ||
825 			    seg->end > phys_avail[i + 1])
826 				continue;
827 
828 			m = seg->first_page;
829 			pagecount = (u_long)atop(seg->end - seg->start);
830 
831 			vmd = VM_DOMAIN(seg->domain);
832 			vm_domain_free_lock(vmd);
833 			vm_phys_free_contig(m, pagecount);
834 			vm_domain_free_unlock(vmd);
835 			vm_domain_freecnt_inc(vmd, pagecount);
836 			vm_cnt.v_page_count += (u_int)pagecount;
837 
838 			vmd = VM_DOMAIN(seg->domain);
839 			vmd->vmd_page_count += (u_int)pagecount;
840 			vmd->vmd_segs |= 1UL << m->segind;
841 			break;
842 		}
843 	}
844 
845 	/*
846 	 * Remove blacklisted pages from the physical memory allocator.
847 	 */
848 	TAILQ_INIT(&blacklist_head);
849 	vm_page_blacklist_load(&list, &listend);
850 	vm_page_blacklist_check(list, listend);
851 
852 	list = kern_getenv("vm.blacklist");
853 	vm_page_blacklist_check(list, NULL);
854 
855 	freeenv(list);
856 #if VM_NRESERVLEVEL > 0
857 	/*
858 	 * Initialize the reservation management system.
859 	 */
860 	vm_reserv_init();
861 #endif
862 
863 	return (vaddr);
864 }
865 
866 void
867 vm_page_reference(vm_page_t m)
868 {
869 
870 	vm_page_aflag_set(m, PGA_REFERENCED);
871 }
872 
873 /*
874  *	vm_page_busy_downgrade:
875  *
876  *	Downgrade an exclusive busy page into a single shared busy page.
877  */
878 void
879 vm_page_busy_downgrade(vm_page_t m)
880 {
881 	u_int x;
882 	bool locked;
883 
884 	vm_page_assert_xbusied(m);
885 	locked = mtx_owned(vm_page_lockptr(m));
886 
887 	for (;;) {
888 		x = m->busy_lock;
889 		x &= VPB_BIT_WAITERS;
890 		if (x != 0 && !locked)
891 			vm_page_lock(m);
892 		if (atomic_cmpset_rel_int(&m->busy_lock,
893 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
894 			break;
895 		if (x != 0 && !locked)
896 			vm_page_unlock(m);
897 	}
898 	if (x != 0) {
899 		wakeup(m);
900 		if (!locked)
901 			vm_page_unlock(m);
902 	}
903 }
904 
905 /*
906  *	vm_page_sbusied:
907  *
908  *	Return a positive value if the page is shared busied, 0 otherwise.
909  */
910 int
911 vm_page_sbusied(vm_page_t m)
912 {
913 	u_int x;
914 
915 	x = m->busy_lock;
916 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
917 }
918 
919 /*
920  *	vm_page_sunbusy:
921  *
922  *	Shared unbusy a page.
923  */
924 void
925 vm_page_sunbusy(vm_page_t m)
926 {
927 	u_int x;
928 
929 	vm_page_lock_assert(m, MA_NOTOWNED);
930 	vm_page_assert_sbusied(m);
931 
932 	for (;;) {
933 		x = m->busy_lock;
934 		if (VPB_SHARERS(x) > 1) {
935 			if (atomic_cmpset_int(&m->busy_lock, x,
936 			    x - VPB_ONE_SHARER))
937 				break;
938 			continue;
939 		}
940 		if ((x & VPB_BIT_WAITERS) == 0) {
941 			KASSERT(x == VPB_SHARERS_WORD(1),
942 			    ("vm_page_sunbusy: invalid lock state"));
943 			if (atomic_cmpset_int(&m->busy_lock,
944 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
945 				break;
946 			continue;
947 		}
948 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
949 		    ("vm_page_sunbusy: invalid lock state for waiters"));
950 
951 		vm_page_lock(m);
952 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
953 			vm_page_unlock(m);
954 			continue;
955 		}
956 		wakeup(m);
957 		vm_page_unlock(m);
958 		break;
959 	}
960 }
961 
962 /*
963  *	vm_page_busy_sleep:
964  *
965  *	Sleep and release the page lock, using the page pointer as wchan.
966  *	This is used to implement the hard-path of busying mechanism.
967  *
968  *	The given page must be locked.
969  *
970  *	If nonshared is true, sleep only if the page is xbusy.
971  */
972 void
973 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
974 {
975 	u_int x;
976 
977 	vm_page_assert_locked(m);
978 
979 	x = m->busy_lock;
980 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
981 	    ((x & VPB_BIT_WAITERS) == 0 &&
982 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
983 		vm_page_unlock(m);
984 		return;
985 	}
986 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
987 }
988 
989 /*
990  *	vm_page_trysbusy:
991  *
992  *	Try to shared busy a page.
993  *	If the operation succeeds 1 is returned otherwise 0.
994  *	The operation never sleeps.
995  */
996 int
997 vm_page_trysbusy(vm_page_t m)
998 {
999 	u_int x;
1000 
1001 	for (;;) {
1002 		x = m->busy_lock;
1003 		if ((x & VPB_BIT_SHARED) == 0)
1004 			return (0);
1005 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
1006 			return (1);
1007 	}
1008 }
1009 
1010 static void
1011 vm_page_xunbusy_locked(vm_page_t m)
1012 {
1013 
1014 	vm_page_assert_xbusied(m);
1015 	vm_page_assert_locked(m);
1016 
1017 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1018 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
1019 	wakeup(m);
1020 }
1021 
1022 void
1023 vm_page_xunbusy_maybelocked(vm_page_t m)
1024 {
1025 	bool lockacq;
1026 
1027 	vm_page_assert_xbusied(m);
1028 
1029 	/*
1030 	 * Fast path for unbusy.  If it succeeds, we know that there
1031 	 * are no waiters, so we do not need a wakeup.
1032 	 */
1033 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
1034 	    VPB_UNBUSIED))
1035 		return;
1036 
1037 	lockacq = !mtx_owned(vm_page_lockptr(m));
1038 	if (lockacq)
1039 		vm_page_lock(m);
1040 	vm_page_xunbusy_locked(m);
1041 	if (lockacq)
1042 		vm_page_unlock(m);
1043 }
1044 
1045 /*
1046  *	vm_page_xunbusy_hard:
1047  *
1048  *	Called after the first try the exclusive unbusy of a page failed.
1049  *	It is assumed that the waiters bit is on.
1050  */
1051 void
1052 vm_page_xunbusy_hard(vm_page_t m)
1053 {
1054 
1055 	vm_page_assert_xbusied(m);
1056 
1057 	vm_page_lock(m);
1058 	vm_page_xunbusy_locked(m);
1059 	vm_page_unlock(m);
1060 }
1061 
1062 /*
1063  *	vm_page_flash:
1064  *
1065  *	Wakeup anyone waiting for the page.
1066  *	The ownership bits do not change.
1067  *
1068  *	The given page must be locked.
1069  */
1070 void
1071 vm_page_flash(vm_page_t m)
1072 {
1073 	u_int x;
1074 
1075 	vm_page_lock_assert(m, MA_OWNED);
1076 
1077 	for (;;) {
1078 		x = m->busy_lock;
1079 		if ((x & VPB_BIT_WAITERS) == 0)
1080 			return;
1081 		if (atomic_cmpset_int(&m->busy_lock, x,
1082 		    x & (~VPB_BIT_WAITERS)))
1083 			break;
1084 	}
1085 	wakeup(m);
1086 }
1087 
1088 /*
1089  * Avoid releasing and reacquiring the same page lock.
1090  */
1091 void
1092 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1093 {
1094 	struct mtx *mtx1;
1095 
1096 	mtx1 = vm_page_lockptr(m);
1097 	if (*mtx == mtx1)
1098 		return;
1099 	if (*mtx != NULL)
1100 		mtx_unlock(*mtx);
1101 	*mtx = mtx1;
1102 	mtx_lock(mtx1);
1103 }
1104 
1105 /*
1106  * Keep page from being freed by the page daemon
1107  * much of the same effect as wiring, except much lower
1108  * overhead and should be used only for *very* temporary
1109  * holding ("wiring").
1110  */
1111 void
1112 vm_page_hold(vm_page_t mem)
1113 {
1114 
1115 	vm_page_lock_assert(mem, MA_OWNED);
1116         mem->hold_count++;
1117 }
1118 
1119 void
1120 vm_page_unhold(vm_page_t mem)
1121 {
1122 
1123 	vm_page_lock_assert(mem, MA_OWNED);
1124 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1125 	--mem->hold_count;
1126 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1127 		vm_page_free_toq(mem);
1128 }
1129 
1130 /*
1131  *	vm_page_unhold_pages:
1132  *
1133  *	Unhold each of the pages that is referenced by the given array.
1134  */
1135 void
1136 vm_page_unhold_pages(vm_page_t *ma, int count)
1137 {
1138 	struct mtx *mtx;
1139 
1140 	mtx = NULL;
1141 	for (; count != 0; count--) {
1142 		vm_page_change_lock(*ma, &mtx);
1143 		vm_page_unhold(*ma);
1144 		ma++;
1145 	}
1146 	if (mtx != NULL)
1147 		mtx_unlock(mtx);
1148 }
1149 
1150 vm_page_t
1151 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1152 {
1153 	vm_page_t m;
1154 
1155 #ifdef VM_PHYSSEG_SPARSE
1156 	m = vm_phys_paddr_to_vm_page(pa);
1157 	if (m == NULL)
1158 		m = vm_phys_fictitious_to_vm_page(pa);
1159 	return (m);
1160 #elif defined(VM_PHYSSEG_DENSE)
1161 	long pi;
1162 
1163 	pi = atop(pa);
1164 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1165 		m = &vm_page_array[pi - first_page];
1166 		return (m);
1167 	}
1168 	return (vm_phys_fictitious_to_vm_page(pa));
1169 #else
1170 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1171 #endif
1172 }
1173 
1174 /*
1175  *	vm_page_getfake:
1176  *
1177  *	Create a fictitious page with the specified physical address and
1178  *	memory attribute.  The memory attribute is the only the machine-
1179  *	dependent aspect of a fictitious page that must be initialized.
1180  */
1181 vm_page_t
1182 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1183 {
1184 	vm_page_t m;
1185 
1186 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1187 	vm_page_initfake(m, paddr, memattr);
1188 	return (m);
1189 }
1190 
1191 void
1192 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1193 {
1194 
1195 	if ((m->flags & PG_FICTITIOUS) != 0) {
1196 		/*
1197 		 * The page's memattr might have changed since the
1198 		 * previous initialization.  Update the pmap to the
1199 		 * new memattr.
1200 		 */
1201 		goto memattr;
1202 	}
1203 	m->phys_addr = paddr;
1204 	m->queue = PQ_NONE;
1205 	/* Fictitious pages don't use "segind". */
1206 	m->flags = PG_FICTITIOUS;
1207 	/* Fictitious pages don't use "order" or "pool". */
1208 	m->oflags = VPO_UNMANAGED;
1209 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1210 	m->wire_count = 1;
1211 	pmap_page_init(m);
1212 memattr:
1213 	pmap_page_set_memattr(m, memattr);
1214 }
1215 
1216 /*
1217  *	vm_page_putfake:
1218  *
1219  *	Release a fictitious page.
1220  */
1221 void
1222 vm_page_putfake(vm_page_t m)
1223 {
1224 
1225 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1226 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1227 	    ("vm_page_putfake: bad page %p", m));
1228 	uma_zfree(fakepg_zone, m);
1229 }
1230 
1231 /*
1232  *	vm_page_updatefake:
1233  *
1234  *	Update the given fictitious page to the specified physical address and
1235  *	memory attribute.
1236  */
1237 void
1238 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1239 {
1240 
1241 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1242 	    ("vm_page_updatefake: bad page %p", m));
1243 	m->phys_addr = paddr;
1244 	pmap_page_set_memattr(m, memattr);
1245 }
1246 
1247 /*
1248  *	vm_page_free:
1249  *
1250  *	Free a page.
1251  */
1252 void
1253 vm_page_free(vm_page_t m)
1254 {
1255 
1256 	m->flags &= ~PG_ZERO;
1257 	vm_page_free_toq(m);
1258 }
1259 
1260 /*
1261  *	vm_page_free_zero:
1262  *
1263  *	Free a page to the zerod-pages queue
1264  */
1265 void
1266 vm_page_free_zero(vm_page_t m)
1267 {
1268 
1269 	m->flags |= PG_ZERO;
1270 	vm_page_free_toq(m);
1271 }
1272 
1273 /*
1274  * Unbusy and handle the page queueing for a page from a getpages request that
1275  * was optionally read ahead or behind.
1276  */
1277 void
1278 vm_page_readahead_finish(vm_page_t m)
1279 {
1280 
1281 	/* We shouldn't put invalid pages on queues. */
1282 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1283 
1284 	/*
1285 	 * Since the page is not the actually needed one, whether it should
1286 	 * be activated or deactivated is not obvious.  Empirical results
1287 	 * have shown that deactivating the page is usually the best choice,
1288 	 * unless the page is wanted by another thread.
1289 	 */
1290 	vm_page_lock(m);
1291 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1292 		vm_page_activate(m);
1293 	else
1294 		vm_page_deactivate(m);
1295 	vm_page_unlock(m);
1296 	vm_page_xunbusy(m);
1297 }
1298 
1299 /*
1300  *	vm_page_sleep_if_busy:
1301  *
1302  *	Sleep and release the page queues lock if the page is busied.
1303  *	Returns TRUE if the thread slept.
1304  *
1305  *	The given page must be unlocked and object containing it must
1306  *	be locked.
1307  */
1308 int
1309 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1310 {
1311 	vm_object_t obj;
1312 
1313 	vm_page_lock_assert(m, MA_NOTOWNED);
1314 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1315 
1316 	if (vm_page_busied(m)) {
1317 		/*
1318 		 * The page-specific object must be cached because page
1319 		 * identity can change during the sleep, causing the
1320 		 * re-lock of a different object.
1321 		 * It is assumed that a reference to the object is already
1322 		 * held by the callers.
1323 		 */
1324 		obj = m->object;
1325 		vm_page_lock(m);
1326 		VM_OBJECT_WUNLOCK(obj);
1327 		vm_page_busy_sleep(m, msg, false);
1328 		VM_OBJECT_WLOCK(obj);
1329 		return (TRUE);
1330 	}
1331 	return (FALSE);
1332 }
1333 
1334 /*
1335  *	vm_page_dirty_KBI:		[ internal use only ]
1336  *
1337  *	Set all bits in the page's dirty field.
1338  *
1339  *	The object containing the specified page must be locked if the
1340  *	call is made from the machine-independent layer.
1341  *
1342  *	See vm_page_clear_dirty_mask().
1343  *
1344  *	This function should only be called by vm_page_dirty().
1345  */
1346 void
1347 vm_page_dirty_KBI(vm_page_t m)
1348 {
1349 
1350 	/* Refer to this operation by its public name. */
1351 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1352 	    ("vm_page_dirty: page is invalid!"));
1353 	m->dirty = VM_PAGE_BITS_ALL;
1354 }
1355 
1356 /*
1357  *	vm_page_insert:		[ internal use only ]
1358  *
1359  *	Inserts the given mem entry into the object and object list.
1360  *
1361  *	The object must be locked.
1362  */
1363 int
1364 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1365 {
1366 	vm_page_t mpred;
1367 
1368 	VM_OBJECT_ASSERT_WLOCKED(object);
1369 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1370 	return (vm_page_insert_after(m, object, pindex, mpred));
1371 }
1372 
1373 /*
1374  *	vm_page_insert_after:
1375  *
1376  *	Inserts the page "m" into the specified object at offset "pindex".
1377  *
1378  *	The page "mpred" must immediately precede the offset "pindex" within
1379  *	the specified object.
1380  *
1381  *	The object must be locked.
1382  */
1383 static int
1384 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1385     vm_page_t mpred)
1386 {
1387 	vm_page_t msucc;
1388 
1389 	VM_OBJECT_ASSERT_WLOCKED(object);
1390 	KASSERT(m->object == NULL,
1391 	    ("vm_page_insert_after: page already inserted"));
1392 	if (mpred != NULL) {
1393 		KASSERT(mpred->object == object,
1394 		    ("vm_page_insert_after: object doesn't contain mpred"));
1395 		KASSERT(mpred->pindex < pindex,
1396 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1397 		msucc = TAILQ_NEXT(mpred, listq);
1398 	} else
1399 		msucc = TAILQ_FIRST(&object->memq);
1400 	if (msucc != NULL)
1401 		KASSERT(msucc->pindex > pindex,
1402 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1403 
1404 	/*
1405 	 * Record the object/offset pair in this page
1406 	 */
1407 	m->object = object;
1408 	m->pindex = pindex;
1409 
1410 	/*
1411 	 * Now link into the object's ordered list of backed pages.
1412 	 */
1413 	if (vm_radix_insert(&object->rtree, m)) {
1414 		m->object = NULL;
1415 		m->pindex = 0;
1416 		return (1);
1417 	}
1418 	vm_page_insert_radixdone(m, object, mpred);
1419 	return (0);
1420 }
1421 
1422 /*
1423  *	vm_page_insert_radixdone:
1424  *
1425  *	Complete page "m" insertion into the specified object after the
1426  *	radix trie hooking.
1427  *
1428  *	The page "mpred" must precede the offset "m->pindex" within the
1429  *	specified object.
1430  *
1431  *	The object must be locked.
1432  */
1433 static void
1434 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1435 {
1436 
1437 	VM_OBJECT_ASSERT_WLOCKED(object);
1438 	KASSERT(object != NULL && m->object == object,
1439 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1440 	if (mpred != NULL) {
1441 		KASSERT(mpred->object == object,
1442 		    ("vm_page_insert_after: object doesn't contain mpred"));
1443 		KASSERT(mpred->pindex < m->pindex,
1444 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1445 	}
1446 
1447 	if (mpred != NULL)
1448 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1449 	else
1450 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1451 
1452 	/*
1453 	 * Show that the object has one more resident page.
1454 	 */
1455 	object->resident_page_count++;
1456 
1457 	/*
1458 	 * Hold the vnode until the last page is released.
1459 	 */
1460 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1461 		vhold(object->handle);
1462 
1463 	/*
1464 	 * Since we are inserting a new and possibly dirty page,
1465 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1466 	 */
1467 	if (pmap_page_is_write_mapped(m))
1468 		vm_object_set_writeable_dirty(object);
1469 }
1470 
1471 /*
1472  *	vm_page_remove:
1473  *
1474  *	Removes the specified page from its containing object, but does not
1475  *	invalidate any backing storage.
1476  *
1477  *	The object must be locked.  The page must be locked if it is managed.
1478  */
1479 void
1480 vm_page_remove(vm_page_t m)
1481 {
1482 	vm_object_t object;
1483 	vm_page_t mrem;
1484 
1485 	if ((m->oflags & VPO_UNMANAGED) == 0)
1486 		vm_page_assert_locked(m);
1487 	if ((object = m->object) == NULL)
1488 		return;
1489 	VM_OBJECT_ASSERT_WLOCKED(object);
1490 	if (vm_page_xbusied(m))
1491 		vm_page_xunbusy_maybelocked(m);
1492 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1493 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1494 
1495 	/*
1496 	 * Now remove from the object's list of backed pages.
1497 	 */
1498 	TAILQ_REMOVE(&object->memq, m, listq);
1499 
1500 	/*
1501 	 * And show that the object has one fewer resident page.
1502 	 */
1503 	object->resident_page_count--;
1504 
1505 	/*
1506 	 * The vnode may now be recycled.
1507 	 */
1508 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1509 		vdrop(object->handle);
1510 
1511 	m->object = NULL;
1512 }
1513 
1514 /*
1515  *	vm_page_lookup:
1516  *
1517  *	Returns the page associated with the object/offset
1518  *	pair specified; if none is found, NULL is returned.
1519  *
1520  *	The object must be locked.
1521  */
1522 vm_page_t
1523 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1524 {
1525 
1526 	VM_OBJECT_ASSERT_LOCKED(object);
1527 	return (vm_radix_lookup(&object->rtree, pindex));
1528 }
1529 
1530 /*
1531  *	vm_page_find_least:
1532  *
1533  *	Returns the page associated with the object with least pindex
1534  *	greater than or equal to the parameter pindex, or NULL.
1535  *
1536  *	The object must be locked.
1537  */
1538 vm_page_t
1539 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1540 {
1541 	vm_page_t m;
1542 
1543 	VM_OBJECT_ASSERT_LOCKED(object);
1544 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1545 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1546 	return (m);
1547 }
1548 
1549 /*
1550  * Returns the given page's successor (by pindex) within the object if it is
1551  * resident; if none is found, NULL is returned.
1552  *
1553  * The object must be locked.
1554  */
1555 vm_page_t
1556 vm_page_next(vm_page_t m)
1557 {
1558 	vm_page_t next;
1559 
1560 	VM_OBJECT_ASSERT_LOCKED(m->object);
1561 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1562 		MPASS(next->object == m->object);
1563 		if (next->pindex != m->pindex + 1)
1564 			next = NULL;
1565 	}
1566 	return (next);
1567 }
1568 
1569 /*
1570  * Returns the given page's predecessor (by pindex) within the object if it is
1571  * resident; if none is found, NULL is returned.
1572  *
1573  * The object must be locked.
1574  */
1575 vm_page_t
1576 vm_page_prev(vm_page_t m)
1577 {
1578 	vm_page_t prev;
1579 
1580 	VM_OBJECT_ASSERT_LOCKED(m->object);
1581 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1582 		MPASS(prev->object == m->object);
1583 		if (prev->pindex != m->pindex - 1)
1584 			prev = NULL;
1585 	}
1586 	return (prev);
1587 }
1588 
1589 /*
1590  * Uses the page mnew as a replacement for an existing page at index
1591  * pindex which must be already present in the object.
1592  *
1593  * The existing page must not be on a paging queue.
1594  */
1595 vm_page_t
1596 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1597 {
1598 	vm_page_t mold;
1599 
1600 	VM_OBJECT_ASSERT_WLOCKED(object);
1601 	KASSERT(mnew->object == NULL,
1602 	    ("vm_page_replace: page %p already in object", mnew));
1603 	KASSERT(mnew->queue == PQ_NONE,
1604 	    ("vm_page_replace: new page %p is on a paging queue", mnew));
1605 
1606 	/*
1607 	 * This function mostly follows vm_page_insert() and
1608 	 * vm_page_remove() without the radix, object count and vnode
1609 	 * dance.  Double check such functions for more comments.
1610 	 */
1611 
1612 	mnew->object = object;
1613 	mnew->pindex = pindex;
1614 	mold = vm_radix_replace(&object->rtree, mnew);
1615 	KASSERT(mold->queue == PQ_NONE,
1616 	    ("vm_page_replace: old page %p is on a paging queue", mold));
1617 
1618 	/* Keep the resident page list in sorted order. */
1619 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1620 	TAILQ_REMOVE(&object->memq, mold, listq);
1621 
1622 	mold->object = NULL;
1623 	vm_page_xunbusy_maybelocked(mold);
1624 
1625 	/*
1626 	 * The object's resident_page_count does not change because we have
1627 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1628 	 */
1629 	if (pmap_page_is_write_mapped(mnew))
1630 		vm_object_set_writeable_dirty(object);
1631 	return (mold);
1632 }
1633 
1634 /*
1635  *	vm_page_rename:
1636  *
1637  *	Move the given memory entry from its
1638  *	current object to the specified target object/offset.
1639  *
1640  *	Note: swap associated with the page must be invalidated by the move.  We
1641  *	      have to do this for several reasons:  (1) we aren't freeing the
1642  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1643  *	      moving the page from object A to B, and will then later move
1644  *	      the backing store from A to B and we can't have a conflict.
1645  *
1646  *	Note: we *always* dirty the page.  It is necessary both for the
1647  *	      fact that we moved it, and because we may be invalidating
1648  *	      swap.
1649  *
1650  *	The objects must be locked.
1651  */
1652 int
1653 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1654 {
1655 	vm_page_t mpred;
1656 	vm_pindex_t opidx;
1657 
1658 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1659 
1660 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1661 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1662 	    ("vm_page_rename: pindex already renamed"));
1663 
1664 	/*
1665 	 * Create a custom version of vm_page_insert() which does not depend
1666 	 * by m_prev and can cheat on the implementation aspects of the
1667 	 * function.
1668 	 */
1669 	opidx = m->pindex;
1670 	m->pindex = new_pindex;
1671 	if (vm_radix_insert(&new_object->rtree, m)) {
1672 		m->pindex = opidx;
1673 		return (1);
1674 	}
1675 
1676 	/*
1677 	 * The operation cannot fail anymore.  The removal must happen before
1678 	 * the listq iterator is tainted.
1679 	 */
1680 	m->pindex = opidx;
1681 	vm_page_lock(m);
1682 	vm_page_remove(m);
1683 
1684 	/* Return back to the new pindex to complete vm_page_insert(). */
1685 	m->pindex = new_pindex;
1686 	m->object = new_object;
1687 	vm_page_unlock(m);
1688 	vm_page_insert_radixdone(m, new_object, mpred);
1689 	vm_page_dirty(m);
1690 	return (0);
1691 }
1692 
1693 /*
1694  *	vm_page_alloc:
1695  *
1696  *	Allocate and return a page that is associated with the specified
1697  *	object and offset pair.  By default, this page is exclusive busied.
1698  *
1699  *	The caller must always specify an allocation class.
1700  *
1701  *	allocation classes:
1702  *	VM_ALLOC_NORMAL		normal process request
1703  *	VM_ALLOC_SYSTEM		system *really* needs a page
1704  *	VM_ALLOC_INTERRUPT	interrupt time request
1705  *
1706  *	optional allocation flags:
1707  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1708  *				intends to allocate
1709  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1710  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1711  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1712  *				should not be exclusive busy
1713  *	VM_ALLOC_SBUSY		shared busy the allocated page
1714  *	VM_ALLOC_WIRED		wire the allocated page
1715  *	VM_ALLOC_ZERO		prefer a zeroed page
1716  */
1717 vm_page_t
1718 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1719 {
1720 
1721 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1722 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1723 }
1724 
1725 vm_page_t
1726 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1727     int req)
1728 {
1729 
1730 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1731 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1732 	    NULL));
1733 }
1734 
1735 /*
1736  * Allocate a page in the specified object with the given page index.  To
1737  * optimize insertion of the page into the object, the caller must also specifiy
1738  * the resident page in the object with largest index smaller than the given
1739  * page index, or NULL if no such page exists.
1740  */
1741 vm_page_t
1742 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1743     int req, vm_page_t mpred)
1744 {
1745 	struct vm_domainset_iter di;
1746 	vm_page_t m;
1747 	int domain;
1748 
1749 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1750 	do {
1751 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1752 		    mpred);
1753 		if (m != NULL)
1754 			break;
1755 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1756 
1757 	return (m);
1758 }
1759 
1760 /*
1761  * Returns true if the number of free pages exceeds the minimum
1762  * for the request class and false otherwise.
1763  */
1764 int
1765 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1766 {
1767 	u_int limit, old, new;
1768 
1769 	req = req & VM_ALLOC_CLASS_MASK;
1770 
1771 	/*
1772 	 * The page daemon is allowed to dig deeper into the free page list.
1773 	 */
1774 	if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1775 		req = VM_ALLOC_SYSTEM;
1776 	if (req == VM_ALLOC_INTERRUPT)
1777 		limit = 0;
1778 	else if (req == VM_ALLOC_SYSTEM)
1779 		limit = vmd->vmd_interrupt_free_min;
1780 	else
1781 		limit = vmd->vmd_free_reserved;
1782 
1783 	/*
1784 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1785 	 */
1786 	limit += npages;
1787 	old = vmd->vmd_free_count;
1788 	do {
1789 		if (old < limit)
1790 			return (0);
1791 		new = old - npages;
1792 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1793 
1794 	/* Wake the page daemon if we've crossed the threshold. */
1795 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1796 		pagedaemon_wakeup(vmd->vmd_domain);
1797 
1798 	/* Only update bitsets on transitions. */
1799 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1800 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1801 		vm_domain_set(vmd);
1802 
1803 	return (1);
1804 }
1805 
1806 vm_page_t
1807 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1808     int req, vm_page_t mpred)
1809 {
1810 	struct vm_domain *vmd;
1811 	vm_page_t m;
1812 	int flags;
1813 
1814 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1815 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1816 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1817 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1818 	    ("inconsistent object(%p)/req(%x)", object, req));
1819 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1820 	    ("Can't sleep and retry object insertion."));
1821 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1822 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1823 	    (uintmax_t)pindex));
1824 	if (object != NULL)
1825 		VM_OBJECT_ASSERT_WLOCKED(object);
1826 
1827 again:
1828 	m = NULL;
1829 #if VM_NRESERVLEVEL > 0
1830 	/*
1831 	 * Can we allocate the page from a reservation?
1832 	 */
1833 	if (vm_object_reserv(object) &&
1834 	    ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL ||
1835 	    (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) {
1836 		domain = vm_phys_domain(m);
1837 		vmd = VM_DOMAIN(domain);
1838 		goto found;
1839 	}
1840 #endif
1841 	vmd = VM_DOMAIN(domain);
1842 	if (object != NULL && vmd->vmd_pgcache != NULL) {
1843 		m = uma_zalloc(vmd->vmd_pgcache, M_NOWAIT);
1844 		if (m != NULL)
1845 			goto found;
1846 	}
1847 	if (vm_domain_allocate(vmd, req, 1)) {
1848 		/*
1849 		 * If not, allocate it from the free page queues.
1850 		 */
1851 		vm_domain_free_lock(vmd);
1852 		m = vm_phys_alloc_pages(domain, object != NULL ?
1853 		    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1854 		vm_domain_free_unlock(vmd);
1855 		if (m == NULL) {
1856 			vm_domain_freecnt_inc(vmd, 1);
1857 #if VM_NRESERVLEVEL > 0
1858 			if (vm_reserv_reclaim_inactive(domain))
1859 				goto again;
1860 #endif
1861 		}
1862 	}
1863 	if (m == NULL) {
1864 		/*
1865 		 * Not allocatable, give up.
1866 		 */
1867 		if (vm_domain_alloc_fail(vmd, object, req))
1868 			goto again;
1869 		return (NULL);
1870 	}
1871 
1872 	/*
1873 	 *  At this point we had better have found a good page.
1874 	 */
1875 	KASSERT(m != NULL, ("missing page"));
1876 
1877 found:
1878 	vm_page_dequeue(m);
1879 	vm_page_alloc_check(m);
1880 
1881 	/*
1882 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1883 	 */
1884 	flags = 0;
1885 	if ((req & VM_ALLOC_ZERO) != 0)
1886 		flags = PG_ZERO;
1887 	flags &= m->flags;
1888 	if ((req & VM_ALLOC_NODUMP) != 0)
1889 		flags |= PG_NODUMP;
1890 	m->flags = flags;
1891 	m->aflags = 0;
1892 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1893 	    VPO_UNMANAGED : 0;
1894 	m->busy_lock = VPB_UNBUSIED;
1895 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1896 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1897 	if ((req & VM_ALLOC_SBUSY) != 0)
1898 		m->busy_lock = VPB_SHARERS_WORD(1);
1899 	if (req & VM_ALLOC_WIRED) {
1900 		/*
1901 		 * The page lock is not required for wiring a page until that
1902 		 * page is inserted into the object.
1903 		 */
1904 		vm_wire_add(1);
1905 		m->wire_count = 1;
1906 	}
1907 	m->act_count = 0;
1908 
1909 	if (object != NULL) {
1910 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1911 			if (req & VM_ALLOC_WIRED) {
1912 				vm_wire_sub(1);
1913 				m->wire_count = 0;
1914 			}
1915 			KASSERT(m->object == NULL, ("page %p has object", m));
1916 			m->oflags = VPO_UNMANAGED;
1917 			m->busy_lock = VPB_UNBUSIED;
1918 			/* Don't change PG_ZERO. */
1919 			vm_page_free_toq(m);
1920 			if (req & VM_ALLOC_WAITFAIL) {
1921 				VM_OBJECT_WUNLOCK(object);
1922 				vm_radix_wait();
1923 				VM_OBJECT_WLOCK(object);
1924 			}
1925 			return (NULL);
1926 		}
1927 
1928 		/* Ignore device objects; the pager sets "memattr" for them. */
1929 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1930 		    (object->flags & OBJ_FICTITIOUS) == 0)
1931 			pmap_page_set_memattr(m, object->memattr);
1932 	} else
1933 		m->pindex = pindex;
1934 
1935 	return (m);
1936 }
1937 
1938 /*
1939  *	vm_page_alloc_contig:
1940  *
1941  *	Allocate a contiguous set of physical pages of the given size "npages"
1942  *	from the free lists.  All of the physical pages must be at or above
1943  *	the given physical address "low" and below the given physical address
1944  *	"high".  The given value "alignment" determines the alignment of the
1945  *	first physical page in the set.  If the given value "boundary" is
1946  *	non-zero, then the set of physical pages cannot cross any physical
1947  *	address boundary that is a multiple of that value.  Both "alignment"
1948  *	and "boundary" must be a power of two.
1949  *
1950  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1951  *	then the memory attribute setting for the physical pages is configured
1952  *	to the object's memory attribute setting.  Otherwise, the memory
1953  *	attribute setting for the physical pages is configured to "memattr",
1954  *	overriding the object's memory attribute setting.  However, if the
1955  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1956  *	memory attribute setting for the physical pages cannot be configured
1957  *	to VM_MEMATTR_DEFAULT.
1958  *
1959  *	The specified object may not contain fictitious pages.
1960  *
1961  *	The caller must always specify an allocation class.
1962  *
1963  *	allocation classes:
1964  *	VM_ALLOC_NORMAL		normal process request
1965  *	VM_ALLOC_SYSTEM		system *really* needs a page
1966  *	VM_ALLOC_INTERRUPT	interrupt time request
1967  *
1968  *	optional allocation flags:
1969  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1970  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1971  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1972  *				should not be exclusive busy
1973  *	VM_ALLOC_SBUSY		shared busy the allocated page
1974  *	VM_ALLOC_WIRED		wire the allocated page
1975  *	VM_ALLOC_ZERO		prefer a zeroed page
1976  */
1977 vm_page_t
1978 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1979     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1980     vm_paddr_t boundary, vm_memattr_t memattr)
1981 {
1982 	struct vm_domainset_iter di;
1983 	vm_page_t m;
1984 	int domain;
1985 
1986 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1987 	do {
1988 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1989 		    npages, low, high, alignment, boundary, memattr);
1990 		if (m != NULL)
1991 			break;
1992 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1993 
1994 	return (m);
1995 }
1996 
1997 vm_page_t
1998 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1999     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2000     vm_paddr_t boundary, vm_memattr_t memattr)
2001 {
2002 	struct vm_domain *vmd;
2003 	vm_page_t m, m_ret, mpred;
2004 	u_int busy_lock, flags, oflags;
2005 
2006 	mpred = NULL;	/* XXX: pacify gcc */
2007 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2008 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2009 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2010 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2011 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2012 	    req));
2013 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2014 	    ("Can't sleep and retry object insertion."));
2015 	if (object != NULL) {
2016 		VM_OBJECT_ASSERT_WLOCKED(object);
2017 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2018 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2019 		    object));
2020 	}
2021 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2022 
2023 	if (object != NULL) {
2024 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2025 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2026 		    ("vm_page_alloc_contig: pindex already allocated"));
2027 	}
2028 
2029 	/*
2030 	 * Can we allocate the pages without the number of free pages falling
2031 	 * below the lower bound for the allocation class?
2032 	 */
2033 again:
2034 #if VM_NRESERVLEVEL > 0
2035 	/*
2036 	 * Can we allocate the pages from a reservation?
2037 	 */
2038 	if (vm_object_reserv(object) &&
2039 	    ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
2040 	    npages, low, high, alignment, boundary, mpred)) != NULL ||
2041 	    (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain,
2042 	    npages, low, high, alignment, boundary, mpred)) != NULL)) {
2043 		domain = vm_phys_domain(m_ret);
2044 		vmd = VM_DOMAIN(domain);
2045 		goto found;
2046 	}
2047 #endif
2048 	m_ret = NULL;
2049 	vmd = VM_DOMAIN(domain);
2050 	if (vm_domain_allocate(vmd, req, npages)) {
2051 		/*
2052 		 * allocate them from the free page queues.
2053 		 */
2054 		vm_domain_free_lock(vmd);
2055 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2056 		    alignment, boundary);
2057 		vm_domain_free_unlock(vmd);
2058 		if (m_ret == NULL) {
2059 			vm_domain_freecnt_inc(vmd, npages);
2060 #if VM_NRESERVLEVEL > 0
2061 			if (vm_reserv_reclaim_contig(domain, npages, low,
2062 			    high, alignment, boundary))
2063 				goto again;
2064 #endif
2065 		}
2066 	}
2067 	if (m_ret == NULL) {
2068 		if (vm_domain_alloc_fail(vmd, object, req))
2069 			goto again;
2070 		return (NULL);
2071 	}
2072 #if VM_NRESERVLEVEL > 0
2073 found:
2074 #endif
2075 	for (m = m_ret; m < &m_ret[npages]; m++) {
2076 		vm_page_dequeue(m);
2077 		vm_page_alloc_check(m);
2078 	}
2079 
2080 	/*
2081 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2082 	 */
2083 	flags = 0;
2084 	if ((req & VM_ALLOC_ZERO) != 0)
2085 		flags = PG_ZERO;
2086 	if ((req & VM_ALLOC_NODUMP) != 0)
2087 		flags |= PG_NODUMP;
2088 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2089 	    VPO_UNMANAGED : 0;
2090 	busy_lock = VPB_UNBUSIED;
2091 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2092 		busy_lock = VPB_SINGLE_EXCLUSIVER;
2093 	if ((req & VM_ALLOC_SBUSY) != 0)
2094 		busy_lock = VPB_SHARERS_WORD(1);
2095 	if ((req & VM_ALLOC_WIRED) != 0)
2096 		vm_wire_add(npages);
2097 	if (object != NULL) {
2098 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2099 		    memattr == VM_MEMATTR_DEFAULT)
2100 			memattr = object->memattr;
2101 	}
2102 	for (m = m_ret; m < &m_ret[npages]; m++) {
2103 		m->aflags = 0;
2104 		m->flags = (m->flags | PG_NODUMP) & flags;
2105 		m->busy_lock = busy_lock;
2106 		if ((req & VM_ALLOC_WIRED) != 0)
2107 			m->wire_count = 1;
2108 		m->act_count = 0;
2109 		m->oflags = oflags;
2110 		if (object != NULL) {
2111 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2112 				if ((req & VM_ALLOC_WIRED) != 0)
2113 					vm_wire_sub(npages);
2114 				KASSERT(m->object == NULL,
2115 				    ("page %p has object", m));
2116 				mpred = m;
2117 				for (m = m_ret; m < &m_ret[npages]; m++) {
2118 					if (m <= mpred &&
2119 					    (req & VM_ALLOC_WIRED) != 0)
2120 						m->wire_count = 0;
2121 					m->oflags = VPO_UNMANAGED;
2122 					m->busy_lock = VPB_UNBUSIED;
2123 					/* Don't change PG_ZERO. */
2124 					vm_page_free_toq(m);
2125 				}
2126 				if (req & VM_ALLOC_WAITFAIL) {
2127 					VM_OBJECT_WUNLOCK(object);
2128 					vm_radix_wait();
2129 					VM_OBJECT_WLOCK(object);
2130 				}
2131 				return (NULL);
2132 			}
2133 			mpred = m;
2134 		} else
2135 			m->pindex = pindex;
2136 		if (memattr != VM_MEMATTR_DEFAULT)
2137 			pmap_page_set_memattr(m, memattr);
2138 		pindex++;
2139 	}
2140 	return (m_ret);
2141 }
2142 
2143 /*
2144  * Check a page that has been freshly dequeued from a freelist.
2145  */
2146 static void
2147 vm_page_alloc_check(vm_page_t m)
2148 {
2149 
2150 	KASSERT(m->object == NULL, ("page %p has object", m));
2151 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2152 	    ("page %p has unexpected queue %d, flags %#x",
2153 	    m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2154 	KASSERT(!vm_page_held(m), ("page %p is held", m));
2155 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2156 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2157 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2158 	    ("page %p has unexpected memattr %d",
2159 	    m, pmap_page_get_memattr(m)));
2160 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2161 }
2162 
2163 /*
2164  * 	vm_page_alloc_freelist:
2165  *
2166  *	Allocate a physical page from the specified free page list.
2167  *
2168  *	The caller must always specify an allocation class.
2169  *
2170  *	allocation classes:
2171  *	VM_ALLOC_NORMAL		normal process request
2172  *	VM_ALLOC_SYSTEM		system *really* needs a page
2173  *	VM_ALLOC_INTERRUPT	interrupt time request
2174  *
2175  *	optional allocation flags:
2176  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2177  *				intends to allocate
2178  *	VM_ALLOC_WIRED		wire the allocated page
2179  *	VM_ALLOC_ZERO		prefer a zeroed page
2180  */
2181 vm_page_t
2182 vm_page_alloc_freelist(int freelist, int req)
2183 {
2184 	struct vm_domainset_iter di;
2185 	vm_page_t m;
2186 	int domain;
2187 
2188 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2189 	do {
2190 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2191 		if (m != NULL)
2192 			break;
2193 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2194 
2195 	return (m);
2196 }
2197 
2198 vm_page_t
2199 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2200 {
2201 	struct vm_domain *vmd;
2202 	vm_page_t m;
2203 	u_int flags;
2204 
2205 	m = NULL;
2206 	vmd = VM_DOMAIN(domain);
2207 again:
2208 	if (vm_domain_allocate(vmd, req, 1)) {
2209 		vm_domain_free_lock(vmd);
2210 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2211 		    VM_FREEPOOL_DIRECT, 0);
2212 		vm_domain_free_unlock(vmd);
2213 		if (m == NULL)
2214 			vm_domain_freecnt_inc(vmd, 1);
2215 	}
2216 	if (m == NULL) {
2217 		if (vm_domain_alloc_fail(vmd, NULL, req))
2218 			goto again;
2219 		return (NULL);
2220 	}
2221 	vm_page_dequeue(m);
2222 	vm_page_alloc_check(m);
2223 
2224 	/*
2225 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2226 	 */
2227 	m->aflags = 0;
2228 	flags = 0;
2229 	if ((req & VM_ALLOC_ZERO) != 0)
2230 		flags = PG_ZERO;
2231 	m->flags &= flags;
2232 	if ((req & VM_ALLOC_WIRED) != 0) {
2233 		/*
2234 		 * The page lock is not required for wiring a page that does
2235 		 * not belong to an object.
2236 		 */
2237 		vm_wire_add(1);
2238 		m->wire_count = 1;
2239 	}
2240 	/* Unmanaged pages don't use "act_count". */
2241 	m->oflags = VPO_UNMANAGED;
2242 	return (m);
2243 }
2244 
2245 static int
2246 vm_page_import(void *arg, void **store, int cnt, int domain, int flags)
2247 {
2248 	struct vm_domain *vmd;
2249 	int i;
2250 
2251 	vmd = arg;
2252 	/* Only import if we can bring in a full bucket. */
2253 	if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2254 		return (0);
2255 	domain = vmd->vmd_domain;
2256 	vm_domain_free_lock(vmd);
2257 	i = vm_phys_alloc_npages(domain, VM_FREEPOOL_DEFAULT, cnt,
2258 	    (vm_page_t *)store);
2259 	vm_domain_free_unlock(vmd);
2260 	if (cnt != i)
2261 		vm_domain_freecnt_inc(vmd, cnt - i);
2262 
2263 	return (i);
2264 }
2265 
2266 static void
2267 vm_page_release(void *arg, void **store, int cnt)
2268 {
2269 	struct vm_domain *vmd;
2270 	vm_page_t m;
2271 	int i;
2272 
2273 	vmd = arg;
2274 	vm_domain_free_lock(vmd);
2275 	for (i = 0; i < cnt; i++) {
2276 		m = (vm_page_t)store[i];
2277 		vm_phys_free_pages(m, 0);
2278 	}
2279 	vm_domain_free_unlock(vmd);
2280 	vm_domain_freecnt_inc(vmd, cnt);
2281 }
2282 
2283 #define	VPSC_ANY	0	/* No restrictions. */
2284 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2285 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2286 
2287 /*
2288  *	vm_page_scan_contig:
2289  *
2290  *	Scan vm_page_array[] between the specified entries "m_start" and
2291  *	"m_end" for a run of contiguous physical pages that satisfy the
2292  *	specified conditions, and return the lowest page in the run.  The
2293  *	specified "alignment" determines the alignment of the lowest physical
2294  *	page in the run.  If the specified "boundary" is non-zero, then the
2295  *	run of physical pages cannot span a physical address that is a
2296  *	multiple of "boundary".
2297  *
2298  *	"m_end" is never dereferenced, so it need not point to a vm_page
2299  *	structure within vm_page_array[].
2300  *
2301  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2302  *	span a hole (or discontiguity) in the physical address space.  Both
2303  *	"alignment" and "boundary" must be a power of two.
2304  */
2305 vm_page_t
2306 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2307     u_long alignment, vm_paddr_t boundary, int options)
2308 {
2309 	struct mtx *m_mtx;
2310 	vm_object_t object;
2311 	vm_paddr_t pa;
2312 	vm_page_t m, m_run;
2313 #if VM_NRESERVLEVEL > 0
2314 	int level;
2315 #endif
2316 	int m_inc, order, run_ext, run_len;
2317 
2318 	KASSERT(npages > 0, ("npages is 0"));
2319 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2320 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2321 	m_run = NULL;
2322 	run_len = 0;
2323 	m_mtx = NULL;
2324 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2325 		KASSERT((m->flags & PG_MARKER) == 0,
2326 		    ("page %p is PG_MARKER", m));
2327 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2328 		    ("fictitious page %p has invalid wire count", m));
2329 
2330 		/*
2331 		 * If the current page would be the start of a run, check its
2332 		 * physical address against the end, alignment, and boundary
2333 		 * conditions.  If it doesn't satisfy these conditions, either
2334 		 * terminate the scan or advance to the next page that
2335 		 * satisfies the failed condition.
2336 		 */
2337 		if (run_len == 0) {
2338 			KASSERT(m_run == NULL, ("m_run != NULL"));
2339 			if (m + npages > m_end)
2340 				break;
2341 			pa = VM_PAGE_TO_PHYS(m);
2342 			if ((pa & (alignment - 1)) != 0) {
2343 				m_inc = atop(roundup2(pa, alignment) - pa);
2344 				continue;
2345 			}
2346 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2347 			    boundary) != 0) {
2348 				m_inc = atop(roundup2(pa, boundary) - pa);
2349 				continue;
2350 			}
2351 		} else
2352 			KASSERT(m_run != NULL, ("m_run == NULL"));
2353 
2354 		vm_page_change_lock(m, &m_mtx);
2355 		m_inc = 1;
2356 retry:
2357 		if (vm_page_held(m))
2358 			run_ext = 0;
2359 #if VM_NRESERVLEVEL > 0
2360 		else if ((level = vm_reserv_level(m)) >= 0 &&
2361 		    (options & VPSC_NORESERV) != 0) {
2362 			run_ext = 0;
2363 			/* Advance to the end of the reservation. */
2364 			pa = VM_PAGE_TO_PHYS(m);
2365 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2366 			    pa);
2367 		}
2368 #endif
2369 		else if ((object = m->object) != NULL) {
2370 			/*
2371 			 * The page is considered eligible for relocation if
2372 			 * and only if it could be laundered or reclaimed by
2373 			 * the page daemon.
2374 			 */
2375 			if (!VM_OBJECT_TRYRLOCK(object)) {
2376 				mtx_unlock(m_mtx);
2377 				VM_OBJECT_RLOCK(object);
2378 				mtx_lock(m_mtx);
2379 				if (m->object != object) {
2380 					/*
2381 					 * The page may have been freed.
2382 					 */
2383 					VM_OBJECT_RUNLOCK(object);
2384 					goto retry;
2385 				} else if (vm_page_held(m)) {
2386 					run_ext = 0;
2387 					goto unlock;
2388 				}
2389 			}
2390 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2391 			    ("page %p is PG_UNHOLDFREE", m));
2392 			/* Don't care: PG_NODUMP, PG_ZERO. */
2393 			if (object->type != OBJT_DEFAULT &&
2394 			    object->type != OBJT_SWAP &&
2395 			    object->type != OBJT_VNODE) {
2396 				run_ext = 0;
2397 #if VM_NRESERVLEVEL > 0
2398 			} else if ((options & VPSC_NOSUPER) != 0 &&
2399 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2400 				run_ext = 0;
2401 				/* Advance to the end of the superpage. */
2402 				pa = VM_PAGE_TO_PHYS(m);
2403 				m_inc = atop(roundup2(pa + 1,
2404 				    vm_reserv_size(level)) - pa);
2405 #endif
2406 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2407 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2408 				/*
2409 				 * The page is allocated but eligible for
2410 				 * relocation.  Extend the current run by one
2411 				 * page.
2412 				 */
2413 				KASSERT(pmap_page_get_memattr(m) ==
2414 				    VM_MEMATTR_DEFAULT,
2415 				    ("page %p has an unexpected memattr", m));
2416 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2417 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2418 				    ("page %p has unexpected oflags", m));
2419 				/* Don't care: VPO_NOSYNC. */
2420 				run_ext = 1;
2421 			} else
2422 				run_ext = 0;
2423 unlock:
2424 			VM_OBJECT_RUNLOCK(object);
2425 #if VM_NRESERVLEVEL > 0
2426 		} else if (level >= 0) {
2427 			/*
2428 			 * The page is reserved but not yet allocated.  In
2429 			 * other words, it is still free.  Extend the current
2430 			 * run by one page.
2431 			 */
2432 			run_ext = 1;
2433 #endif
2434 		} else if ((order = m->order) < VM_NFREEORDER) {
2435 			/*
2436 			 * The page is enqueued in the physical memory
2437 			 * allocator's free page queues.  Moreover, it is the
2438 			 * first page in a power-of-two-sized run of
2439 			 * contiguous free pages.  Add these pages to the end
2440 			 * of the current run, and jump ahead.
2441 			 */
2442 			run_ext = 1 << order;
2443 			m_inc = 1 << order;
2444 		} else {
2445 			/*
2446 			 * Skip the page for one of the following reasons: (1)
2447 			 * It is enqueued in the physical memory allocator's
2448 			 * free page queues.  However, it is not the first
2449 			 * page in a run of contiguous free pages.  (This case
2450 			 * rarely occurs because the scan is performed in
2451 			 * ascending order.) (2) It is not reserved, and it is
2452 			 * transitioning from free to allocated.  (Conversely,
2453 			 * the transition from allocated to free for managed
2454 			 * pages is blocked by the page lock.) (3) It is
2455 			 * allocated but not contained by an object and not
2456 			 * wired, e.g., allocated by Xen's balloon driver.
2457 			 */
2458 			run_ext = 0;
2459 		}
2460 
2461 		/*
2462 		 * Extend or reset the current run of pages.
2463 		 */
2464 		if (run_ext > 0) {
2465 			if (run_len == 0)
2466 				m_run = m;
2467 			run_len += run_ext;
2468 		} else {
2469 			if (run_len > 0) {
2470 				m_run = NULL;
2471 				run_len = 0;
2472 			}
2473 		}
2474 	}
2475 	if (m_mtx != NULL)
2476 		mtx_unlock(m_mtx);
2477 	if (run_len >= npages)
2478 		return (m_run);
2479 	return (NULL);
2480 }
2481 
2482 /*
2483  *	vm_page_reclaim_run:
2484  *
2485  *	Try to relocate each of the allocated virtual pages within the
2486  *	specified run of physical pages to a new physical address.  Free the
2487  *	physical pages underlying the relocated virtual pages.  A virtual page
2488  *	is relocatable if and only if it could be laundered or reclaimed by
2489  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2490  *	physical address above "high".
2491  *
2492  *	Returns 0 if every physical page within the run was already free or
2493  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2494  *	value indicating why the last attempt to relocate a virtual page was
2495  *	unsuccessful.
2496  *
2497  *	"req_class" must be an allocation class.
2498  */
2499 static int
2500 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2501     vm_paddr_t high)
2502 {
2503 	struct vm_domain *vmd;
2504 	struct mtx *m_mtx;
2505 	struct spglist free;
2506 	vm_object_t object;
2507 	vm_paddr_t pa;
2508 	vm_page_t m, m_end, m_new;
2509 	int error, order, req;
2510 
2511 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2512 	    ("req_class is not an allocation class"));
2513 	SLIST_INIT(&free);
2514 	error = 0;
2515 	m = m_run;
2516 	m_end = m_run + npages;
2517 	m_mtx = NULL;
2518 	for (; error == 0 && m < m_end; m++) {
2519 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2520 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2521 
2522 		/*
2523 		 * Avoid releasing and reacquiring the same page lock.
2524 		 */
2525 		vm_page_change_lock(m, &m_mtx);
2526 retry:
2527 		if (vm_page_held(m))
2528 			error = EBUSY;
2529 		else if ((object = m->object) != NULL) {
2530 			/*
2531 			 * The page is relocated if and only if it could be
2532 			 * laundered or reclaimed by the page daemon.
2533 			 */
2534 			if (!VM_OBJECT_TRYWLOCK(object)) {
2535 				mtx_unlock(m_mtx);
2536 				VM_OBJECT_WLOCK(object);
2537 				mtx_lock(m_mtx);
2538 				if (m->object != object) {
2539 					/*
2540 					 * The page may have been freed.
2541 					 */
2542 					VM_OBJECT_WUNLOCK(object);
2543 					goto retry;
2544 				} else if (vm_page_held(m)) {
2545 					error = EBUSY;
2546 					goto unlock;
2547 				}
2548 			}
2549 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2550 			    ("page %p is PG_UNHOLDFREE", m));
2551 			/* Don't care: PG_NODUMP, PG_ZERO. */
2552 			if (object->type != OBJT_DEFAULT &&
2553 			    object->type != OBJT_SWAP &&
2554 			    object->type != OBJT_VNODE)
2555 				error = EINVAL;
2556 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2557 				error = EINVAL;
2558 			else if (vm_page_queue(m) != PQ_NONE &&
2559 			    !vm_page_busied(m)) {
2560 				KASSERT(pmap_page_get_memattr(m) ==
2561 				    VM_MEMATTR_DEFAULT,
2562 				    ("page %p has an unexpected memattr", m));
2563 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2564 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2565 				    ("page %p has unexpected oflags", m));
2566 				/* Don't care: VPO_NOSYNC. */
2567 				if (m->valid != 0) {
2568 					/*
2569 					 * First, try to allocate a new page
2570 					 * that is above "high".  Failing
2571 					 * that, try to allocate a new page
2572 					 * that is below "m_run".  Allocate
2573 					 * the new page between the end of
2574 					 * "m_run" and "high" only as a last
2575 					 * resort.
2576 					 */
2577 					req = req_class | VM_ALLOC_NOOBJ;
2578 					if ((m->flags & PG_NODUMP) != 0)
2579 						req |= VM_ALLOC_NODUMP;
2580 					if (trunc_page(high) !=
2581 					    ~(vm_paddr_t)PAGE_MASK) {
2582 						m_new = vm_page_alloc_contig(
2583 						    NULL, 0, req, 1,
2584 						    round_page(high),
2585 						    ~(vm_paddr_t)0,
2586 						    PAGE_SIZE, 0,
2587 						    VM_MEMATTR_DEFAULT);
2588 					} else
2589 						m_new = NULL;
2590 					if (m_new == NULL) {
2591 						pa = VM_PAGE_TO_PHYS(m_run);
2592 						m_new = vm_page_alloc_contig(
2593 						    NULL, 0, req, 1,
2594 						    0, pa - 1, PAGE_SIZE, 0,
2595 						    VM_MEMATTR_DEFAULT);
2596 					}
2597 					if (m_new == NULL) {
2598 						pa += ptoa(npages);
2599 						m_new = vm_page_alloc_contig(
2600 						    NULL, 0, req, 1,
2601 						    pa, high, PAGE_SIZE, 0,
2602 						    VM_MEMATTR_DEFAULT);
2603 					}
2604 					if (m_new == NULL) {
2605 						error = ENOMEM;
2606 						goto unlock;
2607 					}
2608 					KASSERT(m_new->wire_count == 0,
2609 					    ("page %p is wired", m_new));
2610 
2611 					/*
2612 					 * Replace "m" with the new page.  For
2613 					 * vm_page_replace(), "m" must be busy
2614 					 * and dequeued.  Finally, change "m"
2615 					 * as if vm_page_free() was called.
2616 					 */
2617 					if (object->ref_count != 0)
2618 						pmap_remove_all(m);
2619 					m_new->aflags = m->aflags &
2620 					    ~PGA_QUEUE_STATE_MASK;
2621 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2622 					    ("page %p is managed", m_new));
2623 					m_new->oflags = m->oflags & VPO_NOSYNC;
2624 					pmap_copy_page(m, m_new);
2625 					m_new->valid = m->valid;
2626 					m_new->dirty = m->dirty;
2627 					m->flags &= ~PG_ZERO;
2628 					vm_page_xbusy(m);
2629 					vm_page_dequeue(m);
2630 					vm_page_replace_checked(m_new, object,
2631 					    m->pindex, m);
2632 					if (vm_page_free_prep(m))
2633 						SLIST_INSERT_HEAD(&free, m,
2634 						    plinks.s.ss);
2635 
2636 					/*
2637 					 * The new page must be deactivated
2638 					 * before the object is unlocked.
2639 					 */
2640 					vm_page_change_lock(m_new, &m_mtx);
2641 					vm_page_deactivate(m_new);
2642 				} else {
2643 					m->flags &= ~PG_ZERO;
2644 					vm_page_dequeue(m);
2645 					vm_page_remove(m);
2646 					if (vm_page_free_prep(m))
2647 						SLIST_INSERT_HEAD(&free, m,
2648 						    plinks.s.ss);
2649 					KASSERT(m->dirty == 0,
2650 					    ("page %p is dirty", m));
2651 				}
2652 			} else
2653 				error = EBUSY;
2654 unlock:
2655 			VM_OBJECT_WUNLOCK(object);
2656 		} else {
2657 			MPASS(vm_phys_domain(m) == domain);
2658 			vmd = VM_DOMAIN(domain);
2659 			vm_domain_free_lock(vmd);
2660 			order = m->order;
2661 			if (order < VM_NFREEORDER) {
2662 				/*
2663 				 * The page is enqueued in the physical memory
2664 				 * allocator's free page queues.  Moreover, it
2665 				 * is the first page in a power-of-two-sized
2666 				 * run of contiguous free pages.  Jump ahead
2667 				 * to the last page within that run, and
2668 				 * continue from there.
2669 				 */
2670 				m += (1 << order) - 1;
2671 			}
2672 #if VM_NRESERVLEVEL > 0
2673 			else if (vm_reserv_is_page_free(m))
2674 				order = 0;
2675 #endif
2676 			vm_domain_free_unlock(vmd);
2677 			if (order == VM_NFREEORDER)
2678 				error = EINVAL;
2679 		}
2680 	}
2681 	if (m_mtx != NULL)
2682 		mtx_unlock(m_mtx);
2683 	if ((m = SLIST_FIRST(&free)) != NULL) {
2684 		int cnt;
2685 
2686 		vmd = VM_DOMAIN(domain);
2687 		cnt = 0;
2688 		vm_domain_free_lock(vmd);
2689 		do {
2690 			MPASS(vm_phys_domain(m) == domain);
2691 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2692 			vm_phys_free_pages(m, 0);
2693 			cnt++;
2694 		} while ((m = SLIST_FIRST(&free)) != NULL);
2695 		vm_domain_free_unlock(vmd);
2696 		vm_domain_freecnt_inc(vmd, cnt);
2697 	}
2698 	return (error);
2699 }
2700 
2701 #define	NRUNS	16
2702 
2703 CTASSERT(powerof2(NRUNS));
2704 
2705 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2706 
2707 #define	MIN_RECLAIM	8
2708 
2709 /*
2710  *	vm_page_reclaim_contig:
2711  *
2712  *	Reclaim allocated, contiguous physical memory satisfying the specified
2713  *	conditions by relocating the virtual pages using that physical memory.
2714  *	Returns true if reclamation is successful and false otherwise.  Since
2715  *	relocation requires the allocation of physical pages, reclamation may
2716  *	fail due to a shortage of free pages.  When reclamation fails, callers
2717  *	are expected to perform vm_wait() before retrying a failed allocation
2718  *	operation, e.g., vm_page_alloc_contig().
2719  *
2720  *	The caller must always specify an allocation class through "req".
2721  *
2722  *	allocation classes:
2723  *	VM_ALLOC_NORMAL		normal process request
2724  *	VM_ALLOC_SYSTEM		system *really* needs a page
2725  *	VM_ALLOC_INTERRUPT	interrupt time request
2726  *
2727  *	The optional allocation flags are ignored.
2728  *
2729  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2730  *	must be a power of two.
2731  */
2732 bool
2733 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2734     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2735 {
2736 	struct vm_domain *vmd;
2737 	vm_paddr_t curr_low;
2738 	vm_page_t m_run, m_runs[NRUNS];
2739 	u_long count, reclaimed;
2740 	int error, i, options, req_class;
2741 
2742 	KASSERT(npages > 0, ("npages is 0"));
2743 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2744 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2745 	req_class = req & VM_ALLOC_CLASS_MASK;
2746 
2747 	/*
2748 	 * The page daemon is allowed to dig deeper into the free page list.
2749 	 */
2750 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2751 		req_class = VM_ALLOC_SYSTEM;
2752 
2753 	/*
2754 	 * Return if the number of free pages cannot satisfy the requested
2755 	 * allocation.
2756 	 */
2757 	vmd = VM_DOMAIN(domain);
2758 	count = vmd->vmd_free_count;
2759 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2760 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2761 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2762 		return (false);
2763 
2764 	/*
2765 	 * Scan up to three times, relaxing the restrictions ("options") on
2766 	 * the reclamation of reservations and superpages each time.
2767 	 */
2768 	for (options = VPSC_NORESERV;;) {
2769 		/*
2770 		 * Find the highest runs that satisfy the given constraints
2771 		 * and restrictions, and record them in "m_runs".
2772 		 */
2773 		curr_low = low;
2774 		count = 0;
2775 		for (;;) {
2776 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2777 			    high, alignment, boundary, options);
2778 			if (m_run == NULL)
2779 				break;
2780 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2781 			m_runs[RUN_INDEX(count)] = m_run;
2782 			count++;
2783 		}
2784 
2785 		/*
2786 		 * Reclaim the highest runs in LIFO (descending) order until
2787 		 * the number of reclaimed pages, "reclaimed", is at least
2788 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2789 		 * reclamation is idempotent, and runs will (likely) recur
2790 		 * from one scan to the next as restrictions are relaxed.
2791 		 */
2792 		reclaimed = 0;
2793 		for (i = 0; count > 0 && i < NRUNS; i++) {
2794 			count--;
2795 			m_run = m_runs[RUN_INDEX(count)];
2796 			error = vm_page_reclaim_run(req_class, domain, npages,
2797 			    m_run, high);
2798 			if (error == 0) {
2799 				reclaimed += npages;
2800 				if (reclaimed >= MIN_RECLAIM)
2801 					return (true);
2802 			}
2803 		}
2804 
2805 		/*
2806 		 * Either relax the restrictions on the next scan or return if
2807 		 * the last scan had no restrictions.
2808 		 */
2809 		if (options == VPSC_NORESERV)
2810 			options = VPSC_NOSUPER;
2811 		else if (options == VPSC_NOSUPER)
2812 			options = VPSC_ANY;
2813 		else if (options == VPSC_ANY)
2814 			return (reclaimed != 0);
2815 	}
2816 }
2817 
2818 bool
2819 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2820     u_long alignment, vm_paddr_t boundary)
2821 {
2822 	struct vm_domainset_iter di;
2823 	int domain;
2824 	bool ret;
2825 
2826 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2827 	do {
2828 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2829 		    high, alignment, boundary);
2830 		if (ret)
2831 			break;
2832 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2833 
2834 	return (ret);
2835 }
2836 
2837 /*
2838  * Set the domain in the appropriate page level domainset.
2839  */
2840 void
2841 vm_domain_set(struct vm_domain *vmd)
2842 {
2843 
2844 	mtx_lock(&vm_domainset_lock);
2845 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2846 		vmd->vmd_minset = 1;
2847 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2848 	}
2849 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2850 		vmd->vmd_severeset = 1;
2851 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2852 	}
2853 	mtx_unlock(&vm_domainset_lock);
2854 }
2855 
2856 /*
2857  * Clear the domain from the appropriate page level domainset.
2858  */
2859 void
2860 vm_domain_clear(struct vm_domain *vmd)
2861 {
2862 
2863 	mtx_lock(&vm_domainset_lock);
2864 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2865 		vmd->vmd_minset = 0;
2866 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2867 		if (vm_min_waiters != 0) {
2868 			vm_min_waiters = 0;
2869 			wakeup(&vm_min_domains);
2870 		}
2871 	}
2872 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2873 		vmd->vmd_severeset = 0;
2874 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2875 		if (vm_severe_waiters != 0) {
2876 			vm_severe_waiters = 0;
2877 			wakeup(&vm_severe_domains);
2878 		}
2879 	}
2880 
2881 	/*
2882 	 * If pageout daemon needs pages, then tell it that there are
2883 	 * some free.
2884 	 */
2885 	if (vmd->vmd_pageout_pages_needed &&
2886 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2887 		wakeup(&vmd->vmd_pageout_pages_needed);
2888 		vmd->vmd_pageout_pages_needed = 0;
2889 	}
2890 
2891 	/* See comments in vm_wait_doms(). */
2892 	if (vm_pageproc_waiters) {
2893 		vm_pageproc_waiters = 0;
2894 		wakeup(&vm_pageproc_waiters);
2895 	}
2896 	mtx_unlock(&vm_domainset_lock);
2897 }
2898 
2899 /*
2900  * Wait for free pages to exceed the min threshold globally.
2901  */
2902 void
2903 vm_wait_min(void)
2904 {
2905 
2906 	mtx_lock(&vm_domainset_lock);
2907 	while (vm_page_count_min()) {
2908 		vm_min_waiters++;
2909 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2910 	}
2911 	mtx_unlock(&vm_domainset_lock);
2912 }
2913 
2914 /*
2915  * Wait for free pages to exceed the severe threshold globally.
2916  */
2917 void
2918 vm_wait_severe(void)
2919 {
2920 
2921 	mtx_lock(&vm_domainset_lock);
2922 	while (vm_page_count_severe()) {
2923 		vm_severe_waiters++;
2924 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2925 		    "vmwait", 0);
2926 	}
2927 	mtx_unlock(&vm_domainset_lock);
2928 }
2929 
2930 u_int
2931 vm_wait_count(void)
2932 {
2933 
2934 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2935 }
2936 
2937 void
2938 vm_wait_doms(const domainset_t *wdoms)
2939 {
2940 
2941 	/*
2942 	 * We use racey wakeup synchronization to avoid expensive global
2943 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
2944 	 * To handle this, we only sleep for one tick in this instance.  It
2945 	 * is expected that most allocations for the pageproc will come from
2946 	 * kmem or vm_page_grab* which will use the more specific and
2947 	 * race-free vm_wait_domain().
2948 	 */
2949 	if (curproc == pageproc) {
2950 		mtx_lock(&vm_domainset_lock);
2951 		vm_pageproc_waiters++;
2952 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2953 		    "pageprocwait", 1);
2954 	} else {
2955 		/*
2956 		 * XXX Ideally we would wait only until the allocation could
2957 		 * be satisfied.  This condition can cause new allocators to
2958 		 * consume all freed pages while old allocators wait.
2959 		 */
2960 		mtx_lock(&vm_domainset_lock);
2961 		if (vm_page_count_min_set(wdoms)) {
2962 			vm_min_waiters++;
2963 			msleep(&vm_min_domains, &vm_domainset_lock,
2964 			    PVM | PDROP, "vmwait", 0);
2965 		} else
2966 			mtx_unlock(&vm_domainset_lock);
2967 	}
2968 }
2969 
2970 /*
2971  *	vm_wait_domain:
2972  *
2973  *	Sleep until free pages are available for allocation.
2974  *	- Called in various places after failed memory allocations.
2975  */
2976 void
2977 vm_wait_domain(int domain)
2978 {
2979 	struct vm_domain *vmd;
2980 	domainset_t wdom;
2981 
2982 	vmd = VM_DOMAIN(domain);
2983 	vm_domain_free_assert_unlocked(vmd);
2984 
2985 	if (curproc == pageproc) {
2986 		mtx_lock(&vm_domainset_lock);
2987 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2988 			vmd->vmd_pageout_pages_needed = 1;
2989 			msleep(&vmd->vmd_pageout_pages_needed,
2990 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
2991 		} else
2992 			mtx_unlock(&vm_domainset_lock);
2993 	} else {
2994 		if (pageproc == NULL)
2995 			panic("vm_wait in early boot");
2996 		DOMAINSET_ZERO(&wdom);
2997 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
2998 		vm_wait_doms(&wdom);
2999 	}
3000 }
3001 
3002 /*
3003  *	vm_wait:
3004  *
3005  *	Sleep until free pages are available for allocation in the
3006  *	affinity domains of the obj.  If obj is NULL, the domain set
3007  *	for the calling thread is used.
3008  *	Called in various places after failed memory allocations.
3009  */
3010 void
3011 vm_wait(vm_object_t obj)
3012 {
3013 	struct domainset *d;
3014 
3015 	d = NULL;
3016 
3017 	/*
3018 	 * Carefully fetch pointers only once: the struct domainset
3019 	 * itself is ummutable but the pointer might change.
3020 	 */
3021 	if (obj != NULL)
3022 		d = obj->domain.dr_policy;
3023 	if (d == NULL)
3024 		d = curthread->td_domain.dr_policy;
3025 
3026 	vm_wait_doms(&d->ds_mask);
3027 }
3028 
3029 /*
3030  *	vm_domain_alloc_fail:
3031  *
3032  *	Called when a page allocation function fails.  Informs the
3033  *	pagedaemon and performs the requested wait.  Requires the
3034  *	domain_free and object lock on entry.  Returns with the
3035  *	object lock held and free lock released.  Returns an error when
3036  *	retry is necessary.
3037  *
3038  */
3039 static int
3040 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3041 {
3042 
3043 	vm_domain_free_assert_unlocked(vmd);
3044 
3045 	atomic_add_int(&vmd->vmd_pageout_deficit,
3046 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3047 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3048 		if (object != NULL)
3049 			VM_OBJECT_WUNLOCK(object);
3050 		vm_wait_domain(vmd->vmd_domain);
3051 		if (object != NULL)
3052 			VM_OBJECT_WLOCK(object);
3053 		if (req & VM_ALLOC_WAITOK)
3054 			return (EAGAIN);
3055 	}
3056 
3057 	return (0);
3058 }
3059 
3060 /*
3061  *	vm_waitpfault:
3062  *
3063  *	Sleep until free pages are available for allocation.
3064  *	- Called only in vm_fault so that processes page faulting
3065  *	  can be easily tracked.
3066  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3067  *	  processes will be able to grab memory first.  Do not change
3068  *	  this balance without careful testing first.
3069  */
3070 void
3071 vm_waitpfault(struct domainset *dset)
3072 {
3073 
3074 	/*
3075 	 * XXX Ideally we would wait only until the allocation could
3076 	 * be satisfied.  This condition can cause new allocators to
3077 	 * consume all freed pages while old allocators wait.
3078 	 */
3079 	mtx_lock(&vm_domainset_lock);
3080 	if (vm_page_count_min_set(&dset->ds_mask)) {
3081 		vm_min_waiters++;
3082 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3083 		    "pfault", 0);
3084 	} else
3085 		mtx_unlock(&vm_domainset_lock);
3086 }
3087 
3088 struct vm_pagequeue *
3089 vm_page_pagequeue(vm_page_t m)
3090 {
3091 
3092 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
3093 }
3094 
3095 static struct mtx *
3096 vm_page_pagequeue_lockptr(vm_page_t m)
3097 {
3098 	uint8_t queue;
3099 
3100 	if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3101 		return (NULL);
3102 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex);
3103 }
3104 
3105 static inline void
3106 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3107 {
3108 	struct vm_domain *vmd;
3109 	uint8_t qflags;
3110 
3111 	CRITICAL_ASSERT(curthread);
3112 	vm_pagequeue_assert_locked(pq);
3113 
3114 	/*
3115 	 * The page daemon is allowed to set m->queue = PQ_NONE without
3116 	 * the page queue lock held.  In this case it is about to free the page,
3117 	 * which must not have any queue state.
3118 	 */
3119 	qflags = atomic_load_8(&m->aflags) & PGA_QUEUE_STATE_MASK;
3120 	KASSERT(pq == vm_page_pagequeue(m) || qflags == 0,
3121 	    ("page %p doesn't belong to queue %p but has queue state %#x",
3122 	    m, pq, qflags));
3123 
3124 	if ((qflags & PGA_DEQUEUE) != 0) {
3125 		if (__predict_true((qflags & PGA_ENQUEUED) != 0)) {
3126 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3127 			vm_pagequeue_cnt_dec(pq);
3128 		}
3129 		vm_page_dequeue_complete(m);
3130 	} else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3131 		if ((qflags & PGA_ENQUEUED) != 0)
3132 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3133 		else {
3134 			vm_pagequeue_cnt_inc(pq);
3135 			vm_page_aflag_set(m, PGA_ENQUEUED);
3136 		}
3137 		if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3138 			KASSERT(m->queue == PQ_INACTIVE,
3139 			    ("head enqueue not supported for page %p", m));
3140 			vmd = vm_pagequeue_domain(m);
3141 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3142 		} else
3143 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3144 
3145 		/*
3146 		 * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after
3147 		 * setting PGA_ENQUEUED in order to synchronize with the
3148 		 * page daemon.
3149 		 */
3150 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
3151 	}
3152 }
3153 
3154 static void
3155 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3156     uint8_t queue)
3157 {
3158 	vm_page_t m;
3159 	int i;
3160 
3161 	for (i = 0; i < bq->bq_cnt; i++) {
3162 		m = bq->bq_pa[i];
3163 		if (__predict_false(m->queue != queue))
3164 			continue;
3165 		vm_pqbatch_process_page(pq, m);
3166 	}
3167 	vm_batchqueue_init(bq);
3168 }
3169 
3170 static void
3171 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3172 {
3173 	struct vm_batchqueue *bq;
3174 	struct vm_pagequeue *pq;
3175 	int domain;
3176 
3177 	vm_page_assert_locked(m);
3178 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3179 
3180 	domain = vm_phys_domain(m);
3181 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3182 
3183 	critical_enter();
3184 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3185 	if (vm_batchqueue_insert(bq, m)) {
3186 		critical_exit();
3187 		return;
3188 	}
3189 	if (!vm_pagequeue_trylock(pq)) {
3190 		critical_exit();
3191 		vm_pagequeue_lock(pq);
3192 		critical_enter();
3193 		bq = DPCPU_PTR(pqbatch[domain][queue]);
3194 	}
3195 	vm_pqbatch_process(pq, bq, queue);
3196 
3197 	/*
3198 	 * The page may have been logically dequeued before we acquired the
3199 	 * page queue lock.  In this case, the page lock prevents the page
3200 	 * from being logically enqueued elsewhere.
3201 	 */
3202 	if (__predict_true(m->queue == queue))
3203 		vm_pqbatch_process_page(pq, m);
3204 	else {
3205 		KASSERT(m->queue == PQ_NONE,
3206 		    ("invalid queue transition for page %p", m));
3207 		KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3208 		    ("page %p is enqueued with invalid queue index", m));
3209 		vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3210 	}
3211 	vm_pagequeue_unlock(pq);
3212 	critical_exit();
3213 }
3214 
3215 /*
3216  *	vm_page_drain_pqbatch:		[ internal use only ]
3217  *
3218  *	Force all per-CPU page queue batch queues to be drained.  This is
3219  *	intended for use in severe memory shortages, to ensure that pages
3220  *	do not remain stuck in the batch queues.
3221  */
3222 void
3223 vm_page_drain_pqbatch(void)
3224 {
3225 	struct thread *td;
3226 	struct vm_domain *vmd;
3227 	struct vm_pagequeue *pq;
3228 	int cpu, domain, queue;
3229 
3230 	td = curthread;
3231 	CPU_FOREACH(cpu) {
3232 		thread_lock(td);
3233 		sched_bind(td, cpu);
3234 		thread_unlock(td);
3235 
3236 		for (domain = 0; domain < vm_ndomains; domain++) {
3237 			vmd = VM_DOMAIN(domain);
3238 			for (queue = 0; queue < PQ_COUNT; queue++) {
3239 				pq = &vmd->vmd_pagequeues[queue];
3240 				vm_pagequeue_lock(pq);
3241 				critical_enter();
3242 				vm_pqbatch_process(pq,
3243 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3244 				critical_exit();
3245 				vm_pagequeue_unlock(pq);
3246 			}
3247 		}
3248 	}
3249 	thread_lock(td);
3250 	sched_unbind(td);
3251 	thread_unlock(td);
3252 }
3253 
3254 /*
3255  * Complete the logical removal of a page from a page queue.  We must be
3256  * careful to synchronize with the page daemon, which may be concurrently
3257  * examining the page with only the page lock held.  The page must not be
3258  * in a state where it appears to be logically enqueued.
3259  */
3260 static void
3261 vm_page_dequeue_complete(vm_page_t m)
3262 {
3263 
3264 	m->queue = PQ_NONE;
3265 	atomic_thread_fence_rel();
3266 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3267 }
3268 
3269 /*
3270  *	vm_page_dequeue_deferred:	[ internal use only ]
3271  *
3272  *	Request removal of the given page from its current page
3273  *	queue.  Physical removal from the queue may be deferred
3274  *	indefinitely.
3275  *
3276  *	The page must be locked.
3277  */
3278 void
3279 vm_page_dequeue_deferred(vm_page_t m)
3280 {
3281 	int queue;
3282 
3283 	vm_page_assert_locked(m);
3284 
3285 	queue = atomic_load_8(&m->queue);
3286 	if (queue == PQ_NONE) {
3287 		KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3288 		    ("page %p has queue state", m));
3289 		return;
3290 	}
3291 	if ((m->aflags & PGA_DEQUEUE) == 0)
3292 		vm_page_aflag_set(m, PGA_DEQUEUE);
3293 	vm_pqbatch_submit_page(m, queue);
3294 }
3295 
3296 /*
3297  *	vm_page_dequeue:
3298  *
3299  *	Remove the page from whichever page queue it's in, if any.
3300  *	The page must either be locked or unallocated.  This constraint
3301  *	ensures that the queue state of the page will remain consistent
3302  *	after this function returns.
3303  */
3304 void
3305 vm_page_dequeue(vm_page_t m)
3306 {
3307 	struct mtx *lock, *lock1;
3308 	struct vm_pagequeue *pq;
3309 	uint8_t aflags;
3310 
3311 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->order == VM_NFREEORDER,
3312 	    ("page %p is allocated and unlocked", m));
3313 
3314 	for (;;) {
3315 		lock = vm_page_pagequeue_lockptr(m);
3316 		if (lock == NULL) {
3317 			/*
3318 			 * A thread may be concurrently executing
3319 			 * vm_page_dequeue_complete().  Ensure that all queue
3320 			 * state is cleared before we return.
3321 			 */
3322 			aflags = atomic_load_8(&m->aflags);
3323 			if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3324 				return;
3325 			KASSERT((aflags & PGA_DEQUEUE) != 0,
3326 			    ("page %p has unexpected queue state flags %#x",
3327 			    m, aflags));
3328 
3329 			/*
3330 			 * Busy wait until the thread updating queue state is
3331 			 * finished.  Such a thread must be executing in a
3332 			 * critical section.
3333 			 */
3334 			cpu_spinwait();
3335 			continue;
3336 		}
3337 		mtx_lock(lock);
3338 		if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock)
3339 			break;
3340 		mtx_unlock(lock);
3341 		lock = lock1;
3342 	}
3343 	KASSERT(lock == vm_page_pagequeue_lockptr(m),
3344 	    ("%s: page %p migrated directly between queues", __func__, m));
3345 	KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3346 	    mtx_owned(vm_page_lockptr(m)),
3347 	    ("%s: queued unlocked page %p", __func__, m));
3348 
3349 	if ((m->aflags & PGA_ENQUEUED) != 0) {
3350 		pq = vm_page_pagequeue(m);
3351 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3352 		vm_pagequeue_cnt_dec(pq);
3353 	}
3354 	vm_page_dequeue_complete(m);
3355 	mtx_unlock(lock);
3356 }
3357 
3358 /*
3359  * Schedule the given page for insertion into the specified page queue.
3360  * Physical insertion of the page may be deferred indefinitely.
3361  */
3362 static void
3363 vm_page_enqueue(vm_page_t m, uint8_t queue)
3364 {
3365 
3366 	vm_page_assert_locked(m);
3367 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3368 	    ("%s: page %p is already enqueued", __func__, m));
3369 
3370 	m->queue = queue;
3371 	if ((m->aflags & PGA_REQUEUE) == 0)
3372 		vm_page_aflag_set(m, PGA_REQUEUE);
3373 	vm_pqbatch_submit_page(m, queue);
3374 }
3375 
3376 /*
3377  *	vm_page_requeue:		[ internal use only ]
3378  *
3379  *	Schedule a requeue of the given page.
3380  *
3381  *	The page must be locked.
3382  */
3383 void
3384 vm_page_requeue(vm_page_t m)
3385 {
3386 
3387 	vm_page_assert_locked(m);
3388 	KASSERT(m->queue != PQ_NONE,
3389 	    ("%s: page %p is not logically enqueued", __func__, m));
3390 
3391 	if ((m->aflags & PGA_REQUEUE) == 0)
3392 		vm_page_aflag_set(m, PGA_REQUEUE);
3393 	vm_pqbatch_submit_page(m, atomic_load_8(&m->queue));
3394 }
3395 
3396 /*
3397  *	vm_page_activate:
3398  *
3399  *	Put the specified page on the active list (if appropriate).
3400  *	Ensure that act_count is at least ACT_INIT but do not otherwise
3401  *	mess with it.
3402  *
3403  *	The page must be locked.
3404  */
3405 void
3406 vm_page_activate(vm_page_t m)
3407 {
3408 
3409 	vm_page_assert_locked(m);
3410 
3411 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3412 		return;
3413 	if (vm_page_queue(m) == PQ_ACTIVE) {
3414 		if (m->act_count < ACT_INIT)
3415 			m->act_count = ACT_INIT;
3416 		return;
3417 	}
3418 
3419 	vm_page_dequeue(m);
3420 	if (m->act_count < ACT_INIT)
3421 		m->act_count = ACT_INIT;
3422 	vm_page_enqueue(m, PQ_ACTIVE);
3423 }
3424 
3425 /*
3426  *	vm_page_free_prep:
3427  *
3428  *	Prepares the given page to be put on the free list,
3429  *	disassociating it from any VM object. The caller may return
3430  *	the page to the free list only if this function returns true.
3431  *
3432  *	The object must be locked.  The page must be locked if it is
3433  *	managed.
3434  */
3435 bool
3436 vm_page_free_prep(vm_page_t m)
3437 {
3438 
3439 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3440 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3441 		uint64_t *p;
3442 		int i;
3443 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3444 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3445 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3446 			    m, i, (uintmax_t)*p));
3447 	}
3448 #endif
3449 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3450 		vm_page_lock_assert(m, MA_OWNED);
3451 		KASSERT(!pmap_page_is_mapped(m),
3452 		    ("vm_page_free_prep: freeing mapped page %p", m));
3453 	} else
3454 		KASSERT(m->queue == PQ_NONE,
3455 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3456 	VM_CNT_INC(v_tfree);
3457 
3458 	if (vm_page_sbusied(m))
3459 		panic("vm_page_free_prep: freeing busy page %p", m);
3460 
3461 	vm_page_remove(m);
3462 
3463 	/*
3464 	 * If fictitious remove object association and
3465 	 * return.
3466 	 */
3467 	if ((m->flags & PG_FICTITIOUS) != 0) {
3468 		KASSERT(m->wire_count == 1,
3469 		    ("fictitious page %p is not wired", m));
3470 		KASSERT(m->queue == PQ_NONE,
3471 		    ("fictitious page %p is queued", m));
3472 		return (false);
3473 	}
3474 
3475 	/*
3476 	 * Pages need not be dequeued before they are returned to the physical
3477 	 * memory allocator, but they must at least be marked for a deferred
3478 	 * dequeue.
3479 	 */
3480 	if ((m->oflags & VPO_UNMANAGED) == 0)
3481 		vm_page_dequeue_deferred(m);
3482 
3483 	m->valid = 0;
3484 	vm_page_undirty(m);
3485 
3486 	if (m->wire_count != 0)
3487 		panic("vm_page_free_prep: freeing wired page %p", m);
3488 	if (m->hold_count != 0) {
3489 		m->flags &= ~PG_ZERO;
3490 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3491 		    ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m));
3492 		m->flags |= PG_UNHOLDFREE;
3493 		return (false);
3494 	}
3495 
3496 	/*
3497 	 * Restore the default memory attribute to the page.
3498 	 */
3499 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3500 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3501 
3502 #if VM_NRESERVLEVEL > 0
3503 	if (vm_reserv_free_page(m))
3504 		return (false);
3505 #endif
3506 
3507 	return (true);
3508 }
3509 
3510 /*
3511  *	vm_page_free_toq:
3512  *
3513  *	Returns the given page to the free list, disassociating it
3514  *	from any VM object.
3515  *
3516  *	The object must be locked.  The page must be locked if it is
3517  *	managed.
3518  */
3519 void
3520 vm_page_free_toq(vm_page_t m)
3521 {
3522 	struct vm_domain *vmd;
3523 
3524 	if (!vm_page_free_prep(m))
3525 		return;
3526 
3527 	vmd = vm_pagequeue_domain(m);
3528 	if (m->pool == VM_FREEPOOL_DEFAULT && vmd->vmd_pgcache != NULL) {
3529 		uma_zfree(vmd->vmd_pgcache, m);
3530 		return;
3531 	}
3532 	vm_domain_free_lock(vmd);
3533 	vm_phys_free_pages(m, 0);
3534 	vm_domain_free_unlock(vmd);
3535 	vm_domain_freecnt_inc(vmd, 1);
3536 }
3537 
3538 /*
3539  *	vm_page_free_pages_toq:
3540  *
3541  *	Returns a list of pages to the free list, disassociating it
3542  *	from any VM object.  In other words, this is equivalent to
3543  *	calling vm_page_free_toq() for each page of a list of VM objects.
3544  *
3545  *	The objects must be locked.  The pages must be locked if it is
3546  *	managed.
3547  */
3548 void
3549 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3550 {
3551 	vm_page_t m;
3552 	int count;
3553 
3554 	if (SLIST_EMPTY(free))
3555 		return;
3556 
3557 	count = 0;
3558 	while ((m = SLIST_FIRST(free)) != NULL) {
3559 		count++;
3560 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3561 		vm_page_free_toq(m);
3562 	}
3563 
3564 	if (update_wire_count)
3565 		vm_wire_sub(count);
3566 }
3567 
3568 /*
3569  *	vm_page_wire:
3570  *
3571  * Mark this page as wired down.  If the page is fictitious, then
3572  * its wire count must remain one.
3573  *
3574  * The page must be locked.
3575  */
3576 void
3577 vm_page_wire(vm_page_t m)
3578 {
3579 
3580 	vm_page_assert_locked(m);
3581 	if ((m->flags & PG_FICTITIOUS) != 0) {
3582 		KASSERT(m->wire_count == 1,
3583 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
3584 		    m));
3585 		return;
3586 	}
3587 	if (m->wire_count == 0) {
3588 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3589 		    m->queue == PQ_NONE,
3590 		    ("vm_page_wire: unmanaged page %p is queued", m));
3591 		vm_wire_add(1);
3592 	}
3593 	m->wire_count++;
3594 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3595 }
3596 
3597 /*
3598  * vm_page_unwire:
3599  *
3600  * Release one wiring of the specified page, potentially allowing it to be
3601  * paged out.  Returns TRUE if the number of wirings transitions to zero and
3602  * FALSE otherwise.
3603  *
3604  * Only managed pages belonging to an object can be paged out.  If the number
3605  * of wirings transitions to zero and the page is eligible for page out, then
3606  * the page is added to the specified paging queue (unless PQ_NONE is
3607  * specified, in which case the page is dequeued if it belongs to a paging
3608  * queue).
3609  *
3610  * If a page is fictitious, then its wire count must always be one.
3611  *
3612  * A managed page must be locked.
3613  */
3614 bool
3615 vm_page_unwire(vm_page_t m, uint8_t queue)
3616 {
3617 	bool unwired;
3618 
3619 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3620 	    ("vm_page_unwire: invalid queue %u request for page %p",
3621 	    queue, m));
3622 	if ((m->oflags & VPO_UNMANAGED) == 0)
3623 		vm_page_assert_locked(m);
3624 
3625 	unwired = vm_page_unwire_noq(m);
3626 	if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3627 		return (unwired);
3628 
3629 	if (vm_page_queue(m) == queue) {
3630 		if (queue == PQ_ACTIVE)
3631 			vm_page_reference(m);
3632 		else if (queue != PQ_NONE)
3633 			vm_page_requeue(m);
3634 	} else {
3635 		vm_page_dequeue(m);
3636 		if (queue != PQ_NONE) {
3637 			vm_page_enqueue(m, queue);
3638 			if (queue == PQ_ACTIVE)
3639 				/* Initialize act_count. */
3640 				vm_page_activate(m);
3641 		}
3642 	}
3643 	return (unwired);
3644 }
3645 
3646 /*
3647  *
3648  * vm_page_unwire_noq:
3649  *
3650  * Unwire a page without (re-)inserting it into a page queue.  It is up
3651  * to the caller to enqueue, requeue, or free the page as appropriate.
3652  * In most cases, vm_page_unwire() should be used instead.
3653  */
3654 bool
3655 vm_page_unwire_noq(vm_page_t m)
3656 {
3657 
3658 	if ((m->oflags & VPO_UNMANAGED) == 0)
3659 		vm_page_assert_locked(m);
3660 	if ((m->flags & PG_FICTITIOUS) != 0) {
3661 		KASSERT(m->wire_count == 1,
3662 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3663 		return (false);
3664 	}
3665 	if (m->wire_count == 0)
3666 		panic("vm_page_unwire: page %p's wire count is zero", m);
3667 	m->wire_count--;
3668 	if (m->wire_count == 0) {
3669 		vm_wire_sub(1);
3670 		return (true);
3671 	} else
3672 		return (false);
3673 }
3674 
3675 /*
3676  * Move the specified page to the tail of the inactive queue, or requeue
3677  * the page if it is already in the inactive queue.
3678  *
3679  * The page must be locked.
3680  */
3681 void
3682 vm_page_deactivate(vm_page_t m)
3683 {
3684 
3685 	vm_page_assert_locked(m);
3686 
3687 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3688 		return;
3689 
3690 	if (!vm_page_inactive(m)) {
3691 		vm_page_dequeue(m);
3692 		vm_page_enqueue(m, PQ_INACTIVE);
3693 	} else
3694 		vm_page_requeue(m);
3695 }
3696 
3697 /*
3698  * Move the specified page close to the head of the inactive queue,
3699  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3700  * As with regular enqueues, we use a per-CPU batch queue to reduce
3701  * contention on the page queue lock.
3702  *
3703  * The page must be locked.
3704  */
3705 void
3706 vm_page_deactivate_noreuse(vm_page_t m)
3707 {
3708 
3709 	vm_page_assert_locked(m);
3710 
3711 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3712 		return;
3713 
3714 	if (!vm_page_inactive(m)) {
3715 		vm_page_dequeue(m);
3716 		m->queue = PQ_INACTIVE;
3717 	}
3718 	if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3719 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3720 	vm_pqbatch_submit_page(m, PQ_INACTIVE);
3721 }
3722 
3723 /*
3724  * vm_page_launder
3725  *
3726  * 	Put a page in the laundry, or requeue it if it is already there.
3727  */
3728 void
3729 vm_page_launder(vm_page_t m)
3730 {
3731 
3732 	vm_page_assert_locked(m);
3733 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3734 		return;
3735 
3736 	if (vm_page_in_laundry(m))
3737 		vm_page_requeue(m);
3738 	else {
3739 		vm_page_dequeue(m);
3740 		vm_page_enqueue(m, PQ_LAUNDRY);
3741 	}
3742 }
3743 
3744 /*
3745  * vm_page_unswappable
3746  *
3747  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3748  */
3749 void
3750 vm_page_unswappable(vm_page_t m)
3751 {
3752 
3753 	vm_page_assert_locked(m);
3754 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3755 	    ("page %p already unswappable", m));
3756 
3757 	vm_page_dequeue(m);
3758 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
3759 }
3760 
3761 /*
3762  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3763  * if the page is freed and false otherwise.
3764  *
3765  * The page must be managed.  The page and its containing object must be
3766  * locked.
3767  */
3768 bool
3769 vm_page_try_to_free(vm_page_t m)
3770 {
3771 
3772 	vm_page_assert_locked(m);
3773 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3774 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3775 	if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m))
3776 		return (false);
3777 	if (m->object->ref_count != 0) {
3778 		pmap_remove_all(m);
3779 		if (m->dirty != 0)
3780 			return (false);
3781 	}
3782 	vm_page_free(m);
3783 	return (true);
3784 }
3785 
3786 /*
3787  * vm_page_advise
3788  *
3789  * 	Apply the specified advice to the given page.
3790  *
3791  *	The object and page must be locked.
3792  */
3793 void
3794 vm_page_advise(vm_page_t m, int advice)
3795 {
3796 
3797 	vm_page_assert_locked(m);
3798 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3799 	if (advice == MADV_FREE)
3800 		/*
3801 		 * Mark the page clean.  This will allow the page to be freed
3802 		 * without first paging it out.  MADV_FREE pages are often
3803 		 * quickly reused by malloc(3), so we do not do anything that
3804 		 * would result in a page fault on a later access.
3805 		 */
3806 		vm_page_undirty(m);
3807 	else if (advice != MADV_DONTNEED) {
3808 		if (advice == MADV_WILLNEED)
3809 			vm_page_activate(m);
3810 		return;
3811 	}
3812 
3813 	/*
3814 	 * Clear any references to the page.  Otherwise, the page daemon will
3815 	 * immediately reactivate the page.
3816 	 */
3817 	vm_page_aflag_clear(m, PGA_REFERENCED);
3818 
3819 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3820 		vm_page_dirty(m);
3821 
3822 	/*
3823 	 * Place clean pages near the head of the inactive queue rather than
3824 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3825 	 * the page will be reused quickly.  Dirty pages not already in the
3826 	 * laundry are moved there.
3827 	 */
3828 	if (m->dirty == 0)
3829 		vm_page_deactivate_noreuse(m);
3830 	else if (!vm_page_in_laundry(m))
3831 		vm_page_launder(m);
3832 }
3833 
3834 /*
3835  * Grab a page, waiting until we are waken up due to the page
3836  * changing state.  We keep on waiting, if the page continues
3837  * to be in the object.  If the page doesn't exist, first allocate it
3838  * and then conditionally zero it.
3839  *
3840  * This routine may sleep.
3841  *
3842  * The object must be locked on entry.  The lock will, however, be released
3843  * and reacquired if the routine sleeps.
3844  */
3845 vm_page_t
3846 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3847 {
3848 	vm_page_t m;
3849 	int sleep;
3850 	int pflags;
3851 
3852 	VM_OBJECT_ASSERT_WLOCKED(object);
3853 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3854 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3855 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3856 	pflags = allocflags &
3857 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3858 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3859 		pflags |= VM_ALLOC_WAITFAIL;
3860 retrylookup:
3861 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3862 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3863 		    vm_page_xbusied(m) : vm_page_busied(m);
3864 		if (sleep) {
3865 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3866 				return (NULL);
3867 			/*
3868 			 * Reference the page before unlocking and
3869 			 * sleeping so that the page daemon is less
3870 			 * likely to reclaim it.
3871 			 */
3872 			vm_page_aflag_set(m, PGA_REFERENCED);
3873 			vm_page_lock(m);
3874 			VM_OBJECT_WUNLOCK(object);
3875 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3876 			    VM_ALLOC_IGN_SBUSY) != 0);
3877 			VM_OBJECT_WLOCK(object);
3878 			goto retrylookup;
3879 		} else {
3880 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3881 				vm_page_lock(m);
3882 				vm_page_wire(m);
3883 				vm_page_unlock(m);
3884 			}
3885 			if ((allocflags &
3886 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3887 				vm_page_xbusy(m);
3888 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3889 				vm_page_sbusy(m);
3890 			return (m);
3891 		}
3892 	}
3893 	m = vm_page_alloc(object, pindex, pflags);
3894 	if (m == NULL) {
3895 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3896 			return (NULL);
3897 		goto retrylookup;
3898 	}
3899 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3900 		pmap_zero_page(m);
3901 	return (m);
3902 }
3903 
3904 /*
3905  * Return the specified range of pages from the given object.  For each
3906  * page offset within the range, if a page already exists within the object
3907  * at that offset and it is busy, then wait for it to change state.  If,
3908  * instead, the page doesn't exist, then allocate it.
3909  *
3910  * The caller must always specify an allocation class.
3911  *
3912  * allocation classes:
3913  *	VM_ALLOC_NORMAL		normal process request
3914  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3915  *
3916  * The caller must always specify that the pages are to be busied and/or
3917  * wired.
3918  *
3919  * optional allocation flags:
3920  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3921  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3922  *	VM_ALLOC_NOWAIT		do not sleep
3923  *	VM_ALLOC_SBUSY		set page to sbusy state
3924  *	VM_ALLOC_WIRED		wire the pages
3925  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3926  *
3927  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3928  * may return a partial prefix of the requested range.
3929  */
3930 int
3931 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3932     vm_page_t *ma, int count)
3933 {
3934 	vm_page_t m, mpred;
3935 	int pflags;
3936 	int i;
3937 	bool sleep;
3938 
3939 	VM_OBJECT_ASSERT_WLOCKED(object);
3940 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3941 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3942 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3943 	    (allocflags & VM_ALLOC_WIRED) != 0,
3944 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3945 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3946 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3947 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3948 	if (count == 0)
3949 		return (0);
3950 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3951 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3952 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3953 		pflags |= VM_ALLOC_WAITFAIL;
3954 	i = 0;
3955 retrylookup:
3956 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3957 	if (m == NULL || m->pindex != pindex + i) {
3958 		mpred = m;
3959 		m = NULL;
3960 	} else
3961 		mpred = TAILQ_PREV(m, pglist, listq);
3962 	for (; i < count; i++) {
3963 		if (m != NULL) {
3964 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3965 			    vm_page_xbusied(m) : vm_page_busied(m);
3966 			if (sleep) {
3967 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3968 					break;
3969 				/*
3970 				 * Reference the page before unlocking and
3971 				 * sleeping so that the page daemon is less
3972 				 * likely to reclaim it.
3973 				 */
3974 				vm_page_aflag_set(m, PGA_REFERENCED);
3975 				vm_page_lock(m);
3976 				VM_OBJECT_WUNLOCK(object);
3977 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
3978 				    VM_ALLOC_IGN_SBUSY) != 0);
3979 				VM_OBJECT_WLOCK(object);
3980 				goto retrylookup;
3981 			}
3982 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3983 				vm_page_lock(m);
3984 				vm_page_wire(m);
3985 				vm_page_unlock(m);
3986 			}
3987 			if ((allocflags & (VM_ALLOC_NOBUSY |
3988 			    VM_ALLOC_SBUSY)) == 0)
3989 				vm_page_xbusy(m);
3990 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3991 				vm_page_sbusy(m);
3992 		} else {
3993 			m = vm_page_alloc_after(object, pindex + i,
3994 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
3995 			if (m == NULL) {
3996 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3997 					break;
3998 				goto retrylookup;
3999 			}
4000 		}
4001 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
4002 			if ((m->flags & PG_ZERO) == 0)
4003 				pmap_zero_page(m);
4004 			m->valid = VM_PAGE_BITS_ALL;
4005 		}
4006 		ma[i] = mpred = m;
4007 		m = vm_page_next(m);
4008 	}
4009 	return (i);
4010 }
4011 
4012 /*
4013  * Mapping function for valid or dirty bits in a page.
4014  *
4015  * Inputs are required to range within a page.
4016  */
4017 vm_page_bits_t
4018 vm_page_bits(int base, int size)
4019 {
4020 	int first_bit;
4021 	int last_bit;
4022 
4023 	KASSERT(
4024 	    base + size <= PAGE_SIZE,
4025 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4026 	);
4027 
4028 	if (size == 0)		/* handle degenerate case */
4029 		return (0);
4030 
4031 	first_bit = base >> DEV_BSHIFT;
4032 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4033 
4034 	return (((vm_page_bits_t)2 << last_bit) -
4035 	    ((vm_page_bits_t)1 << first_bit));
4036 }
4037 
4038 /*
4039  *	vm_page_set_valid_range:
4040  *
4041  *	Sets portions of a page valid.  The arguments are expected
4042  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4043  *	of any partial chunks touched by the range.  The invalid portion of
4044  *	such chunks will be zeroed.
4045  *
4046  *	(base + size) must be less then or equal to PAGE_SIZE.
4047  */
4048 void
4049 vm_page_set_valid_range(vm_page_t m, int base, int size)
4050 {
4051 	int endoff, frag;
4052 
4053 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4054 	if (size == 0)	/* handle degenerate case */
4055 		return;
4056 
4057 	/*
4058 	 * If the base is not DEV_BSIZE aligned and the valid
4059 	 * bit is clear, we have to zero out a portion of the
4060 	 * first block.
4061 	 */
4062 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4063 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4064 		pmap_zero_page_area(m, frag, base - frag);
4065 
4066 	/*
4067 	 * If the ending offset is not DEV_BSIZE aligned and the
4068 	 * valid bit is clear, we have to zero out a portion of
4069 	 * the last block.
4070 	 */
4071 	endoff = base + size;
4072 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4073 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4074 		pmap_zero_page_area(m, endoff,
4075 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4076 
4077 	/*
4078 	 * Assert that no previously invalid block that is now being validated
4079 	 * is already dirty.
4080 	 */
4081 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4082 	    ("vm_page_set_valid_range: page %p is dirty", m));
4083 
4084 	/*
4085 	 * Set valid bits inclusive of any overlap.
4086 	 */
4087 	m->valid |= vm_page_bits(base, size);
4088 }
4089 
4090 /*
4091  * Clear the given bits from the specified page's dirty field.
4092  */
4093 static __inline void
4094 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4095 {
4096 	uintptr_t addr;
4097 #if PAGE_SIZE < 16384
4098 	int shift;
4099 #endif
4100 
4101 	/*
4102 	 * If the object is locked and the page is neither exclusive busy nor
4103 	 * write mapped, then the page's dirty field cannot possibly be
4104 	 * set by a concurrent pmap operation.
4105 	 */
4106 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4107 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4108 		m->dirty &= ~pagebits;
4109 	else {
4110 		/*
4111 		 * The pmap layer can call vm_page_dirty() without
4112 		 * holding a distinguished lock.  The combination of
4113 		 * the object's lock and an atomic operation suffice
4114 		 * to guarantee consistency of the page dirty field.
4115 		 *
4116 		 * For PAGE_SIZE == 32768 case, compiler already
4117 		 * properly aligns the dirty field, so no forcible
4118 		 * alignment is needed. Only require existence of
4119 		 * atomic_clear_64 when page size is 32768.
4120 		 */
4121 		addr = (uintptr_t)&m->dirty;
4122 #if PAGE_SIZE == 32768
4123 		atomic_clear_64((uint64_t *)addr, pagebits);
4124 #elif PAGE_SIZE == 16384
4125 		atomic_clear_32((uint32_t *)addr, pagebits);
4126 #else		/* PAGE_SIZE <= 8192 */
4127 		/*
4128 		 * Use a trick to perform a 32-bit atomic on the
4129 		 * containing aligned word, to not depend on the existence
4130 		 * of atomic_clear_{8, 16}.
4131 		 */
4132 		shift = addr & (sizeof(uint32_t) - 1);
4133 #if BYTE_ORDER == BIG_ENDIAN
4134 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4135 #else
4136 		shift *= NBBY;
4137 #endif
4138 		addr &= ~(sizeof(uint32_t) - 1);
4139 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
4140 #endif		/* PAGE_SIZE */
4141 	}
4142 }
4143 
4144 /*
4145  *	vm_page_set_validclean:
4146  *
4147  *	Sets portions of a page valid and clean.  The arguments are expected
4148  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4149  *	of any partial chunks touched by the range.  The invalid portion of
4150  *	such chunks will be zero'd.
4151  *
4152  *	(base + size) must be less then or equal to PAGE_SIZE.
4153  */
4154 void
4155 vm_page_set_validclean(vm_page_t m, int base, int size)
4156 {
4157 	vm_page_bits_t oldvalid, pagebits;
4158 	int endoff, frag;
4159 
4160 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4161 	if (size == 0)	/* handle degenerate case */
4162 		return;
4163 
4164 	/*
4165 	 * If the base is not DEV_BSIZE aligned and the valid
4166 	 * bit is clear, we have to zero out a portion of the
4167 	 * first block.
4168 	 */
4169 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4170 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4171 		pmap_zero_page_area(m, frag, base - frag);
4172 
4173 	/*
4174 	 * If the ending offset is not DEV_BSIZE aligned and the
4175 	 * valid bit is clear, we have to zero out a portion of
4176 	 * the last block.
4177 	 */
4178 	endoff = base + size;
4179 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4180 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4181 		pmap_zero_page_area(m, endoff,
4182 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4183 
4184 	/*
4185 	 * Set valid, clear dirty bits.  If validating the entire
4186 	 * page we can safely clear the pmap modify bit.  We also
4187 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
4188 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4189 	 * be set again.
4190 	 *
4191 	 * We set valid bits inclusive of any overlap, but we can only
4192 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4193 	 * the range.
4194 	 */
4195 	oldvalid = m->valid;
4196 	pagebits = vm_page_bits(base, size);
4197 	m->valid |= pagebits;
4198 #if 0	/* NOT YET */
4199 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4200 		frag = DEV_BSIZE - frag;
4201 		base += frag;
4202 		size -= frag;
4203 		if (size < 0)
4204 			size = 0;
4205 	}
4206 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4207 #endif
4208 	if (base == 0 && size == PAGE_SIZE) {
4209 		/*
4210 		 * The page can only be modified within the pmap if it is
4211 		 * mapped, and it can only be mapped if it was previously
4212 		 * fully valid.
4213 		 */
4214 		if (oldvalid == VM_PAGE_BITS_ALL)
4215 			/*
4216 			 * Perform the pmap_clear_modify() first.  Otherwise,
4217 			 * a concurrent pmap operation, such as
4218 			 * pmap_protect(), could clear a modification in the
4219 			 * pmap and set the dirty field on the page before
4220 			 * pmap_clear_modify() had begun and after the dirty
4221 			 * field was cleared here.
4222 			 */
4223 			pmap_clear_modify(m);
4224 		m->dirty = 0;
4225 		m->oflags &= ~VPO_NOSYNC;
4226 	} else if (oldvalid != VM_PAGE_BITS_ALL)
4227 		m->dirty &= ~pagebits;
4228 	else
4229 		vm_page_clear_dirty_mask(m, pagebits);
4230 }
4231 
4232 void
4233 vm_page_clear_dirty(vm_page_t m, int base, int size)
4234 {
4235 
4236 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4237 }
4238 
4239 /*
4240  *	vm_page_set_invalid:
4241  *
4242  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4243  *	valid and dirty bits for the effected areas are cleared.
4244  */
4245 void
4246 vm_page_set_invalid(vm_page_t m, int base, int size)
4247 {
4248 	vm_page_bits_t bits;
4249 	vm_object_t object;
4250 
4251 	object = m->object;
4252 	VM_OBJECT_ASSERT_WLOCKED(object);
4253 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4254 	    size >= object->un_pager.vnp.vnp_size)
4255 		bits = VM_PAGE_BITS_ALL;
4256 	else
4257 		bits = vm_page_bits(base, size);
4258 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4259 	    bits != 0)
4260 		pmap_remove_all(m);
4261 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4262 	    !pmap_page_is_mapped(m),
4263 	    ("vm_page_set_invalid: page %p is mapped", m));
4264 	m->valid &= ~bits;
4265 	m->dirty &= ~bits;
4266 }
4267 
4268 /*
4269  * vm_page_zero_invalid()
4270  *
4271  *	The kernel assumes that the invalid portions of a page contain
4272  *	garbage, but such pages can be mapped into memory by user code.
4273  *	When this occurs, we must zero out the non-valid portions of the
4274  *	page so user code sees what it expects.
4275  *
4276  *	Pages are most often semi-valid when the end of a file is mapped
4277  *	into memory and the file's size is not page aligned.
4278  */
4279 void
4280 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4281 {
4282 	int b;
4283 	int i;
4284 
4285 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4286 	/*
4287 	 * Scan the valid bits looking for invalid sections that
4288 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4289 	 * valid bit may be set ) have already been zeroed by
4290 	 * vm_page_set_validclean().
4291 	 */
4292 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4293 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4294 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4295 			if (i > b) {
4296 				pmap_zero_page_area(m,
4297 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4298 			}
4299 			b = i + 1;
4300 		}
4301 	}
4302 
4303 	/*
4304 	 * setvalid is TRUE when we can safely set the zero'd areas
4305 	 * as being valid.  We can do this if there are no cache consistancy
4306 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4307 	 */
4308 	if (setvalid)
4309 		m->valid = VM_PAGE_BITS_ALL;
4310 }
4311 
4312 /*
4313  *	vm_page_is_valid:
4314  *
4315  *	Is (partial) page valid?  Note that the case where size == 0
4316  *	will return FALSE in the degenerate case where the page is
4317  *	entirely invalid, and TRUE otherwise.
4318  */
4319 int
4320 vm_page_is_valid(vm_page_t m, int base, int size)
4321 {
4322 	vm_page_bits_t bits;
4323 
4324 	VM_OBJECT_ASSERT_LOCKED(m->object);
4325 	bits = vm_page_bits(base, size);
4326 	return (m->valid != 0 && (m->valid & bits) == bits);
4327 }
4328 
4329 /*
4330  * Returns true if all of the specified predicates are true for the entire
4331  * (super)page and false otherwise.
4332  */
4333 bool
4334 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4335 {
4336 	vm_object_t object;
4337 	int i, npages;
4338 
4339 	object = m->object;
4340 	if (skip_m != NULL && skip_m->object != object)
4341 		return (false);
4342 	VM_OBJECT_ASSERT_LOCKED(object);
4343 	npages = atop(pagesizes[m->psind]);
4344 
4345 	/*
4346 	 * The physically contiguous pages that make up a superpage, i.e., a
4347 	 * page with a page size index ("psind") greater than zero, will
4348 	 * occupy adjacent entries in vm_page_array[].
4349 	 */
4350 	for (i = 0; i < npages; i++) {
4351 		/* Always test object consistency, including "skip_m". */
4352 		if (m[i].object != object)
4353 			return (false);
4354 		if (&m[i] == skip_m)
4355 			continue;
4356 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4357 			return (false);
4358 		if ((flags & PS_ALL_DIRTY) != 0) {
4359 			/*
4360 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4361 			 * might stop this case from spuriously returning
4362 			 * "false".  However, that would require a write lock
4363 			 * on the object containing "m[i]".
4364 			 */
4365 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4366 				return (false);
4367 		}
4368 		if ((flags & PS_ALL_VALID) != 0 &&
4369 		    m[i].valid != VM_PAGE_BITS_ALL)
4370 			return (false);
4371 	}
4372 	return (true);
4373 }
4374 
4375 /*
4376  * Set the page's dirty bits if the page is modified.
4377  */
4378 void
4379 vm_page_test_dirty(vm_page_t m)
4380 {
4381 
4382 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4383 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4384 		vm_page_dirty(m);
4385 }
4386 
4387 void
4388 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4389 {
4390 
4391 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4392 }
4393 
4394 void
4395 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4396 {
4397 
4398 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4399 }
4400 
4401 int
4402 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4403 {
4404 
4405 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4406 }
4407 
4408 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4409 void
4410 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4411 {
4412 
4413 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4414 }
4415 
4416 void
4417 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4418 {
4419 
4420 	mtx_assert_(vm_page_lockptr(m), a, file, line);
4421 }
4422 #endif
4423 
4424 #ifdef INVARIANTS
4425 void
4426 vm_page_object_lock_assert(vm_page_t m)
4427 {
4428 
4429 	/*
4430 	 * Certain of the page's fields may only be modified by the
4431 	 * holder of the containing object's lock or the exclusive busy.
4432 	 * holder.  Unfortunately, the holder of the write busy is
4433 	 * not recorded, and thus cannot be checked here.
4434 	 */
4435 	if (m->object != NULL && !vm_page_xbusied(m))
4436 		VM_OBJECT_ASSERT_WLOCKED(m->object);
4437 }
4438 
4439 void
4440 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4441 {
4442 
4443 	if ((bits & PGA_WRITEABLE) == 0)
4444 		return;
4445 
4446 	/*
4447 	 * The PGA_WRITEABLE flag can only be set if the page is
4448 	 * managed, is exclusively busied or the object is locked.
4449 	 * Currently, this flag is only set by pmap_enter().
4450 	 */
4451 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4452 	    ("PGA_WRITEABLE on unmanaged page"));
4453 	if (!vm_page_xbusied(m))
4454 		VM_OBJECT_ASSERT_LOCKED(m->object);
4455 }
4456 #endif
4457 
4458 #include "opt_ddb.h"
4459 #ifdef DDB
4460 #include <sys/kernel.h>
4461 
4462 #include <ddb/ddb.h>
4463 
4464 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4465 {
4466 
4467 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4468 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4469 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4470 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4471 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4472 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4473 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4474 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4475 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4476 }
4477 
4478 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4479 {
4480 	int dom;
4481 
4482 	db_printf("pq_free %d\n", vm_free_count());
4483 	for (dom = 0; dom < vm_ndomains; dom++) {
4484 		db_printf(
4485     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4486 		    dom,
4487 		    vm_dom[dom].vmd_page_count,
4488 		    vm_dom[dom].vmd_free_count,
4489 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4490 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4491 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4492 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4493 	}
4494 }
4495 
4496 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4497 {
4498 	vm_page_t m;
4499 	boolean_t phys, virt;
4500 
4501 	if (!have_addr) {
4502 		db_printf("show pginfo addr\n");
4503 		return;
4504 	}
4505 
4506 	phys = strchr(modif, 'p') != NULL;
4507 	virt = strchr(modif, 'v') != NULL;
4508 	if (virt)
4509 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
4510 	else if (phys)
4511 		m = PHYS_TO_VM_PAGE(addr);
4512 	else
4513 		m = (vm_page_t)addr;
4514 	db_printf(
4515     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4516     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4517 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4518 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4519 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4520 }
4521 #endif /* DDB */
4522