xref: /freebsd/sys/vm/vm_page.c (revision 00e9473388d3f7ac84d8e51829b532ca476ab14e)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *			GENERAL RULES ON VM_PAGE MANIPULATION
67  *
68  *	- A page queue lock is required when adding or removing a page from a
69  *	  page queue regardless of other locks or the busy state of a page.
70  *
71  *		* In general, no thread besides the page daemon can acquire or
72  *		  hold more than one page queue lock at a time.
73  *
74  *		* The page daemon can acquire and hold any pair of page queue
75  *		  locks in any order.
76  *
77  *	- The object lock is required when inserting or removing
78  *	  pages from an object (vm_page_insert() or vm_page_remove()).
79  *
80  */
81 
82 /*
83  *	Resident memory management module.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 #include "opt_vm.h"
90 
91 #include <sys/param.h>
92 #include <sys/systm.h>
93 #include <sys/lock.h>
94 #include <sys/domainset.h>
95 #include <sys/kernel.h>
96 #include <sys/limits.h>
97 #include <sys/linker.h>
98 #include <sys/malloc.h>
99 #include <sys/mman.h>
100 #include <sys/msgbuf.h>
101 #include <sys/mutex.h>
102 #include <sys/proc.h>
103 #include <sys/rwlock.h>
104 #include <sys/sbuf.h>
105 #include <sys/sched.h>
106 #include <sys/smp.h>
107 #include <sys/sysctl.h>
108 #include <sys/vmmeter.h>
109 #include <sys/vnode.h>
110 
111 #include <vm/vm.h>
112 #include <vm/pmap.h>
113 #include <vm/vm_param.h>
114 #include <vm/vm_domainset.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_map.h>
117 #include <vm/vm_object.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_pageout.h>
120 #include <vm/vm_phys.h>
121 #include <vm/vm_pagequeue.h>
122 #include <vm/vm_pager.h>
123 #include <vm/vm_radix.h>
124 #include <vm/vm_reserv.h>
125 #include <vm/vm_extern.h>
126 #include <vm/uma.h>
127 #include <vm/uma_int.h>
128 
129 #include <machine/md_var.h>
130 
131 extern int	uma_startup_count(int);
132 extern void	uma_startup(void *, int);
133 extern int	vmem_startup_count(void);
134 
135 struct vm_domain vm_dom[MAXMEMDOM];
136 
137 static DPCPU_DEFINE(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
138 
139 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
140 
141 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
142 /* The following fields are protected by the domainset lock. */
143 domainset_t __exclusive_cache_line vm_min_domains;
144 domainset_t __exclusive_cache_line vm_severe_domains;
145 static int vm_min_waiters;
146 static int vm_severe_waiters;
147 static int vm_pageproc_waiters;
148 
149 /*
150  * bogus page -- for I/O to/from partially complete buffers,
151  * or for paging into sparsely invalid regions.
152  */
153 vm_page_t bogus_page;
154 
155 vm_page_t vm_page_array;
156 long vm_page_array_size;
157 long first_page;
158 
159 static int boot_pages;
160 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
161     &boot_pages, 0,
162     "number of pages allocated for bootstrapping the VM system");
163 
164 static int pa_tryrelock_restart;
165 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
166     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
167 
168 static TAILQ_HEAD(, vm_page) blacklist_head;
169 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
170 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
171     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
172 
173 static uma_zone_t fakepg_zone;
174 
175 static void vm_page_alloc_check(vm_page_t m);
176 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
177 static void vm_page_dequeue_complete(vm_page_t m);
178 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
179 static void vm_page_init(void *dummy);
180 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
181     vm_pindex_t pindex, vm_page_t mpred);
182 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
183     vm_page_t mpred);
184 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
185     vm_page_t m_run, vm_paddr_t high);
186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
187     int req);
188 static int vm_page_import(void *arg, void **store, int cnt, int domain,
189     int flags);
190 static void vm_page_release(void *arg, void **store, int cnt);
191 
192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
193 
194 static void
195 vm_page_init(void *dummy)
196 {
197 
198 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
199 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
200 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
201 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
202 }
203 
204 /*
205  * The cache page zone is initialized later since we need to be able to allocate
206  * pages before UMA is fully initialized.
207  */
208 static void
209 vm_page_init_cache_zones(void *dummy __unused)
210 {
211 	struct vm_domain *vmd;
212 	int i;
213 
214 	for (i = 0; i < vm_ndomains; i++) {
215 		vmd = VM_DOMAIN(i);
216 		/*
217 		 * Don't allow the page cache to take up more than .25% of
218 		 * memory.
219 		 */
220 		if (vmd->vmd_page_count / 400 < 256 * mp_ncpus)
221 			continue;
222 		vmd->vmd_pgcache = uma_zcache_create("vm pgcache",
223 		    sizeof(struct vm_page), NULL, NULL, NULL, NULL,
224 		    vm_page_import, vm_page_release, vmd,
225 		    UMA_ZONE_NOBUCKETCACHE | UMA_ZONE_MAXBUCKET | UMA_ZONE_VM);
226 	}
227 }
228 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
229 
230 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
231 #if PAGE_SIZE == 32768
232 #ifdef CTASSERT
233 CTASSERT(sizeof(u_long) >= 8);
234 #endif
235 #endif
236 
237 /*
238  * Try to acquire a physical address lock while a pmap is locked.  If we
239  * fail to trylock we unlock and lock the pmap directly and cache the
240  * locked pa in *locked.  The caller should then restart their loop in case
241  * the virtual to physical mapping has changed.
242  */
243 int
244 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
245 {
246 	vm_paddr_t lockpa;
247 
248 	lockpa = *locked;
249 	*locked = pa;
250 	if (lockpa) {
251 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
252 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
253 			return (0);
254 		PA_UNLOCK(lockpa);
255 	}
256 	if (PA_TRYLOCK(pa))
257 		return (0);
258 	PMAP_UNLOCK(pmap);
259 	atomic_add_int(&pa_tryrelock_restart, 1);
260 	PA_LOCK(pa);
261 	PMAP_LOCK(pmap);
262 	return (EAGAIN);
263 }
264 
265 /*
266  *	vm_set_page_size:
267  *
268  *	Sets the page size, perhaps based upon the memory
269  *	size.  Must be called before any use of page-size
270  *	dependent functions.
271  */
272 void
273 vm_set_page_size(void)
274 {
275 	if (vm_cnt.v_page_size == 0)
276 		vm_cnt.v_page_size = PAGE_SIZE;
277 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
278 		panic("vm_set_page_size: page size not a power of two");
279 }
280 
281 /*
282  *	vm_page_blacklist_next:
283  *
284  *	Find the next entry in the provided string of blacklist
285  *	addresses.  Entries are separated by space, comma, or newline.
286  *	If an invalid integer is encountered then the rest of the
287  *	string is skipped.  Updates the list pointer to the next
288  *	character, or NULL if the string is exhausted or invalid.
289  */
290 static vm_paddr_t
291 vm_page_blacklist_next(char **list, char *end)
292 {
293 	vm_paddr_t bad;
294 	char *cp, *pos;
295 
296 	if (list == NULL || *list == NULL)
297 		return (0);
298 	if (**list =='\0') {
299 		*list = NULL;
300 		return (0);
301 	}
302 
303 	/*
304 	 * If there's no end pointer then the buffer is coming from
305 	 * the kenv and we know it's null-terminated.
306 	 */
307 	if (end == NULL)
308 		end = *list + strlen(*list);
309 
310 	/* Ensure that strtoq() won't walk off the end */
311 	if (*end != '\0') {
312 		if (*end == '\n' || *end == ' ' || *end  == ',')
313 			*end = '\0';
314 		else {
315 			printf("Blacklist not terminated, skipping\n");
316 			*list = NULL;
317 			return (0);
318 		}
319 	}
320 
321 	for (pos = *list; *pos != '\0'; pos = cp) {
322 		bad = strtoq(pos, &cp, 0);
323 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
324 			if (bad == 0) {
325 				if (++cp < end)
326 					continue;
327 				else
328 					break;
329 			}
330 		} else
331 			break;
332 		if (*cp == '\0' || ++cp >= end)
333 			*list = NULL;
334 		else
335 			*list = cp;
336 		return (trunc_page(bad));
337 	}
338 	printf("Garbage in RAM blacklist, skipping\n");
339 	*list = NULL;
340 	return (0);
341 }
342 
343 bool
344 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
345 {
346 	struct vm_domain *vmd;
347 	vm_page_t m;
348 	int ret;
349 
350 	m = vm_phys_paddr_to_vm_page(pa);
351 	if (m == NULL)
352 		return (true); /* page does not exist, no failure */
353 
354 	vmd = vm_pagequeue_domain(m);
355 	vm_domain_free_lock(vmd);
356 	ret = vm_phys_unfree_page(m);
357 	vm_domain_free_unlock(vmd);
358 	if (ret) {
359 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
360 		if (verbose)
361 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
362 	}
363 	return (ret);
364 }
365 
366 /*
367  *	vm_page_blacklist_check:
368  *
369  *	Iterate through the provided string of blacklist addresses, pulling
370  *	each entry out of the physical allocator free list and putting it
371  *	onto a list for reporting via the vm.page_blacklist sysctl.
372  */
373 static void
374 vm_page_blacklist_check(char *list, char *end)
375 {
376 	vm_paddr_t pa;
377 	char *next;
378 
379 	next = list;
380 	while (next != NULL) {
381 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
382 			continue;
383 		vm_page_blacklist_add(pa, bootverbose);
384 	}
385 }
386 
387 /*
388  *	vm_page_blacklist_load:
389  *
390  *	Search for a special module named "ram_blacklist".  It'll be a
391  *	plain text file provided by the user via the loader directive
392  *	of the same name.
393  */
394 static void
395 vm_page_blacklist_load(char **list, char **end)
396 {
397 	void *mod;
398 	u_char *ptr;
399 	u_int len;
400 
401 	mod = NULL;
402 	ptr = NULL;
403 
404 	mod = preload_search_by_type("ram_blacklist");
405 	if (mod != NULL) {
406 		ptr = preload_fetch_addr(mod);
407 		len = preload_fetch_size(mod);
408         }
409 	*list = ptr;
410 	if (ptr != NULL)
411 		*end = ptr + len;
412 	else
413 		*end = NULL;
414 	return;
415 }
416 
417 static int
418 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
419 {
420 	vm_page_t m;
421 	struct sbuf sbuf;
422 	int error, first;
423 
424 	first = 1;
425 	error = sysctl_wire_old_buffer(req, 0);
426 	if (error != 0)
427 		return (error);
428 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
429 	TAILQ_FOREACH(m, &blacklist_head, listq) {
430 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
431 		    (uintmax_t)m->phys_addr);
432 		first = 0;
433 	}
434 	error = sbuf_finish(&sbuf);
435 	sbuf_delete(&sbuf);
436 	return (error);
437 }
438 
439 /*
440  * Initialize a dummy page for use in scans of the specified paging queue.
441  * In principle, this function only needs to set the flag PG_MARKER.
442  * Nonetheless, it write busies and initializes the hold count to one as
443  * safety precautions.
444  */
445 static void
446 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
447 {
448 
449 	bzero(marker, sizeof(*marker));
450 	marker->flags = PG_MARKER;
451 	marker->aflags = aflags;
452 	marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
453 	marker->queue = queue;
454 	marker->hold_count = 1;
455 }
456 
457 static void
458 vm_page_domain_init(int domain)
459 {
460 	struct vm_domain *vmd;
461 	struct vm_pagequeue *pq;
462 	int i;
463 
464 	vmd = VM_DOMAIN(domain);
465 	bzero(vmd, sizeof(*vmd));
466 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
467 	    "vm inactive pagequeue";
468 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
469 	    "vm active pagequeue";
470 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
471 	    "vm laundry pagequeue";
472 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
473 	    "vm unswappable pagequeue";
474 	vmd->vmd_domain = domain;
475 	vmd->vmd_page_count = 0;
476 	vmd->vmd_free_count = 0;
477 	vmd->vmd_segs = 0;
478 	vmd->vmd_oom = FALSE;
479 	for (i = 0; i < PQ_COUNT; i++) {
480 		pq = &vmd->vmd_pagequeues[i];
481 		TAILQ_INIT(&pq->pq_pl);
482 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
483 		    MTX_DEF | MTX_DUPOK);
484 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
485 	}
486 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
487 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
488 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
489 
490 	/*
491 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
492 	 * insertions.
493 	 */
494 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
495 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
496 	    &vmd->vmd_inacthead, plinks.q);
497 
498 	/*
499 	 * The clock pages are used to implement active queue scanning without
500 	 * requeues.  Scans start at clock[0], which is advanced after the scan
501 	 * ends.  When the two clock hands meet, they are reset and scanning
502 	 * resumes from the head of the queue.
503 	 */
504 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
505 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
506 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
507 	    &vmd->vmd_clock[0], plinks.q);
508 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
509 	    &vmd->vmd_clock[1], plinks.q);
510 }
511 
512 /*
513  * Initialize a physical page in preparation for adding it to the free
514  * lists.
515  */
516 static void
517 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
518 {
519 
520 	m->object = NULL;
521 	m->wire_count = 0;
522 	m->busy_lock = VPB_UNBUSIED;
523 	m->hold_count = 0;
524 	m->flags = 0;
525 	m->phys_addr = pa;
526 	m->queue = PQ_NONE;
527 	m->psind = 0;
528 	m->segind = segind;
529 	m->order = VM_NFREEORDER;
530 	m->pool = VM_FREEPOOL_DEFAULT;
531 	m->valid = m->dirty = 0;
532 	pmap_page_init(m);
533 }
534 
535 /*
536  *	vm_page_startup:
537  *
538  *	Initializes the resident memory module.  Allocates physical memory for
539  *	bootstrapping UMA and some data structures that are used to manage
540  *	physical pages.  Initializes these structures, and populates the free
541  *	page queues.
542  */
543 vm_offset_t
544 vm_page_startup(vm_offset_t vaddr)
545 {
546 	struct vm_phys_seg *seg;
547 	vm_page_t m;
548 	char *list, *listend;
549 	vm_offset_t mapped;
550 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
551 	vm_paddr_t biggestsize, last_pa, pa;
552 	u_long pagecount;
553 	int biggestone, i, segind;
554 
555 	biggestsize = 0;
556 	biggestone = 0;
557 	vaddr = round_page(vaddr);
558 
559 	for (i = 0; phys_avail[i + 1]; i += 2) {
560 		phys_avail[i] = round_page(phys_avail[i]);
561 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
562 	}
563 	for (i = 0; phys_avail[i + 1]; i += 2) {
564 		size = phys_avail[i + 1] - phys_avail[i];
565 		if (size > biggestsize) {
566 			biggestone = i;
567 			biggestsize = size;
568 		}
569 	}
570 
571 	end = phys_avail[biggestone+1];
572 
573 	/*
574 	 * Initialize the page and queue locks.
575 	 */
576 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
577 	for (i = 0; i < PA_LOCK_COUNT; i++)
578 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
579 	for (i = 0; i < vm_ndomains; i++)
580 		vm_page_domain_init(i);
581 
582 	/*
583 	 * Allocate memory for use when boot strapping the kernel memory
584 	 * allocator.  Tell UMA how many zones we are going to create
585 	 * before going fully functional.  UMA will add its zones.
586 	 *
587 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
588 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
589 	 */
590 	boot_pages = uma_startup_count(8);
591 
592 #ifndef UMA_MD_SMALL_ALLOC
593 	/* vmem_startup() calls uma_prealloc(). */
594 	boot_pages += vmem_startup_count();
595 	/* vm_map_startup() calls uma_prealloc(). */
596 	boot_pages += howmany(MAX_KMAP,
597 	    UMA_SLAB_SPACE / sizeof(struct vm_map));
598 
599 	/*
600 	 * Before going fully functional kmem_init() does allocation
601 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
602 	 */
603 	boot_pages += 2;
604 #endif
605 	/*
606 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
607 	 * manually fetch the value.
608 	 */
609 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
610 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
611 	new_end = trunc_page(new_end);
612 	mapped = pmap_map(&vaddr, new_end, end,
613 	    VM_PROT_READ | VM_PROT_WRITE);
614 	bzero((void *)mapped, end - new_end);
615 	uma_startup((void *)mapped, boot_pages);
616 
617 #ifdef WITNESS
618 	end = new_end;
619 	new_end = end - round_page(witness_startup_count());
620 	mapped = pmap_map(&vaddr, new_end, end,
621 	    VM_PROT_READ | VM_PROT_WRITE);
622 	bzero((void *)mapped, end - new_end);
623 	witness_startup((void *)mapped);
624 #endif
625 
626 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
627     defined(__i386__) || defined(__mips__)
628 	/*
629 	 * Allocate a bitmap to indicate that a random physical page
630 	 * needs to be included in a minidump.
631 	 *
632 	 * The amd64 port needs this to indicate which direct map pages
633 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
634 	 *
635 	 * However, i386 still needs this workspace internally within the
636 	 * minidump code.  In theory, they are not needed on i386, but are
637 	 * included should the sf_buf code decide to use them.
638 	 */
639 	last_pa = 0;
640 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
641 		if (dump_avail[i + 1] > last_pa)
642 			last_pa = dump_avail[i + 1];
643 	page_range = last_pa / PAGE_SIZE;
644 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
645 	new_end -= vm_page_dump_size;
646 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
647 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
648 	bzero((void *)vm_page_dump, vm_page_dump_size);
649 #else
650 	(void)last_pa;
651 #endif
652 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
653 	/*
654 	 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
655 	 * When pmap_map() uses the direct map, they are not automatically
656 	 * included.
657 	 */
658 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
659 		dump_add_page(pa);
660 #endif
661 	phys_avail[biggestone + 1] = new_end;
662 #ifdef __amd64__
663 	/*
664 	 * Request that the physical pages underlying the message buffer be
665 	 * included in a crash dump.  Since the message buffer is accessed
666 	 * through the direct map, they are not automatically included.
667 	 */
668 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
669 	last_pa = pa + round_page(msgbufsize);
670 	while (pa < last_pa) {
671 		dump_add_page(pa);
672 		pa += PAGE_SIZE;
673 	}
674 #endif
675 	/*
676 	 * Compute the number of pages of memory that will be available for
677 	 * use, taking into account the overhead of a page structure per page.
678 	 * In other words, solve
679 	 *	"available physical memory" - round_page(page_range *
680 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
681 	 * for page_range.
682 	 */
683 	low_avail = phys_avail[0];
684 	high_avail = phys_avail[1];
685 	for (i = 0; i < vm_phys_nsegs; i++) {
686 		if (vm_phys_segs[i].start < low_avail)
687 			low_avail = vm_phys_segs[i].start;
688 		if (vm_phys_segs[i].end > high_avail)
689 			high_avail = vm_phys_segs[i].end;
690 	}
691 	/* Skip the first chunk.  It is already accounted for. */
692 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
693 		if (phys_avail[i] < low_avail)
694 			low_avail = phys_avail[i];
695 		if (phys_avail[i + 1] > high_avail)
696 			high_avail = phys_avail[i + 1];
697 	}
698 	first_page = low_avail / PAGE_SIZE;
699 #ifdef VM_PHYSSEG_SPARSE
700 	size = 0;
701 	for (i = 0; i < vm_phys_nsegs; i++)
702 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
703 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
704 		size += phys_avail[i + 1] - phys_avail[i];
705 #elif defined(VM_PHYSSEG_DENSE)
706 	size = high_avail - low_avail;
707 #else
708 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
709 #endif
710 
711 #ifdef VM_PHYSSEG_DENSE
712 	/*
713 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
714 	 * the overhead of a page structure per page only if vm_page_array is
715 	 * allocated from the last physical memory chunk.  Otherwise, we must
716 	 * allocate page structures representing the physical memory
717 	 * underlying vm_page_array, even though they will not be used.
718 	 */
719 	if (new_end != high_avail)
720 		page_range = size / PAGE_SIZE;
721 	else
722 #endif
723 	{
724 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
725 
726 		/*
727 		 * If the partial bytes remaining are large enough for
728 		 * a page (PAGE_SIZE) without a corresponding
729 		 * 'struct vm_page', then new_end will contain an
730 		 * extra page after subtracting the length of the VM
731 		 * page array.  Compensate by subtracting an extra
732 		 * page from new_end.
733 		 */
734 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
735 			if (new_end == high_avail)
736 				high_avail -= PAGE_SIZE;
737 			new_end -= PAGE_SIZE;
738 		}
739 	}
740 	end = new_end;
741 
742 	/*
743 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
744 	 * However, because this page is allocated from KVM, out-of-bounds
745 	 * accesses using the direct map will not be trapped.
746 	 */
747 	vaddr += PAGE_SIZE;
748 
749 	/*
750 	 * Allocate physical memory for the page structures, and map it.
751 	 */
752 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
753 	mapped = pmap_map(&vaddr, new_end, end,
754 	    VM_PROT_READ | VM_PROT_WRITE);
755 	vm_page_array = (vm_page_t)mapped;
756 	vm_page_array_size = page_range;
757 
758 #if VM_NRESERVLEVEL > 0
759 	/*
760 	 * Allocate physical memory for the reservation management system's
761 	 * data structures, and map it.
762 	 */
763 	if (high_avail == end)
764 		high_avail = new_end;
765 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
766 #endif
767 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
768 	/*
769 	 * Include vm_page_array and vm_reserv_array in a crash dump.
770 	 */
771 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
772 		dump_add_page(pa);
773 #endif
774 	phys_avail[biggestone + 1] = new_end;
775 
776 	/*
777 	 * Add physical memory segments corresponding to the available
778 	 * physical pages.
779 	 */
780 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
781 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
782 
783 	/*
784 	 * Initialize the physical memory allocator.
785 	 */
786 	vm_phys_init();
787 
788 	/*
789 	 * Initialize the page structures and add every available page to the
790 	 * physical memory allocator's free lists.
791 	 */
792 	vm_cnt.v_page_count = 0;
793 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
794 		seg = &vm_phys_segs[segind];
795 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
796 		    m++, pa += PAGE_SIZE)
797 			vm_page_init_page(m, pa, segind);
798 
799 		/*
800 		 * Add the segment to the free lists only if it is covered by
801 		 * one of the ranges in phys_avail.  Because we've added the
802 		 * ranges to the vm_phys_segs array, we can assume that each
803 		 * segment is either entirely contained in one of the ranges,
804 		 * or doesn't overlap any of them.
805 		 */
806 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
807 			struct vm_domain *vmd;
808 
809 			if (seg->start < phys_avail[i] ||
810 			    seg->end > phys_avail[i + 1])
811 				continue;
812 
813 			m = seg->first_page;
814 			pagecount = (u_long)atop(seg->end - seg->start);
815 
816 			vmd = VM_DOMAIN(seg->domain);
817 			vm_domain_free_lock(vmd);
818 			vm_phys_free_contig(m, pagecount);
819 			vm_domain_free_unlock(vmd);
820 			vm_domain_freecnt_inc(vmd, pagecount);
821 			vm_cnt.v_page_count += (u_int)pagecount;
822 
823 			vmd = VM_DOMAIN(seg->domain);
824 			vmd->vmd_page_count += (u_int)pagecount;
825 			vmd->vmd_segs |= 1UL << m->segind;
826 			break;
827 		}
828 	}
829 
830 	/*
831 	 * Remove blacklisted pages from the physical memory allocator.
832 	 */
833 	TAILQ_INIT(&blacklist_head);
834 	vm_page_blacklist_load(&list, &listend);
835 	vm_page_blacklist_check(list, listend);
836 
837 	list = kern_getenv("vm.blacklist");
838 	vm_page_blacklist_check(list, NULL);
839 
840 	freeenv(list);
841 #if VM_NRESERVLEVEL > 0
842 	/*
843 	 * Initialize the reservation management system.
844 	 */
845 	vm_reserv_init();
846 #endif
847 	/*
848 	 * Set an initial domain policy for thread0 so that allocations
849 	 * can work.
850 	 */
851 	domainset_zero();
852 
853 	return (vaddr);
854 }
855 
856 void
857 vm_page_reference(vm_page_t m)
858 {
859 
860 	vm_page_aflag_set(m, PGA_REFERENCED);
861 }
862 
863 /*
864  *	vm_page_busy_downgrade:
865  *
866  *	Downgrade an exclusive busy page into a single shared busy page.
867  */
868 void
869 vm_page_busy_downgrade(vm_page_t m)
870 {
871 	u_int x;
872 	bool locked;
873 
874 	vm_page_assert_xbusied(m);
875 	locked = mtx_owned(vm_page_lockptr(m));
876 
877 	for (;;) {
878 		x = m->busy_lock;
879 		x &= VPB_BIT_WAITERS;
880 		if (x != 0 && !locked)
881 			vm_page_lock(m);
882 		if (atomic_cmpset_rel_int(&m->busy_lock,
883 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
884 			break;
885 		if (x != 0 && !locked)
886 			vm_page_unlock(m);
887 	}
888 	if (x != 0) {
889 		wakeup(m);
890 		if (!locked)
891 			vm_page_unlock(m);
892 	}
893 }
894 
895 /*
896  *	vm_page_sbusied:
897  *
898  *	Return a positive value if the page is shared busied, 0 otherwise.
899  */
900 int
901 vm_page_sbusied(vm_page_t m)
902 {
903 	u_int x;
904 
905 	x = m->busy_lock;
906 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
907 }
908 
909 /*
910  *	vm_page_sunbusy:
911  *
912  *	Shared unbusy a page.
913  */
914 void
915 vm_page_sunbusy(vm_page_t m)
916 {
917 	u_int x;
918 
919 	vm_page_lock_assert(m, MA_NOTOWNED);
920 	vm_page_assert_sbusied(m);
921 
922 	for (;;) {
923 		x = m->busy_lock;
924 		if (VPB_SHARERS(x) > 1) {
925 			if (atomic_cmpset_int(&m->busy_lock, x,
926 			    x - VPB_ONE_SHARER))
927 				break;
928 			continue;
929 		}
930 		if ((x & VPB_BIT_WAITERS) == 0) {
931 			KASSERT(x == VPB_SHARERS_WORD(1),
932 			    ("vm_page_sunbusy: invalid lock state"));
933 			if (atomic_cmpset_int(&m->busy_lock,
934 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
935 				break;
936 			continue;
937 		}
938 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
939 		    ("vm_page_sunbusy: invalid lock state for waiters"));
940 
941 		vm_page_lock(m);
942 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
943 			vm_page_unlock(m);
944 			continue;
945 		}
946 		wakeup(m);
947 		vm_page_unlock(m);
948 		break;
949 	}
950 }
951 
952 /*
953  *	vm_page_busy_sleep:
954  *
955  *	Sleep and release the page lock, using the page pointer as wchan.
956  *	This is used to implement the hard-path of busying mechanism.
957  *
958  *	The given page must be locked.
959  *
960  *	If nonshared is true, sleep only if the page is xbusy.
961  */
962 void
963 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
964 {
965 	u_int x;
966 
967 	vm_page_assert_locked(m);
968 
969 	x = m->busy_lock;
970 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
971 	    ((x & VPB_BIT_WAITERS) == 0 &&
972 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
973 		vm_page_unlock(m);
974 		return;
975 	}
976 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
977 }
978 
979 /*
980  *	vm_page_trysbusy:
981  *
982  *	Try to shared busy a page.
983  *	If the operation succeeds 1 is returned otherwise 0.
984  *	The operation never sleeps.
985  */
986 int
987 vm_page_trysbusy(vm_page_t m)
988 {
989 	u_int x;
990 
991 	for (;;) {
992 		x = m->busy_lock;
993 		if ((x & VPB_BIT_SHARED) == 0)
994 			return (0);
995 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
996 			return (1);
997 	}
998 }
999 
1000 static void
1001 vm_page_xunbusy_locked(vm_page_t m)
1002 {
1003 
1004 	vm_page_assert_xbusied(m);
1005 	vm_page_assert_locked(m);
1006 
1007 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1008 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
1009 	wakeup(m);
1010 }
1011 
1012 void
1013 vm_page_xunbusy_maybelocked(vm_page_t m)
1014 {
1015 	bool lockacq;
1016 
1017 	vm_page_assert_xbusied(m);
1018 
1019 	/*
1020 	 * Fast path for unbusy.  If it succeeds, we know that there
1021 	 * are no waiters, so we do not need a wakeup.
1022 	 */
1023 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
1024 	    VPB_UNBUSIED))
1025 		return;
1026 
1027 	lockacq = !mtx_owned(vm_page_lockptr(m));
1028 	if (lockacq)
1029 		vm_page_lock(m);
1030 	vm_page_xunbusy_locked(m);
1031 	if (lockacq)
1032 		vm_page_unlock(m);
1033 }
1034 
1035 /*
1036  *	vm_page_xunbusy_hard:
1037  *
1038  *	Called after the first try the exclusive unbusy of a page failed.
1039  *	It is assumed that the waiters bit is on.
1040  */
1041 void
1042 vm_page_xunbusy_hard(vm_page_t m)
1043 {
1044 
1045 	vm_page_assert_xbusied(m);
1046 
1047 	vm_page_lock(m);
1048 	vm_page_xunbusy_locked(m);
1049 	vm_page_unlock(m);
1050 }
1051 
1052 /*
1053  *	vm_page_flash:
1054  *
1055  *	Wakeup anyone waiting for the page.
1056  *	The ownership bits do not change.
1057  *
1058  *	The given page must be locked.
1059  */
1060 void
1061 vm_page_flash(vm_page_t m)
1062 {
1063 	u_int x;
1064 
1065 	vm_page_lock_assert(m, MA_OWNED);
1066 
1067 	for (;;) {
1068 		x = m->busy_lock;
1069 		if ((x & VPB_BIT_WAITERS) == 0)
1070 			return;
1071 		if (atomic_cmpset_int(&m->busy_lock, x,
1072 		    x & (~VPB_BIT_WAITERS)))
1073 			break;
1074 	}
1075 	wakeup(m);
1076 }
1077 
1078 /*
1079  * Avoid releasing and reacquiring the same page lock.
1080  */
1081 void
1082 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1083 {
1084 	struct mtx *mtx1;
1085 
1086 	mtx1 = vm_page_lockptr(m);
1087 	if (*mtx == mtx1)
1088 		return;
1089 	if (*mtx != NULL)
1090 		mtx_unlock(*mtx);
1091 	*mtx = mtx1;
1092 	mtx_lock(mtx1);
1093 }
1094 
1095 /*
1096  * Keep page from being freed by the page daemon
1097  * much of the same effect as wiring, except much lower
1098  * overhead and should be used only for *very* temporary
1099  * holding ("wiring").
1100  */
1101 void
1102 vm_page_hold(vm_page_t mem)
1103 {
1104 
1105 	vm_page_lock_assert(mem, MA_OWNED);
1106         mem->hold_count++;
1107 }
1108 
1109 void
1110 vm_page_unhold(vm_page_t mem)
1111 {
1112 
1113 	vm_page_lock_assert(mem, MA_OWNED);
1114 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1115 	--mem->hold_count;
1116 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1117 		vm_page_free_toq(mem);
1118 }
1119 
1120 /*
1121  *	vm_page_unhold_pages:
1122  *
1123  *	Unhold each of the pages that is referenced by the given array.
1124  */
1125 void
1126 vm_page_unhold_pages(vm_page_t *ma, int count)
1127 {
1128 	struct mtx *mtx;
1129 
1130 	mtx = NULL;
1131 	for (; count != 0; count--) {
1132 		vm_page_change_lock(*ma, &mtx);
1133 		vm_page_unhold(*ma);
1134 		ma++;
1135 	}
1136 	if (mtx != NULL)
1137 		mtx_unlock(mtx);
1138 }
1139 
1140 vm_page_t
1141 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1142 {
1143 	vm_page_t m;
1144 
1145 #ifdef VM_PHYSSEG_SPARSE
1146 	m = vm_phys_paddr_to_vm_page(pa);
1147 	if (m == NULL)
1148 		m = vm_phys_fictitious_to_vm_page(pa);
1149 	return (m);
1150 #elif defined(VM_PHYSSEG_DENSE)
1151 	long pi;
1152 
1153 	pi = atop(pa);
1154 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1155 		m = &vm_page_array[pi - first_page];
1156 		return (m);
1157 	}
1158 	return (vm_phys_fictitious_to_vm_page(pa));
1159 #else
1160 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1161 #endif
1162 }
1163 
1164 /*
1165  *	vm_page_getfake:
1166  *
1167  *	Create a fictitious page with the specified physical address and
1168  *	memory attribute.  The memory attribute is the only the machine-
1169  *	dependent aspect of a fictitious page that must be initialized.
1170  */
1171 vm_page_t
1172 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1173 {
1174 	vm_page_t m;
1175 
1176 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1177 	vm_page_initfake(m, paddr, memattr);
1178 	return (m);
1179 }
1180 
1181 void
1182 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1183 {
1184 
1185 	if ((m->flags & PG_FICTITIOUS) != 0) {
1186 		/*
1187 		 * The page's memattr might have changed since the
1188 		 * previous initialization.  Update the pmap to the
1189 		 * new memattr.
1190 		 */
1191 		goto memattr;
1192 	}
1193 	m->phys_addr = paddr;
1194 	m->queue = PQ_NONE;
1195 	/* Fictitious pages don't use "segind". */
1196 	m->flags = PG_FICTITIOUS;
1197 	/* Fictitious pages don't use "order" or "pool". */
1198 	m->oflags = VPO_UNMANAGED;
1199 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1200 	m->wire_count = 1;
1201 	pmap_page_init(m);
1202 memattr:
1203 	pmap_page_set_memattr(m, memattr);
1204 }
1205 
1206 /*
1207  *	vm_page_putfake:
1208  *
1209  *	Release a fictitious page.
1210  */
1211 void
1212 vm_page_putfake(vm_page_t m)
1213 {
1214 
1215 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1216 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1217 	    ("vm_page_putfake: bad page %p", m));
1218 	uma_zfree(fakepg_zone, m);
1219 }
1220 
1221 /*
1222  *	vm_page_updatefake:
1223  *
1224  *	Update the given fictitious page to the specified physical address and
1225  *	memory attribute.
1226  */
1227 void
1228 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1229 {
1230 
1231 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1232 	    ("vm_page_updatefake: bad page %p", m));
1233 	m->phys_addr = paddr;
1234 	pmap_page_set_memattr(m, memattr);
1235 }
1236 
1237 /*
1238  *	vm_page_free:
1239  *
1240  *	Free a page.
1241  */
1242 void
1243 vm_page_free(vm_page_t m)
1244 {
1245 
1246 	m->flags &= ~PG_ZERO;
1247 	vm_page_free_toq(m);
1248 }
1249 
1250 /*
1251  *	vm_page_free_zero:
1252  *
1253  *	Free a page to the zerod-pages queue
1254  */
1255 void
1256 vm_page_free_zero(vm_page_t m)
1257 {
1258 
1259 	m->flags |= PG_ZERO;
1260 	vm_page_free_toq(m);
1261 }
1262 
1263 /*
1264  * Unbusy and handle the page queueing for a page from a getpages request that
1265  * was optionally read ahead or behind.
1266  */
1267 void
1268 vm_page_readahead_finish(vm_page_t m)
1269 {
1270 
1271 	/* We shouldn't put invalid pages on queues. */
1272 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1273 
1274 	/*
1275 	 * Since the page is not the actually needed one, whether it should
1276 	 * be activated or deactivated is not obvious.  Empirical results
1277 	 * have shown that deactivating the page is usually the best choice,
1278 	 * unless the page is wanted by another thread.
1279 	 */
1280 	vm_page_lock(m);
1281 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1282 		vm_page_activate(m);
1283 	else
1284 		vm_page_deactivate(m);
1285 	vm_page_unlock(m);
1286 	vm_page_xunbusy(m);
1287 }
1288 
1289 /*
1290  *	vm_page_sleep_if_busy:
1291  *
1292  *	Sleep and release the page queues lock if the page is busied.
1293  *	Returns TRUE if the thread slept.
1294  *
1295  *	The given page must be unlocked and object containing it must
1296  *	be locked.
1297  */
1298 int
1299 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1300 {
1301 	vm_object_t obj;
1302 
1303 	vm_page_lock_assert(m, MA_NOTOWNED);
1304 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1305 
1306 	if (vm_page_busied(m)) {
1307 		/*
1308 		 * The page-specific object must be cached because page
1309 		 * identity can change during the sleep, causing the
1310 		 * re-lock of a different object.
1311 		 * It is assumed that a reference to the object is already
1312 		 * held by the callers.
1313 		 */
1314 		obj = m->object;
1315 		vm_page_lock(m);
1316 		VM_OBJECT_WUNLOCK(obj);
1317 		vm_page_busy_sleep(m, msg, false);
1318 		VM_OBJECT_WLOCK(obj);
1319 		return (TRUE);
1320 	}
1321 	return (FALSE);
1322 }
1323 
1324 /*
1325  *	vm_page_dirty_KBI:		[ internal use only ]
1326  *
1327  *	Set all bits in the page's dirty field.
1328  *
1329  *	The object containing the specified page must be locked if the
1330  *	call is made from the machine-independent layer.
1331  *
1332  *	See vm_page_clear_dirty_mask().
1333  *
1334  *	This function should only be called by vm_page_dirty().
1335  */
1336 void
1337 vm_page_dirty_KBI(vm_page_t m)
1338 {
1339 
1340 	/* Refer to this operation by its public name. */
1341 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1342 	    ("vm_page_dirty: page is invalid!"));
1343 	m->dirty = VM_PAGE_BITS_ALL;
1344 }
1345 
1346 /*
1347  *	vm_page_insert:		[ internal use only ]
1348  *
1349  *	Inserts the given mem entry into the object and object list.
1350  *
1351  *	The object must be locked.
1352  */
1353 int
1354 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1355 {
1356 	vm_page_t mpred;
1357 
1358 	VM_OBJECT_ASSERT_WLOCKED(object);
1359 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1360 	return (vm_page_insert_after(m, object, pindex, mpred));
1361 }
1362 
1363 /*
1364  *	vm_page_insert_after:
1365  *
1366  *	Inserts the page "m" into the specified object at offset "pindex".
1367  *
1368  *	The page "mpred" must immediately precede the offset "pindex" within
1369  *	the specified object.
1370  *
1371  *	The object must be locked.
1372  */
1373 static int
1374 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1375     vm_page_t mpred)
1376 {
1377 	vm_page_t msucc;
1378 
1379 	VM_OBJECT_ASSERT_WLOCKED(object);
1380 	KASSERT(m->object == NULL,
1381 	    ("vm_page_insert_after: page already inserted"));
1382 	if (mpred != NULL) {
1383 		KASSERT(mpred->object == object,
1384 		    ("vm_page_insert_after: object doesn't contain mpred"));
1385 		KASSERT(mpred->pindex < pindex,
1386 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1387 		msucc = TAILQ_NEXT(mpred, listq);
1388 	} else
1389 		msucc = TAILQ_FIRST(&object->memq);
1390 	if (msucc != NULL)
1391 		KASSERT(msucc->pindex > pindex,
1392 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1393 
1394 	/*
1395 	 * Record the object/offset pair in this page
1396 	 */
1397 	m->object = object;
1398 	m->pindex = pindex;
1399 
1400 	/*
1401 	 * Now link into the object's ordered list of backed pages.
1402 	 */
1403 	if (vm_radix_insert(&object->rtree, m)) {
1404 		m->object = NULL;
1405 		m->pindex = 0;
1406 		return (1);
1407 	}
1408 	vm_page_insert_radixdone(m, object, mpred);
1409 	return (0);
1410 }
1411 
1412 /*
1413  *	vm_page_insert_radixdone:
1414  *
1415  *	Complete page "m" insertion into the specified object after the
1416  *	radix trie hooking.
1417  *
1418  *	The page "mpred" must precede the offset "m->pindex" within the
1419  *	specified object.
1420  *
1421  *	The object must be locked.
1422  */
1423 static void
1424 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1425 {
1426 
1427 	VM_OBJECT_ASSERT_WLOCKED(object);
1428 	KASSERT(object != NULL && m->object == object,
1429 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1430 	if (mpred != NULL) {
1431 		KASSERT(mpred->object == object,
1432 		    ("vm_page_insert_after: object doesn't contain mpred"));
1433 		KASSERT(mpred->pindex < m->pindex,
1434 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1435 	}
1436 
1437 	if (mpred != NULL)
1438 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1439 	else
1440 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1441 
1442 	/*
1443 	 * Show that the object has one more resident page.
1444 	 */
1445 	object->resident_page_count++;
1446 
1447 	/*
1448 	 * Hold the vnode until the last page is released.
1449 	 */
1450 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1451 		vhold(object->handle);
1452 
1453 	/*
1454 	 * Since we are inserting a new and possibly dirty page,
1455 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1456 	 */
1457 	if (pmap_page_is_write_mapped(m))
1458 		vm_object_set_writeable_dirty(object);
1459 }
1460 
1461 /*
1462  *	vm_page_remove:
1463  *
1464  *	Removes the specified page from its containing object, but does not
1465  *	invalidate any backing storage.
1466  *
1467  *	The object must be locked.  The page must be locked if it is managed.
1468  */
1469 void
1470 vm_page_remove(vm_page_t m)
1471 {
1472 	vm_object_t object;
1473 	vm_page_t mrem;
1474 
1475 	if ((m->oflags & VPO_UNMANAGED) == 0)
1476 		vm_page_assert_locked(m);
1477 	if ((object = m->object) == NULL)
1478 		return;
1479 	VM_OBJECT_ASSERT_WLOCKED(object);
1480 	if (vm_page_xbusied(m))
1481 		vm_page_xunbusy_maybelocked(m);
1482 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1483 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1484 
1485 	/*
1486 	 * Now remove from the object's list of backed pages.
1487 	 */
1488 	TAILQ_REMOVE(&object->memq, m, listq);
1489 
1490 	/*
1491 	 * And show that the object has one fewer resident page.
1492 	 */
1493 	object->resident_page_count--;
1494 
1495 	/*
1496 	 * The vnode may now be recycled.
1497 	 */
1498 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1499 		vdrop(object->handle);
1500 
1501 	m->object = NULL;
1502 }
1503 
1504 /*
1505  *	vm_page_lookup:
1506  *
1507  *	Returns the page associated with the object/offset
1508  *	pair specified; if none is found, NULL is returned.
1509  *
1510  *	The object must be locked.
1511  */
1512 vm_page_t
1513 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1514 {
1515 
1516 	VM_OBJECT_ASSERT_LOCKED(object);
1517 	return (vm_radix_lookup(&object->rtree, pindex));
1518 }
1519 
1520 /*
1521  *	vm_page_find_least:
1522  *
1523  *	Returns the page associated with the object with least pindex
1524  *	greater than or equal to the parameter pindex, or NULL.
1525  *
1526  *	The object must be locked.
1527  */
1528 vm_page_t
1529 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1530 {
1531 	vm_page_t m;
1532 
1533 	VM_OBJECT_ASSERT_LOCKED(object);
1534 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1535 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1536 	return (m);
1537 }
1538 
1539 /*
1540  * Returns the given page's successor (by pindex) within the object if it is
1541  * resident; if none is found, NULL is returned.
1542  *
1543  * The object must be locked.
1544  */
1545 vm_page_t
1546 vm_page_next(vm_page_t m)
1547 {
1548 	vm_page_t next;
1549 
1550 	VM_OBJECT_ASSERT_LOCKED(m->object);
1551 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1552 		MPASS(next->object == m->object);
1553 		if (next->pindex != m->pindex + 1)
1554 			next = NULL;
1555 	}
1556 	return (next);
1557 }
1558 
1559 /*
1560  * Returns the given page's predecessor (by pindex) within the object if it is
1561  * resident; if none is found, NULL is returned.
1562  *
1563  * The object must be locked.
1564  */
1565 vm_page_t
1566 vm_page_prev(vm_page_t m)
1567 {
1568 	vm_page_t prev;
1569 
1570 	VM_OBJECT_ASSERT_LOCKED(m->object);
1571 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1572 		MPASS(prev->object == m->object);
1573 		if (prev->pindex != m->pindex - 1)
1574 			prev = NULL;
1575 	}
1576 	return (prev);
1577 }
1578 
1579 /*
1580  * Uses the page mnew as a replacement for an existing page at index
1581  * pindex which must be already present in the object.
1582  *
1583  * The existing page must not be on a paging queue.
1584  */
1585 vm_page_t
1586 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1587 {
1588 	vm_page_t mold;
1589 
1590 	VM_OBJECT_ASSERT_WLOCKED(object);
1591 	KASSERT(mnew->object == NULL,
1592 	    ("vm_page_replace: page %p already in object", mnew));
1593 	KASSERT(mnew->queue == PQ_NONE,
1594 	    ("vm_page_replace: new page %p is on a paging queue", mnew));
1595 
1596 	/*
1597 	 * This function mostly follows vm_page_insert() and
1598 	 * vm_page_remove() without the radix, object count and vnode
1599 	 * dance.  Double check such functions for more comments.
1600 	 */
1601 
1602 	mnew->object = object;
1603 	mnew->pindex = pindex;
1604 	mold = vm_radix_replace(&object->rtree, mnew);
1605 	KASSERT(mold->queue == PQ_NONE,
1606 	    ("vm_page_replace: old page %p is on a paging queue", mold));
1607 
1608 	/* Keep the resident page list in sorted order. */
1609 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1610 	TAILQ_REMOVE(&object->memq, mold, listq);
1611 
1612 	mold->object = NULL;
1613 	vm_page_xunbusy_maybelocked(mold);
1614 
1615 	/*
1616 	 * The object's resident_page_count does not change because we have
1617 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1618 	 */
1619 	if (pmap_page_is_write_mapped(mnew))
1620 		vm_object_set_writeable_dirty(object);
1621 	return (mold);
1622 }
1623 
1624 /*
1625  *	vm_page_rename:
1626  *
1627  *	Move the given memory entry from its
1628  *	current object to the specified target object/offset.
1629  *
1630  *	Note: swap associated with the page must be invalidated by the move.  We
1631  *	      have to do this for several reasons:  (1) we aren't freeing the
1632  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1633  *	      moving the page from object A to B, and will then later move
1634  *	      the backing store from A to B and we can't have a conflict.
1635  *
1636  *	Note: we *always* dirty the page.  It is necessary both for the
1637  *	      fact that we moved it, and because we may be invalidating
1638  *	      swap.
1639  *
1640  *	The objects must be locked.
1641  */
1642 int
1643 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1644 {
1645 	vm_page_t mpred;
1646 	vm_pindex_t opidx;
1647 
1648 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1649 
1650 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1651 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1652 	    ("vm_page_rename: pindex already renamed"));
1653 
1654 	/*
1655 	 * Create a custom version of vm_page_insert() which does not depend
1656 	 * by m_prev and can cheat on the implementation aspects of the
1657 	 * function.
1658 	 */
1659 	opidx = m->pindex;
1660 	m->pindex = new_pindex;
1661 	if (vm_radix_insert(&new_object->rtree, m)) {
1662 		m->pindex = opidx;
1663 		return (1);
1664 	}
1665 
1666 	/*
1667 	 * The operation cannot fail anymore.  The removal must happen before
1668 	 * the listq iterator is tainted.
1669 	 */
1670 	m->pindex = opidx;
1671 	vm_page_lock(m);
1672 	vm_page_remove(m);
1673 
1674 	/* Return back to the new pindex to complete vm_page_insert(). */
1675 	m->pindex = new_pindex;
1676 	m->object = new_object;
1677 	vm_page_unlock(m);
1678 	vm_page_insert_radixdone(m, new_object, mpred);
1679 	vm_page_dirty(m);
1680 	return (0);
1681 }
1682 
1683 /*
1684  *	vm_page_alloc:
1685  *
1686  *	Allocate and return a page that is associated with the specified
1687  *	object and offset pair.  By default, this page is exclusive busied.
1688  *
1689  *	The caller must always specify an allocation class.
1690  *
1691  *	allocation classes:
1692  *	VM_ALLOC_NORMAL		normal process request
1693  *	VM_ALLOC_SYSTEM		system *really* needs a page
1694  *	VM_ALLOC_INTERRUPT	interrupt time request
1695  *
1696  *	optional allocation flags:
1697  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1698  *				intends to allocate
1699  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1700  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1701  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1702  *				should not be exclusive busy
1703  *	VM_ALLOC_SBUSY		shared busy the allocated page
1704  *	VM_ALLOC_WIRED		wire the allocated page
1705  *	VM_ALLOC_ZERO		prefer a zeroed page
1706  */
1707 vm_page_t
1708 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1709 {
1710 
1711 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1712 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1713 }
1714 
1715 vm_page_t
1716 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1717     int req)
1718 {
1719 
1720 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1721 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1722 	    NULL));
1723 }
1724 
1725 /*
1726  * Allocate a page in the specified object with the given page index.  To
1727  * optimize insertion of the page into the object, the caller must also specifiy
1728  * the resident page in the object with largest index smaller than the given
1729  * page index, or NULL if no such page exists.
1730  */
1731 vm_page_t
1732 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1733     int req, vm_page_t mpred)
1734 {
1735 	struct vm_domainset_iter di;
1736 	vm_page_t m;
1737 	int domain;
1738 
1739 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1740 	do {
1741 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1742 		    mpred);
1743 		if (m != NULL)
1744 			break;
1745 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1746 
1747 	return (m);
1748 }
1749 
1750 /*
1751  * Returns true if the number of free pages exceeds the minimum
1752  * for the request class and false otherwise.
1753  */
1754 int
1755 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1756 {
1757 	u_int limit, old, new;
1758 
1759 	req = req & VM_ALLOC_CLASS_MASK;
1760 
1761 	/*
1762 	 * The page daemon is allowed to dig deeper into the free page list.
1763 	 */
1764 	if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1765 		req = VM_ALLOC_SYSTEM;
1766 	if (req == VM_ALLOC_INTERRUPT)
1767 		limit = 0;
1768 	else if (req == VM_ALLOC_SYSTEM)
1769 		limit = vmd->vmd_interrupt_free_min;
1770 	else
1771 		limit = vmd->vmd_free_reserved;
1772 
1773 	/*
1774 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1775 	 */
1776 	limit += npages;
1777 	old = vmd->vmd_free_count;
1778 	do {
1779 		if (old < limit)
1780 			return (0);
1781 		new = old - npages;
1782 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1783 
1784 	/* Wake the page daemon if we've crossed the threshold. */
1785 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1786 		pagedaemon_wakeup(vmd->vmd_domain);
1787 
1788 	/* Only update bitsets on transitions. */
1789 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1790 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1791 		vm_domain_set(vmd);
1792 
1793 	return (1);
1794 }
1795 
1796 vm_page_t
1797 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1798     int req, vm_page_t mpred)
1799 {
1800 	struct vm_domain *vmd;
1801 	vm_page_t m;
1802 	int flags;
1803 
1804 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1805 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1806 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1807 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1808 	    ("inconsistent object(%p)/req(%x)", object, req));
1809 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1810 	    ("Can't sleep and retry object insertion."));
1811 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1812 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1813 	    (uintmax_t)pindex));
1814 	if (object != NULL)
1815 		VM_OBJECT_ASSERT_WLOCKED(object);
1816 
1817 again:
1818 	m = NULL;
1819 #if VM_NRESERVLEVEL > 0
1820 	/*
1821 	 * Can we allocate the page from a reservation?
1822 	 */
1823 	if (vm_object_reserv(object) &&
1824 	    ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL ||
1825 	    (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) {
1826 		domain = vm_phys_domain(m);
1827 		vmd = VM_DOMAIN(domain);
1828 		goto found;
1829 	}
1830 #endif
1831 	vmd = VM_DOMAIN(domain);
1832 	if (object != NULL && vmd->vmd_pgcache != NULL) {
1833 		m = uma_zalloc(vmd->vmd_pgcache, M_NOWAIT);
1834 		if (m != NULL)
1835 			goto found;
1836 	}
1837 	if (vm_domain_allocate(vmd, req, 1)) {
1838 		/*
1839 		 * If not, allocate it from the free page queues.
1840 		 */
1841 		vm_domain_free_lock(vmd);
1842 		m = vm_phys_alloc_pages(domain, object != NULL ?
1843 		    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1844 		vm_domain_free_unlock(vmd);
1845 		if (m == NULL) {
1846 			vm_domain_freecnt_inc(vmd, 1);
1847 #if VM_NRESERVLEVEL > 0
1848 			if (vm_reserv_reclaim_inactive(domain))
1849 				goto again;
1850 #endif
1851 		}
1852 	}
1853 	if (m == NULL) {
1854 		/*
1855 		 * Not allocatable, give up.
1856 		 */
1857 		if (vm_domain_alloc_fail(vmd, object, req))
1858 			goto again;
1859 		return (NULL);
1860 	}
1861 
1862 	/*
1863 	 *  At this point we had better have found a good page.
1864 	 */
1865 	KASSERT(m != NULL, ("missing page"));
1866 
1867 found:
1868 	vm_page_dequeue(m);
1869 	vm_page_alloc_check(m);
1870 
1871 	/*
1872 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1873 	 */
1874 	flags = 0;
1875 	if ((req & VM_ALLOC_ZERO) != 0)
1876 		flags = PG_ZERO;
1877 	flags &= m->flags;
1878 	if ((req & VM_ALLOC_NODUMP) != 0)
1879 		flags |= PG_NODUMP;
1880 	m->flags = flags;
1881 	m->aflags = 0;
1882 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1883 	    VPO_UNMANAGED : 0;
1884 	m->busy_lock = VPB_UNBUSIED;
1885 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1886 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1887 	if ((req & VM_ALLOC_SBUSY) != 0)
1888 		m->busy_lock = VPB_SHARERS_WORD(1);
1889 	if (req & VM_ALLOC_WIRED) {
1890 		/*
1891 		 * The page lock is not required for wiring a page until that
1892 		 * page is inserted into the object.
1893 		 */
1894 		vm_wire_add(1);
1895 		m->wire_count = 1;
1896 	}
1897 	m->act_count = 0;
1898 
1899 	if (object != NULL) {
1900 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1901 			if (req & VM_ALLOC_WIRED) {
1902 				vm_wire_sub(1);
1903 				m->wire_count = 0;
1904 			}
1905 			KASSERT(m->object == NULL, ("page %p has object", m));
1906 			m->oflags = VPO_UNMANAGED;
1907 			m->busy_lock = VPB_UNBUSIED;
1908 			/* Don't change PG_ZERO. */
1909 			vm_page_free_toq(m);
1910 			if (req & VM_ALLOC_WAITFAIL) {
1911 				VM_OBJECT_WUNLOCK(object);
1912 				vm_radix_wait();
1913 				VM_OBJECT_WLOCK(object);
1914 			}
1915 			return (NULL);
1916 		}
1917 
1918 		/* Ignore device objects; the pager sets "memattr" for them. */
1919 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1920 		    (object->flags & OBJ_FICTITIOUS) == 0)
1921 			pmap_page_set_memattr(m, object->memattr);
1922 	} else
1923 		m->pindex = pindex;
1924 
1925 	return (m);
1926 }
1927 
1928 /*
1929  *	vm_page_alloc_contig:
1930  *
1931  *	Allocate a contiguous set of physical pages of the given size "npages"
1932  *	from the free lists.  All of the physical pages must be at or above
1933  *	the given physical address "low" and below the given physical address
1934  *	"high".  The given value "alignment" determines the alignment of the
1935  *	first physical page in the set.  If the given value "boundary" is
1936  *	non-zero, then the set of physical pages cannot cross any physical
1937  *	address boundary that is a multiple of that value.  Both "alignment"
1938  *	and "boundary" must be a power of two.
1939  *
1940  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1941  *	then the memory attribute setting for the physical pages is configured
1942  *	to the object's memory attribute setting.  Otherwise, the memory
1943  *	attribute setting for the physical pages is configured to "memattr",
1944  *	overriding the object's memory attribute setting.  However, if the
1945  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1946  *	memory attribute setting for the physical pages cannot be configured
1947  *	to VM_MEMATTR_DEFAULT.
1948  *
1949  *	The specified object may not contain fictitious pages.
1950  *
1951  *	The caller must always specify an allocation class.
1952  *
1953  *	allocation classes:
1954  *	VM_ALLOC_NORMAL		normal process request
1955  *	VM_ALLOC_SYSTEM		system *really* needs a page
1956  *	VM_ALLOC_INTERRUPT	interrupt time request
1957  *
1958  *	optional allocation flags:
1959  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1960  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1961  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1962  *				should not be exclusive busy
1963  *	VM_ALLOC_SBUSY		shared busy the allocated page
1964  *	VM_ALLOC_WIRED		wire the allocated page
1965  *	VM_ALLOC_ZERO		prefer a zeroed page
1966  */
1967 vm_page_t
1968 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1969     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1970     vm_paddr_t boundary, vm_memattr_t memattr)
1971 {
1972 	struct vm_domainset_iter di;
1973 	vm_page_t m;
1974 	int domain;
1975 
1976 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1977 	do {
1978 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1979 		    npages, low, high, alignment, boundary, memattr);
1980 		if (m != NULL)
1981 			break;
1982 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1983 
1984 	return (m);
1985 }
1986 
1987 vm_page_t
1988 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1989     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1990     vm_paddr_t boundary, vm_memattr_t memattr)
1991 {
1992 	struct vm_domain *vmd;
1993 	vm_page_t m, m_ret, mpred;
1994 	u_int busy_lock, flags, oflags;
1995 
1996 	mpred = NULL;	/* XXX: pacify gcc */
1997 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1998 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1999 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2000 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2001 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2002 	    req));
2003 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2004 	    ("Can't sleep and retry object insertion."));
2005 	if (object != NULL) {
2006 		VM_OBJECT_ASSERT_WLOCKED(object);
2007 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2008 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2009 		    object));
2010 	}
2011 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2012 
2013 	if (object != NULL) {
2014 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2015 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2016 		    ("vm_page_alloc_contig: pindex already allocated"));
2017 	}
2018 
2019 	/*
2020 	 * Can we allocate the pages without the number of free pages falling
2021 	 * below the lower bound for the allocation class?
2022 	 */
2023 again:
2024 #if VM_NRESERVLEVEL > 0
2025 	/*
2026 	 * Can we allocate the pages from a reservation?
2027 	 */
2028 	if (vm_object_reserv(object) &&
2029 	    ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
2030 	    npages, low, high, alignment, boundary, mpred)) != NULL ||
2031 	    (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain,
2032 	    npages, low, high, alignment, boundary, mpred)) != NULL)) {
2033 		domain = vm_phys_domain(m_ret);
2034 		vmd = VM_DOMAIN(domain);
2035 		goto found;
2036 	}
2037 #endif
2038 	m_ret = NULL;
2039 	vmd = VM_DOMAIN(domain);
2040 	if (vm_domain_allocate(vmd, req, npages)) {
2041 		/*
2042 		 * allocate them from the free page queues.
2043 		 */
2044 		vm_domain_free_lock(vmd);
2045 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2046 		    alignment, boundary);
2047 		vm_domain_free_unlock(vmd);
2048 		if (m_ret == NULL) {
2049 			vm_domain_freecnt_inc(vmd, npages);
2050 #if VM_NRESERVLEVEL > 0
2051 			if (vm_reserv_reclaim_contig(domain, npages, low,
2052 			    high, alignment, boundary))
2053 				goto again;
2054 #endif
2055 		}
2056 	}
2057 	if (m_ret == NULL) {
2058 		if (vm_domain_alloc_fail(vmd, object, req))
2059 			goto again;
2060 		return (NULL);
2061 	}
2062 #if VM_NRESERVLEVEL > 0
2063 found:
2064 #endif
2065 	for (m = m_ret; m < &m_ret[npages]; m++) {
2066 		vm_page_dequeue(m);
2067 		vm_page_alloc_check(m);
2068 	}
2069 
2070 	/*
2071 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2072 	 */
2073 	flags = 0;
2074 	if ((req & VM_ALLOC_ZERO) != 0)
2075 		flags = PG_ZERO;
2076 	if ((req & VM_ALLOC_NODUMP) != 0)
2077 		flags |= PG_NODUMP;
2078 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2079 	    VPO_UNMANAGED : 0;
2080 	busy_lock = VPB_UNBUSIED;
2081 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2082 		busy_lock = VPB_SINGLE_EXCLUSIVER;
2083 	if ((req & VM_ALLOC_SBUSY) != 0)
2084 		busy_lock = VPB_SHARERS_WORD(1);
2085 	if ((req & VM_ALLOC_WIRED) != 0)
2086 		vm_wire_add(npages);
2087 	if (object != NULL) {
2088 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2089 		    memattr == VM_MEMATTR_DEFAULT)
2090 			memattr = object->memattr;
2091 	}
2092 	for (m = m_ret; m < &m_ret[npages]; m++) {
2093 		m->aflags = 0;
2094 		m->flags = (m->flags | PG_NODUMP) & flags;
2095 		m->busy_lock = busy_lock;
2096 		if ((req & VM_ALLOC_WIRED) != 0)
2097 			m->wire_count = 1;
2098 		m->act_count = 0;
2099 		m->oflags = oflags;
2100 		if (object != NULL) {
2101 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2102 				if ((req & VM_ALLOC_WIRED) != 0)
2103 					vm_wire_sub(npages);
2104 				KASSERT(m->object == NULL,
2105 				    ("page %p has object", m));
2106 				mpred = m;
2107 				for (m = m_ret; m < &m_ret[npages]; m++) {
2108 					if (m <= mpred &&
2109 					    (req & VM_ALLOC_WIRED) != 0)
2110 						m->wire_count = 0;
2111 					m->oflags = VPO_UNMANAGED;
2112 					m->busy_lock = VPB_UNBUSIED;
2113 					/* Don't change PG_ZERO. */
2114 					vm_page_free_toq(m);
2115 				}
2116 				if (req & VM_ALLOC_WAITFAIL) {
2117 					VM_OBJECT_WUNLOCK(object);
2118 					vm_radix_wait();
2119 					VM_OBJECT_WLOCK(object);
2120 				}
2121 				return (NULL);
2122 			}
2123 			mpred = m;
2124 		} else
2125 			m->pindex = pindex;
2126 		if (memattr != VM_MEMATTR_DEFAULT)
2127 			pmap_page_set_memattr(m, memattr);
2128 		pindex++;
2129 	}
2130 	return (m_ret);
2131 }
2132 
2133 /*
2134  * Check a page that has been freshly dequeued from a freelist.
2135  */
2136 static void
2137 vm_page_alloc_check(vm_page_t m)
2138 {
2139 
2140 	KASSERT(m->object == NULL, ("page %p has object", m));
2141 	KASSERT(m->queue == PQ_NONE,
2142 	    ("page %p has unexpected queue %d", m, m->queue));
2143 	KASSERT(!vm_page_held(m), ("page %p is held", m));
2144 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2145 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2146 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2147 	    ("page %p has unexpected memattr %d",
2148 	    m, pmap_page_get_memattr(m)));
2149 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2150 }
2151 
2152 /*
2153  * 	vm_page_alloc_freelist:
2154  *
2155  *	Allocate a physical page from the specified free page list.
2156  *
2157  *	The caller must always specify an allocation class.
2158  *
2159  *	allocation classes:
2160  *	VM_ALLOC_NORMAL		normal process request
2161  *	VM_ALLOC_SYSTEM		system *really* needs a page
2162  *	VM_ALLOC_INTERRUPT	interrupt time request
2163  *
2164  *	optional allocation flags:
2165  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2166  *				intends to allocate
2167  *	VM_ALLOC_WIRED		wire the allocated page
2168  *	VM_ALLOC_ZERO		prefer a zeroed page
2169  */
2170 vm_page_t
2171 vm_page_alloc_freelist(int freelist, int req)
2172 {
2173 	struct vm_domainset_iter di;
2174 	vm_page_t m;
2175 	int domain;
2176 
2177 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2178 	do {
2179 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2180 		if (m != NULL)
2181 			break;
2182 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2183 
2184 	return (m);
2185 }
2186 
2187 vm_page_t
2188 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2189 {
2190 	struct vm_domain *vmd;
2191 	vm_page_t m;
2192 	u_int flags;
2193 
2194 	m = NULL;
2195 	vmd = VM_DOMAIN(domain);
2196 again:
2197 	if (vm_domain_allocate(vmd, req, 1)) {
2198 		vm_domain_free_lock(vmd);
2199 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2200 		    VM_FREEPOOL_DIRECT, 0);
2201 		vm_domain_free_unlock(vmd);
2202 		if (m == NULL)
2203 			vm_domain_freecnt_inc(vmd, 1);
2204 	}
2205 	if (m == NULL) {
2206 		if (vm_domain_alloc_fail(vmd, NULL, req))
2207 			goto again;
2208 		return (NULL);
2209 	}
2210 	vm_page_dequeue(m);
2211 	vm_page_alloc_check(m);
2212 
2213 	/*
2214 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2215 	 */
2216 	m->aflags = 0;
2217 	flags = 0;
2218 	if ((req & VM_ALLOC_ZERO) != 0)
2219 		flags = PG_ZERO;
2220 	m->flags &= flags;
2221 	if ((req & VM_ALLOC_WIRED) != 0) {
2222 		/*
2223 		 * The page lock is not required for wiring a page that does
2224 		 * not belong to an object.
2225 		 */
2226 		vm_wire_add(1);
2227 		m->wire_count = 1;
2228 	}
2229 	/* Unmanaged pages don't use "act_count". */
2230 	m->oflags = VPO_UNMANAGED;
2231 	return (m);
2232 }
2233 
2234 static int
2235 vm_page_import(void *arg, void **store, int cnt, int domain, int flags)
2236 {
2237 	struct vm_domain *vmd;
2238 	int i;
2239 
2240 	vmd = arg;
2241 	/* Only import if we can bring in a full bucket. */
2242 	if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2243 		return (0);
2244 	domain = vmd->vmd_domain;
2245 	vm_domain_free_lock(vmd);
2246 	i = vm_phys_alloc_npages(domain, VM_FREEPOOL_DEFAULT, cnt,
2247 	    (vm_page_t *)store);
2248 	vm_domain_free_unlock(vmd);
2249 	if (cnt != i)
2250 		vm_domain_freecnt_inc(vmd, cnt - i);
2251 
2252 	return (i);
2253 }
2254 
2255 static void
2256 vm_page_release(void *arg, void **store, int cnt)
2257 {
2258 	struct vm_domain *vmd;
2259 	vm_page_t m;
2260 	int i;
2261 
2262 	vmd = arg;
2263 	vm_domain_free_lock(vmd);
2264 	for (i = 0; i < cnt; i++) {
2265 		m = (vm_page_t)store[i];
2266 		vm_phys_free_pages(m, 0);
2267 	}
2268 	vm_domain_free_unlock(vmd);
2269 	vm_domain_freecnt_inc(vmd, cnt);
2270 }
2271 
2272 #define	VPSC_ANY	0	/* No restrictions. */
2273 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2274 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2275 
2276 /*
2277  *	vm_page_scan_contig:
2278  *
2279  *	Scan vm_page_array[] between the specified entries "m_start" and
2280  *	"m_end" for a run of contiguous physical pages that satisfy the
2281  *	specified conditions, and return the lowest page in the run.  The
2282  *	specified "alignment" determines the alignment of the lowest physical
2283  *	page in the run.  If the specified "boundary" is non-zero, then the
2284  *	run of physical pages cannot span a physical address that is a
2285  *	multiple of "boundary".
2286  *
2287  *	"m_end" is never dereferenced, so it need not point to a vm_page
2288  *	structure within vm_page_array[].
2289  *
2290  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2291  *	span a hole (or discontiguity) in the physical address space.  Both
2292  *	"alignment" and "boundary" must be a power of two.
2293  */
2294 vm_page_t
2295 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2296     u_long alignment, vm_paddr_t boundary, int options)
2297 {
2298 	struct mtx *m_mtx;
2299 	vm_object_t object;
2300 	vm_paddr_t pa;
2301 	vm_page_t m, m_run;
2302 #if VM_NRESERVLEVEL > 0
2303 	int level;
2304 #endif
2305 	int m_inc, order, run_ext, run_len;
2306 
2307 	KASSERT(npages > 0, ("npages is 0"));
2308 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2309 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2310 	m_run = NULL;
2311 	run_len = 0;
2312 	m_mtx = NULL;
2313 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2314 		KASSERT((m->flags & PG_MARKER) == 0,
2315 		    ("page %p is PG_MARKER", m));
2316 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2317 		    ("fictitious page %p has invalid wire count", m));
2318 
2319 		/*
2320 		 * If the current page would be the start of a run, check its
2321 		 * physical address against the end, alignment, and boundary
2322 		 * conditions.  If it doesn't satisfy these conditions, either
2323 		 * terminate the scan or advance to the next page that
2324 		 * satisfies the failed condition.
2325 		 */
2326 		if (run_len == 0) {
2327 			KASSERT(m_run == NULL, ("m_run != NULL"));
2328 			if (m + npages > m_end)
2329 				break;
2330 			pa = VM_PAGE_TO_PHYS(m);
2331 			if ((pa & (alignment - 1)) != 0) {
2332 				m_inc = atop(roundup2(pa, alignment) - pa);
2333 				continue;
2334 			}
2335 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2336 			    boundary) != 0) {
2337 				m_inc = atop(roundup2(pa, boundary) - pa);
2338 				continue;
2339 			}
2340 		} else
2341 			KASSERT(m_run != NULL, ("m_run == NULL"));
2342 
2343 		vm_page_change_lock(m, &m_mtx);
2344 		m_inc = 1;
2345 retry:
2346 		if (vm_page_held(m))
2347 			run_ext = 0;
2348 #if VM_NRESERVLEVEL > 0
2349 		else if ((level = vm_reserv_level(m)) >= 0 &&
2350 		    (options & VPSC_NORESERV) != 0) {
2351 			run_ext = 0;
2352 			/* Advance to the end of the reservation. */
2353 			pa = VM_PAGE_TO_PHYS(m);
2354 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2355 			    pa);
2356 		}
2357 #endif
2358 		else if ((object = m->object) != NULL) {
2359 			/*
2360 			 * The page is considered eligible for relocation if
2361 			 * and only if it could be laundered or reclaimed by
2362 			 * the page daemon.
2363 			 */
2364 			if (!VM_OBJECT_TRYRLOCK(object)) {
2365 				mtx_unlock(m_mtx);
2366 				VM_OBJECT_RLOCK(object);
2367 				mtx_lock(m_mtx);
2368 				if (m->object != object) {
2369 					/*
2370 					 * The page may have been freed.
2371 					 */
2372 					VM_OBJECT_RUNLOCK(object);
2373 					goto retry;
2374 				} else if (vm_page_held(m)) {
2375 					run_ext = 0;
2376 					goto unlock;
2377 				}
2378 			}
2379 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2380 			    ("page %p is PG_UNHOLDFREE", m));
2381 			/* Don't care: PG_NODUMP, PG_ZERO. */
2382 			if (object->type != OBJT_DEFAULT &&
2383 			    object->type != OBJT_SWAP &&
2384 			    object->type != OBJT_VNODE) {
2385 				run_ext = 0;
2386 #if VM_NRESERVLEVEL > 0
2387 			} else if ((options & VPSC_NOSUPER) != 0 &&
2388 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2389 				run_ext = 0;
2390 				/* Advance to the end of the superpage. */
2391 				pa = VM_PAGE_TO_PHYS(m);
2392 				m_inc = atop(roundup2(pa + 1,
2393 				    vm_reserv_size(level)) - pa);
2394 #endif
2395 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2396 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2397 				/*
2398 				 * The page is allocated but eligible for
2399 				 * relocation.  Extend the current run by one
2400 				 * page.
2401 				 */
2402 				KASSERT(pmap_page_get_memattr(m) ==
2403 				    VM_MEMATTR_DEFAULT,
2404 				    ("page %p has an unexpected memattr", m));
2405 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2406 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2407 				    ("page %p has unexpected oflags", m));
2408 				/* Don't care: VPO_NOSYNC. */
2409 				run_ext = 1;
2410 			} else
2411 				run_ext = 0;
2412 unlock:
2413 			VM_OBJECT_RUNLOCK(object);
2414 #if VM_NRESERVLEVEL > 0
2415 		} else if (level >= 0) {
2416 			/*
2417 			 * The page is reserved but not yet allocated.  In
2418 			 * other words, it is still free.  Extend the current
2419 			 * run by one page.
2420 			 */
2421 			run_ext = 1;
2422 #endif
2423 		} else if ((order = m->order) < VM_NFREEORDER) {
2424 			/*
2425 			 * The page is enqueued in the physical memory
2426 			 * allocator's free page queues.  Moreover, it is the
2427 			 * first page in a power-of-two-sized run of
2428 			 * contiguous free pages.  Add these pages to the end
2429 			 * of the current run, and jump ahead.
2430 			 */
2431 			run_ext = 1 << order;
2432 			m_inc = 1 << order;
2433 		} else {
2434 			/*
2435 			 * Skip the page for one of the following reasons: (1)
2436 			 * It is enqueued in the physical memory allocator's
2437 			 * free page queues.  However, it is not the first
2438 			 * page in a run of contiguous free pages.  (This case
2439 			 * rarely occurs because the scan is performed in
2440 			 * ascending order.) (2) It is not reserved, and it is
2441 			 * transitioning from free to allocated.  (Conversely,
2442 			 * the transition from allocated to free for managed
2443 			 * pages is blocked by the page lock.) (3) It is
2444 			 * allocated but not contained by an object and not
2445 			 * wired, e.g., allocated by Xen's balloon driver.
2446 			 */
2447 			run_ext = 0;
2448 		}
2449 
2450 		/*
2451 		 * Extend or reset the current run of pages.
2452 		 */
2453 		if (run_ext > 0) {
2454 			if (run_len == 0)
2455 				m_run = m;
2456 			run_len += run_ext;
2457 		} else {
2458 			if (run_len > 0) {
2459 				m_run = NULL;
2460 				run_len = 0;
2461 			}
2462 		}
2463 	}
2464 	if (m_mtx != NULL)
2465 		mtx_unlock(m_mtx);
2466 	if (run_len >= npages)
2467 		return (m_run);
2468 	return (NULL);
2469 }
2470 
2471 /*
2472  *	vm_page_reclaim_run:
2473  *
2474  *	Try to relocate each of the allocated virtual pages within the
2475  *	specified run of physical pages to a new physical address.  Free the
2476  *	physical pages underlying the relocated virtual pages.  A virtual page
2477  *	is relocatable if and only if it could be laundered or reclaimed by
2478  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2479  *	physical address above "high".
2480  *
2481  *	Returns 0 if every physical page within the run was already free or
2482  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2483  *	value indicating why the last attempt to relocate a virtual page was
2484  *	unsuccessful.
2485  *
2486  *	"req_class" must be an allocation class.
2487  */
2488 static int
2489 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2490     vm_paddr_t high)
2491 {
2492 	struct vm_domain *vmd;
2493 	struct mtx *m_mtx;
2494 	struct spglist free;
2495 	vm_object_t object;
2496 	vm_paddr_t pa;
2497 	vm_page_t m, m_end, m_new;
2498 	int error, order, req;
2499 
2500 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2501 	    ("req_class is not an allocation class"));
2502 	SLIST_INIT(&free);
2503 	error = 0;
2504 	m = m_run;
2505 	m_end = m_run + npages;
2506 	m_mtx = NULL;
2507 	for (; error == 0 && m < m_end; m++) {
2508 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2509 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2510 
2511 		/*
2512 		 * Avoid releasing and reacquiring the same page lock.
2513 		 */
2514 		vm_page_change_lock(m, &m_mtx);
2515 retry:
2516 		if (vm_page_held(m))
2517 			error = EBUSY;
2518 		else if ((object = m->object) != NULL) {
2519 			/*
2520 			 * The page is relocated if and only if it could be
2521 			 * laundered or reclaimed by the page daemon.
2522 			 */
2523 			if (!VM_OBJECT_TRYWLOCK(object)) {
2524 				mtx_unlock(m_mtx);
2525 				VM_OBJECT_WLOCK(object);
2526 				mtx_lock(m_mtx);
2527 				if (m->object != object) {
2528 					/*
2529 					 * The page may have been freed.
2530 					 */
2531 					VM_OBJECT_WUNLOCK(object);
2532 					goto retry;
2533 				} else if (vm_page_held(m)) {
2534 					error = EBUSY;
2535 					goto unlock;
2536 				}
2537 			}
2538 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2539 			    ("page %p is PG_UNHOLDFREE", m));
2540 			/* Don't care: PG_NODUMP, PG_ZERO. */
2541 			if (object->type != OBJT_DEFAULT &&
2542 			    object->type != OBJT_SWAP &&
2543 			    object->type != OBJT_VNODE)
2544 				error = EINVAL;
2545 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2546 				error = EINVAL;
2547 			else if (vm_page_queue(m) != PQ_NONE &&
2548 			    !vm_page_busied(m)) {
2549 				KASSERT(pmap_page_get_memattr(m) ==
2550 				    VM_MEMATTR_DEFAULT,
2551 				    ("page %p has an unexpected memattr", m));
2552 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2553 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2554 				    ("page %p has unexpected oflags", m));
2555 				/* Don't care: VPO_NOSYNC. */
2556 				if (m->valid != 0) {
2557 					/*
2558 					 * First, try to allocate a new page
2559 					 * that is above "high".  Failing
2560 					 * that, try to allocate a new page
2561 					 * that is below "m_run".  Allocate
2562 					 * the new page between the end of
2563 					 * "m_run" and "high" only as a last
2564 					 * resort.
2565 					 */
2566 					req = req_class | VM_ALLOC_NOOBJ;
2567 					if ((m->flags & PG_NODUMP) != 0)
2568 						req |= VM_ALLOC_NODUMP;
2569 					if (trunc_page(high) !=
2570 					    ~(vm_paddr_t)PAGE_MASK) {
2571 						m_new = vm_page_alloc_contig(
2572 						    NULL, 0, req, 1,
2573 						    round_page(high),
2574 						    ~(vm_paddr_t)0,
2575 						    PAGE_SIZE, 0,
2576 						    VM_MEMATTR_DEFAULT);
2577 					} else
2578 						m_new = NULL;
2579 					if (m_new == NULL) {
2580 						pa = VM_PAGE_TO_PHYS(m_run);
2581 						m_new = vm_page_alloc_contig(
2582 						    NULL, 0, req, 1,
2583 						    0, pa - 1, PAGE_SIZE, 0,
2584 						    VM_MEMATTR_DEFAULT);
2585 					}
2586 					if (m_new == NULL) {
2587 						pa += ptoa(npages);
2588 						m_new = vm_page_alloc_contig(
2589 						    NULL, 0, req, 1,
2590 						    pa, high, PAGE_SIZE, 0,
2591 						    VM_MEMATTR_DEFAULT);
2592 					}
2593 					if (m_new == NULL) {
2594 						error = ENOMEM;
2595 						goto unlock;
2596 					}
2597 					KASSERT(m_new->wire_count == 0,
2598 					    ("page %p is wired", m_new));
2599 
2600 					/*
2601 					 * Replace "m" with the new page.  For
2602 					 * vm_page_replace(), "m" must be busy
2603 					 * and dequeued.  Finally, change "m"
2604 					 * as if vm_page_free() was called.
2605 					 */
2606 					if (object->ref_count != 0)
2607 						pmap_remove_all(m);
2608 					m_new->aflags = m->aflags &
2609 					    ~PGA_QUEUE_STATE_MASK;
2610 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2611 					    ("page %p is managed", m_new));
2612 					m_new->oflags = m->oflags & VPO_NOSYNC;
2613 					pmap_copy_page(m, m_new);
2614 					m_new->valid = m->valid;
2615 					m_new->dirty = m->dirty;
2616 					m->flags &= ~PG_ZERO;
2617 					vm_page_xbusy(m);
2618 					vm_page_remque(m);
2619 					vm_page_replace_checked(m_new, object,
2620 					    m->pindex, m);
2621 					if (vm_page_free_prep(m))
2622 						SLIST_INSERT_HEAD(&free, m,
2623 						    plinks.s.ss);
2624 
2625 					/*
2626 					 * The new page must be deactivated
2627 					 * before the object is unlocked.
2628 					 */
2629 					vm_page_change_lock(m_new, &m_mtx);
2630 					vm_page_deactivate(m_new);
2631 				} else {
2632 					m->flags &= ~PG_ZERO;
2633 					vm_page_remque(m);
2634 					vm_page_remove(m);
2635 					if (vm_page_free_prep(m))
2636 						SLIST_INSERT_HEAD(&free, m,
2637 						    plinks.s.ss);
2638 					KASSERT(m->dirty == 0,
2639 					    ("page %p is dirty", m));
2640 				}
2641 			} else
2642 				error = EBUSY;
2643 unlock:
2644 			VM_OBJECT_WUNLOCK(object);
2645 		} else {
2646 			MPASS(vm_phys_domain(m) == domain);
2647 			vmd = VM_DOMAIN(domain);
2648 			vm_domain_free_lock(vmd);
2649 			order = m->order;
2650 			if (order < VM_NFREEORDER) {
2651 				/*
2652 				 * The page is enqueued in the physical memory
2653 				 * allocator's free page queues.  Moreover, it
2654 				 * is the first page in a power-of-two-sized
2655 				 * run of contiguous free pages.  Jump ahead
2656 				 * to the last page within that run, and
2657 				 * continue from there.
2658 				 */
2659 				m += (1 << order) - 1;
2660 			}
2661 #if VM_NRESERVLEVEL > 0
2662 			else if (vm_reserv_is_page_free(m))
2663 				order = 0;
2664 #endif
2665 			vm_domain_free_unlock(vmd);
2666 			if (order == VM_NFREEORDER)
2667 				error = EINVAL;
2668 		}
2669 	}
2670 	if (m_mtx != NULL)
2671 		mtx_unlock(m_mtx);
2672 	if ((m = SLIST_FIRST(&free)) != NULL) {
2673 		int cnt;
2674 
2675 		vmd = VM_DOMAIN(domain);
2676 		cnt = 0;
2677 		vm_domain_free_lock(vmd);
2678 		do {
2679 			MPASS(vm_phys_domain(m) == domain);
2680 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2681 			vm_phys_free_pages(m, 0);
2682 			cnt++;
2683 		} while ((m = SLIST_FIRST(&free)) != NULL);
2684 		vm_domain_free_unlock(vmd);
2685 		vm_domain_freecnt_inc(vmd, cnt);
2686 	}
2687 	return (error);
2688 }
2689 
2690 #define	NRUNS	16
2691 
2692 CTASSERT(powerof2(NRUNS));
2693 
2694 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2695 
2696 #define	MIN_RECLAIM	8
2697 
2698 /*
2699  *	vm_page_reclaim_contig:
2700  *
2701  *	Reclaim allocated, contiguous physical memory satisfying the specified
2702  *	conditions by relocating the virtual pages using that physical memory.
2703  *	Returns true if reclamation is successful and false otherwise.  Since
2704  *	relocation requires the allocation of physical pages, reclamation may
2705  *	fail due to a shortage of free pages.  When reclamation fails, callers
2706  *	are expected to perform vm_wait() before retrying a failed allocation
2707  *	operation, e.g., vm_page_alloc_contig().
2708  *
2709  *	The caller must always specify an allocation class through "req".
2710  *
2711  *	allocation classes:
2712  *	VM_ALLOC_NORMAL		normal process request
2713  *	VM_ALLOC_SYSTEM		system *really* needs a page
2714  *	VM_ALLOC_INTERRUPT	interrupt time request
2715  *
2716  *	The optional allocation flags are ignored.
2717  *
2718  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2719  *	must be a power of two.
2720  */
2721 bool
2722 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2723     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2724 {
2725 	struct vm_domain *vmd;
2726 	vm_paddr_t curr_low;
2727 	vm_page_t m_run, m_runs[NRUNS];
2728 	u_long count, reclaimed;
2729 	int error, i, options, req_class;
2730 
2731 	KASSERT(npages > 0, ("npages is 0"));
2732 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2733 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2734 	req_class = req & VM_ALLOC_CLASS_MASK;
2735 
2736 	/*
2737 	 * The page daemon is allowed to dig deeper into the free page list.
2738 	 */
2739 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2740 		req_class = VM_ALLOC_SYSTEM;
2741 
2742 	/*
2743 	 * Return if the number of free pages cannot satisfy the requested
2744 	 * allocation.
2745 	 */
2746 	vmd = VM_DOMAIN(domain);
2747 	count = vmd->vmd_free_count;
2748 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2749 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2750 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2751 		return (false);
2752 
2753 	/*
2754 	 * Scan up to three times, relaxing the restrictions ("options") on
2755 	 * the reclamation of reservations and superpages each time.
2756 	 */
2757 	for (options = VPSC_NORESERV;;) {
2758 		/*
2759 		 * Find the highest runs that satisfy the given constraints
2760 		 * and restrictions, and record them in "m_runs".
2761 		 */
2762 		curr_low = low;
2763 		count = 0;
2764 		for (;;) {
2765 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2766 			    high, alignment, boundary, options);
2767 			if (m_run == NULL)
2768 				break;
2769 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2770 			m_runs[RUN_INDEX(count)] = m_run;
2771 			count++;
2772 		}
2773 
2774 		/*
2775 		 * Reclaim the highest runs in LIFO (descending) order until
2776 		 * the number of reclaimed pages, "reclaimed", is at least
2777 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2778 		 * reclamation is idempotent, and runs will (likely) recur
2779 		 * from one scan to the next as restrictions are relaxed.
2780 		 */
2781 		reclaimed = 0;
2782 		for (i = 0; count > 0 && i < NRUNS; i++) {
2783 			count--;
2784 			m_run = m_runs[RUN_INDEX(count)];
2785 			error = vm_page_reclaim_run(req_class, domain, npages,
2786 			    m_run, high);
2787 			if (error == 0) {
2788 				reclaimed += npages;
2789 				if (reclaimed >= MIN_RECLAIM)
2790 					return (true);
2791 			}
2792 		}
2793 
2794 		/*
2795 		 * Either relax the restrictions on the next scan or return if
2796 		 * the last scan had no restrictions.
2797 		 */
2798 		if (options == VPSC_NORESERV)
2799 			options = VPSC_NOSUPER;
2800 		else if (options == VPSC_NOSUPER)
2801 			options = VPSC_ANY;
2802 		else if (options == VPSC_ANY)
2803 			return (reclaimed != 0);
2804 	}
2805 }
2806 
2807 bool
2808 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2809     u_long alignment, vm_paddr_t boundary)
2810 {
2811 	struct vm_domainset_iter di;
2812 	int domain;
2813 	bool ret;
2814 
2815 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2816 	do {
2817 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2818 		    high, alignment, boundary);
2819 		if (ret)
2820 			break;
2821 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2822 
2823 	return (ret);
2824 }
2825 
2826 /*
2827  * Set the domain in the appropriate page level domainset.
2828  */
2829 void
2830 vm_domain_set(struct vm_domain *vmd)
2831 {
2832 
2833 	mtx_lock(&vm_domainset_lock);
2834 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2835 		vmd->vmd_minset = 1;
2836 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2837 	}
2838 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2839 		vmd->vmd_severeset = 1;
2840 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2841 	}
2842 	mtx_unlock(&vm_domainset_lock);
2843 }
2844 
2845 /*
2846  * Clear the domain from the appropriate page level domainset.
2847  */
2848 void
2849 vm_domain_clear(struct vm_domain *vmd)
2850 {
2851 
2852 	mtx_lock(&vm_domainset_lock);
2853 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2854 		vmd->vmd_minset = 0;
2855 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2856 		if (vm_min_waiters != 0) {
2857 			vm_min_waiters = 0;
2858 			wakeup(&vm_min_domains);
2859 		}
2860 	}
2861 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2862 		vmd->vmd_severeset = 0;
2863 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2864 		if (vm_severe_waiters != 0) {
2865 			vm_severe_waiters = 0;
2866 			wakeup(&vm_severe_domains);
2867 		}
2868 	}
2869 
2870 	/*
2871 	 * If pageout daemon needs pages, then tell it that there are
2872 	 * some free.
2873 	 */
2874 	if (vmd->vmd_pageout_pages_needed &&
2875 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2876 		wakeup(&vmd->vmd_pageout_pages_needed);
2877 		vmd->vmd_pageout_pages_needed = 0;
2878 	}
2879 
2880 	/* See comments in vm_wait_doms(). */
2881 	if (vm_pageproc_waiters) {
2882 		vm_pageproc_waiters = 0;
2883 		wakeup(&vm_pageproc_waiters);
2884 	}
2885 	mtx_unlock(&vm_domainset_lock);
2886 }
2887 
2888 /*
2889  * Wait for free pages to exceed the min threshold globally.
2890  */
2891 void
2892 vm_wait_min(void)
2893 {
2894 
2895 	mtx_lock(&vm_domainset_lock);
2896 	while (vm_page_count_min()) {
2897 		vm_min_waiters++;
2898 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2899 	}
2900 	mtx_unlock(&vm_domainset_lock);
2901 }
2902 
2903 /*
2904  * Wait for free pages to exceed the severe threshold globally.
2905  */
2906 void
2907 vm_wait_severe(void)
2908 {
2909 
2910 	mtx_lock(&vm_domainset_lock);
2911 	while (vm_page_count_severe()) {
2912 		vm_severe_waiters++;
2913 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2914 		    "vmwait", 0);
2915 	}
2916 	mtx_unlock(&vm_domainset_lock);
2917 }
2918 
2919 u_int
2920 vm_wait_count(void)
2921 {
2922 
2923 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2924 }
2925 
2926 static void
2927 vm_wait_doms(const domainset_t *wdoms)
2928 {
2929 
2930 	/*
2931 	 * We use racey wakeup synchronization to avoid expensive global
2932 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
2933 	 * To handle this, we only sleep for one tick in this instance.  It
2934 	 * is expected that most allocations for the pageproc will come from
2935 	 * kmem or vm_page_grab* which will use the more specific and
2936 	 * race-free vm_wait_domain().
2937 	 */
2938 	if (curproc == pageproc) {
2939 		mtx_lock(&vm_domainset_lock);
2940 		vm_pageproc_waiters++;
2941 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2942 		    "pageprocwait", 1);
2943 	} else {
2944 		/*
2945 		 * XXX Ideally we would wait only until the allocation could
2946 		 * be satisfied.  This condition can cause new allocators to
2947 		 * consume all freed pages while old allocators wait.
2948 		 */
2949 		mtx_lock(&vm_domainset_lock);
2950 		if (DOMAINSET_SUBSET(&vm_min_domains, wdoms)) {
2951 			vm_min_waiters++;
2952 			msleep(&vm_min_domains, &vm_domainset_lock, PVM,
2953 			    "vmwait", 0);
2954 		}
2955 		mtx_unlock(&vm_domainset_lock);
2956 	}
2957 }
2958 
2959 /*
2960  *	vm_wait_domain:
2961  *
2962  *	Sleep until free pages are available for allocation.
2963  *	- Called in various places after failed memory allocations.
2964  */
2965 void
2966 vm_wait_domain(int domain)
2967 {
2968 	struct vm_domain *vmd;
2969 	domainset_t wdom;
2970 
2971 	vmd = VM_DOMAIN(domain);
2972 	vm_domain_free_assert_unlocked(vmd);
2973 
2974 	if (curproc == pageproc) {
2975 		mtx_lock(&vm_domainset_lock);
2976 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2977 			vmd->vmd_pageout_pages_needed = 1;
2978 			msleep(&vmd->vmd_pageout_pages_needed,
2979 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
2980 		} else
2981 			mtx_unlock(&vm_domainset_lock);
2982 	} else {
2983 		if (pageproc == NULL)
2984 			panic("vm_wait in early boot");
2985 		DOMAINSET_ZERO(&wdom);
2986 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
2987 		vm_wait_doms(&wdom);
2988 	}
2989 }
2990 
2991 /*
2992  *	vm_wait:
2993  *
2994  *	Sleep until free pages are available for allocation in the
2995  *	affinity domains of the obj.  If obj is NULL, the domain set
2996  *	for the calling thread is used.
2997  *	Called in various places after failed memory allocations.
2998  */
2999 void
3000 vm_wait(vm_object_t obj)
3001 {
3002 	struct domainset *d;
3003 
3004 	d = NULL;
3005 
3006 	/*
3007 	 * Carefully fetch pointers only once: the struct domainset
3008 	 * itself is ummutable but the pointer might change.
3009 	 */
3010 	if (obj != NULL)
3011 		d = obj->domain.dr_policy;
3012 	if (d == NULL)
3013 		d = curthread->td_domain.dr_policy;
3014 
3015 	vm_wait_doms(&d->ds_mask);
3016 }
3017 
3018 /*
3019  *	vm_domain_alloc_fail:
3020  *
3021  *	Called when a page allocation function fails.  Informs the
3022  *	pagedaemon and performs the requested wait.  Requires the
3023  *	domain_free and object lock on entry.  Returns with the
3024  *	object lock held and free lock released.  Returns an error when
3025  *	retry is necessary.
3026  *
3027  */
3028 static int
3029 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3030 {
3031 
3032 	vm_domain_free_assert_unlocked(vmd);
3033 
3034 	atomic_add_int(&vmd->vmd_pageout_deficit,
3035 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3036 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3037 		if (object != NULL)
3038 			VM_OBJECT_WUNLOCK(object);
3039 		vm_wait_domain(vmd->vmd_domain);
3040 		if (object != NULL)
3041 			VM_OBJECT_WLOCK(object);
3042 		if (req & VM_ALLOC_WAITOK)
3043 			return (EAGAIN);
3044 	}
3045 
3046 	return (0);
3047 }
3048 
3049 /*
3050  *	vm_waitpfault:
3051  *
3052  *	Sleep until free pages are available for allocation.
3053  *	- Called only in vm_fault so that processes page faulting
3054  *	  can be easily tracked.
3055  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3056  *	  processes will be able to grab memory first.  Do not change
3057  *	  this balance without careful testing first.
3058  */
3059 void
3060 vm_waitpfault(void)
3061 {
3062 
3063 	mtx_lock(&vm_domainset_lock);
3064 	if (vm_page_count_min()) {
3065 		vm_min_waiters++;
3066 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER, "pfault", 0);
3067 	}
3068 	mtx_unlock(&vm_domainset_lock);
3069 }
3070 
3071 struct vm_pagequeue *
3072 vm_page_pagequeue(vm_page_t m)
3073 {
3074 
3075 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
3076 }
3077 
3078 static struct mtx *
3079 vm_page_pagequeue_lockptr(vm_page_t m)
3080 {
3081 	uint8_t queue;
3082 
3083 	if ((queue = m->queue) == PQ_NONE)
3084 		return (NULL);
3085 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex);
3086 }
3087 
3088 static inline void
3089 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3090 {
3091 	struct vm_domain *vmd;
3092 	uint8_t aflags;
3093 
3094 	vm_pagequeue_assert_locked(pq);
3095 	KASSERT(pq == vm_page_pagequeue(m),
3096 	    ("page %p doesn't belong to %p", m, pq));
3097 
3098 	aflags = m->aflags;
3099 	if ((aflags & PGA_DEQUEUE) != 0) {
3100 		if (__predict_true((aflags & PGA_ENQUEUED) != 0)) {
3101 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3102 			vm_pagequeue_cnt_dec(pq);
3103 		}
3104 		vm_page_dequeue_complete(m);
3105 	} else if ((aflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3106 		if ((aflags & PGA_ENQUEUED) != 0)
3107 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3108 		else {
3109 			vm_pagequeue_cnt_inc(pq);
3110 			vm_page_aflag_set(m, PGA_ENQUEUED);
3111 		}
3112 		if ((aflags & PGA_REQUEUE_HEAD) != 0) {
3113 			KASSERT(m->queue == PQ_INACTIVE,
3114 			    ("head enqueue not supported for page %p", m));
3115 			vmd = vm_pagequeue_domain(m);
3116 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3117 		} else
3118 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3119 
3120 		/*
3121 		 * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after
3122 		 * setting PGA_ENQUEUED in order to synchronize with the
3123 		 * page daemon.
3124 		 */
3125 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
3126 	}
3127 }
3128 
3129 static void
3130 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3131     uint8_t queue)
3132 {
3133 	vm_page_t m;
3134 	int i;
3135 
3136 	for (i = 0; i < bq->bq_cnt; i++) {
3137 		m = bq->bq_pa[i];
3138 		if (__predict_false(m->queue != queue))
3139 			continue;
3140 		vm_pqbatch_process_page(pq, m);
3141 	}
3142 	vm_batchqueue_init(bq);
3143 }
3144 
3145 static void
3146 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3147 {
3148 	struct vm_batchqueue *bq;
3149 	struct vm_pagequeue *pq;
3150 	int domain;
3151 
3152 	vm_page_assert_locked(m);
3153 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3154 
3155 	domain = vm_phys_domain(m);
3156 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3157 
3158 	critical_enter();
3159 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3160 	if (vm_batchqueue_insert(bq, m)) {
3161 		critical_exit();
3162 		return;
3163 	}
3164 	if (!vm_pagequeue_trylock(pq)) {
3165 		critical_exit();
3166 		vm_pagequeue_lock(pq);
3167 		critical_enter();
3168 		bq = DPCPU_PTR(pqbatch[domain][queue]);
3169 	}
3170 	vm_pqbatch_process(pq, bq, queue);
3171 
3172 	/*
3173 	 * The page may have been logically dequeued before we acquired the
3174 	 * page queue lock.  In this case, the page lock prevents the page
3175 	 * from being logically enqueued elsewhere.
3176 	 */
3177 	if (__predict_true(m->queue == queue))
3178 		vm_pqbatch_process_page(pq, m);
3179 	else {
3180 		KASSERT(m->queue == PQ_NONE,
3181 		    ("invalid queue transition for page %p", m));
3182 		KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3183 		    ("page %p is enqueued with invalid queue index", m));
3184 		vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3185 	}
3186 	vm_pagequeue_unlock(pq);
3187 	critical_exit();
3188 }
3189 
3190 /*
3191  *	vm_page_drain_pqbatch:		[ internal use only ]
3192  *
3193  *	Force all per-CPU page queue batch queues to be drained.  This is
3194  *	intended for use in severe memory shortages, to ensure that pages
3195  *	do not remain stuck in the batch queues.
3196  */
3197 void
3198 vm_page_drain_pqbatch(void)
3199 {
3200 	struct thread *td;
3201 	struct vm_domain *vmd;
3202 	struct vm_pagequeue *pq;
3203 	int cpu, domain, queue;
3204 
3205 	td = curthread;
3206 	CPU_FOREACH(cpu) {
3207 		thread_lock(td);
3208 		sched_bind(td, cpu);
3209 		thread_unlock(td);
3210 
3211 		for (domain = 0; domain < vm_ndomains; domain++) {
3212 			vmd = VM_DOMAIN(domain);
3213 			for (queue = 0; queue < PQ_COUNT; queue++) {
3214 				pq = &vmd->vmd_pagequeues[queue];
3215 				vm_pagequeue_lock(pq);
3216 				critical_enter();
3217 				vm_pqbatch_process(pq,
3218 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3219 				critical_exit();
3220 				vm_pagequeue_unlock(pq);
3221 			}
3222 		}
3223 	}
3224 	thread_lock(td);
3225 	sched_unbind(td);
3226 	thread_unlock(td);
3227 }
3228 
3229 /*
3230  * Complete the logical removal of a page from a page queue.  We must be
3231  * careful to synchronize with the page daemon, which may be concurrently
3232  * examining the page with only the page lock held.  The page must not be
3233  * in a state where it appears to be logically enqueued.
3234  */
3235 static void
3236 vm_page_dequeue_complete(vm_page_t m)
3237 {
3238 
3239 	m->queue = PQ_NONE;
3240 	atomic_thread_fence_rel();
3241 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3242 }
3243 
3244 /*
3245  *	vm_page_dequeue_deferred:	[ internal use only ]
3246  *
3247  *	Request removal of the given page from its current page
3248  *	queue.  Physical removal from the queue may be deferred
3249  *	indefinitely.
3250  *
3251  *	The page must be locked.
3252  */
3253 void
3254 vm_page_dequeue_deferred(vm_page_t m)
3255 {
3256 	int queue;
3257 
3258 	vm_page_assert_locked(m);
3259 
3260 	queue = m->queue;
3261 	if (queue == PQ_NONE) {
3262 		KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3263 		    ("page %p has queue state", m));
3264 		return;
3265 	}
3266 	if ((m->aflags & PGA_DEQUEUE) == 0)
3267 		vm_page_aflag_set(m, PGA_DEQUEUE);
3268 	vm_pqbatch_submit_page(m, queue);
3269 }
3270 
3271 /*
3272  *	vm_page_dequeue_locked:
3273  *
3274  *	Remove the page from its page queue, which must be locked.
3275  *	If the page lock is not held, there is no guarantee that the
3276  *	page will not be enqueued by another thread before this function
3277  *	returns.  In this case, it is up to the caller to ensure that
3278  *	no other threads hold a reference to the page.
3279  *
3280  *	The page queue lock must be held.  If the page is not already
3281  *	logically dequeued, the page lock must be held as well.
3282  */
3283 void
3284 vm_page_dequeue_locked(vm_page_t m)
3285 {
3286 	struct vm_pagequeue *pq;
3287 
3288 	pq = vm_page_pagequeue(m);
3289 
3290 	KASSERT(m->queue != PQ_NONE,
3291 	    ("%s: page %p queue field is PQ_NONE", __func__, m));
3292 	vm_pagequeue_assert_locked(pq);
3293 	KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3294 	    mtx_owned(vm_page_lockptr(m)),
3295 	    ("%s: queued unlocked page %p", __func__, m));
3296 
3297 	if ((m->aflags & PGA_ENQUEUED) != 0) {
3298 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3299 		vm_pagequeue_cnt_dec(pq);
3300 	}
3301 	vm_page_dequeue_complete(m);
3302 }
3303 
3304 /*
3305  *	vm_page_dequeue:
3306  *
3307  *	Remove the page from whichever page queue it's in, if any.
3308  *	If the page lock is not held, there is no guarantee that the
3309  *	page will not be enqueued by another thread before this function
3310  *	returns.  In this case, it is up to the caller to ensure that
3311  *	no other threads hold a reference to the page.
3312  */
3313 void
3314 vm_page_dequeue(vm_page_t m)
3315 {
3316 	struct mtx *lock, *lock1;
3317 
3318 	lock = vm_page_pagequeue_lockptr(m);
3319 	for (;;) {
3320 		if (lock == NULL)
3321 			return;
3322 		mtx_lock(lock);
3323 		if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock)
3324 			break;
3325 		mtx_unlock(lock);
3326 		lock = lock1;
3327 	}
3328 	KASSERT(lock == vm_page_pagequeue_lockptr(m),
3329 	    ("%s: page %p migrated directly between queues", __func__, m));
3330 	vm_page_dequeue_locked(m);
3331 	mtx_unlock(lock);
3332 }
3333 
3334 /*
3335  * Schedule the given page for insertion into the specified page queue.
3336  * Physical insertion of the page may be deferred indefinitely.
3337  */
3338 static void
3339 vm_page_enqueue(vm_page_t m, uint8_t queue)
3340 {
3341 
3342 	vm_page_assert_locked(m);
3343 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3344 	    ("%s: page %p is already enqueued", __func__, m));
3345 
3346 	m->queue = queue;
3347 	if ((m->aflags & PGA_REQUEUE) == 0)
3348 		vm_page_aflag_set(m, PGA_REQUEUE);
3349 	vm_pqbatch_submit_page(m, queue);
3350 }
3351 
3352 /*
3353  *	vm_page_requeue:		[ internal use only ]
3354  *
3355  *	Schedule a requeue of the given page.
3356  *
3357  *	The page must be locked.
3358  */
3359 void
3360 vm_page_requeue(vm_page_t m)
3361 {
3362 
3363 	vm_page_assert_locked(m);
3364 	KASSERT(m->queue != PQ_NONE,
3365 	    ("%s: page %p is not logically enqueued", __func__, m));
3366 
3367 	if ((m->aflags & PGA_REQUEUE) == 0)
3368 		vm_page_aflag_set(m, PGA_REQUEUE);
3369 	vm_pqbatch_submit_page(m, m->queue);
3370 }
3371 
3372 /*
3373  *	vm_page_activate:
3374  *
3375  *	Put the specified page on the active list (if appropriate).
3376  *	Ensure that act_count is at least ACT_INIT but do not otherwise
3377  *	mess with it.
3378  *
3379  *	The page must be locked.
3380  */
3381 void
3382 vm_page_activate(vm_page_t m)
3383 {
3384 	int queue;
3385 
3386 	vm_page_assert_locked(m);
3387 
3388 	if ((queue = vm_page_queue(m)) == PQ_ACTIVE || m->wire_count > 0 ||
3389 	    (m->oflags & VPO_UNMANAGED) != 0) {
3390 		if (queue == PQ_ACTIVE && m->act_count < ACT_INIT)
3391 			m->act_count = ACT_INIT;
3392 		return;
3393 	}
3394 
3395 	vm_page_remque(m);
3396 	if (m->act_count < ACT_INIT)
3397 		m->act_count = ACT_INIT;
3398 	vm_page_enqueue(m, PQ_ACTIVE);
3399 }
3400 
3401 /*
3402  *	vm_page_free_prep:
3403  *
3404  *	Prepares the given page to be put on the free list,
3405  *	disassociating it from any VM object. The caller may return
3406  *	the page to the free list only if this function returns true.
3407  *
3408  *	The object must be locked.  The page must be locked if it is
3409  *	managed.
3410  */
3411 bool
3412 vm_page_free_prep(vm_page_t m)
3413 {
3414 
3415 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3416 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3417 		uint64_t *p;
3418 		int i;
3419 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3420 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3421 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3422 			    m, i, (uintmax_t)*p));
3423 	}
3424 #endif
3425 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3426 		vm_page_lock_assert(m, MA_OWNED);
3427 		KASSERT(!pmap_page_is_mapped(m),
3428 		    ("vm_page_free_prep: freeing mapped page %p", m));
3429 	} else
3430 		KASSERT(m->queue == PQ_NONE,
3431 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3432 	VM_CNT_INC(v_tfree);
3433 
3434 	if (vm_page_sbusied(m))
3435 		panic("vm_page_free_prep: freeing busy page %p", m);
3436 
3437 	vm_page_remove(m);
3438 
3439 	/*
3440 	 * If fictitious remove object association and
3441 	 * return.
3442 	 */
3443 	if ((m->flags & PG_FICTITIOUS) != 0) {
3444 		KASSERT(m->wire_count == 1,
3445 		    ("fictitious page %p is not wired", m));
3446 		KASSERT(m->queue == PQ_NONE,
3447 		    ("fictitious page %p is queued", m));
3448 		return (false);
3449 	}
3450 
3451 	/*
3452 	 * Pages need not be dequeued before they are returned to the physical
3453 	 * memory allocator, but they must at least be marked for a deferred
3454 	 * dequeue.
3455 	 */
3456 	if ((m->oflags & VPO_UNMANAGED) == 0)
3457 		vm_page_dequeue_deferred(m);
3458 
3459 	m->valid = 0;
3460 	vm_page_undirty(m);
3461 
3462 	if (m->wire_count != 0)
3463 		panic("vm_page_free_prep: freeing wired page %p", m);
3464 	if (m->hold_count != 0) {
3465 		m->flags &= ~PG_ZERO;
3466 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3467 		    ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m));
3468 		m->flags |= PG_UNHOLDFREE;
3469 		return (false);
3470 	}
3471 
3472 	/*
3473 	 * Restore the default memory attribute to the page.
3474 	 */
3475 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3476 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3477 
3478 #if VM_NRESERVLEVEL > 0
3479 	if (vm_reserv_free_page(m))
3480 		return (false);
3481 #endif
3482 
3483 	return (true);
3484 }
3485 
3486 /*
3487  *	vm_page_free_toq:
3488  *
3489  *	Returns the given page to the free list, disassociating it
3490  *	from any VM object.
3491  *
3492  *	The object must be locked.  The page must be locked if it is
3493  *	managed.
3494  */
3495 void
3496 vm_page_free_toq(vm_page_t m)
3497 {
3498 	struct vm_domain *vmd;
3499 
3500 	if (!vm_page_free_prep(m))
3501 		return;
3502 
3503 	vmd = vm_pagequeue_domain(m);
3504 	if (m->pool == VM_FREEPOOL_DEFAULT && vmd->vmd_pgcache != NULL) {
3505 		uma_zfree(vmd->vmd_pgcache, m);
3506 		return;
3507 	}
3508 	vm_domain_free_lock(vmd);
3509 	vm_phys_free_pages(m, 0);
3510 	vm_domain_free_unlock(vmd);
3511 	vm_domain_freecnt_inc(vmd, 1);
3512 }
3513 
3514 /*
3515  *	vm_page_free_pages_toq:
3516  *
3517  *	Returns a list of pages to the free list, disassociating it
3518  *	from any VM object.  In other words, this is equivalent to
3519  *	calling vm_page_free_toq() for each page of a list of VM objects.
3520  *
3521  *	The objects must be locked.  The pages must be locked if it is
3522  *	managed.
3523  */
3524 void
3525 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3526 {
3527 	vm_page_t m;
3528 	int count;
3529 
3530 	if (SLIST_EMPTY(free))
3531 		return;
3532 
3533 	count = 0;
3534 	while ((m = SLIST_FIRST(free)) != NULL) {
3535 		count++;
3536 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3537 		vm_page_free_toq(m);
3538 	}
3539 
3540 	if (update_wire_count)
3541 		vm_wire_sub(count);
3542 }
3543 
3544 /*
3545  *	vm_page_wire:
3546  *
3547  * Mark this page as wired down.  If the page is fictitious, then
3548  * its wire count must remain one.
3549  *
3550  * The page must be locked.
3551  */
3552 void
3553 vm_page_wire(vm_page_t m)
3554 {
3555 
3556 	vm_page_assert_locked(m);
3557 	if ((m->flags & PG_FICTITIOUS) != 0) {
3558 		KASSERT(m->wire_count == 1,
3559 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
3560 		    m));
3561 		return;
3562 	}
3563 	if (m->wire_count == 0) {
3564 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3565 		    m->queue == PQ_NONE,
3566 		    ("vm_page_wire: unmanaged page %p is queued", m));
3567 		vm_wire_add(1);
3568 	}
3569 	m->wire_count++;
3570 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3571 }
3572 
3573 /*
3574  * vm_page_unwire:
3575  *
3576  * Release one wiring of the specified page, potentially allowing it to be
3577  * paged out.  Returns TRUE if the number of wirings transitions to zero and
3578  * FALSE otherwise.
3579  *
3580  * Only managed pages belonging to an object can be paged out.  If the number
3581  * of wirings transitions to zero and the page is eligible for page out, then
3582  * the page is added to the specified paging queue (unless PQ_NONE is
3583  * specified, in which case the page is dequeued if it belongs to a paging
3584  * queue).
3585  *
3586  * If a page is fictitious, then its wire count must always be one.
3587  *
3588  * A managed page must be locked.
3589  */
3590 bool
3591 vm_page_unwire(vm_page_t m, uint8_t queue)
3592 {
3593 	bool unwired;
3594 
3595 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3596 	    ("vm_page_unwire: invalid queue %u request for page %p",
3597 	    queue, m));
3598 	if ((m->oflags & VPO_UNMANAGED) == 0)
3599 		vm_page_assert_locked(m);
3600 
3601 	unwired = vm_page_unwire_noq(m);
3602 	if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3603 		return (unwired);
3604 
3605 	if (vm_page_queue(m) == queue) {
3606 		if (queue == PQ_ACTIVE)
3607 			vm_page_reference(m);
3608 		else if (queue != PQ_NONE)
3609 			vm_page_requeue(m);
3610 	} else {
3611 		vm_page_dequeue(m);
3612 		if (queue != PQ_NONE) {
3613 			vm_page_enqueue(m, queue);
3614 			if (queue == PQ_ACTIVE)
3615 				/* Initialize act_count. */
3616 				vm_page_activate(m);
3617 		}
3618 	}
3619 	return (unwired);
3620 }
3621 
3622 /*
3623  *
3624  * vm_page_unwire_noq:
3625  *
3626  * Unwire a page without (re-)inserting it into a page queue.  It is up
3627  * to the caller to enqueue, requeue, or free the page as appropriate.
3628  * In most cases, vm_page_unwire() should be used instead.
3629  */
3630 bool
3631 vm_page_unwire_noq(vm_page_t m)
3632 {
3633 
3634 	if ((m->oflags & VPO_UNMANAGED) == 0)
3635 		vm_page_assert_locked(m);
3636 	if ((m->flags & PG_FICTITIOUS) != 0) {
3637 		KASSERT(m->wire_count == 1,
3638 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3639 		return (false);
3640 	}
3641 	if (m->wire_count == 0)
3642 		panic("vm_page_unwire: page %p's wire count is zero", m);
3643 	m->wire_count--;
3644 	if (m->wire_count == 0) {
3645 		vm_wire_sub(1);
3646 		return (true);
3647 	} else
3648 		return (false);
3649 }
3650 
3651 /*
3652  * Move the specified page to the tail of the inactive queue, or requeue
3653  * the page if it is already in the inactive queue.
3654  *
3655  * The page must be locked.
3656  */
3657 void
3658 vm_page_deactivate(vm_page_t m)
3659 {
3660 
3661 	vm_page_assert_locked(m);
3662 
3663 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3664 		return;
3665 
3666 	if (!vm_page_inactive(m)) {
3667 		vm_page_remque(m);
3668 		vm_page_enqueue(m, PQ_INACTIVE);
3669 	} else
3670 		vm_page_requeue(m);
3671 }
3672 
3673 /*
3674  * Move the specified page close to the head of the inactive queue,
3675  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3676  * As with regular enqueues, we use a per-CPU batch queue to reduce
3677  * contention on the page queue lock.
3678  *
3679  * The page must be locked.
3680  */
3681 void
3682 vm_page_deactivate_noreuse(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_remque(m);
3692 	m->queue = PQ_INACTIVE;
3693 	if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3694 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3695 	vm_pqbatch_submit_page(m, PQ_INACTIVE);
3696 }
3697 
3698 /*
3699  * vm_page_launder
3700  *
3701  * 	Put a page in the laundry, or requeue it if it is already there.
3702  */
3703 void
3704 vm_page_launder(vm_page_t m)
3705 {
3706 
3707 	vm_page_assert_locked(m);
3708 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3709 		return;
3710 
3711 	if (vm_page_in_laundry(m))
3712 		vm_page_requeue(m);
3713 	else {
3714 		vm_page_remque(m);
3715 		vm_page_enqueue(m, PQ_LAUNDRY);
3716 	}
3717 }
3718 
3719 /*
3720  * vm_page_unswappable
3721  *
3722  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3723  */
3724 void
3725 vm_page_unswappable(vm_page_t m)
3726 {
3727 
3728 	vm_page_assert_locked(m);
3729 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3730 	    ("page %p already unswappable", m));
3731 
3732 	vm_page_remque(m);
3733 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
3734 }
3735 
3736 /*
3737  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3738  * if the page is freed and false otherwise.
3739  *
3740  * The page must be managed.  The page and its containing object must be
3741  * locked.
3742  */
3743 bool
3744 vm_page_try_to_free(vm_page_t m)
3745 {
3746 
3747 	vm_page_assert_locked(m);
3748 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3749 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3750 	if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m))
3751 		return (false);
3752 	if (m->object->ref_count != 0) {
3753 		pmap_remove_all(m);
3754 		if (m->dirty != 0)
3755 			return (false);
3756 	}
3757 	vm_page_free(m);
3758 	return (true);
3759 }
3760 
3761 /*
3762  * vm_page_advise
3763  *
3764  * 	Apply the specified advice to the given page.
3765  *
3766  *	The object and page must be locked.
3767  */
3768 void
3769 vm_page_advise(vm_page_t m, int advice)
3770 {
3771 
3772 	vm_page_assert_locked(m);
3773 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3774 	if (advice == MADV_FREE)
3775 		/*
3776 		 * Mark the page clean.  This will allow the page to be freed
3777 		 * without first paging it out.  MADV_FREE pages are often
3778 		 * quickly reused by malloc(3), so we do not do anything that
3779 		 * would result in a page fault on a later access.
3780 		 */
3781 		vm_page_undirty(m);
3782 	else if (advice != MADV_DONTNEED) {
3783 		if (advice == MADV_WILLNEED)
3784 			vm_page_activate(m);
3785 		return;
3786 	}
3787 
3788 	/*
3789 	 * Clear any references to the page.  Otherwise, the page daemon will
3790 	 * immediately reactivate the page.
3791 	 */
3792 	vm_page_aflag_clear(m, PGA_REFERENCED);
3793 
3794 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3795 		vm_page_dirty(m);
3796 
3797 	/*
3798 	 * Place clean pages near the head of the inactive queue rather than
3799 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3800 	 * the page will be reused quickly.  Dirty pages not already in the
3801 	 * laundry are moved there.
3802 	 */
3803 	if (m->dirty == 0)
3804 		vm_page_deactivate_noreuse(m);
3805 	else if (!vm_page_in_laundry(m))
3806 		vm_page_launder(m);
3807 }
3808 
3809 /*
3810  * Grab a page, waiting until we are waken up due to the page
3811  * changing state.  We keep on waiting, if the page continues
3812  * to be in the object.  If the page doesn't exist, first allocate it
3813  * and then conditionally zero it.
3814  *
3815  * This routine may sleep.
3816  *
3817  * The object must be locked on entry.  The lock will, however, be released
3818  * and reacquired if the routine sleeps.
3819  */
3820 vm_page_t
3821 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3822 {
3823 	vm_page_t m;
3824 	int sleep;
3825 	int pflags;
3826 
3827 	VM_OBJECT_ASSERT_WLOCKED(object);
3828 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3829 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3830 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3831 	pflags = allocflags &
3832 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3833 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3834 		pflags |= VM_ALLOC_WAITFAIL;
3835 retrylookup:
3836 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3837 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3838 		    vm_page_xbusied(m) : vm_page_busied(m);
3839 		if (sleep) {
3840 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3841 				return (NULL);
3842 			/*
3843 			 * Reference the page before unlocking and
3844 			 * sleeping so that the page daemon is less
3845 			 * likely to reclaim it.
3846 			 */
3847 			vm_page_aflag_set(m, PGA_REFERENCED);
3848 			vm_page_lock(m);
3849 			VM_OBJECT_WUNLOCK(object);
3850 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3851 			    VM_ALLOC_IGN_SBUSY) != 0);
3852 			VM_OBJECT_WLOCK(object);
3853 			goto retrylookup;
3854 		} else {
3855 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3856 				vm_page_lock(m);
3857 				vm_page_wire(m);
3858 				vm_page_unlock(m);
3859 			}
3860 			if ((allocflags &
3861 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3862 				vm_page_xbusy(m);
3863 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3864 				vm_page_sbusy(m);
3865 			return (m);
3866 		}
3867 	}
3868 	m = vm_page_alloc(object, pindex, pflags);
3869 	if (m == NULL) {
3870 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3871 			return (NULL);
3872 		goto retrylookup;
3873 	}
3874 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3875 		pmap_zero_page(m);
3876 	return (m);
3877 }
3878 
3879 /*
3880  * Return the specified range of pages from the given object.  For each
3881  * page offset within the range, if a page already exists within the object
3882  * at that offset and it is busy, then wait for it to change state.  If,
3883  * instead, the page doesn't exist, then allocate it.
3884  *
3885  * The caller must always specify an allocation class.
3886  *
3887  * allocation classes:
3888  *	VM_ALLOC_NORMAL		normal process request
3889  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3890  *
3891  * The caller must always specify that the pages are to be busied and/or
3892  * wired.
3893  *
3894  * optional allocation flags:
3895  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3896  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3897  *	VM_ALLOC_NOWAIT		do not sleep
3898  *	VM_ALLOC_SBUSY		set page to sbusy state
3899  *	VM_ALLOC_WIRED		wire the pages
3900  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3901  *
3902  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3903  * may return a partial prefix of the requested range.
3904  */
3905 int
3906 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3907     vm_page_t *ma, int count)
3908 {
3909 	vm_page_t m, mpred;
3910 	int pflags;
3911 	int i;
3912 	bool sleep;
3913 
3914 	VM_OBJECT_ASSERT_WLOCKED(object);
3915 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3916 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3917 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3918 	    (allocflags & VM_ALLOC_WIRED) != 0,
3919 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3920 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3921 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3922 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3923 	if (count == 0)
3924 		return (0);
3925 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3926 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3927 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3928 		pflags |= VM_ALLOC_WAITFAIL;
3929 	i = 0;
3930 retrylookup:
3931 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3932 	if (m == NULL || m->pindex != pindex + i) {
3933 		mpred = m;
3934 		m = NULL;
3935 	} else
3936 		mpred = TAILQ_PREV(m, pglist, listq);
3937 	for (; i < count; i++) {
3938 		if (m != NULL) {
3939 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3940 			    vm_page_xbusied(m) : vm_page_busied(m);
3941 			if (sleep) {
3942 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3943 					break;
3944 				/*
3945 				 * Reference the page before unlocking and
3946 				 * sleeping so that the page daemon is less
3947 				 * likely to reclaim it.
3948 				 */
3949 				vm_page_aflag_set(m, PGA_REFERENCED);
3950 				vm_page_lock(m);
3951 				VM_OBJECT_WUNLOCK(object);
3952 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
3953 				    VM_ALLOC_IGN_SBUSY) != 0);
3954 				VM_OBJECT_WLOCK(object);
3955 				goto retrylookup;
3956 			}
3957 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3958 				vm_page_lock(m);
3959 				vm_page_wire(m);
3960 				vm_page_unlock(m);
3961 			}
3962 			if ((allocflags & (VM_ALLOC_NOBUSY |
3963 			    VM_ALLOC_SBUSY)) == 0)
3964 				vm_page_xbusy(m);
3965 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3966 				vm_page_sbusy(m);
3967 		} else {
3968 			m = vm_page_alloc_after(object, pindex + i,
3969 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
3970 			if (m == NULL) {
3971 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3972 					break;
3973 				goto retrylookup;
3974 			}
3975 		}
3976 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3977 			if ((m->flags & PG_ZERO) == 0)
3978 				pmap_zero_page(m);
3979 			m->valid = VM_PAGE_BITS_ALL;
3980 		}
3981 		ma[i] = mpred = m;
3982 		m = vm_page_next(m);
3983 	}
3984 	return (i);
3985 }
3986 
3987 /*
3988  * Mapping function for valid or dirty bits in a page.
3989  *
3990  * Inputs are required to range within a page.
3991  */
3992 vm_page_bits_t
3993 vm_page_bits(int base, int size)
3994 {
3995 	int first_bit;
3996 	int last_bit;
3997 
3998 	KASSERT(
3999 	    base + size <= PAGE_SIZE,
4000 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4001 	);
4002 
4003 	if (size == 0)		/* handle degenerate case */
4004 		return (0);
4005 
4006 	first_bit = base >> DEV_BSHIFT;
4007 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4008 
4009 	return (((vm_page_bits_t)2 << last_bit) -
4010 	    ((vm_page_bits_t)1 << first_bit));
4011 }
4012 
4013 /*
4014  *	vm_page_set_valid_range:
4015  *
4016  *	Sets portions of a page valid.  The arguments are expected
4017  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4018  *	of any partial chunks touched by the range.  The invalid portion of
4019  *	such chunks will be zeroed.
4020  *
4021  *	(base + size) must be less then or equal to PAGE_SIZE.
4022  */
4023 void
4024 vm_page_set_valid_range(vm_page_t m, int base, int size)
4025 {
4026 	int endoff, frag;
4027 
4028 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4029 	if (size == 0)	/* handle degenerate case */
4030 		return;
4031 
4032 	/*
4033 	 * If the base is not DEV_BSIZE aligned and the valid
4034 	 * bit is clear, we have to zero out a portion of the
4035 	 * first block.
4036 	 */
4037 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4038 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4039 		pmap_zero_page_area(m, frag, base - frag);
4040 
4041 	/*
4042 	 * If the ending offset is not DEV_BSIZE aligned and the
4043 	 * valid bit is clear, we have to zero out a portion of
4044 	 * the last block.
4045 	 */
4046 	endoff = base + size;
4047 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4048 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4049 		pmap_zero_page_area(m, endoff,
4050 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4051 
4052 	/*
4053 	 * Assert that no previously invalid block that is now being validated
4054 	 * is already dirty.
4055 	 */
4056 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4057 	    ("vm_page_set_valid_range: page %p is dirty", m));
4058 
4059 	/*
4060 	 * Set valid bits inclusive of any overlap.
4061 	 */
4062 	m->valid |= vm_page_bits(base, size);
4063 }
4064 
4065 /*
4066  * Clear the given bits from the specified page's dirty field.
4067  */
4068 static __inline void
4069 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4070 {
4071 	uintptr_t addr;
4072 #if PAGE_SIZE < 16384
4073 	int shift;
4074 #endif
4075 
4076 	/*
4077 	 * If the object is locked and the page is neither exclusive busy nor
4078 	 * write mapped, then the page's dirty field cannot possibly be
4079 	 * set by a concurrent pmap operation.
4080 	 */
4081 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4082 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4083 		m->dirty &= ~pagebits;
4084 	else {
4085 		/*
4086 		 * The pmap layer can call vm_page_dirty() without
4087 		 * holding a distinguished lock.  The combination of
4088 		 * the object's lock and an atomic operation suffice
4089 		 * to guarantee consistency of the page dirty field.
4090 		 *
4091 		 * For PAGE_SIZE == 32768 case, compiler already
4092 		 * properly aligns the dirty field, so no forcible
4093 		 * alignment is needed. Only require existence of
4094 		 * atomic_clear_64 when page size is 32768.
4095 		 */
4096 		addr = (uintptr_t)&m->dirty;
4097 #if PAGE_SIZE == 32768
4098 		atomic_clear_64((uint64_t *)addr, pagebits);
4099 #elif PAGE_SIZE == 16384
4100 		atomic_clear_32((uint32_t *)addr, pagebits);
4101 #else		/* PAGE_SIZE <= 8192 */
4102 		/*
4103 		 * Use a trick to perform a 32-bit atomic on the
4104 		 * containing aligned word, to not depend on the existence
4105 		 * of atomic_clear_{8, 16}.
4106 		 */
4107 		shift = addr & (sizeof(uint32_t) - 1);
4108 #if BYTE_ORDER == BIG_ENDIAN
4109 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4110 #else
4111 		shift *= NBBY;
4112 #endif
4113 		addr &= ~(sizeof(uint32_t) - 1);
4114 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
4115 #endif		/* PAGE_SIZE */
4116 	}
4117 }
4118 
4119 /*
4120  *	vm_page_set_validclean:
4121  *
4122  *	Sets portions of a page valid and clean.  The arguments are expected
4123  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4124  *	of any partial chunks touched by the range.  The invalid portion of
4125  *	such chunks will be zero'd.
4126  *
4127  *	(base + size) must be less then or equal to PAGE_SIZE.
4128  */
4129 void
4130 vm_page_set_validclean(vm_page_t m, int base, int size)
4131 {
4132 	vm_page_bits_t oldvalid, pagebits;
4133 	int endoff, frag;
4134 
4135 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4136 	if (size == 0)	/* handle degenerate case */
4137 		return;
4138 
4139 	/*
4140 	 * If the base is not DEV_BSIZE aligned and the valid
4141 	 * bit is clear, we have to zero out a portion of the
4142 	 * first block.
4143 	 */
4144 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4145 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4146 		pmap_zero_page_area(m, frag, base - frag);
4147 
4148 	/*
4149 	 * If the ending offset is not DEV_BSIZE aligned and the
4150 	 * valid bit is clear, we have to zero out a portion of
4151 	 * the last block.
4152 	 */
4153 	endoff = base + size;
4154 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4155 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4156 		pmap_zero_page_area(m, endoff,
4157 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4158 
4159 	/*
4160 	 * Set valid, clear dirty bits.  If validating the entire
4161 	 * page we can safely clear the pmap modify bit.  We also
4162 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
4163 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4164 	 * be set again.
4165 	 *
4166 	 * We set valid bits inclusive of any overlap, but we can only
4167 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4168 	 * the range.
4169 	 */
4170 	oldvalid = m->valid;
4171 	pagebits = vm_page_bits(base, size);
4172 	m->valid |= pagebits;
4173 #if 0	/* NOT YET */
4174 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4175 		frag = DEV_BSIZE - frag;
4176 		base += frag;
4177 		size -= frag;
4178 		if (size < 0)
4179 			size = 0;
4180 	}
4181 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4182 #endif
4183 	if (base == 0 && size == PAGE_SIZE) {
4184 		/*
4185 		 * The page can only be modified within the pmap if it is
4186 		 * mapped, and it can only be mapped if it was previously
4187 		 * fully valid.
4188 		 */
4189 		if (oldvalid == VM_PAGE_BITS_ALL)
4190 			/*
4191 			 * Perform the pmap_clear_modify() first.  Otherwise,
4192 			 * a concurrent pmap operation, such as
4193 			 * pmap_protect(), could clear a modification in the
4194 			 * pmap and set the dirty field on the page before
4195 			 * pmap_clear_modify() had begun and after the dirty
4196 			 * field was cleared here.
4197 			 */
4198 			pmap_clear_modify(m);
4199 		m->dirty = 0;
4200 		m->oflags &= ~VPO_NOSYNC;
4201 	} else if (oldvalid != VM_PAGE_BITS_ALL)
4202 		m->dirty &= ~pagebits;
4203 	else
4204 		vm_page_clear_dirty_mask(m, pagebits);
4205 }
4206 
4207 void
4208 vm_page_clear_dirty(vm_page_t m, int base, int size)
4209 {
4210 
4211 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4212 }
4213 
4214 /*
4215  *	vm_page_set_invalid:
4216  *
4217  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4218  *	valid and dirty bits for the effected areas are cleared.
4219  */
4220 void
4221 vm_page_set_invalid(vm_page_t m, int base, int size)
4222 {
4223 	vm_page_bits_t bits;
4224 	vm_object_t object;
4225 
4226 	object = m->object;
4227 	VM_OBJECT_ASSERT_WLOCKED(object);
4228 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4229 	    size >= object->un_pager.vnp.vnp_size)
4230 		bits = VM_PAGE_BITS_ALL;
4231 	else
4232 		bits = vm_page_bits(base, size);
4233 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4234 	    bits != 0)
4235 		pmap_remove_all(m);
4236 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4237 	    !pmap_page_is_mapped(m),
4238 	    ("vm_page_set_invalid: page %p is mapped", m));
4239 	m->valid &= ~bits;
4240 	m->dirty &= ~bits;
4241 }
4242 
4243 /*
4244  * vm_page_zero_invalid()
4245  *
4246  *	The kernel assumes that the invalid portions of a page contain
4247  *	garbage, but such pages can be mapped into memory by user code.
4248  *	When this occurs, we must zero out the non-valid portions of the
4249  *	page so user code sees what it expects.
4250  *
4251  *	Pages are most often semi-valid when the end of a file is mapped
4252  *	into memory and the file's size is not page aligned.
4253  */
4254 void
4255 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4256 {
4257 	int b;
4258 	int i;
4259 
4260 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4261 	/*
4262 	 * Scan the valid bits looking for invalid sections that
4263 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4264 	 * valid bit may be set ) have already been zeroed by
4265 	 * vm_page_set_validclean().
4266 	 */
4267 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4268 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4269 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4270 			if (i > b) {
4271 				pmap_zero_page_area(m,
4272 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4273 			}
4274 			b = i + 1;
4275 		}
4276 	}
4277 
4278 	/*
4279 	 * setvalid is TRUE when we can safely set the zero'd areas
4280 	 * as being valid.  We can do this if there are no cache consistancy
4281 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4282 	 */
4283 	if (setvalid)
4284 		m->valid = VM_PAGE_BITS_ALL;
4285 }
4286 
4287 /*
4288  *	vm_page_is_valid:
4289  *
4290  *	Is (partial) page valid?  Note that the case where size == 0
4291  *	will return FALSE in the degenerate case where the page is
4292  *	entirely invalid, and TRUE otherwise.
4293  */
4294 int
4295 vm_page_is_valid(vm_page_t m, int base, int size)
4296 {
4297 	vm_page_bits_t bits;
4298 
4299 	VM_OBJECT_ASSERT_LOCKED(m->object);
4300 	bits = vm_page_bits(base, size);
4301 	return (m->valid != 0 && (m->valid & bits) == bits);
4302 }
4303 
4304 /*
4305  * Returns true if all of the specified predicates are true for the entire
4306  * (super)page and false otherwise.
4307  */
4308 bool
4309 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4310 {
4311 	vm_object_t object;
4312 	int i, npages;
4313 
4314 	object = m->object;
4315 	if (skip_m != NULL && skip_m->object != object)
4316 		return (false);
4317 	VM_OBJECT_ASSERT_LOCKED(object);
4318 	npages = atop(pagesizes[m->psind]);
4319 
4320 	/*
4321 	 * The physically contiguous pages that make up a superpage, i.e., a
4322 	 * page with a page size index ("psind") greater than zero, will
4323 	 * occupy adjacent entries in vm_page_array[].
4324 	 */
4325 	for (i = 0; i < npages; i++) {
4326 		/* Always test object consistency, including "skip_m". */
4327 		if (m[i].object != object)
4328 			return (false);
4329 		if (&m[i] == skip_m)
4330 			continue;
4331 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4332 			return (false);
4333 		if ((flags & PS_ALL_DIRTY) != 0) {
4334 			/*
4335 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4336 			 * might stop this case from spuriously returning
4337 			 * "false".  However, that would require a write lock
4338 			 * on the object containing "m[i]".
4339 			 */
4340 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4341 				return (false);
4342 		}
4343 		if ((flags & PS_ALL_VALID) != 0 &&
4344 		    m[i].valid != VM_PAGE_BITS_ALL)
4345 			return (false);
4346 	}
4347 	return (true);
4348 }
4349 
4350 /*
4351  * Set the page's dirty bits if the page is modified.
4352  */
4353 void
4354 vm_page_test_dirty(vm_page_t m)
4355 {
4356 
4357 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4358 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4359 		vm_page_dirty(m);
4360 }
4361 
4362 void
4363 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4364 {
4365 
4366 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4367 }
4368 
4369 void
4370 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4371 {
4372 
4373 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4374 }
4375 
4376 int
4377 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4378 {
4379 
4380 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4381 }
4382 
4383 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4384 void
4385 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4386 {
4387 
4388 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4389 }
4390 
4391 void
4392 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4393 {
4394 
4395 	mtx_assert_(vm_page_lockptr(m), a, file, line);
4396 }
4397 #endif
4398 
4399 #ifdef INVARIANTS
4400 void
4401 vm_page_object_lock_assert(vm_page_t m)
4402 {
4403 
4404 	/*
4405 	 * Certain of the page's fields may only be modified by the
4406 	 * holder of the containing object's lock or the exclusive busy.
4407 	 * holder.  Unfortunately, the holder of the write busy is
4408 	 * not recorded, and thus cannot be checked here.
4409 	 */
4410 	if (m->object != NULL && !vm_page_xbusied(m))
4411 		VM_OBJECT_ASSERT_WLOCKED(m->object);
4412 }
4413 
4414 void
4415 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4416 {
4417 
4418 	if ((bits & PGA_WRITEABLE) == 0)
4419 		return;
4420 
4421 	/*
4422 	 * The PGA_WRITEABLE flag can only be set if the page is
4423 	 * managed, is exclusively busied or the object is locked.
4424 	 * Currently, this flag is only set by pmap_enter().
4425 	 */
4426 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4427 	    ("PGA_WRITEABLE on unmanaged page"));
4428 	if (!vm_page_xbusied(m))
4429 		VM_OBJECT_ASSERT_LOCKED(m->object);
4430 }
4431 #endif
4432 
4433 #include "opt_ddb.h"
4434 #ifdef DDB
4435 #include <sys/kernel.h>
4436 
4437 #include <ddb/ddb.h>
4438 
4439 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4440 {
4441 
4442 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4443 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4444 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4445 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4446 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4447 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4448 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4449 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4450 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4451 }
4452 
4453 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4454 {
4455 	int dom;
4456 
4457 	db_printf("pq_free %d\n", vm_free_count());
4458 	for (dom = 0; dom < vm_ndomains; dom++) {
4459 		db_printf(
4460     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4461 		    dom,
4462 		    vm_dom[dom].vmd_page_count,
4463 		    vm_dom[dom].vmd_free_count,
4464 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4465 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4466 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4467 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4468 	}
4469 }
4470 
4471 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4472 {
4473 	vm_page_t m;
4474 	boolean_t phys;
4475 
4476 	if (!have_addr) {
4477 		db_printf("show pginfo addr\n");
4478 		return;
4479 	}
4480 
4481 	phys = strchr(modif, 'p') != NULL;
4482 	if (phys)
4483 		m = PHYS_TO_VM_PAGE(addr);
4484 	else
4485 		m = (vm_page_t)addr;
4486 	db_printf(
4487     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4488     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4489 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4490 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4491 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4492 }
4493 #endif /* DDB */
4494