xref: /freebsd/sys/vm/vm_page.c (revision e7c5d9d3f2a6b76b7528aeaf2e9e284345ba62b0)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *	Resident memory management module.
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_vm.h"
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/lock.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/malloc.h>
82 #include <sys/mman.h>
83 #include <sys/msgbuf.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/rwlock.h>
87 #include <sys/sbuf.h>
88 #include <sys/sched.h>
89 #include <sys/smp.h>
90 #include <sys/sysctl.h>
91 #include <sys/vmmeter.h>
92 #include <sys/vnode.h>
93 
94 #include <vm/vm.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_param.h>
97 #include <vm/vm_domainset.h>
98 #include <vm/vm_kern.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/vm_pager.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_reserv.h>
108 #include <vm/vm_extern.h>
109 #include <vm/uma.h>
110 #include <vm/uma_int.h>
111 
112 #include <machine/md_var.h>
113 
114 extern int	uma_startup_count(int);
115 extern void	uma_startup(void *, int);
116 extern int	vmem_startup_count(void);
117 
118 struct vm_domain vm_dom[MAXMEMDOM];
119 
120 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
121 
122 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
123 
124 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
125 /* The following fields are protected by the domainset lock. */
126 domainset_t __exclusive_cache_line vm_min_domains;
127 domainset_t __exclusive_cache_line vm_severe_domains;
128 static int vm_min_waiters;
129 static int vm_severe_waiters;
130 static int vm_pageproc_waiters;
131 
132 /*
133  * bogus page -- for I/O to/from partially complete buffers,
134  * or for paging into sparsely invalid regions.
135  */
136 vm_page_t bogus_page;
137 
138 vm_page_t vm_page_array;
139 long vm_page_array_size;
140 long first_page;
141 
142 static int boot_pages;
143 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
144     &boot_pages, 0,
145     "number of pages allocated for bootstrapping the VM system");
146 
147 static int pa_tryrelock_restart;
148 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
149     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
150 
151 static TAILQ_HEAD(, vm_page) blacklist_head;
152 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
153 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
154     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
155 
156 static uma_zone_t fakepg_zone;
157 
158 static void vm_page_alloc_check(vm_page_t m);
159 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
160 static void vm_page_dequeue_complete(vm_page_t m);
161 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
162 static void vm_page_init(void *dummy);
163 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
164     vm_pindex_t pindex, vm_page_t mpred);
165 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
166     vm_page_t mpred);
167 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
168     vm_page_t m_run, vm_paddr_t high);
169 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
170     int req);
171 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
172     int flags);
173 static void vm_page_zone_release(void *arg, void **store, int cnt);
174 
175 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
176 
177 static void
178 vm_page_init(void *dummy)
179 {
180 
181 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
182 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
183 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
184 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
185 }
186 
187 /*
188  * The cache page zone is initialized later since we need to be able to allocate
189  * pages before UMA is fully initialized.
190  */
191 static void
192 vm_page_init_cache_zones(void *dummy __unused)
193 {
194 	struct vm_domain *vmd;
195 	struct vm_pgcache *pgcache;
196 	int domain, pool;
197 
198 	for (domain = 0; domain < vm_ndomains; domain++) {
199 		vmd = VM_DOMAIN(domain);
200 
201 		/*
202 		 * Don't allow the page caches to take up more than .25% of
203 		 * memory.
204 		 */
205 		if (vmd->vmd_page_count / 400 < 256 * mp_ncpus * VM_NFREEPOOL)
206 			continue;
207 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
208 			pgcache = &vmd->vmd_pgcache[pool];
209 			pgcache->domain = domain;
210 			pgcache->pool = pool;
211 			pgcache->zone = uma_zcache_create("vm pgcache",
212 			    sizeof(struct vm_page), NULL, NULL, NULL, NULL,
213 			    vm_page_zone_import, vm_page_zone_release, pgcache,
214 			    UMA_ZONE_MAXBUCKET | UMA_ZONE_VM);
215 			(void)uma_zone_set_maxcache(pgcache->zone, 0);
216 		}
217 	}
218 }
219 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
220 
221 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
222 #if PAGE_SIZE == 32768
223 #ifdef CTASSERT
224 CTASSERT(sizeof(u_long) >= 8);
225 #endif
226 #endif
227 
228 /*
229  * Try to acquire a physical address lock while a pmap is locked.  If we
230  * fail to trylock we unlock and lock the pmap directly and cache the
231  * locked pa in *locked.  The caller should then restart their loop in case
232  * the virtual to physical mapping has changed.
233  */
234 int
235 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
236 {
237 	vm_paddr_t lockpa;
238 
239 	lockpa = *locked;
240 	*locked = pa;
241 	if (lockpa) {
242 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
243 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
244 			return (0);
245 		PA_UNLOCK(lockpa);
246 	}
247 	if (PA_TRYLOCK(pa))
248 		return (0);
249 	PMAP_UNLOCK(pmap);
250 	atomic_add_int(&pa_tryrelock_restart, 1);
251 	PA_LOCK(pa);
252 	PMAP_LOCK(pmap);
253 	return (EAGAIN);
254 }
255 
256 /*
257  *	vm_set_page_size:
258  *
259  *	Sets the page size, perhaps based upon the memory
260  *	size.  Must be called before any use of page-size
261  *	dependent functions.
262  */
263 void
264 vm_set_page_size(void)
265 {
266 	if (vm_cnt.v_page_size == 0)
267 		vm_cnt.v_page_size = PAGE_SIZE;
268 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
269 		panic("vm_set_page_size: page size not a power of two");
270 }
271 
272 /*
273  *	vm_page_blacklist_next:
274  *
275  *	Find the next entry in the provided string of blacklist
276  *	addresses.  Entries are separated by space, comma, or newline.
277  *	If an invalid integer is encountered then the rest of the
278  *	string is skipped.  Updates the list pointer to the next
279  *	character, or NULL if the string is exhausted or invalid.
280  */
281 static vm_paddr_t
282 vm_page_blacklist_next(char **list, char *end)
283 {
284 	vm_paddr_t bad;
285 	char *cp, *pos;
286 
287 	if (list == NULL || *list == NULL)
288 		return (0);
289 	if (**list =='\0') {
290 		*list = NULL;
291 		return (0);
292 	}
293 
294 	/*
295 	 * If there's no end pointer then the buffer is coming from
296 	 * the kenv and we know it's null-terminated.
297 	 */
298 	if (end == NULL)
299 		end = *list + strlen(*list);
300 
301 	/* Ensure that strtoq() won't walk off the end */
302 	if (*end != '\0') {
303 		if (*end == '\n' || *end == ' ' || *end  == ',')
304 			*end = '\0';
305 		else {
306 			printf("Blacklist not terminated, skipping\n");
307 			*list = NULL;
308 			return (0);
309 		}
310 	}
311 
312 	for (pos = *list; *pos != '\0'; pos = cp) {
313 		bad = strtoq(pos, &cp, 0);
314 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
315 			if (bad == 0) {
316 				if (++cp < end)
317 					continue;
318 				else
319 					break;
320 			}
321 		} else
322 			break;
323 		if (*cp == '\0' || ++cp >= end)
324 			*list = NULL;
325 		else
326 			*list = cp;
327 		return (trunc_page(bad));
328 	}
329 	printf("Garbage in RAM blacklist, skipping\n");
330 	*list = NULL;
331 	return (0);
332 }
333 
334 bool
335 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
336 {
337 	struct vm_domain *vmd;
338 	vm_page_t m;
339 	int ret;
340 
341 	m = vm_phys_paddr_to_vm_page(pa);
342 	if (m == NULL)
343 		return (true); /* page does not exist, no failure */
344 
345 	vmd = vm_pagequeue_domain(m);
346 	vm_domain_free_lock(vmd);
347 	ret = vm_phys_unfree_page(m);
348 	vm_domain_free_unlock(vmd);
349 	if (ret != 0) {
350 		vm_domain_freecnt_inc(vmd, -1);
351 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
352 		if (verbose)
353 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
354 	}
355 	return (ret);
356 }
357 
358 /*
359  *	vm_page_blacklist_check:
360  *
361  *	Iterate through the provided string of blacklist addresses, pulling
362  *	each entry out of the physical allocator free list and putting it
363  *	onto a list for reporting via the vm.page_blacklist sysctl.
364  */
365 static void
366 vm_page_blacklist_check(char *list, char *end)
367 {
368 	vm_paddr_t pa;
369 	char *next;
370 
371 	next = list;
372 	while (next != NULL) {
373 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
374 			continue;
375 		vm_page_blacklist_add(pa, bootverbose);
376 	}
377 }
378 
379 /*
380  *	vm_page_blacklist_load:
381  *
382  *	Search for a special module named "ram_blacklist".  It'll be a
383  *	plain text file provided by the user via the loader directive
384  *	of the same name.
385  */
386 static void
387 vm_page_blacklist_load(char **list, char **end)
388 {
389 	void *mod;
390 	u_char *ptr;
391 	u_int len;
392 
393 	mod = NULL;
394 	ptr = NULL;
395 
396 	mod = preload_search_by_type("ram_blacklist");
397 	if (mod != NULL) {
398 		ptr = preload_fetch_addr(mod);
399 		len = preload_fetch_size(mod);
400         }
401 	*list = ptr;
402 	if (ptr != NULL)
403 		*end = ptr + len;
404 	else
405 		*end = NULL;
406 	return;
407 }
408 
409 static int
410 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
411 {
412 	vm_page_t m;
413 	struct sbuf sbuf;
414 	int error, first;
415 
416 	first = 1;
417 	error = sysctl_wire_old_buffer(req, 0);
418 	if (error != 0)
419 		return (error);
420 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
421 	TAILQ_FOREACH(m, &blacklist_head, listq) {
422 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
423 		    (uintmax_t)m->phys_addr);
424 		first = 0;
425 	}
426 	error = sbuf_finish(&sbuf);
427 	sbuf_delete(&sbuf);
428 	return (error);
429 }
430 
431 /*
432  * Initialize a dummy page for use in scans of the specified paging queue.
433  * In principle, this function only needs to set the flag PG_MARKER.
434  * Nonetheless, it write busies the page as a safety precaution.
435  */
436 static void
437 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
438 {
439 
440 	bzero(marker, sizeof(*marker));
441 	marker->flags = PG_MARKER;
442 	marker->aflags = aflags;
443 	marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
444 	marker->queue = queue;
445 }
446 
447 static void
448 vm_page_domain_init(int domain)
449 {
450 	struct vm_domain *vmd;
451 	struct vm_pagequeue *pq;
452 	int i;
453 
454 	vmd = VM_DOMAIN(domain);
455 	bzero(vmd, sizeof(*vmd));
456 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
457 	    "vm inactive pagequeue";
458 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
459 	    "vm active pagequeue";
460 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
461 	    "vm laundry pagequeue";
462 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
463 	    "vm unswappable pagequeue";
464 	vmd->vmd_domain = domain;
465 	vmd->vmd_page_count = 0;
466 	vmd->vmd_free_count = 0;
467 	vmd->vmd_segs = 0;
468 	vmd->vmd_oom = FALSE;
469 	for (i = 0; i < PQ_COUNT; i++) {
470 		pq = &vmd->vmd_pagequeues[i];
471 		TAILQ_INIT(&pq->pq_pl);
472 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
473 		    MTX_DEF | MTX_DUPOK);
474 		pq->pq_pdpages = 0;
475 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
476 	}
477 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
478 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
479 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
480 
481 	/*
482 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
483 	 * insertions.
484 	 */
485 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
486 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
487 	    &vmd->vmd_inacthead, plinks.q);
488 
489 	/*
490 	 * The clock pages are used to implement active queue scanning without
491 	 * requeues.  Scans start at clock[0], which is advanced after the scan
492 	 * ends.  When the two clock hands meet, they are reset and scanning
493 	 * resumes from the head of the queue.
494 	 */
495 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
496 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
497 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
498 	    &vmd->vmd_clock[0], plinks.q);
499 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
500 	    &vmd->vmd_clock[1], plinks.q);
501 }
502 
503 /*
504  * Initialize a physical page in preparation for adding it to the free
505  * lists.
506  */
507 static void
508 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
509 {
510 
511 	m->object = NULL;
512 	m->wire_count = 0;
513 	m->busy_lock = VPB_UNBUSIED;
514 	m->flags = m->aflags = 0;
515 	m->phys_addr = pa;
516 	m->queue = PQ_NONE;
517 	m->psind = 0;
518 	m->segind = segind;
519 	m->order = VM_NFREEORDER;
520 	m->pool = VM_FREEPOOL_DEFAULT;
521 	m->valid = m->dirty = 0;
522 	pmap_page_init(m);
523 }
524 
525 /*
526  *	vm_page_startup:
527  *
528  *	Initializes the resident memory module.  Allocates physical memory for
529  *	bootstrapping UMA and some data structures that are used to manage
530  *	physical pages.  Initializes these structures, and populates the free
531  *	page queues.
532  */
533 vm_offset_t
534 vm_page_startup(vm_offset_t vaddr)
535 {
536 	struct vm_phys_seg *seg;
537 	vm_page_t m;
538 	char *list, *listend;
539 	vm_offset_t mapped;
540 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
541 	vm_paddr_t biggestsize, last_pa, pa;
542 	u_long pagecount;
543 	int biggestone, i, segind;
544 #ifdef WITNESS
545 	int witness_size;
546 #endif
547 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
548 	long ii;
549 #endif
550 
551 	biggestsize = 0;
552 	biggestone = 0;
553 	vaddr = round_page(vaddr);
554 
555 	for (i = 0; phys_avail[i + 1]; i += 2) {
556 		phys_avail[i] = round_page(phys_avail[i]);
557 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
558 	}
559 	for (i = 0; phys_avail[i + 1]; i += 2) {
560 		size = phys_avail[i + 1] - phys_avail[i];
561 		if (size > biggestsize) {
562 			biggestone = i;
563 			biggestsize = size;
564 		}
565 	}
566 
567 	end = phys_avail[biggestone+1];
568 
569 	/*
570 	 * Initialize the page and queue locks.
571 	 */
572 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
573 	for (i = 0; i < PA_LOCK_COUNT; i++)
574 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
575 	for (i = 0; i < vm_ndomains; i++)
576 		vm_page_domain_init(i);
577 
578 	/*
579 	 * Allocate memory for use when boot strapping the kernel memory
580 	 * allocator.  Tell UMA how many zones we are going to create
581 	 * before going fully functional.  UMA will add its zones.
582 	 *
583 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
584 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
585 	 */
586 	boot_pages = uma_startup_count(8);
587 
588 #ifndef UMA_MD_SMALL_ALLOC
589 	/* vmem_startup() calls uma_prealloc(). */
590 	boot_pages += vmem_startup_count();
591 	/* vm_map_startup() calls uma_prealloc(). */
592 	boot_pages += howmany(MAX_KMAP,
593 	    UMA_SLAB_SPACE / sizeof(struct vm_map));
594 
595 	/*
596 	 * Before going fully functional kmem_init() does allocation
597 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
598 	 */
599 	boot_pages += 2;
600 #endif
601 	/*
602 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
603 	 * manually fetch the value.
604 	 */
605 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
606 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
607 	new_end = trunc_page(new_end);
608 	mapped = pmap_map(&vaddr, new_end, end,
609 	    VM_PROT_READ | VM_PROT_WRITE);
610 	bzero((void *)mapped, end - new_end);
611 	uma_startup((void *)mapped, boot_pages);
612 
613 #ifdef WITNESS
614 	witness_size = round_page(witness_startup_count());
615 	new_end -= witness_size;
616 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
617 	    VM_PROT_READ | VM_PROT_WRITE);
618 	bzero((void *)mapped, witness_size);
619 	witness_startup((void *)mapped);
620 #endif
621 
622 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
623     defined(__i386__) || defined(__mips__) || defined(__riscv)
624 	/*
625 	 * Allocate a bitmap to indicate that a random physical page
626 	 * needs to be included in a minidump.
627 	 *
628 	 * The amd64 port needs this to indicate which direct map pages
629 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
630 	 *
631 	 * However, i386 still needs this workspace internally within the
632 	 * minidump code.  In theory, they are not needed on i386, but are
633 	 * included should the sf_buf code decide to use them.
634 	 */
635 	last_pa = 0;
636 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
637 		if (dump_avail[i + 1] > last_pa)
638 			last_pa = dump_avail[i + 1];
639 	page_range = last_pa / PAGE_SIZE;
640 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
641 	new_end -= vm_page_dump_size;
642 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
643 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
644 	bzero((void *)vm_page_dump, vm_page_dump_size);
645 #else
646 	(void)last_pa;
647 #endif
648 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
649     defined(__riscv)
650 	/*
651 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
652 	 * in a crash dump.  When pmap_map() uses the direct map, they are
653 	 * not automatically included.
654 	 */
655 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
656 		dump_add_page(pa);
657 #endif
658 	phys_avail[biggestone + 1] = new_end;
659 #ifdef __amd64__
660 	/*
661 	 * Request that the physical pages underlying the message buffer be
662 	 * included in a crash dump.  Since the message buffer is accessed
663 	 * through the direct map, they are not automatically included.
664 	 */
665 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
666 	last_pa = pa + round_page(msgbufsize);
667 	while (pa < last_pa) {
668 		dump_add_page(pa);
669 		pa += PAGE_SIZE;
670 	}
671 #endif
672 	/*
673 	 * Compute the number of pages of memory that will be available for
674 	 * use, taking into account the overhead of a page structure per page.
675 	 * In other words, solve
676 	 *	"available physical memory" - round_page(page_range *
677 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
678 	 * for page_range.
679 	 */
680 	low_avail = phys_avail[0];
681 	high_avail = phys_avail[1];
682 	for (i = 0; i < vm_phys_nsegs; i++) {
683 		if (vm_phys_segs[i].start < low_avail)
684 			low_avail = vm_phys_segs[i].start;
685 		if (vm_phys_segs[i].end > high_avail)
686 			high_avail = vm_phys_segs[i].end;
687 	}
688 	/* Skip the first chunk.  It is already accounted for. */
689 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
690 		if (phys_avail[i] < low_avail)
691 			low_avail = phys_avail[i];
692 		if (phys_avail[i + 1] > high_avail)
693 			high_avail = phys_avail[i + 1];
694 	}
695 	first_page = low_avail / PAGE_SIZE;
696 #ifdef VM_PHYSSEG_SPARSE
697 	size = 0;
698 	for (i = 0; i < vm_phys_nsegs; i++)
699 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
700 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
701 		size += phys_avail[i + 1] - phys_avail[i];
702 #elif defined(VM_PHYSSEG_DENSE)
703 	size = high_avail - low_avail;
704 #else
705 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
706 #endif
707 
708 #ifdef VM_PHYSSEG_DENSE
709 	/*
710 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
711 	 * the overhead of a page structure per page only if vm_page_array is
712 	 * allocated from the last physical memory chunk.  Otherwise, we must
713 	 * allocate page structures representing the physical memory
714 	 * underlying vm_page_array, even though they will not be used.
715 	 */
716 	if (new_end != high_avail)
717 		page_range = size / PAGE_SIZE;
718 	else
719 #endif
720 	{
721 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
722 
723 		/*
724 		 * If the partial bytes remaining are large enough for
725 		 * a page (PAGE_SIZE) without a corresponding
726 		 * 'struct vm_page', then new_end will contain an
727 		 * extra page after subtracting the length of the VM
728 		 * page array.  Compensate by subtracting an extra
729 		 * page from new_end.
730 		 */
731 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
732 			if (new_end == high_avail)
733 				high_avail -= PAGE_SIZE;
734 			new_end -= PAGE_SIZE;
735 		}
736 	}
737 	end = new_end;
738 
739 	/*
740 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
741 	 * However, because this page is allocated from KVM, out-of-bounds
742 	 * accesses using the direct map will not be trapped.
743 	 */
744 	vaddr += PAGE_SIZE;
745 
746 	/*
747 	 * Allocate physical memory for the page structures, and map it.
748 	 */
749 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
750 	mapped = pmap_map(&vaddr, new_end, end,
751 	    VM_PROT_READ | VM_PROT_WRITE);
752 	vm_page_array = (vm_page_t)mapped;
753 	vm_page_array_size = page_range;
754 
755 #if VM_NRESERVLEVEL > 0
756 	/*
757 	 * Allocate physical memory for the reservation management system's
758 	 * data structures, and map it.
759 	 */
760 	if (high_avail == end)
761 		high_avail = new_end;
762 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
763 #endif
764 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
765     defined(__riscv)
766 	/*
767 	 * Include vm_page_array and vm_reserv_array in a crash dump.
768 	 */
769 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
770 		dump_add_page(pa);
771 #endif
772 	phys_avail[biggestone + 1] = new_end;
773 
774 	/*
775 	 * Add physical memory segments corresponding to the available
776 	 * physical pages.
777 	 */
778 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
779 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
780 
781 	/*
782 	 * Initialize the physical memory allocator.
783 	 */
784 	vm_phys_init();
785 
786 	/*
787 	 * Initialize the page structures and add every available page to the
788 	 * physical memory allocator's free lists.
789 	 */
790 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
791 	for (ii = 0; ii < vm_page_array_size; ii++) {
792 		m = &vm_page_array[ii];
793 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
794 		m->flags = PG_FICTITIOUS;
795 	}
796 #endif
797 	vm_cnt.v_page_count = 0;
798 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
799 		seg = &vm_phys_segs[segind];
800 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
801 		    m++, pa += PAGE_SIZE)
802 			vm_page_init_page(m, pa, segind);
803 
804 		/*
805 		 * Add the segment to the free lists only if it is covered by
806 		 * one of the ranges in phys_avail.  Because we've added the
807 		 * ranges to the vm_phys_segs array, we can assume that each
808 		 * segment is either entirely contained in one of the ranges,
809 		 * or doesn't overlap any of them.
810 		 */
811 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
812 			struct vm_domain *vmd;
813 
814 			if (seg->start < phys_avail[i] ||
815 			    seg->end > phys_avail[i + 1])
816 				continue;
817 
818 			m = seg->first_page;
819 			pagecount = (u_long)atop(seg->end - seg->start);
820 
821 			vmd = VM_DOMAIN(seg->domain);
822 			vm_domain_free_lock(vmd);
823 			vm_phys_enqueue_contig(m, pagecount);
824 			vm_domain_free_unlock(vmd);
825 			vm_domain_freecnt_inc(vmd, pagecount);
826 			vm_cnt.v_page_count += (u_int)pagecount;
827 
828 			vmd = VM_DOMAIN(seg->domain);
829 			vmd->vmd_page_count += (u_int)pagecount;
830 			vmd->vmd_segs |= 1UL << m->segind;
831 			break;
832 		}
833 	}
834 
835 	/*
836 	 * Remove blacklisted pages from the physical memory allocator.
837 	 */
838 	TAILQ_INIT(&blacklist_head);
839 	vm_page_blacklist_load(&list, &listend);
840 	vm_page_blacklist_check(list, listend);
841 
842 	list = kern_getenv("vm.blacklist");
843 	vm_page_blacklist_check(list, NULL);
844 
845 	freeenv(list);
846 #if VM_NRESERVLEVEL > 0
847 	/*
848 	 * Initialize the reservation management system.
849 	 */
850 	vm_reserv_init();
851 #endif
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  *	vm_page_unhold_pages:
1097  *
1098  *	Unhold each of the pages that is referenced by the given array.
1099  */
1100 void
1101 vm_page_unhold_pages(vm_page_t *ma, int count)
1102 {
1103 	struct mtx *mtx;
1104 
1105 	mtx = NULL;
1106 	for (; count != 0; count--) {
1107 		vm_page_change_lock(*ma, &mtx);
1108 		if (vm_page_unwire(*ma, PQ_ACTIVE) && (*ma)->object == NULL)
1109 			vm_page_free(*ma);
1110 		ma++;
1111 	}
1112 	if (mtx != NULL)
1113 		mtx_unlock(mtx);
1114 }
1115 
1116 vm_page_t
1117 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1118 {
1119 	vm_page_t m;
1120 
1121 #ifdef VM_PHYSSEG_SPARSE
1122 	m = vm_phys_paddr_to_vm_page(pa);
1123 	if (m == NULL)
1124 		m = vm_phys_fictitious_to_vm_page(pa);
1125 	return (m);
1126 #elif defined(VM_PHYSSEG_DENSE)
1127 	long pi;
1128 
1129 	pi = atop(pa);
1130 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1131 		m = &vm_page_array[pi - first_page];
1132 		return (m);
1133 	}
1134 	return (vm_phys_fictitious_to_vm_page(pa));
1135 #else
1136 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1137 #endif
1138 }
1139 
1140 /*
1141  *	vm_page_getfake:
1142  *
1143  *	Create a fictitious page with the specified physical address and
1144  *	memory attribute.  The memory attribute is the only the machine-
1145  *	dependent aspect of a fictitious page that must be initialized.
1146  */
1147 vm_page_t
1148 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1149 {
1150 	vm_page_t m;
1151 
1152 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1153 	vm_page_initfake(m, paddr, memattr);
1154 	return (m);
1155 }
1156 
1157 void
1158 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1159 {
1160 
1161 	if ((m->flags & PG_FICTITIOUS) != 0) {
1162 		/*
1163 		 * The page's memattr might have changed since the
1164 		 * previous initialization.  Update the pmap to the
1165 		 * new memattr.
1166 		 */
1167 		goto memattr;
1168 	}
1169 	m->phys_addr = paddr;
1170 	m->queue = PQ_NONE;
1171 	/* Fictitious pages don't use "segind". */
1172 	m->flags = PG_FICTITIOUS;
1173 	/* Fictitious pages don't use "order" or "pool". */
1174 	m->oflags = VPO_UNMANAGED;
1175 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1176 	m->wire_count = 1;
1177 	pmap_page_init(m);
1178 memattr:
1179 	pmap_page_set_memattr(m, memattr);
1180 }
1181 
1182 /*
1183  *	vm_page_putfake:
1184  *
1185  *	Release a fictitious page.
1186  */
1187 void
1188 vm_page_putfake(vm_page_t m)
1189 {
1190 
1191 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1192 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1193 	    ("vm_page_putfake: bad page %p", m));
1194 	uma_zfree(fakepg_zone, m);
1195 }
1196 
1197 /*
1198  *	vm_page_updatefake:
1199  *
1200  *	Update the given fictitious page to the specified physical address and
1201  *	memory attribute.
1202  */
1203 void
1204 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1205 {
1206 
1207 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1208 	    ("vm_page_updatefake: bad page %p", m));
1209 	m->phys_addr = paddr;
1210 	pmap_page_set_memattr(m, memattr);
1211 }
1212 
1213 /*
1214  *	vm_page_free:
1215  *
1216  *	Free a page.
1217  */
1218 void
1219 vm_page_free(vm_page_t m)
1220 {
1221 
1222 	m->flags &= ~PG_ZERO;
1223 	vm_page_free_toq(m);
1224 }
1225 
1226 /*
1227  *	vm_page_free_zero:
1228  *
1229  *	Free a page to the zerod-pages queue
1230  */
1231 void
1232 vm_page_free_zero(vm_page_t m)
1233 {
1234 
1235 	m->flags |= PG_ZERO;
1236 	vm_page_free_toq(m);
1237 }
1238 
1239 /*
1240  * Unbusy and handle the page queueing for a page from a getpages request that
1241  * was optionally read ahead or behind.
1242  */
1243 void
1244 vm_page_readahead_finish(vm_page_t m)
1245 {
1246 
1247 	/* We shouldn't put invalid pages on queues. */
1248 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1249 
1250 	/*
1251 	 * Since the page is not the actually needed one, whether it should
1252 	 * be activated or deactivated is not obvious.  Empirical results
1253 	 * have shown that deactivating the page is usually the best choice,
1254 	 * unless the page is wanted by another thread.
1255 	 */
1256 	vm_page_lock(m);
1257 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1258 		vm_page_activate(m);
1259 	else
1260 		vm_page_deactivate(m);
1261 	vm_page_unlock(m);
1262 	vm_page_xunbusy(m);
1263 }
1264 
1265 /*
1266  *	vm_page_sleep_if_busy:
1267  *
1268  *	Sleep and release the page queues lock if the page is busied.
1269  *	Returns TRUE if the thread slept.
1270  *
1271  *	The given page must be unlocked and object containing it must
1272  *	be locked.
1273  */
1274 int
1275 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1276 {
1277 	vm_object_t obj;
1278 
1279 	vm_page_lock_assert(m, MA_NOTOWNED);
1280 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1281 
1282 	if (vm_page_busied(m)) {
1283 		/*
1284 		 * The page-specific object must be cached because page
1285 		 * identity can change during the sleep, causing the
1286 		 * re-lock of a different object.
1287 		 * It is assumed that a reference to the object is already
1288 		 * held by the callers.
1289 		 */
1290 		obj = m->object;
1291 		vm_page_lock(m);
1292 		VM_OBJECT_WUNLOCK(obj);
1293 		vm_page_busy_sleep(m, msg, false);
1294 		VM_OBJECT_WLOCK(obj);
1295 		return (TRUE);
1296 	}
1297 	return (FALSE);
1298 }
1299 
1300 /*
1301  *	vm_page_dirty_KBI:		[ internal use only ]
1302  *
1303  *	Set all bits in the page's dirty field.
1304  *
1305  *	The object containing the specified page must be locked if the
1306  *	call is made from the machine-independent layer.
1307  *
1308  *	See vm_page_clear_dirty_mask().
1309  *
1310  *	This function should only be called by vm_page_dirty().
1311  */
1312 void
1313 vm_page_dirty_KBI(vm_page_t m)
1314 {
1315 
1316 	/* Refer to this operation by its public name. */
1317 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1318 	    ("vm_page_dirty: page is invalid!"));
1319 	m->dirty = VM_PAGE_BITS_ALL;
1320 }
1321 
1322 /*
1323  *	vm_page_insert:		[ internal use only ]
1324  *
1325  *	Inserts the given mem entry into the object and object list.
1326  *
1327  *	The object must be locked.
1328  */
1329 int
1330 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1331 {
1332 	vm_page_t mpred;
1333 
1334 	VM_OBJECT_ASSERT_WLOCKED(object);
1335 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1336 	return (vm_page_insert_after(m, object, pindex, mpred));
1337 }
1338 
1339 /*
1340  *	vm_page_insert_after:
1341  *
1342  *	Inserts the page "m" into the specified object at offset "pindex".
1343  *
1344  *	The page "mpred" must immediately precede the offset "pindex" within
1345  *	the specified object.
1346  *
1347  *	The object must be locked.
1348  */
1349 static int
1350 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1351     vm_page_t mpred)
1352 {
1353 	vm_page_t msucc;
1354 
1355 	VM_OBJECT_ASSERT_WLOCKED(object);
1356 	KASSERT(m->object == NULL,
1357 	    ("vm_page_insert_after: page already inserted"));
1358 	if (mpred != NULL) {
1359 		KASSERT(mpred->object == object,
1360 		    ("vm_page_insert_after: object doesn't contain mpred"));
1361 		KASSERT(mpred->pindex < pindex,
1362 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1363 		msucc = TAILQ_NEXT(mpred, listq);
1364 	} else
1365 		msucc = TAILQ_FIRST(&object->memq);
1366 	if (msucc != NULL)
1367 		KASSERT(msucc->pindex > pindex,
1368 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1369 
1370 	/*
1371 	 * Record the object/offset pair in this page
1372 	 */
1373 	m->object = object;
1374 	m->pindex = pindex;
1375 
1376 	/*
1377 	 * Now link into the object's ordered list of backed pages.
1378 	 */
1379 	if (vm_radix_insert(&object->rtree, m)) {
1380 		m->object = NULL;
1381 		m->pindex = 0;
1382 		return (1);
1383 	}
1384 	vm_page_insert_radixdone(m, object, mpred);
1385 	return (0);
1386 }
1387 
1388 /*
1389  *	vm_page_insert_radixdone:
1390  *
1391  *	Complete page "m" insertion into the specified object after the
1392  *	radix trie hooking.
1393  *
1394  *	The page "mpred" must precede the offset "m->pindex" within the
1395  *	specified object.
1396  *
1397  *	The object must be locked.
1398  */
1399 static void
1400 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1401 {
1402 
1403 	VM_OBJECT_ASSERT_WLOCKED(object);
1404 	KASSERT(object != NULL && m->object == object,
1405 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1406 	if (mpred != NULL) {
1407 		KASSERT(mpred->object == object,
1408 		    ("vm_page_insert_after: object doesn't contain mpred"));
1409 		KASSERT(mpred->pindex < m->pindex,
1410 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1411 	}
1412 
1413 	if (mpred != NULL)
1414 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1415 	else
1416 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1417 
1418 	/*
1419 	 * Show that the object has one more resident page.
1420 	 */
1421 	object->resident_page_count++;
1422 
1423 	/*
1424 	 * Hold the vnode until the last page is released.
1425 	 */
1426 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1427 		vhold(object->handle);
1428 
1429 	/*
1430 	 * Since we are inserting a new and possibly dirty page,
1431 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1432 	 */
1433 	if (pmap_page_is_write_mapped(m))
1434 		vm_object_set_writeable_dirty(object);
1435 }
1436 
1437 /*
1438  *	vm_page_remove:
1439  *
1440  *	Removes the specified page from its containing object, but does not
1441  *	invalidate any backing storage.  Return true if the page may be safely
1442  *	freed and false otherwise.
1443  *
1444  *	The object must be locked.  The page must be locked if it is managed.
1445  */
1446 bool
1447 vm_page_remove(vm_page_t m)
1448 {
1449 	vm_object_t object;
1450 	vm_page_t mrem;
1451 
1452 	object = m->object;
1453 
1454 	if ((m->oflags & VPO_UNMANAGED) == 0)
1455 		vm_page_assert_locked(m);
1456 	VM_OBJECT_ASSERT_WLOCKED(object);
1457 	if (vm_page_xbusied(m))
1458 		vm_page_xunbusy_maybelocked(m);
1459 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1460 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1461 
1462 	/*
1463 	 * Now remove from the object's list of backed pages.
1464 	 */
1465 	TAILQ_REMOVE(&object->memq, m, listq);
1466 
1467 	/*
1468 	 * And show that the object has one fewer resident page.
1469 	 */
1470 	object->resident_page_count--;
1471 
1472 	/*
1473 	 * The vnode may now be recycled.
1474 	 */
1475 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1476 		vdrop(object->handle);
1477 
1478 	m->object = NULL;
1479 	return (!vm_page_wired(m));
1480 }
1481 
1482 /*
1483  *	vm_page_lookup:
1484  *
1485  *	Returns the page associated with the object/offset
1486  *	pair specified; if none is found, NULL is returned.
1487  *
1488  *	The object must be locked.
1489  */
1490 vm_page_t
1491 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1492 {
1493 
1494 	VM_OBJECT_ASSERT_LOCKED(object);
1495 	return (vm_radix_lookup(&object->rtree, pindex));
1496 }
1497 
1498 /*
1499  *	vm_page_find_least:
1500  *
1501  *	Returns the page associated with the object with least pindex
1502  *	greater than or equal to the parameter pindex, or NULL.
1503  *
1504  *	The object must be locked.
1505  */
1506 vm_page_t
1507 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1508 {
1509 	vm_page_t m;
1510 
1511 	VM_OBJECT_ASSERT_LOCKED(object);
1512 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1513 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1514 	return (m);
1515 }
1516 
1517 /*
1518  * Returns the given page's successor (by pindex) within the object if it is
1519  * resident; if none is found, NULL is returned.
1520  *
1521  * The object must be locked.
1522  */
1523 vm_page_t
1524 vm_page_next(vm_page_t m)
1525 {
1526 	vm_page_t next;
1527 
1528 	VM_OBJECT_ASSERT_LOCKED(m->object);
1529 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1530 		MPASS(next->object == m->object);
1531 		if (next->pindex != m->pindex + 1)
1532 			next = NULL;
1533 	}
1534 	return (next);
1535 }
1536 
1537 /*
1538  * Returns the given page's predecessor (by pindex) within the object if it is
1539  * resident; if none is found, NULL is returned.
1540  *
1541  * The object must be locked.
1542  */
1543 vm_page_t
1544 vm_page_prev(vm_page_t m)
1545 {
1546 	vm_page_t prev;
1547 
1548 	VM_OBJECT_ASSERT_LOCKED(m->object);
1549 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1550 		MPASS(prev->object == m->object);
1551 		if (prev->pindex != m->pindex - 1)
1552 			prev = NULL;
1553 	}
1554 	return (prev);
1555 }
1556 
1557 /*
1558  * Uses the page mnew as a replacement for an existing page at index
1559  * pindex which must be already present in the object.
1560  *
1561  * The existing page must not be on a paging queue.
1562  */
1563 vm_page_t
1564 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1565 {
1566 	vm_page_t mold;
1567 
1568 	VM_OBJECT_ASSERT_WLOCKED(object);
1569 	KASSERT(mnew->object == NULL,
1570 	    ("vm_page_replace: page %p already in object", mnew));
1571 	KASSERT(mnew->queue == PQ_NONE || vm_page_wired(mnew),
1572 	    ("vm_page_replace: new page %p is on a paging queue", mnew));
1573 
1574 	/*
1575 	 * This function mostly follows vm_page_insert() and
1576 	 * vm_page_remove() without the radix, object count and vnode
1577 	 * dance.  Double check such functions for more comments.
1578 	 */
1579 
1580 	mnew->object = object;
1581 	mnew->pindex = pindex;
1582 	mold = vm_radix_replace(&object->rtree, mnew);
1583 	KASSERT(mold->queue == PQ_NONE,
1584 	    ("vm_page_replace: old page %p is on a paging queue", mold));
1585 
1586 	/* Keep the resident page list in sorted order. */
1587 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1588 	TAILQ_REMOVE(&object->memq, mold, listq);
1589 
1590 	mold->object = NULL;
1591 	vm_page_xunbusy_maybelocked(mold);
1592 
1593 	/*
1594 	 * The object's resident_page_count does not change because we have
1595 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1596 	 */
1597 	if (pmap_page_is_write_mapped(mnew))
1598 		vm_object_set_writeable_dirty(object);
1599 	return (mold);
1600 }
1601 
1602 /*
1603  *	vm_page_rename:
1604  *
1605  *	Move the given memory entry from its
1606  *	current object to the specified target object/offset.
1607  *
1608  *	Note: swap associated with the page must be invalidated by the move.  We
1609  *	      have to do this for several reasons:  (1) we aren't freeing the
1610  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1611  *	      moving the page from object A to B, and will then later move
1612  *	      the backing store from A to B and we can't have a conflict.
1613  *
1614  *	Note: we *always* dirty the page.  It is necessary both for the
1615  *	      fact that we moved it, and because we may be invalidating
1616  *	      swap.
1617  *
1618  *	The objects must be locked.
1619  */
1620 int
1621 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1622 {
1623 	vm_page_t mpred;
1624 	vm_pindex_t opidx;
1625 
1626 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1627 
1628 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1629 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1630 	    ("vm_page_rename: pindex already renamed"));
1631 
1632 	/*
1633 	 * Create a custom version of vm_page_insert() which does not depend
1634 	 * by m_prev and can cheat on the implementation aspects of the
1635 	 * function.
1636 	 */
1637 	opidx = m->pindex;
1638 	m->pindex = new_pindex;
1639 	if (vm_radix_insert(&new_object->rtree, m)) {
1640 		m->pindex = opidx;
1641 		return (1);
1642 	}
1643 
1644 	/*
1645 	 * The operation cannot fail anymore.  The removal must happen before
1646 	 * the listq iterator is tainted.
1647 	 */
1648 	m->pindex = opidx;
1649 	vm_page_lock(m);
1650 	(void)vm_page_remove(m);
1651 
1652 	/* Return back to the new pindex to complete vm_page_insert(). */
1653 	m->pindex = new_pindex;
1654 	m->object = new_object;
1655 	vm_page_unlock(m);
1656 	vm_page_insert_radixdone(m, new_object, mpred);
1657 	vm_page_dirty(m);
1658 	return (0);
1659 }
1660 
1661 /*
1662  *	vm_page_alloc:
1663  *
1664  *	Allocate and return a page that is associated with the specified
1665  *	object and offset pair.  By default, this page is exclusive busied.
1666  *
1667  *	The caller must always specify an allocation class.
1668  *
1669  *	allocation classes:
1670  *	VM_ALLOC_NORMAL		normal process request
1671  *	VM_ALLOC_SYSTEM		system *really* needs a page
1672  *	VM_ALLOC_INTERRUPT	interrupt time request
1673  *
1674  *	optional allocation flags:
1675  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1676  *				intends to allocate
1677  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1678  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1679  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1680  *				should not be exclusive busy
1681  *	VM_ALLOC_SBUSY		shared busy the allocated page
1682  *	VM_ALLOC_WIRED		wire the allocated page
1683  *	VM_ALLOC_ZERO		prefer a zeroed page
1684  */
1685 vm_page_t
1686 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1687 {
1688 
1689 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1690 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1691 }
1692 
1693 vm_page_t
1694 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1695     int req)
1696 {
1697 
1698 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1699 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1700 	    NULL));
1701 }
1702 
1703 /*
1704  * Allocate a page in the specified object with the given page index.  To
1705  * optimize insertion of the page into the object, the caller must also specifiy
1706  * the resident page in the object with largest index smaller than the given
1707  * page index, or NULL if no such page exists.
1708  */
1709 vm_page_t
1710 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1711     int req, vm_page_t mpred)
1712 {
1713 	struct vm_domainset_iter di;
1714 	vm_page_t m;
1715 	int domain;
1716 
1717 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1718 	do {
1719 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1720 		    mpred);
1721 		if (m != NULL)
1722 			break;
1723 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1724 
1725 	return (m);
1726 }
1727 
1728 /*
1729  * Returns true if the number of free pages exceeds the minimum
1730  * for the request class and false otherwise.
1731  */
1732 int
1733 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1734 {
1735 	u_int limit, old, new;
1736 
1737 	req = req & VM_ALLOC_CLASS_MASK;
1738 
1739 	/*
1740 	 * The page daemon is allowed to dig deeper into the free page list.
1741 	 */
1742 	if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1743 		req = VM_ALLOC_SYSTEM;
1744 	if (req == VM_ALLOC_INTERRUPT)
1745 		limit = 0;
1746 	else if (req == VM_ALLOC_SYSTEM)
1747 		limit = vmd->vmd_interrupt_free_min;
1748 	else
1749 		limit = vmd->vmd_free_reserved;
1750 
1751 	/*
1752 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1753 	 */
1754 	limit += npages;
1755 	old = vmd->vmd_free_count;
1756 	do {
1757 		if (old < limit)
1758 			return (0);
1759 		new = old - npages;
1760 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1761 
1762 	/* Wake the page daemon if we've crossed the threshold. */
1763 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1764 		pagedaemon_wakeup(vmd->vmd_domain);
1765 
1766 	/* Only update bitsets on transitions. */
1767 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1768 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1769 		vm_domain_set(vmd);
1770 
1771 	return (1);
1772 }
1773 
1774 vm_page_t
1775 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1776     int req, vm_page_t mpred)
1777 {
1778 	struct vm_domain *vmd;
1779 	vm_page_t m;
1780 	int flags, pool;
1781 
1782 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1783 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1784 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1785 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1786 	    ("inconsistent object(%p)/req(%x)", object, req));
1787 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1788 	    ("Can't sleep and retry object insertion."));
1789 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1790 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1791 	    (uintmax_t)pindex));
1792 	if (object != NULL)
1793 		VM_OBJECT_ASSERT_WLOCKED(object);
1794 
1795 	flags = 0;
1796 	m = NULL;
1797 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1798 again:
1799 #if VM_NRESERVLEVEL > 0
1800 	/*
1801 	 * Can we allocate the page from a reservation?
1802 	 */
1803 	if (vm_object_reserv(object) &&
1804 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
1805 	    NULL) {
1806 		domain = vm_phys_domain(m);
1807 		vmd = VM_DOMAIN(domain);
1808 		goto found;
1809 	}
1810 #endif
1811 	vmd = VM_DOMAIN(domain);
1812 	if (vmd->vmd_pgcache[pool].zone != NULL) {
1813 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1814 		if (m != NULL) {
1815 			flags |= PG_PCPU_CACHE;
1816 			goto found;
1817 		}
1818 	}
1819 	if (vm_domain_allocate(vmd, req, 1)) {
1820 		/*
1821 		 * If not, allocate it from the free page queues.
1822 		 */
1823 		vm_domain_free_lock(vmd);
1824 		m = vm_phys_alloc_pages(domain, pool, 0);
1825 		vm_domain_free_unlock(vmd);
1826 		if (m == NULL) {
1827 			vm_domain_freecnt_inc(vmd, 1);
1828 #if VM_NRESERVLEVEL > 0
1829 			if (vm_reserv_reclaim_inactive(domain))
1830 				goto again;
1831 #endif
1832 		}
1833 	}
1834 	if (m == NULL) {
1835 		/*
1836 		 * Not allocatable, give up.
1837 		 */
1838 		if (vm_domain_alloc_fail(vmd, object, req))
1839 			goto again;
1840 		return (NULL);
1841 	}
1842 
1843 	/*
1844 	 * At this point we had better have found a good page.
1845 	 */
1846 found:
1847 	vm_page_dequeue(m);
1848 	vm_page_alloc_check(m);
1849 
1850 	/*
1851 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1852 	 */
1853 	if ((req & VM_ALLOC_ZERO) != 0)
1854 		flags |= (m->flags & PG_ZERO);
1855 	if ((req & VM_ALLOC_NODUMP) != 0)
1856 		flags |= PG_NODUMP;
1857 	m->flags = flags;
1858 	m->aflags = 0;
1859 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1860 	    VPO_UNMANAGED : 0;
1861 	m->busy_lock = VPB_UNBUSIED;
1862 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1863 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1864 	if ((req & VM_ALLOC_SBUSY) != 0)
1865 		m->busy_lock = VPB_SHARERS_WORD(1);
1866 	if (req & VM_ALLOC_WIRED) {
1867 		/*
1868 		 * The page lock is not required for wiring a page until that
1869 		 * page is inserted into the object.
1870 		 */
1871 		vm_wire_add(1);
1872 		m->wire_count = 1;
1873 	}
1874 	m->act_count = 0;
1875 
1876 	if (object != NULL) {
1877 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1878 			if (req & VM_ALLOC_WIRED) {
1879 				vm_wire_sub(1);
1880 				m->wire_count = 0;
1881 			}
1882 			KASSERT(m->object == NULL, ("page %p has object", m));
1883 			m->oflags = VPO_UNMANAGED;
1884 			m->busy_lock = VPB_UNBUSIED;
1885 			/* Don't change PG_ZERO. */
1886 			vm_page_free_toq(m);
1887 			if (req & VM_ALLOC_WAITFAIL) {
1888 				VM_OBJECT_WUNLOCK(object);
1889 				vm_radix_wait();
1890 				VM_OBJECT_WLOCK(object);
1891 			}
1892 			return (NULL);
1893 		}
1894 
1895 		/* Ignore device objects; the pager sets "memattr" for them. */
1896 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1897 		    (object->flags & OBJ_FICTITIOUS) == 0)
1898 			pmap_page_set_memattr(m, object->memattr);
1899 	} else
1900 		m->pindex = pindex;
1901 
1902 	return (m);
1903 }
1904 
1905 /*
1906  *	vm_page_alloc_contig:
1907  *
1908  *	Allocate a contiguous set of physical pages of the given size "npages"
1909  *	from the free lists.  All of the physical pages must be at or above
1910  *	the given physical address "low" and below the given physical address
1911  *	"high".  The given value "alignment" determines the alignment of the
1912  *	first physical page in the set.  If the given value "boundary" is
1913  *	non-zero, then the set of physical pages cannot cross any physical
1914  *	address boundary that is a multiple of that value.  Both "alignment"
1915  *	and "boundary" must be a power of two.
1916  *
1917  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1918  *	then the memory attribute setting for the physical pages is configured
1919  *	to the object's memory attribute setting.  Otherwise, the memory
1920  *	attribute setting for the physical pages is configured to "memattr",
1921  *	overriding the object's memory attribute setting.  However, if the
1922  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1923  *	memory attribute setting for the physical pages cannot be configured
1924  *	to VM_MEMATTR_DEFAULT.
1925  *
1926  *	The specified object may not contain fictitious pages.
1927  *
1928  *	The caller must always specify an allocation class.
1929  *
1930  *	allocation classes:
1931  *	VM_ALLOC_NORMAL		normal process request
1932  *	VM_ALLOC_SYSTEM		system *really* needs a page
1933  *	VM_ALLOC_INTERRUPT	interrupt time request
1934  *
1935  *	optional allocation flags:
1936  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1937  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1938  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1939  *				should not be exclusive busy
1940  *	VM_ALLOC_SBUSY		shared busy the allocated page
1941  *	VM_ALLOC_WIRED		wire the allocated page
1942  *	VM_ALLOC_ZERO		prefer a zeroed page
1943  */
1944 vm_page_t
1945 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1946     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1947     vm_paddr_t boundary, vm_memattr_t memattr)
1948 {
1949 	struct vm_domainset_iter di;
1950 	vm_page_t m;
1951 	int domain;
1952 
1953 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1954 	do {
1955 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1956 		    npages, low, high, alignment, boundary, memattr);
1957 		if (m != NULL)
1958 			break;
1959 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1960 
1961 	return (m);
1962 }
1963 
1964 vm_page_t
1965 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1966     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1967     vm_paddr_t boundary, vm_memattr_t memattr)
1968 {
1969 	struct vm_domain *vmd;
1970 	vm_page_t m, m_ret, mpred;
1971 	u_int busy_lock, flags, oflags;
1972 
1973 	mpred = NULL;	/* XXX: pacify gcc */
1974 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1975 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1976 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1977 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1978 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1979 	    req));
1980 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1981 	    ("Can't sleep and retry object insertion."));
1982 	if (object != NULL) {
1983 		VM_OBJECT_ASSERT_WLOCKED(object);
1984 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1985 		    ("vm_page_alloc_contig: object %p has fictitious pages",
1986 		    object));
1987 	}
1988 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1989 
1990 	if (object != NULL) {
1991 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1992 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1993 		    ("vm_page_alloc_contig: pindex already allocated"));
1994 	}
1995 
1996 	/*
1997 	 * Can we allocate the pages without the number of free pages falling
1998 	 * below the lower bound for the allocation class?
1999 	 */
2000 	m_ret = NULL;
2001 again:
2002 #if VM_NRESERVLEVEL > 0
2003 	/*
2004 	 * Can we allocate the pages from a reservation?
2005 	 */
2006 	if (vm_object_reserv(object) &&
2007 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2008 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2009 		domain = vm_phys_domain(m_ret);
2010 		vmd = VM_DOMAIN(domain);
2011 		goto found;
2012 	}
2013 #endif
2014 	vmd = VM_DOMAIN(domain);
2015 	if (vm_domain_allocate(vmd, req, npages)) {
2016 		/*
2017 		 * allocate them from the free page queues.
2018 		 */
2019 		vm_domain_free_lock(vmd);
2020 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2021 		    alignment, boundary);
2022 		vm_domain_free_unlock(vmd);
2023 		if (m_ret == NULL) {
2024 			vm_domain_freecnt_inc(vmd, npages);
2025 #if VM_NRESERVLEVEL > 0
2026 			if (vm_reserv_reclaim_contig(domain, npages, low,
2027 			    high, alignment, boundary))
2028 				goto again;
2029 #endif
2030 		}
2031 	}
2032 	if (m_ret == NULL) {
2033 		if (vm_domain_alloc_fail(vmd, object, req))
2034 			goto again;
2035 		return (NULL);
2036 	}
2037 #if VM_NRESERVLEVEL > 0
2038 found:
2039 #endif
2040 	for (m = m_ret; m < &m_ret[npages]; m++) {
2041 		vm_page_dequeue(m);
2042 		vm_page_alloc_check(m);
2043 	}
2044 
2045 	/*
2046 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2047 	 */
2048 	flags = 0;
2049 	if ((req & VM_ALLOC_ZERO) != 0)
2050 		flags = PG_ZERO;
2051 	if ((req & VM_ALLOC_NODUMP) != 0)
2052 		flags |= PG_NODUMP;
2053 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2054 	    VPO_UNMANAGED : 0;
2055 	busy_lock = VPB_UNBUSIED;
2056 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2057 		busy_lock = VPB_SINGLE_EXCLUSIVER;
2058 	if ((req & VM_ALLOC_SBUSY) != 0)
2059 		busy_lock = VPB_SHARERS_WORD(1);
2060 	if ((req & VM_ALLOC_WIRED) != 0)
2061 		vm_wire_add(npages);
2062 	if (object != NULL) {
2063 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2064 		    memattr == VM_MEMATTR_DEFAULT)
2065 			memattr = object->memattr;
2066 	}
2067 	for (m = m_ret; m < &m_ret[npages]; m++) {
2068 		m->aflags = 0;
2069 		m->flags = (m->flags | PG_NODUMP) & flags;
2070 		m->busy_lock = busy_lock;
2071 		if ((req & VM_ALLOC_WIRED) != 0)
2072 			m->wire_count = 1;
2073 		m->act_count = 0;
2074 		m->oflags = oflags;
2075 		if (object != NULL) {
2076 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2077 				if ((req & VM_ALLOC_WIRED) != 0)
2078 					vm_wire_sub(npages);
2079 				KASSERT(m->object == NULL,
2080 				    ("page %p has object", m));
2081 				mpred = m;
2082 				for (m = m_ret; m < &m_ret[npages]; m++) {
2083 					if (m <= mpred &&
2084 					    (req & VM_ALLOC_WIRED) != 0)
2085 						m->wire_count = 0;
2086 					m->oflags = VPO_UNMANAGED;
2087 					m->busy_lock = VPB_UNBUSIED;
2088 					/* Don't change PG_ZERO. */
2089 					vm_page_free_toq(m);
2090 				}
2091 				if (req & VM_ALLOC_WAITFAIL) {
2092 					VM_OBJECT_WUNLOCK(object);
2093 					vm_radix_wait();
2094 					VM_OBJECT_WLOCK(object);
2095 				}
2096 				return (NULL);
2097 			}
2098 			mpred = m;
2099 		} else
2100 			m->pindex = pindex;
2101 		if (memattr != VM_MEMATTR_DEFAULT)
2102 			pmap_page_set_memattr(m, memattr);
2103 		pindex++;
2104 	}
2105 	return (m_ret);
2106 }
2107 
2108 /*
2109  * Check a page that has been freshly dequeued from a freelist.
2110  */
2111 static void
2112 vm_page_alloc_check(vm_page_t m)
2113 {
2114 
2115 	KASSERT(m->object == NULL, ("page %p has object", m));
2116 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2117 	    ("page %p has unexpected queue %d, flags %#x",
2118 	    m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2119 	KASSERT(!vm_page_wired(m), ("page %p is wired", m));
2120 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2121 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2122 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2123 	    ("page %p has unexpected memattr %d",
2124 	    m, pmap_page_get_memattr(m)));
2125 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2126 }
2127 
2128 /*
2129  * 	vm_page_alloc_freelist:
2130  *
2131  *	Allocate a physical page from the specified free page list.
2132  *
2133  *	The caller must always specify an allocation class.
2134  *
2135  *	allocation classes:
2136  *	VM_ALLOC_NORMAL		normal process request
2137  *	VM_ALLOC_SYSTEM		system *really* needs a page
2138  *	VM_ALLOC_INTERRUPT	interrupt time request
2139  *
2140  *	optional allocation flags:
2141  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2142  *				intends to allocate
2143  *	VM_ALLOC_WIRED		wire the allocated page
2144  *	VM_ALLOC_ZERO		prefer a zeroed page
2145  */
2146 vm_page_t
2147 vm_page_alloc_freelist(int freelist, int req)
2148 {
2149 	struct vm_domainset_iter di;
2150 	vm_page_t m;
2151 	int domain;
2152 
2153 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2154 	do {
2155 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2156 		if (m != NULL)
2157 			break;
2158 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2159 
2160 	return (m);
2161 }
2162 
2163 vm_page_t
2164 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2165 {
2166 	struct vm_domain *vmd;
2167 	vm_page_t m;
2168 	u_int flags;
2169 
2170 	m = NULL;
2171 	vmd = VM_DOMAIN(domain);
2172 again:
2173 	if (vm_domain_allocate(vmd, req, 1)) {
2174 		vm_domain_free_lock(vmd);
2175 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2176 		    VM_FREEPOOL_DIRECT, 0);
2177 		vm_domain_free_unlock(vmd);
2178 		if (m == NULL)
2179 			vm_domain_freecnt_inc(vmd, 1);
2180 	}
2181 	if (m == NULL) {
2182 		if (vm_domain_alloc_fail(vmd, NULL, req))
2183 			goto again;
2184 		return (NULL);
2185 	}
2186 	vm_page_dequeue(m);
2187 	vm_page_alloc_check(m);
2188 
2189 	/*
2190 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2191 	 */
2192 	m->aflags = 0;
2193 	flags = 0;
2194 	if ((req & VM_ALLOC_ZERO) != 0)
2195 		flags = PG_ZERO;
2196 	m->flags &= flags;
2197 	if ((req & VM_ALLOC_WIRED) != 0) {
2198 		/*
2199 		 * The page lock is not required for wiring a page that does
2200 		 * not belong to an object.
2201 		 */
2202 		vm_wire_add(1);
2203 		m->wire_count = 1;
2204 	}
2205 	/* Unmanaged pages don't use "act_count". */
2206 	m->oflags = VPO_UNMANAGED;
2207 	return (m);
2208 }
2209 
2210 static int
2211 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2212 {
2213 	struct vm_domain *vmd;
2214 	struct vm_pgcache *pgcache;
2215 	int i;
2216 
2217 	pgcache = arg;
2218 	vmd = VM_DOMAIN(pgcache->domain);
2219 	/* Only import if we can bring in a full bucket. */
2220 	if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2221 		return (0);
2222 	domain = vmd->vmd_domain;
2223 	vm_domain_free_lock(vmd);
2224 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2225 	    (vm_page_t *)store);
2226 	vm_domain_free_unlock(vmd);
2227 	if (cnt != i)
2228 		vm_domain_freecnt_inc(vmd, cnt - i);
2229 
2230 	return (i);
2231 }
2232 
2233 static void
2234 vm_page_zone_release(void *arg, void **store, int cnt)
2235 {
2236 	struct vm_domain *vmd;
2237 	struct vm_pgcache *pgcache;
2238 	vm_page_t m;
2239 	int i;
2240 
2241 	pgcache = arg;
2242 	vmd = VM_DOMAIN(pgcache->domain);
2243 	vm_domain_free_lock(vmd);
2244 	for (i = 0; i < cnt; i++) {
2245 		m = (vm_page_t)store[i];
2246 		vm_phys_free_pages(m, 0);
2247 	}
2248 	vm_domain_free_unlock(vmd);
2249 	vm_domain_freecnt_inc(vmd, cnt);
2250 }
2251 
2252 #define	VPSC_ANY	0	/* No restrictions. */
2253 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2254 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2255 
2256 /*
2257  *	vm_page_scan_contig:
2258  *
2259  *	Scan vm_page_array[] between the specified entries "m_start" and
2260  *	"m_end" for a run of contiguous physical pages that satisfy the
2261  *	specified conditions, and return the lowest page in the run.  The
2262  *	specified "alignment" determines the alignment of the lowest physical
2263  *	page in the run.  If the specified "boundary" is non-zero, then the
2264  *	run of physical pages cannot span a physical address that is a
2265  *	multiple of "boundary".
2266  *
2267  *	"m_end" is never dereferenced, so it need not point to a vm_page
2268  *	structure within vm_page_array[].
2269  *
2270  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2271  *	span a hole (or discontiguity) in the physical address space.  Both
2272  *	"alignment" and "boundary" must be a power of two.
2273  */
2274 vm_page_t
2275 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2276     u_long alignment, vm_paddr_t boundary, int options)
2277 {
2278 	struct mtx *m_mtx;
2279 	vm_object_t object;
2280 	vm_paddr_t pa;
2281 	vm_page_t m, m_run;
2282 #if VM_NRESERVLEVEL > 0
2283 	int level;
2284 #endif
2285 	int m_inc, order, run_ext, run_len;
2286 
2287 	KASSERT(npages > 0, ("npages is 0"));
2288 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2289 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2290 	m_run = NULL;
2291 	run_len = 0;
2292 	m_mtx = NULL;
2293 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2294 		KASSERT((m->flags & PG_MARKER) == 0,
2295 		    ("page %p is PG_MARKER", m));
2296 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2297 		    ("fictitious page %p has invalid wire count", m));
2298 
2299 		/*
2300 		 * If the current page would be the start of a run, check its
2301 		 * physical address against the end, alignment, and boundary
2302 		 * conditions.  If it doesn't satisfy these conditions, either
2303 		 * terminate the scan or advance to the next page that
2304 		 * satisfies the failed condition.
2305 		 */
2306 		if (run_len == 0) {
2307 			KASSERT(m_run == NULL, ("m_run != NULL"));
2308 			if (m + npages > m_end)
2309 				break;
2310 			pa = VM_PAGE_TO_PHYS(m);
2311 			if ((pa & (alignment - 1)) != 0) {
2312 				m_inc = atop(roundup2(pa, alignment) - pa);
2313 				continue;
2314 			}
2315 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2316 			    boundary) != 0) {
2317 				m_inc = atop(roundup2(pa, boundary) - pa);
2318 				continue;
2319 			}
2320 		} else
2321 			KASSERT(m_run != NULL, ("m_run == NULL"));
2322 
2323 		vm_page_change_lock(m, &m_mtx);
2324 		m_inc = 1;
2325 retry:
2326 		if (vm_page_wired(m))
2327 			run_ext = 0;
2328 #if VM_NRESERVLEVEL > 0
2329 		else if ((level = vm_reserv_level(m)) >= 0 &&
2330 		    (options & VPSC_NORESERV) != 0) {
2331 			run_ext = 0;
2332 			/* Advance to the end of the reservation. */
2333 			pa = VM_PAGE_TO_PHYS(m);
2334 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2335 			    pa);
2336 		}
2337 #endif
2338 		else if ((object = m->object) != NULL) {
2339 			/*
2340 			 * The page is considered eligible for relocation if
2341 			 * and only if it could be laundered or reclaimed by
2342 			 * the page daemon.
2343 			 */
2344 			if (!VM_OBJECT_TRYRLOCK(object)) {
2345 				mtx_unlock(m_mtx);
2346 				VM_OBJECT_RLOCK(object);
2347 				mtx_lock(m_mtx);
2348 				if (m->object != object) {
2349 					/*
2350 					 * The page may have been freed.
2351 					 */
2352 					VM_OBJECT_RUNLOCK(object);
2353 					goto retry;
2354 				} else if (vm_page_wired(m)) {
2355 					run_ext = 0;
2356 					goto unlock;
2357 				}
2358 			}
2359 			/* Don't care: PG_NODUMP, PG_ZERO. */
2360 			if (object->type != OBJT_DEFAULT &&
2361 			    object->type != OBJT_SWAP &&
2362 			    object->type != OBJT_VNODE) {
2363 				run_ext = 0;
2364 #if VM_NRESERVLEVEL > 0
2365 			} else if ((options & VPSC_NOSUPER) != 0 &&
2366 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2367 				run_ext = 0;
2368 				/* Advance to the end of the superpage. */
2369 				pa = VM_PAGE_TO_PHYS(m);
2370 				m_inc = atop(roundup2(pa + 1,
2371 				    vm_reserv_size(level)) - pa);
2372 #endif
2373 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2374 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2375 				/*
2376 				 * The page is allocated but eligible for
2377 				 * relocation.  Extend the current run by one
2378 				 * page.
2379 				 */
2380 				KASSERT(pmap_page_get_memattr(m) ==
2381 				    VM_MEMATTR_DEFAULT,
2382 				    ("page %p has an unexpected memattr", m));
2383 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2384 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2385 				    ("page %p has unexpected oflags", m));
2386 				/* Don't care: VPO_NOSYNC. */
2387 				run_ext = 1;
2388 			} else
2389 				run_ext = 0;
2390 unlock:
2391 			VM_OBJECT_RUNLOCK(object);
2392 #if VM_NRESERVLEVEL > 0
2393 		} else if (level >= 0) {
2394 			/*
2395 			 * The page is reserved but not yet allocated.  In
2396 			 * other words, it is still free.  Extend the current
2397 			 * run by one page.
2398 			 */
2399 			run_ext = 1;
2400 #endif
2401 		} else if ((order = m->order) < VM_NFREEORDER) {
2402 			/*
2403 			 * The page is enqueued in the physical memory
2404 			 * allocator's free page queues.  Moreover, it is the
2405 			 * first page in a power-of-two-sized run of
2406 			 * contiguous free pages.  Add these pages to the end
2407 			 * of the current run, and jump ahead.
2408 			 */
2409 			run_ext = 1 << order;
2410 			m_inc = 1 << order;
2411 		} else {
2412 			/*
2413 			 * Skip the page for one of the following reasons: (1)
2414 			 * It is enqueued in the physical memory allocator's
2415 			 * free page queues.  However, it is not the first
2416 			 * page in a run of contiguous free pages.  (This case
2417 			 * rarely occurs because the scan is performed in
2418 			 * ascending order.) (2) It is not reserved, and it is
2419 			 * transitioning from free to allocated.  (Conversely,
2420 			 * the transition from allocated to free for managed
2421 			 * pages is blocked by the page lock.) (3) It is
2422 			 * allocated but not contained by an object and not
2423 			 * wired, e.g., allocated by Xen's balloon driver.
2424 			 */
2425 			run_ext = 0;
2426 		}
2427 
2428 		/*
2429 		 * Extend or reset the current run of pages.
2430 		 */
2431 		if (run_ext > 0) {
2432 			if (run_len == 0)
2433 				m_run = m;
2434 			run_len += run_ext;
2435 		} else {
2436 			if (run_len > 0) {
2437 				m_run = NULL;
2438 				run_len = 0;
2439 			}
2440 		}
2441 	}
2442 	if (m_mtx != NULL)
2443 		mtx_unlock(m_mtx);
2444 	if (run_len >= npages)
2445 		return (m_run);
2446 	return (NULL);
2447 }
2448 
2449 /*
2450  *	vm_page_reclaim_run:
2451  *
2452  *	Try to relocate each of the allocated virtual pages within the
2453  *	specified run of physical pages to a new physical address.  Free the
2454  *	physical pages underlying the relocated virtual pages.  A virtual page
2455  *	is relocatable if and only if it could be laundered or reclaimed by
2456  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2457  *	physical address above "high".
2458  *
2459  *	Returns 0 if every physical page within the run was already free or
2460  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2461  *	value indicating why the last attempt to relocate a virtual page was
2462  *	unsuccessful.
2463  *
2464  *	"req_class" must be an allocation class.
2465  */
2466 static int
2467 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2468     vm_paddr_t high)
2469 {
2470 	struct vm_domain *vmd;
2471 	struct mtx *m_mtx;
2472 	struct spglist free;
2473 	vm_object_t object;
2474 	vm_paddr_t pa;
2475 	vm_page_t m, m_end, m_new;
2476 	int error, order, req;
2477 
2478 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2479 	    ("req_class is not an allocation class"));
2480 	SLIST_INIT(&free);
2481 	error = 0;
2482 	m = m_run;
2483 	m_end = m_run + npages;
2484 	m_mtx = NULL;
2485 	for (; error == 0 && m < m_end; m++) {
2486 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2487 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2488 
2489 		/*
2490 		 * Avoid releasing and reacquiring the same page lock.
2491 		 */
2492 		vm_page_change_lock(m, &m_mtx);
2493 retry:
2494 		if (vm_page_wired(m))
2495 			error = EBUSY;
2496 		else if ((object = m->object) != NULL) {
2497 			/*
2498 			 * The page is relocated if and only if it could be
2499 			 * laundered or reclaimed by the page daemon.
2500 			 */
2501 			if (!VM_OBJECT_TRYWLOCK(object)) {
2502 				mtx_unlock(m_mtx);
2503 				VM_OBJECT_WLOCK(object);
2504 				mtx_lock(m_mtx);
2505 				if (m->object != object) {
2506 					/*
2507 					 * The page may have been freed.
2508 					 */
2509 					VM_OBJECT_WUNLOCK(object);
2510 					goto retry;
2511 				} else if (vm_page_wired(m)) {
2512 					error = EBUSY;
2513 					goto unlock;
2514 				}
2515 			}
2516 			/* Don't care: PG_NODUMP, PG_ZERO. */
2517 			if (object->type != OBJT_DEFAULT &&
2518 			    object->type != OBJT_SWAP &&
2519 			    object->type != OBJT_VNODE)
2520 				error = EINVAL;
2521 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2522 				error = EINVAL;
2523 			else if (vm_page_queue(m) != PQ_NONE &&
2524 			    !vm_page_busied(m)) {
2525 				KASSERT(pmap_page_get_memattr(m) ==
2526 				    VM_MEMATTR_DEFAULT,
2527 				    ("page %p has an unexpected memattr", m));
2528 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2529 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2530 				    ("page %p has unexpected oflags", m));
2531 				/* Don't care: VPO_NOSYNC. */
2532 				if (m->valid != 0) {
2533 					/*
2534 					 * First, try to allocate a new page
2535 					 * that is above "high".  Failing
2536 					 * that, try to allocate a new page
2537 					 * that is below "m_run".  Allocate
2538 					 * the new page between the end of
2539 					 * "m_run" and "high" only as a last
2540 					 * resort.
2541 					 */
2542 					req = req_class | VM_ALLOC_NOOBJ;
2543 					if ((m->flags & PG_NODUMP) != 0)
2544 						req |= VM_ALLOC_NODUMP;
2545 					if (trunc_page(high) !=
2546 					    ~(vm_paddr_t)PAGE_MASK) {
2547 						m_new = vm_page_alloc_contig(
2548 						    NULL, 0, req, 1,
2549 						    round_page(high),
2550 						    ~(vm_paddr_t)0,
2551 						    PAGE_SIZE, 0,
2552 						    VM_MEMATTR_DEFAULT);
2553 					} else
2554 						m_new = NULL;
2555 					if (m_new == NULL) {
2556 						pa = VM_PAGE_TO_PHYS(m_run);
2557 						m_new = vm_page_alloc_contig(
2558 						    NULL, 0, req, 1,
2559 						    0, pa - 1, PAGE_SIZE, 0,
2560 						    VM_MEMATTR_DEFAULT);
2561 					}
2562 					if (m_new == NULL) {
2563 						pa += ptoa(npages);
2564 						m_new = vm_page_alloc_contig(
2565 						    NULL, 0, req, 1,
2566 						    pa, high, PAGE_SIZE, 0,
2567 						    VM_MEMATTR_DEFAULT);
2568 					}
2569 					if (m_new == NULL) {
2570 						error = ENOMEM;
2571 						goto unlock;
2572 					}
2573 					KASSERT(!vm_page_wired(m_new),
2574 					    ("page %p is wired", m_new));
2575 
2576 					/*
2577 					 * Replace "m" with the new page.  For
2578 					 * vm_page_replace(), "m" must be busy
2579 					 * and dequeued.  Finally, change "m"
2580 					 * as if vm_page_free() was called.
2581 					 */
2582 					if (object->ref_count != 0)
2583 						pmap_remove_all(m);
2584 					m_new->aflags = m->aflags &
2585 					    ~PGA_QUEUE_STATE_MASK;
2586 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2587 					    ("page %p is managed", m_new));
2588 					m_new->oflags = m->oflags & VPO_NOSYNC;
2589 					pmap_copy_page(m, m_new);
2590 					m_new->valid = m->valid;
2591 					m_new->dirty = m->dirty;
2592 					m->flags &= ~PG_ZERO;
2593 					vm_page_xbusy(m);
2594 					vm_page_dequeue(m);
2595 					vm_page_replace_checked(m_new, object,
2596 					    m->pindex, m);
2597 					if (vm_page_free_prep(m))
2598 						SLIST_INSERT_HEAD(&free, m,
2599 						    plinks.s.ss);
2600 
2601 					/*
2602 					 * The new page must be deactivated
2603 					 * before the object is unlocked.
2604 					 */
2605 					vm_page_change_lock(m_new, &m_mtx);
2606 					vm_page_deactivate(m_new);
2607 				} else {
2608 					m->flags &= ~PG_ZERO;
2609 					vm_page_dequeue(m);
2610 					if (vm_page_free_prep(m))
2611 						SLIST_INSERT_HEAD(&free, m,
2612 						    plinks.s.ss);
2613 					KASSERT(m->dirty == 0,
2614 					    ("page %p is dirty", m));
2615 				}
2616 			} else
2617 				error = EBUSY;
2618 unlock:
2619 			VM_OBJECT_WUNLOCK(object);
2620 		} else {
2621 			MPASS(vm_phys_domain(m) == domain);
2622 			vmd = VM_DOMAIN(domain);
2623 			vm_domain_free_lock(vmd);
2624 			order = m->order;
2625 			if (order < VM_NFREEORDER) {
2626 				/*
2627 				 * The page is enqueued in the physical memory
2628 				 * allocator's free page queues.  Moreover, it
2629 				 * is the first page in a power-of-two-sized
2630 				 * run of contiguous free pages.  Jump ahead
2631 				 * to the last page within that run, and
2632 				 * continue from there.
2633 				 */
2634 				m += (1 << order) - 1;
2635 			}
2636 #if VM_NRESERVLEVEL > 0
2637 			else if (vm_reserv_is_page_free(m))
2638 				order = 0;
2639 #endif
2640 			vm_domain_free_unlock(vmd);
2641 			if (order == VM_NFREEORDER)
2642 				error = EINVAL;
2643 		}
2644 	}
2645 	if (m_mtx != NULL)
2646 		mtx_unlock(m_mtx);
2647 	if ((m = SLIST_FIRST(&free)) != NULL) {
2648 		int cnt;
2649 
2650 		vmd = VM_DOMAIN(domain);
2651 		cnt = 0;
2652 		vm_domain_free_lock(vmd);
2653 		do {
2654 			MPASS(vm_phys_domain(m) == domain);
2655 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2656 			vm_phys_free_pages(m, 0);
2657 			cnt++;
2658 		} while ((m = SLIST_FIRST(&free)) != NULL);
2659 		vm_domain_free_unlock(vmd);
2660 		vm_domain_freecnt_inc(vmd, cnt);
2661 	}
2662 	return (error);
2663 }
2664 
2665 #define	NRUNS	16
2666 
2667 CTASSERT(powerof2(NRUNS));
2668 
2669 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2670 
2671 #define	MIN_RECLAIM	8
2672 
2673 /*
2674  *	vm_page_reclaim_contig:
2675  *
2676  *	Reclaim allocated, contiguous physical memory satisfying the specified
2677  *	conditions by relocating the virtual pages using that physical memory.
2678  *	Returns true if reclamation is successful and false otherwise.  Since
2679  *	relocation requires the allocation of physical pages, reclamation may
2680  *	fail due to a shortage of free pages.  When reclamation fails, callers
2681  *	are expected to perform vm_wait() before retrying a failed allocation
2682  *	operation, e.g., vm_page_alloc_contig().
2683  *
2684  *	The caller must always specify an allocation class through "req".
2685  *
2686  *	allocation classes:
2687  *	VM_ALLOC_NORMAL		normal process request
2688  *	VM_ALLOC_SYSTEM		system *really* needs a page
2689  *	VM_ALLOC_INTERRUPT	interrupt time request
2690  *
2691  *	The optional allocation flags are ignored.
2692  *
2693  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2694  *	must be a power of two.
2695  */
2696 bool
2697 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2698     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2699 {
2700 	struct vm_domain *vmd;
2701 	vm_paddr_t curr_low;
2702 	vm_page_t m_run, m_runs[NRUNS];
2703 	u_long count, reclaimed;
2704 	int error, i, options, req_class;
2705 
2706 	KASSERT(npages > 0, ("npages is 0"));
2707 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2708 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2709 	req_class = req & VM_ALLOC_CLASS_MASK;
2710 
2711 	/*
2712 	 * The page daemon is allowed to dig deeper into the free page list.
2713 	 */
2714 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2715 		req_class = VM_ALLOC_SYSTEM;
2716 
2717 	/*
2718 	 * Return if the number of free pages cannot satisfy the requested
2719 	 * allocation.
2720 	 */
2721 	vmd = VM_DOMAIN(domain);
2722 	count = vmd->vmd_free_count;
2723 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2724 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2725 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2726 		return (false);
2727 
2728 	/*
2729 	 * Scan up to three times, relaxing the restrictions ("options") on
2730 	 * the reclamation of reservations and superpages each time.
2731 	 */
2732 	for (options = VPSC_NORESERV;;) {
2733 		/*
2734 		 * Find the highest runs that satisfy the given constraints
2735 		 * and restrictions, and record them in "m_runs".
2736 		 */
2737 		curr_low = low;
2738 		count = 0;
2739 		for (;;) {
2740 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2741 			    high, alignment, boundary, options);
2742 			if (m_run == NULL)
2743 				break;
2744 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2745 			m_runs[RUN_INDEX(count)] = m_run;
2746 			count++;
2747 		}
2748 
2749 		/*
2750 		 * Reclaim the highest runs in LIFO (descending) order until
2751 		 * the number of reclaimed pages, "reclaimed", is at least
2752 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2753 		 * reclamation is idempotent, and runs will (likely) recur
2754 		 * from one scan to the next as restrictions are relaxed.
2755 		 */
2756 		reclaimed = 0;
2757 		for (i = 0; count > 0 && i < NRUNS; i++) {
2758 			count--;
2759 			m_run = m_runs[RUN_INDEX(count)];
2760 			error = vm_page_reclaim_run(req_class, domain, npages,
2761 			    m_run, high);
2762 			if (error == 0) {
2763 				reclaimed += npages;
2764 				if (reclaimed >= MIN_RECLAIM)
2765 					return (true);
2766 			}
2767 		}
2768 
2769 		/*
2770 		 * Either relax the restrictions on the next scan or return if
2771 		 * the last scan had no restrictions.
2772 		 */
2773 		if (options == VPSC_NORESERV)
2774 			options = VPSC_NOSUPER;
2775 		else if (options == VPSC_NOSUPER)
2776 			options = VPSC_ANY;
2777 		else if (options == VPSC_ANY)
2778 			return (reclaimed != 0);
2779 	}
2780 }
2781 
2782 bool
2783 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2784     u_long alignment, vm_paddr_t boundary)
2785 {
2786 	struct vm_domainset_iter di;
2787 	int domain;
2788 	bool ret;
2789 
2790 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2791 	do {
2792 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2793 		    high, alignment, boundary);
2794 		if (ret)
2795 			break;
2796 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2797 
2798 	return (ret);
2799 }
2800 
2801 /*
2802  * Set the domain in the appropriate page level domainset.
2803  */
2804 void
2805 vm_domain_set(struct vm_domain *vmd)
2806 {
2807 
2808 	mtx_lock(&vm_domainset_lock);
2809 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2810 		vmd->vmd_minset = 1;
2811 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2812 	}
2813 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2814 		vmd->vmd_severeset = 1;
2815 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2816 	}
2817 	mtx_unlock(&vm_domainset_lock);
2818 }
2819 
2820 /*
2821  * Clear the domain from the appropriate page level domainset.
2822  */
2823 void
2824 vm_domain_clear(struct vm_domain *vmd)
2825 {
2826 
2827 	mtx_lock(&vm_domainset_lock);
2828 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2829 		vmd->vmd_minset = 0;
2830 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2831 		if (vm_min_waiters != 0) {
2832 			vm_min_waiters = 0;
2833 			wakeup(&vm_min_domains);
2834 		}
2835 	}
2836 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2837 		vmd->vmd_severeset = 0;
2838 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2839 		if (vm_severe_waiters != 0) {
2840 			vm_severe_waiters = 0;
2841 			wakeup(&vm_severe_domains);
2842 		}
2843 	}
2844 
2845 	/*
2846 	 * If pageout daemon needs pages, then tell it that there are
2847 	 * some free.
2848 	 */
2849 	if (vmd->vmd_pageout_pages_needed &&
2850 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2851 		wakeup(&vmd->vmd_pageout_pages_needed);
2852 		vmd->vmd_pageout_pages_needed = 0;
2853 	}
2854 
2855 	/* See comments in vm_wait_doms(). */
2856 	if (vm_pageproc_waiters) {
2857 		vm_pageproc_waiters = 0;
2858 		wakeup(&vm_pageproc_waiters);
2859 	}
2860 	mtx_unlock(&vm_domainset_lock);
2861 }
2862 
2863 /*
2864  * Wait for free pages to exceed the min threshold globally.
2865  */
2866 void
2867 vm_wait_min(void)
2868 {
2869 
2870 	mtx_lock(&vm_domainset_lock);
2871 	while (vm_page_count_min()) {
2872 		vm_min_waiters++;
2873 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2874 	}
2875 	mtx_unlock(&vm_domainset_lock);
2876 }
2877 
2878 /*
2879  * Wait for free pages to exceed the severe threshold globally.
2880  */
2881 void
2882 vm_wait_severe(void)
2883 {
2884 
2885 	mtx_lock(&vm_domainset_lock);
2886 	while (vm_page_count_severe()) {
2887 		vm_severe_waiters++;
2888 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2889 		    "vmwait", 0);
2890 	}
2891 	mtx_unlock(&vm_domainset_lock);
2892 }
2893 
2894 u_int
2895 vm_wait_count(void)
2896 {
2897 
2898 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2899 }
2900 
2901 void
2902 vm_wait_doms(const domainset_t *wdoms)
2903 {
2904 
2905 	/*
2906 	 * We use racey wakeup synchronization to avoid expensive global
2907 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
2908 	 * To handle this, we only sleep for one tick in this instance.  It
2909 	 * is expected that most allocations for the pageproc will come from
2910 	 * kmem or vm_page_grab* which will use the more specific and
2911 	 * race-free vm_wait_domain().
2912 	 */
2913 	if (curproc == pageproc) {
2914 		mtx_lock(&vm_domainset_lock);
2915 		vm_pageproc_waiters++;
2916 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2917 		    "pageprocwait", 1);
2918 	} else {
2919 		/*
2920 		 * XXX Ideally we would wait only until the allocation could
2921 		 * be satisfied.  This condition can cause new allocators to
2922 		 * consume all freed pages while old allocators wait.
2923 		 */
2924 		mtx_lock(&vm_domainset_lock);
2925 		if (vm_page_count_min_set(wdoms)) {
2926 			vm_min_waiters++;
2927 			msleep(&vm_min_domains, &vm_domainset_lock,
2928 			    PVM | PDROP, "vmwait", 0);
2929 		} else
2930 			mtx_unlock(&vm_domainset_lock);
2931 	}
2932 }
2933 
2934 /*
2935  *	vm_wait_domain:
2936  *
2937  *	Sleep until free pages are available for allocation.
2938  *	- Called in various places after failed memory allocations.
2939  */
2940 void
2941 vm_wait_domain(int domain)
2942 {
2943 	struct vm_domain *vmd;
2944 	domainset_t wdom;
2945 
2946 	vmd = VM_DOMAIN(domain);
2947 	vm_domain_free_assert_unlocked(vmd);
2948 
2949 	if (curproc == pageproc) {
2950 		mtx_lock(&vm_domainset_lock);
2951 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2952 			vmd->vmd_pageout_pages_needed = 1;
2953 			msleep(&vmd->vmd_pageout_pages_needed,
2954 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
2955 		} else
2956 			mtx_unlock(&vm_domainset_lock);
2957 	} else {
2958 		if (pageproc == NULL)
2959 			panic("vm_wait in early boot");
2960 		DOMAINSET_ZERO(&wdom);
2961 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
2962 		vm_wait_doms(&wdom);
2963 	}
2964 }
2965 
2966 /*
2967  *	vm_wait:
2968  *
2969  *	Sleep until free pages are available for allocation in the
2970  *	affinity domains of the obj.  If obj is NULL, the domain set
2971  *	for the calling thread is used.
2972  *	Called in various places after failed memory allocations.
2973  */
2974 void
2975 vm_wait(vm_object_t obj)
2976 {
2977 	struct domainset *d;
2978 
2979 	d = NULL;
2980 
2981 	/*
2982 	 * Carefully fetch pointers only once: the struct domainset
2983 	 * itself is ummutable but the pointer might change.
2984 	 */
2985 	if (obj != NULL)
2986 		d = obj->domain.dr_policy;
2987 	if (d == NULL)
2988 		d = curthread->td_domain.dr_policy;
2989 
2990 	vm_wait_doms(&d->ds_mask);
2991 }
2992 
2993 /*
2994  *	vm_domain_alloc_fail:
2995  *
2996  *	Called when a page allocation function fails.  Informs the
2997  *	pagedaemon and performs the requested wait.  Requires the
2998  *	domain_free and object lock on entry.  Returns with the
2999  *	object lock held and free lock released.  Returns an error when
3000  *	retry is necessary.
3001  *
3002  */
3003 static int
3004 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3005 {
3006 
3007 	vm_domain_free_assert_unlocked(vmd);
3008 
3009 	atomic_add_int(&vmd->vmd_pageout_deficit,
3010 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3011 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3012 		if (object != NULL)
3013 			VM_OBJECT_WUNLOCK(object);
3014 		vm_wait_domain(vmd->vmd_domain);
3015 		if (object != NULL)
3016 			VM_OBJECT_WLOCK(object);
3017 		if (req & VM_ALLOC_WAITOK)
3018 			return (EAGAIN);
3019 	}
3020 
3021 	return (0);
3022 }
3023 
3024 /*
3025  *	vm_waitpfault:
3026  *
3027  *	Sleep until free pages are available for allocation.
3028  *	- Called only in vm_fault so that processes page faulting
3029  *	  can be easily tracked.
3030  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3031  *	  processes will be able to grab memory first.  Do not change
3032  *	  this balance without careful testing first.
3033  */
3034 void
3035 vm_waitpfault(struct domainset *dset)
3036 {
3037 
3038 	/*
3039 	 * XXX Ideally we would wait only until the allocation could
3040 	 * be satisfied.  This condition can cause new allocators to
3041 	 * consume all freed pages while old allocators wait.
3042 	 */
3043 	mtx_lock(&vm_domainset_lock);
3044 	if (vm_page_count_min_set(&dset->ds_mask)) {
3045 		vm_min_waiters++;
3046 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3047 		    "pfault", 0);
3048 	} else
3049 		mtx_unlock(&vm_domainset_lock);
3050 }
3051 
3052 struct vm_pagequeue *
3053 vm_page_pagequeue(vm_page_t m)
3054 {
3055 
3056 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
3057 }
3058 
3059 static struct mtx *
3060 vm_page_pagequeue_lockptr(vm_page_t m)
3061 {
3062 	uint8_t queue;
3063 
3064 	if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3065 		return (NULL);
3066 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex);
3067 }
3068 
3069 static inline void
3070 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3071 {
3072 	struct vm_domain *vmd;
3073 	uint8_t qflags;
3074 
3075 	CRITICAL_ASSERT(curthread);
3076 	vm_pagequeue_assert_locked(pq);
3077 
3078 	/*
3079 	 * The page daemon is allowed to set m->queue = PQ_NONE without
3080 	 * the page queue lock held.  In this case it is about to free the page,
3081 	 * which must not have any queue state.
3082 	 */
3083 	qflags = atomic_load_8(&m->aflags) & PGA_QUEUE_STATE_MASK;
3084 	KASSERT(pq == vm_page_pagequeue(m) || qflags == 0,
3085 	    ("page %p doesn't belong to queue %p but has queue state %#x",
3086 	    m, pq, qflags));
3087 
3088 	if ((qflags & PGA_DEQUEUE) != 0) {
3089 		if (__predict_true((qflags & PGA_ENQUEUED) != 0)) {
3090 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3091 			vm_pagequeue_cnt_dec(pq);
3092 		}
3093 		vm_page_dequeue_complete(m);
3094 	} else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3095 		if ((qflags & PGA_ENQUEUED) != 0)
3096 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3097 		else {
3098 			vm_pagequeue_cnt_inc(pq);
3099 			vm_page_aflag_set(m, PGA_ENQUEUED);
3100 		}
3101 		if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3102 			KASSERT(m->queue == PQ_INACTIVE,
3103 			    ("head enqueue not supported for page %p", m));
3104 			vmd = vm_pagequeue_domain(m);
3105 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3106 		} else
3107 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3108 
3109 		/*
3110 		 * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after
3111 		 * setting PGA_ENQUEUED in order to synchronize with the
3112 		 * page daemon.
3113 		 */
3114 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
3115 	}
3116 }
3117 
3118 static void
3119 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3120     uint8_t queue)
3121 {
3122 	vm_page_t m;
3123 	int i;
3124 
3125 	for (i = 0; i < bq->bq_cnt; i++) {
3126 		m = bq->bq_pa[i];
3127 		if (__predict_false(m->queue != queue))
3128 			continue;
3129 		vm_pqbatch_process_page(pq, m);
3130 	}
3131 	vm_batchqueue_init(bq);
3132 }
3133 
3134 static void
3135 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3136 {
3137 	struct vm_batchqueue *bq;
3138 	struct vm_pagequeue *pq;
3139 	int domain;
3140 
3141 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3142 	    ("page %p is unmanaged", m));
3143 	KASSERT(mtx_owned(vm_page_lockptr(m)) ||
3144 	    (m->object == NULL && (m->aflags & PGA_DEQUEUE) != 0),
3145 	    ("missing synchronization for page %p", m));
3146 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3147 
3148 	domain = vm_phys_domain(m);
3149 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3150 
3151 	critical_enter();
3152 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3153 	if (vm_batchqueue_insert(bq, m)) {
3154 		critical_exit();
3155 		return;
3156 	}
3157 	if (!vm_pagequeue_trylock(pq)) {
3158 		critical_exit();
3159 		vm_pagequeue_lock(pq);
3160 		critical_enter();
3161 		bq = DPCPU_PTR(pqbatch[domain][queue]);
3162 	}
3163 	vm_pqbatch_process(pq, bq, queue);
3164 
3165 	/*
3166 	 * The page may have been logically dequeued before we acquired the
3167 	 * page queue lock.  In this case, since we either hold the page lock
3168 	 * or the page is being freed, a different thread cannot be concurrently
3169 	 * enqueuing the page.
3170 	 */
3171 	if (__predict_true(m->queue == queue))
3172 		vm_pqbatch_process_page(pq, m);
3173 	else {
3174 		KASSERT(m->queue == PQ_NONE,
3175 		    ("invalid queue transition for page %p", m));
3176 		KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3177 		    ("page %p is enqueued with invalid queue index", m));
3178 		vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3179 	}
3180 	vm_pagequeue_unlock(pq);
3181 	critical_exit();
3182 }
3183 
3184 /*
3185  *	vm_page_drain_pqbatch:		[ internal use only ]
3186  *
3187  *	Force all per-CPU page queue batch queues to be drained.  This is
3188  *	intended for use in severe memory shortages, to ensure that pages
3189  *	do not remain stuck in the batch queues.
3190  */
3191 void
3192 vm_page_drain_pqbatch(void)
3193 {
3194 	struct thread *td;
3195 	struct vm_domain *vmd;
3196 	struct vm_pagequeue *pq;
3197 	int cpu, domain, queue;
3198 
3199 	td = curthread;
3200 	CPU_FOREACH(cpu) {
3201 		thread_lock(td);
3202 		sched_bind(td, cpu);
3203 		thread_unlock(td);
3204 
3205 		for (domain = 0; domain < vm_ndomains; domain++) {
3206 			vmd = VM_DOMAIN(domain);
3207 			for (queue = 0; queue < PQ_COUNT; queue++) {
3208 				pq = &vmd->vmd_pagequeues[queue];
3209 				vm_pagequeue_lock(pq);
3210 				critical_enter();
3211 				vm_pqbatch_process(pq,
3212 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3213 				critical_exit();
3214 				vm_pagequeue_unlock(pq);
3215 			}
3216 		}
3217 	}
3218 	thread_lock(td);
3219 	sched_unbind(td);
3220 	thread_unlock(td);
3221 }
3222 
3223 /*
3224  * Complete the logical removal of a page from a page queue.  We must be
3225  * careful to synchronize with the page daemon, which may be concurrently
3226  * examining the page with only the page lock held.  The page must not be
3227  * in a state where it appears to be logically enqueued.
3228  */
3229 static void
3230 vm_page_dequeue_complete(vm_page_t m)
3231 {
3232 
3233 	m->queue = PQ_NONE;
3234 	atomic_thread_fence_rel();
3235 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3236 }
3237 
3238 /*
3239  *	vm_page_dequeue_deferred:	[ internal use only ]
3240  *
3241  *	Request removal of the given page from its current page
3242  *	queue.  Physical removal from the queue may be deferred
3243  *	indefinitely.
3244  *
3245  *	The page must be locked.
3246  */
3247 void
3248 vm_page_dequeue_deferred(vm_page_t m)
3249 {
3250 	uint8_t queue;
3251 
3252 	vm_page_assert_locked(m);
3253 
3254 	if ((queue = vm_page_queue(m)) == PQ_NONE)
3255 		return;
3256 	vm_page_aflag_set(m, PGA_DEQUEUE);
3257 	vm_pqbatch_submit_page(m, queue);
3258 }
3259 
3260 /*
3261  * A variant of vm_page_dequeue_deferred() that does not assert the page
3262  * lock and is only to be called from vm_page_free_prep().  It is just an
3263  * open-coded implementation of vm_page_dequeue_deferred().  Because the
3264  * page is being freed, we can assume that nothing else is scheduling queue
3265  * operations on this page, so we get for free the mutual exclusion that
3266  * is otherwise provided by the page lock.
3267  */
3268 static void
3269 vm_page_dequeue_deferred_free(vm_page_t m)
3270 {
3271 	uint8_t queue;
3272 
3273 	KASSERT(m->object == NULL, ("page %p has an object reference", m));
3274 
3275 	if ((m->aflags & PGA_DEQUEUE) != 0)
3276 		return;
3277 	atomic_thread_fence_acq();
3278 	if ((queue = m->queue) == PQ_NONE)
3279 		return;
3280 	vm_page_aflag_set(m, PGA_DEQUEUE);
3281 	vm_pqbatch_submit_page(m, queue);
3282 }
3283 
3284 /*
3285  *	vm_page_dequeue:
3286  *
3287  *	Remove the page from whichever page queue it's in, if any.
3288  *	The page must either be locked or unallocated.  This constraint
3289  *	ensures that the queue state of the page will remain consistent
3290  *	after this function returns.
3291  */
3292 void
3293 vm_page_dequeue(vm_page_t m)
3294 {
3295 	struct mtx *lock, *lock1;
3296 	struct vm_pagequeue *pq;
3297 	uint8_t aflags;
3298 
3299 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->order == VM_NFREEORDER,
3300 	    ("page %p is allocated and unlocked", m));
3301 
3302 	for (;;) {
3303 		lock = vm_page_pagequeue_lockptr(m);
3304 		if (lock == NULL) {
3305 			/*
3306 			 * A thread may be concurrently executing
3307 			 * vm_page_dequeue_complete().  Ensure that all queue
3308 			 * state is cleared before we return.
3309 			 */
3310 			aflags = atomic_load_8(&m->aflags);
3311 			if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3312 				return;
3313 			KASSERT((aflags & PGA_DEQUEUE) != 0,
3314 			    ("page %p has unexpected queue state flags %#x",
3315 			    m, aflags));
3316 
3317 			/*
3318 			 * Busy wait until the thread updating queue state is
3319 			 * finished.  Such a thread must be executing in a
3320 			 * critical section.
3321 			 */
3322 			cpu_spinwait();
3323 			continue;
3324 		}
3325 		mtx_lock(lock);
3326 		if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock)
3327 			break;
3328 		mtx_unlock(lock);
3329 		lock = lock1;
3330 	}
3331 	KASSERT(lock == vm_page_pagequeue_lockptr(m),
3332 	    ("%s: page %p migrated directly between queues", __func__, m));
3333 	KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3334 	    mtx_owned(vm_page_lockptr(m)),
3335 	    ("%s: queued unlocked page %p", __func__, m));
3336 
3337 	if ((m->aflags & PGA_ENQUEUED) != 0) {
3338 		pq = vm_page_pagequeue(m);
3339 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3340 		vm_pagequeue_cnt_dec(pq);
3341 	}
3342 	vm_page_dequeue_complete(m);
3343 	mtx_unlock(lock);
3344 }
3345 
3346 /*
3347  * Schedule the given page for insertion into the specified page queue.
3348  * Physical insertion of the page may be deferred indefinitely.
3349  */
3350 static void
3351 vm_page_enqueue(vm_page_t m, uint8_t queue)
3352 {
3353 
3354 	vm_page_assert_locked(m);
3355 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3356 	    ("%s: page %p is already enqueued", __func__, m));
3357 
3358 	m->queue = queue;
3359 	if ((m->aflags & PGA_REQUEUE) == 0)
3360 		vm_page_aflag_set(m, PGA_REQUEUE);
3361 	vm_pqbatch_submit_page(m, queue);
3362 }
3363 
3364 /*
3365  *	vm_page_requeue:		[ internal use only ]
3366  *
3367  *	Schedule a requeue of the given page.
3368  *
3369  *	The page must be locked.
3370  */
3371 void
3372 vm_page_requeue(vm_page_t m)
3373 {
3374 
3375 	vm_page_assert_locked(m);
3376 	KASSERT(vm_page_queue(m) != PQ_NONE,
3377 	    ("%s: page %p is not logically enqueued", __func__, m));
3378 
3379 	if ((m->aflags & PGA_REQUEUE) == 0)
3380 		vm_page_aflag_set(m, PGA_REQUEUE);
3381 	vm_pqbatch_submit_page(m, atomic_load_8(&m->queue));
3382 }
3383 
3384 /*
3385  *	vm_page_free_prep:
3386  *
3387  *	Prepares the given page to be put on the free list,
3388  *	disassociating it from any VM object. The caller may return
3389  *	the page to the free list only if this function returns true.
3390  *
3391  *	The object must be locked.  The page must be locked if it is
3392  *	managed.
3393  */
3394 bool
3395 vm_page_free_prep(vm_page_t m)
3396 {
3397 
3398 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3399 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3400 		uint64_t *p;
3401 		int i;
3402 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3403 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3404 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3405 			    m, i, (uintmax_t)*p));
3406 	}
3407 #endif
3408 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3409 		vm_page_lock_assert(m, MA_OWNED);
3410 		KASSERT(!pmap_page_is_mapped(m),
3411 		    ("vm_page_free_prep: freeing mapped page %p", m));
3412 	} else
3413 		KASSERT(m->queue == PQ_NONE,
3414 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3415 	VM_CNT_INC(v_tfree);
3416 
3417 	if (vm_page_sbusied(m))
3418 		panic("vm_page_free_prep: freeing busy page %p", m);
3419 
3420 	if (m->object != NULL)
3421 		(void)vm_page_remove(m);
3422 
3423 	/*
3424 	 * If fictitious remove object association and
3425 	 * return.
3426 	 */
3427 	if ((m->flags & PG_FICTITIOUS) != 0) {
3428 		KASSERT(m->wire_count == 1,
3429 		    ("fictitious page %p is not wired", m));
3430 		KASSERT(m->queue == PQ_NONE,
3431 		    ("fictitious page %p is queued", m));
3432 		return (false);
3433 	}
3434 
3435 	/*
3436 	 * Pages need not be dequeued before they are returned to the physical
3437 	 * memory allocator, but they must at least be marked for a deferred
3438 	 * dequeue.
3439 	 */
3440 	if ((m->oflags & VPO_UNMANAGED) == 0)
3441 		vm_page_dequeue_deferred_free(m);
3442 
3443 	m->valid = 0;
3444 	vm_page_undirty(m);
3445 
3446 	if (vm_page_wired(m) != 0)
3447 		panic("vm_page_free_prep: freeing wired page %p", m);
3448 
3449 	/*
3450 	 * Restore the default memory attribute to the page.
3451 	 */
3452 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3453 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3454 
3455 #if VM_NRESERVLEVEL > 0
3456 	/*
3457 	 * Determine whether the page belongs to a reservation.  If the page was
3458 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3459 	 * as an optimization, we avoid the check in that case.
3460 	 */
3461 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3462 		return (false);
3463 #endif
3464 
3465 	return (true);
3466 }
3467 
3468 /*
3469  *	vm_page_free_toq:
3470  *
3471  *	Returns the given page to the free list, disassociating it
3472  *	from any VM object.
3473  *
3474  *	The object must be locked.  The page must be locked if it is
3475  *	managed.
3476  */
3477 void
3478 vm_page_free_toq(vm_page_t m)
3479 {
3480 	struct vm_domain *vmd;
3481 	uma_zone_t zone;
3482 
3483 	if (!vm_page_free_prep(m))
3484 		return;
3485 
3486 	vmd = vm_pagequeue_domain(m);
3487 	zone = vmd->vmd_pgcache[m->pool].zone;
3488 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3489 		uma_zfree(zone, m);
3490 		return;
3491 	}
3492 	vm_domain_free_lock(vmd);
3493 	vm_phys_free_pages(m, 0);
3494 	vm_domain_free_unlock(vmd);
3495 	vm_domain_freecnt_inc(vmd, 1);
3496 }
3497 
3498 /*
3499  *	vm_page_free_pages_toq:
3500  *
3501  *	Returns a list of pages to the free list, disassociating it
3502  *	from any VM object.  In other words, this is equivalent to
3503  *	calling vm_page_free_toq() for each page of a list of VM objects.
3504  *
3505  *	The objects must be locked.  The pages must be locked if it is
3506  *	managed.
3507  */
3508 void
3509 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3510 {
3511 	vm_page_t m;
3512 	int count;
3513 
3514 	if (SLIST_EMPTY(free))
3515 		return;
3516 
3517 	count = 0;
3518 	while ((m = SLIST_FIRST(free)) != NULL) {
3519 		count++;
3520 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3521 		vm_page_free_toq(m);
3522 	}
3523 
3524 	if (update_wire_count)
3525 		vm_wire_sub(count);
3526 }
3527 
3528 /*
3529  *	vm_page_wire:
3530  *
3531  * Mark this page as wired down.  If the page is fictitious, then
3532  * its wire count must remain one.
3533  *
3534  * The page must be locked.
3535  */
3536 void
3537 vm_page_wire(vm_page_t m)
3538 {
3539 
3540 	vm_page_assert_locked(m);
3541 	if ((m->flags & PG_FICTITIOUS) != 0) {
3542 		KASSERT(m->wire_count == 1,
3543 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
3544 		    m));
3545 		return;
3546 	}
3547 	if (!vm_page_wired(m)) {
3548 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3549 		    m->queue == PQ_NONE,
3550 		    ("vm_page_wire: unmanaged page %p is queued", m));
3551 		vm_wire_add(1);
3552 	}
3553 	m->wire_count++;
3554 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3555 }
3556 
3557 /*
3558  * vm_page_unwire:
3559  *
3560  * Release one wiring of the specified page, potentially allowing it to be
3561  * paged out.  Returns TRUE if the number of wirings transitions to zero and
3562  * FALSE otherwise.
3563  *
3564  * Only managed pages belonging to an object can be paged out.  If the number
3565  * of wirings transitions to zero and the page is eligible for page out, then
3566  * the page is added to the specified paging queue (unless PQ_NONE is
3567  * specified, in which case the page is dequeued if it belongs to a paging
3568  * queue).
3569  *
3570  * If a page is fictitious, then its wire count must always be one.
3571  *
3572  * A managed page must be locked.
3573  */
3574 bool
3575 vm_page_unwire(vm_page_t m, uint8_t queue)
3576 {
3577 	bool unwired;
3578 
3579 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3580 	    ("vm_page_unwire: invalid queue %u request for page %p",
3581 	    queue, m));
3582 	if ((m->oflags & VPO_UNMANAGED) == 0)
3583 		vm_page_assert_locked(m);
3584 
3585 	unwired = vm_page_unwire_noq(m);
3586 	if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3587 		return (unwired);
3588 
3589 	if (vm_page_queue(m) == queue) {
3590 		if (queue == PQ_ACTIVE)
3591 			vm_page_reference(m);
3592 		else if (queue != PQ_NONE)
3593 			vm_page_requeue(m);
3594 	} else {
3595 		vm_page_dequeue(m);
3596 		if (queue != PQ_NONE) {
3597 			vm_page_enqueue(m, queue);
3598 			if (queue == PQ_ACTIVE)
3599 				/* Initialize act_count. */
3600 				vm_page_activate(m);
3601 		}
3602 	}
3603 	return (unwired);
3604 }
3605 
3606 /*
3607  *
3608  * vm_page_unwire_noq:
3609  *
3610  * Unwire a page without (re-)inserting it into a page queue.  It is up
3611  * to the caller to enqueue, requeue, or free the page as appropriate.
3612  * In most cases, vm_page_unwire() should be used instead.
3613  */
3614 bool
3615 vm_page_unwire_noq(vm_page_t m)
3616 {
3617 
3618 	if ((m->oflags & VPO_UNMANAGED) == 0)
3619 		vm_page_assert_locked(m);
3620 	if ((m->flags & PG_FICTITIOUS) != 0) {
3621 		KASSERT(m->wire_count == 1,
3622 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3623 		return (false);
3624 	}
3625 	if (!vm_page_wired(m))
3626 		panic("vm_page_unwire: page %p's wire count is zero", m);
3627 	m->wire_count--;
3628 	if (m->wire_count == 0) {
3629 		vm_wire_sub(1);
3630 		return (true);
3631 	} else
3632 		return (false);
3633 }
3634 
3635 /*
3636  *	vm_page_activate:
3637  *
3638  *	Put the specified page on the active list (if appropriate).
3639  *	Ensure that act_count is at least ACT_INIT but do not otherwise
3640  *	mess with it.
3641  *
3642  *	The page must be locked.
3643  */
3644 void
3645 vm_page_activate(vm_page_t m)
3646 {
3647 
3648 	vm_page_assert_locked(m);
3649 
3650 	if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3651 		return;
3652 	if (vm_page_queue(m) == PQ_ACTIVE) {
3653 		if (m->act_count < ACT_INIT)
3654 			m->act_count = ACT_INIT;
3655 		return;
3656 	}
3657 
3658 	vm_page_dequeue(m);
3659 	if (m->act_count < ACT_INIT)
3660 		m->act_count = ACT_INIT;
3661 	vm_page_enqueue(m, PQ_ACTIVE);
3662 }
3663 
3664 /*
3665  * Move the specified page to the tail of the inactive queue, or requeue
3666  * the page if it is already in the inactive queue.
3667  *
3668  * The page must be locked.
3669  */
3670 void
3671 vm_page_deactivate(vm_page_t m)
3672 {
3673 
3674 	vm_page_assert_locked(m);
3675 
3676 	if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3677 		return;
3678 
3679 	if (!vm_page_inactive(m)) {
3680 		vm_page_dequeue(m);
3681 		vm_page_enqueue(m, PQ_INACTIVE);
3682 	} else
3683 		vm_page_requeue(m);
3684 }
3685 
3686 /*
3687  * Move the specified page close to the head of the inactive queue,
3688  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3689  * As with regular enqueues, we use a per-CPU batch queue to reduce
3690  * contention on the page queue lock.
3691  *
3692  * The page must be locked.
3693  */
3694 void
3695 vm_page_deactivate_noreuse(vm_page_t m)
3696 {
3697 
3698 	vm_page_assert_locked(m);
3699 
3700 	if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3701 		return;
3702 
3703 	if (!vm_page_inactive(m)) {
3704 		vm_page_dequeue(m);
3705 		m->queue = PQ_INACTIVE;
3706 	}
3707 	if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3708 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3709 	vm_pqbatch_submit_page(m, PQ_INACTIVE);
3710 }
3711 
3712 /*
3713  * vm_page_launder
3714  *
3715  * 	Put a page in the laundry, or requeue it if it is already there.
3716  */
3717 void
3718 vm_page_launder(vm_page_t m)
3719 {
3720 
3721 	vm_page_assert_locked(m);
3722 	if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3723 		return;
3724 
3725 	if (vm_page_in_laundry(m))
3726 		vm_page_requeue(m);
3727 	else {
3728 		vm_page_dequeue(m);
3729 		vm_page_enqueue(m, PQ_LAUNDRY);
3730 	}
3731 }
3732 
3733 /*
3734  * vm_page_unswappable
3735  *
3736  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3737  */
3738 void
3739 vm_page_unswappable(vm_page_t m)
3740 {
3741 
3742 	vm_page_assert_locked(m);
3743 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
3744 	    ("page %p already unswappable", m));
3745 
3746 	vm_page_dequeue(m);
3747 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
3748 }
3749 
3750 static void
3751 vm_page_release_toq(vm_page_t m, int flags)
3752 {
3753 
3754 	/*
3755 	 * Use a check of the valid bits to determine whether we should
3756 	 * accelerate reclamation of the page.  The object lock might not be
3757 	 * held here, in which case the check is racy.  At worst we will either
3758 	 * accelerate reclamation of a valid page and violate LRU, or
3759 	 * unnecessarily defer reclamation of an invalid page.
3760 	 *
3761 	 * If we were asked to not cache the page, place it near the head of the
3762 	 * inactive queue so that is reclaimed sooner.
3763 	 */
3764 	if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
3765 		vm_page_deactivate_noreuse(m);
3766 	else if (vm_page_active(m))
3767 		vm_page_reference(m);
3768 	else
3769 		vm_page_deactivate(m);
3770 }
3771 
3772 /*
3773  * Unwire a page and either attempt to free it or re-add it to the page queues.
3774  */
3775 void
3776 vm_page_release(vm_page_t m, int flags)
3777 {
3778 	vm_object_t object;
3779 	bool freed;
3780 
3781 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3782 	    ("vm_page_release: page %p is unmanaged", m));
3783 
3784 	vm_page_lock(m);
3785 	if (m->object != NULL)
3786 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
3787 	if (vm_page_unwire_noq(m)) {
3788 		if ((object = m->object) == NULL) {
3789 			vm_page_free(m);
3790 		} else {
3791 			freed = false;
3792 			if ((flags & VPR_TRYFREE) != 0 && !vm_page_busied(m) &&
3793 			    /* Depends on type stability. */
3794 			    VM_OBJECT_TRYWLOCK(object)) {
3795 				/*
3796 				 * Only free unmapped pages.  The busy test from
3797 				 * before the object was locked cannot be relied
3798 				 * upon.
3799 				 */
3800 				if ((object->ref_count == 0 ||
3801 				    !pmap_page_is_mapped(m)) && m->dirty == 0 &&
3802 				    !vm_page_busied(m)) {
3803 					vm_page_free(m);
3804 					freed = true;
3805 				}
3806 				VM_OBJECT_WUNLOCK(object);
3807 			}
3808 
3809 			if (!freed)
3810 				vm_page_release_toq(m, flags);
3811 		}
3812 	}
3813 	vm_page_unlock(m);
3814 }
3815 
3816 /* See vm_page_release(). */
3817 void
3818 vm_page_release_locked(vm_page_t m, int flags)
3819 {
3820 
3821 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3822 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3823 	    ("vm_page_release_locked: page %p is unmanaged", m));
3824 
3825 	vm_page_lock(m);
3826 	if (vm_page_unwire_noq(m)) {
3827 		if ((flags & VPR_TRYFREE) != 0 &&
3828 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
3829 		    m->dirty == 0 && !vm_page_busied(m)) {
3830 			vm_page_free(m);
3831 		} else {
3832 			vm_page_release_toq(m, flags);
3833 		}
3834 	}
3835 	vm_page_unlock(m);
3836 }
3837 
3838 /*
3839  * vm_page_advise
3840  *
3841  * 	Apply the specified advice to the given page.
3842  *
3843  *	The object and page must be locked.
3844  */
3845 void
3846 vm_page_advise(vm_page_t m, int advice)
3847 {
3848 
3849 	vm_page_assert_locked(m);
3850 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3851 	if (advice == MADV_FREE)
3852 		/*
3853 		 * Mark the page clean.  This will allow the page to be freed
3854 		 * without first paging it out.  MADV_FREE pages are often
3855 		 * quickly reused by malloc(3), so we do not do anything that
3856 		 * would result in a page fault on a later access.
3857 		 */
3858 		vm_page_undirty(m);
3859 	else if (advice != MADV_DONTNEED) {
3860 		if (advice == MADV_WILLNEED)
3861 			vm_page_activate(m);
3862 		return;
3863 	}
3864 
3865 	/*
3866 	 * Clear any references to the page.  Otherwise, the page daemon will
3867 	 * immediately reactivate the page.
3868 	 */
3869 	vm_page_aflag_clear(m, PGA_REFERENCED);
3870 
3871 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3872 		vm_page_dirty(m);
3873 
3874 	/*
3875 	 * Place clean pages near the head of the inactive queue rather than
3876 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3877 	 * the page will be reused quickly.  Dirty pages not already in the
3878 	 * laundry are moved there.
3879 	 */
3880 	if (m->dirty == 0)
3881 		vm_page_deactivate_noreuse(m);
3882 	else if (!vm_page_in_laundry(m))
3883 		vm_page_launder(m);
3884 }
3885 
3886 /*
3887  * Grab a page, waiting until we are waken up due to the page
3888  * changing state.  We keep on waiting, if the page continues
3889  * to be in the object.  If the page doesn't exist, first allocate it
3890  * and then conditionally zero it.
3891  *
3892  * This routine may sleep.
3893  *
3894  * The object must be locked on entry.  The lock will, however, be released
3895  * and reacquired if the routine sleeps.
3896  */
3897 vm_page_t
3898 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3899 {
3900 	vm_page_t m;
3901 	int sleep;
3902 	int pflags;
3903 
3904 	VM_OBJECT_ASSERT_WLOCKED(object);
3905 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3906 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3907 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3908 	pflags = allocflags &
3909 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3910 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3911 		pflags |= VM_ALLOC_WAITFAIL;
3912 retrylookup:
3913 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3914 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3915 		    vm_page_xbusied(m) : vm_page_busied(m);
3916 		if (sleep) {
3917 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3918 				return (NULL);
3919 			/*
3920 			 * Reference the page before unlocking and
3921 			 * sleeping so that the page daemon is less
3922 			 * likely to reclaim it.
3923 			 */
3924 			vm_page_aflag_set(m, PGA_REFERENCED);
3925 			vm_page_lock(m);
3926 			VM_OBJECT_WUNLOCK(object);
3927 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3928 			    VM_ALLOC_IGN_SBUSY) != 0);
3929 			VM_OBJECT_WLOCK(object);
3930 			goto retrylookup;
3931 		} else {
3932 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3933 				vm_page_lock(m);
3934 				vm_page_wire(m);
3935 				vm_page_unlock(m);
3936 			}
3937 			if ((allocflags &
3938 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3939 				vm_page_xbusy(m);
3940 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3941 				vm_page_sbusy(m);
3942 			return (m);
3943 		}
3944 	}
3945 	m = vm_page_alloc(object, pindex, pflags);
3946 	if (m == NULL) {
3947 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3948 			return (NULL);
3949 		goto retrylookup;
3950 	}
3951 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3952 		pmap_zero_page(m);
3953 	return (m);
3954 }
3955 
3956 /*
3957  * Return the specified range of pages from the given object.  For each
3958  * page offset within the range, if a page already exists within the object
3959  * at that offset and it is busy, then wait for it to change state.  If,
3960  * instead, the page doesn't exist, then allocate it.
3961  *
3962  * The caller must always specify an allocation class.
3963  *
3964  * allocation classes:
3965  *	VM_ALLOC_NORMAL		normal process request
3966  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3967  *
3968  * The caller must always specify that the pages are to be busied and/or
3969  * wired.
3970  *
3971  * optional allocation flags:
3972  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3973  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3974  *	VM_ALLOC_NOWAIT		do not sleep
3975  *	VM_ALLOC_SBUSY		set page to sbusy state
3976  *	VM_ALLOC_WIRED		wire the pages
3977  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3978  *
3979  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3980  * may return a partial prefix of the requested range.
3981  */
3982 int
3983 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3984     vm_page_t *ma, int count)
3985 {
3986 	vm_page_t m, mpred;
3987 	int pflags;
3988 	int i;
3989 	bool sleep;
3990 
3991 	VM_OBJECT_ASSERT_WLOCKED(object);
3992 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3993 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3994 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3995 	    (allocflags & VM_ALLOC_WIRED) != 0,
3996 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3997 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3998 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3999 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
4000 	if (count == 0)
4001 		return (0);
4002 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
4003 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
4004 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4005 		pflags |= VM_ALLOC_WAITFAIL;
4006 	i = 0;
4007 retrylookup:
4008 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4009 	if (m == NULL || m->pindex != pindex + i) {
4010 		mpred = m;
4011 		m = NULL;
4012 	} else
4013 		mpred = TAILQ_PREV(m, pglist, listq);
4014 	for (; i < count; i++) {
4015 		if (m != NULL) {
4016 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
4017 			    vm_page_xbusied(m) : vm_page_busied(m);
4018 			if (sleep) {
4019 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4020 					break;
4021 				/*
4022 				 * Reference the page before unlocking and
4023 				 * sleeping so that the page daemon is less
4024 				 * likely to reclaim it.
4025 				 */
4026 				vm_page_aflag_set(m, PGA_REFERENCED);
4027 				vm_page_lock(m);
4028 				VM_OBJECT_WUNLOCK(object);
4029 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
4030 				    VM_ALLOC_IGN_SBUSY) != 0);
4031 				VM_OBJECT_WLOCK(object);
4032 				goto retrylookup;
4033 			}
4034 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
4035 				vm_page_lock(m);
4036 				vm_page_wire(m);
4037 				vm_page_unlock(m);
4038 			}
4039 			if ((allocflags & (VM_ALLOC_NOBUSY |
4040 			    VM_ALLOC_SBUSY)) == 0)
4041 				vm_page_xbusy(m);
4042 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
4043 				vm_page_sbusy(m);
4044 		} else {
4045 			m = vm_page_alloc_after(object, pindex + i,
4046 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4047 			if (m == NULL) {
4048 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4049 					break;
4050 				goto retrylookup;
4051 			}
4052 		}
4053 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
4054 			if ((m->flags & PG_ZERO) == 0)
4055 				pmap_zero_page(m);
4056 			m->valid = VM_PAGE_BITS_ALL;
4057 		}
4058 		ma[i] = mpred = m;
4059 		m = vm_page_next(m);
4060 	}
4061 	return (i);
4062 }
4063 
4064 /*
4065  * Mapping function for valid or dirty bits in a page.
4066  *
4067  * Inputs are required to range within a page.
4068  */
4069 vm_page_bits_t
4070 vm_page_bits(int base, int size)
4071 {
4072 	int first_bit;
4073 	int last_bit;
4074 
4075 	KASSERT(
4076 	    base + size <= PAGE_SIZE,
4077 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4078 	);
4079 
4080 	if (size == 0)		/* handle degenerate case */
4081 		return (0);
4082 
4083 	first_bit = base >> DEV_BSHIFT;
4084 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4085 
4086 	return (((vm_page_bits_t)2 << last_bit) -
4087 	    ((vm_page_bits_t)1 << first_bit));
4088 }
4089 
4090 /*
4091  *	vm_page_set_valid_range:
4092  *
4093  *	Sets portions of a page valid.  The arguments are expected
4094  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4095  *	of any partial chunks touched by the range.  The invalid portion of
4096  *	such chunks will be zeroed.
4097  *
4098  *	(base + size) must be less then or equal to PAGE_SIZE.
4099  */
4100 void
4101 vm_page_set_valid_range(vm_page_t m, int base, int size)
4102 {
4103 	int endoff, frag;
4104 
4105 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4106 	if (size == 0)	/* handle degenerate case */
4107 		return;
4108 
4109 	/*
4110 	 * If the base is not DEV_BSIZE aligned and the valid
4111 	 * bit is clear, we have to zero out a portion of the
4112 	 * first block.
4113 	 */
4114 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4115 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4116 		pmap_zero_page_area(m, frag, base - frag);
4117 
4118 	/*
4119 	 * If the ending offset is not DEV_BSIZE aligned and the
4120 	 * valid bit is clear, we have to zero out a portion of
4121 	 * the last block.
4122 	 */
4123 	endoff = base + size;
4124 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4125 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4126 		pmap_zero_page_area(m, endoff,
4127 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4128 
4129 	/*
4130 	 * Assert that no previously invalid block that is now being validated
4131 	 * is already dirty.
4132 	 */
4133 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4134 	    ("vm_page_set_valid_range: page %p is dirty", m));
4135 
4136 	/*
4137 	 * Set valid bits inclusive of any overlap.
4138 	 */
4139 	m->valid |= vm_page_bits(base, size);
4140 }
4141 
4142 /*
4143  * Clear the given bits from the specified page's dirty field.
4144  */
4145 static __inline void
4146 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4147 {
4148 	uintptr_t addr;
4149 #if PAGE_SIZE < 16384
4150 	int shift;
4151 #endif
4152 
4153 	/*
4154 	 * If the object is locked and the page is neither exclusive busy nor
4155 	 * write mapped, then the page's dirty field cannot possibly be
4156 	 * set by a concurrent pmap operation.
4157 	 */
4158 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4159 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4160 		m->dirty &= ~pagebits;
4161 	else {
4162 		/*
4163 		 * The pmap layer can call vm_page_dirty() without
4164 		 * holding a distinguished lock.  The combination of
4165 		 * the object's lock and an atomic operation suffice
4166 		 * to guarantee consistency of the page dirty field.
4167 		 *
4168 		 * For PAGE_SIZE == 32768 case, compiler already
4169 		 * properly aligns the dirty field, so no forcible
4170 		 * alignment is needed. Only require existence of
4171 		 * atomic_clear_64 when page size is 32768.
4172 		 */
4173 		addr = (uintptr_t)&m->dirty;
4174 #if PAGE_SIZE == 32768
4175 		atomic_clear_64((uint64_t *)addr, pagebits);
4176 #elif PAGE_SIZE == 16384
4177 		atomic_clear_32((uint32_t *)addr, pagebits);
4178 #else		/* PAGE_SIZE <= 8192 */
4179 		/*
4180 		 * Use a trick to perform a 32-bit atomic on the
4181 		 * containing aligned word, to not depend on the existence
4182 		 * of atomic_clear_{8, 16}.
4183 		 */
4184 		shift = addr & (sizeof(uint32_t) - 1);
4185 #if BYTE_ORDER == BIG_ENDIAN
4186 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4187 #else
4188 		shift *= NBBY;
4189 #endif
4190 		addr &= ~(sizeof(uint32_t) - 1);
4191 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
4192 #endif		/* PAGE_SIZE */
4193 	}
4194 }
4195 
4196 /*
4197  *	vm_page_set_validclean:
4198  *
4199  *	Sets portions of a page valid and clean.  The arguments are expected
4200  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4201  *	of any partial chunks touched by the range.  The invalid portion of
4202  *	such chunks will be zero'd.
4203  *
4204  *	(base + size) must be less then or equal to PAGE_SIZE.
4205  */
4206 void
4207 vm_page_set_validclean(vm_page_t m, int base, int size)
4208 {
4209 	vm_page_bits_t oldvalid, pagebits;
4210 	int endoff, frag;
4211 
4212 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4213 	if (size == 0)	/* handle degenerate case */
4214 		return;
4215 
4216 	/*
4217 	 * If the base is not DEV_BSIZE aligned and the valid
4218 	 * bit is clear, we have to zero out a portion of the
4219 	 * first block.
4220 	 */
4221 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4222 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4223 		pmap_zero_page_area(m, frag, base - frag);
4224 
4225 	/*
4226 	 * If the ending offset is not DEV_BSIZE aligned and the
4227 	 * valid bit is clear, we have to zero out a portion of
4228 	 * the last block.
4229 	 */
4230 	endoff = base + size;
4231 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4232 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4233 		pmap_zero_page_area(m, endoff,
4234 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4235 
4236 	/*
4237 	 * Set valid, clear dirty bits.  If validating the entire
4238 	 * page we can safely clear the pmap modify bit.  We also
4239 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
4240 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4241 	 * be set again.
4242 	 *
4243 	 * We set valid bits inclusive of any overlap, but we can only
4244 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4245 	 * the range.
4246 	 */
4247 	oldvalid = m->valid;
4248 	pagebits = vm_page_bits(base, size);
4249 	m->valid |= pagebits;
4250 #if 0	/* NOT YET */
4251 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4252 		frag = DEV_BSIZE - frag;
4253 		base += frag;
4254 		size -= frag;
4255 		if (size < 0)
4256 			size = 0;
4257 	}
4258 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4259 #endif
4260 	if (base == 0 && size == PAGE_SIZE) {
4261 		/*
4262 		 * The page can only be modified within the pmap if it is
4263 		 * mapped, and it can only be mapped if it was previously
4264 		 * fully valid.
4265 		 */
4266 		if (oldvalid == VM_PAGE_BITS_ALL)
4267 			/*
4268 			 * Perform the pmap_clear_modify() first.  Otherwise,
4269 			 * a concurrent pmap operation, such as
4270 			 * pmap_protect(), could clear a modification in the
4271 			 * pmap and set the dirty field on the page before
4272 			 * pmap_clear_modify() had begun and after the dirty
4273 			 * field was cleared here.
4274 			 */
4275 			pmap_clear_modify(m);
4276 		m->dirty = 0;
4277 		m->oflags &= ~VPO_NOSYNC;
4278 	} else if (oldvalid != VM_PAGE_BITS_ALL)
4279 		m->dirty &= ~pagebits;
4280 	else
4281 		vm_page_clear_dirty_mask(m, pagebits);
4282 }
4283 
4284 void
4285 vm_page_clear_dirty(vm_page_t m, int base, int size)
4286 {
4287 
4288 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4289 }
4290 
4291 /*
4292  *	vm_page_set_invalid:
4293  *
4294  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4295  *	valid and dirty bits for the effected areas are cleared.
4296  */
4297 void
4298 vm_page_set_invalid(vm_page_t m, int base, int size)
4299 {
4300 	vm_page_bits_t bits;
4301 	vm_object_t object;
4302 
4303 	object = m->object;
4304 	VM_OBJECT_ASSERT_WLOCKED(object);
4305 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4306 	    size >= object->un_pager.vnp.vnp_size)
4307 		bits = VM_PAGE_BITS_ALL;
4308 	else
4309 		bits = vm_page_bits(base, size);
4310 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4311 	    bits != 0)
4312 		pmap_remove_all(m);
4313 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4314 	    !pmap_page_is_mapped(m),
4315 	    ("vm_page_set_invalid: page %p is mapped", m));
4316 	m->valid &= ~bits;
4317 	m->dirty &= ~bits;
4318 }
4319 
4320 /*
4321  * vm_page_zero_invalid()
4322  *
4323  *	The kernel assumes that the invalid portions of a page contain
4324  *	garbage, but such pages can be mapped into memory by user code.
4325  *	When this occurs, we must zero out the non-valid portions of the
4326  *	page so user code sees what it expects.
4327  *
4328  *	Pages are most often semi-valid when the end of a file is mapped
4329  *	into memory and the file's size is not page aligned.
4330  */
4331 void
4332 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4333 {
4334 	int b;
4335 	int i;
4336 
4337 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4338 	/*
4339 	 * Scan the valid bits looking for invalid sections that
4340 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4341 	 * valid bit may be set ) have already been zeroed by
4342 	 * vm_page_set_validclean().
4343 	 */
4344 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4345 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4346 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4347 			if (i > b) {
4348 				pmap_zero_page_area(m,
4349 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4350 			}
4351 			b = i + 1;
4352 		}
4353 	}
4354 
4355 	/*
4356 	 * setvalid is TRUE when we can safely set the zero'd areas
4357 	 * as being valid.  We can do this if there are no cache consistancy
4358 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4359 	 */
4360 	if (setvalid)
4361 		m->valid = VM_PAGE_BITS_ALL;
4362 }
4363 
4364 /*
4365  *	vm_page_is_valid:
4366  *
4367  *	Is (partial) page valid?  Note that the case where size == 0
4368  *	will return FALSE in the degenerate case where the page is
4369  *	entirely invalid, and TRUE otherwise.
4370  */
4371 int
4372 vm_page_is_valid(vm_page_t m, int base, int size)
4373 {
4374 	vm_page_bits_t bits;
4375 
4376 	VM_OBJECT_ASSERT_LOCKED(m->object);
4377 	bits = vm_page_bits(base, size);
4378 	return (m->valid != 0 && (m->valid & bits) == bits);
4379 }
4380 
4381 /*
4382  * Returns true if all of the specified predicates are true for the entire
4383  * (super)page and false otherwise.
4384  */
4385 bool
4386 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4387 {
4388 	vm_object_t object;
4389 	int i, npages;
4390 
4391 	object = m->object;
4392 	if (skip_m != NULL && skip_m->object != object)
4393 		return (false);
4394 	VM_OBJECT_ASSERT_LOCKED(object);
4395 	npages = atop(pagesizes[m->psind]);
4396 
4397 	/*
4398 	 * The physically contiguous pages that make up a superpage, i.e., a
4399 	 * page with a page size index ("psind") greater than zero, will
4400 	 * occupy adjacent entries in vm_page_array[].
4401 	 */
4402 	for (i = 0; i < npages; i++) {
4403 		/* Always test object consistency, including "skip_m". */
4404 		if (m[i].object != object)
4405 			return (false);
4406 		if (&m[i] == skip_m)
4407 			continue;
4408 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4409 			return (false);
4410 		if ((flags & PS_ALL_DIRTY) != 0) {
4411 			/*
4412 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4413 			 * might stop this case from spuriously returning
4414 			 * "false".  However, that would require a write lock
4415 			 * on the object containing "m[i]".
4416 			 */
4417 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4418 				return (false);
4419 		}
4420 		if ((flags & PS_ALL_VALID) != 0 &&
4421 		    m[i].valid != VM_PAGE_BITS_ALL)
4422 			return (false);
4423 	}
4424 	return (true);
4425 }
4426 
4427 /*
4428  * Set the page's dirty bits if the page is modified.
4429  */
4430 void
4431 vm_page_test_dirty(vm_page_t m)
4432 {
4433 
4434 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4435 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4436 		vm_page_dirty(m);
4437 }
4438 
4439 void
4440 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4441 {
4442 
4443 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4444 }
4445 
4446 void
4447 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4448 {
4449 
4450 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4451 }
4452 
4453 int
4454 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4455 {
4456 
4457 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4458 }
4459 
4460 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4461 void
4462 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4463 {
4464 
4465 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4466 }
4467 
4468 void
4469 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4470 {
4471 
4472 	mtx_assert_(vm_page_lockptr(m), a, file, line);
4473 }
4474 #endif
4475 
4476 #ifdef INVARIANTS
4477 void
4478 vm_page_object_lock_assert(vm_page_t m)
4479 {
4480 
4481 	/*
4482 	 * Certain of the page's fields may only be modified by the
4483 	 * holder of the containing object's lock or the exclusive busy.
4484 	 * holder.  Unfortunately, the holder of the write busy is
4485 	 * not recorded, and thus cannot be checked here.
4486 	 */
4487 	if (m->object != NULL && !vm_page_xbusied(m))
4488 		VM_OBJECT_ASSERT_WLOCKED(m->object);
4489 }
4490 
4491 void
4492 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4493 {
4494 
4495 	if ((bits & PGA_WRITEABLE) == 0)
4496 		return;
4497 
4498 	/*
4499 	 * The PGA_WRITEABLE flag can only be set if the page is
4500 	 * managed, is exclusively busied or the object is locked.
4501 	 * Currently, this flag is only set by pmap_enter().
4502 	 */
4503 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4504 	    ("PGA_WRITEABLE on unmanaged page"));
4505 	if (!vm_page_xbusied(m))
4506 		VM_OBJECT_ASSERT_LOCKED(m->object);
4507 }
4508 #endif
4509 
4510 #include "opt_ddb.h"
4511 #ifdef DDB
4512 #include <sys/kernel.h>
4513 
4514 #include <ddb/ddb.h>
4515 
4516 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4517 {
4518 
4519 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4520 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4521 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4522 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4523 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4524 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4525 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4526 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4527 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4528 }
4529 
4530 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4531 {
4532 	int dom;
4533 
4534 	db_printf("pq_free %d\n", vm_free_count());
4535 	for (dom = 0; dom < vm_ndomains; dom++) {
4536 		db_printf(
4537     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4538 		    dom,
4539 		    vm_dom[dom].vmd_page_count,
4540 		    vm_dom[dom].vmd_free_count,
4541 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4542 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4543 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4544 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4545 	}
4546 }
4547 
4548 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4549 {
4550 	vm_page_t m;
4551 	boolean_t phys, virt;
4552 
4553 	if (!have_addr) {
4554 		db_printf("show pginfo addr\n");
4555 		return;
4556 	}
4557 
4558 	phys = strchr(modif, 'p') != NULL;
4559 	virt = strchr(modif, 'v') != NULL;
4560 	if (virt)
4561 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
4562 	else if (phys)
4563 		m = PHYS_TO_VM_PAGE(addr);
4564 	else
4565 		m = (vm_page_t)addr;
4566 	db_printf(
4567     "page %p obj %p pidx 0x%jx phys 0x%jx q %d wire %d\n"
4568     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4569 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4570 	    m->queue, m->wire_count, m->aflags, m->oflags,
4571 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4572 }
4573 #endif /* DDB */
4574