xref: /freebsd/sys/vm/vm_page.c (revision 1171c633fb097a19e1da87128604190bc6d27341)
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/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/lock.h>
82 #include <sys/malloc.h>
83 #include <sys/mman.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
86 #include <sys/proc.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
89 #include <sys/sbuf.h>
90 #include <sys/sched.h>
91 #include <sys/smp.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
95 
96 #include <vm/vm.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
111 #include <vm/uma.h>
112 #include <vm/uma_int.h>
113 
114 #include <machine/md_var.h>
115 
116 extern int	uma_startup_count(int);
117 extern void	uma_startup(void *, int);
118 extern int	vmem_startup_count(void);
119 
120 struct vm_domain vm_dom[MAXMEMDOM];
121 
122 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
123 
124 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
125 
126 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
127 /* The following fields are protected by the domainset lock. */
128 domainset_t __exclusive_cache_line vm_min_domains;
129 domainset_t __exclusive_cache_line vm_severe_domains;
130 static int vm_min_waiters;
131 static int vm_severe_waiters;
132 static int vm_pageproc_waiters;
133 
134 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0,
135     "VM page statistics");
136 
137 static counter_u64_t pqstate_commit_retries = EARLY_COUNTER;
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
139     CTLFLAG_RD, &pqstate_commit_retries,
140     "Number of failed per-page atomic queue state updates");
141 
142 static counter_u64_t queue_ops = EARLY_COUNTER;
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
144     CTLFLAG_RD, &queue_ops,
145     "Number of batched queue operations");
146 
147 static counter_u64_t queue_nops = EARLY_COUNTER;
148 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
149     CTLFLAG_RD, &queue_nops,
150     "Number of batched queue operations with no effects");
151 
152 static void
153 counter_startup(void)
154 {
155 
156 	pqstate_commit_retries = counter_u64_alloc(M_WAITOK);
157 	queue_ops = counter_u64_alloc(M_WAITOK);
158 	queue_nops = counter_u64_alloc(M_WAITOK);
159 }
160 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
161 
162 /*
163  * bogus page -- for I/O to/from partially complete buffers,
164  * or for paging into sparsely invalid regions.
165  */
166 vm_page_t bogus_page;
167 
168 vm_page_t vm_page_array;
169 long vm_page_array_size;
170 long first_page;
171 
172 static int boot_pages;
173 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
174     &boot_pages, 0,
175     "number of pages allocated for bootstrapping the VM system");
176 
177 static TAILQ_HEAD(, vm_page) blacklist_head;
178 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
179 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
180     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
181 
182 static uma_zone_t fakepg_zone;
183 
184 static void vm_page_alloc_check(vm_page_t m);
185 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
186     const char *wmesg, bool nonshared, bool locked);
187 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
188 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
189 static bool vm_page_free_prep(vm_page_t m);
190 static void vm_page_free_toq(vm_page_t m);
191 static void vm_page_init(void *dummy);
192 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
193     vm_pindex_t pindex, vm_page_t mpred);
194 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
195     vm_page_t mpred);
196 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
197     const uint16_t nflag);
198 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
199     vm_page_t m_run, vm_paddr_t high);
200 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
201 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
202     int req);
203 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
204     int flags);
205 static void vm_page_zone_release(void *arg, void **store, int cnt);
206 
207 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
208 
209 static void
210 vm_page_init(void *dummy)
211 {
212 
213 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
214 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
215 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
216 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
217 }
218 
219 /*
220  * The cache page zone is initialized later since we need to be able to allocate
221  * pages before UMA is fully initialized.
222  */
223 static void
224 vm_page_init_cache_zones(void *dummy __unused)
225 {
226 	struct vm_domain *vmd;
227 	struct vm_pgcache *pgcache;
228 	int cache, domain, maxcache, pool;
229 
230 	maxcache = 0;
231 	TUNABLE_INT_FETCH("vm.pgcache_zone_max", &maxcache);
232 	for (domain = 0; domain < vm_ndomains; domain++) {
233 		vmd = VM_DOMAIN(domain);
234 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
235 			pgcache = &vmd->vmd_pgcache[pool];
236 			pgcache->domain = domain;
237 			pgcache->pool = pool;
238 			pgcache->zone = uma_zcache_create("vm pgcache",
239 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
240 			    vm_page_zone_import, vm_page_zone_release, pgcache,
241 			    UMA_ZONE_VM);
242 
243 			/*
244 			 * Limit each pool's zone to 0.1% of the pages in the
245 			 * domain.
246 			 */
247 			cache = maxcache != 0 ? maxcache :
248 			    vmd->vmd_page_count / 1000;
249 			uma_zone_set_maxcache(pgcache->zone, cache);
250 		}
251 	}
252 }
253 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
254 
255 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
256 #if PAGE_SIZE == 32768
257 #ifdef CTASSERT
258 CTASSERT(sizeof(u_long) >= 8);
259 #endif
260 #endif
261 
262 /*
263  *	vm_set_page_size:
264  *
265  *	Sets the page size, perhaps based upon the memory
266  *	size.  Must be called before any use of page-size
267  *	dependent functions.
268  */
269 void
270 vm_set_page_size(void)
271 {
272 	if (vm_cnt.v_page_size == 0)
273 		vm_cnt.v_page_size = PAGE_SIZE;
274 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
275 		panic("vm_set_page_size: page size not a power of two");
276 }
277 
278 /*
279  *	vm_page_blacklist_next:
280  *
281  *	Find the next entry in the provided string of blacklist
282  *	addresses.  Entries are separated by space, comma, or newline.
283  *	If an invalid integer is encountered then the rest of the
284  *	string is skipped.  Updates the list pointer to the next
285  *	character, or NULL if the string is exhausted or invalid.
286  */
287 static vm_paddr_t
288 vm_page_blacklist_next(char **list, char *end)
289 {
290 	vm_paddr_t bad;
291 	char *cp, *pos;
292 
293 	if (list == NULL || *list == NULL)
294 		return (0);
295 	if (**list =='\0') {
296 		*list = NULL;
297 		return (0);
298 	}
299 
300 	/*
301 	 * If there's no end pointer then the buffer is coming from
302 	 * the kenv and we know it's null-terminated.
303 	 */
304 	if (end == NULL)
305 		end = *list + strlen(*list);
306 
307 	/* Ensure that strtoq() won't walk off the end */
308 	if (*end != '\0') {
309 		if (*end == '\n' || *end == ' ' || *end  == ',')
310 			*end = '\0';
311 		else {
312 			printf("Blacklist not terminated, skipping\n");
313 			*list = NULL;
314 			return (0);
315 		}
316 	}
317 
318 	for (pos = *list; *pos != '\0'; pos = cp) {
319 		bad = strtoq(pos, &cp, 0);
320 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
321 			if (bad == 0) {
322 				if (++cp < end)
323 					continue;
324 				else
325 					break;
326 			}
327 		} else
328 			break;
329 		if (*cp == '\0' || ++cp >= end)
330 			*list = NULL;
331 		else
332 			*list = cp;
333 		return (trunc_page(bad));
334 	}
335 	printf("Garbage in RAM blacklist, skipping\n");
336 	*list = NULL;
337 	return (0);
338 }
339 
340 bool
341 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
342 {
343 	struct vm_domain *vmd;
344 	vm_page_t m;
345 	int ret;
346 
347 	m = vm_phys_paddr_to_vm_page(pa);
348 	if (m == NULL)
349 		return (true); /* page does not exist, no failure */
350 
351 	vmd = vm_pagequeue_domain(m);
352 	vm_domain_free_lock(vmd);
353 	ret = vm_phys_unfree_page(m);
354 	vm_domain_free_unlock(vmd);
355 	if (ret != 0) {
356 		vm_domain_freecnt_inc(vmd, -1);
357 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
358 		if (verbose)
359 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
360 	}
361 	return (ret);
362 }
363 
364 /*
365  *	vm_page_blacklist_check:
366  *
367  *	Iterate through the provided string of blacklist addresses, pulling
368  *	each entry out of the physical allocator free list and putting it
369  *	onto a list for reporting via the vm.page_blacklist sysctl.
370  */
371 static void
372 vm_page_blacklist_check(char *list, char *end)
373 {
374 	vm_paddr_t pa;
375 	char *next;
376 
377 	next = list;
378 	while (next != NULL) {
379 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
380 			continue;
381 		vm_page_blacklist_add(pa, bootverbose);
382 	}
383 }
384 
385 /*
386  *	vm_page_blacklist_load:
387  *
388  *	Search for a special module named "ram_blacklist".  It'll be a
389  *	plain text file provided by the user via the loader directive
390  *	of the same name.
391  */
392 static void
393 vm_page_blacklist_load(char **list, char **end)
394 {
395 	void *mod;
396 	u_char *ptr;
397 	u_int len;
398 
399 	mod = NULL;
400 	ptr = NULL;
401 
402 	mod = preload_search_by_type("ram_blacklist");
403 	if (mod != NULL) {
404 		ptr = preload_fetch_addr(mod);
405 		len = preload_fetch_size(mod);
406         }
407 	*list = ptr;
408 	if (ptr != NULL)
409 		*end = ptr + len;
410 	else
411 		*end = NULL;
412 	return;
413 }
414 
415 static int
416 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
417 {
418 	vm_page_t m;
419 	struct sbuf sbuf;
420 	int error, first;
421 
422 	first = 1;
423 	error = sysctl_wire_old_buffer(req, 0);
424 	if (error != 0)
425 		return (error);
426 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
427 	TAILQ_FOREACH(m, &blacklist_head, listq) {
428 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
429 		    (uintmax_t)m->phys_addr);
430 		first = 0;
431 	}
432 	error = sbuf_finish(&sbuf);
433 	sbuf_delete(&sbuf);
434 	return (error);
435 }
436 
437 /*
438  * Initialize a dummy page for use in scans of the specified paging queue.
439  * In principle, this function only needs to set the flag PG_MARKER.
440  * Nonetheless, it write busies the page as a safety precaution.
441  */
442 static void
443 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
444 {
445 
446 	bzero(marker, sizeof(*marker));
447 	marker->flags = PG_MARKER;
448 	marker->a.flags = aflags;
449 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
450 	marker->a.queue = queue;
451 }
452 
453 static void
454 vm_page_domain_init(int domain)
455 {
456 	struct vm_domain *vmd;
457 	struct vm_pagequeue *pq;
458 	int i;
459 
460 	vmd = VM_DOMAIN(domain);
461 	bzero(vmd, sizeof(*vmd));
462 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
463 	    "vm inactive pagequeue";
464 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
465 	    "vm active pagequeue";
466 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
467 	    "vm laundry pagequeue";
468 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
469 	    "vm unswappable pagequeue";
470 	vmd->vmd_domain = domain;
471 	vmd->vmd_page_count = 0;
472 	vmd->vmd_free_count = 0;
473 	vmd->vmd_segs = 0;
474 	vmd->vmd_oom = FALSE;
475 	for (i = 0; i < PQ_COUNT; i++) {
476 		pq = &vmd->vmd_pagequeues[i];
477 		TAILQ_INIT(&pq->pq_pl);
478 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
479 		    MTX_DEF | MTX_DUPOK);
480 		pq->pq_pdpages = 0;
481 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
482 	}
483 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
484 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
485 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
486 
487 	/*
488 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
489 	 * insertions.
490 	 */
491 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
492 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
493 	    &vmd->vmd_inacthead, plinks.q);
494 
495 	/*
496 	 * The clock pages are used to implement active queue scanning without
497 	 * requeues.  Scans start at clock[0], which is advanced after the scan
498 	 * ends.  When the two clock hands meet, they are reset and scanning
499 	 * resumes from the head of the queue.
500 	 */
501 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
502 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
503 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
504 	    &vmd->vmd_clock[0], plinks.q);
505 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
506 	    &vmd->vmd_clock[1], plinks.q);
507 }
508 
509 /*
510  * Initialize a physical page in preparation for adding it to the free
511  * lists.
512  */
513 static void
514 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
515 {
516 
517 	m->object = NULL;
518 	m->ref_count = 0;
519 	m->busy_lock = VPB_UNBUSIED;
520 	m->flags = m->a.flags = 0;
521 	m->phys_addr = pa;
522 	m->a.queue = PQ_NONE;
523 	m->psind = 0;
524 	m->segind = segind;
525 	m->order = VM_NFREEORDER;
526 	m->pool = VM_FREEPOOL_DEFAULT;
527 	m->valid = m->dirty = 0;
528 	pmap_page_init(m);
529 }
530 
531 #ifndef PMAP_HAS_PAGE_ARRAY
532 static vm_paddr_t
533 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
534 {
535 	vm_paddr_t new_end;
536 
537 	/*
538 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
539 	 * However, because this page is allocated from KVM, out-of-bounds
540 	 * accesses using the direct map will not be trapped.
541 	 */
542 	*vaddr += PAGE_SIZE;
543 
544 	/*
545 	 * Allocate physical memory for the page structures, and map it.
546 	 */
547 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
548 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
549 	    VM_PROT_READ | VM_PROT_WRITE);
550 	vm_page_array_size = page_range;
551 
552 	return (new_end);
553 }
554 #endif
555 
556 /*
557  *	vm_page_startup:
558  *
559  *	Initializes the resident memory module.  Allocates physical memory for
560  *	bootstrapping UMA and some data structures that are used to manage
561  *	physical pages.  Initializes these structures, and populates the free
562  *	page queues.
563  */
564 vm_offset_t
565 vm_page_startup(vm_offset_t vaddr)
566 {
567 	struct vm_phys_seg *seg;
568 	vm_page_t m;
569 	char *list, *listend;
570 	vm_offset_t mapped;
571 	vm_paddr_t end, high_avail, low_avail, new_end, size;
572 	vm_paddr_t page_range __unused;
573 	vm_paddr_t last_pa, pa;
574 	u_long pagecount;
575 	int biggestone, i, segind;
576 #ifdef WITNESS
577 	int witness_size;
578 #endif
579 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
580 	long ii;
581 #endif
582 
583 	vaddr = round_page(vaddr);
584 
585 	vm_phys_early_startup();
586 	biggestone = vm_phys_avail_largest();
587 	end = phys_avail[biggestone+1];
588 
589 	/*
590 	 * Initialize the page and queue locks.
591 	 */
592 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
593 	for (i = 0; i < PA_LOCK_COUNT; i++)
594 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
595 	for (i = 0; i < vm_ndomains; i++)
596 		vm_page_domain_init(i);
597 
598 	/*
599 	 * Allocate memory for use when boot strapping the kernel memory
600 	 * allocator.  Tell UMA how many zones we are going to create
601 	 * before going fully functional.  UMA will add its zones.
602 	 *
603 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
604 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
605 	 */
606 	boot_pages = uma_startup_count(8);
607 
608 #ifndef UMA_MD_SMALL_ALLOC
609 	/* vmem_startup() calls uma_prealloc(). */
610 	boot_pages += vmem_startup_count();
611 	/* vm_map_startup() calls uma_prealloc(). */
612 	boot_pages += howmany(MAX_KMAP,
613 	    slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR));
614 
615 	/*
616 	 * Before we are fully boot strapped we need to account for the
617 	 * following allocations:
618 	 *
619 	 * "KMAP ENTRY" from kmem_init()
620 	 * "vmem btag" from vmem_startup()
621 	 * "vmem" from vmem_create()
622 	 * "KMAP" from vm_map_startup()
623 	 *
624 	 * Each needs at least one page per-domain.
625 	 */
626 	boot_pages += 4 * vm_ndomains;
627 #endif
628 	/*
629 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
630 	 * manually fetch the value.
631 	 */
632 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
633 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
634 	new_end = trunc_page(new_end);
635 	mapped = pmap_map(&vaddr, new_end, end,
636 	    VM_PROT_READ | VM_PROT_WRITE);
637 	bzero((void *)mapped, end - new_end);
638 	uma_startup((void *)mapped, boot_pages);
639 
640 #ifdef WITNESS
641 	witness_size = round_page(witness_startup_count());
642 	new_end -= witness_size;
643 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
644 	    VM_PROT_READ | VM_PROT_WRITE);
645 	bzero((void *)mapped, witness_size);
646 	witness_startup((void *)mapped);
647 #endif
648 
649 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
650     defined(__i386__) || defined(__mips__) || defined(__riscv) || \
651     defined(__powerpc64__)
652 	/*
653 	 * Allocate a bitmap to indicate that a random physical page
654 	 * needs to be included in a minidump.
655 	 *
656 	 * The amd64 port needs this to indicate which direct map pages
657 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
658 	 *
659 	 * However, i386 still needs this workspace internally within the
660 	 * minidump code.  In theory, they are not needed on i386, but are
661 	 * included should the sf_buf code decide to use them.
662 	 */
663 	last_pa = 0;
664 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
665 		if (dump_avail[i + 1] > last_pa)
666 			last_pa = dump_avail[i + 1];
667 	page_range = last_pa / PAGE_SIZE;
668 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
669 	new_end -= vm_page_dump_size;
670 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
671 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
672 	bzero((void *)vm_page_dump, vm_page_dump_size);
673 #else
674 	(void)last_pa;
675 #endif
676 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
677     defined(__riscv) || defined(__powerpc64__)
678 	/*
679 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
680 	 * in a crash dump.  When pmap_map() uses the direct map, they are
681 	 * not automatically included.
682 	 */
683 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
684 		dump_add_page(pa);
685 #endif
686 	phys_avail[biggestone + 1] = new_end;
687 #ifdef __amd64__
688 	/*
689 	 * Request that the physical pages underlying the message buffer be
690 	 * included in a crash dump.  Since the message buffer is accessed
691 	 * through the direct map, they are not automatically included.
692 	 */
693 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
694 	last_pa = pa + round_page(msgbufsize);
695 	while (pa < last_pa) {
696 		dump_add_page(pa);
697 		pa += PAGE_SIZE;
698 	}
699 #endif
700 	/*
701 	 * Compute the number of pages of memory that will be available for
702 	 * use, taking into account the overhead of a page structure per page.
703 	 * In other words, solve
704 	 *	"available physical memory" - round_page(page_range *
705 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
706 	 * for page_range.
707 	 */
708 	low_avail = phys_avail[0];
709 	high_avail = phys_avail[1];
710 	for (i = 0; i < vm_phys_nsegs; i++) {
711 		if (vm_phys_segs[i].start < low_avail)
712 			low_avail = vm_phys_segs[i].start;
713 		if (vm_phys_segs[i].end > high_avail)
714 			high_avail = vm_phys_segs[i].end;
715 	}
716 	/* Skip the first chunk.  It is already accounted for. */
717 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
718 		if (phys_avail[i] < low_avail)
719 			low_avail = phys_avail[i];
720 		if (phys_avail[i + 1] > high_avail)
721 			high_avail = phys_avail[i + 1];
722 	}
723 	first_page = low_avail / PAGE_SIZE;
724 #ifdef VM_PHYSSEG_SPARSE
725 	size = 0;
726 	for (i = 0; i < vm_phys_nsegs; i++)
727 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
728 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
729 		size += phys_avail[i + 1] - phys_avail[i];
730 #elif defined(VM_PHYSSEG_DENSE)
731 	size = high_avail - low_avail;
732 #else
733 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
734 #endif
735 
736 #ifdef PMAP_HAS_PAGE_ARRAY
737 	pmap_page_array_startup(size / PAGE_SIZE);
738 	biggestone = vm_phys_avail_largest();
739 	end = new_end = phys_avail[biggestone + 1];
740 #else
741 #ifdef VM_PHYSSEG_DENSE
742 	/*
743 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
744 	 * the overhead of a page structure per page only if vm_page_array is
745 	 * allocated from the last physical memory chunk.  Otherwise, we must
746 	 * allocate page structures representing the physical memory
747 	 * underlying vm_page_array, even though they will not be used.
748 	 */
749 	if (new_end != high_avail)
750 		page_range = size / PAGE_SIZE;
751 	else
752 #endif
753 	{
754 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
755 
756 		/*
757 		 * If the partial bytes remaining are large enough for
758 		 * a page (PAGE_SIZE) without a corresponding
759 		 * 'struct vm_page', then new_end will contain an
760 		 * extra page after subtracting the length of the VM
761 		 * page array.  Compensate by subtracting an extra
762 		 * page from new_end.
763 		 */
764 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
765 			if (new_end == high_avail)
766 				high_avail -= PAGE_SIZE;
767 			new_end -= PAGE_SIZE;
768 		}
769 	}
770 	end = new_end;
771 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
772 #endif
773 
774 #if VM_NRESERVLEVEL > 0
775 	/*
776 	 * Allocate physical memory for the reservation management system's
777 	 * data structures, and map it.
778 	 */
779 	new_end = vm_reserv_startup(&vaddr, new_end);
780 #endif
781 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
782     defined(__riscv) || defined(__powerpc64__)
783 	/*
784 	 * Include vm_page_array and vm_reserv_array in a crash dump.
785 	 */
786 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
787 		dump_add_page(pa);
788 #endif
789 	phys_avail[biggestone + 1] = new_end;
790 
791 	/*
792 	 * Add physical memory segments corresponding to the available
793 	 * physical pages.
794 	 */
795 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
796 		if (vm_phys_avail_size(i) != 0)
797 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
798 
799 	/*
800 	 * Initialize the physical memory allocator.
801 	 */
802 	vm_phys_init();
803 
804 	/*
805 	 * Initialize the page structures and add every available page to the
806 	 * physical memory allocator's free lists.
807 	 */
808 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
809 	for (ii = 0; ii < vm_page_array_size; ii++) {
810 		m = &vm_page_array[ii];
811 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
812 		m->flags = PG_FICTITIOUS;
813 	}
814 #endif
815 	vm_cnt.v_page_count = 0;
816 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
817 		seg = &vm_phys_segs[segind];
818 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
819 		    m++, pa += PAGE_SIZE)
820 			vm_page_init_page(m, pa, segind);
821 
822 		/*
823 		 * Add the segment to the free lists only if it is covered by
824 		 * one of the ranges in phys_avail.  Because we've added the
825 		 * ranges to the vm_phys_segs array, we can assume that each
826 		 * segment is either entirely contained in one of the ranges,
827 		 * or doesn't overlap any of them.
828 		 */
829 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
830 			struct vm_domain *vmd;
831 
832 			if (seg->start < phys_avail[i] ||
833 			    seg->end > phys_avail[i + 1])
834 				continue;
835 
836 			m = seg->first_page;
837 			pagecount = (u_long)atop(seg->end - seg->start);
838 
839 			vmd = VM_DOMAIN(seg->domain);
840 			vm_domain_free_lock(vmd);
841 			vm_phys_enqueue_contig(m, pagecount);
842 			vm_domain_free_unlock(vmd);
843 			vm_domain_freecnt_inc(vmd, pagecount);
844 			vm_cnt.v_page_count += (u_int)pagecount;
845 
846 			vmd = VM_DOMAIN(seg->domain);
847 			vmd->vmd_page_count += (u_int)pagecount;
848 			vmd->vmd_segs |= 1UL << m->segind;
849 			break;
850 		}
851 	}
852 
853 	/*
854 	 * Remove blacklisted pages from the physical memory allocator.
855 	 */
856 	TAILQ_INIT(&blacklist_head);
857 	vm_page_blacklist_load(&list, &listend);
858 	vm_page_blacklist_check(list, listend);
859 
860 	list = kern_getenv("vm.blacklist");
861 	vm_page_blacklist_check(list, NULL);
862 
863 	freeenv(list);
864 #if VM_NRESERVLEVEL > 0
865 	/*
866 	 * Initialize the reservation management system.
867 	 */
868 	vm_reserv_init();
869 #endif
870 
871 	return (vaddr);
872 }
873 
874 void
875 vm_page_reference(vm_page_t m)
876 {
877 
878 	vm_page_aflag_set(m, PGA_REFERENCED);
879 }
880 
881 static bool
882 vm_page_acquire_flags(vm_page_t m, int allocflags)
883 {
884 	bool locked;
885 
886 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
887 		locked = vm_page_trysbusy(m);
888 	else
889 		locked = vm_page_tryxbusy(m);
890 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
891 		vm_page_wire(m);
892 	return (locked);
893 }
894 
895 /*
896  *	vm_page_busy_sleep_flags
897  *
898  *	Sleep for busy according to VM_ALLOC_ parameters.
899  */
900 static bool
901 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg,
902     int allocflags)
903 {
904 
905 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
906 		return (false);
907 	/*
908 	 * Reference the page before unlocking and
909 	 * sleeping so that the page daemon is less
910 	 * likely to reclaim it.
911 	 */
912 	if ((allocflags & VM_ALLOC_NOCREAT) == 0)
913 		vm_page_aflag_set(m, PGA_REFERENCED);
914 	if (_vm_page_busy_sleep(object, m, wmesg, (allocflags &
915 	    VM_ALLOC_IGN_SBUSY) != 0, true))
916 		VM_OBJECT_WLOCK(object);
917 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
918 		return (false);
919 	return (true);
920 }
921 
922 /*
923  *	vm_page_busy_acquire:
924  *
925  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
926  *	and drop the object lock if necessary.
927  */
928 bool
929 vm_page_busy_acquire(vm_page_t m, int allocflags)
930 {
931 	vm_object_t obj;
932 	bool locked;
933 
934 	/*
935 	 * The page-specific object must be cached because page
936 	 * identity can change during the sleep, causing the
937 	 * re-lock of a different object.
938 	 * It is assumed that a reference to the object is already
939 	 * held by the callers.
940 	 */
941 	obj = m->object;
942 	for (;;) {
943 		if (vm_page_acquire_flags(m, allocflags))
944 			return (true);
945 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
946 			return (false);
947 		if (obj != NULL)
948 			locked = VM_OBJECT_WOWNED(obj);
949 		else
950 			locked = false;
951 		MPASS(locked || vm_page_wired(m));
952 		if (_vm_page_busy_sleep(obj, m, "vmpba",
953 		    (allocflags & VM_ALLOC_SBUSY) != 0, locked))
954 			VM_OBJECT_WLOCK(obj);
955 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
956 			return (false);
957 		KASSERT(m->object == obj || m->object == NULL,
958 		    ("vm_page_busy_acquire: page %p does not belong to %p",
959 		    m, obj));
960 	}
961 }
962 
963 /*
964  *	vm_page_busy_downgrade:
965  *
966  *	Downgrade an exclusive busy page into a single shared busy page.
967  */
968 void
969 vm_page_busy_downgrade(vm_page_t m)
970 {
971 	u_int x;
972 
973 	vm_page_assert_xbusied(m);
974 
975 	x = m->busy_lock;
976 	for (;;) {
977 		if (atomic_fcmpset_rel_int(&m->busy_lock,
978 		    &x, VPB_SHARERS_WORD(1)))
979 			break;
980 	}
981 	if ((x & VPB_BIT_WAITERS) != 0)
982 		wakeup(m);
983 }
984 
985 /*
986  *
987  *	vm_page_busy_tryupgrade:
988  *
989  *	Attempt to upgrade a single shared busy into an exclusive busy.
990  */
991 int
992 vm_page_busy_tryupgrade(vm_page_t m)
993 {
994 	u_int ce, x;
995 
996 	vm_page_assert_sbusied(m);
997 
998 	x = m->busy_lock;
999 	ce = VPB_CURTHREAD_EXCLUSIVE;
1000 	for (;;) {
1001 		if (VPB_SHARERS(x) > 1)
1002 			return (0);
1003 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1004 		    ("vm_page_busy_tryupgrade: invalid lock state"));
1005 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1006 		    ce | (x & VPB_BIT_WAITERS)))
1007 			continue;
1008 		return (1);
1009 	}
1010 }
1011 
1012 /*
1013  *	vm_page_sbusied:
1014  *
1015  *	Return a positive value if the page is shared busied, 0 otherwise.
1016  */
1017 int
1018 vm_page_sbusied(vm_page_t m)
1019 {
1020 	u_int x;
1021 
1022 	x = m->busy_lock;
1023 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1024 }
1025 
1026 /*
1027  *	vm_page_sunbusy:
1028  *
1029  *	Shared unbusy a page.
1030  */
1031 void
1032 vm_page_sunbusy(vm_page_t m)
1033 {
1034 	u_int x;
1035 
1036 	vm_page_assert_sbusied(m);
1037 
1038 	x = m->busy_lock;
1039 	for (;;) {
1040 		if (VPB_SHARERS(x) > 1) {
1041 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1042 			    x - VPB_ONE_SHARER))
1043 				break;
1044 			continue;
1045 		}
1046 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1047 		    ("vm_page_sunbusy: invalid lock state"));
1048 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1049 			continue;
1050 		if ((x & VPB_BIT_WAITERS) == 0)
1051 			break;
1052 		wakeup(m);
1053 		break;
1054 	}
1055 }
1056 
1057 /*
1058  *	vm_page_busy_sleep:
1059  *
1060  *	Sleep if the page is busy, using the page pointer as wchan.
1061  *	This is used to implement the hard-path of busying mechanism.
1062  *
1063  *	If nonshared is true, sleep only if the page is xbusy.
1064  *
1065  *	The object lock must be held on entry and will be released on exit.
1066  */
1067 void
1068 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1069 {
1070 	vm_object_t obj;
1071 
1072 	obj = m->object;
1073 	VM_OBJECT_ASSERT_LOCKED(obj);
1074 	vm_page_lock_assert(m, MA_NOTOWNED);
1075 
1076 	if (!_vm_page_busy_sleep(obj, m, wmesg, nonshared, true))
1077 		VM_OBJECT_DROP(obj);
1078 }
1079 
1080 /*
1081  *	_vm_page_busy_sleep:
1082  *
1083  *	Internal busy sleep function.
1084  */
1085 static bool
1086 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg,
1087     bool nonshared, bool locked)
1088 {
1089 	u_int x;
1090 
1091 	/*
1092 	 * If the object is busy we must wait for that to drain to zero
1093 	 * before trying the page again.
1094 	 */
1095 	if (obj != NULL && vm_object_busied(obj)) {
1096 		if (locked)
1097 			VM_OBJECT_DROP(obj);
1098 		vm_object_busy_wait(obj, wmesg);
1099 		return (locked);
1100 	}
1101 	sleepq_lock(m);
1102 	x = m->busy_lock;
1103 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1104 	    ((x & VPB_BIT_WAITERS) == 0 &&
1105 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1106 		sleepq_release(m);
1107 		return (false);
1108 	}
1109 	if (locked)
1110 		VM_OBJECT_DROP(obj);
1111 	DROP_GIANT();
1112 	sleepq_add(m, NULL, wmesg, 0, 0);
1113 	sleepq_wait(m, PVM);
1114 	PICKUP_GIANT();
1115 	return (locked);
1116 }
1117 
1118 /*
1119  *	vm_page_trysbusy:
1120  *
1121  *	Try to shared busy a page.
1122  *	If the operation succeeds 1 is returned otherwise 0.
1123  *	The operation never sleeps.
1124  */
1125 int
1126 vm_page_trysbusy(vm_page_t m)
1127 {
1128 	vm_object_t obj;
1129 	u_int x;
1130 
1131 	obj = m->object;
1132 	x = m->busy_lock;
1133 	for (;;) {
1134 		if ((x & VPB_BIT_SHARED) == 0)
1135 			return (0);
1136 		/*
1137 		 * Reduce the window for transient busies that will trigger
1138 		 * false negatives in vm_page_ps_test().
1139 		 */
1140 		if (obj != NULL && vm_object_busied(obj))
1141 			return (0);
1142 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1143 		    x + VPB_ONE_SHARER))
1144 			break;
1145 	}
1146 
1147 	/* Refetch the object now that we're guaranteed that it is stable. */
1148 	obj = m->object;
1149 	if (obj != NULL && vm_object_busied(obj)) {
1150 		vm_page_sunbusy(m);
1151 		return (0);
1152 	}
1153 	return (1);
1154 }
1155 
1156 /*
1157  *	vm_page_tryxbusy:
1158  *
1159  *	Try to exclusive busy a page.
1160  *	If the operation succeeds 1 is returned otherwise 0.
1161  *	The operation never sleeps.
1162  */
1163 int
1164 vm_page_tryxbusy(vm_page_t m)
1165 {
1166 	vm_object_t obj;
1167 
1168         if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1169             VPB_CURTHREAD_EXCLUSIVE) == 0)
1170 		return (0);
1171 
1172 	obj = m->object;
1173 	if (obj != NULL && vm_object_busied(obj)) {
1174 		vm_page_xunbusy(m);
1175 		return (0);
1176 	}
1177 	return (1);
1178 }
1179 
1180 static void
1181 vm_page_xunbusy_hard_tail(vm_page_t m)
1182 {
1183 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1184 	/* Wake the waiter. */
1185 	wakeup(m);
1186 }
1187 
1188 /*
1189  *	vm_page_xunbusy_hard:
1190  *
1191  *	Called when unbusy has failed because there is a waiter.
1192  */
1193 void
1194 vm_page_xunbusy_hard(vm_page_t m)
1195 {
1196 	vm_page_assert_xbusied(m);
1197 	vm_page_xunbusy_hard_tail(m);
1198 }
1199 
1200 void
1201 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1202 {
1203 	vm_page_assert_xbusied_unchecked(m);
1204 	vm_page_xunbusy_hard_tail(m);
1205 }
1206 
1207 /*
1208  * Avoid releasing and reacquiring the same page lock.
1209  */
1210 void
1211 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1212 {
1213 	struct mtx *mtx1;
1214 
1215 	mtx1 = vm_page_lockptr(m);
1216 	if (*mtx == mtx1)
1217 		return;
1218 	if (*mtx != NULL)
1219 		mtx_unlock(*mtx);
1220 	*mtx = mtx1;
1221 	mtx_lock(mtx1);
1222 }
1223 
1224 /*
1225  *	vm_page_unhold_pages:
1226  *
1227  *	Unhold each of the pages that is referenced by the given array.
1228  */
1229 void
1230 vm_page_unhold_pages(vm_page_t *ma, int count)
1231 {
1232 
1233 	for (; count != 0; count--) {
1234 		vm_page_unwire(*ma, PQ_ACTIVE);
1235 		ma++;
1236 	}
1237 }
1238 
1239 vm_page_t
1240 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1241 {
1242 	vm_page_t m;
1243 
1244 #ifdef VM_PHYSSEG_SPARSE
1245 	m = vm_phys_paddr_to_vm_page(pa);
1246 	if (m == NULL)
1247 		m = vm_phys_fictitious_to_vm_page(pa);
1248 	return (m);
1249 #elif defined(VM_PHYSSEG_DENSE)
1250 	long pi;
1251 
1252 	pi = atop(pa);
1253 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1254 		m = &vm_page_array[pi - first_page];
1255 		return (m);
1256 	}
1257 	return (vm_phys_fictitious_to_vm_page(pa));
1258 #else
1259 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1260 #endif
1261 }
1262 
1263 /*
1264  *	vm_page_getfake:
1265  *
1266  *	Create a fictitious page with the specified physical address and
1267  *	memory attribute.  The memory attribute is the only the machine-
1268  *	dependent aspect of a fictitious page that must be initialized.
1269  */
1270 vm_page_t
1271 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1272 {
1273 	vm_page_t m;
1274 
1275 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1276 	vm_page_initfake(m, paddr, memattr);
1277 	return (m);
1278 }
1279 
1280 void
1281 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1282 {
1283 
1284 	if ((m->flags & PG_FICTITIOUS) != 0) {
1285 		/*
1286 		 * The page's memattr might have changed since the
1287 		 * previous initialization.  Update the pmap to the
1288 		 * new memattr.
1289 		 */
1290 		goto memattr;
1291 	}
1292 	m->phys_addr = paddr;
1293 	m->a.queue = PQ_NONE;
1294 	/* Fictitious pages don't use "segind". */
1295 	m->flags = PG_FICTITIOUS;
1296 	/* Fictitious pages don't use "order" or "pool". */
1297 	m->oflags = VPO_UNMANAGED;
1298 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1299 	/* Fictitious pages are unevictable. */
1300 	m->ref_count = 1;
1301 	pmap_page_init(m);
1302 memattr:
1303 	pmap_page_set_memattr(m, memattr);
1304 }
1305 
1306 /*
1307  *	vm_page_putfake:
1308  *
1309  *	Release a fictitious page.
1310  */
1311 void
1312 vm_page_putfake(vm_page_t m)
1313 {
1314 
1315 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1316 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1317 	    ("vm_page_putfake: bad page %p", m));
1318 	vm_page_xunbusy(m);
1319 	uma_zfree(fakepg_zone, m);
1320 }
1321 
1322 /*
1323  *	vm_page_updatefake:
1324  *
1325  *	Update the given fictitious page to the specified physical address and
1326  *	memory attribute.
1327  */
1328 void
1329 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1330 {
1331 
1332 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1333 	    ("vm_page_updatefake: bad page %p", m));
1334 	m->phys_addr = paddr;
1335 	pmap_page_set_memattr(m, memattr);
1336 }
1337 
1338 /*
1339  *	vm_page_free:
1340  *
1341  *	Free a page.
1342  */
1343 void
1344 vm_page_free(vm_page_t m)
1345 {
1346 
1347 	m->flags &= ~PG_ZERO;
1348 	vm_page_free_toq(m);
1349 }
1350 
1351 /*
1352  *	vm_page_free_zero:
1353  *
1354  *	Free a page to the zerod-pages queue
1355  */
1356 void
1357 vm_page_free_zero(vm_page_t m)
1358 {
1359 
1360 	m->flags |= PG_ZERO;
1361 	vm_page_free_toq(m);
1362 }
1363 
1364 /*
1365  * Unbusy and handle the page queueing for a page from a getpages request that
1366  * was optionally read ahead or behind.
1367  */
1368 void
1369 vm_page_readahead_finish(vm_page_t m)
1370 {
1371 
1372 	/* We shouldn't put invalid pages on queues. */
1373 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1374 
1375 	/*
1376 	 * Since the page is not the actually needed one, whether it should
1377 	 * be activated or deactivated is not obvious.  Empirical results
1378 	 * have shown that deactivating the page is usually the best choice,
1379 	 * unless the page is wanted by another thread.
1380 	 */
1381 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1382 		vm_page_activate(m);
1383 	else
1384 		vm_page_deactivate(m);
1385 	vm_page_xunbusy_unchecked(m);
1386 }
1387 
1388 /*
1389  *	vm_page_sleep_if_busy:
1390  *
1391  *	Sleep and release the object lock if the page is busied.
1392  *	Returns TRUE if the thread slept.
1393  *
1394  *	The given page must be unlocked and object containing it must
1395  *	be locked.
1396  */
1397 int
1398 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1399 {
1400 	vm_object_t obj;
1401 
1402 	vm_page_lock_assert(m, MA_NOTOWNED);
1403 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1404 
1405 	/*
1406 	 * The page-specific object must be cached because page
1407 	 * identity can change during the sleep, causing the
1408 	 * re-lock of a different object.
1409 	 * It is assumed that a reference to the object is already
1410 	 * held by the callers.
1411 	 */
1412 	obj = m->object;
1413 	if (vm_page_busied(m) || (obj != NULL && obj->busy)) {
1414 		vm_page_busy_sleep(m, msg, false);
1415 		VM_OBJECT_WLOCK(obj);
1416 		return (TRUE);
1417 	}
1418 	return (FALSE);
1419 }
1420 
1421 /*
1422  *	vm_page_sleep_if_xbusy:
1423  *
1424  *	Sleep and release the object lock if the page is xbusied.
1425  *	Returns TRUE if the thread slept.
1426  *
1427  *	The given page must be unlocked and object containing it must
1428  *	be locked.
1429  */
1430 int
1431 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1432 {
1433 	vm_object_t obj;
1434 
1435 	vm_page_lock_assert(m, MA_NOTOWNED);
1436 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1437 
1438 	/*
1439 	 * The page-specific object must be cached because page
1440 	 * identity can change during the sleep, causing the
1441 	 * re-lock of a different object.
1442 	 * It is assumed that a reference to the object is already
1443 	 * held by the callers.
1444 	 */
1445 	obj = m->object;
1446 	if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) {
1447 		vm_page_busy_sleep(m, msg, true);
1448 		VM_OBJECT_WLOCK(obj);
1449 		return (TRUE);
1450 	}
1451 	return (FALSE);
1452 }
1453 
1454 /*
1455  *	vm_page_dirty_KBI:		[ internal use only ]
1456  *
1457  *	Set all bits in the page's dirty field.
1458  *
1459  *	The object containing the specified page must be locked if the
1460  *	call is made from the machine-independent layer.
1461  *
1462  *	See vm_page_clear_dirty_mask().
1463  *
1464  *	This function should only be called by vm_page_dirty().
1465  */
1466 void
1467 vm_page_dirty_KBI(vm_page_t m)
1468 {
1469 
1470 	/* Refer to this operation by its public name. */
1471 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1472 	m->dirty = VM_PAGE_BITS_ALL;
1473 }
1474 
1475 /*
1476  *	vm_page_insert:		[ internal use only ]
1477  *
1478  *	Inserts the given mem entry into the object and object list.
1479  *
1480  *	The object must be locked.
1481  */
1482 int
1483 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1484 {
1485 	vm_page_t mpred;
1486 
1487 	VM_OBJECT_ASSERT_WLOCKED(object);
1488 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1489 	return (vm_page_insert_after(m, object, pindex, mpred));
1490 }
1491 
1492 /*
1493  *	vm_page_insert_after:
1494  *
1495  *	Inserts the page "m" into the specified object at offset "pindex".
1496  *
1497  *	The page "mpred" must immediately precede the offset "pindex" within
1498  *	the specified object.
1499  *
1500  *	The object must be locked.
1501  */
1502 static int
1503 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1504     vm_page_t mpred)
1505 {
1506 	vm_page_t msucc;
1507 
1508 	VM_OBJECT_ASSERT_WLOCKED(object);
1509 	KASSERT(m->object == NULL,
1510 	    ("vm_page_insert_after: page already inserted"));
1511 	if (mpred != NULL) {
1512 		KASSERT(mpred->object == object,
1513 		    ("vm_page_insert_after: object doesn't contain mpred"));
1514 		KASSERT(mpred->pindex < pindex,
1515 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1516 		msucc = TAILQ_NEXT(mpred, listq);
1517 	} else
1518 		msucc = TAILQ_FIRST(&object->memq);
1519 	if (msucc != NULL)
1520 		KASSERT(msucc->pindex > pindex,
1521 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1522 
1523 	/*
1524 	 * Record the object/offset pair in this page.
1525 	 */
1526 	m->object = object;
1527 	m->pindex = pindex;
1528 	m->ref_count |= VPRC_OBJREF;
1529 
1530 	/*
1531 	 * Now link into the object's ordered list of backed pages.
1532 	 */
1533 	if (vm_radix_insert(&object->rtree, m)) {
1534 		m->object = NULL;
1535 		m->pindex = 0;
1536 		m->ref_count &= ~VPRC_OBJREF;
1537 		return (1);
1538 	}
1539 	vm_page_insert_radixdone(m, object, mpred);
1540 	return (0);
1541 }
1542 
1543 /*
1544  *	vm_page_insert_radixdone:
1545  *
1546  *	Complete page "m" insertion into the specified object after the
1547  *	radix trie hooking.
1548  *
1549  *	The page "mpred" must precede the offset "m->pindex" within the
1550  *	specified object.
1551  *
1552  *	The object must be locked.
1553  */
1554 static void
1555 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1556 {
1557 
1558 	VM_OBJECT_ASSERT_WLOCKED(object);
1559 	KASSERT(object != NULL && m->object == object,
1560 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1561 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1562 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1563 	if (mpred != NULL) {
1564 		KASSERT(mpred->object == object,
1565 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1566 		KASSERT(mpred->pindex < m->pindex,
1567 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1568 	}
1569 
1570 	if (mpred != NULL)
1571 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1572 	else
1573 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1574 
1575 	/*
1576 	 * Show that the object has one more resident page.
1577 	 */
1578 	object->resident_page_count++;
1579 
1580 	/*
1581 	 * Hold the vnode until the last page is released.
1582 	 */
1583 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1584 		vhold(object->handle);
1585 
1586 	/*
1587 	 * Since we are inserting a new and possibly dirty page,
1588 	 * update the object's generation count.
1589 	 */
1590 	if (pmap_page_is_write_mapped(m))
1591 		vm_object_set_writeable_dirty(object);
1592 }
1593 
1594 /*
1595  * Do the work to remove a page from its object.  The caller is responsible for
1596  * updating the page's fields to reflect this removal.
1597  */
1598 static void
1599 vm_page_object_remove(vm_page_t m)
1600 {
1601 	vm_object_t object;
1602 	vm_page_t mrem;
1603 
1604 	vm_page_assert_xbusied(m);
1605 	object = m->object;
1606 	VM_OBJECT_ASSERT_WLOCKED(object);
1607 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1608 	    ("page %p is missing its object ref", m));
1609 
1610 	/* Deferred free of swap space. */
1611 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1612 		vm_pager_page_unswapped(m);
1613 
1614 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1615 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1616 
1617 	/*
1618 	 * Now remove from the object's list of backed pages.
1619 	 */
1620 	TAILQ_REMOVE(&object->memq, m, listq);
1621 
1622 	/*
1623 	 * And show that the object has one fewer resident page.
1624 	 */
1625 	object->resident_page_count--;
1626 
1627 	/*
1628 	 * The vnode may now be recycled.
1629 	 */
1630 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1631 		vdrop(object->handle);
1632 }
1633 
1634 /*
1635  *	vm_page_remove:
1636  *
1637  *	Removes the specified page from its containing object, but does not
1638  *	invalidate any backing storage.  Returns true if the object's reference
1639  *	was the last reference to the page, and false otherwise.
1640  *
1641  *	The object must be locked and the page must be exclusively busied.
1642  *	The exclusive busy will be released on return.  If this is not the
1643  *	final ref and the caller does not hold a wire reference it may not
1644  *	continue to access the page.
1645  */
1646 bool
1647 vm_page_remove(vm_page_t m)
1648 {
1649 	bool dropped;
1650 
1651 	dropped = vm_page_remove_xbusy(m);
1652 	vm_page_xunbusy(m);
1653 
1654 	return (dropped);
1655 }
1656 
1657 /*
1658  *	vm_page_remove_xbusy
1659  *
1660  *	Removes the page but leaves the xbusy held.  Returns true if this
1661  *	removed the final ref and false otherwise.
1662  */
1663 bool
1664 vm_page_remove_xbusy(vm_page_t m)
1665 {
1666 
1667 	vm_page_object_remove(m);
1668 	m->object = NULL;
1669 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1670 }
1671 
1672 /*
1673  *	vm_page_lookup:
1674  *
1675  *	Returns the page associated with the object/offset
1676  *	pair specified; if none is found, NULL is returned.
1677  *
1678  *	The object must be locked.
1679  */
1680 vm_page_t
1681 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1682 {
1683 
1684 	VM_OBJECT_ASSERT_LOCKED(object);
1685 	return (vm_radix_lookup(&object->rtree, pindex));
1686 }
1687 
1688 /*
1689  *	vm_page_find_least:
1690  *
1691  *	Returns the page associated with the object with least pindex
1692  *	greater than or equal to the parameter pindex, or NULL.
1693  *
1694  *	The object must be locked.
1695  */
1696 vm_page_t
1697 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1698 {
1699 	vm_page_t m;
1700 
1701 	VM_OBJECT_ASSERT_LOCKED(object);
1702 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1703 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1704 	return (m);
1705 }
1706 
1707 /*
1708  * Returns the given page's successor (by pindex) within the object if it is
1709  * resident; if none is found, NULL is returned.
1710  *
1711  * The object must be locked.
1712  */
1713 vm_page_t
1714 vm_page_next(vm_page_t m)
1715 {
1716 	vm_page_t next;
1717 
1718 	VM_OBJECT_ASSERT_LOCKED(m->object);
1719 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1720 		MPASS(next->object == m->object);
1721 		if (next->pindex != m->pindex + 1)
1722 			next = NULL;
1723 	}
1724 	return (next);
1725 }
1726 
1727 /*
1728  * Returns the given page's predecessor (by pindex) within the object if it is
1729  * resident; if none is found, NULL is returned.
1730  *
1731  * The object must be locked.
1732  */
1733 vm_page_t
1734 vm_page_prev(vm_page_t m)
1735 {
1736 	vm_page_t prev;
1737 
1738 	VM_OBJECT_ASSERT_LOCKED(m->object);
1739 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1740 		MPASS(prev->object == m->object);
1741 		if (prev->pindex != m->pindex - 1)
1742 			prev = NULL;
1743 	}
1744 	return (prev);
1745 }
1746 
1747 /*
1748  * Uses the page mnew as a replacement for an existing page at index
1749  * pindex which must be already present in the object.
1750  *
1751  * Both pages must be exclusively busied on enter.  The old page is
1752  * unbusied on exit.
1753  *
1754  * A return value of true means mold is now free.  If this is not the
1755  * final ref and the caller does not hold a wire reference it may not
1756  * continue to access the page.
1757  */
1758 static bool
1759 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1760     vm_page_t mold)
1761 {
1762 	vm_page_t mret;
1763 	bool dropped;
1764 
1765 	VM_OBJECT_ASSERT_WLOCKED(object);
1766 	vm_page_assert_xbusied(mold);
1767 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1768 	    ("vm_page_replace: page %p already in object", mnew));
1769 
1770 	/*
1771 	 * This function mostly follows vm_page_insert() and
1772 	 * vm_page_remove() without the radix, object count and vnode
1773 	 * dance.  Double check such functions for more comments.
1774 	 */
1775 
1776 	mnew->object = object;
1777 	mnew->pindex = pindex;
1778 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1779 	mret = vm_radix_replace(&object->rtree, mnew);
1780 	KASSERT(mret == mold,
1781 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1782 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1783 	    (mnew->oflags & VPO_UNMANAGED),
1784 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1785 
1786 	/* Keep the resident page list in sorted order. */
1787 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1788 	TAILQ_REMOVE(&object->memq, mold, listq);
1789 	mold->object = NULL;
1790 
1791 	/*
1792 	 * The object's resident_page_count does not change because we have
1793 	 * swapped one page for another, but the generation count should
1794 	 * change if the page is dirty.
1795 	 */
1796 	if (pmap_page_is_write_mapped(mnew))
1797 		vm_object_set_writeable_dirty(object);
1798 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1799 	vm_page_xunbusy(mold);
1800 
1801 	return (dropped);
1802 }
1803 
1804 void
1805 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1806     vm_page_t mold)
1807 {
1808 
1809 	vm_page_assert_xbusied(mnew);
1810 
1811 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1812 		vm_page_free(mold);
1813 }
1814 
1815 /*
1816  *	vm_page_rename:
1817  *
1818  *	Move the given memory entry from its
1819  *	current object to the specified target object/offset.
1820  *
1821  *	Note: swap associated with the page must be invalidated by the move.  We
1822  *	      have to do this for several reasons:  (1) we aren't freeing the
1823  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1824  *	      moving the page from object A to B, and will then later move
1825  *	      the backing store from A to B and we can't have a conflict.
1826  *
1827  *	Note: we *always* dirty the page.  It is necessary both for the
1828  *	      fact that we moved it, and because we may be invalidating
1829  *	      swap.
1830  *
1831  *	The objects must be locked.
1832  */
1833 int
1834 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1835 {
1836 	vm_page_t mpred;
1837 	vm_pindex_t opidx;
1838 
1839 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1840 
1841 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1842 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1843 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1844 	    ("vm_page_rename: pindex already renamed"));
1845 
1846 	/*
1847 	 * Create a custom version of vm_page_insert() which does not depend
1848 	 * by m_prev and can cheat on the implementation aspects of the
1849 	 * function.
1850 	 */
1851 	opidx = m->pindex;
1852 	m->pindex = new_pindex;
1853 	if (vm_radix_insert(&new_object->rtree, m)) {
1854 		m->pindex = opidx;
1855 		return (1);
1856 	}
1857 
1858 	/*
1859 	 * The operation cannot fail anymore.  The removal must happen before
1860 	 * the listq iterator is tainted.
1861 	 */
1862 	m->pindex = opidx;
1863 	vm_page_object_remove(m);
1864 
1865 	/* Return back to the new pindex to complete vm_page_insert(). */
1866 	m->pindex = new_pindex;
1867 	m->object = new_object;
1868 
1869 	vm_page_insert_radixdone(m, new_object, mpred);
1870 	vm_page_dirty(m);
1871 	return (0);
1872 }
1873 
1874 /*
1875  *	vm_page_alloc:
1876  *
1877  *	Allocate and return a page that is associated with the specified
1878  *	object and offset pair.  By default, this page is exclusive busied.
1879  *
1880  *	The caller must always specify an allocation class.
1881  *
1882  *	allocation classes:
1883  *	VM_ALLOC_NORMAL		normal process request
1884  *	VM_ALLOC_SYSTEM		system *really* needs a page
1885  *	VM_ALLOC_INTERRUPT	interrupt time request
1886  *
1887  *	optional allocation flags:
1888  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1889  *				intends to allocate
1890  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1891  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1892  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1893  *				should not be exclusive busy
1894  *	VM_ALLOC_SBUSY		shared busy the allocated page
1895  *	VM_ALLOC_WIRED		wire the allocated page
1896  *	VM_ALLOC_ZERO		prefer a zeroed page
1897  */
1898 vm_page_t
1899 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1900 {
1901 
1902 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1903 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1904 }
1905 
1906 vm_page_t
1907 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1908     int req)
1909 {
1910 
1911 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1912 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1913 	    NULL));
1914 }
1915 
1916 /*
1917  * Allocate a page in the specified object with the given page index.  To
1918  * optimize insertion of the page into the object, the caller must also specifiy
1919  * the resident page in the object with largest index smaller than the given
1920  * page index, or NULL if no such page exists.
1921  */
1922 vm_page_t
1923 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1924     int req, vm_page_t mpred)
1925 {
1926 	struct vm_domainset_iter di;
1927 	vm_page_t m;
1928 	int domain;
1929 
1930 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1931 	do {
1932 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1933 		    mpred);
1934 		if (m != NULL)
1935 			break;
1936 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1937 
1938 	return (m);
1939 }
1940 
1941 /*
1942  * Returns true if the number of free pages exceeds the minimum
1943  * for the request class and false otherwise.
1944  */
1945 static int
1946 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1947 {
1948 	u_int limit, old, new;
1949 
1950 	if (req_class == VM_ALLOC_INTERRUPT)
1951 		limit = 0;
1952 	else if (req_class == VM_ALLOC_SYSTEM)
1953 		limit = vmd->vmd_interrupt_free_min;
1954 	else
1955 		limit = vmd->vmd_free_reserved;
1956 
1957 	/*
1958 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1959 	 */
1960 	limit += npages;
1961 	old = vmd->vmd_free_count;
1962 	do {
1963 		if (old < limit)
1964 			return (0);
1965 		new = old - npages;
1966 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1967 
1968 	/* Wake the page daemon if we've crossed the threshold. */
1969 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1970 		pagedaemon_wakeup(vmd->vmd_domain);
1971 
1972 	/* Only update bitsets on transitions. */
1973 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1974 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1975 		vm_domain_set(vmd);
1976 
1977 	return (1);
1978 }
1979 
1980 int
1981 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1982 {
1983 	int req_class;
1984 
1985 	/*
1986 	 * The page daemon is allowed to dig deeper into the free page list.
1987 	 */
1988 	req_class = req & VM_ALLOC_CLASS_MASK;
1989 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1990 		req_class = VM_ALLOC_SYSTEM;
1991 	return (_vm_domain_allocate(vmd, req_class, npages));
1992 }
1993 
1994 vm_page_t
1995 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1996     int req, vm_page_t mpred)
1997 {
1998 	struct vm_domain *vmd;
1999 	vm_page_t m;
2000 	int flags, pool;
2001 
2002 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2003 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2004 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2005 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2006 	    ("inconsistent object(%p)/req(%x)", object, req));
2007 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2008 	    ("Can't sleep and retry object insertion."));
2009 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2010 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2011 	    (uintmax_t)pindex));
2012 	if (object != NULL)
2013 		VM_OBJECT_ASSERT_WLOCKED(object);
2014 
2015 	flags = 0;
2016 	m = NULL;
2017 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2018 again:
2019 #if VM_NRESERVLEVEL > 0
2020 	/*
2021 	 * Can we allocate the page from a reservation?
2022 	 */
2023 	if (vm_object_reserv(object) &&
2024 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2025 	    NULL) {
2026 		domain = vm_phys_domain(m);
2027 		vmd = VM_DOMAIN(domain);
2028 		goto found;
2029 	}
2030 #endif
2031 	vmd = VM_DOMAIN(domain);
2032 	if (vmd->vmd_pgcache[pool].zone != NULL) {
2033 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
2034 		if (m != NULL) {
2035 			flags |= PG_PCPU_CACHE;
2036 			goto found;
2037 		}
2038 	}
2039 	if (vm_domain_allocate(vmd, req, 1)) {
2040 		/*
2041 		 * If not, allocate it from the free page queues.
2042 		 */
2043 		vm_domain_free_lock(vmd);
2044 		m = vm_phys_alloc_pages(domain, pool, 0);
2045 		vm_domain_free_unlock(vmd);
2046 		if (m == NULL) {
2047 			vm_domain_freecnt_inc(vmd, 1);
2048 #if VM_NRESERVLEVEL > 0
2049 			if (vm_reserv_reclaim_inactive(domain))
2050 				goto again;
2051 #endif
2052 		}
2053 	}
2054 	if (m == NULL) {
2055 		/*
2056 		 * Not allocatable, give up.
2057 		 */
2058 		if (vm_domain_alloc_fail(vmd, object, req))
2059 			goto again;
2060 		return (NULL);
2061 	}
2062 
2063 	/*
2064 	 * At this point we had better have found a good page.
2065 	 */
2066 found:
2067 	vm_page_dequeue(m);
2068 	vm_page_alloc_check(m);
2069 
2070 	/*
2071 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2072 	 */
2073 	if ((req & VM_ALLOC_ZERO) != 0)
2074 		flags |= (m->flags & PG_ZERO);
2075 	if ((req & VM_ALLOC_NODUMP) != 0)
2076 		flags |= PG_NODUMP;
2077 	m->flags = flags;
2078 	m->a.flags = 0;
2079 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2080 	    VPO_UNMANAGED : 0;
2081 	m->busy_lock = VPB_UNBUSIED;
2082 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2083 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2084 	if ((req & VM_ALLOC_SBUSY) != 0)
2085 		m->busy_lock = VPB_SHARERS_WORD(1);
2086 	if (req & VM_ALLOC_WIRED) {
2087 		vm_wire_add(1);
2088 		m->ref_count = 1;
2089 	}
2090 	m->a.act_count = 0;
2091 
2092 	if (object != NULL) {
2093 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2094 			if (req & VM_ALLOC_WIRED) {
2095 				vm_wire_sub(1);
2096 				m->ref_count = 0;
2097 			}
2098 			KASSERT(m->object == NULL, ("page %p has object", m));
2099 			m->oflags = VPO_UNMANAGED;
2100 			m->busy_lock = VPB_UNBUSIED;
2101 			/* Don't change PG_ZERO. */
2102 			vm_page_free_toq(m);
2103 			if (req & VM_ALLOC_WAITFAIL) {
2104 				VM_OBJECT_WUNLOCK(object);
2105 				vm_radix_wait();
2106 				VM_OBJECT_WLOCK(object);
2107 			}
2108 			return (NULL);
2109 		}
2110 
2111 		/* Ignore device objects; the pager sets "memattr" for them. */
2112 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2113 		    (object->flags & OBJ_FICTITIOUS) == 0)
2114 			pmap_page_set_memattr(m, object->memattr);
2115 	} else
2116 		m->pindex = pindex;
2117 
2118 	return (m);
2119 }
2120 
2121 /*
2122  *	vm_page_alloc_contig:
2123  *
2124  *	Allocate a contiguous set of physical pages of the given size "npages"
2125  *	from the free lists.  All of the physical pages must be at or above
2126  *	the given physical address "low" and below the given physical address
2127  *	"high".  The given value "alignment" determines the alignment of the
2128  *	first physical page in the set.  If the given value "boundary" is
2129  *	non-zero, then the set of physical pages cannot cross any physical
2130  *	address boundary that is a multiple of that value.  Both "alignment"
2131  *	and "boundary" must be a power of two.
2132  *
2133  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2134  *	then the memory attribute setting for the physical pages is configured
2135  *	to the object's memory attribute setting.  Otherwise, the memory
2136  *	attribute setting for the physical pages is configured to "memattr",
2137  *	overriding the object's memory attribute setting.  However, if the
2138  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2139  *	memory attribute setting for the physical pages cannot be configured
2140  *	to VM_MEMATTR_DEFAULT.
2141  *
2142  *	The specified object may not contain fictitious pages.
2143  *
2144  *	The caller must always specify an allocation class.
2145  *
2146  *	allocation classes:
2147  *	VM_ALLOC_NORMAL		normal process request
2148  *	VM_ALLOC_SYSTEM		system *really* needs a page
2149  *	VM_ALLOC_INTERRUPT	interrupt time request
2150  *
2151  *	optional allocation flags:
2152  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2153  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2154  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2155  *				should not be exclusive busy
2156  *	VM_ALLOC_SBUSY		shared busy the allocated page
2157  *	VM_ALLOC_WIRED		wire the allocated page
2158  *	VM_ALLOC_ZERO		prefer a zeroed page
2159  */
2160 vm_page_t
2161 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2162     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2163     vm_paddr_t boundary, vm_memattr_t memattr)
2164 {
2165 	struct vm_domainset_iter di;
2166 	vm_page_t m;
2167 	int domain;
2168 
2169 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2170 	do {
2171 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2172 		    npages, low, high, alignment, boundary, memattr);
2173 		if (m != NULL)
2174 			break;
2175 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2176 
2177 	return (m);
2178 }
2179 
2180 vm_page_t
2181 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2182     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2183     vm_paddr_t boundary, vm_memattr_t memattr)
2184 {
2185 	struct vm_domain *vmd;
2186 	vm_page_t m, m_ret, mpred;
2187 	u_int busy_lock, flags, oflags;
2188 
2189 	mpred = NULL;	/* XXX: pacify gcc */
2190 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2191 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2192 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2193 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2194 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2195 	    req));
2196 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2197 	    ("Can't sleep and retry object insertion."));
2198 	if (object != NULL) {
2199 		VM_OBJECT_ASSERT_WLOCKED(object);
2200 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2201 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2202 		    object));
2203 	}
2204 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2205 
2206 	if (object != NULL) {
2207 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2208 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2209 		    ("vm_page_alloc_contig: pindex already allocated"));
2210 	}
2211 
2212 	/*
2213 	 * Can we allocate the pages without the number of free pages falling
2214 	 * below the lower bound for the allocation class?
2215 	 */
2216 	m_ret = NULL;
2217 again:
2218 #if VM_NRESERVLEVEL > 0
2219 	/*
2220 	 * Can we allocate the pages from a reservation?
2221 	 */
2222 	if (vm_object_reserv(object) &&
2223 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2224 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2225 		domain = vm_phys_domain(m_ret);
2226 		vmd = VM_DOMAIN(domain);
2227 		goto found;
2228 	}
2229 #endif
2230 	vmd = VM_DOMAIN(domain);
2231 	if (vm_domain_allocate(vmd, req, npages)) {
2232 		/*
2233 		 * allocate them from the free page queues.
2234 		 */
2235 		vm_domain_free_lock(vmd);
2236 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2237 		    alignment, boundary);
2238 		vm_domain_free_unlock(vmd);
2239 		if (m_ret == NULL) {
2240 			vm_domain_freecnt_inc(vmd, npages);
2241 #if VM_NRESERVLEVEL > 0
2242 			if (vm_reserv_reclaim_contig(domain, npages, low,
2243 			    high, alignment, boundary))
2244 				goto again;
2245 #endif
2246 		}
2247 	}
2248 	if (m_ret == NULL) {
2249 		if (vm_domain_alloc_fail(vmd, object, req))
2250 			goto again;
2251 		return (NULL);
2252 	}
2253 #if VM_NRESERVLEVEL > 0
2254 found:
2255 #endif
2256 	for (m = m_ret; m < &m_ret[npages]; m++) {
2257 		vm_page_dequeue(m);
2258 		vm_page_alloc_check(m);
2259 	}
2260 
2261 	/*
2262 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2263 	 */
2264 	flags = 0;
2265 	if ((req & VM_ALLOC_ZERO) != 0)
2266 		flags = PG_ZERO;
2267 	if ((req & VM_ALLOC_NODUMP) != 0)
2268 		flags |= PG_NODUMP;
2269 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2270 	    VPO_UNMANAGED : 0;
2271 	busy_lock = VPB_UNBUSIED;
2272 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2273 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2274 	if ((req & VM_ALLOC_SBUSY) != 0)
2275 		busy_lock = VPB_SHARERS_WORD(1);
2276 	if ((req & VM_ALLOC_WIRED) != 0)
2277 		vm_wire_add(npages);
2278 	if (object != NULL) {
2279 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2280 		    memattr == VM_MEMATTR_DEFAULT)
2281 			memattr = object->memattr;
2282 	}
2283 	for (m = m_ret; m < &m_ret[npages]; m++) {
2284 		m->a.flags = 0;
2285 		m->flags = (m->flags | PG_NODUMP) & flags;
2286 		m->busy_lock = busy_lock;
2287 		if ((req & VM_ALLOC_WIRED) != 0)
2288 			m->ref_count = 1;
2289 		m->a.act_count = 0;
2290 		m->oflags = oflags;
2291 		if (object != NULL) {
2292 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2293 				if ((req & VM_ALLOC_WIRED) != 0)
2294 					vm_wire_sub(npages);
2295 				KASSERT(m->object == NULL,
2296 				    ("page %p has object", m));
2297 				mpred = m;
2298 				for (m = m_ret; m < &m_ret[npages]; m++) {
2299 					if (m <= mpred &&
2300 					    (req & VM_ALLOC_WIRED) != 0)
2301 						m->ref_count = 0;
2302 					m->oflags = VPO_UNMANAGED;
2303 					m->busy_lock = VPB_UNBUSIED;
2304 					/* Don't change PG_ZERO. */
2305 					vm_page_free_toq(m);
2306 				}
2307 				if (req & VM_ALLOC_WAITFAIL) {
2308 					VM_OBJECT_WUNLOCK(object);
2309 					vm_radix_wait();
2310 					VM_OBJECT_WLOCK(object);
2311 				}
2312 				return (NULL);
2313 			}
2314 			mpred = m;
2315 		} else
2316 			m->pindex = pindex;
2317 		if (memattr != VM_MEMATTR_DEFAULT)
2318 			pmap_page_set_memattr(m, memattr);
2319 		pindex++;
2320 	}
2321 	return (m_ret);
2322 }
2323 
2324 /*
2325  * Check a page that has been freshly dequeued from a freelist.
2326  */
2327 static void
2328 vm_page_alloc_check(vm_page_t m)
2329 {
2330 
2331 	KASSERT(m->object == NULL, ("page %p has object", m));
2332 	KASSERT(m->a.queue == PQ_NONE &&
2333 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2334 	    ("page %p has unexpected queue %d, flags %#x",
2335 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2336 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2337 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2338 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2339 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2340 	    ("page %p has unexpected memattr %d",
2341 	    m, pmap_page_get_memattr(m)));
2342 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2343 }
2344 
2345 /*
2346  * 	vm_page_alloc_freelist:
2347  *
2348  *	Allocate a physical page from the specified free page list.
2349  *
2350  *	The caller must always specify an allocation class.
2351  *
2352  *	allocation classes:
2353  *	VM_ALLOC_NORMAL		normal process request
2354  *	VM_ALLOC_SYSTEM		system *really* needs a page
2355  *	VM_ALLOC_INTERRUPT	interrupt time request
2356  *
2357  *	optional allocation flags:
2358  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2359  *				intends to allocate
2360  *	VM_ALLOC_WIRED		wire the allocated page
2361  *	VM_ALLOC_ZERO		prefer a zeroed page
2362  */
2363 vm_page_t
2364 vm_page_alloc_freelist(int freelist, int req)
2365 {
2366 	struct vm_domainset_iter di;
2367 	vm_page_t m;
2368 	int domain;
2369 
2370 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2371 	do {
2372 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2373 		if (m != NULL)
2374 			break;
2375 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2376 
2377 	return (m);
2378 }
2379 
2380 vm_page_t
2381 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2382 {
2383 	struct vm_domain *vmd;
2384 	vm_page_t m;
2385 	u_int flags;
2386 
2387 	m = NULL;
2388 	vmd = VM_DOMAIN(domain);
2389 again:
2390 	if (vm_domain_allocate(vmd, req, 1)) {
2391 		vm_domain_free_lock(vmd);
2392 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2393 		    VM_FREEPOOL_DIRECT, 0);
2394 		vm_domain_free_unlock(vmd);
2395 		if (m == NULL)
2396 			vm_domain_freecnt_inc(vmd, 1);
2397 	}
2398 	if (m == NULL) {
2399 		if (vm_domain_alloc_fail(vmd, NULL, req))
2400 			goto again;
2401 		return (NULL);
2402 	}
2403 	vm_page_dequeue(m);
2404 	vm_page_alloc_check(m);
2405 
2406 	/*
2407 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2408 	 */
2409 	m->a.flags = 0;
2410 	flags = 0;
2411 	if ((req & VM_ALLOC_ZERO) != 0)
2412 		flags = PG_ZERO;
2413 	m->flags &= flags;
2414 	if ((req & VM_ALLOC_WIRED) != 0) {
2415 		vm_wire_add(1);
2416 		m->ref_count = 1;
2417 	}
2418 	/* Unmanaged pages don't use "act_count". */
2419 	m->oflags = VPO_UNMANAGED;
2420 	return (m);
2421 }
2422 
2423 static int
2424 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2425 {
2426 	struct vm_domain *vmd;
2427 	struct vm_pgcache *pgcache;
2428 	int i;
2429 
2430 	pgcache = arg;
2431 	vmd = VM_DOMAIN(pgcache->domain);
2432 
2433 	/*
2434 	 * The page daemon should avoid creating extra memory pressure since its
2435 	 * main purpose is to replenish the store of free pages.
2436 	 */
2437 	if (vmd->vmd_severeset || curproc == pageproc ||
2438 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2439 		return (0);
2440 	domain = vmd->vmd_domain;
2441 	vm_domain_free_lock(vmd);
2442 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2443 	    (vm_page_t *)store);
2444 	vm_domain_free_unlock(vmd);
2445 	if (cnt != i)
2446 		vm_domain_freecnt_inc(vmd, cnt - i);
2447 
2448 	return (i);
2449 }
2450 
2451 static void
2452 vm_page_zone_release(void *arg, void **store, int cnt)
2453 {
2454 	struct vm_domain *vmd;
2455 	struct vm_pgcache *pgcache;
2456 	vm_page_t m;
2457 	int i;
2458 
2459 	pgcache = arg;
2460 	vmd = VM_DOMAIN(pgcache->domain);
2461 	vm_domain_free_lock(vmd);
2462 	for (i = 0; i < cnt; i++) {
2463 		m = (vm_page_t)store[i];
2464 		vm_phys_free_pages(m, 0);
2465 	}
2466 	vm_domain_free_unlock(vmd);
2467 	vm_domain_freecnt_inc(vmd, cnt);
2468 }
2469 
2470 #define	VPSC_ANY	0	/* No restrictions. */
2471 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2472 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2473 
2474 /*
2475  *	vm_page_scan_contig:
2476  *
2477  *	Scan vm_page_array[] between the specified entries "m_start" and
2478  *	"m_end" for a run of contiguous physical pages that satisfy the
2479  *	specified conditions, and return the lowest page in the run.  The
2480  *	specified "alignment" determines the alignment of the lowest physical
2481  *	page in the run.  If the specified "boundary" is non-zero, then the
2482  *	run of physical pages cannot span a physical address that is a
2483  *	multiple of "boundary".
2484  *
2485  *	"m_end" is never dereferenced, so it need not point to a vm_page
2486  *	structure within vm_page_array[].
2487  *
2488  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2489  *	span a hole (or discontiguity) in the physical address space.  Both
2490  *	"alignment" and "boundary" must be a power of two.
2491  */
2492 vm_page_t
2493 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2494     u_long alignment, vm_paddr_t boundary, int options)
2495 {
2496 	struct mtx *m_mtx;
2497 	vm_object_t object;
2498 	vm_paddr_t pa;
2499 	vm_page_t m, m_run;
2500 #if VM_NRESERVLEVEL > 0
2501 	int level;
2502 #endif
2503 	int m_inc, order, run_ext, run_len;
2504 
2505 	KASSERT(npages > 0, ("npages is 0"));
2506 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2507 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2508 	m_run = NULL;
2509 	run_len = 0;
2510 	m_mtx = NULL;
2511 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2512 		KASSERT((m->flags & PG_MARKER) == 0,
2513 		    ("page %p is PG_MARKER", m));
2514 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2515 		    ("fictitious page %p has invalid ref count", m));
2516 
2517 		/*
2518 		 * If the current page would be the start of a run, check its
2519 		 * physical address against the end, alignment, and boundary
2520 		 * conditions.  If it doesn't satisfy these conditions, either
2521 		 * terminate the scan or advance to the next page that
2522 		 * satisfies the failed condition.
2523 		 */
2524 		if (run_len == 0) {
2525 			KASSERT(m_run == NULL, ("m_run != NULL"));
2526 			if (m + npages > m_end)
2527 				break;
2528 			pa = VM_PAGE_TO_PHYS(m);
2529 			if ((pa & (alignment - 1)) != 0) {
2530 				m_inc = atop(roundup2(pa, alignment) - pa);
2531 				continue;
2532 			}
2533 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2534 			    boundary) != 0) {
2535 				m_inc = atop(roundup2(pa, boundary) - pa);
2536 				continue;
2537 			}
2538 		} else
2539 			KASSERT(m_run != NULL, ("m_run == NULL"));
2540 
2541 		vm_page_change_lock(m, &m_mtx);
2542 		m_inc = 1;
2543 retry:
2544 		if (vm_page_wired(m))
2545 			run_ext = 0;
2546 #if VM_NRESERVLEVEL > 0
2547 		else if ((level = vm_reserv_level(m)) >= 0 &&
2548 		    (options & VPSC_NORESERV) != 0) {
2549 			run_ext = 0;
2550 			/* Advance to the end of the reservation. */
2551 			pa = VM_PAGE_TO_PHYS(m);
2552 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2553 			    pa);
2554 		}
2555 #endif
2556 		else if ((object = m->object) != NULL) {
2557 			/*
2558 			 * The page is considered eligible for relocation if
2559 			 * and only if it could be laundered or reclaimed by
2560 			 * the page daemon.
2561 			 */
2562 			if (!VM_OBJECT_TRYRLOCK(object)) {
2563 				mtx_unlock(m_mtx);
2564 				VM_OBJECT_RLOCK(object);
2565 				mtx_lock(m_mtx);
2566 				if (m->object != object) {
2567 					/*
2568 					 * The page may have been freed.
2569 					 */
2570 					VM_OBJECT_RUNLOCK(object);
2571 					goto retry;
2572 				}
2573 			}
2574 			/* Don't care: PG_NODUMP, PG_ZERO. */
2575 			if (object->type != OBJT_DEFAULT &&
2576 			    object->type != OBJT_SWAP &&
2577 			    object->type != OBJT_VNODE) {
2578 				run_ext = 0;
2579 #if VM_NRESERVLEVEL > 0
2580 			} else if ((options & VPSC_NOSUPER) != 0 &&
2581 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2582 				run_ext = 0;
2583 				/* Advance to the end of the superpage. */
2584 				pa = VM_PAGE_TO_PHYS(m);
2585 				m_inc = atop(roundup2(pa + 1,
2586 				    vm_reserv_size(level)) - pa);
2587 #endif
2588 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2589 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2590 			    !vm_page_wired(m)) {
2591 				/*
2592 				 * The page is allocated but eligible for
2593 				 * relocation.  Extend the current run by one
2594 				 * page.
2595 				 */
2596 				KASSERT(pmap_page_get_memattr(m) ==
2597 				    VM_MEMATTR_DEFAULT,
2598 				    ("page %p has an unexpected memattr", m));
2599 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2600 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2601 				    ("page %p has unexpected oflags", m));
2602 				/* Don't care: PGA_NOSYNC. */
2603 				run_ext = 1;
2604 			} else
2605 				run_ext = 0;
2606 			VM_OBJECT_RUNLOCK(object);
2607 #if VM_NRESERVLEVEL > 0
2608 		} else if (level >= 0) {
2609 			/*
2610 			 * The page is reserved but not yet allocated.  In
2611 			 * other words, it is still free.  Extend the current
2612 			 * run by one page.
2613 			 */
2614 			run_ext = 1;
2615 #endif
2616 		} else if ((order = m->order) < VM_NFREEORDER) {
2617 			/*
2618 			 * The page is enqueued in the physical memory
2619 			 * allocator's free page queues.  Moreover, it is the
2620 			 * first page in a power-of-two-sized run of
2621 			 * contiguous free pages.  Add these pages to the end
2622 			 * of the current run, and jump ahead.
2623 			 */
2624 			run_ext = 1 << order;
2625 			m_inc = 1 << order;
2626 		} else {
2627 			/*
2628 			 * Skip the page for one of the following reasons: (1)
2629 			 * It is enqueued in the physical memory allocator's
2630 			 * free page queues.  However, it is not the first
2631 			 * page in a run of contiguous free pages.  (This case
2632 			 * rarely occurs because the scan is performed in
2633 			 * ascending order.) (2) It is not reserved, and it is
2634 			 * transitioning from free to allocated.  (Conversely,
2635 			 * the transition from allocated to free for managed
2636 			 * pages is blocked by the page lock.) (3) It is
2637 			 * allocated but not contained by an object and not
2638 			 * wired, e.g., allocated by Xen's balloon driver.
2639 			 */
2640 			run_ext = 0;
2641 		}
2642 
2643 		/*
2644 		 * Extend or reset the current run of pages.
2645 		 */
2646 		if (run_ext > 0) {
2647 			if (run_len == 0)
2648 				m_run = m;
2649 			run_len += run_ext;
2650 		} else {
2651 			if (run_len > 0) {
2652 				m_run = NULL;
2653 				run_len = 0;
2654 			}
2655 		}
2656 	}
2657 	if (m_mtx != NULL)
2658 		mtx_unlock(m_mtx);
2659 	if (run_len >= npages)
2660 		return (m_run);
2661 	return (NULL);
2662 }
2663 
2664 /*
2665  *	vm_page_reclaim_run:
2666  *
2667  *	Try to relocate each of the allocated virtual pages within the
2668  *	specified run of physical pages to a new physical address.  Free the
2669  *	physical pages underlying the relocated virtual pages.  A virtual page
2670  *	is relocatable if and only if it could be laundered or reclaimed by
2671  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2672  *	physical address above "high".
2673  *
2674  *	Returns 0 if every physical page within the run was already free or
2675  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2676  *	value indicating why the last attempt to relocate a virtual page was
2677  *	unsuccessful.
2678  *
2679  *	"req_class" must be an allocation class.
2680  */
2681 static int
2682 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2683     vm_paddr_t high)
2684 {
2685 	struct vm_domain *vmd;
2686 	struct mtx *m_mtx;
2687 	struct spglist free;
2688 	vm_object_t object;
2689 	vm_paddr_t pa;
2690 	vm_page_t m, m_end, m_new;
2691 	int error, order, req;
2692 
2693 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2694 	    ("req_class is not an allocation class"));
2695 	SLIST_INIT(&free);
2696 	error = 0;
2697 	m = m_run;
2698 	m_end = m_run + npages;
2699 	m_mtx = NULL;
2700 	for (; error == 0 && m < m_end; m++) {
2701 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2702 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2703 
2704 		/*
2705 		 * Avoid releasing and reacquiring the same page lock.
2706 		 */
2707 		vm_page_change_lock(m, &m_mtx);
2708 retry:
2709 		/*
2710 		 * Racily check for wirings.  Races are handled below.
2711 		 */
2712 		if (vm_page_wired(m))
2713 			error = EBUSY;
2714 		else if ((object = m->object) != NULL) {
2715 			/*
2716 			 * The page is relocated if and only if it could be
2717 			 * laundered or reclaimed by the page daemon.
2718 			 */
2719 			if (!VM_OBJECT_TRYWLOCK(object)) {
2720 				mtx_unlock(m_mtx);
2721 				VM_OBJECT_WLOCK(object);
2722 				mtx_lock(m_mtx);
2723 				if (m->object != object) {
2724 					/*
2725 					 * The page may have been freed.
2726 					 */
2727 					VM_OBJECT_WUNLOCK(object);
2728 					goto retry;
2729 				}
2730 			}
2731 			/* Don't care: PG_NODUMP, PG_ZERO. */
2732 			if (object->type != OBJT_DEFAULT &&
2733 			    object->type != OBJT_SWAP &&
2734 			    object->type != OBJT_VNODE)
2735 				error = EINVAL;
2736 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2737 				error = EINVAL;
2738 			else if (vm_page_queue(m) != PQ_NONE &&
2739 			    vm_page_tryxbusy(m) != 0) {
2740 				if (vm_page_wired(m)) {
2741 					vm_page_xunbusy(m);
2742 					error = EBUSY;
2743 					goto unlock;
2744 				}
2745 				KASSERT(pmap_page_get_memattr(m) ==
2746 				    VM_MEMATTR_DEFAULT,
2747 				    ("page %p has an unexpected memattr", m));
2748 				KASSERT(m->oflags == 0,
2749 				    ("page %p has unexpected oflags", m));
2750 				/* Don't care: PGA_NOSYNC. */
2751 				if (!vm_page_none_valid(m)) {
2752 					/*
2753 					 * First, try to allocate a new page
2754 					 * that is above "high".  Failing
2755 					 * that, try to allocate a new page
2756 					 * that is below "m_run".  Allocate
2757 					 * the new page between the end of
2758 					 * "m_run" and "high" only as a last
2759 					 * resort.
2760 					 */
2761 					req = req_class | VM_ALLOC_NOOBJ;
2762 					if ((m->flags & PG_NODUMP) != 0)
2763 						req |= VM_ALLOC_NODUMP;
2764 					if (trunc_page(high) !=
2765 					    ~(vm_paddr_t)PAGE_MASK) {
2766 						m_new = vm_page_alloc_contig(
2767 						    NULL, 0, req, 1,
2768 						    round_page(high),
2769 						    ~(vm_paddr_t)0,
2770 						    PAGE_SIZE, 0,
2771 						    VM_MEMATTR_DEFAULT);
2772 					} else
2773 						m_new = NULL;
2774 					if (m_new == NULL) {
2775 						pa = VM_PAGE_TO_PHYS(m_run);
2776 						m_new = vm_page_alloc_contig(
2777 						    NULL, 0, req, 1,
2778 						    0, pa - 1, PAGE_SIZE, 0,
2779 						    VM_MEMATTR_DEFAULT);
2780 					}
2781 					if (m_new == NULL) {
2782 						pa += ptoa(npages);
2783 						m_new = vm_page_alloc_contig(
2784 						    NULL, 0, req, 1,
2785 						    pa, high, PAGE_SIZE, 0,
2786 						    VM_MEMATTR_DEFAULT);
2787 					}
2788 					if (m_new == NULL) {
2789 						vm_page_xunbusy(m);
2790 						error = ENOMEM;
2791 						goto unlock;
2792 					}
2793 
2794 					/*
2795 					 * Unmap the page and check for new
2796 					 * wirings that may have been acquired
2797 					 * through a pmap lookup.
2798 					 */
2799 					if (object->ref_count != 0 &&
2800 					    !vm_page_try_remove_all(m)) {
2801 						vm_page_xunbusy(m);
2802 						vm_page_free(m_new);
2803 						error = EBUSY;
2804 						goto unlock;
2805 					}
2806 
2807 					/*
2808 					 * Replace "m" with the new page.  For
2809 					 * vm_page_replace(), "m" must be busy
2810 					 * and dequeued.  Finally, change "m"
2811 					 * as if vm_page_free() was called.
2812 					 */
2813 					m_new->a.flags = m->a.flags &
2814 					    ~PGA_QUEUE_STATE_MASK;
2815 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2816 					    ("page %p is managed", m_new));
2817 					m_new->oflags = 0;
2818 					pmap_copy_page(m, m_new);
2819 					m_new->valid = m->valid;
2820 					m_new->dirty = m->dirty;
2821 					m->flags &= ~PG_ZERO;
2822 					vm_page_dequeue(m);
2823 					if (vm_page_replace_hold(m_new, object,
2824 					    m->pindex, m) &&
2825 					    vm_page_free_prep(m))
2826 						SLIST_INSERT_HEAD(&free, m,
2827 						    plinks.s.ss);
2828 
2829 					/*
2830 					 * The new page must be deactivated
2831 					 * before the object is unlocked.
2832 					 */
2833 					vm_page_change_lock(m_new, &m_mtx);
2834 					vm_page_deactivate(m_new);
2835 				} else {
2836 					m->flags &= ~PG_ZERO;
2837 					vm_page_dequeue(m);
2838 					if (vm_page_free_prep(m))
2839 						SLIST_INSERT_HEAD(&free, m,
2840 						    plinks.s.ss);
2841 					KASSERT(m->dirty == 0,
2842 					    ("page %p is dirty", m));
2843 				}
2844 			} else
2845 				error = EBUSY;
2846 unlock:
2847 			VM_OBJECT_WUNLOCK(object);
2848 		} else {
2849 			MPASS(vm_phys_domain(m) == domain);
2850 			vmd = VM_DOMAIN(domain);
2851 			vm_domain_free_lock(vmd);
2852 			order = m->order;
2853 			if (order < VM_NFREEORDER) {
2854 				/*
2855 				 * The page is enqueued in the physical memory
2856 				 * allocator's free page queues.  Moreover, it
2857 				 * is the first page in a power-of-two-sized
2858 				 * run of contiguous free pages.  Jump ahead
2859 				 * to the last page within that run, and
2860 				 * continue from there.
2861 				 */
2862 				m += (1 << order) - 1;
2863 			}
2864 #if VM_NRESERVLEVEL > 0
2865 			else if (vm_reserv_is_page_free(m))
2866 				order = 0;
2867 #endif
2868 			vm_domain_free_unlock(vmd);
2869 			if (order == VM_NFREEORDER)
2870 				error = EINVAL;
2871 		}
2872 	}
2873 	if (m_mtx != NULL)
2874 		mtx_unlock(m_mtx);
2875 	if ((m = SLIST_FIRST(&free)) != NULL) {
2876 		int cnt;
2877 
2878 		vmd = VM_DOMAIN(domain);
2879 		cnt = 0;
2880 		vm_domain_free_lock(vmd);
2881 		do {
2882 			MPASS(vm_phys_domain(m) == domain);
2883 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2884 			vm_phys_free_pages(m, 0);
2885 			cnt++;
2886 		} while ((m = SLIST_FIRST(&free)) != NULL);
2887 		vm_domain_free_unlock(vmd);
2888 		vm_domain_freecnt_inc(vmd, cnt);
2889 	}
2890 	return (error);
2891 }
2892 
2893 #define	NRUNS	16
2894 
2895 CTASSERT(powerof2(NRUNS));
2896 
2897 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2898 
2899 #define	MIN_RECLAIM	8
2900 
2901 /*
2902  *	vm_page_reclaim_contig:
2903  *
2904  *	Reclaim allocated, contiguous physical memory satisfying the specified
2905  *	conditions by relocating the virtual pages using that physical memory.
2906  *	Returns true if reclamation is successful and false otherwise.  Since
2907  *	relocation requires the allocation of physical pages, reclamation may
2908  *	fail due to a shortage of free pages.  When reclamation fails, callers
2909  *	are expected to perform vm_wait() before retrying a failed allocation
2910  *	operation, e.g., vm_page_alloc_contig().
2911  *
2912  *	The caller must always specify an allocation class through "req".
2913  *
2914  *	allocation classes:
2915  *	VM_ALLOC_NORMAL		normal process request
2916  *	VM_ALLOC_SYSTEM		system *really* needs a page
2917  *	VM_ALLOC_INTERRUPT	interrupt time request
2918  *
2919  *	The optional allocation flags are ignored.
2920  *
2921  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2922  *	must be a power of two.
2923  */
2924 bool
2925 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2926     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2927 {
2928 	struct vm_domain *vmd;
2929 	vm_paddr_t curr_low;
2930 	vm_page_t m_run, m_runs[NRUNS];
2931 	u_long count, reclaimed;
2932 	int error, i, options, req_class;
2933 
2934 	KASSERT(npages > 0, ("npages is 0"));
2935 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2936 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2937 	req_class = req & VM_ALLOC_CLASS_MASK;
2938 
2939 	/*
2940 	 * The page daemon is allowed to dig deeper into the free page list.
2941 	 */
2942 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2943 		req_class = VM_ALLOC_SYSTEM;
2944 
2945 	/*
2946 	 * Return if the number of free pages cannot satisfy the requested
2947 	 * allocation.
2948 	 */
2949 	vmd = VM_DOMAIN(domain);
2950 	count = vmd->vmd_free_count;
2951 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2952 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2953 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2954 		return (false);
2955 
2956 	/*
2957 	 * Scan up to three times, relaxing the restrictions ("options") on
2958 	 * the reclamation of reservations and superpages each time.
2959 	 */
2960 	for (options = VPSC_NORESERV;;) {
2961 		/*
2962 		 * Find the highest runs that satisfy the given constraints
2963 		 * and restrictions, and record them in "m_runs".
2964 		 */
2965 		curr_low = low;
2966 		count = 0;
2967 		for (;;) {
2968 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2969 			    high, alignment, boundary, options);
2970 			if (m_run == NULL)
2971 				break;
2972 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2973 			m_runs[RUN_INDEX(count)] = m_run;
2974 			count++;
2975 		}
2976 
2977 		/*
2978 		 * Reclaim the highest runs in LIFO (descending) order until
2979 		 * the number of reclaimed pages, "reclaimed", is at least
2980 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2981 		 * reclamation is idempotent, and runs will (likely) recur
2982 		 * from one scan to the next as restrictions are relaxed.
2983 		 */
2984 		reclaimed = 0;
2985 		for (i = 0; count > 0 && i < NRUNS; i++) {
2986 			count--;
2987 			m_run = m_runs[RUN_INDEX(count)];
2988 			error = vm_page_reclaim_run(req_class, domain, npages,
2989 			    m_run, high);
2990 			if (error == 0) {
2991 				reclaimed += npages;
2992 				if (reclaimed >= MIN_RECLAIM)
2993 					return (true);
2994 			}
2995 		}
2996 
2997 		/*
2998 		 * Either relax the restrictions on the next scan or return if
2999 		 * the last scan had no restrictions.
3000 		 */
3001 		if (options == VPSC_NORESERV)
3002 			options = VPSC_NOSUPER;
3003 		else if (options == VPSC_NOSUPER)
3004 			options = VPSC_ANY;
3005 		else if (options == VPSC_ANY)
3006 			return (reclaimed != 0);
3007 	}
3008 }
3009 
3010 bool
3011 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3012     u_long alignment, vm_paddr_t boundary)
3013 {
3014 	struct vm_domainset_iter di;
3015 	int domain;
3016 	bool ret;
3017 
3018 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3019 	do {
3020 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3021 		    high, alignment, boundary);
3022 		if (ret)
3023 			break;
3024 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3025 
3026 	return (ret);
3027 }
3028 
3029 /*
3030  * Set the domain in the appropriate page level domainset.
3031  */
3032 void
3033 vm_domain_set(struct vm_domain *vmd)
3034 {
3035 
3036 	mtx_lock(&vm_domainset_lock);
3037 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3038 		vmd->vmd_minset = 1;
3039 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3040 	}
3041 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3042 		vmd->vmd_severeset = 1;
3043 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3044 	}
3045 	mtx_unlock(&vm_domainset_lock);
3046 }
3047 
3048 /*
3049  * Clear the domain from the appropriate page level domainset.
3050  */
3051 void
3052 vm_domain_clear(struct vm_domain *vmd)
3053 {
3054 
3055 	mtx_lock(&vm_domainset_lock);
3056 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3057 		vmd->vmd_minset = 0;
3058 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3059 		if (vm_min_waiters != 0) {
3060 			vm_min_waiters = 0;
3061 			wakeup(&vm_min_domains);
3062 		}
3063 	}
3064 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3065 		vmd->vmd_severeset = 0;
3066 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3067 		if (vm_severe_waiters != 0) {
3068 			vm_severe_waiters = 0;
3069 			wakeup(&vm_severe_domains);
3070 		}
3071 	}
3072 
3073 	/*
3074 	 * If pageout daemon needs pages, then tell it that there are
3075 	 * some free.
3076 	 */
3077 	if (vmd->vmd_pageout_pages_needed &&
3078 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3079 		wakeup(&vmd->vmd_pageout_pages_needed);
3080 		vmd->vmd_pageout_pages_needed = 0;
3081 	}
3082 
3083 	/* See comments in vm_wait_doms(). */
3084 	if (vm_pageproc_waiters) {
3085 		vm_pageproc_waiters = 0;
3086 		wakeup(&vm_pageproc_waiters);
3087 	}
3088 	mtx_unlock(&vm_domainset_lock);
3089 }
3090 
3091 /*
3092  * Wait for free pages to exceed the min threshold globally.
3093  */
3094 void
3095 vm_wait_min(void)
3096 {
3097 
3098 	mtx_lock(&vm_domainset_lock);
3099 	while (vm_page_count_min()) {
3100 		vm_min_waiters++;
3101 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3102 	}
3103 	mtx_unlock(&vm_domainset_lock);
3104 }
3105 
3106 /*
3107  * Wait for free pages to exceed the severe threshold globally.
3108  */
3109 void
3110 vm_wait_severe(void)
3111 {
3112 
3113 	mtx_lock(&vm_domainset_lock);
3114 	while (vm_page_count_severe()) {
3115 		vm_severe_waiters++;
3116 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3117 		    "vmwait", 0);
3118 	}
3119 	mtx_unlock(&vm_domainset_lock);
3120 }
3121 
3122 u_int
3123 vm_wait_count(void)
3124 {
3125 
3126 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3127 }
3128 
3129 void
3130 vm_wait_doms(const domainset_t *wdoms)
3131 {
3132 
3133 	/*
3134 	 * We use racey wakeup synchronization to avoid expensive global
3135 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3136 	 * To handle this, we only sleep for one tick in this instance.  It
3137 	 * is expected that most allocations for the pageproc will come from
3138 	 * kmem or vm_page_grab* which will use the more specific and
3139 	 * race-free vm_wait_domain().
3140 	 */
3141 	if (curproc == pageproc) {
3142 		mtx_lock(&vm_domainset_lock);
3143 		vm_pageproc_waiters++;
3144 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3145 		    "pageprocwait", 1);
3146 	} else {
3147 		/*
3148 		 * XXX Ideally we would wait only until the allocation could
3149 		 * be satisfied.  This condition can cause new allocators to
3150 		 * consume all freed pages while old allocators wait.
3151 		 */
3152 		mtx_lock(&vm_domainset_lock);
3153 		if (vm_page_count_min_set(wdoms)) {
3154 			vm_min_waiters++;
3155 			msleep(&vm_min_domains, &vm_domainset_lock,
3156 			    PVM | PDROP, "vmwait", 0);
3157 		} else
3158 			mtx_unlock(&vm_domainset_lock);
3159 	}
3160 }
3161 
3162 /*
3163  *	vm_wait_domain:
3164  *
3165  *	Sleep until free pages are available for allocation.
3166  *	- Called in various places after failed memory allocations.
3167  */
3168 void
3169 vm_wait_domain(int domain)
3170 {
3171 	struct vm_domain *vmd;
3172 	domainset_t wdom;
3173 
3174 	vmd = VM_DOMAIN(domain);
3175 	vm_domain_free_assert_unlocked(vmd);
3176 
3177 	if (curproc == pageproc) {
3178 		mtx_lock(&vm_domainset_lock);
3179 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3180 			vmd->vmd_pageout_pages_needed = 1;
3181 			msleep(&vmd->vmd_pageout_pages_needed,
3182 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3183 		} else
3184 			mtx_unlock(&vm_domainset_lock);
3185 	} else {
3186 		if (pageproc == NULL)
3187 			panic("vm_wait in early boot");
3188 		DOMAINSET_ZERO(&wdom);
3189 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3190 		vm_wait_doms(&wdom);
3191 	}
3192 }
3193 
3194 /*
3195  *	vm_wait:
3196  *
3197  *	Sleep until free pages are available for allocation in the
3198  *	affinity domains of the obj.  If obj is NULL, the domain set
3199  *	for the calling thread is used.
3200  *	Called in various places after failed memory allocations.
3201  */
3202 void
3203 vm_wait(vm_object_t obj)
3204 {
3205 	struct domainset *d;
3206 
3207 	d = NULL;
3208 
3209 	/*
3210 	 * Carefully fetch pointers only once: the struct domainset
3211 	 * itself is ummutable but the pointer might change.
3212 	 */
3213 	if (obj != NULL)
3214 		d = obj->domain.dr_policy;
3215 	if (d == NULL)
3216 		d = curthread->td_domain.dr_policy;
3217 
3218 	vm_wait_doms(&d->ds_mask);
3219 }
3220 
3221 /*
3222  *	vm_domain_alloc_fail:
3223  *
3224  *	Called when a page allocation function fails.  Informs the
3225  *	pagedaemon and performs the requested wait.  Requires the
3226  *	domain_free and object lock on entry.  Returns with the
3227  *	object lock held and free lock released.  Returns an error when
3228  *	retry is necessary.
3229  *
3230  */
3231 static int
3232 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3233 {
3234 
3235 	vm_domain_free_assert_unlocked(vmd);
3236 
3237 	atomic_add_int(&vmd->vmd_pageout_deficit,
3238 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3239 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3240 		if (object != NULL)
3241 			VM_OBJECT_WUNLOCK(object);
3242 		vm_wait_domain(vmd->vmd_domain);
3243 		if (object != NULL)
3244 			VM_OBJECT_WLOCK(object);
3245 		if (req & VM_ALLOC_WAITOK)
3246 			return (EAGAIN);
3247 	}
3248 
3249 	return (0);
3250 }
3251 
3252 /*
3253  *	vm_waitpfault:
3254  *
3255  *	Sleep until free pages are available for allocation.
3256  *	- Called only in vm_fault so that processes page faulting
3257  *	  can be easily tracked.
3258  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3259  *	  processes will be able to grab memory first.  Do not change
3260  *	  this balance without careful testing first.
3261  */
3262 void
3263 vm_waitpfault(struct domainset *dset, int timo)
3264 {
3265 
3266 	/*
3267 	 * XXX Ideally we would wait only until the allocation could
3268 	 * be satisfied.  This condition can cause new allocators to
3269 	 * consume all freed pages while old allocators wait.
3270 	 */
3271 	mtx_lock(&vm_domainset_lock);
3272 	if (vm_page_count_min_set(&dset->ds_mask)) {
3273 		vm_min_waiters++;
3274 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3275 		    "pfault", timo);
3276 	} else
3277 		mtx_unlock(&vm_domainset_lock);
3278 }
3279 
3280 static struct vm_pagequeue *
3281 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3282 {
3283 
3284 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3285 }
3286 
3287 #ifdef INVARIANTS
3288 static struct vm_pagequeue *
3289 vm_page_pagequeue(vm_page_t m)
3290 {
3291 
3292 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3293 }
3294 #endif
3295 
3296 static __always_inline bool
3297 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3298 {
3299 	vm_page_astate_t tmp;
3300 
3301 	tmp = *old;
3302 	do {
3303 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3304 			return (true);
3305 		counter_u64_add(pqstate_commit_retries, 1);
3306 	} while (old->_bits == tmp._bits);
3307 
3308 	return (false);
3309 }
3310 
3311 /*
3312  * Do the work of committing a queue state update that moves the page out of
3313  * its current queue.
3314  */
3315 static bool
3316 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3317     vm_page_astate_t *old, vm_page_astate_t new)
3318 {
3319 	vm_page_t next;
3320 
3321 	vm_pagequeue_assert_locked(pq);
3322 	KASSERT(vm_page_pagequeue(m) == pq,
3323 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3324 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3325 	    ("%s: invalid queue indices %d %d",
3326 	    __func__, old->queue, new.queue));
3327 
3328 	/*
3329 	 * Once the queue index of the page changes there is nothing
3330 	 * synchronizing with further updates to the page's physical
3331 	 * queue state.  Therefore we must speculatively remove the page
3332 	 * from the queue now and be prepared to roll back if the queue
3333 	 * state update fails.  If the page is not physically enqueued then
3334 	 * we just update its queue index.
3335 	 */
3336 	if ((old->flags & PGA_ENQUEUED) != 0) {
3337 		new.flags &= ~PGA_ENQUEUED;
3338 		next = TAILQ_NEXT(m, plinks.q);
3339 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3340 		vm_pagequeue_cnt_dec(pq);
3341 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3342 			if (next == NULL)
3343 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3344 			else
3345 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3346 			vm_pagequeue_cnt_inc(pq);
3347 			return (false);
3348 		} else {
3349 			return (true);
3350 		}
3351 	} else {
3352 		return (vm_page_pqstate_fcmpset(m, old, new));
3353 	}
3354 }
3355 
3356 static bool
3357 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3358     vm_page_astate_t new)
3359 {
3360 	struct vm_pagequeue *pq;
3361 	vm_page_astate_t as;
3362 	bool ret;
3363 
3364 	pq = _vm_page_pagequeue(m, old->queue);
3365 
3366 	/*
3367 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3368 	 * corresponding page queue lock is held.
3369 	 */
3370 	vm_pagequeue_lock(pq);
3371 	as = vm_page_astate_load(m);
3372 	if (__predict_false(as._bits != old->_bits)) {
3373 		*old = as;
3374 		ret = false;
3375 	} else {
3376 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3377 	}
3378 	vm_pagequeue_unlock(pq);
3379 	return (ret);
3380 }
3381 
3382 /*
3383  * Commit a queue state update that enqueues or requeues a page.
3384  */
3385 static bool
3386 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3387     vm_page_astate_t *old, vm_page_astate_t new)
3388 {
3389 	struct vm_domain *vmd;
3390 
3391 	vm_pagequeue_assert_locked(pq);
3392 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3393 	    ("%s: invalid queue indices %d %d",
3394 	    __func__, old->queue, new.queue));
3395 
3396 	new.flags |= PGA_ENQUEUED;
3397 	if (!vm_page_pqstate_fcmpset(m, old, new))
3398 		return (false);
3399 
3400 	if ((old->flags & PGA_ENQUEUED) != 0)
3401 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3402 	else
3403 		vm_pagequeue_cnt_inc(pq);
3404 
3405 	/*
3406 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3407 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3408 	 * applied, even if it was set first.
3409 	 */
3410 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3411 		vmd = vm_pagequeue_domain(m);
3412 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3413 		    ("%s: invalid page queue for page %p", __func__, m));
3414 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3415 	} else {
3416 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3417 	}
3418 	return (true);
3419 }
3420 
3421 /*
3422  * Commit a queue state update that encodes a request for a deferred queue
3423  * operation.
3424  */
3425 static bool
3426 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3427     vm_page_astate_t new)
3428 {
3429 
3430 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3431 	    ("%s: invalid state, queue %d flags %x",
3432 	    __func__, new.queue, new.flags));
3433 
3434 	if (old->_bits != new._bits &&
3435 	    !vm_page_pqstate_fcmpset(m, old, new))
3436 		return (false);
3437 	vm_page_pqbatch_submit(m, new.queue);
3438 	return (true);
3439 }
3440 
3441 /*
3442  * A generic queue state update function.  This handles more cases than the
3443  * specialized functions above.
3444  */
3445 bool
3446 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3447 {
3448 
3449 	if (old->_bits == new._bits)
3450 		return (true);
3451 
3452 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3453 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3454 			return (false);
3455 		if (new.queue != PQ_NONE)
3456 			vm_page_pqbatch_submit(m, new.queue);
3457 	} else {
3458 		if (!vm_page_pqstate_fcmpset(m, old, new))
3459 			return (false);
3460 		if (new.queue != PQ_NONE &&
3461 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3462 			vm_page_pqbatch_submit(m, new.queue);
3463 	}
3464 	return (true);
3465 }
3466 
3467 /*
3468  * Apply deferred queue state updates to a page.
3469  */
3470 static inline void
3471 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3472 {
3473 	vm_page_astate_t new, old;
3474 
3475 	CRITICAL_ASSERT(curthread);
3476 	vm_pagequeue_assert_locked(pq);
3477 	KASSERT(queue < PQ_COUNT,
3478 	    ("%s: invalid queue index %d", __func__, queue));
3479 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3480 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3481 
3482 	for (old = vm_page_astate_load(m);;) {
3483 		if (__predict_false(old.queue != queue ||
3484 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3485 			counter_u64_add(queue_nops, 1);
3486 			break;
3487 		}
3488 		KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3489 		    ("%s: page %p has unexpected queue state", __func__, m));
3490 
3491 		new = old;
3492 		if ((old.flags & PGA_DEQUEUE) != 0) {
3493 			new.flags &= ~PGA_QUEUE_OP_MASK;
3494 			new.queue = PQ_NONE;
3495 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3496 			    m, &old, new))) {
3497 				counter_u64_add(queue_ops, 1);
3498 				break;
3499 			}
3500 		} else {
3501 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3502 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3503 			    m, &old, new))) {
3504 				counter_u64_add(queue_ops, 1);
3505 				break;
3506 			}
3507 		}
3508 	}
3509 }
3510 
3511 static void
3512 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3513     uint8_t queue)
3514 {
3515 	int i;
3516 
3517 	for (i = 0; i < bq->bq_cnt; i++)
3518 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3519 	vm_batchqueue_init(bq);
3520 }
3521 
3522 /*
3523  *	vm_page_pqbatch_submit:		[ internal use only ]
3524  *
3525  *	Enqueue a page in the specified page queue's batched work queue.
3526  *	The caller must have encoded the requested operation in the page
3527  *	structure's a.flags field.
3528  */
3529 void
3530 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3531 {
3532 	struct vm_batchqueue *bq;
3533 	struct vm_pagequeue *pq;
3534 	int domain;
3535 
3536 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3537 	    ("page %p is unmanaged", m));
3538 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3539 
3540 	domain = vm_phys_domain(m);
3541 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3542 
3543 	critical_enter();
3544 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3545 	if (vm_batchqueue_insert(bq, m)) {
3546 		critical_exit();
3547 		return;
3548 	}
3549 	critical_exit();
3550 	vm_pagequeue_lock(pq);
3551 	critical_enter();
3552 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3553 	vm_pqbatch_process(pq, bq, queue);
3554 	vm_pqbatch_process_page(pq, m, queue);
3555 	vm_pagequeue_unlock(pq);
3556 	critical_exit();
3557 }
3558 
3559 /*
3560  *	vm_page_pqbatch_drain:		[ internal use only ]
3561  *
3562  *	Force all per-CPU page queue batch queues to be drained.  This is
3563  *	intended for use in severe memory shortages, to ensure that pages
3564  *	do not remain stuck in the batch queues.
3565  */
3566 void
3567 vm_page_pqbatch_drain(void)
3568 {
3569 	struct thread *td;
3570 	struct vm_domain *vmd;
3571 	struct vm_pagequeue *pq;
3572 	int cpu, domain, queue;
3573 
3574 	td = curthread;
3575 	CPU_FOREACH(cpu) {
3576 		thread_lock(td);
3577 		sched_bind(td, cpu);
3578 		thread_unlock(td);
3579 
3580 		for (domain = 0; domain < vm_ndomains; domain++) {
3581 			vmd = VM_DOMAIN(domain);
3582 			for (queue = 0; queue < PQ_COUNT; queue++) {
3583 				pq = &vmd->vmd_pagequeues[queue];
3584 				vm_pagequeue_lock(pq);
3585 				critical_enter();
3586 				vm_pqbatch_process(pq,
3587 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3588 				critical_exit();
3589 				vm_pagequeue_unlock(pq);
3590 			}
3591 		}
3592 	}
3593 	thread_lock(td);
3594 	sched_unbind(td);
3595 	thread_unlock(td);
3596 }
3597 
3598 /*
3599  *	vm_page_dequeue_deferred:	[ internal use only ]
3600  *
3601  *	Request removal of the given page from its current page
3602  *	queue.  Physical removal from the queue may be deferred
3603  *	indefinitely.
3604  *
3605  *	The page must be locked.
3606  */
3607 void
3608 vm_page_dequeue_deferred(vm_page_t m)
3609 {
3610 	vm_page_astate_t new, old;
3611 
3612 	old = vm_page_astate_load(m);
3613 	do {
3614 		if (old.queue == PQ_NONE) {
3615 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3616 			    ("%s: page %p has unexpected queue state",
3617 			    __func__, m));
3618 			break;
3619 		}
3620 		new = old;
3621 		new.flags |= PGA_DEQUEUE;
3622 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3623 }
3624 
3625 /*
3626  *	vm_page_dequeue:
3627  *
3628  *	Remove the page from whichever page queue it's in, if any, before
3629  *	returning.
3630  */
3631 void
3632 vm_page_dequeue(vm_page_t m)
3633 {
3634 	vm_page_astate_t new, old;
3635 
3636 	old = vm_page_astate_load(m);
3637 	do {
3638 		if (old.queue == PQ_NONE) {
3639 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3640 			    ("%s: page %p has unexpected queue state",
3641 			    __func__, m));
3642 			break;
3643 		}
3644 		new = old;
3645 		new.flags &= ~PGA_QUEUE_OP_MASK;
3646 		new.queue = PQ_NONE;
3647 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3648 
3649 }
3650 
3651 /*
3652  * Schedule the given page for insertion into the specified page queue.
3653  * Physical insertion of the page may be deferred indefinitely.
3654  */
3655 static void
3656 vm_page_enqueue(vm_page_t m, uint8_t queue)
3657 {
3658 
3659 	KASSERT(m->a.queue == PQ_NONE &&
3660 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3661 	    ("%s: page %p is already enqueued", __func__, m));
3662 	KASSERT(m->ref_count > 0,
3663 	    ("%s: page %p does not carry any references", __func__, m));
3664 
3665 	m->a.queue = queue;
3666 	if ((m->a.flags & PGA_REQUEUE) == 0)
3667 		vm_page_aflag_set(m, PGA_REQUEUE);
3668 	vm_page_pqbatch_submit(m, queue);
3669 }
3670 
3671 /*
3672  *	vm_page_free_prep:
3673  *
3674  *	Prepares the given page to be put on the free list,
3675  *	disassociating it from any VM object. The caller may return
3676  *	the page to the free list only if this function returns true.
3677  *
3678  *	The object must be locked.  The page must be locked if it is
3679  *	managed.
3680  */
3681 static bool
3682 vm_page_free_prep(vm_page_t m)
3683 {
3684 
3685 	/*
3686 	 * Synchronize with threads that have dropped a reference to this
3687 	 * page.
3688 	 */
3689 	atomic_thread_fence_acq();
3690 
3691 	if (vm_page_sbusied(m))
3692 		panic("vm_page_free_prep: freeing shared busy page %p", m);
3693 
3694 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3695 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3696 		uint64_t *p;
3697 		int i;
3698 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3699 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3700 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3701 			    m, i, (uintmax_t)*p));
3702 	}
3703 #endif
3704 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3705 		KASSERT(!pmap_page_is_mapped(m),
3706 		    ("vm_page_free_prep: freeing mapped page %p", m));
3707 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3708 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3709 	} else {
3710 		KASSERT(m->a.queue == PQ_NONE,
3711 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3712 	}
3713 	VM_CNT_INC(v_tfree);
3714 
3715 	if (m->object != NULL) {
3716 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3717 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3718 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3719 		    m));
3720 		vm_page_object_remove(m);
3721 
3722 		/*
3723 		 * The object reference can be released without an atomic
3724 		 * operation.
3725 		 */
3726 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3727 		    m->ref_count == VPRC_OBJREF,
3728 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3729 		    m, m->ref_count));
3730 		m->object = NULL;
3731 		m->ref_count -= VPRC_OBJREF;
3732 		vm_page_xunbusy(m);
3733 	}
3734 
3735 	if (vm_page_xbusied(m))
3736 		panic("vm_page_free_prep: freeing exclusive busy page %p", m);
3737 
3738 	/*
3739 	 * If fictitious remove object association and
3740 	 * return.
3741 	 */
3742 	if ((m->flags & PG_FICTITIOUS) != 0) {
3743 		KASSERT(m->ref_count == 1,
3744 		    ("fictitious page %p is referenced", m));
3745 		KASSERT(m->a.queue == PQ_NONE,
3746 		    ("fictitious page %p is queued", m));
3747 		return (false);
3748 	}
3749 
3750 	/*
3751 	 * Pages need not be dequeued before they are returned to the physical
3752 	 * memory allocator, but they must at least be marked for a deferred
3753 	 * dequeue.
3754 	 */
3755 	if ((m->oflags & VPO_UNMANAGED) == 0)
3756 		vm_page_dequeue_deferred(m);
3757 
3758 	m->valid = 0;
3759 	vm_page_undirty(m);
3760 
3761 	if (m->ref_count != 0)
3762 		panic("vm_page_free_prep: page %p has references", m);
3763 
3764 	/*
3765 	 * Restore the default memory attribute to the page.
3766 	 */
3767 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3768 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3769 
3770 #if VM_NRESERVLEVEL > 0
3771 	/*
3772 	 * Determine whether the page belongs to a reservation.  If the page was
3773 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3774 	 * as an optimization, we avoid the check in that case.
3775 	 */
3776 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3777 		return (false);
3778 #endif
3779 
3780 	return (true);
3781 }
3782 
3783 /*
3784  *	vm_page_free_toq:
3785  *
3786  *	Returns the given page to the free list, disassociating it
3787  *	from any VM object.
3788  *
3789  *	The object must be locked.  The page must be locked if it is
3790  *	managed.
3791  */
3792 static void
3793 vm_page_free_toq(vm_page_t m)
3794 {
3795 	struct vm_domain *vmd;
3796 	uma_zone_t zone;
3797 
3798 	if (!vm_page_free_prep(m))
3799 		return;
3800 
3801 	vmd = vm_pagequeue_domain(m);
3802 	zone = vmd->vmd_pgcache[m->pool].zone;
3803 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3804 		uma_zfree(zone, m);
3805 		return;
3806 	}
3807 	vm_domain_free_lock(vmd);
3808 	vm_phys_free_pages(m, 0);
3809 	vm_domain_free_unlock(vmd);
3810 	vm_domain_freecnt_inc(vmd, 1);
3811 }
3812 
3813 /*
3814  *	vm_page_free_pages_toq:
3815  *
3816  *	Returns a list of pages to the free list, disassociating it
3817  *	from any VM object.  In other words, this is equivalent to
3818  *	calling vm_page_free_toq() for each page of a list of VM objects.
3819  *
3820  *	The objects must be locked.  The pages must be locked if it is
3821  *	managed.
3822  */
3823 void
3824 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3825 {
3826 	vm_page_t m;
3827 	int count;
3828 
3829 	if (SLIST_EMPTY(free))
3830 		return;
3831 
3832 	count = 0;
3833 	while ((m = SLIST_FIRST(free)) != NULL) {
3834 		count++;
3835 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3836 		vm_page_free_toq(m);
3837 	}
3838 
3839 	if (update_wire_count)
3840 		vm_wire_sub(count);
3841 }
3842 
3843 /*
3844  * Mark this page as wired down, preventing reclamation by the page daemon
3845  * or when the containing object is destroyed.
3846  */
3847 void
3848 vm_page_wire(vm_page_t m)
3849 {
3850 	u_int old;
3851 
3852 	KASSERT(m->object != NULL,
3853 	    ("vm_page_wire: page %p does not belong to an object", m));
3854 	if (!vm_page_busied(m) && !vm_object_busied(m->object))
3855 		VM_OBJECT_ASSERT_LOCKED(m->object);
3856 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3857 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3858 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3859 
3860 	old = atomic_fetchadd_int(&m->ref_count, 1);
3861 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3862 	    ("vm_page_wire: counter overflow for page %p", m));
3863 	if (VPRC_WIRE_COUNT(old) == 0) {
3864 		if ((m->oflags & VPO_UNMANAGED) == 0)
3865 			vm_page_aflag_set(m, PGA_DEQUEUE);
3866 		vm_wire_add(1);
3867 	}
3868 }
3869 
3870 /*
3871  * Attempt to wire a mapped page following a pmap lookup of that page.
3872  * This may fail if a thread is concurrently tearing down mappings of the page.
3873  * The transient failure is acceptable because it translates to the
3874  * failure of the caller pmap_extract_and_hold(), which should be then
3875  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3876  */
3877 bool
3878 vm_page_wire_mapped(vm_page_t m)
3879 {
3880 	u_int old;
3881 
3882 	old = m->ref_count;
3883 	do {
3884 		KASSERT(old > 0,
3885 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3886 		if ((old & VPRC_BLOCKED) != 0)
3887 			return (false);
3888 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3889 
3890 	if (VPRC_WIRE_COUNT(old) == 0) {
3891 		if ((m->oflags & VPO_UNMANAGED) == 0)
3892 			vm_page_aflag_set(m, PGA_DEQUEUE);
3893 		vm_wire_add(1);
3894 	}
3895 	return (true);
3896 }
3897 
3898 /*
3899  * Release a wiring reference to a managed page.  If the page still belongs to
3900  * an object, update its position in the page queues to reflect the reference.
3901  * If the wiring was the last reference to the page, free the page.
3902  */
3903 static void
3904 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3905 {
3906 	u_int old;
3907 
3908 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3909 	    ("%s: page %p is unmanaged", __func__, m));
3910 
3911 	/*
3912 	 * Update LRU state before releasing the wiring reference.
3913 	 * Use a release store when updating the reference count to
3914 	 * synchronize with vm_page_free_prep().
3915 	 */
3916 	old = m->ref_count;
3917 	do {
3918 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3919 		    ("vm_page_unwire: wire count underflow for page %p", m));
3920 
3921 		if (old > VPRC_OBJREF + 1) {
3922 			/*
3923 			 * The page has at least one other wiring reference.  An
3924 			 * earlier iteration of this loop may have called
3925 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3926 			 * re-set it if necessary.
3927 			 */
3928 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3929 				vm_page_aflag_set(m, PGA_DEQUEUE);
3930 		} else if (old == VPRC_OBJREF + 1) {
3931 			/*
3932 			 * This is the last wiring.  Clear PGA_DEQUEUE and
3933 			 * update the page's queue state to reflect the
3934 			 * reference.  If the page does not belong to an object
3935 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3936 			 * clear leftover queue state.
3937 			 */
3938 			vm_page_release_toq(m, nqueue, false);
3939 		} else if (old == 1) {
3940 			vm_page_aflag_clear(m, PGA_DEQUEUE);
3941 		}
3942 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3943 
3944 	if (VPRC_WIRE_COUNT(old) == 1) {
3945 		vm_wire_sub(1);
3946 		if (old == 1)
3947 			vm_page_free(m);
3948 	}
3949 }
3950 
3951 /*
3952  * Release one wiring of the specified page, potentially allowing it to be
3953  * paged out.
3954  *
3955  * Only managed pages belonging to an object can be paged out.  If the number
3956  * of wirings transitions to zero and the page is eligible for page out, then
3957  * the page is added to the specified paging queue.  If the released wiring
3958  * represented the last reference to the page, the page is freed.
3959  *
3960  * A managed page must be locked.
3961  */
3962 void
3963 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3964 {
3965 
3966 	KASSERT(nqueue < PQ_COUNT,
3967 	    ("vm_page_unwire: invalid queue %u request for page %p",
3968 	    nqueue, m));
3969 
3970 	if ((m->oflags & VPO_UNMANAGED) != 0) {
3971 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
3972 			vm_page_free(m);
3973 		return;
3974 	}
3975 	vm_page_unwire_managed(m, nqueue, false);
3976 }
3977 
3978 /*
3979  * Unwire a page without (re-)inserting it into a page queue.  It is up
3980  * to the caller to enqueue, requeue, or free the page as appropriate.
3981  * In most cases involving managed pages, vm_page_unwire() should be used
3982  * instead.
3983  */
3984 bool
3985 vm_page_unwire_noq(vm_page_t m)
3986 {
3987 	u_int old;
3988 
3989 	old = vm_page_drop(m, 1);
3990 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
3991 	    ("vm_page_unref: counter underflow for page %p", m));
3992 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3993 	    ("vm_page_unref: missing ref on fictitious page %p", m));
3994 
3995 	if (VPRC_WIRE_COUNT(old) > 1)
3996 		return (false);
3997 	if ((m->oflags & VPO_UNMANAGED) == 0)
3998 		vm_page_aflag_clear(m, PGA_DEQUEUE);
3999 	vm_wire_sub(1);
4000 	return (true);
4001 }
4002 
4003 /*
4004  * Ensure that the page ends up in the specified page queue.  If the page is
4005  * active or being moved to the active queue, ensure that its act_count is
4006  * at least ACT_INIT but do not otherwise mess with it.
4007  *
4008  * A managed page must be locked.
4009  */
4010 static __always_inline void
4011 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4012 {
4013 	vm_page_astate_t old, new;
4014 
4015 	KASSERT(m->ref_count > 0,
4016 	    ("%s: page %p does not carry any references", __func__, m));
4017 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4018 	    ("%s: invalid flags %x", __func__, nflag));
4019 
4020 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4021 		return;
4022 
4023 	old = vm_page_astate_load(m);
4024 	do {
4025 		if ((old.flags & PGA_DEQUEUE) != 0)
4026 			break;
4027 		new = old;
4028 		new.flags &= ~PGA_QUEUE_OP_MASK;
4029 		if (nqueue == PQ_ACTIVE)
4030 			new.act_count = max(old.act_count, ACT_INIT);
4031 		if (old.queue == nqueue) {
4032 			if (nqueue != PQ_ACTIVE)
4033 				new.flags |= nflag;
4034 		} else {
4035 			new.flags |= nflag;
4036 			new.queue = nqueue;
4037 		}
4038 	} while (!vm_page_pqstate_commit(m, &old, new));
4039 }
4040 
4041 /*
4042  * Put the specified page on the active list (if appropriate).
4043  */
4044 void
4045 vm_page_activate(vm_page_t m)
4046 {
4047 
4048 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4049 }
4050 
4051 /*
4052  * Move the specified page to the tail of the inactive queue, or requeue
4053  * the page if it is already in the inactive queue.
4054  */
4055 void
4056 vm_page_deactivate(vm_page_t m)
4057 {
4058 
4059 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4060 }
4061 
4062 void
4063 vm_page_deactivate_noreuse(vm_page_t m)
4064 {
4065 
4066 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4067 }
4068 
4069 /*
4070  * Put a page in the laundry, or requeue it if it is already there.
4071  */
4072 void
4073 vm_page_launder(vm_page_t m)
4074 {
4075 
4076 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4077 }
4078 
4079 /*
4080  * Put a page in the PQ_UNSWAPPABLE holding queue.
4081  */
4082 void
4083 vm_page_unswappable(vm_page_t m)
4084 {
4085 
4086 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4087 	    ("page %p already unswappable", m));
4088 
4089 	vm_page_dequeue(m);
4090 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4091 }
4092 
4093 /*
4094  * Release a page back to the page queues in preparation for unwiring.
4095  */
4096 static void
4097 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4098 {
4099 	vm_page_astate_t old, new;
4100 	uint16_t nflag;
4101 
4102 	/*
4103 	 * Use a check of the valid bits to determine whether we should
4104 	 * accelerate reclamation of the page.  The object lock might not be
4105 	 * held here, in which case the check is racy.  At worst we will either
4106 	 * accelerate reclamation of a valid page and violate LRU, or
4107 	 * unnecessarily defer reclamation of an invalid page.
4108 	 *
4109 	 * If we were asked to not cache the page, place it near the head of the
4110 	 * inactive queue so that is reclaimed sooner.
4111 	 */
4112 	if (noreuse || m->valid == 0) {
4113 		nqueue = PQ_INACTIVE;
4114 		nflag = PGA_REQUEUE_HEAD;
4115 	} else {
4116 		nflag = PGA_REQUEUE;
4117 	}
4118 
4119 	old = vm_page_astate_load(m);
4120 	do {
4121 		new = old;
4122 
4123 		/*
4124 		 * If the page is already in the active queue and we are not
4125 		 * trying to accelerate reclamation, simply mark it as
4126 		 * referenced and avoid any queue operations.
4127 		 */
4128 		new.flags &= ~PGA_QUEUE_OP_MASK;
4129 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4130 			new.flags |= PGA_REFERENCED;
4131 		else {
4132 			new.flags |= nflag;
4133 			new.queue = nqueue;
4134 		}
4135 	} while (!vm_page_pqstate_commit(m, &old, new));
4136 }
4137 
4138 /*
4139  * Unwire a page and either attempt to free it or re-add it to the page queues.
4140  */
4141 void
4142 vm_page_release(vm_page_t m, int flags)
4143 {
4144 	vm_object_t object;
4145 
4146 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4147 	    ("vm_page_release: page %p is unmanaged", m));
4148 
4149 	if ((flags & VPR_TRYFREE) != 0) {
4150 		for (;;) {
4151 			object = (vm_object_t)atomic_load_ptr(&m->object);
4152 			if (object == NULL)
4153 				break;
4154 			/* Depends on type-stability. */
4155 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4156 				break;
4157 			if (object == m->object) {
4158 				vm_page_release_locked(m, flags);
4159 				VM_OBJECT_WUNLOCK(object);
4160 				return;
4161 			}
4162 			VM_OBJECT_WUNLOCK(object);
4163 		}
4164 	}
4165 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4166 }
4167 
4168 /* See vm_page_release(). */
4169 void
4170 vm_page_release_locked(vm_page_t m, int flags)
4171 {
4172 
4173 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4174 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4175 	    ("vm_page_release_locked: page %p is unmanaged", m));
4176 
4177 	if (vm_page_unwire_noq(m)) {
4178 		if ((flags & VPR_TRYFREE) != 0 &&
4179 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4180 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4181 			vm_page_free(m);
4182 		} else {
4183 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4184 		}
4185 	}
4186 }
4187 
4188 static bool
4189 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4190 {
4191 	u_int old;
4192 
4193 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4194 	    ("vm_page_try_blocked_op: page %p has no object", m));
4195 	KASSERT(vm_page_busied(m),
4196 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4197 	VM_OBJECT_ASSERT_LOCKED(m->object);
4198 
4199 	old = m->ref_count;
4200 	do {
4201 		KASSERT(old != 0,
4202 		    ("vm_page_try_blocked_op: page %p has no references", m));
4203 		if (VPRC_WIRE_COUNT(old) != 0)
4204 			return (false);
4205 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4206 
4207 	(op)(m);
4208 
4209 	/*
4210 	 * If the object is read-locked, new wirings may be created via an
4211 	 * object lookup.
4212 	 */
4213 	old = vm_page_drop(m, VPRC_BLOCKED);
4214 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4215 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4216 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4217 	    old, m));
4218 	return (true);
4219 }
4220 
4221 /*
4222  * Atomically check for wirings and remove all mappings of the page.
4223  */
4224 bool
4225 vm_page_try_remove_all(vm_page_t m)
4226 {
4227 
4228 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4229 }
4230 
4231 /*
4232  * Atomically check for wirings and remove all writeable mappings of the page.
4233  */
4234 bool
4235 vm_page_try_remove_write(vm_page_t m)
4236 {
4237 
4238 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4239 }
4240 
4241 /*
4242  * vm_page_advise
4243  *
4244  * 	Apply the specified advice to the given page.
4245  *
4246  *	The object and page must be locked.
4247  */
4248 void
4249 vm_page_advise(vm_page_t m, int advice)
4250 {
4251 
4252 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4253 	if (advice == MADV_FREE)
4254 		/*
4255 		 * Mark the page clean.  This will allow the page to be freed
4256 		 * without first paging it out.  MADV_FREE pages are often
4257 		 * quickly reused by malloc(3), so we do not do anything that
4258 		 * would result in a page fault on a later access.
4259 		 */
4260 		vm_page_undirty(m);
4261 	else if (advice != MADV_DONTNEED) {
4262 		if (advice == MADV_WILLNEED)
4263 			vm_page_activate(m);
4264 		return;
4265 	}
4266 
4267 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4268 		vm_page_dirty(m);
4269 
4270 	/*
4271 	 * Clear any references to the page.  Otherwise, the page daemon will
4272 	 * immediately reactivate the page.
4273 	 */
4274 	vm_page_aflag_clear(m, PGA_REFERENCED);
4275 
4276 	/*
4277 	 * Place clean pages near the head of the inactive queue rather than
4278 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4279 	 * the page will be reused quickly.  Dirty pages not already in the
4280 	 * laundry are moved there.
4281 	 */
4282 	if (m->dirty == 0)
4283 		vm_page_deactivate_noreuse(m);
4284 	else if (!vm_page_in_laundry(m))
4285 		vm_page_launder(m);
4286 }
4287 
4288 static inline int
4289 vm_page_grab_pflags(int allocflags)
4290 {
4291 	int pflags;
4292 
4293 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4294 	    (allocflags & VM_ALLOC_WIRED) != 0,
4295 	    ("vm_page_grab_pflags: the pages must be busied or wired"));
4296 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4297 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4298 	    ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4299 	    "mismatch"));
4300 	pflags = allocflags &
4301 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4302 	    VM_ALLOC_NOBUSY);
4303 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4304 		pflags |= VM_ALLOC_WAITFAIL;
4305 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4306 		pflags |= VM_ALLOC_SBUSY;
4307 
4308 	return (pflags);
4309 }
4310 
4311 /*
4312  * Grab a page, waiting until we are waken up due to the page
4313  * changing state.  We keep on waiting, if the page continues
4314  * to be in the object.  If the page doesn't exist, first allocate it
4315  * and then conditionally zero it.
4316  *
4317  * This routine may sleep.
4318  *
4319  * The object must be locked on entry.  The lock will, however, be released
4320  * and reacquired if the routine sleeps.
4321  */
4322 vm_page_t
4323 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4324 {
4325 	vm_page_t m;
4326 	int pflags;
4327 
4328 	VM_OBJECT_ASSERT_WLOCKED(object);
4329 	pflags = vm_page_grab_pflags(allocflags);
4330 retrylookup:
4331 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4332 		if (!vm_page_acquire_flags(m, allocflags)) {
4333 			if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4334 			    allocflags))
4335 				goto retrylookup;
4336 			return (NULL);
4337 		}
4338 		goto out;
4339 	}
4340 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4341 		return (NULL);
4342 	m = vm_page_alloc(object, pindex, pflags);
4343 	if (m == NULL) {
4344 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4345 			return (NULL);
4346 		goto retrylookup;
4347 	}
4348 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4349 		pmap_zero_page(m);
4350 
4351 out:
4352 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4353 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4354 			vm_page_sunbusy(m);
4355 		else
4356 			vm_page_xunbusy(m);
4357 	}
4358 	return (m);
4359 }
4360 
4361 /*
4362  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4363  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4364  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4365  * in simultaneously.  Additional pages will be left on a paging queue but
4366  * will neither be wired nor busy regardless of allocflags.
4367  */
4368 int
4369 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4370 {
4371 	vm_page_t m;
4372 	vm_page_t ma[VM_INITIAL_PAGEIN];
4373 	bool sleep, xbusy;
4374 	int after, i, pflags, rv;
4375 
4376 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4377 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4378 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4379 	KASSERT((allocflags &
4380 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4381 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4382 	VM_OBJECT_ASSERT_WLOCKED(object);
4383 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4384 	pflags |= VM_ALLOC_WAITFAIL;
4385 
4386 retrylookup:
4387 	xbusy = false;
4388 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4389 		/*
4390 		 * If the page is fully valid it can only become invalid
4391 		 * with the object lock held.  If it is not valid it can
4392 		 * become valid with the busy lock held.  Therefore, we
4393 		 * may unnecessarily lock the exclusive busy here if we
4394 		 * race with I/O completion not using the object lock.
4395 		 * However, we will not end up with an invalid page and a
4396 		 * shared lock.
4397 		 */
4398 		if (!vm_page_all_valid(m) ||
4399 		    (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4400 			sleep = !vm_page_tryxbusy(m);
4401 			xbusy = true;
4402 		} else
4403 			sleep = !vm_page_trysbusy(m);
4404 		if (sleep) {
4405 			(void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4406 			    allocflags);
4407 			goto retrylookup;
4408 		}
4409 		if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4410 		   !vm_page_all_valid(m)) {
4411 			if (xbusy)
4412 				vm_page_xunbusy(m);
4413 			else
4414 				vm_page_sunbusy(m);
4415 			*mp = NULL;
4416 			return (VM_PAGER_FAIL);
4417 		}
4418 		if ((allocflags & VM_ALLOC_WIRED) != 0)
4419 			vm_page_wire(m);
4420 		if (vm_page_all_valid(m))
4421 			goto out;
4422 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4423 		*mp = NULL;
4424 		return (VM_PAGER_FAIL);
4425 	} else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4426 		xbusy = true;
4427 	} else {
4428 		goto retrylookup;
4429 	}
4430 
4431 	vm_page_assert_xbusied(m);
4432 	MPASS(xbusy);
4433 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4434 		after = MIN(after, VM_INITIAL_PAGEIN);
4435 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4436 		after = MAX(after, 1);
4437 		ma[0] = m;
4438 		for (i = 1; i < after; i++) {
4439 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4440 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4441 					break;
4442 			} else {
4443 				ma[i] = vm_page_alloc(object, m->pindex + i,
4444 				    VM_ALLOC_NORMAL);
4445 				if (ma[i] == NULL)
4446 					break;
4447 			}
4448 		}
4449 		after = i;
4450 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4451 		/* Pager may have replaced a page. */
4452 		m = ma[0];
4453 		if (rv != VM_PAGER_OK) {
4454 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4455 				vm_page_unwire_noq(m);
4456 			for (i = 0; i < after; i++) {
4457 				if (!vm_page_wired(ma[i]))
4458 					vm_page_free(ma[i]);
4459 				else
4460 					vm_page_xunbusy(ma[i]);
4461 			}
4462 			*mp = NULL;
4463 			return (rv);
4464 		}
4465 		for (i = 1; i < after; i++)
4466 			vm_page_readahead_finish(ma[i]);
4467 		MPASS(vm_page_all_valid(m));
4468 	} else {
4469 		vm_page_zero_invalid(m, TRUE);
4470 	}
4471 out:
4472 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4473 		if (xbusy)
4474 			vm_page_xunbusy(m);
4475 		else
4476 			vm_page_sunbusy(m);
4477 	}
4478 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4479 		vm_page_busy_downgrade(m);
4480 	*mp = m;
4481 	return (VM_PAGER_OK);
4482 }
4483 
4484 /*
4485  * Return the specified range of pages from the given object.  For each
4486  * page offset within the range, if a page already exists within the object
4487  * at that offset and it is busy, then wait for it to change state.  If,
4488  * instead, the page doesn't exist, then allocate it.
4489  *
4490  * The caller must always specify an allocation class.
4491  *
4492  * allocation classes:
4493  *	VM_ALLOC_NORMAL		normal process request
4494  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4495  *
4496  * The caller must always specify that the pages are to be busied and/or
4497  * wired.
4498  *
4499  * optional allocation flags:
4500  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4501  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4502  *	VM_ALLOC_NOWAIT		do not sleep
4503  *	VM_ALLOC_SBUSY		set page to sbusy state
4504  *	VM_ALLOC_WIRED		wire the pages
4505  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4506  *
4507  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4508  * may return a partial prefix of the requested range.
4509  */
4510 int
4511 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4512     vm_page_t *ma, int count)
4513 {
4514 	vm_page_t m, mpred;
4515 	int pflags;
4516 	int i;
4517 
4518 	VM_OBJECT_ASSERT_WLOCKED(object);
4519 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4520 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4521 
4522 	pflags = vm_page_grab_pflags(allocflags);
4523 	if (count == 0)
4524 		return (0);
4525 
4526 	i = 0;
4527 retrylookup:
4528 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4529 	if (m == NULL || m->pindex != pindex + i) {
4530 		mpred = m;
4531 		m = NULL;
4532 	} else
4533 		mpred = TAILQ_PREV(m, pglist, listq);
4534 	for (; i < count; i++) {
4535 		if (m != NULL) {
4536 			if (!vm_page_acquire_flags(m, allocflags)) {
4537 				if (vm_page_busy_sleep_flags(object, m,
4538 				    "grbmaw", allocflags))
4539 					goto retrylookup;
4540 				break;
4541 			}
4542 		} else {
4543 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4544 				break;
4545 			m = vm_page_alloc_after(object, pindex + i,
4546 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4547 			if (m == NULL) {
4548 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4549 					break;
4550 				goto retrylookup;
4551 			}
4552 		}
4553 		if (vm_page_none_valid(m) &&
4554 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4555 			if ((m->flags & PG_ZERO) == 0)
4556 				pmap_zero_page(m);
4557 			vm_page_valid(m);
4558 		}
4559 		if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4560 			if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4561 				vm_page_sunbusy(m);
4562 			else
4563 				vm_page_xunbusy(m);
4564 		}
4565 		ma[i] = mpred = m;
4566 		m = vm_page_next(m);
4567 	}
4568 	return (i);
4569 }
4570 
4571 /*
4572  * Mapping function for valid or dirty bits in a page.
4573  *
4574  * Inputs are required to range within a page.
4575  */
4576 vm_page_bits_t
4577 vm_page_bits(int base, int size)
4578 {
4579 	int first_bit;
4580 	int last_bit;
4581 
4582 	KASSERT(
4583 	    base + size <= PAGE_SIZE,
4584 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4585 	);
4586 
4587 	if (size == 0)		/* handle degenerate case */
4588 		return (0);
4589 
4590 	first_bit = base >> DEV_BSHIFT;
4591 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4592 
4593 	return (((vm_page_bits_t)2 << last_bit) -
4594 	    ((vm_page_bits_t)1 << first_bit));
4595 }
4596 
4597 void
4598 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4599 {
4600 
4601 #if PAGE_SIZE == 32768
4602 	atomic_set_64((uint64_t *)bits, set);
4603 #elif PAGE_SIZE == 16384
4604 	atomic_set_32((uint32_t *)bits, set);
4605 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4606 	atomic_set_16((uint16_t *)bits, set);
4607 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4608 	atomic_set_8((uint8_t *)bits, set);
4609 #else		/* PAGE_SIZE <= 8192 */
4610 	uintptr_t addr;
4611 	int shift;
4612 
4613 	addr = (uintptr_t)bits;
4614 	/*
4615 	 * Use a trick to perform a 32-bit atomic on the
4616 	 * containing aligned word, to not depend on the existence
4617 	 * of atomic_{set, clear}_{8, 16}.
4618 	 */
4619 	shift = addr & (sizeof(uint32_t) - 1);
4620 #if BYTE_ORDER == BIG_ENDIAN
4621 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4622 #else
4623 	shift *= NBBY;
4624 #endif
4625 	addr &= ~(sizeof(uint32_t) - 1);
4626 	atomic_set_32((uint32_t *)addr, set << shift);
4627 #endif		/* PAGE_SIZE */
4628 }
4629 
4630 static inline void
4631 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4632 {
4633 
4634 #if PAGE_SIZE == 32768
4635 	atomic_clear_64((uint64_t *)bits, clear);
4636 #elif PAGE_SIZE == 16384
4637 	atomic_clear_32((uint32_t *)bits, clear);
4638 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4639 	atomic_clear_16((uint16_t *)bits, clear);
4640 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4641 	atomic_clear_8((uint8_t *)bits, clear);
4642 #else		/* PAGE_SIZE <= 8192 */
4643 	uintptr_t addr;
4644 	int shift;
4645 
4646 	addr = (uintptr_t)bits;
4647 	/*
4648 	 * Use a trick to perform a 32-bit atomic on the
4649 	 * containing aligned word, to not depend on the existence
4650 	 * of atomic_{set, clear}_{8, 16}.
4651 	 */
4652 	shift = addr & (sizeof(uint32_t) - 1);
4653 #if BYTE_ORDER == BIG_ENDIAN
4654 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4655 #else
4656 	shift *= NBBY;
4657 #endif
4658 	addr &= ~(sizeof(uint32_t) - 1);
4659 	atomic_clear_32((uint32_t *)addr, clear << shift);
4660 #endif		/* PAGE_SIZE */
4661 }
4662 
4663 static inline vm_page_bits_t
4664 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4665 {
4666 #if PAGE_SIZE == 32768
4667 	uint64_t old;
4668 
4669 	old = *bits;
4670 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4671 	return (old);
4672 #elif PAGE_SIZE == 16384
4673 	uint32_t old;
4674 
4675 	old = *bits;
4676 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4677 	return (old);
4678 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4679 	uint16_t old;
4680 
4681 	old = *bits;
4682 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4683 	return (old);
4684 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4685 	uint8_t old;
4686 
4687 	old = *bits;
4688 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4689 	return (old);
4690 #else		/* PAGE_SIZE <= 4096*/
4691 	uintptr_t addr;
4692 	uint32_t old, new, mask;
4693 	int shift;
4694 
4695 	addr = (uintptr_t)bits;
4696 	/*
4697 	 * Use a trick to perform a 32-bit atomic on the
4698 	 * containing aligned word, to not depend on the existence
4699 	 * of atomic_{set, swap, clear}_{8, 16}.
4700 	 */
4701 	shift = addr & (sizeof(uint32_t) - 1);
4702 #if BYTE_ORDER == BIG_ENDIAN
4703 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4704 #else
4705 	shift *= NBBY;
4706 #endif
4707 	addr &= ~(sizeof(uint32_t) - 1);
4708 	mask = VM_PAGE_BITS_ALL << shift;
4709 
4710 	old = *bits;
4711 	do {
4712 		new = old & ~mask;
4713 		new |= newbits << shift;
4714 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4715 	return (old >> shift);
4716 #endif		/* PAGE_SIZE */
4717 }
4718 
4719 /*
4720  *	vm_page_set_valid_range:
4721  *
4722  *	Sets portions of a page valid.  The arguments are expected
4723  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4724  *	of any partial chunks touched by the range.  The invalid portion of
4725  *	such chunks will be zeroed.
4726  *
4727  *	(base + size) must be less then or equal to PAGE_SIZE.
4728  */
4729 void
4730 vm_page_set_valid_range(vm_page_t m, int base, int size)
4731 {
4732 	int endoff, frag;
4733 	vm_page_bits_t pagebits;
4734 
4735 	vm_page_assert_busied(m);
4736 	if (size == 0)	/* handle degenerate case */
4737 		return;
4738 
4739 	/*
4740 	 * If the base is not DEV_BSIZE aligned and the valid
4741 	 * bit is clear, we have to zero out a portion of the
4742 	 * first block.
4743 	 */
4744 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4745 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4746 		pmap_zero_page_area(m, frag, base - frag);
4747 
4748 	/*
4749 	 * If the ending offset is not DEV_BSIZE aligned and the
4750 	 * valid bit is clear, we have to zero out a portion of
4751 	 * the last block.
4752 	 */
4753 	endoff = base + size;
4754 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4755 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4756 		pmap_zero_page_area(m, endoff,
4757 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4758 
4759 	/*
4760 	 * Assert that no previously invalid block that is now being validated
4761 	 * is already dirty.
4762 	 */
4763 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4764 	    ("vm_page_set_valid_range: page %p is dirty", m));
4765 
4766 	/*
4767 	 * Set valid bits inclusive of any overlap.
4768 	 */
4769 	pagebits = vm_page_bits(base, size);
4770 	if (vm_page_xbusied(m))
4771 		m->valid |= pagebits;
4772 	else
4773 		vm_page_bits_set(m, &m->valid, pagebits);
4774 }
4775 
4776 /*
4777  * Set the page dirty bits and free the invalid swap space if
4778  * present.  Returns the previous dirty bits.
4779  */
4780 vm_page_bits_t
4781 vm_page_set_dirty(vm_page_t m)
4782 {
4783 	vm_page_bits_t old;
4784 
4785 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
4786 
4787 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4788 		old = m->dirty;
4789 		m->dirty = VM_PAGE_BITS_ALL;
4790 	} else
4791 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4792 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4793 		vm_pager_page_unswapped(m);
4794 
4795 	return (old);
4796 }
4797 
4798 /*
4799  * Clear the given bits from the specified page's dirty field.
4800  */
4801 static __inline void
4802 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4803 {
4804 
4805 	vm_page_assert_busied(m);
4806 
4807 	/*
4808 	 * If the page is xbusied and not write mapped we are the
4809 	 * only thread that can modify dirty bits.  Otherwise, The pmap
4810 	 * layer can call vm_page_dirty() without holding a distinguished
4811 	 * lock.  The combination of page busy and atomic operations
4812 	 * suffice to guarantee consistency of the page dirty field.
4813 	 */
4814 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4815 		m->dirty &= ~pagebits;
4816 	else
4817 		vm_page_bits_clear(m, &m->dirty, pagebits);
4818 }
4819 
4820 /*
4821  *	vm_page_set_validclean:
4822  *
4823  *	Sets portions of a page valid and clean.  The arguments are expected
4824  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4825  *	of any partial chunks touched by the range.  The invalid portion of
4826  *	such chunks will be zero'd.
4827  *
4828  *	(base + size) must be less then or equal to PAGE_SIZE.
4829  */
4830 void
4831 vm_page_set_validclean(vm_page_t m, int base, int size)
4832 {
4833 	vm_page_bits_t oldvalid, pagebits;
4834 	int endoff, frag;
4835 
4836 	vm_page_assert_busied(m);
4837 	if (size == 0)	/* handle degenerate case */
4838 		return;
4839 
4840 	/*
4841 	 * If the base is not DEV_BSIZE aligned and the valid
4842 	 * bit is clear, we have to zero out a portion of the
4843 	 * first block.
4844 	 */
4845 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4846 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4847 		pmap_zero_page_area(m, frag, base - frag);
4848 
4849 	/*
4850 	 * If the ending offset is not DEV_BSIZE aligned and the
4851 	 * valid bit is clear, we have to zero out a portion of
4852 	 * the last block.
4853 	 */
4854 	endoff = base + size;
4855 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4856 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4857 		pmap_zero_page_area(m, endoff,
4858 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4859 
4860 	/*
4861 	 * Set valid, clear dirty bits.  If validating the entire
4862 	 * page we can safely clear the pmap modify bit.  We also
4863 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
4864 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4865 	 * be set again.
4866 	 *
4867 	 * We set valid bits inclusive of any overlap, but we can only
4868 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4869 	 * the range.
4870 	 */
4871 	oldvalid = m->valid;
4872 	pagebits = vm_page_bits(base, size);
4873 	if (vm_page_xbusied(m))
4874 		m->valid |= pagebits;
4875 	else
4876 		vm_page_bits_set(m, &m->valid, pagebits);
4877 #if 0	/* NOT YET */
4878 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4879 		frag = DEV_BSIZE - frag;
4880 		base += frag;
4881 		size -= frag;
4882 		if (size < 0)
4883 			size = 0;
4884 	}
4885 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4886 #endif
4887 	if (base == 0 && size == PAGE_SIZE) {
4888 		/*
4889 		 * The page can only be modified within the pmap if it is
4890 		 * mapped, and it can only be mapped if it was previously
4891 		 * fully valid.
4892 		 */
4893 		if (oldvalid == VM_PAGE_BITS_ALL)
4894 			/*
4895 			 * Perform the pmap_clear_modify() first.  Otherwise,
4896 			 * a concurrent pmap operation, such as
4897 			 * pmap_protect(), could clear a modification in the
4898 			 * pmap and set the dirty field on the page before
4899 			 * pmap_clear_modify() had begun and after the dirty
4900 			 * field was cleared here.
4901 			 */
4902 			pmap_clear_modify(m);
4903 		m->dirty = 0;
4904 		vm_page_aflag_clear(m, PGA_NOSYNC);
4905 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4906 		m->dirty &= ~pagebits;
4907 	else
4908 		vm_page_clear_dirty_mask(m, pagebits);
4909 }
4910 
4911 void
4912 vm_page_clear_dirty(vm_page_t m, int base, int size)
4913 {
4914 
4915 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4916 }
4917 
4918 /*
4919  *	vm_page_set_invalid:
4920  *
4921  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4922  *	valid and dirty bits for the effected areas are cleared.
4923  */
4924 void
4925 vm_page_set_invalid(vm_page_t m, int base, int size)
4926 {
4927 	vm_page_bits_t bits;
4928 	vm_object_t object;
4929 
4930 	/*
4931 	 * The object lock is required so that pages can't be mapped
4932 	 * read-only while we're in the process of invalidating them.
4933 	 */
4934 	object = m->object;
4935 	VM_OBJECT_ASSERT_WLOCKED(object);
4936 	vm_page_assert_busied(m);
4937 
4938 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4939 	    size >= object->un_pager.vnp.vnp_size)
4940 		bits = VM_PAGE_BITS_ALL;
4941 	else
4942 		bits = vm_page_bits(base, size);
4943 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4944 		pmap_remove_all(m);
4945 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4946 	    !pmap_page_is_mapped(m),
4947 	    ("vm_page_set_invalid: page %p is mapped", m));
4948 	if (vm_page_xbusied(m)) {
4949 		m->valid &= ~bits;
4950 		m->dirty &= ~bits;
4951 	} else {
4952 		vm_page_bits_clear(m, &m->valid, bits);
4953 		vm_page_bits_clear(m, &m->dirty, bits);
4954 	}
4955 }
4956 
4957 /*
4958  *	vm_page_invalid:
4959  *
4960  *	Invalidates the entire page.  The page must be busy, unmapped, and
4961  *	the enclosing object must be locked.  The object locks protects
4962  *	against concurrent read-only pmap enter which is done without
4963  *	busy.
4964  */
4965 void
4966 vm_page_invalid(vm_page_t m)
4967 {
4968 
4969 	vm_page_assert_busied(m);
4970 	VM_OBJECT_ASSERT_LOCKED(m->object);
4971 	MPASS(!pmap_page_is_mapped(m));
4972 
4973 	if (vm_page_xbusied(m))
4974 		m->valid = 0;
4975 	else
4976 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4977 }
4978 
4979 /*
4980  * vm_page_zero_invalid()
4981  *
4982  *	The kernel assumes that the invalid portions of a page contain
4983  *	garbage, but such pages can be mapped into memory by user code.
4984  *	When this occurs, we must zero out the non-valid portions of the
4985  *	page so user code sees what it expects.
4986  *
4987  *	Pages are most often semi-valid when the end of a file is mapped
4988  *	into memory and the file's size is not page aligned.
4989  */
4990 void
4991 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4992 {
4993 	int b;
4994 	int i;
4995 
4996 	/*
4997 	 * Scan the valid bits looking for invalid sections that
4998 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4999 	 * valid bit may be set ) have already been zeroed by
5000 	 * vm_page_set_validclean().
5001 	 */
5002 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5003 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5004 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5005 			if (i > b) {
5006 				pmap_zero_page_area(m,
5007 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5008 			}
5009 			b = i + 1;
5010 		}
5011 	}
5012 
5013 	/*
5014 	 * setvalid is TRUE when we can safely set the zero'd areas
5015 	 * as being valid.  We can do this if there are no cache consistancy
5016 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5017 	 */
5018 	if (setvalid)
5019 		vm_page_valid(m);
5020 }
5021 
5022 /*
5023  *	vm_page_is_valid:
5024  *
5025  *	Is (partial) page valid?  Note that the case where size == 0
5026  *	will return FALSE in the degenerate case where the page is
5027  *	entirely invalid, and TRUE otherwise.
5028  *
5029  *	Some callers envoke this routine without the busy lock held and
5030  *	handle races via higher level locks.  Typical callers should
5031  *	hold a busy lock to prevent invalidation.
5032  */
5033 int
5034 vm_page_is_valid(vm_page_t m, int base, int size)
5035 {
5036 	vm_page_bits_t bits;
5037 
5038 	bits = vm_page_bits(base, size);
5039 	return (m->valid != 0 && (m->valid & bits) == bits);
5040 }
5041 
5042 /*
5043  * Returns true if all of the specified predicates are true for the entire
5044  * (super)page and false otherwise.
5045  */
5046 bool
5047 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5048 {
5049 	vm_object_t object;
5050 	int i, npages;
5051 
5052 	object = m->object;
5053 	if (skip_m != NULL && skip_m->object != object)
5054 		return (false);
5055 	VM_OBJECT_ASSERT_LOCKED(object);
5056 	npages = atop(pagesizes[m->psind]);
5057 
5058 	/*
5059 	 * The physically contiguous pages that make up a superpage, i.e., a
5060 	 * page with a page size index ("psind") greater than zero, will
5061 	 * occupy adjacent entries in vm_page_array[].
5062 	 */
5063 	for (i = 0; i < npages; i++) {
5064 		/* Always test object consistency, including "skip_m". */
5065 		if (m[i].object != object)
5066 			return (false);
5067 		if (&m[i] == skip_m)
5068 			continue;
5069 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5070 			return (false);
5071 		if ((flags & PS_ALL_DIRTY) != 0) {
5072 			/*
5073 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5074 			 * might stop this case from spuriously returning
5075 			 * "false".  However, that would require a write lock
5076 			 * on the object containing "m[i]".
5077 			 */
5078 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5079 				return (false);
5080 		}
5081 		if ((flags & PS_ALL_VALID) != 0 &&
5082 		    m[i].valid != VM_PAGE_BITS_ALL)
5083 			return (false);
5084 	}
5085 	return (true);
5086 }
5087 
5088 /*
5089  * Set the page's dirty bits if the page is modified.
5090  */
5091 void
5092 vm_page_test_dirty(vm_page_t m)
5093 {
5094 
5095 	vm_page_assert_busied(m);
5096 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5097 		vm_page_dirty(m);
5098 }
5099 
5100 void
5101 vm_page_valid(vm_page_t m)
5102 {
5103 
5104 	vm_page_assert_busied(m);
5105 	if (vm_page_xbusied(m))
5106 		m->valid = VM_PAGE_BITS_ALL;
5107 	else
5108 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5109 }
5110 
5111 void
5112 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5113 {
5114 
5115 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5116 }
5117 
5118 void
5119 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5120 {
5121 
5122 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5123 }
5124 
5125 int
5126 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5127 {
5128 
5129 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5130 }
5131 
5132 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5133 void
5134 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5135 {
5136 
5137 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5138 }
5139 
5140 void
5141 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5142 {
5143 
5144 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5145 }
5146 #endif
5147 
5148 #ifdef INVARIANTS
5149 void
5150 vm_page_object_busy_assert(vm_page_t m)
5151 {
5152 
5153 	/*
5154 	 * Certain of the page's fields may only be modified by the
5155 	 * holder of a page or object busy.
5156 	 */
5157 	if (m->object != NULL && !vm_page_busied(m))
5158 		VM_OBJECT_ASSERT_BUSY(m->object);
5159 }
5160 
5161 void
5162 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5163 {
5164 
5165 	if ((bits & PGA_WRITEABLE) == 0)
5166 		return;
5167 
5168 	/*
5169 	 * The PGA_WRITEABLE flag can only be set if the page is
5170 	 * managed, is exclusively busied or the object is locked.
5171 	 * Currently, this flag is only set by pmap_enter().
5172 	 */
5173 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5174 	    ("PGA_WRITEABLE on unmanaged page"));
5175 	if (!vm_page_xbusied(m))
5176 		VM_OBJECT_ASSERT_BUSY(m->object);
5177 }
5178 #endif
5179 
5180 #include "opt_ddb.h"
5181 #ifdef DDB
5182 #include <sys/kernel.h>
5183 
5184 #include <ddb/ddb.h>
5185 
5186 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5187 {
5188 
5189 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5190 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5191 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5192 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5193 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5194 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5195 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5196 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5197 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5198 }
5199 
5200 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5201 {
5202 	int dom;
5203 
5204 	db_printf("pq_free %d\n", vm_free_count());
5205 	for (dom = 0; dom < vm_ndomains; dom++) {
5206 		db_printf(
5207     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5208 		    dom,
5209 		    vm_dom[dom].vmd_page_count,
5210 		    vm_dom[dom].vmd_free_count,
5211 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5212 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5213 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5214 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5215 	}
5216 }
5217 
5218 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5219 {
5220 	vm_page_t m;
5221 	boolean_t phys, virt;
5222 
5223 	if (!have_addr) {
5224 		db_printf("show pginfo addr\n");
5225 		return;
5226 	}
5227 
5228 	phys = strchr(modif, 'p') != NULL;
5229 	virt = strchr(modif, 'v') != NULL;
5230 	if (virt)
5231 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5232 	else if (phys)
5233 		m = PHYS_TO_VM_PAGE(addr);
5234 	else
5235 		m = (vm_page_t)addr;
5236 	db_printf(
5237     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5238     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5239 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5240 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5241 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5242 }
5243 #endif /* DDB */
5244