xref: /freebsd/sys/vm/vm_page.c (revision ac84ce05c121b1f6ff47f0f3302ac26020231d08)
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 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Resident memory management module.
65  */
66 
67 #include <sys/cdefs.h>
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/counter.h>
73 #include <sys/domainset.h>
74 #include <sys/kernel.h>
75 #include <sys/limits.h>
76 #include <sys/linker.h>
77 #include <sys/lock.h>
78 #include <sys/malloc.h>
79 #include <sys/mman.h>
80 #include <sys/msgbuf.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/rwlock.h>
84 #include <sys/sleepqueue.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/smp.h>
88 #include <sys/sysctl.h>
89 #include <sys/vmmeter.h>
90 #include <sys/vnode.h>
91 
92 #include <vm/vm.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_param.h>
95 #include <vm/vm_domainset.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_dumpset.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 struct vm_domain vm_dom[MAXMEMDOM];
114 
115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
116 
117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
118 
119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
120 /* The following fields are protected by the domainset lock. */
121 domainset_t __exclusive_cache_line vm_min_domains;
122 domainset_t __exclusive_cache_line vm_severe_domains;
123 static int vm_min_waiters;
124 static int vm_severe_waiters;
125 static int vm_pageproc_waiters;
126 
127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
128     "VM page statistics");
129 
130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
132     CTLFLAG_RD, &pqstate_commit_retries,
133     "Number of failed per-page atomic queue state updates");
134 
135 static COUNTER_U64_DEFINE_EARLY(queue_ops);
136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
137     CTLFLAG_RD, &queue_ops,
138     "Number of batched queue operations");
139 
140 static COUNTER_U64_DEFINE_EARLY(queue_nops);
141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
142     CTLFLAG_RD, &queue_nops,
143     "Number of batched queue operations with no effects");
144 
145 /*
146  * bogus page -- for I/O to/from partially complete buffers,
147  * or for paging into sparsely invalid regions.
148  */
149 vm_page_t bogus_page;
150 
151 vm_page_t vm_page_array;
152 long vm_page_array_size;
153 long first_page;
154 
155 struct bitset *vm_page_dump;
156 long vm_page_dump_pages;
157 
158 static TAILQ_HEAD(, vm_page) blacklist_head;
159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
161     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
162 
163 static uma_zone_t fakepg_zone;
164 
165 static vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
166     int req, vm_page_t mpred);
167 static void vm_page_alloc_check(vm_page_t m);
168 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req);
169 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
170     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
171 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
172 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
173 static bool vm_page_free_prep(vm_page_t m);
174 static void vm_page_free_toq(vm_page_t m);
175 static void vm_page_init(void *dummy);
176 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
177     vm_pindex_t pindex, vm_page_t mpred);
178 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
179     vm_page_t mpred);
180 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
181     const uint16_t nflag);
182 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
183     vm_page_t m_run, vm_paddr_t high);
184 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
185 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
186     int req);
187 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
188     int flags);
189 static void vm_page_zone_release(void *arg, void **store, int cnt);
190 
191 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
192 
193 static void
194 vm_page_init(void *dummy)
195 {
196 
197 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
198 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
199 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE);
200 }
201 
202 static int pgcache_zone_max_pcpu;
203 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
204     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
205     "Per-CPU page cache size");
206 
207 /*
208  * The cache page zone is initialized later since we need to be able to allocate
209  * pages before UMA is fully initialized.
210  */
211 static void
212 vm_page_init_cache_zones(void *dummy __unused)
213 {
214 	struct vm_domain *vmd;
215 	struct vm_pgcache *pgcache;
216 	int cache, domain, maxcache, pool;
217 
218 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
219 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
220 	for (domain = 0; domain < vm_ndomains; domain++) {
221 		vmd = VM_DOMAIN(domain);
222 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
223 			pgcache = &vmd->vmd_pgcache[pool];
224 			pgcache->domain = domain;
225 			pgcache->pool = pool;
226 			pgcache->zone = uma_zcache_create("vm pgcache",
227 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
228 			    vm_page_zone_import, vm_page_zone_release, pgcache,
229 			    UMA_ZONE_VM);
230 
231 			/*
232 			 * Limit each pool's zone to 0.1% of the pages in the
233 			 * domain.
234 			 */
235 			cache = maxcache != 0 ? maxcache :
236 			    vmd->vmd_page_count / 1000;
237 			uma_zone_set_maxcache(pgcache->zone, cache);
238 		}
239 	}
240 }
241 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
242 
243 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
244 #if PAGE_SIZE == 32768
245 #ifdef CTASSERT
246 CTASSERT(sizeof(u_long) >= 8);
247 #endif
248 #endif
249 
250 /*
251  *	vm_set_page_size:
252  *
253  *	Sets the page size, perhaps based upon the memory
254  *	size.  Must be called before any use of page-size
255  *	dependent functions.
256  */
257 void
258 vm_set_page_size(void)
259 {
260 	if (vm_cnt.v_page_size == 0)
261 		vm_cnt.v_page_size = PAGE_SIZE;
262 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
263 		panic("vm_set_page_size: page size not a power of two");
264 }
265 
266 /*
267  *	vm_page_blacklist_next:
268  *
269  *	Find the next entry in the provided string of blacklist
270  *	addresses.  Entries are separated by space, comma, or newline.
271  *	If an invalid integer is encountered then the rest of the
272  *	string is skipped.  Updates the list pointer to the next
273  *	character, or NULL if the string is exhausted or invalid.
274  */
275 static vm_paddr_t
276 vm_page_blacklist_next(char **list, char *end)
277 {
278 	vm_paddr_t bad;
279 	char *cp, *pos;
280 
281 	if (list == NULL || *list == NULL)
282 		return (0);
283 	if (**list =='\0') {
284 		*list = NULL;
285 		return (0);
286 	}
287 
288 	/*
289 	 * If there's no end pointer then the buffer is coming from
290 	 * the kenv and we know it's null-terminated.
291 	 */
292 	if (end == NULL)
293 		end = *list + strlen(*list);
294 
295 	/* Ensure that strtoq() won't walk off the end */
296 	if (*end != '\0') {
297 		if (*end == '\n' || *end == ' ' || *end  == ',')
298 			*end = '\0';
299 		else {
300 			printf("Blacklist not terminated, skipping\n");
301 			*list = NULL;
302 			return (0);
303 		}
304 	}
305 
306 	for (pos = *list; *pos != '\0'; pos = cp) {
307 		bad = strtoq(pos, &cp, 0);
308 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
309 			if (bad == 0) {
310 				if (++cp < end)
311 					continue;
312 				else
313 					break;
314 			}
315 		} else
316 			break;
317 		if (*cp == '\0' || ++cp >= end)
318 			*list = NULL;
319 		else
320 			*list = cp;
321 		return (trunc_page(bad));
322 	}
323 	printf("Garbage in RAM blacklist, skipping\n");
324 	*list = NULL;
325 	return (0);
326 }
327 
328 bool
329 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
330 {
331 	struct vm_domain *vmd;
332 	vm_page_t m;
333 	bool found;
334 
335 	m = vm_phys_paddr_to_vm_page(pa);
336 	if (m == NULL)
337 		return (true); /* page does not exist, no failure */
338 
339 	vmd = VM_DOMAIN(vm_phys_domain(pa));
340 	vm_domain_free_lock(vmd);
341 	found = vm_phys_unfree_page(pa);
342 	vm_domain_free_unlock(vmd);
343 	if (found) {
344 		vm_domain_freecnt_inc(vmd, -1);
345 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
346 		if (verbose)
347 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
348 	}
349 	return (found);
350 }
351 
352 /*
353  *	vm_page_blacklist_check:
354  *
355  *	Iterate through the provided string of blacklist addresses, pulling
356  *	each entry out of the physical allocator free list and putting it
357  *	onto a list for reporting via the vm.page_blacklist sysctl.
358  */
359 static void
360 vm_page_blacklist_check(char *list, char *end)
361 {
362 	vm_paddr_t pa;
363 	char *next;
364 
365 	next = list;
366 	while (next != NULL) {
367 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
368 			continue;
369 		vm_page_blacklist_add(pa, bootverbose);
370 	}
371 }
372 
373 /*
374  *	vm_page_blacklist_load:
375  *
376  *	Search for a special module named "ram_blacklist".  It'll be a
377  *	plain text file provided by the user via the loader directive
378  *	of the same name.
379  */
380 static void
381 vm_page_blacklist_load(char **list, char **end)
382 {
383 	void *mod;
384 	u_char *ptr;
385 	u_int len;
386 
387 	mod = NULL;
388 	ptr = NULL;
389 
390 	mod = preload_search_by_type("ram_blacklist");
391 	if (mod != NULL) {
392 		ptr = preload_fetch_addr(mod);
393 		len = preload_fetch_size(mod);
394         }
395 	*list = ptr;
396 	if (ptr != NULL)
397 		*end = ptr + len;
398 	else
399 		*end = NULL;
400 	return;
401 }
402 
403 static int
404 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
405 {
406 	vm_page_t m;
407 	struct sbuf sbuf;
408 	int error, first;
409 
410 	first = 1;
411 	error = sysctl_wire_old_buffer(req, 0);
412 	if (error != 0)
413 		return (error);
414 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
415 	TAILQ_FOREACH(m, &blacklist_head, listq) {
416 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
417 		    (uintmax_t)m->phys_addr);
418 		first = 0;
419 	}
420 	error = sbuf_finish(&sbuf);
421 	sbuf_delete(&sbuf);
422 	return (error);
423 }
424 
425 /*
426  * Initialize a dummy page for use in scans of the specified paging queue.
427  * In principle, this function only needs to set the flag PG_MARKER.
428  * Nonetheless, it write busies the page as a safety precaution.
429  */
430 void
431 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
432 {
433 
434 	bzero(marker, sizeof(*marker));
435 	marker->flags = PG_MARKER;
436 	marker->a.flags = aflags;
437 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
438 	marker->a.queue = queue;
439 }
440 
441 static void
442 vm_page_domain_init(int domain)
443 {
444 	struct vm_domain *vmd;
445 	struct vm_pagequeue *pq;
446 	int i;
447 
448 	vmd = VM_DOMAIN(domain);
449 	bzero(vmd, sizeof(*vmd));
450 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
451 	    "vm inactive pagequeue";
452 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
453 	    "vm active pagequeue";
454 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
455 	    "vm laundry pagequeue";
456 	*__DECONST(const char **,
457 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
458 	    "vm unswappable pagequeue";
459 	vmd->vmd_domain = domain;
460 	vmd->vmd_page_count = 0;
461 	vmd->vmd_free_count = 0;
462 	vmd->vmd_segs = 0;
463 	vmd->vmd_oom = FALSE;
464 	for (i = 0; i < PQ_COUNT; i++) {
465 		pq = &vmd->vmd_pagequeues[i];
466 		TAILQ_INIT(&pq->pq_pl);
467 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
468 		    MTX_DEF | MTX_DUPOK);
469 		pq->pq_pdpages = 0;
470 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
471 	}
472 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
473 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
474 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
475 
476 	/*
477 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
478 	 * insertions.
479 	 */
480 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
481 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
482 	    &vmd->vmd_inacthead, plinks.q);
483 
484 	/*
485 	 * The clock pages are used to implement active queue scanning without
486 	 * requeues.  Scans start at clock[0], which is advanced after the scan
487 	 * ends.  When the two clock hands meet, they are reset and scanning
488 	 * resumes from the head of the queue.
489 	 */
490 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
491 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
492 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
493 	    &vmd->vmd_clock[0], plinks.q);
494 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
495 	    &vmd->vmd_clock[1], plinks.q);
496 }
497 
498 /*
499  * Initialize a physical page in preparation for adding it to the free
500  * lists.
501  */
502 void
503 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool)
504 {
505 	m->object = NULL;
506 	m->ref_count = 0;
507 	m->busy_lock = VPB_FREED;
508 	m->flags = m->a.flags = 0;
509 	m->phys_addr = pa;
510 	m->a.queue = PQ_NONE;
511 	m->psind = 0;
512 	m->segind = segind;
513 	m->order = VM_NFREEORDER;
514 	m->pool = pool;
515 	m->valid = m->dirty = 0;
516 	pmap_page_init(m);
517 }
518 
519 #ifndef PMAP_HAS_PAGE_ARRAY
520 static vm_paddr_t
521 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
522 {
523 	vm_paddr_t new_end;
524 
525 	/*
526 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
527 	 * However, because this page is allocated from KVM, out-of-bounds
528 	 * accesses using the direct map will not be trapped.
529 	 */
530 	*vaddr += PAGE_SIZE;
531 
532 	/*
533 	 * Allocate physical memory for the page structures, and map it.
534 	 */
535 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
536 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
537 	    VM_PROT_READ | VM_PROT_WRITE);
538 	vm_page_array_size = page_range;
539 
540 	return (new_end);
541 }
542 #endif
543 
544 /*
545  *	vm_page_startup:
546  *
547  *	Initializes the resident memory module.  Allocates physical memory for
548  *	bootstrapping UMA and some data structures that are used to manage
549  *	physical pages.  Initializes these structures, and populates the free
550  *	page queues.
551  */
552 vm_offset_t
553 vm_page_startup(vm_offset_t vaddr)
554 {
555 	struct vm_phys_seg *seg;
556 	struct vm_domain *vmd;
557 	vm_page_t m;
558 	char *list, *listend;
559 	vm_paddr_t end, high_avail, low_avail, new_end, size;
560 	vm_paddr_t page_range __unused;
561 	vm_paddr_t last_pa, pa, startp, endp;
562 	u_long pagecount;
563 #if MINIDUMP_PAGE_TRACKING
564 	u_long vm_page_dump_size;
565 #endif
566 	int biggestone, i, segind;
567 #ifdef WITNESS
568 	vm_offset_t mapped;
569 	int witness_size;
570 #endif
571 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
572 	long ii;
573 #endif
574 #ifdef VM_FREEPOOL_LAZYINIT
575 	int lazyinit;
576 #endif
577 
578 	vaddr = round_page(vaddr);
579 
580 	vm_phys_early_startup();
581 	biggestone = vm_phys_avail_largest();
582 	end = phys_avail[biggestone+1];
583 
584 	/*
585 	 * Initialize the page and queue locks.
586 	 */
587 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
588 	for (i = 0; i < PA_LOCK_COUNT; i++)
589 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
590 	for (i = 0; i < vm_ndomains; i++)
591 		vm_page_domain_init(i);
592 
593 	new_end = end;
594 #ifdef WITNESS
595 	witness_size = round_page(witness_startup_count());
596 	new_end -= witness_size;
597 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
598 	    VM_PROT_READ | VM_PROT_WRITE);
599 	bzero((void *)mapped, witness_size);
600 	witness_startup((void *)mapped);
601 #endif
602 
603 #if MINIDUMP_PAGE_TRACKING
604 	/*
605 	 * Allocate a bitmap to indicate that a random physical page
606 	 * needs to be included in a minidump.
607 	 *
608 	 * The amd64 port needs this to indicate which direct map pages
609 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
610 	 *
611 	 * However, i386 still needs this workspace internally within the
612 	 * minidump code.  In theory, they are not needed on i386, but are
613 	 * included should the sf_buf code decide to use them.
614 	 */
615 	last_pa = 0;
616 	vm_page_dump_pages = 0;
617 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
618 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
619 		    dump_avail[i] / PAGE_SIZE;
620 		if (dump_avail[i + 1] > last_pa)
621 			last_pa = dump_avail[i + 1];
622 	}
623 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
624 	new_end -= vm_page_dump_size;
625 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
626 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
627 	bzero((void *)vm_page_dump, vm_page_dump_size);
628 #if MINIDUMP_STARTUP_PAGE_TRACKING
629 	/*
630 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
631 	 * in a crash dump.  When pmap_map() uses the direct map, they are
632 	 * not automatically included.
633 	 */
634 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
635 		dump_add_page(pa);
636 #endif
637 #else
638 	(void)last_pa;
639 #endif
640 	phys_avail[biggestone + 1] = new_end;
641 #ifdef __amd64__
642 	/*
643 	 * Request that the physical pages underlying the message buffer be
644 	 * included in a crash dump.  Since the message buffer is accessed
645 	 * through the direct map, they are not automatically included.
646 	 */
647 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
648 	last_pa = pa + round_page(msgbufsize);
649 	while (pa < last_pa) {
650 		dump_add_page(pa);
651 		pa += PAGE_SIZE;
652 	}
653 #endif
654 	/*
655 	 * Compute the number of pages of memory that will be available for
656 	 * use, taking into account the overhead of a page structure per page.
657 	 * In other words, solve
658 	 *	"available physical memory" - round_page(page_range *
659 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
660 	 * for page_range.
661 	 */
662 	low_avail = phys_avail[0];
663 	high_avail = phys_avail[1];
664 	for (i = 0; i < vm_phys_nsegs; i++) {
665 		if (vm_phys_segs[i].start < low_avail)
666 			low_avail = vm_phys_segs[i].start;
667 		if (vm_phys_segs[i].end > high_avail)
668 			high_avail = vm_phys_segs[i].end;
669 	}
670 	/* Skip the first chunk.  It is already accounted for. */
671 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
672 		if (phys_avail[i] < low_avail)
673 			low_avail = phys_avail[i];
674 		if (phys_avail[i + 1] > high_avail)
675 			high_avail = phys_avail[i + 1];
676 	}
677 	first_page = low_avail / PAGE_SIZE;
678 #ifdef VM_PHYSSEG_SPARSE
679 	size = 0;
680 	for (i = 0; i < vm_phys_nsegs; i++)
681 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
682 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
683 		size += phys_avail[i + 1] - phys_avail[i];
684 #elif defined(VM_PHYSSEG_DENSE)
685 	size = high_avail - low_avail;
686 #else
687 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
688 #endif
689 
690 #ifdef PMAP_HAS_PAGE_ARRAY
691 	pmap_page_array_startup(size / PAGE_SIZE);
692 	biggestone = vm_phys_avail_largest();
693 	end = new_end = phys_avail[biggestone + 1];
694 #else
695 #ifdef VM_PHYSSEG_DENSE
696 	/*
697 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
698 	 * the overhead of a page structure per page only if vm_page_array is
699 	 * allocated from the last physical memory chunk.  Otherwise, we must
700 	 * allocate page structures representing the physical memory
701 	 * underlying vm_page_array, even though they will not be used.
702 	 */
703 	if (new_end != high_avail)
704 		page_range = size / PAGE_SIZE;
705 	else
706 #endif
707 	{
708 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
709 
710 		/*
711 		 * If the partial bytes remaining are large enough for
712 		 * a page (PAGE_SIZE) without a corresponding
713 		 * 'struct vm_page', then new_end will contain an
714 		 * extra page after subtracting the length of the VM
715 		 * page array.  Compensate by subtracting an extra
716 		 * page from new_end.
717 		 */
718 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
719 			if (new_end == high_avail)
720 				high_avail -= PAGE_SIZE;
721 			new_end -= PAGE_SIZE;
722 		}
723 	}
724 	end = new_end;
725 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
726 #endif
727 
728 #if VM_NRESERVLEVEL > 0
729 	/*
730 	 * Allocate physical memory for the reservation management system's
731 	 * data structures, and map it.
732 	 */
733 	new_end = vm_reserv_startup(&vaddr, new_end);
734 #endif
735 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
736 	/*
737 	 * Include vm_page_array and vm_reserv_array in a crash dump.
738 	 */
739 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
740 		dump_add_page(pa);
741 #endif
742 	phys_avail[biggestone + 1] = new_end;
743 
744 	/*
745 	 * Add physical memory segments corresponding to the available
746 	 * physical pages.
747 	 */
748 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
749 		if (vm_phys_avail_size(i) != 0)
750 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
751 
752 	/*
753 	 * Initialize the physical memory allocator.
754 	 */
755 	vm_phys_init();
756 
757 #ifdef VM_FREEPOOL_LAZYINIT
758 	lazyinit = 1;
759 	TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit);
760 #endif
761 
762 	/*
763 	 * Initialize the page structures and add every available page to the
764 	 * physical memory allocator's free lists.
765 	 */
766 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
767 	for (ii = 0; ii < vm_page_array_size; ii++) {
768 		m = &vm_page_array[ii];
769 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0,
770 		    VM_FREEPOOL_DEFAULT);
771 		m->flags = PG_FICTITIOUS;
772 	}
773 #endif
774 	vm_cnt.v_page_count = 0;
775 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
776 		seg = &vm_phys_segs[segind];
777 
778 		/*
779 		 * If lazy vm_page initialization is not enabled, simply
780 		 * initialize all of the pages in the segment.  Otherwise, we
781 		 * only initialize:
782 		 * 1. Pages not covered by phys_avail[], since they might be
783 		 *    freed to the allocator at some future point, e.g., by
784 		 *    kmem_bootstrap_free().
785 		 * 2. The first page of each run of free pages handed to the
786 		 *    vm_phys allocator, which in turn defers initialization
787 		 *    of pages until they are needed.
788 		 * This avoids blocking the boot process for long periods, which
789 		 * may be relevant for VMs (which ought to boot as quickly as
790 		 * possible) and/or systems with large amounts of physical
791 		 * memory.
792 		 */
793 #ifdef VM_FREEPOOL_LAZYINIT
794 		if (lazyinit) {
795 			startp = seg->start;
796 			for (i = 0; phys_avail[i + 1] != 0; i += 2) {
797 				if (startp >= seg->end)
798 					break;
799 
800 				if (phys_avail[i + 1] < startp)
801 					continue;
802 				if (phys_avail[i] <= startp) {
803 					startp = phys_avail[i + 1];
804 					continue;
805 				}
806 
807 				m = vm_phys_seg_paddr_to_vm_page(seg, startp);
808 				for (endp = MIN(phys_avail[i], seg->end);
809 				    startp < endp; startp += PAGE_SIZE, m++) {
810 					vm_page_init_page(m, startp, segind,
811 					    VM_FREEPOOL_DEFAULT);
812 				}
813 			}
814 		} else
815 #endif
816 			for (m = seg->first_page, pa = seg->start;
817 			    pa < seg->end; m++, pa += PAGE_SIZE) {
818 				vm_page_init_page(m, pa, segind,
819 				    VM_FREEPOOL_DEFAULT);
820 			}
821 
822 		/*
823 		 * Add the segment's pages that are covered by one of
824 		 * phys_avail's ranges to the free lists.
825 		 */
826 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
827 			if (seg->end <= phys_avail[i] ||
828 			    seg->start >= phys_avail[i + 1])
829 				continue;
830 
831 			startp = MAX(seg->start, phys_avail[i]);
832 			endp = MIN(seg->end, phys_avail[i + 1]);
833 			pagecount = (u_long)atop(endp - startp);
834 			if (pagecount == 0)
835 				continue;
836 
837 			m = vm_phys_seg_paddr_to_vm_page(seg, startp);
838 #ifdef VM_FREEPOOL_LAZYINIT
839 			if (lazyinit) {
840 				vm_page_init_page(m, startp, segind,
841 				    VM_FREEPOOL_LAZYINIT);
842 			}
843 #endif
844 			vmd = VM_DOMAIN(seg->domain);
845 			vm_domain_free_lock(vmd);
846 			vm_phys_enqueue_contig(m, pagecount);
847 			vm_domain_free_unlock(vmd);
848 			vm_domain_freecnt_inc(vmd, pagecount);
849 			vm_cnt.v_page_count += (u_int)pagecount;
850 			vmd->vmd_page_count += (u_int)pagecount;
851 			vmd->vmd_segs |= 1UL << segind;
852 		}
853 	}
854 
855 	/*
856 	 * Remove blacklisted pages from the physical memory allocator.
857 	 */
858 	TAILQ_INIT(&blacklist_head);
859 	vm_page_blacklist_load(&list, &listend);
860 	vm_page_blacklist_check(list, listend);
861 
862 	list = kern_getenv("vm.blacklist");
863 	vm_page_blacklist_check(list, NULL);
864 
865 	freeenv(list);
866 #if VM_NRESERVLEVEL > 0
867 	/*
868 	 * Initialize the reservation management system.
869 	 */
870 	vm_reserv_init();
871 #endif
872 
873 	return (vaddr);
874 }
875 
876 void
877 vm_page_reference(vm_page_t m)
878 {
879 
880 	vm_page_aflag_set(m, PGA_REFERENCED);
881 }
882 
883 /*
884  *	vm_page_trybusy
885  *
886  *	Helper routine for grab functions to trylock busy.
887  *
888  *	Returns true on success and false on failure.
889  */
890 static bool
891 vm_page_trybusy(vm_page_t m, int allocflags)
892 {
893 
894 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
895 		return (vm_page_trysbusy(m));
896 	else
897 		return (vm_page_tryxbusy(m));
898 }
899 
900 /*
901  *	vm_page_tryacquire
902  *
903  *	Helper routine for grab functions to trylock busy and wire.
904  *
905  *	Returns true on success and false on failure.
906  */
907 static inline bool
908 vm_page_tryacquire(vm_page_t m, int allocflags)
909 {
910 	bool locked;
911 
912 	locked = vm_page_trybusy(m, allocflags);
913 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
914 		vm_page_wire(m);
915 	return (locked);
916 }
917 
918 /*
919  *	vm_page_busy_acquire:
920  *
921  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
922  *	and drop the object lock if necessary.
923  */
924 bool
925 vm_page_busy_acquire(vm_page_t m, int allocflags)
926 {
927 	vm_object_t obj;
928 	bool locked;
929 
930 	/*
931 	 * The page-specific object must be cached because page
932 	 * identity can change during the sleep, causing the
933 	 * re-lock of a different object.
934 	 * It is assumed that a reference to the object is already
935 	 * held by the callers.
936 	 */
937 	obj = atomic_load_ptr(&m->object);
938 	for (;;) {
939 		if (vm_page_tryacquire(m, allocflags))
940 			return (true);
941 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
942 			return (false);
943 		if (obj != NULL)
944 			locked = VM_OBJECT_WOWNED(obj);
945 		else
946 			locked = false;
947 		MPASS(locked || vm_page_wired(m));
948 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
949 		    locked) && locked)
950 			VM_OBJECT_WLOCK(obj);
951 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
952 			return (false);
953 		KASSERT(m->object == obj || m->object == NULL,
954 		    ("vm_page_busy_acquire: page %p does not belong to %p",
955 		    m, obj));
956 	}
957 }
958 
959 /*
960  *	vm_page_busy_downgrade:
961  *
962  *	Downgrade an exclusive busy page into a single shared busy page.
963  */
964 void
965 vm_page_busy_downgrade(vm_page_t m)
966 {
967 	u_int x;
968 
969 	vm_page_assert_xbusied(m);
970 
971 	x = vm_page_busy_fetch(m);
972 	for (;;) {
973 		if (atomic_fcmpset_rel_int(&m->busy_lock,
974 		    &x, VPB_SHARERS_WORD(1)))
975 			break;
976 	}
977 	if ((x & VPB_BIT_WAITERS) != 0)
978 		wakeup(m);
979 }
980 
981 /*
982  *
983  *	vm_page_busy_tryupgrade:
984  *
985  *	Attempt to upgrade a single shared busy into an exclusive busy.
986  */
987 int
988 vm_page_busy_tryupgrade(vm_page_t m)
989 {
990 	u_int ce, x;
991 
992 	vm_page_assert_sbusied(m);
993 
994 	x = vm_page_busy_fetch(m);
995 	ce = VPB_CURTHREAD_EXCLUSIVE;
996 	for (;;) {
997 		if (VPB_SHARERS(x) > 1)
998 			return (0);
999 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1000 		    ("vm_page_busy_tryupgrade: invalid lock state"));
1001 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1002 		    ce | (x & VPB_BIT_WAITERS)))
1003 			continue;
1004 		return (1);
1005 	}
1006 }
1007 
1008 /*
1009  *	vm_page_sbusied:
1010  *
1011  *	Return a positive value if the page is shared busied, 0 otherwise.
1012  */
1013 int
1014 vm_page_sbusied(vm_page_t m)
1015 {
1016 	u_int x;
1017 
1018 	x = vm_page_busy_fetch(m);
1019 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1020 }
1021 
1022 /*
1023  *	vm_page_sunbusy:
1024  *
1025  *	Shared unbusy a page.
1026  */
1027 void
1028 vm_page_sunbusy(vm_page_t m)
1029 {
1030 	u_int x;
1031 
1032 	vm_page_assert_sbusied(m);
1033 
1034 	x = vm_page_busy_fetch(m);
1035 	for (;;) {
1036 		KASSERT(x != VPB_FREED,
1037 		    ("vm_page_sunbusy: Unlocking freed page."));
1038 		if (VPB_SHARERS(x) > 1) {
1039 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1040 			    x - VPB_ONE_SHARER))
1041 				break;
1042 			continue;
1043 		}
1044 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1045 		    ("vm_page_sunbusy: invalid lock state"));
1046 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1047 			continue;
1048 		if ((x & VPB_BIT_WAITERS) == 0)
1049 			break;
1050 		wakeup(m);
1051 		break;
1052 	}
1053 }
1054 
1055 /*
1056  *	vm_page_busy_sleep:
1057  *
1058  *	Sleep if the page is busy, using the page pointer as wchan.
1059  *	This is used to implement the hard-path of the busying mechanism.
1060  *
1061  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1062  *	will not sleep if the page is shared-busy.
1063  *
1064  *	The object lock must be held on entry.
1065  *
1066  *	Returns true if it slept and dropped the object lock, or false
1067  *	if there was no sleep and the lock is still held.
1068  */
1069 bool
1070 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1071 {
1072 	vm_object_t obj;
1073 
1074 	obj = m->object;
1075 	VM_OBJECT_ASSERT_LOCKED(obj);
1076 
1077 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1078 	    true));
1079 }
1080 
1081 /*
1082  *	vm_page_busy_sleep_unlocked:
1083  *
1084  *	Sleep if the page is busy, using the page pointer as wchan.
1085  *	This is used to implement the hard-path of busying mechanism.
1086  *
1087  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1088  *	will not sleep if the page is shared-busy.
1089  *
1090  *	The object lock must not be held on entry.  The operation will
1091  *	return if the page changes identity.
1092  */
1093 void
1094 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1095     const char *wmesg, int allocflags)
1096 {
1097 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1098 
1099 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1100 }
1101 
1102 /*
1103  *	_vm_page_busy_sleep:
1104  *
1105  *	Internal busy sleep function.  Verifies the page identity and
1106  *	lockstate against parameters.  Returns true if it sleeps and
1107  *	false otherwise.
1108  *
1109  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1110  *
1111  *	If locked is true the lock will be dropped for any true returns
1112  *	and held for any false returns.
1113  */
1114 static bool
1115 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1116     const char *wmesg, int allocflags, bool locked)
1117 {
1118 	bool xsleep;
1119 	u_int x;
1120 
1121 	/*
1122 	 * If the object is busy we must wait for that to drain to zero
1123 	 * before trying the page again.
1124 	 */
1125 	if (obj != NULL && vm_object_busied(obj)) {
1126 		if (locked)
1127 			VM_OBJECT_DROP(obj);
1128 		vm_object_busy_wait(obj, wmesg);
1129 		return (true);
1130 	}
1131 
1132 	if (!vm_page_busied(m))
1133 		return (false);
1134 
1135 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1136 	sleepq_lock(m);
1137 	x = vm_page_busy_fetch(m);
1138 	do {
1139 		/*
1140 		 * If the page changes objects or becomes unlocked we can
1141 		 * simply return.
1142 		 */
1143 		if (x == VPB_UNBUSIED ||
1144 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1145 		    m->object != obj || m->pindex != pindex) {
1146 			sleepq_release(m);
1147 			return (false);
1148 		}
1149 		if ((x & VPB_BIT_WAITERS) != 0)
1150 			break;
1151 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1152 	if (locked)
1153 		VM_OBJECT_DROP(obj);
1154 	DROP_GIANT();
1155 	sleepq_add(m, NULL, wmesg, 0, 0);
1156 	sleepq_wait(m, PVM);
1157 	PICKUP_GIANT();
1158 	return (true);
1159 }
1160 
1161 /*
1162  *	vm_page_trysbusy:
1163  *
1164  *	Try to shared busy a page.
1165  *	If the operation succeeds 1 is returned otherwise 0.
1166  *	The operation never sleeps.
1167  */
1168 int
1169 vm_page_trysbusy(vm_page_t m)
1170 {
1171 	vm_object_t obj;
1172 	u_int x;
1173 
1174 	obj = m->object;
1175 	x = vm_page_busy_fetch(m);
1176 	for (;;) {
1177 		if ((x & VPB_BIT_SHARED) == 0)
1178 			return (0);
1179 		/*
1180 		 * Reduce the window for transient busies that will trigger
1181 		 * false negatives in vm_page_ps_test().
1182 		 */
1183 		if (obj != NULL && vm_object_busied(obj))
1184 			return (0);
1185 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1186 		    x + VPB_ONE_SHARER))
1187 			break;
1188 	}
1189 
1190 	/* Refetch the object now that we're guaranteed that it is stable. */
1191 	obj = m->object;
1192 	if (obj != NULL && vm_object_busied(obj)) {
1193 		vm_page_sunbusy(m);
1194 		return (0);
1195 	}
1196 	return (1);
1197 }
1198 
1199 /*
1200  *	vm_page_tryxbusy:
1201  *
1202  *	Try to exclusive busy a page.
1203  *	If the operation succeeds 1 is returned otherwise 0.
1204  *	The operation never sleeps.
1205  */
1206 int
1207 vm_page_tryxbusy(vm_page_t m)
1208 {
1209 	vm_object_t obj;
1210 
1211         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1212             VPB_CURTHREAD_EXCLUSIVE) == 0)
1213 		return (0);
1214 
1215 	obj = m->object;
1216 	if (obj != NULL && vm_object_busied(obj)) {
1217 		vm_page_xunbusy(m);
1218 		return (0);
1219 	}
1220 	return (1);
1221 }
1222 
1223 static void
1224 vm_page_xunbusy_hard_tail(vm_page_t m)
1225 {
1226 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1227 	/* Wake the waiter. */
1228 	wakeup(m);
1229 }
1230 
1231 /*
1232  *	vm_page_xunbusy_hard:
1233  *
1234  *	Called when unbusy has failed because there is a waiter.
1235  */
1236 void
1237 vm_page_xunbusy_hard(vm_page_t m)
1238 {
1239 	vm_page_assert_xbusied(m);
1240 	vm_page_xunbusy_hard_tail(m);
1241 }
1242 
1243 void
1244 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1245 {
1246 	vm_page_assert_xbusied_unchecked(m);
1247 	vm_page_xunbusy_hard_tail(m);
1248 }
1249 
1250 static void
1251 vm_page_busy_free(vm_page_t m)
1252 {
1253 	u_int x;
1254 
1255 	atomic_thread_fence_rel();
1256 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1257 	if ((x & VPB_BIT_WAITERS) != 0)
1258 		wakeup(m);
1259 }
1260 
1261 /*
1262  *	vm_page_unhold_pages:
1263  *
1264  *	Unhold each of the pages that is referenced by the given array.
1265  */
1266 void
1267 vm_page_unhold_pages(vm_page_t *ma, int count)
1268 {
1269 
1270 	for (; count != 0; count--) {
1271 		vm_page_unwire(*ma, PQ_ACTIVE);
1272 		ma++;
1273 	}
1274 }
1275 
1276 vm_page_t
1277 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1278 {
1279 	vm_page_t m;
1280 
1281 #ifdef VM_PHYSSEG_SPARSE
1282 	m = vm_phys_paddr_to_vm_page(pa);
1283 	if (m == NULL)
1284 		m = vm_phys_fictitious_to_vm_page(pa);
1285 	return (m);
1286 #elif defined(VM_PHYSSEG_DENSE)
1287 	long pi;
1288 
1289 	pi = atop(pa);
1290 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1291 		m = &vm_page_array[pi - first_page];
1292 		return (m);
1293 	}
1294 	return (vm_phys_fictitious_to_vm_page(pa));
1295 #else
1296 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1297 #endif
1298 }
1299 
1300 /*
1301  *	vm_page_getfake:
1302  *
1303  *	Create a fictitious page with the specified physical address and
1304  *	memory attribute.  The memory attribute is the only the machine-
1305  *	dependent aspect of a fictitious page that must be initialized.
1306  */
1307 vm_page_t
1308 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1309 {
1310 	vm_page_t m;
1311 
1312 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1313 	vm_page_initfake(m, paddr, memattr);
1314 	return (m);
1315 }
1316 
1317 void
1318 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1319 {
1320 
1321 	if ((m->flags & PG_FICTITIOUS) != 0) {
1322 		/*
1323 		 * The page's memattr might have changed since the
1324 		 * previous initialization.  Update the pmap to the
1325 		 * new memattr.
1326 		 */
1327 		goto memattr;
1328 	}
1329 	m->phys_addr = paddr;
1330 	m->a.queue = PQ_NONE;
1331 	/* Fictitious pages don't use "segind". */
1332 	m->flags = PG_FICTITIOUS;
1333 	/* Fictitious pages don't use "order" or "pool". */
1334 	m->oflags = VPO_UNMANAGED;
1335 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1336 	/* Fictitious pages are unevictable. */
1337 	m->ref_count = 1;
1338 	pmap_page_init(m);
1339 memattr:
1340 	pmap_page_set_memattr(m, memattr);
1341 }
1342 
1343 /*
1344  *	vm_page_putfake:
1345  *
1346  *	Release a fictitious page.
1347  */
1348 void
1349 vm_page_putfake(vm_page_t m)
1350 {
1351 
1352 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1353 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1354 	    ("vm_page_putfake: bad page %p", m));
1355 	vm_page_assert_xbusied(m);
1356 	vm_page_busy_free(m);
1357 	uma_zfree(fakepg_zone, m);
1358 }
1359 
1360 /*
1361  *	vm_page_updatefake:
1362  *
1363  *	Update the given fictitious page to the specified physical address and
1364  *	memory attribute.
1365  */
1366 void
1367 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1368 {
1369 
1370 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1371 	    ("vm_page_updatefake: bad page %p", m));
1372 	m->phys_addr = paddr;
1373 	pmap_page_set_memattr(m, memattr);
1374 }
1375 
1376 /*
1377  *	vm_page_free:
1378  *
1379  *	Free a page.
1380  */
1381 void
1382 vm_page_free(vm_page_t m)
1383 {
1384 
1385 	m->flags &= ~PG_ZERO;
1386 	vm_page_free_toq(m);
1387 }
1388 
1389 /*
1390  *	vm_page_free_zero:
1391  *
1392  *	Free a page to the zerod-pages queue
1393  */
1394 void
1395 vm_page_free_zero(vm_page_t m)
1396 {
1397 
1398 	m->flags |= PG_ZERO;
1399 	vm_page_free_toq(m);
1400 }
1401 
1402 /*
1403  * Unbusy and handle the page queueing for a page from a getpages request that
1404  * was optionally read ahead or behind.
1405  */
1406 void
1407 vm_page_readahead_finish(vm_page_t m)
1408 {
1409 
1410 	/* We shouldn't put invalid pages on queues. */
1411 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1412 
1413 	/*
1414 	 * Since the page is not the actually needed one, whether it should
1415 	 * be activated or deactivated is not obvious.  Empirical results
1416 	 * have shown that deactivating the page is usually the best choice,
1417 	 * unless the page is wanted by another thread.
1418 	 */
1419 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1420 		vm_page_activate(m);
1421 	else
1422 		vm_page_deactivate(m);
1423 	vm_page_xunbusy_unchecked(m);
1424 }
1425 
1426 /*
1427  * Destroy the identity of an invalid page and free it if possible.
1428  * This is intended to be used when reading a page from backing store fails.
1429  */
1430 void
1431 vm_page_free_invalid(vm_page_t m)
1432 {
1433 
1434 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1435 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1436 	KASSERT(m->object != NULL, ("page %p has no object", m));
1437 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1438 
1439 	/*
1440 	 * We may be attempting to free the page as part of the handling for an
1441 	 * I/O error, in which case the page was xbusied by a different thread.
1442 	 */
1443 	vm_page_xbusy_claim(m);
1444 
1445 	/*
1446 	 * If someone has wired this page while the object lock
1447 	 * was not held, then the thread that unwires is responsible
1448 	 * for freeing the page.  Otherwise just free the page now.
1449 	 * The wire count of this unmapped page cannot change while
1450 	 * we have the page xbusy and the page's object wlocked.
1451 	 */
1452 	if (vm_page_remove(m))
1453 		vm_page_free(m);
1454 }
1455 
1456 /*
1457  *	vm_page_dirty_KBI:		[ internal use only ]
1458  *
1459  *	Set all bits in the page's dirty field.
1460  *
1461  *	The object containing the specified page must be locked if the
1462  *	call is made from the machine-independent layer.
1463  *
1464  *	See vm_page_clear_dirty_mask().
1465  *
1466  *	This function should only be called by vm_page_dirty().
1467  */
1468 void
1469 vm_page_dirty_KBI(vm_page_t m)
1470 {
1471 
1472 	/* Refer to this operation by its public name. */
1473 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1474 	m->dirty = VM_PAGE_BITS_ALL;
1475 }
1476 
1477 /*
1478  * Insert the given page into the given object at the given pindex.  mpred is
1479  * used for memq linkage.  From vm_page_insert, lookup is true, mpred is
1480  * initially NULL, and this procedure looks it up.  From vm_page_insert_after
1481  * and vm_page_iter_insert, lookup is false and mpred is known to the caller
1482  * to be valid, and may be NULL if this will be the page with the lowest
1483  * pindex.
1484  *
1485  * The procedure is marked __always_inline to suggest to the compiler to
1486  * eliminate the lookup parameter and the associated alternate branch.
1487  */
1488 static __always_inline int
1489 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1490     struct pctrie_iter *pages, bool iter, vm_page_t mpred, bool lookup)
1491 {
1492 	int error;
1493 
1494 	VM_OBJECT_ASSERT_WLOCKED(object);
1495 	KASSERT(m->object == NULL,
1496 	    ("vm_page_insert: page %p already inserted", m));
1497 
1498 	/*
1499 	 * Record the object/offset pair in this page.
1500 	 */
1501 	m->object = object;
1502 	m->pindex = pindex;
1503 	m->ref_count |= VPRC_OBJREF;
1504 
1505 	/*
1506 	 * Add this page to the object's radix tree, and look up mpred if
1507 	 * needed.
1508 	 */
1509 	if (iter) {
1510 		KASSERT(!lookup, ("%s: cannot lookup mpred", __func__));
1511 		error = vm_radix_iter_insert(pages, m);
1512 	} else if (lookup)
1513 		error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred);
1514 	else
1515 		error = vm_radix_insert(&object->rtree, m);
1516 	if (__predict_false(error != 0)) {
1517 		m->object = NULL;
1518 		m->pindex = 0;
1519 		m->ref_count &= ~VPRC_OBJREF;
1520 		return (1);
1521 	}
1522 
1523 	/*
1524 	 * Now link into the object's ordered list of backed pages.
1525 	 */
1526 	vm_page_insert_radixdone(m, object, mpred);
1527 	vm_pager_page_inserted(object, m);
1528 	return (0);
1529 }
1530 
1531 /*
1532  *	vm_page_insert:		[ internal use only ]
1533  *
1534  *	Inserts the given mem entry into the object and object list.
1535  *
1536  *	The object must be locked.
1537  */
1538 int
1539 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1540 {
1541 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL,
1542 	    true));
1543 }
1544 
1545 /*
1546  *	vm_page_insert_after:
1547  *
1548  *	Inserts the page "m" into the specified object at offset "pindex".
1549  *
1550  *	The page "mpred" must immediately precede the offset "pindex" within
1551  *	the specified object.
1552  *
1553  *	The object must be locked.
1554  */
1555 static int
1556 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1557     vm_page_t mpred)
1558 {
1559 	return (vm_page_insert_lookup(m, object, pindex, NULL, false, mpred,
1560 	    false));
1561 }
1562 
1563 /*
1564  *	vm_page_iter_insert:
1565  *
1566  *	Tries to insert the page "m" into the specified object at offset
1567  *	"pindex" using the iterator "pages".  Returns 0 if the insertion was
1568  *	successful.
1569  *
1570  *	The page "mpred" must immediately precede the offset "pindex" within
1571  *	the specified object.
1572  *
1573  *	The object must be locked.
1574  */
1575 static int
1576 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object,
1577     vm_pindex_t pindex, vm_page_t mpred)
1578 {
1579 	return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred,
1580 	    false));
1581 }
1582 
1583 /*
1584  *	vm_page_insert_radixdone:
1585  *
1586  *	Complete page "m" insertion into the specified object after the
1587  *	radix trie hooking.
1588  *
1589  *	The page "mpred" must precede the offset "m->pindex" within the
1590  *	specified object.
1591  *
1592  *	The object must be locked.
1593  */
1594 static void
1595 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1596 {
1597 
1598 	VM_OBJECT_ASSERT_WLOCKED(object);
1599 	KASSERT(object != NULL && m->object == object,
1600 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1601 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1602 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1603 	if (mpred != NULL) {
1604 		KASSERT(mpred->object == object,
1605 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1606 		KASSERT(mpred->pindex < m->pindex,
1607 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1608 		KASSERT(TAILQ_NEXT(mpred, listq) == NULL ||
1609 		    m->pindex < TAILQ_NEXT(mpred, listq)->pindex,
1610 		    ("vm_page_insert_radixdone: pindex doesn't precede msucc"));
1611 	} else {
1612 		KASSERT(TAILQ_EMPTY(&object->memq) ||
1613 		    m->pindex < TAILQ_FIRST(&object->memq)->pindex,
1614 		    ("vm_page_insert_radixdone: no mpred but not first page"));
1615 	}
1616 
1617 	if (mpred != NULL)
1618 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1619 	else
1620 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1621 
1622 	/*
1623 	 * Show that the object has one more resident page.
1624 	 */
1625 	object->resident_page_count++;
1626 
1627 	/*
1628 	 * Hold the vnode until the last page is released.
1629 	 */
1630 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1631 		vhold(object->handle);
1632 
1633 	/*
1634 	 * Since we are inserting a new and possibly dirty page,
1635 	 * update the object's generation count.
1636 	 */
1637 	if (pmap_page_is_write_mapped(m))
1638 		vm_object_set_writeable_dirty(object);
1639 }
1640 
1641 /*
1642  *	vm_page_remove_radixdone
1643  *
1644  *	Complete page "m" removal from the specified object after the radix trie
1645  *	unhooking.
1646  *
1647  *	The caller is responsible for updating the page's fields to reflect this
1648  *	removal.
1649  */
1650 static void
1651 vm_page_remove_radixdone(vm_page_t m)
1652 {
1653 	vm_object_t object;
1654 
1655 	vm_page_assert_xbusied(m);
1656 	object = m->object;
1657 	VM_OBJECT_ASSERT_WLOCKED(object);
1658 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1659 	    ("page %p is missing its object ref", m));
1660 
1661 	/* Deferred free of swap space. */
1662 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1663 		vm_pager_page_unswapped(m);
1664 
1665 	vm_pager_page_removed(object, m);
1666 	m->object = NULL;
1667 
1668 	/*
1669 	 * Now remove from the object's list of backed pages.
1670 	 */
1671 	TAILQ_REMOVE(&object->memq, m, listq);
1672 
1673 	/*
1674 	 * And show that the object has one fewer resident page.
1675 	 */
1676 	object->resident_page_count--;
1677 
1678 	/*
1679 	 * The vnode may now be recycled.
1680 	 */
1681 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1682 		vdrop(object->handle);
1683 }
1684 
1685 /*
1686  *	vm_page_free_object_prep:
1687  *
1688  *	Disassociates the given page from its VM object.
1689  *
1690  *	The object must be locked, and the page must be xbusy.
1691  */
1692 static void
1693 vm_page_free_object_prep(vm_page_t m)
1694 {
1695 	KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
1696 	    ((m->object->flags & OBJ_UNMANAGED) != 0),
1697 	    ("%s: managed flag mismatch for page %p",
1698 	     __func__, m));
1699 	vm_page_assert_xbusied(m);
1700 
1701 	/*
1702 	 * The object reference can be released without an atomic
1703 	 * operation.
1704 	 */
1705 	KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
1706 	    m->ref_count == VPRC_OBJREF,
1707 	    ("%s: page %p has unexpected ref_count %u",
1708 	    __func__, m, m->ref_count));
1709 	vm_page_remove_radixdone(m);
1710 	m->ref_count -= VPRC_OBJREF;
1711 }
1712 
1713 /*
1714  *	vm_page_iter_free:
1715  *
1716  *	Free the given page, and use the iterator to remove it from the radix
1717  *	tree.
1718  */
1719 void
1720 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m)
1721 {
1722 	vm_radix_iter_remove(pages);
1723 	vm_page_free_object_prep(m);
1724 	vm_page_xunbusy(m);
1725 	m->flags &= ~PG_ZERO;
1726 	vm_page_free_toq(m);
1727 }
1728 
1729 /*
1730  *	vm_page_remove:
1731  *
1732  *	Removes the specified page from its containing object, but does not
1733  *	invalidate any backing storage.  Returns true if the object's reference
1734  *	was the last reference to the page, and false otherwise.
1735  *
1736  *	The object must be locked and the page must be exclusively busied.
1737  *	The exclusive busy will be released on return.  If this is not the
1738  *	final ref and the caller does not hold a wire reference it may not
1739  *	continue to access the page.
1740  */
1741 bool
1742 vm_page_remove(vm_page_t m)
1743 {
1744 	bool dropped;
1745 
1746 	dropped = vm_page_remove_xbusy(m);
1747 	vm_page_xunbusy(m);
1748 
1749 	return (dropped);
1750 }
1751 
1752 /*
1753  *	vm_page_iter_remove:
1754  *
1755  *	Remove the current page, as identified by iterator, and remove it from the
1756  *	radix tree.
1757  */
1758 bool
1759 vm_page_iter_remove(struct pctrie_iter *pages)
1760 {
1761 	vm_page_t m;
1762 	bool dropped;
1763 
1764 	m = vm_radix_iter_page(pages);
1765 	vm_radix_iter_remove(pages);
1766 	vm_page_remove_radixdone(m);
1767 	dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1768 	vm_page_xunbusy(m);
1769 
1770 	return (dropped);
1771 }
1772 
1773 /*
1774  *	vm_page_radix_remove
1775  *
1776  *	Removes the specified page from the radix tree.
1777  */
1778 static void
1779 vm_page_radix_remove(vm_page_t m)
1780 {
1781 	vm_page_t mrem __diagused;
1782 
1783 	mrem = vm_radix_remove(&m->object->rtree, m->pindex);
1784 	KASSERT(mrem == m,
1785 	    ("removed page %p, expected page %p", mrem, m));
1786 }
1787 
1788 /*
1789  *	vm_page_remove_xbusy
1790  *
1791  *	Removes the page but leaves the xbusy held.  Returns true if this
1792  *	removed the final ref and false otherwise.
1793  */
1794 bool
1795 vm_page_remove_xbusy(vm_page_t m)
1796 {
1797 
1798 	vm_page_radix_remove(m);
1799 	vm_page_remove_radixdone(m);
1800 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1801 }
1802 
1803 /*
1804  *	vm_page_lookup:
1805  *
1806  *	Returns the page associated with the object/offset
1807  *	pair specified; if none is found, NULL is returned.
1808  *
1809  *	The object must be locked.
1810  */
1811 vm_page_t
1812 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1813 {
1814 
1815 	VM_OBJECT_ASSERT_LOCKED(object);
1816 	return (vm_radix_lookup(&object->rtree, pindex));
1817 }
1818 
1819 /*
1820  *	vm_page_iter_init:
1821  *
1822  *	Initialize iterator for vm pages.
1823  */
1824 void
1825 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object)
1826 {
1827 
1828 	vm_radix_iter_init(pages, &object->rtree);
1829 }
1830 
1831 /*
1832  *	vm_page_iter_init:
1833  *
1834  *	Initialize iterator for vm pages.
1835  */
1836 void
1837 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object,
1838     vm_pindex_t limit)
1839 {
1840 
1841 	vm_radix_iter_limit_init(pages, &object->rtree, limit);
1842 }
1843 
1844 /*
1845  *	vm_page_iter_lookup:
1846  *
1847  *	Returns the page associated with the object/offset pair specified, and
1848  *	stores the path to its position; if none is found, NULL is returned.
1849  *
1850  *	The iter pctrie must be locked.
1851  */
1852 vm_page_t
1853 vm_page_iter_lookup(struct pctrie_iter *pages, vm_pindex_t pindex)
1854 {
1855 
1856 	return (vm_radix_iter_lookup(pages, pindex));
1857 }
1858 
1859 /*
1860  *	vm_page_lookup_unlocked:
1861  *
1862  *	Returns the page associated with the object/offset pair specified;
1863  *	if none is found, NULL is returned.  The page may be no longer be
1864  *	present in the object at the time that this function returns.  Only
1865  *	useful for opportunistic checks such as inmem().
1866  */
1867 vm_page_t
1868 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1869 {
1870 
1871 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1872 }
1873 
1874 /*
1875  *	vm_page_relookup:
1876  *
1877  *	Returns a page that must already have been busied by
1878  *	the caller.  Used for bogus page replacement.
1879  */
1880 vm_page_t
1881 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1882 {
1883 	vm_page_t m;
1884 
1885 	m = vm_page_lookup_unlocked(object, pindex);
1886 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1887 	    m->object == object && m->pindex == pindex,
1888 	    ("vm_page_relookup: Invalid page %p", m));
1889 	return (m);
1890 }
1891 
1892 /*
1893  * This should only be used by lockless functions for releasing transient
1894  * incorrect acquires.  The page may have been freed after we acquired a
1895  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1896  * further to do.
1897  */
1898 static void
1899 vm_page_busy_release(vm_page_t m)
1900 {
1901 	u_int x;
1902 
1903 	x = vm_page_busy_fetch(m);
1904 	for (;;) {
1905 		if (x == VPB_FREED)
1906 			break;
1907 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1908 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1909 			    x - VPB_ONE_SHARER))
1910 				break;
1911 			continue;
1912 		}
1913 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1914 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1915 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1916 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1917 			continue;
1918 		if ((x & VPB_BIT_WAITERS) != 0)
1919 			wakeup(m);
1920 		break;
1921 	}
1922 }
1923 
1924 /*
1925  *	vm_page_find_least:
1926  *
1927  *	Returns the page associated with the object with least pindex
1928  *	greater than or equal to the parameter pindex, or NULL.
1929  *
1930  *	The object must be locked.
1931  */
1932 vm_page_t
1933 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1934 {
1935 	vm_page_t m;
1936 
1937 	VM_OBJECT_ASSERT_LOCKED(object);
1938 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1939 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1940 	return (m);
1941 }
1942 
1943 /*
1944  *	vm_page_iter_lookup_ge:
1945  *
1946  *	Returns the page associated with the object with least pindex
1947  *	greater than or equal to the parameter pindex, or NULL.  Initializes the
1948  *	iterator to point to that page.
1949  *
1950  *	The iter pctrie must be locked.
1951  */
1952 vm_page_t
1953 vm_page_iter_lookup_ge(struct pctrie_iter *pages, vm_pindex_t pindex)
1954 {
1955 
1956 	return (vm_radix_iter_lookup_ge(pages, pindex));
1957 }
1958 
1959 /*
1960  * Returns the given page's successor (by pindex) within the object if it is
1961  * resident; if none is found, NULL is returned.
1962  *
1963  * The object must be locked.
1964  */
1965 vm_page_t
1966 vm_page_next(vm_page_t m)
1967 {
1968 	vm_page_t next;
1969 
1970 	VM_OBJECT_ASSERT_LOCKED(m->object);
1971 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1972 		MPASS(next->object == m->object);
1973 		if (next->pindex != m->pindex + 1)
1974 			next = NULL;
1975 	}
1976 	return (next);
1977 }
1978 
1979 /*
1980  * Returns the given page's predecessor (by pindex) within the object if it is
1981  * resident; if none is found, NULL is returned.
1982  *
1983  * The object must be locked.
1984  */
1985 vm_page_t
1986 vm_page_prev(vm_page_t m)
1987 {
1988 	vm_page_t prev;
1989 
1990 	VM_OBJECT_ASSERT_LOCKED(m->object);
1991 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1992 		MPASS(prev->object == m->object);
1993 		if (prev->pindex != m->pindex - 1)
1994 			prev = NULL;
1995 	}
1996 	return (prev);
1997 }
1998 
1999 /*
2000  * Uses the page mnew as a replacement for an existing page at index
2001  * pindex which must be already present in the object.
2002  *
2003  * Both pages must be exclusively busied on enter.  The old page is
2004  * unbusied on exit.
2005  *
2006  * A return value of true means mold is now free.  If this is not the
2007  * final ref and the caller does not hold a wire reference it may not
2008  * continue to access the page.
2009  */
2010 static bool
2011 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
2012     vm_page_t mold)
2013 {
2014 	vm_page_t mret __diagused;
2015 	bool dropped;
2016 
2017 	VM_OBJECT_ASSERT_WLOCKED(object);
2018 	vm_page_assert_xbusied(mold);
2019 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
2020 	    ("vm_page_replace: page %p already in object", mnew));
2021 
2022 	/*
2023 	 * This function mostly follows vm_page_insert() and
2024 	 * vm_page_remove() without the radix, object count and vnode
2025 	 * dance.  Double check such functions for more comments.
2026 	 */
2027 
2028 	mnew->object = object;
2029 	mnew->pindex = pindex;
2030 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
2031 	mret = vm_radix_replace(&object->rtree, mnew);
2032 	KASSERT(mret == mold,
2033 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
2034 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
2035 	    (mnew->oflags & VPO_UNMANAGED),
2036 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
2037 
2038 	/* Keep the resident page list in sorted order. */
2039 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
2040 	TAILQ_REMOVE(&object->memq, mold, listq);
2041 	mold->object = NULL;
2042 
2043 	/*
2044 	 * The object's resident_page_count does not change because we have
2045 	 * swapped one page for another, but the generation count should
2046 	 * change if the page is dirty.
2047 	 */
2048 	if (pmap_page_is_write_mapped(mnew))
2049 		vm_object_set_writeable_dirty(object);
2050 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
2051 	vm_page_xunbusy(mold);
2052 
2053 	return (dropped);
2054 }
2055 
2056 void
2057 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
2058     vm_page_t mold)
2059 {
2060 
2061 	vm_page_assert_xbusied(mnew);
2062 
2063 	if (vm_page_replace_hold(mnew, object, pindex, mold))
2064 		vm_page_free(mold);
2065 }
2066 
2067 /*
2068  *	vm_page_iter_rename:
2069  *
2070  *	Tries to move the specified page from its current object to a new object
2071  *	and pindex, using the given iterator to remove the page from its current
2072  *	object.  Returns true if the move was successful, and false if the move
2073  *	was aborted due to a failed memory allocation.
2074  *
2075  *	Panics if a page already resides in the new object at the new pindex.
2076  *
2077  *	Note: swap associated with the page must be invalidated by the move.  We
2078  *	      have to do this for several reasons:  (1) we aren't freeing the
2079  *	      page, (2) we are dirtying the page, (3) the VM system is probably
2080  *	      moving the page from object A to B, and will then later move
2081  *	      the backing store from A to B and we can't have a conflict.
2082  *
2083  *	Note: we *always* dirty the page.  It is necessary both for the
2084  *	      fact that we moved it, and because we may be invalidating
2085  *	      swap.
2086  *
2087  *	The objects must be locked.
2088  */
2089 bool
2090 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m,
2091     vm_object_t new_object, vm_pindex_t new_pindex)
2092 {
2093 	vm_page_t mpred;
2094 	vm_pindex_t opidx;
2095 
2096 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
2097 	    ("%s: page %p is missing object ref", __func__, m));
2098 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2099 	VM_OBJECT_ASSERT_WLOCKED(new_object);
2100 
2101 	/*
2102 	 * Create a custom version of vm_page_insert() which does not depend
2103 	 * by m_prev and can cheat on the implementation aspects of the
2104 	 * function.
2105 	 */
2106 	opidx = m->pindex;
2107 	m->pindex = new_pindex;
2108 	if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) {
2109 		m->pindex = opidx;
2110 		return (false);
2111 	}
2112 
2113 	/*
2114 	 * The operation cannot fail anymore.  The removal must happen before
2115 	 * the listq iterator is tainted.
2116 	 */
2117 	m->pindex = opidx;
2118 	vm_radix_iter_remove(old_pages);
2119 	vm_page_remove_radixdone(m);
2120 
2121 	/* Return back to the new pindex to complete vm_page_insert(). */
2122 	m->pindex = new_pindex;
2123 	m->object = new_object;
2124 
2125 	vm_page_insert_radixdone(m, new_object, mpred);
2126 	vm_page_dirty(m);
2127 	vm_pager_page_inserted(new_object, m);
2128 	return (true);
2129 }
2130 
2131 /*
2132  *	vm_page_mpred:
2133  *
2134  *	Return the greatest page of the object with index <= pindex,
2135  *	or NULL, if there is none.  Assumes object lock is held.
2136  */
2137 vm_page_t
2138 vm_page_mpred(vm_object_t object, vm_pindex_t pindex)
2139 {
2140 	return (vm_radix_lookup_le(&object->rtree, pindex));
2141 }
2142 
2143 /*
2144  *	vm_page_alloc:
2145  *
2146  *	Allocate and return a page that is associated with the specified
2147  *	object and offset pair.  By default, this page is exclusive busied.
2148  *
2149  *	The caller must always specify an allocation class.
2150  *
2151  *	allocation classes:
2152  *	VM_ALLOC_NORMAL		normal process request
2153  *	VM_ALLOC_SYSTEM		system *really* needs a page
2154  *	VM_ALLOC_INTERRUPT	interrupt time request
2155  *
2156  *	optional allocation flags:
2157  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2158  *				intends to allocate
2159  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2160  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2161  *	VM_ALLOC_SBUSY		shared busy the allocated page
2162  *	VM_ALLOC_WIRED		wire the allocated page
2163  *	VM_ALLOC_ZERO		prefer a zeroed page
2164  */
2165 vm_page_t
2166 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
2167 {
2168 
2169 	return (vm_page_alloc_after(object, pindex, req,
2170 	    vm_page_mpred(object, pindex)));
2171 }
2172 
2173 /*
2174  * Allocate a page in the specified object with the given page index.  To
2175  * optimize insertion of the page into the object, the caller must also specify
2176  * the resident page in the object with largest index smaller than the given
2177  * page index, or NULL if no such page exists.
2178  */
2179 static vm_page_t
2180 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
2181     int req, vm_page_t mpred)
2182 {
2183 	struct vm_domainset_iter di;
2184 	vm_page_t m;
2185 	int domain;
2186 
2187 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2188 	do {
2189 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2190 		    mpred);
2191 		if (m != NULL)
2192 			break;
2193 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2194 
2195 	return (m);
2196 }
2197 
2198 /*
2199  * Returns true if the number of free pages exceeds the minimum
2200  * for the request class and false otherwise.
2201  */
2202 static int
2203 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2204 {
2205 	u_int limit, old, new;
2206 
2207 	if (req_class == VM_ALLOC_INTERRUPT)
2208 		limit = 0;
2209 	else if (req_class == VM_ALLOC_SYSTEM)
2210 		limit = vmd->vmd_interrupt_free_min;
2211 	else
2212 		limit = vmd->vmd_free_reserved;
2213 
2214 	/*
2215 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2216 	 */
2217 	limit += npages;
2218 	old = atomic_load_int(&vmd->vmd_free_count);
2219 	do {
2220 		if (old < limit)
2221 			return (0);
2222 		new = old - npages;
2223 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2224 
2225 	/* Wake the page daemon if we've crossed the threshold. */
2226 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2227 		pagedaemon_wakeup(vmd->vmd_domain);
2228 
2229 	/* Only update bitsets on transitions. */
2230 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2231 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2232 		vm_domain_set(vmd);
2233 
2234 	return (1);
2235 }
2236 
2237 int
2238 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2239 {
2240 	int req_class;
2241 
2242 	/*
2243 	 * The page daemon is allowed to dig deeper into the free page list.
2244 	 */
2245 	req_class = req & VM_ALLOC_CLASS_MASK;
2246 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2247 		req_class = VM_ALLOC_SYSTEM;
2248 	return (_vm_domain_allocate(vmd, req_class, npages));
2249 }
2250 
2251 vm_page_t
2252 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2253     int req, vm_page_t mpred)
2254 {
2255 	struct vm_domain *vmd;
2256 	vm_page_t m;
2257 	int flags;
2258 
2259 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2260 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2261 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2262 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2263 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2264 	KASSERT((req & ~VPA_FLAGS) == 0,
2265 	    ("invalid request %#x", req));
2266 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2267 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2268 	    ("invalid request %#x", req));
2269 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2270 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2271 	    (uintmax_t)pindex));
2272 	VM_OBJECT_ASSERT_WLOCKED(object);
2273 
2274 	flags = 0;
2275 	m = NULL;
2276 	if (!vm_pager_can_alloc_page(object, pindex))
2277 		return (NULL);
2278 again:
2279 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2280 		m = vm_page_alloc_nofree_domain(domain, req);
2281 		if (m != NULL)
2282 			goto found;
2283 	}
2284 #if VM_NRESERVLEVEL > 0
2285 	/*
2286 	 * Can we allocate the page from a reservation?
2287 	 */
2288 	if (vm_object_reserv(object) &&
2289 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2290 	    NULL) {
2291 		goto found;
2292 	}
2293 #endif
2294 	vmd = VM_DOMAIN(domain);
2295 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2296 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2297 		    M_NOWAIT | M_NOVM);
2298 		if (m != NULL) {
2299 			flags |= PG_PCPU_CACHE;
2300 			goto found;
2301 		}
2302 	}
2303 	if (vm_domain_allocate(vmd, req, 1)) {
2304 		/*
2305 		 * If not, allocate it from the free page queues.
2306 		 */
2307 		vm_domain_free_lock(vmd);
2308 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2309 		vm_domain_free_unlock(vmd);
2310 		if (m == NULL) {
2311 			vm_domain_freecnt_inc(vmd, 1);
2312 #if VM_NRESERVLEVEL > 0
2313 			if (vm_reserv_reclaim_inactive(domain))
2314 				goto again;
2315 #endif
2316 		}
2317 	}
2318 	if (m == NULL) {
2319 		/*
2320 		 * Not allocatable, give up.
2321 		 */
2322 		if (vm_domain_alloc_fail(vmd, object, req))
2323 			goto again;
2324 		return (NULL);
2325 	}
2326 
2327 	/*
2328 	 * At this point we had better have found a good page.
2329 	 */
2330 found:
2331 	vm_page_dequeue(m);
2332 	vm_page_alloc_check(m);
2333 
2334 	/*
2335 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2336 	 */
2337 	flags |= m->flags & PG_ZERO;
2338 	if ((req & VM_ALLOC_NODUMP) != 0)
2339 		flags |= PG_NODUMP;
2340 	if ((req & VM_ALLOC_NOFREE) != 0)
2341 		flags |= PG_NOFREE;
2342 	m->flags = flags;
2343 	m->a.flags = 0;
2344 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2345 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2346 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2347 	else if ((req & VM_ALLOC_SBUSY) != 0)
2348 		m->busy_lock = VPB_SHARERS_WORD(1);
2349 	else
2350 		m->busy_lock = VPB_UNBUSIED;
2351 	if (req & VM_ALLOC_WIRED) {
2352 		vm_wire_add(1);
2353 		m->ref_count = 1;
2354 	}
2355 	m->a.act_count = 0;
2356 
2357 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2358 		if (req & VM_ALLOC_WIRED) {
2359 			vm_wire_sub(1);
2360 			m->ref_count = 0;
2361 		}
2362 		KASSERT(m->object == NULL, ("page %p has object", m));
2363 		m->oflags = VPO_UNMANAGED;
2364 		m->busy_lock = VPB_UNBUSIED;
2365 		/* Don't change PG_ZERO. */
2366 		vm_page_free_toq(m);
2367 		if (req & VM_ALLOC_WAITFAIL) {
2368 			VM_OBJECT_WUNLOCK(object);
2369 			vm_radix_wait();
2370 			VM_OBJECT_WLOCK(object);
2371 		}
2372 		return (NULL);
2373 	}
2374 
2375 	/* Ignore device objects; the pager sets "memattr" for them. */
2376 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2377 	    (object->flags & OBJ_FICTITIOUS) == 0)
2378 		pmap_page_set_memattr(m, object->memattr);
2379 
2380 	return (m);
2381 }
2382 
2383 /*
2384  *	vm_page_alloc_contig:
2385  *
2386  *	Allocate a contiguous set of physical pages of the given size "npages"
2387  *	from the free lists.  All of the physical pages must be at or above
2388  *	the given physical address "low" and below the given physical address
2389  *	"high".  The given value "alignment" determines the alignment of the
2390  *	first physical page in the set.  If the given value "boundary" is
2391  *	non-zero, then the set of physical pages cannot cross any physical
2392  *	address boundary that is a multiple of that value.  Both "alignment"
2393  *	and "boundary" must be a power of two.
2394  *
2395  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2396  *	then the memory attribute setting for the physical pages is configured
2397  *	to the object's memory attribute setting.  Otherwise, the memory
2398  *	attribute setting for the physical pages is configured to "memattr",
2399  *	overriding the object's memory attribute setting.  However, if the
2400  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2401  *	memory attribute setting for the physical pages cannot be configured
2402  *	to VM_MEMATTR_DEFAULT.
2403  *
2404  *	The specified object may not contain fictitious pages.
2405  *
2406  *	The caller must always specify an allocation class.
2407  *
2408  *	allocation classes:
2409  *	VM_ALLOC_NORMAL		normal process request
2410  *	VM_ALLOC_SYSTEM		system *really* needs a page
2411  *	VM_ALLOC_INTERRUPT	interrupt time request
2412  *
2413  *	optional allocation flags:
2414  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2415  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2416  *	VM_ALLOC_SBUSY		shared busy the allocated page
2417  *	VM_ALLOC_WIRED		wire the allocated page
2418  *	VM_ALLOC_ZERO		prefer a zeroed page
2419  */
2420 vm_page_t
2421 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2422     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2423     vm_paddr_t boundary, vm_memattr_t memattr)
2424 {
2425 	struct vm_domainset_iter di;
2426 	vm_page_t bounds[2];
2427 	vm_page_t m;
2428 	int domain;
2429 	int start_segind;
2430 
2431 	start_segind = -1;
2432 
2433 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2434 	do {
2435 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2436 		    npages, low, high, alignment, boundary, memattr);
2437 		if (m != NULL)
2438 			break;
2439 		if (start_segind == -1)
2440 			start_segind = vm_phys_lookup_segind(low);
2441 		if (vm_phys_find_range(bounds, start_segind, domain,
2442 		    npages, low, high) == -1) {
2443 			vm_domainset_iter_ignore(&di, domain);
2444 		}
2445 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2446 
2447 	return (m);
2448 }
2449 
2450 static vm_page_t
2451 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2452     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2453 {
2454 	struct vm_domain *vmd;
2455 	vm_page_t m_ret;
2456 
2457 	/*
2458 	 * Can we allocate the pages without the number of free pages falling
2459 	 * below the lower bound for the allocation class?
2460 	 */
2461 	vmd = VM_DOMAIN(domain);
2462 	if (!vm_domain_allocate(vmd, req, npages))
2463 		return (NULL);
2464 	/*
2465 	 * Try to allocate the pages from the free page queues.
2466 	 */
2467 	vm_domain_free_lock(vmd);
2468 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2469 	    alignment, boundary);
2470 	vm_domain_free_unlock(vmd);
2471 	if (m_ret != NULL)
2472 		return (m_ret);
2473 #if VM_NRESERVLEVEL > 0
2474 	/*
2475 	 * Try to break a reservation to allocate the pages.
2476 	 */
2477 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2478 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2479 	            high, alignment, boundary);
2480 		if (m_ret != NULL)
2481 			return (m_ret);
2482 	}
2483 #endif
2484 	vm_domain_freecnt_inc(vmd, npages);
2485 	return (NULL);
2486 }
2487 
2488 vm_page_t
2489 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2490     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2491     vm_paddr_t boundary, vm_memattr_t memattr)
2492 {
2493 	struct pctrie_iter pages;
2494 	vm_page_t m, m_ret, mpred;
2495 	u_int busy_lock, flags, oflags;
2496 
2497 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2498 	KASSERT((req & ~VPAC_FLAGS) == 0,
2499 	    ("invalid request %#x", req));
2500 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2501 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2502 	    ("invalid request %#x", req));
2503 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2504 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2505 	    ("invalid request %#x", req));
2506 	VM_OBJECT_ASSERT_WLOCKED(object);
2507 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2508 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2509 	    object));
2510 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2511 
2512 	vm_page_iter_init(&pages, object);
2513 	mpred = vm_radix_iter_lookup_le(&pages, pindex);
2514 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2515 	    ("vm_page_alloc_contig: pindex already allocated"));
2516 	for (;;) {
2517 #if VM_NRESERVLEVEL > 0
2518 		/*
2519 		 * Can we allocate the pages from a reservation?
2520 		 */
2521 		if (vm_object_reserv(object) &&
2522 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2523 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2524 			break;
2525 		}
2526 #endif
2527 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2528 		    low, high, alignment, boundary)) != NULL)
2529 			break;
2530 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2531 			return (NULL);
2532 	}
2533 
2534 	/*
2535 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2536 	 */
2537 	flags = PG_ZERO;
2538 	if ((req & VM_ALLOC_NODUMP) != 0)
2539 		flags |= PG_NODUMP;
2540 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2541 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2542 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2543 	else if ((req & VM_ALLOC_SBUSY) != 0)
2544 		busy_lock = VPB_SHARERS_WORD(1);
2545 	else
2546 		busy_lock = VPB_UNBUSIED;
2547 	if ((req & VM_ALLOC_WIRED) != 0)
2548 		vm_wire_add(npages);
2549 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2550 	    memattr == VM_MEMATTR_DEFAULT)
2551 		memattr = object->memattr;
2552 	for (m = m_ret; m < &m_ret[npages]; m++) {
2553 		vm_page_dequeue(m);
2554 		vm_page_alloc_check(m);
2555 		m->a.flags = 0;
2556 		m->flags = (m->flags | PG_NODUMP) & flags;
2557 		m->busy_lock = busy_lock;
2558 		if ((req & VM_ALLOC_WIRED) != 0)
2559 			m->ref_count = 1;
2560 		m->a.act_count = 0;
2561 		m->oflags = oflags;
2562 		if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) {
2563 			if ((req & VM_ALLOC_WIRED) != 0)
2564 				vm_wire_sub(npages);
2565 			KASSERT(m->object == NULL,
2566 			    ("page %p has object", m));
2567 			mpred = m;
2568 			for (m = m_ret; m < &m_ret[npages]; m++) {
2569 				if (m <= mpred &&
2570 				    (req & VM_ALLOC_WIRED) != 0)
2571 					m->ref_count = 0;
2572 				m->oflags = VPO_UNMANAGED;
2573 				m->busy_lock = VPB_UNBUSIED;
2574 				/* Don't change PG_ZERO. */
2575 				vm_page_free_toq(m);
2576 			}
2577 			if (req & VM_ALLOC_WAITFAIL) {
2578 				VM_OBJECT_WUNLOCK(object);
2579 				vm_radix_wait();
2580 				VM_OBJECT_WLOCK(object);
2581 			}
2582 			return (NULL);
2583 		}
2584 		mpred = m;
2585 		if (memattr != VM_MEMATTR_DEFAULT)
2586 			pmap_page_set_memattr(m, memattr);
2587 		pindex++;
2588 	}
2589 	return (m_ret);
2590 }
2591 
2592 /*
2593  * Allocate a physical page that is not intended to be inserted into a VM
2594  * object.
2595  */
2596 vm_page_t
2597 vm_page_alloc_noobj_domain(int domain, int req)
2598 {
2599 	struct vm_domain *vmd;
2600 	vm_page_t m;
2601 	int flags;
2602 
2603 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2604 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2605 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2606 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO |		\
2607 			 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK)
2608 	KASSERT((req & ~VPAN_FLAGS) == 0,
2609 	    ("invalid request %#x", req));
2610 
2611 	flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) |
2612 	    ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0);
2613 	vmd = VM_DOMAIN(domain);
2614 again:
2615 	if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) {
2616 		m = vm_page_alloc_nofree_domain(domain, req);
2617 		if (m != NULL)
2618 			goto found;
2619 	}
2620 
2621 	if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2622 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2623 		    M_NOWAIT | M_NOVM);
2624 		if (m != NULL) {
2625 			flags |= PG_PCPU_CACHE;
2626 			goto found;
2627 		}
2628 	}
2629 
2630 	if (vm_domain_allocate(vmd, req, 1)) {
2631 		vm_domain_free_lock(vmd);
2632 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2633 		vm_domain_free_unlock(vmd);
2634 		if (m == NULL) {
2635 			vm_domain_freecnt_inc(vmd, 1);
2636 #if VM_NRESERVLEVEL > 0
2637 			if (vm_reserv_reclaim_inactive(domain))
2638 				goto again;
2639 #endif
2640 		}
2641 	}
2642 	if (m == NULL) {
2643 		if (vm_domain_alloc_fail(vmd, NULL, req))
2644 			goto again;
2645 		return (NULL);
2646 	}
2647 
2648 found:
2649 	vm_page_dequeue(m);
2650 	vm_page_alloc_check(m);
2651 
2652 	/*
2653 	 * Consumers should not rely on a useful default pindex value.
2654 	 */
2655 	m->pindex = 0xdeadc0dedeadc0de;
2656 	m->flags = (m->flags & PG_ZERO) | flags;
2657 	m->a.flags = 0;
2658 	m->oflags = VPO_UNMANAGED;
2659 	m->busy_lock = VPB_UNBUSIED;
2660 	if ((req & VM_ALLOC_WIRED) != 0) {
2661 		vm_wire_add(1);
2662 		m->ref_count = 1;
2663 	}
2664 
2665 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2666 		pmap_zero_page(m);
2667 
2668 	return (m);
2669 }
2670 
2671 #if VM_NRESERVLEVEL > 1
2672 #define	VM_NOFREE_IMPORT_ORDER	(VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER)
2673 #elif VM_NRESERVLEVEL > 0
2674 #define	VM_NOFREE_IMPORT_ORDER	VM_LEVEL_0_ORDER
2675 #else
2676 #define	VM_NOFREE_IMPORT_ORDER	8
2677 #endif
2678 
2679 /*
2680  * Allocate a single NOFREE page.
2681  *
2682  * This routine hands out NOFREE pages from higher-order
2683  * physical memory blocks in order to reduce memory fragmentation.
2684  * When a NOFREE for a given domain chunk is used up,
2685  * the routine will try to fetch a new one from the freelists
2686  * and discard the old one.
2687  */
2688 static vm_page_t
2689 vm_page_alloc_nofree_domain(int domain, int req)
2690 {
2691 	vm_page_t m;
2692 	struct vm_domain *vmd;
2693 	struct vm_nofreeq *nqp;
2694 
2695 	KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req));
2696 
2697 	vmd = VM_DOMAIN(domain);
2698 	nqp = &vmd->vmd_nofreeq;
2699 	vm_domain_free_lock(vmd);
2700 	if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) {
2701 		if (!vm_domain_allocate(vmd, req,
2702 		    1 << VM_NOFREE_IMPORT_ORDER)) {
2703 			vm_domain_free_unlock(vmd);
2704 			return (NULL);
2705 		}
2706 		nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
2707 		    VM_NOFREE_IMPORT_ORDER);
2708 		if (nqp->ma == NULL) {
2709 			vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER);
2710 			vm_domain_free_unlock(vmd);
2711 			return (NULL);
2712 		}
2713 		nqp->offs = 0;
2714 	}
2715 	m = &nqp->ma[nqp->offs++];
2716 	vm_domain_free_unlock(vmd);
2717 	VM_CNT_ADD(v_nofree_count, 1);
2718 
2719 	return (m);
2720 }
2721 
2722 vm_page_t
2723 vm_page_alloc_noobj(int req)
2724 {
2725 	struct vm_domainset_iter di;
2726 	vm_page_t m;
2727 	int domain;
2728 
2729 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2730 	do {
2731 		m = vm_page_alloc_noobj_domain(domain, req);
2732 		if (m != NULL)
2733 			break;
2734 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2735 
2736 	return (m);
2737 }
2738 
2739 vm_page_t
2740 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2741     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2742     vm_memattr_t memattr)
2743 {
2744 	struct vm_domainset_iter di;
2745 	vm_page_t m;
2746 	int domain;
2747 
2748 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2749 	do {
2750 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2751 		    high, alignment, boundary, memattr);
2752 		if (m != NULL)
2753 			break;
2754 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2755 
2756 	return (m);
2757 }
2758 
2759 vm_page_t
2760 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2761     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2762     vm_memattr_t memattr)
2763 {
2764 	vm_page_t m, m_ret;
2765 	u_int flags;
2766 
2767 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2768 	KASSERT((req & ~VPANC_FLAGS) == 0,
2769 	    ("invalid request %#x", req));
2770 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2771 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2772 	    ("invalid request %#x", req));
2773 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2774 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2775 	    ("invalid request %#x", req));
2776 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2777 
2778 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2779 	    low, high, alignment, boundary)) == NULL) {
2780 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2781 			return (NULL);
2782 	}
2783 
2784 	/*
2785 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2786 	 */
2787 	flags = PG_ZERO;
2788 	if ((req & VM_ALLOC_NODUMP) != 0)
2789 		flags |= PG_NODUMP;
2790 	if ((req & VM_ALLOC_WIRED) != 0)
2791 		vm_wire_add(npages);
2792 	for (m = m_ret; m < &m_ret[npages]; m++) {
2793 		vm_page_dequeue(m);
2794 		vm_page_alloc_check(m);
2795 
2796 		/*
2797 		 * Consumers should not rely on a useful default pindex value.
2798 		 */
2799 		m->pindex = 0xdeadc0dedeadc0de;
2800 		m->a.flags = 0;
2801 		m->flags = (m->flags | PG_NODUMP) & flags;
2802 		m->busy_lock = VPB_UNBUSIED;
2803 		if ((req & VM_ALLOC_WIRED) != 0)
2804 			m->ref_count = 1;
2805 		m->a.act_count = 0;
2806 		m->oflags = VPO_UNMANAGED;
2807 
2808 		/*
2809 		 * Zero the page before updating any mappings since the page is
2810 		 * not yet shared with any devices which might require the
2811 		 * non-default memory attribute.  pmap_page_set_memattr()
2812 		 * flushes data caches before returning.
2813 		 */
2814 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2815 			pmap_zero_page(m);
2816 		if (memattr != VM_MEMATTR_DEFAULT)
2817 			pmap_page_set_memattr(m, memattr);
2818 	}
2819 	return (m_ret);
2820 }
2821 
2822 /*
2823  * Check a page that has been freshly dequeued from a freelist.
2824  */
2825 static void
2826 vm_page_alloc_check(vm_page_t m)
2827 {
2828 
2829 	KASSERT(m->object == NULL, ("page %p has object", m));
2830 	KASSERT(m->a.queue == PQ_NONE &&
2831 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2832 	    ("page %p has unexpected queue %d, flags %#x",
2833 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2834 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2835 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2836 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2837 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2838 	    ("page %p has unexpected memattr %d",
2839 	    m, pmap_page_get_memattr(m)));
2840 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2841 	pmap_vm_page_alloc_check(m);
2842 }
2843 
2844 static int
2845 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2846 {
2847 	struct vm_domain *vmd;
2848 	struct vm_pgcache *pgcache;
2849 	int i;
2850 
2851 	pgcache = arg;
2852 	vmd = VM_DOMAIN(pgcache->domain);
2853 
2854 	/*
2855 	 * The page daemon should avoid creating extra memory pressure since its
2856 	 * main purpose is to replenish the store of free pages.
2857 	 */
2858 	if (vmd->vmd_severeset || curproc == pageproc ||
2859 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2860 		return (0);
2861 	domain = vmd->vmd_domain;
2862 	vm_domain_free_lock(vmd);
2863 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2864 	    (vm_page_t *)store);
2865 	vm_domain_free_unlock(vmd);
2866 	if (cnt != i)
2867 		vm_domain_freecnt_inc(vmd, cnt - i);
2868 
2869 	return (i);
2870 }
2871 
2872 static void
2873 vm_page_zone_release(void *arg, void **store, int cnt)
2874 {
2875 	struct vm_domain *vmd;
2876 	struct vm_pgcache *pgcache;
2877 	vm_page_t m;
2878 	int i;
2879 
2880 	pgcache = arg;
2881 	vmd = VM_DOMAIN(pgcache->domain);
2882 	vm_domain_free_lock(vmd);
2883 	for (i = 0; i < cnt; i++) {
2884 		m = (vm_page_t)store[i];
2885 		vm_phys_free_pages(m, 0);
2886 	}
2887 	vm_domain_free_unlock(vmd);
2888 	vm_domain_freecnt_inc(vmd, cnt);
2889 }
2890 
2891 #define	VPSC_ANY	0	/* No restrictions. */
2892 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2893 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2894 
2895 /*
2896  *	vm_page_scan_contig:
2897  *
2898  *	Scan vm_page_array[] between the specified entries "m_start" and
2899  *	"m_end" for a run of contiguous physical pages that satisfy the
2900  *	specified conditions, and return the lowest page in the run.  The
2901  *	specified "alignment" determines the alignment of the lowest physical
2902  *	page in the run.  If the specified "boundary" is non-zero, then the
2903  *	run of physical pages cannot span a physical address that is a
2904  *	multiple of "boundary".
2905  *
2906  *	"m_end" is never dereferenced, so it need not point to a vm_page
2907  *	structure within vm_page_array[].
2908  *
2909  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2910  *	span a hole (or discontiguity) in the physical address space.  Both
2911  *	"alignment" and "boundary" must be a power of two.
2912  */
2913 static vm_page_t
2914 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2915     u_long alignment, vm_paddr_t boundary, int options)
2916 {
2917 	vm_object_t object;
2918 	vm_paddr_t pa;
2919 	vm_page_t m, m_run;
2920 #if VM_NRESERVLEVEL > 0
2921 	int level;
2922 #endif
2923 	int m_inc, order, run_ext, run_len;
2924 
2925 	KASSERT(npages > 0, ("npages is 0"));
2926 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2927 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2928 	m_run = NULL;
2929 	run_len = 0;
2930 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2931 		KASSERT((m->flags & PG_MARKER) == 0,
2932 		    ("page %p is PG_MARKER", m));
2933 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2934 		    ("fictitious page %p has invalid ref count", m));
2935 
2936 		/*
2937 		 * If the current page would be the start of a run, check its
2938 		 * physical address against the end, alignment, and boundary
2939 		 * conditions.  If it doesn't satisfy these conditions, either
2940 		 * terminate the scan or advance to the next page that
2941 		 * satisfies the failed condition.
2942 		 */
2943 		if (run_len == 0) {
2944 			KASSERT(m_run == NULL, ("m_run != NULL"));
2945 			if (m + npages > m_end)
2946 				break;
2947 			pa = VM_PAGE_TO_PHYS(m);
2948 			if (!vm_addr_align_ok(pa, alignment)) {
2949 				m_inc = atop(roundup2(pa, alignment) - pa);
2950 				continue;
2951 			}
2952 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2953 				m_inc = atop(roundup2(pa, boundary) - pa);
2954 				continue;
2955 			}
2956 		} else
2957 			KASSERT(m_run != NULL, ("m_run == NULL"));
2958 
2959 retry:
2960 		m_inc = 1;
2961 		if (vm_page_wired(m))
2962 			run_ext = 0;
2963 #if VM_NRESERVLEVEL > 0
2964 		else if ((level = vm_reserv_level(m)) >= 0 &&
2965 		    (options & VPSC_NORESERV) != 0) {
2966 			run_ext = 0;
2967 			/* Advance to the end of the reservation. */
2968 			pa = VM_PAGE_TO_PHYS(m);
2969 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2970 			    pa);
2971 		}
2972 #endif
2973 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2974 			/*
2975 			 * The page is considered eligible for relocation if
2976 			 * and only if it could be laundered or reclaimed by
2977 			 * the page daemon.
2978 			 */
2979 			VM_OBJECT_RLOCK(object);
2980 			if (object != m->object) {
2981 				VM_OBJECT_RUNLOCK(object);
2982 				goto retry;
2983 			}
2984 			/* Don't care: PG_NODUMP, PG_ZERO. */
2985 			if ((object->flags & OBJ_SWAP) == 0 &&
2986 			    object->type != OBJT_VNODE) {
2987 				run_ext = 0;
2988 #if VM_NRESERVLEVEL > 0
2989 			} else if ((options & VPSC_NOSUPER) != 0 &&
2990 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2991 				run_ext = 0;
2992 				/* Advance to the end of the superpage. */
2993 				pa = VM_PAGE_TO_PHYS(m);
2994 				m_inc = atop(roundup2(pa + 1,
2995 				    vm_reserv_size(level)) - pa);
2996 #endif
2997 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2998 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2999 				/*
3000 				 * The page is allocated but eligible for
3001 				 * relocation.  Extend the current run by one
3002 				 * page.
3003 				 */
3004 				KASSERT(pmap_page_get_memattr(m) ==
3005 				    VM_MEMATTR_DEFAULT,
3006 				    ("page %p has an unexpected memattr", m));
3007 				KASSERT((m->oflags & (VPO_SWAPINPROG |
3008 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
3009 				    ("page %p has unexpected oflags", m));
3010 				/* Don't care: PGA_NOSYNC. */
3011 				run_ext = 1;
3012 			} else
3013 				run_ext = 0;
3014 			VM_OBJECT_RUNLOCK(object);
3015 #if VM_NRESERVLEVEL > 0
3016 		} else if (level >= 0) {
3017 			/*
3018 			 * The page is reserved but not yet allocated.  In
3019 			 * other words, it is still free.  Extend the current
3020 			 * run by one page.
3021 			 */
3022 			run_ext = 1;
3023 #endif
3024 		} else if ((order = m->order) < VM_NFREEORDER) {
3025 			/*
3026 			 * The page is enqueued in the physical memory
3027 			 * allocator's free page queues.  Moreover, it is the
3028 			 * first page in a power-of-two-sized run of
3029 			 * contiguous free pages.  Add these pages to the end
3030 			 * of the current run, and jump ahead.
3031 			 */
3032 			run_ext = 1 << order;
3033 			m_inc = 1 << order;
3034 		} else {
3035 			/*
3036 			 * Skip the page for one of the following reasons: (1)
3037 			 * It is enqueued in the physical memory allocator's
3038 			 * free page queues.  However, it is not the first
3039 			 * page in a run of contiguous free pages.  (This case
3040 			 * rarely occurs because the scan is performed in
3041 			 * ascending order.) (2) It is not reserved, and it is
3042 			 * transitioning from free to allocated.  (Conversely,
3043 			 * the transition from allocated to free for managed
3044 			 * pages is blocked by the page busy lock.) (3) It is
3045 			 * allocated but not contained by an object and not
3046 			 * wired, e.g., allocated by Xen's balloon driver.
3047 			 */
3048 			run_ext = 0;
3049 		}
3050 
3051 		/*
3052 		 * Extend or reset the current run of pages.
3053 		 */
3054 		if (run_ext > 0) {
3055 			if (run_len == 0)
3056 				m_run = m;
3057 			run_len += run_ext;
3058 		} else {
3059 			if (run_len > 0) {
3060 				m_run = NULL;
3061 				run_len = 0;
3062 			}
3063 		}
3064 	}
3065 	if (run_len >= npages)
3066 		return (m_run);
3067 	return (NULL);
3068 }
3069 
3070 /*
3071  *	vm_page_reclaim_run:
3072  *
3073  *	Try to relocate each of the allocated virtual pages within the
3074  *	specified run of physical pages to a new physical address.  Free the
3075  *	physical pages underlying the relocated virtual pages.  A virtual page
3076  *	is relocatable if and only if it could be laundered or reclaimed by
3077  *	the page daemon.  Whenever possible, a virtual page is relocated to a
3078  *	physical address above "high".
3079  *
3080  *	Returns 0 if every physical page within the run was already free or
3081  *	just freed by a successful relocation.  Otherwise, returns a non-zero
3082  *	value indicating why the last attempt to relocate a virtual page was
3083  *	unsuccessful.
3084  *
3085  *	"req_class" must be an allocation class.
3086  */
3087 static int
3088 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
3089     vm_paddr_t high)
3090 {
3091 	struct vm_domain *vmd;
3092 	struct spglist free;
3093 	vm_object_t object;
3094 	vm_paddr_t pa;
3095 	vm_page_t m, m_end, m_new;
3096 	int error, order, req;
3097 
3098 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
3099 	    ("req_class is not an allocation class"));
3100 	SLIST_INIT(&free);
3101 	error = 0;
3102 	m = m_run;
3103 	m_end = m_run + npages;
3104 	for (; error == 0 && m < m_end; m++) {
3105 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
3106 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
3107 
3108 		/*
3109 		 * Racily check for wirings.  Races are handled once the object
3110 		 * lock is held and the page is unmapped.
3111 		 */
3112 		if (vm_page_wired(m))
3113 			error = EBUSY;
3114 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
3115 			/*
3116 			 * The page is relocated if and only if it could be
3117 			 * laundered or reclaimed by the page daemon.
3118 			 */
3119 			VM_OBJECT_WLOCK(object);
3120 			/* Don't care: PG_NODUMP, PG_ZERO. */
3121 			if (m->object != object ||
3122 			    ((object->flags & OBJ_SWAP) == 0 &&
3123 			    object->type != OBJT_VNODE))
3124 				error = EINVAL;
3125 			else if (object->memattr != VM_MEMATTR_DEFAULT)
3126 				error = EINVAL;
3127 			else if (vm_page_queue(m) != PQ_NONE &&
3128 			    vm_page_tryxbusy(m) != 0) {
3129 				if (vm_page_wired(m)) {
3130 					vm_page_xunbusy(m);
3131 					error = EBUSY;
3132 					goto unlock;
3133 				}
3134 				KASSERT(pmap_page_get_memattr(m) ==
3135 				    VM_MEMATTR_DEFAULT,
3136 				    ("page %p has an unexpected memattr", m));
3137 				KASSERT(m->oflags == 0,
3138 				    ("page %p has unexpected oflags", m));
3139 				/* Don't care: PGA_NOSYNC. */
3140 				if (!vm_page_none_valid(m)) {
3141 					/*
3142 					 * First, try to allocate a new page
3143 					 * that is above "high".  Failing
3144 					 * that, try to allocate a new page
3145 					 * that is below "m_run".  Allocate
3146 					 * the new page between the end of
3147 					 * "m_run" and "high" only as a last
3148 					 * resort.
3149 					 */
3150 					req = req_class;
3151 					if ((m->flags & PG_NODUMP) != 0)
3152 						req |= VM_ALLOC_NODUMP;
3153 					if (trunc_page(high) !=
3154 					    ~(vm_paddr_t)PAGE_MASK) {
3155 						m_new =
3156 						    vm_page_alloc_noobj_contig(
3157 						    req, 1, round_page(high),
3158 						    ~(vm_paddr_t)0, PAGE_SIZE,
3159 						    0, VM_MEMATTR_DEFAULT);
3160 					} else
3161 						m_new = NULL;
3162 					if (m_new == NULL) {
3163 						pa = VM_PAGE_TO_PHYS(m_run);
3164 						m_new =
3165 						    vm_page_alloc_noobj_contig(
3166 						    req, 1, 0, pa - 1,
3167 						    PAGE_SIZE, 0,
3168 						    VM_MEMATTR_DEFAULT);
3169 					}
3170 					if (m_new == NULL) {
3171 						pa += ptoa(npages);
3172 						m_new =
3173 						    vm_page_alloc_noobj_contig(
3174 						    req, 1, pa, high, PAGE_SIZE,
3175 						    0, VM_MEMATTR_DEFAULT);
3176 					}
3177 					if (m_new == NULL) {
3178 						vm_page_xunbusy(m);
3179 						error = ENOMEM;
3180 						goto unlock;
3181 					}
3182 
3183 					/*
3184 					 * Unmap the page and check for new
3185 					 * wirings that may have been acquired
3186 					 * through a pmap lookup.
3187 					 */
3188 					if (object->ref_count != 0 &&
3189 					    !vm_page_try_remove_all(m)) {
3190 						vm_page_xunbusy(m);
3191 						vm_page_free(m_new);
3192 						error = EBUSY;
3193 						goto unlock;
3194 					}
3195 
3196 					/*
3197 					 * Replace "m" with the new page.  For
3198 					 * vm_page_replace(), "m" must be busy
3199 					 * and dequeued.  Finally, change "m"
3200 					 * as if vm_page_free() was called.
3201 					 */
3202 					m_new->a.flags = m->a.flags &
3203 					    ~PGA_QUEUE_STATE_MASK;
3204 					KASSERT(m_new->oflags == VPO_UNMANAGED,
3205 					    ("page %p is managed", m_new));
3206 					m_new->oflags = 0;
3207 					pmap_copy_page(m, m_new);
3208 					m_new->valid = m->valid;
3209 					m_new->dirty = m->dirty;
3210 					m->flags &= ~PG_ZERO;
3211 					vm_page_dequeue(m);
3212 					if (vm_page_replace_hold(m_new, object,
3213 					    m->pindex, m) &&
3214 					    vm_page_free_prep(m))
3215 						SLIST_INSERT_HEAD(&free, m,
3216 						    plinks.s.ss);
3217 
3218 					/*
3219 					 * The new page must be deactivated
3220 					 * before the object is unlocked.
3221 					 */
3222 					vm_page_deactivate(m_new);
3223 				} else {
3224 					m->flags &= ~PG_ZERO;
3225 					vm_page_dequeue(m);
3226 					if (vm_page_free_prep(m))
3227 						SLIST_INSERT_HEAD(&free, m,
3228 						    plinks.s.ss);
3229 					KASSERT(m->dirty == 0,
3230 					    ("page %p is dirty", m));
3231 				}
3232 			} else
3233 				error = EBUSY;
3234 unlock:
3235 			VM_OBJECT_WUNLOCK(object);
3236 		} else {
3237 			MPASS(vm_page_domain(m) == domain);
3238 			vmd = VM_DOMAIN(domain);
3239 			vm_domain_free_lock(vmd);
3240 			order = m->order;
3241 			if (order < VM_NFREEORDER) {
3242 				/*
3243 				 * The page is enqueued in the physical memory
3244 				 * allocator's free page queues.  Moreover, it
3245 				 * is the first page in a power-of-two-sized
3246 				 * run of contiguous free pages.  Jump ahead
3247 				 * to the last page within that run, and
3248 				 * continue from there.
3249 				 */
3250 				m += (1 << order) - 1;
3251 			}
3252 #if VM_NRESERVLEVEL > 0
3253 			else if (vm_reserv_is_page_free(m))
3254 				order = 0;
3255 #endif
3256 			vm_domain_free_unlock(vmd);
3257 			if (order == VM_NFREEORDER)
3258 				error = EINVAL;
3259 		}
3260 	}
3261 	if ((m = SLIST_FIRST(&free)) != NULL) {
3262 		int cnt;
3263 
3264 		vmd = VM_DOMAIN(domain);
3265 		cnt = 0;
3266 		vm_domain_free_lock(vmd);
3267 		do {
3268 			MPASS(vm_page_domain(m) == domain);
3269 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3270 			vm_phys_free_pages(m, 0);
3271 			cnt++;
3272 		} while ((m = SLIST_FIRST(&free)) != NULL);
3273 		vm_domain_free_unlock(vmd);
3274 		vm_domain_freecnt_inc(vmd, cnt);
3275 	}
3276 	return (error);
3277 }
3278 
3279 #define	NRUNS	16
3280 
3281 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3282 
3283 #define	MIN_RECLAIM	8
3284 
3285 /*
3286  *	vm_page_reclaim_contig:
3287  *
3288  *	Reclaim allocated, contiguous physical memory satisfying the specified
3289  *	conditions by relocating the virtual pages using that physical memory.
3290  *	Returns 0 if reclamation is successful, ERANGE if the specified domain
3291  *	can't possibly satisfy the reclamation request, or ENOMEM if not
3292  *	currently able to reclaim the requested number of pages.  Since
3293  *	relocation requires the allocation of physical pages, reclamation may
3294  *	fail with ENOMEM due to a shortage of free pages.  When reclamation
3295  *	fails in this manner, callers are expected to perform vm_wait() before
3296  *	retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3297  *
3298  *	The caller must always specify an allocation class through "req".
3299  *
3300  *	allocation classes:
3301  *	VM_ALLOC_NORMAL		normal process request
3302  *	VM_ALLOC_SYSTEM		system *really* needs a page
3303  *	VM_ALLOC_INTERRUPT	interrupt time request
3304  *
3305  *	The optional allocation flags are ignored.
3306  *
3307  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3308  *	must be a power of two.
3309  */
3310 int
3311 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3312     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3313     int desired_runs)
3314 {
3315 	struct vm_domain *vmd;
3316 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3317 	u_long count, minalign, reclaimed;
3318 	int error, i, min_reclaim, nruns, options, req_class;
3319 	int segind, start_segind;
3320 	int ret;
3321 
3322 	KASSERT(npages > 0, ("npages is 0"));
3323 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3324 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3325 
3326 	ret = ENOMEM;
3327 
3328 	/*
3329 	 * If the caller wants to reclaim multiple runs, try to allocate
3330 	 * space to store the runs.  If that fails, fall back to the old
3331 	 * behavior of just reclaiming MIN_RECLAIM pages.
3332 	 */
3333 	if (desired_runs > 1)
3334 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3335 		    M_TEMP, M_NOWAIT);
3336 	else
3337 		m_runs = NULL;
3338 
3339 	if (m_runs == NULL) {
3340 		m_runs = _m_runs;
3341 		nruns = NRUNS;
3342 	} else {
3343 		nruns = NRUNS + desired_runs - 1;
3344 	}
3345 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3346 
3347 	/*
3348 	 * The caller will attempt an allocation after some runs have been
3349 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3350 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3351 	 * power of two or maximum chunk size, and ensure that each run is
3352 	 * suitably aligned.
3353 	 */
3354 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3355 	npages = roundup2(npages, minalign);
3356 	if (alignment < ptoa(minalign))
3357 		alignment = ptoa(minalign);
3358 
3359 	/*
3360 	 * The page daemon is allowed to dig deeper into the free page list.
3361 	 */
3362 	req_class = req & VM_ALLOC_CLASS_MASK;
3363 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3364 		req_class = VM_ALLOC_SYSTEM;
3365 
3366 	start_segind = vm_phys_lookup_segind(low);
3367 
3368 	/*
3369 	 * Return if the number of free pages cannot satisfy the requested
3370 	 * allocation.
3371 	 */
3372 	vmd = VM_DOMAIN(domain);
3373 	count = vmd->vmd_free_count;
3374 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3375 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3376 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3377 		goto done;
3378 
3379 	/*
3380 	 * Scan up to three times, relaxing the restrictions ("options") on
3381 	 * the reclamation of reservations and superpages each time.
3382 	 */
3383 	for (options = VPSC_NORESERV;;) {
3384 		bool phys_range_exists = false;
3385 
3386 		/*
3387 		 * Find the highest runs that satisfy the given constraints
3388 		 * and restrictions, and record them in "m_runs".
3389 		 */
3390 		count = 0;
3391 		segind = start_segind;
3392 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3393 		    npages, low, high)) != -1) {
3394 			phys_range_exists = true;
3395 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3396 			    bounds[1], alignment, boundary, options))) {
3397 				bounds[0] = m_run + npages;
3398 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3399 				count++;
3400 			}
3401 			segind++;
3402 		}
3403 
3404 		if (!phys_range_exists) {
3405 			ret = ERANGE;
3406 			goto done;
3407 		}
3408 
3409 		/*
3410 		 * Reclaim the highest runs in LIFO (descending) order until
3411 		 * the number of reclaimed pages, "reclaimed", is at least
3412 		 * "min_reclaim".  Reset "reclaimed" each time because each
3413 		 * reclamation is idempotent, and runs will (likely) recur
3414 		 * from one scan to the next as restrictions are relaxed.
3415 		 */
3416 		reclaimed = 0;
3417 		for (i = 0; count > 0 && i < nruns; i++) {
3418 			count--;
3419 			m_run = m_runs[RUN_INDEX(count, nruns)];
3420 			error = vm_page_reclaim_run(req_class, domain, npages,
3421 			    m_run, high);
3422 			if (error == 0) {
3423 				reclaimed += npages;
3424 				if (reclaimed >= min_reclaim) {
3425 					ret = 0;
3426 					goto done;
3427 				}
3428 			}
3429 		}
3430 
3431 		/*
3432 		 * Either relax the restrictions on the next scan or return if
3433 		 * the last scan had no restrictions.
3434 		 */
3435 		if (options == VPSC_NORESERV)
3436 			options = VPSC_NOSUPER;
3437 		else if (options == VPSC_NOSUPER)
3438 			options = VPSC_ANY;
3439 		else if (options == VPSC_ANY) {
3440 			if (reclaimed != 0)
3441 				ret = 0;
3442 			goto done;
3443 		}
3444 	}
3445 done:
3446 	if (m_runs != _m_runs)
3447 		free(m_runs, M_TEMP);
3448 	return (ret);
3449 }
3450 
3451 int
3452 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3453     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3454 {
3455 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3456 	    alignment, boundary, 1));
3457 }
3458 
3459 int
3460 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3461     u_long alignment, vm_paddr_t boundary)
3462 {
3463 	struct vm_domainset_iter di;
3464 	int domain, ret, status;
3465 
3466 	ret = ERANGE;
3467 
3468 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3469 	do {
3470 		status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3471 		    high, alignment, boundary);
3472 		if (status == 0)
3473 			return (0);
3474 		else if (status == ERANGE)
3475 			vm_domainset_iter_ignore(&di, domain);
3476 		else {
3477 			KASSERT(status == ENOMEM, ("Unrecognized error %d "
3478 			    "from vm_page_reclaim_contig_domain()", status));
3479 			ret = ENOMEM;
3480 		}
3481 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3482 
3483 	return (ret);
3484 }
3485 
3486 /*
3487  * Set the domain in the appropriate page level domainset.
3488  */
3489 void
3490 vm_domain_set(struct vm_domain *vmd)
3491 {
3492 
3493 	mtx_lock(&vm_domainset_lock);
3494 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3495 		vmd->vmd_minset = 1;
3496 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3497 	}
3498 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3499 		vmd->vmd_severeset = 1;
3500 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3501 	}
3502 	mtx_unlock(&vm_domainset_lock);
3503 }
3504 
3505 /*
3506  * Clear the domain from the appropriate page level domainset.
3507  */
3508 void
3509 vm_domain_clear(struct vm_domain *vmd)
3510 {
3511 
3512 	mtx_lock(&vm_domainset_lock);
3513 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3514 		vmd->vmd_minset = 0;
3515 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3516 		if (vm_min_waiters != 0) {
3517 			vm_min_waiters = 0;
3518 			wakeup(&vm_min_domains);
3519 		}
3520 	}
3521 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3522 		vmd->vmd_severeset = 0;
3523 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3524 		if (vm_severe_waiters != 0) {
3525 			vm_severe_waiters = 0;
3526 			wakeup(&vm_severe_domains);
3527 		}
3528 	}
3529 
3530 	/*
3531 	 * If pageout daemon needs pages, then tell it that there are
3532 	 * some free.
3533 	 */
3534 	if (vmd->vmd_pageout_pages_needed &&
3535 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3536 		wakeup(&vmd->vmd_pageout_pages_needed);
3537 		vmd->vmd_pageout_pages_needed = 0;
3538 	}
3539 
3540 	/* See comments in vm_wait_doms(). */
3541 	if (vm_pageproc_waiters) {
3542 		vm_pageproc_waiters = 0;
3543 		wakeup(&vm_pageproc_waiters);
3544 	}
3545 	mtx_unlock(&vm_domainset_lock);
3546 }
3547 
3548 /*
3549  * Wait for free pages to exceed the min threshold globally.
3550  */
3551 void
3552 vm_wait_min(void)
3553 {
3554 
3555 	mtx_lock(&vm_domainset_lock);
3556 	while (vm_page_count_min()) {
3557 		vm_min_waiters++;
3558 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3559 	}
3560 	mtx_unlock(&vm_domainset_lock);
3561 }
3562 
3563 /*
3564  * Wait for free pages to exceed the severe threshold globally.
3565  */
3566 void
3567 vm_wait_severe(void)
3568 {
3569 
3570 	mtx_lock(&vm_domainset_lock);
3571 	while (vm_page_count_severe()) {
3572 		vm_severe_waiters++;
3573 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3574 		    "vmwait", 0);
3575 	}
3576 	mtx_unlock(&vm_domainset_lock);
3577 }
3578 
3579 u_int
3580 vm_wait_count(void)
3581 {
3582 
3583 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3584 }
3585 
3586 int
3587 vm_wait_doms(const domainset_t *wdoms, int mflags)
3588 {
3589 	int error;
3590 
3591 	error = 0;
3592 
3593 	/*
3594 	 * We use racey wakeup synchronization to avoid expensive global
3595 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3596 	 * To handle this, we only sleep for one tick in this instance.  It
3597 	 * is expected that most allocations for the pageproc will come from
3598 	 * kmem or vm_page_grab* which will use the more specific and
3599 	 * race-free vm_wait_domain().
3600 	 */
3601 	if (curproc == pageproc) {
3602 		mtx_lock(&vm_domainset_lock);
3603 		vm_pageproc_waiters++;
3604 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3605 		    PVM | PDROP | mflags, "pageprocwait", 1);
3606 	} else {
3607 		/*
3608 		 * XXX Ideally we would wait only until the allocation could
3609 		 * be satisfied.  This condition can cause new allocators to
3610 		 * consume all freed pages while old allocators wait.
3611 		 */
3612 		mtx_lock(&vm_domainset_lock);
3613 		if (vm_page_count_min_set(wdoms)) {
3614 			if (pageproc == NULL)
3615 				panic("vm_wait in early boot");
3616 			vm_min_waiters++;
3617 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3618 			    PVM | PDROP | mflags, "vmwait", 0);
3619 		} else
3620 			mtx_unlock(&vm_domainset_lock);
3621 	}
3622 	return (error);
3623 }
3624 
3625 /*
3626  *	vm_wait_domain:
3627  *
3628  *	Sleep until free pages are available for allocation.
3629  *	- Called in various places after failed memory allocations.
3630  */
3631 void
3632 vm_wait_domain(int domain)
3633 {
3634 	struct vm_domain *vmd;
3635 	domainset_t wdom;
3636 
3637 	vmd = VM_DOMAIN(domain);
3638 	vm_domain_free_assert_unlocked(vmd);
3639 
3640 	if (curproc == pageproc) {
3641 		mtx_lock(&vm_domainset_lock);
3642 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3643 			vmd->vmd_pageout_pages_needed = 1;
3644 			msleep(&vmd->vmd_pageout_pages_needed,
3645 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3646 		} else
3647 			mtx_unlock(&vm_domainset_lock);
3648 	} else {
3649 		DOMAINSET_ZERO(&wdom);
3650 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3651 		vm_wait_doms(&wdom, 0);
3652 	}
3653 }
3654 
3655 static int
3656 vm_wait_flags(vm_object_t obj, int mflags)
3657 {
3658 	struct domainset *d;
3659 
3660 	d = NULL;
3661 
3662 	/*
3663 	 * Carefully fetch pointers only once: the struct domainset
3664 	 * itself is ummutable but the pointer might change.
3665 	 */
3666 	if (obj != NULL)
3667 		d = obj->domain.dr_policy;
3668 	if (d == NULL)
3669 		d = curthread->td_domain.dr_policy;
3670 
3671 	return (vm_wait_doms(&d->ds_mask, mflags));
3672 }
3673 
3674 /*
3675  *	vm_wait:
3676  *
3677  *	Sleep until free pages are available for allocation in the
3678  *	affinity domains of the obj.  If obj is NULL, the domain set
3679  *	for the calling thread is used.
3680  *	Called in various places after failed memory allocations.
3681  */
3682 void
3683 vm_wait(vm_object_t obj)
3684 {
3685 	(void)vm_wait_flags(obj, 0);
3686 }
3687 
3688 int
3689 vm_wait_intr(vm_object_t obj)
3690 {
3691 	return (vm_wait_flags(obj, PCATCH));
3692 }
3693 
3694 /*
3695  *	vm_domain_alloc_fail:
3696  *
3697  *	Called when a page allocation function fails.  Informs the
3698  *	pagedaemon and performs the requested wait.  Requires the
3699  *	domain_free and object lock on entry.  Returns with the
3700  *	object lock held and free lock released.  Returns an error when
3701  *	retry is necessary.
3702  *
3703  */
3704 static int
3705 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3706 {
3707 
3708 	vm_domain_free_assert_unlocked(vmd);
3709 
3710 	atomic_add_int(&vmd->vmd_pageout_deficit,
3711 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3712 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3713 		if (object != NULL)
3714 			VM_OBJECT_WUNLOCK(object);
3715 		vm_wait_domain(vmd->vmd_domain);
3716 		if (object != NULL)
3717 			VM_OBJECT_WLOCK(object);
3718 		if (req & VM_ALLOC_WAITOK)
3719 			return (EAGAIN);
3720 	}
3721 
3722 	return (0);
3723 }
3724 
3725 /*
3726  *	vm_waitpfault:
3727  *
3728  *	Sleep until free pages are available for allocation.
3729  *	- Called only in vm_fault so that processes page faulting
3730  *	  can be easily tracked.
3731  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3732  *	  processes will be able to grab memory first.  Do not change
3733  *	  this balance without careful testing first.
3734  */
3735 void
3736 vm_waitpfault(struct domainset *dset, int timo)
3737 {
3738 
3739 	/*
3740 	 * XXX Ideally we would wait only until the allocation could
3741 	 * be satisfied.  This condition can cause new allocators to
3742 	 * consume all freed pages while old allocators wait.
3743 	 */
3744 	mtx_lock(&vm_domainset_lock);
3745 	if (vm_page_count_min_set(&dset->ds_mask)) {
3746 		vm_min_waiters++;
3747 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3748 		    "pfault", timo);
3749 	} else
3750 		mtx_unlock(&vm_domainset_lock);
3751 }
3752 
3753 static struct vm_pagequeue *
3754 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3755 {
3756 
3757 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3758 }
3759 
3760 #ifdef INVARIANTS
3761 static struct vm_pagequeue *
3762 vm_page_pagequeue(vm_page_t m)
3763 {
3764 
3765 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3766 }
3767 #endif
3768 
3769 static __always_inline bool
3770 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3771 {
3772 	vm_page_astate_t tmp;
3773 
3774 	tmp = *old;
3775 	do {
3776 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3777 			return (true);
3778 		counter_u64_add(pqstate_commit_retries, 1);
3779 	} while (old->_bits == tmp._bits);
3780 
3781 	return (false);
3782 }
3783 
3784 /*
3785  * Do the work of committing a queue state update that moves the page out of
3786  * its current queue.
3787  */
3788 static bool
3789 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3790     vm_page_astate_t *old, vm_page_astate_t new)
3791 {
3792 	vm_page_t next;
3793 
3794 	vm_pagequeue_assert_locked(pq);
3795 	KASSERT(vm_page_pagequeue(m) == pq,
3796 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3797 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3798 	    ("%s: invalid queue indices %d %d",
3799 	    __func__, old->queue, new.queue));
3800 
3801 	/*
3802 	 * Once the queue index of the page changes there is nothing
3803 	 * synchronizing with further updates to the page's physical
3804 	 * queue state.  Therefore we must speculatively remove the page
3805 	 * from the queue now and be prepared to roll back if the queue
3806 	 * state update fails.  If the page is not physically enqueued then
3807 	 * we just update its queue index.
3808 	 */
3809 	if ((old->flags & PGA_ENQUEUED) != 0) {
3810 		new.flags &= ~PGA_ENQUEUED;
3811 		next = TAILQ_NEXT(m, plinks.q);
3812 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3813 		vm_pagequeue_cnt_dec(pq);
3814 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3815 			if (next == NULL)
3816 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3817 			else
3818 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3819 			vm_pagequeue_cnt_inc(pq);
3820 			return (false);
3821 		} else {
3822 			return (true);
3823 		}
3824 	} else {
3825 		return (vm_page_pqstate_fcmpset(m, old, new));
3826 	}
3827 }
3828 
3829 static bool
3830 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3831     vm_page_astate_t new)
3832 {
3833 	struct vm_pagequeue *pq;
3834 	vm_page_astate_t as;
3835 	bool ret;
3836 
3837 	pq = _vm_page_pagequeue(m, old->queue);
3838 
3839 	/*
3840 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3841 	 * corresponding page queue lock is held.
3842 	 */
3843 	vm_pagequeue_lock(pq);
3844 	as = vm_page_astate_load(m);
3845 	if (__predict_false(as._bits != old->_bits)) {
3846 		*old = as;
3847 		ret = false;
3848 	} else {
3849 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3850 	}
3851 	vm_pagequeue_unlock(pq);
3852 	return (ret);
3853 }
3854 
3855 /*
3856  * Commit a queue state update that enqueues or requeues a page.
3857  */
3858 static bool
3859 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3860     vm_page_astate_t *old, vm_page_astate_t new)
3861 {
3862 	struct vm_domain *vmd;
3863 
3864 	vm_pagequeue_assert_locked(pq);
3865 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3866 	    ("%s: invalid queue indices %d %d",
3867 	    __func__, old->queue, new.queue));
3868 
3869 	new.flags |= PGA_ENQUEUED;
3870 	if (!vm_page_pqstate_fcmpset(m, old, new))
3871 		return (false);
3872 
3873 	if ((old->flags & PGA_ENQUEUED) != 0)
3874 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3875 	else
3876 		vm_pagequeue_cnt_inc(pq);
3877 
3878 	/*
3879 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3880 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3881 	 * applied, even if it was set first.
3882 	 */
3883 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3884 		vmd = vm_pagequeue_domain(m);
3885 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3886 		    ("%s: invalid page queue for page %p", __func__, m));
3887 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3888 	} else {
3889 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3890 	}
3891 	return (true);
3892 }
3893 
3894 /*
3895  * Commit a queue state update that encodes a request for a deferred queue
3896  * operation.
3897  */
3898 static bool
3899 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3900     vm_page_astate_t new)
3901 {
3902 
3903 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3904 	    ("%s: invalid state, queue %d flags %x",
3905 	    __func__, new.queue, new.flags));
3906 
3907 	if (old->_bits != new._bits &&
3908 	    !vm_page_pqstate_fcmpset(m, old, new))
3909 		return (false);
3910 	vm_page_pqbatch_submit(m, new.queue);
3911 	return (true);
3912 }
3913 
3914 /*
3915  * A generic queue state update function.  This handles more cases than the
3916  * specialized functions above.
3917  */
3918 bool
3919 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3920 {
3921 
3922 	if (old->_bits == new._bits)
3923 		return (true);
3924 
3925 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3926 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3927 			return (false);
3928 		if (new.queue != PQ_NONE)
3929 			vm_page_pqbatch_submit(m, new.queue);
3930 	} else {
3931 		if (!vm_page_pqstate_fcmpset(m, old, new))
3932 			return (false);
3933 		if (new.queue != PQ_NONE &&
3934 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3935 			vm_page_pqbatch_submit(m, new.queue);
3936 	}
3937 	return (true);
3938 }
3939 
3940 /*
3941  * Apply deferred queue state updates to a page.
3942  */
3943 static inline void
3944 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3945 {
3946 	vm_page_astate_t new, old;
3947 
3948 	CRITICAL_ASSERT(curthread);
3949 	vm_pagequeue_assert_locked(pq);
3950 	KASSERT(queue < PQ_COUNT,
3951 	    ("%s: invalid queue index %d", __func__, queue));
3952 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3953 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3954 
3955 	for (old = vm_page_astate_load(m);;) {
3956 		if (__predict_false(old.queue != queue ||
3957 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3958 			counter_u64_add(queue_nops, 1);
3959 			break;
3960 		}
3961 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3962 		    ("%s: page %p is unmanaged", __func__, m));
3963 
3964 		new = old;
3965 		if ((old.flags & PGA_DEQUEUE) != 0) {
3966 			new.flags &= ~PGA_QUEUE_OP_MASK;
3967 			new.queue = PQ_NONE;
3968 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3969 			    m, &old, new))) {
3970 				counter_u64_add(queue_ops, 1);
3971 				break;
3972 			}
3973 		} else {
3974 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3975 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3976 			    m, &old, new))) {
3977 				counter_u64_add(queue_ops, 1);
3978 				break;
3979 			}
3980 		}
3981 	}
3982 }
3983 
3984 static void
3985 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3986     uint8_t queue)
3987 {
3988 	int i;
3989 
3990 	for (i = 0; i < bq->bq_cnt; i++)
3991 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3992 	vm_batchqueue_init(bq);
3993 }
3994 
3995 /*
3996  *	vm_page_pqbatch_submit:		[ internal use only ]
3997  *
3998  *	Enqueue a page in the specified page queue's batched work queue.
3999  *	The caller must have encoded the requested operation in the page
4000  *	structure's a.flags field.
4001  */
4002 void
4003 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
4004 {
4005 	struct vm_batchqueue *bq;
4006 	struct vm_pagequeue *pq;
4007 	int domain, slots_remaining;
4008 
4009 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
4010 
4011 	domain = vm_page_domain(m);
4012 	critical_enter();
4013 	bq = DPCPU_PTR(pqbatch[domain][queue]);
4014 	slots_remaining = vm_batchqueue_insert(bq, m);
4015 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
4016 		/* keep building the bq */
4017 		critical_exit();
4018 		return;
4019 	} else if (slots_remaining > 0 ) {
4020 		/* Try to process the bq if we can get the lock */
4021 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
4022 		if (vm_pagequeue_trylock(pq)) {
4023 			vm_pqbatch_process(pq, bq, queue);
4024 			vm_pagequeue_unlock(pq);
4025 		}
4026 		critical_exit();
4027 		return;
4028 	}
4029 	critical_exit();
4030 
4031 	/* if we make it here, the bq is full so wait for the lock */
4032 
4033 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
4034 	vm_pagequeue_lock(pq);
4035 	critical_enter();
4036 	bq = DPCPU_PTR(pqbatch[domain][queue]);
4037 	vm_pqbatch_process(pq, bq, queue);
4038 	vm_pqbatch_process_page(pq, m, queue);
4039 	vm_pagequeue_unlock(pq);
4040 	critical_exit();
4041 }
4042 
4043 /*
4044  *	vm_page_pqbatch_drain:		[ internal use only ]
4045  *
4046  *	Force all per-CPU page queue batch queues to be drained.  This is
4047  *	intended for use in severe memory shortages, to ensure that pages
4048  *	do not remain stuck in the batch queues.
4049  */
4050 void
4051 vm_page_pqbatch_drain(void)
4052 {
4053 	struct thread *td;
4054 	struct vm_domain *vmd;
4055 	struct vm_pagequeue *pq;
4056 	int cpu, domain, queue;
4057 
4058 	td = curthread;
4059 	CPU_FOREACH(cpu) {
4060 		thread_lock(td);
4061 		sched_bind(td, cpu);
4062 		thread_unlock(td);
4063 
4064 		for (domain = 0; domain < vm_ndomains; domain++) {
4065 			vmd = VM_DOMAIN(domain);
4066 			for (queue = 0; queue < PQ_COUNT; queue++) {
4067 				pq = &vmd->vmd_pagequeues[queue];
4068 				vm_pagequeue_lock(pq);
4069 				critical_enter();
4070 				vm_pqbatch_process(pq,
4071 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
4072 				critical_exit();
4073 				vm_pagequeue_unlock(pq);
4074 			}
4075 		}
4076 	}
4077 	thread_lock(td);
4078 	sched_unbind(td);
4079 	thread_unlock(td);
4080 }
4081 
4082 /*
4083  *	vm_page_dequeue_deferred:	[ internal use only ]
4084  *
4085  *	Request removal of the given page from its current page
4086  *	queue.  Physical removal from the queue may be deferred
4087  *	indefinitely.
4088  */
4089 void
4090 vm_page_dequeue_deferred(vm_page_t m)
4091 {
4092 	vm_page_astate_t new, old;
4093 
4094 	old = vm_page_astate_load(m);
4095 	do {
4096 		if (old.queue == PQ_NONE) {
4097 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4098 			    ("%s: page %p has unexpected queue state",
4099 			    __func__, m));
4100 			break;
4101 		}
4102 		new = old;
4103 		new.flags |= PGA_DEQUEUE;
4104 	} while (!vm_page_pqstate_commit_request(m, &old, new));
4105 }
4106 
4107 /*
4108  *	vm_page_dequeue:
4109  *
4110  *	Remove the page from whichever page queue it's in, if any, before
4111  *	returning.
4112  */
4113 void
4114 vm_page_dequeue(vm_page_t m)
4115 {
4116 	vm_page_astate_t new, old;
4117 
4118 	old = vm_page_astate_load(m);
4119 	do {
4120 		if (old.queue == PQ_NONE) {
4121 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
4122 			    ("%s: page %p has unexpected queue state",
4123 			    __func__, m));
4124 			break;
4125 		}
4126 		new = old;
4127 		new.flags &= ~PGA_QUEUE_OP_MASK;
4128 		new.queue = PQ_NONE;
4129 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
4130 
4131 }
4132 
4133 /*
4134  * Schedule the given page for insertion into the specified page queue.
4135  * Physical insertion of the page may be deferred indefinitely.
4136  */
4137 static void
4138 vm_page_enqueue(vm_page_t m, uint8_t queue)
4139 {
4140 
4141 	KASSERT(m->a.queue == PQ_NONE &&
4142 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
4143 	    ("%s: page %p is already enqueued", __func__, m));
4144 	KASSERT(m->ref_count > 0,
4145 	    ("%s: page %p does not carry any references", __func__, m));
4146 
4147 	m->a.queue = queue;
4148 	if ((m->a.flags & PGA_REQUEUE) == 0)
4149 		vm_page_aflag_set(m, PGA_REQUEUE);
4150 	vm_page_pqbatch_submit(m, queue);
4151 }
4152 
4153 /*
4154  *	vm_page_free_prep:
4155  *
4156  *	Prepares the given page to be put on the free list,
4157  *	disassociating it from any VM object. The caller may return
4158  *	the page to the free list only if this function returns true.
4159  *
4160  *	The object, if it exists, must be locked, and then the page must
4161  *	be xbusy.  Otherwise the page must be not busied.  A managed
4162  *	page must be unmapped.
4163  */
4164 static bool
4165 vm_page_free_prep(vm_page_t m)
4166 {
4167 
4168 	/*
4169 	 * Synchronize with threads that have dropped a reference to this
4170 	 * page.
4171 	 */
4172 	atomic_thread_fence_acq();
4173 
4174 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
4175 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
4176 		uint64_t *p;
4177 		int i;
4178 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
4179 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
4180 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
4181 			    m, i, (uintmax_t)*p));
4182 	}
4183 #endif
4184 	KASSERT((m->flags & PG_NOFREE) == 0,
4185 	    ("%s: attempting to free a PG_NOFREE page", __func__));
4186 	if ((m->oflags & VPO_UNMANAGED) == 0) {
4187 		KASSERT(!pmap_page_is_mapped(m),
4188 		    ("vm_page_free_prep: freeing mapped page %p", m));
4189 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
4190 		    ("vm_page_free_prep: mapping flags set in page %p", m));
4191 	} else {
4192 		KASSERT(m->a.queue == PQ_NONE,
4193 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
4194 	}
4195 	VM_CNT_INC(v_tfree);
4196 
4197 	if (m->object != NULL) {
4198 		vm_page_radix_remove(m);
4199 		vm_page_free_object_prep(m);
4200 	} else
4201 		vm_page_assert_unbusied(m);
4202 
4203 	vm_page_busy_free(m);
4204 
4205 	/*
4206 	 * If fictitious remove object association and
4207 	 * return.
4208 	 */
4209 	if ((m->flags & PG_FICTITIOUS) != 0) {
4210 		KASSERT(m->ref_count == 1,
4211 		    ("fictitious page %p is referenced", m));
4212 		KASSERT(m->a.queue == PQ_NONE,
4213 		    ("fictitious page %p is queued", m));
4214 		return (false);
4215 	}
4216 
4217 	/*
4218 	 * Pages need not be dequeued before they are returned to the physical
4219 	 * memory allocator, but they must at least be marked for a deferred
4220 	 * dequeue.
4221 	 */
4222 	if ((m->oflags & VPO_UNMANAGED) == 0)
4223 		vm_page_dequeue_deferred(m);
4224 
4225 	m->valid = 0;
4226 	vm_page_undirty(m);
4227 
4228 	if (m->ref_count != 0)
4229 		panic("vm_page_free_prep: page %p has references", m);
4230 
4231 	/*
4232 	 * Restore the default memory attribute to the page.
4233 	 */
4234 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
4235 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
4236 
4237 #if VM_NRESERVLEVEL > 0
4238 	/*
4239 	 * Determine whether the page belongs to a reservation.  If the page was
4240 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
4241 	 * as an optimization, we avoid the check in that case.
4242 	 */
4243 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
4244 		return (false);
4245 #endif
4246 
4247 	return (true);
4248 }
4249 
4250 /*
4251  *	vm_page_free_toq:
4252  *
4253  *	Returns the given page to the free list, disassociating it
4254  *	from any VM object.
4255  *
4256  *	The object must be locked.  The page must be exclusively busied if it
4257  *	belongs to an object.
4258  */
4259 static void
4260 vm_page_free_toq(vm_page_t m)
4261 {
4262 	struct vm_domain *vmd;
4263 	uma_zone_t zone;
4264 
4265 	if (!vm_page_free_prep(m))
4266 		return;
4267 
4268 	vmd = vm_pagequeue_domain(m);
4269 	zone = vmd->vmd_pgcache[m->pool].zone;
4270 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4271 		uma_zfree(zone, m);
4272 		return;
4273 	}
4274 	vm_domain_free_lock(vmd);
4275 	vm_phys_free_pages(m, 0);
4276 	vm_domain_free_unlock(vmd);
4277 	vm_domain_freecnt_inc(vmd, 1);
4278 }
4279 
4280 /*
4281  *	vm_page_free_pages_toq:
4282  *
4283  *	Returns a list of pages to the free list, disassociating it
4284  *	from any VM object.  In other words, this is equivalent to
4285  *	calling vm_page_free_toq() for each page of a list of VM objects.
4286  */
4287 int
4288 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4289 {
4290 	vm_page_t m;
4291 	int count;
4292 
4293 	if (SLIST_EMPTY(free))
4294 		return (0);
4295 
4296 	count = 0;
4297 	while ((m = SLIST_FIRST(free)) != NULL) {
4298 		count++;
4299 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4300 		vm_page_free_toq(m);
4301 	}
4302 
4303 	if (update_wire_count)
4304 		vm_wire_sub(count);
4305 	return (count);
4306 }
4307 
4308 /*
4309  * Mark this page as wired down.  For managed pages, this prevents reclamation
4310  * by the page daemon, or when the containing object, if any, is destroyed.
4311  */
4312 void
4313 vm_page_wire(vm_page_t m)
4314 {
4315 	u_int old;
4316 
4317 #ifdef INVARIANTS
4318 	if (m->object != NULL && !vm_page_busied(m) &&
4319 	    !vm_object_busied(m->object))
4320 		VM_OBJECT_ASSERT_LOCKED(m->object);
4321 #endif
4322 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4323 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4324 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4325 
4326 	old = atomic_fetchadd_int(&m->ref_count, 1);
4327 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4328 	    ("vm_page_wire: counter overflow for page %p", m));
4329 	if (VPRC_WIRE_COUNT(old) == 0) {
4330 		if ((m->oflags & VPO_UNMANAGED) == 0)
4331 			vm_page_aflag_set(m, PGA_DEQUEUE);
4332 		vm_wire_add(1);
4333 	}
4334 }
4335 
4336 /*
4337  * Attempt to wire a mapped page following a pmap lookup of that page.
4338  * This may fail if a thread is concurrently tearing down mappings of the page.
4339  * The transient failure is acceptable because it translates to the
4340  * failure of the caller pmap_extract_and_hold(), which should be then
4341  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4342  */
4343 bool
4344 vm_page_wire_mapped(vm_page_t m)
4345 {
4346 	u_int old;
4347 
4348 	old = atomic_load_int(&m->ref_count);
4349 	do {
4350 		KASSERT(old > 0,
4351 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4352 		if ((old & VPRC_BLOCKED) != 0)
4353 			return (false);
4354 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4355 
4356 	if (VPRC_WIRE_COUNT(old) == 0) {
4357 		if ((m->oflags & VPO_UNMANAGED) == 0)
4358 			vm_page_aflag_set(m, PGA_DEQUEUE);
4359 		vm_wire_add(1);
4360 	}
4361 	return (true);
4362 }
4363 
4364 /*
4365  * Release a wiring reference to a managed page.  If the page still belongs to
4366  * an object, update its position in the page queues to reflect the reference.
4367  * If the wiring was the last reference to the page, free the page.
4368  */
4369 static void
4370 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4371 {
4372 	u_int old;
4373 
4374 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4375 	    ("%s: page %p is unmanaged", __func__, m));
4376 
4377 	/*
4378 	 * Update LRU state before releasing the wiring reference.
4379 	 * Use a release store when updating the reference count to
4380 	 * synchronize with vm_page_free_prep().
4381 	 */
4382 	old = atomic_load_int(&m->ref_count);
4383 	do {
4384 		u_int count;
4385 
4386 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4387 		    ("vm_page_unwire: wire count underflow for page %p", m));
4388 
4389 		count = old & ~VPRC_BLOCKED;
4390 		if (count > VPRC_OBJREF + 1) {
4391 			/*
4392 			 * The page has at least one other wiring reference.  An
4393 			 * earlier iteration of this loop may have called
4394 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4395 			 * re-set it if necessary.
4396 			 */
4397 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4398 				vm_page_aflag_set(m, PGA_DEQUEUE);
4399 		} else if (count == VPRC_OBJREF + 1) {
4400 			/*
4401 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4402 			 * update the page's queue state to reflect the
4403 			 * reference.  If the page does not belong to an object
4404 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4405 			 * clear leftover queue state.
4406 			 */
4407 			vm_page_release_toq(m, nqueue, noreuse);
4408 		} else if (count == 1) {
4409 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4410 		}
4411 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4412 
4413 	if (VPRC_WIRE_COUNT(old) == 1) {
4414 		vm_wire_sub(1);
4415 		if (old == 1)
4416 			vm_page_free(m);
4417 	}
4418 }
4419 
4420 /*
4421  * Release one wiring of the specified page, potentially allowing it to be
4422  * paged out.
4423  *
4424  * Only managed pages belonging to an object can be paged out.  If the number
4425  * of wirings transitions to zero and the page is eligible for page out, then
4426  * the page is added to the specified paging queue.  If the released wiring
4427  * represented the last reference to the page, the page is freed.
4428  */
4429 void
4430 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4431 {
4432 
4433 	KASSERT(nqueue < PQ_COUNT,
4434 	    ("vm_page_unwire: invalid queue %u request for page %p",
4435 	    nqueue, m));
4436 
4437 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4438 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4439 			vm_page_free(m);
4440 		return;
4441 	}
4442 	vm_page_unwire_managed(m, nqueue, false);
4443 }
4444 
4445 /*
4446  * Unwire a page without (re-)inserting it into a page queue.  It is up
4447  * to the caller to enqueue, requeue, or free the page as appropriate.
4448  * In most cases involving managed pages, vm_page_unwire() should be used
4449  * instead.
4450  */
4451 bool
4452 vm_page_unwire_noq(vm_page_t m)
4453 {
4454 	u_int old;
4455 
4456 	old = vm_page_drop(m, 1);
4457 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4458 	    ("%s: counter underflow for page %p", __func__,  m));
4459 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4460 	    ("%s: missing ref on fictitious page %p", __func__, m));
4461 
4462 	if (VPRC_WIRE_COUNT(old) > 1)
4463 		return (false);
4464 	if ((m->oflags & VPO_UNMANAGED) == 0)
4465 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4466 	vm_wire_sub(1);
4467 	return (true);
4468 }
4469 
4470 /*
4471  * Ensure that the page ends up in the specified page queue.  If the page is
4472  * active or being moved to the active queue, ensure that its act_count is
4473  * at least ACT_INIT but do not otherwise mess with it.
4474  */
4475 static __always_inline void
4476 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4477 {
4478 	vm_page_astate_t old, new;
4479 
4480 	KASSERT(m->ref_count > 0,
4481 	    ("%s: page %p does not carry any references", __func__, m));
4482 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4483 	    ("%s: invalid flags %x", __func__, nflag));
4484 
4485 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4486 		return;
4487 
4488 	old = vm_page_astate_load(m);
4489 	do {
4490 		if ((old.flags & PGA_DEQUEUE) != 0)
4491 			break;
4492 		new = old;
4493 		new.flags &= ~PGA_QUEUE_OP_MASK;
4494 		if (nqueue == PQ_ACTIVE)
4495 			new.act_count = max(old.act_count, ACT_INIT);
4496 		if (old.queue == nqueue) {
4497 			/*
4498 			 * There is no need to requeue pages already in the
4499 			 * active queue.
4500 			 */
4501 			if (nqueue != PQ_ACTIVE ||
4502 			    (old.flags & PGA_ENQUEUED) == 0)
4503 				new.flags |= nflag;
4504 		} else {
4505 			new.flags |= nflag;
4506 			new.queue = nqueue;
4507 		}
4508 	} while (!vm_page_pqstate_commit(m, &old, new));
4509 }
4510 
4511 /*
4512  * Put the specified page on the active list (if appropriate).
4513  */
4514 void
4515 vm_page_activate(vm_page_t m)
4516 {
4517 
4518 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4519 }
4520 
4521 /*
4522  * Move the specified page to the tail of the inactive queue, or requeue
4523  * the page if it is already in the inactive queue.
4524  */
4525 void
4526 vm_page_deactivate(vm_page_t m)
4527 {
4528 
4529 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4530 }
4531 
4532 void
4533 vm_page_deactivate_noreuse(vm_page_t m)
4534 {
4535 
4536 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4537 }
4538 
4539 /*
4540  * Put a page in the laundry, or requeue it if it is already there.
4541  */
4542 void
4543 vm_page_launder(vm_page_t m)
4544 {
4545 
4546 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4547 }
4548 
4549 /*
4550  * Put a page in the PQ_UNSWAPPABLE holding queue.
4551  */
4552 void
4553 vm_page_unswappable(vm_page_t m)
4554 {
4555 
4556 	VM_OBJECT_ASSERT_LOCKED(m->object);
4557 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4558 	    ("page %p already unswappable", m));
4559 
4560 	vm_page_dequeue(m);
4561 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4562 }
4563 
4564 /*
4565  * Release a page back to the page queues in preparation for unwiring.
4566  */
4567 static void
4568 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4569 {
4570 	vm_page_astate_t old, new;
4571 	uint16_t nflag;
4572 
4573 	/*
4574 	 * Use a check of the valid bits to determine whether we should
4575 	 * accelerate reclamation of the page.  The object lock might not be
4576 	 * held here, in which case the check is racy.  At worst we will either
4577 	 * accelerate reclamation of a valid page and violate LRU, or
4578 	 * unnecessarily defer reclamation of an invalid page.
4579 	 *
4580 	 * If we were asked to not cache the page, place it near the head of the
4581 	 * inactive queue so that is reclaimed sooner.
4582 	 */
4583 	if (noreuse || vm_page_none_valid(m)) {
4584 		nqueue = PQ_INACTIVE;
4585 		nflag = PGA_REQUEUE_HEAD;
4586 	} else {
4587 		nflag = PGA_REQUEUE;
4588 	}
4589 
4590 	old = vm_page_astate_load(m);
4591 	do {
4592 		new = old;
4593 
4594 		/*
4595 		 * If the page is already in the active queue and we are not
4596 		 * trying to accelerate reclamation, simply mark it as
4597 		 * referenced and avoid any queue operations.
4598 		 */
4599 		new.flags &= ~PGA_QUEUE_OP_MASK;
4600 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4601 		    (old.flags & PGA_ENQUEUED) != 0)
4602 			new.flags |= PGA_REFERENCED;
4603 		else {
4604 			new.flags |= nflag;
4605 			new.queue = nqueue;
4606 		}
4607 	} while (!vm_page_pqstate_commit(m, &old, new));
4608 }
4609 
4610 /*
4611  * Unwire a page and either attempt to free it or re-add it to the page queues.
4612  */
4613 void
4614 vm_page_release(vm_page_t m, int flags)
4615 {
4616 	vm_object_t object;
4617 
4618 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4619 	    ("vm_page_release: page %p is unmanaged", m));
4620 
4621 	if ((flags & VPR_TRYFREE) != 0) {
4622 		for (;;) {
4623 			object = atomic_load_ptr(&m->object);
4624 			if (object == NULL)
4625 				break;
4626 			/* Depends on type-stability. */
4627 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4628 				break;
4629 			if (object == m->object) {
4630 				vm_page_release_locked(m, flags);
4631 				VM_OBJECT_WUNLOCK(object);
4632 				return;
4633 			}
4634 			VM_OBJECT_WUNLOCK(object);
4635 		}
4636 	}
4637 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4638 }
4639 
4640 /* See vm_page_release(). */
4641 void
4642 vm_page_release_locked(vm_page_t m, int flags)
4643 {
4644 
4645 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4646 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4647 	    ("vm_page_release_locked: page %p is unmanaged", m));
4648 
4649 	if (vm_page_unwire_noq(m)) {
4650 		if ((flags & VPR_TRYFREE) != 0 &&
4651 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4652 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4653 			/*
4654 			 * An unlocked lookup may have wired the page before the
4655 			 * busy lock was acquired, in which case the page must
4656 			 * not be freed.
4657 			 */
4658 			if (__predict_true(!vm_page_wired(m))) {
4659 				vm_page_free(m);
4660 				return;
4661 			}
4662 			vm_page_xunbusy(m);
4663 		} else {
4664 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4665 		}
4666 	}
4667 }
4668 
4669 static bool
4670 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4671 {
4672 	u_int old;
4673 
4674 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4675 	    ("vm_page_try_blocked_op: page %p has no object", m));
4676 	KASSERT(vm_page_busied(m),
4677 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4678 	VM_OBJECT_ASSERT_LOCKED(m->object);
4679 
4680 	old = atomic_load_int(&m->ref_count);
4681 	do {
4682 		KASSERT(old != 0,
4683 		    ("vm_page_try_blocked_op: page %p has no references", m));
4684 		KASSERT((old & VPRC_BLOCKED) == 0,
4685 		    ("vm_page_try_blocked_op: page %p blocks wirings", m));
4686 		if (VPRC_WIRE_COUNT(old) != 0)
4687 			return (false);
4688 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4689 
4690 	(op)(m);
4691 
4692 	/*
4693 	 * If the object is read-locked, new wirings may be created via an
4694 	 * object lookup.
4695 	 */
4696 	old = vm_page_drop(m, VPRC_BLOCKED);
4697 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4698 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4699 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4700 	    old, m));
4701 	return (true);
4702 }
4703 
4704 /*
4705  * Atomically check for wirings and remove all mappings of the page.
4706  */
4707 bool
4708 vm_page_try_remove_all(vm_page_t m)
4709 {
4710 
4711 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4712 }
4713 
4714 /*
4715  * Atomically check for wirings and remove all writeable mappings of the page.
4716  */
4717 bool
4718 vm_page_try_remove_write(vm_page_t m)
4719 {
4720 
4721 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4722 }
4723 
4724 /*
4725  * vm_page_advise
4726  *
4727  * 	Apply the specified advice to the given page.
4728  */
4729 void
4730 vm_page_advise(vm_page_t m, int advice)
4731 {
4732 
4733 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4734 	vm_page_assert_xbusied(m);
4735 
4736 	if (advice == MADV_FREE)
4737 		/*
4738 		 * Mark the page clean.  This will allow the page to be freed
4739 		 * without first paging it out.  MADV_FREE pages are often
4740 		 * quickly reused by malloc(3), so we do not do anything that
4741 		 * would result in a page fault on a later access.
4742 		 */
4743 		vm_page_undirty(m);
4744 	else if (advice != MADV_DONTNEED) {
4745 		if (advice == MADV_WILLNEED)
4746 			vm_page_activate(m);
4747 		return;
4748 	}
4749 
4750 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4751 		vm_page_dirty(m);
4752 
4753 	/*
4754 	 * Clear any references to the page.  Otherwise, the page daemon will
4755 	 * immediately reactivate the page.
4756 	 */
4757 	vm_page_aflag_clear(m, PGA_REFERENCED);
4758 
4759 	/*
4760 	 * Place clean pages near the head of the inactive queue rather than
4761 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4762 	 * the page will be reused quickly.  Dirty pages not already in the
4763 	 * laundry are moved there.
4764 	 */
4765 	if (m->dirty == 0)
4766 		vm_page_deactivate_noreuse(m);
4767 	else if (!vm_page_in_laundry(m))
4768 		vm_page_launder(m);
4769 }
4770 
4771 /*
4772  *	vm_page_grab_release
4773  *
4774  *	Helper routine for grab functions to release busy on return.
4775  */
4776 static inline void
4777 vm_page_grab_release(vm_page_t m, int allocflags)
4778 {
4779 
4780 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4781 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4782 			vm_page_sunbusy(m);
4783 		else
4784 			vm_page_xunbusy(m);
4785 	}
4786 }
4787 
4788 /*
4789  *	vm_page_grab_sleep
4790  *
4791  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4792  *	if the caller should retry and false otherwise.
4793  *
4794  *	If the object is locked on entry the object will be unlocked with
4795  *	false returns and still locked but possibly having been dropped
4796  *	with true returns.
4797  */
4798 static bool
4799 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4800     const char *wmesg, int allocflags, bool locked)
4801 {
4802 
4803 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4804 		return (false);
4805 
4806 	/*
4807 	 * Reference the page before unlocking and sleeping so that
4808 	 * the page daemon is less likely to reclaim it.
4809 	 */
4810 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4811 		vm_page_reference(m);
4812 
4813 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4814 	    locked)
4815 		VM_OBJECT_WLOCK(object);
4816 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4817 		return (false);
4818 
4819 	return (true);
4820 }
4821 
4822 /*
4823  * Assert that the grab flags are valid.
4824  */
4825 static inline void
4826 vm_page_grab_check(int allocflags)
4827 {
4828 
4829 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4830 	    (allocflags & VM_ALLOC_WIRED) != 0,
4831 	    ("vm_page_grab*: the pages must be busied or wired"));
4832 
4833 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4834 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4835 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4836 }
4837 
4838 /*
4839  * Calculate the page allocation flags for grab.
4840  */
4841 static inline int
4842 vm_page_grab_pflags(int allocflags)
4843 {
4844 	int pflags;
4845 
4846 	pflags = allocflags &
4847 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4848 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4849 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4850 		pflags |= VM_ALLOC_WAITFAIL;
4851 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4852 		pflags |= VM_ALLOC_SBUSY;
4853 
4854 	return (pflags);
4855 }
4856 
4857 /*
4858  * Grab a page, waiting until we are waken up due to the page
4859  * changing state.  We keep on waiting, if the page continues
4860  * to be in the object.  If the page doesn't exist, first allocate it
4861  * and then conditionally zero it.
4862  *
4863  * This routine may sleep.
4864  *
4865  * The object must be locked on entry.  The lock will, however, be released
4866  * and reacquired if the routine sleeps.
4867  */
4868 vm_page_t
4869 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4870 {
4871 	vm_page_t m;
4872 
4873 	VM_OBJECT_ASSERT_WLOCKED(object);
4874 	vm_page_grab_check(allocflags);
4875 
4876 retrylookup:
4877 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4878 		if (!vm_page_tryacquire(m, allocflags)) {
4879 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4880 			    allocflags, true))
4881 				goto retrylookup;
4882 			return (NULL);
4883 		}
4884 		goto out;
4885 	}
4886 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4887 		return (NULL);
4888 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4889 	if (m == NULL) {
4890 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4891 			return (NULL);
4892 		goto retrylookup;
4893 	}
4894 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4895 		pmap_zero_page(m);
4896 
4897 out:
4898 	vm_page_grab_release(m, allocflags);
4899 
4900 	return (m);
4901 }
4902 
4903 /*
4904  * Attempt to validate a page, locklessly acquiring it if necessary, given a
4905  * (object, pindex) tuple and either an invalided page or NULL.  The resulting
4906  * page will be validated against the identity tuple, and busied or wired as
4907  * requested.  A NULL page returned guarantees that the page was not in radix at
4908  * the time of the call but callers must perform higher level synchronization or
4909  * retry the operation under a lock if they require an atomic answer.  This is
4910  * the only lock free validation routine, other routines can depend on the
4911  * resulting page state.
4912  *
4913  * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to
4914  * caller flags.
4915  */
4916 #define PAGE_NOT_ACQUIRED ((vm_page_t)1)
4917 static vm_page_t
4918 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m,
4919     int allocflags)
4920 {
4921 	if (m == NULL)
4922 		m = vm_page_lookup_unlocked(object, pindex);
4923 	for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) {
4924 		if (vm_page_trybusy(m, allocflags)) {
4925 			if (m->object == object && m->pindex == pindex) {
4926 				if ((allocflags & VM_ALLOC_WIRED) != 0)
4927 					vm_page_wire(m);
4928 				vm_page_grab_release(m, allocflags);
4929 				break;
4930 			}
4931 			/* relookup. */
4932 			vm_page_busy_release(m);
4933 			cpu_spinwait();
4934 			continue;
4935 		}
4936 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4937 		    allocflags, false))
4938 			return (PAGE_NOT_ACQUIRED);
4939 	}
4940 	return (m);
4941 }
4942 
4943 /*
4944  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4945  * is not set.
4946  */
4947 vm_page_t
4948 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4949 {
4950 	vm_page_t m;
4951 
4952 	vm_page_grab_check(allocflags);
4953 	m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags);
4954 	if (m == PAGE_NOT_ACQUIRED)
4955 		return (NULL);
4956 	if (m != NULL)
4957 		return (m);
4958 
4959 	/*
4960 	 * The radix lockless lookup should never return a false negative
4961 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4962 	 * was no page present at the instant of the call.  A NOCREAT caller
4963 	 * must handle create races gracefully.
4964 	 */
4965 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4966 		return (NULL);
4967 
4968 	VM_OBJECT_WLOCK(object);
4969 	m = vm_page_grab(object, pindex, allocflags);
4970 	VM_OBJECT_WUNLOCK(object);
4971 
4972 	return (m);
4973 }
4974 
4975 /*
4976  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4977  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4978  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4979  * in simultaneously.  Additional pages will be left on a paging queue but
4980  * will neither be wired nor busy regardless of allocflags.
4981  */
4982 int
4983 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4984 {
4985 	vm_page_t m;
4986 	vm_page_t ma[VM_INITIAL_PAGEIN];
4987 	int after, i, pflags, rv;
4988 
4989 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4990 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4991 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4992 	KASSERT((allocflags &
4993 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4994 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4995 	VM_OBJECT_ASSERT_WLOCKED(object);
4996 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4997 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4998 	pflags |= VM_ALLOC_WAITFAIL;
4999 
5000 retrylookup:
5001 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
5002 		/*
5003 		 * If the page is fully valid it can only become invalid
5004 		 * with the object lock held.  If it is not valid it can
5005 		 * become valid with the busy lock held.  Therefore, we
5006 		 * may unnecessarily lock the exclusive busy here if we
5007 		 * race with I/O completion not using the object lock.
5008 		 * However, we will not end up with an invalid page and a
5009 		 * shared lock.
5010 		 */
5011 		if (!vm_page_trybusy(m,
5012 		    vm_page_all_valid(m) ? allocflags : 0)) {
5013 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
5014 			    allocflags, true);
5015 			goto retrylookup;
5016 		}
5017 		if (vm_page_all_valid(m))
5018 			goto out;
5019 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5020 			vm_page_busy_release(m);
5021 			*mp = NULL;
5022 			return (VM_PAGER_FAIL);
5023 		}
5024 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5025 		*mp = NULL;
5026 		return (VM_PAGER_FAIL);
5027 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
5028 		if (!vm_pager_can_alloc_page(object, pindex)) {
5029 			*mp = NULL;
5030 			return (VM_PAGER_AGAIN);
5031 		}
5032 		goto retrylookup;
5033 	}
5034 
5035 	vm_page_assert_xbusied(m);
5036 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
5037 		after = MIN(after, VM_INITIAL_PAGEIN);
5038 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
5039 		after = MAX(after, 1);
5040 		ma[0] = m;
5041 		for (i = 1; i < after; i++) {
5042 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
5043 				if (vm_page_any_valid(ma[i]) ||
5044 				    !vm_page_tryxbusy(ma[i]))
5045 					break;
5046 			} else {
5047 				ma[i] = vm_page_alloc(object, m->pindex + i,
5048 				    VM_ALLOC_NORMAL);
5049 				if (ma[i] == NULL)
5050 					break;
5051 			}
5052 		}
5053 		after = i;
5054 		vm_object_pip_add(object, after);
5055 		VM_OBJECT_WUNLOCK(object);
5056 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
5057 		VM_OBJECT_WLOCK(object);
5058 		vm_object_pip_wakeupn(object, after);
5059 		/* Pager may have replaced a page. */
5060 		m = ma[0];
5061 		if (rv != VM_PAGER_OK) {
5062 			for (i = 0; i < after; i++) {
5063 				if (!vm_page_wired(ma[i]))
5064 					vm_page_free(ma[i]);
5065 				else
5066 					vm_page_xunbusy(ma[i]);
5067 			}
5068 			*mp = NULL;
5069 			return (rv);
5070 		}
5071 		for (i = 1; i < after; i++)
5072 			vm_page_readahead_finish(ma[i]);
5073 		MPASS(vm_page_all_valid(m));
5074 	} else {
5075 		vm_page_zero_invalid(m, TRUE);
5076 	}
5077 out:
5078 	if ((allocflags & VM_ALLOC_WIRED) != 0)
5079 		vm_page_wire(m);
5080 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
5081 		vm_page_busy_downgrade(m);
5082 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
5083 		vm_page_busy_release(m);
5084 	*mp = m;
5085 	return (VM_PAGER_OK);
5086 }
5087 
5088 /*
5089  * Locklessly grab a valid page.  If the page is not valid or not yet
5090  * allocated this will fall back to the object lock method.
5091  */
5092 int
5093 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
5094     vm_pindex_t pindex, int allocflags)
5095 {
5096 	vm_page_t m;
5097 	int flags;
5098 	int error;
5099 
5100 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
5101 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
5102 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
5103 	    "mismatch"));
5104 	KASSERT((allocflags &
5105 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
5106 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
5107 
5108 	/*
5109 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
5110 	 * before we can inspect the valid field and return a wired page.
5111 	 */
5112 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
5113 	vm_page_grab_check(flags);
5114 	m = vm_page_acquire_unlocked(object, pindex, NULL, flags);
5115 	if (m == PAGE_NOT_ACQUIRED)
5116 		return (VM_PAGER_FAIL);
5117 	if (m != NULL) {
5118 		if (vm_page_all_valid(m)) {
5119 			if ((allocflags & VM_ALLOC_WIRED) != 0)
5120 				vm_page_wire(m);
5121 			vm_page_grab_release(m, allocflags);
5122 			*mp = m;
5123 			return (VM_PAGER_OK);
5124 		}
5125 		vm_page_busy_release(m);
5126 	}
5127 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
5128 		*mp = NULL;
5129 		return (VM_PAGER_FAIL);
5130 	}
5131 	VM_OBJECT_WLOCK(object);
5132 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
5133 	VM_OBJECT_WUNLOCK(object);
5134 
5135 	return (error);
5136 }
5137 
5138 /*
5139  * Return the specified range of pages from the given object.  For each
5140  * page offset within the range, if a page already exists within the object
5141  * at that offset and it is busy, then wait for it to change state.  If,
5142  * instead, the page doesn't exist, then allocate it.
5143  *
5144  * The caller must always specify an allocation class.
5145  *
5146  * allocation classes:
5147  *	VM_ALLOC_NORMAL		normal process request
5148  *	VM_ALLOC_SYSTEM		system *really* needs the pages
5149  *
5150  * The caller must always specify that the pages are to be busied and/or
5151  * wired.
5152  *
5153  * optional allocation flags:
5154  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
5155  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
5156  *	VM_ALLOC_NOWAIT		do not sleep
5157  *	VM_ALLOC_SBUSY		set page to sbusy state
5158  *	VM_ALLOC_WIRED		wire the pages
5159  *	VM_ALLOC_ZERO		zero and validate any invalid pages
5160  *
5161  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
5162  * may return a partial prefix of the requested range.
5163  */
5164 int
5165 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
5166     vm_page_t *ma, int count)
5167 {
5168 	vm_page_t m, mpred;
5169 	int pflags;
5170 	int i;
5171 
5172 	VM_OBJECT_ASSERT_WLOCKED(object);
5173 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
5174 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
5175 	KASSERT(count > 0,
5176 	    ("vm_page_grab_pages: invalid page count %d", count));
5177 	vm_page_grab_check(allocflags);
5178 
5179 	pflags = vm_page_grab_pflags(allocflags);
5180 	i = 0;
5181 retrylookup:
5182 	m = vm_page_mpred(object, pindex + i);
5183 	if (m == NULL || m->pindex != pindex + i) {
5184 		mpred = m;
5185 		m = NULL;
5186 	} else
5187 		mpred = TAILQ_PREV(m, pglist, listq);
5188 	for (; i < count; i++) {
5189 		if (m != NULL) {
5190 			if (!vm_page_tryacquire(m, allocflags)) {
5191 				if (vm_page_grab_sleep(object, m, pindex + i,
5192 				    "grbmaw", allocflags, true))
5193 					goto retrylookup;
5194 				break;
5195 			}
5196 		} else {
5197 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
5198 				break;
5199 			m = vm_page_alloc_after(object, pindex + i,
5200 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
5201 			if (m == NULL) {
5202 				if ((allocflags & (VM_ALLOC_NOWAIT |
5203 				    VM_ALLOC_WAITFAIL)) != 0)
5204 					break;
5205 				goto retrylookup;
5206 			}
5207 		}
5208 		if (vm_page_none_valid(m) &&
5209 		    (allocflags & VM_ALLOC_ZERO) != 0) {
5210 			if ((m->flags & PG_ZERO) == 0)
5211 				pmap_zero_page(m);
5212 			vm_page_valid(m);
5213 		}
5214 		vm_page_grab_release(m, allocflags);
5215 		ma[i] = mpred = m;
5216 		m = vm_page_next(m);
5217 	}
5218 	return (i);
5219 }
5220 
5221 /*
5222  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
5223  * and will fall back to the locked variant to handle allocation.
5224  */
5225 int
5226 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
5227     int allocflags, vm_page_t *ma, int count)
5228 {
5229 	vm_page_t m;
5230 	int flags;
5231 	int i;
5232 
5233 	KASSERT(count > 0,
5234 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
5235 	vm_page_grab_check(allocflags);
5236 
5237 	/*
5238 	 * Modify flags for lockless acquire to hold the page until we
5239 	 * set it valid if necessary.
5240 	 */
5241 	flags = allocflags & ~VM_ALLOC_NOBUSY;
5242 	vm_page_grab_check(flags);
5243 	m = NULL;
5244 	for (i = 0; i < count; i++, pindex++) {
5245 		/*
5246 		 * We may see a false NULL here because the previous page has
5247 		 * been removed or just inserted and the list is loaded without
5248 		 * barriers.  Switch to radix to verify.
5249 		 */
5250 		if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex ||
5251 		    atomic_load_ptr(&m->object) != object) {
5252 			/*
5253 			 * This guarantees the result is instantaneously
5254 			 * correct.
5255 			 */
5256 			m = NULL;
5257 		}
5258 		m = vm_page_acquire_unlocked(object, pindex, m, flags);
5259 		if (m == PAGE_NOT_ACQUIRED)
5260 			return (i);
5261 		if (m == NULL)
5262 			break;
5263 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5264 			if ((m->flags & PG_ZERO) == 0)
5265 				pmap_zero_page(m);
5266 			vm_page_valid(m);
5267 		}
5268 		/* m will still be wired or busy according to flags. */
5269 		vm_page_grab_release(m, allocflags);
5270 		ma[i] = m;
5271 		m = TAILQ_NEXT(m, listq);
5272 	}
5273 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5274 		return (i);
5275 	count -= i;
5276 	VM_OBJECT_WLOCK(object);
5277 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5278 	VM_OBJECT_WUNLOCK(object);
5279 
5280 	return (i);
5281 }
5282 
5283 /*
5284  * Mapping function for valid or dirty bits in a page.
5285  *
5286  * Inputs are required to range within a page.
5287  */
5288 vm_page_bits_t
5289 vm_page_bits(int base, int size)
5290 {
5291 	int first_bit;
5292 	int last_bit;
5293 
5294 	KASSERT(
5295 	    base + size <= PAGE_SIZE,
5296 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5297 	);
5298 
5299 	if (size == 0)		/* handle degenerate case */
5300 		return (0);
5301 
5302 	first_bit = base >> DEV_BSHIFT;
5303 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5304 
5305 	return (((vm_page_bits_t)2 << last_bit) -
5306 	    ((vm_page_bits_t)1 << first_bit));
5307 }
5308 
5309 void
5310 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5311 {
5312 
5313 #if PAGE_SIZE == 32768
5314 	atomic_set_64((uint64_t *)bits, set);
5315 #elif PAGE_SIZE == 16384
5316 	atomic_set_32((uint32_t *)bits, set);
5317 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5318 	atomic_set_16((uint16_t *)bits, set);
5319 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5320 	atomic_set_8((uint8_t *)bits, set);
5321 #else		/* PAGE_SIZE <= 8192 */
5322 	uintptr_t addr;
5323 	int shift;
5324 
5325 	addr = (uintptr_t)bits;
5326 	/*
5327 	 * Use a trick to perform a 32-bit atomic on the
5328 	 * containing aligned word, to not depend on the existence
5329 	 * of atomic_{set, clear}_{8, 16}.
5330 	 */
5331 	shift = addr & (sizeof(uint32_t) - 1);
5332 #if BYTE_ORDER == BIG_ENDIAN
5333 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5334 #else
5335 	shift *= NBBY;
5336 #endif
5337 	addr &= ~(sizeof(uint32_t) - 1);
5338 	atomic_set_32((uint32_t *)addr, set << shift);
5339 #endif		/* PAGE_SIZE */
5340 }
5341 
5342 static inline void
5343 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5344 {
5345 
5346 #if PAGE_SIZE == 32768
5347 	atomic_clear_64((uint64_t *)bits, clear);
5348 #elif PAGE_SIZE == 16384
5349 	atomic_clear_32((uint32_t *)bits, clear);
5350 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5351 	atomic_clear_16((uint16_t *)bits, clear);
5352 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5353 	atomic_clear_8((uint8_t *)bits, clear);
5354 #else		/* PAGE_SIZE <= 8192 */
5355 	uintptr_t addr;
5356 	int shift;
5357 
5358 	addr = (uintptr_t)bits;
5359 	/*
5360 	 * Use a trick to perform a 32-bit atomic on the
5361 	 * containing aligned word, to not depend on the existence
5362 	 * of atomic_{set, clear}_{8, 16}.
5363 	 */
5364 	shift = addr & (sizeof(uint32_t) - 1);
5365 #if BYTE_ORDER == BIG_ENDIAN
5366 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5367 #else
5368 	shift *= NBBY;
5369 #endif
5370 	addr &= ~(sizeof(uint32_t) - 1);
5371 	atomic_clear_32((uint32_t *)addr, clear << shift);
5372 #endif		/* PAGE_SIZE */
5373 }
5374 
5375 static inline vm_page_bits_t
5376 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5377 {
5378 #if PAGE_SIZE == 32768
5379 	uint64_t old;
5380 
5381 	old = *bits;
5382 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5383 	return (old);
5384 #elif PAGE_SIZE == 16384
5385 	uint32_t old;
5386 
5387 	old = *bits;
5388 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5389 	return (old);
5390 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5391 	uint16_t old;
5392 
5393 	old = *bits;
5394 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5395 	return (old);
5396 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5397 	uint8_t old;
5398 
5399 	old = *bits;
5400 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5401 	return (old);
5402 #else		/* PAGE_SIZE <= 4096*/
5403 	uintptr_t addr;
5404 	uint32_t old, new, mask;
5405 	int shift;
5406 
5407 	addr = (uintptr_t)bits;
5408 	/*
5409 	 * Use a trick to perform a 32-bit atomic on the
5410 	 * containing aligned word, to not depend on the existence
5411 	 * of atomic_{set, swap, clear}_{8, 16}.
5412 	 */
5413 	shift = addr & (sizeof(uint32_t) - 1);
5414 #if BYTE_ORDER == BIG_ENDIAN
5415 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5416 #else
5417 	shift *= NBBY;
5418 #endif
5419 	addr &= ~(sizeof(uint32_t) - 1);
5420 	mask = VM_PAGE_BITS_ALL << shift;
5421 
5422 	old = *bits;
5423 	do {
5424 		new = old & ~mask;
5425 		new |= newbits << shift;
5426 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5427 	return (old >> shift);
5428 #endif		/* PAGE_SIZE */
5429 }
5430 
5431 /*
5432  *	vm_page_set_valid_range:
5433  *
5434  *	Sets portions of a page valid.  The arguments are expected
5435  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5436  *	of any partial chunks touched by the range.  The invalid portion of
5437  *	such chunks will be zeroed.
5438  *
5439  *	(base + size) must be less then or equal to PAGE_SIZE.
5440  */
5441 void
5442 vm_page_set_valid_range(vm_page_t m, int base, int size)
5443 {
5444 	int endoff, frag;
5445 	vm_page_bits_t pagebits;
5446 
5447 	vm_page_assert_busied(m);
5448 	if (size == 0)	/* handle degenerate case */
5449 		return;
5450 
5451 	/*
5452 	 * If the base is not DEV_BSIZE aligned and the valid
5453 	 * bit is clear, we have to zero out a portion of the
5454 	 * first block.
5455 	 */
5456 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5457 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5458 		pmap_zero_page_area(m, frag, base - frag);
5459 
5460 	/*
5461 	 * If the ending offset is not DEV_BSIZE aligned and the
5462 	 * valid bit is clear, we have to zero out a portion of
5463 	 * the last block.
5464 	 */
5465 	endoff = base + size;
5466 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5467 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5468 		pmap_zero_page_area(m, endoff,
5469 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5470 
5471 	/*
5472 	 * Assert that no previously invalid block that is now being validated
5473 	 * is already dirty.
5474 	 */
5475 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5476 	    ("vm_page_set_valid_range: page %p is dirty", m));
5477 
5478 	/*
5479 	 * Set valid bits inclusive of any overlap.
5480 	 */
5481 	pagebits = vm_page_bits(base, size);
5482 	if (vm_page_xbusied(m))
5483 		m->valid |= pagebits;
5484 	else
5485 		vm_page_bits_set(m, &m->valid, pagebits);
5486 }
5487 
5488 /*
5489  * Set the page dirty bits and free the invalid swap space if
5490  * present.  Returns the previous dirty bits.
5491  */
5492 vm_page_bits_t
5493 vm_page_set_dirty(vm_page_t m)
5494 {
5495 	vm_page_bits_t old;
5496 
5497 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5498 
5499 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5500 		old = m->dirty;
5501 		m->dirty = VM_PAGE_BITS_ALL;
5502 	} else
5503 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5504 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5505 		vm_pager_page_unswapped(m);
5506 
5507 	return (old);
5508 }
5509 
5510 /*
5511  * Clear the given bits from the specified page's dirty field.
5512  */
5513 static __inline void
5514 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5515 {
5516 
5517 	vm_page_assert_busied(m);
5518 
5519 	/*
5520 	 * If the page is xbusied and not write mapped we are the
5521 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5522 	 * layer can call vm_page_dirty() without holding a distinguished
5523 	 * lock.  The combination of page busy and atomic operations
5524 	 * suffice to guarantee consistency of the page dirty field.
5525 	 */
5526 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5527 		m->dirty &= ~pagebits;
5528 	else
5529 		vm_page_bits_clear(m, &m->dirty, pagebits);
5530 }
5531 
5532 /*
5533  *	vm_page_set_validclean:
5534  *
5535  *	Sets portions of a page valid and clean.  The arguments are expected
5536  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5537  *	of any partial chunks touched by the range.  The invalid portion of
5538  *	such chunks will be zero'd.
5539  *
5540  *	(base + size) must be less then or equal to PAGE_SIZE.
5541  */
5542 void
5543 vm_page_set_validclean(vm_page_t m, int base, int size)
5544 {
5545 	vm_page_bits_t oldvalid, pagebits;
5546 	int endoff, frag;
5547 
5548 	vm_page_assert_busied(m);
5549 	if (size == 0)	/* handle degenerate case */
5550 		return;
5551 
5552 	/*
5553 	 * If the base is not DEV_BSIZE aligned and the valid
5554 	 * bit is clear, we have to zero out a portion of the
5555 	 * first block.
5556 	 */
5557 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5558 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5559 		pmap_zero_page_area(m, frag, base - frag);
5560 
5561 	/*
5562 	 * If the ending offset is not DEV_BSIZE aligned and the
5563 	 * valid bit is clear, we have to zero out a portion of
5564 	 * the last block.
5565 	 */
5566 	endoff = base + size;
5567 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5568 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5569 		pmap_zero_page_area(m, endoff,
5570 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5571 
5572 	/*
5573 	 * Set valid, clear dirty bits.  If validating the entire
5574 	 * page we can safely clear the pmap modify bit.  We also
5575 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5576 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5577 	 * be set again.
5578 	 *
5579 	 * We set valid bits inclusive of any overlap, but we can only
5580 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5581 	 * the range.
5582 	 */
5583 	oldvalid = m->valid;
5584 	pagebits = vm_page_bits(base, size);
5585 	if (vm_page_xbusied(m))
5586 		m->valid |= pagebits;
5587 	else
5588 		vm_page_bits_set(m, &m->valid, pagebits);
5589 #if 0	/* NOT YET */
5590 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5591 		frag = DEV_BSIZE - frag;
5592 		base += frag;
5593 		size -= frag;
5594 		if (size < 0)
5595 			size = 0;
5596 	}
5597 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5598 #endif
5599 	if (base == 0 && size == PAGE_SIZE) {
5600 		/*
5601 		 * The page can only be modified within the pmap if it is
5602 		 * mapped, and it can only be mapped if it was previously
5603 		 * fully valid.
5604 		 */
5605 		if (oldvalid == VM_PAGE_BITS_ALL)
5606 			/*
5607 			 * Perform the pmap_clear_modify() first.  Otherwise,
5608 			 * a concurrent pmap operation, such as
5609 			 * pmap_protect(), could clear a modification in the
5610 			 * pmap and set the dirty field on the page before
5611 			 * pmap_clear_modify() had begun and after the dirty
5612 			 * field was cleared here.
5613 			 */
5614 			pmap_clear_modify(m);
5615 		m->dirty = 0;
5616 		vm_page_aflag_clear(m, PGA_NOSYNC);
5617 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5618 		m->dirty &= ~pagebits;
5619 	else
5620 		vm_page_clear_dirty_mask(m, pagebits);
5621 }
5622 
5623 void
5624 vm_page_clear_dirty(vm_page_t m, int base, int size)
5625 {
5626 
5627 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5628 }
5629 
5630 /*
5631  *	vm_page_set_invalid:
5632  *
5633  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5634  *	valid and dirty bits for the effected areas are cleared.
5635  */
5636 void
5637 vm_page_set_invalid(vm_page_t m, int base, int size)
5638 {
5639 	vm_page_bits_t bits;
5640 	vm_object_t object;
5641 
5642 	/*
5643 	 * The object lock is required so that pages can't be mapped
5644 	 * read-only while we're in the process of invalidating them.
5645 	 */
5646 	object = m->object;
5647 	VM_OBJECT_ASSERT_WLOCKED(object);
5648 	vm_page_assert_busied(m);
5649 
5650 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5651 	    size >= object->un_pager.vnp.vnp_size)
5652 		bits = VM_PAGE_BITS_ALL;
5653 	else
5654 		bits = vm_page_bits(base, size);
5655 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5656 		pmap_remove_all(m);
5657 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5658 	    !pmap_page_is_mapped(m),
5659 	    ("vm_page_set_invalid: page %p is mapped", m));
5660 	if (vm_page_xbusied(m)) {
5661 		m->valid &= ~bits;
5662 		m->dirty &= ~bits;
5663 	} else {
5664 		vm_page_bits_clear(m, &m->valid, bits);
5665 		vm_page_bits_clear(m, &m->dirty, bits);
5666 	}
5667 }
5668 
5669 /*
5670  *	vm_page_invalid:
5671  *
5672  *	Invalidates the entire page.  The page must be busy, unmapped, and
5673  *	the enclosing object must be locked.  The object locks protects
5674  *	against concurrent read-only pmap enter which is done without
5675  *	busy.
5676  */
5677 void
5678 vm_page_invalid(vm_page_t m)
5679 {
5680 
5681 	vm_page_assert_busied(m);
5682 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5683 	MPASS(!pmap_page_is_mapped(m));
5684 
5685 	if (vm_page_xbusied(m))
5686 		m->valid = 0;
5687 	else
5688 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5689 }
5690 
5691 /*
5692  * vm_page_zero_invalid()
5693  *
5694  *	The kernel assumes that the invalid portions of a page contain
5695  *	garbage, but such pages can be mapped into memory by user code.
5696  *	When this occurs, we must zero out the non-valid portions of the
5697  *	page so user code sees what it expects.
5698  *
5699  *	Pages are most often semi-valid when the end of a file is mapped
5700  *	into memory and the file's size is not page aligned.
5701  */
5702 void
5703 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5704 {
5705 	int b;
5706 	int i;
5707 
5708 	/*
5709 	 * Scan the valid bits looking for invalid sections that
5710 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5711 	 * valid bit may be set ) have already been zeroed by
5712 	 * vm_page_set_validclean().
5713 	 */
5714 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5715 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5716 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5717 			if (i > b) {
5718 				pmap_zero_page_area(m,
5719 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5720 			}
5721 			b = i + 1;
5722 		}
5723 	}
5724 
5725 	/*
5726 	 * setvalid is TRUE when we can safely set the zero'd areas
5727 	 * as being valid.  We can do this if there are no cache consistency
5728 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5729 	 */
5730 	if (setvalid)
5731 		vm_page_valid(m);
5732 }
5733 
5734 /*
5735  *	vm_page_is_valid:
5736  *
5737  *	Is (partial) page valid?  Note that the case where size == 0
5738  *	will return FALSE in the degenerate case where the page is
5739  *	entirely invalid, and TRUE otherwise.
5740  *
5741  *	Some callers envoke this routine without the busy lock held and
5742  *	handle races via higher level locks.  Typical callers should
5743  *	hold a busy lock to prevent invalidation.
5744  */
5745 int
5746 vm_page_is_valid(vm_page_t m, int base, int size)
5747 {
5748 	vm_page_bits_t bits;
5749 
5750 	bits = vm_page_bits(base, size);
5751 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5752 }
5753 
5754 /*
5755  * Returns true if all of the specified predicates are true for the entire
5756  * (super)page and false otherwise.
5757  */
5758 bool
5759 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m)
5760 {
5761 	vm_object_t object;
5762 	int i, npages;
5763 
5764 	object = m->object;
5765 	if (skip_m != NULL && skip_m->object != object)
5766 		return (false);
5767 	VM_OBJECT_ASSERT_LOCKED(object);
5768 	KASSERT(psind <= m->psind,
5769 	    ("psind %d > psind %d of m %p", psind, m->psind, m));
5770 	npages = atop(pagesizes[psind]);
5771 
5772 	/*
5773 	 * The physically contiguous pages that make up a superpage, i.e., a
5774 	 * page with a page size index ("psind") greater than zero, will
5775 	 * occupy adjacent entries in vm_page_array[].
5776 	 */
5777 	for (i = 0; i < npages; i++) {
5778 		/* Always test object consistency, including "skip_m". */
5779 		if (m[i].object != object)
5780 			return (false);
5781 		if (&m[i] == skip_m)
5782 			continue;
5783 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5784 			return (false);
5785 		if ((flags & PS_ALL_DIRTY) != 0) {
5786 			/*
5787 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5788 			 * might stop this case from spuriously returning
5789 			 * "false".  However, that would require a write lock
5790 			 * on the object containing "m[i]".
5791 			 */
5792 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5793 				return (false);
5794 		}
5795 		if ((flags & PS_ALL_VALID) != 0 &&
5796 		    m[i].valid != VM_PAGE_BITS_ALL)
5797 			return (false);
5798 	}
5799 	return (true);
5800 }
5801 
5802 /*
5803  * Set the page's dirty bits if the page is modified.
5804  */
5805 void
5806 vm_page_test_dirty(vm_page_t m)
5807 {
5808 
5809 	vm_page_assert_busied(m);
5810 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5811 		vm_page_dirty(m);
5812 }
5813 
5814 void
5815 vm_page_valid(vm_page_t m)
5816 {
5817 
5818 	vm_page_assert_busied(m);
5819 	if (vm_page_xbusied(m))
5820 		m->valid = VM_PAGE_BITS_ALL;
5821 	else
5822 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5823 }
5824 
5825 void
5826 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5827 {
5828 
5829 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5830 }
5831 
5832 void
5833 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5834 {
5835 
5836 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5837 }
5838 
5839 int
5840 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5841 {
5842 
5843 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5844 }
5845 
5846 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5847 void
5848 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5849 {
5850 
5851 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5852 }
5853 
5854 void
5855 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5856 {
5857 
5858 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5859 }
5860 #endif
5861 
5862 #ifdef INVARIANTS
5863 void
5864 vm_page_object_busy_assert(vm_page_t m)
5865 {
5866 
5867 	/*
5868 	 * Certain of the page's fields may only be modified by the
5869 	 * holder of a page or object busy.
5870 	 */
5871 	if (m->object != NULL && !vm_page_busied(m))
5872 		VM_OBJECT_ASSERT_BUSY(m->object);
5873 }
5874 
5875 void
5876 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5877 {
5878 
5879 	if ((bits & PGA_WRITEABLE) == 0)
5880 		return;
5881 
5882 	/*
5883 	 * The PGA_WRITEABLE flag can only be set if the page is
5884 	 * managed, is exclusively busied or the object is locked.
5885 	 * Currently, this flag is only set by pmap_enter().
5886 	 */
5887 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5888 	    ("PGA_WRITEABLE on unmanaged page"));
5889 	if (!vm_page_xbusied(m))
5890 		VM_OBJECT_ASSERT_BUSY(m->object);
5891 }
5892 #endif
5893 
5894 #include "opt_ddb.h"
5895 #ifdef DDB
5896 #include <sys/kernel.h>
5897 
5898 #include <ddb/ddb.h>
5899 
5900 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5901 {
5902 
5903 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5904 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5905 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5906 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5907 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5908 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5909 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5910 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5911 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5912 }
5913 
5914 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5915 {
5916 	int dom;
5917 
5918 	db_printf("pq_free %d\n", vm_free_count());
5919 	for (dom = 0; dom < vm_ndomains; dom++) {
5920 		db_printf(
5921     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5922 		    dom,
5923 		    vm_dom[dom].vmd_page_count,
5924 		    vm_dom[dom].vmd_free_count,
5925 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5926 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5927 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5928 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5929 	}
5930 }
5931 
5932 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5933 {
5934 	vm_page_t m;
5935 	boolean_t phys, virt;
5936 
5937 	if (!have_addr) {
5938 		db_printf("show pginfo addr\n");
5939 		return;
5940 	}
5941 
5942 	phys = strchr(modif, 'p') != NULL;
5943 	virt = strchr(modif, 'v') != NULL;
5944 	if (virt)
5945 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5946 	else if (phys)
5947 		m = PHYS_TO_VM_PAGE(addr);
5948 	else
5949 		m = (vm_page_t)addr;
5950 	db_printf(
5951     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5952     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5953 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5954 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5955 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5956 }
5957 #endif /* DDB */
5958