xref: /freebsd/sys/vm/vm_page.c (revision e95923a2288fb2845c7be4822b1b92b2fc106d18)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *			GENERAL RULES ON VM_PAGE MANIPULATION
67  *
68  *	- A page queue lock is required when adding or removing a page from a
69  *	  page queue regardless of other locks or the busy state of a page.
70  *
71  *		* In general, no thread besides the page daemon can acquire or
72  *		  hold more than one page queue lock at a time.
73  *
74  *		* The page daemon can acquire and hold any pair of page queue
75  *		  locks in any order.
76  *
77  *	- The object lock is required when inserting or removing
78  *	  pages from an object (vm_page_insert() or vm_page_remove()).
79  *
80  */
81 
82 /*
83  *	Resident memory management module.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 #include "opt_vm.h"
90 
91 #include <sys/param.h>
92 #include <sys/systm.h>
93 #include <sys/lock.h>
94 #include <sys/domainset.h>
95 #include <sys/kernel.h>
96 #include <sys/limits.h>
97 #include <sys/linker.h>
98 #include <sys/malloc.h>
99 #include <sys/mman.h>
100 #include <sys/msgbuf.h>
101 #include <sys/mutex.h>
102 #include <sys/proc.h>
103 #include <sys/rwlock.h>
104 #include <sys/sbuf.h>
105 #include <sys/sched.h>
106 #include <sys/smp.h>
107 #include <sys/sysctl.h>
108 #include <sys/vmmeter.h>
109 #include <sys/vnode.h>
110 
111 #include <vm/vm.h>
112 #include <vm/pmap.h>
113 #include <vm/vm_param.h>
114 #include <vm/vm_domainset.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_map.h>
117 #include <vm/vm_object.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_pageout.h>
120 #include <vm/vm_phys.h>
121 #include <vm/vm_pagequeue.h>
122 #include <vm/vm_pager.h>
123 #include <vm/vm_radix.h>
124 #include <vm/vm_reserv.h>
125 #include <vm/vm_extern.h>
126 #include <vm/uma.h>
127 #include <vm/uma_int.h>
128 
129 #include <machine/md_var.h>
130 
131 extern int	uma_startup_count(int);
132 extern void	uma_startup(void *, int);
133 extern int	vmem_startup_count(void);
134 
135 struct vm_domain vm_dom[MAXMEMDOM];
136 
137 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
138 
139 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
140 
141 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
142 /* The following fields are protected by the domainset lock. */
143 domainset_t __exclusive_cache_line vm_min_domains;
144 domainset_t __exclusive_cache_line vm_severe_domains;
145 static int vm_min_waiters;
146 static int vm_severe_waiters;
147 static int vm_pageproc_waiters;
148 
149 /*
150  * bogus page -- for I/O to/from partially complete buffers,
151  * or for paging into sparsely invalid regions.
152  */
153 vm_page_t bogus_page;
154 
155 vm_page_t vm_page_array;
156 long vm_page_array_size;
157 long first_page;
158 
159 static int boot_pages;
160 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
161     &boot_pages, 0,
162     "number of pages allocated for bootstrapping the VM system");
163 
164 static int pa_tryrelock_restart;
165 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
166     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
167 
168 static TAILQ_HEAD(, vm_page) blacklist_head;
169 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
170 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
171     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
172 
173 static uma_zone_t fakepg_zone;
174 
175 static void vm_page_alloc_check(vm_page_t m);
176 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
177 static void vm_page_dequeue_complete(vm_page_t m);
178 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
179 static void vm_page_init(void *dummy);
180 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
181     vm_pindex_t pindex, vm_page_t mpred);
182 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
183     vm_page_t mpred);
184 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
185     vm_page_t m_run, vm_paddr_t high);
186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
187     int req);
188 static int vm_page_import(void *arg, void **store, int cnt, int domain,
189     int flags);
190 static void vm_page_release(void *arg, void **store, int cnt);
191 
192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
193 
194 static void
195 vm_page_init(void *dummy)
196 {
197 
198 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
199 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
200 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
201 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
202 }
203 
204 /*
205  * The cache page zone is initialized later since we need to be able to allocate
206  * pages before UMA is fully initialized.
207  */
208 static void
209 vm_page_init_cache_zones(void *dummy __unused)
210 {
211 	struct vm_domain *vmd;
212 	int i;
213 
214 	for (i = 0; i < vm_ndomains; i++) {
215 		vmd = VM_DOMAIN(i);
216 		/*
217 		 * Don't allow the page cache to take up more than .25% of
218 		 * memory.
219 		 */
220 		if (vmd->vmd_page_count / 400 < 256 * mp_ncpus)
221 			continue;
222 		vmd->vmd_pgcache = uma_zcache_create("vm pgcache",
223 		    sizeof(struct vm_page), NULL, NULL, NULL, NULL,
224 		    vm_page_import, vm_page_release, vmd,
225 		    UMA_ZONE_MAXBUCKET | UMA_ZONE_VM);
226 		(void )uma_zone_set_maxcache(vmd->vmd_pgcache, 0);
227 	}
228 }
229 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
230 
231 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
232 #if PAGE_SIZE == 32768
233 #ifdef CTASSERT
234 CTASSERT(sizeof(u_long) >= 8);
235 #endif
236 #endif
237 
238 /*
239  * Try to acquire a physical address lock while a pmap is locked.  If we
240  * fail to trylock we unlock and lock the pmap directly and cache the
241  * locked pa in *locked.  The caller should then restart their loop in case
242  * the virtual to physical mapping has changed.
243  */
244 int
245 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
246 {
247 	vm_paddr_t lockpa;
248 
249 	lockpa = *locked;
250 	*locked = pa;
251 	if (lockpa) {
252 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
253 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
254 			return (0);
255 		PA_UNLOCK(lockpa);
256 	}
257 	if (PA_TRYLOCK(pa))
258 		return (0);
259 	PMAP_UNLOCK(pmap);
260 	atomic_add_int(&pa_tryrelock_restart, 1);
261 	PA_LOCK(pa);
262 	PMAP_LOCK(pmap);
263 	return (EAGAIN);
264 }
265 
266 /*
267  *	vm_set_page_size:
268  *
269  *	Sets the page size, perhaps based upon the memory
270  *	size.  Must be called before any use of page-size
271  *	dependent functions.
272  */
273 void
274 vm_set_page_size(void)
275 {
276 	if (vm_cnt.v_page_size == 0)
277 		vm_cnt.v_page_size = PAGE_SIZE;
278 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
279 		panic("vm_set_page_size: page size not a power of two");
280 }
281 
282 /*
283  *	vm_page_blacklist_next:
284  *
285  *	Find the next entry in the provided string of blacklist
286  *	addresses.  Entries are separated by space, comma, or newline.
287  *	If an invalid integer is encountered then the rest of the
288  *	string is skipped.  Updates the list pointer to the next
289  *	character, or NULL if the string is exhausted or invalid.
290  */
291 static vm_paddr_t
292 vm_page_blacklist_next(char **list, char *end)
293 {
294 	vm_paddr_t bad;
295 	char *cp, *pos;
296 
297 	if (list == NULL || *list == NULL)
298 		return (0);
299 	if (**list =='\0') {
300 		*list = NULL;
301 		return (0);
302 	}
303 
304 	/*
305 	 * If there's no end pointer then the buffer is coming from
306 	 * the kenv and we know it's null-terminated.
307 	 */
308 	if (end == NULL)
309 		end = *list + strlen(*list);
310 
311 	/* Ensure that strtoq() won't walk off the end */
312 	if (*end != '\0') {
313 		if (*end == '\n' || *end == ' ' || *end  == ',')
314 			*end = '\0';
315 		else {
316 			printf("Blacklist not terminated, skipping\n");
317 			*list = NULL;
318 			return (0);
319 		}
320 	}
321 
322 	for (pos = *list; *pos != '\0'; pos = cp) {
323 		bad = strtoq(pos, &cp, 0);
324 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
325 			if (bad == 0) {
326 				if (++cp < end)
327 					continue;
328 				else
329 					break;
330 			}
331 		} else
332 			break;
333 		if (*cp == '\0' || ++cp >= end)
334 			*list = NULL;
335 		else
336 			*list = cp;
337 		return (trunc_page(bad));
338 	}
339 	printf("Garbage in RAM blacklist, skipping\n");
340 	*list = NULL;
341 	return (0);
342 }
343 
344 bool
345 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
346 {
347 	struct vm_domain *vmd;
348 	vm_page_t m;
349 	int ret;
350 
351 	m = vm_phys_paddr_to_vm_page(pa);
352 	if (m == NULL)
353 		return (true); /* page does not exist, no failure */
354 
355 	vmd = vm_pagequeue_domain(m);
356 	vm_domain_free_lock(vmd);
357 	ret = vm_phys_unfree_page(m);
358 	vm_domain_free_unlock(vmd);
359 	if (ret != 0) {
360 		vm_domain_freecnt_inc(vmd, -1);
361 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
362 		if (verbose)
363 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
364 	}
365 	return (ret);
366 }
367 
368 /*
369  *	vm_page_blacklist_check:
370  *
371  *	Iterate through the provided string of blacklist addresses, pulling
372  *	each entry out of the physical allocator free list and putting it
373  *	onto a list for reporting via the vm.page_blacklist sysctl.
374  */
375 static void
376 vm_page_blacklist_check(char *list, char *end)
377 {
378 	vm_paddr_t pa;
379 	char *next;
380 
381 	next = list;
382 	while (next != NULL) {
383 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
384 			continue;
385 		vm_page_blacklist_add(pa, bootverbose);
386 	}
387 }
388 
389 /*
390  *	vm_page_blacklist_load:
391  *
392  *	Search for a special module named "ram_blacklist".  It'll be a
393  *	plain text file provided by the user via the loader directive
394  *	of the same name.
395  */
396 static void
397 vm_page_blacklist_load(char **list, char **end)
398 {
399 	void *mod;
400 	u_char *ptr;
401 	u_int len;
402 
403 	mod = NULL;
404 	ptr = NULL;
405 
406 	mod = preload_search_by_type("ram_blacklist");
407 	if (mod != NULL) {
408 		ptr = preload_fetch_addr(mod);
409 		len = preload_fetch_size(mod);
410         }
411 	*list = ptr;
412 	if (ptr != NULL)
413 		*end = ptr + len;
414 	else
415 		*end = NULL;
416 	return;
417 }
418 
419 static int
420 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
421 {
422 	vm_page_t m;
423 	struct sbuf sbuf;
424 	int error, first;
425 
426 	first = 1;
427 	error = sysctl_wire_old_buffer(req, 0);
428 	if (error != 0)
429 		return (error);
430 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
431 	TAILQ_FOREACH(m, &blacklist_head, listq) {
432 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
433 		    (uintmax_t)m->phys_addr);
434 		first = 0;
435 	}
436 	error = sbuf_finish(&sbuf);
437 	sbuf_delete(&sbuf);
438 	return (error);
439 }
440 
441 /*
442  * Initialize a dummy page for use in scans of the specified paging queue.
443  * In principle, this function only needs to set the flag PG_MARKER.
444  * Nonetheless, it write busies and initializes the hold count to one as
445  * safety precautions.
446  */
447 static void
448 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
449 {
450 
451 	bzero(marker, sizeof(*marker));
452 	marker->flags = PG_MARKER;
453 	marker->aflags = aflags;
454 	marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
455 	marker->queue = queue;
456 	marker->hold_count = 1;
457 }
458 
459 static void
460 vm_page_domain_init(int domain)
461 {
462 	struct vm_domain *vmd;
463 	struct vm_pagequeue *pq;
464 	int i;
465 
466 	vmd = VM_DOMAIN(domain);
467 	bzero(vmd, sizeof(*vmd));
468 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
469 	    "vm inactive pagequeue";
470 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
471 	    "vm active pagequeue";
472 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
473 	    "vm laundry pagequeue";
474 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
475 	    "vm unswappable pagequeue";
476 	vmd->vmd_domain = domain;
477 	vmd->vmd_page_count = 0;
478 	vmd->vmd_free_count = 0;
479 	vmd->vmd_segs = 0;
480 	vmd->vmd_oom = FALSE;
481 	for (i = 0; i < PQ_COUNT; i++) {
482 		pq = &vmd->vmd_pagequeues[i];
483 		TAILQ_INIT(&pq->pq_pl);
484 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
485 		    MTX_DEF | MTX_DUPOK);
486 		pq->pq_pdpages = 0;
487 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
488 	}
489 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
490 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
491 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
492 
493 	/*
494 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
495 	 * insertions.
496 	 */
497 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
498 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
499 	    &vmd->vmd_inacthead, plinks.q);
500 
501 	/*
502 	 * The clock pages are used to implement active queue scanning without
503 	 * requeues.  Scans start at clock[0], which is advanced after the scan
504 	 * ends.  When the two clock hands meet, they are reset and scanning
505 	 * resumes from the head of the queue.
506 	 */
507 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
508 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
509 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
510 	    &vmd->vmd_clock[0], plinks.q);
511 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
512 	    &vmd->vmd_clock[1], plinks.q);
513 }
514 
515 /*
516  * Initialize a physical page in preparation for adding it to the free
517  * lists.
518  */
519 static void
520 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
521 {
522 
523 	m->object = NULL;
524 	m->wire_count = 0;
525 	m->busy_lock = VPB_UNBUSIED;
526 	m->hold_count = 0;
527 	m->flags = m->aflags = 0;
528 	m->phys_addr = pa;
529 	m->queue = PQ_NONE;
530 	m->psind = 0;
531 	m->segind = segind;
532 	m->order = VM_NFREEORDER;
533 	m->pool = VM_FREEPOOL_DEFAULT;
534 	m->valid = m->dirty = 0;
535 	pmap_page_init(m);
536 }
537 
538 /*
539  *	vm_page_startup:
540  *
541  *	Initializes the resident memory module.  Allocates physical memory for
542  *	bootstrapping UMA and some data structures that are used to manage
543  *	physical pages.  Initializes these structures, and populates the free
544  *	page queues.
545  */
546 vm_offset_t
547 vm_page_startup(vm_offset_t vaddr)
548 {
549 	struct vm_phys_seg *seg;
550 	vm_page_t m;
551 	char *list, *listend;
552 	vm_offset_t mapped;
553 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
554 	vm_paddr_t biggestsize, last_pa, pa;
555 	u_long pagecount;
556 	int biggestone, i, segind;
557 #ifdef WITNESS
558 	int witness_size;
559 #endif
560 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
561 	long ii;
562 #endif
563 
564 	biggestsize = 0;
565 	biggestone = 0;
566 	vaddr = round_page(vaddr);
567 
568 	for (i = 0; phys_avail[i + 1]; i += 2) {
569 		phys_avail[i] = round_page(phys_avail[i]);
570 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
571 	}
572 	for (i = 0; phys_avail[i + 1]; i += 2) {
573 		size = phys_avail[i + 1] - phys_avail[i];
574 		if (size > biggestsize) {
575 			biggestone = i;
576 			biggestsize = size;
577 		}
578 	}
579 
580 	end = phys_avail[biggestone+1];
581 
582 	/*
583 	 * Initialize the page and queue locks.
584 	 */
585 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
586 	for (i = 0; i < PA_LOCK_COUNT; i++)
587 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
588 	for (i = 0; i < vm_ndomains; i++)
589 		vm_page_domain_init(i);
590 
591 	/*
592 	 * Allocate memory for use when boot strapping the kernel memory
593 	 * allocator.  Tell UMA how many zones we are going to create
594 	 * before going fully functional.  UMA will add its zones.
595 	 *
596 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
597 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
598 	 */
599 	boot_pages = uma_startup_count(8);
600 
601 #ifndef UMA_MD_SMALL_ALLOC
602 	/* vmem_startup() calls uma_prealloc(). */
603 	boot_pages += vmem_startup_count();
604 	/* vm_map_startup() calls uma_prealloc(). */
605 	boot_pages += howmany(MAX_KMAP,
606 	    UMA_SLAB_SPACE / sizeof(struct vm_map));
607 
608 	/*
609 	 * Before going fully functional kmem_init() does allocation
610 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
611 	 */
612 	boot_pages += 2;
613 #endif
614 	/*
615 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
616 	 * manually fetch the value.
617 	 */
618 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
619 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
620 	new_end = trunc_page(new_end);
621 	mapped = pmap_map(&vaddr, new_end, end,
622 	    VM_PROT_READ | VM_PROT_WRITE);
623 	bzero((void *)mapped, end - new_end);
624 	uma_startup((void *)mapped, boot_pages);
625 
626 #ifdef WITNESS
627 	witness_size = round_page(witness_startup_count());
628 	new_end -= witness_size;
629 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
630 	    VM_PROT_READ | VM_PROT_WRITE);
631 	bzero((void *)mapped, witness_size);
632 	witness_startup((void *)mapped);
633 #endif
634 
635 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
636     defined(__i386__) || defined(__mips__) || defined(__riscv)
637 	/*
638 	 * Allocate a bitmap to indicate that a random physical page
639 	 * needs to be included in a minidump.
640 	 *
641 	 * The amd64 port needs this to indicate which direct map pages
642 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
643 	 *
644 	 * However, i386 still needs this workspace internally within the
645 	 * minidump code.  In theory, they are not needed on i386, but are
646 	 * included should the sf_buf code decide to use them.
647 	 */
648 	last_pa = 0;
649 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
650 		if (dump_avail[i + 1] > last_pa)
651 			last_pa = dump_avail[i + 1];
652 	page_range = last_pa / PAGE_SIZE;
653 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
654 	new_end -= vm_page_dump_size;
655 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
656 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
657 	bzero((void *)vm_page_dump, vm_page_dump_size);
658 #else
659 	(void)last_pa;
660 #endif
661 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
662     defined(__riscv)
663 	/*
664 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
665 	 * in a crash dump.  When pmap_map() uses the direct map, they are
666 	 * not automatically included.
667 	 */
668 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
669 		dump_add_page(pa);
670 #endif
671 	phys_avail[biggestone + 1] = new_end;
672 #ifdef __amd64__
673 	/*
674 	 * Request that the physical pages underlying the message buffer be
675 	 * included in a crash dump.  Since the message buffer is accessed
676 	 * through the direct map, they are not automatically included.
677 	 */
678 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
679 	last_pa = pa + round_page(msgbufsize);
680 	while (pa < last_pa) {
681 		dump_add_page(pa);
682 		pa += PAGE_SIZE;
683 	}
684 #endif
685 	/*
686 	 * Compute the number of pages of memory that will be available for
687 	 * use, taking into account the overhead of a page structure per page.
688 	 * In other words, solve
689 	 *	"available physical memory" - round_page(page_range *
690 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
691 	 * for page_range.
692 	 */
693 	low_avail = phys_avail[0];
694 	high_avail = phys_avail[1];
695 	for (i = 0; i < vm_phys_nsegs; i++) {
696 		if (vm_phys_segs[i].start < low_avail)
697 			low_avail = vm_phys_segs[i].start;
698 		if (vm_phys_segs[i].end > high_avail)
699 			high_avail = vm_phys_segs[i].end;
700 	}
701 	/* Skip the first chunk.  It is already accounted for. */
702 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
703 		if (phys_avail[i] < low_avail)
704 			low_avail = phys_avail[i];
705 		if (phys_avail[i + 1] > high_avail)
706 			high_avail = phys_avail[i + 1];
707 	}
708 	first_page = low_avail / PAGE_SIZE;
709 #ifdef VM_PHYSSEG_SPARSE
710 	size = 0;
711 	for (i = 0; i < vm_phys_nsegs; i++)
712 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
713 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
714 		size += phys_avail[i + 1] - phys_avail[i];
715 #elif defined(VM_PHYSSEG_DENSE)
716 	size = high_avail - low_avail;
717 #else
718 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
719 #endif
720 
721 #ifdef VM_PHYSSEG_DENSE
722 	/*
723 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
724 	 * the overhead of a page structure per page only if vm_page_array is
725 	 * allocated from the last physical memory chunk.  Otherwise, we must
726 	 * allocate page structures representing the physical memory
727 	 * underlying vm_page_array, even though they will not be used.
728 	 */
729 	if (new_end != high_avail)
730 		page_range = size / PAGE_SIZE;
731 	else
732 #endif
733 	{
734 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
735 
736 		/*
737 		 * If the partial bytes remaining are large enough for
738 		 * a page (PAGE_SIZE) without a corresponding
739 		 * 'struct vm_page', then new_end will contain an
740 		 * extra page after subtracting the length of the VM
741 		 * page array.  Compensate by subtracting an extra
742 		 * page from new_end.
743 		 */
744 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
745 			if (new_end == high_avail)
746 				high_avail -= PAGE_SIZE;
747 			new_end -= PAGE_SIZE;
748 		}
749 	}
750 	end = new_end;
751 
752 	/*
753 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
754 	 * However, because this page is allocated from KVM, out-of-bounds
755 	 * accesses using the direct map will not be trapped.
756 	 */
757 	vaddr += PAGE_SIZE;
758 
759 	/*
760 	 * Allocate physical memory for the page structures, and map it.
761 	 */
762 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
763 	mapped = pmap_map(&vaddr, new_end, end,
764 	    VM_PROT_READ | VM_PROT_WRITE);
765 	vm_page_array = (vm_page_t)mapped;
766 	vm_page_array_size = page_range;
767 
768 #if VM_NRESERVLEVEL > 0
769 	/*
770 	 * Allocate physical memory for the reservation management system's
771 	 * data structures, and map it.
772 	 */
773 	if (high_avail == end)
774 		high_avail = new_end;
775 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
776 #endif
777 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
778     defined(__riscv)
779 	/*
780 	 * Include vm_page_array and vm_reserv_array in a crash dump.
781 	 */
782 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
783 		dump_add_page(pa);
784 #endif
785 	phys_avail[biggestone + 1] = new_end;
786 
787 	/*
788 	 * Add physical memory segments corresponding to the available
789 	 * physical pages.
790 	 */
791 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
792 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
793 
794 	/*
795 	 * Initialize the physical memory allocator.
796 	 */
797 	vm_phys_init();
798 
799 	/*
800 	 * Initialize the page structures and add every available page to the
801 	 * physical memory allocator's free lists.
802 	 */
803 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
804 	for (ii = 0; ii < vm_page_array_size; ii++) {
805 		m = &vm_page_array[ii];
806 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
807 		m->flags = PG_FICTITIOUS;
808 	}
809 #endif
810 	vm_cnt.v_page_count = 0;
811 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
812 		seg = &vm_phys_segs[segind];
813 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
814 		    m++, pa += PAGE_SIZE)
815 			vm_page_init_page(m, pa, segind);
816 
817 		/*
818 		 * Add the segment to the free lists only if it is covered by
819 		 * one of the ranges in phys_avail.  Because we've added the
820 		 * ranges to the vm_phys_segs array, we can assume that each
821 		 * segment is either entirely contained in one of the ranges,
822 		 * or doesn't overlap any of them.
823 		 */
824 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
825 			struct vm_domain *vmd;
826 
827 			if (seg->start < phys_avail[i] ||
828 			    seg->end > phys_avail[i + 1])
829 				continue;
830 
831 			m = seg->first_page;
832 			pagecount = (u_long)atop(seg->end - seg->start);
833 
834 			vmd = VM_DOMAIN(seg->domain);
835 			vm_domain_free_lock(vmd);
836 			vm_phys_free_contig(m, pagecount);
837 			vm_domain_free_unlock(vmd);
838 			vm_domain_freecnt_inc(vmd, pagecount);
839 			vm_cnt.v_page_count += (u_int)pagecount;
840 
841 			vmd = VM_DOMAIN(seg->domain);
842 			vmd->vmd_page_count += (u_int)pagecount;
843 			vmd->vmd_segs |= 1UL << m->segind;
844 			break;
845 		}
846 	}
847 
848 	/*
849 	 * Remove blacklisted pages from the physical memory allocator.
850 	 */
851 	TAILQ_INIT(&blacklist_head);
852 	vm_page_blacklist_load(&list, &listend);
853 	vm_page_blacklist_check(list, listend);
854 
855 	list = kern_getenv("vm.blacklist");
856 	vm_page_blacklist_check(list, NULL);
857 
858 	freeenv(list);
859 #if VM_NRESERVLEVEL > 0
860 	/*
861 	 * Initialize the reservation management system.
862 	 */
863 	vm_reserv_init();
864 #endif
865 
866 	return (vaddr);
867 }
868 
869 void
870 vm_page_reference(vm_page_t m)
871 {
872 
873 	vm_page_aflag_set(m, PGA_REFERENCED);
874 }
875 
876 /*
877  *	vm_page_busy_downgrade:
878  *
879  *	Downgrade an exclusive busy page into a single shared busy page.
880  */
881 void
882 vm_page_busy_downgrade(vm_page_t m)
883 {
884 	u_int x;
885 	bool locked;
886 
887 	vm_page_assert_xbusied(m);
888 	locked = mtx_owned(vm_page_lockptr(m));
889 
890 	for (;;) {
891 		x = m->busy_lock;
892 		x &= VPB_BIT_WAITERS;
893 		if (x != 0 && !locked)
894 			vm_page_lock(m);
895 		if (atomic_cmpset_rel_int(&m->busy_lock,
896 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
897 			break;
898 		if (x != 0 && !locked)
899 			vm_page_unlock(m);
900 	}
901 	if (x != 0) {
902 		wakeup(m);
903 		if (!locked)
904 			vm_page_unlock(m);
905 	}
906 }
907 
908 /*
909  *	vm_page_sbusied:
910  *
911  *	Return a positive value if the page is shared busied, 0 otherwise.
912  */
913 int
914 vm_page_sbusied(vm_page_t m)
915 {
916 	u_int x;
917 
918 	x = m->busy_lock;
919 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
920 }
921 
922 /*
923  *	vm_page_sunbusy:
924  *
925  *	Shared unbusy a page.
926  */
927 void
928 vm_page_sunbusy(vm_page_t m)
929 {
930 	u_int x;
931 
932 	vm_page_lock_assert(m, MA_NOTOWNED);
933 	vm_page_assert_sbusied(m);
934 
935 	for (;;) {
936 		x = m->busy_lock;
937 		if (VPB_SHARERS(x) > 1) {
938 			if (atomic_cmpset_int(&m->busy_lock, x,
939 			    x - VPB_ONE_SHARER))
940 				break;
941 			continue;
942 		}
943 		if ((x & VPB_BIT_WAITERS) == 0) {
944 			KASSERT(x == VPB_SHARERS_WORD(1),
945 			    ("vm_page_sunbusy: invalid lock state"));
946 			if (atomic_cmpset_int(&m->busy_lock,
947 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
948 				break;
949 			continue;
950 		}
951 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
952 		    ("vm_page_sunbusy: invalid lock state for waiters"));
953 
954 		vm_page_lock(m);
955 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
956 			vm_page_unlock(m);
957 			continue;
958 		}
959 		wakeup(m);
960 		vm_page_unlock(m);
961 		break;
962 	}
963 }
964 
965 /*
966  *	vm_page_busy_sleep:
967  *
968  *	Sleep and release the page lock, using the page pointer as wchan.
969  *	This is used to implement the hard-path of busying mechanism.
970  *
971  *	The given page must be locked.
972  *
973  *	If nonshared is true, sleep only if the page is xbusy.
974  */
975 void
976 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
977 {
978 	u_int x;
979 
980 	vm_page_assert_locked(m);
981 
982 	x = m->busy_lock;
983 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
984 	    ((x & VPB_BIT_WAITERS) == 0 &&
985 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
986 		vm_page_unlock(m);
987 		return;
988 	}
989 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
990 }
991 
992 /*
993  *	vm_page_trysbusy:
994  *
995  *	Try to shared busy a page.
996  *	If the operation succeeds 1 is returned otherwise 0.
997  *	The operation never sleeps.
998  */
999 int
1000 vm_page_trysbusy(vm_page_t m)
1001 {
1002 	u_int x;
1003 
1004 	for (;;) {
1005 		x = m->busy_lock;
1006 		if ((x & VPB_BIT_SHARED) == 0)
1007 			return (0);
1008 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
1009 			return (1);
1010 	}
1011 }
1012 
1013 static void
1014 vm_page_xunbusy_locked(vm_page_t m)
1015 {
1016 
1017 	vm_page_assert_xbusied(m);
1018 	vm_page_assert_locked(m);
1019 
1020 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1021 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
1022 	wakeup(m);
1023 }
1024 
1025 void
1026 vm_page_xunbusy_maybelocked(vm_page_t m)
1027 {
1028 	bool lockacq;
1029 
1030 	vm_page_assert_xbusied(m);
1031 
1032 	/*
1033 	 * Fast path for unbusy.  If it succeeds, we know that there
1034 	 * are no waiters, so we do not need a wakeup.
1035 	 */
1036 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
1037 	    VPB_UNBUSIED))
1038 		return;
1039 
1040 	lockacq = !mtx_owned(vm_page_lockptr(m));
1041 	if (lockacq)
1042 		vm_page_lock(m);
1043 	vm_page_xunbusy_locked(m);
1044 	if (lockacq)
1045 		vm_page_unlock(m);
1046 }
1047 
1048 /*
1049  *	vm_page_xunbusy_hard:
1050  *
1051  *	Called after the first try the exclusive unbusy of a page failed.
1052  *	It is assumed that the waiters bit is on.
1053  */
1054 void
1055 vm_page_xunbusy_hard(vm_page_t m)
1056 {
1057 
1058 	vm_page_assert_xbusied(m);
1059 
1060 	vm_page_lock(m);
1061 	vm_page_xunbusy_locked(m);
1062 	vm_page_unlock(m);
1063 }
1064 
1065 /*
1066  *	vm_page_flash:
1067  *
1068  *	Wakeup anyone waiting for the page.
1069  *	The ownership bits do not change.
1070  *
1071  *	The given page must be locked.
1072  */
1073 void
1074 vm_page_flash(vm_page_t m)
1075 {
1076 	u_int x;
1077 
1078 	vm_page_lock_assert(m, MA_OWNED);
1079 
1080 	for (;;) {
1081 		x = m->busy_lock;
1082 		if ((x & VPB_BIT_WAITERS) == 0)
1083 			return;
1084 		if (atomic_cmpset_int(&m->busy_lock, x,
1085 		    x & (~VPB_BIT_WAITERS)))
1086 			break;
1087 	}
1088 	wakeup(m);
1089 }
1090 
1091 /*
1092  * Avoid releasing and reacquiring the same page lock.
1093  */
1094 void
1095 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1096 {
1097 	struct mtx *mtx1;
1098 
1099 	mtx1 = vm_page_lockptr(m);
1100 	if (*mtx == mtx1)
1101 		return;
1102 	if (*mtx != NULL)
1103 		mtx_unlock(*mtx);
1104 	*mtx = mtx1;
1105 	mtx_lock(mtx1);
1106 }
1107 
1108 /*
1109  * Keep page from being freed by the page daemon
1110  * much of the same effect as wiring, except much lower
1111  * overhead and should be used only for *very* temporary
1112  * holding ("wiring").
1113  */
1114 void
1115 vm_page_hold(vm_page_t mem)
1116 {
1117 
1118 	vm_page_lock_assert(mem, MA_OWNED);
1119         mem->hold_count++;
1120 }
1121 
1122 void
1123 vm_page_unhold(vm_page_t mem)
1124 {
1125 
1126 	vm_page_lock_assert(mem, MA_OWNED);
1127 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1128 	--mem->hold_count;
1129 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1130 		vm_page_free_toq(mem);
1131 }
1132 
1133 /*
1134  *	vm_page_unhold_pages:
1135  *
1136  *	Unhold each of the pages that is referenced by the given array.
1137  */
1138 void
1139 vm_page_unhold_pages(vm_page_t *ma, int count)
1140 {
1141 	struct mtx *mtx;
1142 
1143 	mtx = NULL;
1144 	for (; count != 0; count--) {
1145 		vm_page_change_lock(*ma, &mtx);
1146 		vm_page_unhold(*ma);
1147 		ma++;
1148 	}
1149 	if (mtx != NULL)
1150 		mtx_unlock(mtx);
1151 }
1152 
1153 vm_page_t
1154 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1155 {
1156 	vm_page_t m;
1157 
1158 #ifdef VM_PHYSSEG_SPARSE
1159 	m = vm_phys_paddr_to_vm_page(pa);
1160 	if (m == NULL)
1161 		m = vm_phys_fictitious_to_vm_page(pa);
1162 	return (m);
1163 #elif defined(VM_PHYSSEG_DENSE)
1164 	long pi;
1165 
1166 	pi = atop(pa);
1167 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1168 		m = &vm_page_array[pi - first_page];
1169 		return (m);
1170 	}
1171 	return (vm_phys_fictitious_to_vm_page(pa));
1172 #else
1173 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1174 #endif
1175 }
1176 
1177 /*
1178  *	vm_page_getfake:
1179  *
1180  *	Create a fictitious page with the specified physical address and
1181  *	memory attribute.  The memory attribute is the only the machine-
1182  *	dependent aspect of a fictitious page that must be initialized.
1183  */
1184 vm_page_t
1185 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1186 {
1187 	vm_page_t m;
1188 
1189 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1190 	vm_page_initfake(m, paddr, memattr);
1191 	return (m);
1192 }
1193 
1194 void
1195 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1196 {
1197 
1198 	if ((m->flags & PG_FICTITIOUS) != 0) {
1199 		/*
1200 		 * The page's memattr might have changed since the
1201 		 * previous initialization.  Update the pmap to the
1202 		 * new memattr.
1203 		 */
1204 		goto memattr;
1205 	}
1206 	m->phys_addr = paddr;
1207 	m->queue = PQ_NONE;
1208 	/* Fictitious pages don't use "segind". */
1209 	m->flags = PG_FICTITIOUS;
1210 	/* Fictitious pages don't use "order" or "pool". */
1211 	m->oflags = VPO_UNMANAGED;
1212 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1213 	m->wire_count = 1;
1214 	pmap_page_init(m);
1215 memattr:
1216 	pmap_page_set_memattr(m, memattr);
1217 }
1218 
1219 /*
1220  *	vm_page_putfake:
1221  *
1222  *	Release a fictitious page.
1223  */
1224 void
1225 vm_page_putfake(vm_page_t m)
1226 {
1227 
1228 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1229 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1230 	    ("vm_page_putfake: bad page %p", m));
1231 	uma_zfree(fakepg_zone, m);
1232 }
1233 
1234 /*
1235  *	vm_page_updatefake:
1236  *
1237  *	Update the given fictitious page to the specified physical address and
1238  *	memory attribute.
1239  */
1240 void
1241 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1242 {
1243 
1244 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1245 	    ("vm_page_updatefake: bad page %p", m));
1246 	m->phys_addr = paddr;
1247 	pmap_page_set_memattr(m, memattr);
1248 }
1249 
1250 /*
1251  *	vm_page_free:
1252  *
1253  *	Free a page.
1254  */
1255 void
1256 vm_page_free(vm_page_t m)
1257 {
1258 
1259 	m->flags &= ~PG_ZERO;
1260 	vm_page_free_toq(m);
1261 }
1262 
1263 /*
1264  *	vm_page_free_zero:
1265  *
1266  *	Free a page to the zerod-pages queue
1267  */
1268 void
1269 vm_page_free_zero(vm_page_t m)
1270 {
1271 
1272 	m->flags |= PG_ZERO;
1273 	vm_page_free_toq(m);
1274 }
1275 
1276 /*
1277  * Unbusy and handle the page queueing for a page from a getpages request that
1278  * was optionally read ahead or behind.
1279  */
1280 void
1281 vm_page_readahead_finish(vm_page_t m)
1282 {
1283 
1284 	/* We shouldn't put invalid pages on queues. */
1285 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1286 
1287 	/*
1288 	 * Since the page is not the actually needed one, whether it should
1289 	 * be activated or deactivated is not obvious.  Empirical results
1290 	 * have shown that deactivating the page is usually the best choice,
1291 	 * unless the page is wanted by another thread.
1292 	 */
1293 	vm_page_lock(m);
1294 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1295 		vm_page_activate(m);
1296 	else
1297 		vm_page_deactivate(m);
1298 	vm_page_unlock(m);
1299 	vm_page_xunbusy(m);
1300 }
1301 
1302 /*
1303  *	vm_page_sleep_if_busy:
1304  *
1305  *	Sleep and release the page queues lock if the page is busied.
1306  *	Returns TRUE if the thread slept.
1307  *
1308  *	The given page must be unlocked and object containing it must
1309  *	be locked.
1310  */
1311 int
1312 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1313 {
1314 	vm_object_t obj;
1315 
1316 	vm_page_lock_assert(m, MA_NOTOWNED);
1317 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1318 
1319 	if (vm_page_busied(m)) {
1320 		/*
1321 		 * The page-specific object must be cached because page
1322 		 * identity can change during the sleep, causing the
1323 		 * re-lock of a different object.
1324 		 * It is assumed that a reference to the object is already
1325 		 * held by the callers.
1326 		 */
1327 		obj = m->object;
1328 		vm_page_lock(m);
1329 		VM_OBJECT_WUNLOCK(obj);
1330 		vm_page_busy_sleep(m, msg, false);
1331 		VM_OBJECT_WLOCK(obj);
1332 		return (TRUE);
1333 	}
1334 	return (FALSE);
1335 }
1336 
1337 /*
1338  *	vm_page_dirty_KBI:		[ internal use only ]
1339  *
1340  *	Set all bits in the page's dirty field.
1341  *
1342  *	The object containing the specified page must be locked if the
1343  *	call is made from the machine-independent layer.
1344  *
1345  *	See vm_page_clear_dirty_mask().
1346  *
1347  *	This function should only be called by vm_page_dirty().
1348  */
1349 void
1350 vm_page_dirty_KBI(vm_page_t m)
1351 {
1352 
1353 	/* Refer to this operation by its public name. */
1354 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1355 	    ("vm_page_dirty: page is invalid!"));
1356 	m->dirty = VM_PAGE_BITS_ALL;
1357 }
1358 
1359 /*
1360  *	vm_page_insert:		[ internal use only ]
1361  *
1362  *	Inserts the given mem entry into the object and object list.
1363  *
1364  *	The object must be locked.
1365  */
1366 int
1367 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1368 {
1369 	vm_page_t mpred;
1370 
1371 	VM_OBJECT_ASSERT_WLOCKED(object);
1372 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1373 	return (vm_page_insert_after(m, object, pindex, mpred));
1374 }
1375 
1376 /*
1377  *	vm_page_insert_after:
1378  *
1379  *	Inserts the page "m" into the specified object at offset "pindex".
1380  *
1381  *	The page "mpred" must immediately precede the offset "pindex" within
1382  *	the specified object.
1383  *
1384  *	The object must be locked.
1385  */
1386 static int
1387 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1388     vm_page_t mpred)
1389 {
1390 	vm_page_t msucc;
1391 
1392 	VM_OBJECT_ASSERT_WLOCKED(object);
1393 	KASSERT(m->object == NULL,
1394 	    ("vm_page_insert_after: page already inserted"));
1395 	if (mpred != NULL) {
1396 		KASSERT(mpred->object == object,
1397 		    ("vm_page_insert_after: object doesn't contain mpred"));
1398 		KASSERT(mpred->pindex < pindex,
1399 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1400 		msucc = TAILQ_NEXT(mpred, listq);
1401 	} else
1402 		msucc = TAILQ_FIRST(&object->memq);
1403 	if (msucc != NULL)
1404 		KASSERT(msucc->pindex > pindex,
1405 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1406 
1407 	/*
1408 	 * Record the object/offset pair in this page
1409 	 */
1410 	m->object = object;
1411 	m->pindex = pindex;
1412 
1413 	/*
1414 	 * Now link into the object's ordered list of backed pages.
1415 	 */
1416 	if (vm_radix_insert(&object->rtree, m)) {
1417 		m->object = NULL;
1418 		m->pindex = 0;
1419 		return (1);
1420 	}
1421 	vm_page_insert_radixdone(m, object, mpred);
1422 	return (0);
1423 }
1424 
1425 /*
1426  *	vm_page_insert_radixdone:
1427  *
1428  *	Complete page "m" insertion into the specified object after the
1429  *	radix trie hooking.
1430  *
1431  *	The page "mpred" must precede the offset "m->pindex" within the
1432  *	specified object.
1433  *
1434  *	The object must be locked.
1435  */
1436 static void
1437 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1438 {
1439 
1440 	VM_OBJECT_ASSERT_WLOCKED(object);
1441 	KASSERT(object != NULL && m->object == object,
1442 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1443 	if (mpred != NULL) {
1444 		KASSERT(mpred->object == object,
1445 		    ("vm_page_insert_after: object doesn't contain mpred"));
1446 		KASSERT(mpred->pindex < m->pindex,
1447 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1448 	}
1449 
1450 	if (mpred != NULL)
1451 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1452 	else
1453 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1454 
1455 	/*
1456 	 * Show that the object has one more resident page.
1457 	 */
1458 	object->resident_page_count++;
1459 
1460 	/*
1461 	 * Hold the vnode until the last page is released.
1462 	 */
1463 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1464 		vhold(object->handle);
1465 
1466 	/*
1467 	 * Since we are inserting a new and possibly dirty page,
1468 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1469 	 */
1470 	if (pmap_page_is_write_mapped(m))
1471 		vm_object_set_writeable_dirty(object);
1472 }
1473 
1474 /*
1475  *	vm_page_remove:
1476  *
1477  *	Removes the specified page from its containing object, but does not
1478  *	invalidate any backing storage.
1479  *
1480  *	The object must be locked.  The page must be locked if it is managed.
1481  */
1482 void
1483 vm_page_remove(vm_page_t m)
1484 {
1485 	vm_object_t object;
1486 	vm_page_t mrem;
1487 
1488 	if ((m->oflags & VPO_UNMANAGED) == 0)
1489 		vm_page_assert_locked(m);
1490 	if ((object = m->object) == NULL)
1491 		return;
1492 	VM_OBJECT_ASSERT_WLOCKED(object);
1493 	if (vm_page_xbusied(m))
1494 		vm_page_xunbusy_maybelocked(m);
1495 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1496 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1497 
1498 	/*
1499 	 * Now remove from the object's list of backed pages.
1500 	 */
1501 	TAILQ_REMOVE(&object->memq, m, listq);
1502 
1503 	/*
1504 	 * And show that the object has one fewer resident page.
1505 	 */
1506 	object->resident_page_count--;
1507 
1508 	/*
1509 	 * The vnode may now be recycled.
1510 	 */
1511 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1512 		vdrop(object->handle);
1513 
1514 	m->object = NULL;
1515 }
1516 
1517 /*
1518  *	vm_page_lookup:
1519  *
1520  *	Returns the page associated with the object/offset
1521  *	pair specified; if none is found, NULL is returned.
1522  *
1523  *	The object must be locked.
1524  */
1525 vm_page_t
1526 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1527 {
1528 
1529 	VM_OBJECT_ASSERT_LOCKED(object);
1530 	return (vm_radix_lookup(&object->rtree, pindex));
1531 }
1532 
1533 /*
1534  *	vm_page_find_least:
1535  *
1536  *	Returns the page associated with the object with least pindex
1537  *	greater than or equal to the parameter pindex, or NULL.
1538  *
1539  *	The object must be locked.
1540  */
1541 vm_page_t
1542 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1543 {
1544 	vm_page_t m;
1545 
1546 	VM_OBJECT_ASSERT_LOCKED(object);
1547 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1548 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1549 	return (m);
1550 }
1551 
1552 /*
1553  * Returns the given page's successor (by pindex) within the object if it is
1554  * resident; if none is found, NULL is returned.
1555  *
1556  * The object must be locked.
1557  */
1558 vm_page_t
1559 vm_page_next(vm_page_t m)
1560 {
1561 	vm_page_t next;
1562 
1563 	VM_OBJECT_ASSERT_LOCKED(m->object);
1564 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1565 		MPASS(next->object == m->object);
1566 		if (next->pindex != m->pindex + 1)
1567 			next = NULL;
1568 	}
1569 	return (next);
1570 }
1571 
1572 /*
1573  * Returns the given page's predecessor (by pindex) within the object if it is
1574  * resident; if none is found, NULL is returned.
1575  *
1576  * The object must be locked.
1577  */
1578 vm_page_t
1579 vm_page_prev(vm_page_t m)
1580 {
1581 	vm_page_t prev;
1582 
1583 	VM_OBJECT_ASSERT_LOCKED(m->object);
1584 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1585 		MPASS(prev->object == m->object);
1586 		if (prev->pindex != m->pindex - 1)
1587 			prev = NULL;
1588 	}
1589 	return (prev);
1590 }
1591 
1592 /*
1593  * Uses the page mnew as a replacement for an existing page at index
1594  * pindex which must be already present in the object.
1595  *
1596  * The existing page must not be on a paging queue.
1597  */
1598 vm_page_t
1599 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1600 {
1601 	vm_page_t mold;
1602 
1603 	VM_OBJECT_ASSERT_WLOCKED(object);
1604 	KASSERT(mnew->object == NULL,
1605 	    ("vm_page_replace: page %p already in object", mnew));
1606 	KASSERT(mnew->queue == PQ_NONE,
1607 	    ("vm_page_replace: new page %p is on a paging queue", mnew));
1608 
1609 	/*
1610 	 * This function mostly follows vm_page_insert() and
1611 	 * vm_page_remove() without the radix, object count and vnode
1612 	 * dance.  Double check such functions for more comments.
1613 	 */
1614 
1615 	mnew->object = object;
1616 	mnew->pindex = pindex;
1617 	mold = vm_radix_replace(&object->rtree, mnew);
1618 	KASSERT(mold->queue == PQ_NONE,
1619 	    ("vm_page_replace: old page %p is on a paging queue", mold));
1620 
1621 	/* Keep the resident page list in sorted order. */
1622 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1623 	TAILQ_REMOVE(&object->memq, mold, listq);
1624 
1625 	mold->object = NULL;
1626 	vm_page_xunbusy_maybelocked(mold);
1627 
1628 	/*
1629 	 * The object's resident_page_count does not change because we have
1630 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1631 	 */
1632 	if (pmap_page_is_write_mapped(mnew))
1633 		vm_object_set_writeable_dirty(object);
1634 	return (mold);
1635 }
1636 
1637 /*
1638  *	vm_page_rename:
1639  *
1640  *	Move the given memory entry from its
1641  *	current object to the specified target object/offset.
1642  *
1643  *	Note: swap associated with the page must be invalidated by the move.  We
1644  *	      have to do this for several reasons:  (1) we aren't freeing the
1645  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1646  *	      moving the page from object A to B, and will then later move
1647  *	      the backing store from A to B and we can't have a conflict.
1648  *
1649  *	Note: we *always* dirty the page.  It is necessary both for the
1650  *	      fact that we moved it, and because we may be invalidating
1651  *	      swap.
1652  *
1653  *	The objects must be locked.
1654  */
1655 int
1656 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1657 {
1658 	vm_page_t mpred;
1659 	vm_pindex_t opidx;
1660 
1661 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1662 
1663 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1664 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1665 	    ("vm_page_rename: pindex already renamed"));
1666 
1667 	/*
1668 	 * Create a custom version of vm_page_insert() which does not depend
1669 	 * by m_prev and can cheat on the implementation aspects of the
1670 	 * function.
1671 	 */
1672 	opidx = m->pindex;
1673 	m->pindex = new_pindex;
1674 	if (vm_radix_insert(&new_object->rtree, m)) {
1675 		m->pindex = opidx;
1676 		return (1);
1677 	}
1678 
1679 	/*
1680 	 * The operation cannot fail anymore.  The removal must happen before
1681 	 * the listq iterator is tainted.
1682 	 */
1683 	m->pindex = opidx;
1684 	vm_page_lock(m);
1685 	vm_page_remove(m);
1686 
1687 	/* Return back to the new pindex to complete vm_page_insert(). */
1688 	m->pindex = new_pindex;
1689 	m->object = new_object;
1690 	vm_page_unlock(m);
1691 	vm_page_insert_radixdone(m, new_object, mpred);
1692 	vm_page_dirty(m);
1693 	return (0);
1694 }
1695 
1696 /*
1697  *	vm_page_alloc:
1698  *
1699  *	Allocate and return a page that is associated with the specified
1700  *	object and offset pair.  By default, this page is exclusive busied.
1701  *
1702  *	The caller must always specify an allocation class.
1703  *
1704  *	allocation classes:
1705  *	VM_ALLOC_NORMAL		normal process request
1706  *	VM_ALLOC_SYSTEM		system *really* needs a page
1707  *	VM_ALLOC_INTERRUPT	interrupt time request
1708  *
1709  *	optional allocation flags:
1710  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1711  *				intends to allocate
1712  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1713  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1714  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1715  *				should not be exclusive busy
1716  *	VM_ALLOC_SBUSY		shared busy the allocated page
1717  *	VM_ALLOC_WIRED		wire the allocated page
1718  *	VM_ALLOC_ZERO		prefer a zeroed page
1719  */
1720 vm_page_t
1721 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1722 {
1723 
1724 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1725 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1726 }
1727 
1728 vm_page_t
1729 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1730     int req)
1731 {
1732 
1733 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1734 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1735 	    NULL));
1736 }
1737 
1738 /*
1739  * Allocate a page in the specified object with the given page index.  To
1740  * optimize insertion of the page into the object, the caller must also specifiy
1741  * the resident page in the object with largest index smaller than the given
1742  * page index, or NULL if no such page exists.
1743  */
1744 vm_page_t
1745 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1746     int req, vm_page_t mpred)
1747 {
1748 	struct vm_domainset_iter di;
1749 	vm_page_t m;
1750 	int domain;
1751 
1752 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1753 	do {
1754 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1755 		    mpred);
1756 		if (m != NULL)
1757 			break;
1758 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1759 
1760 	return (m);
1761 }
1762 
1763 /*
1764  * Returns true if the number of free pages exceeds the minimum
1765  * for the request class and false otherwise.
1766  */
1767 int
1768 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1769 {
1770 	u_int limit, old, new;
1771 
1772 	req = req & VM_ALLOC_CLASS_MASK;
1773 
1774 	/*
1775 	 * The page daemon is allowed to dig deeper into the free page list.
1776 	 */
1777 	if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1778 		req = VM_ALLOC_SYSTEM;
1779 	if (req == VM_ALLOC_INTERRUPT)
1780 		limit = 0;
1781 	else if (req == VM_ALLOC_SYSTEM)
1782 		limit = vmd->vmd_interrupt_free_min;
1783 	else
1784 		limit = vmd->vmd_free_reserved;
1785 
1786 	/*
1787 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1788 	 */
1789 	limit += npages;
1790 	old = vmd->vmd_free_count;
1791 	do {
1792 		if (old < limit)
1793 			return (0);
1794 		new = old - npages;
1795 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1796 
1797 	/* Wake the page daemon if we've crossed the threshold. */
1798 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1799 		pagedaemon_wakeup(vmd->vmd_domain);
1800 
1801 	/* Only update bitsets on transitions. */
1802 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1803 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1804 		vm_domain_set(vmd);
1805 
1806 	return (1);
1807 }
1808 
1809 vm_page_t
1810 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1811     int req, vm_page_t mpred)
1812 {
1813 	struct vm_domain *vmd;
1814 	vm_page_t m;
1815 	int flags;
1816 
1817 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1818 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1819 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1820 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1821 	    ("inconsistent object(%p)/req(%x)", object, req));
1822 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1823 	    ("Can't sleep and retry object insertion."));
1824 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1825 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1826 	    (uintmax_t)pindex));
1827 	if (object != NULL)
1828 		VM_OBJECT_ASSERT_WLOCKED(object);
1829 
1830 again:
1831 	m = NULL;
1832 #if VM_NRESERVLEVEL > 0
1833 	/*
1834 	 * Can we allocate the page from a reservation?
1835 	 */
1836 	if (vm_object_reserv(object) &&
1837 	    ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL ||
1838 	    (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) {
1839 		domain = vm_phys_domain(m);
1840 		vmd = VM_DOMAIN(domain);
1841 		goto found;
1842 	}
1843 #endif
1844 	vmd = VM_DOMAIN(domain);
1845 	if (object != NULL && vmd->vmd_pgcache != NULL) {
1846 		m = uma_zalloc(vmd->vmd_pgcache, M_NOWAIT);
1847 		if (m != NULL)
1848 			goto found;
1849 	}
1850 	if (vm_domain_allocate(vmd, req, 1)) {
1851 		/*
1852 		 * If not, allocate it from the free page queues.
1853 		 */
1854 		vm_domain_free_lock(vmd);
1855 		m = vm_phys_alloc_pages(domain, object != NULL ?
1856 		    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1857 		vm_domain_free_unlock(vmd);
1858 		if (m == NULL) {
1859 			vm_domain_freecnt_inc(vmd, 1);
1860 #if VM_NRESERVLEVEL > 0
1861 			if (vm_reserv_reclaim_inactive(domain))
1862 				goto again;
1863 #endif
1864 		}
1865 	}
1866 	if (m == NULL) {
1867 		/*
1868 		 * Not allocatable, give up.
1869 		 */
1870 		if (vm_domain_alloc_fail(vmd, object, req))
1871 			goto again;
1872 		return (NULL);
1873 	}
1874 
1875 	/*
1876 	 *  At this point we had better have found a good page.
1877 	 */
1878 	KASSERT(m != NULL, ("missing page"));
1879 
1880 found:
1881 	vm_page_dequeue(m);
1882 	vm_page_alloc_check(m);
1883 
1884 	/*
1885 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1886 	 */
1887 	flags = 0;
1888 	if ((req & VM_ALLOC_ZERO) != 0)
1889 		flags = PG_ZERO;
1890 	flags &= m->flags;
1891 	if ((req & VM_ALLOC_NODUMP) != 0)
1892 		flags |= PG_NODUMP;
1893 	m->flags = flags;
1894 	m->aflags = 0;
1895 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1896 	    VPO_UNMANAGED : 0;
1897 	m->busy_lock = VPB_UNBUSIED;
1898 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1899 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1900 	if ((req & VM_ALLOC_SBUSY) != 0)
1901 		m->busy_lock = VPB_SHARERS_WORD(1);
1902 	if (req & VM_ALLOC_WIRED) {
1903 		/*
1904 		 * The page lock is not required for wiring a page until that
1905 		 * page is inserted into the object.
1906 		 */
1907 		vm_wire_add(1);
1908 		m->wire_count = 1;
1909 	}
1910 	m->act_count = 0;
1911 
1912 	if (object != NULL) {
1913 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1914 			if (req & VM_ALLOC_WIRED) {
1915 				vm_wire_sub(1);
1916 				m->wire_count = 0;
1917 			}
1918 			KASSERT(m->object == NULL, ("page %p has object", m));
1919 			m->oflags = VPO_UNMANAGED;
1920 			m->busy_lock = VPB_UNBUSIED;
1921 			/* Don't change PG_ZERO. */
1922 			vm_page_free_toq(m);
1923 			if (req & VM_ALLOC_WAITFAIL) {
1924 				VM_OBJECT_WUNLOCK(object);
1925 				vm_radix_wait();
1926 				VM_OBJECT_WLOCK(object);
1927 			}
1928 			return (NULL);
1929 		}
1930 
1931 		/* Ignore device objects; the pager sets "memattr" for them. */
1932 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1933 		    (object->flags & OBJ_FICTITIOUS) == 0)
1934 			pmap_page_set_memattr(m, object->memattr);
1935 	} else
1936 		m->pindex = pindex;
1937 
1938 	return (m);
1939 }
1940 
1941 /*
1942  *	vm_page_alloc_contig:
1943  *
1944  *	Allocate a contiguous set of physical pages of the given size "npages"
1945  *	from the free lists.  All of the physical pages must be at or above
1946  *	the given physical address "low" and below the given physical address
1947  *	"high".  The given value "alignment" determines the alignment of the
1948  *	first physical page in the set.  If the given value "boundary" is
1949  *	non-zero, then the set of physical pages cannot cross any physical
1950  *	address boundary that is a multiple of that value.  Both "alignment"
1951  *	and "boundary" must be a power of two.
1952  *
1953  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1954  *	then the memory attribute setting for the physical pages is configured
1955  *	to the object's memory attribute setting.  Otherwise, the memory
1956  *	attribute setting for the physical pages is configured to "memattr",
1957  *	overriding the object's memory attribute setting.  However, if the
1958  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1959  *	memory attribute setting for the physical pages cannot be configured
1960  *	to VM_MEMATTR_DEFAULT.
1961  *
1962  *	The specified object may not contain fictitious pages.
1963  *
1964  *	The caller must always specify an allocation class.
1965  *
1966  *	allocation classes:
1967  *	VM_ALLOC_NORMAL		normal process request
1968  *	VM_ALLOC_SYSTEM		system *really* needs a page
1969  *	VM_ALLOC_INTERRUPT	interrupt time request
1970  *
1971  *	optional allocation flags:
1972  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1973  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1974  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1975  *				should not be exclusive busy
1976  *	VM_ALLOC_SBUSY		shared busy the allocated page
1977  *	VM_ALLOC_WIRED		wire the allocated page
1978  *	VM_ALLOC_ZERO		prefer a zeroed page
1979  */
1980 vm_page_t
1981 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1982     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1983     vm_paddr_t boundary, vm_memattr_t memattr)
1984 {
1985 	struct vm_domainset_iter di;
1986 	vm_page_t m;
1987 	int domain;
1988 
1989 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1990 	do {
1991 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1992 		    npages, low, high, alignment, boundary, memattr);
1993 		if (m != NULL)
1994 			break;
1995 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1996 
1997 	return (m);
1998 }
1999 
2000 vm_page_t
2001 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2002     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2003     vm_paddr_t boundary, vm_memattr_t memattr)
2004 {
2005 	struct vm_domain *vmd;
2006 	vm_page_t m, m_ret, mpred;
2007 	u_int busy_lock, flags, oflags;
2008 
2009 	mpred = NULL;	/* XXX: pacify gcc */
2010 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2011 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2012 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2013 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2014 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2015 	    req));
2016 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2017 	    ("Can't sleep and retry object insertion."));
2018 	if (object != NULL) {
2019 		VM_OBJECT_ASSERT_WLOCKED(object);
2020 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2021 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2022 		    object));
2023 	}
2024 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2025 
2026 	if (object != NULL) {
2027 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2028 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2029 		    ("vm_page_alloc_contig: pindex already allocated"));
2030 	}
2031 
2032 	/*
2033 	 * Can we allocate the pages without the number of free pages falling
2034 	 * below the lower bound for the allocation class?
2035 	 */
2036 again:
2037 #if VM_NRESERVLEVEL > 0
2038 	/*
2039 	 * Can we allocate the pages from a reservation?
2040 	 */
2041 	if (vm_object_reserv(object) &&
2042 	    ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
2043 	    npages, low, high, alignment, boundary, mpred)) != NULL ||
2044 	    (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain,
2045 	    npages, low, high, alignment, boundary, mpred)) != NULL)) {
2046 		domain = vm_phys_domain(m_ret);
2047 		vmd = VM_DOMAIN(domain);
2048 		goto found;
2049 	}
2050 #endif
2051 	m_ret = NULL;
2052 	vmd = VM_DOMAIN(domain);
2053 	if (vm_domain_allocate(vmd, req, npages)) {
2054 		/*
2055 		 * allocate them from the free page queues.
2056 		 */
2057 		vm_domain_free_lock(vmd);
2058 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2059 		    alignment, boundary);
2060 		vm_domain_free_unlock(vmd);
2061 		if (m_ret == NULL) {
2062 			vm_domain_freecnt_inc(vmd, npages);
2063 #if VM_NRESERVLEVEL > 0
2064 			if (vm_reserv_reclaim_contig(domain, npages, low,
2065 			    high, alignment, boundary))
2066 				goto again;
2067 #endif
2068 		}
2069 	}
2070 	if (m_ret == NULL) {
2071 		if (vm_domain_alloc_fail(vmd, object, req))
2072 			goto again;
2073 		return (NULL);
2074 	}
2075 #if VM_NRESERVLEVEL > 0
2076 found:
2077 #endif
2078 	for (m = m_ret; m < &m_ret[npages]; m++) {
2079 		vm_page_dequeue(m);
2080 		vm_page_alloc_check(m);
2081 	}
2082 
2083 	/*
2084 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2085 	 */
2086 	flags = 0;
2087 	if ((req & VM_ALLOC_ZERO) != 0)
2088 		flags = PG_ZERO;
2089 	if ((req & VM_ALLOC_NODUMP) != 0)
2090 		flags |= PG_NODUMP;
2091 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2092 	    VPO_UNMANAGED : 0;
2093 	busy_lock = VPB_UNBUSIED;
2094 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2095 		busy_lock = VPB_SINGLE_EXCLUSIVER;
2096 	if ((req & VM_ALLOC_SBUSY) != 0)
2097 		busy_lock = VPB_SHARERS_WORD(1);
2098 	if ((req & VM_ALLOC_WIRED) != 0)
2099 		vm_wire_add(npages);
2100 	if (object != NULL) {
2101 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2102 		    memattr == VM_MEMATTR_DEFAULT)
2103 			memattr = object->memattr;
2104 	}
2105 	for (m = m_ret; m < &m_ret[npages]; m++) {
2106 		m->aflags = 0;
2107 		m->flags = (m->flags | PG_NODUMP) & flags;
2108 		m->busy_lock = busy_lock;
2109 		if ((req & VM_ALLOC_WIRED) != 0)
2110 			m->wire_count = 1;
2111 		m->act_count = 0;
2112 		m->oflags = oflags;
2113 		if (object != NULL) {
2114 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2115 				if ((req & VM_ALLOC_WIRED) != 0)
2116 					vm_wire_sub(npages);
2117 				KASSERT(m->object == NULL,
2118 				    ("page %p has object", m));
2119 				mpred = m;
2120 				for (m = m_ret; m < &m_ret[npages]; m++) {
2121 					if (m <= mpred &&
2122 					    (req & VM_ALLOC_WIRED) != 0)
2123 						m->wire_count = 0;
2124 					m->oflags = VPO_UNMANAGED;
2125 					m->busy_lock = VPB_UNBUSIED;
2126 					/* Don't change PG_ZERO. */
2127 					vm_page_free_toq(m);
2128 				}
2129 				if (req & VM_ALLOC_WAITFAIL) {
2130 					VM_OBJECT_WUNLOCK(object);
2131 					vm_radix_wait();
2132 					VM_OBJECT_WLOCK(object);
2133 				}
2134 				return (NULL);
2135 			}
2136 			mpred = m;
2137 		} else
2138 			m->pindex = pindex;
2139 		if (memattr != VM_MEMATTR_DEFAULT)
2140 			pmap_page_set_memattr(m, memattr);
2141 		pindex++;
2142 	}
2143 	return (m_ret);
2144 }
2145 
2146 /*
2147  * Check a page that has been freshly dequeued from a freelist.
2148  */
2149 static void
2150 vm_page_alloc_check(vm_page_t m)
2151 {
2152 
2153 	KASSERT(m->object == NULL, ("page %p has object", m));
2154 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2155 	    ("page %p has unexpected queue %d, flags %#x",
2156 	    m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2157 	KASSERT(!vm_page_held(m), ("page %p is held", m));
2158 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2159 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2160 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2161 	    ("page %p has unexpected memattr %d",
2162 	    m, pmap_page_get_memattr(m)));
2163 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2164 }
2165 
2166 /*
2167  * 	vm_page_alloc_freelist:
2168  *
2169  *	Allocate a physical page from the specified free page list.
2170  *
2171  *	The caller must always specify an allocation class.
2172  *
2173  *	allocation classes:
2174  *	VM_ALLOC_NORMAL		normal process request
2175  *	VM_ALLOC_SYSTEM		system *really* needs a page
2176  *	VM_ALLOC_INTERRUPT	interrupt time request
2177  *
2178  *	optional allocation flags:
2179  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2180  *				intends to allocate
2181  *	VM_ALLOC_WIRED		wire the allocated page
2182  *	VM_ALLOC_ZERO		prefer a zeroed page
2183  */
2184 vm_page_t
2185 vm_page_alloc_freelist(int freelist, int req)
2186 {
2187 	struct vm_domainset_iter di;
2188 	vm_page_t m;
2189 	int domain;
2190 
2191 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2192 	do {
2193 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2194 		if (m != NULL)
2195 			break;
2196 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2197 
2198 	return (m);
2199 }
2200 
2201 vm_page_t
2202 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2203 {
2204 	struct vm_domain *vmd;
2205 	vm_page_t m;
2206 	u_int flags;
2207 
2208 	m = NULL;
2209 	vmd = VM_DOMAIN(domain);
2210 again:
2211 	if (vm_domain_allocate(vmd, req, 1)) {
2212 		vm_domain_free_lock(vmd);
2213 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2214 		    VM_FREEPOOL_DIRECT, 0);
2215 		vm_domain_free_unlock(vmd);
2216 		if (m == NULL)
2217 			vm_domain_freecnt_inc(vmd, 1);
2218 	}
2219 	if (m == NULL) {
2220 		if (vm_domain_alloc_fail(vmd, NULL, req))
2221 			goto again;
2222 		return (NULL);
2223 	}
2224 	vm_page_dequeue(m);
2225 	vm_page_alloc_check(m);
2226 
2227 	/*
2228 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2229 	 */
2230 	m->aflags = 0;
2231 	flags = 0;
2232 	if ((req & VM_ALLOC_ZERO) != 0)
2233 		flags = PG_ZERO;
2234 	m->flags &= flags;
2235 	if ((req & VM_ALLOC_WIRED) != 0) {
2236 		/*
2237 		 * The page lock is not required for wiring a page that does
2238 		 * not belong to an object.
2239 		 */
2240 		vm_wire_add(1);
2241 		m->wire_count = 1;
2242 	}
2243 	/* Unmanaged pages don't use "act_count". */
2244 	m->oflags = VPO_UNMANAGED;
2245 	return (m);
2246 }
2247 
2248 static int
2249 vm_page_import(void *arg, void **store, int cnt, int domain, int flags)
2250 {
2251 	struct vm_domain *vmd;
2252 	int i;
2253 
2254 	vmd = arg;
2255 	/* Only import if we can bring in a full bucket. */
2256 	if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2257 		return (0);
2258 	domain = vmd->vmd_domain;
2259 	vm_domain_free_lock(vmd);
2260 	i = vm_phys_alloc_npages(domain, VM_FREEPOOL_DEFAULT, cnt,
2261 	    (vm_page_t *)store);
2262 	vm_domain_free_unlock(vmd);
2263 	if (cnt != i)
2264 		vm_domain_freecnt_inc(vmd, cnt - i);
2265 
2266 	return (i);
2267 }
2268 
2269 static void
2270 vm_page_release(void *arg, void **store, int cnt)
2271 {
2272 	struct vm_domain *vmd;
2273 	vm_page_t m;
2274 	int i;
2275 
2276 	vmd = arg;
2277 	vm_domain_free_lock(vmd);
2278 	for (i = 0; i < cnt; i++) {
2279 		m = (vm_page_t)store[i];
2280 		vm_phys_free_pages(m, 0);
2281 	}
2282 	vm_domain_free_unlock(vmd);
2283 	vm_domain_freecnt_inc(vmd, cnt);
2284 }
2285 
2286 #define	VPSC_ANY	0	/* No restrictions. */
2287 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2288 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2289 
2290 /*
2291  *	vm_page_scan_contig:
2292  *
2293  *	Scan vm_page_array[] between the specified entries "m_start" and
2294  *	"m_end" for a run of contiguous physical pages that satisfy the
2295  *	specified conditions, and return the lowest page in the run.  The
2296  *	specified "alignment" determines the alignment of the lowest physical
2297  *	page in the run.  If the specified "boundary" is non-zero, then the
2298  *	run of physical pages cannot span a physical address that is a
2299  *	multiple of "boundary".
2300  *
2301  *	"m_end" is never dereferenced, so it need not point to a vm_page
2302  *	structure within vm_page_array[].
2303  *
2304  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2305  *	span a hole (or discontiguity) in the physical address space.  Both
2306  *	"alignment" and "boundary" must be a power of two.
2307  */
2308 vm_page_t
2309 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2310     u_long alignment, vm_paddr_t boundary, int options)
2311 {
2312 	struct mtx *m_mtx;
2313 	vm_object_t object;
2314 	vm_paddr_t pa;
2315 	vm_page_t m, m_run;
2316 #if VM_NRESERVLEVEL > 0
2317 	int level;
2318 #endif
2319 	int m_inc, order, run_ext, run_len;
2320 
2321 	KASSERT(npages > 0, ("npages is 0"));
2322 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2323 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2324 	m_run = NULL;
2325 	run_len = 0;
2326 	m_mtx = NULL;
2327 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2328 		KASSERT((m->flags & PG_MARKER) == 0,
2329 		    ("page %p is PG_MARKER", m));
2330 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2331 		    ("fictitious page %p has invalid wire count", m));
2332 
2333 		/*
2334 		 * If the current page would be the start of a run, check its
2335 		 * physical address against the end, alignment, and boundary
2336 		 * conditions.  If it doesn't satisfy these conditions, either
2337 		 * terminate the scan or advance to the next page that
2338 		 * satisfies the failed condition.
2339 		 */
2340 		if (run_len == 0) {
2341 			KASSERT(m_run == NULL, ("m_run != NULL"));
2342 			if (m + npages > m_end)
2343 				break;
2344 			pa = VM_PAGE_TO_PHYS(m);
2345 			if ((pa & (alignment - 1)) != 0) {
2346 				m_inc = atop(roundup2(pa, alignment) - pa);
2347 				continue;
2348 			}
2349 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2350 			    boundary) != 0) {
2351 				m_inc = atop(roundup2(pa, boundary) - pa);
2352 				continue;
2353 			}
2354 		} else
2355 			KASSERT(m_run != NULL, ("m_run == NULL"));
2356 
2357 		vm_page_change_lock(m, &m_mtx);
2358 		m_inc = 1;
2359 retry:
2360 		if (vm_page_held(m))
2361 			run_ext = 0;
2362 #if VM_NRESERVLEVEL > 0
2363 		else if ((level = vm_reserv_level(m)) >= 0 &&
2364 		    (options & VPSC_NORESERV) != 0) {
2365 			run_ext = 0;
2366 			/* Advance to the end of the reservation. */
2367 			pa = VM_PAGE_TO_PHYS(m);
2368 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2369 			    pa);
2370 		}
2371 #endif
2372 		else if ((object = m->object) != NULL) {
2373 			/*
2374 			 * The page is considered eligible for relocation if
2375 			 * and only if it could be laundered or reclaimed by
2376 			 * the page daemon.
2377 			 */
2378 			if (!VM_OBJECT_TRYRLOCK(object)) {
2379 				mtx_unlock(m_mtx);
2380 				VM_OBJECT_RLOCK(object);
2381 				mtx_lock(m_mtx);
2382 				if (m->object != object) {
2383 					/*
2384 					 * The page may have been freed.
2385 					 */
2386 					VM_OBJECT_RUNLOCK(object);
2387 					goto retry;
2388 				} else if (vm_page_held(m)) {
2389 					run_ext = 0;
2390 					goto unlock;
2391 				}
2392 			}
2393 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2394 			    ("page %p is PG_UNHOLDFREE", m));
2395 			/* Don't care: PG_NODUMP, PG_ZERO. */
2396 			if (object->type != OBJT_DEFAULT &&
2397 			    object->type != OBJT_SWAP &&
2398 			    object->type != OBJT_VNODE) {
2399 				run_ext = 0;
2400 #if VM_NRESERVLEVEL > 0
2401 			} else if ((options & VPSC_NOSUPER) != 0 &&
2402 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2403 				run_ext = 0;
2404 				/* Advance to the end of the superpage. */
2405 				pa = VM_PAGE_TO_PHYS(m);
2406 				m_inc = atop(roundup2(pa + 1,
2407 				    vm_reserv_size(level)) - pa);
2408 #endif
2409 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2410 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2411 				/*
2412 				 * The page is allocated but eligible for
2413 				 * relocation.  Extend the current run by one
2414 				 * page.
2415 				 */
2416 				KASSERT(pmap_page_get_memattr(m) ==
2417 				    VM_MEMATTR_DEFAULT,
2418 				    ("page %p has an unexpected memattr", m));
2419 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2420 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2421 				    ("page %p has unexpected oflags", m));
2422 				/* Don't care: VPO_NOSYNC. */
2423 				run_ext = 1;
2424 			} else
2425 				run_ext = 0;
2426 unlock:
2427 			VM_OBJECT_RUNLOCK(object);
2428 #if VM_NRESERVLEVEL > 0
2429 		} else if (level >= 0) {
2430 			/*
2431 			 * The page is reserved but not yet allocated.  In
2432 			 * other words, it is still free.  Extend the current
2433 			 * run by one page.
2434 			 */
2435 			run_ext = 1;
2436 #endif
2437 		} else if ((order = m->order) < VM_NFREEORDER) {
2438 			/*
2439 			 * The page is enqueued in the physical memory
2440 			 * allocator's free page queues.  Moreover, it is the
2441 			 * first page in a power-of-two-sized run of
2442 			 * contiguous free pages.  Add these pages to the end
2443 			 * of the current run, and jump ahead.
2444 			 */
2445 			run_ext = 1 << order;
2446 			m_inc = 1 << order;
2447 		} else {
2448 			/*
2449 			 * Skip the page for one of the following reasons: (1)
2450 			 * It is enqueued in the physical memory allocator's
2451 			 * free page queues.  However, it is not the first
2452 			 * page in a run of contiguous free pages.  (This case
2453 			 * rarely occurs because the scan is performed in
2454 			 * ascending order.) (2) It is not reserved, and it is
2455 			 * transitioning from free to allocated.  (Conversely,
2456 			 * the transition from allocated to free for managed
2457 			 * pages is blocked by the page lock.) (3) It is
2458 			 * allocated but not contained by an object and not
2459 			 * wired, e.g., allocated by Xen's balloon driver.
2460 			 */
2461 			run_ext = 0;
2462 		}
2463 
2464 		/*
2465 		 * Extend or reset the current run of pages.
2466 		 */
2467 		if (run_ext > 0) {
2468 			if (run_len == 0)
2469 				m_run = m;
2470 			run_len += run_ext;
2471 		} else {
2472 			if (run_len > 0) {
2473 				m_run = NULL;
2474 				run_len = 0;
2475 			}
2476 		}
2477 	}
2478 	if (m_mtx != NULL)
2479 		mtx_unlock(m_mtx);
2480 	if (run_len >= npages)
2481 		return (m_run);
2482 	return (NULL);
2483 }
2484 
2485 /*
2486  *	vm_page_reclaim_run:
2487  *
2488  *	Try to relocate each of the allocated virtual pages within the
2489  *	specified run of physical pages to a new physical address.  Free the
2490  *	physical pages underlying the relocated virtual pages.  A virtual page
2491  *	is relocatable if and only if it could be laundered or reclaimed by
2492  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2493  *	physical address above "high".
2494  *
2495  *	Returns 0 if every physical page within the run was already free or
2496  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2497  *	value indicating why the last attempt to relocate a virtual page was
2498  *	unsuccessful.
2499  *
2500  *	"req_class" must be an allocation class.
2501  */
2502 static int
2503 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2504     vm_paddr_t high)
2505 {
2506 	struct vm_domain *vmd;
2507 	struct mtx *m_mtx;
2508 	struct spglist free;
2509 	vm_object_t object;
2510 	vm_paddr_t pa;
2511 	vm_page_t m, m_end, m_new;
2512 	int error, order, req;
2513 
2514 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2515 	    ("req_class is not an allocation class"));
2516 	SLIST_INIT(&free);
2517 	error = 0;
2518 	m = m_run;
2519 	m_end = m_run + npages;
2520 	m_mtx = NULL;
2521 	for (; error == 0 && m < m_end; m++) {
2522 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2523 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2524 
2525 		/*
2526 		 * Avoid releasing and reacquiring the same page lock.
2527 		 */
2528 		vm_page_change_lock(m, &m_mtx);
2529 retry:
2530 		if (vm_page_held(m))
2531 			error = EBUSY;
2532 		else if ((object = m->object) != NULL) {
2533 			/*
2534 			 * The page is relocated if and only if it could be
2535 			 * laundered or reclaimed by the page daemon.
2536 			 */
2537 			if (!VM_OBJECT_TRYWLOCK(object)) {
2538 				mtx_unlock(m_mtx);
2539 				VM_OBJECT_WLOCK(object);
2540 				mtx_lock(m_mtx);
2541 				if (m->object != object) {
2542 					/*
2543 					 * The page may have been freed.
2544 					 */
2545 					VM_OBJECT_WUNLOCK(object);
2546 					goto retry;
2547 				} else if (vm_page_held(m)) {
2548 					error = EBUSY;
2549 					goto unlock;
2550 				}
2551 			}
2552 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2553 			    ("page %p is PG_UNHOLDFREE", m));
2554 			/* Don't care: PG_NODUMP, PG_ZERO. */
2555 			if (object->type != OBJT_DEFAULT &&
2556 			    object->type != OBJT_SWAP &&
2557 			    object->type != OBJT_VNODE)
2558 				error = EINVAL;
2559 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2560 				error = EINVAL;
2561 			else if (vm_page_queue(m) != PQ_NONE &&
2562 			    !vm_page_busied(m)) {
2563 				KASSERT(pmap_page_get_memattr(m) ==
2564 				    VM_MEMATTR_DEFAULT,
2565 				    ("page %p has an unexpected memattr", m));
2566 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2567 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2568 				    ("page %p has unexpected oflags", m));
2569 				/* Don't care: VPO_NOSYNC. */
2570 				if (m->valid != 0) {
2571 					/*
2572 					 * First, try to allocate a new page
2573 					 * that is above "high".  Failing
2574 					 * that, try to allocate a new page
2575 					 * that is below "m_run".  Allocate
2576 					 * the new page between the end of
2577 					 * "m_run" and "high" only as a last
2578 					 * resort.
2579 					 */
2580 					req = req_class | VM_ALLOC_NOOBJ;
2581 					if ((m->flags & PG_NODUMP) != 0)
2582 						req |= VM_ALLOC_NODUMP;
2583 					if (trunc_page(high) !=
2584 					    ~(vm_paddr_t)PAGE_MASK) {
2585 						m_new = vm_page_alloc_contig(
2586 						    NULL, 0, req, 1,
2587 						    round_page(high),
2588 						    ~(vm_paddr_t)0,
2589 						    PAGE_SIZE, 0,
2590 						    VM_MEMATTR_DEFAULT);
2591 					} else
2592 						m_new = NULL;
2593 					if (m_new == NULL) {
2594 						pa = VM_PAGE_TO_PHYS(m_run);
2595 						m_new = vm_page_alloc_contig(
2596 						    NULL, 0, req, 1,
2597 						    0, pa - 1, PAGE_SIZE, 0,
2598 						    VM_MEMATTR_DEFAULT);
2599 					}
2600 					if (m_new == NULL) {
2601 						pa += ptoa(npages);
2602 						m_new = vm_page_alloc_contig(
2603 						    NULL, 0, req, 1,
2604 						    pa, high, PAGE_SIZE, 0,
2605 						    VM_MEMATTR_DEFAULT);
2606 					}
2607 					if (m_new == NULL) {
2608 						error = ENOMEM;
2609 						goto unlock;
2610 					}
2611 					KASSERT(m_new->wire_count == 0,
2612 					    ("page %p is wired", m_new));
2613 
2614 					/*
2615 					 * Replace "m" with the new page.  For
2616 					 * vm_page_replace(), "m" must be busy
2617 					 * and dequeued.  Finally, change "m"
2618 					 * as if vm_page_free() was called.
2619 					 */
2620 					if (object->ref_count != 0)
2621 						pmap_remove_all(m);
2622 					m_new->aflags = m->aflags &
2623 					    ~PGA_QUEUE_STATE_MASK;
2624 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2625 					    ("page %p is managed", m_new));
2626 					m_new->oflags = m->oflags & VPO_NOSYNC;
2627 					pmap_copy_page(m, m_new);
2628 					m_new->valid = m->valid;
2629 					m_new->dirty = m->dirty;
2630 					m->flags &= ~PG_ZERO;
2631 					vm_page_xbusy(m);
2632 					vm_page_dequeue(m);
2633 					vm_page_replace_checked(m_new, object,
2634 					    m->pindex, m);
2635 					if (vm_page_free_prep(m))
2636 						SLIST_INSERT_HEAD(&free, m,
2637 						    plinks.s.ss);
2638 
2639 					/*
2640 					 * The new page must be deactivated
2641 					 * before the object is unlocked.
2642 					 */
2643 					vm_page_change_lock(m_new, &m_mtx);
2644 					vm_page_deactivate(m_new);
2645 				} else {
2646 					m->flags &= ~PG_ZERO;
2647 					vm_page_dequeue(m);
2648 					vm_page_remove(m);
2649 					if (vm_page_free_prep(m))
2650 						SLIST_INSERT_HEAD(&free, m,
2651 						    plinks.s.ss);
2652 					KASSERT(m->dirty == 0,
2653 					    ("page %p is dirty", m));
2654 				}
2655 			} else
2656 				error = EBUSY;
2657 unlock:
2658 			VM_OBJECT_WUNLOCK(object);
2659 		} else {
2660 			MPASS(vm_phys_domain(m) == domain);
2661 			vmd = VM_DOMAIN(domain);
2662 			vm_domain_free_lock(vmd);
2663 			order = m->order;
2664 			if (order < VM_NFREEORDER) {
2665 				/*
2666 				 * The page is enqueued in the physical memory
2667 				 * allocator's free page queues.  Moreover, it
2668 				 * is the first page in a power-of-two-sized
2669 				 * run of contiguous free pages.  Jump ahead
2670 				 * to the last page within that run, and
2671 				 * continue from there.
2672 				 */
2673 				m += (1 << order) - 1;
2674 			}
2675 #if VM_NRESERVLEVEL > 0
2676 			else if (vm_reserv_is_page_free(m))
2677 				order = 0;
2678 #endif
2679 			vm_domain_free_unlock(vmd);
2680 			if (order == VM_NFREEORDER)
2681 				error = EINVAL;
2682 		}
2683 	}
2684 	if (m_mtx != NULL)
2685 		mtx_unlock(m_mtx);
2686 	if ((m = SLIST_FIRST(&free)) != NULL) {
2687 		int cnt;
2688 
2689 		vmd = VM_DOMAIN(domain);
2690 		cnt = 0;
2691 		vm_domain_free_lock(vmd);
2692 		do {
2693 			MPASS(vm_phys_domain(m) == domain);
2694 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2695 			vm_phys_free_pages(m, 0);
2696 			cnt++;
2697 		} while ((m = SLIST_FIRST(&free)) != NULL);
2698 		vm_domain_free_unlock(vmd);
2699 		vm_domain_freecnt_inc(vmd, cnt);
2700 	}
2701 	return (error);
2702 }
2703 
2704 #define	NRUNS	16
2705 
2706 CTASSERT(powerof2(NRUNS));
2707 
2708 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2709 
2710 #define	MIN_RECLAIM	8
2711 
2712 /*
2713  *	vm_page_reclaim_contig:
2714  *
2715  *	Reclaim allocated, contiguous physical memory satisfying the specified
2716  *	conditions by relocating the virtual pages using that physical memory.
2717  *	Returns true if reclamation is successful and false otherwise.  Since
2718  *	relocation requires the allocation of physical pages, reclamation may
2719  *	fail due to a shortage of free pages.  When reclamation fails, callers
2720  *	are expected to perform vm_wait() before retrying a failed allocation
2721  *	operation, e.g., vm_page_alloc_contig().
2722  *
2723  *	The caller must always specify an allocation class through "req".
2724  *
2725  *	allocation classes:
2726  *	VM_ALLOC_NORMAL		normal process request
2727  *	VM_ALLOC_SYSTEM		system *really* needs a page
2728  *	VM_ALLOC_INTERRUPT	interrupt time request
2729  *
2730  *	The optional allocation flags are ignored.
2731  *
2732  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2733  *	must be a power of two.
2734  */
2735 bool
2736 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2737     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2738 {
2739 	struct vm_domain *vmd;
2740 	vm_paddr_t curr_low;
2741 	vm_page_t m_run, m_runs[NRUNS];
2742 	u_long count, reclaimed;
2743 	int error, i, options, req_class;
2744 
2745 	KASSERT(npages > 0, ("npages is 0"));
2746 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2747 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2748 	req_class = req & VM_ALLOC_CLASS_MASK;
2749 
2750 	/*
2751 	 * The page daemon is allowed to dig deeper into the free page list.
2752 	 */
2753 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2754 		req_class = VM_ALLOC_SYSTEM;
2755 
2756 	/*
2757 	 * Return if the number of free pages cannot satisfy the requested
2758 	 * allocation.
2759 	 */
2760 	vmd = VM_DOMAIN(domain);
2761 	count = vmd->vmd_free_count;
2762 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2763 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2764 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2765 		return (false);
2766 
2767 	/*
2768 	 * Scan up to three times, relaxing the restrictions ("options") on
2769 	 * the reclamation of reservations and superpages each time.
2770 	 */
2771 	for (options = VPSC_NORESERV;;) {
2772 		/*
2773 		 * Find the highest runs that satisfy the given constraints
2774 		 * and restrictions, and record them in "m_runs".
2775 		 */
2776 		curr_low = low;
2777 		count = 0;
2778 		for (;;) {
2779 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2780 			    high, alignment, boundary, options);
2781 			if (m_run == NULL)
2782 				break;
2783 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2784 			m_runs[RUN_INDEX(count)] = m_run;
2785 			count++;
2786 		}
2787 
2788 		/*
2789 		 * Reclaim the highest runs in LIFO (descending) order until
2790 		 * the number of reclaimed pages, "reclaimed", is at least
2791 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2792 		 * reclamation is idempotent, and runs will (likely) recur
2793 		 * from one scan to the next as restrictions are relaxed.
2794 		 */
2795 		reclaimed = 0;
2796 		for (i = 0; count > 0 && i < NRUNS; i++) {
2797 			count--;
2798 			m_run = m_runs[RUN_INDEX(count)];
2799 			error = vm_page_reclaim_run(req_class, domain, npages,
2800 			    m_run, high);
2801 			if (error == 0) {
2802 				reclaimed += npages;
2803 				if (reclaimed >= MIN_RECLAIM)
2804 					return (true);
2805 			}
2806 		}
2807 
2808 		/*
2809 		 * Either relax the restrictions on the next scan or return if
2810 		 * the last scan had no restrictions.
2811 		 */
2812 		if (options == VPSC_NORESERV)
2813 			options = VPSC_NOSUPER;
2814 		else if (options == VPSC_NOSUPER)
2815 			options = VPSC_ANY;
2816 		else if (options == VPSC_ANY)
2817 			return (reclaimed != 0);
2818 	}
2819 }
2820 
2821 bool
2822 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2823     u_long alignment, vm_paddr_t boundary)
2824 {
2825 	struct vm_domainset_iter di;
2826 	int domain;
2827 	bool ret;
2828 
2829 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2830 	do {
2831 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2832 		    high, alignment, boundary);
2833 		if (ret)
2834 			break;
2835 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2836 
2837 	return (ret);
2838 }
2839 
2840 /*
2841  * Set the domain in the appropriate page level domainset.
2842  */
2843 void
2844 vm_domain_set(struct vm_domain *vmd)
2845 {
2846 
2847 	mtx_lock(&vm_domainset_lock);
2848 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2849 		vmd->vmd_minset = 1;
2850 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2851 	}
2852 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2853 		vmd->vmd_severeset = 1;
2854 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2855 	}
2856 	mtx_unlock(&vm_domainset_lock);
2857 }
2858 
2859 /*
2860  * Clear the domain from the appropriate page level domainset.
2861  */
2862 void
2863 vm_domain_clear(struct vm_domain *vmd)
2864 {
2865 
2866 	mtx_lock(&vm_domainset_lock);
2867 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2868 		vmd->vmd_minset = 0;
2869 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2870 		if (vm_min_waiters != 0) {
2871 			vm_min_waiters = 0;
2872 			wakeup(&vm_min_domains);
2873 		}
2874 	}
2875 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2876 		vmd->vmd_severeset = 0;
2877 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2878 		if (vm_severe_waiters != 0) {
2879 			vm_severe_waiters = 0;
2880 			wakeup(&vm_severe_domains);
2881 		}
2882 	}
2883 
2884 	/*
2885 	 * If pageout daemon needs pages, then tell it that there are
2886 	 * some free.
2887 	 */
2888 	if (vmd->vmd_pageout_pages_needed &&
2889 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2890 		wakeup(&vmd->vmd_pageout_pages_needed);
2891 		vmd->vmd_pageout_pages_needed = 0;
2892 	}
2893 
2894 	/* See comments in vm_wait_doms(). */
2895 	if (vm_pageproc_waiters) {
2896 		vm_pageproc_waiters = 0;
2897 		wakeup(&vm_pageproc_waiters);
2898 	}
2899 	mtx_unlock(&vm_domainset_lock);
2900 }
2901 
2902 /*
2903  * Wait for free pages to exceed the min threshold globally.
2904  */
2905 void
2906 vm_wait_min(void)
2907 {
2908 
2909 	mtx_lock(&vm_domainset_lock);
2910 	while (vm_page_count_min()) {
2911 		vm_min_waiters++;
2912 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2913 	}
2914 	mtx_unlock(&vm_domainset_lock);
2915 }
2916 
2917 /*
2918  * Wait for free pages to exceed the severe threshold globally.
2919  */
2920 void
2921 vm_wait_severe(void)
2922 {
2923 
2924 	mtx_lock(&vm_domainset_lock);
2925 	while (vm_page_count_severe()) {
2926 		vm_severe_waiters++;
2927 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2928 		    "vmwait", 0);
2929 	}
2930 	mtx_unlock(&vm_domainset_lock);
2931 }
2932 
2933 u_int
2934 vm_wait_count(void)
2935 {
2936 
2937 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2938 }
2939 
2940 void
2941 vm_wait_doms(const domainset_t *wdoms)
2942 {
2943 
2944 	/*
2945 	 * We use racey wakeup synchronization to avoid expensive global
2946 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
2947 	 * To handle this, we only sleep for one tick in this instance.  It
2948 	 * is expected that most allocations for the pageproc will come from
2949 	 * kmem or vm_page_grab* which will use the more specific and
2950 	 * race-free vm_wait_domain().
2951 	 */
2952 	if (curproc == pageproc) {
2953 		mtx_lock(&vm_domainset_lock);
2954 		vm_pageproc_waiters++;
2955 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2956 		    "pageprocwait", 1);
2957 	} else {
2958 		/*
2959 		 * XXX Ideally we would wait only until the allocation could
2960 		 * be satisfied.  This condition can cause new allocators to
2961 		 * consume all freed pages while old allocators wait.
2962 		 */
2963 		mtx_lock(&vm_domainset_lock);
2964 		if (vm_page_count_min_set(wdoms)) {
2965 			vm_min_waiters++;
2966 			msleep(&vm_min_domains, &vm_domainset_lock,
2967 			    PVM | PDROP, "vmwait", 0);
2968 		} else
2969 			mtx_unlock(&vm_domainset_lock);
2970 	}
2971 }
2972 
2973 /*
2974  *	vm_wait_domain:
2975  *
2976  *	Sleep until free pages are available for allocation.
2977  *	- Called in various places after failed memory allocations.
2978  */
2979 void
2980 vm_wait_domain(int domain)
2981 {
2982 	struct vm_domain *vmd;
2983 	domainset_t wdom;
2984 
2985 	vmd = VM_DOMAIN(domain);
2986 	vm_domain_free_assert_unlocked(vmd);
2987 
2988 	if (curproc == pageproc) {
2989 		mtx_lock(&vm_domainset_lock);
2990 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2991 			vmd->vmd_pageout_pages_needed = 1;
2992 			msleep(&vmd->vmd_pageout_pages_needed,
2993 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
2994 		} else
2995 			mtx_unlock(&vm_domainset_lock);
2996 	} else {
2997 		if (pageproc == NULL)
2998 			panic("vm_wait in early boot");
2999 		DOMAINSET_ZERO(&wdom);
3000 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3001 		vm_wait_doms(&wdom);
3002 	}
3003 }
3004 
3005 /*
3006  *	vm_wait:
3007  *
3008  *	Sleep until free pages are available for allocation in the
3009  *	affinity domains of the obj.  If obj is NULL, the domain set
3010  *	for the calling thread is used.
3011  *	Called in various places after failed memory allocations.
3012  */
3013 void
3014 vm_wait(vm_object_t obj)
3015 {
3016 	struct domainset *d;
3017 
3018 	d = NULL;
3019 
3020 	/*
3021 	 * Carefully fetch pointers only once: the struct domainset
3022 	 * itself is ummutable but the pointer might change.
3023 	 */
3024 	if (obj != NULL)
3025 		d = obj->domain.dr_policy;
3026 	if (d == NULL)
3027 		d = curthread->td_domain.dr_policy;
3028 
3029 	vm_wait_doms(&d->ds_mask);
3030 }
3031 
3032 /*
3033  *	vm_domain_alloc_fail:
3034  *
3035  *	Called when a page allocation function fails.  Informs the
3036  *	pagedaemon and performs the requested wait.  Requires the
3037  *	domain_free and object lock on entry.  Returns with the
3038  *	object lock held and free lock released.  Returns an error when
3039  *	retry is necessary.
3040  *
3041  */
3042 static int
3043 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3044 {
3045 
3046 	vm_domain_free_assert_unlocked(vmd);
3047 
3048 	atomic_add_int(&vmd->vmd_pageout_deficit,
3049 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3050 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3051 		if (object != NULL)
3052 			VM_OBJECT_WUNLOCK(object);
3053 		vm_wait_domain(vmd->vmd_domain);
3054 		if (object != NULL)
3055 			VM_OBJECT_WLOCK(object);
3056 		if (req & VM_ALLOC_WAITOK)
3057 			return (EAGAIN);
3058 	}
3059 
3060 	return (0);
3061 }
3062 
3063 /*
3064  *	vm_waitpfault:
3065  *
3066  *	Sleep until free pages are available for allocation.
3067  *	- Called only in vm_fault so that processes page faulting
3068  *	  can be easily tracked.
3069  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3070  *	  processes will be able to grab memory first.  Do not change
3071  *	  this balance without careful testing first.
3072  */
3073 void
3074 vm_waitpfault(struct domainset *dset)
3075 {
3076 
3077 	/*
3078 	 * XXX Ideally we would wait only until the allocation could
3079 	 * be satisfied.  This condition can cause new allocators to
3080 	 * consume all freed pages while old allocators wait.
3081 	 */
3082 	mtx_lock(&vm_domainset_lock);
3083 	if (vm_page_count_min_set(&dset->ds_mask)) {
3084 		vm_min_waiters++;
3085 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3086 		    "pfault", 0);
3087 	} else
3088 		mtx_unlock(&vm_domainset_lock);
3089 }
3090 
3091 struct vm_pagequeue *
3092 vm_page_pagequeue(vm_page_t m)
3093 {
3094 
3095 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
3096 }
3097 
3098 static struct mtx *
3099 vm_page_pagequeue_lockptr(vm_page_t m)
3100 {
3101 	uint8_t queue;
3102 
3103 	if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3104 		return (NULL);
3105 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex);
3106 }
3107 
3108 static inline void
3109 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3110 {
3111 	struct vm_domain *vmd;
3112 	uint8_t qflags;
3113 
3114 	CRITICAL_ASSERT(curthread);
3115 	vm_pagequeue_assert_locked(pq);
3116 
3117 	/*
3118 	 * The page daemon is allowed to set m->queue = PQ_NONE without
3119 	 * the page queue lock held.  In this case it is about to free the page,
3120 	 * which must not have any queue state.
3121 	 */
3122 	qflags = atomic_load_8(&m->aflags) & PGA_QUEUE_STATE_MASK;
3123 	KASSERT(pq == vm_page_pagequeue(m) || qflags == 0,
3124 	    ("page %p doesn't belong to queue %p but has queue state %#x",
3125 	    m, pq, qflags));
3126 
3127 	if ((qflags & PGA_DEQUEUE) != 0) {
3128 		if (__predict_true((qflags & PGA_ENQUEUED) != 0)) {
3129 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3130 			vm_pagequeue_cnt_dec(pq);
3131 		}
3132 		vm_page_dequeue_complete(m);
3133 	} else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3134 		if ((qflags & PGA_ENQUEUED) != 0)
3135 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3136 		else {
3137 			vm_pagequeue_cnt_inc(pq);
3138 			vm_page_aflag_set(m, PGA_ENQUEUED);
3139 		}
3140 		if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3141 			KASSERT(m->queue == PQ_INACTIVE,
3142 			    ("head enqueue not supported for page %p", m));
3143 			vmd = vm_pagequeue_domain(m);
3144 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3145 		} else
3146 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3147 
3148 		/*
3149 		 * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after
3150 		 * setting PGA_ENQUEUED in order to synchronize with the
3151 		 * page daemon.
3152 		 */
3153 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
3154 	}
3155 }
3156 
3157 static void
3158 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3159     uint8_t queue)
3160 {
3161 	vm_page_t m;
3162 	int i;
3163 
3164 	for (i = 0; i < bq->bq_cnt; i++) {
3165 		m = bq->bq_pa[i];
3166 		if (__predict_false(m->queue != queue))
3167 			continue;
3168 		vm_pqbatch_process_page(pq, m);
3169 	}
3170 	vm_batchqueue_init(bq);
3171 }
3172 
3173 static void
3174 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3175 {
3176 	struct vm_batchqueue *bq;
3177 	struct vm_pagequeue *pq;
3178 	int domain;
3179 
3180 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3181 	    ("page %p is unmanaged", m));
3182 	KASSERT(mtx_owned(vm_page_lockptr(m)) ||
3183 	    (m->object == NULL && (m->aflags & PGA_DEQUEUE) != 0),
3184 	    ("missing synchronization for page %p", m));
3185 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3186 
3187 	domain = vm_phys_domain(m);
3188 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3189 
3190 	critical_enter();
3191 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3192 	if (vm_batchqueue_insert(bq, m)) {
3193 		critical_exit();
3194 		return;
3195 	}
3196 	if (!vm_pagequeue_trylock(pq)) {
3197 		critical_exit();
3198 		vm_pagequeue_lock(pq);
3199 		critical_enter();
3200 		bq = DPCPU_PTR(pqbatch[domain][queue]);
3201 	}
3202 	vm_pqbatch_process(pq, bq, queue);
3203 
3204 	/*
3205 	 * The page may have been logically dequeued before we acquired the
3206 	 * page queue lock.  In this case, since we either hold the page lock
3207 	 * or the page is being freed, a different thread cannot be concurrently
3208 	 * enqueuing the page.
3209 	 */
3210 	if (__predict_true(m->queue == queue))
3211 		vm_pqbatch_process_page(pq, m);
3212 	else {
3213 		KASSERT(m->queue == PQ_NONE,
3214 		    ("invalid queue transition for page %p", m));
3215 		KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3216 		    ("page %p is enqueued with invalid queue index", m));
3217 		vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3218 	}
3219 	vm_pagequeue_unlock(pq);
3220 	critical_exit();
3221 }
3222 
3223 /*
3224  *	vm_page_drain_pqbatch:		[ internal use only ]
3225  *
3226  *	Force all per-CPU page queue batch queues to be drained.  This is
3227  *	intended for use in severe memory shortages, to ensure that pages
3228  *	do not remain stuck in the batch queues.
3229  */
3230 void
3231 vm_page_drain_pqbatch(void)
3232 {
3233 	struct thread *td;
3234 	struct vm_domain *vmd;
3235 	struct vm_pagequeue *pq;
3236 	int cpu, domain, queue;
3237 
3238 	td = curthread;
3239 	CPU_FOREACH(cpu) {
3240 		thread_lock(td);
3241 		sched_bind(td, cpu);
3242 		thread_unlock(td);
3243 
3244 		for (domain = 0; domain < vm_ndomains; domain++) {
3245 			vmd = VM_DOMAIN(domain);
3246 			for (queue = 0; queue < PQ_COUNT; queue++) {
3247 				pq = &vmd->vmd_pagequeues[queue];
3248 				vm_pagequeue_lock(pq);
3249 				critical_enter();
3250 				vm_pqbatch_process(pq,
3251 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3252 				critical_exit();
3253 				vm_pagequeue_unlock(pq);
3254 			}
3255 		}
3256 	}
3257 	thread_lock(td);
3258 	sched_unbind(td);
3259 	thread_unlock(td);
3260 }
3261 
3262 /*
3263  * Complete the logical removal of a page from a page queue.  We must be
3264  * careful to synchronize with the page daemon, which may be concurrently
3265  * examining the page with only the page lock held.  The page must not be
3266  * in a state where it appears to be logically enqueued.
3267  */
3268 static void
3269 vm_page_dequeue_complete(vm_page_t m)
3270 {
3271 
3272 	m->queue = PQ_NONE;
3273 	atomic_thread_fence_rel();
3274 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3275 }
3276 
3277 /*
3278  *	vm_page_dequeue_deferred:	[ internal use only ]
3279  *
3280  *	Request removal of the given page from its current page
3281  *	queue.  Physical removal from the queue may be deferred
3282  *	indefinitely.
3283  *
3284  *	The page must be locked.
3285  */
3286 void
3287 vm_page_dequeue_deferred(vm_page_t m)
3288 {
3289 	uint8_t queue;
3290 
3291 	vm_page_assert_locked(m);
3292 
3293 	if ((queue = vm_page_queue(m)) == PQ_NONE)
3294 		return;
3295 	vm_page_aflag_set(m, PGA_DEQUEUE);
3296 	vm_pqbatch_submit_page(m, queue);
3297 }
3298 
3299 /*
3300  * A variant of vm_page_dequeue_deferred() that does not assert the page
3301  * lock and is only to be called from vm_page_free_prep().  It is just an
3302  * open-coded implementation of vm_page_dequeue_deferred().  Because the
3303  * page is being freed, we can assume that nothing else is scheduling queue
3304  * operations on this page, so we get for free the mutual exclusion that
3305  * is otherwise provided by the page lock.
3306  */
3307 static void
3308 vm_page_dequeue_deferred_free(vm_page_t m)
3309 {
3310 	uint8_t queue;
3311 
3312 	KASSERT(m->object == NULL, ("page %p has an object reference", m));
3313 
3314 	if ((m->aflags & PGA_DEQUEUE) != 0)
3315 		return;
3316 	atomic_thread_fence_acq();
3317 	if ((queue = m->queue) == PQ_NONE)
3318 		return;
3319 	vm_page_aflag_set(m, PGA_DEQUEUE);
3320 	vm_pqbatch_submit_page(m, queue);
3321 }
3322 
3323 /*
3324  *	vm_page_dequeue:
3325  *
3326  *	Remove the page from whichever page queue it's in, if any.
3327  *	The page must either be locked or unallocated.  This constraint
3328  *	ensures that the queue state of the page will remain consistent
3329  *	after this function returns.
3330  */
3331 void
3332 vm_page_dequeue(vm_page_t m)
3333 {
3334 	struct mtx *lock, *lock1;
3335 	struct vm_pagequeue *pq;
3336 	uint8_t aflags;
3337 
3338 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->order == VM_NFREEORDER,
3339 	    ("page %p is allocated and unlocked", m));
3340 
3341 	for (;;) {
3342 		lock = vm_page_pagequeue_lockptr(m);
3343 		if (lock == NULL) {
3344 			/*
3345 			 * A thread may be concurrently executing
3346 			 * vm_page_dequeue_complete().  Ensure that all queue
3347 			 * state is cleared before we return.
3348 			 */
3349 			aflags = atomic_load_8(&m->aflags);
3350 			if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3351 				return;
3352 			KASSERT((aflags & PGA_DEQUEUE) != 0,
3353 			    ("page %p has unexpected queue state flags %#x",
3354 			    m, aflags));
3355 
3356 			/*
3357 			 * Busy wait until the thread updating queue state is
3358 			 * finished.  Such a thread must be executing in a
3359 			 * critical section.
3360 			 */
3361 			cpu_spinwait();
3362 			continue;
3363 		}
3364 		mtx_lock(lock);
3365 		if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock)
3366 			break;
3367 		mtx_unlock(lock);
3368 		lock = lock1;
3369 	}
3370 	KASSERT(lock == vm_page_pagequeue_lockptr(m),
3371 	    ("%s: page %p migrated directly between queues", __func__, m));
3372 	KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3373 	    mtx_owned(vm_page_lockptr(m)),
3374 	    ("%s: queued unlocked page %p", __func__, m));
3375 
3376 	if ((m->aflags & PGA_ENQUEUED) != 0) {
3377 		pq = vm_page_pagequeue(m);
3378 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3379 		vm_pagequeue_cnt_dec(pq);
3380 	}
3381 	vm_page_dequeue_complete(m);
3382 	mtx_unlock(lock);
3383 }
3384 
3385 /*
3386  * Schedule the given page for insertion into the specified page queue.
3387  * Physical insertion of the page may be deferred indefinitely.
3388  */
3389 static void
3390 vm_page_enqueue(vm_page_t m, uint8_t queue)
3391 {
3392 
3393 	vm_page_assert_locked(m);
3394 	KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3395 	    ("%s: page %p is already enqueued", __func__, m));
3396 
3397 	m->queue = queue;
3398 	if ((m->aflags & PGA_REQUEUE) == 0)
3399 		vm_page_aflag_set(m, PGA_REQUEUE);
3400 	vm_pqbatch_submit_page(m, queue);
3401 }
3402 
3403 /*
3404  *	vm_page_requeue:		[ internal use only ]
3405  *
3406  *	Schedule a requeue of the given page.
3407  *
3408  *	The page must be locked.
3409  */
3410 void
3411 vm_page_requeue(vm_page_t m)
3412 {
3413 
3414 	vm_page_assert_locked(m);
3415 	KASSERT(vm_page_queue(m) != PQ_NONE,
3416 	    ("%s: page %p is not logically enqueued", __func__, m));
3417 
3418 	if ((m->aflags & PGA_REQUEUE) == 0)
3419 		vm_page_aflag_set(m, PGA_REQUEUE);
3420 	vm_pqbatch_submit_page(m, atomic_load_8(&m->queue));
3421 }
3422 
3423 /*
3424  *	vm_page_activate:
3425  *
3426  *	Put the specified page on the active list (if appropriate).
3427  *	Ensure that act_count is at least ACT_INIT but do not otherwise
3428  *	mess with it.
3429  *
3430  *	The page must be locked.
3431  */
3432 void
3433 vm_page_activate(vm_page_t m)
3434 {
3435 
3436 	vm_page_assert_locked(m);
3437 
3438 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3439 		return;
3440 	if (vm_page_queue(m) == PQ_ACTIVE) {
3441 		if (m->act_count < ACT_INIT)
3442 			m->act_count = ACT_INIT;
3443 		return;
3444 	}
3445 
3446 	vm_page_dequeue(m);
3447 	if (m->act_count < ACT_INIT)
3448 		m->act_count = ACT_INIT;
3449 	vm_page_enqueue(m, PQ_ACTIVE);
3450 }
3451 
3452 /*
3453  *	vm_page_free_prep:
3454  *
3455  *	Prepares the given page to be put on the free list,
3456  *	disassociating it from any VM object. The caller may return
3457  *	the page to the free list only if this function returns true.
3458  *
3459  *	The object must be locked.  The page must be locked if it is
3460  *	managed.
3461  */
3462 bool
3463 vm_page_free_prep(vm_page_t m)
3464 {
3465 
3466 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3467 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3468 		uint64_t *p;
3469 		int i;
3470 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3471 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3472 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3473 			    m, i, (uintmax_t)*p));
3474 	}
3475 #endif
3476 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3477 		vm_page_lock_assert(m, MA_OWNED);
3478 		KASSERT(!pmap_page_is_mapped(m),
3479 		    ("vm_page_free_prep: freeing mapped page %p", m));
3480 	} else
3481 		KASSERT(m->queue == PQ_NONE,
3482 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3483 	VM_CNT_INC(v_tfree);
3484 
3485 	if (vm_page_sbusied(m))
3486 		panic("vm_page_free_prep: freeing busy page %p", m);
3487 
3488 	vm_page_remove(m);
3489 
3490 	/*
3491 	 * If fictitious remove object association and
3492 	 * return.
3493 	 */
3494 	if ((m->flags & PG_FICTITIOUS) != 0) {
3495 		KASSERT(m->wire_count == 1,
3496 		    ("fictitious page %p is not wired", m));
3497 		KASSERT(m->queue == PQ_NONE,
3498 		    ("fictitious page %p is queued", m));
3499 		return (false);
3500 	}
3501 
3502 	/*
3503 	 * Pages need not be dequeued before they are returned to the physical
3504 	 * memory allocator, but they must at least be marked for a deferred
3505 	 * dequeue.
3506 	 */
3507 	if ((m->oflags & VPO_UNMANAGED) == 0)
3508 		vm_page_dequeue_deferred_free(m);
3509 
3510 	m->valid = 0;
3511 	vm_page_undirty(m);
3512 
3513 	if (m->wire_count != 0)
3514 		panic("vm_page_free_prep: freeing wired page %p", m);
3515 	if (m->hold_count != 0) {
3516 		m->flags &= ~PG_ZERO;
3517 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3518 		    ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m));
3519 		m->flags |= PG_UNHOLDFREE;
3520 		return (false);
3521 	}
3522 
3523 	/*
3524 	 * Restore the default memory attribute to the page.
3525 	 */
3526 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3527 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3528 
3529 #if VM_NRESERVLEVEL > 0
3530 	if (vm_reserv_free_page(m))
3531 		return (false);
3532 #endif
3533 
3534 	return (true);
3535 }
3536 
3537 /*
3538  *	vm_page_free_toq:
3539  *
3540  *	Returns the given page to the free list, disassociating it
3541  *	from any VM object.
3542  *
3543  *	The object must be locked.  The page must be locked if it is
3544  *	managed.
3545  */
3546 void
3547 vm_page_free_toq(vm_page_t m)
3548 {
3549 	struct vm_domain *vmd;
3550 
3551 	if (!vm_page_free_prep(m))
3552 		return;
3553 
3554 	vmd = vm_pagequeue_domain(m);
3555 	if (m->pool == VM_FREEPOOL_DEFAULT && vmd->vmd_pgcache != NULL) {
3556 		uma_zfree(vmd->vmd_pgcache, m);
3557 		return;
3558 	}
3559 	vm_domain_free_lock(vmd);
3560 	vm_phys_free_pages(m, 0);
3561 	vm_domain_free_unlock(vmd);
3562 	vm_domain_freecnt_inc(vmd, 1);
3563 }
3564 
3565 /*
3566  *	vm_page_free_pages_toq:
3567  *
3568  *	Returns a list of pages to the free list, disassociating it
3569  *	from any VM object.  In other words, this is equivalent to
3570  *	calling vm_page_free_toq() for each page of a list of VM objects.
3571  *
3572  *	The objects must be locked.  The pages must be locked if it is
3573  *	managed.
3574  */
3575 void
3576 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3577 {
3578 	vm_page_t m;
3579 	int count;
3580 
3581 	if (SLIST_EMPTY(free))
3582 		return;
3583 
3584 	count = 0;
3585 	while ((m = SLIST_FIRST(free)) != NULL) {
3586 		count++;
3587 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3588 		vm_page_free_toq(m);
3589 	}
3590 
3591 	if (update_wire_count)
3592 		vm_wire_sub(count);
3593 }
3594 
3595 /*
3596  *	vm_page_wire:
3597  *
3598  * Mark this page as wired down.  If the page is fictitious, then
3599  * its wire count must remain one.
3600  *
3601  * The page must be locked.
3602  */
3603 void
3604 vm_page_wire(vm_page_t m)
3605 {
3606 
3607 	vm_page_assert_locked(m);
3608 	if ((m->flags & PG_FICTITIOUS) != 0) {
3609 		KASSERT(m->wire_count == 1,
3610 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
3611 		    m));
3612 		return;
3613 	}
3614 	if (m->wire_count == 0) {
3615 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3616 		    m->queue == PQ_NONE,
3617 		    ("vm_page_wire: unmanaged page %p is queued", m));
3618 		vm_wire_add(1);
3619 	}
3620 	m->wire_count++;
3621 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3622 }
3623 
3624 /*
3625  * vm_page_unwire:
3626  *
3627  * Release one wiring of the specified page, potentially allowing it to be
3628  * paged out.  Returns TRUE if the number of wirings transitions to zero and
3629  * FALSE otherwise.
3630  *
3631  * Only managed pages belonging to an object can be paged out.  If the number
3632  * of wirings transitions to zero and the page is eligible for page out, then
3633  * the page is added to the specified paging queue (unless PQ_NONE is
3634  * specified, in which case the page is dequeued if it belongs to a paging
3635  * queue).
3636  *
3637  * If a page is fictitious, then its wire count must always be one.
3638  *
3639  * A managed page must be locked.
3640  */
3641 bool
3642 vm_page_unwire(vm_page_t m, uint8_t queue)
3643 {
3644 	bool unwired;
3645 
3646 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3647 	    ("vm_page_unwire: invalid queue %u request for page %p",
3648 	    queue, m));
3649 	if ((m->oflags & VPO_UNMANAGED) == 0)
3650 		vm_page_assert_locked(m);
3651 
3652 	unwired = vm_page_unwire_noq(m);
3653 	if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3654 		return (unwired);
3655 
3656 	if (vm_page_queue(m) == queue) {
3657 		if (queue == PQ_ACTIVE)
3658 			vm_page_reference(m);
3659 		else if (queue != PQ_NONE)
3660 			vm_page_requeue(m);
3661 	} else {
3662 		vm_page_dequeue(m);
3663 		if (queue != PQ_NONE) {
3664 			vm_page_enqueue(m, queue);
3665 			if (queue == PQ_ACTIVE)
3666 				/* Initialize act_count. */
3667 				vm_page_activate(m);
3668 		}
3669 	}
3670 	return (unwired);
3671 }
3672 
3673 /*
3674  *
3675  * vm_page_unwire_noq:
3676  *
3677  * Unwire a page without (re-)inserting it into a page queue.  It is up
3678  * to the caller to enqueue, requeue, or free the page as appropriate.
3679  * In most cases, vm_page_unwire() should be used instead.
3680  */
3681 bool
3682 vm_page_unwire_noq(vm_page_t m)
3683 {
3684 
3685 	if ((m->oflags & VPO_UNMANAGED) == 0)
3686 		vm_page_assert_locked(m);
3687 	if ((m->flags & PG_FICTITIOUS) != 0) {
3688 		KASSERT(m->wire_count == 1,
3689 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3690 		return (false);
3691 	}
3692 	if (m->wire_count == 0)
3693 		panic("vm_page_unwire: page %p's wire count is zero", m);
3694 	m->wire_count--;
3695 	if (m->wire_count == 0) {
3696 		vm_wire_sub(1);
3697 		return (true);
3698 	} else
3699 		return (false);
3700 }
3701 
3702 /*
3703  * Move the specified page to the tail of the inactive queue, or requeue
3704  * the page if it is already in the inactive queue.
3705  *
3706  * The page must be locked.
3707  */
3708 void
3709 vm_page_deactivate(vm_page_t m)
3710 {
3711 
3712 	vm_page_assert_locked(m);
3713 
3714 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3715 		return;
3716 
3717 	if (!vm_page_inactive(m)) {
3718 		vm_page_dequeue(m);
3719 		vm_page_enqueue(m, PQ_INACTIVE);
3720 	} else
3721 		vm_page_requeue(m);
3722 }
3723 
3724 /*
3725  * Move the specified page close to the head of the inactive queue,
3726  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3727  * As with regular enqueues, we use a per-CPU batch queue to reduce
3728  * contention on the page queue lock.
3729  *
3730  * The page must be locked.
3731  */
3732 void
3733 vm_page_deactivate_noreuse(vm_page_t m)
3734 {
3735 
3736 	vm_page_assert_locked(m);
3737 
3738 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3739 		return;
3740 
3741 	if (!vm_page_inactive(m)) {
3742 		vm_page_dequeue(m);
3743 		m->queue = PQ_INACTIVE;
3744 	}
3745 	if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3746 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3747 	vm_pqbatch_submit_page(m, PQ_INACTIVE);
3748 }
3749 
3750 /*
3751  * vm_page_launder
3752  *
3753  * 	Put a page in the laundry, or requeue it if it is already there.
3754  */
3755 void
3756 vm_page_launder(vm_page_t m)
3757 {
3758 
3759 	vm_page_assert_locked(m);
3760 	if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0)
3761 		return;
3762 
3763 	if (vm_page_in_laundry(m))
3764 		vm_page_requeue(m);
3765 	else {
3766 		vm_page_dequeue(m);
3767 		vm_page_enqueue(m, PQ_LAUNDRY);
3768 	}
3769 }
3770 
3771 /*
3772  * vm_page_unswappable
3773  *
3774  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3775  */
3776 void
3777 vm_page_unswappable(vm_page_t m)
3778 {
3779 
3780 	vm_page_assert_locked(m);
3781 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3782 	    ("page %p already unswappable", m));
3783 
3784 	vm_page_dequeue(m);
3785 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
3786 }
3787 
3788 /*
3789  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3790  * if the page is freed and false otherwise.
3791  *
3792  * The page must be managed.  The page and its containing object must be
3793  * locked.
3794  */
3795 bool
3796 vm_page_try_to_free(vm_page_t m)
3797 {
3798 
3799 	vm_page_assert_locked(m);
3800 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3801 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3802 	if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m))
3803 		return (false);
3804 	if (m->object->ref_count != 0) {
3805 		pmap_remove_all(m);
3806 		if (m->dirty != 0)
3807 			return (false);
3808 	}
3809 	vm_page_free(m);
3810 	return (true);
3811 }
3812 
3813 /*
3814  * vm_page_advise
3815  *
3816  * 	Apply the specified advice to the given page.
3817  *
3818  *	The object and page must be locked.
3819  */
3820 void
3821 vm_page_advise(vm_page_t m, int advice)
3822 {
3823 
3824 	vm_page_assert_locked(m);
3825 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3826 	if (advice == MADV_FREE)
3827 		/*
3828 		 * Mark the page clean.  This will allow the page to be freed
3829 		 * without first paging it out.  MADV_FREE pages are often
3830 		 * quickly reused by malloc(3), so we do not do anything that
3831 		 * would result in a page fault on a later access.
3832 		 */
3833 		vm_page_undirty(m);
3834 	else if (advice != MADV_DONTNEED) {
3835 		if (advice == MADV_WILLNEED)
3836 			vm_page_activate(m);
3837 		return;
3838 	}
3839 
3840 	/*
3841 	 * Clear any references to the page.  Otherwise, the page daemon will
3842 	 * immediately reactivate the page.
3843 	 */
3844 	vm_page_aflag_clear(m, PGA_REFERENCED);
3845 
3846 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3847 		vm_page_dirty(m);
3848 
3849 	/*
3850 	 * Place clean pages near the head of the inactive queue rather than
3851 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3852 	 * the page will be reused quickly.  Dirty pages not already in the
3853 	 * laundry are moved there.
3854 	 */
3855 	if (m->dirty == 0)
3856 		vm_page_deactivate_noreuse(m);
3857 	else if (!vm_page_in_laundry(m))
3858 		vm_page_launder(m);
3859 }
3860 
3861 /*
3862  * Grab a page, waiting until we are waken up due to the page
3863  * changing state.  We keep on waiting, if the page continues
3864  * to be in the object.  If the page doesn't exist, first allocate it
3865  * and then conditionally zero it.
3866  *
3867  * This routine may sleep.
3868  *
3869  * The object must be locked on entry.  The lock will, however, be released
3870  * and reacquired if the routine sleeps.
3871  */
3872 vm_page_t
3873 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3874 {
3875 	vm_page_t m;
3876 	int sleep;
3877 	int pflags;
3878 
3879 	VM_OBJECT_ASSERT_WLOCKED(object);
3880 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3881 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3882 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3883 	pflags = allocflags &
3884 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3885 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3886 		pflags |= VM_ALLOC_WAITFAIL;
3887 retrylookup:
3888 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3889 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3890 		    vm_page_xbusied(m) : vm_page_busied(m);
3891 		if (sleep) {
3892 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3893 				return (NULL);
3894 			/*
3895 			 * Reference the page before unlocking and
3896 			 * sleeping so that the page daemon is less
3897 			 * likely to reclaim it.
3898 			 */
3899 			vm_page_aflag_set(m, PGA_REFERENCED);
3900 			vm_page_lock(m);
3901 			VM_OBJECT_WUNLOCK(object);
3902 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3903 			    VM_ALLOC_IGN_SBUSY) != 0);
3904 			VM_OBJECT_WLOCK(object);
3905 			goto retrylookup;
3906 		} else {
3907 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3908 				vm_page_lock(m);
3909 				vm_page_wire(m);
3910 				vm_page_unlock(m);
3911 			}
3912 			if ((allocflags &
3913 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3914 				vm_page_xbusy(m);
3915 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3916 				vm_page_sbusy(m);
3917 			return (m);
3918 		}
3919 	}
3920 	m = vm_page_alloc(object, pindex, pflags);
3921 	if (m == NULL) {
3922 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3923 			return (NULL);
3924 		goto retrylookup;
3925 	}
3926 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3927 		pmap_zero_page(m);
3928 	return (m);
3929 }
3930 
3931 /*
3932  * Return the specified range of pages from the given object.  For each
3933  * page offset within the range, if a page already exists within the object
3934  * at that offset and it is busy, then wait for it to change state.  If,
3935  * instead, the page doesn't exist, then allocate it.
3936  *
3937  * The caller must always specify an allocation class.
3938  *
3939  * allocation classes:
3940  *	VM_ALLOC_NORMAL		normal process request
3941  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3942  *
3943  * The caller must always specify that the pages are to be busied and/or
3944  * wired.
3945  *
3946  * optional allocation flags:
3947  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3948  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3949  *	VM_ALLOC_NOWAIT		do not sleep
3950  *	VM_ALLOC_SBUSY		set page to sbusy state
3951  *	VM_ALLOC_WIRED		wire the pages
3952  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3953  *
3954  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3955  * may return a partial prefix of the requested range.
3956  */
3957 int
3958 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3959     vm_page_t *ma, int count)
3960 {
3961 	vm_page_t m, mpred;
3962 	int pflags;
3963 	int i;
3964 	bool sleep;
3965 
3966 	VM_OBJECT_ASSERT_WLOCKED(object);
3967 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3968 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3969 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3970 	    (allocflags & VM_ALLOC_WIRED) != 0,
3971 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3972 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3973 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3974 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3975 	if (count == 0)
3976 		return (0);
3977 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3978 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3979 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3980 		pflags |= VM_ALLOC_WAITFAIL;
3981 	i = 0;
3982 retrylookup:
3983 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3984 	if (m == NULL || m->pindex != pindex + i) {
3985 		mpred = m;
3986 		m = NULL;
3987 	} else
3988 		mpred = TAILQ_PREV(m, pglist, listq);
3989 	for (; i < count; i++) {
3990 		if (m != NULL) {
3991 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3992 			    vm_page_xbusied(m) : vm_page_busied(m);
3993 			if (sleep) {
3994 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3995 					break;
3996 				/*
3997 				 * Reference the page before unlocking and
3998 				 * sleeping so that the page daemon is less
3999 				 * likely to reclaim it.
4000 				 */
4001 				vm_page_aflag_set(m, PGA_REFERENCED);
4002 				vm_page_lock(m);
4003 				VM_OBJECT_WUNLOCK(object);
4004 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
4005 				    VM_ALLOC_IGN_SBUSY) != 0);
4006 				VM_OBJECT_WLOCK(object);
4007 				goto retrylookup;
4008 			}
4009 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
4010 				vm_page_lock(m);
4011 				vm_page_wire(m);
4012 				vm_page_unlock(m);
4013 			}
4014 			if ((allocflags & (VM_ALLOC_NOBUSY |
4015 			    VM_ALLOC_SBUSY)) == 0)
4016 				vm_page_xbusy(m);
4017 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
4018 				vm_page_sbusy(m);
4019 		} else {
4020 			m = vm_page_alloc_after(object, pindex + i,
4021 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4022 			if (m == NULL) {
4023 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4024 					break;
4025 				goto retrylookup;
4026 			}
4027 		}
4028 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
4029 			if ((m->flags & PG_ZERO) == 0)
4030 				pmap_zero_page(m);
4031 			m->valid = VM_PAGE_BITS_ALL;
4032 		}
4033 		ma[i] = mpred = m;
4034 		m = vm_page_next(m);
4035 	}
4036 	return (i);
4037 }
4038 
4039 /*
4040  * Mapping function for valid or dirty bits in a page.
4041  *
4042  * Inputs are required to range within a page.
4043  */
4044 vm_page_bits_t
4045 vm_page_bits(int base, int size)
4046 {
4047 	int first_bit;
4048 	int last_bit;
4049 
4050 	KASSERT(
4051 	    base + size <= PAGE_SIZE,
4052 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4053 	);
4054 
4055 	if (size == 0)		/* handle degenerate case */
4056 		return (0);
4057 
4058 	first_bit = base >> DEV_BSHIFT;
4059 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4060 
4061 	return (((vm_page_bits_t)2 << last_bit) -
4062 	    ((vm_page_bits_t)1 << first_bit));
4063 }
4064 
4065 /*
4066  *	vm_page_set_valid_range:
4067  *
4068  *	Sets portions of a page valid.  The arguments are expected
4069  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4070  *	of any partial chunks touched by the range.  The invalid portion of
4071  *	such chunks will be zeroed.
4072  *
4073  *	(base + size) must be less then or equal to PAGE_SIZE.
4074  */
4075 void
4076 vm_page_set_valid_range(vm_page_t m, int base, int size)
4077 {
4078 	int endoff, frag;
4079 
4080 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4081 	if (size == 0)	/* handle degenerate case */
4082 		return;
4083 
4084 	/*
4085 	 * If the base is not DEV_BSIZE aligned and the valid
4086 	 * bit is clear, we have to zero out a portion of the
4087 	 * first block.
4088 	 */
4089 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4090 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4091 		pmap_zero_page_area(m, frag, base - frag);
4092 
4093 	/*
4094 	 * If the ending offset is not DEV_BSIZE aligned and the
4095 	 * valid bit is clear, we have to zero out a portion of
4096 	 * the last block.
4097 	 */
4098 	endoff = base + size;
4099 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4100 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4101 		pmap_zero_page_area(m, endoff,
4102 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4103 
4104 	/*
4105 	 * Assert that no previously invalid block that is now being validated
4106 	 * is already dirty.
4107 	 */
4108 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4109 	    ("vm_page_set_valid_range: page %p is dirty", m));
4110 
4111 	/*
4112 	 * Set valid bits inclusive of any overlap.
4113 	 */
4114 	m->valid |= vm_page_bits(base, size);
4115 }
4116 
4117 /*
4118  * Clear the given bits from the specified page's dirty field.
4119  */
4120 static __inline void
4121 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4122 {
4123 	uintptr_t addr;
4124 #if PAGE_SIZE < 16384
4125 	int shift;
4126 #endif
4127 
4128 	/*
4129 	 * If the object is locked and the page is neither exclusive busy nor
4130 	 * write mapped, then the page's dirty field cannot possibly be
4131 	 * set by a concurrent pmap operation.
4132 	 */
4133 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4134 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4135 		m->dirty &= ~pagebits;
4136 	else {
4137 		/*
4138 		 * The pmap layer can call vm_page_dirty() without
4139 		 * holding a distinguished lock.  The combination of
4140 		 * the object's lock and an atomic operation suffice
4141 		 * to guarantee consistency of the page dirty field.
4142 		 *
4143 		 * For PAGE_SIZE == 32768 case, compiler already
4144 		 * properly aligns the dirty field, so no forcible
4145 		 * alignment is needed. Only require existence of
4146 		 * atomic_clear_64 when page size is 32768.
4147 		 */
4148 		addr = (uintptr_t)&m->dirty;
4149 #if PAGE_SIZE == 32768
4150 		atomic_clear_64((uint64_t *)addr, pagebits);
4151 #elif PAGE_SIZE == 16384
4152 		atomic_clear_32((uint32_t *)addr, pagebits);
4153 #else		/* PAGE_SIZE <= 8192 */
4154 		/*
4155 		 * Use a trick to perform a 32-bit atomic on the
4156 		 * containing aligned word, to not depend on the existence
4157 		 * of atomic_clear_{8, 16}.
4158 		 */
4159 		shift = addr & (sizeof(uint32_t) - 1);
4160 #if BYTE_ORDER == BIG_ENDIAN
4161 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4162 #else
4163 		shift *= NBBY;
4164 #endif
4165 		addr &= ~(sizeof(uint32_t) - 1);
4166 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
4167 #endif		/* PAGE_SIZE */
4168 	}
4169 }
4170 
4171 /*
4172  *	vm_page_set_validclean:
4173  *
4174  *	Sets portions of a page valid and clean.  The arguments are expected
4175  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4176  *	of any partial chunks touched by the range.  The invalid portion of
4177  *	such chunks will be zero'd.
4178  *
4179  *	(base + size) must be less then or equal to PAGE_SIZE.
4180  */
4181 void
4182 vm_page_set_validclean(vm_page_t m, int base, int size)
4183 {
4184 	vm_page_bits_t oldvalid, pagebits;
4185 	int endoff, frag;
4186 
4187 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4188 	if (size == 0)	/* handle degenerate case */
4189 		return;
4190 
4191 	/*
4192 	 * If the base is not DEV_BSIZE aligned and the valid
4193 	 * bit is clear, we have to zero out a portion of the
4194 	 * first block.
4195 	 */
4196 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4197 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4198 		pmap_zero_page_area(m, frag, base - frag);
4199 
4200 	/*
4201 	 * If the ending offset is not DEV_BSIZE aligned and the
4202 	 * valid bit is clear, we have to zero out a portion of
4203 	 * the last block.
4204 	 */
4205 	endoff = base + size;
4206 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4207 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4208 		pmap_zero_page_area(m, endoff,
4209 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4210 
4211 	/*
4212 	 * Set valid, clear dirty bits.  If validating the entire
4213 	 * page we can safely clear the pmap modify bit.  We also
4214 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
4215 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4216 	 * be set again.
4217 	 *
4218 	 * We set valid bits inclusive of any overlap, but we can only
4219 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4220 	 * the range.
4221 	 */
4222 	oldvalid = m->valid;
4223 	pagebits = vm_page_bits(base, size);
4224 	m->valid |= pagebits;
4225 #if 0	/* NOT YET */
4226 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4227 		frag = DEV_BSIZE - frag;
4228 		base += frag;
4229 		size -= frag;
4230 		if (size < 0)
4231 			size = 0;
4232 	}
4233 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4234 #endif
4235 	if (base == 0 && size == PAGE_SIZE) {
4236 		/*
4237 		 * The page can only be modified within the pmap if it is
4238 		 * mapped, and it can only be mapped if it was previously
4239 		 * fully valid.
4240 		 */
4241 		if (oldvalid == VM_PAGE_BITS_ALL)
4242 			/*
4243 			 * Perform the pmap_clear_modify() first.  Otherwise,
4244 			 * a concurrent pmap operation, such as
4245 			 * pmap_protect(), could clear a modification in the
4246 			 * pmap and set the dirty field on the page before
4247 			 * pmap_clear_modify() had begun and after the dirty
4248 			 * field was cleared here.
4249 			 */
4250 			pmap_clear_modify(m);
4251 		m->dirty = 0;
4252 		m->oflags &= ~VPO_NOSYNC;
4253 	} else if (oldvalid != VM_PAGE_BITS_ALL)
4254 		m->dirty &= ~pagebits;
4255 	else
4256 		vm_page_clear_dirty_mask(m, pagebits);
4257 }
4258 
4259 void
4260 vm_page_clear_dirty(vm_page_t m, int base, int size)
4261 {
4262 
4263 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4264 }
4265 
4266 /*
4267  *	vm_page_set_invalid:
4268  *
4269  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4270  *	valid and dirty bits for the effected areas are cleared.
4271  */
4272 void
4273 vm_page_set_invalid(vm_page_t m, int base, int size)
4274 {
4275 	vm_page_bits_t bits;
4276 	vm_object_t object;
4277 
4278 	object = m->object;
4279 	VM_OBJECT_ASSERT_WLOCKED(object);
4280 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4281 	    size >= object->un_pager.vnp.vnp_size)
4282 		bits = VM_PAGE_BITS_ALL;
4283 	else
4284 		bits = vm_page_bits(base, size);
4285 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4286 	    bits != 0)
4287 		pmap_remove_all(m);
4288 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4289 	    !pmap_page_is_mapped(m),
4290 	    ("vm_page_set_invalid: page %p is mapped", m));
4291 	m->valid &= ~bits;
4292 	m->dirty &= ~bits;
4293 }
4294 
4295 /*
4296  * vm_page_zero_invalid()
4297  *
4298  *	The kernel assumes that the invalid portions of a page contain
4299  *	garbage, but such pages can be mapped into memory by user code.
4300  *	When this occurs, we must zero out the non-valid portions of the
4301  *	page so user code sees what it expects.
4302  *
4303  *	Pages are most often semi-valid when the end of a file is mapped
4304  *	into memory and the file's size is not page aligned.
4305  */
4306 void
4307 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4308 {
4309 	int b;
4310 	int i;
4311 
4312 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4313 	/*
4314 	 * Scan the valid bits looking for invalid sections that
4315 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4316 	 * valid bit may be set ) have already been zeroed by
4317 	 * vm_page_set_validclean().
4318 	 */
4319 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4320 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4321 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4322 			if (i > b) {
4323 				pmap_zero_page_area(m,
4324 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4325 			}
4326 			b = i + 1;
4327 		}
4328 	}
4329 
4330 	/*
4331 	 * setvalid is TRUE when we can safely set the zero'd areas
4332 	 * as being valid.  We can do this if there are no cache consistancy
4333 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4334 	 */
4335 	if (setvalid)
4336 		m->valid = VM_PAGE_BITS_ALL;
4337 }
4338 
4339 /*
4340  *	vm_page_is_valid:
4341  *
4342  *	Is (partial) page valid?  Note that the case where size == 0
4343  *	will return FALSE in the degenerate case where the page is
4344  *	entirely invalid, and TRUE otherwise.
4345  */
4346 int
4347 vm_page_is_valid(vm_page_t m, int base, int size)
4348 {
4349 	vm_page_bits_t bits;
4350 
4351 	VM_OBJECT_ASSERT_LOCKED(m->object);
4352 	bits = vm_page_bits(base, size);
4353 	return (m->valid != 0 && (m->valid & bits) == bits);
4354 }
4355 
4356 /*
4357  * Returns true if all of the specified predicates are true for the entire
4358  * (super)page and false otherwise.
4359  */
4360 bool
4361 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4362 {
4363 	vm_object_t object;
4364 	int i, npages;
4365 
4366 	object = m->object;
4367 	if (skip_m != NULL && skip_m->object != object)
4368 		return (false);
4369 	VM_OBJECT_ASSERT_LOCKED(object);
4370 	npages = atop(pagesizes[m->psind]);
4371 
4372 	/*
4373 	 * The physically contiguous pages that make up a superpage, i.e., a
4374 	 * page with a page size index ("psind") greater than zero, will
4375 	 * occupy adjacent entries in vm_page_array[].
4376 	 */
4377 	for (i = 0; i < npages; i++) {
4378 		/* Always test object consistency, including "skip_m". */
4379 		if (m[i].object != object)
4380 			return (false);
4381 		if (&m[i] == skip_m)
4382 			continue;
4383 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4384 			return (false);
4385 		if ((flags & PS_ALL_DIRTY) != 0) {
4386 			/*
4387 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4388 			 * might stop this case from spuriously returning
4389 			 * "false".  However, that would require a write lock
4390 			 * on the object containing "m[i]".
4391 			 */
4392 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4393 				return (false);
4394 		}
4395 		if ((flags & PS_ALL_VALID) != 0 &&
4396 		    m[i].valid != VM_PAGE_BITS_ALL)
4397 			return (false);
4398 	}
4399 	return (true);
4400 }
4401 
4402 /*
4403  * Set the page's dirty bits if the page is modified.
4404  */
4405 void
4406 vm_page_test_dirty(vm_page_t m)
4407 {
4408 
4409 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4410 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4411 		vm_page_dirty(m);
4412 }
4413 
4414 void
4415 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4416 {
4417 
4418 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4419 }
4420 
4421 void
4422 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4423 {
4424 
4425 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4426 }
4427 
4428 int
4429 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4430 {
4431 
4432 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4433 }
4434 
4435 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4436 void
4437 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4438 {
4439 
4440 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4441 }
4442 
4443 void
4444 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4445 {
4446 
4447 	mtx_assert_(vm_page_lockptr(m), a, file, line);
4448 }
4449 #endif
4450 
4451 #ifdef INVARIANTS
4452 void
4453 vm_page_object_lock_assert(vm_page_t m)
4454 {
4455 
4456 	/*
4457 	 * Certain of the page's fields may only be modified by the
4458 	 * holder of the containing object's lock or the exclusive busy.
4459 	 * holder.  Unfortunately, the holder of the write busy is
4460 	 * not recorded, and thus cannot be checked here.
4461 	 */
4462 	if (m->object != NULL && !vm_page_xbusied(m))
4463 		VM_OBJECT_ASSERT_WLOCKED(m->object);
4464 }
4465 
4466 void
4467 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4468 {
4469 
4470 	if ((bits & PGA_WRITEABLE) == 0)
4471 		return;
4472 
4473 	/*
4474 	 * The PGA_WRITEABLE flag can only be set if the page is
4475 	 * managed, is exclusively busied or the object is locked.
4476 	 * Currently, this flag is only set by pmap_enter().
4477 	 */
4478 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4479 	    ("PGA_WRITEABLE on unmanaged page"));
4480 	if (!vm_page_xbusied(m))
4481 		VM_OBJECT_ASSERT_LOCKED(m->object);
4482 }
4483 #endif
4484 
4485 #include "opt_ddb.h"
4486 #ifdef DDB
4487 #include <sys/kernel.h>
4488 
4489 #include <ddb/ddb.h>
4490 
4491 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4492 {
4493 
4494 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4495 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4496 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4497 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4498 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4499 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4500 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4501 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4502 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4503 }
4504 
4505 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4506 {
4507 	int dom;
4508 
4509 	db_printf("pq_free %d\n", vm_free_count());
4510 	for (dom = 0; dom < vm_ndomains; dom++) {
4511 		db_printf(
4512     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4513 		    dom,
4514 		    vm_dom[dom].vmd_page_count,
4515 		    vm_dom[dom].vmd_free_count,
4516 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4517 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4518 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4519 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4520 	}
4521 }
4522 
4523 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4524 {
4525 	vm_page_t m;
4526 	boolean_t phys, virt;
4527 
4528 	if (!have_addr) {
4529 		db_printf("show pginfo addr\n");
4530 		return;
4531 	}
4532 
4533 	phys = strchr(modif, 'p') != NULL;
4534 	virt = strchr(modif, 'v') != NULL;
4535 	if (virt)
4536 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
4537 	else if (phys)
4538 		m = PHYS_TO_VM_PAGE(addr);
4539 	else
4540 		m = (vm_page_t)addr;
4541 	db_printf(
4542     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4543     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4544 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4545 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4546 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4547 }
4548 #endif /* DDB */
4549