xref: /freebsd/sys/vm/vm_page.c (revision 0183e0151669735d62584fbba9125ed90716af5e)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * The Mach Operating System project at Carnegie-Mellon University.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *			GENERAL RULES ON VM_PAGE MANIPULATION
65  *
66  *	- A page queue lock is required when adding or removing a page from a
67  *	  page queue regardless of other locks or the busy state of a page.
68  *
69  *		* In general, no thread besides the page daemon can acquire or
70  *		  hold more than one page queue lock at a time.
71  *
72  *		* The page daemon can acquire and hold any pair of page queue
73  *		  locks in any order.
74  *
75  *	- The object lock is required when inserting or removing
76  *	  pages from an object (vm_page_insert() or vm_page_remove()).
77  *
78  */
79 
80 /*
81  *	Resident memory management module.
82  */
83 
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
86 
87 #include "opt_vm.h"
88 
89 #include <sys/param.h>
90 #include <sys/systm.h>
91 #include <sys/lock.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/linker.h>
95 #include <sys/malloc.h>
96 #include <sys/mman.h>
97 #include <sys/msgbuf.h>
98 #include <sys/mutex.h>
99 #include <sys/proc.h>
100 #include <sys/rwlock.h>
101 #include <sys/sbuf.h>
102 #include <sys/smp.h>
103 #include <sys/sysctl.h>
104 #include <sys/vmmeter.h>
105 #include <sys/vnode.h>
106 
107 #include <vm/vm.h>
108 #include <vm/pmap.h>
109 #include <vm/vm_param.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pageout.h>
114 #include <vm/vm_pager.h>
115 #include <vm/vm_phys.h>
116 #include <vm/vm_radix.h>
117 #include <vm/vm_reserv.h>
118 #include <vm/vm_extern.h>
119 #include <vm/uma.h>
120 #include <vm/uma_int.h>
121 
122 #include <machine/md_var.h>
123 
124 /*
125  *	Associated with page of user-allocatable memory is a
126  *	page structure.
127  */
128 
129 struct vm_domain vm_dom[MAXMEMDOM];
130 struct mtx_padalign vm_page_queue_free_mtx;
131 
132 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
133 
134 /*
135  * bogus page -- for I/O to/from partially complete buffers,
136  * or for paging into sparsely invalid regions.
137  */
138 vm_page_t bogus_page;
139 
140 vm_page_t vm_page_array;
141 long vm_page_array_size;
142 long first_page;
143 
144 static int boot_pages = UMA_BOOT_PAGES;
145 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
146     &boot_pages, 0,
147     "number of pages allocated for bootstrapping the VM system");
148 
149 static int pa_tryrelock_restart;
150 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
151     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
152 
153 static TAILQ_HEAD(, vm_page) blacklist_head;
154 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
155 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
156     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
157 
158 /* Is the page daemon waiting for free pages? */
159 static int vm_pageout_pages_needed;
160 
161 static uma_zone_t fakepg_zone;
162 
163 static void vm_page_alloc_check(vm_page_t m);
164 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
165 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
166 static void vm_page_free_wakeup(void);
167 static void vm_page_init(void *dummy);
168 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
169     vm_pindex_t pindex, vm_page_t mpred);
170 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
171     vm_page_t mpred);
172 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
173     vm_paddr_t high);
174 
175 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
176 
177 static void
178 vm_page_init(void *dummy)
179 {
180 
181 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
182 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
183 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
184 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
185 }
186 
187 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
188 #if PAGE_SIZE == 32768
189 #ifdef CTASSERT
190 CTASSERT(sizeof(u_long) >= 8);
191 #endif
192 #endif
193 
194 /*
195  * Try to acquire a physical address lock while a pmap is locked.  If we
196  * fail to trylock we unlock and lock the pmap directly and cache the
197  * locked pa in *locked.  The caller should then restart their loop in case
198  * the virtual to physical mapping has changed.
199  */
200 int
201 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
202 {
203 	vm_paddr_t lockpa;
204 
205 	lockpa = *locked;
206 	*locked = pa;
207 	if (lockpa) {
208 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
209 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
210 			return (0);
211 		PA_UNLOCK(lockpa);
212 	}
213 	if (PA_TRYLOCK(pa))
214 		return (0);
215 	PMAP_UNLOCK(pmap);
216 	atomic_add_int(&pa_tryrelock_restart, 1);
217 	PA_LOCK(pa);
218 	PMAP_LOCK(pmap);
219 	return (EAGAIN);
220 }
221 
222 /*
223  *	vm_set_page_size:
224  *
225  *	Sets the page size, perhaps based upon the memory
226  *	size.  Must be called before any use of page-size
227  *	dependent functions.
228  */
229 void
230 vm_set_page_size(void)
231 {
232 	if (vm_cnt.v_page_size == 0)
233 		vm_cnt.v_page_size = PAGE_SIZE;
234 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
235 		panic("vm_set_page_size: page size not a power of two");
236 }
237 
238 /*
239  *	vm_page_blacklist_next:
240  *
241  *	Find the next entry in the provided string of blacklist
242  *	addresses.  Entries are separated by space, comma, or newline.
243  *	If an invalid integer is encountered then the rest of the
244  *	string is skipped.  Updates the list pointer to the next
245  *	character, or NULL if the string is exhausted or invalid.
246  */
247 static vm_paddr_t
248 vm_page_blacklist_next(char **list, char *end)
249 {
250 	vm_paddr_t bad;
251 	char *cp, *pos;
252 
253 	if (list == NULL || *list == NULL)
254 		return (0);
255 	if (**list =='\0') {
256 		*list = NULL;
257 		return (0);
258 	}
259 
260 	/*
261 	 * If there's no end pointer then the buffer is coming from
262 	 * the kenv and we know it's null-terminated.
263 	 */
264 	if (end == NULL)
265 		end = *list + strlen(*list);
266 
267 	/* Ensure that strtoq() won't walk off the end */
268 	if (*end != '\0') {
269 		if (*end == '\n' || *end == ' ' || *end  == ',')
270 			*end = '\0';
271 		else {
272 			printf("Blacklist not terminated, skipping\n");
273 			*list = NULL;
274 			return (0);
275 		}
276 	}
277 
278 	for (pos = *list; *pos != '\0'; pos = cp) {
279 		bad = strtoq(pos, &cp, 0);
280 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
281 			if (bad == 0) {
282 				if (++cp < end)
283 					continue;
284 				else
285 					break;
286 			}
287 		} else
288 			break;
289 		if (*cp == '\0' || ++cp >= end)
290 			*list = NULL;
291 		else
292 			*list = cp;
293 		return (trunc_page(bad));
294 	}
295 	printf("Garbage in RAM blacklist, skipping\n");
296 	*list = NULL;
297 	return (0);
298 }
299 
300 /*
301  *	vm_page_blacklist_check:
302  *
303  *	Iterate through the provided string of blacklist addresses, pulling
304  *	each entry out of the physical allocator free list and putting it
305  *	onto a list for reporting via the vm.page_blacklist sysctl.
306  */
307 static void
308 vm_page_blacklist_check(char *list, char *end)
309 {
310 	vm_paddr_t pa;
311 	vm_page_t m;
312 	char *next;
313 	int ret;
314 
315 	next = list;
316 	while (next != NULL) {
317 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
318 			continue;
319 		m = vm_phys_paddr_to_vm_page(pa);
320 		if (m == NULL)
321 			continue;
322 		mtx_lock(&vm_page_queue_free_mtx);
323 		ret = vm_phys_unfree_page(m);
324 		mtx_unlock(&vm_page_queue_free_mtx);
325 		if (ret == TRUE) {
326 			TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
327 			if (bootverbose)
328 				printf("Skipping page with pa 0x%jx\n",
329 				    (uintmax_t)pa);
330 		}
331 	}
332 }
333 
334 /*
335  *	vm_page_blacklist_load:
336  *
337  *	Search for a special module named "ram_blacklist".  It'll be a
338  *	plain text file provided by the user via the loader directive
339  *	of the same name.
340  */
341 static void
342 vm_page_blacklist_load(char **list, char **end)
343 {
344 	void *mod;
345 	u_char *ptr;
346 	u_int len;
347 
348 	mod = NULL;
349 	ptr = NULL;
350 
351 	mod = preload_search_by_type("ram_blacklist");
352 	if (mod != NULL) {
353 		ptr = preload_fetch_addr(mod);
354 		len = preload_fetch_size(mod);
355         }
356 	*list = ptr;
357 	if (ptr != NULL)
358 		*end = ptr + len;
359 	else
360 		*end = NULL;
361 	return;
362 }
363 
364 static int
365 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
366 {
367 	vm_page_t m;
368 	struct sbuf sbuf;
369 	int error, first;
370 
371 	first = 1;
372 	error = sysctl_wire_old_buffer(req, 0);
373 	if (error != 0)
374 		return (error);
375 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
376 	TAILQ_FOREACH(m, &blacklist_head, listq) {
377 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
378 		    (uintmax_t)m->phys_addr);
379 		first = 0;
380 	}
381 	error = sbuf_finish(&sbuf);
382 	sbuf_delete(&sbuf);
383 	return (error);
384 }
385 
386 static void
387 vm_page_domain_init(struct vm_domain *vmd)
388 {
389 	struct vm_pagequeue *pq;
390 	int i;
391 
392 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
393 	    "vm inactive pagequeue";
394 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
395 	    &vm_cnt.v_inactive_count;
396 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
397 	    "vm active pagequeue";
398 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
399 	    &vm_cnt.v_active_count;
400 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
401 	    "vm laundry pagequeue";
402 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
403 	    &vm_cnt.v_laundry_count;
404 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
405 	    "vm unswappable pagequeue";
406 	/* Unswappable dirty pages are counted as being in the laundry. */
407 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_vcnt) =
408 	    &vm_cnt.v_laundry_count;
409 	vmd->vmd_page_count = 0;
410 	vmd->vmd_free_count = 0;
411 	vmd->vmd_segs = 0;
412 	vmd->vmd_oom = FALSE;
413 	for (i = 0; i < PQ_COUNT; i++) {
414 		pq = &vmd->vmd_pagequeues[i];
415 		TAILQ_INIT(&pq->pq_pl);
416 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
417 		    MTX_DEF | MTX_DUPOK);
418 	}
419 }
420 
421 /*
422  *	vm_page_startup:
423  *
424  *	Initializes the resident memory module.  Allocates physical memory for
425  *	bootstrapping UMA and some data structures that are used to manage
426  *	physical pages.  Initializes these structures, and populates the free
427  *	page queues.
428  */
429 vm_offset_t
430 vm_page_startup(vm_offset_t vaddr)
431 {
432 	vm_offset_t mapped;
433 	vm_paddr_t high_avail, low_avail, page_range, size;
434 	vm_paddr_t new_end;
435 	int i;
436 	vm_paddr_t pa;
437 	vm_paddr_t last_pa;
438 	char *list, *listend;
439 	vm_paddr_t end;
440 	vm_paddr_t biggestsize;
441 	int biggestone;
442 	int pages_per_zone;
443 
444 	biggestsize = 0;
445 	biggestone = 0;
446 	vaddr = round_page(vaddr);
447 
448 	for (i = 0; phys_avail[i + 1]; i += 2) {
449 		phys_avail[i] = round_page(phys_avail[i]);
450 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
451 	}
452 	for (i = 0; phys_avail[i + 1]; i += 2) {
453 		size = phys_avail[i + 1] - phys_avail[i];
454 		if (size > biggestsize) {
455 			biggestone = i;
456 			biggestsize = size;
457 		}
458 	}
459 
460 	end = phys_avail[biggestone+1];
461 
462 	/*
463 	 * Initialize the page and queue locks.
464 	 */
465 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
466 	for (i = 0; i < PA_LOCK_COUNT; i++)
467 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
468 	for (i = 0; i < vm_ndomains; i++)
469 		vm_page_domain_init(&vm_dom[i]);
470 
471 	/*
472 	 * Almost all of the pages needed for bootstrapping UMA are used
473 	 * for zone structures, so if the number of CPUs results in those
474 	 * structures taking more than one page each, we set aside more pages
475 	 * in proportion to the zone structure size.
476 	 */
477 	pages_per_zone = howmany(sizeof(struct uma_zone) +
478 	    sizeof(struct uma_cache) * (mp_maxid + 1), UMA_SLAB_SIZE);
479 	if (pages_per_zone > 1) {
480 		/* Reserve more pages so that we don't run out. */
481 		boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
482 	}
483 
484 	/*
485 	 * Allocate memory for use when boot strapping the kernel memory
486 	 * allocator.
487 	 *
488 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
489 	 * manually fetch the value.
490 	 */
491 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
492 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
493 	new_end = trunc_page(new_end);
494 	mapped = pmap_map(&vaddr, new_end, end,
495 	    VM_PROT_READ | VM_PROT_WRITE);
496 	bzero((void *)mapped, end - new_end);
497 	uma_startup((void *)mapped, boot_pages);
498 
499 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
500     defined(__i386__) || defined(__mips__)
501 	/*
502 	 * Allocate a bitmap to indicate that a random physical page
503 	 * needs to be included in a minidump.
504 	 *
505 	 * The amd64 port needs this to indicate which direct map pages
506 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
507 	 *
508 	 * However, i386 still needs this workspace internally within the
509 	 * minidump code.  In theory, they are not needed on i386, but are
510 	 * included should the sf_buf code decide to use them.
511 	 */
512 	last_pa = 0;
513 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
514 		if (dump_avail[i + 1] > last_pa)
515 			last_pa = dump_avail[i + 1];
516 	page_range = last_pa / PAGE_SIZE;
517 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
518 	new_end -= vm_page_dump_size;
519 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
520 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
521 	bzero((void *)vm_page_dump, vm_page_dump_size);
522 #endif
523 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
524 	/*
525 	 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
526 	 * When pmap_map() uses the direct map, they are not automatically
527 	 * included.
528 	 */
529 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
530 		dump_add_page(pa);
531 #endif
532 	phys_avail[biggestone + 1] = new_end;
533 #ifdef __amd64__
534 	/*
535 	 * Request that the physical pages underlying the message buffer be
536 	 * included in a crash dump.  Since the message buffer is accessed
537 	 * through the direct map, they are not automatically included.
538 	 */
539 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
540 	last_pa = pa + round_page(msgbufsize);
541 	while (pa < last_pa) {
542 		dump_add_page(pa);
543 		pa += PAGE_SIZE;
544 	}
545 #endif
546 	/*
547 	 * Compute the number of pages of memory that will be available for
548 	 * use, taking into account the overhead of a page structure per page.
549 	 * In other words, solve
550 	 *	"available physical memory" - round_page(page_range *
551 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
552 	 * for page_range.
553 	 */
554 	low_avail = phys_avail[0];
555 	high_avail = phys_avail[1];
556 	for (i = 0; i < vm_phys_nsegs; i++) {
557 		if (vm_phys_segs[i].start < low_avail)
558 			low_avail = vm_phys_segs[i].start;
559 		if (vm_phys_segs[i].end > high_avail)
560 			high_avail = vm_phys_segs[i].end;
561 	}
562 	/* Skip the first chunk.  It is already accounted for. */
563 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
564 		if (phys_avail[i] < low_avail)
565 			low_avail = phys_avail[i];
566 		if (phys_avail[i + 1] > high_avail)
567 			high_avail = phys_avail[i + 1];
568 	}
569 	first_page = low_avail / PAGE_SIZE;
570 #ifdef VM_PHYSSEG_SPARSE
571 	size = 0;
572 	for (i = 0; i < vm_phys_nsegs; i++)
573 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
574 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
575 		size += phys_avail[i + 1] - phys_avail[i];
576 	page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
577 #elif defined(VM_PHYSSEG_DENSE)
578 	/*
579 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
580 	 * the overhead of a page structure per page only if vm_page_array is
581 	 * allocated from the last physical memory chunk.  Otherwise, we must
582 	 * allocate page structures representing the physical memory
583 	 * underlying vm_page_array, even though they will not be used.
584 	 */
585 	if (new_end == high_avail)
586 		page_range = (high_avail - low_avail) / (PAGE_SIZE +
587 		    sizeof(struct vm_page));
588 	else
589 		page_range = high_avail / PAGE_SIZE - first_page;
590 #else
591 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
592 #endif
593 	end = new_end;
594 
595 	/*
596 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
597 	 * However, because this page is allocated from KVM, out-of-bounds
598 	 * accesses using the direct map will not be trapped.
599 	 */
600 	vaddr += PAGE_SIZE;
601 
602 	/*
603 	 * Allocate physical memory for the page structures, and map it.
604 	 */
605 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
606 	mapped = pmap_map(&vaddr, new_end, end,
607 	    VM_PROT_READ | VM_PROT_WRITE);
608 	vm_page_array = (vm_page_t) mapped;
609 #if VM_NRESERVLEVEL > 0
610 	/*
611 	 * Allocate physical memory for the reservation management system's
612 	 * data structures, and map it.
613 	 */
614 	if (high_avail == end)
615 		high_avail = new_end;
616 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
617 #endif
618 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
619 	/*
620 	 * Include vm_page_array and vm_reserv_array in a crash dump.
621 	 */
622 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
623 		dump_add_page(pa);
624 #endif
625 	phys_avail[biggestone + 1] = new_end;
626 
627 	/*
628 	 * Add physical memory segments corresponding to the available
629 	 * physical pages.
630 	 */
631 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
632 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
633 
634 	/*
635 	 * Clear all of the page structures
636 	 */
637 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
638 	for (i = 0; i < page_range; i++)
639 		vm_page_array[i].order = VM_NFREEORDER;
640 	vm_page_array_size = page_range;
641 
642 	/*
643 	 * Initialize the physical memory allocator.
644 	 */
645 	vm_phys_init();
646 
647 	/*
648 	 * Add every available physical page that is not blacklisted to
649 	 * the free lists.
650 	 */
651 	vm_cnt.v_page_count = 0;
652 	vm_cnt.v_free_count = 0;
653 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
654 		pa = phys_avail[i];
655 		last_pa = phys_avail[i + 1];
656 		while (pa < last_pa) {
657 			vm_phys_add_page(pa);
658 			pa += PAGE_SIZE;
659 		}
660 	}
661 
662 	TAILQ_INIT(&blacklist_head);
663 	vm_page_blacklist_load(&list, &listend);
664 	vm_page_blacklist_check(list, listend);
665 
666 	list = kern_getenv("vm.blacklist");
667 	vm_page_blacklist_check(list, NULL);
668 
669 	freeenv(list);
670 #if VM_NRESERVLEVEL > 0
671 	/*
672 	 * Initialize the reservation management system.
673 	 */
674 	vm_reserv_init();
675 #endif
676 	return (vaddr);
677 }
678 
679 void
680 vm_page_reference(vm_page_t m)
681 {
682 
683 	vm_page_aflag_set(m, PGA_REFERENCED);
684 }
685 
686 /*
687  *	vm_page_busy_downgrade:
688  *
689  *	Downgrade an exclusive busy page into a single shared busy page.
690  */
691 void
692 vm_page_busy_downgrade(vm_page_t m)
693 {
694 	u_int x;
695 	bool locked;
696 
697 	vm_page_assert_xbusied(m);
698 	locked = mtx_owned(vm_page_lockptr(m));
699 
700 	for (;;) {
701 		x = m->busy_lock;
702 		x &= VPB_BIT_WAITERS;
703 		if (x != 0 && !locked)
704 			vm_page_lock(m);
705 		if (atomic_cmpset_rel_int(&m->busy_lock,
706 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
707 			break;
708 		if (x != 0 && !locked)
709 			vm_page_unlock(m);
710 	}
711 	if (x != 0) {
712 		wakeup(m);
713 		if (!locked)
714 			vm_page_unlock(m);
715 	}
716 }
717 
718 /*
719  *	vm_page_sbusied:
720  *
721  *	Return a positive value if the page is shared busied, 0 otherwise.
722  */
723 int
724 vm_page_sbusied(vm_page_t m)
725 {
726 	u_int x;
727 
728 	x = m->busy_lock;
729 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
730 }
731 
732 /*
733  *	vm_page_sunbusy:
734  *
735  *	Shared unbusy a page.
736  */
737 void
738 vm_page_sunbusy(vm_page_t m)
739 {
740 	u_int x;
741 
742 	vm_page_assert_sbusied(m);
743 
744 	for (;;) {
745 		x = m->busy_lock;
746 		if (VPB_SHARERS(x) > 1) {
747 			if (atomic_cmpset_int(&m->busy_lock, x,
748 			    x - VPB_ONE_SHARER))
749 				break;
750 			continue;
751 		}
752 		if ((x & VPB_BIT_WAITERS) == 0) {
753 			KASSERT(x == VPB_SHARERS_WORD(1),
754 			    ("vm_page_sunbusy: invalid lock state"));
755 			if (atomic_cmpset_int(&m->busy_lock,
756 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
757 				break;
758 			continue;
759 		}
760 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
761 		    ("vm_page_sunbusy: invalid lock state for waiters"));
762 
763 		vm_page_lock(m);
764 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
765 			vm_page_unlock(m);
766 			continue;
767 		}
768 		wakeup(m);
769 		vm_page_unlock(m);
770 		break;
771 	}
772 }
773 
774 /*
775  *	vm_page_busy_sleep:
776  *
777  *	Sleep and release the page lock, using the page pointer as wchan.
778  *	This is used to implement the hard-path of busying mechanism.
779  *
780  *	The given page must be locked.
781  *
782  *	If nonshared is true, sleep only if the page is xbusy.
783  */
784 void
785 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
786 {
787 	u_int x;
788 
789 	vm_page_assert_locked(m);
790 
791 	x = m->busy_lock;
792 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
793 	    ((x & VPB_BIT_WAITERS) == 0 &&
794 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
795 		vm_page_unlock(m);
796 		return;
797 	}
798 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
799 }
800 
801 /*
802  *	vm_page_trysbusy:
803  *
804  *	Try to shared busy a page.
805  *	If the operation succeeds 1 is returned otherwise 0.
806  *	The operation never sleeps.
807  */
808 int
809 vm_page_trysbusy(vm_page_t m)
810 {
811 	u_int x;
812 
813 	for (;;) {
814 		x = m->busy_lock;
815 		if ((x & VPB_BIT_SHARED) == 0)
816 			return (0);
817 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
818 			return (1);
819 	}
820 }
821 
822 static void
823 vm_page_xunbusy_locked(vm_page_t m)
824 {
825 
826 	vm_page_assert_xbusied(m);
827 	vm_page_assert_locked(m);
828 
829 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
830 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
831 	wakeup(m);
832 }
833 
834 void
835 vm_page_xunbusy_maybelocked(vm_page_t m)
836 {
837 	bool lockacq;
838 
839 	vm_page_assert_xbusied(m);
840 
841 	/*
842 	 * Fast path for unbusy.  If it succeeds, we know that there
843 	 * are no waiters, so we do not need a wakeup.
844 	 */
845 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
846 	    VPB_UNBUSIED))
847 		return;
848 
849 	lockacq = !mtx_owned(vm_page_lockptr(m));
850 	if (lockacq)
851 		vm_page_lock(m);
852 	vm_page_xunbusy_locked(m);
853 	if (lockacq)
854 		vm_page_unlock(m);
855 }
856 
857 /*
858  *	vm_page_xunbusy_hard:
859  *
860  *	Called after the first try the exclusive unbusy of a page failed.
861  *	It is assumed that the waiters bit is on.
862  */
863 void
864 vm_page_xunbusy_hard(vm_page_t m)
865 {
866 
867 	vm_page_assert_xbusied(m);
868 
869 	vm_page_lock(m);
870 	vm_page_xunbusy_locked(m);
871 	vm_page_unlock(m);
872 }
873 
874 /*
875  *	vm_page_flash:
876  *
877  *	Wakeup anyone waiting for the page.
878  *	The ownership bits do not change.
879  *
880  *	The given page must be locked.
881  */
882 void
883 vm_page_flash(vm_page_t m)
884 {
885 	u_int x;
886 
887 	vm_page_lock_assert(m, MA_OWNED);
888 
889 	for (;;) {
890 		x = m->busy_lock;
891 		if ((x & VPB_BIT_WAITERS) == 0)
892 			return;
893 		if (atomic_cmpset_int(&m->busy_lock, x,
894 		    x & (~VPB_BIT_WAITERS)))
895 			break;
896 	}
897 	wakeup(m);
898 }
899 
900 /*
901  * Keep page from being freed by the page daemon
902  * much of the same effect as wiring, except much lower
903  * overhead and should be used only for *very* temporary
904  * holding ("wiring").
905  */
906 void
907 vm_page_hold(vm_page_t mem)
908 {
909 
910 	vm_page_lock_assert(mem, MA_OWNED);
911         mem->hold_count++;
912 }
913 
914 void
915 vm_page_unhold(vm_page_t mem)
916 {
917 
918 	vm_page_lock_assert(mem, MA_OWNED);
919 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
920 	--mem->hold_count;
921 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
922 		vm_page_free_toq(mem);
923 }
924 
925 /*
926  *	vm_page_unhold_pages:
927  *
928  *	Unhold each of the pages that is referenced by the given array.
929  */
930 void
931 vm_page_unhold_pages(vm_page_t *ma, int count)
932 {
933 	struct mtx *mtx, *new_mtx;
934 
935 	mtx = NULL;
936 	for (; count != 0; count--) {
937 		/*
938 		 * Avoid releasing and reacquiring the same page lock.
939 		 */
940 		new_mtx = vm_page_lockptr(*ma);
941 		if (mtx != new_mtx) {
942 			if (mtx != NULL)
943 				mtx_unlock(mtx);
944 			mtx = new_mtx;
945 			mtx_lock(mtx);
946 		}
947 		vm_page_unhold(*ma);
948 		ma++;
949 	}
950 	if (mtx != NULL)
951 		mtx_unlock(mtx);
952 }
953 
954 vm_page_t
955 PHYS_TO_VM_PAGE(vm_paddr_t pa)
956 {
957 	vm_page_t m;
958 
959 #ifdef VM_PHYSSEG_SPARSE
960 	m = vm_phys_paddr_to_vm_page(pa);
961 	if (m == NULL)
962 		m = vm_phys_fictitious_to_vm_page(pa);
963 	return (m);
964 #elif defined(VM_PHYSSEG_DENSE)
965 	long pi;
966 
967 	pi = atop(pa);
968 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
969 		m = &vm_page_array[pi - first_page];
970 		return (m);
971 	}
972 	return (vm_phys_fictitious_to_vm_page(pa));
973 #else
974 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
975 #endif
976 }
977 
978 /*
979  *	vm_page_getfake:
980  *
981  *	Create a fictitious page with the specified physical address and
982  *	memory attribute.  The memory attribute is the only the machine-
983  *	dependent aspect of a fictitious page that must be initialized.
984  */
985 vm_page_t
986 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
987 {
988 	vm_page_t m;
989 
990 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
991 	vm_page_initfake(m, paddr, memattr);
992 	return (m);
993 }
994 
995 void
996 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
997 {
998 
999 	if ((m->flags & PG_FICTITIOUS) != 0) {
1000 		/*
1001 		 * The page's memattr might have changed since the
1002 		 * previous initialization.  Update the pmap to the
1003 		 * new memattr.
1004 		 */
1005 		goto memattr;
1006 	}
1007 	m->phys_addr = paddr;
1008 	m->queue = PQ_NONE;
1009 	/* Fictitious pages don't use "segind". */
1010 	m->flags = PG_FICTITIOUS;
1011 	/* Fictitious pages don't use "order" or "pool". */
1012 	m->oflags = VPO_UNMANAGED;
1013 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1014 	m->wire_count = 1;
1015 	pmap_page_init(m);
1016 memattr:
1017 	pmap_page_set_memattr(m, memattr);
1018 }
1019 
1020 /*
1021  *	vm_page_putfake:
1022  *
1023  *	Release a fictitious page.
1024  */
1025 void
1026 vm_page_putfake(vm_page_t m)
1027 {
1028 
1029 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1030 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1031 	    ("vm_page_putfake: bad page %p", m));
1032 	uma_zfree(fakepg_zone, m);
1033 }
1034 
1035 /*
1036  *	vm_page_updatefake:
1037  *
1038  *	Update the given fictitious page to the specified physical address and
1039  *	memory attribute.
1040  */
1041 void
1042 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1043 {
1044 
1045 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1046 	    ("vm_page_updatefake: bad page %p", m));
1047 	m->phys_addr = paddr;
1048 	pmap_page_set_memattr(m, memattr);
1049 }
1050 
1051 /*
1052  *	vm_page_free:
1053  *
1054  *	Free a page.
1055  */
1056 void
1057 vm_page_free(vm_page_t m)
1058 {
1059 
1060 	m->flags &= ~PG_ZERO;
1061 	vm_page_free_toq(m);
1062 }
1063 
1064 /*
1065  *	vm_page_free_zero:
1066  *
1067  *	Free a page to the zerod-pages queue
1068  */
1069 void
1070 vm_page_free_zero(vm_page_t m)
1071 {
1072 
1073 	m->flags |= PG_ZERO;
1074 	vm_page_free_toq(m);
1075 }
1076 
1077 /*
1078  * Unbusy and handle the page queueing for a page from a getpages request that
1079  * was optionally read ahead or behind.
1080  */
1081 void
1082 vm_page_readahead_finish(vm_page_t m)
1083 {
1084 
1085 	/* We shouldn't put invalid pages on queues. */
1086 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1087 
1088 	/*
1089 	 * Since the page is not the actually needed one, whether it should
1090 	 * be activated or deactivated is not obvious.  Empirical results
1091 	 * have shown that deactivating the page is usually the best choice,
1092 	 * unless the page is wanted by another thread.
1093 	 */
1094 	vm_page_lock(m);
1095 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1096 		vm_page_activate(m);
1097 	else
1098 		vm_page_deactivate(m);
1099 	vm_page_unlock(m);
1100 	vm_page_xunbusy(m);
1101 }
1102 
1103 /*
1104  *	vm_page_sleep_if_busy:
1105  *
1106  *	Sleep and release the page queues lock if the page is busied.
1107  *	Returns TRUE if the thread slept.
1108  *
1109  *	The given page must be unlocked and object containing it must
1110  *	be locked.
1111  */
1112 int
1113 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1114 {
1115 	vm_object_t obj;
1116 
1117 	vm_page_lock_assert(m, MA_NOTOWNED);
1118 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1119 
1120 	if (vm_page_busied(m)) {
1121 		/*
1122 		 * The page-specific object must be cached because page
1123 		 * identity can change during the sleep, causing the
1124 		 * re-lock of a different object.
1125 		 * It is assumed that a reference to the object is already
1126 		 * held by the callers.
1127 		 */
1128 		obj = m->object;
1129 		vm_page_lock(m);
1130 		VM_OBJECT_WUNLOCK(obj);
1131 		vm_page_busy_sleep(m, msg, false);
1132 		VM_OBJECT_WLOCK(obj);
1133 		return (TRUE);
1134 	}
1135 	return (FALSE);
1136 }
1137 
1138 /*
1139  *	vm_page_dirty_KBI:		[ internal use only ]
1140  *
1141  *	Set all bits in the page's dirty field.
1142  *
1143  *	The object containing the specified page must be locked if the
1144  *	call is made from the machine-independent layer.
1145  *
1146  *	See vm_page_clear_dirty_mask().
1147  *
1148  *	This function should only be called by vm_page_dirty().
1149  */
1150 void
1151 vm_page_dirty_KBI(vm_page_t m)
1152 {
1153 
1154 	/* Refer to this operation by its public name. */
1155 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1156 	    ("vm_page_dirty: page is invalid!"));
1157 	m->dirty = VM_PAGE_BITS_ALL;
1158 }
1159 
1160 /*
1161  *	vm_page_insert:		[ internal use only ]
1162  *
1163  *	Inserts the given mem entry into the object and object list.
1164  *
1165  *	The object must be locked.
1166  */
1167 int
1168 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1169 {
1170 	vm_page_t mpred;
1171 
1172 	VM_OBJECT_ASSERT_WLOCKED(object);
1173 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1174 	return (vm_page_insert_after(m, object, pindex, mpred));
1175 }
1176 
1177 /*
1178  *	vm_page_insert_after:
1179  *
1180  *	Inserts the page "m" into the specified object at offset "pindex".
1181  *
1182  *	The page "mpred" must immediately precede the offset "pindex" within
1183  *	the specified object.
1184  *
1185  *	The object must be locked.
1186  */
1187 static int
1188 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1189     vm_page_t mpred)
1190 {
1191 	vm_page_t msucc;
1192 
1193 	VM_OBJECT_ASSERT_WLOCKED(object);
1194 	KASSERT(m->object == NULL,
1195 	    ("vm_page_insert_after: page already inserted"));
1196 	if (mpred != NULL) {
1197 		KASSERT(mpred->object == object,
1198 		    ("vm_page_insert_after: object doesn't contain mpred"));
1199 		KASSERT(mpred->pindex < pindex,
1200 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1201 		msucc = TAILQ_NEXT(mpred, listq);
1202 	} else
1203 		msucc = TAILQ_FIRST(&object->memq);
1204 	if (msucc != NULL)
1205 		KASSERT(msucc->pindex > pindex,
1206 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1207 
1208 	/*
1209 	 * Record the object/offset pair in this page
1210 	 */
1211 	m->object = object;
1212 	m->pindex = pindex;
1213 
1214 	/*
1215 	 * Now link into the object's ordered list of backed pages.
1216 	 */
1217 	if (vm_radix_insert(&object->rtree, m)) {
1218 		m->object = NULL;
1219 		m->pindex = 0;
1220 		return (1);
1221 	}
1222 	vm_page_insert_radixdone(m, object, mpred);
1223 	return (0);
1224 }
1225 
1226 /*
1227  *	vm_page_insert_radixdone:
1228  *
1229  *	Complete page "m" insertion into the specified object after the
1230  *	radix trie hooking.
1231  *
1232  *	The page "mpred" must precede the offset "m->pindex" within the
1233  *	specified object.
1234  *
1235  *	The object must be locked.
1236  */
1237 static void
1238 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1239 {
1240 
1241 	VM_OBJECT_ASSERT_WLOCKED(object);
1242 	KASSERT(object != NULL && m->object == object,
1243 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1244 	if (mpred != NULL) {
1245 		KASSERT(mpred->object == object,
1246 		    ("vm_page_insert_after: object doesn't contain mpred"));
1247 		KASSERT(mpred->pindex < m->pindex,
1248 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1249 	}
1250 
1251 	if (mpred != NULL)
1252 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1253 	else
1254 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1255 
1256 	/*
1257 	 * Show that the object has one more resident page.
1258 	 */
1259 	object->resident_page_count++;
1260 
1261 	/*
1262 	 * Hold the vnode until the last page is released.
1263 	 */
1264 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1265 		vhold(object->handle);
1266 
1267 	/*
1268 	 * Since we are inserting a new and possibly dirty page,
1269 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1270 	 */
1271 	if (pmap_page_is_write_mapped(m))
1272 		vm_object_set_writeable_dirty(object);
1273 }
1274 
1275 /*
1276  *	vm_page_remove:
1277  *
1278  *	Removes the specified page from its containing object, but does not
1279  *	invalidate any backing storage.
1280  *
1281  *	The object must be locked.  The page must be locked if it is managed.
1282  */
1283 void
1284 vm_page_remove(vm_page_t m)
1285 {
1286 	vm_object_t object;
1287 	vm_page_t mrem;
1288 
1289 	if ((m->oflags & VPO_UNMANAGED) == 0)
1290 		vm_page_assert_locked(m);
1291 	if ((object = m->object) == NULL)
1292 		return;
1293 	VM_OBJECT_ASSERT_WLOCKED(object);
1294 	if (vm_page_xbusied(m))
1295 		vm_page_xunbusy_maybelocked(m);
1296 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1297 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1298 
1299 	/*
1300 	 * Now remove from the object's list of backed pages.
1301 	 */
1302 	TAILQ_REMOVE(&object->memq, m, listq);
1303 
1304 	/*
1305 	 * And show that the object has one fewer resident page.
1306 	 */
1307 	object->resident_page_count--;
1308 
1309 	/*
1310 	 * The vnode may now be recycled.
1311 	 */
1312 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1313 		vdrop(object->handle);
1314 
1315 	m->object = NULL;
1316 }
1317 
1318 /*
1319  *	vm_page_lookup:
1320  *
1321  *	Returns the page associated with the object/offset
1322  *	pair specified; if none is found, NULL is returned.
1323  *
1324  *	The object must be locked.
1325  */
1326 vm_page_t
1327 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1328 {
1329 
1330 	VM_OBJECT_ASSERT_LOCKED(object);
1331 	return (vm_radix_lookup(&object->rtree, pindex));
1332 }
1333 
1334 /*
1335  *	vm_page_find_least:
1336  *
1337  *	Returns the page associated with the object with least pindex
1338  *	greater than or equal to the parameter pindex, or NULL.
1339  *
1340  *	The object must be locked.
1341  */
1342 vm_page_t
1343 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1344 {
1345 	vm_page_t m;
1346 
1347 	VM_OBJECT_ASSERT_LOCKED(object);
1348 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1349 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1350 	return (m);
1351 }
1352 
1353 /*
1354  * Returns the given page's successor (by pindex) within the object if it is
1355  * resident; if none is found, NULL is returned.
1356  *
1357  * The object must be locked.
1358  */
1359 vm_page_t
1360 vm_page_next(vm_page_t m)
1361 {
1362 	vm_page_t next;
1363 
1364 	VM_OBJECT_ASSERT_LOCKED(m->object);
1365 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1366 		MPASS(next->object == m->object);
1367 		if (next->pindex != m->pindex + 1)
1368 			next = NULL;
1369 	}
1370 	return (next);
1371 }
1372 
1373 /*
1374  * Returns the given page's predecessor (by pindex) within the object if it is
1375  * resident; if none is found, NULL is returned.
1376  *
1377  * The object must be locked.
1378  */
1379 vm_page_t
1380 vm_page_prev(vm_page_t m)
1381 {
1382 	vm_page_t prev;
1383 
1384 	VM_OBJECT_ASSERT_LOCKED(m->object);
1385 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1386 		MPASS(prev->object == m->object);
1387 		if (prev->pindex != m->pindex - 1)
1388 			prev = NULL;
1389 	}
1390 	return (prev);
1391 }
1392 
1393 /*
1394  * Uses the page mnew as a replacement for an existing page at index
1395  * pindex which must be already present in the object.
1396  *
1397  * The existing page must not be on a paging queue.
1398  */
1399 vm_page_t
1400 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1401 {
1402 	vm_page_t mold;
1403 
1404 	VM_OBJECT_ASSERT_WLOCKED(object);
1405 	KASSERT(mnew->object == NULL,
1406 	    ("vm_page_replace: page already in object"));
1407 
1408 	/*
1409 	 * This function mostly follows vm_page_insert() and
1410 	 * vm_page_remove() without the radix, object count and vnode
1411 	 * dance.  Double check such functions for more comments.
1412 	 */
1413 
1414 	mnew->object = object;
1415 	mnew->pindex = pindex;
1416 	mold = vm_radix_replace(&object->rtree, mnew);
1417 	KASSERT(mold->queue == PQ_NONE,
1418 	    ("vm_page_replace: mold is on a paging queue"));
1419 
1420 	/* Keep the resident page list in sorted order. */
1421 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1422 	TAILQ_REMOVE(&object->memq, mold, listq);
1423 
1424 	mold->object = NULL;
1425 	vm_page_xunbusy_maybelocked(mold);
1426 
1427 	/*
1428 	 * The object's resident_page_count does not change because we have
1429 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1430 	 */
1431 	if (pmap_page_is_write_mapped(mnew))
1432 		vm_object_set_writeable_dirty(object);
1433 	return (mold);
1434 }
1435 
1436 /*
1437  *	vm_page_rename:
1438  *
1439  *	Move the given memory entry from its
1440  *	current object to the specified target object/offset.
1441  *
1442  *	Note: swap associated with the page must be invalidated by the move.  We
1443  *	      have to do this for several reasons:  (1) we aren't freeing the
1444  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1445  *	      moving the page from object A to B, and will then later move
1446  *	      the backing store from A to B and we can't have a conflict.
1447  *
1448  *	Note: we *always* dirty the page.  It is necessary both for the
1449  *	      fact that we moved it, and because we may be invalidating
1450  *	      swap.
1451  *
1452  *	The objects must be locked.
1453  */
1454 int
1455 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1456 {
1457 	vm_page_t mpred;
1458 	vm_pindex_t opidx;
1459 
1460 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1461 
1462 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1463 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1464 	    ("vm_page_rename: pindex already renamed"));
1465 
1466 	/*
1467 	 * Create a custom version of vm_page_insert() which does not depend
1468 	 * by m_prev and can cheat on the implementation aspects of the
1469 	 * function.
1470 	 */
1471 	opidx = m->pindex;
1472 	m->pindex = new_pindex;
1473 	if (vm_radix_insert(&new_object->rtree, m)) {
1474 		m->pindex = opidx;
1475 		return (1);
1476 	}
1477 
1478 	/*
1479 	 * The operation cannot fail anymore.  The removal must happen before
1480 	 * the listq iterator is tainted.
1481 	 */
1482 	m->pindex = opidx;
1483 	vm_page_lock(m);
1484 	vm_page_remove(m);
1485 
1486 	/* Return back to the new pindex to complete vm_page_insert(). */
1487 	m->pindex = new_pindex;
1488 	m->object = new_object;
1489 	vm_page_unlock(m);
1490 	vm_page_insert_radixdone(m, new_object, mpred);
1491 	vm_page_dirty(m);
1492 	return (0);
1493 }
1494 
1495 /*
1496  *	vm_page_alloc:
1497  *
1498  *	Allocate and return a page that is associated with the specified
1499  *	object and offset pair.  By default, this page is exclusive busied.
1500  *
1501  *	The caller must always specify an allocation class.
1502  *
1503  *	allocation classes:
1504  *	VM_ALLOC_NORMAL		normal process request
1505  *	VM_ALLOC_SYSTEM		system *really* needs a page
1506  *	VM_ALLOC_INTERRUPT	interrupt time request
1507  *
1508  *	optional allocation flags:
1509  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1510  *				intends to allocate
1511  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1512  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1513  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1514  *				should not be exclusive busy
1515  *	VM_ALLOC_SBUSY		shared busy the allocated page
1516  *	VM_ALLOC_WIRED		wire the allocated page
1517  *	VM_ALLOC_ZERO		prefer a zeroed page
1518  *
1519  *	This routine may not sleep.
1520  */
1521 vm_page_t
1522 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1523 {
1524 	vm_page_t m, mpred;
1525 	int flags, req_class;
1526 
1527 	mpred = NULL;	/* XXX: pacify gcc */
1528 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1529 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1530 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1531 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1532 	    ("vm_page_alloc: inconsistent object(%p)/req(%x)", object, req));
1533 	if (object != NULL)
1534 		VM_OBJECT_ASSERT_WLOCKED(object);
1535 
1536 	req_class = req & VM_ALLOC_CLASS_MASK;
1537 
1538 	/*
1539 	 * The page daemon is allowed to dig deeper into the free page list.
1540 	 */
1541 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1542 		req_class = VM_ALLOC_SYSTEM;
1543 
1544 	if (object != NULL) {
1545 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1546 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1547 		   ("vm_page_alloc: pindex already allocated"));
1548 	}
1549 
1550 	/*
1551 	 * Allocate a page if the number of free pages exceeds the minimum
1552 	 * for the request class.
1553 	 */
1554 	mtx_lock(&vm_page_queue_free_mtx);
1555 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1556 	    (req_class == VM_ALLOC_SYSTEM &&
1557 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1558 	    (req_class == VM_ALLOC_INTERRUPT &&
1559 	    vm_cnt.v_free_count > 0)) {
1560 		/*
1561 		 * Can we allocate the page from a reservation?
1562 		 */
1563 #if VM_NRESERVLEVEL > 0
1564 		if (object == NULL || (object->flags & (OBJ_COLORED |
1565 		    OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1566 		    vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1567 #endif
1568 		{
1569 			/*
1570 			 * If not, allocate it from the free page queues.
1571 			 */
1572 			m = vm_phys_alloc_pages(object != NULL ?
1573 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1574 #if VM_NRESERVLEVEL > 0
1575 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1576 				m = vm_phys_alloc_pages(object != NULL ?
1577 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1578 				    0);
1579 			}
1580 #endif
1581 		}
1582 	} else {
1583 		/*
1584 		 * Not allocatable, give up.
1585 		 */
1586 		mtx_unlock(&vm_page_queue_free_mtx);
1587 		atomic_add_int(&vm_pageout_deficit,
1588 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1589 		pagedaemon_wakeup();
1590 		return (NULL);
1591 	}
1592 
1593 	/*
1594 	 *  At this point we had better have found a good page.
1595 	 */
1596 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1597 	vm_phys_freecnt_adj(m, -1);
1598 	mtx_unlock(&vm_page_queue_free_mtx);
1599 	vm_page_alloc_check(m);
1600 
1601 	/*
1602 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1603 	 */
1604 	flags = 0;
1605 	if ((req & VM_ALLOC_ZERO) != 0)
1606 		flags = PG_ZERO;
1607 	flags &= m->flags;
1608 	if ((req & VM_ALLOC_NODUMP) != 0)
1609 		flags |= PG_NODUMP;
1610 	m->flags = flags;
1611 	m->aflags = 0;
1612 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1613 	    VPO_UNMANAGED : 0;
1614 	m->busy_lock = VPB_UNBUSIED;
1615 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1616 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1617 	if ((req & VM_ALLOC_SBUSY) != 0)
1618 		m->busy_lock = VPB_SHARERS_WORD(1);
1619 	if (req & VM_ALLOC_WIRED) {
1620 		/*
1621 		 * The page lock is not required for wiring a page until that
1622 		 * page is inserted into the object.
1623 		 */
1624 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1625 		m->wire_count = 1;
1626 	}
1627 	m->act_count = 0;
1628 
1629 	if (object != NULL) {
1630 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1631 			pagedaemon_wakeup();
1632 			if (req & VM_ALLOC_WIRED) {
1633 				atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1634 				m->wire_count = 0;
1635 			}
1636 			KASSERT(m->object == NULL, ("page %p has object", m));
1637 			m->oflags = VPO_UNMANAGED;
1638 			m->busy_lock = VPB_UNBUSIED;
1639 			/* Don't change PG_ZERO. */
1640 			vm_page_free_toq(m);
1641 			return (NULL);
1642 		}
1643 
1644 		/* Ignore device objects; the pager sets "memattr" for them. */
1645 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1646 		    (object->flags & OBJ_FICTITIOUS) == 0)
1647 			pmap_page_set_memattr(m, object->memattr);
1648 	} else
1649 		m->pindex = pindex;
1650 
1651 	/*
1652 	 * Don't wakeup too often - wakeup the pageout daemon when
1653 	 * we would be nearly out of memory.
1654 	 */
1655 	if (vm_paging_needed())
1656 		pagedaemon_wakeup();
1657 
1658 	return (m);
1659 }
1660 
1661 /*
1662  *	vm_page_alloc_contig:
1663  *
1664  *	Allocate a contiguous set of physical pages of the given size "npages"
1665  *	from the free lists.  All of the physical pages must be at or above
1666  *	the given physical address "low" and below the given physical address
1667  *	"high".  The given value "alignment" determines the alignment of the
1668  *	first physical page in the set.  If the given value "boundary" is
1669  *	non-zero, then the set of physical pages cannot cross any physical
1670  *	address boundary that is a multiple of that value.  Both "alignment"
1671  *	and "boundary" must be a power of two.
1672  *
1673  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1674  *	then the memory attribute setting for the physical pages is configured
1675  *	to the object's memory attribute setting.  Otherwise, the memory
1676  *	attribute setting for the physical pages is configured to "memattr",
1677  *	overriding the object's memory attribute setting.  However, if the
1678  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1679  *	memory attribute setting for the physical pages cannot be configured
1680  *	to VM_MEMATTR_DEFAULT.
1681  *
1682  *	The specified object may not contain fictitious pages.
1683  *
1684  *	The caller must always specify an allocation class.
1685  *
1686  *	allocation classes:
1687  *	VM_ALLOC_NORMAL		normal process request
1688  *	VM_ALLOC_SYSTEM		system *really* needs a page
1689  *	VM_ALLOC_INTERRUPT	interrupt time request
1690  *
1691  *	optional allocation flags:
1692  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1693  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1694  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1695  *				should not be exclusive busy
1696  *	VM_ALLOC_SBUSY		shared busy the allocated page
1697  *	VM_ALLOC_WIRED		wire the allocated page
1698  *	VM_ALLOC_ZERO		prefer a zeroed page
1699  *
1700  *	This routine may not sleep.
1701  */
1702 vm_page_t
1703 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1704     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1705     vm_paddr_t boundary, vm_memattr_t memattr)
1706 {
1707 	vm_page_t m, m_ret, mpred;
1708 	u_int busy_lock, flags, oflags;
1709 	int req_class;
1710 
1711 	mpred = NULL;	/* XXX: pacify gcc */
1712 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1713 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1714 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1715 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1716 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1717 	    req));
1718 	if (object != NULL) {
1719 		VM_OBJECT_ASSERT_WLOCKED(object);
1720 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1721 		    ("vm_page_alloc_contig: object %p has fictitious pages",
1722 		    object));
1723 	}
1724 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1725 	req_class = req & VM_ALLOC_CLASS_MASK;
1726 
1727 	/*
1728 	 * The page daemon is allowed to dig deeper into the free page list.
1729 	 */
1730 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1731 		req_class = VM_ALLOC_SYSTEM;
1732 
1733 	if (object != NULL) {
1734 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1735 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1736 		    ("vm_page_alloc_contig: pindex already allocated"));
1737 	}
1738 
1739 	/*
1740 	 * Can we allocate the pages without the number of free pages falling
1741 	 * below the lower bound for the allocation class?
1742 	 */
1743 	mtx_lock(&vm_page_queue_free_mtx);
1744 	if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1745 	    (req_class == VM_ALLOC_SYSTEM &&
1746 	    vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1747 	    (req_class == VM_ALLOC_INTERRUPT &&
1748 	    vm_cnt.v_free_count >= npages)) {
1749 		/*
1750 		 * Can we allocate the pages from a reservation?
1751 		 */
1752 #if VM_NRESERVLEVEL > 0
1753 retry:
1754 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1755 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1756 		    low, high, alignment, boundary, mpred)) == NULL)
1757 #endif
1758 			/*
1759 			 * If not, allocate them from the free page queues.
1760 			 */
1761 			m_ret = vm_phys_alloc_contig(npages, low, high,
1762 			    alignment, boundary);
1763 	} else {
1764 		mtx_unlock(&vm_page_queue_free_mtx);
1765 		atomic_add_int(&vm_pageout_deficit, npages);
1766 		pagedaemon_wakeup();
1767 		return (NULL);
1768 	}
1769 	if (m_ret != NULL)
1770 		vm_phys_freecnt_adj(m_ret, -npages);
1771 	else {
1772 #if VM_NRESERVLEVEL > 0
1773 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1774 		    boundary))
1775 			goto retry;
1776 #endif
1777 	}
1778 	mtx_unlock(&vm_page_queue_free_mtx);
1779 	if (m_ret == NULL)
1780 		return (NULL);
1781 	for (m = m_ret; m < &m_ret[npages]; m++)
1782 		vm_page_alloc_check(m);
1783 
1784 	/*
1785 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1786 	 */
1787 	flags = 0;
1788 	if ((req & VM_ALLOC_ZERO) != 0)
1789 		flags = PG_ZERO;
1790 	if ((req & VM_ALLOC_NODUMP) != 0)
1791 		flags |= PG_NODUMP;
1792 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1793 	    VPO_UNMANAGED : 0;
1794 	busy_lock = VPB_UNBUSIED;
1795 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1796 		busy_lock = VPB_SINGLE_EXCLUSIVER;
1797 	if ((req & VM_ALLOC_SBUSY) != 0)
1798 		busy_lock = VPB_SHARERS_WORD(1);
1799 	if ((req & VM_ALLOC_WIRED) != 0)
1800 		atomic_add_int(&vm_cnt.v_wire_count, npages);
1801 	if (object != NULL) {
1802 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1803 		    memattr == VM_MEMATTR_DEFAULT)
1804 			memattr = object->memattr;
1805 	}
1806 	for (m = m_ret; m < &m_ret[npages]; m++) {
1807 		m->aflags = 0;
1808 		m->flags = (m->flags | PG_NODUMP) & flags;
1809 		m->busy_lock = busy_lock;
1810 		if ((req & VM_ALLOC_WIRED) != 0)
1811 			m->wire_count = 1;
1812 		m->act_count = 0;
1813 		m->oflags = oflags;
1814 		if (object != NULL) {
1815 			if (vm_page_insert_after(m, object, pindex, mpred)) {
1816 				pagedaemon_wakeup();
1817 				if ((req & VM_ALLOC_WIRED) != 0)
1818 					atomic_subtract_int(
1819 					    &vm_cnt.v_wire_count, npages);
1820 				KASSERT(m->object == NULL,
1821 				    ("page %p has object", m));
1822 				mpred = m;
1823 				for (m = m_ret; m < &m_ret[npages]; m++) {
1824 					if (m <= mpred &&
1825 					    (req & VM_ALLOC_WIRED) != 0)
1826 						m->wire_count = 0;
1827 					m->oflags = VPO_UNMANAGED;
1828 					m->busy_lock = VPB_UNBUSIED;
1829 					/* Don't change PG_ZERO. */
1830 					vm_page_free_toq(m);
1831 				}
1832 				return (NULL);
1833 			}
1834 			mpred = m;
1835 		} else
1836 			m->pindex = pindex;
1837 		if (memattr != VM_MEMATTR_DEFAULT)
1838 			pmap_page_set_memattr(m, memattr);
1839 		pindex++;
1840 	}
1841 	if (vm_paging_needed())
1842 		pagedaemon_wakeup();
1843 	return (m_ret);
1844 }
1845 
1846 /*
1847  * Check a page that has been freshly dequeued from a freelist.
1848  */
1849 static void
1850 vm_page_alloc_check(vm_page_t m)
1851 {
1852 
1853 	KASSERT(m->object == NULL, ("page %p has object", m));
1854 	KASSERT(m->queue == PQ_NONE,
1855 	    ("page %p has unexpected queue %d", m, m->queue));
1856 	KASSERT(m->wire_count == 0, ("page %p is wired", m));
1857 	KASSERT(m->hold_count == 0, ("page %p is held", m));
1858 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1859 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
1860 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1861 	    ("page %p has unexpected memattr %d",
1862 	    m, pmap_page_get_memattr(m)));
1863 	KASSERT(m->valid == 0, ("free page %p is valid", m));
1864 }
1865 
1866 /*
1867  * 	vm_page_alloc_freelist:
1868  *
1869  *	Allocate a physical page from the specified free page list.
1870  *
1871  *	The caller must always specify an allocation class.
1872  *
1873  *	allocation classes:
1874  *	VM_ALLOC_NORMAL		normal process request
1875  *	VM_ALLOC_SYSTEM		system *really* needs a page
1876  *	VM_ALLOC_INTERRUPT	interrupt time request
1877  *
1878  *	optional allocation flags:
1879  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1880  *				intends to allocate
1881  *	VM_ALLOC_WIRED		wire the allocated page
1882  *	VM_ALLOC_ZERO		prefer a zeroed page
1883  *
1884  *	This routine may not sleep.
1885  */
1886 vm_page_t
1887 vm_page_alloc_freelist(int flind, int req)
1888 {
1889 	vm_page_t m;
1890 	u_int flags;
1891 	int req_class;
1892 
1893 	req_class = req & VM_ALLOC_CLASS_MASK;
1894 
1895 	/*
1896 	 * The page daemon is allowed to dig deeper into the free page list.
1897 	 */
1898 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1899 		req_class = VM_ALLOC_SYSTEM;
1900 
1901 	/*
1902 	 * Do not allocate reserved pages unless the req has asked for it.
1903 	 */
1904 	mtx_lock(&vm_page_queue_free_mtx);
1905 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1906 	    (req_class == VM_ALLOC_SYSTEM &&
1907 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1908 	    (req_class == VM_ALLOC_INTERRUPT &&
1909 	    vm_cnt.v_free_count > 0))
1910 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1911 	else {
1912 		mtx_unlock(&vm_page_queue_free_mtx);
1913 		atomic_add_int(&vm_pageout_deficit,
1914 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1915 		pagedaemon_wakeup();
1916 		return (NULL);
1917 	}
1918 	if (m == NULL) {
1919 		mtx_unlock(&vm_page_queue_free_mtx);
1920 		return (NULL);
1921 	}
1922 	vm_phys_freecnt_adj(m, -1);
1923 	mtx_unlock(&vm_page_queue_free_mtx);
1924 	vm_page_alloc_check(m);
1925 
1926 	/*
1927 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1928 	 */
1929 	m->aflags = 0;
1930 	flags = 0;
1931 	if ((req & VM_ALLOC_ZERO) != 0)
1932 		flags = PG_ZERO;
1933 	m->flags &= flags;
1934 	if ((req & VM_ALLOC_WIRED) != 0) {
1935 		/*
1936 		 * The page lock is not required for wiring a page that does
1937 		 * not belong to an object.
1938 		 */
1939 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1940 		m->wire_count = 1;
1941 	}
1942 	/* Unmanaged pages don't use "act_count". */
1943 	m->oflags = VPO_UNMANAGED;
1944 	if (vm_paging_needed())
1945 		pagedaemon_wakeup();
1946 	return (m);
1947 }
1948 
1949 #define	VPSC_ANY	0	/* No restrictions. */
1950 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
1951 #define	VPSC_NOSUPER	2	/* Skip superpages. */
1952 
1953 /*
1954  *	vm_page_scan_contig:
1955  *
1956  *	Scan vm_page_array[] between the specified entries "m_start" and
1957  *	"m_end" for a run of contiguous physical pages that satisfy the
1958  *	specified conditions, and return the lowest page in the run.  The
1959  *	specified "alignment" determines the alignment of the lowest physical
1960  *	page in the run.  If the specified "boundary" is non-zero, then the
1961  *	run of physical pages cannot span a physical address that is a
1962  *	multiple of "boundary".
1963  *
1964  *	"m_end" is never dereferenced, so it need not point to a vm_page
1965  *	structure within vm_page_array[].
1966  *
1967  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
1968  *	span a hole (or discontiguity) in the physical address space.  Both
1969  *	"alignment" and "boundary" must be a power of two.
1970  */
1971 vm_page_t
1972 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
1973     u_long alignment, vm_paddr_t boundary, int options)
1974 {
1975 	struct mtx *m_mtx, *new_mtx;
1976 	vm_object_t object;
1977 	vm_paddr_t pa;
1978 	vm_page_t m, m_run;
1979 #if VM_NRESERVLEVEL > 0
1980 	int level;
1981 #endif
1982 	int m_inc, order, run_ext, run_len;
1983 
1984 	KASSERT(npages > 0, ("npages is 0"));
1985 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1986 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1987 	m_run = NULL;
1988 	run_len = 0;
1989 	m_mtx = NULL;
1990 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
1991 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
1992 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
1993 
1994 		/*
1995 		 * If the current page would be the start of a run, check its
1996 		 * physical address against the end, alignment, and boundary
1997 		 * conditions.  If it doesn't satisfy these conditions, either
1998 		 * terminate the scan or advance to the next page that
1999 		 * satisfies the failed condition.
2000 		 */
2001 		if (run_len == 0) {
2002 			KASSERT(m_run == NULL, ("m_run != NULL"));
2003 			if (m + npages > m_end)
2004 				break;
2005 			pa = VM_PAGE_TO_PHYS(m);
2006 			if ((pa & (alignment - 1)) != 0) {
2007 				m_inc = atop(roundup2(pa, alignment) - pa);
2008 				continue;
2009 			}
2010 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2011 			    boundary) != 0) {
2012 				m_inc = atop(roundup2(pa, boundary) - pa);
2013 				continue;
2014 			}
2015 		} else
2016 			KASSERT(m_run != NULL, ("m_run == NULL"));
2017 
2018 		/*
2019 		 * Avoid releasing and reacquiring the same page lock.
2020 		 */
2021 		new_mtx = vm_page_lockptr(m);
2022 		if (m_mtx != new_mtx) {
2023 			if (m_mtx != NULL)
2024 				mtx_unlock(m_mtx);
2025 			m_mtx = new_mtx;
2026 			mtx_lock(m_mtx);
2027 		}
2028 		m_inc = 1;
2029 retry:
2030 		if (m->wire_count != 0 || m->hold_count != 0)
2031 			run_ext = 0;
2032 #if VM_NRESERVLEVEL > 0
2033 		else if ((level = vm_reserv_level(m)) >= 0 &&
2034 		    (options & VPSC_NORESERV) != 0) {
2035 			run_ext = 0;
2036 			/* Advance to the end of the reservation. */
2037 			pa = VM_PAGE_TO_PHYS(m);
2038 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2039 			    pa);
2040 		}
2041 #endif
2042 		else if ((object = m->object) != NULL) {
2043 			/*
2044 			 * The page is considered eligible for relocation if
2045 			 * and only if it could be laundered or reclaimed by
2046 			 * the page daemon.
2047 			 */
2048 			if (!VM_OBJECT_TRYRLOCK(object)) {
2049 				mtx_unlock(m_mtx);
2050 				VM_OBJECT_RLOCK(object);
2051 				mtx_lock(m_mtx);
2052 				if (m->object != object) {
2053 					/*
2054 					 * The page may have been freed.
2055 					 */
2056 					VM_OBJECT_RUNLOCK(object);
2057 					goto retry;
2058 				} else if (m->wire_count != 0 ||
2059 				    m->hold_count != 0) {
2060 					run_ext = 0;
2061 					goto unlock;
2062 				}
2063 			}
2064 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2065 			    ("page %p is PG_UNHOLDFREE", m));
2066 			/* Don't care: PG_NODUMP, PG_ZERO. */
2067 			if (object->type != OBJT_DEFAULT &&
2068 			    object->type != OBJT_SWAP &&
2069 			    object->type != OBJT_VNODE) {
2070 				run_ext = 0;
2071 #if VM_NRESERVLEVEL > 0
2072 			} else if ((options & VPSC_NOSUPER) != 0 &&
2073 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2074 				run_ext = 0;
2075 				/* Advance to the end of the superpage. */
2076 				pa = VM_PAGE_TO_PHYS(m);
2077 				m_inc = atop(roundup2(pa + 1,
2078 				    vm_reserv_size(level)) - pa);
2079 #endif
2080 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2081 			    m->queue != PQ_NONE && !vm_page_busied(m)) {
2082 				/*
2083 				 * The page is allocated but eligible for
2084 				 * relocation.  Extend the current run by one
2085 				 * page.
2086 				 */
2087 				KASSERT(pmap_page_get_memattr(m) ==
2088 				    VM_MEMATTR_DEFAULT,
2089 				    ("page %p has an unexpected memattr", m));
2090 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2091 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2092 				    ("page %p has unexpected oflags", m));
2093 				/* Don't care: VPO_NOSYNC. */
2094 				run_ext = 1;
2095 			} else
2096 				run_ext = 0;
2097 unlock:
2098 			VM_OBJECT_RUNLOCK(object);
2099 #if VM_NRESERVLEVEL > 0
2100 		} else if (level >= 0) {
2101 			/*
2102 			 * The page is reserved but not yet allocated.  In
2103 			 * other words, it is still free.  Extend the current
2104 			 * run by one page.
2105 			 */
2106 			run_ext = 1;
2107 #endif
2108 		} else if ((order = m->order) < VM_NFREEORDER) {
2109 			/*
2110 			 * The page is enqueued in the physical memory
2111 			 * allocator's free page queues.  Moreover, it is the
2112 			 * first page in a power-of-two-sized run of
2113 			 * contiguous free pages.  Add these pages to the end
2114 			 * of the current run, and jump ahead.
2115 			 */
2116 			run_ext = 1 << order;
2117 			m_inc = 1 << order;
2118 		} else {
2119 			/*
2120 			 * Skip the page for one of the following reasons: (1)
2121 			 * It is enqueued in the physical memory allocator's
2122 			 * free page queues.  However, it is not the first
2123 			 * page in a run of contiguous free pages.  (This case
2124 			 * rarely occurs because the scan is performed in
2125 			 * ascending order.) (2) It is not reserved, and it is
2126 			 * transitioning from free to allocated.  (Conversely,
2127 			 * the transition from allocated to free for managed
2128 			 * pages is blocked by the page lock.) (3) It is
2129 			 * allocated but not contained by an object and not
2130 			 * wired, e.g., allocated by Xen's balloon driver.
2131 			 */
2132 			run_ext = 0;
2133 		}
2134 
2135 		/*
2136 		 * Extend or reset the current run of pages.
2137 		 */
2138 		if (run_ext > 0) {
2139 			if (run_len == 0)
2140 				m_run = m;
2141 			run_len += run_ext;
2142 		} else {
2143 			if (run_len > 0) {
2144 				m_run = NULL;
2145 				run_len = 0;
2146 			}
2147 		}
2148 	}
2149 	if (m_mtx != NULL)
2150 		mtx_unlock(m_mtx);
2151 	if (run_len >= npages)
2152 		return (m_run);
2153 	return (NULL);
2154 }
2155 
2156 /*
2157  *	vm_page_reclaim_run:
2158  *
2159  *	Try to relocate each of the allocated virtual pages within the
2160  *	specified run of physical pages to a new physical address.  Free the
2161  *	physical pages underlying the relocated virtual pages.  A virtual page
2162  *	is relocatable if and only if it could be laundered or reclaimed by
2163  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2164  *	physical address above "high".
2165  *
2166  *	Returns 0 if every physical page within the run was already free or
2167  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2168  *	value indicating why the last attempt to relocate a virtual page was
2169  *	unsuccessful.
2170  *
2171  *	"req_class" must be an allocation class.
2172  */
2173 static int
2174 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2175     vm_paddr_t high)
2176 {
2177 	struct mtx *m_mtx, *new_mtx;
2178 	struct spglist free;
2179 	vm_object_t object;
2180 	vm_paddr_t pa;
2181 	vm_page_t m, m_end, m_new;
2182 	int error, order, req;
2183 
2184 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2185 	    ("req_class is not an allocation class"));
2186 	SLIST_INIT(&free);
2187 	error = 0;
2188 	m = m_run;
2189 	m_end = m_run + npages;
2190 	m_mtx = NULL;
2191 	for (; error == 0 && m < m_end; m++) {
2192 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2193 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2194 
2195 		/*
2196 		 * Avoid releasing and reacquiring the same page lock.
2197 		 */
2198 		new_mtx = vm_page_lockptr(m);
2199 		if (m_mtx != new_mtx) {
2200 			if (m_mtx != NULL)
2201 				mtx_unlock(m_mtx);
2202 			m_mtx = new_mtx;
2203 			mtx_lock(m_mtx);
2204 		}
2205 retry:
2206 		if (m->wire_count != 0 || m->hold_count != 0)
2207 			error = EBUSY;
2208 		else if ((object = m->object) != NULL) {
2209 			/*
2210 			 * The page is relocated if and only if it could be
2211 			 * laundered or reclaimed by the page daemon.
2212 			 */
2213 			if (!VM_OBJECT_TRYWLOCK(object)) {
2214 				mtx_unlock(m_mtx);
2215 				VM_OBJECT_WLOCK(object);
2216 				mtx_lock(m_mtx);
2217 				if (m->object != object) {
2218 					/*
2219 					 * The page may have been freed.
2220 					 */
2221 					VM_OBJECT_WUNLOCK(object);
2222 					goto retry;
2223 				} else if (m->wire_count != 0 ||
2224 				    m->hold_count != 0) {
2225 					error = EBUSY;
2226 					goto unlock;
2227 				}
2228 			}
2229 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2230 			    ("page %p is PG_UNHOLDFREE", m));
2231 			/* Don't care: PG_NODUMP, PG_ZERO. */
2232 			if (object->type != OBJT_DEFAULT &&
2233 			    object->type != OBJT_SWAP &&
2234 			    object->type != OBJT_VNODE)
2235 				error = EINVAL;
2236 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2237 				error = EINVAL;
2238 			else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2239 				KASSERT(pmap_page_get_memattr(m) ==
2240 				    VM_MEMATTR_DEFAULT,
2241 				    ("page %p has an unexpected memattr", m));
2242 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2243 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2244 				    ("page %p has unexpected oflags", m));
2245 				/* Don't care: VPO_NOSYNC. */
2246 				if (m->valid != 0) {
2247 					/*
2248 					 * First, try to allocate a new page
2249 					 * that is above "high".  Failing
2250 					 * that, try to allocate a new page
2251 					 * that is below "m_run".  Allocate
2252 					 * the new page between the end of
2253 					 * "m_run" and "high" only as a last
2254 					 * resort.
2255 					 */
2256 					req = req_class | VM_ALLOC_NOOBJ;
2257 					if ((m->flags & PG_NODUMP) != 0)
2258 						req |= VM_ALLOC_NODUMP;
2259 					if (trunc_page(high) !=
2260 					    ~(vm_paddr_t)PAGE_MASK) {
2261 						m_new = vm_page_alloc_contig(
2262 						    NULL, 0, req, 1,
2263 						    round_page(high),
2264 						    ~(vm_paddr_t)0,
2265 						    PAGE_SIZE, 0,
2266 						    VM_MEMATTR_DEFAULT);
2267 					} else
2268 						m_new = NULL;
2269 					if (m_new == NULL) {
2270 						pa = VM_PAGE_TO_PHYS(m_run);
2271 						m_new = vm_page_alloc_contig(
2272 						    NULL, 0, req, 1,
2273 						    0, pa - 1, PAGE_SIZE, 0,
2274 						    VM_MEMATTR_DEFAULT);
2275 					}
2276 					if (m_new == NULL) {
2277 						pa += ptoa(npages);
2278 						m_new = vm_page_alloc_contig(
2279 						    NULL, 0, req, 1,
2280 						    pa, high, PAGE_SIZE, 0,
2281 						    VM_MEMATTR_DEFAULT);
2282 					}
2283 					if (m_new == NULL) {
2284 						error = ENOMEM;
2285 						goto unlock;
2286 					}
2287 					KASSERT(m_new->wire_count == 0,
2288 					    ("page %p is wired", m));
2289 
2290 					/*
2291 					 * Replace "m" with the new page.  For
2292 					 * vm_page_replace(), "m" must be busy
2293 					 * and dequeued.  Finally, change "m"
2294 					 * as if vm_page_free() was called.
2295 					 */
2296 					if (object->ref_count != 0)
2297 						pmap_remove_all(m);
2298 					m_new->aflags = m->aflags;
2299 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2300 					    ("page %p is managed", m));
2301 					m_new->oflags = m->oflags & VPO_NOSYNC;
2302 					pmap_copy_page(m, m_new);
2303 					m_new->valid = m->valid;
2304 					m_new->dirty = m->dirty;
2305 					m->flags &= ~PG_ZERO;
2306 					vm_page_xbusy(m);
2307 					vm_page_remque(m);
2308 					vm_page_replace_checked(m_new, object,
2309 					    m->pindex, m);
2310 					m->valid = 0;
2311 					vm_page_undirty(m);
2312 
2313 					/*
2314 					 * The new page must be deactivated
2315 					 * before the object is unlocked.
2316 					 */
2317 					new_mtx = vm_page_lockptr(m_new);
2318 					if (m_mtx != new_mtx) {
2319 						mtx_unlock(m_mtx);
2320 						m_mtx = new_mtx;
2321 						mtx_lock(m_mtx);
2322 					}
2323 					vm_page_deactivate(m_new);
2324 				} else {
2325 					m->flags &= ~PG_ZERO;
2326 					vm_page_remque(m);
2327 					vm_page_remove(m);
2328 					KASSERT(m->dirty == 0,
2329 					    ("page %p is dirty", m));
2330 				}
2331 				SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2332 			} else
2333 				error = EBUSY;
2334 unlock:
2335 			VM_OBJECT_WUNLOCK(object);
2336 		} else {
2337 			mtx_lock(&vm_page_queue_free_mtx);
2338 			order = m->order;
2339 			if (order < VM_NFREEORDER) {
2340 				/*
2341 				 * The page is enqueued in the physical memory
2342 				 * allocator's free page queues.  Moreover, it
2343 				 * is the first page in a power-of-two-sized
2344 				 * run of contiguous free pages.  Jump ahead
2345 				 * to the last page within that run, and
2346 				 * continue from there.
2347 				 */
2348 				m += (1 << order) - 1;
2349 			}
2350 #if VM_NRESERVLEVEL > 0
2351 			else if (vm_reserv_is_page_free(m))
2352 				order = 0;
2353 #endif
2354 			mtx_unlock(&vm_page_queue_free_mtx);
2355 			if (order == VM_NFREEORDER)
2356 				error = EINVAL;
2357 		}
2358 	}
2359 	if (m_mtx != NULL)
2360 		mtx_unlock(m_mtx);
2361 	if ((m = SLIST_FIRST(&free)) != NULL) {
2362 		mtx_lock(&vm_page_queue_free_mtx);
2363 		do {
2364 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2365 			vm_phys_freecnt_adj(m, 1);
2366 #if VM_NRESERVLEVEL > 0
2367 			if (!vm_reserv_free_page(m))
2368 #else
2369 			if (true)
2370 #endif
2371 				vm_phys_free_pages(m, 0);
2372 		} while ((m = SLIST_FIRST(&free)) != NULL);
2373 		vm_page_free_wakeup();
2374 		mtx_unlock(&vm_page_queue_free_mtx);
2375 	}
2376 	return (error);
2377 }
2378 
2379 #define	NRUNS	16
2380 
2381 CTASSERT(powerof2(NRUNS));
2382 
2383 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2384 
2385 #define	MIN_RECLAIM	8
2386 
2387 /*
2388  *	vm_page_reclaim_contig:
2389  *
2390  *	Reclaim allocated, contiguous physical memory satisfying the specified
2391  *	conditions by relocating the virtual pages using that physical memory.
2392  *	Returns true if reclamation is successful and false otherwise.  Since
2393  *	relocation requires the allocation of physical pages, reclamation may
2394  *	fail due to a shortage of free pages.  When reclamation fails, callers
2395  *	are expected to perform VM_WAIT before retrying a failed allocation
2396  *	operation, e.g., vm_page_alloc_contig().
2397  *
2398  *	The caller must always specify an allocation class through "req".
2399  *
2400  *	allocation classes:
2401  *	VM_ALLOC_NORMAL		normal process request
2402  *	VM_ALLOC_SYSTEM		system *really* needs a page
2403  *	VM_ALLOC_INTERRUPT	interrupt time request
2404  *
2405  *	The optional allocation flags are ignored.
2406  *
2407  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2408  *	must be a power of two.
2409  */
2410 bool
2411 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2412     u_long alignment, vm_paddr_t boundary)
2413 {
2414 	vm_paddr_t curr_low;
2415 	vm_page_t m_run, m_runs[NRUNS];
2416 	u_long count, reclaimed;
2417 	int error, i, options, req_class;
2418 
2419 	KASSERT(npages > 0, ("npages is 0"));
2420 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2421 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2422 	req_class = req & VM_ALLOC_CLASS_MASK;
2423 
2424 	/*
2425 	 * The page daemon is allowed to dig deeper into the free page list.
2426 	 */
2427 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2428 		req_class = VM_ALLOC_SYSTEM;
2429 
2430 	/*
2431 	 * Return if the number of free pages cannot satisfy the requested
2432 	 * allocation.
2433 	 */
2434 	count = vm_cnt.v_free_count;
2435 	if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2436 	    vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2437 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2438 		return (false);
2439 
2440 	/*
2441 	 * Scan up to three times, relaxing the restrictions ("options") on
2442 	 * the reclamation of reservations and superpages each time.
2443 	 */
2444 	for (options = VPSC_NORESERV;;) {
2445 		/*
2446 		 * Find the highest runs that satisfy the given constraints
2447 		 * and restrictions, and record them in "m_runs".
2448 		 */
2449 		curr_low = low;
2450 		count = 0;
2451 		for (;;) {
2452 			m_run = vm_phys_scan_contig(npages, curr_low, high,
2453 			    alignment, boundary, options);
2454 			if (m_run == NULL)
2455 				break;
2456 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2457 			m_runs[RUN_INDEX(count)] = m_run;
2458 			count++;
2459 		}
2460 
2461 		/*
2462 		 * Reclaim the highest runs in LIFO (descending) order until
2463 		 * the number of reclaimed pages, "reclaimed", is at least
2464 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2465 		 * reclamation is idempotent, and runs will (likely) recur
2466 		 * from one scan to the next as restrictions are relaxed.
2467 		 */
2468 		reclaimed = 0;
2469 		for (i = 0; count > 0 && i < NRUNS; i++) {
2470 			count--;
2471 			m_run = m_runs[RUN_INDEX(count)];
2472 			error = vm_page_reclaim_run(req_class, npages, m_run,
2473 			    high);
2474 			if (error == 0) {
2475 				reclaimed += npages;
2476 				if (reclaimed >= MIN_RECLAIM)
2477 					return (true);
2478 			}
2479 		}
2480 
2481 		/*
2482 		 * Either relax the restrictions on the next scan or return if
2483 		 * the last scan had no restrictions.
2484 		 */
2485 		if (options == VPSC_NORESERV)
2486 			options = VPSC_NOSUPER;
2487 		else if (options == VPSC_NOSUPER)
2488 			options = VPSC_ANY;
2489 		else if (options == VPSC_ANY)
2490 			return (reclaimed != 0);
2491 	}
2492 }
2493 
2494 /*
2495  *	vm_wait:	(also see VM_WAIT macro)
2496  *
2497  *	Sleep until free pages are available for allocation.
2498  *	- Called in various places before memory allocations.
2499  */
2500 void
2501 vm_wait(void)
2502 {
2503 
2504 	mtx_lock(&vm_page_queue_free_mtx);
2505 	if (curproc == pageproc) {
2506 		vm_pageout_pages_needed = 1;
2507 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2508 		    PDROP | PSWP, "VMWait", 0);
2509 	} else {
2510 		if (__predict_false(pageproc == NULL))
2511 			panic("vm_wait in early boot");
2512 		if (!vm_pageout_wanted) {
2513 			vm_pageout_wanted = true;
2514 			wakeup(&vm_pageout_wanted);
2515 		}
2516 		vm_pages_needed = true;
2517 		msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2518 		    "vmwait", 0);
2519 	}
2520 }
2521 
2522 /*
2523  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
2524  *
2525  *	Sleep until free pages are available for allocation.
2526  *	- Called only in vm_fault so that processes page faulting
2527  *	  can be easily tracked.
2528  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2529  *	  processes will be able to grab memory first.  Do not change
2530  *	  this balance without careful testing first.
2531  */
2532 void
2533 vm_waitpfault(void)
2534 {
2535 
2536 	mtx_lock(&vm_page_queue_free_mtx);
2537 	if (!vm_pageout_wanted) {
2538 		vm_pageout_wanted = true;
2539 		wakeup(&vm_pageout_wanted);
2540 	}
2541 	vm_pages_needed = true;
2542 	msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2543 	    "pfault", 0);
2544 }
2545 
2546 struct vm_pagequeue *
2547 vm_page_pagequeue(vm_page_t m)
2548 {
2549 
2550 	if (vm_page_in_laundry(m))
2551 		return (&vm_dom[0].vmd_pagequeues[m->queue]);
2552 	else
2553 		return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2554 }
2555 
2556 /*
2557  *	vm_page_dequeue:
2558  *
2559  *	Remove the given page from its current page queue.
2560  *
2561  *	The page must be locked.
2562  */
2563 void
2564 vm_page_dequeue(vm_page_t m)
2565 {
2566 	struct vm_pagequeue *pq;
2567 
2568 	vm_page_assert_locked(m);
2569 	KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2570 	    m));
2571 	pq = vm_page_pagequeue(m);
2572 	vm_pagequeue_lock(pq);
2573 	m->queue = PQ_NONE;
2574 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2575 	vm_pagequeue_cnt_dec(pq);
2576 	vm_pagequeue_unlock(pq);
2577 }
2578 
2579 /*
2580  *	vm_page_dequeue_locked:
2581  *
2582  *	Remove the given page from its current page queue.
2583  *
2584  *	The page and page queue must be locked.
2585  */
2586 void
2587 vm_page_dequeue_locked(vm_page_t m)
2588 {
2589 	struct vm_pagequeue *pq;
2590 
2591 	vm_page_lock_assert(m, MA_OWNED);
2592 	pq = vm_page_pagequeue(m);
2593 	vm_pagequeue_assert_locked(pq);
2594 	m->queue = PQ_NONE;
2595 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2596 	vm_pagequeue_cnt_dec(pq);
2597 }
2598 
2599 /*
2600  *	vm_page_enqueue:
2601  *
2602  *	Add the given page to the specified page queue.
2603  *
2604  *	The page must be locked.
2605  */
2606 static void
2607 vm_page_enqueue(uint8_t queue, vm_page_t m)
2608 {
2609 	struct vm_pagequeue *pq;
2610 
2611 	vm_page_lock_assert(m, MA_OWNED);
2612 	KASSERT(queue < PQ_COUNT,
2613 	    ("vm_page_enqueue: invalid queue %u request for page %p",
2614 	    queue, m));
2615 	if (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE)
2616 		pq = &vm_dom[0].vmd_pagequeues[queue];
2617 	else
2618 		pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2619 	vm_pagequeue_lock(pq);
2620 	m->queue = queue;
2621 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2622 	vm_pagequeue_cnt_inc(pq);
2623 	vm_pagequeue_unlock(pq);
2624 }
2625 
2626 /*
2627  *	vm_page_requeue:
2628  *
2629  *	Move the given page to the tail of its current page queue.
2630  *
2631  *	The page must be locked.
2632  */
2633 void
2634 vm_page_requeue(vm_page_t m)
2635 {
2636 	struct vm_pagequeue *pq;
2637 
2638 	vm_page_lock_assert(m, MA_OWNED);
2639 	KASSERT(m->queue != PQ_NONE,
2640 	    ("vm_page_requeue: page %p is not queued", m));
2641 	pq = vm_page_pagequeue(m);
2642 	vm_pagequeue_lock(pq);
2643 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2644 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2645 	vm_pagequeue_unlock(pq);
2646 }
2647 
2648 /*
2649  *	vm_page_requeue_locked:
2650  *
2651  *	Move the given page to the tail of its current page queue.
2652  *
2653  *	The page queue must be locked.
2654  */
2655 void
2656 vm_page_requeue_locked(vm_page_t m)
2657 {
2658 	struct vm_pagequeue *pq;
2659 
2660 	KASSERT(m->queue != PQ_NONE,
2661 	    ("vm_page_requeue_locked: page %p is not queued", m));
2662 	pq = vm_page_pagequeue(m);
2663 	vm_pagequeue_assert_locked(pq);
2664 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2665 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2666 }
2667 
2668 /*
2669  *	vm_page_activate:
2670  *
2671  *	Put the specified page on the active list (if appropriate).
2672  *	Ensure that act_count is at least ACT_INIT but do not otherwise
2673  *	mess with it.
2674  *
2675  *	The page must be locked.
2676  */
2677 void
2678 vm_page_activate(vm_page_t m)
2679 {
2680 	int queue;
2681 
2682 	vm_page_lock_assert(m, MA_OWNED);
2683 	if ((queue = m->queue) != PQ_ACTIVE) {
2684 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2685 			if (m->act_count < ACT_INIT)
2686 				m->act_count = ACT_INIT;
2687 			if (queue != PQ_NONE)
2688 				vm_page_dequeue(m);
2689 			vm_page_enqueue(PQ_ACTIVE, m);
2690 		} else
2691 			KASSERT(queue == PQ_NONE,
2692 			    ("vm_page_activate: wired page %p is queued", m));
2693 	} else {
2694 		if (m->act_count < ACT_INIT)
2695 			m->act_count = ACT_INIT;
2696 	}
2697 }
2698 
2699 /*
2700  *	vm_page_free_wakeup:
2701  *
2702  *	Helper routine for vm_page_free_toq().  This routine is called
2703  *	when a page is added to the free queues.
2704  *
2705  *	The page queues must be locked.
2706  */
2707 static inline void
2708 vm_page_free_wakeup(void)
2709 {
2710 
2711 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2712 	/*
2713 	 * if pageout daemon needs pages, then tell it that there are
2714 	 * some free.
2715 	 */
2716 	if (vm_pageout_pages_needed &&
2717 	    vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2718 		wakeup(&vm_pageout_pages_needed);
2719 		vm_pageout_pages_needed = 0;
2720 	}
2721 	/*
2722 	 * wakeup processes that are waiting on memory if we hit a
2723 	 * high water mark. And wakeup scheduler process if we have
2724 	 * lots of memory. this process will swapin processes.
2725 	 */
2726 	if (vm_pages_needed && !vm_page_count_min()) {
2727 		vm_pages_needed = false;
2728 		wakeup(&vm_cnt.v_free_count);
2729 	}
2730 }
2731 
2732 /*
2733  *	vm_page_free_toq:
2734  *
2735  *	Returns the given page to the free list,
2736  *	disassociating it with any VM object.
2737  *
2738  *	The object must be locked.  The page must be locked if it is managed.
2739  */
2740 void
2741 vm_page_free_toq(vm_page_t m)
2742 {
2743 
2744 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2745 		vm_page_lock_assert(m, MA_OWNED);
2746 		KASSERT(!pmap_page_is_mapped(m),
2747 		    ("vm_page_free_toq: freeing mapped page %p", m));
2748 	} else
2749 		KASSERT(m->queue == PQ_NONE,
2750 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
2751 	VM_CNT_INC(v_tfree);
2752 
2753 	if (vm_page_sbusied(m))
2754 		panic("vm_page_free: freeing busy page %p", m);
2755 
2756 	/*
2757 	 * Unqueue, then remove page.  Note that we cannot destroy
2758 	 * the page here because we do not want to call the pager's
2759 	 * callback routine until after we've put the page on the
2760 	 * appropriate free queue.
2761 	 */
2762 	vm_page_remque(m);
2763 	vm_page_remove(m);
2764 
2765 	/*
2766 	 * If fictitious remove object association and
2767 	 * return, otherwise delay object association removal.
2768 	 */
2769 	if ((m->flags & PG_FICTITIOUS) != 0) {
2770 		return;
2771 	}
2772 
2773 	m->valid = 0;
2774 	vm_page_undirty(m);
2775 
2776 	if (m->wire_count != 0)
2777 		panic("vm_page_free: freeing wired page %p", m);
2778 	if (m->hold_count != 0) {
2779 		m->flags &= ~PG_ZERO;
2780 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2781 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2782 		m->flags |= PG_UNHOLDFREE;
2783 	} else {
2784 		/*
2785 		 * Restore the default memory attribute to the page.
2786 		 */
2787 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2788 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2789 
2790 		/*
2791 		 * Insert the page into the physical memory allocator's free
2792 		 * page queues.
2793 		 */
2794 		mtx_lock(&vm_page_queue_free_mtx);
2795 		vm_phys_freecnt_adj(m, 1);
2796 #if VM_NRESERVLEVEL > 0
2797 		if (!vm_reserv_free_page(m))
2798 #else
2799 		if (TRUE)
2800 #endif
2801 			vm_phys_free_pages(m, 0);
2802 		vm_page_free_wakeup();
2803 		mtx_unlock(&vm_page_queue_free_mtx);
2804 	}
2805 }
2806 
2807 /*
2808  *	vm_page_wire:
2809  *
2810  *	Mark this page as wired down by yet
2811  *	another map, removing it from paging queues
2812  *	as necessary.
2813  *
2814  *	If the page is fictitious, then its wire count must remain one.
2815  *
2816  *	The page must be locked.
2817  */
2818 void
2819 vm_page_wire(vm_page_t m)
2820 {
2821 
2822 	/*
2823 	 * Only bump the wire statistics if the page is not already wired,
2824 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2825 	 * it is already off the queues).
2826 	 */
2827 	vm_page_lock_assert(m, MA_OWNED);
2828 	if ((m->flags & PG_FICTITIOUS) != 0) {
2829 		KASSERT(m->wire_count == 1,
2830 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2831 		    m));
2832 		return;
2833 	}
2834 	if (m->wire_count == 0) {
2835 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2836 		    m->queue == PQ_NONE,
2837 		    ("vm_page_wire: unmanaged page %p is queued", m));
2838 		vm_page_remque(m);
2839 		atomic_add_int(&vm_cnt.v_wire_count, 1);
2840 	}
2841 	m->wire_count++;
2842 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2843 }
2844 
2845 /*
2846  * vm_page_unwire:
2847  *
2848  * Release one wiring of the specified page, potentially allowing it to be
2849  * paged out.  Returns TRUE if the number of wirings transitions to zero and
2850  * FALSE otherwise.
2851  *
2852  * Only managed pages belonging to an object can be paged out.  If the number
2853  * of wirings transitions to zero and the page is eligible for page out, then
2854  * the page is added to the specified paging queue (unless PQ_NONE is
2855  * specified).
2856  *
2857  * If a page is fictitious, then its wire count must always be one.
2858  *
2859  * A managed page must be locked.
2860  */
2861 boolean_t
2862 vm_page_unwire(vm_page_t m, uint8_t queue)
2863 {
2864 
2865 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
2866 	    ("vm_page_unwire: invalid queue %u request for page %p",
2867 	    queue, m));
2868 	if ((m->oflags & VPO_UNMANAGED) == 0)
2869 		vm_page_assert_locked(m);
2870 	if ((m->flags & PG_FICTITIOUS) != 0) {
2871 		KASSERT(m->wire_count == 1,
2872 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2873 		return (FALSE);
2874 	}
2875 	if (m->wire_count > 0) {
2876 		m->wire_count--;
2877 		if (m->wire_count == 0) {
2878 			atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2879 			if ((m->oflags & VPO_UNMANAGED) == 0 &&
2880 			    m->object != NULL && queue != PQ_NONE)
2881 				vm_page_enqueue(queue, m);
2882 			return (TRUE);
2883 		} else
2884 			return (FALSE);
2885 	} else
2886 		panic("vm_page_unwire: page %p's wire count is zero", m);
2887 }
2888 
2889 /*
2890  * Move the specified page to the inactive queue.
2891  *
2892  * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
2893  * queue.  However, setting "noreuse" to TRUE will accelerate the specified
2894  * page's reclamation, but it will not unmap the page from any address space.
2895  * This is implemented by inserting the page near the head of the inactive
2896  * queue, using a marker page to guide FIFO insertion ordering.
2897  *
2898  * The page must be locked.
2899  */
2900 static inline void
2901 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
2902 {
2903 	struct vm_pagequeue *pq;
2904 	int queue;
2905 
2906 	vm_page_assert_locked(m);
2907 
2908 	/*
2909 	 * Ignore if the page is already inactive, unless it is unlikely to be
2910 	 * reactivated.
2911 	 */
2912 	if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
2913 		return;
2914 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2915 		pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2916 		/* Avoid multiple acquisitions of the inactive queue lock. */
2917 		if (queue == PQ_INACTIVE) {
2918 			vm_pagequeue_lock(pq);
2919 			vm_page_dequeue_locked(m);
2920 		} else {
2921 			if (queue != PQ_NONE)
2922 				vm_page_dequeue(m);
2923 			vm_pagequeue_lock(pq);
2924 		}
2925 		m->queue = PQ_INACTIVE;
2926 		if (noreuse)
2927 			TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
2928 			    m, plinks.q);
2929 		else
2930 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2931 		vm_pagequeue_cnt_inc(pq);
2932 		vm_pagequeue_unlock(pq);
2933 	}
2934 }
2935 
2936 /*
2937  * Move the specified page to the inactive queue.
2938  *
2939  * The page must be locked.
2940  */
2941 void
2942 vm_page_deactivate(vm_page_t m)
2943 {
2944 
2945 	_vm_page_deactivate(m, FALSE);
2946 }
2947 
2948 /*
2949  * Move the specified page to the inactive queue with the expectation
2950  * that it is unlikely to be reused.
2951  *
2952  * The page must be locked.
2953  */
2954 void
2955 vm_page_deactivate_noreuse(vm_page_t m)
2956 {
2957 
2958 	_vm_page_deactivate(m, TRUE);
2959 }
2960 
2961 /*
2962  * vm_page_launder
2963  *
2964  * 	Put a page in the laundry.
2965  */
2966 void
2967 vm_page_launder(vm_page_t m)
2968 {
2969 	int queue;
2970 
2971 	vm_page_assert_locked(m);
2972 	if ((queue = m->queue) != PQ_LAUNDRY) {
2973 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2974 			if (queue != PQ_NONE)
2975 				vm_page_dequeue(m);
2976 			vm_page_enqueue(PQ_LAUNDRY, m);
2977 		} else
2978 			KASSERT(queue == PQ_NONE,
2979 			    ("wired page %p is queued", m));
2980 	}
2981 }
2982 
2983 /*
2984  * vm_page_unswappable
2985  *
2986  *	Put a page in the PQ_UNSWAPPABLE holding queue.
2987  */
2988 void
2989 vm_page_unswappable(vm_page_t m)
2990 {
2991 
2992 	vm_page_assert_locked(m);
2993 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
2994 	    ("page %p already unswappable", m));
2995 	if (m->queue != PQ_NONE)
2996 		vm_page_dequeue(m);
2997 	vm_page_enqueue(PQ_UNSWAPPABLE, m);
2998 }
2999 
3000 /*
3001  * vm_page_try_to_free()
3002  *
3003  *	Attempt to free the page.  If we cannot free it, we do nothing.
3004  *	1 is returned on success, 0 on failure.
3005  */
3006 int
3007 vm_page_try_to_free(vm_page_t m)
3008 {
3009 
3010 	vm_page_lock_assert(m, MA_OWNED);
3011 	if (m->object != NULL)
3012 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3013 	if (m->dirty || m->hold_count || m->wire_count ||
3014 	    (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3015 		return (0);
3016 	pmap_remove_all(m);
3017 	if (m->dirty)
3018 		return (0);
3019 	vm_page_free(m);
3020 	return (1);
3021 }
3022 
3023 /*
3024  * vm_page_advise
3025  *
3026  * 	Apply the specified advice to the given page.
3027  *
3028  *	The object and page must be locked.
3029  */
3030 void
3031 vm_page_advise(vm_page_t m, int advice)
3032 {
3033 
3034 	vm_page_assert_locked(m);
3035 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3036 	if (advice == MADV_FREE)
3037 		/*
3038 		 * Mark the page clean.  This will allow the page to be freed
3039 		 * without first paging it out.  MADV_FREE pages are often
3040 		 * quickly reused by malloc(3), so we do not do anything that
3041 		 * would result in a page fault on a later access.
3042 		 */
3043 		vm_page_undirty(m);
3044 	else if (advice != MADV_DONTNEED) {
3045 		if (advice == MADV_WILLNEED)
3046 			vm_page_activate(m);
3047 		return;
3048 	}
3049 
3050 	/*
3051 	 * Clear any references to the page.  Otherwise, the page daemon will
3052 	 * immediately reactivate the page.
3053 	 */
3054 	vm_page_aflag_clear(m, PGA_REFERENCED);
3055 
3056 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3057 		vm_page_dirty(m);
3058 
3059 	/*
3060 	 * Place clean pages near the head of the inactive queue rather than
3061 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3062 	 * the page will be reused quickly.  Dirty pages not already in the
3063 	 * laundry are moved there.
3064 	 */
3065 	if (m->dirty == 0)
3066 		vm_page_deactivate_noreuse(m);
3067 	else
3068 		vm_page_launder(m);
3069 }
3070 
3071 /*
3072  * Grab a page, waiting until we are waken up due to the page
3073  * changing state.  We keep on waiting, if the page continues
3074  * to be in the object.  If the page doesn't exist, first allocate it
3075  * and then conditionally zero it.
3076  *
3077  * This routine may sleep.
3078  *
3079  * The object must be locked on entry.  The lock will, however, be released
3080  * and reacquired if the routine sleeps.
3081  */
3082 vm_page_t
3083 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3084 {
3085 	vm_page_t m;
3086 	int sleep;
3087 
3088 	VM_OBJECT_ASSERT_WLOCKED(object);
3089 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3090 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3091 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3092 retrylookup:
3093 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3094 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3095 		    vm_page_xbusied(m) : vm_page_busied(m);
3096 		if (sleep) {
3097 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3098 				return (NULL);
3099 			/*
3100 			 * Reference the page before unlocking and
3101 			 * sleeping so that the page daemon is less
3102 			 * likely to reclaim it.
3103 			 */
3104 			vm_page_aflag_set(m, PGA_REFERENCED);
3105 			vm_page_lock(m);
3106 			VM_OBJECT_WUNLOCK(object);
3107 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3108 			    VM_ALLOC_IGN_SBUSY) != 0);
3109 			VM_OBJECT_WLOCK(object);
3110 			goto retrylookup;
3111 		} else {
3112 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3113 				vm_page_lock(m);
3114 				vm_page_wire(m);
3115 				vm_page_unlock(m);
3116 			}
3117 			if ((allocflags &
3118 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3119 				vm_page_xbusy(m);
3120 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3121 				vm_page_sbusy(m);
3122 			return (m);
3123 		}
3124 	}
3125 	m = vm_page_alloc(object, pindex, allocflags);
3126 	if (m == NULL) {
3127 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3128 			return (NULL);
3129 		VM_OBJECT_WUNLOCK(object);
3130 		VM_WAIT;
3131 		VM_OBJECT_WLOCK(object);
3132 		goto retrylookup;
3133 	}
3134 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3135 		pmap_zero_page(m);
3136 	return (m);
3137 }
3138 
3139 /*
3140  * Mapping function for valid or dirty bits in a page.
3141  *
3142  * Inputs are required to range within a page.
3143  */
3144 vm_page_bits_t
3145 vm_page_bits(int base, int size)
3146 {
3147 	int first_bit;
3148 	int last_bit;
3149 
3150 	KASSERT(
3151 	    base + size <= PAGE_SIZE,
3152 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
3153 	);
3154 
3155 	if (size == 0)		/* handle degenerate case */
3156 		return (0);
3157 
3158 	first_bit = base >> DEV_BSHIFT;
3159 	last_bit = (base + size - 1) >> DEV_BSHIFT;
3160 
3161 	return (((vm_page_bits_t)2 << last_bit) -
3162 	    ((vm_page_bits_t)1 << first_bit));
3163 }
3164 
3165 /*
3166  *	vm_page_set_valid_range:
3167  *
3168  *	Sets portions of a page valid.  The arguments are expected
3169  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3170  *	of any partial chunks touched by the range.  The invalid portion of
3171  *	such chunks will be zeroed.
3172  *
3173  *	(base + size) must be less then or equal to PAGE_SIZE.
3174  */
3175 void
3176 vm_page_set_valid_range(vm_page_t m, int base, int size)
3177 {
3178 	int endoff, frag;
3179 
3180 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3181 	if (size == 0)	/* handle degenerate case */
3182 		return;
3183 
3184 	/*
3185 	 * If the base is not DEV_BSIZE aligned and the valid
3186 	 * bit is clear, we have to zero out a portion of the
3187 	 * first block.
3188 	 */
3189 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3190 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3191 		pmap_zero_page_area(m, frag, base - frag);
3192 
3193 	/*
3194 	 * If the ending offset is not DEV_BSIZE aligned and the
3195 	 * valid bit is clear, we have to zero out a portion of
3196 	 * the last block.
3197 	 */
3198 	endoff = base + size;
3199 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3200 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3201 		pmap_zero_page_area(m, endoff,
3202 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3203 
3204 	/*
3205 	 * Assert that no previously invalid block that is now being validated
3206 	 * is already dirty.
3207 	 */
3208 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3209 	    ("vm_page_set_valid_range: page %p is dirty", m));
3210 
3211 	/*
3212 	 * Set valid bits inclusive of any overlap.
3213 	 */
3214 	m->valid |= vm_page_bits(base, size);
3215 }
3216 
3217 /*
3218  * Clear the given bits from the specified page's dirty field.
3219  */
3220 static __inline void
3221 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3222 {
3223 	uintptr_t addr;
3224 #if PAGE_SIZE < 16384
3225 	int shift;
3226 #endif
3227 
3228 	/*
3229 	 * If the object is locked and the page is neither exclusive busy nor
3230 	 * write mapped, then the page's dirty field cannot possibly be
3231 	 * set by a concurrent pmap operation.
3232 	 */
3233 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3234 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3235 		m->dirty &= ~pagebits;
3236 	else {
3237 		/*
3238 		 * The pmap layer can call vm_page_dirty() without
3239 		 * holding a distinguished lock.  The combination of
3240 		 * the object's lock and an atomic operation suffice
3241 		 * to guarantee consistency of the page dirty field.
3242 		 *
3243 		 * For PAGE_SIZE == 32768 case, compiler already
3244 		 * properly aligns the dirty field, so no forcible
3245 		 * alignment is needed. Only require existence of
3246 		 * atomic_clear_64 when page size is 32768.
3247 		 */
3248 		addr = (uintptr_t)&m->dirty;
3249 #if PAGE_SIZE == 32768
3250 		atomic_clear_64((uint64_t *)addr, pagebits);
3251 #elif PAGE_SIZE == 16384
3252 		atomic_clear_32((uint32_t *)addr, pagebits);
3253 #else		/* PAGE_SIZE <= 8192 */
3254 		/*
3255 		 * Use a trick to perform a 32-bit atomic on the
3256 		 * containing aligned word, to not depend on the existence
3257 		 * of atomic_clear_{8, 16}.
3258 		 */
3259 		shift = addr & (sizeof(uint32_t) - 1);
3260 #if BYTE_ORDER == BIG_ENDIAN
3261 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3262 #else
3263 		shift *= NBBY;
3264 #endif
3265 		addr &= ~(sizeof(uint32_t) - 1);
3266 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
3267 #endif		/* PAGE_SIZE */
3268 	}
3269 }
3270 
3271 /*
3272  *	vm_page_set_validclean:
3273  *
3274  *	Sets portions of a page valid and clean.  The arguments are expected
3275  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3276  *	of any partial chunks touched by the range.  The invalid portion of
3277  *	such chunks will be zero'd.
3278  *
3279  *	(base + size) must be less then or equal to PAGE_SIZE.
3280  */
3281 void
3282 vm_page_set_validclean(vm_page_t m, int base, int size)
3283 {
3284 	vm_page_bits_t oldvalid, pagebits;
3285 	int endoff, frag;
3286 
3287 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3288 	if (size == 0)	/* handle degenerate case */
3289 		return;
3290 
3291 	/*
3292 	 * If the base is not DEV_BSIZE aligned and the valid
3293 	 * bit is clear, we have to zero out a portion of the
3294 	 * first block.
3295 	 */
3296 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3297 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3298 		pmap_zero_page_area(m, frag, base - frag);
3299 
3300 	/*
3301 	 * If the ending offset is not DEV_BSIZE aligned and the
3302 	 * valid bit is clear, we have to zero out a portion of
3303 	 * the last block.
3304 	 */
3305 	endoff = base + size;
3306 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3307 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3308 		pmap_zero_page_area(m, endoff,
3309 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3310 
3311 	/*
3312 	 * Set valid, clear dirty bits.  If validating the entire
3313 	 * page we can safely clear the pmap modify bit.  We also
3314 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
3315 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
3316 	 * be set again.
3317 	 *
3318 	 * We set valid bits inclusive of any overlap, but we can only
3319 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
3320 	 * the range.
3321 	 */
3322 	oldvalid = m->valid;
3323 	pagebits = vm_page_bits(base, size);
3324 	m->valid |= pagebits;
3325 #if 0	/* NOT YET */
3326 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3327 		frag = DEV_BSIZE - frag;
3328 		base += frag;
3329 		size -= frag;
3330 		if (size < 0)
3331 			size = 0;
3332 	}
3333 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3334 #endif
3335 	if (base == 0 && size == PAGE_SIZE) {
3336 		/*
3337 		 * The page can only be modified within the pmap if it is
3338 		 * mapped, and it can only be mapped if it was previously
3339 		 * fully valid.
3340 		 */
3341 		if (oldvalid == VM_PAGE_BITS_ALL)
3342 			/*
3343 			 * Perform the pmap_clear_modify() first.  Otherwise,
3344 			 * a concurrent pmap operation, such as
3345 			 * pmap_protect(), could clear a modification in the
3346 			 * pmap and set the dirty field on the page before
3347 			 * pmap_clear_modify() had begun and after the dirty
3348 			 * field was cleared here.
3349 			 */
3350 			pmap_clear_modify(m);
3351 		m->dirty = 0;
3352 		m->oflags &= ~VPO_NOSYNC;
3353 	} else if (oldvalid != VM_PAGE_BITS_ALL)
3354 		m->dirty &= ~pagebits;
3355 	else
3356 		vm_page_clear_dirty_mask(m, pagebits);
3357 }
3358 
3359 void
3360 vm_page_clear_dirty(vm_page_t m, int base, int size)
3361 {
3362 
3363 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3364 }
3365 
3366 /*
3367  *	vm_page_set_invalid:
3368  *
3369  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
3370  *	valid and dirty bits for the effected areas are cleared.
3371  */
3372 void
3373 vm_page_set_invalid(vm_page_t m, int base, int size)
3374 {
3375 	vm_page_bits_t bits;
3376 	vm_object_t object;
3377 
3378 	object = m->object;
3379 	VM_OBJECT_ASSERT_WLOCKED(object);
3380 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3381 	    size >= object->un_pager.vnp.vnp_size)
3382 		bits = VM_PAGE_BITS_ALL;
3383 	else
3384 		bits = vm_page_bits(base, size);
3385 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3386 	    bits != 0)
3387 		pmap_remove_all(m);
3388 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3389 	    !pmap_page_is_mapped(m),
3390 	    ("vm_page_set_invalid: page %p is mapped", m));
3391 	m->valid &= ~bits;
3392 	m->dirty &= ~bits;
3393 }
3394 
3395 /*
3396  * vm_page_zero_invalid()
3397  *
3398  *	The kernel assumes that the invalid portions of a page contain
3399  *	garbage, but such pages can be mapped into memory by user code.
3400  *	When this occurs, we must zero out the non-valid portions of the
3401  *	page so user code sees what it expects.
3402  *
3403  *	Pages are most often semi-valid when the end of a file is mapped
3404  *	into memory and the file's size is not page aligned.
3405  */
3406 void
3407 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3408 {
3409 	int b;
3410 	int i;
3411 
3412 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3413 	/*
3414 	 * Scan the valid bits looking for invalid sections that
3415 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
3416 	 * valid bit may be set ) have already been zeroed by
3417 	 * vm_page_set_validclean().
3418 	 */
3419 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3420 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
3421 		    (m->valid & ((vm_page_bits_t)1 << i))) {
3422 			if (i > b) {
3423 				pmap_zero_page_area(m,
3424 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3425 			}
3426 			b = i + 1;
3427 		}
3428 	}
3429 
3430 	/*
3431 	 * setvalid is TRUE when we can safely set the zero'd areas
3432 	 * as being valid.  We can do this if there are no cache consistancy
3433 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
3434 	 */
3435 	if (setvalid)
3436 		m->valid = VM_PAGE_BITS_ALL;
3437 }
3438 
3439 /*
3440  *	vm_page_is_valid:
3441  *
3442  *	Is (partial) page valid?  Note that the case where size == 0
3443  *	will return FALSE in the degenerate case where the page is
3444  *	entirely invalid, and TRUE otherwise.
3445  */
3446 int
3447 vm_page_is_valid(vm_page_t m, int base, int size)
3448 {
3449 	vm_page_bits_t bits;
3450 
3451 	VM_OBJECT_ASSERT_LOCKED(m->object);
3452 	bits = vm_page_bits(base, size);
3453 	return (m->valid != 0 && (m->valid & bits) == bits);
3454 }
3455 
3456 /*
3457  *	vm_page_ps_is_valid:
3458  *
3459  *	Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3460  */
3461 boolean_t
3462 vm_page_ps_is_valid(vm_page_t m)
3463 {
3464 	int i, npages;
3465 
3466 	VM_OBJECT_ASSERT_LOCKED(m->object);
3467 	npages = atop(pagesizes[m->psind]);
3468 
3469 	/*
3470 	 * The physically contiguous pages that make up a superpage, i.e., a
3471 	 * page with a page size index ("psind") greater than zero, will
3472 	 * occupy adjacent entries in vm_page_array[].
3473 	 */
3474 	for (i = 0; i < npages; i++) {
3475 		if (m[i].valid != VM_PAGE_BITS_ALL)
3476 			return (FALSE);
3477 	}
3478 	return (TRUE);
3479 }
3480 
3481 /*
3482  * Set the page's dirty bits if the page is modified.
3483  */
3484 void
3485 vm_page_test_dirty(vm_page_t m)
3486 {
3487 
3488 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3489 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3490 		vm_page_dirty(m);
3491 }
3492 
3493 void
3494 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3495 {
3496 
3497 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3498 }
3499 
3500 void
3501 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3502 {
3503 
3504 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3505 }
3506 
3507 int
3508 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3509 {
3510 
3511 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3512 }
3513 
3514 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3515 void
3516 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3517 {
3518 
3519 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3520 }
3521 
3522 void
3523 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3524 {
3525 
3526 	mtx_assert_(vm_page_lockptr(m), a, file, line);
3527 }
3528 #endif
3529 
3530 #ifdef INVARIANTS
3531 void
3532 vm_page_object_lock_assert(vm_page_t m)
3533 {
3534 
3535 	/*
3536 	 * Certain of the page's fields may only be modified by the
3537 	 * holder of the containing object's lock or the exclusive busy.
3538 	 * holder.  Unfortunately, the holder of the write busy is
3539 	 * not recorded, and thus cannot be checked here.
3540 	 */
3541 	if (m->object != NULL && !vm_page_xbusied(m))
3542 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3543 }
3544 
3545 void
3546 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3547 {
3548 
3549 	if ((bits & PGA_WRITEABLE) == 0)
3550 		return;
3551 
3552 	/*
3553 	 * The PGA_WRITEABLE flag can only be set if the page is
3554 	 * managed, is exclusively busied or the object is locked.
3555 	 * Currently, this flag is only set by pmap_enter().
3556 	 */
3557 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3558 	    ("PGA_WRITEABLE on unmanaged page"));
3559 	if (!vm_page_xbusied(m))
3560 		VM_OBJECT_ASSERT_LOCKED(m->object);
3561 }
3562 #endif
3563 
3564 #include "opt_ddb.h"
3565 #ifdef DDB
3566 #include <sys/kernel.h>
3567 
3568 #include <ddb/ddb.h>
3569 
3570 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3571 {
3572 
3573 	db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3574 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3575 	db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3576 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3577 	db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3578 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3579 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3580 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3581 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3582 }
3583 
3584 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3585 {
3586 	int dom;
3587 
3588 	db_printf("pq_free %d\n", vm_cnt.v_free_count);
3589 	for (dom = 0; dom < vm_ndomains; dom++) {
3590 		db_printf(
3591     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
3592 		    dom,
3593 		    vm_dom[dom].vmd_page_count,
3594 		    vm_dom[dom].vmd_free_count,
3595 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3596 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3597 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
3598 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
3599 	}
3600 }
3601 
3602 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3603 {
3604 	vm_page_t m;
3605 	boolean_t phys;
3606 
3607 	if (!have_addr) {
3608 		db_printf("show pginfo addr\n");
3609 		return;
3610 	}
3611 
3612 	phys = strchr(modif, 'p') != NULL;
3613 	if (phys)
3614 		m = PHYS_TO_VM_PAGE(addr);
3615 	else
3616 		m = (vm_page_t)addr;
3617 	db_printf(
3618     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3619     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3620 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3621 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3622 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3623 }
3624 #endif /* DDB */
3625