xref: /freebsd/sys/vm/vm_page.c (revision 8a16b7a18f5d0b031f09832fd7752fba717e2a97)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *			GENERAL RULES ON VM_PAGE MANIPULATION
67  *
68  *	- A page queue lock is required when adding or removing a page from a
69  *	  page queue regardless of other locks or the busy state of a page.
70  *
71  *		* In general, no thread besides the page daemon can acquire or
72  *		  hold more than one page queue lock at a time.
73  *
74  *		* The page daemon can acquire and hold any pair of page queue
75  *		  locks in any order.
76  *
77  *	- The object lock is required when inserting or removing
78  *	  pages from an object (vm_page_insert() or vm_page_remove()).
79  *
80  */
81 
82 /*
83  *	Resident memory management module.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 #include "opt_vm.h"
90 
91 #include <sys/param.h>
92 #include <sys/systm.h>
93 #include <sys/lock.h>
94 #include <sys/kernel.h>
95 #include <sys/limits.h>
96 #include <sys/linker.h>
97 #include <sys/malloc.h>
98 #include <sys/mman.h>
99 #include <sys/msgbuf.h>
100 #include <sys/mutex.h>
101 #include <sys/proc.h>
102 #include <sys/rwlock.h>
103 #include <sys/sbuf.h>
104 #include <sys/smp.h>
105 #include <sys/sysctl.h>
106 #include <sys/vmmeter.h>
107 #include <sys/vnode.h>
108 
109 #include <vm/vm.h>
110 #include <vm/pmap.h>
111 #include <vm/vm_param.h>
112 #include <vm/vm_kern.h>
113 #include <vm/vm_object.h>
114 #include <vm/vm_page.h>
115 #include <vm/vm_pageout.h>
116 #include <vm/vm_pager.h>
117 #include <vm/vm_phys.h>
118 #include <vm/vm_radix.h>
119 #include <vm/vm_reserv.h>
120 #include <vm/vm_extern.h>
121 #include <vm/uma.h>
122 #include <vm/uma_int.h>
123 
124 #include <machine/md_var.h>
125 
126 /*
127  *	Associated with page of user-allocatable memory is a
128  *	page structure.
129  */
130 
131 struct vm_domain vm_dom[MAXMEMDOM];
132 struct mtx_padalign __exclusive_cache_line vm_page_queue_free_mtx;
133 
134 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
135 
136 /*
137  * bogus page -- for I/O to/from partially complete buffers,
138  * or for paging into sparsely invalid regions.
139  */
140 vm_page_t bogus_page;
141 
142 vm_page_t vm_page_array;
143 long vm_page_array_size;
144 long first_page;
145 
146 static int boot_pages = UMA_BOOT_PAGES;
147 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
148     &boot_pages, 0,
149     "number of pages allocated for bootstrapping the VM system");
150 
151 static int pa_tryrelock_restart;
152 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
153     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
154 
155 static TAILQ_HEAD(, vm_page) blacklist_head;
156 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
157 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
158     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
159 
160 /* Is the page daemon waiting for free pages? */
161 static int vm_pageout_pages_needed;
162 
163 static uma_zone_t fakepg_zone;
164 
165 static void vm_page_alloc_check(vm_page_t m);
166 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
167 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
168 static void vm_page_free_phys(vm_page_t m);
169 static void vm_page_free_wakeup(void);
170 static void vm_page_init(void *dummy);
171 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
172     vm_pindex_t pindex, vm_page_t mpred);
173 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
174     vm_page_t mpred);
175 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
176     vm_paddr_t high);
177 static int vm_page_alloc_fail(vm_object_t object, int req);
178 
179 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
180 
181 static void
182 vm_page_init(void *dummy)
183 {
184 
185 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
186 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
187 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
188 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
189 }
190 
191 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
192 #if PAGE_SIZE == 32768
193 #ifdef CTASSERT
194 CTASSERT(sizeof(u_long) >= 8);
195 #endif
196 #endif
197 
198 /*
199  * Try to acquire a physical address lock while a pmap is locked.  If we
200  * fail to trylock we unlock and lock the pmap directly and cache the
201  * locked pa in *locked.  The caller should then restart their loop in case
202  * the virtual to physical mapping has changed.
203  */
204 int
205 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
206 {
207 	vm_paddr_t lockpa;
208 
209 	lockpa = *locked;
210 	*locked = pa;
211 	if (lockpa) {
212 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
213 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
214 			return (0);
215 		PA_UNLOCK(lockpa);
216 	}
217 	if (PA_TRYLOCK(pa))
218 		return (0);
219 	PMAP_UNLOCK(pmap);
220 	atomic_add_int(&pa_tryrelock_restart, 1);
221 	PA_LOCK(pa);
222 	PMAP_LOCK(pmap);
223 	return (EAGAIN);
224 }
225 
226 /*
227  *	vm_set_page_size:
228  *
229  *	Sets the page size, perhaps based upon the memory
230  *	size.  Must be called before any use of page-size
231  *	dependent functions.
232  */
233 void
234 vm_set_page_size(void)
235 {
236 	if (vm_cnt.v_page_size == 0)
237 		vm_cnt.v_page_size = PAGE_SIZE;
238 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
239 		panic("vm_set_page_size: page size not a power of two");
240 }
241 
242 /*
243  *	vm_page_blacklist_next:
244  *
245  *	Find the next entry in the provided string of blacklist
246  *	addresses.  Entries are separated by space, comma, or newline.
247  *	If an invalid integer is encountered then the rest of the
248  *	string is skipped.  Updates the list pointer to the next
249  *	character, or NULL if the string is exhausted or invalid.
250  */
251 static vm_paddr_t
252 vm_page_blacklist_next(char **list, char *end)
253 {
254 	vm_paddr_t bad;
255 	char *cp, *pos;
256 
257 	if (list == NULL || *list == NULL)
258 		return (0);
259 	if (**list =='\0') {
260 		*list = NULL;
261 		return (0);
262 	}
263 
264 	/*
265 	 * If there's no end pointer then the buffer is coming from
266 	 * the kenv and we know it's null-terminated.
267 	 */
268 	if (end == NULL)
269 		end = *list + strlen(*list);
270 
271 	/* Ensure that strtoq() won't walk off the end */
272 	if (*end != '\0') {
273 		if (*end == '\n' || *end == ' ' || *end  == ',')
274 			*end = '\0';
275 		else {
276 			printf("Blacklist not terminated, skipping\n");
277 			*list = NULL;
278 			return (0);
279 		}
280 	}
281 
282 	for (pos = *list; *pos != '\0'; pos = cp) {
283 		bad = strtoq(pos, &cp, 0);
284 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
285 			if (bad == 0) {
286 				if (++cp < end)
287 					continue;
288 				else
289 					break;
290 			}
291 		} else
292 			break;
293 		if (*cp == '\0' || ++cp >= end)
294 			*list = NULL;
295 		else
296 			*list = cp;
297 		return (trunc_page(bad));
298 	}
299 	printf("Garbage in RAM blacklist, skipping\n");
300 	*list = NULL;
301 	return (0);
302 }
303 
304 /*
305  *	vm_page_blacklist_check:
306  *
307  *	Iterate through the provided string of blacklist addresses, pulling
308  *	each entry out of the physical allocator free list and putting it
309  *	onto a list for reporting via the vm.page_blacklist sysctl.
310  */
311 static void
312 vm_page_blacklist_check(char *list, char *end)
313 {
314 	vm_paddr_t pa;
315 	vm_page_t m;
316 	char *next;
317 	int ret;
318 
319 	next = list;
320 	while (next != NULL) {
321 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
322 			continue;
323 		m = vm_phys_paddr_to_vm_page(pa);
324 		if (m == NULL)
325 			continue;
326 		mtx_lock(&vm_page_queue_free_mtx);
327 		ret = vm_phys_unfree_page(m);
328 		mtx_unlock(&vm_page_queue_free_mtx);
329 		if (ret == TRUE) {
330 			TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
331 			if (bootverbose)
332 				printf("Skipping page with pa 0x%jx\n",
333 				    (uintmax_t)pa);
334 		}
335 	}
336 }
337 
338 /*
339  *	vm_page_blacklist_load:
340  *
341  *	Search for a special module named "ram_blacklist".  It'll be a
342  *	plain text file provided by the user via the loader directive
343  *	of the same name.
344  */
345 static void
346 vm_page_blacklist_load(char **list, char **end)
347 {
348 	void *mod;
349 	u_char *ptr;
350 	u_int len;
351 
352 	mod = NULL;
353 	ptr = NULL;
354 
355 	mod = preload_search_by_type("ram_blacklist");
356 	if (mod != NULL) {
357 		ptr = preload_fetch_addr(mod);
358 		len = preload_fetch_size(mod);
359         }
360 	*list = ptr;
361 	if (ptr != NULL)
362 		*end = ptr + len;
363 	else
364 		*end = NULL;
365 	return;
366 }
367 
368 static int
369 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
370 {
371 	vm_page_t m;
372 	struct sbuf sbuf;
373 	int error, first;
374 
375 	first = 1;
376 	error = sysctl_wire_old_buffer(req, 0);
377 	if (error != 0)
378 		return (error);
379 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
380 	TAILQ_FOREACH(m, &blacklist_head, listq) {
381 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
382 		    (uintmax_t)m->phys_addr);
383 		first = 0;
384 	}
385 	error = sbuf_finish(&sbuf);
386 	sbuf_delete(&sbuf);
387 	return (error);
388 }
389 
390 static void
391 vm_page_domain_init(struct vm_domain *vmd)
392 {
393 	struct vm_pagequeue *pq;
394 	int i;
395 
396 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
397 	    "vm inactive pagequeue";
398 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
399 	    &vm_cnt.v_inactive_count;
400 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
401 	    "vm active pagequeue";
402 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
403 	    &vm_cnt.v_active_count;
404 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
405 	    "vm laundry pagequeue";
406 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
407 	    &vm_cnt.v_laundry_count;
408 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
409 	    "vm unswappable pagequeue";
410 	/* Unswappable dirty pages are counted as being in the laundry. */
411 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_vcnt) =
412 	    &vm_cnt.v_laundry_count;
413 	vmd->vmd_page_count = 0;
414 	vmd->vmd_free_count = 0;
415 	vmd->vmd_segs = 0;
416 	vmd->vmd_oom = FALSE;
417 	for (i = 0; i < PQ_COUNT; i++) {
418 		pq = &vmd->vmd_pagequeues[i];
419 		TAILQ_INIT(&pq->pq_pl);
420 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
421 		    MTX_DEF | MTX_DUPOK);
422 	}
423 }
424 
425 /*
426  *	vm_page_startup:
427  *
428  *	Initializes the resident memory module.  Allocates physical memory for
429  *	bootstrapping UMA and some data structures that are used to manage
430  *	physical pages.  Initializes these structures, and populates the free
431  *	page queues.
432  */
433 vm_offset_t
434 vm_page_startup(vm_offset_t vaddr)
435 {
436 	struct vm_domain *vmd;
437 	struct vm_phys_seg *seg;
438 	vm_page_t m;
439 	char *list, *listend;
440 	vm_offset_t mapped;
441 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
442 	vm_paddr_t biggestsize, last_pa, pa;
443 	u_long pagecount;
444 	int biggestone, i, pages_per_zone, segind;
445 
446 	biggestsize = 0;
447 	biggestone = 0;
448 	vaddr = round_page(vaddr);
449 
450 	for (i = 0; phys_avail[i + 1]; i += 2) {
451 		phys_avail[i] = round_page(phys_avail[i]);
452 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
453 	}
454 	for (i = 0; phys_avail[i + 1]; i += 2) {
455 		size = phys_avail[i + 1] - phys_avail[i];
456 		if (size > biggestsize) {
457 			biggestone = i;
458 			biggestsize = size;
459 		}
460 	}
461 
462 	end = phys_avail[biggestone+1];
463 
464 	/*
465 	 * Initialize the page and queue locks.
466 	 */
467 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
468 	for (i = 0; i < PA_LOCK_COUNT; i++)
469 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
470 	for (i = 0; i < vm_ndomains; i++)
471 		vm_page_domain_init(&vm_dom[i]);
472 
473 	/*
474 	 * Almost all of the pages needed for bootstrapping UMA are used
475 	 * for zone structures, so if the number of CPUs results in those
476 	 * structures taking more than one page each, we set aside more pages
477 	 * in proportion to the zone structure size.
478 	 */
479 	pages_per_zone = howmany(sizeof(struct uma_zone) +
480 	    sizeof(struct uma_cache) * (mp_maxid + 1) +
481 	    roundup2(sizeof(struct uma_slab), sizeof(void *)), UMA_SLAB_SIZE);
482 	if (pages_per_zone > 1) {
483 		/* Reserve more pages so that we don't run out. */
484 		boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
485 	}
486 
487 	/*
488 	 * Allocate memory for use when boot strapping the kernel memory
489 	 * allocator.
490 	 *
491 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
492 	 * manually fetch the value.
493 	 */
494 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
495 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
496 	new_end = trunc_page(new_end);
497 	mapped = pmap_map(&vaddr, new_end, end,
498 	    VM_PROT_READ | VM_PROT_WRITE);
499 	bzero((void *)mapped, end - new_end);
500 	uma_startup((void *)mapped, boot_pages);
501 
502 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
503     defined(__i386__) || defined(__mips__)
504 	/*
505 	 * Allocate a bitmap to indicate that a random physical page
506 	 * needs to be included in a minidump.
507 	 *
508 	 * The amd64 port needs this to indicate which direct map pages
509 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
510 	 *
511 	 * However, i386 still needs this workspace internally within the
512 	 * minidump code.  In theory, they are not needed on i386, but are
513 	 * included should the sf_buf code decide to use them.
514 	 */
515 	last_pa = 0;
516 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
517 		if (dump_avail[i + 1] > last_pa)
518 			last_pa = dump_avail[i + 1];
519 	page_range = last_pa / PAGE_SIZE;
520 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
521 	new_end -= vm_page_dump_size;
522 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
523 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
524 	bzero((void *)vm_page_dump, vm_page_dump_size);
525 #else
526 	(void)last_pa;
527 #endif
528 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
529 	/*
530 	 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
531 	 * When pmap_map() uses the direct map, they are not automatically
532 	 * included.
533 	 */
534 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
535 		dump_add_page(pa);
536 #endif
537 	phys_avail[biggestone + 1] = new_end;
538 #ifdef __amd64__
539 	/*
540 	 * Request that the physical pages underlying the message buffer be
541 	 * included in a crash dump.  Since the message buffer is accessed
542 	 * through the direct map, they are not automatically included.
543 	 */
544 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
545 	last_pa = pa + round_page(msgbufsize);
546 	while (pa < last_pa) {
547 		dump_add_page(pa);
548 		pa += PAGE_SIZE;
549 	}
550 #endif
551 	/*
552 	 * Compute the number of pages of memory that will be available for
553 	 * use, taking into account the overhead of a page structure per page.
554 	 * In other words, solve
555 	 *	"available physical memory" - round_page(page_range *
556 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
557 	 * for page_range.
558 	 */
559 	low_avail = phys_avail[0];
560 	high_avail = phys_avail[1];
561 	for (i = 0; i < vm_phys_nsegs; i++) {
562 		if (vm_phys_segs[i].start < low_avail)
563 			low_avail = vm_phys_segs[i].start;
564 		if (vm_phys_segs[i].end > high_avail)
565 			high_avail = vm_phys_segs[i].end;
566 	}
567 	/* Skip the first chunk.  It is already accounted for. */
568 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
569 		if (phys_avail[i] < low_avail)
570 			low_avail = phys_avail[i];
571 		if (phys_avail[i + 1] > high_avail)
572 			high_avail = phys_avail[i + 1];
573 	}
574 	first_page = low_avail / PAGE_SIZE;
575 #ifdef VM_PHYSSEG_SPARSE
576 	size = 0;
577 	for (i = 0; i < vm_phys_nsegs; i++)
578 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
579 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
580 		size += phys_avail[i + 1] - phys_avail[i];
581 #elif defined(VM_PHYSSEG_DENSE)
582 	size = high_avail - low_avail;
583 #else
584 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
585 #endif
586 
587 #ifdef VM_PHYSSEG_DENSE
588 	/*
589 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
590 	 * the overhead of a page structure per page only if vm_page_array is
591 	 * allocated from the last physical memory chunk.  Otherwise, we must
592 	 * allocate page structures representing the physical memory
593 	 * underlying vm_page_array, even though they will not be used.
594 	 */
595 	if (new_end != high_avail)
596 		page_range = size / PAGE_SIZE;
597 	else
598 #endif
599 	{
600 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
601 
602 		/*
603 		 * If the partial bytes remaining are large enough for
604 		 * a page (PAGE_SIZE) without a corresponding
605 		 * 'struct vm_page', then new_end will contain an
606 		 * extra page after subtracting the length of the VM
607 		 * page array.  Compensate by subtracting an extra
608 		 * page from new_end.
609 		 */
610 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
611 			if (new_end == high_avail)
612 				high_avail -= PAGE_SIZE;
613 			new_end -= PAGE_SIZE;
614 		}
615 	}
616 	end = new_end;
617 
618 	/*
619 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
620 	 * However, because this page is allocated from KVM, out-of-bounds
621 	 * accesses using the direct map will not be trapped.
622 	 */
623 	vaddr += PAGE_SIZE;
624 
625 	/*
626 	 * Allocate physical memory for the page structures, and map it.
627 	 */
628 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
629 	mapped = pmap_map(&vaddr, new_end, end,
630 	    VM_PROT_READ | VM_PROT_WRITE);
631 	vm_page_array = (vm_page_t)mapped;
632 	vm_page_array_size = page_range;
633 
634 #if VM_NRESERVLEVEL > 0
635 	/*
636 	 * Allocate physical memory for the reservation management system's
637 	 * data structures, and map it.
638 	 */
639 	if (high_avail == end)
640 		high_avail = new_end;
641 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
642 #endif
643 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
644 	/*
645 	 * Include vm_page_array and vm_reserv_array in a crash dump.
646 	 */
647 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
648 		dump_add_page(pa);
649 #endif
650 	phys_avail[biggestone + 1] = new_end;
651 
652 	/*
653 	 * Add physical memory segments corresponding to the available
654 	 * physical pages.
655 	 */
656 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
657 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
658 
659 	/*
660 	 * Initialize the physical memory allocator.
661 	 */
662 	vm_phys_init();
663 
664 	/*
665 	 * Initialize the page structures and add every available page to the
666 	 * physical memory allocator's free lists.
667 	 */
668 	vm_cnt.v_page_count = 0;
669 	vm_cnt.v_free_count = 0;
670 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
671 		seg = &vm_phys_segs[segind];
672 		for (pa = seg->start; pa < seg->end; pa += PAGE_SIZE)
673 			vm_phys_init_page(pa);
674 
675 		/*
676 		 * Add the segment to the free lists only if it is covered by
677 		 * one of the ranges in phys_avail.  Because we've added the
678 		 * ranges to the vm_phys_segs array, we can assume that each
679 		 * segment is either entirely contained in one of the ranges,
680 		 * or doesn't overlap any of them.
681 		 */
682 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
683 			if (seg->start < phys_avail[i] ||
684 			    seg->end > phys_avail[i + 1])
685 				continue;
686 
687 			m = seg->first_page;
688 			pagecount = (u_long)atop(seg->end - seg->start);
689 
690 			mtx_lock(&vm_page_queue_free_mtx);
691 			vm_phys_free_contig(m, pagecount);
692 			vm_phys_freecnt_adj(m, (int)pagecount);
693 			mtx_unlock(&vm_page_queue_free_mtx);
694 			vm_cnt.v_page_count += (u_int)pagecount;
695 
696 			vmd = &vm_dom[seg->domain];
697 			vmd->vmd_page_count += (u_int)pagecount;
698 			vmd->vmd_segs |= 1UL << m->segind;
699 			break;
700 		}
701 	}
702 
703 	/*
704 	 * Remove blacklisted pages from the physical memory allocator.
705 	 */
706 	TAILQ_INIT(&blacklist_head);
707 	vm_page_blacklist_load(&list, &listend);
708 	vm_page_blacklist_check(list, listend);
709 
710 	list = kern_getenv("vm.blacklist");
711 	vm_page_blacklist_check(list, NULL);
712 
713 	freeenv(list);
714 #if VM_NRESERVLEVEL > 0
715 	/*
716 	 * Initialize the reservation management system.
717 	 */
718 	vm_reserv_init();
719 #endif
720 	return (vaddr);
721 }
722 
723 void
724 vm_page_reference(vm_page_t m)
725 {
726 
727 	vm_page_aflag_set(m, PGA_REFERENCED);
728 }
729 
730 /*
731  *	vm_page_busy_downgrade:
732  *
733  *	Downgrade an exclusive busy page into a single shared busy page.
734  */
735 void
736 vm_page_busy_downgrade(vm_page_t m)
737 {
738 	u_int x;
739 	bool locked;
740 
741 	vm_page_assert_xbusied(m);
742 	locked = mtx_owned(vm_page_lockptr(m));
743 
744 	for (;;) {
745 		x = m->busy_lock;
746 		x &= VPB_BIT_WAITERS;
747 		if (x != 0 && !locked)
748 			vm_page_lock(m);
749 		if (atomic_cmpset_rel_int(&m->busy_lock,
750 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
751 			break;
752 		if (x != 0 && !locked)
753 			vm_page_unlock(m);
754 	}
755 	if (x != 0) {
756 		wakeup(m);
757 		if (!locked)
758 			vm_page_unlock(m);
759 	}
760 }
761 
762 /*
763  *	vm_page_sbusied:
764  *
765  *	Return a positive value if the page is shared busied, 0 otherwise.
766  */
767 int
768 vm_page_sbusied(vm_page_t m)
769 {
770 	u_int x;
771 
772 	x = m->busy_lock;
773 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
774 }
775 
776 /*
777  *	vm_page_sunbusy:
778  *
779  *	Shared unbusy a page.
780  */
781 void
782 vm_page_sunbusy(vm_page_t m)
783 {
784 	u_int x;
785 
786 	vm_page_lock_assert(m, MA_NOTOWNED);
787 	vm_page_assert_sbusied(m);
788 
789 	for (;;) {
790 		x = m->busy_lock;
791 		if (VPB_SHARERS(x) > 1) {
792 			if (atomic_cmpset_int(&m->busy_lock, x,
793 			    x - VPB_ONE_SHARER))
794 				break;
795 			continue;
796 		}
797 		if ((x & VPB_BIT_WAITERS) == 0) {
798 			KASSERT(x == VPB_SHARERS_WORD(1),
799 			    ("vm_page_sunbusy: invalid lock state"));
800 			if (atomic_cmpset_int(&m->busy_lock,
801 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
802 				break;
803 			continue;
804 		}
805 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
806 		    ("vm_page_sunbusy: invalid lock state for waiters"));
807 
808 		vm_page_lock(m);
809 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
810 			vm_page_unlock(m);
811 			continue;
812 		}
813 		wakeup(m);
814 		vm_page_unlock(m);
815 		break;
816 	}
817 }
818 
819 /*
820  *	vm_page_busy_sleep:
821  *
822  *	Sleep and release the page lock, using the page pointer as wchan.
823  *	This is used to implement the hard-path of busying mechanism.
824  *
825  *	The given page must be locked.
826  *
827  *	If nonshared is true, sleep only if the page is xbusy.
828  */
829 void
830 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
831 {
832 	u_int x;
833 
834 	vm_page_assert_locked(m);
835 
836 	x = m->busy_lock;
837 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
838 	    ((x & VPB_BIT_WAITERS) == 0 &&
839 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
840 		vm_page_unlock(m);
841 		return;
842 	}
843 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
844 }
845 
846 /*
847  *	vm_page_trysbusy:
848  *
849  *	Try to shared busy a page.
850  *	If the operation succeeds 1 is returned otherwise 0.
851  *	The operation never sleeps.
852  */
853 int
854 vm_page_trysbusy(vm_page_t m)
855 {
856 	u_int x;
857 
858 	for (;;) {
859 		x = m->busy_lock;
860 		if ((x & VPB_BIT_SHARED) == 0)
861 			return (0);
862 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
863 			return (1);
864 	}
865 }
866 
867 static void
868 vm_page_xunbusy_locked(vm_page_t m)
869 {
870 
871 	vm_page_assert_xbusied(m);
872 	vm_page_assert_locked(m);
873 
874 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
875 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
876 	wakeup(m);
877 }
878 
879 void
880 vm_page_xunbusy_maybelocked(vm_page_t m)
881 {
882 	bool lockacq;
883 
884 	vm_page_assert_xbusied(m);
885 
886 	/*
887 	 * Fast path for unbusy.  If it succeeds, we know that there
888 	 * are no waiters, so we do not need a wakeup.
889 	 */
890 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
891 	    VPB_UNBUSIED))
892 		return;
893 
894 	lockacq = !mtx_owned(vm_page_lockptr(m));
895 	if (lockacq)
896 		vm_page_lock(m);
897 	vm_page_xunbusy_locked(m);
898 	if (lockacq)
899 		vm_page_unlock(m);
900 }
901 
902 /*
903  *	vm_page_xunbusy_hard:
904  *
905  *	Called after the first try the exclusive unbusy of a page failed.
906  *	It is assumed that the waiters bit is on.
907  */
908 void
909 vm_page_xunbusy_hard(vm_page_t m)
910 {
911 
912 	vm_page_assert_xbusied(m);
913 
914 	vm_page_lock(m);
915 	vm_page_xunbusy_locked(m);
916 	vm_page_unlock(m);
917 }
918 
919 /*
920  *	vm_page_flash:
921  *
922  *	Wakeup anyone waiting for the page.
923  *	The ownership bits do not change.
924  *
925  *	The given page must be locked.
926  */
927 void
928 vm_page_flash(vm_page_t m)
929 {
930 	u_int x;
931 
932 	vm_page_lock_assert(m, MA_OWNED);
933 
934 	for (;;) {
935 		x = m->busy_lock;
936 		if ((x & VPB_BIT_WAITERS) == 0)
937 			return;
938 		if (atomic_cmpset_int(&m->busy_lock, x,
939 		    x & (~VPB_BIT_WAITERS)))
940 			break;
941 	}
942 	wakeup(m);
943 }
944 
945 /*
946  * Avoid releasing and reacquiring the same page lock.
947  */
948 void
949 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
950 {
951 	struct mtx *mtx1;
952 
953 	mtx1 = vm_page_lockptr(m);
954 	if (*mtx == mtx1)
955 		return;
956 	if (*mtx != NULL)
957 		mtx_unlock(*mtx);
958 	*mtx = mtx1;
959 	mtx_lock(mtx1);
960 }
961 
962 /*
963  * Keep page from being freed by the page daemon
964  * much of the same effect as wiring, except much lower
965  * overhead and should be used only for *very* temporary
966  * holding ("wiring").
967  */
968 void
969 vm_page_hold(vm_page_t mem)
970 {
971 
972 	vm_page_lock_assert(mem, MA_OWNED);
973         mem->hold_count++;
974 }
975 
976 void
977 vm_page_unhold(vm_page_t mem)
978 {
979 
980 	vm_page_lock_assert(mem, MA_OWNED);
981 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
982 	--mem->hold_count;
983 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
984 		vm_page_free_toq(mem);
985 }
986 
987 /*
988  *	vm_page_unhold_pages:
989  *
990  *	Unhold each of the pages that is referenced by the given array.
991  */
992 void
993 vm_page_unhold_pages(vm_page_t *ma, int count)
994 {
995 	struct mtx *mtx;
996 
997 	mtx = NULL;
998 	for (; count != 0; count--) {
999 		vm_page_change_lock(*ma, &mtx);
1000 		vm_page_unhold(*ma);
1001 		ma++;
1002 	}
1003 	if (mtx != NULL)
1004 		mtx_unlock(mtx);
1005 }
1006 
1007 vm_page_t
1008 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1009 {
1010 	vm_page_t m;
1011 
1012 #ifdef VM_PHYSSEG_SPARSE
1013 	m = vm_phys_paddr_to_vm_page(pa);
1014 	if (m == NULL)
1015 		m = vm_phys_fictitious_to_vm_page(pa);
1016 	return (m);
1017 #elif defined(VM_PHYSSEG_DENSE)
1018 	long pi;
1019 
1020 	pi = atop(pa);
1021 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1022 		m = &vm_page_array[pi - first_page];
1023 		return (m);
1024 	}
1025 	return (vm_phys_fictitious_to_vm_page(pa));
1026 #else
1027 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1028 #endif
1029 }
1030 
1031 /*
1032  *	vm_page_getfake:
1033  *
1034  *	Create a fictitious page with the specified physical address and
1035  *	memory attribute.  The memory attribute is the only the machine-
1036  *	dependent aspect of a fictitious page that must be initialized.
1037  */
1038 vm_page_t
1039 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1040 {
1041 	vm_page_t m;
1042 
1043 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1044 	vm_page_initfake(m, paddr, memattr);
1045 	return (m);
1046 }
1047 
1048 void
1049 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1050 {
1051 
1052 	if ((m->flags & PG_FICTITIOUS) != 0) {
1053 		/*
1054 		 * The page's memattr might have changed since the
1055 		 * previous initialization.  Update the pmap to the
1056 		 * new memattr.
1057 		 */
1058 		goto memattr;
1059 	}
1060 	m->phys_addr = paddr;
1061 	m->queue = PQ_NONE;
1062 	/* Fictitious pages don't use "segind". */
1063 	m->flags = PG_FICTITIOUS;
1064 	/* Fictitious pages don't use "order" or "pool". */
1065 	m->oflags = VPO_UNMANAGED;
1066 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1067 	m->wire_count = 1;
1068 	pmap_page_init(m);
1069 memattr:
1070 	pmap_page_set_memattr(m, memattr);
1071 }
1072 
1073 /*
1074  *	vm_page_putfake:
1075  *
1076  *	Release a fictitious page.
1077  */
1078 void
1079 vm_page_putfake(vm_page_t m)
1080 {
1081 
1082 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1083 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1084 	    ("vm_page_putfake: bad page %p", m));
1085 	uma_zfree(fakepg_zone, m);
1086 }
1087 
1088 /*
1089  *	vm_page_updatefake:
1090  *
1091  *	Update the given fictitious page to the specified physical address and
1092  *	memory attribute.
1093  */
1094 void
1095 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1096 {
1097 
1098 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1099 	    ("vm_page_updatefake: bad page %p", m));
1100 	m->phys_addr = paddr;
1101 	pmap_page_set_memattr(m, memattr);
1102 }
1103 
1104 /*
1105  *	vm_page_free:
1106  *
1107  *	Free a page.
1108  */
1109 void
1110 vm_page_free(vm_page_t m)
1111 {
1112 
1113 	m->flags &= ~PG_ZERO;
1114 	vm_page_free_toq(m);
1115 }
1116 
1117 /*
1118  *	vm_page_free_zero:
1119  *
1120  *	Free a page to the zerod-pages queue
1121  */
1122 void
1123 vm_page_free_zero(vm_page_t m)
1124 {
1125 
1126 	m->flags |= PG_ZERO;
1127 	vm_page_free_toq(m);
1128 }
1129 
1130 /*
1131  * Unbusy and handle the page queueing for a page from a getpages request that
1132  * was optionally read ahead or behind.
1133  */
1134 void
1135 vm_page_readahead_finish(vm_page_t m)
1136 {
1137 
1138 	/* We shouldn't put invalid pages on queues. */
1139 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1140 
1141 	/*
1142 	 * Since the page is not the actually needed one, whether it should
1143 	 * be activated or deactivated is not obvious.  Empirical results
1144 	 * have shown that deactivating the page is usually the best choice,
1145 	 * unless the page is wanted by another thread.
1146 	 */
1147 	vm_page_lock(m);
1148 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1149 		vm_page_activate(m);
1150 	else
1151 		vm_page_deactivate(m);
1152 	vm_page_unlock(m);
1153 	vm_page_xunbusy(m);
1154 }
1155 
1156 /*
1157  *	vm_page_sleep_if_busy:
1158  *
1159  *	Sleep and release the page queues lock if the page is busied.
1160  *	Returns TRUE if the thread slept.
1161  *
1162  *	The given page must be unlocked and object containing it must
1163  *	be locked.
1164  */
1165 int
1166 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1167 {
1168 	vm_object_t obj;
1169 
1170 	vm_page_lock_assert(m, MA_NOTOWNED);
1171 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1172 
1173 	if (vm_page_busied(m)) {
1174 		/*
1175 		 * The page-specific object must be cached because page
1176 		 * identity can change during the sleep, causing the
1177 		 * re-lock of a different object.
1178 		 * It is assumed that a reference to the object is already
1179 		 * held by the callers.
1180 		 */
1181 		obj = m->object;
1182 		vm_page_lock(m);
1183 		VM_OBJECT_WUNLOCK(obj);
1184 		vm_page_busy_sleep(m, msg, false);
1185 		VM_OBJECT_WLOCK(obj);
1186 		return (TRUE);
1187 	}
1188 	return (FALSE);
1189 }
1190 
1191 /*
1192  *	vm_page_dirty_KBI:		[ internal use only ]
1193  *
1194  *	Set all bits in the page's dirty field.
1195  *
1196  *	The object containing the specified page must be locked if the
1197  *	call is made from the machine-independent layer.
1198  *
1199  *	See vm_page_clear_dirty_mask().
1200  *
1201  *	This function should only be called by vm_page_dirty().
1202  */
1203 void
1204 vm_page_dirty_KBI(vm_page_t m)
1205 {
1206 
1207 	/* Refer to this operation by its public name. */
1208 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1209 	    ("vm_page_dirty: page is invalid!"));
1210 	m->dirty = VM_PAGE_BITS_ALL;
1211 }
1212 
1213 /*
1214  *	vm_page_insert:		[ internal use only ]
1215  *
1216  *	Inserts the given mem entry into the object and object list.
1217  *
1218  *	The object must be locked.
1219  */
1220 int
1221 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1222 {
1223 	vm_page_t mpred;
1224 
1225 	VM_OBJECT_ASSERT_WLOCKED(object);
1226 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1227 	return (vm_page_insert_after(m, object, pindex, mpred));
1228 }
1229 
1230 /*
1231  *	vm_page_insert_after:
1232  *
1233  *	Inserts the page "m" into the specified object at offset "pindex".
1234  *
1235  *	The page "mpred" must immediately precede the offset "pindex" within
1236  *	the specified object.
1237  *
1238  *	The object must be locked.
1239  */
1240 static int
1241 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1242     vm_page_t mpred)
1243 {
1244 	vm_page_t msucc;
1245 
1246 	VM_OBJECT_ASSERT_WLOCKED(object);
1247 	KASSERT(m->object == NULL,
1248 	    ("vm_page_insert_after: page already inserted"));
1249 	if (mpred != NULL) {
1250 		KASSERT(mpred->object == object,
1251 		    ("vm_page_insert_after: object doesn't contain mpred"));
1252 		KASSERT(mpred->pindex < pindex,
1253 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1254 		msucc = TAILQ_NEXT(mpred, listq);
1255 	} else
1256 		msucc = TAILQ_FIRST(&object->memq);
1257 	if (msucc != NULL)
1258 		KASSERT(msucc->pindex > pindex,
1259 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1260 
1261 	/*
1262 	 * Record the object/offset pair in this page
1263 	 */
1264 	m->object = object;
1265 	m->pindex = pindex;
1266 
1267 	/*
1268 	 * Now link into the object's ordered list of backed pages.
1269 	 */
1270 	if (vm_radix_insert(&object->rtree, m)) {
1271 		m->object = NULL;
1272 		m->pindex = 0;
1273 		return (1);
1274 	}
1275 	vm_page_insert_radixdone(m, object, mpred);
1276 	return (0);
1277 }
1278 
1279 /*
1280  *	vm_page_insert_radixdone:
1281  *
1282  *	Complete page "m" insertion into the specified object after the
1283  *	radix trie hooking.
1284  *
1285  *	The page "mpred" must precede the offset "m->pindex" within the
1286  *	specified object.
1287  *
1288  *	The object must be locked.
1289  */
1290 static void
1291 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1292 {
1293 
1294 	VM_OBJECT_ASSERT_WLOCKED(object);
1295 	KASSERT(object != NULL && m->object == object,
1296 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1297 	if (mpred != NULL) {
1298 		KASSERT(mpred->object == object,
1299 		    ("vm_page_insert_after: object doesn't contain mpred"));
1300 		KASSERT(mpred->pindex < m->pindex,
1301 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1302 	}
1303 
1304 	if (mpred != NULL)
1305 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1306 	else
1307 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1308 
1309 	/*
1310 	 * Show that the object has one more resident page.
1311 	 */
1312 	object->resident_page_count++;
1313 
1314 	/*
1315 	 * Hold the vnode until the last page is released.
1316 	 */
1317 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1318 		vhold(object->handle);
1319 
1320 	/*
1321 	 * Since we are inserting a new and possibly dirty page,
1322 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1323 	 */
1324 	if (pmap_page_is_write_mapped(m))
1325 		vm_object_set_writeable_dirty(object);
1326 }
1327 
1328 /*
1329  *	vm_page_remove:
1330  *
1331  *	Removes the specified page from its containing object, but does not
1332  *	invalidate any backing storage.
1333  *
1334  *	The object must be locked.  The page must be locked if it is managed.
1335  */
1336 void
1337 vm_page_remove(vm_page_t m)
1338 {
1339 	vm_object_t object;
1340 	vm_page_t mrem;
1341 
1342 	if ((m->oflags & VPO_UNMANAGED) == 0)
1343 		vm_page_assert_locked(m);
1344 	if ((object = m->object) == NULL)
1345 		return;
1346 	VM_OBJECT_ASSERT_WLOCKED(object);
1347 	if (vm_page_xbusied(m))
1348 		vm_page_xunbusy_maybelocked(m);
1349 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1350 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1351 
1352 	/*
1353 	 * Now remove from the object's list of backed pages.
1354 	 */
1355 	TAILQ_REMOVE(&object->memq, m, listq);
1356 
1357 	/*
1358 	 * And show that the object has one fewer resident page.
1359 	 */
1360 	object->resident_page_count--;
1361 
1362 	/*
1363 	 * The vnode may now be recycled.
1364 	 */
1365 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1366 		vdrop(object->handle);
1367 
1368 	m->object = NULL;
1369 }
1370 
1371 /*
1372  *	vm_page_lookup:
1373  *
1374  *	Returns the page associated with the object/offset
1375  *	pair specified; if none is found, NULL is returned.
1376  *
1377  *	The object must be locked.
1378  */
1379 vm_page_t
1380 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1381 {
1382 
1383 	VM_OBJECT_ASSERT_LOCKED(object);
1384 	return (vm_radix_lookup(&object->rtree, pindex));
1385 }
1386 
1387 /*
1388  *	vm_page_find_least:
1389  *
1390  *	Returns the page associated with the object with least pindex
1391  *	greater than or equal to the parameter pindex, or NULL.
1392  *
1393  *	The object must be locked.
1394  */
1395 vm_page_t
1396 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1397 {
1398 	vm_page_t m;
1399 
1400 	VM_OBJECT_ASSERT_LOCKED(object);
1401 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1402 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1403 	return (m);
1404 }
1405 
1406 /*
1407  * Returns the given page's successor (by pindex) within the object if it is
1408  * resident; if none is found, NULL is returned.
1409  *
1410  * The object must be locked.
1411  */
1412 vm_page_t
1413 vm_page_next(vm_page_t m)
1414 {
1415 	vm_page_t next;
1416 
1417 	VM_OBJECT_ASSERT_LOCKED(m->object);
1418 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1419 		MPASS(next->object == m->object);
1420 		if (next->pindex != m->pindex + 1)
1421 			next = NULL;
1422 	}
1423 	return (next);
1424 }
1425 
1426 /*
1427  * Returns the given page's predecessor (by pindex) within the object if it is
1428  * resident; if none is found, NULL is returned.
1429  *
1430  * The object must be locked.
1431  */
1432 vm_page_t
1433 vm_page_prev(vm_page_t m)
1434 {
1435 	vm_page_t prev;
1436 
1437 	VM_OBJECT_ASSERT_LOCKED(m->object);
1438 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1439 		MPASS(prev->object == m->object);
1440 		if (prev->pindex != m->pindex - 1)
1441 			prev = NULL;
1442 	}
1443 	return (prev);
1444 }
1445 
1446 /*
1447  * Uses the page mnew as a replacement for an existing page at index
1448  * pindex which must be already present in the object.
1449  *
1450  * The existing page must not be on a paging queue.
1451  */
1452 vm_page_t
1453 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1454 {
1455 	vm_page_t mold;
1456 
1457 	VM_OBJECT_ASSERT_WLOCKED(object);
1458 	KASSERT(mnew->object == NULL,
1459 	    ("vm_page_replace: page already in object"));
1460 
1461 	/*
1462 	 * This function mostly follows vm_page_insert() and
1463 	 * vm_page_remove() without the radix, object count and vnode
1464 	 * dance.  Double check such functions for more comments.
1465 	 */
1466 
1467 	mnew->object = object;
1468 	mnew->pindex = pindex;
1469 	mold = vm_radix_replace(&object->rtree, mnew);
1470 	KASSERT(mold->queue == PQ_NONE,
1471 	    ("vm_page_replace: mold is on a paging queue"));
1472 
1473 	/* Keep the resident page list in sorted order. */
1474 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1475 	TAILQ_REMOVE(&object->memq, mold, listq);
1476 
1477 	mold->object = NULL;
1478 	vm_page_xunbusy_maybelocked(mold);
1479 
1480 	/*
1481 	 * The object's resident_page_count does not change because we have
1482 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1483 	 */
1484 	if (pmap_page_is_write_mapped(mnew))
1485 		vm_object_set_writeable_dirty(object);
1486 	return (mold);
1487 }
1488 
1489 /*
1490  *	vm_page_rename:
1491  *
1492  *	Move the given memory entry from its
1493  *	current object to the specified target object/offset.
1494  *
1495  *	Note: swap associated with the page must be invalidated by the move.  We
1496  *	      have to do this for several reasons:  (1) we aren't freeing the
1497  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1498  *	      moving the page from object A to B, and will then later move
1499  *	      the backing store from A to B and we can't have a conflict.
1500  *
1501  *	Note: we *always* dirty the page.  It is necessary both for the
1502  *	      fact that we moved it, and because we may be invalidating
1503  *	      swap.
1504  *
1505  *	The objects must be locked.
1506  */
1507 int
1508 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1509 {
1510 	vm_page_t mpred;
1511 	vm_pindex_t opidx;
1512 
1513 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1514 
1515 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1516 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1517 	    ("vm_page_rename: pindex already renamed"));
1518 
1519 	/*
1520 	 * Create a custom version of vm_page_insert() which does not depend
1521 	 * by m_prev and can cheat on the implementation aspects of the
1522 	 * function.
1523 	 */
1524 	opidx = m->pindex;
1525 	m->pindex = new_pindex;
1526 	if (vm_radix_insert(&new_object->rtree, m)) {
1527 		m->pindex = opidx;
1528 		return (1);
1529 	}
1530 
1531 	/*
1532 	 * The operation cannot fail anymore.  The removal must happen before
1533 	 * the listq iterator is tainted.
1534 	 */
1535 	m->pindex = opidx;
1536 	vm_page_lock(m);
1537 	vm_page_remove(m);
1538 
1539 	/* Return back to the new pindex to complete vm_page_insert(). */
1540 	m->pindex = new_pindex;
1541 	m->object = new_object;
1542 	vm_page_unlock(m);
1543 	vm_page_insert_radixdone(m, new_object, mpred);
1544 	vm_page_dirty(m);
1545 	return (0);
1546 }
1547 
1548 /*
1549  *	vm_page_alloc:
1550  *
1551  *	Allocate and return a page that is associated with the specified
1552  *	object and offset pair.  By default, this page is exclusive busied.
1553  *
1554  *	The caller must always specify an allocation class.
1555  *
1556  *	allocation classes:
1557  *	VM_ALLOC_NORMAL		normal process request
1558  *	VM_ALLOC_SYSTEM		system *really* needs a page
1559  *	VM_ALLOC_INTERRUPT	interrupt time request
1560  *
1561  *	optional allocation flags:
1562  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1563  *				intends to allocate
1564  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1565  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1566  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1567  *				should not be exclusive busy
1568  *	VM_ALLOC_SBUSY		shared busy the allocated page
1569  *	VM_ALLOC_WIRED		wire the allocated page
1570  *	VM_ALLOC_ZERO		prefer a zeroed page
1571  *
1572  *	This routine may not sleep.
1573  */
1574 vm_page_t
1575 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1576 {
1577 
1578 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1579 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1580 }
1581 
1582 /*
1583  * Allocate a page in the specified object with the given page index.  To
1584  * optimize insertion of the page into the object, the caller must also specifiy
1585  * the resident page in the object with largest index smaller than the given
1586  * page index, or NULL if no such page exists.
1587  */
1588 vm_page_t
1589 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req,
1590     vm_page_t mpred)
1591 {
1592 	vm_page_t m;
1593 	int flags, req_class;
1594 	u_int free_count;
1595 
1596 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1597 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1598 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1599 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1600 	    ("inconsistent object(%p)/req(%x)", object, req));
1601 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1602 	    ("Can't sleep and retry object insertion."));
1603 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1604 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1605 	    (uintmax_t)pindex));
1606 	if (object != NULL)
1607 		VM_OBJECT_ASSERT_WLOCKED(object);
1608 
1609 	req_class = req & VM_ALLOC_CLASS_MASK;
1610 
1611 	/*
1612 	 * The page daemon is allowed to dig deeper into the free page list.
1613 	 */
1614 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1615 		req_class = VM_ALLOC_SYSTEM;
1616 
1617 	/*
1618 	 * Allocate a page if the number of free pages exceeds the minimum
1619 	 * for the request class.
1620 	 */
1621 again:
1622 	mtx_lock(&vm_page_queue_free_mtx);
1623 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1624 	    (req_class == VM_ALLOC_SYSTEM &&
1625 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1626 	    (req_class == VM_ALLOC_INTERRUPT &&
1627 	    vm_cnt.v_free_count > 0)) {
1628 		/*
1629 		 * Can we allocate the page from a reservation?
1630 		 */
1631 #if VM_NRESERVLEVEL > 0
1632 		if (object == NULL || (object->flags & (OBJ_COLORED |
1633 		    OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1634 		    vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1635 #endif
1636 		{
1637 			/*
1638 			 * If not, allocate it from the free page queues.
1639 			 */
1640 			m = vm_phys_alloc_pages(object != NULL ?
1641 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1642 #if VM_NRESERVLEVEL > 0
1643 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1644 				m = vm_phys_alloc_pages(object != NULL ?
1645 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1646 				    0);
1647 			}
1648 #endif
1649 		}
1650 	} else {
1651 		/*
1652 		 * Not allocatable, give up.
1653 		 */
1654 		if (vm_page_alloc_fail(object, req))
1655 			goto again;
1656 		return (NULL);
1657 	}
1658 
1659 	/*
1660 	 *  At this point we had better have found a good page.
1661 	 */
1662 	KASSERT(m != NULL, ("missing page"));
1663 	free_count = vm_phys_freecnt_adj(m, -1);
1664 	mtx_unlock(&vm_page_queue_free_mtx);
1665 	vm_page_alloc_check(m);
1666 
1667 	/*
1668 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1669 	 */
1670 	flags = 0;
1671 	if ((req & VM_ALLOC_ZERO) != 0)
1672 		flags = PG_ZERO;
1673 	flags &= m->flags;
1674 	if ((req & VM_ALLOC_NODUMP) != 0)
1675 		flags |= PG_NODUMP;
1676 	m->flags = flags;
1677 	m->aflags = 0;
1678 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1679 	    VPO_UNMANAGED : 0;
1680 	m->busy_lock = VPB_UNBUSIED;
1681 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1682 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1683 	if ((req & VM_ALLOC_SBUSY) != 0)
1684 		m->busy_lock = VPB_SHARERS_WORD(1);
1685 	if (req & VM_ALLOC_WIRED) {
1686 		/*
1687 		 * The page lock is not required for wiring a page until that
1688 		 * page is inserted into the object.
1689 		 */
1690 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1691 		m->wire_count = 1;
1692 	}
1693 	m->act_count = 0;
1694 
1695 	if (object != NULL) {
1696 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1697 			pagedaemon_wakeup();
1698 			if (req & VM_ALLOC_WIRED) {
1699 				atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1700 				m->wire_count = 0;
1701 			}
1702 			KASSERT(m->object == NULL, ("page %p has object", m));
1703 			m->oflags = VPO_UNMANAGED;
1704 			m->busy_lock = VPB_UNBUSIED;
1705 			/* Don't change PG_ZERO. */
1706 			vm_page_free_toq(m);
1707 			if (req & VM_ALLOC_WAITFAIL) {
1708 				VM_OBJECT_WUNLOCK(object);
1709 				vm_radix_wait();
1710 				VM_OBJECT_WLOCK(object);
1711 			}
1712 			return (NULL);
1713 		}
1714 
1715 		/* Ignore device objects; the pager sets "memattr" for them. */
1716 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1717 		    (object->flags & OBJ_FICTITIOUS) == 0)
1718 			pmap_page_set_memattr(m, object->memattr);
1719 	} else
1720 		m->pindex = pindex;
1721 
1722 	/*
1723 	 * Don't wakeup too often - wakeup the pageout daemon when
1724 	 * we would be nearly out of memory.
1725 	 */
1726 	if (vm_paging_needed(free_count))
1727 		pagedaemon_wakeup();
1728 
1729 	return (m);
1730 }
1731 
1732 /*
1733  *	vm_page_alloc_contig:
1734  *
1735  *	Allocate a contiguous set of physical pages of the given size "npages"
1736  *	from the free lists.  All of the physical pages must be at or above
1737  *	the given physical address "low" and below the given physical address
1738  *	"high".  The given value "alignment" determines the alignment of the
1739  *	first physical page in the set.  If the given value "boundary" is
1740  *	non-zero, then the set of physical pages cannot cross any physical
1741  *	address boundary that is a multiple of that value.  Both "alignment"
1742  *	and "boundary" must be a power of two.
1743  *
1744  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1745  *	then the memory attribute setting for the physical pages is configured
1746  *	to the object's memory attribute setting.  Otherwise, the memory
1747  *	attribute setting for the physical pages is configured to "memattr",
1748  *	overriding the object's memory attribute setting.  However, if the
1749  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1750  *	memory attribute setting for the physical pages cannot be configured
1751  *	to VM_MEMATTR_DEFAULT.
1752  *
1753  *	The specified object may not contain fictitious pages.
1754  *
1755  *	The caller must always specify an allocation class.
1756  *
1757  *	allocation classes:
1758  *	VM_ALLOC_NORMAL		normal process request
1759  *	VM_ALLOC_SYSTEM		system *really* needs a page
1760  *	VM_ALLOC_INTERRUPT	interrupt time request
1761  *
1762  *	optional allocation flags:
1763  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1764  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1765  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1766  *				should not be exclusive busy
1767  *	VM_ALLOC_SBUSY		shared busy the allocated page
1768  *	VM_ALLOC_WIRED		wire the allocated page
1769  *	VM_ALLOC_ZERO		prefer a zeroed page
1770  *
1771  *	This routine may not sleep.
1772  */
1773 vm_page_t
1774 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1775     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1776     vm_paddr_t boundary, vm_memattr_t memattr)
1777 {
1778 	vm_page_t m, m_ret, mpred;
1779 	u_int busy_lock, flags, oflags;
1780 	int req_class;
1781 
1782 	mpred = NULL;	/* XXX: pacify gcc */
1783 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1784 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1785 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1786 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1787 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1788 	    req));
1789 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1790 	    ("Can't sleep and retry object insertion."));
1791 	if (object != NULL) {
1792 		VM_OBJECT_ASSERT_WLOCKED(object);
1793 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1794 		    ("vm_page_alloc_contig: object %p has fictitious pages",
1795 		    object));
1796 	}
1797 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1798 	req_class = req & VM_ALLOC_CLASS_MASK;
1799 
1800 	/*
1801 	 * The page daemon is allowed to dig deeper into the free page list.
1802 	 */
1803 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1804 		req_class = VM_ALLOC_SYSTEM;
1805 
1806 	if (object != NULL) {
1807 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1808 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1809 		    ("vm_page_alloc_contig: pindex already allocated"));
1810 	}
1811 
1812 	/*
1813 	 * Can we allocate the pages without the number of free pages falling
1814 	 * below the lower bound for the allocation class?
1815 	 */
1816 again:
1817 	mtx_lock(&vm_page_queue_free_mtx);
1818 	if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1819 	    (req_class == VM_ALLOC_SYSTEM &&
1820 	    vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1821 	    (req_class == VM_ALLOC_INTERRUPT &&
1822 	    vm_cnt.v_free_count >= npages)) {
1823 		/*
1824 		 * Can we allocate the pages from a reservation?
1825 		 */
1826 #if VM_NRESERVLEVEL > 0
1827 retry:
1828 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1829 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1830 		    low, high, alignment, boundary, mpred)) == NULL)
1831 #endif
1832 			/*
1833 			 * If not, allocate them from the free page queues.
1834 			 */
1835 			m_ret = vm_phys_alloc_contig(npages, low, high,
1836 			    alignment, boundary);
1837 	} else {
1838 		if (vm_page_alloc_fail(object, req))
1839 			goto again;
1840 		return (NULL);
1841 	}
1842 	if (m_ret != NULL)
1843 		vm_phys_freecnt_adj(m_ret, -npages);
1844 	else {
1845 #if VM_NRESERVLEVEL > 0
1846 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1847 		    boundary))
1848 			goto retry;
1849 #endif
1850 	}
1851 	mtx_unlock(&vm_page_queue_free_mtx);
1852 	if (m_ret == NULL)
1853 		return (NULL);
1854 	for (m = m_ret; m < &m_ret[npages]; m++)
1855 		vm_page_alloc_check(m);
1856 
1857 	/*
1858 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1859 	 */
1860 	flags = 0;
1861 	if ((req & VM_ALLOC_ZERO) != 0)
1862 		flags = PG_ZERO;
1863 	if ((req & VM_ALLOC_NODUMP) != 0)
1864 		flags |= PG_NODUMP;
1865 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1866 	    VPO_UNMANAGED : 0;
1867 	busy_lock = VPB_UNBUSIED;
1868 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1869 		busy_lock = VPB_SINGLE_EXCLUSIVER;
1870 	if ((req & VM_ALLOC_SBUSY) != 0)
1871 		busy_lock = VPB_SHARERS_WORD(1);
1872 	if ((req & VM_ALLOC_WIRED) != 0)
1873 		atomic_add_int(&vm_cnt.v_wire_count, npages);
1874 	if (object != NULL) {
1875 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1876 		    memattr == VM_MEMATTR_DEFAULT)
1877 			memattr = object->memattr;
1878 	}
1879 	for (m = m_ret; m < &m_ret[npages]; m++) {
1880 		m->aflags = 0;
1881 		m->flags = (m->flags | PG_NODUMP) & flags;
1882 		m->busy_lock = busy_lock;
1883 		if ((req & VM_ALLOC_WIRED) != 0)
1884 			m->wire_count = 1;
1885 		m->act_count = 0;
1886 		m->oflags = oflags;
1887 		if (object != NULL) {
1888 			if (vm_page_insert_after(m, object, pindex, mpred)) {
1889 				pagedaemon_wakeup();
1890 				if ((req & VM_ALLOC_WIRED) != 0)
1891 					atomic_subtract_int(
1892 					    &vm_cnt.v_wire_count, npages);
1893 				KASSERT(m->object == NULL,
1894 				    ("page %p has object", m));
1895 				mpred = m;
1896 				for (m = m_ret; m < &m_ret[npages]; m++) {
1897 					if (m <= mpred &&
1898 					    (req & VM_ALLOC_WIRED) != 0)
1899 						m->wire_count = 0;
1900 					m->oflags = VPO_UNMANAGED;
1901 					m->busy_lock = VPB_UNBUSIED;
1902 					/* Don't change PG_ZERO. */
1903 					vm_page_free_toq(m);
1904 				}
1905 				if (req & VM_ALLOC_WAITFAIL) {
1906 					VM_OBJECT_WUNLOCK(object);
1907 					vm_radix_wait();
1908 					VM_OBJECT_WLOCK(object);
1909 				}
1910 				return (NULL);
1911 			}
1912 			mpred = m;
1913 		} else
1914 			m->pindex = pindex;
1915 		if (memattr != VM_MEMATTR_DEFAULT)
1916 			pmap_page_set_memattr(m, memattr);
1917 		pindex++;
1918 	}
1919 	if (vm_paging_needed(vm_cnt.v_free_count))
1920 		pagedaemon_wakeup();
1921 	return (m_ret);
1922 }
1923 
1924 /*
1925  * Check a page that has been freshly dequeued from a freelist.
1926  */
1927 static void
1928 vm_page_alloc_check(vm_page_t m)
1929 {
1930 
1931 	KASSERT(m->object == NULL, ("page %p has object", m));
1932 	KASSERT(m->queue == PQ_NONE,
1933 	    ("page %p has unexpected queue %d", m, m->queue));
1934 	KASSERT(m->wire_count == 0, ("page %p is wired", m));
1935 	KASSERT(m->hold_count == 0, ("page %p is held", m));
1936 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1937 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
1938 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1939 	    ("page %p has unexpected memattr %d",
1940 	    m, pmap_page_get_memattr(m)));
1941 	KASSERT(m->valid == 0, ("free page %p is valid", m));
1942 }
1943 
1944 /*
1945  * 	vm_page_alloc_freelist:
1946  *
1947  *	Allocate a physical page from the specified free page list.
1948  *
1949  *	The caller must always specify an allocation class.
1950  *
1951  *	allocation classes:
1952  *	VM_ALLOC_NORMAL		normal process request
1953  *	VM_ALLOC_SYSTEM		system *really* needs a page
1954  *	VM_ALLOC_INTERRUPT	interrupt time request
1955  *
1956  *	optional allocation flags:
1957  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1958  *				intends to allocate
1959  *	VM_ALLOC_WIRED		wire the allocated page
1960  *	VM_ALLOC_ZERO		prefer a zeroed page
1961  *
1962  *	This routine may not sleep.
1963  */
1964 vm_page_t
1965 vm_page_alloc_freelist(int flind, int req)
1966 {
1967 	vm_page_t m;
1968 	u_int flags, free_count;
1969 	int req_class;
1970 
1971 	req_class = req & VM_ALLOC_CLASS_MASK;
1972 
1973 	/*
1974 	 * The page daemon is allowed to dig deeper into the free page list.
1975 	 */
1976 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1977 		req_class = VM_ALLOC_SYSTEM;
1978 
1979 	/*
1980 	 * Do not allocate reserved pages unless the req has asked for it.
1981 	 */
1982 again:
1983 	mtx_lock(&vm_page_queue_free_mtx);
1984 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1985 	    (req_class == VM_ALLOC_SYSTEM &&
1986 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1987 	    (req_class == VM_ALLOC_INTERRUPT &&
1988 	    vm_cnt.v_free_count > 0)) {
1989 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1990 	} else {
1991 		if (vm_page_alloc_fail(NULL, req))
1992 			goto again;
1993 		return (NULL);
1994 	}
1995 	if (m == NULL) {
1996 		mtx_unlock(&vm_page_queue_free_mtx);
1997 		return (NULL);
1998 	}
1999 	free_count = vm_phys_freecnt_adj(m, -1);
2000 	mtx_unlock(&vm_page_queue_free_mtx);
2001 	vm_page_alloc_check(m);
2002 
2003 	/*
2004 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2005 	 */
2006 	m->aflags = 0;
2007 	flags = 0;
2008 	if ((req & VM_ALLOC_ZERO) != 0)
2009 		flags = PG_ZERO;
2010 	m->flags &= flags;
2011 	if ((req & VM_ALLOC_WIRED) != 0) {
2012 		/*
2013 		 * The page lock is not required for wiring a page that does
2014 		 * not belong to an object.
2015 		 */
2016 		atomic_add_int(&vm_cnt.v_wire_count, 1);
2017 		m->wire_count = 1;
2018 	}
2019 	/* Unmanaged pages don't use "act_count". */
2020 	m->oflags = VPO_UNMANAGED;
2021 	if (vm_paging_needed(free_count))
2022 		pagedaemon_wakeup();
2023 	return (m);
2024 }
2025 
2026 #define	VPSC_ANY	0	/* No restrictions. */
2027 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2028 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2029 
2030 /*
2031  *	vm_page_scan_contig:
2032  *
2033  *	Scan vm_page_array[] between the specified entries "m_start" and
2034  *	"m_end" for a run of contiguous physical pages that satisfy the
2035  *	specified conditions, and return the lowest page in the run.  The
2036  *	specified "alignment" determines the alignment of the lowest physical
2037  *	page in the run.  If the specified "boundary" is non-zero, then the
2038  *	run of physical pages cannot span a physical address that is a
2039  *	multiple of "boundary".
2040  *
2041  *	"m_end" is never dereferenced, so it need not point to a vm_page
2042  *	structure within vm_page_array[].
2043  *
2044  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2045  *	span a hole (or discontiguity) in the physical address space.  Both
2046  *	"alignment" and "boundary" must be a power of two.
2047  */
2048 vm_page_t
2049 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2050     u_long alignment, vm_paddr_t boundary, int options)
2051 {
2052 	struct mtx *m_mtx;
2053 	vm_object_t object;
2054 	vm_paddr_t pa;
2055 	vm_page_t m, m_run;
2056 #if VM_NRESERVLEVEL > 0
2057 	int level;
2058 #endif
2059 	int m_inc, order, run_ext, run_len;
2060 
2061 	KASSERT(npages > 0, ("npages is 0"));
2062 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2063 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2064 	m_run = NULL;
2065 	run_len = 0;
2066 	m_mtx = NULL;
2067 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2068 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2069 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2070 
2071 		/*
2072 		 * If the current page would be the start of a run, check its
2073 		 * physical address against the end, alignment, and boundary
2074 		 * conditions.  If it doesn't satisfy these conditions, either
2075 		 * terminate the scan or advance to the next page that
2076 		 * satisfies the failed condition.
2077 		 */
2078 		if (run_len == 0) {
2079 			KASSERT(m_run == NULL, ("m_run != NULL"));
2080 			if (m + npages > m_end)
2081 				break;
2082 			pa = VM_PAGE_TO_PHYS(m);
2083 			if ((pa & (alignment - 1)) != 0) {
2084 				m_inc = atop(roundup2(pa, alignment) - pa);
2085 				continue;
2086 			}
2087 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2088 			    boundary) != 0) {
2089 				m_inc = atop(roundup2(pa, boundary) - pa);
2090 				continue;
2091 			}
2092 		} else
2093 			KASSERT(m_run != NULL, ("m_run == NULL"));
2094 
2095 		vm_page_change_lock(m, &m_mtx);
2096 		m_inc = 1;
2097 retry:
2098 		if (m->wire_count != 0 || m->hold_count != 0)
2099 			run_ext = 0;
2100 #if VM_NRESERVLEVEL > 0
2101 		else if ((level = vm_reserv_level(m)) >= 0 &&
2102 		    (options & VPSC_NORESERV) != 0) {
2103 			run_ext = 0;
2104 			/* Advance to the end of the reservation. */
2105 			pa = VM_PAGE_TO_PHYS(m);
2106 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2107 			    pa);
2108 		}
2109 #endif
2110 		else if ((object = m->object) != NULL) {
2111 			/*
2112 			 * The page is considered eligible for relocation if
2113 			 * and only if it could be laundered or reclaimed by
2114 			 * the page daemon.
2115 			 */
2116 			if (!VM_OBJECT_TRYRLOCK(object)) {
2117 				mtx_unlock(m_mtx);
2118 				VM_OBJECT_RLOCK(object);
2119 				mtx_lock(m_mtx);
2120 				if (m->object != object) {
2121 					/*
2122 					 * The page may have been freed.
2123 					 */
2124 					VM_OBJECT_RUNLOCK(object);
2125 					goto retry;
2126 				} else if (m->wire_count != 0 ||
2127 				    m->hold_count != 0) {
2128 					run_ext = 0;
2129 					goto unlock;
2130 				}
2131 			}
2132 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2133 			    ("page %p is PG_UNHOLDFREE", m));
2134 			/* Don't care: PG_NODUMP, PG_ZERO. */
2135 			if (object->type != OBJT_DEFAULT &&
2136 			    object->type != OBJT_SWAP &&
2137 			    object->type != OBJT_VNODE) {
2138 				run_ext = 0;
2139 #if VM_NRESERVLEVEL > 0
2140 			} else if ((options & VPSC_NOSUPER) != 0 &&
2141 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2142 				run_ext = 0;
2143 				/* Advance to the end of the superpage. */
2144 				pa = VM_PAGE_TO_PHYS(m);
2145 				m_inc = atop(roundup2(pa + 1,
2146 				    vm_reserv_size(level)) - pa);
2147 #endif
2148 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2149 			    m->queue != PQ_NONE && !vm_page_busied(m)) {
2150 				/*
2151 				 * The page is allocated but eligible for
2152 				 * relocation.  Extend the current run by one
2153 				 * page.
2154 				 */
2155 				KASSERT(pmap_page_get_memattr(m) ==
2156 				    VM_MEMATTR_DEFAULT,
2157 				    ("page %p has an unexpected memattr", m));
2158 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2159 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2160 				    ("page %p has unexpected oflags", m));
2161 				/* Don't care: VPO_NOSYNC. */
2162 				run_ext = 1;
2163 			} else
2164 				run_ext = 0;
2165 unlock:
2166 			VM_OBJECT_RUNLOCK(object);
2167 #if VM_NRESERVLEVEL > 0
2168 		} else if (level >= 0) {
2169 			/*
2170 			 * The page is reserved but not yet allocated.  In
2171 			 * other words, it is still free.  Extend the current
2172 			 * run by one page.
2173 			 */
2174 			run_ext = 1;
2175 #endif
2176 		} else if ((order = m->order) < VM_NFREEORDER) {
2177 			/*
2178 			 * The page is enqueued in the physical memory
2179 			 * allocator's free page queues.  Moreover, it is the
2180 			 * first page in a power-of-two-sized run of
2181 			 * contiguous free pages.  Add these pages to the end
2182 			 * of the current run, and jump ahead.
2183 			 */
2184 			run_ext = 1 << order;
2185 			m_inc = 1 << order;
2186 		} else {
2187 			/*
2188 			 * Skip the page for one of the following reasons: (1)
2189 			 * It is enqueued in the physical memory allocator's
2190 			 * free page queues.  However, it is not the first
2191 			 * page in a run of contiguous free pages.  (This case
2192 			 * rarely occurs because the scan is performed in
2193 			 * ascending order.) (2) It is not reserved, and it is
2194 			 * transitioning from free to allocated.  (Conversely,
2195 			 * the transition from allocated to free for managed
2196 			 * pages is blocked by the page lock.) (3) It is
2197 			 * allocated but not contained by an object and not
2198 			 * wired, e.g., allocated by Xen's balloon driver.
2199 			 */
2200 			run_ext = 0;
2201 		}
2202 
2203 		/*
2204 		 * Extend or reset the current run of pages.
2205 		 */
2206 		if (run_ext > 0) {
2207 			if (run_len == 0)
2208 				m_run = m;
2209 			run_len += run_ext;
2210 		} else {
2211 			if (run_len > 0) {
2212 				m_run = NULL;
2213 				run_len = 0;
2214 			}
2215 		}
2216 	}
2217 	if (m_mtx != NULL)
2218 		mtx_unlock(m_mtx);
2219 	if (run_len >= npages)
2220 		return (m_run);
2221 	return (NULL);
2222 }
2223 
2224 /*
2225  *	vm_page_reclaim_run:
2226  *
2227  *	Try to relocate each of the allocated virtual pages within the
2228  *	specified run of physical pages to a new physical address.  Free the
2229  *	physical pages underlying the relocated virtual pages.  A virtual page
2230  *	is relocatable if and only if it could be laundered or reclaimed by
2231  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2232  *	physical address above "high".
2233  *
2234  *	Returns 0 if every physical page within the run was already free or
2235  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2236  *	value indicating why the last attempt to relocate a virtual page was
2237  *	unsuccessful.
2238  *
2239  *	"req_class" must be an allocation class.
2240  */
2241 static int
2242 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2243     vm_paddr_t high)
2244 {
2245 	struct mtx *m_mtx;
2246 	struct spglist free;
2247 	vm_object_t object;
2248 	vm_paddr_t pa;
2249 	vm_page_t m, m_end, m_new;
2250 	int error, order, req;
2251 
2252 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2253 	    ("req_class is not an allocation class"));
2254 	SLIST_INIT(&free);
2255 	error = 0;
2256 	m = m_run;
2257 	m_end = m_run + npages;
2258 	m_mtx = NULL;
2259 	for (; error == 0 && m < m_end; m++) {
2260 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2261 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2262 
2263 		/*
2264 		 * Avoid releasing and reacquiring the same page lock.
2265 		 */
2266 		vm_page_change_lock(m, &m_mtx);
2267 retry:
2268 		if (m->wire_count != 0 || m->hold_count != 0)
2269 			error = EBUSY;
2270 		else if ((object = m->object) != NULL) {
2271 			/*
2272 			 * The page is relocated if and only if it could be
2273 			 * laundered or reclaimed by the page daemon.
2274 			 */
2275 			if (!VM_OBJECT_TRYWLOCK(object)) {
2276 				mtx_unlock(m_mtx);
2277 				VM_OBJECT_WLOCK(object);
2278 				mtx_lock(m_mtx);
2279 				if (m->object != object) {
2280 					/*
2281 					 * The page may have been freed.
2282 					 */
2283 					VM_OBJECT_WUNLOCK(object);
2284 					goto retry;
2285 				} else if (m->wire_count != 0 ||
2286 				    m->hold_count != 0) {
2287 					error = EBUSY;
2288 					goto unlock;
2289 				}
2290 			}
2291 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2292 			    ("page %p is PG_UNHOLDFREE", m));
2293 			/* Don't care: PG_NODUMP, PG_ZERO. */
2294 			if (object->type != OBJT_DEFAULT &&
2295 			    object->type != OBJT_SWAP &&
2296 			    object->type != OBJT_VNODE)
2297 				error = EINVAL;
2298 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2299 				error = EINVAL;
2300 			else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2301 				KASSERT(pmap_page_get_memattr(m) ==
2302 				    VM_MEMATTR_DEFAULT,
2303 				    ("page %p has an unexpected memattr", m));
2304 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2305 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2306 				    ("page %p has unexpected oflags", m));
2307 				/* Don't care: VPO_NOSYNC. */
2308 				if (m->valid != 0) {
2309 					/*
2310 					 * First, try to allocate a new page
2311 					 * that is above "high".  Failing
2312 					 * that, try to allocate a new page
2313 					 * that is below "m_run".  Allocate
2314 					 * the new page between the end of
2315 					 * "m_run" and "high" only as a last
2316 					 * resort.
2317 					 */
2318 					req = req_class | VM_ALLOC_NOOBJ;
2319 					if ((m->flags & PG_NODUMP) != 0)
2320 						req |= VM_ALLOC_NODUMP;
2321 					if (trunc_page(high) !=
2322 					    ~(vm_paddr_t)PAGE_MASK) {
2323 						m_new = vm_page_alloc_contig(
2324 						    NULL, 0, req, 1,
2325 						    round_page(high),
2326 						    ~(vm_paddr_t)0,
2327 						    PAGE_SIZE, 0,
2328 						    VM_MEMATTR_DEFAULT);
2329 					} else
2330 						m_new = NULL;
2331 					if (m_new == NULL) {
2332 						pa = VM_PAGE_TO_PHYS(m_run);
2333 						m_new = vm_page_alloc_contig(
2334 						    NULL, 0, req, 1,
2335 						    0, pa - 1, PAGE_SIZE, 0,
2336 						    VM_MEMATTR_DEFAULT);
2337 					}
2338 					if (m_new == NULL) {
2339 						pa += ptoa(npages);
2340 						m_new = vm_page_alloc_contig(
2341 						    NULL, 0, req, 1,
2342 						    pa, high, PAGE_SIZE, 0,
2343 						    VM_MEMATTR_DEFAULT);
2344 					}
2345 					if (m_new == NULL) {
2346 						error = ENOMEM;
2347 						goto unlock;
2348 					}
2349 					KASSERT(m_new->wire_count == 0,
2350 					    ("page %p is wired", m));
2351 
2352 					/*
2353 					 * Replace "m" with the new page.  For
2354 					 * vm_page_replace(), "m" must be busy
2355 					 * and dequeued.  Finally, change "m"
2356 					 * as if vm_page_free() was called.
2357 					 */
2358 					if (object->ref_count != 0)
2359 						pmap_remove_all(m);
2360 					m_new->aflags = m->aflags;
2361 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2362 					    ("page %p is managed", m));
2363 					m_new->oflags = m->oflags & VPO_NOSYNC;
2364 					pmap_copy_page(m, m_new);
2365 					m_new->valid = m->valid;
2366 					m_new->dirty = m->dirty;
2367 					m->flags &= ~PG_ZERO;
2368 					vm_page_xbusy(m);
2369 					vm_page_remque(m);
2370 					vm_page_replace_checked(m_new, object,
2371 					    m->pindex, m);
2372 					m->valid = 0;
2373 					vm_page_undirty(m);
2374 
2375 					/*
2376 					 * The new page must be deactivated
2377 					 * before the object is unlocked.
2378 					 */
2379 					vm_page_change_lock(m_new, &m_mtx);
2380 					vm_page_deactivate(m_new);
2381 				} else {
2382 					m->flags &= ~PG_ZERO;
2383 					vm_page_remque(m);
2384 					vm_page_remove(m);
2385 					KASSERT(m->dirty == 0,
2386 					    ("page %p is dirty", m));
2387 				}
2388 				SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2389 			} else
2390 				error = EBUSY;
2391 unlock:
2392 			VM_OBJECT_WUNLOCK(object);
2393 		} else {
2394 			mtx_lock(&vm_page_queue_free_mtx);
2395 			order = m->order;
2396 			if (order < VM_NFREEORDER) {
2397 				/*
2398 				 * The page is enqueued in the physical memory
2399 				 * allocator's free page queues.  Moreover, it
2400 				 * is the first page in a power-of-two-sized
2401 				 * run of contiguous free pages.  Jump ahead
2402 				 * to the last page within that run, and
2403 				 * continue from there.
2404 				 */
2405 				m += (1 << order) - 1;
2406 			}
2407 #if VM_NRESERVLEVEL > 0
2408 			else if (vm_reserv_is_page_free(m))
2409 				order = 0;
2410 #endif
2411 			mtx_unlock(&vm_page_queue_free_mtx);
2412 			if (order == VM_NFREEORDER)
2413 				error = EINVAL;
2414 		}
2415 	}
2416 	if (m_mtx != NULL)
2417 		mtx_unlock(m_mtx);
2418 	if ((m = SLIST_FIRST(&free)) != NULL) {
2419 		mtx_lock(&vm_page_queue_free_mtx);
2420 		do {
2421 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2422 			vm_page_free_phys(m);
2423 		} while ((m = SLIST_FIRST(&free)) != NULL);
2424 		vm_page_free_wakeup();
2425 		mtx_unlock(&vm_page_queue_free_mtx);
2426 	}
2427 	return (error);
2428 }
2429 
2430 #define	NRUNS	16
2431 
2432 CTASSERT(powerof2(NRUNS));
2433 
2434 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2435 
2436 #define	MIN_RECLAIM	8
2437 
2438 /*
2439  *	vm_page_reclaim_contig:
2440  *
2441  *	Reclaim allocated, contiguous physical memory satisfying the specified
2442  *	conditions by relocating the virtual pages using that physical memory.
2443  *	Returns true if reclamation is successful and false otherwise.  Since
2444  *	relocation requires the allocation of physical pages, reclamation may
2445  *	fail due to a shortage of free pages.  When reclamation fails, callers
2446  *	are expected to perform VM_WAIT before retrying a failed allocation
2447  *	operation, e.g., vm_page_alloc_contig().
2448  *
2449  *	The caller must always specify an allocation class through "req".
2450  *
2451  *	allocation classes:
2452  *	VM_ALLOC_NORMAL		normal process request
2453  *	VM_ALLOC_SYSTEM		system *really* needs a page
2454  *	VM_ALLOC_INTERRUPT	interrupt time request
2455  *
2456  *	The optional allocation flags are ignored.
2457  *
2458  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2459  *	must be a power of two.
2460  */
2461 bool
2462 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2463     u_long alignment, vm_paddr_t boundary)
2464 {
2465 	vm_paddr_t curr_low;
2466 	vm_page_t m_run, m_runs[NRUNS];
2467 	u_long count, reclaimed;
2468 	int error, i, options, req_class;
2469 
2470 	KASSERT(npages > 0, ("npages is 0"));
2471 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2472 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2473 	req_class = req & VM_ALLOC_CLASS_MASK;
2474 
2475 	/*
2476 	 * The page daemon is allowed to dig deeper into the free page list.
2477 	 */
2478 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2479 		req_class = VM_ALLOC_SYSTEM;
2480 
2481 	/*
2482 	 * Return if the number of free pages cannot satisfy the requested
2483 	 * allocation.
2484 	 */
2485 	count = vm_cnt.v_free_count;
2486 	if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2487 	    vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2488 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2489 		return (false);
2490 
2491 	/*
2492 	 * Scan up to three times, relaxing the restrictions ("options") on
2493 	 * the reclamation of reservations and superpages each time.
2494 	 */
2495 	for (options = VPSC_NORESERV;;) {
2496 		/*
2497 		 * Find the highest runs that satisfy the given constraints
2498 		 * and restrictions, and record them in "m_runs".
2499 		 */
2500 		curr_low = low;
2501 		count = 0;
2502 		for (;;) {
2503 			m_run = vm_phys_scan_contig(npages, curr_low, high,
2504 			    alignment, boundary, options);
2505 			if (m_run == NULL)
2506 				break;
2507 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2508 			m_runs[RUN_INDEX(count)] = m_run;
2509 			count++;
2510 		}
2511 
2512 		/*
2513 		 * Reclaim the highest runs in LIFO (descending) order until
2514 		 * the number of reclaimed pages, "reclaimed", is at least
2515 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2516 		 * reclamation is idempotent, and runs will (likely) recur
2517 		 * from one scan to the next as restrictions are relaxed.
2518 		 */
2519 		reclaimed = 0;
2520 		for (i = 0; count > 0 && i < NRUNS; i++) {
2521 			count--;
2522 			m_run = m_runs[RUN_INDEX(count)];
2523 			error = vm_page_reclaim_run(req_class, npages, m_run,
2524 			    high);
2525 			if (error == 0) {
2526 				reclaimed += npages;
2527 				if (reclaimed >= MIN_RECLAIM)
2528 					return (true);
2529 			}
2530 		}
2531 
2532 		/*
2533 		 * Either relax the restrictions on the next scan or return if
2534 		 * the last scan had no restrictions.
2535 		 */
2536 		if (options == VPSC_NORESERV)
2537 			options = VPSC_NOSUPER;
2538 		else if (options == VPSC_NOSUPER)
2539 			options = VPSC_ANY;
2540 		else if (options == VPSC_ANY)
2541 			return (reclaimed != 0);
2542 	}
2543 }
2544 
2545 /*
2546  *	vm_wait:	(also see VM_WAIT macro)
2547  *
2548  *	Sleep until free pages are available for allocation.
2549  *	- Called in various places before memory allocations.
2550  */
2551 static void
2552 _vm_wait(void)
2553 {
2554 
2555 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2556 	if (curproc == pageproc) {
2557 		vm_pageout_pages_needed = 1;
2558 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2559 		    PDROP | PSWP, "VMWait", 0);
2560 	} else {
2561 		if (__predict_false(pageproc == NULL))
2562 			panic("vm_wait in early boot");
2563 		if (!vm_pageout_wanted) {
2564 			vm_pageout_wanted = true;
2565 			wakeup(&vm_pageout_wanted);
2566 		}
2567 		vm_pages_needed = true;
2568 		msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2569 		    "vmwait", 0);
2570 	}
2571 }
2572 
2573 void
2574 vm_wait(void)
2575 {
2576 
2577 	mtx_lock(&vm_page_queue_free_mtx);
2578 	_vm_wait();
2579 }
2580 
2581 /*
2582  *	vm_page_alloc_fail:
2583  *
2584  *	Called when a page allocation function fails.  Informs the
2585  *	pagedaemon and performs the requested wait.  Requires the
2586  *	page_queue_free and object lock on entry.  Returns with the
2587  *	object lock held and free lock released.  Returns an error when
2588  *	retry is necessary.
2589  *
2590  */
2591 static int
2592 vm_page_alloc_fail(vm_object_t object, int req)
2593 {
2594 
2595 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2596 
2597 	atomic_add_int(&vm_pageout_deficit,
2598 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2599 	pagedaemon_wakeup();
2600 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
2601 		if (object != NULL)
2602 			VM_OBJECT_WUNLOCK(object);
2603 		_vm_wait();
2604 		if (object != NULL)
2605 			VM_OBJECT_WLOCK(object);
2606 		if (req & VM_ALLOC_WAITOK)
2607 			return (EAGAIN);
2608 	} else
2609 		mtx_unlock(&vm_page_queue_free_mtx);
2610 	return (0);
2611 }
2612 
2613 /*
2614  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
2615  *
2616  *	Sleep until free pages are available for allocation.
2617  *	- Called only in vm_fault so that processes page faulting
2618  *	  can be easily tracked.
2619  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2620  *	  processes will be able to grab memory first.  Do not change
2621  *	  this balance without careful testing first.
2622  */
2623 void
2624 vm_waitpfault(void)
2625 {
2626 
2627 	mtx_lock(&vm_page_queue_free_mtx);
2628 	if (!vm_pageout_wanted) {
2629 		vm_pageout_wanted = true;
2630 		wakeup(&vm_pageout_wanted);
2631 	}
2632 	vm_pages_needed = true;
2633 	msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2634 	    "pfault", 0);
2635 }
2636 
2637 struct vm_pagequeue *
2638 vm_page_pagequeue(vm_page_t m)
2639 {
2640 
2641 	if (vm_page_in_laundry(m))
2642 		return (&vm_dom[0].vmd_pagequeues[m->queue]);
2643 	else
2644 		return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2645 }
2646 
2647 /*
2648  *	vm_page_dequeue:
2649  *
2650  *	Remove the given page from its current page queue.
2651  *
2652  *	The page must be locked.
2653  */
2654 void
2655 vm_page_dequeue(vm_page_t m)
2656 {
2657 	struct vm_pagequeue *pq;
2658 
2659 	vm_page_assert_locked(m);
2660 	KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2661 	    m));
2662 	pq = vm_page_pagequeue(m);
2663 	vm_pagequeue_lock(pq);
2664 	m->queue = PQ_NONE;
2665 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2666 	vm_pagequeue_cnt_dec(pq);
2667 	vm_pagequeue_unlock(pq);
2668 }
2669 
2670 /*
2671  *	vm_page_dequeue_locked:
2672  *
2673  *	Remove the given page from its current page queue.
2674  *
2675  *	The page and page queue must be locked.
2676  */
2677 void
2678 vm_page_dequeue_locked(vm_page_t m)
2679 {
2680 	struct vm_pagequeue *pq;
2681 
2682 	vm_page_lock_assert(m, MA_OWNED);
2683 	pq = vm_page_pagequeue(m);
2684 	vm_pagequeue_assert_locked(pq);
2685 	m->queue = PQ_NONE;
2686 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2687 	vm_pagequeue_cnt_dec(pq);
2688 }
2689 
2690 /*
2691  *	vm_page_enqueue:
2692  *
2693  *	Add the given page to the specified page queue.
2694  *
2695  *	The page must be locked.
2696  */
2697 static void
2698 vm_page_enqueue(uint8_t queue, vm_page_t m)
2699 {
2700 	struct vm_pagequeue *pq;
2701 
2702 	vm_page_lock_assert(m, MA_OWNED);
2703 	KASSERT(queue < PQ_COUNT,
2704 	    ("vm_page_enqueue: invalid queue %u request for page %p",
2705 	    queue, m));
2706 	if (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE)
2707 		pq = &vm_dom[0].vmd_pagequeues[queue];
2708 	else
2709 		pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2710 	vm_pagequeue_lock(pq);
2711 	m->queue = queue;
2712 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2713 	vm_pagequeue_cnt_inc(pq);
2714 	vm_pagequeue_unlock(pq);
2715 }
2716 
2717 /*
2718  *	vm_page_requeue:
2719  *
2720  *	Move the given page to the tail of its current page queue.
2721  *
2722  *	The page must be locked.
2723  */
2724 void
2725 vm_page_requeue(vm_page_t m)
2726 {
2727 	struct vm_pagequeue *pq;
2728 
2729 	vm_page_lock_assert(m, MA_OWNED);
2730 	KASSERT(m->queue != PQ_NONE,
2731 	    ("vm_page_requeue: page %p is not queued", m));
2732 	pq = vm_page_pagequeue(m);
2733 	vm_pagequeue_lock(pq);
2734 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2735 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2736 	vm_pagequeue_unlock(pq);
2737 }
2738 
2739 /*
2740  *	vm_page_requeue_locked:
2741  *
2742  *	Move the given page to the tail of its current page queue.
2743  *
2744  *	The page queue must be locked.
2745  */
2746 void
2747 vm_page_requeue_locked(vm_page_t m)
2748 {
2749 	struct vm_pagequeue *pq;
2750 
2751 	KASSERT(m->queue != PQ_NONE,
2752 	    ("vm_page_requeue_locked: page %p is not queued", m));
2753 	pq = vm_page_pagequeue(m);
2754 	vm_pagequeue_assert_locked(pq);
2755 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2756 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2757 }
2758 
2759 /*
2760  *	vm_page_activate:
2761  *
2762  *	Put the specified page on the active list (if appropriate).
2763  *	Ensure that act_count is at least ACT_INIT but do not otherwise
2764  *	mess with it.
2765  *
2766  *	The page must be locked.
2767  */
2768 void
2769 vm_page_activate(vm_page_t m)
2770 {
2771 	int queue;
2772 
2773 	vm_page_lock_assert(m, MA_OWNED);
2774 	if ((queue = m->queue) != PQ_ACTIVE) {
2775 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2776 			if (m->act_count < ACT_INIT)
2777 				m->act_count = ACT_INIT;
2778 			if (queue != PQ_NONE)
2779 				vm_page_dequeue(m);
2780 			vm_page_enqueue(PQ_ACTIVE, m);
2781 		} else
2782 			KASSERT(queue == PQ_NONE,
2783 			    ("vm_page_activate: wired page %p is queued", m));
2784 	} else {
2785 		if (m->act_count < ACT_INIT)
2786 			m->act_count = ACT_INIT;
2787 	}
2788 }
2789 
2790 /*
2791  *	vm_page_free_wakeup:
2792  *
2793  *	Helper routine for vm_page_free_toq().  This routine is called
2794  *	when a page is added to the free queues.
2795  *
2796  *	The page queues must be locked.
2797  */
2798 static void
2799 vm_page_free_wakeup(void)
2800 {
2801 
2802 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2803 	/*
2804 	 * if pageout daemon needs pages, then tell it that there are
2805 	 * some free.
2806 	 */
2807 	if (vm_pageout_pages_needed &&
2808 	    vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2809 		wakeup(&vm_pageout_pages_needed);
2810 		vm_pageout_pages_needed = 0;
2811 	}
2812 	/*
2813 	 * wakeup processes that are waiting on memory if we hit a
2814 	 * high water mark. And wakeup scheduler process if we have
2815 	 * lots of memory. this process will swapin processes.
2816 	 */
2817 	if (vm_pages_needed && !vm_page_count_min()) {
2818 		vm_pages_needed = false;
2819 		wakeup(&vm_cnt.v_free_count);
2820 	}
2821 }
2822 
2823 /*
2824  *	vm_page_free_prep:
2825  *
2826  *	Prepares the given page to be put on the free list,
2827  *	disassociating it from any VM object. The caller may return
2828  *	the page to the free list only if this function returns true.
2829  *
2830  *	The object must be locked.  The page must be locked if it is
2831  *	managed.  For a queued managed page, the pagequeue_locked
2832  *	argument specifies whether the page queue is already locked.
2833  */
2834 bool
2835 vm_page_free_prep(vm_page_t m, bool pagequeue_locked)
2836 {
2837 
2838 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
2839 	if ((m->flags & PG_ZERO) != 0) {
2840 		uint64_t *p;
2841 		int i;
2842 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
2843 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
2844 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
2845 			    m, i, (uintmax_t)*p));
2846 	}
2847 #endif
2848 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2849 		vm_page_lock_assert(m, MA_OWNED);
2850 		KASSERT(!pmap_page_is_mapped(m),
2851 		    ("vm_page_free_toq: freeing mapped page %p", m));
2852 	} else
2853 		KASSERT(m->queue == PQ_NONE,
2854 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
2855 	VM_CNT_INC(v_tfree);
2856 
2857 	if (vm_page_sbusied(m))
2858 		panic("vm_page_free: freeing busy page %p", m);
2859 
2860 	vm_page_remove(m);
2861 
2862 	/*
2863 	 * If fictitious remove object association and
2864 	 * return.
2865 	 */
2866 	if ((m->flags & PG_FICTITIOUS) != 0) {
2867 		KASSERT(m->wire_count == 1,
2868 		    ("fictitious page %p is not wired", m));
2869 		KASSERT(m->queue == PQ_NONE,
2870 		    ("fictitious page %p is queued", m));
2871 		return (false);
2872 	}
2873 
2874 	if (m->queue != PQ_NONE) {
2875 		if (pagequeue_locked)
2876 			vm_page_dequeue_locked(m);
2877 		else
2878 			vm_page_dequeue(m);
2879 	}
2880 	m->valid = 0;
2881 	vm_page_undirty(m);
2882 
2883 	if (m->wire_count != 0)
2884 		panic("vm_page_free: freeing wired page %p", m);
2885 	if (m->hold_count != 0) {
2886 		m->flags &= ~PG_ZERO;
2887 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2888 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2889 		m->flags |= PG_UNHOLDFREE;
2890 		return (false);
2891 	}
2892 
2893 	/*
2894 	 * Restore the default memory attribute to the page.
2895 	 */
2896 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2897 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2898 
2899 	return (true);
2900 }
2901 
2902 /*
2903  * Insert the page into the physical memory allocator's free page
2904  * queues.  This is the last step to free a page.
2905  */
2906 static void
2907 vm_page_free_phys(vm_page_t m)
2908 {
2909 
2910 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2911 
2912 	vm_phys_freecnt_adj(m, 1);
2913 #if VM_NRESERVLEVEL > 0
2914 	if (!vm_reserv_free_page(m))
2915 #endif
2916 		vm_phys_free_pages(m, 0);
2917 }
2918 
2919 void
2920 vm_page_free_phys_pglist(struct pglist *tq)
2921 {
2922 	vm_page_t m;
2923 
2924 	if (TAILQ_EMPTY(tq))
2925 		return;
2926 	mtx_lock(&vm_page_queue_free_mtx);
2927 	TAILQ_FOREACH(m, tq, listq)
2928 		vm_page_free_phys(m);
2929 	vm_page_free_wakeup();
2930 	mtx_unlock(&vm_page_queue_free_mtx);
2931 }
2932 
2933 /*
2934  *	vm_page_free_toq:
2935  *
2936  *	Returns the given page to the free list, disassociating it
2937  *	from any VM object.
2938  *
2939  *	The object must be locked.  The page must be locked if it is
2940  *	managed.
2941  */
2942 void
2943 vm_page_free_toq(vm_page_t m)
2944 {
2945 
2946 	if (!vm_page_free_prep(m, false))
2947 		return;
2948 	mtx_lock(&vm_page_queue_free_mtx);
2949 	vm_page_free_phys(m);
2950 	vm_page_free_wakeup();
2951 	mtx_unlock(&vm_page_queue_free_mtx);
2952 }
2953 
2954 /*
2955  *	vm_page_wire:
2956  *
2957  *	Mark this page as wired down by yet
2958  *	another map, removing it from paging queues
2959  *	as necessary.
2960  *
2961  *	If the page is fictitious, then its wire count must remain one.
2962  *
2963  *	The page must be locked.
2964  */
2965 void
2966 vm_page_wire(vm_page_t m)
2967 {
2968 
2969 	/*
2970 	 * Only bump the wire statistics if the page is not already wired,
2971 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2972 	 * it is already off the queues).
2973 	 */
2974 	vm_page_lock_assert(m, MA_OWNED);
2975 	if ((m->flags & PG_FICTITIOUS) != 0) {
2976 		KASSERT(m->wire_count == 1,
2977 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2978 		    m));
2979 		return;
2980 	}
2981 	if (m->wire_count == 0) {
2982 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2983 		    m->queue == PQ_NONE,
2984 		    ("vm_page_wire: unmanaged page %p is queued", m));
2985 		vm_page_remque(m);
2986 		atomic_add_int(&vm_cnt.v_wire_count, 1);
2987 	}
2988 	m->wire_count++;
2989 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2990 }
2991 
2992 /*
2993  * vm_page_unwire:
2994  *
2995  * Release one wiring of the specified page, potentially allowing it to be
2996  * paged out.  Returns TRUE if the number of wirings transitions to zero and
2997  * FALSE otherwise.
2998  *
2999  * Only managed pages belonging to an object can be paged out.  If the number
3000  * of wirings transitions to zero and the page is eligible for page out, then
3001  * the page is added to the specified paging queue (unless PQ_NONE is
3002  * specified).
3003  *
3004  * If a page is fictitious, then its wire count must always be one.
3005  *
3006  * A managed page must be locked.
3007  */
3008 boolean_t
3009 vm_page_unwire(vm_page_t m, uint8_t queue)
3010 {
3011 
3012 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3013 	    ("vm_page_unwire: invalid queue %u request for page %p",
3014 	    queue, m));
3015 	if ((m->oflags & VPO_UNMANAGED) == 0)
3016 		vm_page_assert_locked(m);
3017 	if ((m->flags & PG_FICTITIOUS) != 0) {
3018 		KASSERT(m->wire_count == 1,
3019 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3020 		return (FALSE);
3021 	}
3022 	if (m->wire_count > 0) {
3023 		m->wire_count--;
3024 		if (m->wire_count == 0) {
3025 			atomic_subtract_int(&vm_cnt.v_wire_count, 1);
3026 			if ((m->oflags & VPO_UNMANAGED) == 0 &&
3027 			    m->object != NULL && queue != PQ_NONE)
3028 				vm_page_enqueue(queue, m);
3029 			return (TRUE);
3030 		} else
3031 			return (FALSE);
3032 	} else
3033 		panic("vm_page_unwire: page %p's wire count is zero", m);
3034 }
3035 
3036 /*
3037  * Move the specified page to the inactive queue.
3038  *
3039  * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
3040  * queue.  However, setting "noreuse" to TRUE will accelerate the specified
3041  * page's reclamation, but it will not unmap the page from any address space.
3042  * This is implemented by inserting the page near the head of the inactive
3043  * queue, using a marker page to guide FIFO insertion ordering.
3044  *
3045  * The page must be locked.
3046  */
3047 static inline void
3048 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3049 {
3050 	struct vm_pagequeue *pq;
3051 	int queue;
3052 
3053 	vm_page_assert_locked(m);
3054 
3055 	/*
3056 	 * Ignore if the page is already inactive, unless it is unlikely to be
3057 	 * reactivated.
3058 	 */
3059 	if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
3060 		return;
3061 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3062 		pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3063 		/* Avoid multiple acquisitions of the inactive queue lock. */
3064 		if (queue == PQ_INACTIVE) {
3065 			vm_pagequeue_lock(pq);
3066 			vm_page_dequeue_locked(m);
3067 		} else {
3068 			if (queue != PQ_NONE)
3069 				vm_page_dequeue(m);
3070 			vm_pagequeue_lock(pq);
3071 		}
3072 		m->queue = PQ_INACTIVE;
3073 		if (noreuse)
3074 			TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
3075 			    m, plinks.q);
3076 		else
3077 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3078 		vm_pagequeue_cnt_inc(pq);
3079 		vm_pagequeue_unlock(pq);
3080 	}
3081 }
3082 
3083 /*
3084  * Move the specified page to the inactive queue.
3085  *
3086  * The page must be locked.
3087  */
3088 void
3089 vm_page_deactivate(vm_page_t m)
3090 {
3091 
3092 	_vm_page_deactivate(m, FALSE);
3093 }
3094 
3095 /*
3096  * Move the specified page to the inactive queue with the expectation
3097  * that it is unlikely to be reused.
3098  *
3099  * The page must be locked.
3100  */
3101 void
3102 vm_page_deactivate_noreuse(vm_page_t m)
3103 {
3104 
3105 	_vm_page_deactivate(m, TRUE);
3106 }
3107 
3108 /*
3109  * vm_page_launder
3110  *
3111  * 	Put a page in the laundry.
3112  */
3113 void
3114 vm_page_launder(vm_page_t m)
3115 {
3116 	int queue;
3117 
3118 	vm_page_assert_locked(m);
3119 	if ((queue = m->queue) != PQ_LAUNDRY) {
3120 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3121 			if (queue != PQ_NONE)
3122 				vm_page_dequeue(m);
3123 			vm_page_enqueue(PQ_LAUNDRY, m);
3124 		} else
3125 			KASSERT(queue == PQ_NONE,
3126 			    ("wired page %p is queued", m));
3127 	}
3128 }
3129 
3130 /*
3131  * vm_page_unswappable
3132  *
3133  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3134  */
3135 void
3136 vm_page_unswappable(vm_page_t m)
3137 {
3138 
3139 	vm_page_assert_locked(m);
3140 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3141 	    ("page %p already unswappable", m));
3142 	if (m->queue != PQ_NONE)
3143 		vm_page_dequeue(m);
3144 	vm_page_enqueue(PQ_UNSWAPPABLE, m);
3145 }
3146 
3147 /*
3148  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3149  * if the page is freed and false otherwise.
3150  *
3151  * The page must be managed.  The page and its containing object must be
3152  * locked.
3153  */
3154 bool
3155 vm_page_try_to_free(vm_page_t m)
3156 {
3157 
3158 	vm_page_assert_locked(m);
3159 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3160 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3161 	if (m->dirty != 0 || m->hold_count != 0 || m->wire_count != 0 ||
3162 	    vm_page_busied(m))
3163 		return (false);
3164 	if (m->object->ref_count != 0) {
3165 		pmap_remove_all(m);
3166 		if (m->dirty != 0)
3167 			return (false);
3168 	}
3169 	vm_page_free(m);
3170 	return (true);
3171 }
3172 
3173 /*
3174  * vm_page_advise
3175  *
3176  * 	Apply the specified advice to the given page.
3177  *
3178  *	The object and page must be locked.
3179  */
3180 void
3181 vm_page_advise(vm_page_t m, int advice)
3182 {
3183 
3184 	vm_page_assert_locked(m);
3185 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3186 	if (advice == MADV_FREE)
3187 		/*
3188 		 * Mark the page clean.  This will allow the page to be freed
3189 		 * without first paging it out.  MADV_FREE pages are often
3190 		 * quickly reused by malloc(3), so we do not do anything that
3191 		 * would result in a page fault on a later access.
3192 		 */
3193 		vm_page_undirty(m);
3194 	else if (advice != MADV_DONTNEED) {
3195 		if (advice == MADV_WILLNEED)
3196 			vm_page_activate(m);
3197 		return;
3198 	}
3199 
3200 	/*
3201 	 * Clear any references to the page.  Otherwise, the page daemon will
3202 	 * immediately reactivate the page.
3203 	 */
3204 	vm_page_aflag_clear(m, PGA_REFERENCED);
3205 
3206 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3207 		vm_page_dirty(m);
3208 
3209 	/*
3210 	 * Place clean pages near the head of the inactive queue rather than
3211 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3212 	 * the page will be reused quickly.  Dirty pages not already in the
3213 	 * laundry are moved there.
3214 	 */
3215 	if (m->dirty == 0)
3216 		vm_page_deactivate_noreuse(m);
3217 	else
3218 		vm_page_launder(m);
3219 }
3220 
3221 /*
3222  * Grab a page, waiting until we are waken up due to the page
3223  * changing state.  We keep on waiting, if the page continues
3224  * to be in the object.  If the page doesn't exist, first allocate it
3225  * and then conditionally zero it.
3226  *
3227  * This routine may sleep.
3228  *
3229  * The object must be locked on entry.  The lock will, however, be released
3230  * and reacquired if the routine sleeps.
3231  */
3232 vm_page_t
3233 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3234 {
3235 	vm_page_t m;
3236 	int sleep;
3237 	int pflags;
3238 
3239 	VM_OBJECT_ASSERT_WLOCKED(object);
3240 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3241 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3242 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3243 	pflags = allocflags &
3244 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3245 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3246 		pflags |= VM_ALLOC_WAITFAIL;
3247 retrylookup:
3248 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3249 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3250 		    vm_page_xbusied(m) : vm_page_busied(m);
3251 		if (sleep) {
3252 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3253 				return (NULL);
3254 			/*
3255 			 * Reference the page before unlocking and
3256 			 * sleeping so that the page daemon is less
3257 			 * likely to reclaim it.
3258 			 */
3259 			vm_page_aflag_set(m, PGA_REFERENCED);
3260 			vm_page_lock(m);
3261 			VM_OBJECT_WUNLOCK(object);
3262 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3263 			    VM_ALLOC_IGN_SBUSY) != 0);
3264 			VM_OBJECT_WLOCK(object);
3265 			goto retrylookup;
3266 		} else {
3267 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3268 				vm_page_lock(m);
3269 				vm_page_wire(m);
3270 				vm_page_unlock(m);
3271 			}
3272 			if ((allocflags &
3273 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3274 				vm_page_xbusy(m);
3275 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3276 				vm_page_sbusy(m);
3277 			return (m);
3278 		}
3279 	}
3280 	m = vm_page_alloc(object, pindex, pflags);
3281 	if (m == NULL) {
3282 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3283 			return (NULL);
3284 		goto retrylookup;
3285 	}
3286 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3287 		pmap_zero_page(m);
3288 	return (m);
3289 }
3290 
3291 /*
3292  * Return the specified range of pages from the given object.  For each
3293  * page offset within the range, if a page already exists within the object
3294  * at that offset and it is busy, then wait for it to change state.  If,
3295  * instead, the page doesn't exist, then allocate it.
3296  *
3297  * The caller must always specify an allocation class.
3298  *
3299  * allocation classes:
3300  *	VM_ALLOC_NORMAL		normal process request
3301  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3302  *
3303  * The caller must always specify that the pages are to be busied and/or
3304  * wired.
3305  *
3306  * optional allocation flags:
3307  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3308  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3309  *	VM_ALLOC_NOWAIT		do not sleep
3310  *	VM_ALLOC_SBUSY		set page to sbusy state
3311  *	VM_ALLOC_WIRED		wire the pages
3312  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3313  *
3314  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3315  * may return a partial prefix of the requested range.
3316  */
3317 int
3318 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3319     vm_page_t *ma, int count)
3320 {
3321 	vm_page_t m, mpred;
3322 	int pflags;
3323 	int i;
3324 	bool sleep;
3325 
3326 	VM_OBJECT_ASSERT_WLOCKED(object);
3327 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3328 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3329 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3330 	    (allocflags & VM_ALLOC_WIRED) != 0,
3331 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3332 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3333 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3334 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3335 	if (count == 0)
3336 		return (0);
3337 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3338 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3339 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3340 		pflags |= VM_ALLOC_WAITFAIL;
3341 	i = 0;
3342 retrylookup:
3343 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3344 	if (m == NULL || m->pindex != pindex + i) {
3345 		mpred = m;
3346 		m = NULL;
3347 	} else
3348 		mpred = TAILQ_PREV(m, pglist, listq);
3349 	for (; i < count; i++) {
3350 		if (m != NULL) {
3351 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3352 			    vm_page_xbusied(m) : vm_page_busied(m);
3353 			if (sleep) {
3354 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3355 					break;
3356 				/*
3357 				 * Reference the page before unlocking and
3358 				 * sleeping so that the page daemon is less
3359 				 * likely to reclaim it.
3360 				 */
3361 				vm_page_aflag_set(m, PGA_REFERENCED);
3362 				vm_page_lock(m);
3363 				VM_OBJECT_WUNLOCK(object);
3364 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
3365 				    VM_ALLOC_IGN_SBUSY) != 0);
3366 				VM_OBJECT_WLOCK(object);
3367 				goto retrylookup;
3368 			}
3369 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3370 				vm_page_lock(m);
3371 				vm_page_wire(m);
3372 				vm_page_unlock(m);
3373 			}
3374 			if ((allocflags & (VM_ALLOC_NOBUSY |
3375 			    VM_ALLOC_SBUSY)) == 0)
3376 				vm_page_xbusy(m);
3377 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3378 				vm_page_sbusy(m);
3379 		} else {
3380 			m = vm_page_alloc_after(object, pindex + i,
3381 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
3382 			if (m == NULL) {
3383 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3384 					break;
3385 				goto retrylookup;
3386 			}
3387 		}
3388 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3389 			if ((m->flags & PG_ZERO) == 0)
3390 				pmap_zero_page(m);
3391 			m->valid = VM_PAGE_BITS_ALL;
3392 		}
3393 		ma[i] = mpred = m;
3394 		m = vm_page_next(m);
3395 	}
3396 	return (i);
3397 }
3398 
3399 /*
3400  * Mapping function for valid or dirty bits in a page.
3401  *
3402  * Inputs are required to range within a page.
3403  */
3404 vm_page_bits_t
3405 vm_page_bits(int base, int size)
3406 {
3407 	int first_bit;
3408 	int last_bit;
3409 
3410 	KASSERT(
3411 	    base + size <= PAGE_SIZE,
3412 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
3413 	);
3414 
3415 	if (size == 0)		/* handle degenerate case */
3416 		return (0);
3417 
3418 	first_bit = base >> DEV_BSHIFT;
3419 	last_bit = (base + size - 1) >> DEV_BSHIFT;
3420 
3421 	return (((vm_page_bits_t)2 << last_bit) -
3422 	    ((vm_page_bits_t)1 << first_bit));
3423 }
3424 
3425 /*
3426  *	vm_page_set_valid_range:
3427  *
3428  *	Sets portions of a page valid.  The arguments are expected
3429  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3430  *	of any partial chunks touched by the range.  The invalid portion of
3431  *	such chunks will be zeroed.
3432  *
3433  *	(base + size) must be less then or equal to PAGE_SIZE.
3434  */
3435 void
3436 vm_page_set_valid_range(vm_page_t m, int base, int size)
3437 {
3438 	int endoff, frag;
3439 
3440 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3441 	if (size == 0)	/* handle degenerate case */
3442 		return;
3443 
3444 	/*
3445 	 * If the base is not DEV_BSIZE aligned and the valid
3446 	 * bit is clear, we have to zero out a portion of the
3447 	 * first block.
3448 	 */
3449 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3450 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3451 		pmap_zero_page_area(m, frag, base - frag);
3452 
3453 	/*
3454 	 * If the ending offset is not DEV_BSIZE aligned and the
3455 	 * valid bit is clear, we have to zero out a portion of
3456 	 * the last block.
3457 	 */
3458 	endoff = base + size;
3459 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3460 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3461 		pmap_zero_page_area(m, endoff,
3462 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3463 
3464 	/*
3465 	 * Assert that no previously invalid block that is now being validated
3466 	 * is already dirty.
3467 	 */
3468 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3469 	    ("vm_page_set_valid_range: page %p is dirty", m));
3470 
3471 	/*
3472 	 * Set valid bits inclusive of any overlap.
3473 	 */
3474 	m->valid |= vm_page_bits(base, size);
3475 }
3476 
3477 /*
3478  * Clear the given bits from the specified page's dirty field.
3479  */
3480 static __inline void
3481 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3482 {
3483 	uintptr_t addr;
3484 #if PAGE_SIZE < 16384
3485 	int shift;
3486 #endif
3487 
3488 	/*
3489 	 * If the object is locked and the page is neither exclusive busy nor
3490 	 * write mapped, then the page's dirty field cannot possibly be
3491 	 * set by a concurrent pmap operation.
3492 	 */
3493 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3494 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3495 		m->dirty &= ~pagebits;
3496 	else {
3497 		/*
3498 		 * The pmap layer can call vm_page_dirty() without
3499 		 * holding a distinguished lock.  The combination of
3500 		 * the object's lock and an atomic operation suffice
3501 		 * to guarantee consistency of the page dirty field.
3502 		 *
3503 		 * For PAGE_SIZE == 32768 case, compiler already
3504 		 * properly aligns the dirty field, so no forcible
3505 		 * alignment is needed. Only require existence of
3506 		 * atomic_clear_64 when page size is 32768.
3507 		 */
3508 		addr = (uintptr_t)&m->dirty;
3509 #if PAGE_SIZE == 32768
3510 		atomic_clear_64((uint64_t *)addr, pagebits);
3511 #elif PAGE_SIZE == 16384
3512 		atomic_clear_32((uint32_t *)addr, pagebits);
3513 #else		/* PAGE_SIZE <= 8192 */
3514 		/*
3515 		 * Use a trick to perform a 32-bit atomic on the
3516 		 * containing aligned word, to not depend on the existence
3517 		 * of atomic_clear_{8, 16}.
3518 		 */
3519 		shift = addr & (sizeof(uint32_t) - 1);
3520 #if BYTE_ORDER == BIG_ENDIAN
3521 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3522 #else
3523 		shift *= NBBY;
3524 #endif
3525 		addr &= ~(sizeof(uint32_t) - 1);
3526 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
3527 #endif		/* PAGE_SIZE */
3528 	}
3529 }
3530 
3531 /*
3532  *	vm_page_set_validclean:
3533  *
3534  *	Sets portions of a page valid and clean.  The arguments are expected
3535  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3536  *	of any partial chunks touched by the range.  The invalid portion of
3537  *	such chunks will be zero'd.
3538  *
3539  *	(base + size) must be less then or equal to PAGE_SIZE.
3540  */
3541 void
3542 vm_page_set_validclean(vm_page_t m, int base, int size)
3543 {
3544 	vm_page_bits_t oldvalid, pagebits;
3545 	int endoff, frag;
3546 
3547 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3548 	if (size == 0)	/* handle degenerate case */
3549 		return;
3550 
3551 	/*
3552 	 * If the base is not DEV_BSIZE aligned and the valid
3553 	 * bit is clear, we have to zero out a portion of the
3554 	 * first block.
3555 	 */
3556 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3557 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3558 		pmap_zero_page_area(m, frag, base - frag);
3559 
3560 	/*
3561 	 * If the ending offset is not DEV_BSIZE aligned and the
3562 	 * valid bit is clear, we have to zero out a portion of
3563 	 * the last block.
3564 	 */
3565 	endoff = base + size;
3566 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3567 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3568 		pmap_zero_page_area(m, endoff,
3569 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3570 
3571 	/*
3572 	 * Set valid, clear dirty bits.  If validating the entire
3573 	 * page we can safely clear the pmap modify bit.  We also
3574 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
3575 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
3576 	 * be set again.
3577 	 *
3578 	 * We set valid bits inclusive of any overlap, but we can only
3579 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
3580 	 * the range.
3581 	 */
3582 	oldvalid = m->valid;
3583 	pagebits = vm_page_bits(base, size);
3584 	m->valid |= pagebits;
3585 #if 0	/* NOT YET */
3586 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3587 		frag = DEV_BSIZE - frag;
3588 		base += frag;
3589 		size -= frag;
3590 		if (size < 0)
3591 			size = 0;
3592 	}
3593 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3594 #endif
3595 	if (base == 0 && size == PAGE_SIZE) {
3596 		/*
3597 		 * The page can only be modified within the pmap if it is
3598 		 * mapped, and it can only be mapped if it was previously
3599 		 * fully valid.
3600 		 */
3601 		if (oldvalid == VM_PAGE_BITS_ALL)
3602 			/*
3603 			 * Perform the pmap_clear_modify() first.  Otherwise,
3604 			 * a concurrent pmap operation, such as
3605 			 * pmap_protect(), could clear a modification in the
3606 			 * pmap and set the dirty field on the page before
3607 			 * pmap_clear_modify() had begun and after the dirty
3608 			 * field was cleared here.
3609 			 */
3610 			pmap_clear_modify(m);
3611 		m->dirty = 0;
3612 		m->oflags &= ~VPO_NOSYNC;
3613 	} else if (oldvalid != VM_PAGE_BITS_ALL)
3614 		m->dirty &= ~pagebits;
3615 	else
3616 		vm_page_clear_dirty_mask(m, pagebits);
3617 }
3618 
3619 void
3620 vm_page_clear_dirty(vm_page_t m, int base, int size)
3621 {
3622 
3623 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3624 }
3625 
3626 /*
3627  *	vm_page_set_invalid:
3628  *
3629  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
3630  *	valid and dirty bits for the effected areas are cleared.
3631  */
3632 void
3633 vm_page_set_invalid(vm_page_t m, int base, int size)
3634 {
3635 	vm_page_bits_t bits;
3636 	vm_object_t object;
3637 
3638 	object = m->object;
3639 	VM_OBJECT_ASSERT_WLOCKED(object);
3640 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3641 	    size >= object->un_pager.vnp.vnp_size)
3642 		bits = VM_PAGE_BITS_ALL;
3643 	else
3644 		bits = vm_page_bits(base, size);
3645 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3646 	    bits != 0)
3647 		pmap_remove_all(m);
3648 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3649 	    !pmap_page_is_mapped(m),
3650 	    ("vm_page_set_invalid: page %p is mapped", m));
3651 	m->valid &= ~bits;
3652 	m->dirty &= ~bits;
3653 }
3654 
3655 /*
3656  * vm_page_zero_invalid()
3657  *
3658  *	The kernel assumes that the invalid portions of a page contain
3659  *	garbage, but such pages can be mapped into memory by user code.
3660  *	When this occurs, we must zero out the non-valid portions of the
3661  *	page so user code sees what it expects.
3662  *
3663  *	Pages are most often semi-valid when the end of a file is mapped
3664  *	into memory and the file's size is not page aligned.
3665  */
3666 void
3667 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3668 {
3669 	int b;
3670 	int i;
3671 
3672 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3673 	/*
3674 	 * Scan the valid bits looking for invalid sections that
3675 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
3676 	 * valid bit may be set ) have already been zeroed by
3677 	 * vm_page_set_validclean().
3678 	 */
3679 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3680 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
3681 		    (m->valid & ((vm_page_bits_t)1 << i))) {
3682 			if (i > b) {
3683 				pmap_zero_page_area(m,
3684 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3685 			}
3686 			b = i + 1;
3687 		}
3688 	}
3689 
3690 	/*
3691 	 * setvalid is TRUE when we can safely set the zero'd areas
3692 	 * as being valid.  We can do this if there are no cache consistancy
3693 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
3694 	 */
3695 	if (setvalid)
3696 		m->valid = VM_PAGE_BITS_ALL;
3697 }
3698 
3699 /*
3700  *	vm_page_is_valid:
3701  *
3702  *	Is (partial) page valid?  Note that the case where size == 0
3703  *	will return FALSE in the degenerate case where the page is
3704  *	entirely invalid, and TRUE otherwise.
3705  */
3706 int
3707 vm_page_is_valid(vm_page_t m, int base, int size)
3708 {
3709 	vm_page_bits_t bits;
3710 
3711 	VM_OBJECT_ASSERT_LOCKED(m->object);
3712 	bits = vm_page_bits(base, size);
3713 	return (m->valid != 0 && (m->valid & bits) == bits);
3714 }
3715 
3716 /*
3717  * Returns true if all of the specified predicates are true for the entire
3718  * (super)page and false otherwise.
3719  */
3720 bool
3721 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
3722 {
3723 	vm_object_t object;
3724 	int i, npages;
3725 
3726 	object = m->object;
3727 	VM_OBJECT_ASSERT_LOCKED(object);
3728 	npages = atop(pagesizes[m->psind]);
3729 
3730 	/*
3731 	 * The physically contiguous pages that make up a superpage, i.e., a
3732 	 * page with a page size index ("psind") greater than zero, will
3733 	 * occupy adjacent entries in vm_page_array[].
3734 	 */
3735 	for (i = 0; i < npages; i++) {
3736 		/* Always test object consistency, including "skip_m". */
3737 		if (m[i].object != object)
3738 			return (false);
3739 		if (&m[i] == skip_m)
3740 			continue;
3741 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
3742 			return (false);
3743 		if ((flags & PS_ALL_DIRTY) != 0) {
3744 			/*
3745 			 * Calling vm_page_test_dirty() or pmap_is_modified()
3746 			 * might stop this case from spuriously returning
3747 			 * "false".  However, that would require a write lock
3748 			 * on the object containing "m[i]".
3749 			 */
3750 			if (m[i].dirty != VM_PAGE_BITS_ALL)
3751 				return (false);
3752 		}
3753 		if ((flags & PS_ALL_VALID) != 0 &&
3754 		    m[i].valid != VM_PAGE_BITS_ALL)
3755 			return (false);
3756 	}
3757 	return (true);
3758 }
3759 
3760 /*
3761  * Set the page's dirty bits if the page is modified.
3762  */
3763 void
3764 vm_page_test_dirty(vm_page_t m)
3765 {
3766 
3767 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3768 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3769 		vm_page_dirty(m);
3770 }
3771 
3772 void
3773 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3774 {
3775 
3776 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3777 }
3778 
3779 void
3780 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3781 {
3782 
3783 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3784 }
3785 
3786 int
3787 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3788 {
3789 
3790 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3791 }
3792 
3793 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3794 void
3795 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3796 {
3797 
3798 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3799 }
3800 
3801 void
3802 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3803 {
3804 
3805 	mtx_assert_(vm_page_lockptr(m), a, file, line);
3806 }
3807 #endif
3808 
3809 #ifdef INVARIANTS
3810 void
3811 vm_page_object_lock_assert(vm_page_t m)
3812 {
3813 
3814 	/*
3815 	 * Certain of the page's fields may only be modified by the
3816 	 * holder of the containing object's lock or the exclusive busy.
3817 	 * holder.  Unfortunately, the holder of the write busy is
3818 	 * not recorded, and thus cannot be checked here.
3819 	 */
3820 	if (m->object != NULL && !vm_page_xbusied(m))
3821 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3822 }
3823 
3824 void
3825 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3826 {
3827 
3828 	if ((bits & PGA_WRITEABLE) == 0)
3829 		return;
3830 
3831 	/*
3832 	 * The PGA_WRITEABLE flag can only be set if the page is
3833 	 * managed, is exclusively busied or the object is locked.
3834 	 * Currently, this flag is only set by pmap_enter().
3835 	 */
3836 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3837 	    ("PGA_WRITEABLE on unmanaged page"));
3838 	if (!vm_page_xbusied(m))
3839 		VM_OBJECT_ASSERT_LOCKED(m->object);
3840 }
3841 #endif
3842 
3843 #include "opt_ddb.h"
3844 #ifdef DDB
3845 #include <sys/kernel.h>
3846 
3847 #include <ddb/ddb.h>
3848 
3849 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3850 {
3851 
3852 	db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3853 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3854 	db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3855 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3856 	db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3857 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3858 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3859 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3860 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3861 }
3862 
3863 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3864 {
3865 	int dom;
3866 
3867 	db_printf("pq_free %d\n", vm_cnt.v_free_count);
3868 	for (dom = 0; dom < vm_ndomains; dom++) {
3869 		db_printf(
3870     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
3871 		    dom,
3872 		    vm_dom[dom].vmd_page_count,
3873 		    vm_dom[dom].vmd_free_count,
3874 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3875 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3876 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
3877 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
3878 	}
3879 }
3880 
3881 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3882 {
3883 	vm_page_t m;
3884 	boolean_t phys;
3885 
3886 	if (!have_addr) {
3887 		db_printf("show pginfo addr\n");
3888 		return;
3889 	}
3890 
3891 	phys = strchr(modif, 'p') != NULL;
3892 	if (phys)
3893 		m = PHYS_TO_VM_PAGE(addr);
3894 	else
3895 		m = (vm_page_t)addr;
3896 	db_printf(
3897     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3898     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3899 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3900 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3901 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3902 }
3903 #endif /* DDB */
3904