xref: /freebsd/sys/vm/vm_page.c (revision f1f890804985a1043da42a5def13c79dc005f5e9)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * The Mach Operating System project at Carnegie-Mellon University.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *			GENERAL RULES ON VM_PAGE MANIPULATION
65  *
66  *	- A page queue lock is required when adding or removing a page from a
67  *	  page queue (vm_pagequeues[]), regardless of other locks or the
68  *	  busy state of a page.
69  *
70  *		* In general, no thread besides the page daemon can acquire or
71  *		  hold more than one page queue lock at a time.
72  *
73  *		* The page daemon can acquire and hold any pair of page queue
74  *		  locks in any order.
75  *
76  *	- The object lock is required when inserting or removing
77  *	  pages from an object (vm_page_insert() or vm_page_remove()).
78  *
79  */
80 
81 /*
82  *	Resident memory management module.
83  */
84 
85 #include <sys/cdefs.h>
86 __FBSDID("$FreeBSD$");
87 
88 #include "opt_vm.h"
89 
90 #include <sys/param.h>
91 #include <sys/systm.h>
92 #include <sys/lock.h>
93 #include <sys/kernel.h>
94 #include <sys/limits.h>
95 #include <sys/malloc.h>
96 #include <sys/msgbuf.h>
97 #include <sys/mutex.h>
98 #include <sys/proc.h>
99 #include <sys/rwlock.h>
100 #include <sys/sysctl.h>
101 #include <sys/vmmeter.h>
102 #include <sys/vnode.h>
103 
104 #include <vm/vm.h>
105 #include <vm/pmap.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_kern.h>
108 #include <vm/vm_object.h>
109 #include <vm/vm_page.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/vm_reserv.h>
114 #include <vm/vm_extern.h>
115 #include <vm/uma.h>
116 #include <vm/uma_int.h>
117 
118 #include <machine/md_var.h>
119 
120 /*
121  *	Associated with page of user-allocatable memory is a
122  *	page structure.
123  */
124 
125 struct vm_pagequeue vm_pagequeues[PQ_COUNT] = {
126 	[PQ_INACTIVE] = {
127 		.pq_pl = TAILQ_HEAD_INITIALIZER(
128 		    vm_pagequeues[PQ_INACTIVE].pq_pl),
129 		.pq_cnt = &cnt.v_inactive_count,
130 		.pq_name = "vm inactive pagequeue"
131 	},
132 	[PQ_ACTIVE] = {
133 		.pq_pl = TAILQ_HEAD_INITIALIZER(
134 		    vm_pagequeues[PQ_ACTIVE].pq_pl),
135 		.pq_cnt = &cnt.v_active_count,
136 		.pq_name = "vm active pagequeue"
137 	}
138 };
139 struct mtx_padalign vm_page_queue_free_mtx;
140 
141 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
142 
143 vm_page_t vm_page_array;
144 long vm_page_array_size;
145 long first_page;
146 int vm_page_zero_count;
147 
148 static int boot_pages = UMA_BOOT_PAGES;
149 TUNABLE_INT("vm.boot_pages", &boot_pages);
150 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
151 	"number of pages allocated for bootstrapping the VM system");
152 
153 static int pa_tryrelock_restart;
154 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
155     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
156 
157 static uma_zone_t fakepg_zone;
158 
159 static struct vnode *vm_page_alloc_init(vm_page_t m);
160 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
161 static void vm_page_enqueue(int queue, vm_page_t m);
162 static void vm_page_init_fakepg(void *dummy);
163 
164 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
165 
166 static void
167 vm_page_init_fakepg(void *dummy)
168 {
169 
170 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
171 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
172 }
173 
174 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
175 #if PAGE_SIZE == 32768
176 #ifdef CTASSERT
177 CTASSERT(sizeof(u_long) >= 8);
178 #endif
179 #endif
180 
181 /*
182  * Try to acquire a physical address lock while a pmap is locked.  If we
183  * fail to trylock we unlock and lock the pmap directly and cache the
184  * locked pa in *locked.  The caller should then restart their loop in case
185  * the virtual to physical mapping has changed.
186  */
187 int
188 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
189 {
190 	vm_paddr_t lockpa;
191 
192 	lockpa = *locked;
193 	*locked = pa;
194 	if (lockpa) {
195 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
196 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
197 			return (0);
198 		PA_UNLOCK(lockpa);
199 	}
200 	if (PA_TRYLOCK(pa))
201 		return (0);
202 	PMAP_UNLOCK(pmap);
203 	atomic_add_int(&pa_tryrelock_restart, 1);
204 	PA_LOCK(pa);
205 	PMAP_LOCK(pmap);
206 	return (EAGAIN);
207 }
208 
209 /*
210  *	vm_set_page_size:
211  *
212  *	Sets the page size, perhaps based upon the memory
213  *	size.  Must be called before any use of page-size
214  *	dependent functions.
215  */
216 void
217 vm_set_page_size(void)
218 {
219 	if (cnt.v_page_size == 0)
220 		cnt.v_page_size = PAGE_SIZE;
221 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
222 		panic("vm_set_page_size: page size not a power of two");
223 }
224 
225 /*
226  *	vm_page_blacklist_lookup:
227  *
228  *	See if a physical address in this page has been listed
229  *	in the blacklist tunable.  Entries in the tunable are
230  *	separated by spaces or commas.  If an invalid integer is
231  *	encountered then the rest of the string is skipped.
232  */
233 static int
234 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
235 {
236 	vm_paddr_t bad;
237 	char *cp, *pos;
238 
239 	for (pos = list; *pos != '\0'; pos = cp) {
240 		bad = strtoq(pos, &cp, 0);
241 		if (*cp != '\0') {
242 			if (*cp == ' ' || *cp == ',') {
243 				cp++;
244 				if (cp == pos)
245 					continue;
246 			} else
247 				break;
248 		}
249 		if (pa == trunc_page(bad))
250 			return (1);
251 	}
252 	return (0);
253 }
254 
255 /*
256  *	vm_page_startup:
257  *
258  *	Initializes the resident memory module.
259  *
260  *	Allocates memory for the page cells, and
261  *	for the object/offset-to-page hash table headers.
262  *	Each page cell is initialized and placed on the free list.
263  */
264 vm_offset_t
265 vm_page_startup(vm_offset_t vaddr)
266 {
267 	vm_offset_t mapped;
268 	vm_paddr_t page_range;
269 	vm_paddr_t new_end;
270 	int i;
271 	vm_paddr_t pa;
272 	vm_paddr_t last_pa;
273 	char *list;
274 
275 	/* the biggest memory array is the second group of pages */
276 	vm_paddr_t end;
277 	vm_paddr_t biggestsize;
278 	vm_paddr_t low_water, high_water;
279 	int biggestone;
280 
281 	biggestsize = 0;
282 	biggestone = 0;
283 	vaddr = round_page(vaddr);
284 
285 	for (i = 0; phys_avail[i + 1]; i += 2) {
286 		phys_avail[i] = round_page(phys_avail[i]);
287 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
288 	}
289 
290 	low_water = phys_avail[0];
291 	high_water = phys_avail[1];
292 
293 	for (i = 0; phys_avail[i + 1]; i += 2) {
294 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
295 
296 		if (size > biggestsize) {
297 			biggestone = i;
298 			biggestsize = size;
299 		}
300 		if (phys_avail[i] < low_water)
301 			low_water = phys_avail[i];
302 		if (phys_avail[i + 1] > high_water)
303 			high_water = phys_avail[i + 1];
304 	}
305 
306 #ifdef XEN
307 	low_water = 0;
308 #endif
309 
310 	end = phys_avail[biggestone+1];
311 
312 	/*
313 	 * Initialize the page and queue locks.
314 	 */
315 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
316 	for (i = 0; i < PA_LOCK_COUNT; i++)
317 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
318 	for (i = 0; i < PQ_COUNT; i++)
319 		vm_pagequeue_init_lock(&vm_pagequeues[i]);
320 
321 	/*
322 	 * Allocate memory for use when boot strapping the kernel memory
323 	 * allocator.
324 	 */
325 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
326 	new_end = trunc_page(new_end);
327 	mapped = pmap_map(&vaddr, new_end, end,
328 	    VM_PROT_READ | VM_PROT_WRITE);
329 	bzero((void *)mapped, end - new_end);
330 	uma_startup((void *)mapped, boot_pages);
331 
332 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
333     defined(__mips__)
334 	/*
335 	 * Allocate a bitmap to indicate that a random physical page
336 	 * needs to be included in a minidump.
337 	 *
338 	 * The amd64 port needs this to indicate which direct map pages
339 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
340 	 *
341 	 * However, i386 still needs this workspace internally within the
342 	 * minidump code.  In theory, they are not needed on i386, but are
343 	 * included should the sf_buf code decide to use them.
344 	 */
345 	last_pa = 0;
346 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
347 		if (dump_avail[i + 1] > last_pa)
348 			last_pa = dump_avail[i + 1];
349 	page_range = last_pa / PAGE_SIZE;
350 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
351 	new_end -= vm_page_dump_size;
352 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
353 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
354 	bzero((void *)vm_page_dump, vm_page_dump_size);
355 #endif
356 #ifdef __amd64__
357 	/*
358 	 * Request that the physical pages underlying the message buffer be
359 	 * included in a crash dump.  Since the message buffer is accessed
360 	 * through the direct map, they are not automatically included.
361 	 */
362 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
363 	last_pa = pa + round_page(msgbufsize);
364 	while (pa < last_pa) {
365 		dump_add_page(pa);
366 		pa += PAGE_SIZE;
367 	}
368 #endif
369 	/*
370 	 * Compute the number of pages of memory that will be available for
371 	 * use (taking into account the overhead of a page structure per
372 	 * page).
373 	 */
374 	first_page = low_water / PAGE_SIZE;
375 #ifdef VM_PHYSSEG_SPARSE
376 	page_range = 0;
377 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
378 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
379 #elif defined(VM_PHYSSEG_DENSE)
380 	page_range = high_water / PAGE_SIZE - first_page;
381 #else
382 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
383 #endif
384 	end = new_end;
385 
386 	/*
387 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
388 	 */
389 	vaddr += PAGE_SIZE;
390 
391 	/*
392 	 * Initialize the mem entry structures now, and put them in the free
393 	 * queue.
394 	 */
395 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
396 	mapped = pmap_map(&vaddr, new_end, end,
397 	    VM_PROT_READ | VM_PROT_WRITE);
398 	vm_page_array = (vm_page_t) mapped;
399 #if VM_NRESERVLEVEL > 0
400 	/*
401 	 * Allocate memory for the reservation management system's data
402 	 * structures.
403 	 */
404 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
405 #endif
406 #if defined(__amd64__) || defined(__mips__)
407 	/*
408 	 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
409 	 * like i386, so the pages must be tracked for a crashdump to include
410 	 * this data.  This includes the vm_page_array and the early UMA
411 	 * bootstrap pages.
412 	 */
413 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
414 		dump_add_page(pa);
415 #endif
416 	phys_avail[biggestone + 1] = new_end;
417 
418 	/*
419 	 * Clear all of the page structures
420 	 */
421 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
422 	for (i = 0; i < page_range; i++)
423 		vm_page_array[i].order = VM_NFREEORDER;
424 	vm_page_array_size = page_range;
425 
426 	/*
427 	 * Initialize the physical memory allocator.
428 	 */
429 	vm_phys_init();
430 
431 	/*
432 	 * Add every available physical page that is not blacklisted to
433 	 * the free lists.
434 	 */
435 	cnt.v_page_count = 0;
436 	cnt.v_free_count = 0;
437 	list = getenv("vm.blacklist");
438 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
439 		pa = phys_avail[i];
440 		last_pa = phys_avail[i + 1];
441 		while (pa < last_pa) {
442 			if (list != NULL &&
443 			    vm_page_blacklist_lookup(list, pa))
444 				printf("Skipping page with pa 0x%jx\n",
445 				    (uintmax_t)pa);
446 			else
447 				vm_phys_add_page(pa);
448 			pa += PAGE_SIZE;
449 		}
450 	}
451 	freeenv(list);
452 #if VM_NRESERVLEVEL > 0
453 	/*
454 	 * Initialize the reservation management system.
455 	 */
456 	vm_reserv_init();
457 #endif
458 	return (vaddr);
459 }
460 
461 void
462 vm_page_reference(vm_page_t m)
463 {
464 
465 	vm_page_aflag_set(m, PGA_REFERENCED);
466 }
467 
468 void
469 vm_page_busy(vm_page_t m)
470 {
471 
472 	VM_OBJECT_ASSERT_WLOCKED(m->object);
473 	KASSERT((m->oflags & VPO_BUSY) == 0,
474 	    ("vm_page_busy: page already busy!!!"));
475 	m->oflags |= VPO_BUSY;
476 }
477 
478 /*
479  *      vm_page_flash:
480  *
481  *      wakeup anyone waiting for the page.
482  */
483 void
484 vm_page_flash(vm_page_t m)
485 {
486 
487 	VM_OBJECT_ASSERT_WLOCKED(m->object);
488 	if (m->oflags & VPO_WANTED) {
489 		m->oflags &= ~VPO_WANTED;
490 		wakeup(m);
491 	}
492 }
493 
494 /*
495  *      vm_page_wakeup:
496  *
497  *      clear the VPO_BUSY flag and wakeup anyone waiting for the
498  *      page.
499  *
500  */
501 void
502 vm_page_wakeup(vm_page_t m)
503 {
504 
505 	VM_OBJECT_ASSERT_WLOCKED(m->object);
506 	KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
507 	m->oflags &= ~VPO_BUSY;
508 	vm_page_flash(m);
509 }
510 
511 void
512 vm_page_io_start(vm_page_t m)
513 {
514 
515 	VM_OBJECT_ASSERT_WLOCKED(m->object);
516 	m->busy++;
517 }
518 
519 void
520 vm_page_io_finish(vm_page_t m)
521 {
522 
523 	VM_OBJECT_ASSERT_WLOCKED(m->object);
524 	KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
525 	m->busy--;
526 	if (m->busy == 0)
527 		vm_page_flash(m);
528 }
529 
530 /*
531  * Keep page from being freed by the page daemon
532  * much of the same effect as wiring, except much lower
533  * overhead and should be used only for *very* temporary
534  * holding ("wiring").
535  */
536 void
537 vm_page_hold(vm_page_t mem)
538 {
539 
540 	vm_page_lock_assert(mem, MA_OWNED);
541         mem->hold_count++;
542 }
543 
544 void
545 vm_page_unhold(vm_page_t mem)
546 {
547 
548 	vm_page_lock_assert(mem, MA_OWNED);
549 	--mem->hold_count;
550 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
551 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
552 		vm_page_free_toq(mem);
553 }
554 
555 /*
556  *	vm_page_unhold_pages:
557  *
558  *	Unhold each of the pages that is referenced by the given array.
559  */
560 void
561 vm_page_unhold_pages(vm_page_t *ma, int count)
562 {
563 	struct mtx *mtx, *new_mtx;
564 
565 	mtx = NULL;
566 	for (; count != 0; count--) {
567 		/*
568 		 * Avoid releasing and reacquiring the same page lock.
569 		 */
570 		new_mtx = vm_page_lockptr(*ma);
571 		if (mtx != new_mtx) {
572 			if (mtx != NULL)
573 				mtx_unlock(mtx);
574 			mtx = new_mtx;
575 			mtx_lock(mtx);
576 		}
577 		vm_page_unhold(*ma);
578 		ma++;
579 	}
580 	if (mtx != NULL)
581 		mtx_unlock(mtx);
582 }
583 
584 vm_page_t
585 PHYS_TO_VM_PAGE(vm_paddr_t pa)
586 {
587 	vm_page_t m;
588 
589 #ifdef VM_PHYSSEG_SPARSE
590 	m = vm_phys_paddr_to_vm_page(pa);
591 	if (m == NULL)
592 		m = vm_phys_fictitious_to_vm_page(pa);
593 	return (m);
594 #elif defined(VM_PHYSSEG_DENSE)
595 	long pi;
596 
597 	pi = atop(pa);
598 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
599 		m = &vm_page_array[pi - first_page];
600 		return (m);
601 	}
602 	return (vm_phys_fictitious_to_vm_page(pa));
603 #else
604 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
605 #endif
606 }
607 
608 /*
609  *	vm_page_getfake:
610  *
611  *	Create a fictitious page with the specified physical address and
612  *	memory attribute.  The memory attribute is the only the machine-
613  *	dependent aspect of a fictitious page that must be initialized.
614  */
615 vm_page_t
616 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
617 {
618 	vm_page_t m;
619 
620 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
621 	vm_page_initfake(m, paddr, memattr);
622 	return (m);
623 }
624 
625 void
626 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
627 {
628 
629 	if ((m->flags & PG_FICTITIOUS) != 0) {
630 		/*
631 		 * The page's memattr might have changed since the
632 		 * previous initialization.  Update the pmap to the
633 		 * new memattr.
634 		 */
635 		goto memattr;
636 	}
637 	m->phys_addr = paddr;
638 	m->queue = PQ_NONE;
639 	/* Fictitious pages don't use "segind". */
640 	m->flags = PG_FICTITIOUS;
641 	/* Fictitious pages don't use "order" or "pool". */
642 	m->oflags = VPO_BUSY | VPO_UNMANAGED;
643 	m->wire_count = 1;
644 memattr:
645 	pmap_page_set_memattr(m, memattr);
646 }
647 
648 /*
649  *	vm_page_putfake:
650  *
651  *	Release a fictitious page.
652  */
653 void
654 vm_page_putfake(vm_page_t m)
655 {
656 
657 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
658 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
659 	    ("vm_page_putfake: bad page %p", m));
660 	uma_zfree(fakepg_zone, m);
661 }
662 
663 /*
664  *	vm_page_updatefake:
665  *
666  *	Update the given fictitious page to the specified physical address and
667  *	memory attribute.
668  */
669 void
670 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
671 {
672 
673 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
674 	    ("vm_page_updatefake: bad page %p", m));
675 	m->phys_addr = paddr;
676 	pmap_page_set_memattr(m, memattr);
677 }
678 
679 /*
680  *	vm_page_free:
681  *
682  *	Free a page.
683  */
684 void
685 vm_page_free(vm_page_t m)
686 {
687 
688 	m->flags &= ~PG_ZERO;
689 	vm_page_free_toq(m);
690 }
691 
692 /*
693  *	vm_page_free_zero:
694  *
695  *	Free a page to the zerod-pages queue
696  */
697 void
698 vm_page_free_zero(vm_page_t m)
699 {
700 
701 	m->flags |= PG_ZERO;
702 	vm_page_free_toq(m);
703 }
704 
705 /*
706  * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
707  * array which is not the request page.
708  */
709 void
710 vm_page_readahead_finish(vm_page_t m)
711 {
712 
713 	if (m->valid != 0) {
714 		/*
715 		 * Since the page is not the requested page, whether
716 		 * it should be activated or deactivated is not
717 		 * obvious.  Empirical results have shown that
718 		 * deactivating the page is usually the best choice,
719 		 * unless the page is wanted by another thread.
720 		 */
721 		if (m->oflags & VPO_WANTED) {
722 			vm_page_lock(m);
723 			vm_page_activate(m);
724 			vm_page_unlock(m);
725 		} else {
726 			vm_page_lock(m);
727 			vm_page_deactivate(m);
728 			vm_page_unlock(m);
729 		}
730 		vm_page_wakeup(m);
731 	} else {
732 		/*
733 		 * Free the completely invalid page.  Such page state
734 		 * occurs due to the short read operation which did
735 		 * not covered our page at all, or in case when a read
736 		 * error happens.
737 		 */
738 		vm_page_lock(m);
739 		vm_page_free(m);
740 		vm_page_unlock(m);
741 	}
742 }
743 
744 /*
745  *	vm_page_sleep:
746  *
747  *	Sleep and release the page lock.
748  *
749  *	The object containing the given page must be locked.
750  */
751 void
752 vm_page_sleep(vm_page_t m, const char *msg)
753 {
754 
755 	VM_OBJECT_ASSERT_WLOCKED(m->object);
756 	if (mtx_owned(vm_page_lockptr(m)))
757 		vm_page_unlock(m);
758 
759 	/*
760 	 * It's possible that while we sleep, the page will get
761 	 * unbusied and freed.  If we are holding the object
762 	 * lock, we will assume we hold a reference to the object
763 	 * such that even if m->object changes, we can re-lock
764 	 * it.
765 	 */
766 	m->oflags |= VPO_WANTED;
767 	VM_OBJECT_SLEEP(m->object, m, PVM, msg, 0);
768 }
769 
770 /*
771  *	vm_page_dirty_KBI:		[ internal use only ]
772  *
773  *	Set all bits in the page's dirty field.
774  *
775  *	The object containing the specified page must be locked if the
776  *	call is made from the machine-independent layer.
777  *
778  *	See vm_page_clear_dirty_mask().
779  *
780  *	This function should only be called by vm_page_dirty().
781  */
782 void
783 vm_page_dirty_KBI(vm_page_t m)
784 {
785 
786 	/* These assertions refer to this operation by its public name. */
787 	KASSERT((m->flags & PG_CACHED) == 0,
788 	    ("vm_page_dirty: page in cache!"));
789 	KASSERT(!VM_PAGE_IS_FREE(m),
790 	    ("vm_page_dirty: page is free!"));
791 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
792 	    ("vm_page_dirty: page is invalid!"));
793 	m->dirty = VM_PAGE_BITS_ALL;
794 }
795 
796 /*
797  *	vm_page_splay:
798  *
799  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
800  *	the vm_page containing the given pindex.  If, however, that
801  *	pindex is not found in the vm_object, returns a vm_page that is
802  *	adjacent to the pindex, coming before or after it.
803  */
804 vm_page_t
805 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
806 {
807 	struct vm_page dummy;
808 	vm_page_t lefttreemax, righttreemin, y;
809 
810 	if (root == NULL)
811 		return (root);
812 	lefttreemax = righttreemin = &dummy;
813 	for (;; root = y) {
814 		if (pindex < root->pindex) {
815 			if ((y = root->left) == NULL)
816 				break;
817 			if (pindex < y->pindex) {
818 				/* Rotate right. */
819 				root->left = y->right;
820 				y->right = root;
821 				root = y;
822 				if ((y = root->left) == NULL)
823 					break;
824 			}
825 			/* Link into the new root's right tree. */
826 			righttreemin->left = root;
827 			righttreemin = root;
828 		} else if (pindex > root->pindex) {
829 			if ((y = root->right) == NULL)
830 				break;
831 			if (pindex > y->pindex) {
832 				/* Rotate left. */
833 				root->right = y->left;
834 				y->left = root;
835 				root = y;
836 				if ((y = root->right) == NULL)
837 					break;
838 			}
839 			/* Link into the new root's left tree. */
840 			lefttreemax->right = root;
841 			lefttreemax = root;
842 		} else
843 			break;
844 	}
845 	/* Assemble the new root. */
846 	lefttreemax->right = root->left;
847 	righttreemin->left = root->right;
848 	root->left = dummy.right;
849 	root->right = dummy.left;
850 	return (root);
851 }
852 
853 /*
854  *	vm_page_insert:		[ internal use only ]
855  *
856  *	Inserts the given mem entry into the object and object list.
857  *
858  *	The pagetables are not updated but will presumably fault the page
859  *	in if necessary, or if a kernel page the caller will at some point
860  *	enter the page into the kernel's pmap.  We are not allowed to sleep
861  *	here so we *can't* do this anyway.
862  *
863  *	The object must be locked.
864  */
865 void
866 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
867 {
868 	vm_page_t root;
869 
870 	VM_OBJECT_ASSERT_WLOCKED(object);
871 	if (m->object != NULL)
872 		panic("vm_page_insert: page already inserted");
873 
874 	/*
875 	 * Record the object/offset pair in this page
876 	 */
877 	m->object = object;
878 	m->pindex = pindex;
879 
880 	/*
881 	 * Now link into the object's ordered list of backed pages.
882 	 */
883 	root = object->root;
884 	if (root == NULL) {
885 		m->left = NULL;
886 		m->right = NULL;
887 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
888 	} else {
889 		root = vm_page_splay(pindex, root);
890 		if (pindex < root->pindex) {
891 			m->left = root->left;
892 			m->right = root;
893 			root->left = NULL;
894 			TAILQ_INSERT_BEFORE(root, m, listq);
895 		} else if (pindex == root->pindex)
896 			panic("vm_page_insert: offset already allocated");
897 		else {
898 			m->right = root->right;
899 			m->left = root;
900 			root->right = NULL;
901 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
902 		}
903 	}
904 	object->root = m;
905 
906 	/*
907 	 * Show that the object has one more resident page.
908 	 */
909 	object->resident_page_count++;
910 
911 	/*
912 	 * Hold the vnode until the last page is released.
913 	 */
914 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
915 		vhold(object->handle);
916 
917 	/*
918 	 * Since we are inserting a new and possibly dirty page,
919 	 * update the object's OBJ_MIGHTBEDIRTY flag.
920 	 */
921 	if (pmap_page_is_write_mapped(m))
922 		vm_object_set_writeable_dirty(object);
923 }
924 
925 /*
926  *	vm_page_remove:
927  *
928  *	Removes the given mem entry from the object/offset-page
929  *	table and the object page list, but do not invalidate/terminate
930  *	the backing store.
931  *
932  *	The underlying pmap entry (if any) is NOT removed here.
933  *
934  *	The object must be locked.  The page must be locked if it is managed.
935  */
936 void
937 vm_page_remove(vm_page_t m)
938 {
939 	vm_object_t object;
940 	vm_page_t next, prev, root;
941 
942 	if ((m->oflags & VPO_UNMANAGED) == 0)
943 		vm_page_lock_assert(m, MA_OWNED);
944 	if ((object = m->object) == NULL)
945 		return;
946 	VM_OBJECT_ASSERT_WLOCKED(object);
947 	if (m->oflags & VPO_BUSY) {
948 		m->oflags &= ~VPO_BUSY;
949 		vm_page_flash(m);
950 	}
951 
952 	/*
953 	 * Now remove from the object's list of backed pages.
954 	 */
955 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
956 		/*
957 		 * Since the page's successor in the list is also its parent
958 		 * in the tree, its right subtree must be empty.
959 		 */
960 		next->left = m->left;
961 		KASSERT(m->right == NULL,
962 		    ("vm_page_remove: page %p has right child", m));
963 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
964 	    prev->right == m) {
965 		/*
966 		 * Since the page's predecessor in the list is also its parent
967 		 * in the tree, its left subtree must be empty.
968 		 */
969 		KASSERT(m->left == NULL,
970 		    ("vm_page_remove: page %p has left child", m));
971 		prev->right = m->right;
972 	} else {
973 		if (m != object->root)
974 			vm_page_splay(m->pindex, object->root);
975 		if (m->left == NULL)
976 			root = m->right;
977 		else if (m->right == NULL)
978 			root = m->left;
979 		else {
980 			/*
981 			 * Move the page's successor to the root, because
982 			 * pages are usually removed in ascending order.
983 			 */
984 			if (m->right != next)
985 				vm_page_splay(m->pindex, m->right);
986 			next->left = m->left;
987 			root = next;
988 		}
989 		object->root = root;
990 	}
991 	TAILQ_REMOVE(&object->memq, m, listq);
992 
993 	/*
994 	 * And show that the object has one fewer resident page.
995 	 */
996 	object->resident_page_count--;
997 
998 	/*
999 	 * The vnode may now be recycled.
1000 	 */
1001 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1002 		vdrop(object->handle);
1003 
1004 	m->object = NULL;
1005 }
1006 
1007 /*
1008  *	vm_page_lookup:
1009  *
1010  *	Returns the page associated with the object/offset
1011  *	pair specified; if none is found, NULL is returned.
1012  *
1013  *	The object must be locked.
1014  */
1015 vm_page_t
1016 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1017 {
1018 	vm_page_t m;
1019 
1020 	VM_OBJECT_ASSERT_WLOCKED(object);
1021 	if ((m = object->root) != NULL && m->pindex != pindex) {
1022 		m = vm_page_splay(pindex, m);
1023 		if ((object->root = m)->pindex != pindex)
1024 			m = NULL;
1025 	}
1026 	return (m);
1027 }
1028 
1029 /*
1030  *	vm_page_find_least:
1031  *
1032  *	Returns the page associated with the object with least pindex
1033  *	greater than or equal to the parameter pindex, or NULL.
1034  *
1035  *	The object must be locked.
1036  */
1037 vm_page_t
1038 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1039 {
1040 	vm_page_t m;
1041 
1042 	VM_OBJECT_ASSERT_WLOCKED(object);
1043 	if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
1044 		if (m->pindex < pindex) {
1045 			m = vm_page_splay(pindex, object->root);
1046 			if ((object->root = m)->pindex < pindex)
1047 				m = TAILQ_NEXT(m, listq);
1048 		}
1049 	}
1050 	return (m);
1051 }
1052 
1053 /*
1054  * Returns the given page's successor (by pindex) within the object if it is
1055  * resident; if none is found, NULL is returned.
1056  *
1057  * The object must be locked.
1058  */
1059 vm_page_t
1060 vm_page_next(vm_page_t m)
1061 {
1062 	vm_page_t next;
1063 
1064 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1065 	if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1066 	    next->pindex != m->pindex + 1)
1067 		next = NULL;
1068 	return (next);
1069 }
1070 
1071 /*
1072  * Returns the given page's predecessor (by pindex) within the object if it is
1073  * resident; if none is found, NULL is returned.
1074  *
1075  * The object must be locked.
1076  */
1077 vm_page_t
1078 vm_page_prev(vm_page_t m)
1079 {
1080 	vm_page_t prev;
1081 
1082 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1083 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1084 	    prev->pindex != m->pindex - 1)
1085 		prev = NULL;
1086 	return (prev);
1087 }
1088 
1089 /*
1090  *	vm_page_rename:
1091  *
1092  *	Move the given memory entry from its
1093  *	current object to the specified target object/offset.
1094  *
1095  *	Note: swap associated with the page must be invalidated by the move.  We
1096  *	      have to do this for several reasons:  (1) we aren't freeing the
1097  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1098  *	      moving the page from object A to B, and will then later move
1099  *	      the backing store from A to B and we can't have a conflict.
1100  *
1101  *	Note: we *always* dirty the page.  It is necessary both for the
1102  *	      fact that we moved it, and because we may be invalidating
1103  *	      swap.  If the page is on the cache, we have to deactivate it
1104  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
1105  *	      on the cache.
1106  *
1107  *	The objects must be locked.  The page must be locked if it is managed.
1108  */
1109 void
1110 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1111 {
1112 
1113 	vm_page_remove(m);
1114 	vm_page_insert(m, new_object, new_pindex);
1115 	vm_page_dirty(m);
1116 }
1117 
1118 /*
1119  *	Convert all of the given object's cached pages that have a
1120  *	pindex within the given range into free pages.  If the value
1121  *	zero is given for "end", then the range's upper bound is
1122  *	infinity.  If the given object is backed by a vnode and it
1123  *	transitions from having one or more cached pages to none, the
1124  *	vnode's hold count is reduced.
1125  */
1126 void
1127 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1128 {
1129 	vm_page_t m, m_next;
1130 	boolean_t empty;
1131 
1132 	mtx_lock(&vm_page_queue_free_mtx);
1133 	if (__predict_false(vm_object_cache_is_empty(object))) {
1134 		mtx_unlock(&vm_page_queue_free_mtx);
1135 		return;
1136 	}
1137 	m = object->cache = vm_page_splay(start, object->cache);
1138 	if (m->pindex < start) {
1139 		if (m->right == NULL)
1140 			m = NULL;
1141 		else {
1142 			m_next = vm_page_splay(start, m->right);
1143 			m_next->left = m;
1144 			m->right = NULL;
1145 			m = object->cache = m_next;
1146 		}
1147 	}
1148 
1149 	/*
1150 	 * At this point, "m" is either (1) a reference to the page
1151 	 * with the least pindex that is greater than or equal to
1152 	 * "start" or (2) NULL.
1153 	 */
1154 	for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
1155 		/*
1156 		 * Find "m"'s successor and remove "m" from the
1157 		 * object's cache.
1158 		 */
1159 		if (m->right == NULL) {
1160 			object->cache = m->left;
1161 			m_next = NULL;
1162 		} else {
1163 			m_next = vm_page_splay(start, m->right);
1164 			m_next->left = m->left;
1165 			object->cache = m_next;
1166 		}
1167 		/* Convert "m" to a free page. */
1168 		m->object = NULL;
1169 		m->valid = 0;
1170 		/* Clear PG_CACHED and set PG_FREE. */
1171 		m->flags ^= PG_CACHED | PG_FREE;
1172 		KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1173 		    ("vm_page_cache_free: page %p has inconsistent flags", m));
1174 		cnt.v_cache_count--;
1175 		cnt.v_free_count++;
1176 	}
1177 	empty = vm_object_cache_is_empty(object);
1178 	mtx_unlock(&vm_page_queue_free_mtx);
1179 	if (object->type == OBJT_VNODE && empty)
1180 		vdrop(object->handle);
1181 }
1182 
1183 /*
1184  *	Returns the cached page that is associated with the given
1185  *	object and offset.  If, however, none exists, returns NULL.
1186  *
1187  *	The free page queue must be locked.
1188  */
1189 static inline vm_page_t
1190 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1191 {
1192 	vm_page_t m;
1193 
1194 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1195 	if ((m = object->cache) != NULL && m->pindex != pindex) {
1196 		m = vm_page_splay(pindex, m);
1197 		if ((object->cache = m)->pindex != pindex)
1198 			m = NULL;
1199 	}
1200 	return (m);
1201 }
1202 
1203 /*
1204  *	Remove the given cached page from its containing object's
1205  *	collection of cached pages.
1206  *
1207  *	The free page queue must be locked.
1208  */
1209 static void
1210 vm_page_cache_remove(vm_page_t m)
1211 {
1212 	vm_object_t object;
1213 	vm_page_t root;
1214 
1215 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1216 	KASSERT((m->flags & PG_CACHED) != 0,
1217 	    ("vm_page_cache_remove: page %p is not cached", m));
1218 	object = m->object;
1219 	if (m != object->cache) {
1220 		root = vm_page_splay(m->pindex, object->cache);
1221 		KASSERT(root == m,
1222 		    ("vm_page_cache_remove: page %p is not cached in object %p",
1223 		    m, object));
1224 	}
1225 	if (m->left == NULL)
1226 		root = m->right;
1227 	else if (m->right == NULL)
1228 		root = m->left;
1229 	else {
1230 		root = vm_page_splay(m->pindex, m->left);
1231 		root->right = m->right;
1232 	}
1233 	object->cache = root;
1234 	m->object = NULL;
1235 	cnt.v_cache_count--;
1236 }
1237 
1238 /*
1239  *	Transfer all of the cached pages with offset greater than or
1240  *	equal to 'offidxstart' from the original object's cache to the
1241  *	new object's cache.  However, any cached pages with offset
1242  *	greater than or equal to the new object's size are kept in the
1243  *	original object.  Initially, the new object's cache must be
1244  *	empty.  Offset 'offidxstart' in the original object must
1245  *	correspond to offset zero in the new object.
1246  *
1247  *	The new object must be locked.
1248  */
1249 void
1250 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1251     vm_object_t new_object)
1252 {
1253 	vm_page_t m, m_next;
1254 
1255 	/*
1256 	 * Insertion into an object's collection of cached pages
1257 	 * requires the object to be locked.  In contrast, removal does
1258 	 * not.
1259 	 */
1260 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1261 	KASSERT(vm_object_cache_is_empty(new_object),
1262 	    ("vm_page_cache_transfer: object %p has cached pages",
1263 	    new_object));
1264 	mtx_lock(&vm_page_queue_free_mtx);
1265 	if ((m = orig_object->cache) != NULL) {
1266 		/*
1267 		 * Transfer all of the pages with offset greater than or
1268 		 * equal to 'offidxstart' from the original object's
1269 		 * cache to the new object's cache.
1270 		 */
1271 		m = vm_page_splay(offidxstart, m);
1272 		if (m->pindex < offidxstart) {
1273 			orig_object->cache = m;
1274 			new_object->cache = m->right;
1275 			m->right = NULL;
1276 		} else {
1277 			orig_object->cache = m->left;
1278 			new_object->cache = m;
1279 			m->left = NULL;
1280 		}
1281 		while ((m = new_object->cache) != NULL) {
1282 			if ((m->pindex - offidxstart) >= new_object->size) {
1283 				/*
1284 				 * Return all of the cached pages with
1285 				 * offset greater than or equal to the
1286 				 * new object's size to the original
1287 				 * object's cache.
1288 				 */
1289 				new_object->cache = m->left;
1290 				m->left = orig_object->cache;
1291 				orig_object->cache = m;
1292 				break;
1293 			}
1294 			m_next = vm_page_splay(m->pindex, m->right);
1295 			/* Update the page's object and offset. */
1296 			m->object = new_object;
1297 			m->pindex -= offidxstart;
1298 			if (m_next == NULL)
1299 				break;
1300 			m->right = NULL;
1301 			m_next->left = m;
1302 			new_object->cache = m_next;
1303 		}
1304 		KASSERT(vm_object_cache_is_empty(new_object) ||
1305 		    new_object->type == OBJT_SWAP,
1306 		    ("vm_page_cache_transfer: object %p's type is incompatible"
1307 		    " with cached pages", new_object));
1308 	}
1309 	mtx_unlock(&vm_page_queue_free_mtx);
1310 }
1311 
1312 /*
1313  *	Returns TRUE if a cached page is associated with the given object and
1314  *	offset, and FALSE otherwise.
1315  *
1316  *	The object must be locked.
1317  */
1318 boolean_t
1319 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1320 {
1321 	vm_page_t m;
1322 
1323 	/*
1324 	 * Insertion into an object's collection of cached pages requires the
1325 	 * object to be locked.  Therefore, if the object is locked and the
1326 	 * object's collection is empty, there is no need to acquire the free
1327 	 * page queues lock in order to prove that the specified page doesn't
1328 	 * exist.
1329 	 */
1330 	VM_OBJECT_ASSERT_WLOCKED(object);
1331 	if (__predict_true(vm_object_cache_is_empty(object)))
1332 		return (FALSE);
1333 	mtx_lock(&vm_page_queue_free_mtx);
1334 	m = vm_page_cache_lookup(object, pindex);
1335 	mtx_unlock(&vm_page_queue_free_mtx);
1336 	return (m != NULL);
1337 }
1338 
1339 /*
1340  *	vm_page_alloc:
1341  *
1342  *	Allocate and return a page that is associated with the specified
1343  *	object and offset pair.  By default, this page has the flag VPO_BUSY
1344  *	set.
1345  *
1346  *	The caller must always specify an allocation class.
1347  *
1348  *	allocation classes:
1349  *	VM_ALLOC_NORMAL		normal process request
1350  *	VM_ALLOC_SYSTEM		system *really* needs a page
1351  *	VM_ALLOC_INTERRUPT	interrupt time request
1352  *
1353  *	optional allocation flags:
1354  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1355  *				intends to allocate
1356  *	VM_ALLOC_IFCACHED	return page only if it is cached
1357  *	VM_ALLOC_IFNOTCACHED	return NULL, do not reactivate if the page
1358  *				is cached
1359  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1360  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1361  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1362  *				should not have the flag VPO_BUSY set
1363  *	VM_ALLOC_WIRED		wire the allocated page
1364  *	VM_ALLOC_ZERO		prefer a zeroed page
1365  *
1366  *	This routine may not sleep.
1367  */
1368 vm_page_t
1369 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1370 {
1371 	struct vnode *vp = NULL;
1372 	vm_object_t m_object;
1373 	vm_page_t m;
1374 	int flags, req_class;
1375 
1376 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1377 	    ("vm_page_alloc: inconsistent object/req"));
1378 	if (object != NULL)
1379 		VM_OBJECT_ASSERT_WLOCKED(object);
1380 
1381 	req_class = req & VM_ALLOC_CLASS_MASK;
1382 
1383 	/*
1384 	 * The page daemon is allowed to dig deeper into the free page list.
1385 	 */
1386 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1387 		req_class = VM_ALLOC_SYSTEM;
1388 
1389 	mtx_lock(&vm_page_queue_free_mtx);
1390 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1391 	    (req_class == VM_ALLOC_SYSTEM &&
1392 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1393 	    (req_class == VM_ALLOC_INTERRUPT &&
1394 	    cnt.v_free_count + cnt.v_cache_count > 0)) {
1395 		/*
1396 		 * Allocate from the free queue if the number of free pages
1397 		 * exceeds the minimum for the request class.
1398 		 */
1399 		if (object != NULL &&
1400 		    (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1401 			if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1402 				mtx_unlock(&vm_page_queue_free_mtx);
1403 				return (NULL);
1404 			}
1405 			if (vm_phys_unfree_page(m))
1406 				vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1407 #if VM_NRESERVLEVEL > 0
1408 			else if (!vm_reserv_reactivate_page(m))
1409 #else
1410 			else
1411 #endif
1412 				panic("vm_page_alloc: cache page %p is missing"
1413 				    " from the free queue", m);
1414 		} else if ((req & VM_ALLOC_IFCACHED) != 0) {
1415 			mtx_unlock(&vm_page_queue_free_mtx);
1416 			return (NULL);
1417 #if VM_NRESERVLEVEL > 0
1418 		} else if (object == NULL || (object->flags & (OBJ_COLORED |
1419 		    OBJ_FICTITIOUS)) != OBJ_COLORED ||
1420 		    (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1421 #else
1422 		} else {
1423 #endif
1424 			m = vm_phys_alloc_pages(object != NULL ?
1425 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1426 #if VM_NRESERVLEVEL > 0
1427 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1428 				m = vm_phys_alloc_pages(object != NULL ?
1429 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1430 				    0);
1431 			}
1432 #endif
1433 		}
1434 	} else {
1435 		/*
1436 		 * Not allocatable, give up.
1437 		 */
1438 		mtx_unlock(&vm_page_queue_free_mtx);
1439 		atomic_add_int(&vm_pageout_deficit,
1440 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1441 		pagedaemon_wakeup();
1442 		return (NULL);
1443 	}
1444 
1445 	/*
1446 	 *  At this point we had better have found a good page.
1447 	 */
1448 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1449 	KASSERT(m->queue == PQ_NONE,
1450 	    ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1451 	KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1452 	KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1453 	KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1454 	KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1455 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1456 	    ("vm_page_alloc: page %p has unexpected memattr %d", m,
1457 	    pmap_page_get_memattr(m)));
1458 	if ((m->flags & PG_CACHED) != 0) {
1459 		KASSERT((m->flags & PG_ZERO) == 0,
1460 		    ("vm_page_alloc: cached page %p is PG_ZERO", m));
1461 		KASSERT(m->valid != 0,
1462 		    ("vm_page_alloc: cached page %p is invalid", m));
1463 		if (m->object == object && m->pindex == pindex)
1464 	  		cnt.v_reactivated++;
1465 		else
1466 			m->valid = 0;
1467 		m_object = m->object;
1468 		vm_page_cache_remove(m);
1469 		if (m_object->type == OBJT_VNODE &&
1470 		    vm_object_cache_is_empty(m_object))
1471 			vp = m_object->handle;
1472 	} else {
1473 		KASSERT(VM_PAGE_IS_FREE(m),
1474 		    ("vm_page_alloc: page %p is not free", m));
1475 		KASSERT(m->valid == 0,
1476 		    ("vm_page_alloc: free page %p is valid", m));
1477 		cnt.v_free_count--;
1478 	}
1479 
1480 	/*
1481 	 * Only the PG_ZERO flag is inherited.  The PG_CACHED or PG_FREE flag
1482 	 * must be cleared before the free page queues lock is released.
1483 	 */
1484 	flags = 0;
1485 	if (m->flags & PG_ZERO) {
1486 		vm_page_zero_count--;
1487 		if (req & VM_ALLOC_ZERO)
1488 			flags = PG_ZERO;
1489 	}
1490 	if (req & VM_ALLOC_NODUMP)
1491 		flags |= PG_NODUMP;
1492 	m->flags = flags;
1493 	mtx_unlock(&vm_page_queue_free_mtx);
1494 	m->aflags = 0;
1495 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1496 	    VPO_UNMANAGED : 0;
1497 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1498 		m->oflags |= VPO_BUSY;
1499 	if (req & VM_ALLOC_WIRED) {
1500 		/*
1501 		 * The page lock is not required for wiring a page until that
1502 		 * page is inserted into the object.
1503 		 */
1504 		atomic_add_int(&cnt.v_wire_count, 1);
1505 		m->wire_count = 1;
1506 	}
1507 	m->act_count = 0;
1508 
1509 	if (object != NULL) {
1510 		/* Ignore device objects; the pager sets "memattr" for them. */
1511 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1512 		    (object->flags & OBJ_FICTITIOUS) == 0)
1513 			pmap_page_set_memattr(m, object->memattr);
1514 		vm_page_insert(m, object, pindex);
1515 	} else
1516 		m->pindex = pindex;
1517 
1518 	/*
1519 	 * The following call to vdrop() must come after the above call
1520 	 * to vm_page_insert() in case both affect the same object and
1521 	 * vnode.  Otherwise, the affected vnode's hold count could
1522 	 * temporarily become zero.
1523 	 */
1524 	if (vp != NULL)
1525 		vdrop(vp);
1526 
1527 	/*
1528 	 * Don't wakeup too often - wakeup the pageout daemon when
1529 	 * we would be nearly out of memory.
1530 	 */
1531 	if (vm_paging_needed())
1532 		pagedaemon_wakeup();
1533 
1534 	return (m);
1535 }
1536 
1537 /*
1538  *	vm_page_alloc_contig:
1539  *
1540  *	Allocate a contiguous set of physical pages of the given size "npages"
1541  *	from the free lists.  All of the physical pages must be at or above
1542  *	the given physical address "low" and below the given physical address
1543  *	"high".  The given value "alignment" determines the alignment of the
1544  *	first physical page in the set.  If the given value "boundary" is
1545  *	non-zero, then the set of physical pages cannot cross any physical
1546  *	address boundary that is a multiple of that value.  Both "alignment"
1547  *	and "boundary" must be a power of two.
1548  *
1549  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1550  *	then the memory attribute setting for the physical pages is configured
1551  *	to the object's memory attribute setting.  Otherwise, the memory
1552  *	attribute setting for the physical pages is configured to "memattr",
1553  *	overriding the object's memory attribute setting.  However, if the
1554  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1555  *	memory attribute setting for the physical pages cannot be configured
1556  *	to VM_MEMATTR_DEFAULT.
1557  *
1558  *	The caller must always specify an allocation class.
1559  *
1560  *	allocation classes:
1561  *	VM_ALLOC_NORMAL		normal process request
1562  *	VM_ALLOC_SYSTEM		system *really* needs a page
1563  *	VM_ALLOC_INTERRUPT	interrupt time request
1564  *
1565  *	optional allocation flags:
1566  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1567  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1568  *				should not have the flag VPO_BUSY set
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_contig(vm_object_t object, vm_pindex_t pindex, int req,
1576     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1577     vm_paddr_t boundary, vm_memattr_t memattr)
1578 {
1579 	struct vnode *drop;
1580 	vm_page_t deferred_vdrop_list, m, m_ret;
1581 	u_int flags, oflags;
1582 	int req_class;
1583 
1584 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1585 	    ("vm_page_alloc_contig: inconsistent object/req"));
1586 	if (object != NULL) {
1587 		VM_OBJECT_ASSERT_WLOCKED(object);
1588 		KASSERT(object->type == OBJT_PHYS,
1589 		    ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1590 		    object));
1591 	}
1592 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1593 	req_class = req & VM_ALLOC_CLASS_MASK;
1594 
1595 	/*
1596 	 * The page daemon is allowed to dig deeper into the free page list.
1597 	 */
1598 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1599 		req_class = VM_ALLOC_SYSTEM;
1600 
1601 	deferred_vdrop_list = NULL;
1602 	mtx_lock(&vm_page_queue_free_mtx);
1603 	if (cnt.v_free_count + cnt.v_cache_count >= npages +
1604 	    cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1605 	    cnt.v_free_count + cnt.v_cache_count >= npages +
1606 	    cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1607 	    cnt.v_free_count + cnt.v_cache_count >= npages)) {
1608 #if VM_NRESERVLEVEL > 0
1609 retry:
1610 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1611 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1612 		    low, high, alignment, boundary)) == NULL)
1613 #endif
1614 			m_ret = vm_phys_alloc_contig(npages, low, high,
1615 			    alignment, boundary);
1616 	} else {
1617 		mtx_unlock(&vm_page_queue_free_mtx);
1618 		atomic_add_int(&vm_pageout_deficit, npages);
1619 		pagedaemon_wakeup();
1620 		return (NULL);
1621 	}
1622 	if (m_ret != NULL)
1623 		for (m = m_ret; m < &m_ret[npages]; m++) {
1624 			drop = vm_page_alloc_init(m);
1625 			if (drop != NULL) {
1626 				/*
1627 				 * Enqueue the vnode for deferred vdrop().
1628 				 *
1629 				 * Once the pages are removed from the free
1630 				 * page list, "pageq" can be safely abused to
1631 				 * construct a short-lived list of vnodes.
1632 				 */
1633 				m->pageq.tqe_prev = (void *)drop;
1634 				m->pageq.tqe_next = deferred_vdrop_list;
1635 				deferred_vdrop_list = m;
1636 			}
1637 		}
1638 	else {
1639 #if VM_NRESERVLEVEL > 0
1640 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1641 		    boundary))
1642 			goto retry;
1643 #endif
1644 	}
1645 	mtx_unlock(&vm_page_queue_free_mtx);
1646 	if (m_ret == NULL)
1647 		return (NULL);
1648 
1649 	/*
1650 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1651 	 */
1652 	flags = 0;
1653 	if ((req & VM_ALLOC_ZERO) != 0)
1654 		flags = PG_ZERO;
1655 	if ((req & VM_ALLOC_NODUMP) != 0)
1656 		flags |= PG_NODUMP;
1657 	if ((req & VM_ALLOC_WIRED) != 0)
1658 		atomic_add_int(&cnt.v_wire_count, npages);
1659 	oflags = VPO_UNMANAGED;
1660 	if (object != NULL) {
1661 		if ((req & VM_ALLOC_NOBUSY) == 0)
1662 			oflags |= VPO_BUSY;
1663 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1664 		    memattr == VM_MEMATTR_DEFAULT)
1665 			memattr = object->memattr;
1666 	}
1667 	for (m = m_ret; m < &m_ret[npages]; m++) {
1668 		m->aflags = 0;
1669 		m->flags = (m->flags | PG_NODUMP) & flags;
1670 		if ((req & VM_ALLOC_WIRED) != 0)
1671 			m->wire_count = 1;
1672 		/* Unmanaged pages don't use "act_count". */
1673 		m->oflags = oflags;
1674 		if (memattr != VM_MEMATTR_DEFAULT)
1675 			pmap_page_set_memattr(m, memattr);
1676 		if (object != NULL)
1677 			vm_page_insert(m, object, pindex);
1678 		else
1679 			m->pindex = pindex;
1680 		pindex++;
1681 	}
1682 	while (deferred_vdrop_list != NULL) {
1683 		vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1684 		deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1685 	}
1686 	if (vm_paging_needed())
1687 		pagedaemon_wakeup();
1688 	return (m_ret);
1689 }
1690 
1691 /*
1692  * Initialize a page that has been freshly dequeued from a freelist.
1693  * The caller has to drop the vnode returned, if it is not NULL.
1694  *
1695  * This function may only be used to initialize unmanaged pages.
1696  *
1697  * To be called with vm_page_queue_free_mtx held.
1698  */
1699 static struct vnode *
1700 vm_page_alloc_init(vm_page_t m)
1701 {
1702 	struct vnode *drop;
1703 	vm_object_t m_object;
1704 
1705 	KASSERT(m->queue == PQ_NONE,
1706 	    ("vm_page_alloc_init: page %p has unexpected queue %d",
1707 	    m, m->queue));
1708 	KASSERT(m->wire_count == 0,
1709 	    ("vm_page_alloc_init: page %p is wired", m));
1710 	KASSERT(m->hold_count == 0,
1711 	    ("vm_page_alloc_init: page %p is held", m));
1712 	KASSERT(m->busy == 0,
1713 	    ("vm_page_alloc_init: page %p is busy", m));
1714 	KASSERT(m->dirty == 0,
1715 	    ("vm_page_alloc_init: page %p is dirty", m));
1716 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1717 	    ("vm_page_alloc_init: page %p has unexpected memattr %d",
1718 	    m, pmap_page_get_memattr(m)));
1719 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1720 	drop = NULL;
1721 	if ((m->flags & PG_CACHED) != 0) {
1722 		KASSERT((m->flags & PG_ZERO) == 0,
1723 		    ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1724 		m->valid = 0;
1725 		m_object = m->object;
1726 		vm_page_cache_remove(m);
1727 		if (m_object->type == OBJT_VNODE &&
1728 		    vm_object_cache_is_empty(m_object))
1729 			drop = m_object->handle;
1730 	} else {
1731 		KASSERT(VM_PAGE_IS_FREE(m),
1732 		    ("vm_page_alloc_init: page %p is not free", m));
1733 		KASSERT(m->valid == 0,
1734 		    ("vm_page_alloc_init: free page %p is valid", m));
1735 		cnt.v_free_count--;
1736 		if ((m->flags & PG_ZERO) != 0)
1737 			vm_page_zero_count--;
1738 	}
1739 	/* Don't clear the PG_ZERO flag; we'll need it later. */
1740 	m->flags &= PG_ZERO;
1741 	return (drop);
1742 }
1743 
1744 /*
1745  * 	vm_page_alloc_freelist:
1746  *
1747  *	Allocate a physical page from the specified free page list.
1748  *
1749  *	The caller must always specify an allocation class.
1750  *
1751  *	allocation classes:
1752  *	VM_ALLOC_NORMAL		normal process request
1753  *	VM_ALLOC_SYSTEM		system *really* needs a page
1754  *	VM_ALLOC_INTERRUPT	interrupt time request
1755  *
1756  *	optional allocation flags:
1757  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1758  *				intends to allocate
1759  *	VM_ALLOC_WIRED		wire the allocated page
1760  *	VM_ALLOC_ZERO		prefer a zeroed page
1761  *
1762  *	This routine may not sleep.
1763  */
1764 vm_page_t
1765 vm_page_alloc_freelist(int flind, int req)
1766 {
1767 	struct vnode *drop;
1768 	vm_page_t m;
1769 	u_int flags;
1770 	int req_class;
1771 
1772 	req_class = req & VM_ALLOC_CLASS_MASK;
1773 
1774 	/*
1775 	 * The page daemon is allowed to dig deeper into the free page list.
1776 	 */
1777 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1778 		req_class = VM_ALLOC_SYSTEM;
1779 
1780 	/*
1781 	 * Do not allocate reserved pages unless the req has asked for it.
1782 	 */
1783 	mtx_lock(&vm_page_queue_free_mtx);
1784 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1785 	    (req_class == VM_ALLOC_SYSTEM &&
1786 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1787 	    (req_class == VM_ALLOC_INTERRUPT &&
1788 	    cnt.v_free_count + cnt.v_cache_count > 0))
1789 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1790 	else {
1791 		mtx_unlock(&vm_page_queue_free_mtx);
1792 		atomic_add_int(&vm_pageout_deficit,
1793 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1794 		pagedaemon_wakeup();
1795 		return (NULL);
1796 	}
1797 	if (m == NULL) {
1798 		mtx_unlock(&vm_page_queue_free_mtx);
1799 		return (NULL);
1800 	}
1801 	drop = vm_page_alloc_init(m);
1802 	mtx_unlock(&vm_page_queue_free_mtx);
1803 
1804 	/*
1805 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1806 	 */
1807 	m->aflags = 0;
1808 	flags = 0;
1809 	if ((req & VM_ALLOC_ZERO) != 0)
1810 		flags = PG_ZERO;
1811 	m->flags &= flags;
1812 	if ((req & VM_ALLOC_WIRED) != 0) {
1813 		/*
1814 		 * The page lock is not required for wiring a page that does
1815 		 * not belong to an object.
1816 		 */
1817 		atomic_add_int(&cnt.v_wire_count, 1);
1818 		m->wire_count = 1;
1819 	}
1820 	/* Unmanaged pages don't use "act_count". */
1821 	m->oflags = VPO_UNMANAGED;
1822 	if (drop != NULL)
1823 		vdrop(drop);
1824 	if (vm_paging_needed())
1825 		pagedaemon_wakeup();
1826 	return (m);
1827 }
1828 
1829 /*
1830  *	vm_wait:	(also see VM_WAIT macro)
1831  *
1832  *	Sleep until free pages are available for allocation.
1833  *	- Called in various places before memory allocations.
1834  */
1835 void
1836 vm_wait(void)
1837 {
1838 
1839 	mtx_lock(&vm_page_queue_free_mtx);
1840 	if (curproc == pageproc) {
1841 		vm_pageout_pages_needed = 1;
1842 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1843 		    PDROP | PSWP, "VMWait", 0);
1844 	} else {
1845 		if (!vm_pages_needed) {
1846 			vm_pages_needed = 1;
1847 			wakeup(&vm_pages_needed);
1848 		}
1849 		msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1850 		    "vmwait", 0);
1851 	}
1852 }
1853 
1854 /*
1855  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
1856  *
1857  *	Sleep until free pages are available for allocation.
1858  *	- Called only in vm_fault so that processes page faulting
1859  *	  can be easily tracked.
1860  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1861  *	  processes will be able to grab memory first.  Do not change
1862  *	  this balance without careful testing first.
1863  */
1864 void
1865 vm_waitpfault(void)
1866 {
1867 
1868 	mtx_lock(&vm_page_queue_free_mtx);
1869 	if (!vm_pages_needed) {
1870 		vm_pages_needed = 1;
1871 		wakeup(&vm_pages_needed);
1872 	}
1873 	msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1874 	    "pfault", 0);
1875 }
1876 
1877 /*
1878  *	vm_page_dequeue:
1879  *
1880  *	Remove the given page from its current page queue.
1881  *
1882  *	The page must be locked.
1883  */
1884 void
1885 vm_page_dequeue(vm_page_t m)
1886 {
1887 	struct vm_pagequeue *pq;
1888 
1889 	vm_page_lock_assert(m, MA_OWNED);
1890 	KASSERT(m->queue != PQ_NONE,
1891 	    ("vm_page_dequeue: page %p is not queued", m));
1892 	pq = &vm_pagequeues[m->queue];
1893 	vm_pagequeue_lock(pq);
1894 	m->queue = PQ_NONE;
1895 	TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1896 	(*pq->pq_cnt)--;
1897 	vm_pagequeue_unlock(pq);
1898 }
1899 
1900 /*
1901  *	vm_page_dequeue_locked:
1902  *
1903  *	Remove the given page from its current page queue.
1904  *
1905  *	The page and page queue must be locked.
1906  */
1907 void
1908 vm_page_dequeue_locked(vm_page_t m)
1909 {
1910 	struct vm_pagequeue *pq;
1911 
1912 	vm_page_lock_assert(m, MA_OWNED);
1913 	pq = &vm_pagequeues[m->queue];
1914 	vm_pagequeue_assert_locked(pq);
1915 	m->queue = PQ_NONE;
1916 	TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1917 	(*pq->pq_cnt)--;
1918 }
1919 
1920 /*
1921  *	vm_page_enqueue:
1922  *
1923  *	Add the given page to the specified page queue.
1924  *
1925  *	The page must be locked.
1926  */
1927 static void
1928 vm_page_enqueue(int queue, vm_page_t m)
1929 {
1930 	struct vm_pagequeue *pq;
1931 
1932 	vm_page_lock_assert(m, MA_OWNED);
1933 	pq = &vm_pagequeues[queue];
1934 	vm_pagequeue_lock(pq);
1935 	m->queue = queue;
1936 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1937 	++*pq->pq_cnt;
1938 	vm_pagequeue_unlock(pq);
1939 }
1940 
1941 /*
1942  *	vm_page_requeue:
1943  *
1944  *	Move the given page to the tail of its current page queue.
1945  *
1946  *	The page must be locked.
1947  */
1948 void
1949 vm_page_requeue(vm_page_t m)
1950 {
1951 	struct vm_pagequeue *pq;
1952 
1953 	vm_page_lock_assert(m, MA_OWNED);
1954 	KASSERT(m->queue != PQ_NONE,
1955 	    ("vm_page_requeue: page %p is not queued", m));
1956 	pq = &vm_pagequeues[m->queue];
1957 	vm_pagequeue_lock(pq);
1958 	TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1959 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1960 	vm_pagequeue_unlock(pq);
1961 }
1962 
1963 /*
1964  *	vm_page_requeue_locked:
1965  *
1966  *	Move the given page to the tail of its current page queue.
1967  *
1968  *	The page queue must be locked.
1969  */
1970 void
1971 vm_page_requeue_locked(vm_page_t m)
1972 {
1973 	struct vm_pagequeue *pq;
1974 
1975 	KASSERT(m->queue != PQ_NONE,
1976 	    ("vm_page_requeue_locked: page %p is not queued", m));
1977 	pq = &vm_pagequeues[m->queue];
1978 	vm_pagequeue_assert_locked(pq);
1979 	TAILQ_REMOVE(&pq->pq_pl, m, pageq);
1980 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
1981 }
1982 
1983 /*
1984  *	vm_page_activate:
1985  *
1986  *	Put the specified page on the active list (if appropriate).
1987  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1988  *	mess with it.
1989  *
1990  *	The page must be locked.
1991  */
1992 void
1993 vm_page_activate(vm_page_t m)
1994 {
1995 	int queue;
1996 
1997 	vm_page_lock_assert(m, MA_OWNED);
1998 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1999 	if ((queue = m->queue) != PQ_ACTIVE) {
2000 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2001 			if (m->act_count < ACT_INIT)
2002 				m->act_count = ACT_INIT;
2003 			if (queue != PQ_NONE)
2004 				vm_page_dequeue(m);
2005 			vm_page_enqueue(PQ_ACTIVE, m);
2006 		} else
2007 			KASSERT(queue == PQ_NONE,
2008 			    ("vm_page_activate: wired page %p is queued", m));
2009 	} else {
2010 		if (m->act_count < ACT_INIT)
2011 			m->act_count = ACT_INIT;
2012 	}
2013 }
2014 
2015 /*
2016  *	vm_page_free_wakeup:
2017  *
2018  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
2019  *	routine is called when a page has been added to the cache or free
2020  *	queues.
2021  *
2022  *	The page queues must be locked.
2023  */
2024 static inline void
2025 vm_page_free_wakeup(void)
2026 {
2027 
2028 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2029 	/*
2030 	 * if pageout daemon needs pages, then tell it that there are
2031 	 * some free.
2032 	 */
2033 	if (vm_pageout_pages_needed &&
2034 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
2035 		wakeup(&vm_pageout_pages_needed);
2036 		vm_pageout_pages_needed = 0;
2037 	}
2038 	/*
2039 	 * wakeup processes that are waiting on memory if we hit a
2040 	 * high water mark. And wakeup scheduler process if we have
2041 	 * lots of memory. this process will swapin processes.
2042 	 */
2043 	if (vm_pages_needed && !vm_page_count_min()) {
2044 		vm_pages_needed = 0;
2045 		wakeup(&cnt.v_free_count);
2046 	}
2047 }
2048 
2049 /*
2050  *	vm_page_free_toq:
2051  *
2052  *	Returns the given page to the free list,
2053  *	disassociating it with any VM object.
2054  *
2055  *	The object must be locked.  The page must be locked if it is managed.
2056  */
2057 void
2058 vm_page_free_toq(vm_page_t m)
2059 {
2060 
2061 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2062 		vm_page_lock_assert(m, MA_OWNED);
2063 		KASSERT(!pmap_page_is_mapped(m),
2064 		    ("vm_page_free_toq: freeing mapped page %p", m));
2065 	} else
2066 		KASSERT(m->queue == PQ_NONE,
2067 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
2068 	PCPU_INC(cnt.v_tfree);
2069 
2070 	if (VM_PAGE_IS_FREE(m))
2071 		panic("vm_page_free: freeing free page %p", m);
2072 	else if (m->busy != 0)
2073 		panic("vm_page_free: freeing busy page %p", m);
2074 
2075 	/*
2076 	 * Unqueue, then remove page.  Note that we cannot destroy
2077 	 * the page here because we do not want to call the pager's
2078 	 * callback routine until after we've put the page on the
2079 	 * appropriate free queue.
2080 	 */
2081 	vm_page_remque(m);
2082 	vm_page_remove(m);
2083 
2084 	/*
2085 	 * If fictitious remove object association and
2086 	 * return, otherwise delay object association removal.
2087 	 */
2088 	if ((m->flags & PG_FICTITIOUS) != 0) {
2089 		return;
2090 	}
2091 
2092 	m->valid = 0;
2093 	vm_page_undirty(m);
2094 
2095 	if (m->wire_count != 0)
2096 		panic("vm_page_free: freeing wired page %p", m);
2097 	if (m->hold_count != 0) {
2098 		m->flags &= ~PG_ZERO;
2099 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2100 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2101 		m->flags |= PG_UNHOLDFREE;
2102 	} else {
2103 		/*
2104 		 * Restore the default memory attribute to the page.
2105 		 */
2106 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2107 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2108 
2109 		/*
2110 		 * Insert the page into the physical memory allocator's
2111 		 * cache/free page queues.
2112 		 */
2113 		mtx_lock(&vm_page_queue_free_mtx);
2114 		m->flags |= PG_FREE;
2115 		cnt.v_free_count++;
2116 #if VM_NRESERVLEVEL > 0
2117 		if (!vm_reserv_free_page(m))
2118 #else
2119 		if (TRUE)
2120 #endif
2121 			vm_phys_free_pages(m, 0);
2122 		if ((m->flags & PG_ZERO) != 0)
2123 			++vm_page_zero_count;
2124 		else
2125 			vm_page_zero_idle_wakeup();
2126 		vm_page_free_wakeup();
2127 		mtx_unlock(&vm_page_queue_free_mtx);
2128 	}
2129 }
2130 
2131 /*
2132  *	vm_page_wire:
2133  *
2134  *	Mark this page as wired down by yet
2135  *	another map, removing it from paging queues
2136  *	as necessary.
2137  *
2138  *	If the page is fictitious, then its wire count must remain one.
2139  *
2140  *	The page must be locked.
2141  */
2142 void
2143 vm_page_wire(vm_page_t m)
2144 {
2145 
2146 	/*
2147 	 * Only bump the wire statistics if the page is not already wired,
2148 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2149 	 * it is already off the queues).
2150 	 */
2151 	vm_page_lock_assert(m, MA_OWNED);
2152 	if ((m->flags & PG_FICTITIOUS) != 0) {
2153 		KASSERT(m->wire_count == 1,
2154 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2155 		    m));
2156 		return;
2157 	}
2158 	if (m->wire_count == 0) {
2159 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2160 		    m->queue == PQ_NONE,
2161 		    ("vm_page_wire: unmanaged page %p is queued", m));
2162 		vm_page_remque(m);
2163 		atomic_add_int(&cnt.v_wire_count, 1);
2164 	}
2165 	m->wire_count++;
2166 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2167 }
2168 
2169 /*
2170  * vm_page_unwire:
2171  *
2172  * Release one wiring of the specified page, potentially enabling it to be
2173  * paged again.  If paging is enabled, then the value of the parameter
2174  * "activate" determines to which queue the page is added.  If "activate" is
2175  * non-zero, then the page is added to the active queue.  Otherwise, it is
2176  * added to the inactive queue.
2177  *
2178  * However, unless the page belongs to an object, it is not enqueued because
2179  * it cannot be paged out.
2180  *
2181  * If a page is fictitious, then its wire count must alway be one.
2182  *
2183  * A managed page must be locked.
2184  */
2185 void
2186 vm_page_unwire(vm_page_t m, int activate)
2187 {
2188 
2189 	if ((m->oflags & VPO_UNMANAGED) == 0)
2190 		vm_page_lock_assert(m, MA_OWNED);
2191 	if ((m->flags & PG_FICTITIOUS) != 0) {
2192 		KASSERT(m->wire_count == 1,
2193 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2194 		return;
2195 	}
2196 	if (m->wire_count > 0) {
2197 		m->wire_count--;
2198 		if (m->wire_count == 0) {
2199 			atomic_subtract_int(&cnt.v_wire_count, 1);
2200 			if ((m->oflags & VPO_UNMANAGED) != 0 ||
2201 			    m->object == NULL)
2202 				return;
2203 			if (!activate)
2204 				m->flags &= ~PG_WINATCFLS;
2205 			vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2206 		}
2207 	} else
2208 		panic("vm_page_unwire: page %p's wire count is zero", m);
2209 }
2210 
2211 /*
2212  * Move the specified page to the inactive queue.
2213  *
2214  * Many pages placed on the inactive queue should actually go
2215  * into the cache, but it is difficult to figure out which.  What
2216  * we do instead, if the inactive target is well met, is to put
2217  * clean pages at the head of the inactive queue instead of the tail.
2218  * This will cause them to be moved to the cache more quickly and
2219  * if not actively re-referenced, reclaimed more quickly.  If we just
2220  * stick these pages at the end of the inactive queue, heavy filesystem
2221  * meta-data accesses can cause an unnecessary paging load on memory bound
2222  * processes.  This optimization causes one-time-use metadata to be
2223  * reused more quickly.
2224  *
2225  * Normally athead is 0 resulting in LRU operation.  athead is set
2226  * to 1 if we want this page to be 'as if it were placed in the cache',
2227  * except without unmapping it from the process address space.
2228  *
2229  * The page must be locked.
2230  */
2231 static inline void
2232 _vm_page_deactivate(vm_page_t m, int athead)
2233 {
2234 	struct vm_pagequeue *pq;
2235 	int queue;
2236 
2237 	vm_page_lock_assert(m, MA_OWNED);
2238 
2239 	/*
2240 	 * Ignore if already inactive.
2241 	 */
2242 	if ((queue = m->queue) == PQ_INACTIVE)
2243 		return;
2244 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2245 		if (queue != PQ_NONE)
2246 			vm_page_dequeue(m);
2247 		m->flags &= ~PG_WINATCFLS;
2248 		pq = &vm_pagequeues[PQ_INACTIVE];
2249 		vm_pagequeue_lock(pq);
2250 		m->queue = PQ_INACTIVE;
2251 		if (athead)
2252 			TAILQ_INSERT_HEAD(&pq->pq_pl, m, pageq);
2253 		else
2254 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq);
2255 		cnt.v_inactive_count++;
2256 		vm_pagequeue_unlock(pq);
2257 	}
2258 }
2259 
2260 /*
2261  * Move the specified page to the inactive queue.
2262  *
2263  * The page must be locked.
2264  */
2265 void
2266 vm_page_deactivate(vm_page_t m)
2267 {
2268 
2269 	_vm_page_deactivate(m, 0);
2270 }
2271 
2272 /*
2273  * vm_page_try_to_cache:
2274  *
2275  * Returns 0 on failure, 1 on success
2276  */
2277 int
2278 vm_page_try_to_cache(vm_page_t m)
2279 {
2280 
2281 	vm_page_lock_assert(m, MA_OWNED);
2282 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2283 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2284 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2285 		return (0);
2286 	pmap_remove_all(m);
2287 	if (m->dirty)
2288 		return (0);
2289 	vm_page_cache(m);
2290 	return (1);
2291 }
2292 
2293 /*
2294  * vm_page_try_to_free()
2295  *
2296  *	Attempt to free the page.  If we cannot free it, we do nothing.
2297  *	1 is returned on success, 0 on failure.
2298  */
2299 int
2300 vm_page_try_to_free(vm_page_t m)
2301 {
2302 
2303 	vm_page_lock_assert(m, MA_OWNED);
2304 	if (m->object != NULL)
2305 		VM_OBJECT_ASSERT_WLOCKED(m->object);
2306 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2307 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2308 		return (0);
2309 	pmap_remove_all(m);
2310 	if (m->dirty)
2311 		return (0);
2312 	vm_page_free(m);
2313 	return (1);
2314 }
2315 
2316 /*
2317  * vm_page_cache
2318  *
2319  * Put the specified page onto the page cache queue (if appropriate).
2320  *
2321  * The object and page must be locked.
2322  */
2323 void
2324 vm_page_cache(vm_page_t m)
2325 {
2326 	vm_object_t object;
2327 	vm_page_t next, prev, root;
2328 
2329 	vm_page_lock_assert(m, MA_OWNED);
2330 	object = m->object;
2331 	VM_OBJECT_ASSERT_WLOCKED(object);
2332 	if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2333 	    m->hold_count || m->wire_count)
2334 		panic("vm_page_cache: attempting to cache busy page");
2335 	KASSERT(!pmap_page_is_mapped(m),
2336 	    ("vm_page_cache: page %p is mapped", m));
2337 	KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2338 	if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2339 	    (object->type == OBJT_SWAP &&
2340 	    !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2341 		/*
2342 		 * Hypothesis: A cache-elgible page belonging to a
2343 		 * default object or swap object but without a backing
2344 		 * store must be zero filled.
2345 		 */
2346 		vm_page_free(m);
2347 		return;
2348 	}
2349 	KASSERT((m->flags & PG_CACHED) == 0,
2350 	    ("vm_page_cache: page %p is already cached", m));
2351 	PCPU_INC(cnt.v_tcached);
2352 
2353 	/*
2354 	 * Remove the page from the paging queues.
2355 	 */
2356 	vm_page_remque(m);
2357 
2358 	/*
2359 	 * Remove the page from the object's collection of resident
2360 	 * pages.
2361 	 */
2362 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
2363 		/*
2364 		 * Since the page's successor in the list is also its parent
2365 		 * in the tree, its right subtree must be empty.
2366 		 */
2367 		next->left = m->left;
2368 		KASSERT(m->right == NULL,
2369 		    ("vm_page_cache: page %p has right child", m));
2370 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
2371 	    prev->right == m) {
2372 		/*
2373 		 * Since the page's predecessor in the list is also its parent
2374 		 * in the tree, its left subtree must be empty.
2375 		 */
2376 		KASSERT(m->left == NULL,
2377 		    ("vm_page_cache: page %p has left child", m));
2378 		prev->right = m->right;
2379 	} else {
2380 		if (m != object->root)
2381 			vm_page_splay(m->pindex, object->root);
2382 		if (m->left == NULL)
2383 			root = m->right;
2384 		else if (m->right == NULL)
2385 			root = m->left;
2386 		else {
2387 			/*
2388 			 * Move the page's successor to the root, because
2389 			 * pages are usually removed in ascending order.
2390 			 */
2391 			if (m->right != next)
2392 				vm_page_splay(m->pindex, m->right);
2393 			next->left = m->left;
2394 			root = next;
2395 		}
2396 		object->root = root;
2397 	}
2398 	TAILQ_REMOVE(&object->memq, m, listq);
2399 	object->resident_page_count--;
2400 
2401 	/*
2402 	 * Restore the default memory attribute to the page.
2403 	 */
2404 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2405 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2406 
2407 	/*
2408 	 * Insert the page into the object's collection of cached pages
2409 	 * and the physical memory allocator's cache/free page queues.
2410 	 */
2411 	m->flags &= ~PG_ZERO;
2412 	mtx_lock(&vm_page_queue_free_mtx);
2413 	m->flags |= PG_CACHED;
2414 	cnt.v_cache_count++;
2415 	root = object->cache;
2416 	if (root == NULL) {
2417 		m->left = NULL;
2418 		m->right = NULL;
2419 	} else {
2420 		root = vm_page_splay(m->pindex, root);
2421 		if (m->pindex < root->pindex) {
2422 			m->left = root->left;
2423 			m->right = root;
2424 			root->left = NULL;
2425 		} else if (__predict_false(m->pindex == root->pindex))
2426 			panic("vm_page_cache: offset already cached");
2427 		else {
2428 			m->right = root->right;
2429 			m->left = root;
2430 			root->right = NULL;
2431 		}
2432 	}
2433 	object->cache = m;
2434 #if VM_NRESERVLEVEL > 0
2435 	if (!vm_reserv_free_page(m)) {
2436 #else
2437 	if (TRUE) {
2438 #endif
2439 		vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2440 		vm_phys_free_pages(m, 0);
2441 	}
2442 	vm_page_free_wakeup();
2443 	mtx_unlock(&vm_page_queue_free_mtx);
2444 
2445 	/*
2446 	 * Increment the vnode's hold count if this is the object's only
2447 	 * cached page.  Decrement the vnode's hold count if this was
2448 	 * the object's only resident page.
2449 	 */
2450 	if (object->type == OBJT_VNODE) {
2451 		if (root == NULL && object->resident_page_count != 0)
2452 			vhold(object->handle);
2453 		else if (root != NULL && object->resident_page_count == 0)
2454 			vdrop(object->handle);
2455 	}
2456 }
2457 
2458 /*
2459  * vm_page_dontneed
2460  *
2461  *	Cache, deactivate, or do nothing as appropriate.  This routine
2462  *	is typically used by madvise() MADV_DONTNEED.
2463  *
2464  *	Generally speaking we want to move the page into the cache so
2465  *	it gets reused quickly.  However, this can result in a silly syndrome
2466  *	due to the page recycling too quickly.  Small objects will not be
2467  *	fully cached.  On the otherhand, if we move the page to the inactive
2468  *	queue we wind up with a problem whereby very large objects
2469  *	unnecessarily blow away our inactive and cache queues.
2470  *
2471  *	The solution is to move the pages based on a fixed weighting.  We
2472  *	either leave them alone, deactivate them, or move them to the cache,
2473  *	where moving them to the cache has the highest weighting.
2474  *	By forcing some pages into other queues we eventually force the
2475  *	system to balance the queues, potentially recovering other unrelated
2476  *	space from active.  The idea is to not force this to happen too
2477  *	often.
2478  *
2479  *	The object and page must be locked.
2480  */
2481 void
2482 vm_page_dontneed(vm_page_t m)
2483 {
2484 	int dnw;
2485 	int head;
2486 
2487 	vm_page_lock_assert(m, MA_OWNED);
2488 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2489 	dnw = PCPU_GET(dnweight);
2490 	PCPU_INC(dnweight);
2491 
2492 	/*
2493 	 * Occasionally leave the page alone.
2494 	 */
2495 	if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2496 		if (m->act_count >= ACT_INIT)
2497 			--m->act_count;
2498 		return;
2499 	}
2500 
2501 	/*
2502 	 * Clear any references to the page.  Otherwise, the page daemon will
2503 	 * immediately reactivate the page.
2504 	 *
2505 	 * Perform the pmap_clear_reference() first.  Otherwise, a concurrent
2506 	 * pmap operation, such as pmap_remove(), could clear a reference in
2507 	 * the pmap and set PGA_REFERENCED on the page before the
2508 	 * pmap_clear_reference() had completed.  Consequently, the page would
2509 	 * appear referenced based upon an old reference that occurred before
2510 	 * this function ran.
2511 	 */
2512 	pmap_clear_reference(m);
2513 	vm_page_aflag_clear(m, PGA_REFERENCED);
2514 
2515 	if (m->dirty == 0 && pmap_is_modified(m))
2516 		vm_page_dirty(m);
2517 
2518 	if (m->dirty || (dnw & 0x0070) == 0) {
2519 		/*
2520 		 * Deactivate the page 3 times out of 32.
2521 		 */
2522 		head = 0;
2523 	} else {
2524 		/*
2525 		 * Cache the page 28 times out of every 32.  Note that
2526 		 * the page is deactivated instead of cached, but placed
2527 		 * at the head of the queue instead of the tail.
2528 		 */
2529 		head = 1;
2530 	}
2531 	_vm_page_deactivate(m, head);
2532 }
2533 
2534 /*
2535  * Grab a page, waiting until we are waken up due to the page
2536  * changing state.  We keep on waiting, if the page continues
2537  * to be in the object.  If the page doesn't exist, first allocate it
2538  * and then conditionally zero it.
2539  *
2540  * The caller must always specify the VM_ALLOC_RETRY flag.  This is intended
2541  * to facilitate its eventual removal.
2542  *
2543  * This routine may sleep.
2544  *
2545  * The object must be locked on entry.  The lock will, however, be released
2546  * and reacquired if the routine sleeps.
2547  */
2548 vm_page_t
2549 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2550 {
2551 	vm_page_t m;
2552 
2553 	VM_OBJECT_ASSERT_WLOCKED(object);
2554 	KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2555 	    ("vm_page_grab: VM_ALLOC_RETRY is required"));
2556 retrylookup:
2557 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
2558 		if ((m->oflags & VPO_BUSY) != 0 ||
2559 		    ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2560 			/*
2561 			 * Reference the page before unlocking and
2562 			 * sleeping so that the page daemon is less
2563 			 * likely to reclaim it.
2564 			 */
2565 			vm_page_aflag_set(m, PGA_REFERENCED);
2566 			vm_page_sleep(m, "pgrbwt");
2567 			goto retrylookup;
2568 		} else {
2569 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
2570 				vm_page_lock(m);
2571 				vm_page_wire(m);
2572 				vm_page_unlock(m);
2573 			}
2574 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2575 				vm_page_busy(m);
2576 			return (m);
2577 		}
2578 	}
2579 	m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2580 	    VM_ALLOC_IGN_SBUSY));
2581 	if (m == NULL) {
2582 		VM_OBJECT_WUNLOCK(object);
2583 		VM_WAIT;
2584 		VM_OBJECT_WLOCK(object);
2585 		goto retrylookup;
2586 	} else if (m->valid != 0)
2587 		return (m);
2588 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2589 		pmap_zero_page(m);
2590 	return (m);
2591 }
2592 
2593 /*
2594  * Mapping function for valid or dirty bits in a page.
2595  *
2596  * Inputs are required to range within a page.
2597  */
2598 vm_page_bits_t
2599 vm_page_bits(int base, int size)
2600 {
2601 	int first_bit;
2602 	int last_bit;
2603 
2604 	KASSERT(
2605 	    base + size <= PAGE_SIZE,
2606 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
2607 	);
2608 
2609 	if (size == 0)		/* handle degenerate case */
2610 		return (0);
2611 
2612 	first_bit = base >> DEV_BSHIFT;
2613 	last_bit = (base + size - 1) >> DEV_BSHIFT;
2614 
2615 	return (((vm_page_bits_t)2 << last_bit) -
2616 	    ((vm_page_bits_t)1 << first_bit));
2617 }
2618 
2619 /*
2620  *	vm_page_set_valid_range:
2621  *
2622  *	Sets portions of a page valid.  The arguments are expected
2623  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2624  *	of any partial chunks touched by the range.  The invalid portion of
2625  *	such chunks will be zeroed.
2626  *
2627  *	(base + size) must be less then or equal to PAGE_SIZE.
2628  */
2629 void
2630 vm_page_set_valid_range(vm_page_t m, int base, int size)
2631 {
2632 	int endoff, frag;
2633 
2634 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2635 	if (size == 0)	/* handle degenerate case */
2636 		return;
2637 
2638 	/*
2639 	 * If the base is not DEV_BSIZE aligned and the valid
2640 	 * bit is clear, we have to zero out a portion of the
2641 	 * first block.
2642 	 */
2643 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2644 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2645 		pmap_zero_page_area(m, frag, base - frag);
2646 
2647 	/*
2648 	 * If the ending offset is not DEV_BSIZE aligned and the
2649 	 * valid bit is clear, we have to zero out a portion of
2650 	 * the last block.
2651 	 */
2652 	endoff = base + size;
2653 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2654 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2655 		pmap_zero_page_area(m, endoff,
2656 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2657 
2658 	/*
2659 	 * Assert that no previously invalid block that is now being validated
2660 	 * is already dirty.
2661 	 */
2662 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2663 	    ("vm_page_set_valid_range: page %p is dirty", m));
2664 
2665 	/*
2666 	 * Set valid bits inclusive of any overlap.
2667 	 */
2668 	m->valid |= vm_page_bits(base, size);
2669 }
2670 
2671 /*
2672  * Clear the given bits from the specified page's dirty field.
2673  */
2674 static __inline void
2675 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2676 {
2677 	uintptr_t addr;
2678 #if PAGE_SIZE < 16384
2679 	int shift;
2680 #endif
2681 
2682 	/*
2683 	 * If the object is locked and the page is neither VPO_BUSY nor
2684 	 * write mapped, then the page's dirty field cannot possibly be
2685 	 * set by a concurrent pmap operation.
2686 	 */
2687 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2688 	if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m))
2689 		m->dirty &= ~pagebits;
2690 	else {
2691 		/*
2692 		 * The pmap layer can call vm_page_dirty() without
2693 		 * holding a distinguished lock.  The combination of
2694 		 * the object's lock and an atomic operation suffice
2695 		 * to guarantee consistency of the page dirty field.
2696 		 *
2697 		 * For PAGE_SIZE == 32768 case, compiler already
2698 		 * properly aligns the dirty field, so no forcible
2699 		 * alignment is needed. Only require existence of
2700 		 * atomic_clear_64 when page size is 32768.
2701 		 */
2702 		addr = (uintptr_t)&m->dirty;
2703 #if PAGE_SIZE == 32768
2704 		atomic_clear_64((uint64_t *)addr, pagebits);
2705 #elif PAGE_SIZE == 16384
2706 		atomic_clear_32((uint32_t *)addr, pagebits);
2707 #else		/* PAGE_SIZE <= 8192 */
2708 		/*
2709 		 * Use a trick to perform a 32-bit atomic on the
2710 		 * containing aligned word, to not depend on the existence
2711 		 * of atomic_clear_{8, 16}.
2712 		 */
2713 		shift = addr & (sizeof(uint32_t) - 1);
2714 #if BYTE_ORDER == BIG_ENDIAN
2715 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2716 #else
2717 		shift *= NBBY;
2718 #endif
2719 		addr &= ~(sizeof(uint32_t) - 1);
2720 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
2721 #endif		/* PAGE_SIZE */
2722 	}
2723 }
2724 
2725 /*
2726  *	vm_page_set_validclean:
2727  *
2728  *	Sets portions of a page valid and clean.  The arguments are expected
2729  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2730  *	of any partial chunks touched by the range.  The invalid portion of
2731  *	such chunks will be zero'd.
2732  *
2733  *	(base + size) must be less then or equal to PAGE_SIZE.
2734  */
2735 void
2736 vm_page_set_validclean(vm_page_t m, int base, int size)
2737 {
2738 	vm_page_bits_t oldvalid, pagebits;
2739 	int endoff, frag;
2740 
2741 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2742 	if (size == 0)	/* handle degenerate case */
2743 		return;
2744 
2745 	/*
2746 	 * If the base is not DEV_BSIZE aligned and the valid
2747 	 * bit is clear, we have to zero out a portion of the
2748 	 * first block.
2749 	 */
2750 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2751 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2752 		pmap_zero_page_area(m, frag, base - frag);
2753 
2754 	/*
2755 	 * If the ending offset is not DEV_BSIZE aligned and the
2756 	 * valid bit is clear, we have to zero out a portion of
2757 	 * the last block.
2758 	 */
2759 	endoff = base + size;
2760 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2761 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2762 		pmap_zero_page_area(m, endoff,
2763 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2764 
2765 	/*
2766 	 * Set valid, clear dirty bits.  If validating the entire
2767 	 * page we can safely clear the pmap modify bit.  We also
2768 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
2769 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
2770 	 * be set again.
2771 	 *
2772 	 * We set valid bits inclusive of any overlap, but we can only
2773 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
2774 	 * the range.
2775 	 */
2776 	oldvalid = m->valid;
2777 	pagebits = vm_page_bits(base, size);
2778 	m->valid |= pagebits;
2779 #if 0	/* NOT YET */
2780 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2781 		frag = DEV_BSIZE - frag;
2782 		base += frag;
2783 		size -= frag;
2784 		if (size < 0)
2785 			size = 0;
2786 	}
2787 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2788 #endif
2789 	if (base == 0 && size == PAGE_SIZE) {
2790 		/*
2791 		 * The page can only be modified within the pmap if it is
2792 		 * mapped, and it can only be mapped if it was previously
2793 		 * fully valid.
2794 		 */
2795 		if (oldvalid == VM_PAGE_BITS_ALL)
2796 			/*
2797 			 * Perform the pmap_clear_modify() first.  Otherwise,
2798 			 * a concurrent pmap operation, such as
2799 			 * pmap_protect(), could clear a modification in the
2800 			 * pmap and set the dirty field on the page before
2801 			 * pmap_clear_modify() had begun and after the dirty
2802 			 * field was cleared here.
2803 			 */
2804 			pmap_clear_modify(m);
2805 		m->dirty = 0;
2806 		m->oflags &= ~VPO_NOSYNC;
2807 	} else if (oldvalid != VM_PAGE_BITS_ALL)
2808 		m->dirty &= ~pagebits;
2809 	else
2810 		vm_page_clear_dirty_mask(m, pagebits);
2811 }
2812 
2813 void
2814 vm_page_clear_dirty(vm_page_t m, int base, int size)
2815 {
2816 
2817 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2818 }
2819 
2820 /*
2821  *	vm_page_set_invalid:
2822  *
2823  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
2824  *	valid and dirty bits for the effected areas are cleared.
2825  */
2826 void
2827 vm_page_set_invalid(vm_page_t m, int base, int size)
2828 {
2829 	vm_page_bits_t bits;
2830 
2831 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2832 	KASSERT((m->oflags & VPO_BUSY) == 0,
2833 	    ("vm_page_set_invalid: page %p is busy", m));
2834 	bits = vm_page_bits(base, size);
2835 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2836 		pmap_remove_all(m);
2837 	KASSERT(!pmap_page_is_mapped(m),
2838 	    ("vm_page_set_invalid: page %p is mapped", m));
2839 	m->valid &= ~bits;
2840 	m->dirty &= ~bits;
2841 }
2842 
2843 /*
2844  * vm_page_zero_invalid()
2845  *
2846  *	The kernel assumes that the invalid portions of a page contain
2847  *	garbage, but such pages can be mapped into memory by user code.
2848  *	When this occurs, we must zero out the non-valid portions of the
2849  *	page so user code sees what it expects.
2850  *
2851  *	Pages are most often semi-valid when the end of a file is mapped
2852  *	into memory and the file's size is not page aligned.
2853  */
2854 void
2855 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2856 {
2857 	int b;
2858 	int i;
2859 
2860 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2861 	/*
2862 	 * Scan the valid bits looking for invalid sections that
2863 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
2864 	 * valid bit may be set ) have already been zerod by
2865 	 * vm_page_set_validclean().
2866 	 */
2867 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2868 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
2869 		    (m->valid & ((vm_page_bits_t)1 << i))) {
2870 			if (i > b) {
2871 				pmap_zero_page_area(m,
2872 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2873 			}
2874 			b = i + 1;
2875 		}
2876 	}
2877 
2878 	/*
2879 	 * setvalid is TRUE when we can safely set the zero'd areas
2880 	 * as being valid.  We can do this if there are no cache consistancy
2881 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
2882 	 */
2883 	if (setvalid)
2884 		m->valid = VM_PAGE_BITS_ALL;
2885 }
2886 
2887 /*
2888  *	vm_page_is_valid:
2889  *
2890  *	Is (partial) page valid?  Note that the case where size == 0
2891  *	will return FALSE in the degenerate case where the page is
2892  *	entirely invalid, and TRUE otherwise.
2893  */
2894 int
2895 vm_page_is_valid(vm_page_t m, int base, int size)
2896 {
2897 	vm_page_bits_t bits;
2898 
2899 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2900 	bits = vm_page_bits(base, size);
2901 	return (m->valid != 0 && (m->valid & bits) == bits);
2902 }
2903 
2904 /*
2905  * Set the page's dirty bits if the page is modified.
2906  */
2907 void
2908 vm_page_test_dirty(vm_page_t m)
2909 {
2910 
2911 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2912 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2913 		vm_page_dirty(m);
2914 }
2915 
2916 void
2917 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2918 {
2919 
2920 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2921 }
2922 
2923 void
2924 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2925 {
2926 
2927 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2928 }
2929 
2930 int
2931 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2932 {
2933 
2934 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2935 }
2936 
2937 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2938 void
2939 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2940 {
2941 
2942 	mtx_assert_(vm_page_lockptr(m), a, file, line);
2943 }
2944 #endif
2945 
2946 int so_zerocp_fullpage = 0;
2947 
2948 /*
2949  *	Replace the given page with a copy.  The copied page assumes
2950  *	the portion of the given page's "wire_count" that is not the
2951  *	responsibility of this copy-on-write mechanism.
2952  *
2953  *	The object containing the given page must have a non-zero
2954  *	paging-in-progress count and be locked.
2955  */
2956 void
2957 vm_page_cowfault(vm_page_t m)
2958 {
2959 	vm_page_t mnew;
2960 	vm_object_t object;
2961 	vm_pindex_t pindex;
2962 
2963 	vm_page_lock_assert(m, MA_OWNED);
2964 	object = m->object;
2965 	VM_OBJECT_ASSERT_WLOCKED(object);
2966 	KASSERT(object->paging_in_progress != 0,
2967 	    ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2968 	    object));
2969 	pindex = m->pindex;
2970 
2971  retry_alloc:
2972 	pmap_remove_all(m);
2973 	vm_page_remove(m);
2974 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2975 	if (mnew == NULL) {
2976 		vm_page_insert(m, object, pindex);
2977 		vm_page_unlock(m);
2978 		VM_OBJECT_WUNLOCK(object);
2979 		VM_WAIT;
2980 		VM_OBJECT_WLOCK(object);
2981 		if (m == vm_page_lookup(object, pindex)) {
2982 			vm_page_lock(m);
2983 			goto retry_alloc;
2984 		} else {
2985 			/*
2986 			 * Page disappeared during the wait.
2987 			 */
2988 			return;
2989 		}
2990 	}
2991 
2992 	if (m->cow == 0) {
2993 		/*
2994 		 * check to see if we raced with an xmit complete when
2995 		 * waiting to allocate a page.  If so, put things back
2996 		 * the way they were
2997 		 */
2998 		vm_page_unlock(m);
2999 		vm_page_lock(mnew);
3000 		vm_page_free(mnew);
3001 		vm_page_unlock(mnew);
3002 		vm_page_insert(m, object, pindex);
3003 	} else { /* clear COW & copy page */
3004 		if (!so_zerocp_fullpage)
3005 			pmap_copy_page(m, mnew);
3006 		mnew->valid = VM_PAGE_BITS_ALL;
3007 		vm_page_dirty(mnew);
3008 		mnew->wire_count = m->wire_count - m->cow;
3009 		m->wire_count = m->cow;
3010 		vm_page_unlock(m);
3011 	}
3012 }
3013 
3014 void
3015 vm_page_cowclear(vm_page_t m)
3016 {
3017 
3018 	vm_page_lock_assert(m, MA_OWNED);
3019 	if (m->cow) {
3020 		m->cow--;
3021 		/*
3022 		 * let vm_fault add back write permission  lazily
3023 		 */
3024 	}
3025 	/*
3026 	 *  sf_buf_free() will free the page, so we needn't do it here
3027 	 */
3028 }
3029 
3030 int
3031 vm_page_cowsetup(vm_page_t m)
3032 {
3033 
3034 	vm_page_lock_assert(m, MA_OWNED);
3035 	if ((m->flags & PG_FICTITIOUS) != 0 ||
3036 	    (m->oflags & VPO_UNMANAGED) != 0 ||
3037 	    m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYWLOCK(m->object))
3038 		return (EBUSY);
3039 	m->cow++;
3040 	pmap_remove_write(m);
3041 	VM_OBJECT_WUNLOCK(m->object);
3042 	return (0);
3043 }
3044 
3045 #ifdef INVARIANTS
3046 void
3047 vm_page_object_lock_assert(vm_page_t m)
3048 {
3049 
3050 	/*
3051 	 * Certain of the page's fields may only be modified by the
3052 	 * holder of the containing object's lock or the setter of the
3053 	 * page's VPO_BUSY flag.  Unfortunately, the setter of the
3054 	 * VPO_BUSY flag is not recorded, and thus cannot be checked
3055 	 * here.
3056 	 */
3057 	if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
3058 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3059 }
3060 #endif
3061 
3062 #include "opt_ddb.h"
3063 #ifdef DDB
3064 #include <sys/kernel.h>
3065 
3066 #include <ddb/ddb.h>
3067 
3068 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3069 {
3070 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
3071 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
3072 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
3073 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
3074 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
3075 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
3076 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
3077 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
3078 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
3079 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
3080 }
3081 
3082 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3083 {
3084 
3085 	db_printf("PQ_FREE:");
3086 	db_printf(" %d", cnt.v_free_count);
3087 	db_printf("\n");
3088 
3089 	db_printf("PQ_CACHE:");
3090 	db_printf(" %d", cnt.v_cache_count);
3091 	db_printf("\n");
3092 
3093 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
3094 		*vm_pagequeues[PQ_ACTIVE].pq_cnt,
3095 		*vm_pagequeues[PQ_INACTIVE].pq_cnt);
3096 }
3097 #endif /* DDB */
3098