xref: /freebsd/sys/vm/vm_page.c (revision 21d30ec18d673020b701750a4d498a4e75cde516)
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 pageq mutex is required when adding or removing a page from a
67  *	  page queue (vm_page_queue[]), regardless of other mutexes or the
68  *	  busy state of a page.
69  *
70  *	- a hash chain mutex is required when associating or disassociating
71  *	  a page from the VM PAGE CACHE hash table (vm_page_buckets),
72  *	  regardless of other mutexes or the busy state of a page.
73  *
74  *	- either a hash chain mutex OR a busied page is required in order
75  *	  to modify the page flags.  A hash chain mutex must be obtained in
76  *	  order to busy a page.  A page's flags cannot be modified by a
77  *	  hash chain mutex if the page is marked busy.
78  *
79  *	- The object memq mutex is held when inserting or removing
80  *	  pages from an object (vm_page_insert() or vm_page_remove()).  This
81  *	  is different from the object's main mutex.
82  *
83  *	Generally speaking, you have to be aware of side effects when running
84  *	vm_page ops.  A vm_page_lookup() will return with the hash chain
85  *	locked, whether it was able to lookup the page or not.  vm_page_free(),
86  *	vm_page_cache(), vm_page_activate(), and a number of other routines
87  *	will release the hash chain mutex for you.  Intermediate manipulation
88  *	routines such as vm_page_flag_set() expect the hash chain to be held
89  *	on entry and the hash chain will remain held on return.
90  *
91  *	pageq scanning can only occur with the pageq in question locked.
92  *	We have a known bottleneck with the active queue, but the cache
93  *	and free queues are actually arrays already.
94  */
95 
96 /*
97  *	Resident memory management module.
98  */
99 
100 #include <sys/cdefs.h>
101 __FBSDID("$FreeBSD$");
102 
103 #include "opt_vm.h"
104 
105 #include <sys/param.h>
106 #include <sys/systm.h>
107 #include <sys/lock.h>
108 #include <sys/kernel.h>
109 #include <sys/limits.h>
110 #include <sys/malloc.h>
111 #include <sys/msgbuf.h>
112 #include <sys/mutex.h>
113 #include <sys/proc.h>
114 #include <sys/sysctl.h>
115 #include <sys/vmmeter.h>
116 #include <sys/vnode.h>
117 
118 #include <vm/vm.h>
119 #include <vm/pmap.h>
120 #include <vm/vm_param.h>
121 #include <vm/vm_kern.h>
122 #include <vm/vm_object.h>
123 #include <vm/vm_page.h>
124 #include <vm/vm_pageout.h>
125 #include <vm/vm_pager.h>
126 #include <vm/vm_phys.h>
127 #include <vm/vm_reserv.h>
128 #include <vm/vm_extern.h>
129 #include <vm/uma.h>
130 #include <vm/uma_int.h>
131 
132 #include <machine/md_var.h>
133 
134 #if defined(__amd64__) || defined (__i386__)
135 extern struct sysctl_oid_list sysctl__vm_pmap_children;
136 #else
137 SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters");
138 #endif
139 
140 static uint64_t pmap_tryrelock_calls;
141 SYSCTL_QUAD(_vm_pmap, OID_AUTO, tryrelock_calls, CTLFLAG_RD,
142     &pmap_tryrelock_calls, 0, "Number of tryrelock calls");
143 
144 static int pmap_tryrelock_restart;
145 SYSCTL_INT(_vm_pmap, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
146     &pmap_tryrelock_restart, 0, "Number of tryrelock restarts");
147 
148 static int pmap_tryrelock_race;
149 SYSCTL_INT(_vm_pmap, OID_AUTO, tryrelock_race, CTLFLAG_RD,
150     &pmap_tryrelock_race, 0, "Number of tryrelock pmap race cases");
151 
152 /*
153  *	Associated with page of user-allocatable memory is a
154  *	page structure.
155  */
156 
157 struct vpgqueues vm_page_queues[PQ_COUNT];
158 struct vpglocks vm_page_queue_lock;
159 struct vpglocks vm_page_queue_free_lock;
160 
161 struct vpglocks	pa_lock[PA_LOCK_COUNT] __aligned(CACHE_LINE_SIZE);
162 
163 vm_page_t vm_page_array = 0;
164 int vm_page_array_size = 0;
165 long first_page = 0;
166 int vm_page_zero_count = 0;
167 
168 static int boot_pages = UMA_BOOT_PAGES;
169 TUNABLE_INT("vm.boot_pages", &boot_pages);
170 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
171 	"number of pages allocated for bootstrapping the VM system");
172 
173 static void vm_page_clear_dirty_mask(vm_page_t m, int pagebits);
174 static void vm_page_queue_remove(int queue, vm_page_t m);
175 static void vm_page_enqueue(int queue, vm_page_t m);
176 
177 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
178 #if PAGE_SIZE == 32768
179 #ifdef CTASSERT
180 CTASSERT(sizeof(u_long) >= 8);
181 #endif
182 #endif
183 
184 /*
185  * Try to acquire a physical address lock while a pmap is locked.  If we
186  * fail to trylock we unlock and lock the pmap directly and cache the
187  * locked pa in *locked.  The caller should then restart their loop in case
188  * the virtual to physical mapping has changed.
189  */
190 int
191 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
192 {
193 	vm_paddr_t lockpa;
194 	uint32_t gen_count;
195 
196 	gen_count = pmap->pm_gen_count;
197 	atomic_add_long((volatile long *)&pmap_tryrelock_calls, 1);
198 	lockpa = *locked;
199 	*locked = pa;
200 	if (lockpa) {
201 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
202 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
203 			return (0);
204 		PA_UNLOCK(lockpa);
205 	}
206 	if (PA_TRYLOCK(pa))
207 		return (0);
208 	PMAP_UNLOCK(pmap);
209 	atomic_add_int((volatile int *)&pmap_tryrelock_restart, 1);
210 	PA_LOCK(pa);
211 	PMAP_LOCK(pmap);
212 
213 	if (pmap->pm_gen_count != gen_count + 1) {
214 		pmap->pm_retries++;
215 		atomic_add_int((volatile int *)&pmap_tryrelock_race, 1);
216 		return (EAGAIN);
217 	}
218 	return (0);
219 }
220 
221 /*
222  *	vm_set_page_size:
223  *
224  *	Sets the page size, perhaps based upon the memory
225  *	size.  Must be called before any use of page-size
226  *	dependent functions.
227  */
228 void
229 vm_set_page_size(void)
230 {
231 	if (cnt.v_page_size == 0)
232 		cnt.v_page_size = PAGE_SIZE;
233 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
234 		panic("vm_set_page_size: page size not a power of two");
235 }
236 
237 /*
238  *	vm_page_blacklist_lookup:
239  *
240  *	See if a physical address in this page has been listed
241  *	in the blacklist tunable.  Entries in the tunable are
242  *	separated by spaces or commas.  If an invalid integer is
243  *	encountered then the rest of the string is skipped.
244  */
245 static int
246 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
247 {
248 	vm_paddr_t bad;
249 	char *cp, *pos;
250 
251 	for (pos = list; *pos != '\0'; pos = cp) {
252 		bad = strtoq(pos, &cp, 0);
253 		if (*cp != '\0') {
254 			if (*cp == ' ' || *cp == ',') {
255 				cp++;
256 				if (cp == pos)
257 					continue;
258 			} else
259 				break;
260 		}
261 		if (pa == trunc_page(bad))
262 			return (1);
263 	}
264 	return (0);
265 }
266 
267 /*
268  *	vm_page_startup:
269  *
270  *	Initializes the resident memory module.
271  *
272  *	Allocates memory for the page cells, and
273  *	for the object/offset-to-page hash table headers.
274  *	Each page cell is initialized and placed on the free list.
275  */
276 vm_offset_t
277 vm_page_startup(vm_offset_t vaddr)
278 {
279 	vm_offset_t mapped;
280 	vm_paddr_t page_range;
281 	vm_paddr_t new_end;
282 	int i;
283 	vm_paddr_t pa;
284 	int nblocks;
285 	vm_paddr_t last_pa;
286 	char *list;
287 
288 	/* the biggest memory array is the second group of pages */
289 	vm_paddr_t end;
290 	vm_paddr_t biggestsize;
291 	vm_paddr_t low_water, high_water;
292 	int biggestone;
293 
294 	biggestsize = 0;
295 	biggestone = 0;
296 	nblocks = 0;
297 	vaddr = round_page(vaddr);
298 
299 	for (i = 0; phys_avail[i + 1]; i += 2) {
300 		phys_avail[i] = round_page(phys_avail[i]);
301 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
302 	}
303 
304 	low_water = phys_avail[0];
305 	high_water = phys_avail[1];
306 
307 	for (i = 0; phys_avail[i + 1]; i += 2) {
308 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
309 
310 		if (size > biggestsize) {
311 			biggestone = i;
312 			biggestsize = size;
313 		}
314 		if (phys_avail[i] < low_water)
315 			low_water = phys_avail[i];
316 		if (phys_avail[i + 1] > high_water)
317 			high_water = phys_avail[i + 1];
318 		++nblocks;
319 	}
320 
321 #ifdef XEN
322 	low_water = 0;
323 #endif
324 
325 	end = phys_avail[biggestone+1];
326 
327 	/*
328 	 * Initialize the locks.
329 	 */
330 	mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF |
331 	    MTX_RECURSE);
332 	mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
333 	    MTX_DEF);
334 
335 	/* Setup page locks. */
336 	for (i = 0; i < PA_LOCK_COUNT; i++)
337 		mtx_init(&pa_lock[i].data, "page lock", NULL,
338 		    MTX_DEF | MTX_RECURSE | MTX_DUPOK);
339 
340 	/*
341 	 * Initialize the queue headers for the hold queue, the active queue,
342 	 * and the inactive queue.
343 	 */
344 	for (i = 0; i < PQ_COUNT; i++)
345 		TAILQ_INIT(&vm_page_queues[i].pl);
346 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
347 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
348 	vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
349 
350 	/*
351 	 * Allocate memory for use when boot strapping the kernel memory
352 	 * allocator.
353 	 */
354 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
355 	new_end = trunc_page(new_end);
356 	mapped = pmap_map(&vaddr, new_end, end,
357 	    VM_PROT_READ | VM_PROT_WRITE);
358 	bzero((void *)mapped, end - new_end);
359 	uma_startup((void *)mapped, boot_pages);
360 
361 #if defined(__amd64__) || defined(__i386__) || defined(__arm__)
362 	/*
363 	 * Allocate a bitmap to indicate that a random physical page
364 	 * needs to be included in a minidump.
365 	 *
366 	 * The amd64 port needs this to indicate which direct map pages
367 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
368 	 *
369 	 * However, i386 still needs this workspace internally within the
370 	 * minidump code.  In theory, they are not needed on i386, but are
371 	 * included should the sf_buf code decide to use them.
372 	 */
373 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
374 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
375 	new_end -= vm_page_dump_size;
376 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
377 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
378 	bzero((void *)vm_page_dump, vm_page_dump_size);
379 #endif
380 #ifdef __amd64__
381 	/*
382 	 * Request that the physical pages underlying the message buffer be
383 	 * included in a crash dump.  Since the message buffer is accessed
384 	 * through the direct map, they are not automatically included.
385 	 */
386 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
387 	last_pa = pa + round_page(MSGBUF_SIZE);
388 	while (pa < last_pa) {
389 		dump_add_page(pa);
390 		pa += PAGE_SIZE;
391 	}
392 #endif
393 	/*
394 	 * Compute the number of pages of memory that will be available for
395 	 * use (taking into account the overhead of a page structure per
396 	 * page).
397 	 */
398 	first_page = low_water / PAGE_SIZE;
399 #ifdef VM_PHYSSEG_SPARSE
400 	page_range = 0;
401 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
402 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
403 #elif defined(VM_PHYSSEG_DENSE)
404 	page_range = high_water / PAGE_SIZE - first_page;
405 #else
406 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
407 #endif
408 	end = new_end;
409 
410 	/*
411 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
412 	 */
413 	vaddr += PAGE_SIZE;
414 
415 	/*
416 	 * Initialize the mem entry structures now, and put them in the free
417 	 * queue.
418 	 */
419 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
420 	mapped = pmap_map(&vaddr, new_end, end,
421 	    VM_PROT_READ | VM_PROT_WRITE);
422 	vm_page_array = (vm_page_t) mapped;
423 #if VM_NRESERVLEVEL > 0
424 	/*
425 	 * Allocate memory for the reservation management system's data
426 	 * structures.
427 	 */
428 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
429 #endif
430 #ifdef __amd64__
431 	/*
432 	 * pmap_map on amd64 comes out of the direct-map, not kvm like i386,
433 	 * so the pages must be tracked for a crashdump to include this data.
434 	 * This includes the vm_page_array and the early UMA bootstrap pages.
435 	 */
436 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
437 		dump_add_page(pa);
438 #endif
439 	phys_avail[biggestone + 1] = new_end;
440 
441 	/*
442 	 * Clear all of the page structures
443 	 */
444 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
445 	for (i = 0; i < page_range; i++)
446 		vm_page_array[i].order = VM_NFREEORDER;
447 	vm_page_array_size = page_range;
448 
449 	/*
450 	 * Initialize the physical memory allocator.
451 	 */
452 	vm_phys_init();
453 
454 	/*
455 	 * Add every available physical page that is not blacklisted to
456 	 * the free lists.
457 	 */
458 	cnt.v_page_count = 0;
459 	cnt.v_free_count = 0;
460 	list = getenv("vm.blacklist");
461 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
462 		pa = phys_avail[i];
463 		last_pa = phys_avail[i + 1];
464 		while (pa < last_pa) {
465 			if (list != NULL &&
466 			    vm_page_blacklist_lookup(list, pa))
467 				printf("Skipping page with pa 0x%jx\n",
468 				    (uintmax_t)pa);
469 			else
470 				vm_phys_add_page(pa);
471 			pa += PAGE_SIZE;
472 		}
473 	}
474 	freeenv(list);
475 #if VM_NRESERVLEVEL > 0
476 	/*
477 	 * Initialize the reservation management system.
478 	 */
479 	vm_reserv_init();
480 #endif
481 	return (vaddr);
482 }
483 
484 void
485 vm_page_flag_set(vm_page_t m, unsigned short bits)
486 {
487 
488 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
489 	m->flags |= bits;
490 }
491 
492 void
493 vm_page_flag_clear(vm_page_t m, unsigned short bits)
494 {
495 
496 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
497 	m->flags &= ~bits;
498 }
499 
500 void
501 vm_page_busy(vm_page_t m)
502 {
503 
504 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
505 	KASSERT((m->oflags & VPO_BUSY) == 0,
506 	    ("vm_page_busy: page already busy!!!"));
507 	m->oflags |= VPO_BUSY;
508 }
509 
510 /*
511  *      vm_page_flash:
512  *
513  *      wakeup anyone waiting for the page.
514  */
515 void
516 vm_page_flash(vm_page_t m)
517 {
518 
519 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
520 	if (m->oflags & VPO_WANTED) {
521 		m->oflags &= ~VPO_WANTED;
522 		wakeup(m);
523 	}
524 }
525 
526 /*
527  *      vm_page_wakeup:
528  *
529  *      clear the VPO_BUSY flag and wakeup anyone waiting for the
530  *      page.
531  *
532  */
533 void
534 vm_page_wakeup(vm_page_t m)
535 {
536 
537 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
538 	KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
539 	m->oflags &= ~VPO_BUSY;
540 	vm_page_flash(m);
541 }
542 
543 void
544 vm_page_io_start(vm_page_t m)
545 {
546 
547 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
548 	m->busy++;
549 }
550 
551 void
552 vm_page_io_finish(vm_page_t m)
553 {
554 
555 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
556 	m->busy--;
557 	if (m->busy == 0)
558 		vm_page_flash(m);
559 }
560 
561 /*
562  * Keep page from being freed by the page daemon
563  * much of the same effect as wiring, except much lower
564  * overhead and should be used only for *very* temporary
565  * holding ("wiring").
566  */
567 void
568 vm_page_hold(vm_page_t mem)
569 {
570 
571 	vm_page_lock_assert(mem, MA_OWNED);
572         mem->hold_count++;
573 }
574 
575 void
576 vm_page_unhold(vm_page_t mem)
577 {
578 
579 	vm_page_lock_assert(mem, MA_OWNED);
580 	--mem->hold_count;
581 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
582 	if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD))
583 		vm_page_free_toq(mem);
584 }
585 
586 /*
587  *	vm_page_free:
588  *
589  *	Free a page.
590  */
591 void
592 vm_page_free(vm_page_t m)
593 {
594 
595 	m->flags &= ~PG_ZERO;
596 	vm_page_free_toq(m);
597 }
598 
599 /*
600  *	vm_page_free_zero:
601  *
602  *	Free a page to the zerod-pages queue
603  */
604 void
605 vm_page_free_zero(vm_page_t m)
606 {
607 
608 	m->flags |= PG_ZERO;
609 	vm_page_free_toq(m);
610 }
611 
612 /*
613  *	vm_page_sleep:
614  *
615  *	Sleep and release the page and page queues locks.
616  *
617  *	The object containing the given page must be locked.
618  */
619 void
620 vm_page_sleep(vm_page_t m, const char *msg)
621 {
622 
623 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
624 	if (mtx_owned(&vm_page_queue_mtx))
625 		vm_page_unlock_queues();
626 	if (mtx_owned(vm_page_lockptr(m)))
627 		vm_page_unlock(m);
628 
629 	/*
630 	 * It's possible that while we sleep, the page will get
631 	 * unbusied and freed.  If we are holding the object
632 	 * lock, we will assume we hold a reference to the object
633 	 * such that even if m->object changes, we can re-lock
634 	 * it.
635 	 */
636 	m->oflags |= VPO_WANTED;
637 	msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
638 }
639 
640 /*
641  *	vm_page_dirty:
642  *
643  *	make page all dirty
644  */
645 void
646 vm_page_dirty(vm_page_t m)
647 {
648 
649 	KASSERT((m->flags & PG_CACHED) == 0,
650 	    ("vm_page_dirty: page in cache!"));
651 	KASSERT(!VM_PAGE_IS_FREE(m),
652 	    ("vm_page_dirty: page is free!"));
653 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
654 	    ("vm_page_dirty: page is invalid!"));
655 	m->dirty = VM_PAGE_BITS_ALL;
656 }
657 
658 /*
659  *	vm_page_splay:
660  *
661  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
662  *	the vm_page containing the given pindex.  If, however, that
663  *	pindex is not found in the vm_object, returns a vm_page that is
664  *	adjacent to the pindex, coming before or after it.
665  */
666 vm_page_t
667 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
668 {
669 	struct vm_page dummy;
670 	vm_page_t lefttreemax, righttreemin, y;
671 
672 	if (root == NULL)
673 		return (root);
674 	lefttreemax = righttreemin = &dummy;
675 	for (;; root = y) {
676 		if (pindex < root->pindex) {
677 			if ((y = root->left) == NULL)
678 				break;
679 			if (pindex < y->pindex) {
680 				/* Rotate right. */
681 				root->left = y->right;
682 				y->right = root;
683 				root = y;
684 				if ((y = root->left) == NULL)
685 					break;
686 			}
687 			/* Link into the new root's right tree. */
688 			righttreemin->left = root;
689 			righttreemin = root;
690 		} else if (pindex > root->pindex) {
691 			if ((y = root->right) == NULL)
692 				break;
693 			if (pindex > y->pindex) {
694 				/* Rotate left. */
695 				root->right = y->left;
696 				y->left = root;
697 				root = y;
698 				if ((y = root->right) == NULL)
699 					break;
700 			}
701 			/* Link into the new root's left tree. */
702 			lefttreemax->right = root;
703 			lefttreemax = root;
704 		} else
705 			break;
706 	}
707 	/* Assemble the new root. */
708 	lefttreemax->right = root->left;
709 	righttreemin->left = root->right;
710 	root->left = dummy.right;
711 	root->right = dummy.left;
712 	return (root);
713 }
714 
715 /*
716  *	vm_page_insert:		[ internal use only ]
717  *
718  *	Inserts the given mem entry into the object and object list.
719  *
720  *	The pagetables are not updated but will presumably fault the page
721  *	in if necessary, or if a kernel page the caller will at some point
722  *	enter the page into the kernel's pmap.  We are not allowed to block
723  *	here so we *can't* do this anyway.
724  *
725  *	The object and page must be locked.
726  *	This routine may not block.
727  */
728 void
729 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
730 {
731 	vm_page_t root;
732 
733 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
734 	if (m->object != NULL)
735 		panic("vm_page_insert: page already inserted");
736 
737 	/*
738 	 * Record the object/offset pair in this page
739 	 */
740 	m->object = object;
741 	m->pindex = pindex;
742 
743 	/*
744 	 * Now link into the object's ordered list of backed pages.
745 	 */
746 	root = object->root;
747 	if (root == NULL) {
748 		m->left = NULL;
749 		m->right = NULL;
750 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
751 	} else {
752 		root = vm_page_splay(pindex, root);
753 		if (pindex < root->pindex) {
754 			m->left = root->left;
755 			m->right = root;
756 			root->left = NULL;
757 			TAILQ_INSERT_BEFORE(root, m, listq);
758 		} else if (pindex == root->pindex)
759 			panic("vm_page_insert: offset already allocated");
760 		else {
761 			m->right = root->right;
762 			m->left = root;
763 			root->right = NULL;
764 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
765 		}
766 	}
767 	object->root = m;
768 	object->generation++;
769 
770 	/*
771 	 * show that the object has one more resident page.
772 	 */
773 	object->resident_page_count++;
774 	/*
775 	 * Hold the vnode until the last page is released.
776 	 */
777 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
778 		vhold((struct vnode *)object->handle);
779 
780 	/*
781 	 * Since we are inserting a new and possibly dirty page,
782 	 * update the object's OBJ_MIGHTBEDIRTY flag.
783 	 */
784 	if (m->flags & PG_WRITEABLE)
785 		vm_object_set_writeable_dirty(object);
786 }
787 
788 /*
789  *	vm_page_remove:
790  *				NOTE: used by device pager as well -wfj
791  *
792  *	Removes the given mem entry from the object/offset-page
793  *	table and the object page list, but do not invalidate/terminate
794  *	the backing store.
795  *
796  *	The object and page must be locked.
797  *	The underlying pmap entry (if any) is NOT removed here.
798  *	This routine may not block.
799  */
800 void
801 vm_page_remove(vm_page_t m)
802 {
803 	vm_object_t object;
804 	vm_page_t root;
805 
806 	if ((m->flags & PG_UNMANAGED) == 0)
807 		vm_page_lock_assert(m, MA_OWNED);
808 	if ((object = m->object) == NULL)
809 		return;
810 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
811 	if (m->oflags & VPO_BUSY) {
812 		m->oflags &= ~VPO_BUSY;
813 		vm_page_flash(m);
814 	}
815 
816 	/*
817 	 * Now remove from the object's list of backed pages.
818 	 */
819 	if (m != object->root)
820 		vm_page_splay(m->pindex, object->root);
821 	if (m->left == NULL)
822 		root = m->right;
823 	else {
824 		root = vm_page_splay(m->pindex, m->left);
825 		root->right = m->right;
826 	}
827 	object->root = root;
828 	TAILQ_REMOVE(&object->memq, m, listq);
829 
830 	/*
831 	 * And show that the object has one fewer resident page.
832 	 */
833 	object->resident_page_count--;
834 	object->generation++;
835 	/*
836 	 * The vnode may now be recycled.
837 	 */
838 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
839 		vdrop((struct vnode *)object->handle);
840 
841 	m->object = NULL;
842 }
843 
844 /*
845  *	vm_page_lookup:
846  *
847  *	Returns the page associated with the object/offset
848  *	pair specified; if none is found, NULL is returned.
849  *
850  *	The object must be locked.
851  *	This routine may not block.
852  *	This is a critical path routine
853  */
854 vm_page_t
855 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
856 {
857 	vm_page_t m;
858 
859 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
860 	if ((m = object->root) != NULL && m->pindex != pindex) {
861 		m = vm_page_splay(pindex, m);
862 		if ((object->root = m)->pindex != pindex)
863 			m = NULL;
864 	}
865 	return (m);
866 }
867 
868 /*
869  *	vm_page_rename:
870  *
871  *	Move the given memory entry from its
872  *	current object to the specified target object/offset.
873  *
874  *	The object must be locked.
875  *	This routine may not block.
876  *
877  *	Note: swap associated with the page must be invalidated by the move.  We
878  *	      have to do this for several reasons:  (1) we aren't freeing the
879  *	      page, (2) we are dirtying the page, (3) the VM system is probably
880  *	      moving the page from object A to B, and will then later move
881  *	      the backing store from A to B and we can't have a conflict.
882  *
883  *	Note: we *always* dirty the page.  It is necessary both for the
884  *	      fact that we moved it, and because we may be invalidating
885  *	      swap.  If the page is on the cache, we have to deactivate it
886  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
887  *	      on the cache.
888  */
889 void
890 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
891 {
892 
893 	vm_page_remove(m);
894 	vm_page_insert(m, new_object, new_pindex);
895 	vm_page_dirty(m);
896 }
897 
898 /*
899  *	Convert all of the given object's cached pages that have a
900  *	pindex within the given range into free pages.  If the value
901  *	zero is given for "end", then the range's upper bound is
902  *	infinity.  If the given object is backed by a vnode and it
903  *	transitions from having one or more cached pages to none, the
904  *	vnode's hold count is reduced.
905  */
906 void
907 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
908 {
909 	vm_page_t m, m_next;
910 	boolean_t empty;
911 
912 	mtx_lock(&vm_page_queue_free_mtx);
913 	if (__predict_false(object->cache == NULL)) {
914 		mtx_unlock(&vm_page_queue_free_mtx);
915 		return;
916 	}
917 	m = object->cache = vm_page_splay(start, object->cache);
918 	if (m->pindex < start) {
919 		if (m->right == NULL)
920 			m = NULL;
921 		else {
922 			m_next = vm_page_splay(start, m->right);
923 			m_next->left = m;
924 			m->right = NULL;
925 			m = object->cache = m_next;
926 		}
927 	}
928 
929 	/*
930 	 * At this point, "m" is either (1) a reference to the page
931 	 * with the least pindex that is greater than or equal to
932 	 * "start" or (2) NULL.
933 	 */
934 	for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
935 		/*
936 		 * Find "m"'s successor and remove "m" from the
937 		 * object's cache.
938 		 */
939 		if (m->right == NULL) {
940 			object->cache = m->left;
941 			m_next = NULL;
942 		} else {
943 			m_next = vm_page_splay(start, m->right);
944 			m_next->left = m->left;
945 			object->cache = m_next;
946 		}
947 		/* Convert "m" to a free page. */
948 		m->object = NULL;
949 		m->valid = 0;
950 		/* Clear PG_CACHED and set PG_FREE. */
951 		m->flags ^= PG_CACHED | PG_FREE;
952 		KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
953 		    ("vm_page_cache_free: page %p has inconsistent flags", m));
954 		cnt.v_cache_count--;
955 		cnt.v_free_count++;
956 	}
957 	empty = object->cache == NULL;
958 	mtx_unlock(&vm_page_queue_free_mtx);
959 	if (object->type == OBJT_VNODE && empty)
960 		vdrop(object->handle);
961 }
962 
963 /*
964  *	Returns the cached page that is associated with the given
965  *	object and offset.  If, however, none exists, returns NULL.
966  *
967  *	The free page queue must be locked.
968  */
969 static inline vm_page_t
970 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
971 {
972 	vm_page_t m;
973 
974 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
975 	if ((m = object->cache) != NULL && m->pindex != pindex) {
976 		m = vm_page_splay(pindex, m);
977 		if ((object->cache = m)->pindex != pindex)
978 			m = NULL;
979 	}
980 	return (m);
981 }
982 
983 /*
984  *	Remove the given cached page from its containing object's
985  *	collection of cached pages.
986  *
987  *	The free page queue must be locked.
988  */
989 void
990 vm_page_cache_remove(vm_page_t m)
991 {
992 	vm_object_t object;
993 	vm_page_t root;
994 
995 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
996 	KASSERT((m->flags & PG_CACHED) != 0,
997 	    ("vm_page_cache_remove: page %p is not cached", m));
998 	object = m->object;
999 	if (m != object->cache) {
1000 		root = vm_page_splay(m->pindex, object->cache);
1001 		KASSERT(root == m,
1002 		    ("vm_page_cache_remove: page %p is not cached in object %p",
1003 		    m, object));
1004 	}
1005 	if (m->left == NULL)
1006 		root = m->right;
1007 	else if (m->right == NULL)
1008 		root = m->left;
1009 	else {
1010 		root = vm_page_splay(m->pindex, m->left);
1011 		root->right = m->right;
1012 	}
1013 	object->cache = root;
1014 	m->object = NULL;
1015 	cnt.v_cache_count--;
1016 }
1017 
1018 /*
1019  *	Transfer all of the cached pages with offset greater than or
1020  *	equal to 'offidxstart' from the original object's cache to the
1021  *	new object's cache.  However, any cached pages with offset
1022  *	greater than or equal to the new object's size are kept in the
1023  *	original object.  Initially, the new object's cache must be
1024  *	empty.  Offset 'offidxstart' in the original object must
1025  *	correspond to offset zero in the new object.
1026  *
1027  *	The new object must be locked.
1028  */
1029 void
1030 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1031     vm_object_t new_object)
1032 {
1033 	vm_page_t m, m_next;
1034 
1035 	/*
1036 	 * Insertion into an object's collection of cached pages
1037 	 * requires the object to be locked.  In contrast, removal does
1038 	 * not.
1039 	 */
1040 	VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1041 	KASSERT(new_object->cache == NULL,
1042 	    ("vm_page_cache_transfer: object %p has cached pages",
1043 	    new_object));
1044 	mtx_lock(&vm_page_queue_free_mtx);
1045 	if ((m = orig_object->cache) != NULL) {
1046 		/*
1047 		 * Transfer all of the pages with offset greater than or
1048 		 * equal to 'offidxstart' from the original object's
1049 		 * cache to the new object's cache.
1050 		 */
1051 		m = vm_page_splay(offidxstart, m);
1052 		if (m->pindex < offidxstart) {
1053 			orig_object->cache = m;
1054 			new_object->cache = m->right;
1055 			m->right = NULL;
1056 		} else {
1057 			orig_object->cache = m->left;
1058 			new_object->cache = m;
1059 			m->left = NULL;
1060 		}
1061 		while ((m = new_object->cache) != NULL) {
1062 			if ((m->pindex - offidxstart) >= new_object->size) {
1063 				/*
1064 				 * Return all of the cached pages with
1065 				 * offset greater than or equal to the
1066 				 * new object's size to the original
1067 				 * object's cache.
1068 				 */
1069 				new_object->cache = m->left;
1070 				m->left = orig_object->cache;
1071 				orig_object->cache = m;
1072 				break;
1073 			}
1074 			m_next = vm_page_splay(m->pindex, m->right);
1075 			/* Update the page's object and offset. */
1076 			m->object = new_object;
1077 			m->pindex -= offidxstart;
1078 			if (m_next == NULL)
1079 				break;
1080 			m->right = NULL;
1081 			m_next->left = m;
1082 			new_object->cache = m_next;
1083 		}
1084 		KASSERT(new_object->cache == NULL ||
1085 		    new_object->type == OBJT_SWAP,
1086 		    ("vm_page_cache_transfer: object %p's type is incompatible"
1087 		    " with cached pages", new_object));
1088 	}
1089 	mtx_unlock(&vm_page_queue_free_mtx);
1090 }
1091 
1092 /*
1093  *	vm_page_alloc:
1094  *
1095  *	Allocate and return a memory cell associated
1096  *	with this VM object/offset pair.
1097  *
1098  *	page_req classes:
1099  *	VM_ALLOC_NORMAL		normal process request
1100  *	VM_ALLOC_SYSTEM		system *really* needs a page
1101  *	VM_ALLOC_INTERRUPT	interrupt time request
1102  *	VM_ALLOC_ZERO		zero page
1103  *	VM_ALLOC_WIRED		wire the allocated page
1104  *	VM_ALLOC_NOOBJ		page is not associated with a vm object
1105  *	VM_ALLOC_NOBUSY		do not set the page busy
1106  *	VM_ALLOC_IFNOTCACHED	return NULL, do not reactivate if the page
1107  *				is cached
1108  *
1109  *	This routine may not sleep.
1110  */
1111 vm_page_t
1112 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1113 {
1114 	struct vnode *vp = NULL;
1115 	vm_object_t m_object;
1116 	vm_page_t m;
1117 	int flags, page_req;
1118 
1119 	page_req = req & VM_ALLOC_CLASS_MASK;
1120 	KASSERT(curthread->td_intr_nesting_level == 0 ||
1121 	    page_req == VM_ALLOC_INTERRUPT,
1122 	    ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context"));
1123 
1124 	if ((req & VM_ALLOC_NOOBJ) == 0) {
1125 		KASSERT(object != NULL,
1126 		    ("vm_page_alloc: NULL object."));
1127 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1128 	}
1129 
1130 	/*
1131 	 * The pager is allowed to eat deeper into the free page list.
1132 	 */
1133 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
1134 		page_req = VM_ALLOC_SYSTEM;
1135 	};
1136 
1137 	mtx_lock(&vm_page_queue_free_mtx);
1138 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1139 	    (page_req == VM_ALLOC_SYSTEM &&
1140 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1141 	    (page_req == VM_ALLOC_INTERRUPT &&
1142 	    cnt.v_free_count + cnt.v_cache_count > 0)) {
1143 		/*
1144 		 * Allocate from the free queue if the number of free pages
1145 		 * exceeds the minimum for the request class.
1146 		 */
1147 		if (object != NULL &&
1148 		    (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1149 			if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1150 				mtx_unlock(&vm_page_queue_free_mtx);
1151 				return (NULL);
1152 			}
1153 			if (vm_phys_unfree_page(m))
1154 				vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1155 #if VM_NRESERVLEVEL > 0
1156 			else if (!vm_reserv_reactivate_page(m))
1157 #else
1158 			else
1159 #endif
1160 				panic("vm_page_alloc: cache page %p is missing"
1161 				    " from the free queue", m);
1162 		} else if ((req & VM_ALLOC_IFCACHED) != 0) {
1163 			mtx_unlock(&vm_page_queue_free_mtx);
1164 			return (NULL);
1165 #if VM_NRESERVLEVEL > 0
1166 		} else if (object == NULL || object->type == OBJT_DEVICE ||
1167 		    object->type == OBJT_SG ||
1168 		    (object->flags & OBJ_COLORED) == 0 ||
1169 		    (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1170 #else
1171 		} else {
1172 #endif
1173 			m = vm_phys_alloc_pages(object != NULL ?
1174 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1175 #if VM_NRESERVLEVEL > 0
1176 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1177 				m = vm_phys_alloc_pages(object != NULL ?
1178 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1179 				    0);
1180 			}
1181 #endif
1182 		}
1183 	} else {
1184 		/*
1185 		 * Not allocatable, give up.
1186 		 */
1187 		mtx_unlock(&vm_page_queue_free_mtx);
1188 		atomic_add_int(&vm_pageout_deficit, 1);
1189 		pagedaemon_wakeup();
1190 		return (NULL);
1191 	}
1192 
1193 	/*
1194 	 *  At this point we had better have found a good page.
1195 	 */
1196 
1197 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1198 	KASSERT(m->queue == PQ_NONE,
1199 	    ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1200 	KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1201 	KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1202 	KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1203 	KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1204 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1205 	    ("vm_page_alloc: page %p has unexpected memattr %d", m,
1206 	    pmap_page_get_memattr(m)));
1207 	if ((m->flags & PG_CACHED) != 0) {
1208 		KASSERT(m->valid != 0,
1209 		    ("vm_page_alloc: cached page %p is invalid", m));
1210 		if (m->object == object && m->pindex == pindex)
1211 	  		cnt.v_reactivated++;
1212 		else
1213 			m->valid = 0;
1214 		m_object = m->object;
1215 		vm_page_cache_remove(m);
1216 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1217 			vp = m_object->handle;
1218 	} else {
1219 		KASSERT(VM_PAGE_IS_FREE(m),
1220 		    ("vm_page_alloc: page %p is not free", m));
1221 		KASSERT(m->valid == 0,
1222 		    ("vm_page_alloc: free page %p is valid", m));
1223 		cnt.v_free_count--;
1224 	}
1225 
1226 	/*
1227 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
1228 	 */
1229 	flags = 0;
1230 	if (m->flags & PG_ZERO) {
1231 		vm_page_zero_count--;
1232 		if (req & VM_ALLOC_ZERO)
1233 			flags = PG_ZERO;
1234 	}
1235 	if (object == NULL || object->type == OBJT_PHYS)
1236 		flags |= PG_UNMANAGED;
1237 	m->flags = flags;
1238 	if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ))
1239 		m->oflags = 0;
1240 	else
1241 		m->oflags = VPO_BUSY;
1242 	if (req & VM_ALLOC_WIRED) {
1243 		atomic_add_int(&cnt.v_wire_count, 1);
1244 		m->wire_count = 1;
1245 	}
1246 	m->act_count = 0;
1247 	mtx_unlock(&vm_page_queue_free_mtx);
1248 
1249 	if (object != NULL) {
1250 		/* Ignore device objects; the pager sets "memattr" for them. */
1251 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1252 		    object->type != OBJT_DEVICE && object->type != OBJT_SG)
1253 			pmap_page_set_memattr(m, object->memattr);
1254 		vm_page_insert(m, object, pindex);
1255 	} else
1256 		m->pindex = pindex;
1257 
1258 	/*
1259 	 * The following call to vdrop() must come after the above call
1260 	 * to vm_page_insert() in case both affect the same object and
1261 	 * vnode.  Otherwise, the affected vnode's hold count could
1262 	 * temporarily become zero.
1263 	 */
1264 	if (vp != NULL)
1265 		vdrop(vp);
1266 
1267 	/*
1268 	 * Don't wakeup too often - wakeup the pageout daemon when
1269 	 * we would be nearly out of memory.
1270 	 */
1271 	if (vm_paging_needed())
1272 		pagedaemon_wakeup();
1273 
1274 	return (m);
1275 }
1276 
1277 /*
1278  *	vm_wait:	(also see VM_WAIT macro)
1279  *
1280  *	Block until free pages are available for allocation
1281  *	- Called in various places before memory allocations.
1282  */
1283 void
1284 vm_wait(void)
1285 {
1286 
1287 	mtx_lock(&vm_page_queue_free_mtx);
1288 	if (curproc == pageproc) {
1289 		vm_pageout_pages_needed = 1;
1290 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1291 		    PDROP | PSWP, "VMWait", 0);
1292 	} else {
1293 		if (!vm_pages_needed) {
1294 			vm_pages_needed = 1;
1295 			wakeup(&vm_pages_needed);
1296 		}
1297 		msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1298 		    "vmwait", 0);
1299 	}
1300 }
1301 
1302 /*
1303  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
1304  *
1305  *	Block until free pages are available for allocation
1306  *	- Called only in vm_fault so that processes page faulting
1307  *	  can be easily tracked.
1308  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1309  *	  processes will be able to grab memory first.  Do not change
1310  *	  this balance without careful testing first.
1311  */
1312 void
1313 vm_waitpfault(void)
1314 {
1315 
1316 	mtx_lock(&vm_page_queue_free_mtx);
1317 	if (!vm_pages_needed) {
1318 		vm_pages_needed = 1;
1319 		wakeup(&vm_pages_needed);
1320 	}
1321 	msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1322 	    "pfault", 0);
1323 }
1324 
1325 /*
1326  *	vm_page_requeue:
1327  *
1328  *	If the given page is contained within a page queue, move it to the tail
1329  *	of that queue.
1330  *
1331  *	The page queues must be locked.
1332  */
1333 void
1334 vm_page_requeue(vm_page_t m)
1335 {
1336 	int queue = VM_PAGE_GETQUEUE(m);
1337 	struct vpgqueues *vpq;
1338 
1339 	if (queue != PQ_NONE) {
1340 		vpq = &vm_page_queues[queue];
1341 		TAILQ_REMOVE(&vpq->pl, m, pageq);
1342 		TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1343 	}
1344 }
1345 
1346 /*
1347  *	vm_page_queue_remove:
1348  *
1349  *	Remove the given page from the specified queue.
1350  *
1351  *	The page and page queues must be locked.
1352  */
1353 static __inline void
1354 vm_page_queue_remove(int queue, vm_page_t m)
1355 {
1356 	struct vpgqueues *pq;
1357 
1358 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1359 	vm_page_lock_assert(m, MA_OWNED);
1360 	pq = &vm_page_queues[queue];
1361 	TAILQ_REMOVE(&pq->pl, m, pageq);
1362 	(*pq->cnt)--;
1363 }
1364 
1365 /*
1366  *	vm_pageq_remove:
1367  *
1368  *	Remove a page from its queue.
1369  *
1370  *	The given page must be locked.
1371  *	This routine may not block.
1372  */
1373 void
1374 vm_pageq_remove(vm_page_t m)
1375 {
1376 	int queue = VM_PAGE_GETQUEUE(m);
1377 
1378 	vm_page_lock_assert(m, MA_OWNED);
1379 	if (queue != PQ_NONE) {
1380 		vm_page_lock_queues();
1381 		VM_PAGE_SETQUEUE2(m, PQ_NONE);
1382 		vm_page_queue_remove(queue, m);
1383 		vm_page_unlock_queues();
1384 	}
1385 }
1386 
1387 /*
1388  *	vm_page_enqueue:
1389  *
1390  *	Add the given page to the specified queue.
1391  *
1392  *	The page queues must be locked.
1393  */
1394 static void
1395 vm_page_enqueue(int queue, vm_page_t m)
1396 {
1397 	struct vpgqueues *vpq;
1398 
1399 	vpq = &vm_page_queues[queue];
1400 	VM_PAGE_SETQUEUE2(m, queue);
1401 	TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1402 	++*vpq->cnt;
1403 }
1404 
1405 /*
1406  *	vm_page_activate:
1407  *
1408  *	Put the specified page on the active list (if appropriate).
1409  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1410  *	mess with it.
1411  *
1412  *	The page must be locked.
1413  *	This routine may not block.
1414  */
1415 void
1416 vm_page_activate(vm_page_t m)
1417 {
1418 	int queue;
1419 
1420 	vm_page_lock_assert(m, MA_OWNED);
1421 	if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) != PQ_ACTIVE) {
1422 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1423 			if (m->act_count < ACT_INIT)
1424 				m->act_count = ACT_INIT;
1425 			vm_page_lock_queues();
1426 			if (queue != PQ_NONE)
1427 				vm_page_queue_remove(queue, m);
1428 			vm_page_enqueue(PQ_ACTIVE, m);
1429 			vm_page_unlock_queues();
1430 		} else
1431 			KASSERT(queue == PQ_NONE,
1432 			    ("vm_page_activate: wired page %p is queued", m));
1433 	} else {
1434 		if (m->act_count < ACT_INIT)
1435 			m->act_count = ACT_INIT;
1436 	}
1437 }
1438 
1439 /*
1440  *	vm_page_free_wakeup:
1441  *
1442  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1443  *	routine is called when a page has been added to the cache or free
1444  *	queues.
1445  *
1446  *	The page queues must be locked.
1447  *	This routine may not block.
1448  */
1449 static inline void
1450 vm_page_free_wakeup(void)
1451 {
1452 
1453 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1454 	/*
1455 	 * if pageout daemon needs pages, then tell it that there are
1456 	 * some free.
1457 	 */
1458 	if (vm_pageout_pages_needed &&
1459 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1460 		wakeup(&vm_pageout_pages_needed);
1461 		vm_pageout_pages_needed = 0;
1462 	}
1463 	/*
1464 	 * wakeup processes that are waiting on memory if we hit a
1465 	 * high water mark. And wakeup scheduler process if we have
1466 	 * lots of memory. this process will swapin processes.
1467 	 */
1468 	if (vm_pages_needed && !vm_page_count_min()) {
1469 		vm_pages_needed = 0;
1470 		wakeup(&cnt.v_free_count);
1471 	}
1472 }
1473 
1474 /*
1475  *	vm_page_free_toq:
1476  *
1477  *	Returns the given page to the free list,
1478  *	disassociating it with any VM object.
1479  *
1480  *	Object and page must be locked prior to entry.
1481  *	This routine may not block.
1482  */
1483 
1484 void
1485 vm_page_free_toq(vm_page_t m)
1486 {
1487 
1488 	if ((m->flags & PG_UNMANAGED) == 0) {
1489 		vm_page_lock_assert(m, MA_OWNED);
1490 		KASSERT(!pmap_page_is_mapped(m),
1491 		    ("vm_page_free_toq: freeing mapped page %p", m));
1492 	}
1493 	PCPU_INC(cnt.v_tfree);
1494 
1495 	if (m->busy || VM_PAGE_IS_FREE(m)) {
1496 		printf(
1497 		"vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n",
1498 		    (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0,
1499 		    m->hold_count);
1500 		if (VM_PAGE_IS_FREE(m))
1501 			panic("vm_page_free: freeing free page");
1502 		else
1503 			panic("vm_page_free: freeing busy page");
1504 	}
1505 
1506 	/*
1507 	 * unqueue, then remove page.  Note that we cannot destroy
1508 	 * the page here because we do not want to call the pager's
1509 	 * callback routine until after we've put the page on the
1510 	 * appropriate free queue.
1511 	 */
1512 	if ((m->flags & PG_UNMANAGED) == 0)
1513 		vm_pageq_remove(m);
1514 	vm_page_remove(m);
1515 
1516 	/*
1517 	 * If fictitious remove object association and
1518 	 * return, otherwise delay object association removal.
1519 	 */
1520 	if ((m->flags & PG_FICTITIOUS) != 0) {
1521 		return;
1522 	}
1523 
1524 	m->valid = 0;
1525 	vm_page_undirty(m);
1526 
1527 	if (m->wire_count != 0) {
1528 		if (m->wire_count > 1) {
1529 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1530 				m->wire_count, (long)m->pindex);
1531 		}
1532 		panic("vm_page_free: freeing wired page");
1533 	}
1534 	if (m->hold_count != 0) {
1535 		m->flags &= ~PG_ZERO;
1536 		vm_page_lock_queues();
1537 		vm_page_enqueue(PQ_HOLD, m);
1538 		vm_page_unlock_queues();
1539 	} else {
1540 		/*
1541 		 * Restore the default memory attribute to the page.
1542 		 */
1543 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1544 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1545 
1546 		/*
1547 		 * Insert the page into the physical memory allocator's
1548 		 * cache/free page queues.
1549 		 */
1550 		mtx_lock(&vm_page_queue_free_mtx);
1551 		m->flags |= PG_FREE;
1552 		cnt.v_free_count++;
1553 #if VM_NRESERVLEVEL > 0
1554 		if (!vm_reserv_free_page(m))
1555 #else
1556 		if (TRUE)
1557 #endif
1558 			vm_phys_free_pages(m, 0);
1559 		if ((m->flags & PG_ZERO) != 0)
1560 			++vm_page_zero_count;
1561 		else
1562 			vm_page_zero_idle_wakeup();
1563 		vm_page_free_wakeup();
1564 		mtx_unlock(&vm_page_queue_free_mtx);
1565 	}
1566 }
1567 
1568 /*
1569  *	vm_page_wire:
1570  *
1571  *	Mark this page as wired down by yet
1572  *	another map, removing it from paging queues
1573  *	as necessary.
1574  *
1575  *	The page must be locked.
1576  *	This routine may not block.
1577  */
1578 void
1579 vm_page_wire(vm_page_t m)
1580 {
1581 
1582 	/*
1583 	 * Only bump the wire statistics if the page is not already wired,
1584 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1585 	 * it is already off the queues).
1586 	 */
1587 	vm_page_lock_assert(m, MA_OWNED);
1588 	if (m->flags & PG_FICTITIOUS)
1589 		return;
1590 	if (m->wire_count == 0) {
1591 		if ((m->flags & PG_UNMANAGED) == 0)
1592 			vm_pageq_remove(m);
1593 		atomic_add_int(&cnt.v_wire_count, 1);
1594 	}
1595 	m->wire_count++;
1596 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1597 }
1598 
1599 /*
1600  *	vm_page_unwire:
1601  *
1602  *	Release one wiring of this page, potentially
1603  *	enabling it to be paged again.
1604  *
1605  *	Many pages placed on the inactive queue should actually go
1606  *	into the cache, but it is difficult to figure out which.  What
1607  *	we do instead, if the inactive target is well met, is to put
1608  *	clean pages at the head of the inactive queue instead of the tail.
1609  *	This will cause them to be moved to the cache more quickly and
1610  *	if not actively re-referenced, freed more quickly.  If we just
1611  *	stick these pages at the end of the inactive queue, heavy filesystem
1612  *	meta-data accesses can cause an unnecessary paging load on memory bound
1613  *	processes.  This optimization causes one-time-use metadata to be
1614  *	reused more quickly.
1615  *
1616  *	BUT, if we are in a low-memory situation we have no choice but to
1617  *	put clean pages on the cache queue.
1618  *
1619  *	A number of routines use vm_page_unwire() to guarantee that the page
1620  *	will go into either the inactive or active queues, and will NEVER
1621  *	be placed in the cache - for example, just after dirtying a page.
1622  *	dirty pages in the cache are not allowed.
1623  *
1624  *	The page must be locked.
1625  *	This routine may not block.
1626  */
1627 void
1628 vm_page_unwire(vm_page_t m, int activate)
1629 {
1630 
1631 	if ((m->flags & PG_UNMANAGED) == 0)
1632 		vm_page_lock_assert(m, MA_OWNED);
1633 	if (m->flags & PG_FICTITIOUS)
1634 		return;
1635 	if (m->wire_count > 0) {
1636 		m->wire_count--;
1637 		if (m->wire_count == 0) {
1638 			atomic_subtract_int(&cnt.v_wire_count, 1);
1639 			if ((m->flags & PG_UNMANAGED) != 0)
1640 				return;
1641 			vm_page_lock_queues();
1642 			if (activate)
1643 				vm_page_enqueue(PQ_ACTIVE, m);
1644 			else {
1645 				vm_page_flag_clear(m, PG_WINATCFLS);
1646 				vm_page_enqueue(PQ_INACTIVE, m);
1647 			}
1648 			vm_page_unlock_queues();
1649 		}
1650 	} else {
1651 		panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1652 	}
1653 }
1654 
1655 /*
1656  * Move the specified page to the inactive queue.
1657  *
1658  * Normally athead is 0 resulting in LRU operation.  athead is set
1659  * to 1 if we want this page to be 'as if it were placed in the cache',
1660  * except without unmapping it from the process address space.
1661  *
1662  * This routine may not block.
1663  */
1664 static inline void
1665 _vm_page_deactivate(vm_page_t m, int athead)
1666 {
1667 	int queue;
1668 
1669 	vm_page_lock_assert(m, MA_OWNED);
1670 
1671 	/*
1672 	 * Ignore if already inactive.
1673 	 */
1674 	if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) == PQ_INACTIVE)
1675 		return;
1676 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1677 		vm_page_lock_queues();
1678 		vm_page_flag_clear(m, PG_WINATCFLS);
1679 		if (queue != PQ_NONE)
1680 			vm_page_queue_remove(queue, m);
1681 		if (athead)
1682 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m,
1683 			    pageq);
1684 		else
1685 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m,
1686 			    pageq);
1687 		VM_PAGE_SETQUEUE2(m, PQ_INACTIVE);
1688 		cnt.v_inactive_count++;
1689 		vm_page_unlock_queues();
1690 	}
1691 }
1692 
1693 /*
1694  * Move the specified page to the inactive queue.
1695  *
1696  * The page must be locked.
1697  */
1698 void
1699 vm_page_deactivate(vm_page_t m)
1700 {
1701 
1702 	_vm_page_deactivate(m, 0);
1703 }
1704 
1705 /*
1706  * vm_page_try_to_cache:
1707  *
1708  * Returns 0 on failure, 1 on success
1709  */
1710 int
1711 vm_page_try_to_cache(vm_page_t m)
1712 {
1713 
1714 	vm_page_lock_assert(m, MA_OWNED);
1715 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1716 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1717 	    (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED))
1718 		return (0);
1719 	pmap_remove_all(m);
1720 	if (m->dirty)
1721 		return (0);
1722 	vm_page_cache(m);
1723 	return (1);
1724 }
1725 
1726 /*
1727  * vm_page_try_to_free()
1728  *
1729  *	Attempt to free the page.  If we cannot free it, we do nothing.
1730  *	1 is returned on success, 0 on failure.
1731  */
1732 int
1733 vm_page_try_to_free(vm_page_t m)
1734 {
1735 
1736 	vm_page_lock_assert(m, MA_OWNED);
1737 	if (m->object != NULL)
1738 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1739 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1740 	    (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED))
1741 		return (0);
1742 	pmap_remove_all(m);
1743 	if (m->dirty)
1744 		return (0);
1745 	vm_page_free(m);
1746 	return (1);
1747 }
1748 
1749 /*
1750  * vm_page_cache
1751  *
1752  * Put the specified page onto the page cache queue (if appropriate).
1753  *
1754  * This routine may not block.
1755  */
1756 void
1757 vm_page_cache(vm_page_t m)
1758 {
1759 	vm_object_t object;
1760 	vm_page_t root;
1761 
1762 	vm_page_lock_assert(m, MA_OWNED);
1763 	object = m->object;
1764 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1765 	if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy ||
1766 	    m->hold_count || m->wire_count)
1767 		panic("vm_page_cache: attempting to cache busy page");
1768 	pmap_remove_all(m);
1769 	if (m->dirty != 0)
1770 		panic("vm_page_cache: page %p is dirty", m);
1771 	if (m->valid == 0 || object->type == OBJT_DEFAULT ||
1772 	    (object->type == OBJT_SWAP &&
1773 	    !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
1774 		/*
1775 		 * Hypothesis: A cache-elgible page belonging to a
1776 		 * default object or swap object but without a backing
1777 		 * store must be zero filled.
1778 		 */
1779 		vm_page_free(m);
1780 		return;
1781 	}
1782 	KASSERT((m->flags & PG_CACHED) == 0,
1783 	    ("vm_page_cache: page %p is already cached", m));
1784 	PCPU_INC(cnt.v_tcached);
1785 
1786 	/*
1787 	 * Remove the page from the paging queues.
1788 	 */
1789 	vm_pageq_remove(m);
1790 
1791 	/*
1792 	 * Remove the page from the object's collection of resident
1793 	 * pages.
1794 	 */
1795 	if (m != object->root)
1796 		vm_page_splay(m->pindex, object->root);
1797 	if (m->left == NULL)
1798 		root = m->right;
1799 	else {
1800 		root = vm_page_splay(m->pindex, m->left);
1801 		root->right = m->right;
1802 	}
1803 	object->root = root;
1804 	TAILQ_REMOVE(&object->memq, m, listq);
1805 	object->resident_page_count--;
1806 	object->generation++;
1807 
1808 	/*
1809 	 * Restore the default memory attribute to the page.
1810 	 */
1811 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1812 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1813 
1814 	/*
1815 	 * Insert the page into the object's collection of cached pages
1816 	 * and the physical memory allocator's cache/free page queues.
1817 	 */
1818 	m->flags &= ~PG_ZERO;
1819 	mtx_lock(&vm_page_queue_free_mtx);
1820 	m->flags |= PG_CACHED;
1821 	cnt.v_cache_count++;
1822 	root = object->cache;
1823 	if (root == NULL) {
1824 		m->left = NULL;
1825 		m->right = NULL;
1826 	} else {
1827 		root = vm_page_splay(m->pindex, root);
1828 		if (m->pindex < root->pindex) {
1829 			m->left = root->left;
1830 			m->right = root;
1831 			root->left = NULL;
1832 		} else if (__predict_false(m->pindex == root->pindex))
1833 			panic("vm_page_cache: offset already cached");
1834 		else {
1835 			m->right = root->right;
1836 			m->left = root;
1837 			root->right = NULL;
1838 		}
1839 	}
1840 	object->cache = m;
1841 #if VM_NRESERVLEVEL > 0
1842 	if (!vm_reserv_free_page(m)) {
1843 #else
1844 	if (TRUE) {
1845 #endif
1846 		vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
1847 		vm_phys_free_pages(m, 0);
1848 	}
1849 	vm_page_free_wakeup();
1850 	mtx_unlock(&vm_page_queue_free_mtx);
1851 
1852 	/*
1853 	 * Increment the vnode's hold count if this is the object's only
1854 	 * cached page.  Decrement the vnode's hold count if this was
1855 	 * the object's only resident page.
1856 	 */
1857 	if (object->type == OBJT_VNODE) {
1858 		if (root == NULL && object->resident_page_count != 0)
1859 			vhold(object->handle);
1860 		else if (root != NULL && object->resident_page_count == 0)
1861 			vdrop(object->handle);
1862 	}
1863 }
1864 
1865 /*
1866  * vm_page_dontneed
1867  *
1868  *	Cache, deactivate, or do nothing as appropriate.  This routine
1869  *	is typically used by madvise() MADV_DONTNEED.
1870  *
1871  *	Generally speaking we want to move the page into the cache so
1872  *	it gets reused quickly.  However, this can result in a silly syndrome
1873  *	due to the page recycling too quickly.  Small objects will not be
1874  *	fully cached.  On the otherhand, if we move the page to the inactive
1875  *	queue we wind up with a problem whereby very large objects
1876  *	unnecessarily blow away our inactive and cache queues.
1877  *
1878  *	The solution is to move the pages based on a fixed weighting.  We
1879  *	either leave them alone, deactivate them, or move them to the cache,
1880  *	where moving them to the cache has the highest weighting.
1881  *	By forcing some pages into other queues we eventually force the
1882  *	system to balance the queues, potentially recovering other unrelated
1883  *	space from active.  The idea is to not force this to happen too
1884  *	often.
1885  */
1886 void
1887 vm_page_dontneed(vm_page_t m)
1888 {
1889 	int dnw;
1890 	int head;
1891 
1892 	vm_page_lock_assert(m, MA_OWNED);
1893 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1894 	dnw = PCPU_GET(dnweight);
1895 	PCPU_INC(dnweight);
1896 
1897 	/*
1898 	 * Occasionally leave the page alone.
1899 	 */
1900 	if ((dnw & 0x01F0) == 0 ||
1901 	    VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) {
1902 		if (m->act_count >= ACT_INIT)
1903 			--m->act_count;
1904 		return;
1905 	}
1906 
1907 	/*
1908 	 * Clear any references to the page.  Otherwise, the page daemon will
1909 	 * immediately reactivate the page.
1910 	 *
1911 	 * Perform the pmap_clear_reference() first.  Otherwise, a concurrent
1912 	 * pmap operation, such as pmap_remove(), could clear a reference in
1913 	 * the pmap and set PG_REFERENCED on the page before the
1914 	 * pmap_clear_reference() had completed.  Consequently, the page would
1915 	 * appear referenced based upon an old reference that occurred before
1916 	 * this function ran.
1917 	 */
1918 	pmap_clear_reference(m);
1919 	vm_page_lock_queues();
1920 	vm_page_flag_clear(m, PG_REFERENCED);
1921 	vm_page_unlock_queues();
1922 
1923 	if (m->dirty == 0 && pmap_is_modified(m))
1924 		vm_page_dirty(m);
1925 
1926 	if (m->dirty || (dnw & 0x0070) == 0) {
1927 		/*
1928 		 * Deactivate the page 3 times out of 32.
1929 		 */
1930 		head = 0;
1931 	} else {
1932 		/*
1933 		 * Cache the page 28 times out of every 32.  Note that
1934 		 * the page is deactivated instead of cached, but placed
1935 		 * at the head of the queue instead of the tail.
1936 		 */
1937 		head = 1;
1938 	}
1939 	_vm_page_deactivate(m, head);
1940 }
1941 
1942 /*
1943  * Grab a page, waiting until we are waken up due to the page
1944  * changing state.  We keep on waiting, if the page continues
1945  * to be in the object.  If the page doesn't exist, first allocate it
1946  * and then conditionally zero it.
1947  *
1948  * This routine may block.
1949  */
1950 vm_page_t
1951 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1952 {
1953 	vm_page_t m;
1954 
1955 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1956 retrylookup:
1957 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1958 		if ((m->oflags & VPO_BUSY) != 0 || m->busy != 0) {
1959 			if ((allocflags & VM_ALLOC_RETRY) != 0) {
1960 				/*
1961 				 * Reference the page before unlocking and
1962 				 * sleeping so that the page daemon is less
1963 				 * likely to reclaim it.
1964 				 */
1965 				vm_page_lock_queues();
1966 				vm_page_flag_set(m, PG_REFERENCED);
1967 			}
1968 			vm_page_sleep(m, "pgrbwt");
1969 			if ((allocflags & VM_ALLOC_RETRY) == 0)
1970 				return (NULL);
1971 			goto retrylookup;
1972 		} else {
1973 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
1974 				vm_page_lock(m);
1975 				vm_page_wire(m);
1976 				vm_page_unlock(m);
1977 			}
1978 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
1979 				vm_page_busy(m);
1980 			return (m);
1981 		}
1982 	}
1983 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1984 	if (m == NULL) {
1985 		VM_OBJECT_UNLOCK(object);
1986 		VM_WAIT;
1987 		VM_OBJECT_LOCK(object);
1988 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1989 			return (NULL);
1990 		goto retrylookup;
1991 	} else if (m->valid != 0)
1992 		return (m);
1993 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
1994 		pmap_zero_page(m);
1995 	return (m);
1996 }
1997 
1998 /*
1999  * Mapping function for valid bits or for dirty bits in
2000  * a page.  May not block.
2001  *
2002  * Inputs are required to range within a page.
2003  */
2004 int
2005 vm_page_bits(int base, int size)
2006 {
2007 	int first_bit;
2008 	int last_bit;
2009 
2010 	KASSERT(
2011 	    base + size <= PAGE_SIZE,
2012 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
2013 	);
2014 
2015 	if (size == 0)		/* handle degenerate case */
2016 		return (0);
2017 
2018 	first_bit = base >> DEV_BSHIFT;
2019 	last_bit = (base + size - 1) >> DEV_BSHIFT;
2020 
2021 	return ((2 << last_bit) - (1 << first_bit));
2022 }
2023 
2024 /*
2025  *	vm_page_set_valid:
2026  *
2027  *	Sets portions of a page valid.  The arguments are expected
2028  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2029  *	of any partial chunks touched by the range.  The invalid portion of
2030  *	such chunks will be zeroed.
2031  *
2032  *	(base + size) must be less then or equal to PAGE_SIZE.
2033  */
2034 void
2035 vm_page_set_valid(vm_page_t m, int base, int size)
2036 {
2037 	int endoff, frag;
2038 
2039 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2040 	if (size == 0)	/* handle degenerate case */
2041 		return;
2042 
2043 	/*
2044 	 * If the base is not DEV_BSIZE aligned and the valid
2045 	 * bit is clear, we have to zero out a portion of the
2046 	 * first block.
2047 	 */
2048 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2049 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2050 		pmap_zero_page_area(m, frag, base - frag);
2051 
2052 	/*
2053 	 * If the ending offset is not DEV_BSIZE aligned and the
2054 	 * valid bit is clear, we have to zero out a portion of
2055 	 * the last block.
2056 	 */
2057 	endoff = base + size;
2058 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2059 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2060 		pmap_zero_page_area(m, endoff,
2061 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2062 
2063 	/*
2064 	 * Assert that no previously invalid block that is now being validated
2065 	 * is already dirty.
2066 	 */
2067 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2068 	    ("vm_page_set_valid: page %p is dirty", m));
2069 
2070 	/*
2071 	 * Set valid bits inclusive of any overlap.
2072 	 */
2073 	m->valid |= vm_page_bits(base, size);
2074 }
2075 
2076 /*
2077  * Clear the given bits from the specified page's dirty field.
2078  */
2079 static __inline void
2080 vm_page_clear_dirty_mask(vm_page_t m, int pagebits)
2081 {
2082 
2083 	/*
2084 	 * If the object is locked and the page is neither VPO_BUSY nor
2085 	 * PG_WRITEABLE, then the page's dirty field cannot possibly be
2086 	 * modified by a concurrent pmap operation.
2087 	 */
2088 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2089 	if ((m->oflags & VPO_BUSY) == 0 && (m->flags & PG_WRITEABLE) == 0)
2090 		m->dirty &= ~pagebits;
2091 	else {
2092 		vm_page_lock_queues();
2093 		m->dirty &= ~pagebits;
2094 		vm_page_unlock_queues();
2095 	}
2096 }
2097 
2098 /*
2099  *	vm_page_set_validclean:
2100  *
2101  *	Sets portions of a page valid and clean.  The arguments are expected
2102  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2103  *	of any partial chunks touched by the range.  The invalid portion of
2104  *	such chunks will be zero'd.
2105  *
2106  *	This routine may not block.
2107  *
2108  *	(base + size) must be less then or equal to PAGE_SIZE.
2109  */
2110 void
2111 vm_page_set_validclean(vm_page_t m, int base, int size)
2112 {
2113 	u_int oldvalid;
2114 	int endoff, frag, pagebits;
2115 
2116 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2117 	if (size == 0)	/* handle degenerate case */
2118 		return;
2119 
2120 	/*
2121 	 * If the base is not DEV_BSIZE aligned and the valid
2122 	 * bit is clear, we have to zero out a portion of the
2123 	 * first block.
2124 	 */
2125 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2126 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2127 		pmap_zero_page_area(m, frag, base - frag);
2128 
2129 	/*
2130 	 * If the ending offset is not DEV_BSIZE aligned and the
2131 	 * valid bit is clear, we have to zero out a portion of
2132 	 * the last block.
2133 	 */
2134 	endoff = base + size;
2135 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2136 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2137 		pmap_zero_page_area(m, endoff,
2138 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2139 
2140 	/*
2141 	 * Set valid, clear dirty bits.  If validating the entire
2142 	 * page we can safely clear the pmap modify bit.  We also
2143 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
2144 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
2145 	 * be set again.
2146 	 *
2147 	 * We set valid bits inclusive of any overlap, but we can only
2148 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
2149 	 * the range.
2150 	 */
2151 	oldvalid = m->valid;
2152 	pagebits = vm_page_bits(base, size);
2153 	m->valid |= pagebits;
2154 #if 0	/* NOT YET */
2155 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2156 		frag = DEV_BSIZE - frag;
2157 		base += frag;
2158 		size -= frag;
2159 		if (size < 0)
2160 			size = 0;
2161 	}
2162 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2163 #endif
2164 	if (base == 0 && size == PAGE_SIZE) {
2165 		/*
2166 		 * The page can only be modified within the pmap if it is
2167 		 * mapped, and it can only be mapped if it was previously
2168 		 * fully valid.
2169 		 */
2170 		if (oldvalid == VM_PAGE_BITS_ALL)
2171 			/*
2172 			 * Perform the pmap_clear_modify() first.  Otherwise,
2173 			 * a concurrent pmap operation, such as
2174 			 * pmap_protect(), could clear a modification in the
2175 			 * pmap and set the dirty field on the page before
2176 			 * pmap_clear_modify() had begun and after the dirty
2177 			 * field was cleared here.
2178 			 */
2179 			pmap_clear_modify(m);
2180 		m->dirty = 0;
2181 		m->oflags &= ~VPO_NOSYNC;
2182 	} else if (oldvalid != VM_PAGE_BITS_ALL)
2183 		m->dirty &= ~pagebits;
2184 	else
2185 		vm_page_clear_dirty_mask(m, pagebits);
2186 }
2187 
2188 void
2189 vm_page_clear_dirty(vm_page_t m, int base, int size)
2190 {
2191 
2192 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2193 }
2194 
2195 /*
2196  *	vm_page_set_invalid:
2197  *
2198  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
2199  *	valid and dirty bits for the effected areas are cleared.
2200  *
2201  *	May not block.
2202  */
2203 void
2204 vm_page_set_invalid(vm_page_t m, int base, int size)
2205 {
2206 	int bits;
2207 
2208 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2209 	KASSERT((m->oflags & VPO_BUSY) == 0,
2210 	    ("vm_page_set_invalid: page %p is busy", m));
2211 	bits = vm_page_bits(base, size);
2212 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2213 		pmap_remove_all(m);
2214 	KASSERT(!pmap_page_is_mapped(m),
2215 	    ("vm_page_set_invalid: page %p is mapped", m));
2216 	m->valid &= ~bits;
2217 	m->dirty &= ~bits;
2218 	m->object->generation++;
2219 }
2220 
2221 /*
2222  * vm_page_zero_invalid()
2223  *
2224  *	The kernel assumes that the invalid portions of a page contain
2225  *	garbage, but such pages can be mapped into memory by user code.
2226  *	When this occurs, we must zero out the non-valid portions of the
2227  *	page so user code sees what it expects.
2228  *
2229  *	Pages are most often semi-valid when the end of a file is mapped
2230  *	into memory and the file's size is not page aligned.
2231  */
2232 void
2233 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2234 {
2235 	int b;
2236 	int i;
2237 
2238 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2239 	/*
2240 	 * Scan the valid bits looking for invalid sections that
2241 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
2242 	 * valid bit may be set ) have already been zerod by
2243 	 * vm_page_set_validclean().
2244 	 */
2245 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2246 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
2247 		    (m->valid & (1 << i))
2248 		) {
2249 			if (i > b) {
2250 				pmap_zero_page_area(m,
2251 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2252 			}
2253 			b = i + 1;
2254 		}
2255 	}
2256 
2257 	/*
2258 	 * setvalid is TRUE when we can safely set the zero'd areas
2259 	 * as being valid.  We can do this if there are no cache consistancy
2260 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
2261 	 */
2262 	if (setvalid)
2263 		m->valid = VM_PAGE_BITS_ALL;
2264 }
2265 
2266 /*
2267  *	vm_page_is_valid:
2268  *
2269  *	Is (partial) page valid?  Note that the case where size == 0
2270  *	will return FALSE in the degenerate case where the page is
2271  *	entirely invalid, and TRUE otherwise.
2272  *
2273  *	May not block.
2274  */
2275 int
2276 vm_page_is_valid(vm_page_t m, int base, int size)
2277 {
2278 	int bits = vm_page_bits(base, size);
2279 
2280 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2281 	if (m->valid && ((m->valid & bits) == bits))
2282 		return 1;
2283 	else
2284 		return 0;
2285 }
2286 
2287 /*
2288  * update dirty bits from pmap/mmu.  May not block.
2289  */
2290 void
2291 vm_page_test_dirty(vm_page_t m)
2292 {
2293 
2294 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2295 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2296 		vm_page_dirty(m);
2297 }
2298 
2299 int so_zerocp_fullpage = 0;
2300 
2301 /*
2302  *	Replace the given page with a copy.  The copied page assumes
2303  *	the portion of the given page's "wire_count" that is not the
2304  *	responsibility of this copy-on-write mechanism.
2305  *
2306  *	The object containing the given page must have a non-zero
2307  *	paging-in-progress count and be locked.
2308  */
2309 void
2310 vm_page_cowfault(vm_page_t m)
2311 {
2312 	vm_page_t mnew;
2313 	vm_object_t object;
2314 	vm_pindex_t pindex;
2315 
2316 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
2317 	vm_page_lock_assert(m, MA_OWNED);
2318 	object = m->object;
2319 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2320 	KASSERT(object->paging_in_progress != 0,
2321 	    ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2322 	    object));
2323 	pindex = m->pindex;
2324 
2325  retry_alloc:
2326 	pmap_remove_all(m);
2327 	vm_page_remove(m);
2328 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2329 	if (mnew == NULL) {
2330 		vm_page_insert(m, object, pindex);
2331 		vm_page_unlock(m);
2332 		VM_OBJECT_UNLOCK(object);
2333 		VM_WAIT;
2334 		VM_OBJECT_LOCK(object);
2335 		if (m == vm_page_lookup(object, pindex)) {
2336 			vm_page_lock(m);
2337 			goto retry_alloc;
2338 		} else {
2339 			/*
2340 			 * Page disappeared during the wait.
2341 			 */
2342 			return;
2343 		}
2344 	}
2345 
2346 	if (m->cow == 0) {
2347 		/*
2348 		 * check to see if we raced with an xmit complete when
2349 		 * waiting to allocate a page.  If so, put things back
2350 		 * the way they were
2351 		 */
2352 		vm_page_unlock(m);
2353 		vm_page_lock(mnew);
2354 		vm_page_free(mnew);
2355 		vm_page_unlock(mnew);
2356 		vm_page_insert(m, object, pindex);
2357 	} else { /* clear COW & copy page */
2358 		if (!so_zerocp_fullpage)
2359 			pmap_copy_page(m, mnew);
2360 		mnew->valid = VM_PAGE_BITS_ALL;
2361 		vm_page_dirty(mnew);
2362 		mnew->wire_count = m->wire_count - m->cow;
2363 		m->wire_count = m->cow;
2364 		vm_page_unlock(m);
2365 	}
2366 }
2367 
2368 void
2369 vm_page_cowclear(vm_page_t m)
2370 {
2371 
2372 	vm_page_lock_assert(m, MA_OWNED);
2373 	if (m->cow) {
2374 		m->cow--;
2375 		/*
2376 		 * let vm_fault add back write permission  lazily
2377 		 */
2378 	}
2379 	/*
2380 	 *  sf_buf_free() will free the page, so we needn't do it here
2381 	 */
2382 }
2383 
2384 int
2385 vm_page_cowsetup(vm_page_t m)
2386 {
2387 
2388 	vm_page_lock_assert(m, MA_OWNED);
2389 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 ||
2390 	    m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
2391 		return (EBUSY);
2392 	m->cow++;
2393 	pmap_remove_write(m);
2394 	VM_OBJECT_UNLOCK(m->object);
2395 	return (0);
2396 }
2397 
2398 #include "opt_ddb.h"
2399 #ifdef DDB
2400 #include <sys/kernel.h>
2401 
2402 #include <ddb/ddb.h>
2403 
2404 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2405 {
2406 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
2407 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
2408 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
2409 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
2410 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
2411 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
2412 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
2413 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
2414 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
2415 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
2416 }
2417 
2418 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2419 {
2420 
2421 	db_printf("PQ_FREE:");
2422 	db_printf(" %d", cnt.v_free_count);
2423 	db_printf("\n");
2424 
2425 	db_printf("PQ_CACHE:");
2426 	db_printf(" %d", cnt.v_cache_count);
2427 	db_printf("\n");
2428 
2429 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2430 		*vm_page_queues[PQ_ACTIVE].cnt,
2431 		*vm_page_queues[PQ_INACTIVE].cnt);
2432 }
2433 #endif /* DDB */
2434