xref: /freebsd/sys/vm/vm_page.c (revision acd3428b7d3e94cef0e1881c868cb4b131d4ff41)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
33  */
34 
35 /*-
36  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
37  * All rights reserved.
38  *
39  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40  *
41  * Permission to use, copy, modify and distribute this software and
42  * its documentation is hereby granted, provided that both the copyright
43  * notice and this permission notice appear in all copies of the
44  * software, derivative works or modified versions, and any portions
45  * thereof, and that both notices appear in supporting documentation.
46  *
47  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
48  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
49  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50  *
51  * Carnegie Mellon requests users of this software to return to
52  *
53  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
54  *  School of Computer Science
55  *  Carnegie Mellon University
56  *  Pittsburgh PA 15213-3890
57  *
58  * any improvements or extensions that they make and grant Carnegie the
59  * rights to redistribute these changes.
60  */
61 
62 /*
63  *			GENERAL RULES ON VM_PAGE MANIPULATION
64  *
65  *	- a pageq mutex is required when adding or removing a page from a
66  *	  page queue (vm_page_queue[]), regardless of other mutexes or the
67  *	  busy state of a page.
68  *
69  *	- a hash chain mutex is required when associating or disassociating
70  *	  a page from the VM PAGE CACHE hash table (vm_page_buckets),
71  *	  regardless of other mutexes or the busy state of a page.
72  *
73  *	- either a hash chain mutex OR a busied page is required in order
74  *	  to modify the page flags.  A hash chain mutex must be obtained in
75  *	  order to busy a page.  A page's flags cannot be modified by a
76  *	  hash chain mutex if the page is marked busy.
77  *
78  *	- The object memq mutex is held when inserting or removing
79  *	  pages from an object (vm_page_insert() or vm_page_remove()).  This
80  *	  is different from the object's main mutex.
81  *
82  *	Generally speaking, you have to be aware of side effects when running
83  *	vm_page ops.  A vm_page_lookup() will return with the hash chain
84  *	locked, whether it was able to lookup the page or not.  vm_page_free(),
85  *	vm_page_cache(), vm_page_activate(), and a number of other routines
86  *	will release the hash chain mutex for you.  Intermediate manipulation
87  *	routines such as vm_page_flag_set() expect the hash chain to be held
88  *	on entry and the hash chain will remain held on return.
89  *
90  *	pageq scanning can only occur with the pageq in question locked.
91  *	We have a known bottleneck with the active queue, but the cache
92  *	and free queues are actually arrays already.
93  */
94 
95 /*
96  *	Resident memory management module.
97  */
98 
99 #include <sys/cdefs.h>
100 __FBSDID("$FreeBSD$");
101 
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/lock.h>
105 #include <sys/kernel.h>
106 #include <sys/malloc.h>
107 #include <sys/mutex.h>
108 #include <sys/proc.h>
109 #include <sys/sysctl.h>
110 #include <sys/vmmeter.h>
111 #include <sys/vnode.h>
112 
113 #include <vm/vm.h>
114 #include <vm/vm_param.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_object.h>
117 #include <vm/vm_page.h>
118 #include <vm/vm_pageout.h>
119 #include <vm/vm_pager.h>
120 #include <vm/vm_extern.h>
121 #include <vm/uma.h>
122 #include <vm/uma_int.h>
123 
124 #include <machine/md_var.h>
125 
126 /*
127  *	Associated with page of user-allocatable memory is a
128  *	page structure.
129  */
130 
131 struct mtx vm_page_queue_mtx;
132 struct mtx vm_page_queue_free_mtx;
133 
134 vm_page_t vm_page_array = 0;
135 int vm_page_array_size = 0;
136 long first_page = 0;
137 int vm_page_zero_count = 0;
138 
139 static int boot_pages = UMA_BOOT_PAGES;
140 TUNABLE_INT("vm.boot_pages", &boot_pages);
141 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
142 	"number of pages allocated for bootstrapping the VM system");
143 
144 /*
145  *	vm_set_page_size:
146  *
147  *	Sets the page size, perhaps based upon the memory
148  *	size.  Must be called before any use of page-size
149  *	dependent functions.
150  */
151 void
152 vm_set_page_size(void)
153 {
154 	if (cnt.v_page_size == 0)
155 		cnt.v_page_size = PAGE_SIZE;
156 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
157 		panic("vm_set_page_size: page size not a power of two");
158 }
159 
160 /*
161  *	vm_page_blacklist_lookup:
162  *
163  *	See if a physical address in this page has been listed
164  *	in the blacklist tunable.  Entries in the tunable are
165  *	separated by spaces or commas.  If an invalid integer is
166  *	encountered then the rest of the string is skipped.
167  */
168 static int
169 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
170 {
171 	vm_paddr_t bad;
172 	char *cp, *pos;
173 
174 	for (pos = list; *pos != '\0'; pos = cp) {
175 		bad = strtoq(pos, &cp, 0);
176 		if (*cp != '\0') {
177 			if (*cp == ' ' || *cp == ',') {
178 				cp++;
179 				if (cp == pos)
180 					continue;
181 			} else
182 				break;
183 		}
184 		if (pa == trunc_page(bad))
185 			return (1);
186 	}
187 	return (0);
188 }
189 
190 /*
191  *	vm_page_startup:
192  *
193  *	Initializes the resident memory module.
194  *
195  *	Allocates memory for the page cells, and
196  *	for the object/offset-to-page hash table headers.
197  *	Each page cell is initialized and placed on the free list.
198  */
199 vm_offset_t
200 vm_page_startup(vm_offset_t vaddr)
201 {
202 	vm_offset_t mapped;
203 	vm_size_t npages;
204 	vm_paddr_t page_range;
205 	vm_paddr_t new_end;
206 	int i;
207 	vm_paddr_t pa;
208 	int nblocks;
209 	vm_paddr_t last_pa;
210 	char *list;
211 
212 	/* the biggest memory array is the second group of pages */
213 	vm_paddr_t end;
214 	vm_paddr_t biggestsize;
215 	int biggestone;
216 
217 	vm_paddr_t total;
218 
219 	total = 0;
220 	biggestsize = 0;
221 	biggestone = 0;
222 	nblocks = 0;
223 	vaddr = round_page(vaddr);
224 
225 	for (i = 0; phys_avail[i + 1]; i += 2) {
226 		phys_avail[i] = round_page(phys_avail[i]);
227 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
228 	}
229 
230 	for (i = 0; phys_avail[i + 1]; i += 2) {
231 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
232 
233 		if (size > biggestsize) {
234 			biggestone = i;
235 			biggestsize = size;
236 		}
237 		++nblocks;
238 		total += size;
239 	}
240 
241 	end = phys_avail[biggestone+1];
242 
243 	/*
244 	 * Initialize the locks.
245 	 */
246 	mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF |
247 	    MTX_RECURSE);
248 	mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
249 	    MTX_SPIN);
250 
251 	/*
252 	 * Initialize the queue headers for the free queue, the active queue
253 	 * and the inactive queue.
254 	 */
255 	vm_pageq_init();
256 
257 	/*
258 	 * Allocate memory for use when boot strapping the kernel memory
259 	 * allocator.
260 	 */
261 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
262 	new_end = trunc_page(new_end);
263 	mapped = pmap_map(&vaddr, new_end, end,
264 	    VM_PROT_READ | VM_PROT_WRITE);
265 	bzero((void *)mapped, end - new_end);
266 	uma_startup((void *)mapped, boot_pages);
267 
268 #if defined(__amd64__) || defined(__i386__)
269 	/*
270 	 * Allocate a bitmap to indicate that a random physical page
271 	 * needs to be included in a minidump.
272 	 *
273 	 * The amd64 port needs this to indicate which direct map pages
274 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
275 	 *
276 	 * However, i386 still needs this workspace internally within the
277 	 * minidump code.  In theory, they are not needed on i386, but are
278 	 * included should the sf_buf code decide to use them.
279 	 */
280 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
281 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
282 	new_end -= vm_page_dump_size;
283 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
284 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
285 	bzero((void *)vm_page_dump, vm_page_dump_size);
286 #endif
287 	/*
288 	 * Compute the number of pages of memory that will be available for
289 	 * use (taking into account the overhead of a page structure per
290 	 * page).
291 	 */
292 	first_page = phys_avail[0] / PAGE_SIZE;
293 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
294 	npages = (total - (page_range * sizeof(struct vm_page)) -
295 	    (end - new_end)) / PAGE_SIZE;
296 	end = new_end;
297 
298 	/*
299 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
300 	 */
301 	vaddr += PAGE_SIZE;
302 
303 	/*
304 	 * Initialize the mem entry structures now, and put them in the free
305 	 * queue.
306 	 */
307 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
308 	mapped = pmap_map(&vaddr, new_end, end,
309 	    VM_PROT_READ | VM_PROT_WRITE);
310 	vm_page_array = (vm_page_t) mapped;
311 #ifdef __amd64__
312 	/*
313 	 * pmap_map on amd64 comes out of the direct-map, not kvm like i386,
314 	 * so the pages must be tracked for a crashdump to include this data.
315 	 * This includes the vm_page_array and the early UMA bootstrap pages.
316 	 */
317 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
318 		dump_add_page(pa);
319 #endif
320 	phys_avail[biggestone + 1] = new_end;
321 
322 	/*
323 	 * Clear all of the page structures
324 	 */
325 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
326 	vm_page_array_size = page_range;
327 
328 	/*
329 	 * Construct the free queue(s) in descending order (by physical
330 	 * address) so that the first 16MB of physical memory is allocated
331 	 * last rather than first.  On large-memory machines, this avoids
332 	 * the exhaustion of low physical memory before isa_dma_init has run.
333 	 */
334 	cnt.v_page_count = 0;
335 	cnt.v_free_count = 0;
336 	list = getenv("vm.blacklist");
337 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
338 		pa = phys_avail[i];
339 		last_pa = phys_avail[i + 1];
340 		while (pa < last_pa && npages-- > 0) {
341 			if (list != NULL &&
342 			    vm_page_blacklist_lookup(list, pa))
343 				printf("Skipping page with pa 0x%jx\n",
344 				    (uintmax_t)pa);
345 			else
346 				vm_pageq_add_new_page(pa);
347 			pa += PAGE_SIZE;
348 		}
349 	}
350 	freeenv(list);
351 	return (vaddr);
352 }
353 
354 void
355 vm_page_flag_set(vm_page_t m, unsigned short bits)
356 {
357 
358 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
359 	m->flags |= bits;
360 }
361 
362 void
363 vm_page_flag_clear(vm_page_t m, unsigned short bits)
364 {
365 
366 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
367 	m->flags &= ~bits;
368 }
369 
370 void
371 vm_page_busy(vm_page_t m)
372 {
373 
374 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
375 	KASSERT((m->oflags & VPO_BUSY) == 0,
376 	    ("vm_page_busy: page already busy!!!"));
377 	m->oflags |= VPO_BUSY;
378 }
379 
380 /*
381  *      vm_page_flash:
382  *
383  *      wakeup anyone waiting for the page.
384  */
385 void
386 vm_page_flash(vm_page_t m)
387 {
388 
389 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
390 	if (m->oflags & VPO_WANTED) {
391 		m->oflags &= ~VPO_WANTED;
392 		wakeup(m);
393 	}
394 }
395 
396 /*
397  *      vm_page_wakeup:
398  *
399  *      clear the VPO_BUSY flag and wakeup anyone waiting for the
400  *      page.
401  *
402  */
403 void
404 vm_page_wakeup(vm_page_t m)
405 {
406 
407 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
408 	KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
409 	m->oflags &= ~VPO_BUSY;
410 	vm_page_flash(m);
411 }
412 
413 void
414 vm_page_io_start(vm_page_t m)
415 {
416 
417 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
418 	m->busy++;
419 }
420 
421 void
422 vm_page_io_finish(vm_page_t m)
423 {
424 
425 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
426 	m->busy--;
427 	if (m->busy == 0)
428 		vm_page_flash(m);
429 }
430 
431 /*
432  * Keep page from being freed by the page daemon
433  * much of the same effect as wiring, except much lower
434  * overhead and should be used only for *very* temporary
435  * holding ("wiring").
436  */
437 void
438 vm_page_hold(vm_page_t mem)
439 {
440 
441 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
442         mem->hold_count++;
443 }
444 
445 void
446 vm_page_unhold(vm_page_t mem)
447 {
448 
449 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
450 	--mem->hold_count;
451 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
452 	if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD))
453 		vm_page_free_toq(mem);
454 }
455 
456 /*
457  *	vm_page_free:
458  *
459  *	Free a page
460  *
461  *	The clearing of PG_ZERO is a temporary safety until the code can be
462  *	reviewed to determine that PG_ZERO is being properly cleared on
463  *	write faults or maps.  PG_ZERO was previously cleared in
464  *	vm_page_alloc().
465  */
466 void
467 vm_page_free(vm_page_t m)
468 {
469 	vm_page_flag_clear(m, PG_ZERO);
470 	vm_page_free_toq(m);
471 	vm_page_zero_idle_wakeup();
472 }
473 
474 /*
475  *	vm_page_free_zero:
476  *
477  *	Free a page to the zerod-pages queue
478  */
479 void
480 vm_page_free_zero(vm_page_t m)
481 {
482 	vm_page_flag_set(m, PG_ZERO);
483 	vm_page_free_toq(m);
484 }
485 
486 /*
487  *	vm_page_sleep:
488  *
489  *	Sleep and release the page queues lock.
490  *
491  *	The object containing the given page must be locked.
492  */
493 void
494 vm_page_sleep(vm_page_t m, const char *msg)
495 {
496 
497 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
498 	if (!mtx_owned(&vm_page_queue_mtx))
499 		vm_page_lock_queues();
500 	vm_page_flag_set(m, PG_REFERENCED);
501 	vm_page_unlock_queues();
502 
503 	/*
504 	 * It's possible that while we sleep, the page will get
505 	 * unbusied and freed.  If we are holding the object
506 	 * lock, we will assume we hold a reference to the object
507 	 * such that even if m->object changes, we can re-lock
508 	 * it.
509 	 */
510 	m->oflags |= VPO_WANTED;
511 	msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
512 }
513 
514 /*
515  *	vm_page_dirty:
516  *
517  *	make page all dirty
518  */
519 void
520 vm_page_dirty(vm_page_t m)
521 {
522 	KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_CACHE,
523 	    ("vm_page_dirty: page in cache!"));
524 	KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_FREE,
525 	    ("vm_page_dirty: page is free!"));
526 	m->dirty = VM_PAGE_BITS_ALL;
527 }
528 
529 /*
530  *	vm_page_splay:
531  *
532  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
533  *	the vm_page containing the given pindex.  If, however, that
534  *	pindex is not found in the vm_object, returns a vm_page that is
535  *	adjacent to the pindex, coming before or after it.
536  */
537 vm_page_t
538 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
539 {
540 	struct vm_page dummy;
541 	vm_page_t lefttreemax, righttreemin, y;
542 
543 	if (root == NULL)
544 		return (root);
545 	lefttreemax = righttreemin = &dummy;
546 	for (;; root = y) {
547 		if (pindex < root->pindex) {
548 			if ((y = root->left) == NULL)
549 				break;
550 			if (pindex < y->pindex) {
551 				/* Rotate right. */
552 				root->left = y->right;
553 				y->right = root;
554 				root = y;
555 				if ((y = root->left) == NULL)
556 					break;
557 			}
558 			/* Link into the new root's right tree. */
559 			righttreemin->left = root;
560 			righttreemin = root;
561 		} else if (pindex > root->pindex) {
562 			if ((y = root->right) == NULL)
563 				break;
564 			if (pindex > y->pindex) {
565 				/* Rotate left. */
566 				root->right = y->left;
567 				y->left = root;
568 				root = y;
569 				if ((y = root->right) == NULL)
570 					break;
571 			}
572 			/* Link into the new root's left tree. */
573 			lefttreemax->right = root;
574 			lefttreemax = root;
575 		} else
576 			break;
577 	}
578 	/* Assemble the new root. */
579 	lefttreemax->right = root->left;
580 	righttreemin->left = root->right;
581 	root->left = dummy.right;
582 	root->right = dummy.left;
583 	return (root);
584 }
585 
586 /*
587  *	vm_page_insert:		[ internal use only ]
588  *
589  *	Inserts the given mem entry into the object and object list.
590  *
591  *	The pagetables are not updated but will presumably fault the page
592  *	in if necessary, or if a kernel page the caller will at some point
593  *	enter the page into the kernel's pmap.  We are not allowed to block
594  *	here so we *can't* do this anyway.
595  *
596  *	The object and page must be locked.
597  *	This routine may not block.
598  */
599 void
600 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
601 {
602 	vm_page_t root;
603 
604 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
605 	if (m->object != NULL)
606 		panic("vm_page_insert: page already inserted");
607 
608 	/*
609 	 * Record the object/offset pair in this page
610 	 */
611 	m->object = object;
612 	m->pindex = pindex;
613 
614 	/*
615 	 * Now link into the object's ordered list of backed pages.
616 	 */
617 	root = object->root;
618 	if (root == NULL) {
619 		m->left = NULL;
620 		m->right = NULL;
621 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
622 	} else {
623 		root = vm_page_splay(pindex, root);
624 		if (pindex < root->pindex) {
625 			m->left = root->left;
626 			m->right = root;
627 			root->left = NULL;
628 			TAILQ_INSERT_BEFORE(root, m, listq);
629 		} else if (pindex == root->pindex)
630 			panic("vm_page_insert: offset already allocated");
631 		else {
632 			m->right = root->right;
633 			m->left = root;
634 			root->right = NULL;
635 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
636 		}
637 	}
638 	object->root = m;
639 	object->generation++;
640 
641 	/*
642 	 * show that the object has one more resident page.
643 	 */
644 	object->resident_page_count++;
645 	/*
646 	 * Hold the vnode until the last page is released.
647 	 */
648 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
649 		vhold((struct vnode *)object->handle);
650 
651 	/*
652 	 * Since we are inserting a new and possibly dirty page,
653 	 * update the object's OBJ_MIGHTBEDIRTY flag.
654 	 */
655 	if (m->flags & PG_WRITEABLE)
656 		vm_object_set_writeable_dirty(object);
657 }
658 
659 /*
660  *	vm_page_remove:
661  *				NOTE: used by device pager as well -wfj
662  *
663  *	Removes the given mem entry from the object/offset-page
664  *	table and the object page list, but do not invalidate/terminate
665  *	the backing store.
666  *
667  *	The object and page must be locked.
668  *	The underlying pmap entry (if any) is NOT removed here.
669  *	This routine may not block.
670  */
671 void
672 vm_page_remove(vm_page_t m)
673 {
674 	vm_object_t object;
675 	vm_page_t root;
676 
677 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
678 	if ((object = m->object) == NULL)
679 		return;
680 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
681 	if (m->oflags & VPO_BUSY) {
682 		m->oflags &= ~VPO_BUSY;
683 		vm_page_flash(m);
684 	}
685 
686 	/*
687 	 * Now remove from the object's list of backed pages.
688 	 */
689 	if (m != object->root)
690 		vm_page_splay(m->pindex, object->root);
691 	if (m->left == NULL)
692 		root = m->right;
693 	else {
694 		root = vm_page_splay(m->pindex, m->left);
695 		root->right = m->right;
696 	}
697 	object->root = root;
698 	TAILQ_REMOVE(&object->memq, m, listq);
699 
700 	/*
701 	 * And show that the object has one fewer resident page.
702 	 */
703 	object->resident_page_count--;
704 	object->generation++;
705 	/*
706 	 * The vnode may now be recycled.
707 	 */
708 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
709 		vdrop((struct vnode *)object->handle);
710 
711 	m->object = NULL;
712 }
713 
714 /*
715  *	vm_page_lookup:
716  *
717  *	Returns the page associated with the object/offset
718  *	pair specified; if none is found, NULL is returned.
719  *
720  *	The object must be locked.
721  *	This routine may not block.
722  *	This is a critical path routine
723  */
724 vm_page_t
725 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
726 {
727 	vm_page_t m;
728 
729 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
730 	if ((m = object->root) != NULL && m->pindex != pindex) {
731 		m = vm_page_splay(pindex, m);
732 		if ((object->root = m)->pindex != pindex)
733 			m = NULL;
734 	}
735 	return (m);
736 }
737 
738 /*
739  *	vm_page_rename:
740  *
741  *	Move the given memory entry from its
742  *	current object to the specified target object/offset.
743  *
744  *	The object must be locked.
745  *	This routine may not block.
746  *
747  *	Note: swap associated with the page must be invalidated by the move.  We
748  *	      have to do this for several reasons:  (1) we aren't freeing the
749  *	      page, (2) we are dirtying the page, (3) the VM system is probably
750  *	      moving the page from object A to B, and will then later move
751  *	      the backing store from A to B and we can't have a conflict.
752  *
753  *	Note: we *always* dirty the page.  It is necessary both for the
754  *	      fact that we moved it, and because we may be invalidating
755  *	      swap.  If the page is on the cache, we have to deactivate it
756  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
757  *	      on the cache.
758  */
759 void
760 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
761 {
762 
763 	vm_page_remove(m);
764 	vm_page_insert(m, new_object, new_pindex);
765 	if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
766 		vm_page_deactivate(m);
767 	vm_page_dirty(m);
768 }
769 
770 /*
771  *	vm_page_select_cache:
772  *
773  *	Move a page of the given color from the cache queue to the free
774  *	queue.  As pages might be found, but are not applicable, they are
775  *	deactivated.
776  *
777  *	This routine may not block.
778  */
779 vm_page_t
780 vm_page_select_cache(int color)
781 {
782 	vm_object_t object;
783 	vm_page_t m;
784 	boolean_t was_trylocked;
785 
786 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
787 	while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) {
788 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
789 		KASSERT(!pmap_page_is_mapped(m),
790 		    ("Found mapped cache page %p", m));
791 		KASSERT((m->flags & PG_UNMANAGED) == 0,
792 		    ("Found unmanaged cache page %p", m));
793 		KASSERT(m->wire_count == 0, ("Found wired cache page %p", m));
794 		if (m->hold_count == 0 && (object = m->object,
795 		    (was_trylocked = VM_OBJECT_TRYLOCK(object)) ||
796 		    VM_OBJECT_LOCKED(object))) {
797 			KASSERT((m->oflags & VPO_BUSY) == 0 && m->busy == 0,
798 			    ("Found busy cache page %p", m));
799 			vm_page_free(m);
800 			if (was_trylocked)
801 				VM_OBJECT_UNLOCK(object);
802 			break;
803 		}
804 		vm_page_deactivate(m);
805 	}
806 	return (m);
807 }
808 
809 /*
810  *	vm_page_alloc:
811  *
812  *	Allocate and return a memory cell associated
813  *	with this VM object/offset pair.
814  *
815  *	page_req classes:
816  *	VM_ALLOC_NORMAL		normal process request
817  *	VM_ALLOC_SYSTEM		system *really* needs a page
818  *	VM_ALLOC_INTERRUPT	interrupt time request
819  *	VM_ALLOC_ZERO		zero page
820  *
821  *	This routine may not block.
822  *
823  *	Additional special handling is required when called from an
824  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
825  *	the page cache in this case.
826  */
827 vm_page_t
828 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
829 {
830 	vm_page_t m = NULL;
831 	int color, flags, page_req;
832 
833 	page_req = req & VM_ALLOC_CLASS_MASK;
834 	KASSERT(curthread->td_intr_nesting_level == 0 ||
835 	    page_req == VM_ALLOC_INTERRUPT,
836 	    ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context"));
837 
838 	if ((req & VM_ALLOC_NOOBJ) == 0) {
839 		KASSERT(object != NULL,
840 		    ("vm_page_alloc: NULL object."));
841 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
842 		color = (pindex + object->pg_color) & PQ_COLORMASK;
843 	} else
844 		color = pindex & PQ_COLORMASK;
845 
846 	/*
847 	 * The pager is allowed to eat deeper into the free page list.
848 	 */
849 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
850 		page_req = VM_ALLOC_SYSTEM;
851 	};
852 
853 loop:
854 	mtx_lock_spin(&vm_page_queue_free_mtx);
855 	if (cnt.v_free_count > cnt.v_free_reserved ||
856 	    (page_req == VM_ALLOC_SYSTEM &&
857 	     cnt.v_cache_count == 0 &&
858 	     cnt.v_free_count > cnt.v_interrupt_free_min) ||
859 	    (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) {
860 		/*
861 		 * Allocate from the free queue if the number of free pages
862 		 * exceeds the minimum for the request class.
863 		 */
864 		m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0);
865 	} else if (page_req != VM_ALLOC_INTERRUPT) {
866 		mtx_unlock_spin(&vm_page_queue_free_mtx);
867 		/*
868 		 * Allocatable from cache (non-interrupt only).  On success,
869 		 * we must free the page and try again, thus ensuring that
870 		 * cnt.v_*_free_min counters are replenished.
871 		 */
872 		vm_page_lock_queues();
873 		if ((m = vm_page_select_cache(color)) == NULL) {
874 			KASSERT(cnt.v_cache_count == 0,
875 			    ("vm_page_alloc: cache queue is missing %d pages",
876 			    cnt.v_cache_count));
877 			vm_page_unlock_queues();
878 			atomic_add_int(&vm_pageout_deficit, 1);
879 			pagedaemon_wakeup();
880 
881 			if (page_req != VM_ALLOC_SYSTEM)
882 				return (NULL);
883 
884 			mtx_lock_spin(&vm_page_queue_free_mtx);
885 			if (cnt.v_free_count <= cnt.v_interrupt_free_min) {
886 				mtx_unlock_spin(&vm_page_queue_free_mtx);
887 				return (NULL);
888 			}
889 			m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0);
890 		} else {
891 			vm_page_unlock_queues();
892 			goto loop;
893 		}
894 	} else {
895 		/*
896 		 * Not allocatable from cache from interrupt, give up.
897 		 */
898 		mtx_unlock_spin(&vm_page_queue_free_mtx);
899 		atomic_add_int(&vm_pageout_deficit, 1);
900 		pagedaemon_wakeup();
901 		return (NULL);
902 	}
903 
904 	/*
905 	 *  At this point we had better have found a good page.
906 	 */
907 
908 	KASSERT(
909 	    m != NULL,
910 	    ("vm_page_alloc(): missing page on free queue")
911 	);
912 
913 	/*
914 	 * Remove from free queue
915 	 */
916 	vm_pageq_remove_nowakeup(m);
917 
918 	/*
919 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
920 	 */
921 	flags = 0;
922 	if (m->flags & PG_ZERO) {
923 		vm_page_zero_count--;
924 		if (req & VM_ALLOC_ZERO)
925 			flags = PG_ZERO;
926 	}
927 	m->flags = flags;
928 	if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ))
929 		m->oflags = 0;
930 	else
931 		m->oflags = VPO_BUSY;
932 	if (req & VM_ALLOC_WIRED) {
933 		atomic_add_int(&cnt.v_wire_count, 1);
934 		m->wire_count = 1;
935 	} else
936 		m->wire_count = 0;
937 	m->hold_count = 0;
938 	m->act_count = 0;
939 	m->busy = 0;
940 	m->valid = 0;
941 	KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
942 	mtx_unlock_spin(&vm_page_queue_free_mtx);
943 
944 	if ((req & VM_ALLOC_NOOBJ) == 0)
945 		vm_page_insert(m, object, pindex);
946 	else
947 		m->pindex = pindex;
948 
949 	/*
950 	 * Don't wakeup too often - wakeup the pageout daemon when
951 	 * we would be nearly out of memory.
952 	 */
953 	if (vm_paging_needed())
954 		pagedaemon_wakeup();
955 
956 	return (m);
957 }
958 
959 /*
960  *	vm_wait:	(also see VM_WAIT macro)
961  *
962  *	Block until free pages are available for allocation
963  *	- Called in various places before memory allocations.
964  */
965 void
966 vm_wait(void)
967 {
968 
969 	vm_page_lock_queues();
970 	if (curproc == pageproc) {
971 		vm_pageout_pages_needed = 1;
972 		msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx,
973 		    PDROP | PSWP, "VMWait", 0);
974 	} else {
975 		if (!vm_pages_needed) {
976 			vm_pages_needed = 1;
977 			wakeup(&vm_pages_needed);
978 		}
979 		msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM,
980 		    "vmwait", 0);
981 	}
982 }
983 
984 /*
985  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
986  *
987  *	Block until free pages are available for allocation
988  *	- Called only in vm_fault so that processes page faulting
989  *	  can be easily tracked.
990  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
991  *	  processes will be able to grab memory first.  Do not change
992  *	  this balance without careful testing first.
993  */
994 void
995 vm_waitpfault(void)
996 {
997 
998 	vm_page_lock_queues();
999 	if (!vm_pages_needed) {
1000 		vm_pages_needed = 1;
1001 		wakeup(&vm_pages_needed);
1002 	}
1003 	msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER,
1004 	    "pfault", 0);
1005 }
1006 
1007 /*
1008  *	vm_page_activate:
1009  *
1010  *	Put the specified page on the active list (if appropriate).
1011  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1012  *	mess with it.
1013  *
1014  *	The page queues must be locked.
1015  *	This routine may not block.
1016  */
1017 void
1018 vm_page_activate(vm_page_t m)
1019 {
1020 
1021 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1022 	if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) {
1023 		if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1024 			cnt.v_reactivated++;
1025 		vm_pageq_remove(m);
1026 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1027 			if (m->act_count < ACT_INIT)
1028 				m->act_count = ACT_INIT;
1029 			vm_pageq_enqueue(PQ_ACTIVE, m);
1030 		}
1031 	} else {
1032 		if (m->act_count < ACT_INIT)
1033 			m->act_count = ACT_INIT;
1034 	}
1035 }
1036 
1037 /*
1038  *	vm_page_free_wakeup:
1039  *
1040  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1041  *	routine is called when a page has been added to the cache or free
1042  *	queues.
1043  *
1044  *	The page queues must be locked.
1045  *	This routine may not block.
1046  */
1047 static inline void
1048 vm_page_free_wakeup(void)
1049 {
1050 
1051 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1052 	/*
1053 	 * if pageout daemon needs pages, then tell it that there are
1054 	 * some free.
1055 	 */
1056 	if (vm_pageout_pages_needed &&
1057 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1058 		wakeup(&vm_pageout_pages_needed);
1059 		vm_pageout_pages_needed = 0;
1060 	}
1061 	/*
1062 	 * wakeup processes that are waiting on memory if we hit a
1063 	 * high water mark. And wakeup scheduler process if we have
1064 	 * lots of memory. this process will swapin processes.
1065 	 */
1066 	if (vm_pages_needed && !vm_page_count_min()) {
1067 		vm_pages_needed = 0;
1068 		wakeup(&cnt.v_free_count);
1069 	}
1070 }
1071 
1072 /*
1073  *	vm_page_free_toq:
1074  *
1075  *	Returns the given page to the PQ_FREE list,
1076  *	disassociating it with any VM object.
1077  *
1078  *	Object and page must be locked prior to entry.
1079  *	This routine may not block.
1080  */
1081 
1082 void
1083 vm_page_free_toq(vm_page_t m)
1084 {
1085 	struct vpgqueues *pq;
1086 
1087 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1088 	KASSERT(!pmap_page_is_mapped(m),
1089 	    ("vm_page_free_toq: freeing mapped page %p", m));
1090 	cnt.v_tfree++;
1091 
1092 	if (m->busy || VM_PAGE_INQUEUE1(m, PQ_FREE)) {
1093 		printf(
1094 		"vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n",
1095 		    (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0,
1096 		    m->hold_count);
1097 		if (VM_PAGE_INQUEUE1(m, PQ_FREE))
1098 			panic("vm_page_free: freeing free page");
1099 		else
1100 			panic("vm_page_free: freeing busy page");
1101 	}
1102 
1103 	/*
1104 	 * unqueue, then remove page.  Note that we cannot destroy
1105 	 * the page here because we do not want to call the pager's
1106 	 * callback routine until after we've put the page on the
1107 	 * appropriate free queue.
1108 	 */
1109 	vm_pageq_remove_nowakeup(m);
1110 	vm_page_remove(m);
1111 
1112 	/*
1113 	 * If fictitious remove object association and
1114 	 * return, otherwise delay object association removal.
1115 	 */
1116 	if ((m->flags & PG_FICTITIOUS) != 0) {
1117 		return;
1118 	}
1119 
1120 	m->valid = 0;
1121 	vm_page_undirty(m);
1122 
1123 	if (m->wire_count != 0) {
1124 		if (m->wire_count > 1) {
1125 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1126 				m->wire_count, (long)m->pindex);
1127 		}
1128 		panic("vm_page_free: freeing wired page");
1129 	}
1130 	if (m->hold_count != 0) {
1131 		m->flags &= ~PG_ZERO;
1132 		VM_PAGE_SETQUEUE2(m, PQ_HOLD);
1133 	} else
1134 		VM_PAGE_SETQUEUE1(m, PQ_FREE);
1135 	pq = &vm_page_queues[VM_PAGE_GETQUEUE(m)];
1136 	mtx_lock_spin(&vm_page_queue_free_mtx);
1137 	pq->lcnt++;
1138 	++(*pq->cnt);
1139 
1140 	/*
1141 	 * Put zero'd pages on the end ( where we look for zero'd pages
1142 	 * first ) and non-zerod pages at the head.
1143 	 */
1144 	if (m->flags & PG_ZERO) {
1145 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1146 		++vm_page_zero_count;
1147 	} else {
1148 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1149 	}
1150 	mtx_unlock_spin(&vm_page_queue_free_mtx);
1151 	vm_page_free_wakeup();
1152 }
1153 
1154 /*
1155  *	vm_page_unmanage:
1156  *
1157  * 	Prevent PV management from being done on the page.  The page is
1158  *	removed from the paging queues as if it were wired, and as a
1159  *	consequence of no longer being managed the pageout daemon will not
1160  *	touch it (since there is no way to locate the pte mappings for the
1161  *	page).  madvise() calls that mess with the pmap will also no longer
1162  *	operate on the page.
1163  *
1164  *	Beyond that the page is still reasonably 'normal'.  Freeing the page
1165  *	will clear the flag.
1166  *
1167  *	This routine is used by OBJT_PHYS objects - objects using unswappable
1168  *	physical memory as backing store rather then swap-backed memory and
1169  *	will eventually be extended to support 4MB unmanaged physical
1170  *	mappings.
1171  */
1172 void
1173 vm_page_unmanage(vm_page_t m)
1174 {
1175 
1176 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1177 	if ((m->flags & PG_UNMANAGED) == 0) {
1178 		if (m->wire_count == 0)
1179 			vm_pageq_remove(m);
1180 	}
1181 	vm_page_flag_set(m, PG_UNMANAGED);
1182 }
1183 
1184 /*
1185  *	vm_page_wire:
1186  *
1187  *	Mark this page as wired down by yet
1188  *	another map, removing it from paging queues
1189  *	as necessary.
1190  *
1191  *	The page queues must be locked.
1192  *	This routine may not block.
1193  */
1194 void
1195 vm_page_wire(vm_page_t m)
1196 {
1197 
1198 	/*
1199 	 * Only bump the wire statistics if the page is not already wired,
1200 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1201 	 * it is already off the queues).
1202 	 */
1203 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1204 	if (m->flags & PG_FICTITIOUS)
1205 		return;
1206 	if (m->wire_count == 0) {
1207 		if ((m->flags & PG_UNMANAGED) == 0)
1208 			vm_pageq_remove(m);
1209 		atomic_add_int(&cnt.v_wire_count, 1);
1210 	}
1211 	m->wire_count++;
1212 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1213 }
1214 
1215 /*
1216  *	vm_page_unwire:
1217  *
1218  *	Release one wiring of this page, potentially
1219  *	enabling it to be paged again.
1220  *
1221  *	Many pages placed on the inactive queue should actually go
1222  *	into the cache, but it is difficult to figure out which.  What
1223  *	we do instead, if the inactive target is well met, is to put
1224  *	clean pages at the head of the inactive queue instead of the tail.
1225  *	This will cause them to be moved to the cache more quickly and
1226  *	if not actively re-referenced, freed more quickly.  If we just
1227  *	stick these pages at the end of the inactive queue, heavy filesystem
1228  *	meta-data accesses can cause an unnecessary paging load on memory bound
1229  *	processes.  This optimization causes one-time-use metadata to be
1230  *	reused more quickly.
1231  *
1232  *	BUT, if we are in a low-memory situation we have no choice but to
1233  *	put clean pages on the cache queue.
1234  *
1235  *	A number of routines use vm_page_unwire() to guarantee that the page
1236  *	will go into either the inactive or active queues, and will NEVER
1237  *	be placed in the cache - for example, just after dirtying a page.
1238  *	dirty pages in the cache are not allowed.
1239  *
1240  *	The page queues must be locked.
1241  *	This routine may not block.
1242  */
1243 void
1244 vm_page_unwire(vm_page_t m, int activate)
1245 {
1246 
1247 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1248 	if (m->flags & PG_FICTITIOUS)
1249 		return;
1250 	if (m->wire_count > 0) {
1251 		m->wire_count--;
1252 		if (m->wire_count == 0) {
1253 			atomic_subtract_int(&cnt.v_wire_count, 1);
1254 			if (m->flags & PG_UNMANAGED) {
1255 				;
1256 			} else if (activate)
1257 				vm_pageq_enqueue(PQ_ACTIVE, m);
1258 			else {
1259 				vm_page_flag_clear(m, PG_WINATCFLS);
1260 				vm_pageq_enqueue(PQ_INACTIVE, m);
1261 			}
1262 		}
1263 	} else {
1264 		panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1265 	}
1266 }
1267 
1268 
1269 /*
1270  * Move the specified page to the inactive queue.  If the page has
1271  * any associated swap, the swap is deallocated.
1272  *
1273  * Normally athead is 0 resulting in LRU operation.  athead is set
1274  * to 1 if we want this page to be 'as if it were placed in the cache',
1275  * except without unmapping it from the process address space.
1276  *
1277  * This routine may not block.
1278  */
1279 static inline void
1280 _vm_page_deactivate(vm_page_t m, int athead)
1281 {
1282 
1283 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1284 
1285 	/*
1286 	 * Ignore if already inactive.
1287 	 */
1288 	if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE))
1289 		return;
1290 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1291 		if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1292 			cnt.v_reactivated++;
1293 		vm_page_flag_clear(m, PG_WINATCFLS);
1294 		vm_pageq_remove(m);
1295 		if (athead)
1296 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1297 		else
1298 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1299 		VM_PAGE_SETQUEUE2(m, PQ_INACTIVE);
1300 		vm_page_queues[PQ_INACTIVE].lcnt++;
1301 		cnt.v_inactive_count++;
1302 	}
1303 }
1304 
1305 void
1306 vm_page_deactivate(vm_page_t m)
1307 {
1308     _vm_page_deactivate(m, 0);
1309 }
1310 
1311 /*
1312  * vm_page_try_to_cache:
1313  *
1314  * Returns 0 on failure, 1 on success
1315  */
1316 int
1317 vm_page_try_to_cache(vm_page_t m)
1318 {
1319 
1320 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1321 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1322 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1323 	    (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) {
1324 		return (0);
1325 	}
1326 	pmap_remove_all(m);
1327 	if (m->dirty)
1328 		return (0);
1329 	vm_page_cache(m);
1330 	return (1);
1331 }
1332 
1333 /*
1334  * vm_page_try_to_free()
1335  *
1336  *	Attempt to free the page.  If we cannot free it, we do nothing.
1337  *	1 is returned on success, 0 on failure.
1338  */
1339 int
1340 vm_page_try_to_free(vm_page_t m)
1341 {
1342 
1343 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1344 	if (m->object != NULL)
1345 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1346 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1347 	    (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) {
1348 		return (0);
1349 	}
1350 	pmap_remove_all(m);
1351 	if (m->dirty)
1352 		return (0);
1353 	vm_page_free(m);
1354 	return (1);
1355 }
1356 
1357 /*
1358  * vm_page_cache
1359  *
1360  * Put the specified page onto the page cache queue (if appropriate).
1361  *
1362  * This routine may not block.
1363  */
1364 void
1365 vm_page_cache(vm_page_t m)
1366 {
1367 
1368 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1369 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1370 	if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy ||
1371 	    m->hold_count || m->wire_count) {
1372 		printf("vm_page_cache: attempting to cache busy page\n");
1373 		return;
1374 	}
1375 	if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1376 		return;
1377 
1378 	/*
1379 	 * Remove all pmaps and indicate that the page is not
1380 	 * writeable or mapped.
1381 	 */
1382 	pmap_remove_all(m);
1383 	if (m->dirty != 0) {
1384 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1385 			(long)m->pindex);
1386 	}
1387 	vm_pageq_remove_nowakeup(m);
1388 	vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1389 	vm_page_free_wakeup();
1390 }
1391 
1392 /*
1393  * vm_page_dontneed
1394  *
1395  *	Cache, deactivate, or do nothing as appropriate.  This routine
1396  *	is typically used by madvise() MADV_DONTNEED.
1397  *
1398  *	Generally speaking we want to move the page into the cache so
1399  *	it gets reused quickly.  However, this can result in a silly syndrome
1400  *	due to the page recycling too quickly.  Small objects will not be
1401  *	fully cached.  On the otherhand, if we move the page to the inactive
1402  *	queue we wind up with a problem whereby very large objects
1403  *	unnecessarily blow away our inactive and cache queues.
1404  *
1405  *	The solution is to move the pages based on a fixed weighting.  We
1406  *	either leave them alone, deactivate them, or move them to the cache,
1407  *	where moving them to the cache has the highest weighting.
1408  *	By forcing some pages into other queues we eventually force the
1409  *	system to balance the queues, potentially recovering other unrelated
1410  *	space from active.  The idea is to not force this to happen too
1411  *	often.
1412  */
1413 void
1414 vm_page_dontneed(vm_page_t m)
1415 {
1416 	static int dnweight;
1417 	int dnw;
1418 	int head;
1419 
1420 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1421 	dnw = ++dnweight;
1422 
1423 	/*
1424 	 * occassionally leave the page alone
1425 	 */
1426 	if ((dnw & 0x01F0) == 0 ||
1427 	    VM_PAGE_INQUEUE2(m, PQ_INACTIVE) ||
1428 	    VM_PAGE_INQUEUE1(m, PQ_CACHE)
1429 	) {
1430 		if (m->act_count >= ACT_INIT)
1431 			--m->act_count;
1432 		return;
1433 	}
1434 
1435 	if (m->dirty == 0 && pmap_is_modified(m))
1436 		vm_page_dirty(m);
1437 
1438 	if (m->dirty || (dnw & 0x0070) == 0) {
1439 		/*
1440 		 * Deactivate the page 3 times out of 32.
1441 		 */
1442 		head = 0;
1443 	} else {
1444 		/*
1445 		 * Cache the page 28 times out of every 32.  Note that
1446 		 * the page is deactivated instead of cached, but placed
1447 		 * at the head of the queue instead of the tail.
1448 		 */
1449 		head = 1;
1450 	}
1451 	_vm_page_deactivate(m, head);
1452 }
1453 
1454 /*
1455  * Grab a page, waiting until we are waken up due to the page
1456  * changing state.  We keep on waiting, if the page continues
1457  * to be in the object.  If the page doesn't exist, first allocate it
1458  * and then conditionally zero it.
1459  *
1460  * This routine may block.
1461  */
1462 vm_page_t
1463 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1464 {
1465 	vm_page_t m;
1466 
1467 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1468 retrylookup:
1469 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1470 		if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) {
1471 			if ((allocflags & VM_ALLOC_RETRY) == 0)
1472 				return (NULL);
1473 			goto retrylookup;
1474 		} else {
1475 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
1476 				vm_page_lock_queues();
1477 				vm_page_wire(m);
1478 				vm_page_unlock_queues();
1479 			}
1480 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
1481 				vm_page_busy(m);
1482 			return (m);
1483 		}
1484 	}
1485 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1486 	if (m == NULL) {
1487 		VM_OBJECT_UNLOCK(object);
1488 		VM_WAIT;
1489 		VM_OBJECT_LOCK(object);
1490 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1491 			return (NULL);
1492 		goto retrylookup;
1493 	}
1494 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
1495 		pmap_zero_page(m);
1496 	return (m);
1497 }
1498 
1499 /*
1500  * Mapping function for valid bits or for dirty bits in
1501  * a page.  May not block.
1502  *
1503  * Inputs are required to range within a page.
1504  */
1505 inline int
1506 vm_page_bits(int base, int size)
1507 {
1508 	int first_bit;
1509 	int last_bit;
1510 
1511 	KASSERT(
1512 	    base + size <= PAGE_SIZE,
1513 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1514 	);
1515 
1516 	if (size == 0)		/* handle degenerate case */
1517 		return (0);
1518 
1519 	first_bit = base >> DEV_BSHIFT;
1520 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1521 
1522 	return ((2 << last_bit) - (1 << first_bit));
1523 }
1524 
1525 /*
1526  *	vm_page_set_validclean:
1527  *
1528  *	Sets portions of a page valid and clean.  The arguments are expected
1529  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1530  *	of any partial chunks touched by the range.  The invalid portion of
1531  *	such chunks will be zero'd.
1532  *
1533  *	This routine may not block.
1534  *
1535  *	(base + size) must be less then or equal to PAGE_SIZE.
1536  */
1537 void
1538 vm_page_set_validclean(vm_page_t m, int base, int size)
1539 {
1540 	int pagebits;
1541 	int frag;
1542 	int endoff;
1543 
1544 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1545 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1546 	if (size == 0)	/* handle degenerate case */
1547 		return;
1548 
1549 	/*
1550 	 * If the base is not DEV_BSIZE aligned and the valid
1551 	 * bit is clear, we have to zero out a portion of the
1552 	 * first block.
1553 	 */
1554 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1555 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1556 		pmap_zero_page_area(m, frag, base - frag);
1557 
1558 	/*
1559 	 * If the ending offset is not DEV_BSIZE aligned and the
1560 	 * valid bit is clear, we have to zero out a portion of
1561 	 * the last block.
1562 	 */
1563 	endoff = base + size;
1564 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1565 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1566 		pmap_zero_page_area(m, endoff,
1567 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1568 
1569 	/*
1570 	 * Set valid, clear dirty bits.  If validating the entire
1571 	 * page we can safely clear the pmap modify bit.  We also
1572 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
1573 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1574 	 * be set again.
1575 	 *
1576 	 * We set valid bits inclusive of any overlap, but we can only
1577 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1578 	 * the range.
1579 	 */
1580 	pagebits = vm_page_bits(base, size);
1581 	m->valid |= pagebits;
1582 #if 0	/* NOT YET */
1583 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1584 		frag = DEV_BSIZE - frag;
1585 		base += frag;
1586 		size -= frag;
1587 		if (size < 0)
1588 			size = 0;
1589 	}
1590 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1591 #endif
1592 	m->dirty &= ~pagebits;
1593 	if (base == 0 && size == PAGE_SIZE) {
1594 		pmap_clear_modify(m);
1595 		m->oflags &= ~VPO_NOSYNC;
1596 	}
1597 }
1598 
1599 void
1600 vm_page_clear_dirty(vm_page_t m, int base, int size)
1601 {
1602 
1603 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1604 	m->dirty &= ~vm_page_bits(base, size);
1605 }
1606 
1607 /*
1608  *	vm_page_set_invalid:
1609  *
1610  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1611  *	valid and dirty bits for the effected areas are cleared.
1612  *
1613  *	May not block.
1614  */
1615 void
1616 vm_page_set_invalid(vm_page_t m, int base, int size)
1617 {
1618 	int bits;
1619 
1620 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1621 	bits = vm_page_bits(base, size);
1622 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1623 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
1624 		pmap_remove_all(m);
1625 	m->valid &= ~bits;
1626 	m->dirty &= ~bits;
1627 	m->object->generation++;
1628 }
1629 
1630 /*
1631  * vm_page_zero_invalid()
1632  *
1633  *	The kernel assumes that the invalid portions of a page contain
1634  *	garbage, but such pages can be mapped into memory by user code.
1635  *	When this occurs, we must zero out the non-valid portions of the
1636  *	page so user code sees what it expects.
1637  *
1638  *	Pages are most often semi-valid when the end of a file is mapped
1639  *	into memory and the file's size is not page aligned.
1640  */
1641 void
1642 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1643 {
1644 	int b;
1645 	int i;
1646 
1647 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1648 	/*
1649 	 * Scan the valid bits looking for invalid sections that
1650 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1651 	 * valid bit may be set ) have already been zerod by
1652 	 * vm_page_set_validclean().
1653 	 */
1654 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1655 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1656 		    (m->valid & (1 << i))
1657 		) {
1658 			if (i > b) {
1659 				pmap_zero_page_area(m,
1660 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1661 			}
1662 			b = i + 1;
1663 		}
1664 	}
1665 
1666 	/*
1667 	 * setvalid is TRUE when we can safely set the zero'd areas
1668 	 * as being valid.  We can do this if there are no cache consistancy
1669 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1670 	 */
1671 	if (setvalid)
1672 		m->valid = VM_PAGE_BITS_ALL;
1673 }
1674 
1675 /*
1676  *	vm_page_is_valid:
1677  *
1678  *	Is (partial) page valid?  Note that the case where size == 0
1679  *	will return FALSE in the degenerate case where the page is
1680  *	entirely invalid, and TRUE otherwise.
1681  *
1682  *	May not block.
1683  */
1684 int
1685 vm_page_is_valid(vm_page_t m, int base, int size)
1686 {
1687 	int bits = vm_page_bits(base, size);
1688 
1689 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1690 	if (m->valid && ((m->valid & bits) == bits))
1691 		return 1;
1692 	else
1693 		return 0;
1694 }
1695 
1696 /*
1697  * update dirty bits from pmap/mmu.  May not block.
1698  */
1699 void
1700 vm_page_test_dirty(vm_page_t m)
1701 {
1702 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1703 		vm_page_dirty(m);
1704 	}
1705 }
1706 
1707 int so_zerocp_fullpage = 0;
1708 
1709 void
1710 vm_page_cowfault(vm_page_t m)
1711 {
1712 	vm_page_t mnew;
1713 	vm_object_t object;
1714 	vm_pindex_t pindex;
1715 
1716 	object = m->object;
1717 	pindex = m->pindex;
1718 
1719  retry_alloc:
1720 	pmap_remove_all(m);
1721 	vm_page_remove(m);
1722 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
1723 	if (mnew == NULL) {
1724 		vm_page_insert(m, object, pindex);
1725 		vm_page_unlock_queues();
1726 		VM_OBJECT_UNLOCK(object);
1727 		VM_WAIT;
1728 		VM_OBJECT_LOCK(object);
1729 		vm_page_lock_queues();
1730 		goto retry_alloc;
1731 	}
1732 
1733 	if (m->cow == 0) {
1734 		/*
1735 		 * check to see if we raced with an xmit complete when
1736 		 * waiting to allocate a page.  If so, put things back
1737 		 * the way they were
1738 		 */
1739 		vm_page_free(mnew);
1740 		vm_page_insert(m, object, pindex);
1741 	} else { /* clear COW & copy page */
1742 		if (!so_zerocp_fullpage)
1743 			pmap_copy_page(m, mnew);
1744 		mnew->valid = VM_PAGE_BITS_ALL;
1745 		vm_page_dirty(mnew);
1746 		mnew->wire_count = m->wire_count - m->cow;
1747 		m->wire_count = m->cow;
1748 	}
1749 }
1750 
1751 void
1752 vm_page_cowclear(vm_page_t m)
1753 {
1754 
1755 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1756 	if (m->cow) {
1757 		m->cow--;
1758 		/*
1759 		 * let vm_fault add back write permission  lazily
1760 		 */
1761 	}
1762 	/*
1763 	 *  sf_buf_free() will free the page, so we needn't do it here
1764 	 */
1765 }
1766 
1767 void
1768 vm_page_cowsetup(vm_page_t m)
1769 {
1770 
1771 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1772 	m->cow++;
1773 	pmap_remove_write(m);
1774 }
1775 
1776 #include "opt_ddb.h"
1777 #ifdef DDB
1778 #include <sys/kernel.h>
1779 
1780 #include <ddb/ddb.h>
1781 
1782 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1783 {
1784 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1785 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1786 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1787 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1788 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1789 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1790 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1791 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1792 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1793 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1794 }
1795 
1796 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1797 {
1798 	int i;
1799 	db_printf("PQ_FREE:");
1800 	for (i = 0; i < PQ_NUMCOLORS; i++) {
1801 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1802 	}
1803 	db_printf("\n");
1804 
1805 	db_printf("PQ_CACHE:");
1806 	for (i = 0; i < PQ_NUMCOLORS; i++) {
1807 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1808 	}
1809 	db_printf("\n");
1810 
1811 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1812 		vm_page_queues[PQ_ACTIVE].lcnt,
1813 		vm_page_queues[PQ_INACTIVE].lcnt);
1814 }
1815 #endif /* DDB */
1816