xref: /freebsd/sys/vm/vm_page.c (revision a4eb85b6acb49cb60c72c2cab0d0d3f00eaa6d46)
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->flags & PG_BUSY) == 0,
376 	    ("vm_page_busy: page already busy!!!"));
377 	vm_page_flag_set(m, PG_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->flags & PG_WANTED) {
391 		vm_page_flag_clear(m, PG_WANTED);
392 		wakeup(m);
393 	}
394 }
395 
396 /*
397  *      vm_page_wakeup:
398  *
399  *      clear the PG_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->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
409 	vm_page_flag_clear(m, PG_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 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
427 	m->busy--;
428 	if (m->busy == 0)
429 		vm_page_flash(m);
430 }
431 
432 /*
433  * Keep page from being freed by the page daemon
434  * much of the same effect as wiring, except much lower
435  * overhead and should be used only for *very* temporary
436  * holding ("wiring").
437  */
438 void
439 vm_page_hold(vm_page_t mem)
440 {
441 
442 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
443         mem->hold_count++;
444 }
445 
446 void
447 vm_page_unhold(vm_page_t mem)
448 {
449 
450 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
451 	--mem->hold_count;
452 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
453 	if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD))
454 		vm_page_free_toq(mem);
455 }
456 
457 /*
458  *	vm_page_free:
459  *
460  *	Free a page
461  *
462  *	The clearing of PG_ZERO is a temporary safety until the code can be
463  *	reviewed to determine that PG_ZERO is being properly cleared on
464  *	write faults or maps.  PG_ZERO was previously cleared in
465  *	vm_page_alloc().
466  */
467 void
468 vm_page_free(vm_page_t m)
469 {
470 	vm_page_flag_clear(m, PG_ZERO);
471 	vm_page_free_toq(m);
472 	vm_page_zero_idle_wakeup();
473 }
474 
475 /*
476  *	vm_page_free_zero:
477  *
478  *	Free a page to the zerod-pages queue
479  */
480 void
481 vm_page_free_zero(vm_page_t m)
482 {
483 	vm_page_flag_set(m, PG_ZERO);
484 	vm_page_free_toq(m);
485 }
486 
487 /*
488  *	vm_page_sleep_if_busy:
489  *
490  *	Sleep and release the page queues lock if PG_BUSY is set or,
491  *	if also_m_busy is TRUE, busy is non-zero.  Returns TRUE if the
492  *	thread slept and the page queues lock was released.
493  *	Otherwise, retains the page queues lock and returns FALSE.
494  */
495 int
496 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
497 {
498 	vm_object_t object;
499 
500 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
501 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
502 	if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
503 		vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
504 		/*
505 		 * It's possible that while we sleep, the page will get
506 		 * unbusied and freed.  If we are holding the object
507 		 * lock, we will assume we hold a reference to the object
508 		 * such that even if m->object changes, we can re-lock
509 		 * it.
510 		 */
511 		object = m->object;
512 		VM_OBJECT_UNLOCK(object);
513 		msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
514 		VM_OBJECT_LOCK(object);
515 		return (TRUE);
516 	}
517 	return (FALSE);
518 }
519 
520 /*
521  *	vm_page_dirty:
522  *
523  *	make page all dirty
524  */
525 void
526 vm_page_dirty(vm_page_t m)
527 {
528 	KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_CACHE,
529 	    ("vm_page_dirty: page in cache!"));
530 	KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_FREE,
531 	    ("vm_page_dirty: page is free!"));
532 	m->dirty = VM_PAGE_BITS_ALL;
533 }
534 
535 /*
536  *	vm_page_splay:
537  *
538  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
539  *	the vm_page containing the given pindex.  If, however, that
540  *	pindex is not found in the vm_object, returns a vm_page that is
541  *	adjacent to the pindex, coming before or after it.
542  */
543 vm_page_t
544 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
545 {
546 	struct vm_page dummy;
547 	vm_page_t lefttreemax, righttreemin, y;
548 
549 	if (root == NULL)
550 		return (root);
551 	lefttreemax = righttreemin = &dummy;
552 	for (;; root = y) {
553 		if (pindex < root->pindex) {
554 			if ((y = root->left) == NULL)
555 				break;
556 			if (pindex < y->pindex) {
557 				/* Rotate right. */
558 				root->left = y->right;
559 				y->right = root;
560 				root = y;
561 				if ((y = root->left) == NULL)
562 					break;
563 			}
564 			/* Link into the new root's right tree. */
565 			righttreemin->left = root;
566 			righttreemin = root;
567 		} else if (pindex > root->pindex) {
568 			if ((y = root->right) == NULL)
569 				break;
570 			if (pindex > y->pindex) {
571 				/* Rotate left. */
572 				root->right = y->left;
573 				y->left = root;
574 				root = y;
575 				if ((y = root->right) == NULL)
576 					break;
577 			}
578 			/* Link into the new root's left tree. */
579 			lefttreemax->right = root;
580 			lefttreemax = root;
581 		} else
582 			break;
583 	}
584 	/* Assemble the new root. */
585 	lefttreemax->right = root->left;
586 	righttreemin->left = root->right;
587 	root->left = dummy.right;
588 	root->right = dummy.left;
589 	return (root);
590 }
591 
592 /*
593  *	vm_page_insert:		[ internal use only ]
594  *
595  *	Inserts the given mem entry into the object and object list.
596  *
597  *	The pagetables are not updated but will presumably fault the page
598  *	in if necessary, or if a kernel page the caller will at some point
599  *	enter the page into the kernel's pmap.  We are not allowed to block
600  *	here so we *can't* do this anyway.
601  *
602  *	The object and page must be locked.
603  *	This routine may not block.
604  */
605 void
606 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
607 {
608 	vm_page_t root;
609 
610 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
611 	if (m->object != NULL)
612 		panic("vm_page_insert: page already inserted");
613 
614 	/*
615 	 * Record the object/offset pair in this page
616 	 */
617 	m->object = object;
618 	m->pindex = pindex;
619 
620 	/*
621 	 * Now link into the object's ordered list of backed pages.
622 	 */
623 	root = object->root;
624 	if (root == NULL) {
625 		m->left = NULL;
626 		m->right = NULL;
627 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
628 	} else {
629 		root = vm_page_splay(pindex, root);
630 		if (pindex < root->pindex) {
631 			m->left = root->left;
632 			m->right = root;
633 			root->left = NULL;
634 			TAILQ_INSERT_BEFORE(root, m, listq);
635 		} else if (pindex == root->pindex)
636 			panic("vm_page_insert: offset already allocated");
637 		else {
638 			m->right = root->right;
639 			m->left = root;
640 			root->right = NULL;
641 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
642 		}
643 	}
644 	object->root = m;
645 	object->generation++;
646 
647 	/*
648 	 * show that the object has one more resident page.
649 	 */
650 	object->resident_page_count++;
651 	/*
652 	 * Hold the vnode until the last page is released.
653 	 */
654 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
655 		vhold((struct vnode *)object->handle);
656 
657 	/*
658 	 * Since we are inserting a new and possibly dirty page,
659 	 * update the object's OBJ_MIGHTBEDIRTY flag.
660 	 */
661 	if (m->flags & PG_WRITEABLE)
662 		vm_object_set_writeable_dirty(object);
663 }
664 
665 /*
666  *	vm_page_remove:
667  *				NOTE: used by device pager as well -wfj
668  *
669  *	Removes the given mem entry from the object/offset-page
670  *	table and the object page list, but do not invalidate/terminate
671  *	the backing store.
672  *
673  *	The object and page must be locked.
674  *	The underlying pmap entry (if any) is NOT removed here.
675  *	This routine may not block.
676  */
677 void
678 vm_page_remove(vm_page_t m)
679 {
680 	vm_object_t object;
681 	vm_page_t root;
682 
683 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
684 	if ((object = m->object) == NULL)
685 		return;
686 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
687 	if (m->flags & PG_BUSY) {
688 		vm_page_flag_clear(m, PG_BUSY);
689 		vm_page_flash(m);
690 	}
691 
692 	/*
693 	 * Now remove from the object's list of backed pages.
694 	 */
695 	if (m != object->root)
696 		vm_page_splay(m->pindex, object->root);
697 	if (m->left == NULL)
698 		root = m->right;
699 	else {
700 		root = vm_page_splay(m->pindex, m->left);
701 		root->right = m->right;
702 	}
703 	object->root = root;
704 	TAILQ_REMOVE(&object->memq, m, listq);
705 
706 	/*
707 	 * And show that the object has one fewer resident page.
708 	 */
709 	object->resident_page_count--;
710 	object->generation++;
711 	/*
712 	 * The vnode may now be recycled.
713 	 */
714 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
715 		vdrop((struct vnode *)object->handle);
716 
717 	m->object = NULL;
718 }
719 
720 /*
721  *	vm_page_lookup:
722  *
723  *	Returns the page associated with the object/offset
724  *	pair specified; if none is found, NULL is returned.
725  *
726  *	The object must be locked.
727  *	This routine may not block.
728  *	This is a critical path routine
729  */
730 vm_page_t
731 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
732 {
733 	vm_page_t m;
734 
735 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
736 	if ((m = object->root) != NULL && m->pindex != pindex) {
737 		m = vm_page_splay(pindex, m);
738 		if ((object->root = m)->pindex != pindex)
739 			m = NULL;
740 	}
741 	return (m);
742 }
743 
744 /*
745  *	vm_page_rename:
746  *
747  *	Move the given memory entry from its
748  *	current object to the specified target object/offset.
749  *
750  *	The object must be locked.
751  *	This routine may not block.
752  *
753  *	Note: swap associated with the page must be invalidated by the move.  We
754  *	      have to do this for several reasons:  (1) we aren't freeing the
755  *	      page, (2) we are dirtying the page, (3) the VM system is probably
756  *	      moving the page from object A to B, and will then later move
757  *	      the backing store from A to B and we can't have a conflict.
758  *
759  *	Note: we *always* dirty the page.  It is necessary both for the
760  *	      fact that we moved it, and because we may be invalidating
761  *	      swap.  If the page is on the cache, we have to deactivate it
762  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
763  *	      on the cache.
764  */
765 void
766 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
767 {
768 
769 	vm_page_remove(m);
770 	vm_page_insert(m, new_object, new_pindex);
771 	if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
772 		vm_page_deactivate(m);
773 	vm_page_dirty(m);
774 }
775 
776 /*
777  *	vm_page_select_cache:
778  *
779  *	Move a page of the given color from the cache queue to the free
780  *	queue.  As pages might be found, but are not applicable, they are
781  *	deactivated.
782  *
783  *	This routine may not block.
784  */
785 vm_page_t
786 vm_page_select_cache(int color)
787 {
788 	vm_object_t object;
789 	vm_page_t m;
790 	boolean_t was_trylocked;
791 
792 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
793 	while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) {
794 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
795 		KASSERT(!pmap_page_is_mapped(m),
796 		    ("Found mapped cache page %p", m));
797 		KASSERT((m->flags & PG_UNMANAGED) == 0,
798 		    ("Found unmanaged cache page %p", m));
799 		KASSERT(m->wire_count == 0, ("Found wired cache page %p", m));
800 		if (m->hold_count == 0 && (object = m->object,
801 		    (was_trylocked = VM_OBJECT_TRYLOCK(object)) ||
802 		    VM_OBJECT_LOCKED(object))) {
803 			KASSERT((m->flags & PG_BUSY) == 0 && m->busy == 0,
804 			    ("Found busy cache page %p", m));
805 			vm_page_free(m);
806 			if (was_trylocked)
807 				VM_OBJECT_UNLOCK(object);
808 			break;
809 		}
810 		vm_page_deactivate(m);
811 	}
812 	return (m);
813 }
814 
815 /*
816  *	vm_page_alloc:
817  *
818  *	Allocate and return a memory cell associated
819  *	with this VM object/offset pair.
820  *
821  *	page_req classes:
822  *	VM_ALLOC_NORMAL		normal process request
823  *	VM_ALLOC_SYSTEM		system *really* needs a page
824  *	VM_ALLOC_INTERRUPT	interrupt time request
825  *	VM_ALLOC_ZERO		zero page
826  *
827  *	This routine may not block.
828  *
829  *	Additional special handling is required when called from an
830  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
831  *	the page cache in this case.
832  */
833 vm_page_t
834 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
835 {
836 	vm_page_t m = NULL;
837 	int color, flags, page_req;
838 
839 	page_req = req & VM_ALLOC_CLASS_MASK;
840 	KASSERT(curthread->td_intr_nesting_level == 0 ||
841 	    page_req == VM_ALLOC_INTERRUPT,
842 	    ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context"));
843 
844 	if ((req & VM_ALLOC_NOOBJ) == 0) {
845 		KASSERT(object != NULL,
846 		    ("vm_page_alloc: NULL object."));
847 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
848 		color = (pindex + object->pg_color) & PQ_COLORMASK;
849 	} else
850 		color = pindex & PQ_COLORMASK;
851 
852 	/*
853 	 * The pager is allowed to eat deeper into the free page list.
854 	 */
855 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
856 		page_req = VM_ALLOC_SYSTEM;
857 	};
858 
859 loop:
860 	mtx_lock_spin(&vm_page_queue_free_mtx);
861 	if (cnt.v_free_count > cnt.v_free_reserved ||
862 	    (page_req == VM_ALLOC_SYSTEM &&
863 	     cnt.v_cache_count == 0 &&
864 	     cnt.v_free_count > cnt.v_interrupt_free_min) ||
865 	    (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) {
866 		/*
867 		 * Allocate from the free queue if the number of free pages
868 		 * exceeds the minimum for the request class.
869 		 */
870 		m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0);
871 	} else if (page_req != VM_ALLOC_INTERRUPT) {
872 		mtx_unlock_spin(&vm_page_queue_free_mtx);
873 		/*
874 		 * Allocatable from cache (non-interrupt only).  On success,
875 		 * we must free the page and try again, thus ensuring that
876 		 * cnt.v_*_free_min counters are replenished.
877 		 */
878 		vm_page_lock_queues();
879 		if ((m = vm_page_select_cache(color)) == NULL) {
880 			KASSERT(cnt.v_cache_count == 0,
881 			    ("vm_page_alloc: cache queue is missing %d pages",
882 			    cnt.v_cache_count));
883 			vm_page_unlock_queues();
884 			atomic_add_int(&vm_pageout_deficit, 1);
885 			pagedaemon_wakeup();
886 
887 			if (page_req != VM_ALLOC_SYSTEM)
888 				return NULL;
889 
890 			mtx_lock_spin(&vm_page_queue_free_mtx);
891 			if (cnt.v_free_count <=  cnt.v_interrupt_free_min) {
892 				mtx_unlock_spin(&vm_page_queue_free_mtx);
893 				return (NULL);
894 			}
895 			m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0);
896 		} else {
897 			vm_page_unlock_queues();
898 			goto loop;
899 		}
900 	} else {
901 		/*
902 		 * Not allocatable from cache from interrupt, give up.
903 		 */
904 		mtx_unlock_spin(&vm_page_queue_free_mtx);
905 		atomic_add_int(&vm_pageout_deficit, 1);
906 		pagedaemon_wakeup();
907 		return (NULL);
908 	}
909 
910 	/*
911 	 *  At this point we had better have found a good page.
912 	 */
913 
914 	KASSERT(
915 	    m != NULL,
916 	    ("vm_page_alloc(): missing page on free queue")
917 	);
918 
919 	/*
920 	 * Remove from free queue
921 	 */
922 	vm_pageq_remove_nowakeup(m);
923 
924 	/*
925 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
926 	 */
927 	flags = PG_BUSY;
928 	if (m->flags & PG_ZERO) {
929 		vm_page_zero_count--;
930 		if (req & VM_ALLOC_ZERO)
931 			flags = PG_ZERO | PG_BUSY;
932 	}
933 	if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ))
934 		flags &= ~PG_BUSY;
935 	m->flags = flags;
936 	if (req & VM_ALLOC_WIRED) {
937 		atomic_add_int(&cnt.v_wire_count, 1);
938 		m->wire_count = 1;
939 	} else
940 		m->wire_count = 0;
941 	m->hold_count = 0;
942 	m->act_count = 0;
943 	m->busy = 0;
944 	m->valid = 0;
945 	KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
946 	mtx_unlock_spin(&vm_page_queue_free_mtx);
947 
948 	if ((req & VM_ALLOC_NOOBJ) == 0)
949 		vm_page_insert(m, object, pindex);
950 	else
951 		m->pindex = pindex;
952 
953 	/*
954 	 * Don't wakeup too often - wakeup the pageout daemon when
955 	 * we would be nearly out of memory.
956 	 */
957 	if (vm_paging_needed())
958 		pagedaemon_wakeup();
959 
960 	return (m);
961 }
962 
963 /*
964  *	vm_wait:	(also see VM_WAIT macro)
965  *
966  *	Block until free pages are available for allocation
967  *	- Called in various places before memory allocations.
968  */
969 void
970 vm_wait(void)
971 {
972 
973 	vm_page_lock_queues();
974 	if (curproc == pageproc) {
975 		vm_pageout_pages_needed = 1;
976 		msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx,
977 		    PDROP | PSWP, "VMWait", 0);
978 	} else {
979 		if (!vm_pages_needed) {
980 			vm_pages_needed = 1;
981 			wakeup(&vm_pages_needed);
982 		}
983 		msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM,
984 		    "vmwait", 0);
985 	}
986 }
987 
988 /*
989  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
990  *
991  *	Block until free pages are available for allocation
992  *	- Called only in vm_fault so that processes page faulting
993  *	  can be easily tracked.
994  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
995  *	  processes will be able to grab memory first.  Do not change
996  *	  this balance without careful testing first.
997  */
998 void
999 vm_waitpfault(void)
1000 {
1001 
1002 	vm_page_lock_queues();
1003 	if (!vm_pages_needed) {
1004 		vm_pages_needed = 1;
1005 		wakeup(&vm_pages_needed);
1006 	}
1007 	msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER,
1008 	    "pfault", 0);
1009 }
1010 
1011 /*
1012  *	vm_page_activate:
1013  *
1014  *	Put the specified page on the active list (if appropriate).
1015  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1016  *	mess with it.
1017  *
1018  *	The page queues must be locked.
1019  *	This routine may not block.
1020  */
1021 void
1022 vm_page_activate(vm_page_t m)
1023 {
1024 
1025 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1026 	if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) {
1027 		if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1028 			cnt.v_reactivated++;
1029 		vm_pageq_remove(m);
1030 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1031 			if (m->act_count < ACT_INIT)
1032 				m->act_count = ACT_INIT;
1033 			vm_pageq_enqueue(PQ_ACTIVE, m);
1034 		}
1035 	} else {
1036 		if (m->act_count < ACT_INIT)
1037 			m->act_count = ACT_INIT;
1038 	}
1039 }
1040 
1041 /*
1042  *	vm_page_free_wakeup:
1043  *
1044  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1045  *	routine is called when a page has been added to the cache or free
1046  *	queues.
1047  *
1048  *	The page queues must be locked.
1049  *	This routine may not block.
1050  */
1051 static inline void
1052 vm_page_free_wakeup(void)
1053 {
1054 
1055 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1056 	/*
1057 	 * if pageout daemon needs pages, then tell it that there are
1058 	 * some free.
1059 	 */
1060 	if (vm_pageout_pages_needed &&
1061 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1062 		wakeup(&vm_pageout_pages_needed);
1063 		vm_pageout_pages_needed = 0;
1064 	}
1065 	/*
1066 	 * wakeup processes that are waiting on memory if we hit a
1067 	 * high water mark. And wakeup scheduler process if we have
1068 	 * lots of memory. this process will swapin processes.
1069 	 */
1070 	if (vm_pages_needed && !vm_page_count_min()) {
1071 		vm_pages_needed = 0;
1072 		wakeup(&cnt.v_free_count);
1073 	}
1074 }
1075 
1076 /*
1077  *	vm_page_free_toq:
1078  *
1079  *	Returns the given page to the PQ_FREE list,
1080  *	disassociating it with any VM object.
1081  *
1082  *	Object and page must be locked prior to entry.
1083  *	This routine may not block.
1084  */
1085 
1086 void
1087 vm_page_free_toq(vm_page_t m)
1088 {
1089 	struct vpgqueues *pq;
1090 
1091 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1092 	KASSERT(!pmap_page_is_mapped(m),
1093 	    ("vm_page_free_toq: freeing mapped page %p", m));
1094 	cnt.v_tfree++;
1095 
1096 	if (m->busy || VM_PAGE_INQUEUE1(m, PQ_FREE)) {
1097 		printf(
1098 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1099 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1100 		    m->hold_count);
1101 		if (VM_PAGE_INQUEUE1(m, PQ_FREE))
1102 			panic("vm_page_free: freeing free page");
1103 		else
1104 			panic("vm_page_free: freeing busy page");
1105 	}
1106 
1107 	/*
1108 	 * unqueue, then remove page.  Note that we cannot destroy
1109 	 * the page here because we do not want to call the pager's
1110 	 * callback routine until after we've put the page on the
1111 	 * appropriate free queue.
1112 	 */
1113 	vm_pageq_remove_nowakeup(m);
1114 	vm_page_remove(m);
1115 
1116 	/*
1117 	 * If fictitious remove object association and
1118 	 * return, otherwise delay object association removal.
1119 	 */
1120 	if ((m->flags & PG_FICTITIOUS) != 0) {
1121 		return;
1122 	}
1123 
1124 	m->valid = 0;
1125 	vm_page_undirty(m);
1126 
1127 	if (m->wire_count != 0) {
1128 		if (m->wire_count > 1) {
1129 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1130 				m->wire_count, (long)m->pindex);
1131 		}
1132 		panic("vm_page_free: freeing wired page");
1133 	}
1134 
1135 	/*
1136 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1137 	 */
1138 	if (m->flags & PG_UNMANAGED) {
1139 		m->flags &= ~PG_UNMANAGED;
1140 	}
1141 
1142 	if (m->hold_count != 0) {
1143 		m->flags &= ~PG_ZERO;
1144 		VM_PAGE_SETQUEUE2(m, PQ_HOLD);
1145 	} else
1146 		VM_PAGE_SETQUEUE1(m, PQ_FREE);
1147 	pq = &vm_page_queues[VM_PAGE_GETQUEUE(m)];
1148 	mtx_lock_spin(&vm_page_queue_free_mtx);
1149 	pq->lcnt++;
1150 	++(*pq->cnt);
1151 
1152 	/*
1153 	 * Put zero'd pages on the end ( where we look for zero'd pages
1154 	 * first ) and non-zerod pages at the head.
1155 	 */
1156 	if (m->flags & PG_ZERO) {
1157 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1158 		++vm_page_zero_count;
1159 	} else {
1160 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1161 	}
1162 	mtx_unlock_spin(&vm_page_queue_free_mtx);
1163 	vm_page_free_wakeup();
1164 }
1165 
1166 /*
1167  *	vm_page_unmanage:
1168  *
1169  * 	Prevent PV management from being done on the page.  The page is
1170  *	removed from the paging queues as if it were wired, and as a
1171  *	consequence of no longer being managed the pageout daemon will not
1172  *	touch it (since there is no way to locate the pte mappings for the
1173  *	page).  madvise() calls that mess with the pmap will also no longer
1174  *	operate on the page.
1175  *
1176  *	Beyond that the page is still reasonably 'normal'.  Freeing the page
1177  *	will clear the flag.
1178  *
1179  *	This routine is used by OBJT_PHYS objects - objects using unswappable
1180  *	physical memory as backing store rather then swap-backed memory and
1181  *	will eventually be extended to support 4MB unmanaged physical
1182  *	mappings.
1183  */
1184 void
1185 vm_page_unmanage(vm_page_t m)
1186 {
1187 
1188 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1189 	if ((m->flags & PG_UNMANAGED) == 0) {
1190 		if (m->wire_count == 0)
1191 			vm_pageq_remove(m);
1192 	}
1193 	vm_page_flag_set(m, PG_UNMANAGED);
1194 }
1195 
1196 /*
1197  *	vm_page_wire:
1198  *
1199  *	Mark this page as wired down by yet
1200  *	another map, removing it from paging queues
1201  *	as necessary.
1202  *
1203  *	The page queues must be locked.
1204  *	This routine may not block.
1205  */
1206 void
1207 vm_page_wire(vm_page_t m)
1208 {
1209 
1210 	/*
1211 	 * Only bump the wire statistics if the page is not already wired,
1212 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1213 	 * it is already off the queues).
1214 	 */
1215 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1216 	if (m->flags & PG_FICTITIOUS)
1217 		return;
1218 	if (m->wire_count == 0) {
1219 		if ((m->flags & PG_UNMANAGED) == 0)
1220 			vm_pageq_remove(m);
1221 		atomic_add_int(&cnt.v_wire_count, 1);
1222 	}
1223 	m->wire_count++;
1224 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1225 }
1226 
1227 /*
1228  *	vm_page_unwire:
1229  *
1230  *	Release one wiring of this page, potentially
1231  *	enabling it to be paged again.
1232  *
1233  *	Many pages placed on the inactive queue should actually go
1234  *	into the cache, but it is difficult to figure out which.  What
1235  *	we do instead, if the inactive target is well met, is to put
1236  *	clean pages at the head of the inactive queue instead of the tail.
1237  *	This will cause them to be moved to the cache more quickly and
1238  *	if not actively re-referenced, freed more quickly.  If we just
1239  *	stick these pages at the end of the inactive queue, heavy filesystem
1240  *	meta-data accesses can cause an unnecessary paging load on memory bound
1241  *	processes.  This optimization causes one-time-use metadata to be
1242  *	reused more quickly.
1243  *
1244  *	BUT, if we are in a low-memory situation we have no choice but to
1245  *	put clean pages on the cache queue.
1246  *
1247  *	A number of routines use vm_page_unwire() to guarantee that the page
1248  *	will go into either the inactive or active queues, and will NEVER
1249  *	be placed in the cache - for example, just after dirtying a page.
1250  *	dirty pages in the cache are not allowed.
1251  *
1252  *	The page queues must be locked.
1253  *	This routine may not block.
1254  */
1255 void
1256 vm_page_unwire(vm_page_t m, int activate)
1257 {
1258 
1259 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1260 	if (m->flags & PG_FICTITIOUS)
1261 		return;
1262 	if (m->wire_count > 0) {
1263 		m->wire_count--;
1264 		if (m->wire_count == 0) {
1265 			atomic_subtract_int(&cnt.v_wire_count, 1);
1266 			if (m->flags & PG_UNMANAGED) {
1267 				;
1268 			} else if (activate)
1269 				vm_pageq_enqueue(PQ_ACTIVE, m);
1270 			else {
1271 				vm_page_flag_clear(m, PG_WINATCFLS);
1272 				vm_pageq_enqueue(PQ_INACTIVE, m);
1273 			}
1274 		}
1275 	} else {
1276 		panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1277 	}
1278 }
1279 
1280 
1281 /*
1282  * Move the specified page to the inactive queue.  If the page has
1283  * any associated swap, the swap is deallocated.
1284  *
1285  * Normally athead is 0 resulting in LRU operation.  athead is set
1286  * to 1 if we want this page to be 'as if it were placed in the cache',
1287  * except without unmapping it from the process address space.
1288  *
1289  * This routine may not block.
1290  */
1291 static inline void
1292 _vm_page_deactivate(vm_page_t m, int athead)
1293 {
1294 
1295 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1296 
1297 	/*
1298 	 * Ignore if already inactive.
1299 	 */
1300 	if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE))
1301 		return;
1302 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1303 		if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1304 			cnt.v_reactivated++;
1305 		vm_page_flag_clear(m, PG_WINATCFLS);
1306 		vm_pageq_remove(m);
1307 		if (athead)
1308 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1309 		else
1310 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1311 		VM_PAGE_SETQUEUE2(m, PQ_INACTIVE);
1312 		vm_page_queues[PQ_INACTIVE].lcnt++;
1313 		cnt.v_inactive_count++;
1314 	}
1315 }
1316 
1317 void
1318 vm_page_deactivate(vm_page_t m)
1319 {
1320     _vm_page_deactivate(m, 0);
1321 }
1322 
1323 /*
1324  * vm_page_try_to_cache:
1325  *
1326  * Returns 0 on failure, 1 on success
1327  */
1328 int
1329 vm_page_try_to_cache(vm_page_t m)
1330 {
1331 
1332 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1333 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1334 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1335 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1336 		return (0);
1337 	}
1338 	pmap_remove_all(m);
1339 	if (m->dirty)
1340 		return (0);
1341 	vm_page_cache(m);
1342 	return (1);
1343 }
1344 
1345 /*
1346  * vm_page_try_to_free()
1347  *
1348  *	Attempt to free the page.  If we cannot free it, we do nothing.
1349  *	1 is returned on success, 0 on failure.
1350  */
1351 int
1352 vm_page_try_to_free(vm_page_t m)
1353 {
1354 
1355 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1356 	if (m->object != NULL)
1357 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1358 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1359 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1360 		return (0);
1361 	}
1362 	pmap_remove_all(m);
1363 	if (m->dirty)
1364 		return (0);
1365 	vm_page_free(m);
1366 	return (1);
1367 }
1368 
1369 /*
1370  * vm_page_cache
1371  *
1372  * Put the specified page onto the page cache queue (if appropriate).
1373  *
1374  * This routine may not block.
1375  */
1376 void
1377 vm_page_cache(vm_page_t m)
1378 {
1379 
1380 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1381 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1382 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1383 	    m->hold_count || m->wire_count) {
1384 		printf("vm_page_cache: attempting to cache busy page\n");
1385 		return;
1386 	}
1387 	if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1388 		return;
1389 
1390 	/*
1391 	 * Remove all pmaps and indicate that the page is not
1392 	 * writeable or mapped.
1393 	 */
1394 	pmap_remove_all(m);
1395 	if (m->dirty != 0) {
1396 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1397 			(long)m->pindex);
1398 	}
1399 	vm_pageq_remove_nowakeup(m);
1400 	vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1401 	vm_page_free_wakeup();
1402 }
1403 
1404 /*
1405  * vm_page_dontneed
1406  *
1407  *	Cache, deactivate, or do nothing as appropriate.  This routine
1408  *	is typically used by madvise() MADV_DONTNEED.
1409  *
1410  *	Generally speaking we want to move the page into the cache so
1411  *	it gets reused quickly.  However, this can result in a silly syndrome
1412  *	due to the page recycling too quickly.  Small objects will not be
1413  *	fully cached.  On the otherhand, if we move the page to the inactive
1414  *	queue we wind up with a problem whereby very large objects
1415  *	unnecessarily blow away our inactive and cache queues.
1416  *
1417  *	The solution is to move the pages based on a fixed weighting.  We
1418  *	either leave them alone, deactivate them, or move them to the cache,
1419  *	where moving them to the cache has the highest weighting.
1420  *	By forcing some pages into other queues we eventually force the
1421  *	system to balance the queues, potentially recovering other unrelated
1422  *	space from active.  The idea is to not force this to happen too
1423  *	often.
1424  */
1425 void
1426 vm_page_dontneed(vm_page_t m)
1427 {
1428 	static int dnweight;
1429 	int dnw;
1430 	int head;
1431 
1432 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1433 	dnw = ++dnweight;
1434 
1435 	/*
1436 	 * occassionally leave the page alone
1437 	 */
1438 	if ((dnw & 0x01F0) == 0 ||
1439 	    VM_PAGE_INQUEUE2(m, PQ_INACTIVE) ||
1440 	    VM_PAGE_INQUEUE1(m, PQ_CACHE)
1441 	) {
1442 		if (m->act_count >= ACT_INIT)
1443 			--m->act_count;
1444 		return;
1445 	}
1446 
1447 	if (m->dirty == 0 && pmap_is_modified(m))
1448 		vm_page_dirty(m);
1449 
1450 	if (m->dirty || (dnw & 0x0070) == 0) {
1451 		/*
1452 		 * Deactivate the page 3 times out of 32.
1453 		 */
1454 		head = 0;
1455 	} else {
1456 		/*
1457 		 * Cache the page 28 times out of every 32.  Note that
1458 		 * the page is deactivated instead of cached, but placed
1459 		 * at the head of the queue instead of the tail.
1460 		 */
1461 		head = 1;
1462 	}
1463 	_vm_page_deactivate(m, head);
1464 }
1465 
1466 /*
1467  * Grab a page, waiting until we are waken up due to the page
1468  * changing state.  We keep on waiting, if the page continues
1469  * to be in the object.  If the page doesn't exist, first allocate it
1470  * and then conditionally zero it.
1471  *
1472  * This routine may block.
1473  */
1474 vm_page_t
1475 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1476 {
1477 	vm_page_t m;
1478 
1479 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1480 retrylookup:
1481 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1482 		vm_page_lock_queues();
1483 		if (m->busy || (m->flags & PG_BUSY)) {
1484 			vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1485 			VM_OBJECT_UNLOCK(object);
1486 			msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0);
1487 			VM_OBJECT_LOCK(object);
1488 			if ((allocflags & VM_ALLOC_RETRY) == 0)
1489 				return (NULL);
1490 			goto retrylookup;
1491 		} else {
1492 			if (allocflags & VM_ALLOC_WIRED)
1493 				vm_page_wire(m);
1494 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
1495 				vm_page_busy(m);
1496 			vm_page_unlock_queues();
1497 			return (m);
1498 		}
1499 	}
1500 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1501 	if (m == NULL) {
1502 		VM_OBJECT_UNLOCK(object);
1503 		VM_WAIT;
1504 		VM_OBJECT_LOCK(object);
1505 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1506 			return (NULL);
1507 		goto retrylookup;
1508 	}
1509 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
1510 		pmap_zero_page(m);
1511 	return (m);
1512 }
1513 
1514 /*
1515  * Mapping function for valid bits or for dirty bits in
1516  * a page.  May not block.
1517  *
1518  * Inputs are required to range within a page.
1519  */
1520 inline int
1521 vm_page_bits(int base, int size)
1522 {
1523 	int first_bit;
1524 	int last_bit;
1525 
1526 	KASSERT(
1527 	    base + size <= PAGE_SIZE,
1528 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1529 	);
1530 
1531 	if (size == 0)		/* handle degenerate case */
1532 		return (0);
1533 
1534 	first_bit = base >> DEV_BSHIFT;
1535 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1536 
1537 	return ((2 << last_bit) - (1 << first_bit));
1538 }
1539 
1540 /*
1541  *	vm_page_set_validclean:
1542  *
1543  *	Sets portions of a page valid and clean.  The arguments are expected
1544  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1545  *	of any partial chunks touched by the range.  The invalid portion of
1546  *	such chunks will be zero'd.
1547  *
1548  *	This routine may not block.
1549  *
1550  *	(base + size) must be less then or equal to PAGE_SIZE.
1551  */
1552 void
1553 vm_page_set_validclean(vm_page_t m, int base, int size)
1554 {
1555 	int pagebits;
1556 	int frag;
1557 	int endoff;
1558 
1559 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1560 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1561 	if (size == 0)	/* handle degenerate case */
1562 		return;
1563 
1564 	/*
1565 	 * If the base is not DEV_BSIZE aligned and the valid
1566 	 * bit is clear, we have to zero out a portion of the
1567 	 * first block.
1568 	 */
1569 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1570 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1571 		pmap_zero_page_area(m, frag, base - frag);
1572 
1573 	/*
1574 	 * If the ending offset is not DEV_BSIZE aligned and the
1575 	 * valid bit is clear, we have to zero out a portion of
1576 	 * the last block.
1577 	 */
1578 	endoff = base + size;
1579 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1580 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1581 		pmap_zero_page_area(m, endoff,
1582 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1583 
1584 	/*
1585 	 * Set valid, clear dirty bits.  If validating the entire
1586 	 * page we can safely clear the pmap modify bit.  We also
1587 	 * use this opportunity to clear the PG_NOSYNC flag.  If a process
1588 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1589 	 * be set again.
1590 	 *
1591 	 * We set valid bits inclusive of any overlap, but we can only
1592 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1593 	 * the range.
1594 	 */
1595 	pagebits = vm_page_bits(base, size);
1596 	m->valid |= pagebits;
1597 #if 0	/* NOT YET */
1598 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1599 		frag = DEV_BSIZE - frag;
1600 		base += frag;
1601 		size -= frag;
1602 		if (size < 0)
1603 			size = 0;
1604 	}
1605 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1606 #endif
1607 	m->dirty &= ~pagebits;
1608 	if (base == 0 && size == PAGE_SIZE) {
1609 		pmap_clear_modify(m);
1610 		vm_page_flag_clear(m, PG_NOSYNC);
1611 	}
1612 }
1613 
1614 void
1615 vm_page_clear_dirty(vm_page_t m, int base, int size)
1616 {
1617 
1618 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1619 	m->dirty &= ~vm_page_bits(base, size);
1620 }
1621 
1622 /*
1623  *	vm_page_set_invalid:
1624  *
1625  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1626  *	valid and dirty bits for the effected areas are cleared.
1627  *
1628  *	May not block.
1629  */
1630 void
1631 vm_page_set_invalid(vm_page_t m, int base, int size)
1632 {
1633 	int bits;
1634 
1635 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1636 	bits = vm_page_bits(base, size);
1637 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1638 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
1639 		pmap_remove_all(m);
1640 	m->valid &= ~bits;
1641 	m->dirty &= ~bits;
1642 	m->object->generation++;
1643 }
1644 
1645 /*
1646  * vm_page_zero_invalid()
1647  *
1648  *	The kernel assumes that the invalid portions of a page contain
1649  *	garbage, but such pages can be mapped into memory by user code.
1650  *	When this occurs, we must zero out the non-valid portions of the
1651  *	page so user code sees what it expects.
1652  *
1653  *	Pages are most often semi-valid when the end of a file is mapped
1654  *	into memory and the file's size is not page aligned.
1655  */
1656 void
1657 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1658 {
1659 	int b;
1660 	int i;
1661 
1662 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1663 	/*
1664 	 * Scan the valid bits looking for invalid sections that
1665 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1666 	 * valid bit may be set ) have already been zerod by
1667 	 * vm_page_set_validclean().
1668 	 */
1669 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1670 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1671 		    (m->valid & (1 << i))
1672 		) {
1673 			if (i > b) {
1674 				pmap_zero_page_area(m,
1675 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1676 			}
1677 			b = i + 1;
1678 		}
1679 	}
1680 
1681 	/*
1682 	 * setvalid is TRUE when we can safely set the zero'd areas
1683 	 * as being valid.  We can do this if there are no cache consistancy
1684 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1685 	 */
1686 	if (setvalid)
1687 		m->valid = VM_PAGE_BITS_ALL;
1688 }
1689 
1690 /*
1691  *	vm_page_is_valid:
1692  *
1693  *	Is (partial) page valid?  Note that the case where size == 0
1694  *	will return FALSE in the degenerate case where the page is
1695  *	entirely invalid, and TRUE otherwise.
1696  *
1697  *	May not block.
1698  */
1699 int
1700 vm_page_is_valid(vm_page_t m, int base, int size)
1701 {
1702 	int bits = vm_page_bits(base, size);
1703 
1704 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1705 	if (m->valid && ((m->valid & bits) == bits))
1706 		return 1;
1707 	else
1708 		return 0;
1709 }
1710 
1711 /*
1712  * update dirty bits from pmap/mmu.  May not block.
1713  */
1714 void
1715 vm_page_test_dirty(vm_page_t m)
1716 {
1717 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1718 		vm_page_dirty(m);
1719 	}
1720 }
1721 
1722 int so_zerocp_fullpage = 0;
1723 
1724 void
1725 vm_page_cowfault(vm_page_t m)
1726 {
1727 	vm_page_t mnew;
1728 	vm_object_t object;
1729 	vm_pindex_t pindex;
1730 
1731 	object = m->object;
1732 	pindex = m->pindex;
1733 
1734  retry_alloc:
1735 	pmap_remove_all(m);
1736 	vm_page_remove(m);
1737 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
1738 	if (mnew == NULL) {
1739 		vm_page_insert(m, object, pindex);
1740 		vm_page_unlock_queues();
1741 		VM_OBJECT_UNLOCK(object);
1742 		VM_WAIT;
1743 		VM_OBJECT_LOCK(object);
1744 		vm_page_lock_queues();
1745 		goto retry_alloc;
1746 	}
1747 
1748 	if (m->cow == 0) {
1749 		/*
1750 		 * check to see if we raced with an xmit complete when
1751 		 * waiting to allocate a page.  If so, put things back
1752 		 * the way they were
1753 		 */
1754 		vm_page_free(mnew);
1755 		vm_page_insert(m, object, pindex);
1756 	} else { /* clear COW & copy page */
1757 		if (!so_zerocp_fullpage)
1758 			pmap_copy_page(m, mnew);
1759 		mnew->valid = VM_PAGE_BITS_ALL;
1760 		vm_page_dirty(mnew);
1761 		vm_page_flag_clear(mnew, PG_BUSY);
1762 		mnew->wire_count = m->wire_count - m->cow;
1763 		m->wire_count = m->cow;
1764 	}
1765 }
1766 
1767 void
1768 vm_page_cowclear(vm_page_t m)
1769 {
1770 
1771 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1772 	if (m->cow) {
1773 		m->cow--;
1774 		/*
1775 		 * let vm_fault add back write permission  lazily
1776 		 */
1777 	}
1778 	/*
1779 	 *  sf_buf_free() will free the page, so we needn't do it here
1780 	 */
1781 }
1782 
1783 void
1784 vm_page_cowsetup(vm_page_t m)
1785 {
1786 
1787 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1788 	m->cow++;
1789 	pmap_page_protect(m, VM_PROT_READ);
1790 }
1791 
1792 #include "opt_ddb.h"
1793 #ifdef DDB
1794 #include <sys/kernel.h>
1795 
1796 #include <ddb/ddb.h>
1797 
1798 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1799 {
1800 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1801 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1802 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1803 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1804 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1805 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1806 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1807 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1808 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1809 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1810 }
1811 
1812 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1813 {
1814 	int i;
1815 	db_printf("PQ_FREE:");
1816 	for (i = 0; i < PQ_NUMCOLORS; i++) {
1817 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1818 	}
1819 	db_printf("\n");
1820 
1821 	db_printf("PQ_CACHE:");
1822 	for (i = 0; i < PQ_NUMCOLORS; i++) {
1823 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1824 	}
1825 	db_printf("\n");
1826 
1827 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1828 		vm_page_queues[PQ_ACTIVE].lcnt,
1829 		vm_page_queues[PQ_INACTIVE].lcnt);
1830 }
1831 #endif /* DDB */
1832