FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_page.c
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: src/sys/vm/vm_page.c,v 1.290.2.6 2005/03/02 04:00:04 alc Exp $");
101
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/lock.h>
105 #include <sys/malloc.h>
106 #include <sys/mutex.h>
107 #include <sys/proc.h>
108 #include <sys/vmmeter.h>
109 #include <sys/vnode.h>
110
111 #include <vm/vm.h>
112 #include <vm/vm_param.h>
113 #include <vm/vm_kern.h>
114 #include <vm/vm_object.h>
115 #include <vm/vm_page.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/vm_pager.h>
118 #include <vm/vm_extern.h>
119 #include <vm/uma.h>
120 #include <vm/uma_int.h>
121
122 /*
123 * Associated with page of user-allocatable memory is a
124 * page structure.
125 */
126
127 struct mtx vm_page_queue_mtx;
128 struct mtx vm_page_queue_free_mtx;
129
130 vm_page_t vm_page_array = 0;
131 int vm_page_array_size = 0;
132 long first_page = 0;
133 int vm_page_zero_count = 0;
134
135 /*
136 * vm_set_page_size:
137 *
138 * Sets the page size, perhaps based upon the memory
139 * size. Must be called before any use of page-size
140 * dependent functions.
141 */
142 void
143 vm_set_page_size(void)
144 {
145 if (cnt.v_page_size == 0)
146 cnt.v_page_size = PAGE_SIZE;
147 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
148 panic("vm_set_page_size: page size not a power of two");
149 }
150
151 /*
152 * vm_page_startup:
153 *
154 * Initializes the resident memory module.
155 *
156 * Allocates memory for the page cells, and
157 * for the object/offset-to-page hash table headers.
158 * Each page cell is initialized and placed on the free list.
159 */
160 vm_offset_t
161 vm_page_startup(vm_offset_t vaddr)
162 {
163 vm_offset_t mapped;
164 vm_size_t npages;
165 vm_paddr_t page_range;
166 vm_paddr_t new_end;
167 int i;
168 vm_paddr_t pa;
169 int nblocks;
170 vm_paddr_t last_pa;
171
172 /* the biggest memory array is the second group of pages */
173 vm_paddr_t end;
174 vm_paddr_t biggestsize;
175 int biggestone;
176
177 vm_paddr_t total;
178 vm_size_t bootpages;
179
180 total = 0;
181 biggestsize = 0;
182 biggestone = 0;
183 nblocks = 0;
184 vaddr = round_page(vaddr);
185
186 for (i = 0; phys_avail[i + 1]; i += 2) {
187 phys_avail[i] = round_page(phys_avail[i]);
188 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
189 }
190
191 for (i = 0; phys_avail[i + 1]; i += 2) {
192 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
193
194 if (size > biggestsize) {
195 biggestone = i;
196 biggestsize = size;
197 }
198 ++nblocks;
199 total += size;
200 }
201
202 end = phys_avail[biggestone+1];
203
204 /*
205 * Initialize the locks.
206 */
207 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF |
208 MTX_RECURSE);
209 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
210 MTX_SPIN);
211
212 /*
213 * Initialize the queue headers for the free queue, the active queue
214 * and the inactive queue.
215 */
216 vm_pageq_init();
217
218 /*
219 * Allocate memory for use when boot strapping the kernel memory
220 * allocator.
221 */
222 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
223 new_end = end - bootpages;
224 new_end = trunc_page(new_end);
225 mapped = pmap_map(&vaddr, new_end, end,
226 VM_PROT_READ | VM_PROT_WRITE);
227 bzero((caddr_t) mapped, end - new_end);
228 uma_startup((caddr_t)mapped);
229
230 /*
231 * Compute the number of pages of memory that will be available for
232 * use (taking into account the overhead of a page structure per
233 * page).
234 */
235 first_page = phys_avail[0] / PAGE_SIZE;
236 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
237 npages = (total - (page_range * sizeof(struct vm_page)) -
238 (end - new_end)) / PAGE_SIZE;
239 end = new_end;
240
241 /*
242 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
243 */
244 vaddr += PAGE_SIZE;
245
246 /*
247 * Initialize the mem entry structures now, and put them in the free
248 * queue.
249 */
250 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
251 mapped = pmap_map(&vaddr, new_end, end,
252 VM_PROT_READ | VM_PROT_WRITE);
253 vm_page_array = (vm_page_t) mapped;
254 phys_avail[biggestone + 1] = new_end;
255
256 /*
257 * Clear all of the page structures
258 */
259 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
260 vm_page_array_size = page_range;
261
262 /*
263 * Construct the free queue(s) in descending order (by physical
264 * address) so that the first 16MB of physical memory is allocated
265 * last rather than first. On large-memory machines, this avoids
266 * the exhaustion of low physical memory before isa_dmainit has run.
267 */
268 cnt.v_page_count = 0;
269 cnt.v_free_count = 0;
270 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
271 pa = phys_avail[i];
272 last_pa = phys_avail[i + 1];
273 while (pa < last_pa && npages-- > 0) {
274 vm_pageq_add_new_page(pa);
275 pa += PAGE_SIZE;
276 }
277 }
278 return (vaddr);
279 }
280
281 void
282 vm_page_flag_set(vm_page_t m, unsigned short bits)
283 {
284
285 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
286 m->flags |= bits;
287 }
288
289 void
290 vm_page_flag_clear(vm_page_t m, unsigned short bits)
291 {
292
293 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
294 m->flags &= ~bits;
295 }
296
297 void
298 vm_page_busy(vm_page_t m)
299 {
300
301 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
302 KASSERT((m->flags & PG_BUSY) == 0,
303 ("vm_page_busy: page already busy!!!"));
304 vm_page_flag_set(m, PG_BUSY);
305 }
306
307 /*
308 * vm_page_flash:
309 *
310 * wakeup anyone waiting for the page.
311 */
312 void
313 vm_page_flash(vm_page_t m)
314 {
315
316 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
317 if (m->flags & PG_WANTED) {
318 vm_page_flag_clear(m, PG_WANTED);
319 wakeup(m);
320 }
321 }
322
323 /*
324 * vm_page_wakeup:
325 *
326 * clear the PG_BUSY flag and wakeup anyone waiting for the
327 * page.
328 *
329 */
330 void
331 vm_page_wakeup(vm_page_t m)
332 {
333
334 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
335 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
336 vm_page_flag_clear(m, PG_BUSY);
337 vm_page_flash(m);
338 }
339
340 void
341 vm_page_io_start(vm_page_t m)
342 {
343
344 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
345 m->busy++;
346 }
347
348 void
349 vm_page_io_finish(vm_page_t m)
350 {
351
352 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
353 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
354 m->busy--;
355 if (m->busy == 0)
356 vm_page_flash(m);
357 }
358
359 /*
360 * Keep page from being freed by the page daemon
361 * much of the same effect as wiring, except much lower
362 * overhead and should be used only for *very* temporary
363 * holding ("wiring").
364 */
365 void
366 vm_page_hold(vm_page_t mem)
367 {
368
369 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
370 mem->hold_count++;
371 }
372
373 void
374 vm_page_unhold(vm_page_t mem)
375 {
376
377 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
378 --mem->hold_count;
379 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
380 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
381 vm_page_free_toq(mem);
382 }
383
384 /*
385 * vm_page_free:
386 *
387 * Free a page
388 *
389 * The clearing of PG_ZERO is a temporary safety until the code can be
390 * reviewed to determine that PG_ZERO is being properly cleared on
391 * write faults or maps. PG_ZERO was previously cleared in
392 * vm_page_alloc().
393 */
394 void
395 vm_page_free(vm_page_t m)
396 {
397 vm_page_flag_clear(m, PG_ZERO);
398 vm_page_free_toq(m);
399 vm_page_zero_idle_wakeup();
400 }
401
402 /*
403 * vm_page_free_zero:
404 *
405 * Free a page to the zerod-pages queue
406 */
407 void
408 vm_page_free_zero(vm_page_t m)
409 {
410 vm_page_flag_set(m, PG_ZERO);
411 vm_page_free_toq(m);
412 }
413
414 /*
415 * vm_page_sleep_if_busy:
416 *
417 * Sleep and release the page queues lock if PG_BUSY is set or,
418 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the
419 * thread slept and the page queues lock was released.
420 * Otherwise, retains the page queues lock and returns FALSE.
421 */
422 int
423 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
424 {
425 vm_object_t object;
426
427 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
428 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
429 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
430 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
431 /*
432 * It's possible that while we sleep, the page will get
433 * unbusied and freed. If we are holding the object
434 * lock, we will assume we hold a reference to the object
435 * such that even if m->object changes, we can re-lock
436 * it.
437 */
438 object = m->object;
439 VM_OBJECT_UNLOCK(object);
440 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
441 VM_OBJECT_LOCK(object);
442 return (TRUE);
443 }
444 return (FALSE);
445 }
446
447 /*
448 * vm_page_dirty:
449 *
450 * make page all dirty
451 */
452 void
453 vm_page_dirty(vm_page_t m)
454 {
455 KASSERT(m->queue - m->pc != PQ_CACHE,
456 ("vm_page_dirty: page in cache!"));
457 KASSERT(m->queue - m->pc != PQ_FREE,
458 ("vm_page_dirty: page is free!"));
459 m->dirty = VM_PAGE_BITS_ALL;
460 }
461
462 /*
463 * vm_page_splay:
464 *
465 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
466 * the vm_page containing the given pindex. If, however, that
467 * pindex is not found in the vm_object, returns a vm_page that is
468 * adjacent to the pindex, coming before or after it.
469 */
470 vm_page_t
471 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
472 {
473 struct vm_page dummy;
474 vm_page_t lefttreemax, righttreemin, y;
475
476 if (root == NULL)
477 return (root);
478 lefttreemax = righttreemin = &dummy;
479 for (;; root = y) {
480 if (pindex < root->pindex) {
481 if ((y = root->left) == NULL)
482 break;
483 if (pindex < y->pindex) {
484 /* Rotate right. */
485 root->left = y->right;
486 y->right = root;
487 root = y;
488 if ((y = root->left) == NULL)
489 break;
490 }
491 /* Link into the new root's right tree. */
492 righttreemin->left = root;
493 righttreemin = root;
494 } else if (pindex > root->pindex) {
495 if ((y = root->right) == NULL)
496 break;
497 if (pindex > y->pindex) {
498 /* Rotate left. */
499 root->right = y->left;
500 y->left = root;
501 root = y;
502 if ((y = root->right) == NULL)
503 break;
504 }
505 /* Link into the new root's left tree. */
506 lefttreemax->right = root;
507 lefttreemax = root;
508 } else
509 break;
510 }
511 /* Assemble the new root. */
512 lefttreemax->right = root->left;
513 righttreemin->left = root->right;
514 root->left = dummy.right;
515 root->right = dummy.left;
516 return (root);
517 }
518
519 /*
520 * vm_page_insert: [ internal use only ]
521 *
522 * Inserts the given mem entry into the object and object list.
523 *
524 * The pagetables are not updated but will presumably fault the page
525 * in if necessary, or if a kernel page the caller will at some point
526 * enter the page into the kernel's pmap. We are not allowed to block
527 * here so we *can't* do this anyway.
528 *
529 * The object and page must be locked.
530 * This routine may not block.
531 */
532 void
533 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
534 {
535 vm_page_t root;
536
537 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
538 if (m->object != NULL)
539 panic("vm_page_insert: page already inserted");
540
541 /*
542 * Record the object/offset pair in this page
543 */
544 m->object = object;
545 m->pindex = pindex;
546
547 /*
548 * Now link into the object's ordered list of backed pages.
549 */
550 root = object->root;
551 if (root == NULL) {
552 m->left = NULL;
553 m->right = NULL;
554 TAILQ_INSERT_TAIL(&object->memq, m, listq);
555 } else {
556 root = vm_page_splay(pindex, root);
557 if (pindex < root->pindex) {
558 m->left = root->left;
559 m->right = root;
560 root->left = NULL;
561 TAILQ_INSERT_BEFORE(root, m, listq);
562 } else if (pindex == root->pindex)
563 panic("vm_page_insert: offset already allocated");
564 else {
565 m->right = root->right;
566 m->left = root;
567 root->right = NULL;
568 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
569 }
570 }
571 object->root = m;
572 object->generation++;
573
574 /*
575 * show that the object has one more resident page.
576 */
577 object->resident_page_count++;
578
579 /*
580 * Since we are inserting a new and possibly dirty page,
581 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
582 */
583 if (m->flags & PG_WRITEABLE)
584 vm_object_set_writeable_dirty(object);
585 }
586
587 /*
588 * vm_page_remove:
589 * NOTE: used by device pager as well -wfj
590 *
591 * Removes the given mem entry from the object/offset-page
592 * table and the object page list, but do not invalidate/terminate
593 * the backing store.
594 *
595 * The object and page must be locked.
596 * The underlying pmap entry (if any) is NOT removed here.
597 * This routine may not block.
598 */
599 void
600 vm_page_remove(vm_page_t m)
601 {
602 vm_object_t object;
603 vm_page_t root;
604
605 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
606 if ((object = m->object) == NULL)
607 return;
608 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
609 if (m->flags & PG_BUSY) {
610 vm_page_flag_clear(m, PG_BUSY);
611 vm_page_flash(m);
612 }
613
614 /*
615 * Now remove from the object's list of backed pages.
616 */
617 if (m != object->root)
618 vm_page_splay(m->pindex, object->root);
619 if (m->left == NULL)
620 root = m->right;
621 else {
622 root = vm_page_splay(m->pindex, m->left);
623 root->right = m->right;
624 }
625 object->root = root;
626 TAILQ_REMOVE(&object->memq, m, listq);
627
628 /*
629 * And show that the object has one fewer resident page.
630 */
631 object->resident_page_count--;
632 object->generation++;
633
634 m->object = NULL;
635 }
636
637 /*
638 * vm_page_lookup:
639 *
640 * Returns the page associated with the object/offset
641 * pair specified; if none is found, NULL is returned.
642 *
643 * The object must be locked.
644 * This routine may not block.
645 * This is a critical path routine
646 */
647 vm_page_t
648 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
649 {
650 vm_page_t m;
651
652 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
653 if ((m = object->root) != NULL && m->pindex != pindex) {
654 m = vm_page_splay(pindex, m);
655 if ((object->root = m)->pindex != pindex)
656 m = NULL;
657 }
658 return (m);
659 }
660
661 /*
662 * vm_page_rename:
663 *
664 * Move the given memory entry from its
665 * current object to the specified target object/offset.
666 *
667 * The object must be locked.
668 * This routine may not block.
669 *
670 * Note: swap associated with the page must be invalidated by the move. We
671 * have to do this for several reasons: (1) we aren't freeing the
672 * page, (2) we are dirtying the page, (3) the VM system is probably
673 * moving the page from object A to B, and will then later move
674 * the backing store from A to B and we can't have a conflict.
675 *
676 * Note: we *always* dirty the page. It is necessary both for the
677 * fact that we moved it, and because we may be invalidating
678 * swap. If the page is on the cache, we have to deactivate it
679 * or vm_page_dirty() will panic. Dirty pages are not allowed
680 * on the cache.
681 */
682 void
683 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
684 {
685
686 vm_page_remove(m);
687 vm_page_insert(m, new_object, new_pindex);
688 if (m->queue - m->pc == PQ_CACHE)
689 vm_page_deactivate(m);
690 vm_page_dirty(m);
691 }
692
693 /*
694 * vm_page_select_cache:
695 *
696 * Move a page of the given color from the cache queue to the free
697 * queue. As pages might be found, but are not applicable, they are
698 * deactivated.
699 *
700 * This routine may not block.
701 */
702 vm_page_t
703 vm_page_select_cache(int color)
704 {
705 vm_object_t object;
706 vm_page_t m;
707 boolean_t was_trylocked;
708
709 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
710 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) {
711 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
712 KASSERT(!pmap_page_is_mapped(m),
713 ("Found mapped cache page %p", m));
714 KASSERT((m->flags & PG_UNMANAGED) == 0,
715 ("Found unmanaged cache page %p", m));
716 KASSERT(m->wire_count == 0, ("Found wired cache page %p", m));
717 if (m->hold_count == 0 && (object = m->object,
718 (was_trylocked = VM_OBJECT_TRYLOCK(object)) ||
719 VM_OBJECT_LOCKED(object))) {
720 KASSERT((m->flags & PG_BUSY) == 0 && m->busy == 0,
721 ("Found busy cache page %p", m));
722 vm_page_free(m);
723 if (was_trylocked)
724 VM_OBJECT_UNLOCK(object);
725 break;
726 }
727 vm_page_deactivate(m);
728 }
729 return (m);
730 }
731
732 /*
733 * vm_page_alloc:
734 *
735 * Allocate and return a memory cell associated
736 * with this VM object/offset pair.
737 *
738 * page_req classes:
739 * VM_ALLOC_NORMAL normal process request
740 * VM_ALLOC_SYSTEM system *really* needs a page
741 * VM_ALLOC_INTERRUPT interrupt time request
742 * VM_ALLOC_ZERO zero page
743 *
744 * This routine may not block.
745 *
746 * Additional special handling is required when called from an
747 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
748 * the page cache in this case.
749 */
750 vm_page_t
751 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
752 {
753 vm_page_t m = NULL;
754 int color, flags, page_req;
755
756 page_req = req & VM_ALLOC_CLASS_MASK;
757 KASSERT(curthread->td_intr_nesting_level == 0 ||
758 page_req == VM_ALLOC_INTERRUPT,
759 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context"));
760
761 if ((req & VM_ALLOC_NOOBJ) == 0) {
762 KASSERT(object != NULL,
763 ("vm_page_alloc: NULL object."));
764 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
765 color = (pindex + object->pg_color) & PQ_L2_MASK;
766 } else
767 color = pindex & PQ_L2_MASK;
768
769 /*
770 * The pager is allowed to eat deeper into the free page list.
771 */
772 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
773 page_req = VM_ALLOC_SYSTEM;
774 };
775
776 loop:
777 mtx_lock_spin(&vm_page_queue_free_mtx);
778 if (cnt.v_free_count > cnt.v_free_reserved ||
779 (page_req == VM_ALLOC_SYSTEM &&
780 cnt.v_cache_count == 0 &&
781 cnt.v_free_count > cnt.v_interrupt_free_min) ||
782 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) {
783 /*
784 * Allocate from the free queue if the number of free pages
785 * exceeds the minimum for the request class.
786 */
787 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0);
788 } else if (page_req != VM_ALLOC_INTERRUPT) {
789 mtx_unlock_spin(&vm_page_queue_free_mtx);
790 /*
791 * Allocatable from cache (non-interrupt only). On success,
792 * we must free the page and try again, thus ensuring that
793 * cnt.v_*_free_min counters are replenished.
794 */
795 vm_page_lock_queues();
796 if ((m = vm_page_select_cache(color)) == NULL) {
797 #if defined(DIAGNOSTIC)
798 if (cnt.v_cache_count > 0)
799 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
800 #endif
801 vm_page_unlock_queues();
802 atomic_add_int(&vm_pageout_deficit, 1);
803 pagedaemon_wakeup();
804 return (NULL);
805 }
806 vm_page_unlock_queues();
807 goto loop;
808 } else {
809 /*
810 * Not allocatable from cache from interrupt, give up.
811 */
812 mtx_unlock_spin(&vm_page_queue_free_mtx);
813 atomic_add_int(&vm_pageout_deficit, 1);
814 pagedaemon_wakeup();
815 return (NULL);
816 }
817
818 /*
819 * At this point we had better have found a good page.
820 */
821
822 KASSERT(
823 m != NULL,
824 ("vm_page_alloc(): missing page on free queue")
825 );
826
827 /*
828 * Remove from free queue
829 */
830 vm_pageq_remove_nowakeup(m);
831
832 /*
833 * Initialize structure. Only the PG_ZERO flag is inherited.
834 */
835 flags = PG_BUSY;
836 if (m->flags & PG_ZERO) {
837 vm_page_zero_count--;
838 if (req & VM_ALLOC_ZERO)
839 flags = PG_ZERO | PG_BUSY;
840 }
841 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ))
842 flags &= ~PG_BUSY;
843 m->flags = flags;
844 if (req & VM_ALLOC_WIRED) {
845 atomic_add_int(&cnt.v_wire_count, 1);
846 m->wire_count = 1;
847 } else
848 m->wire_count = 0;
849 m->hold_count = 0;
850 m->act_count = 0;
851 m->busy = 0;
852 m->valid = 0;
853 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
854 mtx_unlock_spin(&vm_page_queue_free_mtx);
855
856 if ((req & VM_ALLOC_NOOBJ) == 0)
857 vm_page_insert(m, object, pindex);
858 else
859 m->pindex = pindex;
860
861 /*
862 * Don't wakeup too often - wakeup the pageout daemon when
863 * we would be nearly out of memory.
864 */
865 if (vm_paging_needed())
866 pagedaemon_wakeup();
867
868 return (m);
869 }
870
871 /*
872 * vm_wait: (also see VM_WAIT macro)
873 *
874 * Block until free pages are available for allocation
875 * - Called in various places before memory allocations.
876 */
877 void
878 vm_wait(void)
879 {
880
881 vm_page_lock_queues();
882 if (curproc == pageproc) {
883 vm_pageout_pages_needed = 1;
884 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx,
885 PDROP | PSWP, "VMWait", 0);
886 } else {
887 if (!vm_pages_needed) {
888 vm_pages_needed = 1;
889 wakeup(&vm_pages_needed);
890 }
891 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM,
892 "vmwait", 0);
893 }
894 }
895
896 /*
897 * vm_waitpfault: (also see VM_WAITPFAULT macro)
898 *
899 * Block until free pages are available for allocation
900 * - Called only in vm_fault so that processes page faulting
901 * can be easily tracked.
902 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
903 * processes will be able to grab memory first. Do not change
904 * this balance without careful testing first.
905 */
906 void
907 vm_waitpfault(void)
908 {
909
910 vm_page_lock_queues();
911 if (!vm_pages_needed) {
912 vm_pages_needed = 1;
913 wakeup(&vm_pages_needed);
914 }
915 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER,
916 "pfault", 0);
917 }
918
919 /*
920 * vm_page_activate:
921 *
922 * Put the specified page on the active list (if appropriate).
923 * Ensure that act_count is at least ACT_INIT but do not otherwise
924 * mess with it.
925 *
926 * The page queues must be locked.
927 * This routine may not block.
928 */
929 void
930 vm_page_activate(vm_page_t m)
931 {
932
933 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
934 if (m->queue != PQ_ACTIVE) {
935 if ((m->queue - m->pc) == PQ_CACHE)
936 cnt.v_reactivated++;
937 vm_pageq_remove(m);
938 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
939 if (m->act_count < ACT_INIT)
940 m->act_count = ACT_INIT;
941 vm_pageq_enqueue(PQ_ACTIVE, m);
942 }
943 } else {
944 if (m->act_count < ACT_INIT)
945 m->act_count = ACT_INIT;
946 }
947 }
948
949 /*
950 * vm_page_free_wakeup:
951 *
952 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
953 * routine is called when a page has been added to the cache or free
954 * queues.
955 *
956 * The page queues must be locked.
957 * This routine may not block.
958 */
959 static __inline void
960 vm_page_free_wakeup(void)
961 {
962
963 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
964 /*
965 * if pageout daemon needs pages, then tell it that there are
966 * some free.
967 */
968 if (vm_pageout_pages_needed &&
969 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
970 wakeup(&vm_pageout_pages_needed);
971 vm_pageout_pages_needed = 0;
972 }
973 /*
974 * wakeup processes that are waiting on memory if we hit a
975 * high water mark. And wakeup scheduler process if we have
976 * lots of memory. this process will swapin processes.
977 */
978 if (vm_pages_needed && !vm_page_count_min()) {
979 vm_pages_needed = 0;
980 wakeup(&cnt.v_free_count);
981 }
982 }
983
984 /*
985 * vm_page_free_toq:
986 *
987 * Returns the given page to the PQ_FREE list,
988 * disassociating it with any VM object.
989 *
990 * Object and page must be locked prior to entry.
991 * This routine may not block.
992 */
993
994 void
995 vm_page_free_toq(vm_page_t m)
996 {
997 struct vpgqueues *pq;
998 vm_object_t object = m->object;
999
1000 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1001 cnt.v_tfree++;
1002
1003 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1004 printf(
1005 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1006 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1007 m->hold_count);
1008 if ((m->queue - m->pc) == PQ_FREE)
1009 panic("vm_page_free: freeing free page");
1010 else
1011 panic("vm_page_free: freeing busy page");
1012 }
1013
1014 /*
1015 * unqueue, then remove page. Note that we cannot destroy
1016 * the page here because we do not want to call the pager's
1017 * callback routine until after we've put the page on the
1018 * appropriate free queue.
1019 */
1020 vm_pageq_remove_nowakeup(m);
1021 vm_page_remove(m);
1022
1023 /*
1024 * If fictitious remove object association and
1025 * return, otherwise delay object association removal.
1026 */
1027 if ((m->flags & PG_FICTITIOUS) != 0) {
1028 return;
1029 }
1030
1031 m->valid = 0;
1032 vm_page_undirty(m);
1033
1034 if (m->wire_count != 0) {
1035 if (m->wire_count > 1) {
1036 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1037 m->wire_count, (long)m->pindex);
1038 }
1039 panic("vm_page_free: freeing wired page");
1040 }
1041
1042 /*
1043 * If we've exhausted the object's resident pages we want to free
1044 * it up.
1045 */
1046 if (object &&
1047 (object->type == OBJT_VNODE) &&
1048 ((object->flags & OBJ_DEAD) == 0)
1049 ) {
1050 struct vnode *vp = (struct vnode *)object->handle;
1051
1052 if (vp) {
1053 VI_LOCK(vp);
1054 if (VSHOULDFREE(vp))
1055 vfree(vp);
1056 VI_UNLOCK(vp);
1057 }
1058 }
1059
1060 /*
1061 * Clear the UNMANAGED flag when freeing an unmanaged page.
1062 */
1063 if (m->flags & PG_UNMANAGED) {
1064 m->flags &= ~PG_UNMANAGED;
1065 }
1066
1067 if (m->hold_count != 0) {
1068 m->flags &= ~PG_ZERO;
1069 m->queue = PQ_HOLD;
1070 } else
1071 m->queue = PQ_FREE + m->pc;
1072 pq = &vm_page_queues[m->queue];
1073 mtx_lock_spin(&vm_page_queue_free_mtx);
1074 pq->lcnt++;
1075 ++(*pq->cnt);
1076
1077 /*
1078 * Put zero'd pages on the end ( where we look for zero'd pages
1079 * first ) and non-zerod pages at the head.
1080 */
1081 if (m->flags & PG_ZERO) {
1082 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1083 ++vm_page_zero_count;
1084 } else {
1085 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1086 }
1087 mtx_unlock_spin(&vm_page_queue_free_mtx);
1088 vm_page_free_wakeup();
1089 }
1090
1091 /*
1092 * vm_page_unmanage:
1093 *
1094 * Prevent PV management from being done on the page. The page is
1095 * removed from the paging queues as if it were wired, and as a
1096 * consequence of no longer being managed the pageout daemon will not
1097 * touch it (since there is no way to locate the pte mappings for the
1098 * page). madvise() calls that mess with the pmap will also no longer
1099 * operate on the page.
1100 *
1101 * Beyond that the page is still reasonably 'normal'. Freeing the page
1102 * will clear the flag.
1103 *
1104 * This routine is used by OBJT_PHYS objects - objects using unswappable
1105 * physical memory as backing store rather then swap-backed memory and
1106 * will eventually be extended to support 4MB unmanaged physical
1107 * mappings.
1108 */
1109 void
1110 vm_page_unmanage(vm_page_t m)
1111 {
1112
1113 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1114 if ((m->flags & PG_UNMANAGED) == 0) {
1115 if (m->wire_count == 0)
1116 vm_pageq_remove(m);
1117 }
1118 vm_page_flag_set(m, PG_UNMANAGED);
1119 }
1120
1121 /*
1122 * vm_page_wire:
1123 *
1124 * Mark this page as wired down by yet
1125 * another map, removing it from paging queues
1126 * as necessary.
1127 *
1128 * The page queues must be locked.
1129 * This routine may not block.
1130 */
1131 void
1132 vm_page_wire(vm_page_t m)
1133 {
1134
1135 /*
1136 * Only bump the wire statistics if the page is not already wired,
1137 * and only unqueue the page if it is on some queue (if it is unmanaged
1138 * it is already off the queues).
1139 */
1140 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1141 if (m->flags & PG_FICTITIOUS)
1142 return;
1143 if (m->wire_count == 0) {
1144 if ((m->flags & PG_UNMANAGED) == 0)
1145 vm_pageq_remove(m);
1146 atomic_add_int(&cnt.v_wire_count, 1);
1147 }
1148 m->wire_count++;
1149 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1150 }
1151
1152 /*
1153 * vm_page_unwire:
1154 *
1155 * Release one wiring of this page, potentially
1156 * enabling it to be paged again.
1157 *
1158 * Many pages placed on the inactive queue should actually go
1159 * into the cache, but it is difficult to figure out which. What
1160 * we do instead, if the inactive target is well met, is to put
1161 * clean pages at the head of the inactive queue instead of the tail.
1162 * This will cause them to be moved to the cache more quickly and
1163 * if not actively re-referenced, freed more quickly. If we just
1164 * stick these pages at the end of the inactive queue, heavy filesystem
1165 * meta-data accesses can cause an unnecessary paging load on memory bound
1166 * processes. This optimization causes one-time-use metadata to be
1167 * reused more quickly.
1168 *
1169 * BUT, if we are in a low-memory situation we have no choice but to
1170 * put clean pages on the cache queue.
1171 *
1172 * A number of routines use vm_page_unwire() to guarantee that the page
1173 * will go into either the inactive or active queues, and will NEVER
1174 * be placed in the cache - for example, just after dirtying a page.
1175 * dirty pages in the cache are not allowed.
1176 *
1177 * The page queues must be locked.
1178 * This routine may not block.
1179 */
1180 void
1181 vm_page_unwire(vm_page_t m, int activate)
1182 {
1183
1184 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1185 if (m->flags & PG_FICTITIOUS)
1186 return;
1187 if (m->wire_count > 0) {
1188 m->wire_count--;
1189 if (m->wire_count == 0) {
1190 atomic_subtract_int(&cnt.v_wire_count, 1);
1191 if (m->flags & PG_UNMANAGED) {
1192 ;
1193 } else if (activate)
1194 vm_pageq_enqueue(PQ_ACTIVE, m);
1195 else {
1196 vm_page_flag_clear(m, PG_WINATCFLS);
1197 vm_pageq_enqueue(PQ_INACTIVE, m);
1198 }
1199 }
1200 } else {
1201 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1202 }
1203 }
1204
1205
1206 /*
1207 * Move the specified page to the inactive queue. If the page has
1208 * any associated swap, the swap is deallocated.
1209 *
1210 * Normally athead is 0 resulting in LRU operation. athead is set
1211 * to 1 if we want this page to be 'as if it were placed in the cache',
1212 * except without unmapping it from the process address space.
1213 *
1214 * This routine may not block.
1215 */
1216 static __inline void
1217 _vm_page_deactivate(vm_page_t m, int athead)
1218 {
1219
1220 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1221
1222 /*
1223 * Ignore if already inactive.
1224 */
1225 if (m->queue == PQ_INACTIVE)
1226 return;
1227 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1228 if ((m->queue - m->pc) == PQ_CACHE)
1229 cnt.v_reactivated++;
1230 vm_page_flag_clear(m, PG_WINATCFLS);
1231 vm_pageq_remove(m);
1232 if (athead)
1233 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1234 else
1235 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1236 m->queue = PQ_INACTIVE;
1237 vm_page_queues[PQ_INACTIVE].lcnt++;
1238 cnt.v_inactive_count++;
1239 }
1240 }
1241
1242 void
1243 vm_page_deactivate(vm_page_t m)
1244 {
1245 _vm_page_deactivate(m, 0);
1246 }
1247
1248 /*
1249 * vm_page_try_to_cache:
1250 *
1251 * Returns 0 on failure, 1 on success
1252 */
1253 int
1254 vm_page_try_to_cache(vm_page_t m)
1255 {
1256
1257 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1258 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1259 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1260 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1261 return (0);
1262 }
1263 pmap_remove_all(m);
1264 if (m->dirty)
1265 return (0);
1266 vm_page_cache(m);
1267 return (1);
1268 }
1269
1270 /*
1271 * vm_page_try_to_free()
1272 *
1273 * Attempt to free the page. If we cannot free it, we do nothing.
1274 * 1 is returned on success, 0 on failure.
1275 */
1276 int
1277 vm_page_try_to_free(vm_page_t m)
1278 {
1279
1280 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1281 if (m->object != NULL)
1282 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1283 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1284 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1285 return (0);
1286 }
1287 pmap_remove_all(m);
1288 if (m->dirty)
1289 return (0);
1290 vm_page_free(m);
1291 return (1);
1292 }
1293
1294 /*
1295 * vm_page_cache
1296 *
1297 * Put the specified page onto the page cache queue (if appropriate).
1298 *
1299 * This routine may not block.
1300 */
1301 void
1302 vm_page_cache(vm_page_t m)
1303 {
1304
1305 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1306 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1307 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1308 m->hold_count || m->wire_count) {
1309 printf("vm_page_cache: attempting to cache busy page\n");
1310 return;
1311 }
1312 if ((m->queue - m->pc) == PQ_CACHE)
1313 return;
1314
1315 /*
1316 * Remove all pmaps and indicate that the page is not
1317 * writeable or mapped.
1318 */
1319 pmap_remove_all(m);
1320 if (m->dirty != 0) {
1321 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1322 (long)m->pindex);
1323 }
1324 vm_pageq_remove_nowakeup(m);
1325 vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1326 vm_page_free_wakeup();
1327 }
1328
1329 /*
1330 * vm_page_dontneed
1331 *
1332 * Cache, deactivate, or do nothing as appropriate. This routine
1333 * is typically used by madvise() MADV_DONTNEED.
1334 *
1335 * Generally speaking we want to move the page into the cache so
1336 * it gets reused quickly. However, this can result in a silly syndrome
1337 * due to the page recycling too quickly. Small objects will not be
1338 * fully cached. On the otherhand, if we move the page to the inactive
1339 * queue we wind up with a problem whereby very large objects
1340 * unnecessarily blow away our inactive and cache queues.
1341 *
1342 * The solution is to move the pages based on a fixed weighting. We
1343 * either leave them alone, deactivate them, or move them to the cache,
1344 * where moving them to the cache has the highest weighting.
1345 * By forcing some pages into other queues we eventually force the
1346 * system to balance the queues, potentially recovering other unrelated
1347 * space from active. The idea is to not force this to happen too
1348 * often.
1349 */
1350 void
1351 vm_page_dontneed(vm_page_t m)
1352 {
1353 static int dnweight;
1354 int dnw;
1355 int head;
1356
1357 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1358 dnw = ++dnweight;
1359
1360 /*
1361 * occassionally leave the page alone
1362 */
1363 if ((dnw & 0x01F0) == 0 ||
1364 m->queue == PQ_INACTIVE ||
1365 m->queue - m->pc == PQ_CACHE
1366 ) {
1367 if (m->act_count >= ACT_INIT)
1368 --m->act_count;
1369 return;
1370 }
1371
1372 if (m->dirty == 0 && pmap_is_modified(m))
1373 vm_page_dirty(m);
1374
1375 if (m->dirty || (dnw & 0x0070) == 0) {
1376 /*
1377 * Deactivate the page 3 times out of 32.
1378 */
1379 head = 0;
1380 } else {
1381 /*
1382 * Cache the page 28 times out of every 32. Note that
1383 * the page is deactivated instead of cached, but placed
1384 * at the head of the queue instead of the tail.
1385 */
1386 head = 1;
1387 }
1388 _vm_page_deactivate(m, head);
1389 }
1390
1391 /*
1392 * Grab a page, waiting until we are waken up due to the page
1393 * changing state. We keep on waiting, if the page continues
1394 * to be in the object. If the page doesn't exist, first allocate it
1395 * and then conditionally zero it.
1396 *
1397 * This routine may block.
1398 */
1399 vm_page_t
1400 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1401 {
1402 vm_page_t m;
1403
1404 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1405 retrylookup:
1406 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1407 vm_page_lock_queues();
1408 if (m->busy || (m->flags & PG_BUSY)) {
1409 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1410 VM_OBJECT_UNLOCK(object);
1411 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0);
1412 VM_OBJECT_LOCK(object);
1413 if ((allocflags & VM_ALLOC_RETRY) == 0)
1414 return (NULL);
1415 goto retrylookup;
1416 } else {
1417 if (allocflags & VM_ALLOC_WIRED)
1418 vm_page_wire(m);
1419 if ((allocflags & VM_ALLOC_NOBUSY) == 0)
1420 vm_page_busy(m);
1421 vm_page_unlock_queues();
1422 return (m);
1423 }
1424 }
1425 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1426 if (m == NULL) {
1427 VM_OBJECT_UNLOCK(object);
1428 VM_WAIT;
1429 VM_OBJECT_LOCK(object);
1430 if ((allocflags & VM_ALLOC_RETRY) == 0)
1431 return (NULL);
1432 goto retrylookup;
1433 }
1434 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
1435 pmap_zero_page(m);
1436 return (m);
1437 }
1438
1439 /*
1440 * Mapping function for valid bits or for dirty bits in
1441 * a page. May not block.
1442 *
1443 * Inputs are required to range within a page.
1444 */
1445 __inline int
1446 vm_page_bits(int base, int size)
1447 {
1448 int first_bit;
1449 int last_bit;
1450
1451 KASSERT(
1452 base + size <= PAGE_SIZE,
1453 ("vm_page_bits: illegal base/size %d/%d", base, size)
1454 );
1455
1456 if (size == 0) /* handle degenerate case */
1457 return (0);
1458
1459 first_bit = base >> DEV_BSHIFT;
1460 last_bit = (base + size - 1) >> DEV_BSHIFT;
1461
1462 return ((2 << last_bit) - (1 << first_bit));
1463 }
1464
1465 /*
1466 * vm_page_set_validclean:
1467 *
1468 * Sets portions of a page valid and clean. The arguments are expected
1469 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1470 * of any partial chunks touched by the range. The invalid portion of
1471 * such chunks will be zero'd.
1472 *
1473 * This routine may not block.
1474 *
1475 * (base + size) must be less then or equal to PAGE_SIZE.
1476 */
1477 void
1478 vm_page_set_validclean(vm_page_t m, int base, int size)
1479 {
1480 int pagebits;
1481 int frag;
1482 int endoff;
1483
1484 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1485 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1486 if (size == 0) /* handle degenerate case */
1487 return;
1488
1489 /*
1490 * If the base is not DEV_BSIZE aligned and the valid
1491 * bit is clear, we have to zero out a portion of the
1492 * first block.
1493 */
1494 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1495 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1496 pmap_zero_page_area(m, frag, base - frag);
1497
1498 /*
1499 * If the ending offset is not DEV_BSIZE aligned and the
1500 * valid bit is clear, we have to zero out a portion of
1501 * the last block.
1502 */
1503 endoff = base + size;
1504 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1505 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1506 pmap_zero_page_area(m, endoff,
1507 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1508
1509 /*
1510 * Set valid, clear dirty bits. If validating the entire
1511 * page we can safely clear the pmap modify bit. We also
1512 * use this opportunity to clear the PG_NOSYNC flag. If a process
1513 * takes a write fault on a MAP_NOSYNC memory area the flag will
1514 * be set again.
1515 *
1516 * We set valid bits inclusive of any overlap, but we can only
1517 * clear dirty bits for DEV_BSIZE chunks that are fully within
1518 * the range.
1519 */
1520 pagebits = vm_page_bits(base, size);
1521 m->valid |= pagebits;
1522 #if 0 /* NOT YET */
1523 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1524 frag = DEV_BSIZE - frag;
1525 base += frag;
1526 size -= frag;
1527 if (size < 0)
1528 size = 0;
1529 }
1530 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1531 #endif
1532 m->dirty &= ~pagebits;
1533 if (base == 0 && size == PAGE_SIZE) {
1534 pmap_clear_modify(m);
1535 vm_page_flag_clear(m, PG_NOSYNC);
1536 }
1537 }
1538
1539 void
1540 vm_page_clear_dirty(vm_page_t m, int base, int size)
1541 {
1542
1543 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1544 m->dirty &= ~vm_page_bits(base, size);
1545 }
1546
1547 /*
1548 * vm_page_set_invalid:
1549 *
1550 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1551 * valid and dirty bits for the effected areas are cleared.
1552 *
1553 * May not block.
1554 */
1555 void
1556 vm_page_set_invalid(vm_page_t m, int base, int size)
1557 {
1558 int bits;
1559
1560 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1561 bits = vm_page_bits(base, size);
1562 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1563 m->valid &= ~bits;
1564 m->dirty &= ~bits;
1565 m->object->generation++;
1566 }
1567
1568 /*
1569 * vm_page_zero_invalid()
1570 *
1571 * The kernel assumes that the invalid portions of a page contain
1572 * garbage, but such pages can be mapped into memory by user code.
1573 * When this occurs, we must zero out the non-valid portions of the
1574 * page so user code sees what it expects.
1575 *
1576 * Pages are most often semi-valid when the end of a file is mapped
1577 * into memory and the file's size is not page aligned.
1578 */
1579 void
1580 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1581 {
1582 int b;
1583 int i;
1584
1585 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1586 /*
1587 * Scan the valid bits looking for invalid sections that
1588 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1589 * valid bit may be set ) have already been zerod by
1590 * vm_page_set_validclean().
1591 */
1592 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1593 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1594 (m->valid & (1 << i))
1595 ) {
1596 if (i > b) {
1597 pmap_zero_page_area(m,
1598 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1599 }
1600 b = i + 1;
1601 }
1602 }
1603
1604 /*
1605 * setvalid is TRUE when we can safely set the zero'd areas
1606 * as being valid. We can do this if there are no cache consistancy
1607 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1608 */
1609 if (setvalid)
1610 m->valid = VM_PAGE_BITS_ALL;
1611 }
1612
1613 /*
1614 * vm_page_is_valid:
1615 *
1616 * Is (partial) page valid? Note that the case where size == 0
1617 * will return FALSE in the degenerate case where the page is
1618 * entirely invalid, and TRUE otherwise.
1619 *
1620 * May not block.
1621 */
1622 int
1623 vm_page_is_valid(vm_page_t m, int base, int size)
1624 {
1625 int bits = vm_page_bits(base, size);
1626
1627 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1628 if (m->valid && ((m->valid & bits) == bits))
1629 return 1;
1630 else
1631 return 0;
1632 }
1633
1634 /*
1635 * update dirty bits from pmap/mmu. May not block.
1636 */
1637 void
1638 vm_page_test_dirty(vm_page_t m)
1639 {
1640 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1641 vm_page_dirty(m);
1642 }
1643 }
1644
1645 int so_zerocp_fullpage = 0;
1646
1647 void
1648 vm_page_cowfault(vm_page_t m)
1649 {
1650 vm_page_t mnew;
1651 vm_object_t object;
1652 vm_pindex_t pindex;
1653
1654 object = m->object;
1655 pindex = m->pindex;
1656
1657 retry_alloc:
1658 vm_page_remove(m);
1659 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
1660 if (mnew == NULL) {
1661 vm_page_insert(m, object, pindex);
1662 vm_page_unlock_queues();
1663 VM_OBJECT_UNLOCK(object);
1664 VM_WAIT;
1665 VM_OBJECT_LOCK(object);
1666 vm_page_lock_queues();
1667 goto retry_alloc;
1668 }
1669
1670 if (m->cow == 0) {
1671 /*
1672 * check to see if we raced with an xmit complete when
1673 * waiting to allocate a page. If so, put things back
1674 * the way they were
1675 */
1676 vm_page_free(mnew);
1677 vm_page_insert(m, object, pindex);
1678 } else { /* clear COW & copy page */
1679 if (!so_zerocp_fullpage)
1680 pmap_copy_page(m, mnew);
1681 mnew->valid = VM_PAGE_BITS_ALL;
1682 vm_page_dirty(mnew);
1683 vm_page_flag_clear(mnew, PG_BUSY);
1684 }
1685 }
1686
1687 void
1688 vm_page_cowclear(vm_page_t m)
1689 {
1690
1691 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1692 if (m->cow) {
1693 m->cow--;
1694 /*
1695 * let vm_fault add back write permission lazily
1696 */
1697 }
1698 /*
1699 * sf_buf_free() will free the page, so we needn't do it here
1700 */
1701 }
1702
1703 void
1704 vm_page_cowsetup(vm_page_t m)
1705 {
1706
1707 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1708 m->cow++;
1709 pmap_page_protect(m, VM_PROT_READ);
1710 }
1711
1712 #include "opt_ddb.h"
1713 #ifdef DDB
1714 #include <sys/kernel.h>
1715
1716 #include <ddb/ddb.h>
1717
1718 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1719 {
1720 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1721 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1722 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1723 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1724 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1725 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1726 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1727 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1728 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1729 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1730 }
1731
1732 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1733 {
1734 int i;
1735 db_printf("PQ_FREE:");
1736 for (i = 0; i < PQ_L2_SIZE; i++) {
1737 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1738 }
1739 db_printf("\n");
1740
1741 db_printf("PQ_CACHE:");
1742 for (i = 0; i < PQ_L2_SIZE; i++) {
1743 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1744 }
1745 db_printf("\n");
1746
1747 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1748 vm_page_queues[PQ_ACTIVE].lcnt,
1749 vm_page_queues[PQ_INACTIVE].lcnt);
1750 }
1751 #endif /* DDB */
Cache object: 98ef0627452d13bd2de3d9b7fe6ad27f
|