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 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 */
38
39 /*
40 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
41 * All rights reserved.
42 *
43 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 *
45 * Permission to use, copy, modify and distribute this software and
46 * its documentation is hereby granted, provided that both the copyright
47 * notice and this permission notice appear in all copies of the
48 * software, derivative works or modified versions, and any portions
49 * thereof, and that both notices appear in supporting documentation.
50 *
51 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
52 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
53 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 *
55 * Carnegie Mellon requests users of this software to return to
56 *
57 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
58 * School of Computer Science
59 * Carnegie Mellon University
60 * Pittsburgh PA 15213-3890
61 *
62 * any improvements or extensions that they make and grant Carnegie the
63 * rights to redistribute these changes.
64 */
65
66 /*
67 * GENERAL RULES ON VM_PAGE MANIPULATION
68 *
69 * - a pageq mutex is required when adding or removing a page from a
70 * page queue (vm_page_queue[]), regardless of other mutexes or the
71 * busy state of a page.
72 *
73 * - a hash chain mutex is required when associating or disassociating
74 * a page from the VM PAGE CACHE hash table (vm_page_buckets),
75 * regardless of other mutexes or the busy state of a page.
76 *
77 * - either a hash chain mutex OR a busied page is required in order
78 * to modify the page flags. A hash chain mutex must be obtained in
79 * order to busy a page. A page's flags cannot be modified by a
80 * hash chain mutex if the page is marked busy.
81 *
82 * - The object memq mutex is held when inserting or removing
83 * pages from an object (vm_page_insert() or vm_page_remove()). This
84 * is different from the object's main mutex.
85 *
86 * Generally speaking, you have to be aware of side effects when running
87 * vm_page ops. A vm_page_lookup() will return with the hash chain
88 * locked, whether it was able to lookup the page or not. vm_page_free(),
89 * vm_page_cache(), vm_page_activate(), and a number of other routines
90 * will release the hash chain mutex for you. Intermediate manipulation
91 * routines such as vm_page_flag_set() expect the hash chain to be held
92 * on entry and the hash chain will remain held on return.
93 *
94 * pageq scanning can only occur with the pageq in question locked.
95 * We have a known bottleneck with the active queue, but the cache
96 * and free queues are actually arrays already.
97 */
98
99 /*
100 * Resident memory management module.
101 */
102
103 #include <sys/cdefs.h>
104 __FBSDID("$FreeBSD: releng/5.2/sys/vm/vm_page.c 121844 2003-11-01 04:54:23Z alc $");
105
106 #include <sys/param.h>
107 #include <sys/systm.h>
108 #include <sys/lock.h>
109 #include <sys/malloc.h>
110 #include <sys/mutex.h>
111 #include <sys/proc.h>
112 #include <sys/vmmeter.h>
113 #include <sys/vnode.h>
114
115 #include <vm/vm.h>
116 #include <vm/vm_param.h>
117 #include <vm/vm_kern.h>
118 #include <vm/vm_object.h>
119 #include <vm/vm_page.h>
120 #include <vm/vm_pageout.h>
121 #include <vm/vm_pager.h>
122 #include <vm/vm_extern.h>
123 #include <vm/uma.h>
124 #include <vm/uma_int.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 /*
140 * vm_set_page_size:
141 *
142 * Sets the page size, perhaps based upon the memory
143 * size. Must be called before any use of page-size
144 * dependent functions.
145 */
146 void
147 vm_set_page_size(void)
148 {
149 if (cnt.v_page_size == 0)
150 cnt.v_page_size = PAGE_SIZE;
151 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
152 panic("vm_set_page_size: page size not a power of two");
153 }
154
155 /*
156 * vm_page_startup:
157 *
158 * Initializes the resident memory module.
159 *
160 * Allocates memory for the page cells, and
161 * for the object/offset-to-page hash table headers.
162 * Each page cell is initialized and placed on the free list.
163 */
164 vm_offset_t
165 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
166 {
167 vm_offset_t mapped;
168 vm_size_t npages;
169 vm_paddr_t page_range;
170 vm_paddr_t new_end;
171 int i;
172 vm_paddr_t pa;
173 int nblocks;
174 vm_paddr_t last_pa;
175
176 /* the biggest memory array is the second group of pages */
177 vm_paddr_t end;
178 vm_paddr_t biggestsize;
179 int biggestone;
180
181 vm_paddr_t total;
182 vm_size_t bootpages;
183
184 total = 0;
185 biggestsize = 0;
186 biggestone = 0;
187 nblocks = 0;
188 vaddr = round_page(vaddr);
189
190 for (i = 0; phys_avail[i + 1]; i += 2) {
191 phys_avail[i] = round_page(phys_avail[i]);
192 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
193 }
194
195 for (i = 0; phys_avail[i + 1]; i += 2) {
196 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
197
198 if (size > biggestsize) {
199 biggestone = i;
200 biggestsize = size;
201 }
202 ++nblocks;
203 total += size;
204 }
205
206 end = phys_avail[biggestone+1];
207
208 /*
209 * Initialize the locks.
210 */
211 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF);
212 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
213 MTX_SPIN);
214
215 /*
216 * Initialize the queue headers for the free queue, the active queue
217 * and the inactive queue.
218 */
219 vm_pageq_init();
220
221 /*
222 * Allocate memory for use when boot strapping the kernel memory
223 * allocator.
224 */
225 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
226 new_end = end - bootpages;
227 new_end = trunc_page(new_end);
228 mapped = pmap_map(&vaddr, new_end, end,
229 VM_PROT_READ | VM_PROT_WRITE);
230 bzero((caddr_t) mapped, end - new_end);
231 uma_startup((caddr_t)mapped);
232
233 /*
234 * Compute the number of pages of memory that will be available for
235 * use (taking into account the overhead of a page structure per
236 * page).
237 */
238 first_page = phys_avail[0] / PAGE_SIZE;
239 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
240 npages = (total - (page_range * sizeof(struct vm_page)) -
241 (end - new_end)) / PAGE_SIZE;
242 end = new_end;
243
244 /*
245 * Initialize the mem entry structures now, and put them in the free
246 * queue.
247 */
248 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
249 mapped = pmap_map(&vaddr, new_end, end,
250 VM_PROT_READ | VM_PROT_WRITE);
251 vm_page_array = (vm_page_t) mapped;
252 phys_avail[biggestone + 1] = new_end;
253
254 /*
255 * Clear all of the page structures
256 */
257 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
258 vm_page_array_size = page_range;
259
260 /*
261 * Construct the free queue(s) in descending order (by physical
262 * address) so that the first 16MB of physical memory is allocated
263 * last rather than first. On large-memory machines, this avoids
264 * the exhaustion of low physical memory before isa_dmainit has run.
265 */
266 cnt.v_page_count = 0;
267 cnt.v_free_count = 0;
268 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
269 pa = phys_avail[i];
270 last_pa = phys_avail[i + 1];
271 while (pa < last_pa && npages-- > 0) {
272 vm_pageq_add_new_page(pa);
273 pa += PAGE_SIZE;
274 }
275 }
276 return (vaddr);
277 }
278
279 void
280 vm_page_flag_set(vm_page_t m, unsigned short bits)
281 {
282
283 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
284 m->flags |= bits;
285 }
286
287 void
288 vm_page_flag_clear(vm_page_t m, unsigned short bits)
289 {
290
291 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
292 m->flags &= ~bits;
293 }
294
295 void
296 vm_page_busy(vm_page_t m)
297 {
298 KASSERT((m->flags & PG_BUSY) == 0,
299 ("vm_page_busy: page already busy!!!"));
300 vm_page_flag_set(m, PG_BUSY);
301 }
302
303 /*
304 * vm_page_flash:
305 *
306 * wakeup anyone waiting for the page.
307 */
308 void
309 vm_page_flash(vm_page_t m)
310 {
311 if (m->flags & PG_WANTED) {
312 vm_page_flag_clear(m, PG_WANTED);
313 wakeup(m);
314 }
315 }
316
317 /*
318 * vm_page_wakeup:
319 *
320 * clear the PG_BUSY flag and wakeup anyone waiting for the
321 * page.
322 *
323 */
324 void
325 vm_page_wakeup(vm_page_t m)
326 {
327 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
328 vm_page_flag_clear(m, PG_BUSY);
329 vm_page_flash(m);
330 }
331
332 void
333 vm_page_io_start(vm_page_t m)
334 {
335
336 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
337 m->busy++;
338 }
339
340 void
341 vm_page_io_finish(vm_page_t m)
342 {
343
344 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
345 m->busy--;
346 if (m->busy == 0)
347 vm_page_flash(m);
348 }
349
350 /*
351 * Keep page from being freed by the page daemon
352 * much of the same effect as wiring, except much lower
353 * overhead and should be used only for *very* temporary
354 * holding ("wiring").
355 */
356 void
357 vm_page_hold(vm_page_t mem)
358 {
359
360 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
361 mem->hold_count++;
362 }
363
364 void
365 vm_page_unhold(vm_page_t mem)
366 {
367
368 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
369 --mem->hold_count;
370 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
371 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
372 vm_page_free_toq(mem);
373 }
374
375 /*
376 * vm_page_free:
377 *
378 * Free a page
379 *
380 * The clearing of PG_ZERO is a temporary safety until the code can be
381 * reviewed to determine that PG_ZERO is being properly cleared on
382 * write faults or maps. PG_ZERO was previously cleared in
383 * vm_page_alloc().
384 */
385 void
386 vm_page_free(vm_page_t m)
387 {
388 vm_page_flag_clear(m, PG_ZERO);
389 vm_page_free_toq(m);
390 vm_page_zero_idle_wakeup();
391 }
392
393 /*
394 * vm_page_free_zero:
395 *
396 * Free a page to the zerod-pages queue
397 */
398 void
399 vm_page_free_zero(vm_page_t m)
400 {
401 vm_page_flag_set(m, PG_ZERO);
402 vm_page_free_toq(m);
403 }
404
405 /*
406 * vm_page_sleep_if_busy:
407 *
408 * Sleep and release the page queues lock if PG_BUSY is set or,
409 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the
410 * thread slept and the page queues lock was released.
411 * Otherwise, retains the page queues lock and returns FALSE.
412 */
413 int
414 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
415 {
416 int is_object_locked;
417
418 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
419 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
420 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
421 /*
422 * Remove mtx_owned() after vm_object locking is finished.
423 */
424 if ((is_object_locked = m->object != NULL &&
425 mtx_owned(&m->object->mtx)))
426 mtx_unlock(&m->object->mtx);
427 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
428 if (is_object_locked)
429 mtx_lock(&m->object->mtx);
430 return (TRUE);
431 }
432 return (FALSE);
433 }
434
435 /*
436 * vm_page_dirty:
437 *
438 * make page all dirty
439 */
440 void
441 vm_page_dirty(vm_page_t m)
442 {
443 KASSERT(m->queue - m->pc != PQ_CACHE,
444 ("vm_page_dirty: page in cache!"));
445 KASSERT(m->queue - m->pc != PQ_FREE,
446 ("vm_page_dirty: page is free!"));
447 m->dirty = VM_PAGE_BITS_ALL;
448 }
449
450 /*
451 * vm_page_splay:
452 *
453 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
454 * the vm_page containing the given pindex. If, however, that
455 * pindex is not found in the vm_object, returns a vm_page that is
456 * adjacent to the pindex, coming before or after it.
457 */
458 vm_page_t
459 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
460 {
461 struct vm_page dummy;
462 vm_page_t lefttreemax, righttreemin, y;
463
464 if (root == NULL)
465 return (root);
466 lefttreemax = righttreemin = &dummy;
467 for (;; root = y) {
468 if (pindex < root->pindex) {
469 if ((y = root->left) == NULL)
470 break;
471 if (pindex < y->pindex) {
472 /* Rotate right. */
473 root->left = y->right;
474 y->right = root;
475 root = y;
476 if ((y = root->left) == NULL)
477 break;
478 }
479 /* Link into the new root's right tree. */
480 righttreemin->left = root;
481 righttreemin = root;
482 } else if (pindex > root->pindex) {
483 if ((y = root->right) == NULL)
484 break;
485 if (pindex > y->pindex) {
486 /* Rotate left. */
487 root->right = y->left;
488 y->left = root;
489 root = y;
490 if ((y = root->right) == NULL)
491 break;
492 }
493 /* Link into the new root's left tree. */
494 lefttreemax->right = root;
495 lefttreemax = root;
496 } else
497 break;
498 }
499 /* Assemble the new root. */
500 lefttreemax->right = root->left;
501 righttreemin->left = root->right;
502 root->left = dummy.right;
503 root->right = dummy.left;
504 return (root);
505 }
506
507 /*
508 * vm_page_insert: [ internal use only ]
509 *
510 * Inserts the given mem entry into the object and object list.
511 *
512 * The pagetables are not updated but will presumably fault the page
513 * in if necessary, or if a kernel page the caller will at some point
514 * enter the page into the kernel's pmap. We are not allowed to block
515 * here so we *can't* do this anyway.
516 *
517 * The object and page must be locked, and must be splhigh.
518 * This routine may not block.
519 */
520 void
521 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
522 {
523 vm_page_t root;
524
525 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
526 if (m->object != NULL)
527 panic("vm_page_insert: already inserted");
528
529 /*
530 * Record the object/offset pair in this page
531 */
532 m->object = object;
533 m->pindex = pindex;
534
535 /*
536 * Now link into the object's ordered list of backed pages.
537 */
538 root = object->root;
539 if (root == NULL) {
540 m->left = NULL;
541 m->right = NULL;
542 TAILQ_INSERT_TAIL(&object->memq, m, listq);
543 } else {
544 root = vm_page_splay(pindex, root);
545 if (pindex < root->pindex) {
546 m->left = root->left;
547 m->right = root;
548 root->left = NULL;
549 TAILQ_INSERT_BEFORE(root, m, listq);
550 } else {
551 m->right = root->right;
552 m->left = root;
553 root->right = NULL;
554 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
555 }
556 }
557 object->root = m;
558 object->generation++;
559
560 /*
561 * show that the object has one more resident page.
562 */
563 object->resident_page_count++;
564
565 /*
566 * Since we are inserting a new and possibly dirty page,
567 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
568 */
569 if (m->flags & PG_WRITEABLE)
570 vm_object_set_writeable_dirty(object);
571 }
572
573 /*
574 * vm_page_remove:
575 * NOTE: used by device pager as well -wfj
576 *
577 * Removes the given mem entry from the object/offset-page
578 * table and the object page list, but do not invalidate/terminate
579 * the backing store.
580 *
581 * The object and page must be locked, and at splhigh.
582 * The underlying pmap entry (if any) is NOT removed here.
583 * This routine may not block.
584 */
585 void
586 vm_page_remove(vm_page_t m)
587 {
588 vm_object_t object;
589 vm_page_t root;
590
591 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
592 if (m->object == NULL)
593 return;
594 #ifndef __alpha__
595 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
596 #endif
597 if ((m->flags & PG_BUSY) == 0) {
598 panic("vm_page_remove: page not busy");
599 }
600
601 /*
602 * Basically destroy the page.
603 */
604 vm_page_wakeup(m);
605
606 object = m->object;
607
608 /*
609 * Now remove from the object's list of backed pages.
610 */
611 if (m != object->root)
612 vm_page_splay(m->pindex, object->root);
613 if (m->left == NULL)
614 root = m->right;
615 else {
616 root = vm_page_splay(m->pindex, m->left);
617 root->right = m->right;
618 }
619 object->root = root;
620 TAILQ_REMOVE(&object->memq, m, listq);
621
622 /*
623 * And show that the object has one fewer resident page.
624 */
625 object->resident_page_count--;
626 object->generation++;
627
628 m->object = NULL;
629 }
630
631 /*
632 * vm_page_lookup:
633 *
634 * Returns the page associated with the object/offset
635 * pair specified; if none is found, NULL is returned.
636 *
637 * The object must be locked.
638 * This routine may not block.
639 * This is a critical path routine
640 */
641 vm_page_t
642 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
643 {
644 vm_page_t m;
645
646 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
647 m = vm_page_splay(pindex, object->root);
648 if ((object->root = m) != NULL && m->pindex != pindex)
649 m = NULL;
650 return (m);
651 }
652
653 /*
654 * vm_page_rename:
655 *
656 * Move the given memory entry from its
657 * current object to the specified target object/offset.
658 *
659 * The object must be locked.
660 * This routine may not block.
661 *
662 * Note: this routine will raise itself to splvm(), the caller need not.
663 *
664 * Note: swap associated with the page must be invalidated by the move. We
665 * have to do this for several reasons: (1) we aren't freeing the
666 * page, (2) we are dirtying the page, (3) the VM system is probably
667 * moving the page from object A to B, and will then later move
668 * the backing store from A to B and we can't have a conflict.
669 *
670 * Note: we *always* dirty the page. It is necessary both for the
671 * fact that we moved it, and because we may be invalidating
672 * swap. If the page is on the cache, we have to deactivate it
673 * or vm_page_dirty() will panic. Dirty pages are not allowed
674 * on the cache.
675 */
676 void
677 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
678 {
679 int s;
680
681 s = splvm();
682 vm_page_remove(m);
683 vm_page_insert(m, new_object, new_pindex);
684 if (m->queue - m->pc == PQ_CACHE)
685 vm_page_deactivate(m);
686 vm_page_dirty(m);
687 splx(s);
688 }
689
690 /*
691 * vm_page_select_cache:
692 *
693 * Find a page on the cache queue with color optimization. As pages
694 * might be found, but not applicable, they are deactivated. This
695 * keeps us from using potentially busy cached pages.
696 *
697 * This routine must be called at splvm().
698 * This routine may not block.
699 */
700 vm_page_t
701 vm_page_select_cache(int color)
702 {
703 vm_page_t m;
704
705 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
706 while (TRUE) {
707 m = vm_pageq_find(PQ_CACHE, color, FALSE);
708 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
709 m->hold_count || m->wire_count ||
710 (!VM_OBJECT_TRYLOCK(m->object) &&
711 !VM_OBJECT_LOCKED(m->object)))) {
712 vm_page_deactivate(m);
713 continue;
714 }
715 return m;
716 }
717 }
718
719 /*
720 * vm_page_alloc:
721 *
722 * Allocate and return a memory cell associated
723 * with this VM object/offset pair.
724 *
725 * page_req classes:
726 * VM_ALLOC_NORMAL normal process request
727 * VM_ALLOC_SYSTEM system *really* needs a page
728 * VM_ALLOC_INTERRUPT interrupt time request
729 * VM_ALLOC_ZERO zero page
730 *
731 * This routine may not block.
732 *
733 * Additional special handling is required when called from an
734 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with
735 * the page cache in this case.
736 */
737 vm_page_t
738 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
739 {
740 vm_object_t m_object;
741 vm_page_t m = NULL;
742 int color, flags, page_req, s;
743
744 page_req = req & VM_ALLOC_CLASS_MASK;
745
746 if ((req & VM_ALLOC_NOOBJ) == 0) {
747 KASSERT(object != NULL,
748 ("vm_page_alloc: NULL object."));
749 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
750 KASSERT(!vm_page_lookup(object, pindex),
751 ("vm_page_alloc: page already allocated"));
752 color = (pindex + object->pg_color) & PQ_L2_MASK;
753 } else
754 color = pindex & PQ_L2_MASK;
755
756 /*
757 * The pager is allowed to eat deeper into the free page list.
758 */
759 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
760 page_req = VM_ALLOC_SYSTEM;
761 };
762
763 s = splvm();
764 loop:
765 mtx_lock_spin(&vm_page_queue_free_mtx);
766 if (cnt.v_free_count > cnt.v_free_reserved ||
767 (page_req == VM_ALLOC_SYSTEM &&
768 cnt.v_cache_count == 0 &&
769 cnt.v_free_count > cnt.v_interrupt_free_min) ||
770 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) {
771 /*
772 * Allocate from the free queue if the number of free pages
773 * exceeds the minimum for the request class.
774 */
775 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0);
776 } else if (page_req != VM_ALLOC_INTERRUPT) {
777 mtx_unlock_spin(&vm_page_queue_free_mtx);
778 /*
779 * Allocatable from cache (non-interrupt only). On success,
780 * we must free the page and try again, thus ensuring that
781 * cnt.v_*_free_min counters are replenished.
782 */
783 vm_page_lock_queues();
784 if ((m = vm_page_select_cache(color)) == NULL) {
785 vm_page_unlock_queues();
786 splx(s);
787 #if defined(DIAGNOSTIC)
788 if (cnt.v_cache_count > 0)
789 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
790 #endif
791 atomic_add_int(&vm_pageout_deficit, 1);
792 pagedaemon_wakeup();
793 return (NULL);
794 }
795 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
796 m_object = m->object;
797 VM_OBJECT_LOCK_ASSERT(m_object, MA_OWNED);
798 vm_page_busy(m);
799 pmap_remove_all(m);
800 vm_page_free(m);
801 vm_page_unlock_queues();
802 if (m_object != object)
803 VM_OBJECT_UNLOCK(m_object);
804 goto loop;
805 } else {
806 /*
807 * Not allocatable from cache from interrupt, give up.
808 */
809 mtx_unlock_spin(&vm_page_queue_free_mtx);
810 splx(s);
811 atomic_add_int(&vm_pageout_deficit, 1);
812 pagedaemon_wakeup();
813 return (NULL);
814 }
815
816 /*
817 * At this point we had better have found a good page.
818 */
819
820 KASSERT(
821 m != NULL,
822 ("vm_page_alloc(): missing page on free queue\n")
823 );
824
825 /*
826 * Remove from free queue
827 */
828
829 vm_pageq_remove_nowakeup(m);
830
831 /*
832 * Initialize structure. Only the PG_ZERO flag is inherited.
833 */
834 flags = PG_BUSY;
835 if (m->flags & PG_ZERO) {
836 vm_page_zero_count--;
837 if (req & VM_ALLOC_ZERO)
838 flags = PG_ZERO | PG_BUSY;
839 }
840 m->flags = flags;
841 if (req & VM_ALLOC_WIRED) {
842 atomic_add_int(&cnt.v_wire_count, 1);
843 m->wire_count = 1;
844 } else
845 m->wire_count = 0;
846 m->hold_count = 0;
847 m->act_count = 0;
848 m->busy = 0;
849 m->valid = 0;
850 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
851 mtx_unlock_spin(&vm_page_queue_free_mtx);
852
853 /*
854 * vm_page_insert() is safe prior to the splx(). Note also that
855 * inserting a page here does not insert it into the pmap (which
856 * could cause us to block allocating memory). We cannot block
857 * anywhere.
858 */
859 if ((req & VM_ALLOC_NOOBJ) == 0)
860 vm_page_insert(m, object, pindex);
861 else
862 m->pindex = pindex;
863
864 /*
865 * Don't wakeup too often - wakeup the pageout daemon when
866 * we would be nearly out of memory.
867 */
868 if (vm_paging_needed())
869 pagedaemon_wakeup();
870
871 splx(s);
872 return (m);
873 }
874
875 /*
876 * vm_wait: (also see VM_WAIT macro)
877 *
878 * Block until free pages are available for allocation
879 * - Called in various places before memory allocations.
880 */
881 void
882 vm_wait(void)
883 {
884 int s;
885
886 s = splvm();
887 vm_page_lock_queues();
888 if (curproc == pageproc) {
889 vm_pageout_pages_needed = 1;
890 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx,
891 PDROP | PSWP, "VMWait", 0);
892 } else {
893 if (!vm_pages_needed) {
894 vm_pages_needed = 1;
895 wakeup(&vm_pages_needed);
896 }
897 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM,
898 "vmwait", 0);
899 }
900 splx(s);
901 }
902
903 /*
904 * vm_waitpfault: (also see VM_WAITPFAULT macro)
905 *
906 * Block until free pages are available for allocation
907 * - Called only in vm_fault so that processes page faulting
908 * can be easily tracked.
909 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
910 * processes will be able to grab memory first. Do not change
911 * this balance without careful testing first.
912 */
913 void
914 vm_waitpfault(void)
915 {
916 int s;
917
918 s = splvm();
919 vm_page_lock_queues();
920 if (!vm_pages_needed) {
921 vm_pages_needed = 1;
922 wakeup(&vm_pages_needed);
923 }
924 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER,
925 "pfault", 0);
926 splx(s);
927 }
928
929 /*
930 * vm_page_activate:
931 *
932 * Put the specified page on the active list (if appropriate).
933 * Ensure that act_count is at least ACT_INIT but do not otherwise
934 * mess with it.
935 *
936 * The page queues must be locked.
937 * This routine may not block.
938 */
939 void
940 vm_page_activate(vm_page_t m)
941 {
942 int s;
943
944 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
945 s = splvm();
946 if (m->queue != PQ_ACTIVE) {
947 if ((m->queue - m->pc) == PQ_CACHE)
948 cnt.v_reactivated++;
949 vm_pageq_remove(m);
950 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
951 if (m->act_count < ACT_INIT)
952 m->act_count = ACT_INIT;
953 vm_pageq_enqueue(PQ_ACTIVE, m);
954 }
955 } else {
956 if (m->act_count < ACT_INIT)
957 m->act_count = ACT_INIT;
958 }
959 splx(s);
960 }
961
962 /*
963 * vm_page_free_wakeup:
964 *
965 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
966 * routine is called when a page has been added to the cache or free
967 * queues.
968 *
969 * This routine may not block.
970 * This routine must be called at splvm()
971 */
972 static __inline void
973 vm_page_free_wakeup(void)
974 {
975
976 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
977 /*
978 * if pageout daemon needs pages, then tell it that there are
979 * some free.
980 */
981 if (vm_pageout_pages_needed &&
982 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
983 wakeup(&vm_pageout_pages_needed);
984 vm_pageout_pages_needed = 0;
985 }
986 /*
987 * wakeup processes that are waiting on memory if we hit a
988 * high water mark. And wakeup scheduler process if we have
989 * lots of memory. this process will swapin processes.
990 */
991 if (vm_pages_needed && !vm_page_count_min()) {
992 vm_pages_needed = 0;
993 wakeup(&cnt.v_free_count);
994 }
995 }
996
997 /*
998 * vm_page_free_toq:
999 *
1000 * Returns the given page to the PQ_FREE list,
1001 * disassociating it with any VM object.
1002 *
1003 * Object and page must be locked prior to entry.
1004 * This routine may not block.
1005 */
1006
1007 void
1008 vm_page_free_toq(vm_page_t m)
1009 {
1010 int s;
1011 struct vpgqueues *pq;
1012 vm_object_t object = m->object;
1013
1014 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1015 s = splvm();
1016 cnt.v_tfree++;
1017
1018 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1019 printf(
1020 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1021 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1022 m->hold_count);
1023 if ((m->queue - m->pc) == PQ_FREE)
1024 panic("vm_page_free: freeing free page");
1025 else
1026 panic("vm_page_free: freeing busy page");
1027 }
1028
1029 /*
1030 * unqueue, then remove page. Note that we cannot destroy
1031 * the page here because we do not want to call the pager's
1032 * callback routine until after we've put the page on the
1033 * appropriate free queue.
1034 */
1035 vm_pageq_remove_nowakeup(m);
1036 vm_page_remove(m);
1037
1038 /*
1039 * If fictitious remove object association and
1040 * return, otherwise delay object association removal.
1041 */
1042 if ((m->flags & PG_FICTITIOUS) != 0) {
1043 splx(s);
1044 return;
1045 }
1046
1047 m->valid = 0;
1048 vm_page_undirty(m);
1049
1050 if (m->wire_count != 0) {
1051 if (m->wire_count > 1) {
1052 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1053 m->wire_count, (long)m->pindex);
1054 }
1055 panic("vm_page_free: freeing wired page\n");
1056 }
1057
1058 /*
1059 * If we've exhausted the object's resident pages we want to free
1060 * it up.
1061 */
1062 if (object &&
1063 (object->type == OBJT_VNODE) &&
1064 ((object->flags & OBJ_DEAD) == 0)
1065 ) {
1066 struct vnode *vp = (struct vnode *)object->handle;
1067
1068 if (vp) {
1069 VI_LOCK(vp);
1070 if (VSHOULDFREE(vp))
1071 vfree(vp);
1072 VI_UNLOCK(vp);
1073 }
1074 }
1075
1076 /*
1077 * Clear the UNMANAGED flag when freeing an unmanaged page.
1078 */
1079 if (m->flags & PG_UNMANAGED) {
1080 m->flags &= ~PG_UNMANAGED;
1081 }
1082
1083 if (m->hold_count != 0) {
1084 m->flags &= ~PG_ZERO;
1085 m->queue = PQ_HOLD;
1086 } else
1087 m->queue = PQ_FREE + m->pc;
1088 pq = &vm_page_queues[m->queue];
1089 mtx_lock_spin(&vm_page_queue_free_mtx);
1090 pq->lcnt++;
1091 ++(*pq->cnt);
1092
1093 /*
1094 * Put zero'd pages on the end ( where we look for zero'd pages
1095 * first ) and non-zerod pages at the head.
1096 */
1097 if (m->flags & PG_ZERO) {
1098 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1099 ++vm_page_zero_count;
1100 } else {
1101 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1102 }
1103 mtx_unlock_spin(&vm_page_queue_free_mtx);
1104 vm_page_free_wakeup();
1105 splx(s);
1106 }
1107
1108 /*
1109 * vm_page_unmanage:
1110 *
1111 * Prevent PV management from being done on the page. The page is
1112 * removed from the paging queues as if it were wired, and as a
1113 * consequence of no longer being managed the pageout daemon will not
1114 * touch it (since there is no way to locate the pte mappings for the
1115 * page). madvise() calls that mess with the pmap will also no longer
1116 * operate on the page.
1117 *
1118 * Beyond that the page is still reasonably 'normal'. Freeing the page
1119 * will clear the flag.
1120 *
1121 * This routine is used by OBJT_PHYS objects - objects using unswappable
1122 * physical memory as backing store rather then swap-backed memory and
1123 * will eventually be extended to support 4MB unmanaged physical
1124 * mappings.
1125 */
1126 void
1127 vm_page_unmanage(vm_page_t m)
1128 {
1129 int s;
1130
1131 s = splvm();
1132 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1133 if ((m->flags & PG_UNMANAGED) == 0) {
1134 if (m->wire_count == 0)
1135 vm_pageq_remove(m);
1136 }
1137 vm_page_flag_set(m, PG_UNMANAGED);
1138 splx(s);
1139 }
1140
1141 /*
1142 * vm_page_wire:
1143 *
1144 * Mark this page as wired down by yet
1145 * another map, removing it from paging queues
1146 * as necessary.
1147 *
1148 * The page queues must be locked.
1149 * This routine may not block.
1150 */
1151 void
1152 vm_page_wire(vm_page_t m)
1153 {
1154 int s;
1155
1156 /*
1157 * Only bump the wire statistics if the page is not already wired,
1158 * and only unqueue the page if it is on some queue (if it is unmanaged
1159 * it is already off the queues).
1160 */
1161 s = splvm();
1162 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1163 if (m->wire_count == 0) {
1164 if ((m->flags & PG_UNMANAGED) == 0)
1165 vm_pageq_remove(m);
1166 atomic_add_int(&cnt.v_wire_count, 1);
1167 }
1168 m->wire_count++;
1169 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1170 splx(s);
1171 }
1172
1173 /*
1174 * vm_page_unwire:
1175 *
1176 * Release one wiring of this page, potentially
1177 * enabling it to be paged again.
1178 *
1179 * Many pages placed on the inactive queue should actually go
1180 * into the cache, but it is difficult to figure out which. What
1181 * we do instead, if the inactive target is well met, is to put
1182 * clean pages at the head of the inactive queue instead of the tail.
1183 * This will cause them to be moved to the cache more quickly and
1184 * if not actively re-referenced, freed more quickly. If we just
1185 * stick these pages at the end of the inactive queue, heavy filesystem
1186 * meta-data accesses can cause an unnecessary paging load on memory bound
1187 * processes. This optimization causes one-time-use metadata to be
1188 * reused more quickly.
1189 *
1190 * BUT, if we are in a low-memory situation we have no choice but to
1191 * put clean pages on the cache queue.
1192 *
1193 * A number of routines use vm_page_unwire() to guarantee that the page
1194 * will go into either the inactive or active queues, and will NEVER
1195 * be placed in the cache - for example, just after dirtying a page.
1196 * dirty pages in the cache are not allowed.
1197 *
1198 * The page queues must be locked.
1199 * This routine may not block.
1200 */
1201 void
1202 vm_page_unwire(vm_page_t m, int activate)
1203 {
1204 int s;
1205
1206 s = splvm();
1207 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1208 if (m->wire_count > 0) {
1209 m->wire_count--;
1210 if (m->wire_count == 0) {
1211 atomic_subtract_int(&cnt.v_wire_count, 1);
1212 if (m->flags & PG_UNMANAGED) {
1213 ;
1214 } else if (activate)
1215 vm_pageq_enqueue(PQ_ACTIVE, m);
1216 else {
1217 vm_page_flag_clear(m, PG_WINATCFLS);
1218 vm_pageq_enqueue(PQ_INACTIVE, m);
1219 }
1220 }
1221 } else {
1222 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1223 }
1224 splx(s);
1225 }
1226
1227
1228 /*
1229 * Move the specified page to the inactive queue. If the page has
1230 * any associated swap, the swap is deallocated.
1231 *
1232 * Normally athead is 0 resulting in LRU operation. athead is set
1233 * to 1 if we want this page to be 'as if it were placed in the cache',
1234 * except without unmapping it from the process address space.
1235 *
1236 * This routine may not block.
1237 */
1238 static __inline void
1239 _vm_page_deactivate(vm_page_t m, int athead)
1240 {
1241 int s;
1242
1243 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1244 /*
1245 * Ignore if already inactive.
1246 */
1247 if (m->queue == PQ_INACTIVE)
1248 return;
1249
1250 s = splvm();
1251 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1252 if ((m->queue - m->pc) == PQ_CACHE)
1253 cnt.v_reactivated++;
1254 vm_page_flag_clear(m, PG_WINATCFLS);
1255 vm_pageq_remove(m);
1256 if (athead)
1257 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1258 else
1259 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1260 m->queue = PQ_INACTIVE;
1261 vm_page_queues[PQ_INACTIVE].lcnt++;
1262 cnt.v_inactive_count++;
1263 }
1264 splx(s);
1265 }
1266
1267 void
1268 vm_page_deactivate(vm_page_t m)
1269 {
1270 _vm_page_deactivate(m, 0);
1271 }
1272
1273 /*
1274 * vm_page_try_to_cache:
1275 *
1276 * Returns 0 on failure, 1 on success
1277 */
1278 int
1279 vm_page_try_to_cache(vm_page_t m)
1280 {
1281
1282 mtx_assert(&vm_page_queue_mtx, 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 vm_page_test_dirty(m);
1288 if (m->dirty)
1289 return (0);
1290 vm_page_cache(m);
1291 return (1);
1292 }
1293
1294 /*
1295 * vm_page_try_to_free()
1296 *
1297 * Attempt to free the page. If we cannot free it, we do nothing.
1298 * 1 is returned on success, 0 on failure.
1299 */
1300 int
1301 vm_page_try_to_free(vm_page_t m)
1302 {
1303
1304 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1305 if (m->object != NULL)
1306 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1307 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1308 (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1309 return (0);
1310 }
1311 vm_page_test_dirty(m);
1312 if (m->dirty)
1313 return (0);
1314 vm_page_busy(m);
1315 pmap_remove_all(m);
1316 vm_page_free(m);
1317 return (1);
1318 }
1319
1320 /*
1321 * vm_page_cache
1322 *
1323 * Put the specified page onto the page cache queue (if appropriate).
1324 *
1325 * This routine may not block.
1326 */
1327 void
1328 vm_page_cache(vm_page_t m)
1329 {
1330 int s;
1331
1332 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1333 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1334 m->hold_count || m->wire_count) {
1335 printf("vm_page_cache: attempting to cache busy page\n");
1336 return;
1337 }
1338 if ((m->queue - m->pc) == PQ_CACHE)
1339 return;
1340
1341 /*
1342 * Remove all pmaps and indicate that the page is not
1343 * writeable or mapped.
1344 */
1345 pmap_remove_all(m);
1346 if (m->dirty != 0) {
1347 panic("vm_page_cache: caching a dirty page, pindex: %ld",
1348 (long)m->pindex);
1349 }
1350 s = splvm();
1351 vm_pageq_remove_nowakeup(m);
1352 vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1353 vm_page_free_wakeup();
1354 splx(s);
1355 }
1356
1357 /*
1358 * vm_page_dontneed
1359 *
1360 * Cache, deactivate, or do nothing as appropriate. This routine
1361 * is typically used by madvise() MADV_DONTNEED.
1362 *
1363 * Generally speaking we want to move the page into the cache so
1364 * it gets reused quickly. However, this can result in a silly syndrome
1365 * due to the page recycling too quickly. Small objects will not be
1366 * fully cached. On the otherhand, if we move the page to the inactive
1367 * queue we wind up with a problem whereby very large objects
1368 * unnecessarily blow away our inactive and cache queues.
1369 *
1370 * The solution is to move the pages based on a fixed weighting. We
1371 * either leave them alone, deactivate them, or move them to the cache,
1372 * where moving them to the cache has the highest weighting.
1373 * By forcing some pages into other queues we eventually force the
1374 * system to balance the queues, potentially recovering other unrelated
1375 * space from active. The idea is to not force this to happen too
1376 * often.
1377 */
1378 void
1379 vm_page_dontneed(vm_page_t m)
1380 {
1381 static int dnweight;
1382 int dnw;
1383 int head;
1384
1385 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1386 dnw = ++dnweight;
1387
1388 /*
1389 * occassionally leave the page alone
1390 */
1391 if ((dnw & 0x01F0) == 0 ||
1392 m->queue == PQ_INACTIVE ||
1393 m->queue - m->pc == PQ_CACHE
1394 ) {
1395 if (m->act_count >= ACT_INIT)
1396 --m->act_count;
1397 return;
1398 }
1399
1400 if (m->dirty == 0)
1401 vm_page_test_dirty(m);
1402
1403 if (m->dirty || (dnw & 0x0070) == 0) {
1404 /*
1405 * Deactivate the page 3 times out of 32.
1406 */
1407 head = 0;
1408 } else {
1409 /*
1410 * Cache the page 28 times out of every 32. Note that
1411 * the page is deactivated instead of cached, but placed
1412 * at the head of the queue instead of the tail.
1413 */
1414 head = 1;
1415 }
1416 _vm_page_deactivate(m, head);
1417 }
1418
1419 /*
1420 * Grab a page, waiting until we are waken up due to the page
1421 * changing state. We keep on waiting, if the page continues
1422 * to be in the object. If the page doesn't exist, allocate it.
1423 *
1424 * This routine may block.
1425 */
1426 vm_page_t
1427 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1428 {
1429 vm_page_t m;
1430 int s, generation;
1431
1432 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1433 retrylookup:
1434 if ((m = vm_page_lookup(object, pindex)) != NULL) {
1435 vm_page_lock_queues();
1436 if (m->busy || (m->flags & PG_BUSY)) {
1437 generation = object->generation;
1438
1439 s = splvm();
1440 while ((object->generation == generation) &&
1441 (m->busy || (m->flags & PG_BUSY))) {
1442 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1443 VM_OBJECT_UNLOCK(object);
1444 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0);
1445 VM_OBJECT_LOCK(object);
1446 if ((allocflags & VM_ALLOC_RETRY) == 0) {
1447 splx(s);
1448 return NULL;
1449 }
1450 vm_page_lock_queues();
1451 }
1452 vm_page_unlock_queues();
1453 splx(s);
1454 goto retrylookup;
1455 } else {
1456 if (allocflags & VM_ALLOC_WIRED)
1457 vm_page_wire(m);
1458 vm_page_busy(m);
1459 vm_page_unlock_queues();
1460 return m;
1461 }
1462 }
1463
1464 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1465 if (m == NULL) {
1466 VM_OBJECT_UNLOCK(object);
1467 VM_WAIT;
1468 VM_OBJECT_LOCK(object);
1469 if ((allocflags & VM_ALLOC_RETRY) == 0)
1470 return NULL;
1471 goto retrylookup;
1472 }
1473
1474 return m;
1475 }
1476
1477 /*
1478 * Mapping function for valid bits or for dirty bits in
1479 * a page. May not block.
1480 *
1481 * Inputs are required to range within a page.
1482 */
1483 __inline int
1484 vm_page_bits(int base, int size)
1485 {
1486 int first_bit;
1487 int last_bit;
1488
1489 KASSERT(
1490 base + size <= PAGE_SIZE,
1491 ("vm_page_bits: illegal base/size %d/%d", base, size)
1492 );
1493
1494 if (size == 0) /* handle degenerate case */
1495 return (0);
1496
1497 first_bit = base >> DEV_BSHIFT;
1498 last_bit = (base + size - 1) >> DEV_BSHIFT;
1499
1500 return ((2 << last_bit) - (1 << first_bit));
1501 }
1502
1503 /*
1504 * vm_page_set_validclean:
1505 *
1506 * Sets portions of a page valid and clean. The arguments are expected
1507 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1508 * of any partial chunks touched by the range. The invalid portion of
1509 * such chunks will be zero'd.
1510 *
1511 * This routine may not block.
1512 *
1513 * (base + size) must be less then or equal to PAGE_SIZE.
1514 */
1515 void
1516 vm_page_set_validclean(vm_page_t m, int base, int size)
1517 {
1518 int pagebits;
1519 int frag;
1520 int endoff;
1521
1522 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1523 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1524 if (size == 0) /* handle degenerate case */
1525 return;
1526
1527 /*
1528 * If the base is not DEV_BSIZE aligned and the valid
1529 * bit is clear, we have to zero out a portion of the
1530 * first block.
1531 */
1532 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1533 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1534 pmap_zero_page_area(m, frag, base - frag);
1535
1536 /*
1537 * If the ending offset is not DEV_BSIZE aligned and the
1538 * valid bit is clear, we have to zero out a portion of
1539 * the last block.
1540 */
1541 endoff = base + size;
1542 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1543 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1544 pmap_zero_page_area(m, endoff,
1545 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1546
1547 /*
1548 * Set valid, clear dirty bits. If validating the entire
1549 * page we can safely clear the pmap modify bit. We also
1550 * use this opportunity to clear the PG_NOSYNC flag. If a process
1551 * takes a write fault on a MAP_NOSYNC memory area the flag will
1552 * be set again.
1553 *
1554 * We set valid bits inclusive of any overlap, but we can only
1555 * clear dirty bits for DEV_BSIZE chunks that are fully within
1556 * the range.
1557 */
1558 pagebits = vm_page_bits(base, size);
1559 m->valid |= pagebits;
1560 #if 0 /* NOT YET */
1561 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1562 frag = DEV_BSIZE - frag;
1563 base += frag;
1564 size -= frag;
1565 if (size < 0)
1566 size = 0;
1567 }
1568 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1569 #endif
1570 m->dirty &= ~pagebits;
1571 if (base == 0 && size == PAGE_SIZE) {
1572 pmap_clear_modify(m);
1573 vm_page_flag_clear(m, PG_NOSYNC);
1574 }
1575 }
1576
1577 void
1578 vm_page_clear_dirty(vm_page_t m, int base, int size)
1579 {
1580
1581 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1582 m->dirty &= ~vm_page_bits(base, size);
1583 }
1584
1585 /*
1586 * vm_page_set_invalid:
1587 *
1588 * Invalidates DEV_BSIZE'd chunks within a page. Both the
1589 * valid and dirty bits for the effected areas are cleared.
1590 *
1591 * May not block.
1592 */
1593 void
1594 vm_page_set_invalid(vm_page_t m, int base, int size)
1595 {
1596 int bits;
1597
1598 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1599 bits = vm_page_bits(base, size);
1600 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1601 m->valid &= ~bits;
1602 m->dirty &= ~bits;
1603 m->object->generation++;
1604 }
1605
1606 /*
1607 * vm_page_zero_invalid()
1608 *
1609 * The kernel assumes that the invalid portions of a page contain
1610 * garbage, but such pages can be mapped into memory by user code.
1611 * When this occurs, we must zero out the non-valid portions of the
1612 * page so user code sees what it expects.
1613 *
1614 * Pages are most often semi-valid when the end of a file is mapped
1615 * into memory and the file's size is not page aligned.
1616 */
1617 void
1618 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1619 {
1620 int b;
1621 int i;
1622
1623 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1624 /*
1625 * Scan the valid bits looking for invalid sections that
1626 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
1627 * valid bit may be set ) have already been zerod by
1628 * vm_page_set_validclean().
1629 */
1630 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1631 if (i == (PAGE_SIZE / DEV_BSIZE) ||
1632 (m->valid & (1 << i))
1633 ) {
1634 if (i > b) {
1635 pmap_zero_page_area(m,
1636 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1637 }
1638 b = i + 1;
1639 }
1640 }
1641
1642 /*
1643 * setvalid is TRUE when we can safely set the zero'd areas
1644 * as being valid. We can do this if there are no cache consistancy
1645 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
1646 */
1647 if (setvalid)
1648 m->valid = VM_PAGE_BITS_ALL;
1649 }
1650
1651 /*
1652 * vm_page_is_valid:
1653 *
1654 * Is (partial) page valid? Note that the case where size == 0
1655 * will return FALSE in the degenerate case where the page is
1656 * entirely invalid, and TRUE otherwise.
1657 *
1658 * May not block.
1659 */
1660 int
1661 vm_page_is_valid(vm_page_t m, int base, int size)
1662 {
1663 int bits = vm_page_bits(base, size);
1664
1665 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1666 if (m->valid && ((m->valid & bits) == bits))
1667 return 1;
1668 else
1669 return 0;
1670 }
1671
1672 /*
1673 * update dirty bits from pmap/mmu. May not block.
1674 */
1675 void
1676 vm_page_test_dirty(vm_page_t m)
1677 {
1678 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1679 vm_page_dirty(m);
1680 }
1681 }
1682
1683 int so_zerocp_fullpage = 0;
1684
1685 void
1686 vm_page_cowfault(vm_page_t m)
1687 {
1688 vm_page_t mnew;
1689 vm_object_t object;
1690 vm_pindex_t pindex;
1691
1692 object = m->object;
1693 pindex = m->pindex;
1694 vm_page_busy(m);
1695
1696 retry_alloc:
1697 vm_page_remove(m);
1698 /*
1699 * An interrupt allocation is requested because the page
1700 * queues lock is held.
1701 */
1702 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT);
1703 if (mnew == NULL) {
1704 vm_page_insert(m, object, pindex);
1705 vm_page_unlock_queues();
1706 VM_OBJECT_UNLOCK(object);
1707 VM_WAIT;
1708 VM_OBJECT_LOCK(object);
1709 vm_page_lock_queues();
1710 goto retry_alloc;
1711 }
1712
1713 if (m->cow == 0) {
1714 /*
1715 * check to see if we raced with an xmit complete when
1716 * waiting to allocate a page. If so, put things back
1717 * the way they were
1718 */
1719 vm_page_busy(mnew);
1720 vm_page_free(mnew);
1721 vm_page_insert(m, object, pindex);
1722 } else { /* clear COW & copy page */
1723 if (!so_zerocp_fullpage)
1724 pmap_copy_page(m, mnew);
1725 mnew->valid = VM_PAGE_BITS_ALL;
1726 vm_page_dirty(mnew);
1727 vm_page_flag_clear(mnew, PG_BUSY);
1728 }
1729 }
1730
1731 void
1732 vm_page_cowclear(vm_page_t m)
1733 {
1734
1735 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1736 if (m->cow) {
1737 m->cow--;
1738 /*
1739 * let vm_fault add back write permission lazily
1740 */
1741 }
1742 /*
1743 * sf_buf_free() will free the page, so we needn't do it here
1744 */
1745 }
1746
1747 void
1748 vm_page_cowsetup(vm_page_t m)
1749 {
1750
1751 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1752 m->cow++;
1753 pmap_page_protect(m, VM_PROT_READ);
1754 }
1755
1756 #include "opt_ddb.h"
1757 #ifdef DDB
1758 #include <sys/kernel.h>
1759
1760 #include <ddb/ddb.h>
1761
1762 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1763 {
1764 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1765 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1766 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1767 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1768 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1769 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1770 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1771 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1772 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1773 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1774 }
1775
1776 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1777 {
1778 int i;
1779 db_printf("PQ_FREE:");
1780 for (i = 0; i < PQ_L2_SIZE; i++) {
1781 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1782 }
1783 db_printf("\n");
1784
1785 db_printf("PQ_CACHE:");
1786 for (i = 0; i < PQ_L2_SIZE; i++) {
1787 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1788 }
1789 db_printf("\n");
1790
1791 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1792 vm_page_queues[PQ_ACTIVE].lcnt,
1793 vm_page_queues[PQ_INACTIVE].lcnt);
1794 }
1795 #endif /* DDB */
Cache object: 6ba888cc45b16d7a7b197626487ffb5b
|