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 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
5 *
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 *
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
34 */
35
36 /*-
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
39 *
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41 *
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
47 *
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51 *
52 * Carnegie Mellon requests users of this software to return to
53 *
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
58 *
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
61 */
62
63 /*
64 * GENERAL RULES ON VM_PAGE MANIPULATION
65 *
66 * - a pageq mutex is required when adding or removing a page from a
67 * page queue (vm_page_queue[]), regardless of other mutexes or the
68 * busy state of a page.
69 *
70 * - The object mutex is held when inserting or removing
71 * pages from an object (vm_page_insert() or vm_page_remove()).
72 *
73 */
74
75 /*
76 * Resident memory management module.
77 */
78
79 #include <sys/cdefs.h>
80 __FBSDID("$FreeBSD: releng/9.2/sys/vm/vm_page.c 251450 2013-06-06 05:28:28Z kib $");
81
82 #include "opt_vm.h"
83
84 #include <sys/param.h>
85 #include <sys/systm.h>
86 #include <sys/lock.h>
87 #include <sys/kernel.h>
88 #include <sys/limits.h>
89 #include <sys/malloc.h>
90 #include <sys/msgbuf.h>
91 #include <sys/mutex.h>
92 #include <sys/proc.h>
93 #include <sys/sysctl.h>
94 #include <sys/vmmeter.h>
95 #include <sys/vnode.h>
96
97 #include <vm/vm.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_param.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_pager.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/vm_extern.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110
111 #include <machine/md_var.h>
112
113 /*
114 * Associated with page of user-allocatable memory is a
115 * page structure.
116 */
117
118 struct vpgqueues vm_page_queues[PQ_COUNT];
119 struct vpglocks vm_page_queue_lock;
120 struct vpglocks vm_page_queue_free_lock;
121
122 struct vpglocks pa_lock[PA_LOCK_COUNT];
123
124 vm_page_t vm_page_array;
125 long vm_page_array_size;
126 long first_page;
127 int vm_page_zero_count;
128
129 static int boot_pages = UMA_BOOT_PAGES;
130 TUNABLE_INT("vm.boot_pages", &boot_pages);
131 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
132 "number of pages allocated for bootstrapping the VM system");
133
134 int pa_tryrelock_restart;
135 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
136 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
137
138 static uma_zone_t fakepg_zone;
139
140 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
141 static void vm_page_queue_remove(int queue, vm_page_t m);
142 static void vm_page_enqueue(int queue, vm_page_t m);
143 static void vm_page_init_fakepg(void *dummy);
144
145 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
146
147 static void
148 vm_page_init_fakepg(void *dummy)
149 {
150
151 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
152 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
153 }
154
155 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
156 #if PAGE_SIZE == 32768
157 #ifdef CTASSERT
158 CTASSERT(sizeof(u_long) >= 8);
159 #endif
160 #endif
161
162 /*
163 * Try to acquire a physical address lock while a pmap is locked. If we
164 * fail to trylock we unlock and lock the pmap directly and cache the
165 * locked pa in *locked. The caller should then restart their loop in case
166 * the virtual to physical mapping has changed.
167 */
168 int
169 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
170 {
171 vm_paddr_t lockpa;
172
173 lockpa = *locked;
174 *locked = pa;
175 if (lockpa) {
176 PA_LOCK_ASSERT(lockpa, MA_OWNED);
177 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
178 return (0);
179 PA_UNLOCK(lockpa);
180 }
181 if (PA_TRYLOCK(pa))
182 return (0);
183 PMAP_UNLOCK(pmap);
184 atomic_add_int(&pa_tryrelock_restart, 1);
185 PA_LOCK(pa);
186 PMAP_LOCK(pmap);
187 return (EAGAIN);
188 }
189
190 /*
191 * vm_set_page_size:
192 *
193 * Sets the page size, perhaps based upon the memory
194 * size. Must be called before any use of page-size
195 * dependent functions.
196 */
197 void
198 vm_set_page_size(void)
199 {
200 if (cnt.v_page_size == 0)
201 cnt.v_page_size = PAGE_SIZE;
202 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
203 panic("vm_set_page_size: page size not a power of two");
204 }
205
206 /*
207 * vm_page_blacklist_lookup:
208 *
209 * See if a physical address in this page has been listed
210 * in the blacklist tunable. Entries in the tunable are
211 * separated by spaces or commas. If an invalid integer is
212 * encountered then the rest of the string is skipped.
213 */
214 static int
215 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
216 {
217 vm_paddr_t bad;
218 char *cp, *pos;
219
220 for (pos = list; *pos != '\0'; pos = cp) {
221 bad = strtoq(pos, &cp, 0);
222 if (*cp != '\0') {
223 if (*cp == ' ' || *cp == ',') {
224 cp++;
225 if (cp == pos)
226 continue;
227 } else
228 break;
229 }
230 if (pa == trunc_page(bad))
231 return (1);
232 }
233 return (0);
234 }
235
236 /*
237 * vm_page_startup:
238 *
239 * Initializes the resident memory module.
240 *
241 * Allocates memory for the page cells, and
242 * for the object/offset-to-page hash table headers.
243 * Each page cell is initialized and placed on the free list.
244 */
245 vm_offset_t
246 vm_page_startup(vm_offset_t vaddr)
247 {
248 vm_offset_t mapped;
249 vm_paddr_t page_range;
250 vm_paddr_t new_end;
251 int i;
252 vm_paddr_t pa;
253 vm_paddr_t last_pa;
254 char *list;
255
256 /* the biggest memory array is the second group of pages */
257 vm_paddr_t end;
258 vm_paddr_t biggestsize;
259 vm_paddr_t low_water, high_water;
260 int biggestone;
261
262 biggestsize = 0;
263 biggestone = 0;
264 vaddr = round_page(vaddr);
265
266 for (i = 0; phys_avail[i + 1]; i += 2) {
267 phys_avail[i] = round_page(phys_avail[i]);
268 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
269 }
270
271 low_water = phys_avail[0];
272 high_water = phys_avail[1];
273
274 for (i = 0; phys_avail[i + 1]; i += 2) {
275 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
276
277 if (size > biggestsize) {
278 biggestone = i;
279 biggestsize = size;
280 }
281 if (phys_avail[i] < low_water)
282 low_water = phys_avail[i];
283 if (phys_avail[i + 1] > high_water)
284 high_water = phys_avail[i + 1];
285 }
286
287 #ifdef XEN
288 low_water = 0;
289 #endif
290
291 end = phys_avail[biggestone+1];
292
293 /*
294 * Initialize the page and queue locks.
295 */
296 mtx_init(&vm_page_queue_mtx, "vm page queue", NULL, MTX_DEF |
297 MTX_RECURSE);
298 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
299 for (i = 0; i < PA_LOCK_COUNT; i++)
300 mtx_init(&pa_lock[i].data, "vm page", NULL, MTX_DEF);
301
302 /*
303 * Initialize the queue headers for the hold queue, the active queue,
304 * and the inactive queue.
305 */
306 for (i = 0; i < PQ_COUNT; i++)
307 TAILQ_INIT(&vm_page_queues[i].pl);
308 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
309 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
310 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
311
312 /*
313 * Allocate memory for use when boot strapping the kernel memory
314 * allocator.
315 */
316 new_end = end - (boot_pages * UMA_SLAB_SIZE);
317 new_end = trunc_page(new_end);
318 mapped = pmap_map(&vaddr, new_end, end,
319 VM_PROT_READ | VM_PROT_WRITE);
320 bzero((void *)mapped, end - new_end);
321 uma_startup((void *)mapped, boot_pages);
322
323 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
324 defined(__mips__)
325 /*
326 * Allocate a bitmap to indicate that a random physical page
327 * needs to be included in a minidump.
328 *
329 * The amd64 port needs this to indicate which direct map pages
330 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
331 *
332 * However, i386 still needs this workspace internally within the
333 * minidump code. In theory, they are not needed on i386, but are
334 * included should the sf_buf code decide to use them.
335 */
336 last_pa = 0;
337 for (i = 0; dump_avail[i + 1] != 0; i += 2)
338 if (dump_avail[i + 1] > last_pa)
339 last_pa = dump_avail[i + 1];
340 page_range = last_pa / PAGE_SIZE;
341 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
342 new_end -= vm_page_dump_size;
343 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
344 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
345 bzero((void *)vm_page_dump, vm_page_dump_size);
346 #endif
347 #ifdef __amd64__
348 /*
349 * Request that the physical pages underlying the message buffer be
350 * included in a crash dump. Since the message buffer is accessed
351 * through the direct map, they are not automatically included.
352 */
353 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
354 last_pa = pa + round_page(msgbufsize);
355 while (pa < last_pa) {
356 dump_add_page(pa);
357 pa += PAGE_SIZE;
358 }
359 #endif
360 /*
361 * Compute the number of pages of memory that will be available for
362 * use (taking into account the overhead of a page structure per
363 * page).
364 */
365 first_page = low_water / PAGE_SIZE;
366 #ifdef VM_PHYSSEG_SPARSE
367 page_range = 0;
368 for (i = 0; phys_avail[i + 1] != 0; i += 2)
369 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
370 #elif defined(VM_PHYSSEG_DENSE)
371 page_range = high_water / PAGE_SIZE - first_page;
372 #else
373 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
374 #endif
375 end = new_end;
376
377 /*
378 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
379 */
380 vaddr += PAGE_SIZE;
381
382 /*
383 * Initialize the mem entry structures now, and put them in the free
384 * queue.
385 */
386 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
387 mapped = pmap_map(&vaddr, new_end, end,
388 VM_PROT_READ | VM_PROT_WRITE);
389 vm_page_array = (vm_page_t) mapped;
390 #if VM_NRESERVLEVEL > 0
391 /*
392 * Allocate memory for the reservation management system's data
393 * structures.
394 */
395 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
396 #endif
397 #if defined(__amd64__) || defined(__mips__)
398 /*
399 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
400 * like i386, so the pages must be tracked for a crashdump to include
401 * this data. This includes the vm_page_array and the early UMA
402 * bootstrap pages.
403 */
404 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
405 dump_add_page(pa);
406 #endif
407 phys_avail[biggestone + 1] = new_end;
408
409 /*
410 * Clear all of the page structures
411 */
412 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
413 for (i = 0; i < page_range; i++)
414 vm_page_array[i].order = VM_NFREEORDER;
415 vm_page_array_size = page_range;
416
417 /*
418 * Initialize the physical memory allocator.
419 */
420 vm_phys_init();
421
422 /*
423 * Add every available physical page that is not blacklisted to
424 * the free lists.
425 */
426 cnt.v_page_count = 0;
427 cnt.v_free_count = 0;
428 list = getenv("vm.blacklist");
429 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
430 pa = phys_avail[i];
431 last_pa = phys_avail[i + 1];
432 while (pa < last_pa) {
433 if (list != NULL &&
434 vm_page_blacklist_lookup(list, pa))
435 printf("Skipping page with pa 0x%jx\n",
436 (uintmax_t)pa);
437 else
438 vm_phys_add_page(pa);
439 pa += PAGE_SIZE;
440 }
441 }
442 freeenv(list);
443 #if VM_NRESERVLEVEL > 0
444 /*
445 * Initialize the reservation management system.
446 */
447 vm_reserv_init();
448 #endif
449 return (vaddr);
450 }
451
452
453 CTASSERT(offsetof(struct vm_page, aflags) % sizeof(uint32_t) == 0);
454
455 void
456 vm_page_aflag_set(vm_page_t m, uint8_t bits)
457 {
458 uint32_t *addr, val;
459
460 /*
461 * The PGA_WRITEABLE flag can only be set if the page is managed and
462 * VPO_BUSY. Currently, this flag is only set by pmap_enter().
463 */
464 KASSERT((bits & PGA_WRITEABLE) == 0 ||
465 (m->oflags & (VPO_UNMANAGED | VPO_BUSY)) == VPO_BUSY,
466 ("PGA_WRITEABLE and !VPO_BUSY"));
467
468 /*
469 * We want to use atomic updates for m->aflags, which is a
470 * byte wide. Not all architectures provide atomic operations
471 * on the single-byte destination. Punt and access the whole
472 * 4-byte word with an atomic update. Parallel non-atomic
473 * updates to the fields included in the update by proximity
474 * are handled properly by atomics.
475 */
476 addr = (void *)&m->aflags;
477 MPASS(((uintptr_t)addr & (sizeof(uint32_t) - 1)) == 0);
478 val = bits;
479 #if BYTE_ORDER == BIG_ENDIAN
480 val <<= 24;
481 #endif
482 atomic_set_32(addr, val);
483 }
484
485 void
486 vm_page_aflag_clear(vm_page_t m, uint8_t bits)
487 {
488 uint32_t *addr, val;
489
490 /*
491 * The PGA_REFERENCED flag can only be cleared if the object
492 * containing the page is locked.
493 */
494 KASSERT((bits & PGA_REFERENCED) == 0 || VM_OBJECT_LOCKED(m->object),
495 ("PGA_REFERENCED and !VM_OBJECT_LOCKED"));
496
497 /*
498 * See the comment in vm_page_aflag_set().
499 */
500 addr = (void *)&m->aflags;
501 MPASS(((uintptr_t)addr & (sizeof(uint32_t) - 1)) == 0);
502 val = bits;
503 #if BYTE_ORDER == BIG_ENDIAN
504 val <<= 24;
505 #endif
506 atomic_clear_32(addr, val);
507 }
508
509 void
510 vm_page_reference(vm_page_t m)
511 {
512
513 vm_page_aflag_set(m, PGA_REFERENCED);
514 }
515
516 void
517 vm_page_busy(vm_page_t m)
518 {
519
520 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
521 KASSERT((m->oflags & VPO_BUSY) == 0,
522 ("vm_page_busy: page already busy!!!"));
523 m->oflags |= VPO_BUSY;
524 }
525
526 /*
527 * vm_page_flash:
528 *
529 * wakeup anyone waiting for the page.
530 */
531 void
532 vm_page_flash(vm_page_t m)
533 {
534
535 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
536 if (m->oflags & VPO_WANTED) {
537 m->oflags &= ~VPO_WANTED;
538 wakeup(m);
539 }
540 }
541
542 /*
543 * vm_page_wakeup:
544 *
545 * clear the VPO_BUSY flag and wakeup anyone waiting for the
546 * page.
547 *
548 */
549 void
550 vm_page_wakeup(vm_page_t m)
551 {
552
553 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
554 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
555 m->oflags &= ~VPO_BUSY;
556 vm_page_flash(m);
557 }
558
559 void
560 vm_page_io_start(vm_page_t m)
561 {
562
563 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
564 m->busy++;
565 }
566
567 void
568 vm_page_io_finish(vm_page_t m)
569 {
570
571 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
572 KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
573 m->busy--;
574 if (m->busy == 0)
575 vm_page_flash(m);
576 }
577
578 /*
579 * Keep page from being freed by the page daemon
580 * much of the same effect as wiring, except much lower
581 * overhead and should be used only for *very* temporary
582 * holding ("wiring").
583 */
584 void
585 vm_page_hold(vm_page_t mem)
586 {
587
588 vm_page_lock_assert(mem, MA_OWNED);
589 mem->hold_count++;
590 }
591
592 void
593 vm_page_unhold(vm_page_t mem)
594 {
595
596 vm_page_lock_assert(mem, MA_OWNED);
597 --mem->hold_count;
598 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
599 if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
600 vm_page_free_toq(mem);
601 }
602
603 /*
604 * vm_page_unhold_pages:
605 *
606 * Unhold each of the pages that is referenced by the given array.
607 */
608 void
609 vm_page_unhold_pages(vm_page_t *ma, int count)
610 {
611 struct mtx *mtx, *new_mtx;
612
613 mtx = NULL;
614 for (; count != 0; count--) {
615 /*
616 * Avoid releasing and reacquiring the same page lock.
617 */
618 new_mtx = vm_page_lockptr(*ma);
619 if (mtx != new_mtx) {
620 if (mtx != NULL)
621 mtx_unlock(mtx);
622 mtx = new_mtx;
623 mtx_lock(mtx);
624 }
625 vm_page_unhold(*ma);
626 ma++;
627 }
628 if (mtx != NULL)
629 mtx_unlock(mtx);
630 }
631
632 vm_page_t
633 PHYS_TO_VM_PAGE(vm_paddr_t pa)
634 {
635 vm_page_t m;
636
637 #ifdef VM_PHYSSEG_SPARSE
638 m = vm_phys_paddr_to_vm_page(pa);
639 if (m == NULL)
640 m = vm_phys_fictitious_to_vm_page(pa);
641 return (m);
642 #elif defined(VM_PHYSSEG_DENSE)
643 long pi;
644
645 pi = atop(pa);
646 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
647 m = &vm_page_array[pi - first_page];
648 return (m);
649 }
650 return (vm_phys_fictitious_to_vm_page(pa));
651 #else
652 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
653 #endif
654 }
655
656 /*
657 * vm_page_getfake:
658 *
659 * Create a fictitious page with the specified physical address and
660 * memory attribute. The memory attribute is the only the machine-
661 * dependent aspect of a fictitious page that must be initialized.
662 */
663 vm_page_t
664 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
665 {
666 vm_page_t m;
667
668 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
669 vm_page_initfake(m, paddr, memattr);
670 return (m);
671 }
672
673 void
674 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
675 {
676
677 if ((m->flags & PG_FICTITIOUS) != 0) {
678 /*
679 * The page's memattr might have changed since the
680 * previous initialization. Update the pmap to the
681 * new memattr.
682 */
683 goto memattr;
684 }
685 m->phys_addr = paddr;
686 m->queue = PQ_NONE;
687 /* Fictitious pages don't use "segind". */
688 m->flags = PG_FICTITIOUS;
689 /* Fictitious pages don't use "order" or "pool". */
690 m->oflags = VPO_BUSY | VPO_UNMANAGED;
691 m->wire_count = 1;
692 memattr:
693 pmap_page_set_memattr(m, memattr);
694 }
695
696 /*
697 * vm_page_putfake:
698 *
699 * Release a fictitious page.
700 */
701 void
702 vm_page_putfake(vm_page_t m)
703 {
704
705 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
706 KASSERT((m->flags & PG_FICTITIOUS) != 0,
707 ("vm_page_putfake: bad page %p", m));
708 uma_zfree(fakepg_zone, m);
709 }
710
711 /*
712 * vm_page_updatefake:
713 *
714 * Update the given fictitious page to the specified physical address and
715 * memory attribute.
716 */
717 void
718 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
719 {
720
721 KASSERT((m->flags & PG_FICTITIOUS) != 0,
722 ("vm_page_updatefake: bad page %p", m));
723 m->phys_addr = paddr;
724 pmap_page_set_memattr(m, memattr);
725 }
726
727 /*
728 * vm_page_free:
729 *
730 * Free a page.
731 */
732 void
733 vm_page_free(vm_page_t m)
734 {
735
736 m->flags &= ~PG_ZERO;
737 vm_page_free_toq(m);
738 }
739
740 /*
741 * vm_page_free_zero:
742 *
743 * Free a page to the zerod-pages queue
744 */
745 void
746 vm_page_free_zero(vm_page_t m)
747 {
748
749 m->flags |= PG_ZERO;
750 vm_page_free_toq(m);
751 }
752
753 /*
754 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
755 * array which is not the request page.
756 */
757 void
758 vm_page_readahead_finish(vm_page_t m)
759 {
760
761 if (m->valid != 0) {
762 /*
763 * Since the page is not the requested page, whether
764 * it should be activated or deactivated is not
765 * obvious. Empirical results have shown that
766 * deactivating the page is usually the best choice,
767 * unless the page is wanted by another thread.
768 */
769 if (m->oflags & VPO_WANTED) {
770 vm_page_lock(m);
771 vm_page_activate(m);
772 vm_page_unlock(m);
773 } else {
774 vm_page_lock(m);
775 vm_page_deactivate(m);
776 vm_page_unlock(m);
777 }
778 vm_page_wakeup(m);
779 } else {
780 /*
781 * Free the completely invalid page. Such page state
782 * occurs due to the short read operation which did
783 * not covered our page at all, or in case when a read
784 * error happens.
785 */
786 vm_page_lock(m);
787 vm_page_free(m);
788 vm_page_unlock(m);
789 }
790 }
791
792 /*
793 * vm_page_sleep:
794 *
795 * Sleep and release the page and page queues locks.
796 *
797 * The object containing the given page must be locked.
798 */
799 void
800 vm_page_sleep(vm_page_t m, const char *msg)
801 {
802
803 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
804 if (mtx_owned(&vm_page_queue_mtx))
805 vm_page_unlock_queues();
806 if (mtx_owned(vm_page_lockptr(m)))
807 vm_page_unlock(m);
808
809 /*
810 * It's possible that while we sleep, the page will get
811 * unbusied and freed. If we are holding the object
812 * lock, we will assume we hold a reference to the object
813 * such that even if m->object changes, we can re-lock
814 * it.
815 */
816 m->oflags |= VPO_WANTED;
817 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
818 }
819
820 /*
821 * vm_page_dirty:
822 *
823 * Set all bits in the page's dirty field.
824 *
825 * The object containing the specified page must be locked if the
826 * call is made from the machine-independent layer.
827 *
828 * See vm_page_clear_dirty_mask().
829 */
830 void
831 vm_page_dirty(vm_page_t m)
832 {
833
834 KASSERT((m->flags & PG_CACHED) == 0,
835 ("vm_page_dirty: page in cache!"));
836 KASSERT(!VM_PAGE_IS_FREE(m),
837 ("vm_page_dirty: page is free!"));
838 KASSERT(m->valid == VM_PAGE_BITS_ALL,
839 ("vm_page_dirty: page is invalid!"));
840 m->dirty = VM_PAGE_BITS_ALL;
841 }
842
843 /*
844 * vm_page_splay:
845 *
846 * Implements Sleator and Tarjan's top-down splay algorithm. Returns
847 * the vm_page containing the given pindex. If, however, that
848 * pindex is not found in the vm_object, returns a vm_page that is
849 * adjacent to the pindex, coming before or after it.
850 */
851 vm_page_t
852 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
853 {
854 struct vm_page dummy;
855 vm_page_t lefttreemax, righttreemin, y;
856
857 if (root == NULL)
858 return (root);
859 lefttreemax = righttreemin = &dummy;
860 for (;; root = y) {
861 if (pindex < root->pindex) {
862 if ((y = root->left) == NULL)
863 break;
864 if (pindex < y->pindex) {
865 /* Rotate right. */
866 root->left = y->right;
867 y->right = root;
868 root = y;
869 if ((y = root->left) == NULL)
870 break;
871 }
872 /* Link into the new root's right tree. */
873 righttreemin->left = root;
874 righttreemin = root;
875 } else if (pindex > root->pindex) {
876 if ((y = root->right) == NULL)
877 break;
878 if (pindex > y->pindex) {
879 /* Rotate left. */
880 root->right = y->left;
881 y->left = root;
882 root = y;
883 if ((y = root->right) == NULL)
884 break;
885 }
886 /* Link into the new root's left tree. */
887 lefttreemax->right = root;
888 lefttreemax = root;
889 } else
890 break;
891 }
892 /* Assemble the new root. */
893 lefttreemax->right = root->left;
894 righttreemin->left = root->right;
895 root->left = dummy.right;
896 root->right = dummy.left;
897 return (root);
898 }
899
900 /*
901 * vm_page_insert: [ internal use only ]
902 *
903 * Inserts the given mem entry into the object and object list.
904 *
905 * The object must be locked.
906 */
907 void
908 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
909 {
910 vm_page_t root;
911
912 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
913 if (m->object != NULL)
914 panic("vm_page_insert: page already inserted");
915
916 /*
917 * Record the object/offset pair in this page
918 */
919 m->object = object;
920 m->pindex = pindex;
921
922 /*
923 * Now link into the object's ordered list of backed pages.
924 */
925 root = object->root;
926 if (root == NULL) {
927 m->left = NULL;
928 m->right = NULL;
929 TAILQ_INSERT_TAIL(&object->memq, m, listq);
930 } else {
931 root = vm_page_splay(pindex, root);
932 if (pindex < root->pindex) {
933 m->left = root->left;
934 m->right = root;
935 root->left = NULL;
936 TAILQ_INSERT_BEFORE(root, m, listq);
937 } else if (pindex == root->pindex)
938 panic("vm_page_insert: offset already allocated");
939 else {
940 m->right = root->right;
941 m->left = root;
942 root->right = NULL;
943 TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
944 }
945 }
946 object->root = m;
947
948 /*
949 * Show that the object has one more resident page.
950 */
951 object->resident_page_count++;
952
953 /*
954 * Hold the vnode until the last page is released.
955 */
956 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
957 vhold(object->handle);
958
959 /*
960 * Since we are inserting a new and possibly dirty page,
961 * update the object's OBJ_MIGHTBEDIRTY flag.
962 */
963 if (pmap_page_is_write_mapped(m))
964 vm_object_set_writeable_dirty(object);
965 }
966
967 /*
968 * vm_page_remove:
969 *
970 * Removes the given mem entry from the object/offset-page
971 * table and the object page list, but do not invalidate/terminate
972 * the backing store.
973 *
974 * The object must be locked. The page must be locked if it is managed.
975 */
976 void
977 vm_page_remove(vm_page_t m)
978 {
979 vm_object_t object;
980 vm_page_t next, prev, root;
981
982 if ((m->oflags & VPO_UNMANAGED) == 0)
983 vm_page_lock_assert(m, MA_OWNED);
984 if ((object = m->object) == NULL)
985 return;
986 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
987 if (m->oflags & VPO_BUSY) {
988 m->oflags &= ~VPO_BUSY;
989 vm_page_flash(m);
990 }
991
992 /*
993 * Now remove from the object's list of backed pages.
994 */
995 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
996 /*
997 * Since the page's successor in the list is also its parent
998 * in the tree, its right subtree must be empty.
999 */
1000 next->left = m->left;
1001 KASSERT(m->right == NULL,
1002 ("vm_page_remove: page %p has right child", m));
1003 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1004 prev->right == m) {
1005 /*
1006 * Since the page's predecessor in the list is also its parent
1007 * in the tree, its left subtree must be empty.
1008 */
1009 KASSERT(m->left == NULL,
1010 ("vm_page_remove: page %p has left child", m));
1011 prev->right = m->right;
1012 } else {
1013 if (m != object->root)
1014 vm_page_splay(m->pindex, object->root);
1015 if (m->left == NULL)
1016 root = m->right;
1017 else if (m->right == NULL)
1018 root = m->left;
1019 else {
1020 /*
1021 * Move the page's successor to the root, because
1022 * pages are usually removed in ascending order.
1023 */
1024 if (m->right != next)
1025 vm_page_splay(m->pindex, m->right);
1026 next->left = m->left;
1027 root = next;
1028 }
1029 object->root = root;
1030 }
1031 TAILQ_REMOVE(&object->memq, m, listq);
1032
1033 /*
1034 * And show that the object has one fewer resident page.
1035 */
1036 object->resident_page_count--;
1037
1038 /*
1039 * The vnode may now be recycled.
1040 */
1041 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1042 vdrop(object->handle);
1043
1044 m->object = NULL;
1045 }
1046
1047 /*
1048 * vm_page_lookup:
1049 *
1050 * Returns the page associated with the object/offset
1051 * pair specified; if none is found, NULL is returned.
1052 *
1053 * The object must be locked.
1054 */
1055 vm_page_t
1056 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1057 {
1058 vm_page_t m;
1059
1060 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1061 if ((m = object->root) != NULL && m->pindex != pindex) {
1062 m = vm_page_splay(pindex, m);
1063 if ((object->root = m)->pindex != pindex)
1064 m = NULL;
1065 }
1066 return (m);
1067 }
1068
1069 /*
1070 * vm_page_find_least:
1071 *
1072 * Returns the page associated with the object with least pindex
1073 * greater than or equal to the parameter pindex, or NULL.
1074 *
1075 * The object must be locked.
1076 */
1077 vm_page_t
1078 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1079 {
1080 vm_page_t m;
1081
1082 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1083 if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
1084 if (m->pindex < pindex) {
1085 m = vm_page_splay(pindex, object->root);
1086 if ((object->root = m)->pindex < pindex)
1087 m = TAILQ_NEXT(m, listq);
1088 }
1089 }
1090 return (m);
1091 }
1092
1093 /*
1094 * Returns the given page's successor (by pindex) within the object if it is
1095 * resident; if none is found, NULL is returned.
1096 *
1097 * The object must be locked.
1098 */
1099 vm_page_t
1100 vm_page_next(vm_page_t m)
1101 {
1102 vm_page_t next;
1103
1104 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1105 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1106 next->pindex != m->pindex + 1)
1107 next = NULL;
1108 return (next);
1109 }
1110
1111 /*
1112 * Returns the given page's predecessor (by pindex) within the object if it is
1113 * resident; if none is found, NULL is returned.
1114 *
1115 * The object must be locked.
1116 */
1117 vm_page_t
1118 vm_page_prev(vm_page_t m)
1119 {
1120 vm_page_t prev;
1121
1122 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1123 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1124 prev->pindex != m->pindex - 1)
1125 prev = NULL;
1126 return (prev);
1127 }
1128
1129 /*
1130 * vm_page_rename:
1131 *
1132 * Move the given memory entry from its
1133 * current object to the specified target object/offset.
1134 *
1135 * Note: swap associated with the page must be invalidated by the move. We
1136 * have to do this for several reasons: (1) we aren't freeing the
1137 * page, (2) we are dirtying the page, (3) the VM system is probably
1138 * moving the page from object A to B, and will then later move
1139 * the backing store from A to B and we can't have a conflict.
1140 *
1141 * Note: we *always* dirty the page. It is necessary both for the
1142 * fact that we moved it, and because we may be invalidating
1143 * swap. If the page is on the cache, we have to deactivate it
1144 * or vm_page_dirty() will panic. Dirty pages are not allowed
1145 * on the cache.
1146 *
1147 * The objects must be locked. The page must be locked if it is managed.
1148 */
1149 void
1150 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1151 {
1152
1153 vm_page_remove(m);
1154 vm_page_insert(m, new_object, new_pindex);
1155 vm_page_dirty(m);
1156 }
1157
1158 /*
1159 * Convert all of the given object's cached pages that have a
1160 * pindex within the given range into free pages. If the value
1161 * zero is given for "end", then the range's upper bound is
1162 * infinity. If the given object is backed by a vnode and it
1163 * transitions from having one or more cached pages to none, the
1164 * vnode's hold count is reduced.
1165 */
1166 void
1167 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1168 {
1169 vm_page_t m, m_next;
1170 boolean_t empty;
1171
1172 mtx_lock(&vm_page_queue_free_mtx);
1173 if (__predict_false(object->cache == NULL)) {
1174 mtx_unlock(&vm_page_queue_free_mtx);
1175 return;
1176 }
1177 m = object->cache = vm_page_splay(start, object->cache);
1178 if (m->pindex < start) {
1179 if (m->right == NULL)
1180 m = NULL;
1181 else {
1182 m_next = vm_page_splay(start, m->right);
1183 m_next->left = m;
1184 m->right = NULL;
1185 m = object->cache = m_next;
1186 }
1187 }
1188
1189 /*
1190 * At this point, "m" is either (1) a reference to the page
1191 * with the least pindex that is greater than or equal to
1192 * "start" or (2) NULL.
1193 */
1194 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
1195 /*
1196 * Find "m"'s successor and remove "m" from the
1197 * object's cache.
1198 */
1199 if (m->right == NULL) {
1200 object->cache = m->left;
1201 m_next = NULL;
1202 } else {
1203 m_next = vm_page_splay(start, m->right);
1204 m_next->left = m->left;
1205 object->cache = m_next;
1206 }
1207 /* Convert "m" to a free page. */
1208 m->object = NULL;
1209 m->valid = 0;
1210 /* Clear PG_CACHED and set PG_FREE. */
1211 m->flags ^= PG_CACHED | PG_FREE;
1212 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1213 ("vm_page_cache_free: page %p has inconsistent flags", m));
1214 cnt.v_cache_count--;
1215 cnt.v_free_count++;
1216 }
1217 empty = object->cache == NULL;
1218 mtx_unlock(&vm_page_queue_free_mtx);
1219 if (object->type == OBJT_VNODE && empty)
1220 vdrop(object->handle);
1221 }
1222
1223 /*
1224 * Returns the cached page that is associated with the given
1225 * object and offset. If, however, none exists, returns NULL.
1226 *
1227 * The free page queue must be locked.
1228 */
1229 static inline vm_page_t
1230 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1231 {
1232 vm_page_t m;
1233
1234 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1235 if ((m = object->cache) != NULL && m->pindex != pindex) {
1236 m = vm_page_splay(pindex, m);
1237 if ((object->cache = m)->pindex != pindex)
1238 m = NULL;
1239 }
1240 return (m);
1241 }
1242
1243 /*
1244 * Remove the given cached page from its containing object's
1245 * collection of cached pages.
1246 *
1247 * The free page queue must be locked.
1248 */
1249 void
1250 vm_page_cache_remove(vm_page_t m)
1251 {
1252 vm_object_t object;
1253 vm_page_t root;
1254
1255 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1256 KASSERT((m->flags & PG_CACHED) != 0,
1257 ("vm_page_cache_remove: page %p is not cached", m));
1258 object = m->object;
1259 if (m != object->cache) {
1260 root = vm_page_splay(m->pindex, object->cache);
1261 KASSERT(root == m,
1262 ("vm_page_cache_remove: page %p is not cached in object %p",
1263 m, object));
1264 }
1265 if (m->left == NULL)
1266 root = m->right;
1267 else if (m->right == NULL)
1268 root = m->left;
1269 else {
1270 root = vm_page_splay(m->pindex, m->left);
1271 root->right = m->right;
1272 }
1273 object->cache = root;
1274 m->object = NULL;
1275 cnt.v_cache_count--;
1276 }
1277
1278 /*
1279 * Transfer all of the cached pages with offset greater than or
1280 * equal to 'offidxstart' from the original object's cache to the
1281 * new object's cache. However, any cached pages with offset
1282 * greater than or equal to the new object's size are kept in the
1283 * original object. Initially, the new object's cache must be
1284 * empty. Offset 'offidxstart' in the original object must
1285 * correspond to offset zero in the new object.
1286 *
1287 * The new object must be locked.
1288 */
1289 void
1290 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1291 vm_object_t new_object)
1292 {
1293 vm_page_t m, m_next;
1294
1295 /*
1296 * Insertion into an object's collection of cached pages
1297 * requires the object to be locked. In contrast, removal does
1298 * not.
1299 */
1300 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1301 KASSERT(new_object->cache == NULL,
1302 ("vm_page_cache_transfer: object %p has cached pages",
1303 new_object));
1304 mtx_lock(&vm_page_queue_free_mtx);
1305 if ((m = orig_object->cache) != NULL) {
1306 /*
1307 * Transfer all of the pages with offset greater than or
1308 * equal to 'offidxstart' from the original object's
1309 * cache to the new object's cache.
1310 */
1311 m = vm_page_splay(offidxstart, m);
1312 if (m->pindex < offidxstart) {
1313 orig_object->cache = m;
1314 new_object->cache = m->right;
1315 m->right = NULL;
1316 } else {
1317 orig_object->cache = m->left;
1318 new_object->cache = m;
1319 m->left = NULL;
1320 }
1321 while ((m = new_object->cache) != NULL) {
1322 if ((m->pindex - offidxstart) >= new_object->size) {
1323 /*
1324 * Return all of the cached pages with
1325 * offset greater than or equal to the
1326 * new object's size to the original
1327 * object's cache.
1328 */
1329 new_object->cache = m->left;
1330 m->left = orig_object->cache;
1331 orig_object->cache = m;
1332 break;
1333 }
1334 m_next = vm_page_splay(m->pindex, m->right);
1335 /* Update the page's object and offset. */
1336 m->object = new_object;
1337 m->pindex -= offidxstart;
1338 if (m_next == NULL)
1339 break;
1340 m->right = NULL;
1341 m_next->left = m;
1342 new_object->cache = m_next;
1343 }
1344 KASSERT(new_object->cache == NULL ||
1345 new_object->type == OBJT_SWAP,
1346 ("vm_page_cache_transfer: object %p's type is incompatible"
1347 " with cached pages", new_object));
1348 }
1349 mtx_unlock(&vm_page_queue_free_mtx);
1350 }
1351
1352 /*
1353 * Returns TRUE if a cached page is associated with the given object and
1354 * offset, and FALSE otherwise.
1355 *
1356 * The object must be locked.
1357 */
1358 boolean_t
1359 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1360 {
1361 vm_page_t m;
1362
1363 /*
1364 * Insertion into an object's collection of cached pages requires the
1365 * object to be locked. Therefore, if the object is locked and the
1366 * object's collection is empty, there is no need to acquire the free
1367 * page queues lock in order to prove that the specified page doesn't
1368 * exist.
1369 */
1370 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1371 if (object->cache == NULL)
1372 return (FALSE);
1373 mtx_lock(&vm_page_queue_free_mtx);
1374 m = vm_page_cache_lookup(object, pindex);
1375 mtx_unlock(&vm_page_queue_free_mtx);
1376 return (m != NULL);
1377 }
1378
1379 /*
1380 * vm_page_alloc:
1381 *
1382 * Allocate and return a memory cell associated
1383 * with this VM object/offset pair.
1384 *
1385 * The caller must always specify an allocation class.
1386 *
1387 * allocation classes:
1388 * VM_ALLOC_NORMAL normal process request
1389 * VM_ALLOC_SYSTEM system *really* needs a page
1390 * VM_ALLOC_INTERRUPT interrupt time request
1391 *
1392 * optional allocation flags:
1393 * VM_ALLOC_ZERO prefer a zeroed page
1394 * VM_ALLOC_WIRED wire the allocated page
1395 * VM_ALLOC_NOOBJ page is not associated with a vm object
1396 * VM_ALLOC_NOBUSY do not set the page busy
1397 * VM_ALLOC_IFCACHED return page only if it is cached
1398 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1399 * is cached
1400 *
1401 * This routine may not sleep.
1402 */
1403 vm_page_t
1404 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1405 {
1406 struct vnode *vp = NULL;
1407 vm_object_t m_object;
1408 vm_page_t m;
1409 int flags, page_req;
1410
1411 if ((req & VM_ALLOC_NOOBJ) == 0) {
1412 KASSERT(object != NULL,
1413 ("vm_page_alloc: NULL object."));
1414 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1415 }
1416
1417 page_req = req & VM_ALLOC_CLASS_MASK;
1418
1419 /*
1420 * The pager is allowed to eat deeper into the free page list.
1421 */
1422 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT))
1423 page_req = VM_ALLOC_SYSTEM;
1424
1425 mtx_lock(&vm_page_queue_free_mtx);
1426 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1427 (page_req == VM_ALLOC_SYSTEM &&
1428 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1429 (page_req == VM_ALLOC_INTERRUPT &&
1430 cnt.v_free_count + cnt.v_cache_count > 0)) {
1431 /*
1432 * Allocate from the free queue if the number of free pages
1433 * exceeds the minimum for the request class.
1434 */
1435 if (object != NULL &&
1436 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1437 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1438 mtx_unlock(&vm_page_queue_free_mtx);
1439 return (NULL);
1440 }
1441 if (vm_phys_unfree_page(m))
1442 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1443 #if VM_NRESERVLEVEL > 0
1444 else if (!vm_reserv_reactivate_page(m))
1445 #else
1446 else
1447 #endif
1448 panic("vm_page_alloc: cache page %p is missing"
1449 " from the free queue", m);
1450 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1451 mtx_unlock(&vm_page_queue_free_mtx);
1452 return (NULL);
1453 #if VM_NRESERVLEVEL > 0
1454 } else if (object == NULL || object->type == OBJT_DEVICE ||
1455 object->type == OBJT_SG ||
1456 (object->flags & OBJ_COLORED) == 0 ||
1457 (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1458 #else
1459 } else {
1460 #endif
1461 m = vm_phys_alloc_pages(object != NULL ?
1462 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1463 #if VM_NRESERVLEVEL > 0
1464 if (m == NULL && vm_reserv_reclaim_inactive()) {
1465 m = vm_phys_alloc_pages(object != NULL ?
1466 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1467 0);
1468 }
1469 #endif
1470 }
1471 } else {
1472 /*
1473 * Not allocatable, give up.
1474 */
1475 mtx_unlock(&vm_page_queue_free_mtx);
1476 atomic_add_int(&vm_pageout_deficit,
1477 MAX((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1478 pagedaemon_wakeup();
1479 return (NULL);
1480 }
1481
1482 /*
1483 * At this point we had better have found a good page.
1484 */
1485
1486 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1487 KASSERT(m->queue == PQ_NONE,
1488 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1489 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1490 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1491 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1492 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1493 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1494 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1495 pmap_page_get_memattr(m)));
1496 if ((m->flags & PG_CACHED) != 0) {
1497 KASSERT(m->valid != 0,
1498 ("vm_page_alloc: cached page %p is invalid", m));
1499 if (m->object == object && m->pindex == pindex)
1500 cnt.v_reactivated++;
1501 else
1502 m->valid = 0;
1503 m_object = m->object;
1504 vm_page_cache_remove(m);
1505 if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1506 vp = m_object->handle;
1507 } else {
1508 KASSERT(VM_PAGE_IS_FREE(m),
1509 ("vm_page_alloc: page %p is not free", m));
1510 KASSERT(m->valid == 0,
1511 ("vm_page_alloc: free page %p is valid", m));
1512 cnt.v_free_count--;
1513 }
1514
1515 /*
1516 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag
1517 * must be cleared before the free page queues lock is released.
1518 */
1519 flags = 0;
1520 if (req & VM_ALLOC_NODUMP)
1521 flags |= PG_NODUMP;
1522 if (m->flags & PG_ZERO) {
1523 vm_page_zero_count--;
1524 if (req & VM_ALLOC_ZERO)
1525 flags = PG_ZERO;
1526 }
1527 m->flags = flags;
1528 mtx_unlock(&vm_page_queue_free_mtx);
1529 m->aflags = 0;
1530 if (object == NULL || object->type == OBJT_PHYS)
1531 m->oflags = VPO_UNMANAGED;
1532 else
1533 m->oflags = 0;
1534 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1535 m->oflags |= VPO_BUSY;
1536 if (req & VM_ALLOC_WIRED) {
1537 /*
1538 * The page lock is not required for wiring a page until that
1539 * page is inserted into the object.
1540 */
1541 atomic_add_int(&cnt.v_wire_count, 1);
1542 m->wire_count = 1;
1543 }
1544 m->act_count = 0;
1545
1546 if (object != NULL) {
1547 /* Ignore device objects; the pager sets "memattr" for them. */
1548 if (object->memattr != VM_MEMATTR_DEFAULT &&
1549 object->type != OBJT_DEVICE && object->type != OBJT_SG)
1550 pmap_page_set_memattr(m, object->memattr);
1551 vm_page_insert(m, object, pindex);
1552 } else
1553 m->pindex = pindex;
1554
1555 /*
1556 * The following call to vdrop() must come after the above call
1557 * to vm_page_insert() in case both affect the same object and
1558 * vnode. Otherwise, the affected vnode's hold count could
1559 * temporarily become zero.
1560 */
1561 if (vp != NULL)
1562 vdrop(vp);
1563
1564 /*
1565 * Don't wakeup too often - wakeup the pageout daemon when
1566 * we would be nearly out of memory.
1567 */
1568 if (vm_paging_needed())
1569 pagedaemon_wakeup();
1570
1571 return (m);
1572 }
1573
1574 /*
1575 * Initialize a page that has been freshly dequeued from a freelist.
1576 * The caller has to drop the vnode returned, if it is not NULL.
1577 *
1578 * To be called with vm_page_queue_free_mtx held.
1579 */
1580 struct vnode *
1581 vm_page_alloc_init(vm_page_t m)
1582 {
1583 struct vnode *drop;
1584 vm_object_t m_object;
1585
1586 KASSERT(m->queue == PQ_NONE,
1587 ("vm_page_alloc_init: page %p has unexpected queue %d",
1588 m, m->queue));
1589 KASSERT(m->wire_count == 0,
1590 ("vm_page_alloc_init: page %p is wired", m));
1591 KASSERT(m->hold_count == 0,
1592 ("vm_page_alloc_init: page %p is held", m));
1593 KASSERT(m->busy == 0,
1594 ("vm_page_alloc_init: page %p is busy", m));
1595 KASSERT(m->dirty == 0,
1596 ("vm_page_alloc_init: page %p is dirty", m));
1597 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1598 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1599 m, pmap_page_get_memattr(m)));
1600 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1601 drop = NULL;
1602 if ((m->flags & PG_CACHED) != 0) {
1603 m->valid = 0;
1604 m_object = m->object;
1605 vm_page_cache_remove(m);
1606 if (m_object->type == OBJT_VNODE &&
1607 m_object->cache == NULL)
1608 drop = m_object->handle;
1609 } else {
1610 KASSERT(VM_PAGE_IS_FREE(m),
1611 ("vm_page_alloc_init: page %p is not free", m));
1612 KASSERT(m->valid == 0,
1613 ("vm_page_alloc_init: free page %p is valid", m));
1614 cnt.v_free_count--;
1615 }
1616 if (m->flags & PG_ZERO)
1617 vm_page_zero_count--;
1618 /* Don't clear the PG_ZERO flag; we'll need it later. */
1619 m->flags &= PG_ZERO;
1620 m->aflags = 0;
1621 m->oflags = VPO_UNMANAGED;
1622 /* Unmanaged pages don't use "act_count". */
1623 return (drop);
1624 }
1625
1626 /*
1627 * vm_page_alloc_freelist:
1628 *
1629 * Allocate a page from the specified freelist.
1630 * Only the ALLOC_CLASS values in req are honored, other request flags
1631 * are ignored.
1632 */
1633 vm_page_t
1634 vm_page_alloc_freelist(int flind, int req)
1635 {
1636 struct vnode *drop;
1637 vm_page_t m;
1638 int page_req;
1639
1640 m = NULL;
1641 page_req = req & VM_ALLOC_CLASS_MASK;
1642 mtx_lock(&vm_page_queue_free_mtx);
1643 /*
1644 * Do not allocate reserved pages unless the req has asked for it.
1645 */
1646 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1647 (page_req == VM_ALLOC_SYSTEM &&
1648 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1649 (page_req == VM_ALLOC_INTERRUPT &&
1650 cnt.v_free_count + cnt.v_cache_count > 0)) {
1651 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1652 }
1653 if (m == NULL) {
1654 mtx_unlock(&vm_page_queue_free_mtx);
1655 return (NULL);
1656 }
1657 drop = vm_page_alloc_init(m);
1658 mtx_unlock(&vm_page_queue_free_mtx);
1659 if (drop)
1660 vdrop(drop);
1661 return (m);
1662 }
1663
1664 /*
1665 * vm_wait: (also see VM_WAIT macro)
1666 *
1667 * Sleep until free pages are available for allocation.
1668 * - Called in various places before memory allocations.
1669 */
1670 void
1671 vm_wait(void)
1672 {
1673
1674 mtx_lock(&vm_page_queue_free_mtx);
1675 if (curproc == pageproc) {
1676 vm_pageout_pages_needed = 1;
1677 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1678 PDROP | PSWP, "VMWait", 0);
1679 } else {
1680 if (!vm_pages_needed) {
1681 vm_pages_needed = 1;
1682 wakeup(&vm_pages_needed);
1683 }
1684 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1685 "vmwait", 0);
1686 }
1687 }
1688
1689 /*
1690 * vm_waitpfault: (also see VM_WAITPFAULT macro)
1691 *
1692 * Sleep until free pages are available for allocation.
1693 * - Called only in vm_fault so that processes page faulting
1694 * can be easily tracked.
1695 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1696 * processes will be able to grab memory first. Do not change
1697 * this balance without careful testing first.
1698 */
1699 void
1700 vm_waitpfault(void)
1701 {
1702
1703 mtx_lock(&vm_page_queue_free_mtx);
1704 if (!vm_pages_needed) {
1705 vm_pages_needed = 1;
1706 wakeup(&vm_pages_needed);
1707 }
1708 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1709 "pfault", 0);
1710 }
1711
1712 /*
1713 * vm_page_requeue:
1714 *
1715 * Move the given page to the tail of its present page queue.
1716 *
1717 * The page queues must be locked.
1718 */
1719 void
1720 vm_page_requeue(vm_page_t m)
1721 {
1722 struct vpgqueues *vpq;
1723 int queue;
1724
1725 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1726 queue = m->queue;
1727 KASSERT(queue != PQ_NONE,
1728 ("vm_page_requeue: page %p is not queued", m));
1729 vpq = &vm_page_queues[queue];
1730 TAILQ_REMOVE(&vpq->pl, m, pageq);
1731 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1732 }
1733
1734 /*
1735 * vm_page_queue_remove:
1736 *
1737 * Remove the given page from the specified queue.
1738 *
1739 * The page and page queues must be locked.
1740 */
1741 static __inline void
1742 vm_page_queue_remove(int queue, vm_page_t m)
1743 {
1744 struct vpgqueues *pq;
1745
1746 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1747 vm_page_lock_assert(m, MA_OWNED);
1748 pq = &vm_page_queues[queue];
1749 TAILQ_REMOVE(&pq->pl, m, pageq);
1750 (*pq->cnt)--;
1751 }
1752
1753 /*
1754 * vm_pageq_remove:
1755 *
1756 * Remove a page from its queue.
1757 *
1758 * The given page must be locked.
1759 */
1760 void
1761 vm_pageq_remove(vm_page_t m)
1762 {
1763 int queue;
1764
1765 vm_page_lock_assert(m, MA_OWNED);
1766 if ((queue = m->queue) != PQ_NONE) {
1767 vm_page_lock_queues();
1768 m->queue = PQ_NONE;
1769 vm_page_queue_remove(queue, m);
1770 vm_page_unlock_queues();
1771 }
1772 }
1773
1774 /*
1775 * vm_page_enqueue:
1776 *
1777 * Add the given page to the specified queue.
1778 *
1779 * The page queues must be locked.
1780 */
1781 static void
1782 vm_page_enqueue(int queue, vm_page_t m)
1783 {
1784 struct vpgqueues *vpq;
1785
1786 vpq = &vm_page_queues[queue];
1787 m->queue = queue;
1788 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1789 ++*vpq->cnt;
1790 }
1791
1792 /*
1793 * vm_page_activate:
1794 *
1795 * Put the specified page on the active list (if appropriate).
1796 * Ensure that act_count is at least ACT_INIT but do not otherwise
1797 * mess with it.
1798 *
1799 * The page must be locked.
1800 */
1801 void
1802 vm_page_activate(vm_page_t m)
1803 {
1804 int queue;
1805
1806 vm_page_lock_assert(m, MA_OWNED);
1807 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1808 if ((queue = m->queue) != PQ_ACTIVE) {
1809 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1810 if (m->act_count < ACT_INIT)
1811 m->act_count = ACT_INIT;
1812 vm_page_lock_queues();
1813 if (queue != PQ_NONE)
1814 vm_page_queue_remove(queue, m);
1815 vm_page_enqueue(PQ_ACTIVE, m);
1816 vm_page_unlock_queues();
1817 } else
1818 KASSERT(queue == PQ_NONE,
1819 ("vm_page_activate: wired page %p is queued", m));
1820 } else {
1821 if (m->act_count < ACT_INIT)
1822 m->act_count = ACT_INIT;
1823 }
1824 }
1825
1826 /*
1827 * vm_page_free_wakeup:
1828 *
1829 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
1830 * routine is called when a page has been added to the cache or free
1831 * queues.
1832 *
1833 * The page queues must be locked.
1834 */
1835 static inline void
1836 vm_page_free_wakeup(void)
1837 {
1838
1839 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1840 /*
1841 * if pageout daemon needs pages, then tell it that there are
1842 * some free.
1843 */
1844 if (vm_pageout_pages_needed &&
1845 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1846 wakeup(&vm_pageout_pages_needed);
1847 vm_pageout_pages_needed = 0;
1848 }
1849 /*
1850 * wakeup processes that are waiting on memory if we hit a
1851 * high water mark. And wakeup scheduler process if we have
1852 * lots of memory. this process will swapin processes.
1853 */
1854 if (vm_pages_needed && !vm_page_count_min()) {
1855 vm_pages_needed = 0;
1856 wakeup(&cnt.v_free_count);
1857 }
1858 }
1859
1860 /*
1861 * vm_page_free_toq:
1862 *
1863 * Returns the given page to the free list,
1864 * disassociating it with any VM object.
1865 *
1866 * The object must be locked. The page must be locked if it is managed.
1867 */
1868 void
1869 vm_page_free_toq(vm_page_t m)
1870 {
1871
1872 if ((m->oflags & VPO_UNMANAGED) == 0) {
1873 vm_page_lock_assert(m, MA_OWNED);
1874 KASSERT(!pmap_page_is_mapped(m),
1875 ("vm_page_free_toq: freeing mapped page %p", m));
1876 }
1877 PCPU_INC(cnt.v_tfree);
1878
1879 if (VM_PAGE_IS_FREE(m))
1880 panic("vm_page_free: freeing free page %p", m);
1881 else if (m->busy != 0)
1882 panic("vm_page_free: freeing busy page %p", m);
1883
1884 /*
1885 * Unqueue, then remove page. Note that we cannot destroy
1886 * the page here because we do not want to call the pager's
1887 * callback routine until after we've put the page on the
1888 * appropriate free queue.
1889 */
1890 if ((m->oflags & VPO_UNMANAGED) == 0)
1891 vm_pageq_remove(m);
1892 vm_page_remove(m);
1893
1894 /*
1895 * If fictitious remove object association and
1896 * return, otherwise delay object association removal.
1897 */
1898 if ((m->flags & PG_FICTITIOUS) != 0) {
1899 return;
1900 }
1901
1902 m->valid = 0;
1903 vm_page_undirty(m);
1904
1905 if (m->wire_count != 0)
1906 panic("vm_page_free: freeing wired page %p", m);
1907 if (m->hold_count != 0) {
1908 m->flags &= ~PG_ZERO;
1909 vm_page_lock_queues();
1910 vm_page_enqueue(PQ_HOLD, m);
1911 vm_page_unlock_queues();
1912 } else {
1913 /*
1914 * Restore the default memory attribute to the page.
1915 */
1916 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
1917 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
1918
1919 /*
1920 * Insert the page into the physical memory allocator's
1921 * cache/free page queues.
1922 */
1923 mtx_lock(&vm_page_queue_free_mtx);
1924 m->flags |= PG_FREE;
1925 cnt.v_free_count++;
1926 #if VM_NRESERVLEVEL > 0
1927 if (!vm_reserv_free_page(m))
1928 #else
1929 if (TRUE)
1930 #endif
1931 vm_phys_free_pages(m, 0);
1932 if ((m->flags & PG_ZERO) != 0)
1933 ++vm_page_zero_count;
1934 else
1935 vm_page_zero_idle_wakeup();
1936 vm_page_free_wakeup();
1937 mtx_unlock(&vm_page_queue_free_mtx);
1938 }
1939 }
1940
1941 /*
1942 * vm_page_wire:
1943 *
1944 * Mark this page as wired down by yet
1945 * another map, removing it from paging queues
1946 * as necessary.
1947 *
1948 * If the page is fictitious, then its wire count must remain one.
1949 *
1950 * The page must be locked.
1951 */
1952 void
1953 vm_page_wire(vm_page_t m)
1954 {
1955
1956 /*
1957 * Only bump the wire statistics if the page is not already wired,
1958 * and only unqueue the page if it is on some queue (if it is unmanaged
1959 * it is already off the queues).
1960 */
1961 vm_page_lock_assert(m, MA_OWNED);
1962 if ((m->flags & PG_FICTITIOUS) != 0) {
1963 KASSERT(m->wire_count == 1,
1964 ("vm_page_wire: fictitious page %p's wire count isn't one",
1965 m));
1966 return;
1967 }
1968 if (m->wire_count == 0) {
1969 if ((m->oflags & VPO_UNMANAGED) == 0)
1970 vm_pageq_remove(m);
1971 atomic_add_int(&cnt.v_wire_count, 1);
1972 }
1973 m->wire_count++;
1974 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1975 }
1976
1977 /*
1978 * vm_page_unwire:
1979 *
1980 * Release one wiring of the specified page, potentially enabling it to be
1981 * paged again. If paging is enabled, then the value of the parameter
1982 * "activate" determines to which queue the page is added. If "activate" is
1983 * non-zero, then the page is added to the active queue. Otherwise, it is
1984 * added to the inactive queue.
1985 *
1986 * However, unless the page belongs to an object, it is not enqueued because
1987 * it cannot be paged out.
1988 *
1989 * If a page is fictitious, then its wire count must alway be one.
1990 *
1991 * A managed page must be locked.
1992 */
1993 void
1994 vm_page_unwire(vm_page_t m, int activate)
1995 {
1996
1997 if ((m->oflags & VPO_UNMANAGED) == 0)
1998 vm_page_lock_assert(m, MA_OWNED);
1999 if ((m->flags & PG_FICTITIOUS) != 0) {
2000 KASSERT(m->wire_count == 1,
2001 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2002 return;
2003 }
2004 if (m->wire_count > 0) {
2005 m->wire_count--;
2006 if (m->wire_count == 0) {
2007 atomic_subtract_int(&cnt.v_wire_count, 1);
2008 if ((m->oflags & VPO_UNMANAGED) != 0 ||
2009 m->object == NULL)
2010 return;
2011 if (!activate)
2012 m->flags &= ~PG_WINATCFLS;
2013 vm_page_lock_queues();
2014 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2015 vm_page_unlock_queues();
2016 }
2017 } else
2018 panic("vm_page_unwire: page %p's wire count is zero", m);
2019 }
2020
2021 /*
2022 * Move the specified page to the inactive queue.
2023 *
2024 * Many pages placed on the inactive queue should actually go
2025 * into the cache, but it is difficult to figure out which. What
2026 * we do instead, if the inactive target is well met, is to put
2027 * clean pages at the head of the inactive queue instead of the tail.
2028 * This will cause them to be moved to the cache more quickly and
2029 * if not actively re-referenced, reclaimed more quickly. If we just
2030 * stick these pages at the end of the inactive queue, heavy filesystem
2031 * meta-data accesses can cause an unnecessary paging load on memory bound
2032 * processes. This optimization causes one-time-use metadata to be
2033 * reused more quickly.
2034 *
2035 * Normally athead is 0 resulting in LRU operation. athead is set
2036 * to 1 if we want this page to be 'as if it were placed in the cache',
2037 * except without unmapping it from the process address space.
2038 *
2039 * The page must be locked.
2040 */
2041 static inline void
2042 _vm_page_deactivate(vm_page_t m, int athead)
2043 {
2044 int queue;
2045
2046 vm_page_lock_assert(m, MA_OWNED);
2047
2048 /*
2049 * Ignore if already inactive.
2050 */
2051 if ((queue = m->queue) == PQ_INACTIVE)
2052 return;
2053 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2054 m->flags &= ~PG_WINATCFLS;
2055 vm_page_lock_queues();
2056 if (queue != PQ_NONE)
2057 vm_page_queue_remove(queue, m);
2058 if (athead)
2059 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m,
2060 pageq);
2061 else
2062 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m,
2063 pageq);
2064 m->queue = PQ_INACTIVE;
2065 cnt.v_inactive_count++;
2066 vm_page_unlock_queues();
2067 }
2068 }
2069
2070 /*
2071 * Move the specified page to the inactive queue.
2072 *
2073 * The page must be locked.
2074 */
2075 void
2076 vm_page_deactivate(vm_page_t m)
2077 {
2078
2079 _vm_page_deactivate(m, 0);
2080 }
2081
2082 /*
2083 * vm_page_try_to_cache:
2084 *
2085 * Returns 0 on failure, 1 on success
2086 */
2087 int
2088 vm_page_try_to_cache(vm_page_t m)
2089 {
2090
2091 vm_page_lock_assert(m, MA_OWNED);
2092 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2093 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2094 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2095 return (0);
2096 pmap_remove_all(m);
2097 if (m->dirty)
2098 return (0);
2099 vm_page_cache(m);
2100 return (1);
2101 }
2102
2103 /*
2104 * vm_page_try_to_free()
2105 *
2106 * Attempt to free the page. If we cannot free it, we do nothing.
2107 * 1 is returned on success, 0 on failure.
2108 */
2109 int
2110 vm_page_try_to_free(vm_page_t m)
2111 {
2112
2113 vm_page_lock_assert(m, MA_OWNED);
2114 if (m->object != NULL)
2115 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2116 if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2117 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2118 return (0);
2119 pmap_remove_all(m);
2120 if (m->dirty)
2121 return (0);
2122 vm_page_free(m);
2123 return (1);
2124 }
2125
2126 /*
2127 * vm_page_cache
2128 *
2129 * Put the specified page onto the page cache queue (if appropriate).
2130 *
2131 * The object and page must be locked.
2132 */
2133 void
2134 vm_page_cache(vm_page_t m)
2135 {
2136 vm_object_t object;
2137 vm_page_t next, prev, root;
2138
2139 vm_page_lock_assert(m, MA_OWNED);
2140 object = m->object;
2141 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2142 if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2143 m->hold_count || m->wire_count)
2144 panic("vm_page_cache: attempting to cache busy page");
2145 pmap_remove_all(m);
2146 if (m->dirty != 0)
2147 panic("vm_page_cache: page %p is dirty", m);
2148 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2149 (object->type == OBJT_SWAP &&
2150 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2151 /*
2152 * Hypothesis: A cache-elgible page belonging to a
2153 * default object or swap object but without a backing
2154 * store must be zero filled.
2155 */
2156 vm_page_free(m);
2157 return;
2158 }
2159 KASSERT((m->flags & PG_CACHED) == 0,
2160 ("vm_page_cache: page %p is already cached", m));
2161 PCPU_INC(cnt.v_tcached);
2162
2163 /*
2164 * Remove the page from the paging queues.
2165 */
2166 vm_pageq_remove(m);
2167
2168 /*
2169 * Remove the page from the object's collection of resident
2170 * pages.
2171 */
2172 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
2173 /*
2174 * Since the page's successor in the list is also its parent
2175 * in the tree, its right subtree must be empty.
2176 */
2177 next->left = m->left;
2178 KASSERT(m->right == NULL,
2179 ("vm_page_cache: page %p has right child", m));
2180 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
2181 prev->right == m) {
2182 /*
2183 * Since the page's predecessor in the list is also its parent
2184 * in the tree, its left subtree must be empty.
2185 */
2186 KASSERT(m->left == NULL,
2187 ("vm_page_cache: page %p has left child", m));
2188 prev->right = m->right;
2189 } else {
2190 if (m != object->root)
2191 vm_page_splay(m->pindex, object->root);
2192 if (m->left == NULL)
2193 root = m->right;
2194 else if (m->right == NULL)
2195 root = m->left;
2196 else {
2197 /*
2198 * Move the page's successor to the root, because
2199 * pages are usually removed in ascending order.
2200 */
2201 if (m->right != next)
2202 vm_page_splay(m->pindex, m->right);
2203 next->left = m->left;
2204 root = next;
2205 }
2206 object->root = root;
2207 }
2208 TAILQ_REMOVE(&object->memq, m, listq);
2209 object->resident_page_count--;
2210
2211 /*
2212 * Restore the default memory attribute to the page.
2213 */
2214 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2215 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2216
2217 /*
2218 * Insert the page into the object's collection of cached pages
2219 * and the physical memory allocator's cache/free page queues.
2220 */
2221 m->flags &= ~PG_ZERO;
2222 mtx_lock(&vm_page_queue_free_mtx);
2223 m->flags |= PG_CACHED;
2224 cnt.v_cache_count++;
2225 root = object->cache;
2226 if (root == NULL) {
2227 m->left = NULL;
2228 m->right = NULL;
2229 } else {
2230 root = vm_page_splay(m->pindex, root);
2231 if (m->pindex < root->pindex) {
2232 m->left = root->left;
2233 m->right = root;
2234 root->left = NULL;
2235 } else if (__predict_false(m->pindex == root->pindex))
2236 panic("vm_page_cache: offset already cached");
2237 else {
2238 m->right = root->right;
2239 m->left = root;
2240 root->right = NULL;
2241 }
2242 }
2243 object->cache = m;
2244 #if VM_NRESERVLEVEL > 0
2245 if (!vm_reserv_free_page(m)) {
2246 #else
2247 if (TRUE) {
2248 #endif
2249 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2250 vm_phys_free_pages(m, 0);
2251 }
2252 vm_page_free_wakeup();
2253 mtx_unlock(&vm_page_queue_free_mtx);
2254
2255 /*
2256 * Increment the vnode's hold count if this is the object's only
2257 * cached page. Decrement the vnode's hold count if this was
2258 * the object's only resident page.
2259 */
2260 if (object->type == OBJT_VNODE) {
2261 if (root == NULL && object->resident_page_count != 0)
2262 vhold(object->handle);
2263 else if (root != NULL && object->resident_page_count == 0)
2264 vdrop(object->handle);
2265 }
2266 }
2267
2268 /*
2269 * vm_page_dontneed
2270 *
2271 * Cache, deactivate, or do nothing as appropriate. This routine
2272 * is typically used by madvise() MADV_DONTNEED.
2273 *
2274 * Generally speaking we want to move the page into the cache so
2275 * it gets reused quickly. However, this can result in a silly syndrome
2276 * due to the page recycling too quickly. Small objects will not be
2277 * fully cached. On the otherhand, if we move the page to the inactive
2278 * queue we wind up with a problem whereby very large objects
2279 * unnecessarily blow away our inactive and cache queues.
2280 *
2281 * The solution is to move the pages based on a fixed weighting. We
2282 * either leave them alone, deactivate them, or move them to the cache,
2283 * where moving them to the cache has the highest weighting.
2284 * By forcing some pages into other queues we eventually force the
2285 * system to balance the queues, potentially recovering other unrelated
2286 * space from active. The idea is to not force this to happen too
2287 * often.
2288 *
2289 * The object and page must be locked.
2290 */
2291 void
2292 vm_page_dontneed(vm_page_t m)
2293 {
2294 int dnw;
2295 int head;
2296
2297 vm_page_lock_assert(m, MA_OWNED);
2298 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2299 dnw = PCPU_GET(dnweight);
2300 PCPU_INC(dnweight);
2301
2302 /*
2303 * Occasionally leave the page alone.
2304 */
2305 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2306 if (m->act_count >= ACT_INIT)
2307 --m->act_count;
2308 return;
2309 }
2310
2311 /*
2312 * Clear any references to the page. Otherwise, the page daemon will
2313 * immediately reactivate the page.
2314 *
2315 * Perform the pmap_clear_reference() first. Otherwise, a concurrent
2316 * pmap operation, such as pmap_remove(), could clear a reference in
2317 * the pmap and set PGA_REFERENCED on the page before the
2318 * pmap_clear_reference() had completed. Consequently, the page would
2319 * appear referenced based upon an old reference that occurred before
2320 * this function ran.
2321 */
2322 pmap_clear_reference(m);
2323 vm_page_aflag_clear(m, PGA_REFERENCED);
2324
2325 if (m->dirty == 0 && pmap_is_modified(m))
2326 vm_page_dirty(m);
2327
2328 if (m->dirty || (dnw & 0x0070) == 0) {
2329 /*
2330 * Deactivate the page 3 times out of 32.
2331 */
2332 head = 0;
2333 } else {
2334 /*
2335 * Cache the page 28 times out of every 32. Note that
2336 * the page is deactivated instead of cached, but placed
2337 * at the head of the queue instead of the tail.
2338 */
2339 head = 1;
2340 }
2341 _vm_page_deactivate(m, head);
2342 }
2343
2344 /*
2345 * Grab a page, waiting until we are waken up due to the page
2346 * changing state. We keep on waiting, if the page continues
2347 * to be in the object. If the page doesn't exist, first allocate it
2348 * and then conditionally zero it.
2349 *
2350 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended
2351 * to facilitate its eventual removal.
2352 *
2353 * This routine may sleep.
2354 *
2355 * The object must be locked on entry. The lock will, however, be released
2356 * and reacquired if the routine sleeps.
2357 */
2358 vm_page_t
2359 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2360 {
2361 vm_page_t m;
2362
2363 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2364 KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2365 ("vm_page_grab: VM_ALLOC_RETRY is required"));
2366 retrylookup:
2367 if ((m = vm_page_lookup(object, pindex)) != NULL) {
2368 if ((m->oflags & VPO_BUSY) != 0 ||
2369 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2370 /*
2371 * Reference the page before unlocking and
2372 * sleeping so that the page daemon is less
2373 * likely to reclaim it.
2374 */
2375 vm_page_aflag_set(m, PGA_REFERENCED);
2376 vm_page_sleep(m, "pgrbwt");
2377 goto retrylookup;
2378 } else {
2379 if ((allocflags & VM_ALLOC_WIRED) != 0) {
2380 vm_page_lock(m);
2381 vm_page_wire(m);
2382 vm_page_unlock(m);
2383 }
2384 if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2385 vm_page_busy(m);
2386 return (m);
2387 }
2388 }
2389 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2390 VM_ALLOC_IGN_SBUSY));
2391 if (m == NULL) {
2392 VM_OBJECT_UNLOCK(object);
2393 VM_WAIT;
2394 VM_OBJECT_LOCK(object);
2395 goto retrylookup;
2396 } else if (m->valid != 0)
2397 return (m);
2398 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2399 pmap_zero_page(m);
2400 return (m);
2401 }
2402
2403 /*
2404 * Mapping function for valid or dirty bits in a page.
2405 *
2406 * Inputs are required to range within a page.
2407 */
2408 vm_page_bits_t
2409 vm_page_bits(int base, int size)
2410 {
2411 int first_bit;
2412 int last_bit;
2413
2414 KASSERT(
2415 base + size <= PAGE_SIZE,
2416 ("vm_page_bits: illegal base/size %d/%d", base, size)
2417 );
2418
2419 if (size == 0) /* handle degenerate case */
2420 return (0);
2421
2422 first_bit = base >> DEV_BSHIFT;
2423 last_bit = (base + size - 1) >> DEV_BSHIFT;
2424
2425 return (((vm_page_bits_t)2 << last_bit) -
2426 ((vm_page_bits_t)1 << first_bit));
2427 }
2428
2429 /*
2430 * vm_page_set_valid:
2431 *
2432 * Sets portions of a page valid. The arguments are expected
2433 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2434 * of any partial chunks touched by the range. The invalid portion of
2435 * such chunks will be zeroed.
2436 *
2437 * (base + size) must be less then or equal to PAGE_SIZE.
2438 */
2439 void
2440 vm_page_set_valid(vm_page_t m, int base, int size)
2441 {
2442 int endoff, frag;
2443
2444 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2445 if (size == 0) /* handle degenerate case */
2446 return;
2447
2448 /*
2449 * If the base is not DEV_BSIZE aligned and the valid
2450 * bit is clear, we have to zero out a portion of the
2451 * first block.
2452 */
2453 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2454 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2455 pmap_zero_page_area(m, frag, base - frag);
2456
2457 /*
2458 * If the ending offset is not DEV_BSIZE aligned and the
2459 * valid bit is clear, we have to zero out a portion of
2460 * the last block.
2461 */
2462 endoff = base + size;
2463 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2464 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2465 pmap_zero_page_area(m, endoff,
2466 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2467
2468 /*
2469 * Assert that no previously invalid block that is now being validated
2470 * is already dirty.
2471 */
2472 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2473 ("vm_page_set_valid: page %p is dirty", m));
2474
2475 /*
2476 * Set valid bits inclusive of any overlap.
2477 */
2478 m->valid |= vm_page_bits(base, size);
2479 }
2480
2481 /*
2482 * Clear the given bits from the specified page's dirty field.
2483 */
2484 static __inline void
2485 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2486 {
2487 uintptr_t addr;
2488 #if PAGE_SIZE < 16384
2489 int shift;
2490 #endif
2491
2492 /*
2493 * If the object is locked and the page is neither VPO_BUSY nor
2494 * write mapped, then the page's dirty field cannot possibly be
2495 * set by a concurrent pmap operation.
2496 */
2497 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2498 if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m))
2499 m->dirty &= ~pagebits;
2500 else {
2501 /*
2502 * The pmap layer can call vm_page_dirty() without
2503 * holding a distinguished lock. The combination of
2504 * the object's lock and an atomic operation suffice
2505 * to guarantee consistency of the page dirty field.
2506 *
2507 * For PAGE_SIZE == 32768 case, compiler already
2508 * properly aligns the dirty field, so no forcible
2509 * alignment is needed. Only require existence of
2510 * atomic_clear_64 when page size is 32768.
2511 */
2512 addr = (uintptr_t)&m->dirty;
2513 #if PAGE_SIZE == 32768
2514 atomic_clear_64((uint64_t *)addr, pagebits);
2515 #elif PAGE_SIZE == 16384
2516 atomic_clear_32((uint32_t *)addr, pagebits);
2517 #else /* PAGE_SIZE <= 8192 */
2518 /*
2519 * Use a trick to perform a 32-bit atomic on the
2520 * containing aligned word, to not depend on the existence
2521 * of atomic_clear_{8, 16}.
2522 */
2523 shift = addr & (sizeof(uint32_t) - 1);
2524 #if BYTE_ORDER == BIG_ENDIAN
2525 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2526 #else
2527 shift *= NBBY;
2528 #endif
2529 addr &= ~(sizeof(uint32_t) - 1);
2530 atomic_clear_32((uint32_t *)addr, pagebits << shift);
2531 #endif /* PAGE_SIZE */
2532 }
2533 }
2534
2535 /*
2536 * vm_page_set_validclean:
2537 *
2538 * Sets portions of a page valid and clean. The arguments are expected
2539 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2540 * of any partial chunks touched by the range. The invalid portion of
2541 * such chunks will be zero'd.
2542 *
2543 * (base + size) must be less then or equal to PAGE_SIZE.
2544 */
2545 void
2546 vm_page_set_validclean(vm_page_t m, int base, int size)
2547 {
2548 vm_page_bits_t oldvalid, pagebits;
2549 int endoff, frag;
2550
2551 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2552 if (size == 0) /* handle degenerate case */
2553 return;
2554
2555 /*
2556 * If the base is not DEV_BSIZE aligned and the valid
2557 * bit is clear, we have to zero out a portion of the
2558 * first block.
2559 */
2560 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2561 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2562 pmap_zero_page_area(m, frag, base - frag);
2563
2564 /*
2565 * If the ending offset is not DEV_BSIZE aligned and the
2566 * valid bit is clear, we have to zero out a portion of
2567 * the last block.
2568 */
2569 endoff = base + size;
2570 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2571 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2572 pmap_zero_page_area(m, endoff,
2573 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2574
2575 /*
2576 * Set valid, clear dirty bits. If validating the entire
2577 * page we can safely clear the pmap modify bit. We also
2578 * use this opportunity to clear the VPO_NOSYNC flag. If a process
2579 * takes a write fault on a MAP_NOSYNC memory area the flag will
2580 * be set again.
2581 *
2582 * We set valid bits inclusive of any overlap, but we can only
2583 * clear dirty bits for DEV_BSIZE chunks that are fully within
2584 * the range.
2585 */
2586 oldvalid = m->valid;
2587 pagebits = vm_page_bits(base, size);
2588 m->valid |= pagebits;
2589 #if 0 /* NOT YET */
2590 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2591 frag = DEV_BSIZE - frag;
2592 base += frag;
2593 size -= frag;
2594 if (size < 0)
2595 size = 0;
2596 }
2597 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2598 #endif
2599 if (base == 0 && size == PAGE_SIZE) {
2600 /*
2601 * The page can only be modified within the pmap if it is
2602 * mapped, and it can only be mapped if it was previously
2603 * fully valid.
2604 */
2605 if (oldvalid == VM_PAGE_BITS_ALL)
2606 /*
2607 * Perform the pmap_clear_modify() first. Otherwise,
2608 * a concurrent pmap operation, such as
2609 * pmap_protect(), could clear a modification in the
2610 * pmap and set the dirty field on the page before
2611 * pmap_clear_modify() had begun and after the dirty
2612 * field was cleared here.
2613 */
2614 pmap_clear_modify(m);
2615 m->dirty = 0;
2616 m->oflags &= ~VPO_NOSYNC;
2617 } else if (oldvalid != VM_PAGE_BITS_ALL)
2618 m->dirty &= ~pagebits;
2619 else
2620 vm_page_clear_dirty_mask(m, pagebits);
2621 }
2622
2623 void
2624 vm_page_clear_dirty(vm_page_t m, int base, int size)
2625 {
2626
2627 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2628 }
2629
2630 /*
2631 * vm_page_set_invalid:
2632 *
2633 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2634 * valid and dirty bits for the effected areas are cleared.
2635 */
2636 void
2637 vm_page_set_invalid(vm_page_t m, int base, int size)
2638 {
2639 vm_page_bits_t bits;
2640
2641 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2642 KASSERT((m->oflags & VPO_BUSY) == 0,
2643 ("vm_page_set_invalid: page %p is busy", m));
2644 bits = vm_page_bits(base, size);
2645 if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2646 pmap_remove_all(m);
2647 KASSERT(!pmap_page_is_mapped(m),
2648 ("vm_page_set_invalid: page %p is mapped", m));
2649 m->valid &= ~bits;
2650 m->dirty &= ~bits;
2651 }
2652
2653 /*
2654 * vm_page_zero_invalid()
2655 *
2656 * The kernel assumes that the invalid portions of a page contain
2657 * garbage, but such pages can be mapped into memory by user code.
2658 * When this occurs, we must zero out the non-valid portions of the
2659 * page so user code sees what it expects.
2660 *
2661 * Pages are most often semi-valid when the end of a file is mapped
2662 * into memory and the file's size is not page aligned.
2663 */
2664 void
2665 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2666 {
2667 int b;
2668 int i;
2669
2670 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2671 /*
2672 * Scan the valid bits looking for invalid sections that
2673 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
2674 * valid bit may be set ) have already been zerod by
2675 * vm_page_set_validclean().
2676 */
2677 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2678 if (i == (PAGE_SIZE / DEV_BSIZE) ||
2679 (m->valid & ((vm_page_bits_t)1 << i))) {
2680 if (i > b) {
2681 pmap_zero_page_area(m,
2682 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2683 }
2684 b = i + 1;
2685 }
2686 }
2687
2688 /*
2689 * setvalid is TRUE when we can safely set the zero'd areas
2690 * as being valid. We can do this if there are no cache consistancy
2691 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
2692 */
2693 if (setvalid)
2694 m->valid = VM_PAGE_BITS_ALL;
2695 }
2696
2697 /*
2698 * vm_page_is_valid:
2699 *
2700 * Is (partial) page valid? Note that the case where size == 0
2701 * will return FALSE in the degenerate case where the page is
2702 * entirely invalid, and TRUE otherwise.
2703 */
2704 int
2705 vm_page_is_valid(vm_page_t m, int base, int size)
2706 {
2707 vm_page_bits_t bits;
2708
2709 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2710 bits = vm_page_bits(base, size);
2711 if (m->valid && ((m->valid & bits) == bits))
2712 return 1;
2713 else
2714 return 0;
2715 }
2716
2717 /*
2718 * Set the page's dirty bits if the page is modified.
2719 */
2720 void
2721 vm_page_test_dirty(vm_page_t m)
2722 {
2723
2724 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2725 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2726 vm_page_dirty(m);
2727 }
2728
2729 void
2730 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2731 {
2732
2733 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2734 }
2735
2736 void
2737 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2738 {
2739
2740 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2741 }
2742
2743 int
2744 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2745 {
2746
2747 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2748 }
2749
2750 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2751 void
2752 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2753 {
2754
2755 mtx_assert_(vm_page_lockptr(m), a, file, line);
2756 }
2757 #endif
2758
2759 int so_zerocp_fullpage = 0;
2760
2761 /*
2762 * Replace the given page with a copy. The copied page assumes
2763 * the portion of the given page's "wire_count" that is not the
2764 * responsibility of this copy-on-write mechanism.
2765 *
2766 * The object containing the given page must have a non-zero
2767 * paging-in-progress count and be locked.
2768 */
2769 void
2770 vm_page_cowfault(vm_page_t m)
2771 {
2772 vm_page_t mnew;
2773 vm_object_t object;
2774 vm_pindex_t pindex;
2775
2776 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
2777 vm_page_lock_assert(m, MA_OWNED);
2778 object = m->object;
2779 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2780 KASSERT(object->paging_in_progress != 0,
2781 ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2782 object));
2783 pindex = m->pindex;
2784
2785 retry_alloc:
2786 pmap_remove_all(m);
2787 vm_page_remove(m);
2788 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2789 if (mnew == NULL) {
2790 vm_page_insert(m, object, pindex);
2791 vm_page_unlock(m);
2792 VM_OBJECT_UNLOCK(object);
2793 VM_WAIT;
2794 VM_OBJECT_LOCK(object);
2795 if (m == vm_page_lookup(object, pindex)) {
2796 vm_page_lock(m);
2797 goto retry_alloc;
2798 } else {
2799 /*
2800 * Page disappeared during the wait.
2801 */
2802 return;
2803 }
2804 }
2805
2806 if (m->cow == 0) {
2807 /*
2808 * check to see if we raced with an xmit complete when
2809 * waiting to allocate a page. If so, put things back
2810 * the way they were
2811 */
2812 vm_page_unlock(m);
2813 vm_page_lock(mnew);
2814 vm_page_free(mnew);
2815 vm_page_unlock(mnew);
2816 vm_page_insert(m, object, pindex);
2817 } else { /* clear COW & copy page */
2818 if (!so_zerocp_fullpage)
2819 pmap_copy_page(m, mnew);
2820 mnew->valid = VM_PAGE_BITS_ALL;
2821 vm_page_dirty(mnew);
2822 mnew->wire_count = m->wire_count - m->cow;
2823 m->wire_count = m->cow;
2824 vm_page_unlock(m);
2825 }
2826 }
2827
2828 void
2829 vm_page_cowclear(vm_page_t m)
2830 {
2831
2832 vm_page_lock_assert(m, MA_OWNED);
2833 if (m->cow) {
2834 m->cow--;
2835 /*
2836 * let vm_fault add back write permission lazily
2837 */
2838 }
2839 /*
2840 * sf_buf_free() will free the page, so we needn't do it here
2841 */
2842 }
2843
2844 int
2845 vm_page_cowsetup(vm_page_t m)
2846 {
2847
2848 vm_page_lock_assert(m, MA_OWNED);
2849 if ((m->flags & PG_FICTITIOUS) != 0 ||
2850 (m->oflags & VPO_UNMANAGED) != 0 ||
2851 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
2852 return (EBUSY);
2853 m->cow++;
2854 pmap_remove_write(m);
2855 VM_OBJECT_UNLOCK(m->object);
2856 return (0);
2857 }
2858
2859 #ifdef INVARIANTS
2860 void
2861 vm_page_object_lock_assert(vm_page_t m)
2862 {
2863
2864 /*
2865 * Certain of the page's fields may only be modified by the
2866 * holder of the containing object's lock or the setter of the
2867 * page's VPO_BUSY flag. Unfortunately, the setter of the
2868 * VPO_BUSY flag is not recorded, and thus cannot be checked
2869 * here.
2870 */
2871 if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
2872 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2873 }
2874 #endif
2875
2876 #include "opt_ddb.h"
2877 #ifdef DDB
2878 #include <sys/kernel.h>
2879
2880 #include <ddb/ddb.h>
2881
2882 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2883 {
2884 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
2885 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
2886 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
2887 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
2888 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
2889 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
2890 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
2891 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
2892 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
2893 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
2894 }
2895
2896 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2897 {
2898
2899 db_printf("PQ_FREE:");
2900 db_printf(" %d", cnt.v_free_count);
2901 db_printf("\n");
2902
2903 db_printf("PQ_CACHE:");
2904 db_printf(" %d", cnt.v_cache_count);
2905 db_printf("\n");
2906
2907 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2908 *vm_page_queues[PQ_ACTIVE].cnt,
2909 *vm_page_queues[PQ_INACTIVE].cnt);
2910 }
2911
2912 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
2913 {
2914 vm_page_t m;
2915 boolean_t phys;
2916
2917 if (!have_addr) {
2918 db_printf("show pginfo addr\n");
2919 return;
2920 }
2921
2922 phys = strchr(modif, 'p') != NULL;
2923 if (phys)
2924 m = PHYS_TO_VM_PAGE(addr);
2925 else
2926 m = (vm_page_t)addr;
2927 db_printf(
2928 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
2929 " af 0x%x of 0x%x f 0x%x act %d busy %d valid 0x%x dirty 0x%x\n",
2930 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
2931 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
2932 m->flags, m->act_count, m->busy, m->valid, m->dirty);
2933 }
2934 #endif /* DDB */
Cache object: b8176c28ac85a840865dac9047a77273
|