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 page queue lock is required when adding or removing a page from a
67 * page queue regardless of other locks or the busy state of a page.
68 *
69 * * In general, no thread besides the page daemon can acquire or
70 * hold more than one page queue lock at a time.
71 *
72 * * The page daemon can acquire and hold any pair of page queue
73 * locks in any order.
74 *
75 * - The object lock is required when inserting or removing
76 * pages from an object (vm_page_insert() or vm_page_remove()).
77 *
78 */
79
80/*
81 * Resident memory management module.
82 */
83
84#include <sys/cdefs.h>
85__FBSDID("$FreeBSD: head/sys/vm/vm_page.c 295222 2016-02-03 23:30:17Z glebius $");
86
87#include "opt_vm.h"
88
89#include <sys/param.h>
90#include <sys/systm.h>
91#include <sys/lock.h>
92#include <sys/kernel.h>
93#include <sys/limits.h>
94#include <sys/linker.h>
95#include <sys/malloc.h>
96#include <sys/mman.h>
97#include <sys/msgbuf.h>
98#include <sys/mutex.h>
99#include <sys/proc.h>
100#include <sys/rwlock.h>
101#include <sys/sbuf.h>
102#include <sys/sysctl.h>
103#include <sys/vmmeter.h>
104#include <sys/vnode.h>
105#include <sys/taskqueue.h>
106
107#include <vm/vm.h>
108#include <vm/pmap.h>
109#include <vm/vm_param.h>
110#include <vm/vm_kern.h>
111#include <vm/vm_object.h>
112#include <vm/vm_page.h>
113#include <vm/vm_pageout.h>
114#include <vm/vm_pager.h>
115#include <vm/vm_phys.h>
116#include <vm/vm_radix.h>
117#include <vm/vm_reserv.h>
118#include <vm/vm_extern.h>
119#include <vm/uma.h>
120#include <vm/uma_int.h>
121
122#include <machine/md_var.h>
123
124/*
125 * Associated with page of user-allocatable memory is a
126 * page structure.
127 */
128
129struct vm_domain vm_dom[MAXMEMDOM];
130struct mtx_padalign vm_page_queue_free_mtx;
131
132struct mtx_padalign pa_lock[PA_LOCK_COUNT];
133
134vm_page_t vm_page_array;
135long vm_page_array_size;
136long first_page;
137int vm_page_zero_count;
138
139static int boot_pages = UMA_BOOT_PAGES;
140SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
141 &boot_pages, 0,
142 "number of pages allocated for bootstrapping the VM system");
143
144static int pa_tryrelock_restart;
145SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
146 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
147
148static TAILQ_HEAD(, vm_page) blacklist_head;
149static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
150SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
151 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
152
153/* Is the page daemon waiting for free pages? */
154static int vm_pageout_pages_needed;
155
156static uma_zone_t fakepg_zone;
157
158static struct vnode *vm_page_alloc_init(vm_page_t m);
159static void vm_page_cache_turn_free(vm_page_t m);
160static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
161static void vm_page_enqueue(uint8_t queue, vm_page_t m);
162static void vm_page_free_wakeup(void);
163static void vm_page_init_fakepg(void *dummy);
164static int vm_page_insert_after(vm_page_t m, vm_object_t object,
165 vm_pindex_t pindex, vm_page_t mpred);
166static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
167 vm_page_t mpred);
168static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
169 vm_paddr_t high);
170
171SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
172
173static void
174vm_page_init_fakepg(void *dummy)
175{
176
177 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
178 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
179}
180
181/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
182#if PAGE_SIZE == 32768
183#ifdef CTASSERT
184CTASSERT(sizeof(u_long) >= 8);
185#endif
186#endif
187
188/*
189 * Try to acquire a physical address lock while a pmap is locked. If we
190 * fail to trylock we unlock and lock the pmap directly and cache the
191 * locked pa in *locked. The caller should then restart their loop in case
192 * the virtual to physical mapping has changed.
193 */
194int
195vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
196{
197 vm_paddr_t lockpa;
198
199 lockpa = *locked;
200 *locked = pa;
201 if (lockpa) {
202 PA_LOCK_ASSERT(lockpa, MA_OWNED);
203 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
204 return (0);
205 PA_UNLOCK(lockpa);
206 }
207 if (PA_TRYLOCK(pa))
208 return (0);
209 PMAP_UNLOCK(pmap);
210 atomic_add_int(&pa_tryrelock_restart, 1);
211 PA_LOCK(pa);
212 PMAP_LOCK(pmap);
213 return (EAGAIN);
214}
215
216/*
217 * vm_set_page_size:
218 *
219 * Sets the page size, perhaps based upon the memory
220 * size. Must be called before any use of page-size
221 * dependent functions.
222 */
223void
224vm_set_page_size(void)
225{
226 if (vm_cnt.v_page_size == 0)
227 vm_cnt.v_page_size = PAGE_SIZE;
228 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
229 panic("vm_set_page_size: page size not a power of two");
230}
231
232/*
233 * vm_page_blacklist_next:
234 *
235 * Find the next entry in the provided string of blacklist
236 * addresses. Entries are separated by space, comma, or newline.
237 * If an invalid integer is encountered then the rest of the
238 * string is skipped. Updates the list pointer to the next
239 * character, or NULL if the string is exhausted or invalid.
240 */
241static vm_paddr_t
242vm_page_blacklist_next(char **list, char *end)
243{
244 vm_paddr_t bad;
245 char *cp, *pos;
246
247 if (list == NULL || *list == NULL)
248 return (0);
249 if (**list =='\0') {
250 *list = NULL;
251 return (0);
252 }
253
254 /*
255 * If there's no end pointer then the buffer is coming from
256 * the kenv and we know it's null-terminated.
257 */
258 if (end == NULL)
259 end = *list + strlen(*list);
260
261 /* Ensure that strtoq() won't walk off the end */
262 if (*end != '\0') {
263 if (*end == '\n' || *end == ' ' || *end == ',')
264 *end = '\0';
265 else {
266 printf("Blacklist not terminated, skipping\n");
267 *list = NULL;
268 return (0);
269 }
270 }
271
272 for (pos = *list; *pos != '\0'; pos = cp) {
273 bad = strtoq(pos, &cp, 0);
274 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
275 if (bad == 0) {
276 if (++cp < end)
277 continue;
278 else
279 break;
280 }
281 } else
282 break;
283 if (*cp == '\0' || ++cp >= end)
284 *list = NULL;
285 else
286 *list = cp;
287 return (trunc_page(bad));
288 }
289 printf("Garbage in RAM blacklist, skipping\n");
290 *list = NULL;
291 return (0);
292}
293
294/*
295 * vm_page_blacklist_check:
296 *
297 * Iterate through the provided string of blacklist addresses, pulling
298 * each entry out of the physical allocator free list and putting it
299 * onto a list for reporting via the vm.page_blacklist sysctl.
300 */
301static void
302vm_page_blacklist_check(char *list, char *end)
303{
304 vm_paddr_t pa;
305 vm_page_t m;
306 char *next;
307 int ret;
308
309 next = list;
310 while (next != NULL) {
311 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
312 continue;
313 m = vm_phys_paddr_to_vm_page(pa);
314 if (m == NULL)
315 continue;
316 mtx_lock(&vm_page_queue_free_mtx);
317 ret = vm_phys_unfree_page(m);
318 mtx_unlock(&vm_page_queue_free_mtx);
319 if (ret == TRUE) {
320 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
321 if (bootverbose)
322 printf("Skipping page with pa 0x%jx\n",
323 (uintmax_t)pa);
324 }
325 }
326}
327
328/*
329 * vm_page_blacklist_load:
330 *
331 * Search for a special module named "ram_blacklist". It'll be a
332 * plain text file provided by the user via the loader directive
333 * of the same name.
334 */
335static void
336vm_page_blacklist_load(char **list, char **end)
337{
338 void *mod;
339 u_char *ptr;
340 u_int len;
341
342 mod = NULL;
343 ptr = NULL;
344
345 mod = preload_search_by_type("ram_blacklist");
346 if (mod != NULL) {
347 ptr = preload_fetch_addr(mod);
348 len = preload_fetch_size(mod);
349 }
350 *list = ptr;
351 if (ptr != NULL)
352 *end = ptr + len;
353 else
354 *end = NULL;
355 return;
356}
357
358static int
359sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
360{
361 vm_page_t m;
362 struct sbuf sbuf;
363 int error, first;
364
365 first = 1;
366 error = sysctl_wire_old_buffer(req, 0);
367 if (error != 0)
368 return (error);
369 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
370 TAILQ_FOREACH(m, &blacklist_head, listq) {
371 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
372 (uintmax_t)m->phys_addr);
373 first = 0;
374 }
375 error = sbuf_finish(&sbuf);
376 sbuf_delete(&sbuf);
377 return (error);
378}
379
380static void
381vm_page_domain_init(struct vm_domain *vmd)
382{
383 struct vm_pagequeue *pq;
384 int i;
385
386 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
387 "vm inactive pagequeue";
388 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
389 &vm_cnt.v_inactive_count;
390 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
391 "vm active pagequeue";
392 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
393 &vm_cnt.v_active_count;
394 vmd->vmd_page_count = 0;
395 vmd->vmd_free_count = 0;
396 vmd->vmd_segs = 0;
397 vmd->vmd_oom = FALSE;
398 vmd->vmd_pass = 0;
399 for (i = 0; i < PQ_COUNT; i++) {
400 pq = &vmd->vmd_pagequeues[i];
401 TAILQ_INIT(&pq->pq_pl);
402 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
403 MTX_DEF | MTX_DUPOK);
404 }
405}
406
407/*
408 * vm_page_startup:
409 *
410 * Initializes the resident memory module.
411 *
412 * Allocates memory for the page cells, and
413 * for the object/offset-to-page hash table headers.
414 * Each page cell is initialized and placed on the free list.
415 */
416vm_offset_t
417vm_page_startup(vm_offset_t vaddr)
418{
419 vm_offset_t mapped;
420 vm_paddr_t page_range;
421 vm_paddr_t new_end;
422 int i;
423 vm_paddr_t pa;
424 vm_paddr_t last_pa;
425 char *list, *listend;
426 vm_paddr_t end;
427 vm_paddr_t biggestsize;
428 vm_paddr_t low_water, high_water;
429 int biggestone;
430
431 biggestsize = 0;
432 biggestone = 0;
433 vaddr = round_page(vaddr);
434
435 for (i = 0; phys_avail[i + 1]; i += 2) {
436 phys_avail[i] = round_page(phys_avail[i]);
437 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
438 }
439
440 low_water = phys_avail[0];
441 high_water = phys_avail[1];
442
443 for (i = 0; i < vm_phys_nsegs; i++) {
444 if (vm_phys_segs[i].start < low_water)
445 low_water = vm_phys_segs[i].start;
446 if (vm_phys_segs[i].end > high_water)
447 high_water = vm_phys_segs[i].end;
448 }
449 for (i = 0; phys_avail[i + 1]; i += 2) {
450 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
451
452 if (size > biggestsize) {
453 biggestone = i;
454 biggestsize = size;
455 }
456 if (phys_avail[i] < low_water)
457 low_water = phys_avail[i];
458 if (phys_avail[i + 1] > high_water)
459 high_water = phys_avail[i + 1];
460 }
461
462 end = phys_avail[biggestone+1];
463
464 /*
465 * Initialize the page and queue locks.
466 */
467 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
468 for (i = 0; i < PA_LOCK_COUNT; i++)
469 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
470 for (i = 0; i < vm_ndomains; i++)
471 vm_page_domain_init(&vm_dom[i]);
472
473 /*
474 * Allocate memory for use when boot strapping the kernel memory
475 * allocator.
476 *
477 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
478 * manually fetch the value.
479 */
480 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
481 new_end = end - (boot_pages * UMA_SLAB_SIZE);
482 new_end = trunc_page(new_end);
483 mapped = pmap_map(&vaddr, new_end, end,
484 VM_PROT_READ | VM_PROT_WRITE);
485 bzero((void *)mapped, end - new_end);
486 uma_startup((void *)mapped, boot_pages);
487
488#if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
489 defined(__i386__) || defined(__mips__)
490 /*
491 * Allocate a bitmap to indicate that a random physical page
492 * needs to be included in a minidump.
493 *
494 * The amd64 port needs this to indicate which direct map pages
495 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
496 *
497 * However, i386 still needs this workspace internally within the
498 * minidump code. In theory, they are not needed on i386, but are
499 * included should the sf_buf code decide to use them.
500 */
501 last_pa = 0;
502 for (i = 0; dump_avail[i + 1] != 0; i += 2)
503 if (dump_avail[i + 1] > last_pa)
504 last_pa = dump_avail[i + 1];
505 page_range = last_pa / PAGE_SIZE;
506 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
507 new_end -= vm_page_dump_size;
508 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
509 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
510 bzero((void *)vm_page_dump, vm_page_dump_size);
511#endif
512#ifdef __amd64__
513 /*
514 * Request that the physical pages underlying the message buffer be
515 * included in a crash dump. Since the message buffer is accessed
516 * through the direct map, they are not automatically included.
517 */
518 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
519 last_pa = pa + round_page(msgbufsize);
520 while (pa < last_pa) {
521 dump_add_page(pa);
522 pa += PAGE_SIZE;
523 }
524#endif
525 /*
526 * Compute the number of pages of memory that will be available for
527 * use (taking into account the overhead of a page structure per
528 * page).
529 */
530 first_page = low_water / PAGE_SIZE;
531#ifdef VM_PHYSSEG_SPARSE
532 page_range = 0;
533 for (i = 0; i < vm_phys_nsegs; i++) {
534 page_range += atop(vm_phys_segs[i].end -
535 vm_phys_segs[i].start);
536 }
537 for (i = 0; phys_avail[i + 1] != 0; i += 2)
538 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
539#elif defined(VM_PHYSSEG_DENSE)
540 page_range = high_water / PAGE_SIZE - first_page;
541#else
542#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
543#endif
544 end = new_end;
545
546 /*
547 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
548 */
549 vaddr += PAGE_SIZE;
550
551 /*
552 * Initialize the mem entry structures now, and put them in the free
553 * queue.
554 */
555 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
556 mapped = pmap_map(&vaddr, new_end, end,
557 VM_PROT_READ | VM_PROT_WRITE);
558 vm_page_array = (vm_page_t) mapped;
559#if VM_NRESERVLEVEL > 0
560 /*
561 * Allocate memory for the reservation management system's data
562 * structures.
563 */
564 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
565#endif
566#if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
567 /*
568 * pmap_map on arm64, amd64, and mips can come out of the direct-map,
569 * not kvm like i386, so the pages must be tracked for a crashdump to
570 * include this data. This includes the vm_page_array and the early
571 * UMA bootstrap pages.
572 */
573 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
574 dump_add_page(pa);
575#endif
576 phys_avail[biggestone + 1] = new_end;
577
578 /*
579 * Add physical memory segments corresponding to the available
580 * physical pages.
581 */
582 for (i = 0; phys_avail[i + 1] != 0; i += 2)
583 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
584
585 /*
586 * Clear all of the page structures
587 */
588 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
589 for (i = 0; i < page_range; i++)
590 vm_page_array[i].order = VM_NFREEORDER;
591 vm_page_array_size = page_range;
592
593 /*
594 * Initialize the physical memory allocator.
595 */
596 vm_phys_init();
597
598 /*
599 * Add every available physical page that is not blacklisted to
600 * the free lists.
601 */
602 vm_cnt.v_page_count = 0;
603 vm_cnt.v_free_count = 0;
604 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
605 pa = phys_avail[i];
606 last_pa = phys_avail[i + 1];
607 while (pa < last_pa) {
608 vm_phys_add_page(pa);
609 pa += PAGE_SIZE;
610 }
611 }
612
613 TAILQ_INIT(&blacklist_head);
614 vm_page_blacklist_load(&list, &listend);
615 vm_page_blacklist_check(list, listend);
616
617 list = kern_getenv("vm.blacklist");
618 vm_page_blacklist_check(list, NULL);
619
620 freeenv(list);
621#if VM_NRESERVLEVEL > 0
622 /*
623 * Initialize the reservation management system.
624 */
625 vm_reserv_init();
626#endif
627 return (vaddr);
628}
629
630void
631vm_page_reference(vm_page_t m)
632{
633
634 vm_page_aflag_set(m, PGA_REFERENCED);
635}
636
637/*
638 * vm_page_busy_downgrade:
639 *
640 * Downgrade an exclusive busy page into a single shared busy page.
641 */
642void
643vm_page_busy_downgrade(vm_page_t m)
644{
645 u_int x;
646
647 vm_page_assert_xbusied(m);
648
649 for (;;) {
650 x = m->busy_lock;
651 x &= VPB_BIT_WAITERS;
652 if (atomic_cmpset_rel_int(&m->busy_lock,
653 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x))
654 break;
655 }
656}
657
658/*
659 * vm_page_sbusied:
660 *
661 * Return a positive value if the page is shared busied, 0 otherwise.
662 */
663int
664vm_page_sbusied(vm_page_t m)
665{
666 u_int x;
667
668 x = m->busy_lock;
669 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
670}
671
672/*
673 * vm_page_sunbusy:
674 *
675 * Shared unbusy a page.
676 */
677void
678vm_page_sunbusy(vm_page_t m)
679{
680 u_int x;
681
682 vm_page_assert_sbusied(m);
683
684 for (;;) {
685 x = m->busy_lock;
686 if (VPB_SHARERS(x) > 1) {
687 if (atomic_cmpset_int(&m->busy_lock, x,
688 x - VPB_ONE_SHARER))
689 break;
690 continue;
691 }
692 if ((x & VPB_BIT_WAITERS) == 0) {
693 KASSERT(x == VPB_SHARERS_WORD(1),
694 ("vm_page_sunbusy: invalid lock state"));
695 if (atomic_cmpset_int(&m->busy_lock,
696 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
697 break;
698 continue;
699 }
700 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
701 ("vm_page_sunbusy: invalid lock state for waiters"));
702
703 vm_page_lock(m);
704 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
705 vm_page_unlock(m);
706 continue;
707 }
708 wakeup(m);
709 vm_page_unlock(m);
710 break;
711 }
712}
713
714/*
715 * vm_page_busy_sleep:
716 *
717 * Sleep and release the page lock, using the page pointer as wchan.
718 * This is used to implement the hard-path of busying mechanism.
719 *
720 * The given page must be locked.
721 */
722void
723vm_page_busy_sleep(vm_page_t m, const char *wmesg)
724{
725 u_int x;
726
727 vm_page_lock_assert(m, MA_OWNED);
728
729 x = m->busy_lock;
730 if (x == VPB_UNBUSIED) {
731 vm_page_unlock(m);
732 return;
733 }
734 if ((x & VPB_BIT_WAITERS) == 0 &&
735 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) {
736 vm_page_unlock(m);
737 return;
738 }
739 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
740}
741
742/*
743 * vm_page_trysbusy:
744 *
745 * Try to shared busy a page.
746 * If the operation succeeds 1 is returned otherwise 0.
747 * The operation never sleeps.
748 */
749int
750vm_page_trysbusy(vm_page_t m)
751{
752 u_int x;
753
754 for (;;) {
755 x = m->busy_lock;
756 if ((x & VPB_BIT_SHARED) == 0)
757 return (0);
758 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
759 return (1);
760 }
761}
762
763/*
764 * vm_page_xunbusy_hard:
765 *
766 * Called after the first try the exclusive unbusy of a page failed.
767 * It is assumed that the waiters bit is on.
768 */
769void
770vm_page_xunbusy_hard(vm_page_t m)
771{
772
773 vm_page_assert_xbusied(m);
774
775 vm_page_lock(m);
776 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
777 wakeup(m);
778 vm_page_unlock(m);
779}
780
781/*
782 * vm_page_flash:
783 *
784 * Wakeup anyone waiting for the page.
785 * The ownership bits do not change.
786 *
787 * The given page must be locked.
788 */
789void
790vm_page_flash(vm_page_t m)
791{
792 u_int x;
793
794 vm_page_lock_assert(m, MA_OWNED);
795
796 for (;;) {
797 x = m->busy_lock;
798 if ((x & VPB_BIT_WAITERS) == 0)
799 return;
800 if (atomic_cmpset_int(&m->busy_lock, x,
801 x & (~VPB_BIT_WAITERS)))
802 break;
803 }
804 wakeup(m);
805}
806
807/*
808 * Keep page from being freed by the page daemon
809 * much of the same effect as wiring, except much lower
810 * overhead and should be used only for *very* temporary
811 * holding ("wiring").
812 */
813void
814vm_page_hold(vm_page_t mem)
815{
816
817 vm_page_lock_assert(mem, MA_OWNED);
818 mem->hold_count++;
819}
820
821void
822vm_page_unhold(vm_page_t mem)
823{
824
825 vm_page_lock_assert(mem, MA_OWNED);
826 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
827 --mem->hold_count;
828 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
829 vm_page_free_toq(mem);
830}
831
832/*
833 * vm_page_unhold_pages:
834 *
835 * Unhold each of the pages that is referenced by the given array.
836 */
837void
838vm_page_unhold_pages(vm_page_t *ma, int count)
839{
840 struct mtx *mtx, *new_mtx;
841
842 mtx = NULL;
843 for (; count != 0; count--) {
844 /*
845 * Avoid releasing and reacquiring the same page lock.
846 */
847 new_mtx = vm_page_lockptr(*ma);
848 if (mtx != new_mtx) {
849 if (mtx != NULL)
850 mtx_unlock(mtx);
851 mtx = new_mtx;
852 mtx_lock(mtx);
853 }
854 vm_page_unhold(*ma);
855 ma++;
856 }
857 if (mtx != NULL)
858 mtx_unlock(mtx);
859}
860
861vm_page_t
862PHYS_TO_VM_PAGE(vm_paddr_t pa)
863{
864 vm_page_t m;
865
866#ifdef VM_PHYSSEG_SPARSE
867 m = vm_phys_paddr_to_vm_page(pa);
868 if (m == NULL)
869 m = vm_phys_fictitious_to_vm_page(pa);
870 return (m);
871#elif defined(VM_PHYSSEG_DENSE)
872 long pi;
873
874 pi = atop(pa);
875 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
876 m = &vm_page_array[pi - first_page];
877 return (m);
878 }
879 return (vm_phys_fictitious_to_vm_page(pa));
880#else
881#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
882#endif
883}
884
885/*
886 * vm_page_getfake:
887 *
888 * Create a fictitious page with the specified physical address and
889 * memory attribute. The memory attribute is the only the machine-
890 * dependent aspect of a fictitious page that must be initialized.
891 */
892vm_page_t
893vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
894{
895 vm_page_t m;
896
897 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
898 vm_page_initfake(m, paddr, memattr);
899 return (m);
900}
901
902void
903vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
904{
905
906 if ((m->flags & PG_FICTITIOUS) != 0) {
907 /*
908 * The page's memattr might have changed since the
909 * previous initialization. Update the pmap to the
910 * new memattr.
911 */
912 goto memattr;
913 }
914 m->phys_addr = paddr;
915 m->queue = PQ_NONE;
916 /* Fictitious pages don't use "segind". */
917 m->flags = PG_FICTITIOUS;
918 /* Fictitious pages don't use "order" or "pool". */
919 m->oflags = VPO_UNMANAGED;
920 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
921 m->wire_count = 1;
922 pmap_page_init(m);
923memattr:
924 pmap_page_set_memattr(m, memattr);
925}
926
927/*
928 * vm_page_putfake:
929 *
930 * Release a fictitious page.
931 */
932void
933vm_page_putfake(vm_page_t m)
934{
935
936 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
937 KASSERT((m->flags & PG_FICTITIOUS) != 0,
938 ("vm_page_putfake: bad page %p", m));
939 uma_zfree(fakepg_zone, m);
940}
941
942/*
943 * vm_page_updatefake:
944 *
945 * Update the given fictitious page to the specified physical address and
946 * memory attribute.
947 */
948void
949vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
950{
951
952 KASSERT((m->flags & PG_FICTITIOUS) != 0,
953 ("vm_page_updatefake: bad page %p", m));
954 m->phys_addr = paddr;
955 pmap_page_set_memattr(m, memattr);
956}
957
958/*
959 * vm_page_free:
960 *
961 * Free a page.
962 */
963void
964vm_page_free(vm_page_t m)
965{
966
967 m->flags &= ~PG_ZERO;
968 vm_page_free_toq(m);
969}
970
971/*
972 * vm_page_free_zero:
973 *
974 * Free a page to the zerod-pages queue
975 */
976void
977vm_page_free_zero(vm_page_t m)
978{
979
980 m->flags |= PG_ZERO;
981 vm_page_free_toq(m);
982}
983
984/*
985 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
986 * array which was optionally read ahead or behind.
987 */
988void
989vm_page_readahead_finish(vm_page_t m)
990{
991
992 /* We shouldn't put invalid pages on queues. */
993 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
994
995 /*
996 * Since the page is not the actually needed one, whether it should
997 * be activated or deactivated is not obvious. Empirical results
998 * have shown that deactivating the page is usually the best choice,
999 * unless the page is wanted by another thread.
1000 */
1001 vm_page_lock(m);
1002 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1003 vm_page_activate(m);
1004 else
1005 vm_page_deactivate(m);
1006 vm_page_unlock(m);
1007 vm_page_xunbusy(m);
1008}
1009
1010/*
1011 * vm_page_sleep_if_busy:
1012 *
1013 * Sleep and release the page queues lock if the page is busied.
1014 * Returns TRUE if the thread slept.
1015 *
1016 * The given page must be unlocked and object containing it must
1017 * be locked.
1018 */
1019int
1020vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1021{
1022 vm_object_t obj;
1023
1024 vm_page_lock_assert(m, MA_NOTOWNED);
1025 VM_OBJECT_ASSERT_WLOCKED(m->object);
1026
1027 if (vm_page_busied(m)) {
1028 /*
1029 * The page-specific object must be cached because page
1030 * identity can change during the sleep, causing the
1031 * re-lock of a different object.
1032 * It is assumed that a reference to the object is already
1033 * held by the callers.
1034 */
1035 obj = m->object;
1036 vm_page_lock(m);
1037 VM_OBJECT_WUNLOCK(obj);
1038 vm_page_busy_sleep(m, msg);
1039 VM_OBJECT_WLOCK(obj);
1040 return (TRUE);
1041 }
1042 return (FALSE);
1043}
1044
1045/*
1046 * vm_page_dirty_KBI: [ internal use only ]
1047 *
1048 * Set all bits in the page's dirty field.
1049 *
1050 * The object containing the specified page must be locked if the
1051 * call is made from the machine-independent layer.
1052 *
1053 * See vm_page_clear_dirty_mask().
1054 *
1055 * This function should only be called by vm_page_dirty().
1056 */
1057void
1058vm_page_dirty_KBI(vm_page_t m)
1059{
1060
1061 /* These assertions refer to this operation by its public name. */
1062 KASSERT((m->flags & PG_CACHED) == 0,
1063 ("vm_page_dirty: page in cache!"));
1064 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1065 ("vm_page_dirty: page is invalid!"));
1066 m->dirty = VM_PAGE_BITS_ALL;
1067}
1068
1069/*
1070 * vm_page_insert: [ internal use only ]
1071 *
1072 * Inserts the given mem entry into the object and object list.
1073 *
1074 * The object must be locked.
1075 */
1076int
1077vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1078{
1079 vm_page_t mpred;
1080
1081 VM_OBJECT_ASSERT_WLOCKED(object);
1082 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1083 return (vm_page_insert_after(m, object, pindex, mpred));
1084}
1085
1086/*
1087 * vm_page_insert_after:
1088 *
1089 * Inserts the page "m" into the specified object at offset "pindex".
1090 *
1091 * The page "mpred" must immediately precede the offset "pindex" within
1092 * the specified object.
1093 *
1094 * The object must be locked.
1095 */
1096static int
1097vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1098 vm_page_t mpred)
1099{
1100 vm_pindex_t sidx;
1101 vm_object_t sobj;
1102 vm_page_t msucc;
1103
1104 VM_OBJECT_ASSERT_WLOCKED(object);
1105 KASSERT(m->object == NULL,
1106 ("vm_page_insert_after: page already inserted"));
1107 if (mpred != NULL) {
1108 KASSERT(mpred->object == object,
1109 ("vm_page_insert_after: object doesn't contain mpred"));
1110 KASSERT(mpred->pindex < pindex,
1111 ("vm_page_insert_after: mpred doesn't precede pindex"));
1112 msucc = TAILQ_NEXT(mpred, listq);
1113 } else
1114 msucc = TAILQ_FIRST(&object->memq);
1115 if (msucc != NULL)
1116 KASSERT(msucc->pindex > pindex,
1117 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1118
1119 /*
1120 * Record the object/offset pair in this page
1121 */
1122 sobj = m->object;
1123 sidx = m->pindex;
1124 m->object = object;
1125 m->pindex = pindex;
1126
1127 /*
1128 * Now link into the object's ordered list of backed pages.
1129 */
1130 if (vm_radix_insert(&object->rtree, m)) {
1131 m->object = sobj;
1132 m->pindex = sidx;
1133 return (1);
1134 }
1135 vm_page_insert_radixdone(m, object, mpred);
1136 return (0);
1137}
1138
1139/*
1140 * vm_page_insert_radixdone:
1141 *
1142 * Complete page "m" insertion into the specified object after the
1143 * radix trie hooking.
1144 *
1145 * The page "mpred" must precede the offset "m->pindex" within the
1146 * specified object.
1147 *
1148 * The object must be locked.
1149 */
1150static void
1151vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1152{
1153
1154 VM_OBJECT_ASSERT_WLOCKED(object);
1155 KASSERT(object != NULL && m->object == object,
1156 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1157 if (mpred != NULL) {
1158 KASSERT(mpred->object == object,
1159 ("vm_page_insert_after: object doesn't contain mpred"));
1160 KASSERT(mpred->pindex < m->pindex,
1161 ("vm_page_insert_after: mpred doesn't precede pindex"));
1162 }
1163
1164 if (mpred != NULL)
1165 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1166 else
1167 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1168
1169 /*
1170 * Show that the object has one more resident page.
1171 */
1172 object->resident_page_count++;
1173
1174 /*
1175 * Hold the vnode until the last page is released.
1176 */
1177 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1178 vhold(object->handle);
1179
1180 /*
1181 * Since we are inserting a new and possibly dirty page,
1182 * update the object's OBJ_MIGHTBEDIRTY flag.
1183 */
1184 if (pmap_page_is_write_mapped(m))
1185 vm_object_set_writeable_dirty(object);
1186}
1187
1188/*
1189 * vm_page_remove:
1190 *
1191 * Removes the given mem entry from the object/offset-page
1192 * table and the object page list, but do not invalidate/terminate
1193 * the backing store.
1194 *
1195 * The object must be locked. The page must be locked if it is managed.
1196 */
1197void
1198vm_page_remove(vm_page_t m)
1199{
1200 vm_object_t object;
1201 boolean_t lockacq;
1202
1203 if ((m->oflags & VPO_UNMANAGED) == 0)
1204 vm_page_lock_assert(m, MA_OWNED);
1205 if ((object = m->object) == NULL)
1206 return;
1207 VM_OBJECT_ASSERT_WLOCKED(object);
1208 if (vm_page_xbusied(m)) {
1209 lockacq = FALSE;
1210 if ((m->oflags & VPO_UNMANAGED) != 0 &&
1211 !mtx_owned(vm_page_lockptr(m))) {
1212 lockacq = TRUE;
1213 vm_page_lock(m);
1214 }
1215 vm_page_flash(m);
1216 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1217 if (lockacq)
1218 vm_page_unlock(m);
1219 }
1220
1221 /*
1222 * Now remove from the object's list of backed pages.
1223 */
1224 vm_radix_remove(&object->rtree, m->pindex);
1225 TAILQ_REMOVE(&object->memq, m, listq);
1226
1227 /*
1228 * And show that the object has one fewer resident page.
1229 */
1230 object->resident_page_count--;
1231
1232 /*
1233 * The vnode may now be recycled.
1234 */
1235 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1236 vdrop(object->handle);
1237
1238 m->object = NULL;
1239}
1240
1241/*
1242 * vm_page_lookup:
1243 *
1244 * Returns the page associated with the object/offset
1245 * pair specified; if none is found, NULL is returned.
1246 *
1247 * The object must be locked.
1248 */
1249vm_page_t
1250vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1251{
1252
1253 VM_OBJECT_ASSERT_LOCKED(object);
1254 return (vm_radix_lookup(&object->rtree, pindex));
1255}
1256
1257/*
1258 * vm_page_find_least:
1259 *
1260 * Returns the page associated with the object with least pindex
1261 * greater than or equal to the parameter pindex, or NULL.
1262 *
1263 * The object must be locked.
1264 */
1265vm_page_t
1266vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1267{
1268 vm_page_t m;
1269
1270 VM_OBJECT_ASSERT_LOCKED(object);
1271 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1272 m = vm_radix_lookup_ge(&object->rtree, pindex);
1273 return (m);
1274}
1275
1276/*
1277 * Returns the given page's successor (by pindex) within the object if it is
1278 * resident; if none is found, NULL is returned.
1279 *
1280 * The object must be locked.
1281 */
1282vm_page_t
1283vm_page_next(vm_page_t m)
1284{
1285 vm_page_t next;
1286
1287 VM_OBJECT_ASSERT_WLOCKED(m->object);
1288 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1289 next->pindex != m->pindex + 1)
1290 next = NULL;
1291 return (next);
1292}
1293
1294/*
1295 * Returns the given page's predecessor (by pindex) within the object if it is
1296 * resident; if none is found, NULL is returned.
1297 *
1298 * The object must be locked.
1299 */
1300vm_page_t
1301vm_page_prev(vm_page_t m)
1302{
1303 vm_page_t prev;
1304
1305 VM_OBJECT_ASSERT_WLOCKED(m->object);
1306 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1307 prev->pindex != m->pindex - 1)
1308 prev = NULL;
1309 return (prev);
1310}
1311
1312/*
1313 * Uses the page mnew as a replacement for an existing page at index
1314 * pindex which must be already present in the object.
1315 *
1316 * The existing page must not be on a paging queue.
1317 */
1318vm_page_t
1319vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1320{
1321 vm_page_t mold;
1322
1323 VM_OBJECT_ASSERT_WLOCKED(object);
1324 KASSERT(mnew->object == NULL,
1325 ("vm_page_replace: page already in object"));
1326
1327 /*
1328 * This function mostly follows vm_page_insert() and
1329 * vm_page_remove() without the radix, object count and vnode
1330 * dance. Double check such functions for more comments.
1331 */
1332
1333 mnew->object = object;
1334 mnew->pindex = pindex;
1335 mold = vm_radix_replace(&object->rtree, mnew);
1336 KASSERT(mold->queue == PQ_NONE,
1337 ("vm_page_replace: mold is on a paging queue"));
1338
1339 /* Keep the resident page list in sorted order. */
1340 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1341 TAILQ_REMOVE(&object->memq, mold, listq);
1342
1343 mold->object = NULL;
1344 vm_page_xunbusy(mold);
1345
1346 /*
1347 * The object's resident_page_count does not change because we have
1348 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1349 */
1350 if (pmap_page_is_write_mapped(mnew))
1351 vm_object_set_writeable_dirty(object);
1352 return (mold);
1353}
1354
1355/*
1356 * vm_page_rename:
1357 *
1358 * Move the given memory entry from its
1359 * current object to the specified target object/offset.
1360 *
1361 * Note: swap associated with the page must be invalidated by the move. We
1362 * have to do this for several reasons: (1) we aren't freeing the
1363 * page, (2) we are dirtying the page, (3) the VM system is probably
1364 * moving the page from object A to B, and will then later move
1365 * the backing store from A to B and we can't have a conflict.
1366 *
1367 * Note: we *always* dirty the page. It is necessary both for the
1368 * fact that we moved it, and because we may be invalidating
1369 * swap. If the page is on the cache, we have to deactivate it
1370 * or vm_page_dirty() will panic. Dirty pages are not allowed
1371 * on the cache.
1372 *
1373 * The objects must be locked.
1374 */
1375int
1376vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1377{
1378 vm_page_t mpred;
1379 vm_pindex_t opidx;
1380
1381 VM_OBJECT_ASSERT_WLOCKED(new_object);
1382
1383 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1384 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1385 ("vm_page_rename: pindex already renamed"));
1386
1387 /*
1388 * Create a custom version of vm_page_insert() which does not depend
1389 * by m_prev and can cheat on the implementation aspects of the
1390 * function.
1391 */
1392 opidx = m->pindex;
1393 m->pindex = new_pindex;
1394 if (vm_radix_insert(&new_object->rtree, m)) {
1395 m->pindex = opidx;
1396 return (1);
1397 }
1398
1399 /*
1400 * The operation cannot fail anymore. The removal must happen before
1401 * the listq iterator is tainted.
1402 */
1403 m->pindex = opidx;
1404 vm_page_lock(m);
1405 vm_page_remove(m);
1406
1407 /* Return back to the new pindex to complete vm_page_insert(). */
1408 m->pindex = new_pindex;
1409 m->object = new_object;
1410 vm_page_unlock(m);
1411 vm_page_insert_radixdone(m, new_object, mpred);
1412 vm_page_dirty(m);
1413 return (0);
1414}
1415
1416/*
1417 * Convert all of the given object's cached pages that have a
1418 * pindex within the given range into free pages. If the value
1419 * zero is given for "end", then the range's upper bound is
1420 * infinity. If the given object is backed by a vnode and it
1421 * transitions from having one or more cached pages to none, the
1422 * vnode's hold count is reduced.
1423 */
1424void
1425vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1426{
1427 vm_page_t m;
1428 boolean_t empty;
1429
1430 mtx_lock(&vm_page_queue_free_mtx);
1431 if (__predict_false(vm_radix_is_empty(&object->cache))) {
1432 mtx_unlock(&vm_page_queue_free_mtx);
1433 return;
1434 }
1435 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1436 if (end != 0 && m->pindex >= end)
1437 break;
1438 vm_radix_remove(&object->cache, m->pindex);
1439 vm_page_cache_turn_free(m);
1440 }
1441 empty = vm_radix_is_empty(&object->cache);
1442 mtx_unlock(&vm_page_queue_free_mtx);
1443 if (object->type == OBJT_VNODE && empty)
1444 vdrop(object->handle);
1445}
1446
1447/*
1448 * Returns the cached page that is associated with the given
1449 * object and offset. If, however, none exists, returns NULL.
1450 *
1451 * The free page queue must be locked.
1452 */
1453static inline vm_page_t
1454vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1455{
1456
1457 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1458 return (vm_radix_lookup(&object->cache, pindex));
1459}
1460
1461/*
1462 * Remove the given cached page from its containing object's
1463 * collection of cached pages.
1464 *
1465 * The free page queue must be locked.
1466 */
1467static void
1468vm_page_cache_remove(vm_page_t m)
1469{
1470
1471 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1472 KASSERT((m->flags & PG_CACHED) != 0,
1473 ("vm_page_cache_remove: page %p is not cached", m));
1474 vm_radix_remove(&m->object->cache, m->pindex);
1475 m->object = NULL;
1476 vm_cnt.v_cache_count--;
1477}
1478
1479/*
1480 * Transfer all of the cached pages with offset greater than or
1481 * equal to 'offidxstart' from the original object's cache to the
1482 * new object's cache. However, any cached pages with offset
1483 * greater than or equal to the new object's size are kept in the
1484 * original object. Initially, the new object's cache must be
1485 * empty. Offset 'offidxstart' in the original object must
1486 * correspond to offset zero in the new object.
1487 *
1488 * The new object must be locked.
1489 */
1490void
1491vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1492 vm_object_t new_object)
1493{
1494 vm_page_t m;
1495
1496 /*
1497 * Insertion into an object's collection of cached pages
1498 * requires the object to be locked. In contrast, removal does
1499 * not.
1500 */
1501 VM_OBJECT_ASSERT_WLOCKED(new_object);
1502 KASSERT(vm_radix_is_empty(&new_object->cache),
1503 ("vm_page_cache_transfer: object %p has cached pages",
1504 new_object));
1505 mtx_lock(&vm_page_queue_free_mtx);
1506 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1507 offidxstart)) != NULL) {
1508 /*
1509 * Transfer all of the pages with offset greater than or
1510 * equal to 'offidxstart' from the original object's
1511 * cache to the new object's cache.
1512 */
1513 if ((m->pindex - offidxstart) >= new_object->size)
1514 break;
1515 vm_radix_remove(&orig_object->cache, m->pindex);
1516 /* Update the page's object and offset. */
1517 m->object = new_object;
1518 m->pindex -= offidxstart;
1519 if (vm_radix_insert(&new_object->cache, m))
1520 vm_page_cache_turn_free(m);
1521 }
1522 mtx_unlock(&vm_page_queue_free_mtx);
1523}
1524
1525/*
1526 * Returns TRUE if a cached page is associated with the given object and
1527 * offset, and FALSE otherwise.
1528 *
1529 * The object must be locked.
1530 */
1531boolean_t
1532vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1533{
1534 vm_page_t m;
1535
1536 /*
1537 * Insertion into an object's collection of cached pages requires the
1538 * object to be locked. Therefore, if the object is locked and the
1539 * object's collection is empty, there is no need to acquire the free
1540 * page queues lock in order to prove that the specified page doesn't
1541 * exist.
1542 */
1543 VM_OBJECT_ASSERT_WLOCKED(object);
1544 if (__predict_true(vm_object_cache_is_empty(object)))
1545 return (FALSE);
1546 mtx_lock(&vm_page_queue_free_mtx);
1547 m = vm_page_cache_lookup(object, pindex);
1548 mtx_unlock(&vm_page_queue_free_mtx);
1549 return (m != NULL);
1550}
1551
1552/*
1553 * vm_page_alloc:
1554 *
1555 * Allocate and return a page that is associated with the specified
1556 * object and offset pair. By default, this page is exclusive busied.
1557 *
1558 * The caller must always specify an allocation class.
1559 *
1560 * allocation classes:
1561 * VM_ALLOC_NORMAL normal process request
1562 * VM_ALLOC_SYSTEM system *really* needs a page
1563 * VM_ALLOC_INTERRUPT interrupt time request
1564 *
1565 * optional allocation flags:
1566 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1567 * intends to allocate
1568 * VM_ALLOC_IFCACHED return page only if it is cached
1569 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1570 * is cached
1571 * VM_ALLOC_NOBUSY do not exclusive busy the page
1572 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1573 * VM_ALLOC_NOOBJ page is not associated with an object and
1574 * should not be exclusive busy
1575 * VM_ALLOC_SBUSY shared busy the allocated page
1576 * VM_ALLOC_WIRED wire the allocated page
1577 * VM_ALLOC_ZERO prefer a zeroed page
1578 *
1579 * This routine may not sleep.
1580 */
1581vm_page_t
1582vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1583{
1584 struct vnode *vp = NULL;
1585 vm_object_t m_object;
1586 vm_page_t m, mpred;
1587 int flags, req_class;
1588
1589 mpred = 0; /* XXX: pacify gcc */
1590 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1591 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1592 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1593 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1594 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1595 req));
1596 if (object != NULL)
1597 VM_OBJECT_ASSERT_WLOCKED(object);
1598
1599 req_class = req & VM_ALLOC_CLASS_MASK;
1600
1601 /*
1602 * The page daemon is allowed to dig deeper into the free page list.
1603 */
1604 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1605 req_class = VM_ALLOC_SYSTEM;
1606
1607 if (object != NULL) {
1608 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1609 KASSERT(mpred == NULL || mpred->pindex != pindex,
1610 ("vm_page_alloc: pindex already allocated"));
1611 }
1612
1613 /*
1614 * The page allocation request can came from consumers which already
1615 * hold the free page queue mutex, like vm_page_insert() in
1616 * vm_page_cache().
1617 */
1618 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1619 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1620 (req_class == VM_ALLOC_SYSTEM &&
1621 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1622 (req_class == VM_ALLOC_INTERRUPT &&
1623 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) {
1624 /*
1625 * Allocate from the free queue if the number of free pages
1626 * exceeds the minimum for the request class.
1627 */
1628 if (object != NULL &&
1629 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1630 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1631 mtx_unlock(&vm_page_queue_free_mtx);
1632 return (NULL);
1633 }
1634 if (vm_phys_unfree_page(m))
1635 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1636#if VM_NRESERVLEVEL > 0
1637 else if (!vm_reserv_reactivate_page(m))
1638#else
1639 else
1640#endif
1641 panic("vm_page_alloc: cache page %p is missing"
1642 " from the free queue", m);
1643 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1644 mtx_unlock(&vm_page_queue_free_mtx);
1645 return (NULL);
1646#if VM_NRESERVLEVEL > 0
1647 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1648 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1649 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1650#else
1651 } else {
1652#endif
1653 m = vm_phys_alloc_pages(object != NULL ?
1654 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1655#if VM_NRESERVLEVEL > 0
1656 if (m == NULL && vm_reserv_reclaim_inactive()) {
1657 m = vm_phys_alloc_pages(object != NULL ?
1658 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1659 0);
1660 }
1661#endif
1662 }
1663 } else {
1664 /*
1665 * Not allocatable, give up.
1666 */
1667 mtx_unlock(&vm_page_queue_free_mtx);
1668 atomic_add_int(&vm_pageout_deficit,
1669 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1670 pagedaemon_wakeup();
1671 return (NULL);
1672 }
1673
1674 /*
1675 * At this point we had better have found a good page.
1676 */
1677 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1678 KASSERT(m->queue == PQ_NONE,
1679 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1680 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1681 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1682 KASSERT(!vm_page_sbusied(m),
1683 ("vm_page_alloc: page %p is busy", m));
1684 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1685 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1686 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1687 pmap_page_get_memattr(m)));
1688 if ((m->flags & PG_CACHED) != 0) {
1689 KASSERT((m->flags & PG_ZERO) == 0,
1690 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1691 KASSERT(m->valid != 0,
1692 ("vm_page_alloc: cached page %p is invalid", m));
1693 if (m->object == object && m->pindex == pindex)
1694 vm_cnt.v_reactivated++;
1695 else
1696 m->valid = 0;
1697 m_object = m->object;
1698 vm_page_cache_remove(m);
1699 if (m_object->type == OBJT_VNODE &&
1700 vm_object_cache_is_empty(m_object))
1701 vp = m_object->handle;
1702 } else {
1703 KASSERT(m->valid == 0,
1704 ("vm_page_alloc: free page %p is valid", m));
1705 vm_phys_freecnt_adj(m, -1);
1706 if ((m->flags & PG_ZERO) != 0)
1707 vm_page_zero_count--;
1708 }
1709 mtx_unlock(&vm_page_queue_free_mtx);
1710
1711 /*
1712 * Initialize the page. Only the PG_ZERO flag is inherited.
1713 */
1714 flags = 0;
1715 if ((req & VM_ALLOC_ZERO) != 0)
1716 flags = PG_ZERO;
1717 flags &= m->flags;
1718 if ((req & VM_ALLOC_NODUMP) != 0)
1719 flags |= PG_NODUMP;
1720 m->flags = flags;
1721 m->aflags = 0;
1722 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1723 VPO_UNMANAGED : 0;
1724 m->busy_lock = VPB_UNBUSIED;
1725 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1726 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1727 if ((req & VM_ALLOC_SBUSY) != 0)
1728 m->busy_lock = VPB_SHARERS_WORD(1);
1729 if (req & VM_ALLOC_WIRED) {
1730 /*
1731 * The page lock is not required for wiring a page until that
1732 * page is inserted into the object.
1733 */
1734 atomic_add_int(&vm_cnt.v_wire_count, 1);
1735 m->wire_count = 1;
1736 }
1737 m->act_count = 0;
1738
1739 if (object != NULL) {
1740 if (vm_page_insert_after(m, object, pindex, mpred)) {
1741 /* See the comment below about hold count. */
1742 if (vp != NULL)
1743 vdrop(vp);
1744 pagedaemon_wakeup();
1745 if (req & VM_ALLOC_WIRED) {
1746 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1747 m->wire_count = 0;
1748 }
1749 m->object = NULL;
1750 m->oflags = VPO_UNMANAGED;
1751 vm_page_free(m);
1752 return (NULL);
1753 }
1754
1755 /* Ignore device objects; the pager sets "memattr" for them. */
1756 if (object->memattr != VM_MEMATTR_DEFAULT &&
1757 (object->flags & OBJ_FICTITIOUS) == 0)
1758 pmap_page_set_memattr(m, object->memattr);
1759 } else
1760 m->pindex = pindex;
1761
1762 /*
1763 * The following call to vdrop() must come after the above call
1764 * to vm_page_insert() in case both affect the same object and
1765 * vnode. Otherwise, the affected vnode's hold count could
1766 * temporarily become zero.
1767 */
1768 if (vp != NULL)
1769 vdrop(vp);
1770
1771 /*
1772 * Don't wakeup too often - wakeup the pageout daemon when
1773 * we would be nearly out of memory.
1774 */
1775 if (vm_paging_needed())
1776 pagedaemon_wakeup();
1777
1778 return (m);
1779}
1780
1781static void
1782vm_page_alloc_contig_vdrop(struct spglist *lst)
1783{
1784
1785 while (!SLIST_EMPTY(lst)) {
1786 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1787 SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1788 }
1789}
1790
1791/*
1792 * vm_page_alloc_contig:
1793 *
1794 * Allocate a contiguous set of physical pages of the given size "npages"
1795 * from the free lists. All of the physical pages must be at or above
1796 * the given physical address "low" and below the given physical address
1797 * "high". The given value "alignment" determines the alignment of the
1798 * first physical page in the set. If the given value "boundary" is
1799 * non-zero, then the set of physical pages cannot cross any physical
1800 * address boundary that is a multiple of that value. Both "alignment"
1801 * and "boundary" must be a power of two.
1802 *
1803 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1804 * then the memory attribute setting for the physical pages is configured
1805 * to the object's memory attribute setting. Otherwise, the memory
1806 * attribute setting for the physical pages is configured to "memattr",
1807 * overriding the object's memory attribute setting. However, if the
1808 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1809 * memory attribute setting for the physical pages cannot be configured
1810 * to VM_MEMATTR_DEFAULT.
1811 *
1812 * The caller must always specify an allocation class.
1813 *
1814 * allocation classes:
1815 * VM_ALLOC_NORMAL normal process request
1816 * VM_ALLOC_SYSTEM system *really* needs a page
1817 * VM_ALLOC_INTERRUPT interrupt time request
1818 *
1819 * optional allocation flags:
1820 * VM_ALLOC_NOBUSY do not exclusive busy the page
1821 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1822 * VM_ALLOC_NOOBJ page is not associated with an object and
1823 * should not be exclusive busy
1824 * VM_ALLOC_SBUSY shared busy the allocated page
1825 * VM_ALLOC_WIRED wire the allocated page
1826 * VM_ALLOC_ZERO prefer a zeroed page
1827 *
1828 * This routine may not sleep.
1829 */
1830vm_page_t
1831vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1832 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1833 vm_paddr_t boundary, vm_memattr_t memattr)
1834{
1835 struct vnode *drop;
1836 struct spglist deferred_vdrop_list;
1837 vm_page_t m, m_tmp, m_ret;
1838 u_int flags;
1839 int req_class;
1840
1841 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1842 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1843 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1844 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1845 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1846 req));
1847 if (object != NULL) {
1848 VM_OBJECT_ASSERT_WLOCKED(object);
1849 KASSERT(object->type == OBJT_PHYS,
1850 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1851 object));
1852 }
1853 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1854 req_class = req & VM_ALLOC_CLASS_MASK;
1855
1856 /*
1857 * The page daemon is allowed to dig deeper into the free page list.
1858 */
1859 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1860 req_class = VM_ALLOC_SYSTEM;
1861
1862 SLIST_INIT(&deferred_vdrop_list);
1863 mtx_lock(&vm_page_queue_free_mtx);
1864 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1865 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1866 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1867 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1868 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) {
1869#if VM_NRESERVLEVEL > 0
1870retry:
1871 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1872 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1873 low, high, alignment, boundary)) == NULL)
1874#endif
1875 m_ret = vm_phys_alloc_contig(npages, low, high,
1876 alignment, boundary);
1877 } else {
1878 mtx_unlock(&vm_page_queue_free_mtx);
1879 atomic_add_int(&vm_pageout_deficit, npages);
1880 pagedaemon_wakeup();
1881 return (NULL);
1882 }
1883 if (m_ret != NULL)
1884 for (m = m_ret; m < &m_ret[npages]; m++) {
1885 drop = vm_page_alloc_init(m);
1886 if (drop != NULL) {
1887 /*
1888 * Enqueue the vnode for deferred vdrop().
1889 */
1890 m->plinks.s.pv = drop;
1891 SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1892 plinks.s.ss);
1893 }
1894 }
1895 else {
1896#if VM_NRESERVLEVEL > 0
1897 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1898 boundary))
1899 goto retry;
1900#endif
1901 }
1902 mtx_unlock(&vm_page_queue_free_mtx);
1903 if (m_ret == NULL)
1904 return (NULL);
1905
1906 /*
1907 * Initialize the pages. Only the PG_ZERO flag is inherited.
1908 */
1909 flags = 0;
1910 if ((req & VM_ALLOC_ZERO) != 0)
1911 flags = PG_ZERO;
1912 if ((req & VM_ALLOC_NODUMP) != 0)
1913 flags |= PG_NODUMP;
1914 if ((req & VM_ALLOC_WIRED) != 0)
1915 atomic_add_int(&vm_cnt.v_wire_count, npages);
1916 if (object != NULL) {
1917 if (object->memattr != VM_MEMATTR_DEFAULT &&
1918 memattr == VM_MEMATTR_DEFAULT)
1919 memattr = object->memattr;
1920 }
1921 for (m = m_ret; m < &m_ret[npages]; m++) {
1922 m->aflags = 0;
1923 m->flags = (m->flags | PG_NODUMP) & flags;
1924 m->busy_lock = VPB_UNBUSIED;
1925 if (object != NULL) {
1926 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1927 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1928 if ((req & VM_ALLOC_SBUSY) != 0)
1929 m->busy_lock = VPB_SHARERS_WORD(1);
1930 }
1931 if ((req & VM_ALLOC_WIRED) != 0)
1932 m->wire_count = 1;
1933 /* Unmanaged pages don't use "act_count". */
1934 m->oflags = VPO_UNMANAGED;
1935 if (object != NULL) {
1936 if (vm_page_insert(m, object, pindex)) {
1937 vm_page_alloc_contig_vdrop(
1938 &deferred_vdrop_list);
1939 if (vm_paging_needed())
1940 pagedaemon_wakeup();
1941 if ((req & VM_ALLOC_WIRED) != 0)
1942 atomic_subtract_int(&vm_cnt.v_wire_count,
1943 npages);
1944 for (m_tmp = m, m = m_ret;
1945 m < &m_ret[npages]; m++) {
1946 if ((req & VM_ALLOC_WIRED) != 0)
1947 m->wire_count = 0;
1948 if (m >= m_tmp)
1949 m->object = NULL;
1950 vm_page_free(m);
1951 }
1952 return (NULL);
1953 }
1954 } else
1955 m->pindex = pindex;
1956 if (memattr != VM_MEMATTR_DEFAULT)
1957 pmap_page_set_memattr(m, memattr);
1958 pindex++;
1959 }
1960 vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1961 if (vm_paging_needed())
1962 pagedaemon_wakeup();
1963 return (m_ret);
1964}
1965
1966/*
1967 * Initialize a page that has been freshly dequeued from a freelist.
1968 * The caller has to drop the vnode returned, if it is not NULL.
1969 *
1970 * This function may only be used to initialize unmanaged pages.
1971 *
1972 * To be called with vm_page_queue_free_mtx held.
1973 */
1974static struct vnode *
1975vm_page_alloc_init(vm_page_t m)
1976{
1977 struct vnode *drop;
1978 vm_object_t m_object;
1979
1980 KASSERT(m->queue == PQ_NONE,
1981 ("vm_page_alloc_init: page %p has unexpected queue %d",
1982 m, m->queue));
1983 KASSERT(m->wire_count == 0,
1984 ("vm_page_alloc_init: page %p is wired", m));
1985 KASSERT(m->hold_count == 0,
1986 ("vm_page_alloc_init: page %p is held", m));
1987 KASSERT(!vm_page_sbusied(m),
1988 ("vm_page_alloc_init: page %p is busy", m));
1989 KASSERT(m->dirty == 0,
1990 ("vm_page_alloc_init: page %p is dirty", m));
1991 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1992 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1993 m, pmap_page_get_memattr(m)));
1994 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1995 drop = NULL;
1996 if ((m->flags & PG_CACHED) != 0) {
1997 KASSERT((m->flags & PG_ZERO) == 0,
1998 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1999 m->valid = 0;
2000 m_object = m->object;
2001 vm_page_cache_remove(m);
2002 if (m_object->type == OBJT_VNODE &&
2003 vm_object_cache_is_empty(m_object))
2004 drop = m_object->handle;
2005 } else {
2006 KASSERT(m->valid == 0,
2007 ("vm_page_alloc_init: free page %p is valid", m));
2008 vm_phys_freecnt_adj(m, -1);
2009 if ((m->flags & PG_ZERO) != 0)
2010 vm_page_zero_count--;
2011 }
2012 return (drop);
2013}
2014
2015/*
2016 * vm_page_alloc_freelist:
2017 *
2018 * Allocate a physical page from the specified free page list.
2019 *
2020 * The caller must always specify an allocation class.
2021 *
2022 * allocation classes:
2023 * VM_ALLOC_NORMAL normal process request
2024 * VM_ALLOC_SYSTEM system *really* needs a page
2025 * VM_ALLOC_INTERRUPT interrupt time request
2026 *
2027 * optional allocation flags:
2028 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2029 * intends to allocate
2030 * VM_ALLOC_WIRED wire the allocated page
2031 * VM_ALLOC_ZERO prefer a zeroed page
2032 *
2033 * This routine may not sleep.
2034 */
2035vm_page_t
2036vm_page_alloc_freelist(int flind, int req)
2037{
2038 struct vnode *drop;
2039 vm_page_t m;
2040 u_int flags;
2041 int req_class;
2042
2043 req_class = req & VM_ALLOC_CLASS_MASK;
2044
2045 /*
2046 * The page daemon is allowed to dig deeper into the free page list.
2047 */
2048 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2049 req_class = VM_ALLOC_SYSTEM;
2050
2051 /*
2052 * Do not allocate reserved pages unless the req has asked for it.
2053 */
2054 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
2055 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
2056 (req_class == VM_ALLOC_SYSTEM &&
2057 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
2058 (req_class == VM_ALLOC_INTERRUPT &&
2059 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0))
2060 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
2061 else {
2062 mtx_unlock(&vm_page_queue_free_mtx);
2063 atomic_add_int(&vm_pageout_deficit,
2064 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2065 pagedaemon_wakeup();
2066 return (NULL);
2067 }
2068 if (m == NULL) {
2069 mtx_unlock(&vm_page_queue_free_mtx);
2070 return (NULL);
2071 }
2072 drop = vm_page_alloc_init(m);
2073 mtx_unlock(&vm_page_queue_free_mtx);
2074
2075 /*
2076 * Initialize the page. Only the PG_ZERO flag is inherited.
2077 */
2078 m->aflags = 0;
2079 flags = 0;
2080 if ((req & VM_ALLOC_ZERO) != 0)
2081 flags = PG_ZERO;
2082 m->flags &= flags;
2083 if ((req & VM_ALLOC_WIRED) != 0) {
2084 /*
2085 * The page lock is not required for wiring a page that does
2086 * not belong to an object.
2087 */
2088 atomic_add_int(&vm_cnt.v_wire_count, 1);
2089 m->wire_count = 1;
2090 }
2091 /* Unmanaged pages don't use "act_count". */
2092 m->oflags = VPO_UNMANAGED;
2093 if (drop != NULL)
2094 vdrop(drop);
2095 if (vm_paging_needed())
2096 pagedaemon_wakeup();
2097 return (m);
2098}
2099
2100#define VPSC_ANY 0 /* No restrictions. */
2101#define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2102#define VPSC_NOSUPER 2 /* Skip superpages. */
2103
2104/*
2105 * vm_page_scan_contig:
2106 *
2107 * Scan vm_page_array[] between the specified entries "m_start" and
2108 * "m_end" for a run of contiguous physical pages that satisfy the
2109 * specified conditions, and return the lowest page in the run. The
2110 * specified "alignment" determines the alignment of the lowest physical
2111 * page in the run. If the specified "boundary" is non-zero, then the
2112 * run of physical pages cannot span a physical address that is a
2113 * multiple of "boundary".
2114 *
2115 * "m_end" is never dereferenced, so it need not point to a vm_page
2116 * structure within vm_page_array[].
2117 *
2118 * "npages" must be greater than zero. "m_start" and "m_end" must not
2119 * span a hole (or discontiguity) in the physical address space. Both
2120 * "alignment" and "boundary" must be a power of two.
2121 */
2122vm_page_t
2123vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2124 u_long alignment, vm_paddr_t boundary, int options)
2125{
2126 struct mtx *m_mtx, *new_mtx;
2127 vm_object_t object;
2128 vm_paddr_t pa;
2129 vm_page_t m, m_run;
2130#if VM_NRESERVLEVEL > 0
2131 int level;
2132#endif
2133 int m_inc, order, run_ext, run_len;
2134
2135 KASSERT(npages > 0, ("npages is 0"));
2136 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2137 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2138 m_run = NULL;
2139 run_len = 0;
2140 m_mtx = NULL;
2141 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2142 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2143 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2144
2145 /*
2146 * If the current page would be the start of a run, check its
2147 * physical address against the end, alignment, and boundary
2148 * conditions. If it doesn't satisfy these conditions, either
2149 * terminate the scan or advance to the next page that
2150 * satisfies the failed condition.
2151 */
2152 if (run_len == 0) {
2153 KASSERT(m_run == NULL, ("m_run != NULL"));
2154 if (m + npages > m_end)
2155 break;
2156 pa = VM_PAGE_TO_PHYS(m);
2157 if ((pa & (alignment - 1)) != 0) {
2158 m_inc = atop(roundup2(pa, alignment) - pa);
2159 continue;
2160 }
2161 if (((pa ^ (pa + ptoa(npages) - 1)) & ~(boundary -
2162 1)) != 0) {
2163 m_inc = atop(roundup2(pa, boundary) - pa);
2164 continue;
2165 }
2166 } else
2167 KASSERT(m_run != NULL, ("m_run == NULL"));
2168
2169 /*
2170 * Avoid releasing and reacquiring the same page lock.
2171 */
2172 new_mtx = vm_page_lockptr(m);
2173 if (m_mtx != new_mtx) {
2174 if (m_mtx != NULL)
2175 mtx_unlock(m_mtx);
2176 m_mtx = new_mtx;
2177 mtx_lock(m_mtx);
2178 }
2179 m_inc = 1;
2180retry:
2181 if (m->wire_count != 0 || m->hold_count != 0)
2182 run_ext = 0;
2183#if VM_NRESERVLEVEL > 0
2184 else if ((level = vm_reserv_level(m)) >= 0 &&
2185 (options & VPSC_NORESERV) != 0) {
2186 run_ext = 0;
2187 /* Advance to the end of the reservation. */
2188 pa = VM_PAGE_TO_PHYS(m);
2189 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2190 pa);
2191 }
2192#endif
2193 else if ((object = m->object) != NULL) {
2194 /*
2195 * The page is considered eligible for relocation if
2196 * and only if it could be laundered or reclaimed by
2197 * the page daemon.
2198 */
2199 if (!VM_OBJECT_TRYRLOCK(object)) {
2200 mtx_unlock(m_mtx);
2201 VM_OBJECT_RLOCK(object);
2202 mtx_lock(m_mtx);
2203 if (m->object != object) {
2204 /*
2205 * The page may have been freed.
2206 */
2207 VM_OBJECT_RUNLOCK(object);
2208 goto retry;
2209 } else if (m->wire_count != 0 ||
2210 m->hold_count != 0) {
2211 run_ext = 0;
2212 goto unlock;
2213 }
2214 }
2215 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2216 ("page %p is PG_UNHOLDFREE", m));
2217 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2218 if (object->type != OBJT_DEFAULT &&
2219 object->type != OBJT_SWAP &&
2220 object->type != OBJT_VNODE)
2221 run_ext = 0;
2222 else if ((m->flags & PG_CACHED) != 0 ||
2223 m != vm_page_lookup(object, m->pindex)) {
2224 /*
2225 * The page is cached or recently converted
2226 * from cached to free.
2227 */
2228#if VM_NRESERVLEVEL > 0
2229 if (level >= 0) {
2230 /*
2231 * The page is reserved. Extend the
2232 * current run by one page.
2233 */
2234 run_ext = 1;
2235 } else
2236#endif
2237 if ((order = m->order) < VM_NFREEORDER) {
2238 /*
2239 * The page is enqueued in the
2240 * physical memory allocator's cache/
2241 * free page queues. Moreover, it is
2242 * the first page in a power-of-two-
2243 * sized run of contiguous cache/free
2244 * pages. Add these pages to the end
2245 * of the current run, and jump
2246 * ahead.
2247 */
2248 run_ext = 1 << order;
2249 m_inc = 1 << order;
2250 } else
2251 run_ext = 0;
2252#if VM_NRESERVLEVEL > 0
2253 } else if ((options & VPSC_NOSUPER) != 0 &&
2254 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2255 run_ext = 0;
2256 /* Advance to the end of the superpage. */
2257 pa = VM_PAGE_TO_PHYS(m);
2258 m_inc = atop(roundup2(pa + 1,
2259 vm_reserv_size(level)) - pa);
2260#endif
2261 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2262 m->queue != PQ_NONE && !vm_page_busied(m)) {
2263 /*
2264 * The page is allocated but eligible for
2265 * relocation. Extend the current run by one
2266 * page.
2267 */
2268 KASSERT(pmap_page_get_memattr(m) ==
2269 VM_MEMATTR_DEFAULT,
2270 ("page %p has an unexpected memattr", m));
2271 KASSERT((m->oflags & (VPO_SWAPINPROG |
2272 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2273 ("page %p has unexpected oflags", m));
2274 /* Don't care: VPO_NOSYNC. */
2275 run_ext = 1;
2276 } else
2277 run_ext = 0;
2278unlock:
2279 VM_OBJECT_RUNLOCK(object);
2280#if VM_NRESERVLEVEL > 0
2281 } else if (level >= 0) {
2282 /*
2283 * The page is reserved but not yet allocated. In
2284 * other words, it is still cached or free. Extend
2285 * the current run by one page.
2286 */
2287 run_ext = 1;
2288#endif
2289 } else if ((order = m->order) < VM_NFREEORDER) {
2290 /*
2291 * The page is enqueued in the physical memory
2292 * allocator's cache/free page queues. Moreover, it
2293 * is the first page in a power-of-two-sized run of
2294 * contiguous cache/free pages. Add these pages to
2295 * the end of the current run, and jump ahead.
2296 */
2297 run_ext = 1 << order;
2298 m_inc = 1 << order;
2299 } else {
2300 /*
2301 * Skip the page for one of the following reasons: (1)
2302 * It is enqueued in the physical memory allocator's
2303 * cache/free page queues. However, it is not the
2304 * first page in a run of contiguous cache/free pages.
2305 * (This case rarely occurs because the scan is
2306 * performed in ascending order.) (2) It is not
2307 * reserved, and it is transitioning from free to
2308 * allocated. (Conversely, the transition from
2309 * allocated to free for managed pages is blocked by
2310 * the page lock.) (3) It is allocated but not
2311 * contained by an object and not wired, e.g.,
2312 * allocated by Xen's balloon driver.
2313 */
2314 run_ext = 0;
2315 }
2316
2317 /*
2318 * Extend or reset the current run of pages.
2319 */
2320 if (run_ext > 0) {
2321 if (run_len == 0)
2322 m_run = m;
2323 run_len += run_ext;
2324 } else {
2325 if (run_len > 0) {
2326 m_run = NULL;
2327 run_len = 0;
2328 }
2329 }
2330 }
2331 if (m_mtx != NULL)
2332 mtx_unlock(m_mtx);
2333 if (run_len >= npages)
2334 return (m_run);
2335 return (NULL);
2336}
2337
2338/*
2339 * vm_page_reclaim_run:
2340 *
2341 * Try to relocate each of the allocated virtual pages within the
2342 * specified run of physical pages to a new physical address. Free the
2343 * physical pages underlying the relocated virtual pages. A virtual page
2344 * is relocatable if and only if it could be laundered or reclaimed by
2345 * the page daemon. Whenever possible, a virtual page is relocated to a
2346 * physical address above "high".
2347 *
2348 * Returns 0 if every physical page within the run was already free or
2349 * just freed by a successful relocation. Otherwise, returns a non-zero
2350 * value indicating why the last attempt to relocate a virtual page was
2351 * unsuccessful.
2352 *
2353 * "req_class" must be an allocation class.
2354 */
2355static int
2356vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2357 vm_paddr_t high)
2358{
2359 struct mtx *m_mtx, *new_mtx;
2360 struct spglist free;
2361 vm_object_t object;
2362 vm_paddr_t pa;
2363 vm_page_t m, m_end, m_new;
2364 int error, order, req;
2365
2366 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2367 ("req_class is not an allocation class"));
2368 SLIST_INIT(&free);
2369 error = 0;
2370 m = m_run;
2371 m_end = m_run + npages;
2372 m_mtx = NULL;
2373 for (; error == 0 && m < m_end; m++) {
2374 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2375 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2376
2377 /*
2378 * Avoid releasing and reacquiring the same page lock.
2379 */
2380 new_mtx = vm_page_lockptr(m);
2381 if (m_mtx != new_mtx) {
2382 if (m_mtx != NULL)
2383 mtx_unlock(m_mtx);
2384 m_mtx = new_mtx;
2385 mtx_lock(m_mtx);
2386 }
2387retry:
2388 if (m->wire_count != 0 || m->hold_count != 0)
2389 error = EBUSY;
2390 else if ((object = m->object) != NULL) {
2391 /*
2392 * The page is relocated if and only if it could be
2393 * laundered or reclaimed by the page daemon.
2394 */
2395 if (!VM_OBJECT_TRYWLOCK(object)) {
2396 mtx_unlock(m_mtx);
2397 VM_OBJECT_WLOCK(object);
2398 mtx_lock(m_mtx);
2399 if (m->object != object) {
2400 /*
2401 * The page may have been freed.
2402 */
2403 VM_OBJECT_WUNLOCK(object);
2404 goto retry;
2405 } else if (m->wire_count != 0 ||
2406 m->hold_count != 0) {
2407 error = EBUSY;
2408 goto unlock;
2409 }
2410 }
2411 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2412 ("page %p is PG_UNHOLDFREE", m));
2413 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2414 if (object->type != OBJT_DEFAULT &&
2415 object->type != OBJT_SWAP &&
2416 object->type != OBJT_VNODE)
2417 error = EINVAL;
2418 else if ((m->flags & PG_CACHED) != 0 ||
2419 m != vm_page_lookup(object, m->pindex)) {
2420 /*
2421 * The page is cached or recently converted
2422 * from cached to free.
2423 */
2424 VM_OBJECT_WUNLOCK(object);
2425 goto cached;
2426 } else if (object->memattr != VM_MEMATTR_DEFAULT)
2427 error = EINVAL;
2428 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2429 KASSERT(pmap_page_get_memattr(m) ==
2430 VM_MEMATTR_DEFAULT,
2431 ("page %p has an unexpected memattr", m));
2432 KASSERT((m->oflags & (VPO_SWAPINPROG |
2433 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2434 ("page %p has unexpected oflags", m));
2435 /* Don't care: VPO_NOSYNC. */
2436 if (m->valid != 0) {
2437 /*
2438 * First, try to allocate a new page
2439 * that is above "high". Failing
2440 * that, try to allocate a new page
2441 * that is below "m_run". Allocate
2442 * the new page between the end of
2443 * "m_run" and "high" only as a last
2444 * resort.
2445 */
2446 req = req_class | VM_ALLOC_NOOBJ;
2447 if ((m->flags & PG_NODUMP) != 0)
2448 req |= VM_ALLOC_NODUMP;
2449 if (trunc_page(high) !=
2450 ~(vm_paddr_t)PAGE_MASK) {
2451 m_new = vm_page_alloc_contig(
2452 NULL, 0, req, 1,
2453 round_page(high),
2454 ~(vm_paddr_t)0,
2455 PAGE_SIZE, 0,
2456 VM_MEMATTR_DEFAULT);
2457 } else
2458 m_new = NULL;
2459 if (m_new == NULL) {
2460 pa = VM_PAGE_TO_PHYS(m_run);
2461 m_new = vm_page_alloc_contig(
2462 NULL, 0, req, 1,
2463 0, pa - 1, PAGE_SIZE, 0,
2464 VM_MEMATTR_DEFAULT);
2465 }
2466 if (m_new == NULL) {
2467 pa += ptoa(npages);
2468 m_new = vm_page_alloc_contig(
2469 NULL, 0, req, 1,
2470 pa, high, PAGE_SIZE, 0,
2471 VM_MEMATTR_DEFAULT);
2472 }
2473 if (m_new == NULL) {
2474 error = ENOMEM;
2475 goto unlock;
2476 }
2477 KASSERT(m_new->wire_count == 0,
2478 ("page %p is wired", m));
2479
2480 /*
2481 * Replace "m" with the new page. For
2482 * vm_page_replace(), "m" must be busy
2483 * and dequeued. Finally, change "m"
2484 * as if vm_page_free() was called.
2485 */
2486 if (object->ref_count != 0)
2487 pmap_remove_all(m);
2488 m_new->aflags = m->aflags;
2489 KASSERT(m_new->oflags == VPO_UNMANAGED,
2490 ("page %p is managed", m));
2491 m_new->oflags = m->oflags & VPO_NOSYNC;
2492 pmap_copy_page(m, m_new);
2493 m_new->valid = m->valid;
2494 m_new->dirty = m->dirty;
2495 m->flags &= ~PG_ZERO;
2496 vm_page_xbusy(m);
2497 vm_page_remque(m);
2498 vm_page_replace_checked(m_new, object,
2499 m->pindex, m);
2500 m->valid = 0;
2501 vm_page_undirty(m);
2502
2503 /*
2504 * The new page must be deactivated
2505 * before the object is unlocked.
2506 */
2507 new_mtx = vm_page_lockptr(m_new);
2508 if (m_mtx != new_mtx) {
2509 mtx_unlock(m_mtx);
2510 m_mtx = new_mtx;
2511 mtx_lock(m_mtx);
2512 }
2513 vm_page_deactivate(m_new);
2514 } else {
2515 m->flags &= ~PG_ZERO;
2516 vm_page_remque(m);
2517 vm_page_remove(m);
2518 KASSERT(m->dirty == 0,
2519 ("page %p is dirty", m));
2520 }
2521 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2522 } else
2523 error = EBUSY;
2524unlock:
2525 VM_OBJECT_WUNLOCK(object);
2526 } else {
2527cached:
2528 mtx_lock(&vm_page_queue_free_mtx);
2529 order = m->order;
2530 if (order < VM_NFREEORDER) {
2531 /*
2532 * The page is enqueued in the physical memory
2533 * allocator's cache/free page queues.
2534 * Moreover, it is the first page in a power-
2535 * of-two-sized run of contiguous cache/free
2536 * pages. Jump ahead to the last page within
2537 * that run, and continue from there.
2538 */
2539 m += (1 << order) - 1;
2540 }
2541#if VM_NRESERVLEVEL > 0
2542 else if (vm_reserv_is_page_free(m))
2543 order = 0;
2544#endif
2545 mtx_unlock(&vm_page_queue_free_mtx);
2546 if (order == VM_NFREEORDER)
2547 error = EINVAL;
2548 }
2549 }
2550 if (m_mtx != NULL)
2551 mtx_unlock(m_mtx);
2552 if ((m = SLIST_FIRST(&free)) != NULL) {
2553 mtx_lock(&vm_page_queue_free_mtx);
2554 do {
2555 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2556 vm_phys_freecnt_adj(m, 1);
2557#if VM_NRESERVLEVEL > 0
2558 if (!vm_reserv_free_page(m))
2559#else
2560 if (true)
2561#endif
2562 vm_phys_free_pages(m, 0);
2563 } while ((m = SLIST_FIRST(&free)) != NULL);
2564 vm_page_zero_idle_wakeup();
2565 vm_page_free_wakeup();
2566 mtx_unlock(&vm_page_queue_free_mtx);
2567 }
2568 return (error);
2569}
2570
2571#define NRUNS 16
2572
2573CTASSERT(powerof2(NRUNS));
2574
2575#define RUN_INDEX(count) ((count) & (NRUNS - 1))
2576
2577#define MIN_RECLAIM 8
2578
2579/*
2580 * vm_page_reclaim_contig:
2581 *
2582 * Reclaim allocated, contiguous physical memory satisfying the specified
2583 * conditions by relocating the virtual pages using that physical memory.
2584 * Returns true if reclamation is successful and false otherwise. Since
2585 * relocation requires the allocation of physical pages, reclamation may
2586 * fail due to a shortage of cache/free pages. When reclamation fails,
2587 * callers are expected to perform VM_WAIT before retrying a failed
2588 * allocation operation, e.g., vm_page_alloc_contig().
2589 *
2590 * The caller must always specify an allocation class through "req".
2591 *
2592 * allocation classes:
2593 * VM_ALLOC_NORMAL normal process request
2594 * VM_ALLOC_SYSTEM system *really* needs a page
2595 * VM_ALLOC_INTERRUPT interrupt time request
2596 *
2597 * The optional allocation flags are ignored.
2598 *
2599 * "npages" must be greater than zero. Both "alignment" and "boundary"
2600 * must be a power of two.
2601 */
2602bool
2603vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2604 u_long alignment, vm_paddr_t boundary)
2605{
2606 vm_paddr_t curr_low;
2607 vm_page_t m_run, m_runs[NRUNS];
2608 u_long count, reclaimed;
2609 int error, i, options, req_class;
2610
2611 KASSERT(npages > 0, ("npages is 0"));
2612 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2613 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2614 req_class = req & VM_ALLOC_CLASS_MASK;
2615
2616 /*
2617 * The page daemon is allowed to dig deeper into the free page list.
2618 */
2619 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2620 req_class = VM_ALLOC_SYSTEM;
2621
2622 /*
2623 * Return if the number of cached and free pages cannot satisfy the
2624 * requested allocation.
2625 */
2626 count = vm_cnt.v_free_count + vm_cnt.v_cache_count;
2627 if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2628 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2629 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2630 return (false);
2631
2632 /*
2633 * Scan up to three times, relaxing the restrictions ("options") on
2634 * the reclamation of reservations and superpages each time.
2635 */
2636 for (options = VPSC_NORESERV;;) {
2637 /*
2638 * Find the highest runs that satisfy the given constraints
2639 * and restrictions, and record them in "m_runs".
2640 */
2641 curr_low = low;
2642 count = 0;
2643 for (;;) {
2644 m_run = vm_phys_scan_contig(npages, curr_low, high,
2645 alignment, boundary, options);
2646 if (m_run == NULL)
2647 break;
2648 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2649 m_runs[RUN_INDEX(count)] = m_run;
2650 count++;
2651 }
2652
2653 /*
2654 * Reclaim the highest runs in LIFO (descending) order until
2655 * the number of reclaimed pages, "reclaimed", is at least
2656 * MIN_RECLAIM. Reset "reclaimed" each time because each
2657 * reclamation is idempotent, and runs will (likely) recur
2658 * from one scan to the next as restrictions are relaxed.
2659 */
2660 reclaimed = 0;
2661 for (i = 0; count > 0 && i < NRUNS; i++) {
2662 count--;
2663 m_run = m_runs[RUN_INDEX(count)];
2664 error = vm_page_reclaim_run(req_class, npages, m_run,
2665 high);
2666 if (error == 0) {
2667 reclaimed += npages;
2668 if (reclaimed >= MIN_RECLAIM)
2669 return (true);
2670 }
2671 }
2672
2673 /*
2674 * Either relax the restrictions on the next scan or return if
2675 * the last scan had no restrictions.
2676 */
2677 if (options == VPSC_NORESERV)
2678 options = VPSC_NOSUPER;
2679 else if (options == VPSC_NOSUPER)
2680 options = VPSC_ANY;
2681 else if (options == VPSC_ANY)
2682 return (reclaimed != 0);
2683 }
2684}
2685
2686/*
2687 * vm_wait: (also see VM_WAIT macro)
2688 *
2689 * Sleep until free pages are available for allocation.
2690 * - Called in various places before memory allocations.
2691 */
2692void
2693vm_wait(void)
2694{
2695
2696 mtx_lock(&vm_page_queue_free_mtx);
2697 if (curproc == pageproc) {
2698 vm_pageout_pages_needed = 1;
2699 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2700 PDROP | PSWP, "VMWait", 0);
2701 } else {
2702 if (!vm_pages_needed) {
2703 vm_pages_needed = 1;
2704 wakeup(&vm_pages_needed);
2705 }
2706 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2707 "vmwait", 0);
2708 }
2709}
2710
2711/*
2712 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2713 *
2714 * Sleep until free pages are available for allocation.
2715 * - Called only in vm_fault so that processes page faulting
2716 * can be easily tracked.
2717 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2718 * processes will be able to grab memory first. Do not change
2719 * this balance without careful testing first.
2720 */
2721void
2722vm_waitpfault(void)
2723{
2724
2725 mtx_lock(&vm_page_queue_free_mtx);
2726 if (!vm_pages_needed) {
2727 vm_pages_needed = 1;
2728 wakeup(&vm_pages_needed);
2729 }
2730 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2731 "pfault", 0);
2732}
2733
2734struct vm_pagequeue *
2735vm_page_pagequeue(vm_page_t m)
2736{
2737
2738 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2739}
2740
2741/*
2742 * vm_page_dequeue:
2743 *
2744 * Remove the given page from its current page queue.
2745 *
2746 * The page must be locked.
2747 */
2748void
2749vm_page_dequeue(vm_page_t m)
2750{
2751 struct vm_pagequeue *pq;
2752
2753 vm_page_assert_locked(m);
2754 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2755 m));
2756 pq = vm_page_pagequeue(m);
2757 vm_pagequeue_lock(pq);
2758 m->queue = PQ_NONE;
2759 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2760 vm_pagequeue_cnt_dec(pq);
2761 vm_pagequeue_unlock(pq);
2762}
2763
2764/*
2765 * vm_page_dequeue_locked:
2766 *
2767 * Remove the given page from its current page queue.
2768 *
2769 * The page and page queue must be locked.
2770 */
2771void
2772vm_page_dequeue_locked(vm_page_t m)
2773{
2774 struct vm_pagequeue *pq;
2775
2776 vm_page_lock_assert(m, MA_OWNED);
2777 pq = vm_page_pagequeue(m);
2778 vm_pagequeue_assert_locked(pq);
2779 m->queue = PQ_NONE;
2780 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2781 vm_pagequeue_cnt_dec(pq);
2782}
2783
2784/*
2785 * vm_page_enqueue:
2786 *
2787 * Add the given page to the specified page queue.
2788 *
2789 * The page must be locked.
2790 */
2791static void
2792vm_page_enqueue(uint8_t queue, vm_page_t m)
2793{
2794 struct vm_pagequeue *pq;
2795
2796 vm_page_lock_assert(m, MA_OWNED);
2797 KASSERT(queue < PQ_COUNT,
2798 ("vm_page_enqueue: invalid queue %u request for page %p",
2799 queue, m));
2800 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2801 vm_pagequeue_lock(pq);
2802 m->queue = queue;
2803 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2804 vm_pagequeue_cnt_inc(pq);
2805 vm_pagequeue_unlock(pq);
2806}
2807
2808/*
2809 * vm_page_requeue:
2810 *
2811 * Move the given page to the tail of its current page queue.
2812 *
2813 * The page must be locked.
2814 */
2815void
2816vm_page_requeue(vm_page_t m)
2817{
2818 struct vm_pagequeue *pq;
2819
2820 vm_page_lock_assert(m, MA_OWNED);
2821 KASSERT(m->queue != PQ_NONE,
2822 ("vm_page_requeue: page %p is not queued", m));
2823 pq = vm_page_pagequeue(m);
2824 vm_pagequeue_lock(pq);
2825 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2826 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2827 vm_pagequeue_unlock(pq);
2828}
2829
2830/*
2831 * vm_page_requeue_locked:
2832 *
2833 * Move the given page to the tail of its current page queue.
2834 *
2835 * The page queue must be locked.
2836 */
2837void
2838vm_page_requeue_locked(vm_page_t m)
2839{
2840 struct vm_pagequeue *pq;
2841
2842 KASSERT(m->queue != PQ_NONE,
2843 ("vm_page_requeue_locked: page %p is not queued", m));
2844 pq = vm_page_pagequeue(m);
2845 vm_pagequeue_assert_locked(pq);
2846 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2847 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2848}
2849
2850/*
2851 * vm_page_activate:
2852 *
2853 * Put the specified page on the active list (if appropriate).
2854 * Ensure that act_count is at least ACT_INIT but do not otherwise
2855 * mess with it.
2856 *
2857 * The page must be locked.
2858 */
2859void
2860vm_page_activate(vm_page_t m)
2861{
2862 int queue;
2863
2864 vm_page_lock_assert(m, MA_OWNED);
2865 if ((queue = m->queue) != PQ_ACTIVE) {
2866 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2867 if (m->act_count < ACT_INIT)
2868 m->act_count = ACT_INIT;
2869 if (queue != PQ_NONE)
2870 vm_page_dequeue(m);
2871 vm_page_enqueue(PQ_ACTIVE, m);
2872 } else
2873 KASSERT(queue == PQ_NONE,
2874 ("vm_page_activate: wired page %p is queued", m));
2875 } else {
2876 if (m->act_count < ACT_INIT)
2877 m->act_count = ACT_INIT;
2878 }
2879}
2880
2881/*
2882 * vm_page_free_wakeup:
2883 *
2884 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2885 * routine is called when a page has been added to the cache or free
2886 * queues.
2887 *
2888 * The page queues must be locked.
2889 */
2890static inline void
2891vm_page_free_wakeup(void)
2892{
2893
2894 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2895 /*
2896 * if pageout daemon needs pages, then tell it that there are
2897 * some free.
2898 */
2899 if (vm_pageout_pages_needed &&
2900 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2901 wakeup(&vm_pageout_pages_needed);
2902 vm_pageout_pages_needed = 0;
2903 }
2904 /*
2905 * wakeup processes that are waiting on memory if we hit a
2906 * high water mark. And wakeup scheduler process if we have
2907 * lots of memory. this process will swapin processes.
2908 */
2909 if (vm_pages_needed && !vm_page_count_min()) {
2910 vm_pages_needed = 0;
2911 wakeup(&vm_cnt.v_free_count);
2912 }
2913}
2914
2915/*
2916 * Turn a cached page into a free page, by changing its attributes.
2917 * Keep the statistics up-to-date.
2918 *
2919 * The free page queue must be locked.
2920 */
2921static void
2922vm_page_cache_turn_free(vm_page_t m)
2923{
2924
2925 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2926
2927 m->object = NULL;
2928 m->valid = 0;
2929 KASSERT((m->flags & PG_CACHED) != 0,
2930 ("vm_page_cache_turn_free: page %p is not cached", m));
2931 m->flags &= ~PG_CACHED;
2932 vm_cnt.v_cache_count--;
2933 vm_phys_freecnt_adj(m, 1);
2934}
2935
2936/*
2937 * vm_page_free_toq:
2938 *
2939 * Returns the given page to the free list,
2940 * disassociating it with any VM object.
2941 *
2942 * The object must be locked. The page must be locked if it is managed.
2943 */
2944void
2945vm_page_free_toq(vm_page_t m)
2946{
2947
2948 if ((m->oflags & VPO_UNMANAGED) == 0) {
2949 vm_page_lock_assert(m, MA_OWNED);
2950 KASSERT(!pmap_page_is_mapped(m),
2951 ("vm_page_free_toq: freeing mapped page %p", m));
2952 } else
2953 KASSERT(m->queue == PQ_NONE,
2954 ("vm_page_free_toq: unmanaged page %p is queued", m));
2955 PCPU_INC(cnt.v_tfree);
2956
2957 if (vm_page_sbusied(m))
2958 panic("vm_page_free: freeing busy page %p", m);
2959
2960 /*
2961 * Unqueue, then remove page. Note that we cannot destroy
2962 * the page here because we do not want to call the pager's
2963 * callback routine until after we've put the page on the
2964 * appropriate free queue.
2965 */
2966 vm_page_remque(m);
2967 vm_page_remove(m);
2968
2969 /*
2970 * If fictitious remove object association and
2971 * return, otherwise delay object association removal.
2972 */
2973 if ((m->flags & PG_FICTITIOUS) != 0) {
2974 return;
2975 }
2976
2977 m->valid = 0;
2978 vm_page_undirty(m);
2979
2980 if (m->wire_count != 0)
2981 panic("vm_page_free: freeing wired page %p", m);
2982 if (m->hold_count != 0) {
2983 m->flags &= ~PG_ZERO;
2984 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2985 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2986 m->flags |= PG_UNHOLDFREE;
2987 } else {
2988 /*
2989 * Restore the default memory attribute to the page.
2990 */
2991 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2992 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2993
2994 /*
2995 * Insert the page into the physical memory allocator's
2996 * cache/free page queues.
2997 */
2998 mtx_lock(&vm_page_queue_free_mtx);
2999 vm_phys_freecnt_adj(m, 1);
3000#if VM_NRESERVLEVEL > 0
3001 if (!vm_reserv_free_page(m))
3002#else
3003 if (TRUE)
3004#endif
3005 vm_phys_free_pages(m, 0);
3006 if ((m->flags & PG_ZERO) != 0)
3007 ++vm_page_zero_count;
3008 else
3009 vm_page_zero_idle_wakeup();
3010 vm_page_free_wakeup();
3011 mtx_unlock(&vm_page_queue_free_mtx);
3012 }
3013}
3014
3015/*
3016 * vm_page_wire:
3017 *
3018 * Mark this page as wired down by yet
3019 * another map, removing it from paging queues
3020 * as necessary.
3021 *
3022 * If the page is fictitious, then its wire count must remain one.
3023 *
3024 * The page must be locked.
3025 */
3026void
3027vm_page_wire(vm_page_t m)
3028{
3029
3030 /*
3031 * Only bump the wire statistics if the page is not already wired,
3032 * and only unqueue the page if it is on some queue (if it is unmanaged
3033 * it is already off the queues).
3034 */
3035 vm_page_lock_assert(m, MA_OWNED);
3036 if ((m->flags & PG_FICTITIOUS) != 0) {
3037 KASSERT(m->wire_count == 1,
3038 ("vm_page_wire: fictitious page %p's wire count isn't one",
3039 m));
3040 return;
3041 }
3042 if (m->wire_count == 0) {
3043 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3044 m->queue == PQ_NONE,
3045 ("vm_page_wire: unmanaged page %p is queued", m));
3046 vm_page_remque(m);
3047 atomic_add_int(&vm_cnt.v_wire_count, 1);
3048 }
3049 m->wire_count++;
3050 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3051}
3052
3053/*
3054 * vm_page_unwire:
3055 *
3056 * Release one wiring of the specified page, potentially allowing it to be
3057 * paged out. Returns TRUE if the number of wirings transitions to zero and
3058 * FALSE otherwise.
3059 *
3060 * Only managed pages belonging to an object can be paged out. If the number
3061 * of wirings transitions to zero and the page is eligible for page out, then
3062 * the page is added to the specified paging queue (unless PQ_NONE is
3063 * specified).
3064 *
3065 * If a page is fictitious, then its wire count must always be one.
3066 *
3067 * A managed page must be locked.
3068 */
3069boolean_t
3070vm_page_unwire(vm_page_t m, uint8_t queue)
3071{
3072
3073 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3074 ("vm_page_unwire: invalid queue %u request for page %p",
3075 queue, m));
3076 if ((m->oflags & VPO_UNMANAGED) == 0)
3077 vm_page_assert_locked(m);
3078 if ((m->flags & PG_FICTITIOUS) != 0) {
3079 KASSERT(m->wire_count == 1,
3080 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3081 return (FALSE);
3082 }
3083 if (m->wire_count > 0) {
3084 m->wire_count--;
3085 if (m->wire_count == 0) {
3086 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
3087 if ((m->oflags & VPO_UNMANAGED) == 0 &&
3088 m->object != NULL && queue != PQ_NONE) {
3089 if (queue == PQ_INACTIVE)
3090 m->flags &= ~PG_WINATCFLS;
3091 vm_page_enqueue(queue, m);
3092 }
3093 return (TRUE);
3094 } else
3095 return (FALSE);
3096 } else
3097 panic("vm_page_unwire: page %p's wire count is zero", m);
3098}
3099
3100/*
3101 * Move the specified page to the inactive queue.
3102 *
3103 * Many pages placed on the inactive queue should actually go
3104 * into the cache, but it is difficult to figure out which. What
3105 * we do instead, if the inactive target is well met, is to put
3106 * clean pages at the head of the inactive queue instead of the tail.
3107 * This will cause them to be moved to the cache more quickly and
3108 * if not actively re-referenced, reclaimed more quickly. If we just
3109 * stick these pages at the end of the inactive queue, heavy filesystem
3110 * meta-data accesses can cause an unnecessary paging load on memory bound
3111 * processes. This optimization causes one-time-use metadata to be
3112 * reused more quickly.
3113 *
3114 * Normally noreuse is FALSE, resulting in LRU operation. noreuse is set
3115 * to TRUE if we want this page to be 'as if it were placed in the cache',
3116 * except without unmapping it from the process address space. In
3117 * practice this is implemented by inserting the page at the head of the
3118 * queue, using a marker page to guide FIFO insertion ordering.
3119 *
3120 * The page must be locked.
3121 */
3122static inline void
3123_vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3124{
3125 struct vm_pagequeue *pq;
3126 int queue;
3127
3128 vm_page_assert_locked(m);
3129
3130 /*
3131 * Ignore if the page is already inactive, unless it is unlikely to be
3132 * reactivated.
3133 */
3134 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
3135 return;
3136 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3137 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3138 /* Avoid multiple acquisitions of the inactive queue lock. */
3139 if (queue == PQ_INACTIVE) {
3140 vm_pagequeue_lock(pq);
3141 vm_page_dequeue_locked(m);
3142 } else {
3143 if (queue != PQ_NONE)
3144 vm_page_dequeue(m);
3145 m->flags &= ~PG_WINATCFLS;
3146 vm_pagequeue_lock(pq);
3147 }
3148 m->queue = PQ_INACTIVE;
3149 if (noreuse)
3150 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
3151 m, plinks.q);
3152 else
3153 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3154 vm_pagequeue_cnt_inc(pq);
3155 vm_pagequeue_unlock(pq);
3156 }
3157}
3158
3159/*
3160 * Move the specified page to the inactive queue.
3161 *
3162 * The page must be locked.
3163 */
3164void
3165vm_page_deactivate(vm_page_t m)
3166{
3167
3168 _vm_page_deactivate(m, FALSE);
3169}
3170
3171/*
3172 * Move the specified page to the inactive queue with the expectation
3173 * that it is unlikely to be reused.
3174 *
3175 * The page must be locked.
3176 */
3177void
3178vm_page_deactivate_noreuse(vm_page_t m)
3179{
3180
3181 _vm_page_deactivate(m, TRUE);
3182}
3183
3184/*
3185 * vm_page_try_to_cache:
3186 *
3187 * Returns 0 on failure, 1 on success
3188 */
3189int
3190vm_page_try_to_cache(vm_page_t m)
3191{
3192
3193 vm_page_lock_assert(m, MA_OWNED);
3194 VM_OBJECT_ASSERT_WLOCKED(m->object);
3195 if (m->dirty || m->hold_count || m->wire_count ||
3196 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3197 return (0);
3198 pmap_remove_all(m);
3199 if (m->dirty)
3200 return (0);
3201 vm_page_cache(m);
3202 return (1);
3203}
3204
3205/*
3206 * vm_page_try_to_free()
3207 *
3208 * Attempt to free the page. If we cannot free it, we do nothing.
3209 * 1 is returned on success, 0 on failure.
3210 */
3211int
3212vm_page_try_to_free(vm_page_t m)
3213{
3214
3215 vm_page_lock_assert(m, MA_OWNED);
3216 if (m->object != NULL)
3217 VM_OBJECT_ASSERT_WLOCKED(m->object);
3218 if (m->dirty || m->hold_count || m->wire_count ||
3219 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3220 return (0);
3221 pmap_remove_all(m);
3222 if (m->dirty)
3223 return (0);
3224 vm_page_free(m);
3225 return (1);
3226}
3227
3228/*
3229 * vm_page_cache
3230 *
3231 * Put the specified page onto the page cache queue (if appropriate).
3232 *
3233 * The object and page must be locked.
3234 */
3235void
3236vm_page_cache(vm_page_t m)
3237{
3238 vm_object_t object;
3239 boolean_t cache_was_empty;
3240
3241 vm_page_lock_assert(m, MA_OWNED);
3242 object = m->object;
3243 VM_OBJECT_ASSERT_WLOCKED(object);
3244 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) ||
3245 m->hold_count || m->wire_count)
3246 panic("vm_page_cache: attempting to cache busy page");
3247 KASSERT(!pmap_page_is_mapped(m),
3248 ("vm_page_cache: page %p is mapped", m));
3249 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
3250 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
3251 (object->type == OBJT_SWAP &&
3252 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
3253 /*
3254 * Hypothesis: A cache-eligible page belonging to a
3255 * default object or swap object but without a backing
3256 * store must be zero filled.
3257 */
3258 vm_page_free(m);
3259 return;
3260 }
3261 KASSERT((m->flags & PG_CACHED) == 0,
3262 ("vm_page_cache: page %p is already cached", m));
3263
3264 /*
3265 * Remove the page from the paging queues.
3266 */
3267 vm_page_remque(m);
3268
3269 /*
3270 * Remove the page from the object's collection of resident
3271 * pages.
3272 */
3273 vm_radix_remove(&object->rtree, m->pindex);
3274 TAILQ_REMOVE(&object->memq, m, listq);
3275 object->resident_page_count--;
3276
3277 /*
3278 * Restore the default memory attribute to the page.
3279 */
3280 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3281 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3282
3283 /*
3284 * Insert the page into the object's collection of cached pages
3285 * and the physical memory allocator's cache/free page queues.
3286 */
3287 m->flags &= ~PG_ZERO;
3288 mtx_lock(&vm_page_queue_free_mtx);
3289 cache_was_empty = vm_radix_is_empty(&object->cache);
3290 if (vm_radix_insert(&object->cache, m)) {
3291 mtx_unlock(&vm_page_queue_free_mtx);
3292 if (object->resident_page_count == 0)
3293 vdrop(object->handle);
3294 m->object = NULL;
3295 vm_page_free(m);
3296 return;
3297 }
3298
3299 /*
3300 * The above call to vm_radix_insert() could reclaim the one pre-
3301 * existing cached page from this object, resulting in a call to
3302 * vdrop().
3303 */
3304 if (!cache_was_empty)
3305 cache_was_empty = vm_radix_is_singleton(&object->cache);
3306
3307 m->flags |= PG_CACHED;
3308 vm_cnt.v_cache_count++;
3309 PCPU_INC(cnt.v_tcached);
3310#if VM_NRESERVLEVEL > 0
3311 if (!vm_reserv_free_page(m)) {
3312#else
3313 if (TRUE) {
3314#endif
3315 vm_phys_free_pages(m, 0);
3316 }
3317 vm_page_free_wakeup();
3318 mtx_unlock(&vm_page_queue_free_mtx);
3319
3320 /*
3321 * Increment the vnode's hold count if this is the object's only
3322 * cached page. Decrement the vnode's hold count if this was
3323 * the object's only resident page.
3324 */
3325 if (object->type == OBJT_VNODE) {
3326 if (cache_was_empty && object->resident_page_count != 0)
3327 vhold(object->handle);
3328 else if (!cache_was_empty && object->resident_page_count == 0)
3329 vdrop(object->handle);
3330 }
3331}
3332
3333/*
3334 * vm_page_advise
3335 *
3336 * Deactivate or do nothing, as appropriate.
3337 *
3338 * The object and page must be locked.
3339 */
3340void
3341vm_page_advise(vm_page_t m, int advice)
3342{
3343
3344 vm_page_assert_locked(m);
3345 VM_OBJECT_ASSERT_WLOCKED(m->object);
3346 if (advice == MADV_FREE)
3347 /*
3348 * Mark the page clean. This will allow the page to be freed
3349 * up by the system. However, such pages are often reused
3350 * quickly by malloc() so we do not do anything that would
3351 * cause a page fault if we can help it.
3352 *
3353 * Specifically, we do not try to actually free the page now
3354 * nor do we try to put it in the cache (which would cause a
3355 * page fault on reuse).
3356 *
3357 * But we do make the page as freeable as we can without
3358 * actually taking the step of unmapping it.
3359 */
3360 m->dirty = 0;
3361 else if (advice != MADV_DONTNEED)
3362 return;
3363
3364 /*
3365 * Clear any references to the page. Otherwise, the page daemon will
3366 * immediately reactivate the page.
3367 */
3368 vm_page_aflag_clear(m, PGA_REFERENCED);
3369
3370 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3371 vm_page_dirty(m);
3372
3373 /*
3374 * Place clean pages at the head of the inactive queue rather than the
3375 * tail, thus defeating the queue's LRU operation and ensuring that the
3376 * page will be reused quickly.
3377 */
3378 _vm_page_deactivate(m, m->dirty == 0);
3379}
3380
3381/*
3382 * Grab a page, waiting until we are waken up due to the page
3383 * changing state. We keep on waiting, if the page continues
3384 * to be in the object. If the page doesn't exist, first allocate it
3385 * and then conditionally zero it.
3386 *
3387 * This routine may sleep.
3388 *
3389 * The object must be locked on entry. The lock will, however, be released
3390 * and reacquired if the routine sleeps.
3391 */
3392vm_page_t
3393vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3394{
3395 vm_page_t m;
3396 int sleep;
3397
3398 VM_OBJECT_ASSERT_WLOCKED(object);
3399 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3400 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3401 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3402retrylookup:
3403 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3404 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3405 vm_page_xbusied(m) : vm_page_busied(m);
3406 if (sleep) {
3407 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3408 return (NULL);
3409 /*
3410 * Reference the page before unlocking and
3411 * sleeping so that the page daemon is less
3412 * likely to reclaim it.
3413 */
3414 vm_page_aflag_set(m, PGA_REFERENCED);
3415 vm_page_lock(m);
3416 VM_OBJECT_WUNLOCK(object);
3417 vm_page_busy_sleep(m, "pgrbwt");
3418 VM_OBJECT_WLOCK(object);
3419 goto retrylookup;
3420 } else {
3421 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3422 vm_page_lock(m);
3423 vm_page_wire(m);
3424 vm_page_unlock(m);
3425 }
3426 if ((allocflags &
3427 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3428 vm_page_xbusy(m);
3429 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3430 vm_page_sbusy(m);
3431 return (m);
3432 }
3433 }
3434 m = vm_page_alloc(object, pindex, allocflags);
3435 if (m == NULL) {
3436 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3437 return (NULL);
3438 VM_OBJECT_WUNLOCK(object);
3439 VM_WAIT;
3440 VM_OBJECT_WLOCK(object);
3441 goto retrylookup;
3442 } else if (m->valid != 0)
3443 return (m);
3444 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3445 pmap_zero_page(m);
3446 return (m);
3447}
3448
3449/*
3450 * Mapping function for valid or dirty bits in a page.
3451 *
3452 * Inputs are required to range within a page.
3453 */
3454vm_page_bits_t
3455vm_page_bits(int base, int size)
3456{
3457 int first_bit;
3458 int last_bit;
3459
3460 KASSERT(
3461 base + size <= PAGE_SIZE,
3462 ("vm_page_bits: illegal base/size %d/%d", base, size)
3463 );
3464
3465 if (size == 0) /* handle degenerate case */
3466 return (0);
3467
3468 first_bit = base >> DEV_BSHIFT;
3469 last_bit = (base + size - 1) >> DEV_BSHIFT;
3470
3471 return (((vm_page_bits_t)2 << last_bit) -
3472 ((vm_page_bits_t)1 << first_bit));
3473}
3474
3475/*
3476 * vm_page_set_valid_range:
3477 *
3478 * Sets portions of a page valid. The arguments are expected
3479 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3480 * of any partial chunks touched by the range. The invalid portion of
3481 * such chunks will be zeroed.
3482 *
3483 * (base + size) must be less then or equal to PAGE_SIZE.
3484 */
3485void
3486vm_page_set_valid_range(vm_page_t m, int base, int size)
3487{
3488 int endoff, frag;
3489
3490 VM_OBJECT_ASSERT_WLOCKED(m->object);
3491 if (size == 0) /* handle degenerate case */
3492 return;
3493
3494 /*
3495 * If the base is not DEV_BSIZE aligned and the valid
3496 * bit is clear, we have to zero out a portion of the
3497 * first block.
3498 */
3499 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3500 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3501 pmap_zero_page_area(m, frag, base - frag);
3502
3503 /*
3504 * If the ending offset is not DEV_BSIZE aligned and the
3505 * valid bit is clear, we have to zero out a portion of
3506 * the last block.
3507 */
3508 endoff = base + size;
3509 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3510 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3511 pmap_zero_page_area(m, endoff,
3512 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3513
3514 /*
3515 * Assert that no previously invalid block that is now being validated
3516 * is already dirty.
3517 */
3518 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3519 ("vm_page_set_valid_range: page %p is dirty", m));
3520
3521 /*
3522 * Set valid bits inclusive of any overlap.
3523 */
3524 m->valid |= vm_page_bits(base, size);
3525}
3526
3527/*
3528 * Clear the given bits from the specified page's dirty field.
3529 */
3530static __inline void
3531vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3532{
3533 uintptr_t addr;
3534#if PAGE_SIZE < 16384
3535 int shift;
3536#endif
3537
3538 /*
3539 * If the object is locked and the page is neither exclusive busy nor
3540 * write mapped, then the page's dirty field cannot possibly be
3541 * set by a concurrent pmap operation.
3542 */
3543 VM_OBJECT_ASSERT_WLOCKED(m->object);
3544 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3545 m->dirty &= ~pagebits;
3546 else {
3547 /*
3548 * The pmap layer can call vm_page_dirty() without
3549 * holding a distinguished lock. The combination of
3550 * the object's lock and an atomic operation suffice
3551 * to guarantee consistency of the page dirty field.
3552 *
3553 * For PAGE_SIZE == 32768 case, compiler already
3554 * properly aligns the dirty field, so no forcible
3555 * alignment is needed. Only require existence of
3556 * atomic_clear_64 when page size is 32768.
3557 */
3558 addr = (uintptr_t)&m->dirty;
3559#if PAGE_SIZE == 32768
3560 atomic_clear_64((uint64_t *)addr, pagebits);
3561#elif PAGE_SIZE == 16384
3562 atomic_clear_32((uint32_t *)addr, pagebits);
3563#else /* PAGE_SIZE <= 8192 */
3564 /*
3565 * Use a trick to perform a 32-bit atomic on the
3566 * containing aligned word, to not depend on the existence
3567 * of atomic_clear_{8, 16}.
3568 */
3569 shift = addr & (sizeof(uint32_t) - 1);
3570#if BYTE_ORDER == BIG_ENDIAN
3571 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3572#else
3573 shift *= NBBY;
3574#endif
3575 addr &= ~(sizeof(uint32_t) - 1);
3576 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3577#endif /* PAGE_SIZE */
3578 }
3579}
3580
3581/*
3582 * vm_page_set_validclean:
3583 *
3584 * Sets portions of a page valid and clean. The arguments are expected
3585 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3586 * of any partial chunks touched by the range. The invalid portion of
3587 * such chunks will be zero'd.
3588 *
3589 * (base + size) must be less then or equal to PAGE_SIZE.
3590 */
3591void
3592vm_page_set_validclean(vm_page_t m, int base, int size)
3593{
3594 vm_page_bits_t oldvalid, pagebits;
3595 int endoff, frag;
3596
3597 VM_OBJECT_ASSERT_WLOCKED(m->object);
3598 if (size == 0) /* handle degenerate case */
3599 return;
3600
3601 /*
3602 * If the base is not DEV_BSIZE aligned and the valid
3603 * bit is clear, we have to zero out a portion of the
3604 * first block.
3605 */
3606 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3607 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3608 pmap_zero_page_area(m, frag, base - frag);
3609
3610 /*
3611 * If the ending offset is not DEV_BSIZE aligned and the
3612 * valid bit is clear, we have to zero out a portion of
3613 * the last block.
3614 */
3615 endoff = base + size;
3616 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3617 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3618 pmap_zero_page_area(m, endoff,
3619 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3620
3621 /*
3622 * Set valid, clear dirty bits. If validating the entire
3623 * page we can safely clear the pmap modify bit. We also
3624 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3625 * takes a write fault on a MAP_NOSYNC memory area the flag will
3626 * be set again.
3627 *
3628 * We set valid bits inclusive of any overlap, but we can only
3629 * clear dirty bits for DEV_BSIZE chunks that are fully within
3630 * the range.
3631 */
3632 oldvalid = m->valid;
3633 pagebits = vm_page_bits(base, size);
3634 m->valid |= pagebits;
3635#if 0 /* NOT YET */
3636 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3637 frag = DEV_BSIZE - frag;
3638 base += frag;
3639 size -= frag;
3640 if (size < 0)
3641 size = 0;
3642 }
3643 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3644#endif
3645 if (base == 0 && size == PAGE_SIZE) {
3646 /*
3647 * The page can only be modified within the pmap if it is
3648 * mapped, and it can only be mapped if it was previously
3649 * fully valid.
3650 */
3651 if (oldvalid == VM_PAGE_BITS_ALL)
3652 /*
3653 * Perform the pmap_clear_modify() first. Otherwise,
3654 * a concurrent pmap operation, such as
3655 * pmap_protect(), could clear a modification in the
3656 * pmap and set the dirty field on the page before
3657 * pmap_clear_modify() had begun and after the dirty
3658 * field was cleared here.
3659 */
3660 pmap_clear_modify(m);
3661 m->dirty = 0;
3662 m->oflags &= ~VPO_NOSYNC;
3663 } else if (oldvalid != VM_PAGE_BITS_ALL)
3664 m->dirty &= ~pagebits;
3665 else
3666 vm_page_clear_dirty_mask(m, pagebits);
3667}
3668
3669void
3670vm_page_clear_dirty(vm_page_t m, int base, int size)
3671{
3672
3673 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3674}
3675
3676/*
3677 * vm_page_set_invalid:
3678 *
3679 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3680 * valid and dirty bits for the effected areas are cleared.
3681 */
3682void
3683vm_page_set_invalid(vm_page_t m, int base, int size)
3684{
3685 vm_page_bits_t bits;
3686 vm_object_t object;
3687
3688 object = m->object;
3689 VM_OBJECT_ASSERT_WLOCKED(object);
3690 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3691 size >= object->un_pager.vnp.vnp_size)
3692 bits = VM_PAGE_BITS_ALL;
3693 else
3694 bits = vm_page_bits(base, size);
3695 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3696 bits != 0)
3697 pmap_remove_all(m);
3698 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3699 !pmap_page_is_mapped(m),
3700 ("vm_page_set_invalid: page %p is mapped", m));
3701 m->valid &= ~bits;
3702 m->dirty &= ~bits;
3703}
3704
3705/*
3706 * vm_page_zero_invalid()
3707 *
3708 * The kernel assumes that the invalid portions of a page contain
3709 * garbage, but such pages can be mapped into memory by user code.
3710 * When this occurs, we must zero out the non-valid portions of the
3711 * page so user code sees what it expects.
3712 *
3713 * Pages are most often semi-valid when the end of a file is mapped
3714 * into memory and the file's size is not page aligned.
3715 */
3716void
3717vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3718{
3719 int b;
3720 int i;
3721
3722 VM_OBJECT_ASSERT_WLOCKED(m->object);
3723 /*
3724 * Scan the valid bits looking for invalid sections that
3725 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3726 * valid bit may be set ) have already been zeroed by
3727 * vm_page_set_validclean().
3728 */
3729 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3730 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3731 (m->valid & ((vm_page_bits_t)1 << i))) {
3732 if (i > b) {
3733 pmap_zero_page_area(m,
3734 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3735 }
3736 b = i + 1;
3737 }
3738 }
3739
3740 /*
3741 * setvalid is TRUE when we can safely set the zero'd areas
3742 * as being valid. We can do this if there are no cache consistancy
3743 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3744 */
3745 if (setvalid)
3746 m->valid = VM_PAGE_BITS_ALL;
3747}
3748
3749/*
3750 * vm_page_is_valid:
3751 *
3752 * Is (partial) page valid? Note that the case where size == 0
3753 * will return FALSE in the degenerate case where the page is
3754 * entirely invalid, and TRUE otherwise.
3755 */
3756int
3757vm_page_is_valid(vm_page_t m, int base, int size)
3758{
3759 vm_page_bits_t bits;
3760
3761 VM_OBJECT_ASSERT_LOCKED(m->object);
3762 bits = vm_page_bits(base, size);
3763 return (m->valid != 0 && (m->valid & bits) == bits);
3764}
3765
3766/*
3767 * vm_page_ps_is_valid:
3768 *
3769 * Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3770 */
3771boolean_t
3772vm_page_ps_is_valid(vm_page_t m)
3773{
3774 int i, npages;
3775
3776 VM_OBJECT_ASSERT_LOCKED(m->object);
3777 npages = atop(pagesizes[m->psind]);
3778
3779 /*
3780 * The physically contiguous pages that make up a superpage, i.e., a
3781 * page with a page size index ("psind") greater than zero, will
3782 * occupy adjacent entries in vm_page_array[].
3783 */
3784 for (i = 0; i < npages; i++) {
3785 if (m[i].valid != VM_PAGE_BITS_ALL)
3786 return (FALSE);
3787 }
3788 return (TRUE);
3789}
3790
3791/*
3792 * Set the page's dirty bits if the page is modified.
3793 */
3794void
3795vm_page_test_dirty(vm_page_t m)
3796{
3797
3798 VM_OBJECT_ASSERT_WLOCKED(m->object);
3799 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3800 vm_page_dirty(m);
3801}
3802
3803void
3804vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3805{
3806
3807 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3808}
3809
3810void
3811vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3812{
3813
3814 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3815}
3816
3817int
3818vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3819{
3820
3821 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3822}
3823
3824#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3825void
3826vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3827{
3828
3829 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3830}
3831
3832void
3833vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3834{
3835
3836 mtx_assert_(vm_page_lockptr(m), a, file, line);
3837}
3838#endif
3839
3840#ifdef INVARIANTS
3841void
3842vm_page_object_lock_assert(vm_page_t m)
3843{
3844
3845 /*
3846 * Certain of the page's fields may only be modified by the
3847 * holder of the containing object's lock or the exclusive busy.
3848 * holder. Unfortunately, the holder of the write busy is
3849 * not recorded, and thus cannot be checked here.
3850 */
3851 if (m->object != NULL && !vm_page_xbusied(m))
3852 VM_OBJECT_ASSERT_WLOCKED(m->object);
3853}
3854
3855void
3856vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3857{
3858
3859 if ((bits & PGA_WRITEABLE) == 0)
3860 return;
3861
3862 /*
3863 * The PGA_WRITEABLE flag can only be set if the page is
3864 * managed, is exclusively busied or the object is locked.
3865 * Currently, this flag is only set by pmap_enter().
3866 */
3867 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3868 ("PGA_WRITEABLE on unmanaged page"));
3869 if (!vm_page_xbusied(m))
3870 VM_OBJECT_ASSERT_LOCKED(m->object);
3871}
3872#endif
3873
3874#include "opt_ddb.h"
3875#ifdef DDB
3876#include <sys/kernel.h>
3877
3878#include <ddb/ddb.h>
3879
3880DB_SHOW_COMMAND(page, vm_page_print_page_info)
3881{
3882 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3883 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count);
3884 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3885 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3886 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3887 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3888 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3889 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3890 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3891}
3892
3893DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3894{
3895 int dom;
3896
3897 db_printf("pq_free %d pq_cache %d\n",
3898 vm_cnt.v_free_count, vm_cnt.v_cache_count);
3899 for (dom = 0; dom < vm_ndomains; dom++) {
3900 db_printf(
3901 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3902 dom,
3903 vm_dom[dom].vmd_page_count,
3904 vm_dom[dom].vmd_free_count,
3905 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3906 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3907 vm_dom[dom].vmd_pass);
3908 }
3909}
3910
3911DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3912{
3913 vm_page_t m;
3914 boolean_t phys;
3915
3916 if (!have_addr) {
3917 db_printf("show pginfo addr\n");
3918 return;
3919 }
3920
3921 phys = strchr(modif, 'p') != NULL;
3922 if (phys)
3923 m = PHYS_TO_VM_PAGE(addr);
3924 else
3925 m = (vm_page_t)addr;
3926 db_printf(
3927 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3928 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3929 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3930 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3931 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3932}
3933#endif /* DDB */