FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_page.c
1 /*-
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
5 *
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 *
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
34 */
35
36 /*-
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
39 *
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41 *
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
47 *
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51 *
52 * Carnegie Mellon requests users of this software to return to
53 *
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
58 *
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
61 */
62
63 /*
64 * GENERAL RULES ON VM_PAGE MANIPULATION
65 *
66 * - A 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: releng/11.1/sys/vm/vm_page.c 337828 2018-08-15 02:30:11Z delphij $");
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/smp.h>
103 #include <sys/sysctl.h>
104 #include <sys/vmmeter.h>
105 #include <sys/vnode.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
129 struct vm_domain vm_dom[MAXMEMDOM];
130 struct mtx_padalign vm_page_queue_free_mtx;
131
132 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
133
134 vm_page_t vm_page_array;
135 long vm_page_array_size;
136 long first_page;
137 int vm_page_zero_count;
138
139 static int boot_pages = UMA_BOOT_PAGES;
140 SYSCTL_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
144 static int pa_tryrelock_restart;
145 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
146 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
147
148 static TAILQ_HEAD(, vm_page) blacklist_head;
149 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
150 SYSCTL_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? */
154 static int vm_pageout_pages_needed;
155
156 static uma_zone_t fakepg_zone;
157
158 static void vm_page_alloc_check(vm_page_t m);
159 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
160 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
161 static void vm_page_free_wakeup(void);
162 static void vm_page_init_fakepg(void *dummy);
163 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
164 vm_pindex_t pindex, vm_page_t mpred);
165 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
166 vm_page_t mpred);
167 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
168 vm_paddr_t high);
169
170 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
171
172 static void
173 vm_page_init_fakepg(void *dummy)
174 {
175
176 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
177 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
178 }
179
180 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
181 #if PAGE_SIZE == 32768
182 #ifdef CTASSERT
183 CTASSERT(sizeof(u_long) >= 8);
184 #endif
185 #endif
186
187 /*
188 * Try to acquire a physical address lock while a pmap is locked. If we
189 * fail to trylock we unlock and lock the pmap directly and cache the
190 * locked pa in *locked. The caller should then restart their loop in case
191 * the virtual to physical mapping has changed.
192 */
193 int
194 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
195 {
196 vm_paddr_t lockpa;
197
198 lockpa = *locked;
199 *locked = pa;
200 if (lockpa) {
201 PA_LOCK_ASSERT(lockpa, MA_OWNED);
202 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
203 return (0);
204 PA_UNLOCK(lockpa);
205 }
206 if (PA_TRYLOCK(pa))
207 return (0);
208 PMAP_UNLOCK(pmap);
209 atomic_add_int(&pa_tryrelock_restart, 1);
210 PA_LOCK(pa);
211 PMAP_LOCK(pmap);
212 return (EAGAIN);
213 }
214
215 /*
216 * vm_set_page_size:
217 *
218 * Sets the page size, perhaps based upon the memory
219 * size. Must be called before any use of page-size
220 * dependent functions.
221 */
222 void
223 vm_set_page_size(void)
224 {
225 if (vm_cnt.v_page_size == 0)
226 vm_cnt.v_page_size = PAGE_SIZE;
227 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
228 panic("vm_set_page_size: page size not a power of two");
229 }
230
231 /*
232 * vm_page_blacklist_next:
233 *
234 * Find the next entry in the provided string of blacklist
235 * addresses. Entries are separated by space, comma, or newline.
236 * If an invalid integer is encountered then the rest of the
237 * string is skipped. Updates the list pointer to the next
238 * character, or NULL if the string is exhausted or invalid.
239 */
240 static vm_paddr_t
241 vm_page_blacklist_next(char **list, char *end)
242 {
243 vm_paddr_t bad;
244 char *cp, *pos;
245
246 if (list == NULL || *list == NULL)
247 return (0);
248 if (**list =='\0') {
249 *list = NULL;
250 return (0);
251 }
252
253 /*
254 * If there's no end pointer then the buffer is coming from
255 * the kenv and we know it's null-terminated.
256 */
257 if (end == NULL)
258 end = *list + strlen(*list);
259
260 /* Ensure that strtoq() won't walk off the end */
261 if (*end != '\0') {
262 if (*end == '\n' || *end == ' ' || *end == ',')
263 *end = '\0';
264 else {
265 printf("Blacklist not terminated, skipping\n");
266 *list = NULL;
267 return (0);
268 }
269 }
270
271 for (pos = *list; *pos != '\0'; pos = cp) {
272 bad = strtoq(pos, &cp, 0);
273 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
274 if (bad == 0) {
275 if (++cp < end)
276 continue;
277 else
278 break;
279 }
280 } else
281 break;
282 if (*cp == '\0' || ++cp >= end)
283 *list = NULL;
284 else
285 *list = cp;
286 return (trunc_page(bad));
287 }
288 printf("Garbage in RAM blacklist, skipping\n");
289 *list = NULL;
290 return (0);
291 }
292
293 bool
294 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
295 {
296 vm_page_t m;
297 int ret;
298
299 m = vm_phys_paddr_to_vm_page(pa);
300 if (m == NULL)
301 return (true); /* page does not exist, no failure */
302
303 mtx_lock(&vm_page_queue_free_mtx);
304 ret = vm_phys_unfree_page(m);
305 mtx_unlock(&vm_page_queue_free_mtx);
306 if (ret) {
307 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
308 if (verbose)
309 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
310 }
311 return (ret);
312 }
313
314 /*
315 * vm_page_blacklist_check:
316 *
317 * Iterate through the provided string of blacklist addresses, pulling
318 * each entry out of the physical allocator free list and putting it
319 * onto a list for reporting via the vm.page_blacklist sysctl.
320 */
321 static void
322 vm_page_blacklist_check(char *list, char *end)
323 {
324 vm_paddr_t pa;
325 char *next;
326
327 next = list;
328 while (next != NULL) {
329 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
330 continue;
331 vm_page_blacklist_add(pa, bootverbose);
332 }
333 }
334
335 /*
336 * vm_page_blacklist_load:
337 *
338 * Search for a special module named "ram_blacklist". It'll be a
339 * plain text file provided by the user via the loader directive
340 * of the same name.
341 */
342 static void
343 vm_page_blacklist_load(char **list, char **end)
344 {
345 void *mod;
346 u_char *ptr;
347 u_int len;
348
349 mod = NULL;
350 ptr = NULL;
351
352 mod = preload_search_by_type("ram_blacklist");
353 if (mod != NULL) {
354 ptr = preload_fetch_addr(mod);
355 len = preload_fetch_size(mod);
356 }
357 *list = ptr;
358 if (ptr != NULL)
359 *end = ptr + len;
360 else
361 *end = NULL;
362 return;
363 }
364
365 static int
366 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
367 {
368 vm_page_t m;
369 struct sbuf sbuf;
370 int error, first;
371
372 first = 1;
373 error = sysctl_wire_old_buffer(req, 0);
374 if (error != 0)
375 return (error);
376 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
377 TAILQ_FOREACH(m, &blacklist_head, listq) {
378 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
379 (uintmax_t)m->phys_addr);
380 first = 0;
381 }
382 error = sbuf_finish(&sbuf);
383 sbuf_delete(&sbuf);
384 return (error);
385 }
386
387 static void
388 vm_page_domain_init(struct vm_domain *vmd)
389 {
390 struct vm_pagequeue *pq;
391 int i;
392
393 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
394 "vm inactive pagequeue";
395 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
396 &vm_cnt.v_inactive_count;
397 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
398 "vm active pagequeue";
399 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
400 &vm_cnt.v_active_count;
401 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
402 "vm laundry pagequeue";
403 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
404 &vm_cnt.v_laundry_count;
405 vmd->vmd_page_count = 0;
406 vmd->vmd_free_count = 0;
407 vmd->vmd_segs = 0;
408 vmd->vmd_oom = FALSE;
409 for (i = 0; i < PQ_COUNT; i++) {
410 pq = &vmd->vmd_pagequeues[i];
411 TAILQ_INIT(&pq->pq_pl);
412 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
413 MTX_DEF | MTX_DUPOK);
414 }
415 }
416
417 /*
418 * vm_page_startup:
419 *
420 * Initializes the resident memory module. Allocates physical memory for
421 * bootstrapping UMA and some data structures that are used to manage
422 * physical pages. Initializes these structures, and populates the free
423 * page queues.
424 */
425 vm_offset_t
426 vm_page_startup(vm_offset_t vaddr)
427 {
428 vm_offset_t mapped;
429 vm_paddr_t high_avail, low_avail, page_range, size;
430 vm_paddr_t new_end;
431 int i;
432 vm_paddr_t pa;
433 vm_paddr_t last_pa;
434 char *list, *listend;
435 vm_paddr_t end;
436 vm_paddr_t biggestsize;
437 int biggestone;
438 int pages_per_zone;
439
440 biggestsize = 0;
441 biggestone = 0;
442 vaddr = round_page(vaddr);
443
444 for (i = 0; phys_avail[i + 1]; i += 2) {
445 phys_avail[i] = round_page(phys_avail[i]);
446 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
447 }
448 for (i = 0; phys_avail[i + 1]; i += 2) {
449 size = phys_avail[i + 1] - phys_avail[i];
450 if (size > biggestsize) {
451 biggestone = i;
452 biggestsize = size;
453 }
454 }
455
456 end = phys_avail[biggestone+1];
457
458 /*
459 * Initialize the page and queue locks.
460 */
461 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
462 for (i = 0; i < PA_LOCK_COUNT; i++)
463 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
464 for (i = 0; i < vm_ndomains; i++)
465 vm_page_domain_init(&vm_dom[i]);
466
467 /*
468 * Almost all of the pages needed for bootstrapping UMA are used
469 * for zone structures, so if the number of CPUs results in those
470 * structures taking more than one page each, we set aside more pages
471 * in proportion to the zone structure size.
472 */
473 pages_per_zone = howmany(sizeof(struct uma_zone) +
474 sizeof(struct uma_cache) * (mp_maxid + 1), UMA_SLAB_SIZE);
475 if (pages_per_zone > 1) {
476 /* Reserve more pages so that we don't run out. */
477 boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
478 }
479
480 /*
481 * Allocate memory for use when boot strapping the kernel memory
482 * allocator.
483 *
484 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
485 * manually fetch the value.
486 */
487 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
488 new_end = end - (boot_pages * UMA_SLAB_SIZE);
489 new_end = trunc_page(new_end);
490 mapped = pmap_map(&vaddr, new_end, end,
491 VM_PROT_READ | VM_PROT_WRITE);
492 bzero((void *)mapped, end - new_end);
493 uma_startup((void *)mapped, boot_pages);
494
495 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
496 defined(__i386__) || defined(__mips__)
497 /*
498 * Allocate a bitmap to indicate that a random physical page
499 * needs to be included in a minidump.
500 *
501 * The amd64 port needs this to indicate which direct map pages
502 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
503 *
504 * However, i386 still needs this workspace internally within the
505 * minidump code. In theory, they are not needed on i386, but are
506 * included should the sf_buf code decide to use them.
507 */
508 last_pa = 0;
509 for (i = 0; dump_avail[i + 1] != 0; i += 2)
510 if (dump_avail[i + 1] > last_pa)
511 last_pa = dump_avail[i + 1];
512 page_range = last_pa / PAGE_SIZE;
513 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
514 new_end -= vm_page_dump_size;
515 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
516 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
517 bzero((void *)vm_page_dump, vm_page_dump_size);
518 #endif
519 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
520 /*
521 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
522 * When pmap_map() uses the direct map, they are not automatically
523 * included.
524 */
525 for (pa = new_end; pa < end; pa += PAGE_SIZE)
526 dump_add_page(pa);
527 #endif
528 phys_avail[biggestone + 1] = new_end;
529 #ifdef __amd64__
530 /*
531 * Request that the physical pages underlying the message buffer be
532 * included in a crash dump. Since the message buffer is accessed
533 * through the direct map, they are not automatically included.
534 */
535 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
536 last_pa = pa + round_page(msgbufsize);
537 while (pa < last_pa) {
538 dump_add_page(pa);
539 pa += PAGE_SIZE;
540 }
541 #endif
542 /*
543 * Compute the number of pages of memory that will be available for
544 * use, taking into account the overhead of a page structure per page.
545 * In other words, solve
546 * "available physical memory" - round_page(page_range *
547 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
548 * for page_range.
549 */
550 low_avail = phys_avail[0];
551 high_avail = phys_avail[1];
552 for (i = 0; i < vm_phys_nsegs; i++) {
553 if (vm_phys_segs[i].start < low_avail)
554 low_avail = vm_phys_segs[i].start;
555 if (vm_phys_segs[i].end > high_avail)
556 high_avail = vm_phys_segs[i].end;
557 }
558 /* Skip the first chunk. It is already accounted for. */
559 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
560 if (phys_avail[i] < low_avail)
561 low_avail = phys_avail[i];
562 if (phys_avail[i + 1] > high_avail)
563 high_avail = phys_avail[i + 1];
564 }
565 first_page = low_avail / PAGE_SIZE;
566 #ifdef VM_PHYSSEG_SPARSE
567 size = 0;
568 for (i = 0; i < vm_phys_nsegs; i++)
569 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
570 for (i = 0; phys_avail[i + 1] != 0; i += 2)
571 size += phys_avail[i + 1] - phys_avail[i];
572 #elif defined(VM_PHYSSEG_DENSE)
573 size = high_avail - low_avail;
574 #else
575 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
576 #endif
577
578 #ifdef VM_PHYSSEG_DENSE
579 /*
580 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
581 * the overhead of a page structure per page only if vm_page_array is
582 * allocated from the last physical memory chunk. Otherwise, we must
583 * allocate page structures representing the physical memory
584 * underlying vm_page_array, even though they will not be used.
585 */
586 if (new_end != high_avail)
587 page_range = size / PAGE_SIZE;
588 else
589 #endif
590 {
591 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
592
593 /*
594 * If the partial bytes remaining are large enough for
595 * a page (PAGE_SIZE) without a corresponding
596 * 'struct vm_page', then new_end will contain an
597 * extra page after subtracting the length of the VM
598 * page array. Compensate by subtracting an extra
599 * page from new_end.
600 */
601 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
602 if (new_end == high_avail)
603 high_avail -= PAGE_SIZE;
604 new_end -= PAGE_SIZE;
605 }
606 }
607 end = new_end;
608
609 /*
610 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
611 * However, because this page is allocated from KVM, out-of-bounds
612 * accesses using the direct map will not be trapped.
613 */
614 vaddr += PAGE_SIZE;
615
616 /*
617 * Allocate physical memory for the page structures, and map it.
618 */
619 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
620 mapped = pmap_map(&vaddr, new_end, end,
621 VM_PROT_READ | VM_PROT_WRITE);
622 vm_page_array = (vm_page_t) mapped;
623 #if VM_NRESERVLEVEL > 0
624 /*
625 * Allocate physical memory for the reservation management system's
626 * data structures, and map it.
627 */
628 if (high_avail == end)
629 high_avail = new_end;
630 new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
631 #endif
632 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
633 /*
634 * Include vm_page_array and vm_reserv_array in a crash dump.
635 */
636 for (pa = new_end; pa < end; pa += PAGE_SIZE)
637 dump_add_page(pa);
638 #endif
639 phys_avail[biggestone + 1] = new_end;
640
641 /*
642 * Add physical memory segments corresponding to the available
643 * physical pages.
644 */
645 for (i = 0; phys_avail[i + 1] != 0; i += 2)
646 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
647
648 /*
649 * Clear all of the page structures
650 */
651 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
652 for (i = 0; i < page_range; i++)
653 vm_page_array[i].order = VM_NFREEORDER;
654 vm_page_array_size = page_range;
655
656 /*
657 * Initialize the physical memory allocator.
658 */
659 vm_phys_init();
660
661 /*
662 * Add every available physical page that is not blacklisted to
663 * the free lists.
664 */
665 vm_cnt.v_page_count = 0;
666 vm_cnt.v_free_count = 0;
667 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
668 pa = phys_avail[i];
669 last_pa = phys_avail[i + 1];
670 while (pa < last_pa) {
671 vm_phys_add_page(pa);
672 pa += PAGE_SIZE;
673 }
674 }
675
676 TAILQ_INIT(&blacklist_head);
677 vm_page_blacklist_load(&list, &listend);
678 vm_page_blacklist_check(list, listend);
679
680 list = kern_getenv("vm.blacklist");
681 vm_page_blacklist_check(list, NULL);
682
683 freeenv(list);
684 #if VM_NRESERVLEVEL > 0
685 /*
686 * Initialize the reservation management system.
687 */
688 vm_reserv_init();
689 #endif
690 return (vaddr);
691 }
692
693 void
694 vm_page_reference(vm_page_t m)
695 {
696
697 vm_page_aflag_set(m, PGA_REFERENCED);
698 }
699
700 /*
701 * vm_page_busy_downgrade:
702 *
703 * Downgrade an exclusive busy page into a single shared busy page.
704 */
705 void
706 vm_page_busy_downgrade(vm_page_t m)
707 {
708 u_int x;
709 bool locked;
710
711 vm_page_assert_xbusied(m);
712 locked = mtx_owned(vm_page_lockptr(m));
713
714 for (;;) {
715 x = m->busy_lock;
716 x &= VPB_BIT_WAITERS;
717 if (x != 0 && !locked)
718 vm_page_lock(m);
719 if (atomic_cmpset_rel_int(&m->busy_lock,
720 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
721 break;
722 if (x != 0 && !locked)
723 vm_page_unlock(m);
724 }
725 if (x != 0) {
726 wakeup(m);
727 if (!locked)
728 vm_page_unlock(m);
729 }
730 }
731
732 /*
733 * vm_page_sbusied:
734 *
735 * Return a positive value if the page is shared busied, 0 otherwise.
736 */
737 int
738 vm_page_sbusied(vm_page_t m)
739 {
740 u_int x;
741
742 x = m->busy_lock;
743 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
744 }
745
746 /*
747 * vm_page_sunbusy:
748 *
749 * Shared unbusy a page.
750 */
751 void
752 vm_page_sunbusy(vm_page_t m)
753 {
754 u_int x;
755
756 vm_page_assert_sbusied(m);
757
758 for (;;) {
759 x = m->busy_lock;
760 if (VPB_SHARERS(x) > 1) {
761 if (atomic_cmpset_int(&m->busy_lock, x,
762 x - VPB_ONE_SHARER))
763 break;
764 continue;
765 }
766 if ((x & VPB_BIT_WAITERS) == 0) {
767 KASSERT(x == VPB_SHARERS_WORD(1),
768 ("vm_page_sunbusy: invalid lock state"));
769 if (atomic_cmpset_int(&m->busy_lock,
770 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
771 break;
772 continue;
773 }
774 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
775 ("vm_page_sunbusy: invalid lock state for waiters"));
776
777 vm_page_lock(m);
778 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
779 vm_page_unlock(m);
780 continue;
781 }
782 wakeup(m);
783 vm_page_unlock(m);
784 break;
785 }
786 }
787
788 /*
789 * vm_page_busy_sleep:
790 *
791 * Sleep and release the page lock, using the page pointer as wchan.
792 * This is used to implement the hard-path of busying mechanism.
793 *
794 * The given page must be locked.
795 *
796 * If nonshared is true, sleep only if the page is xbusy.
797 */
798 void
799 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
800 {
801 u_int x;
802
803 vm_page_assert_locked(m);
804
805 x = m->busy_lock;
806 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
807 ((x & VPB_BIT_WAITERS) == 0 &&
808 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
809 vm_page_unlock(m);
810 return;
811 }
812 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
813 }
814
815 /*
816 * vm_page_trysbusy:
817 *
818 * Try to shared busy a page.
819 * If the operation succeeds 1 is returned otherwise 0.
820 * The operation never sleeps.
821 */
822 int
823 vm_page_trysbusy(vm_page_t m)
824 {
825 u_int x;
826
827 for (;;) {
828 x = m->busy_lock;
829 if ((x & VPB_BIT_SHARED) == 0)
830 return (0);
831 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
832 return (1);
833 }
834 }
835
836 static void
837 vm_page_xunbusy_locked(vm_page_t m)
838 {
839
840 vm_page_assert_xbusied(m);
841 vm_page_assert_locked(m);
842
843 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
844 /* There is a waiter, do wakeup() instead of vm_page_flash(). */
845 wakeup(m);
846 }
847
848 void
849 vm_page_xunbusy_maybelocked(vm_page_t m)
850 {
851 bool lockacq;
852
853 vm_page_assert_xbusied(m);
854
855 /*
856 * Fast path for unbusy. If it succeeds, we know that there
857 * are no waiters, so we do not need a wakeup.
858 */
859 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
860 VPB_UNBUSIED))
861 return;
862
863 lockacq = !mtx_owned(vm_page_lockptr(m));
864 if (lockacq)
865 vm_page_lock(m);
866 vm_page_xunbusy_locked(m);
867 if (lockacq)
868 vm_page_unlock(m);
869 }
870
871 /*
872 * vm_page_xunbusy_hard:
873 *
874 * Called after the first try the exclusive unbusy of a page failed.
875 * It is assumed that the waiters bit is on.
876 */
877 void
878 vm_page_xunbusy_hard(vm_page_t m)
879 {
880
881 vm_page_assert_xbusied(m);
882
883 vm_page_lock(m);
884 vm_page_xunbusy_locked(m);
885 vm_page_unlock(m);
886 }
887
888 /*
889 * vm_page_flash:
890 *
891 * Wakeup anyone waiting for the page.
892 * The ownership bits do not change.
893 *
894 * The given page must be locked.
895 */
896 void
897 vm_page_flash(vm_page_t m)
898 {
899 u_int x;
900
901 vm_page_lock_assert(m, MA_OWNED);
902
903 for (;;) {
904 x = m->busy_lock;
905 if ((x & VPB_BIT_WAITERS) == 0)
906 return;
907 if (atomic_cmpset_int(&m->busy_lock, x,
908 x & (~VPB_BIT_WAITERS)))
909 break;
910 }
911 wakeup(m);
912 }
913
914 /*
915 * Keep page from being freed by the page daemon
916 * much of the same effect as wiring, except much lower
917 * overhead and should be used only for *very* temporary
918 * holding ("wiring").
919 */
920 void
921 vm_page_hold(vm_page_t mem)
922 {
923
924 vm_page_lock_assert(mem, MA_OWNED);
925 mem->hold_count++;
926 }
927
928 void
929 vm_page_unhold(vm_page_t mem)
930 {
931
932 vm_page_lock_assert(mem, MA_OWNED);
933 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
934 --mem->hold_count;
935 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
936 vm_page_free_toq(mem);
937 }
938
939 /*
940 * vm_page_unhold_pages:
941 *
942 * Unhold each of the pages that is referenced by the given array.
943 */
944 void
945 vm_page_unhold_pages(vm_page_t *ma, int count)
946 {
947 struct mtx *mtx, *new_mtx;
948
949 mtx = NULL;
950 for (; count != 0; count--) {
951 /*
952 * Avoid releasing and reacquiring the same page lock.
953 */
954 new_mtx = vm_page_lockptr(*ma);
955 if (mtx != new_mtx) {
956 if (mtx != NULL)
957 mtx_unlock(mtx);
958 mtx = new_mtx;
959 mtx_lock(mtx);
960 }
961 vm_page_unhold(*ma);
962 ma++;
963 }
964 if (mtx != NULL)
965 mtx_unlock(mtx);
966 }
967
968 vm_page_t
969 PHYS_TO_VM_PAGE(vm_paddr_t pa)
970 {
971 vm_page_t m;
972
973 #ifdef VM_PHYSSEG_SPARSE
974 m = vm_phys_paddr_to_vm_page(pa);
975 if (m == NULL)
976 m = vm_phys_fictitious_to_vm_page(pa);
977 return (m);
978 #elif defined(VM_PHYSSEG_DENSE)
979 long pi;
980
981 pi = atop(pa);
982 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
983 m = &vm_page_array[pi - first_page];
984 return (m);
985 }
986 return (vm_phys_fictitious_to_vm_page(pa));
987 #else
988 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
989 #endif
990 }
991
992 /*
993 * vm_page_getfake:
994 *
995 * Create a fictitious page with the specified physical address and
996 * memory attribute. The memory attribute is the only the machine-
997 * dependent aspect of a fictitious page that must be initialized.
998 */
999 vm_page_t
1000 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1001 {
1002 vm_page_t m;
1003
1004 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1005 vm_page_initfake(m, paddr, memattr);
1006 return (m);
1007 }
1008
1009 void
1010 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1011 {
1012
1013 if ((m->flags & PG_FICTITIOUS) != 0) {
1014 /*
1015 * The page's memattr might have changed since the
1016 * previous initialization. Update the pmap to the
1017 * new memattr.
1018 */
1019 goto memattr;
1020 }
1021 m->phys_addr = paddr;
1022 m->queue = PQ_NONE;
1023 /* Fictitious pages don't use "segind". */
1024 m->flags = PG_FICTITIOUS;
1025 /* Fictitious pages don't use "order" or "pool". */
1026 m->oflags = VPO_UNMANAGED;
1027 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1028 m->wire_count = 1;
1029 pmap_page_init(m);
1030 memattr:
1031 pmap_page_set_memattr(m, memattr);
1032 }
1033
1034 /*
1035 * vm_page_putfake:
1036 *
1037 * Release a fictitious page.
1038 */
1039 void
1040 vm_page_putfake(vm_page_t m)
1041 {
1042
1043 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1044 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1045 ("vm_page_putfake: bad page %p", m));
1046 uma_zfree(fakepg_zone, m);
1047 }
1048
1049 /*
1050 * vm_page_updatefake:
1051 *
1052 * Update the given fictitious page to the specified physical address and
1053 * memory attribute.
1054 */
1055 void
1056 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1057 {
1058
1059 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1060 ("vm_page_updatefake: bad page %p", m));
1061 m->phys_addr = paddr;
1062 pmap_page_set_memattr(m, memattr);
1063 }
1064
1065 /*
1066 * vm_page_free:
1067 *
1068 * Free a page.
1069 */
1070 void
1071 vm_page_free(vm_page_t m)
1072 {
1073
1074 m->flags &= ~PG_ZERO;
1075 vm_page_free_toq(m);
1076 }
1077
1078 /*
1079 * vm_page_free_zero:
1080 *
1081 * Free a page to the zerod-pages queue
1082 */
1083 void
1084 vm_page_free_zero(vm_page_t m)
1085 {
1086
1087 m->flags |= PG_ZERO;
1088 vm_page_free_toq(m);
1089 }
1090
1091 /*
1092 * Unbusy and handle the page queueing for a page from a getpages request that
1093 * was optionally read ahead or behind.
1094 */
1095 void
1096 vm_page_readahead_finish(vm_page_t m)
1097 {
1098
1099 /* We shouldn't put invalid pages on queues. */
1100 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1101
1102 /*
1103 * Since the page is not the actually needed one, whether it should
1104 * be activated or deactivated is not obvious. Empirical results
1105 * have shown that deactivating the page is usually the best choice,
1106 * unless the page is wanted by another thread.
1107 */
1108 vm_page_lock(m);
1109 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1110 vm_page_activate(m);
1111 else
1112 vm_page_deactivate(m);
1113 vm_page_unlock(m);
1114 vm_page_xunbusy(m);
1115 }
1116
1117 /*
1118 * vm_page_sleep_if_busy:
1119 *
1120 * Sleep and release the page queues lock if the page is busied.
1121 * Returns TRUE if the thread slept.
1122 *
1123 * The given page must be unlocked and object containing it must
1124 * be locked.
1125 */
1126 int
1127 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1128 {
1129 vm_object_t obj;
1130
1131 vm_page_lock_assert(m, MA_NOTOWNED);
1132 VM_OBJECT_ASSERT_WLOCKED(m->object);
1133
1134 if (vm_page_busied(m)) {
1135 /*
1136 * The page-specific object must be cached because page
1137 * identity can change during the sleep, causing the
1138 * re-lock of a different object.
1139 * It is assumed that a reference to the object is already
1140 * held by the callers.
1141 */
1142 obj = m->object;
1143 vm_page_lock(m);
1144 VM_OBJECT_WUNLOCK(obj);
1145 vm_page_busy_sleep(m, msg, false);
1146 VM_OBJECT_WLOCK(obj);
1147 return (TRUE);
1148 }
1149 return (FALSE);
1150 }
1151
1152 /*
1153 * vm_page_dirty_KBI: [ internal use only ]
1154 *
1155 * Set all bits in the page's dirty field.
1156 *
1157 * The object containing the specified page must be locked if the
1158 * call is made from the machine-independent layer.
1159 *
1160 * See vm_page_clear_dirty_mask().
1161 *
1162 * This function should only be called by vm_page_dirty().
1163 */
1164 void
1165 vm_page_dirty_KBI(vm_page_t m)
1166 {
1167
1168 /* Refer to this operation by its public name. */
1169 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1170 ("vm_page_dirty: page is invalid!"));
1171 m->dirty = VM_PAGE_BITS_ALL;
1172 }
1173
1174 /*
1175 * vm_page_insert: [ internal use only ]
1176 *
1177 * Inserts the given mem entry into the object and object list.
1178 *
1179 * The object must be locked.
1180 */
1181 int
1182 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1183 {
1184 vm_page_t mpred;
1185
1186 VM_OBJECT_ASSERT_WLOCKED(object);
1187 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1188 return (vm_page_insert_after(m, object, pindex, mpred));
1189 }
1190
1191 /*
1192 * vm_page_insert_after:
1193 *
1194 * Inserts the page "m" into the specified object at offset "pindex".
1195 *
1196 * The page "mpred" must immediately precede the offset "pindex" within
1197 * the specified object.
1198 *
1199 * The object must be locked.
1200 */
1201 static int
1202 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1203 vm_page_t mpred)
1204 {
1205 vm_page_t msucc;
1206
1207 VM_OBJECT_ASSERT_WLOCKED(object);
1208 KASSERT(m->object == NULL,
1209 ("vm_page_insert_after: page already inserted"));
1210 if (mpred != NULL) {
1211 KASSERT(mpred->object == object,
1212 ("vm_page_insert_after: object doesn't contain mpred"));
1213 KASSERT(mpred->pindex < pindex,
1214 ("vm_page_insert_after: mpred doesn't precede pindex"));
1215 msucc = TAILQ_NEXT(mpred, listq);
1216 } else
1217 msucc = TAILQ_FIRST(&object->memq);
1218 if (msucc != NULL)
1219 KASSERT(msucc->pindex > pindex,
1220 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1221
1222 /*
1223 * Record the object/offset pair in this page
1224 */
1225 m->object = object;
1226 m->pindex = pindex;
1227
1228 /*
1229 * Now link into the object's ordered list of backed pages.
1230 */
1231 if (vm_radix_insert(&object->rtree, m)) {
1232 m->object = NULL;
1233 m->pindex = 0;
1234 return (1);
1235 }
1236 vm_page_insert_radixdone(m, object, mpred);
1237 return (0);
1238 }
1239
1240 /*
1241 * vm_page_insert_radixdone:
1242 *
1243 * Complete page "m" insertion into the specified object after the
1244 * radix trie hooking.
1245 *
1246 * The page "mpred" must precede the offset "m->pindex" within the
1247 * specified object.
1248 *
1249 * The object must be locked.
1250 */
1251 static void
1252 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1253 {
1254
1255 VM_OBJECT_ASSERT_WLOCKED(object);
1256 KASSERT(object != NULL && m->object == object,
1257 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1258 if (mpred != NULL) {
1259 KASSERT(mpred->object == object,
1260 ("vm_page_insert_after: object doesn't contain mpred"));
1261 KASSERT(mpred->pindex < m->pindex,
1262 ("vm_page_insert_after: mpred doesn't precede pindex"));
1263 }
1264
1265 if (mpred != NULL)
1266 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1267 else
1268 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1269
1270 /*
1271 * Show that the object has one more resident page.
1272 */
1273 object->resident_page_count++;
1274
1275 /*
1276 * Hold the vnode until the last page is released.
1277 */
1278 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1279 vhold(object->handle);
1280
1281 /*
1282 * Since we are inserting a new and possibly dirty page,
1283 * update the object's OBJ_MIGHTBEDIRTY flag.
1284 */
1285 if (pmap_page_is_write_mapped(m))
1286 vm_object_set_writeable_dirty(object);
1287 }
1288
1289 /*
1290 * vm_page_remove:
1291 *
1292 * Removes the specified page from its containing object, but does not
1293 * invalidate any backing storage.
1294 *
1295 * The object must be locked. The page must be locked if it is managed.
1296 */
1297 void
1298 vm_page_remove(vm_page_t m)
1299 {
1300 vm_object_t object;
1301 vm_page_t mrem;
1302
1303 if ((m->oflags & VPO_UNMANAGED) == 0)
1304 vm_page_assert_locked(m);
1305 if ((object = m->object) == NULL)
1306 return;
1307 VM_OBJECT_ASSERT_WLOCKED(object);
1308 if (vm_page_xbusied(m))
1309 vm_page_xunbusy_maybelocked(m);
1310 mrem = vm_radix_remove(&object->rtree, m->pindex);
1311 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1312
1313 /*
1314 * Now remove from the object's list of backed pages.
1315 */
1316 TAILQ_REMOVE(&object->memq, m, listq);
1317
1318 /*
1319 * And show that the object has one fewer resident page.
1320 */
1321 object->resident_page_count--;
1322
1323 /*
1324 * The vnode may now be recycled.
1325 */
1326 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1327 vdrop(object->handle);
1328
1329 m->object = NULL;
1330 }
1331
1332 /*
1333 * vm_page_lookup:
1334 *
1335 * Returns the page associated with the object/offset
1336 * pair specified; if none is found, NULL is returned.
1337 *
1338 * The object must be locked.
1339 */
1340 vm_page_t
1341 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1342 {
1343
1344 VM_OBJECT_ASSERT_LOCKED(object);
1345 return (vm_radix_lookup(&object->rtree, pindex));
1346 }
1347
1348 /*
1349 * vm_page_find_least:
1350 *
1351 * Returns the page associated with the object with least pindex
1352 * greater than or equal to the parameter pindex, or NULL.
1353 *
1354 * The object must be locked.
1355 */
1356 vm_page_t
1357 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1358 {
1359 vm_page_t m;
1360
1361 VM_OBJECT_ASSERT_LOCKED(object);
1362 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1363 m = vm_radix_lookup_ge(&object->rtree, pindex);
1364 return (m);
1365 }
1366
1367 /*
1368 * Returns the given page's successor (by pindex) within the object if it is
1369 * resident; if none is found, NULL is returned.
1370 *
1371 * The object must be locked.
1372 */
1373 vm_page_t
1374 vm_page_next(vm_page_t m)
1375 {
1376 vm_page_t next;
1377
1378 VM_OBJECT_ASSERT_LOCKED(m->object);
1379 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1380 MPASS(next->object == m->object);
1381 if (next->pindex != m->pindex + 1)
1382 next = NULL;
1383 }
1384 return (next);
1385 }
1386
1387 /*
1388 * Returns the given page's predecessor (by pindex) within the object if it is
1389 * resident; if none is found, NULL is returned.
1390 *
1391 * The object must be locked.
1392 */
1393 vm_page_t
1394 vm_page_prev(vm_page_t m)
1395 {
1396 vm_page_t prev;
1397
1398 VM_OBJECT_ASSERT_LOCKED(m->object);
1399 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1400 MPASS(prev->object == m->object);
1401 if (prev->pindex != m->pindex - 1)
1402 prev = NULL;
1403 }
1404 return (prev);
1405 }
1406
1407 /*
1408 * Uses the page mnew as a replacement for an existing page at index
1409 * pindex which must be already present in the object.
1410 *
1411 * The existing page must not be on a paging queue.
1412 */
1413 vm_page_t
1414 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1415 {
1416 vm_page_t mold;
1417
1418 VM_OBJECT_ASSERT_WLOCKED(object);
1419 KASSERT(mnew->object == NULL,
1420 ("vm_page_replace: page already in object"));
1421
1422 /*
1423 * This function mostly follows vm_page_insert() and
1424 * vm_page_remove() without the radix, object count and vnode
1425 * dance. Double check such functions for more comments.
1426 */
1427
1428 mnew->object = object;
1429 mnew->pindex = pindex;
1430 mold = vm_radix_replace(&object->rtree, mnew);
1431 KASSERT(mold->queue == PQ_NONE,
1432 ("vm_page_replace: mold is on a paging queue"));
1433
1434 /* Keep the resident page list in sorted order. */
1435 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1436 TAILQ_REMOVE(&object->memq, mold, listq);
1437
1438 mold->object = NULL;
1439 vm_page_xunbusy_maybelocked(mold);
1440
1441 /*
1442 * The object's resident_page_count does not change because we have
1443 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1444 */
1445 if (pmap_page_is_write_mapped(mnew))
1446 vm_object_set_writeable_dirty(object);
1447 return (mold);
1448 }
1449
1450 /*
1451 * vm_page_rename:
1452 *
1453 * Move the given memory entry from its
1454 * current object to the specified target object/offset.
1455 *
1456 * Note: swap associated with the page must be invalidated by the move. We
1457 * have to do this for several reasons: (1) we aren't freeing the
1458 * page, (2) we are dirtying the page, (3) the VM system is probably
1459 * moving the page from object A to B, and will then later move
1460 * the backing store from A to B and we can't have a conflict.
1461 *
1462 * Note: we *always* dirty the page. It is necessary both for the
1463 * fact that we moved it, and because we may be invalidating
1464 * swap.
1465 *
1466 * The objects must be locked.
1467 */
1468 int
1469 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1470 {
1471 vm_page_t mpred;
1472 vm_pindex_t opidx;
1473
1474 VM_OBJECT_ASSERT_WLOCKED(new_object);
1475
1476 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1477 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1478 ("vm_page_rename: pindex already renamed"));
1479
1480 /*
1481 * Create a custom version of vm_page_insert() which does not depend
1482 * by m_prev and can cheat on the implementation aspects of the
1483 * function.
1484 */
1485 opidx = m->pindex;
1486 m->pindex = new_pindex;
1487 if (vm_radix_insert(&new_object->rtree, m)) {
1488 m->pindex = opidx;
1489 return (1);
1490 }
1491
1492 /*
1493 * The operation cannot fail anymore. The removal must happen before
1494 * the listq iterator is tainted.
1495 */
1496 m->pindex = opidx;
1497 vm_page_lock(m);
1498 vm_page_remove(m);
1499
1500 /* Return back to the new pindex to complete vm_page_insert(). */
1501 m->pindex = new_pindex;
1502 m->object = new_object;
1503 vm_page_unlock(m);
1504 vm_page_insert_radixdone(m, new_object, mpred);
1505 vm_page_dirty(m);
1506 return (0);
1507 }
1508
1509 /*
1510 * vm_page_alloc:
1511 *
1512 * Allocate and return a page that is associated with the specified
1513 * object and offset pair. By default, this page is exclusive busied.
1514 *
1515 * The caller must always specify an allocation class.
1516 *
1517 * allocation classes:
1518 * VM_ALLOC_NORMAL normal process request
1519 * VM_ALLOC_SYSTEM system *really* needs a page
1520 * VM_ALLOC_INTERRUPT interrupt time request
1521 *
1522 * optional allocation flags:
1523 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1524 * intends to allocate
1525 * VM_ALLOC_NOBUSY do not exclusive busy the page
1526 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1527 * VM_ALLOC_NOOBJ page is not associated with an object and
1528 * should not be exclusive busy
1529 * VM_ALLOC_SBUSY shared busy the allocated page
1530 * VM_ALLOC_WIRED wire the allocated page
1531 * VM_ALLOC_ZERO prefer a zeroed page
1532 *
1533 * This routine may not sleep.
1534 */
1535 vm_page_t
1536 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1537 {
1538 vm_page_t m, mpred;
1539 int flags, req_class;
1540
1541 mpred = NULL; /* XXX: pacify gcc */
1542 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1543 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1544 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1545 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1546 ("vm_page_alloc: inconsistent object(%p)/req(%x)", object, req));
1547 if (object != NULL)
1548 VM_OBJECT_ASSERT_WLOCKED(object);
1549
1550 if (__predict_false((req & VM_ALLOC_IFCACHED) != 0))
1551 return (NULL);
1552
1553 req_class = req & VM_ALLOC_CLASS_MASK;
1554
1555 /*
1556 * The page daemon is allowed to dig deeper into the free page list.
1557 */
1558 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1559 req_class = VM_ALLOC_SYSTEM;
1560
1561 if (object != NULL) {
1562 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1563 KASSERT(mpred == NULL || mpred->pindex != pindex,
1564 ("vm_page_alloc: pindex already allocated"));
1565 }
1566
1567 /*
1568 * Allocate a page if the number of free pages exceeds the minimum
1569 * for the request class.
1570 */
1571 mtx_lock(&vm_page_queue_free_mtx);
1572 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1573 (req_class == VM_ALLOC_SYSTEM &&
1574 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1575 (req_class == VM_ALLOC_INTERRUPT &&
1576 vm_cnt.v_free_count > 0)) {
1577 /*
1578 * Can we allocate the page from a reservation?
1579 */
1580 #if VM_NRESERVLEVEL > 0
1581 if (object == NULL || (object->flags & (OBJ_COLORED |
1582 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1583 vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1584 #endif
1585 {
1586 /*
1587 * If not, allocate it from the free page queues.
1588 */
1589 m = vm_phys_alloc_pages(object != NULL ?
1590 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1591 #if VM_NRESERVLEVEL > 0
1592 if (m == NULL && vm_reserv_reclaim_inactive()) {
1593 m = vm_phys_alloc_pages(object != NULL ?
1594 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1595 0);
1596 }
1597 #endif
1598 }
1599 } else {
1600 /*
1601 * Not allocatable, give up.
1602 */
1603 mtx_unlock(&vm_page_queue_free_mtx);
1604 atomic_add_int(&vm_pageout_deficit,
1605 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1606 pagedaemon_wakeup();
1607 return (NULL);
1608 }
1609
1610 /*
1611 * At this point we had better have found a good page.
1612 */
1613 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1614 vm_phys_freecnt_adj(m, -1);
1615 if ((m->flags & PG_ZERO) != 0)
1616 vm_page_zero_count--;
1617 mtx_unlock(&vm_page_queue_free_mtx);
1618 vm_page_alloc_check(m);
1619
1620 /*
1621 * Initialize the page. Only the PG_ZERO flag is inherited.
1622 */
1623 flags = 0;
1624 if ((req & VM_ALLOC_ZERO) != 0)
1625 flags = PG_ZERO;
1626 flags &= m->flags;
1627 if ((req & VM_ALLOC_NODUMP) != 0)
1628 flags |= PG_NODUMP;
1629 m->flags = flags;
1630 m->aflags = 0;
1631 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1632 VPO_UNMANAGED : 0;
1633 m->busy_lock = VPB_UNBUSIED;
1634 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1635 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1636 if ((req & VM_ALLOC_SBUSY) != 0)
1637 m->busy_lock = VPB_SHARERS_WORD(1);
1638 if (req & VM_ALLOC_WIRED) {
1639 /*
1640 * The page lock is not required for wiring a page until that
1641 * page is inserted into the object.
1642 */
1643 atomic_add_int(&vm_cnt.v_wire_count, 1);
1644 m->wire_count = 1;
1645 }
1646 m->act_count = 0;
1647
1648 if (object != NULL) {
1649 if (vm_page_insert_after(m, object, pindex, mpred)) {
1650 pagedaemon_wakeup();
1651 if (req & VM_ALLOC_WIRED) {
1652 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1653 m->wire_count = 0;
1654 }
1655 KASSERT(m->object == NULL, ("page %p has object", m));
1656 m->oflags = VPO_UNMANAGED;
1657 m->busy_lock = VPB_UNBUSIED;
1658 /* Don't change PG_ZERO. */
1659 vm_page_free_toq(m);
1660 return (NULL);
1661 }
1662
1663 /* Ignore device objects; the pager sets "memattr" for them. */
1664 if (object->memattr != VM_MEMATTR_DEFAULT &&
1665 (object->flags & OBJ_FICTITIOUS) == 0)
1666 pmap_page_set_memattr(m, object->memattr);
1667 } else
1668 m->pindex = pindex;
1669
1670 /*
1671 * Don't wakeup too often - wakeup the pageout daemon when
1672 * we would be nearly out of memory.
1673 */
1674 if (vm_paging_needed())
1675 pagedaemon_wakeup();
1676
1677 return (m);
1678 }
1679
1680 /*
1681 * vm_page_alloc_contig:
1682 *
1683 * Allocate a contiguous set of physical pages of the given size "npages"
1684 * from the free lists. All of the physical pages must be at or above
1685 * the given physical address "low" and below the given physical address
1686 * "high". The given value "alignment" determines the alignment of the
1687 * first physical page in the set. If the given value "boundary" is
1688 * non-zero, then the set of physical pages cannot cross any physical
1689 * address boundary that is a multiple of that value. Both "alignment"
1690 * and "boundary" must be a power of two.
1691 *
1692 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1693 * then the memory attribute setting for the physical pages is configured
1694 * to the object's memory attribute setting. Otherwise, the memory
1695 * attribute setting for the physical pages is configured to "memattr",
1696 * overriding the object's memory attribute setting. However, if the
1697 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1698 * memory attribute setting for the physical pages cannot be configured
1699 * to VM_MEMATTR_DEFAULT.
1700 *
1701 * The specified object may not contain fictitious pages.
1702 *
1703 * The caller must always specify an allocation class.
1704 *
1705 * allocation classes:
1706 * VM_ALLOC_NORMAL normal process request
1707 * VM_ALLOC_SYSTEM system *really* needs a page
1708 * VM_ALLOC_INTERRUPT interrupt time request
1709 *
1710 * optional allocation flags:
1711 * VM_ALLOC_NOBUSY do not exclusive busy the page
1712 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1713 * VM_ALLOC_NOOBJ page is not associated with an object and
1714 * should not be exclusive busy
1715 * VM_ALLOC_SBUSY shared busy the allocated page
1716 * VM_ALLOC_WIRED wire the allocated page
1717 * VM_ALLOC_ZERO prefer a zeroed page
1718 *
1719 * This routine may not sleep.
1720 */
1721 vm_page_t
1722 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1723 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1724 vm_paddr_t boundary, vm_memattr_t memattr)
1725 {
1726 vm_page_t m, m_ret, mpred;
1727 u_int busy_lock, flags, oflags;
1728 int req_class;
1729
1730 mpred = NULL; /* XXX: pacify gcc */
1731 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1732 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1733 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1734 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1735 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1736 req));
1737 if (object != NULL) {
1738 VM_OBJECT_ASSERT_WLOCKED(object);
1739 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1740 ("vm_page_alloc_contig: object %p has fictitious pages",
1741 object));
1742 }
1743 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1744 req_class = req & VM_ALLOC_CLASS_MASK;
1745
1746 /*
1747 * The page daemon is allowed to dig deeper into the free page list.
1748 */
1749 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1750 req_class = VM_ALLOC_SYSTEM;
1751
1752 if (object != NULL) {
1753 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1754 KASSERT(mpred == NULL || mpred->pindex != pindex,
1755 ("vm_page_alloc_contig: pindex already allocated"));
1756 }
1757
1758 /*
1759 * Can we allocate the pages without the number of free pages falling
1760 * below the lower bound for the allocation class?
1761 */
1762 mtx_lock(&vm_page_queue_free_mtx);
1763 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1764 (req_class == VM_ALLOC_SYSTEM &&
1765 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1766 (req_class == VM_ALLOC_INTERRUPT &&
1767 vm_cnt.v_free_count >= npages)) {
1768 /*
1769 * Can we allocate the pages from a reservation?
1770 */
1771 #if VM_NRESERVLEVEL > 0
1772 retry:
1773 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1774 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1775 low, high, alignment, boundary, mpred)) == NULL)
1776 #endif
1777 /*
1778 * If not, allocate them from the free page queues.
1779 */
1780 m_ret = vm_phys_alloc_contig(npages, low, high,
1781 alignment, boundary);
1782 } else {
1783 mtx_unlock(&vm_page_queue_free_mtx);
1784 atomic_add_int(&vm_pageout_deficit, npages);
1785 pagedaemon_wakeup();
1786 return (NULL);
1787 }
1788 if (m_ret != NULL) {
1789 vm_phys_freecnt_adj(m_ret, -npages);
1790 for (m = m_ret; m < &m_ret[npages]; m++)
1791 if ((m->flags & PG_ZERO) != 0)
1792 vm_page_zero_count--;
1793 } else {
1794 #if VM_NRESERVLEVEL > 0
1795 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1796 boundary))
1797 goto retry;
1798 #endif
1799 }
1800 mtx_unlock(&vm_page_queue_free_mtx);
1801 if (m_ret == NULL)
1802 return (NULL);
1803 for (m = m_ret; m < &m_ret[npages]; m++)
1804 vm_page_alloc_check(m);
1805
1806 /*
1807 * Initialize the pages. Only the PG_ZERO flag is inherited.
1808 */
1809 flags = 0;
1810 if ((req & VM_ALLOC_ZERO) != 0)
1811 flags = PG_ZERO;
1812 if ((req & VM_ALLOC_NODUMP) != 0)
1813 flags |= PG_NODUMP;
1814 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1815 VPO_UNMANAGED : 0;
1816 busy_lock = VPB_UNBUSIED;
1817 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1818 busy_lock = VPB_SINGLE_EXCLUSIVER;
1819 if ((req & VM_ALLOC_SBUSY) != 0)
1820 busy_lock = VPB_SHARERS_WORD(1);
1821 if ((req & VM_ALLOC_WIRED) != 0)
1822 atomic_add_int(&vm_cnt.v_wire_count, npages);
1823 if (object != NULL) {
1824 if (object->memattr != VM_MEMATTR_DEFAULT &&
1825 memattr == VM_MEMATTR_DEFAULT)
1826 memattr = object->memattr;
1827 }
1828 for (m = m_ret; m < &m_ret[npages]; m++) {
1829 m->aflags = 0;
1830 m->flags = (m->flags | PG_NODUMP) & flags;
1831 m->busy_lock = busy_lock;
1832 if ((req & VM_ALLOC_WIRED) != 0)
1833 m->wire_count = 1;
1834 m->act_count = 0;
1835 m->oflags = oflags;
1836 if (object != NULL) {
1837 if (vm_page_insert_after(m, object, pindex, mpred)) {
1838 pagedaemon_wakeup();
1839 if ((req & VM_ALLOC_WIRED) != 0)
1840 atomic_subtract_int(
1841 &vm_cnt.v_wire_count, npages);
1842 KASSERT(m->object == NULL,
1843 ("page %p has object", m));
1844 mpred = m;
1845 for (m = m_ret; m < &m_ret[npages]; m++) {
1846 if (m <= mpred &&
1847 (req & VM_ALLOC_WIRED) != 0)
1848 m->wire_count = 0;
1849 m->oflags = VPO_UNMANAGED;
1850 m->busy_lock = VPB_UNBUSIED;
1851 /* Don't change PG_ZERO. */
1852 vm_page_free_toq(m);
1853 }
1854 return (NULL);
1855 }
1856 mpred = m;
1857 } else
1858 m->pindex = pindex;
1859 if (memattr != VM_MEMATTR_DEFAULT)
1860 pmap_page_set_memattr(m, memattr);
1861 pindex++;
1862 }
1863 if (vm_paging_needed())
1864 pagedaemon_wakeup();
1865 return (m_ret);
1866 }
1867
1868 /*
1869 * Check a page that has been freshly dequeued from a freelist.
1870 */
1871 static void
1872 vm_page_alloc_check(vm_page_t m)
1873 {
1874
1875 KASSERT(m->object == NULL, ("page %p has object", m));
1876 KASSERT(m->queue == PQ_NONE,
1877 ("page %p has unexpected queue %d", m, m->queue));
1878 KASSERT(m->wire_count == 0, ("page %p is wired", m));
1879 KASSERT(m->hold_count == 0, ("page %p is held", m));
1880 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1881 KASSERT(m->dirty == 0, ("page %p is dirty", m));
1882 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1883 ("page %p has unexpected memattr %d",
1884 m, pmap_page_get_memattr(m)));
1885 KASSERT(m->valid == 0, ("free page %p is valid", m));
1886 }
1887
1888 /*
1889 * vm_page_alloc_freelist:
1890 *
1891 * Allocate a physical page from the specified free page list.
1892 *
1893 * The caller must always specify an allocation class.
1894 *
1895 * allocation classes:
1896 * VM_ALLOC_NORMAL normal process request
1897 * VM_ALLOC_SYSTEM system *really* needs a page
1898 * VM_ALLOC_INTERRUPT interrupt time request
1899 *
1900 * optional allocation flags:
1901 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1902 * intends to allocate
1903 * VM_ALLOC_WIRED wire the allocated page
1904 * VM_ALLOC_ZERO prefer a zeroed page
1905 *
1906 * This routine may not sleep.
1907 */
1908 vm_page_t
1909 vm_page_alloc_freelist(int flind, int req)
1910 {
1911 vm_page_t m;
1912 u_int flags;
1913 int req_class;
1914
1915 req_class = req & VM_ALLOC_CLASS_MASK;
1916
1917 /*
1918 * The page daemon is allowed to dig deeper into the free page list.
1919 */
1920 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1921 req_class = VM_ALLOC_SYSTEM;
1922
1923 /*
1924 * Do not allocate reserved pages unless the req has asked for it.
1925 */
1926 mtx_lock(&vm_page_queue_free_mtx);
1927 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1928 (req_class == VM_ALLOC_SYSTEM &&
1929 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1930 (req_class == VM_ALLOC_INTERRUPT &&
1931 vm_cnt.v_free_count > 0))
1932 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1933 else {
1934 mtx_unlock(&vm_page_queue_free_mtx);
1935 atomic_add_int(&vm_pageout_deficit,
1936 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1937 pagedaemon_wakeup();
1938 return (NULL);
1939 }
1940 if (m == NULL) {
1941 mtx_unlock(&vm_page_queue_free_mtx);
1942 return (NULL);
1943 }
1944 vm_phys_freecnt_adj(m, -1);
1945 if ((m->flags & PG_ZERO) != 0)
1946 vm_page_zero_count--;
1947 mtx_unlock(&vm_page_queue_free_mtx);
1948 vm_page_alloc_check(m);
1949
1950 /*
1951 * Initialize the page. Only the PG_ZERO flag is inherited.
1952 */
1953 m->aflags = 0;
1954 flags = 0;
1955 if ((req & VM_ALLOC_ZERO) != 0)
1956 flags = PG_ZERO;
1957 m->flags &= flags;
1958 if ((req & VM_ALLOC_WIRED) != 0) {
1959 /*
1960 * The page lock is not required for wiring a page that does
1961 * not belong to an object.
1962 */
1963 atomic_add_int(&vm_cnt.v_wire_count, 1);
1964 m->wire_count = 1;
1965 }
1966 /* Unmanaged pages don't use "act_count". */
1967 m->oflags = VPO_UNMANAGED;
1968 if (vm_paging_needed())
1969 pagedaemon_wakeup();
1970 return (m);
1971 }
1972
1973 #define VPSC_ANY 0 /* No restrictions. */
1974 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
1975 #define VPSC_NOSUPER 2 /* Skip superpages. */
1976
1977 /*
1978 * vm_page_scan_contig:
1979 *
1980 * Scan vm_page_array[] between the specified entries "m_start" and
1981 * "m_end" for a run of contiguous physical pages that satisfy the
1982 * specified conditions, and return the lowest page in the run. The
1983 * specified "alignment" determines the alignment of the lowest physical
1984 * page in the run. If the specified "boundary" is non-zero, then the
1985 * run of physical pages cannot span a physical address that is a
1986 * multiple of "boundary".
1987 *
1988 * "m_end" is never dereferenced, so it need not point to a vm_page
1989 * structure within vm_page_array[].
1990 *
1991 * "npages" must be greater than zero. "m_start" and "m_end" must not
1992 * span a hole (or discontiguity) in the physical address space. Both
1993 * "alignment" and "boundary" must be a power of two.
1994 */
1995 vm_page_t
1996 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
1997 u_long alignment, vm_paddr_t boundary, int options)
1998 {
1999 struct mtx *m_mtx, *new_mtx;
2000 vm_object_t object;
2001 vm_paddr_t pa;
2002 vm_page_t m, m_run;
2003 #if VM_NRESERVLEVEL > 0
2004 int level;
2005 #endif
2006 int m_inc, order, run_ext, run_len;
2007
2008 KASSERT(npages > 0, ("npages is 0"));
2009 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2010 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2011 m_run = NULL;
2012 run_len = 0;
2013 m_mtx = NULL;
2014 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2015 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2016 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2017
2018 /*
2019 * If the current page would be the start of a run, check its
2020 * physical address against the end, alignment, and boundary
2021 * conditions. If it doesn't satisfy these conditions, either
2022 * terminate the scan or advance to the next page that
2023 * satisfies the failed condition.
2024 */
2025 if (run_len == 0) {
2026 KASSERT(m_run == NULL, ("m_run != NULL"));
2027 if (m + npages > m_end)
2028 break;
2029 pa = VM_PAGE_TO_PHYS(m);
2030 if ((pa & (alignment - 1)) != 0) {
2031 m_inc = atop(roundup2(pa, alignment) - pa);
2032 continue;
2033 }
2034 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2035 boundary) != 0) {
2036 m_inc = atop(roundup2(pa, boundary) - pa);
2037 continue;
2038 }
2039 } else
2040 KASSERT(m_run != NULL, ("m_run == NULL"));
2041
2042 /*
2043 * Avoid releasing and reacquiring the same page lock.
2044 */
2045 new_mtx = vm_page_lockptr(m);
2046 if (m_mtx != new_mtx) {
2047 if (m_mtx != NULL)
2048 mtx_unlock(m_mtx);
2049 m_mtx = new_mtx;
2050 mtx_lock(m_mtx);
2051 }
2052 m_inc = 1;
2053 retry:
2054 if (m->wire_count != 0 || m->hold_count != 0)
2055 run_ext = 0;
2056 #if VM_NRESERVLEVEL > 0
2057 else if ((level = vm_reserv_level(m)) >= 0 &&
2058 (options & VPSC_NORESERV) != 0) {
2059 run_ext = 0;
2060 /* Advance to the end of the reservation. */
2061 pa = VM_PAGE_TO_PHYS(m);
2062 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2063 pa);
2064 }
2065 #endif
2066 else if ((object = m->object) != NULL) {
2067 /*
2068 * The page is considered eligible for relocation if
2069 * and only if it could be laundered or reclaimed by
2070 * the page daemon.
2071 */
2072 if (!VM_OBJECT_TRYRLOCK(object)) {
2073 mtx_unlock(m_mtx);
2074 VM_OBJECT_RLOCK(object);
2075 mtx_lock(m_mtx);
2076 if (m->object != object) {
2077 /*
2078 * The page may have been freed.
2079 */
2080 VM_OBJECT_RUNLOCK(object);
2081 goto retry;
2082 } else if (m->wire_count != 0 ||
2083 m->hold_count != 0) {
2084 run_ext = 0;
2085 goto unlock;
2086 }
2087 }
2088 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2089 ("page %p is PG_UNHOLDFREE", m));
2090 /* Don't care: PG_NODUMP, PG_ZERO. */
2091 if (object->type != OBJT_DEFAULT &&
2092 object->type != OBJT_SWAP &&
2093 object->type != OBJT_VNODE) {
2094 run_ext = 0;
2095 #if VM_NRESERVLEVEL > 0
2096 } else if ((options & VPSC_NOSUPER) != 0 &&
2097 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2098 run_ext = 0;
2099 /* Advance to the end of the superpage. */
2100 pa = VM_PAGE_TO_PHYS(m);
2101 m_inc = atop(roundup2(pa + 1,
2102 vm_reserv_size(level)) - pa);
2103 #endif
2104 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2105 m->queue != PQ_NONE && !vm_page_busied(m)) {
2106 /*
2107 * The page is allocated but eligible for
2108 * relocation. Extend the current run by one
2109 * page.
2110 */
2111 KASSERT(pmap_page_get_memattr(m) ==
2112 VM_MEMATTR_DEFAULT,
2113 ("page %p has an unexpected memattr", m));
2114 KASSERT((m->oflags & (VPO_SWAPINPROG |
2115 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2116 ("page %p has unexpected oflags", m));
2117 /* Don't care: VPO_NOSYNC. */
2118 run_ext = 1;
2119 } else
2120 run_ext = 0;
2121 unlock:
2122 VM_OBJECT_RUNLOCK(object);
2123 #if VM_NRESERVLEVEL > 0
2124 } else if (level >= 0) {
2125 /*
2126 * The page is reserved but not yet allocated. In
2127 * other words, it is still free. Extend the current
2128 * run by one page.
2129 */
2130 run_ext = 1;
2131 #endif
2132 } else if ((order = m->order) < VM_NFREEORDER) {
2133 /*
2134 * The page is enqueued in the physical memory
2135 * allocator's free page queues. Moreover, it is the
2136 * first page in a power-of-two-sized run of
2137 * contiguous free pages. Add these pages to the end
2138 * of the current run, and jump ahead.
2139 */
2140 run_ext = 1 << order;
2141 m_inc = 1 << order;
2142 } else {
2143 /*
2144 * Skip the page for one of the following reasons: (1)
2145 * It is enqueued in the physical memory allocator's
2146 * free page queues. However, it is not the first
2147 * page in a run of contiguous free pages. (This case
2148 * rarely occurs because the scan is performed in
2149 * ascending order.) (2) It is not reserved, and it is
2150 * transitioning from free to allocated. (Conversely,
2151 * the transition from allocated to free for managed
2152 * pages is blocked by the page lock.) (3) It is
2153 * allocated but not contained by an object and not
2154 * wired, e.g., allocated by Xen's balloon driver.
2155 */
2156 run_ext = 0;
2157 }
2158
2159 /*
2160 * Extend or reset the current run of pages.
2161 */
2162 if (run_ext > 0) {
2163 if (run_len == 0)
2164 m_run = m;
2165 run_len += run_ext;
2166 } else {
2167 if (run_len > 0) {
2168 m_run = NULL;
2169 run_len = 0;
2170 }
2171 }
2172 }
2173 if (m_mtx != NULL)
2174 mtx_unlock(m_mtx);
2175 if (run_len >= npages)
2176 return (m_run);
2177 return (NULL);
2178 }
2179
2180 /*
2181 * vm_page_reclaim_run:
2182 *
2183 * Try to relocate each of the allocated virtual pages within the
2184 * specified run of physical pages to a new physical address. Free the
2185 * physical pages underlying the relocated virtual pages. A virtual page
2186 * is relocatable if and only if it could be laundered or reclaimed by
2187 * the page daemon. Whenever possible, a virtual page is relocated to a
2188 * physical address above "high".
2189 *
2190 * Returns 0 if every physical page within the run was already free or
2191 * just freed by a successful relocation. Otherwise, returns a non-zero
2192 * value indicating why the last attempt to relocate a virtual page was
2193 * unsuccessful.
2194 *
2195 * "req_class" must be an allocation class.
2196 */
2197 static int
2198 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2199 vm_paddr_t high)
2200 {
2201 struct mtx *m_mtx, *new_mtx;
2202 struct spglist free;
2203 vm_object_t object;
2204 vm_paddr_t pa;
2205 vm_page_t m, m_end, m_new;
2206 int error, order, req;
2207
2208 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2209 ("req_class is not an allocation class"));
2210 SLIST_INIT(&free);
2211 error = 0;
2212 m = m_run;
2213 m_end = m_run + npages;
2214 m_mtx = NULL;
2215 for (; error == 0 && m < m_end; m++) {
2216 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2217 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2218
2219 /*
2220 * Avoid releasing and reacquiring the same page lock.
2221 */
2222 new_mtx = vm_page_lockptr(m);
2223 if (m_mtx != new_mtx) {
2224 if (m_mtx != NULL)
2225 mtx_unlock(m_mtx);
2226 m_mtx = new_mtx;
2227 mtx_lock(m_mtx);
2228 }
2229 retry:
2230 if (m->wire_count != 0 || m->hold_count != 0)
2231 error = EBUSY;
2232 else if ((object = m->object) != NULL) {
2233 /*
2234 * The page is relocated if and only if it could be
2235 * laundered or reclaimed by the page daemon.
2236 */
2237 if (!VM_OBJECT_TRYWLOCK(object)) {
2238 mtx_unlock(m_mtx);
2239 VM_OBJECT_WLOCK(object);
2240 mtx_lock(m_mtx);
2241 if (m->object != object) {
2242 /*
2243 * The page may have been freed.
2244 */
2245 VM_OBJECT_WUNLOCK(object);
2246 goto retry;
2247 } else if (m->wire_count != 0 ||
2248 m->hold_count != 0) {
2249 error = EBUSY;
2250 goto unlock;
2251 }
2252 }
2253 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2254 ("page %p is PG_UNHOLDFREE", m));
2255 /* Don't care: PG_NODUMP, PG_ZERO. */
2256 if (object->type != OBJT_DEFAULT &&
2257 object->type != OBJT_SWAP &&
2258 object->type != OBJT_VNODE)
2259 error = EINVAL;
2260 else if (object->memattr != VM_MEMATTR_DEFAULT)
2261 error = EINVAL;
2262 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2263 KASSERT(pmap_page_get_memattr(m) ==
2264 VM_MEMATTR_DEFAULT,
2265 ("page %p has an unexpected memattr", m));
2266 KASSERT((m->oflags & (VPO_SWAPINPROG |
2267 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2268 ("page %p has unexpected oflags", m));
2269 /* Don't care: VPO_NOSYNC. */
2270 if (m->valid != 0) {
2271 /*
2272 * First, try to allocate a new page
2273 * that is above "high". Failing
2274 * that, try to allocate a new page
2275 * that is below "m_run". Allocate
2276 * the new page between the end of
2277 * "m_run" and "high" only as a last
2278 * resort.
2279 */
2280 req = req_class | VM_ALLOC_NOOBJ;
2281 if ((m->flags & PG_NODUMP) != 0)
2282 req |= VM_ALLOC_NODUMP;
2283 if (trunc_page(high) !=
2284 ~(vm_paddr_t)PAGE_MASK) {
2285 m_new = vm_page_alloc_contig(
2286 NULL, 0, req, 1,
2287 round_page(high),
2288 ~(vm_paddr_t)0,
2289 PAGE_SIZE, 0,
2290 VM_MEMATTR_DEFAULT);
2291 } else
2292 m_new = NULL;
2293 if (m_new == NULL) {
2294 pa = VM_PAGE_TO_PHYS(m_run);
2295 m_new = vm_page_alloc_contig(
2296 NULL, 0, req, 1,
2297 0, pa - 1, PAGE_SIZE, 0,
2298 VM_MEMATTR_DEFAULT);
2299 }
2300 if (m_new == NULL) {
2301 pa += ptoa(npages);
2302 m_new = vm_page_alloc_contig(
2303 NULL, 0, req, 1,
2304 pa, high, PAGE_SIZE, 0,
2305 VM_MEMATTR_DEFAULT);
2306 }
2307 if (m_new == NULL) {
2308 error = ENOMEM;
2309 goto unlock;
2310 }
2311 KASSERT(m_new->wire_count == 0,
2312 ("page %p is wired", m));
2313
2314 /*
2315 * Replace "m" with the new page. For
2316 * vm_page_replace(), "m" must be busy
2317 * and dequeued. Finally, change "m"
2318 * as if vm_page_free() was called.
2319 */
2320 if (object->ref_count != 0)
2321 pmap_remove_all(m);
2322 m_new->aflags = m->aflags;
2323 KASSERT(m_new->oflags == VPO_UNMANAGED,
2324 ("page %p is managed", m));
2325 m_new->oflags = m->oflags & VPO_NOSYNC;
2326 pmap_copy_page(m, m_new);
2327 m_new->valid = m->valid;
2328 m_new->dirty = m->dirty;
2329 m->flags &= ~PG_ZERO;
2330 vm_page_xbusy(m);
2331 vm_page_remque(m);
2332 vm_page_replace_checked(m_new, object,
2333 m->pindex, m);
2334 m->valid = 0;
2335 vm_page_undirty(m);
2336
2337 /*
2338 * The new page must be deactivated
2339 * before the object is unlocked.
2340 */
2341 new_mtx = vm_page_lockptr(m_new);
2342 if (m_mtx != new_mtx) {
2343 mtx_unlock(m_mtx);
2344 m_mtx = new_mtx;
2345 mtx_lock(m_mtx);
2346 }
2347 vm_page_deactivate(m_new);
2348 } else {
2349 m->flags &= ~PG_ZERO;
2350 vm_page_remque(m);
2351 vm_page_remove(m);
2352 KASSERT(m->dirty == 0,
2353 ("page %p is dirty", m));
2354 }
2355 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2356 } else
2357 error = EBUSY;
2358 unlock:
2359 VM_OBJECT_WUNLOCK(object);
2360 } else {
2361 mtx_lock(&vm_page_queue_free_mtx);
2362 order = m->order;
2363 if (order < VM_NFREEORDER) {
2364 /*
2365 * The page is enqueued in the physical memory
2366 * allocator's free page queues. Moreover, it
2367 * is the first page in a power-of-two-sized
2368 * run of contiguous free pages. Jump ahead
2369 * to the last page within that run, and
2370 * continue from there.
2371 */
2372 m += (1 << order) - 1;
2373 }
2374 #if VM_NRESERVLEVEL > 0
2375 else if (vm_reserv_is_page_free(m))
2376 order = 0;
2377 #endif
2378 mtx_unlock(&vm_page_queue_free_mtx);
2379 if (order == VM_NFREEORDER)
2380 error = EINVAL;
2381 }
2382 }
2383 if (m_mtx != NULL)
2384 mtx_unlock(m_mtx);
2385 if ((m = SLIST_FIRST(&free)) != NULL) {
2386 mtx_lock(&vm_page_queue_free_mtx);
2387 do {
2388 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2389 vm_phys_freecnt_adj(m, 1);
2390 #if VM_NRESERVLEVEL > 0
2391 if (!vm_reserv_free_page(m))
2392 #else
2393 if (true)
2394 #endif
2395 vm_phys_free_pages(m, 0);
2396 } while ((m = SLIST_FIRST(&free)) != NULL);
2397 vm_page_zero_idle_wakeup();
2398 vm_page_free_wakeup();
2399 mtx_unlock(&vm_page_queue_free_mtx);
2400 }
2401 return (error);
2402 }
2403
2404 #define NRUNS 16
2405
2406 CTASSERT(powerof2(NRUNS));
2407
2408 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2409
2410 #define MIN_RECLAIM 8
2411
2412 /*
2413 * vm_page_reclaim_contig:
2414 *
2415 * Reclaim allocated, contiguous physical memory satisfying the specified
2416 * conditions by relocating the virtual pages using that physical memory.
2417 * Returns true if reclamation is successful and false otherwise. Since
2418 * relocation requires the allocation of physical pages, reclamation may
2419 * fail due to a shortage of free pages. When reclamation fails, callers
2420 * are expected to perform VM_WAIT before retrying a failed allocation
2421 * operation, e.g., vm_page_alloc_contig().
2422 *
2423 * The caller must always specify an allocation class through "req".
2424 *
2425 * allocation classes:
2426 * VM_ALLOC_NORMAL normal process request
2427 * VM_ALLOC_SYSTEM system *really* needs a page
2428 * VM_ALLOC_INTERRUPT interrupt time request
2429 *
2430 * The optional allocation flags are ignored.
2431 *
2432 * "npages" must be greater than zero. Both "alignment" and "boundary"
2433 * must be a power of two.
2434 */
2435 bool
2436 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2437 u_long alignment, vm_paddr_t boundary)
2438 {
2439 vm_paddr_t curr_low;
2440 vm_page_t m_run, m_runs[NRUNS];
2441 u_long count, reclaimed;
2442 int error, i, options, req_class;
2443
2444 KASSERT(npages > 0, ("npages is 0"));
2445 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2446 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2447 req_class = req & VM_ALLOC_CLASS_MASK;
2448
2449 /*
2450 * The page daemon is allowed to dig deeper into the free page list.
2451 */
2452 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2453 req_class = VM_ALLOC_SYSTEM;
2454
2455 /*
2456 * Return if the number of free pages cannot satisfy the requested
2457 * allocation.
2458 */
2459 count = vm_cnt.v_free_count;
2460 if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2461 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2462 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2463 return (false);
2464
2465 /*
2466 * Scan up to three times, relaxing the restrictions ("options") on
2467 * the reclamation of reservations and superpages each time.
2468 */
2469 for (options = VPSC_NORESERV;;) {
2470 /*
2471 * Find the highest runs that satisfy the given constraints
2472 * and restrictions, and record them in "m_runs".
2473 */
2474 curr_low = low;
2475 count = 0;
2476 for (;;) {
2477 m_run = vm_phys_scan_contig(npages, curr_low, high,
2478 alignment, boundary, options);
2479 if (m_run == NULL)
2480 break;
2481 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2482 m_runs[RUN_INDEX(count)] = m_run;
2483 count++;
2484 }
2485
2486 /*
2487 * Reclaim the highest runs in LIFO (descending) order until
2488 * the number of reclaimed pages, "reclaimed", is at least
2489 * MIN_RECLAIM. Reset "reclaimed" each time because each
2490 * reclamation is idempotent, and runs will (likely) recur
2491 * from one scan to the next as restrictions are relaxed.
2492 */
2493 reclaimed = 0;
2494 for (i = 0; count > 0 && i < NRUNS; i++) {
2495 count--;
2496 m_run = m_runs[RUN_INDEX(count)];
2497 error = vm_page_reclaim_run(req_class, npages, m_run,
2498 high);
2499 if (error == 0) {
2500 reclaimed += npages;
2501 if (reclaimed >= MIN_RECLAIM)
2502 return (true);
2503 }
2504 }
2505
2506 /*
2507 * Either relax the restrictions on the next scan or return if
2508 * the last scan had no restrictions.
2509 */
2510 if (options == VPSC_NORESERV)
2511 options = VPSC_NOSUPER;
2512 else if (options == VPSC_NOSUPER)
2513 options = VPSC_ANY;
2514 else if (options == VPSC_ANY)
2515 return (reclaimed != 0);
2516 }
2517 }
2518
2519 /*
2520 * vm_wait: (also see VM_WAIT macro)
2521 *
2522 * Sleep until free pages are available for allocation.
2523 * - Called in various places before memory allocations.
2524 */
2525 void
2526 vm_wait(void)
2527 {
2528
2529 mtx_lock(&vm_page_queue_free_mtx);
2530 if (curproc == pageproc) {
2531 vm_pageout_pages_needed = 1;
2532 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2533 PDROP | PSWP, "VMWait", 0);
2534 } else {
2535 if (__predict_false(pageproc == NULL))
2536 panic("vm_wait in early boot");
2537 if (!vm_pageout_wanted) {
2538 vm_pageout_wanted = true;
2539 wakeup(&vm_pageout_wanted);
2540 }
2541 vm_pages_needed = true;
2542 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2543 "vmwait", 0);
2544 }
2545 }
2546
2547 /*
2548 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2549 *
2550 * Sleep until free pages are available for allocation.
2551 * - Called only in vm_fault so that processes page faulting
2552 * can be easily tracked.
2553 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2554 * processes will be able to grab memory first. Do not change
2555 * this balance without careful testing first.
2556 */
2557 void
2558 vm_waitpfault(void)
2559 {
2560
2561 mtx_lock(&vm_page_queue_free_mtx);
2562 if (!vm_pageout_wanted) {
2563 vm_pageout_wanted = true;
2564 wakeup(&vm_pageout_wanted);
2565 }
2566 vm_pages_needed = true;
2567 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2568 "pfault", 0);
2569 }
2570
2571 struct vm_pagequeue *
2572 vm_page_pagequeue(vm_page_t m)
2573 {
2574
2575 if (vm_page_in_laundry(m))
2576 return (&vm_dom[0].vmd_pagequeues[m->queue]);
2577 else
2578 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2579 }
2580
2581 /*
2582 * vm_page_dequeue:
2583 *
2584 * Remove the given page from its current page queue.
2585 *
2586 * The page must be locked.
2587 */
2588 void
2589 vm_page_dequeue(vm_page_t m)
2590 {
2591 struct vm_pagequeue *pq;
2592
2593 vm_page_assert_locked(m);
2594 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2595 m));
2596 pq = vm_page_pagequeue(m);
2597 vm_pagequeue_lock(pq);
2598 m->queue = PQ_NONE;
2599 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2600 vm_pagequeue_cnt_dec(pq);
2601 vm_pagequeue_unlock(pq);
2602 }
2603
2604 /*
2605 * vm_page_dequeue_locked:
2606 *
2607 * Remove the given page from its current page queue.
2608 *
2609 * The page and page queue must be locked.
2610 */
2611 void
2612 vm_page_dequeue_locked(vm_page_t m)
2613 {
2614 struct vm_pagequeue *pq;
2615
2616 vm_page_lock_assert(m, MA_OWNED);
2617 pq = vm_page_pagequeue(m);
2618 vm_pagequeue_assert_locked(pq);
2619 m->queue = PQ_NONE;
2620 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2621 vm_pagequeue_cnt_dec(pq);
2622 }
2623
2624 /*
2625 * vm_page_enqueue:
2626 *
2627 * Add the given page to the specified page queue.
2628 *
2629 * The page must be locked.
2630 */
2631 static void
2632 vm_page_enqueue(uint8_t queue, vm_page_t m)
2633 {
2634 struct vm_pagequeue *pq;
2635
2636 vm_page_lock_assert(m, MA_OWNED);
2637 KASSERT(queue < PQ_COUNT,
2638 ("vm_page_enqueue: invalid queue %u request for page %p",
2639 queue, m));
2640 if (queue == PQ_LAUNDRY)
2641 pq = &vm_dom[0].vmd_pagequeues[queue];
2642 else
2643 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2644 vm_pagequeue_lock(pq);
2645 m->queue = queue;
2646 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2647 vm_pagequeue_cnt_inc(pq);
2648 vm_pagequeue_unlock(pq);
2649 }
2650
2651 /*
2652 * vm_page_requeue:
2653 *
2654 * Move the given page to the tail of its current page queue.
2655 *
2656 * The page must be locked.
2657 */
2658 void
2659 vm_page_requeue(vm_page_t m)
2660 {
2661 struct vm_pagequeue *pq;
2662
2663 vm_page_lock_assert(m, MA_OWNED);
2664 KASSERT(m->queue != PQ_NONE,
2665 ("vm_page_requeue: page %p is not queued", m));
2666 pq = vm_page_pagequeue(m);
2667 vm_pagequeue_lock(pq);
2668 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2669 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2670 vm_pagequeue_unlock(pq);
2671 }
2672
2673 /*
2674 * vm_page_requeue_locked:
2675 *
2676 * Move the given page to the tail of its current page queue.
2677 *
2678 * The page queue must be locked.
2679 */
2680 void
2681 vm_page_requeue_locked(vm_page_t m)
2682 {
2683 struct vm_pagequeue *pq;
2684
2685 KASSERT(m->queue != PQ_NONE,
2686 ("vm_page_requeue_locked: page %p is not queued", m));
2687 pq = vm_page_pagequeue(m);
2688 vm_pagequeue_assert_locked(pq);
2689 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2690 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2691 }
2692
2693 /*
2694 * vm_page_activate:
2695 *
2696 * Put the specified page on the active list (if appropriate).
2697 * Ensure that act_count is at least ACT_INIT but do not otherwise
2698 * mess with it.
2699 *
2700 * The page must be locked.
2701 */
2702 void
2703 vm_page_activate(vm_page_t m)
2704 {
2705 int queue;
2706
2707 vm_page_lock_assert(m, MA_OWNED);
2708 if ((queue = m->queue) != PQ_ACTIVE) {
2709 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2710 if (m->act_count < ACT_INIT)
2711 m->act_count = ACT_INIT;
2712 if (queue != PQ_NONE)
2713 vm_page_dequeue(m);
2714 vm_page_enqueue(PQ_ACTIVE, m);
2715 } else
2716 KASSERT(queue == PQ_NONE,
2717 ("vm_page_activate: wired page %p is queued", m));
2718 } else {
2719 if (m->act_count < ACT_INIT)
2720 m->act_count = ACT_INIT;
2721 }
2722 }
2723
2724 /*
2725 * vm_page_free_wakeup:
2726 *
2727 * Helper routine for vm_page_free_toq(). This routine is called
2728 * when a page is added to the free queues.
2729 *
2730 * The page queues must be locked.
2731 */
2732 static inline void
2733 vm_page_free_wakeup(void)
2734 {
2735
2736 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2737 /*
2738 * if pageout daemon needs pages, then tell it that there are
2739 * some free.
2740 */
2741 if (vm_pageout_pages_needed &&
2742 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2743 wakeup(&vm_pageout_pages_needed);
2744 vm_pageout_pages_needed = 0;
2745 }
2746 /*
2747 * wakeup processes that are waiting on memory if we hit a
2748 * high water mark. And wakeup scheduler process if we have
2749 * lots of memory. this process will swapin processes.
2750 */
2751 if (vm_pages_needed && !vm_page_count_min()) {
2752 vm_pages_needed = false;
2753 wakeup(&vm_cnt.v_free_count);
2754 }
2755 }
2756
2757 /*
2758 * vm_page_free_toq:
2759 *
2760 * Returns the given page to the free list,
2761 * disassociating it with any VM object.
2762 *
2763 * The object must be locked. The page must be locked if it is managed.
2764 */
2765 void
2766 vm_page_free_toq(vm_page_t m)
2767 {
2768
2769 if ((m->oflags & VPO_UNMANAGED) == 0) {
2770 vm_page_lock_assert(m, MA_OWNED);
2771 KASSERT(!pmap_page_is_mapped(m),
2772 ("vm_page_free_toq: freeing mapped page %p", m));
2773 } else
2774 KASSERT(m->queue == PQ_NONE,
2775 ("vm_page_free_toq: unmanaged page %p is queued", m));
2776 PCPU_INC(cnt.v_tfree);
2777
2778 if (vm_page_sbusied(m))
2779 panic("vm_page_free: freeing busy page %p", m);
2780
2781 /*
2782 * Unqueue, then remove page. Note that we cannot destroy
2783 * the page here because we do not want to call the pager's
2784 * callback routine until after we've put the page on the
2785 * appropriate free queue.
2786 */
2787 vm_page_remque(m);
2788 vm_page_remove(m);
2789
2790 /*
2791 * If fictitious remove object association and
2792 * return, otherwise delay object association removal.
2793 */
2794 if ((m->flags & PG_FICTITIOUS) != 0) {
2795 return;
2796 }
2797
2798 m->valid = 0;
2799 vm_page_undirty(m);
2800
2801 if (m->wire_count != 0)
2802 panic("vm_page_free: freeing wired page %p", m);
2803 if (m->hold_count != 0) {
2804 m->flags &= ~PG_ZERO;
2805 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2806 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2807 m->flags |= PG_UNHOLDFREE;
2808 } else {
2809 /*
2810 * Restore the default memory attribute to the page.
2811 */
2812 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2813 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2814
2815 /*
2816 * Insert the page into the physical memory allocator's free
2817 * page queues.
2818 */
2819 mtx_lock(&vm_page_queue_free_mtx);
2820 vm_phys_freecnt_adj(m, 1);
2821 #if VM_NRESERVLEVEL > 0
2822 if (!vm_reserv_free_page(m))
2823 #else
2824 if (TRUE)
2825 #endif
2826 vm_phys_free_pages(m, 0);
2827 if ((m->flags & PG_ZERO) != 0)
2828 ++vm_page_zero_count;
2829 else
2830 vm_page_zero_idle_wakeup();
2831 vm_page_free_wakeup();
2832 mtx_unlock(&vm_page_queue_free_mtx);
2833 }
2834 }
2835
2836 /*
2837 * vm_page_wire:
2838 *
2839 * Mark this page as wired down by yet
2840 * another map, removing it from paging queues
2841 * as necessary.
2842 *
2843 * If the page is fictitious, then its wire count must remain one.
2844 *
2845 * The page must be locked.
2846 */
2847 void
2848 vm_page_wire(vm_page_t m)
2849 {
2850
2851 /*
2852 * Only bump the wire statistics if the page is not already wired,
2853 * and only unqueue the page if it is on some queue (if it is unmanaged
2854 * it is already off the queues).
2855 */
2856 vm_page_lock_assert(m, MA_OWNED);
2857 if ((m->flags & PG_FICTITIOUS) != 0) {
2858 KASSERT(m->wire_count == 1,
2859 ("vm_page_wire: fictitious page %p's wire count isn't one",
2860 m));
2861 return;
2862 }
2863 if (m->wire_count == 0) {
2864 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2865 m->queue == PQ_NONE,
2866 ("vm_page_wire: unmanaged page %p is queued", m));
2867 vm_page_remque(m);
2868 atomic_add_int(&vm_cnt.v_wire_count, 1);
2869 }
2870 m->wire_count++;
2871 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2872 }
2873
2874 /*
2875 * vm_page_unwire:
2876 *
2877 * Release one wiring of the specified page, potentially allowing it to be
2878 * paged out. Returns TRUE if the number of wirings transitions to zero and
2879 * FALSE otherwise.
2880 *
2881 * Only managed pages belonging to an object can be paged out. If the number
2882 * of wirings transitions to zero and the page is eligible for page out, then
2883 * the page is added to the specified paging queue (unless PQ_NONE is
2884 * specified).
2885 *
2886 * If a page is fictitious, then its wire count must always be one.
2887 *
2888 * A managed page must be locked.
2889 */
2890 boolean_t
2891 vm_page_unwire(vm_page_t m, uint8_t queue)
2892 {
2893
2894 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
2895 ("vm_page_unwire: invalid queue %u request for page %p",
2896 queue, m));
2897 if ((m->oflags & VPO_UNMANAGED) == 0)
2898 vm_page_assert_locked(m);
2899 if ((m->flags & PG_FICTITIOUS) != 0) {
2900 KASSERT(m->wire_count == 1,
2901 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2902 return (FALSE);
2903 }
2904 if (m->wire_count > 0) {
2905 m->wire_count--;
2906 if (m->wire_count == 0) {
2907 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2908 if ((m->oflags & VPO_UNMANAGED) == 0 &&
2909 m->object != NULL && queue != PQ_NONE)
2910 vm_page_enqueue(queue, m);
2911 return (TRUE);
2912 } else
2913 return (FALSE);
2914 } else
2915 panic("vm_page_unwire: page %p's wire count is zero", m);
2916 }
2917
2918 /*
2919 * Move the specified page to the inactive queue.
2920 *
2921 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
2922 * queue. However, setting "noreuse" to TRUE will accelerate the specified
2923 * page's reclamation, but it will not unmap the page from any address space.
2924 * This is implemented by inserting the page near the head of the inactive
2925 * queue, using a marker page to guide FIFO insertion ordering.
2926 *
2927 * The page must be locked.
2928 */
2929 static inline void
2930 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
2931 {
2932 struct vm_pagequeue *pq;
2933 int queue;
2934
2935 vm_page_assert_locked(m);
2936
2937 /*
2938 * Ignore if the page is already inactive, unless it is unlikely to be
2939 * reactivated.
2940 */
2941 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
2942 return;
2943 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2944 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2945 /* Avoid multiple acquisitions of the inactive queue lock. */
2946 if (queue == PQ_INACTIVE) {
2947 vm_pagequeue_lock(pq);
2948 vm_page_dequeue_locked(m);
2949 } else {
2950 if (queue != PQ_NONE)
2951 vm_page_dequeue(m);
2952 vm_pagequeue_lock(pq);
2953 }
2954 m->queue = PQ_INACTIVE;
2955 if (noreuse)
2956 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
2957 m, plinks.q);
2958 else
2959 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2960 vm_pagequeue_cnt_inc(pq);
2961 vm_pagequeue_unlock(pq);
2962 }
2963 }
2964
2965 /*
2966 * Move the specified page to the inactive queue.
2967 *
2968 * The page must be locked.
2969 */
2970 void
2971 vm_page_deactivate(vm_page_t m)
2972 {
2973
2974 _vm_page_deactivate(m, FALSE);
2975 }
2976
2977 /*
2978 * Move the specified page to the inactive queue with the expectation
2979 * that it is unlikely to be reused.
2980 *
2981 * The page must be locked.
2982 */
2983 void
2984 vm_page_deactivate_noreuse(vm_page_t m)
2985 {
2986
2987 _vm_page_deactivate(m, TRUE);
2988 }
2989
2990 /*
2991 * vm_page_launder
2992 *
2993 * Put a page in the laundry.
2994 */
2995 void
2996 vm_page_launder(vm_page_t m)
2997 {
2998 int queue;
2999
3000 vm_page_assert_locked(m);
3001 if ((queue = m->queue) != PQ_LAUNDRY) {
3002 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3003 if (queue != PQ_NONE)
3004 vm_page_dequeue(m);
3005 vm_page_enqueue(PQ_LAUNDRY, m);
3006 } else
3007 KASSERT(queue == PQ_NONE,
3008 ("wired page %p is queued", m));
3009 }
3010 }
3011
3012 /*
3013 * vm_page_try_to_free()
3014 *
3015 * Attempt to free the page. If we cannot free it, we do nothing.
3016 * 1 is returned on success, 0 on failure.
3017 */
3018 int
3019 vm_page_try_to_free(vm_page_t m)
3020 {
3021
3022 vm_page_lock_assert(m, MA_OWNED);
3023 if (m->object != NULL)
3024 VM_OBJECT_ASSERT_WLOCKED(m->object);
3025 if (m->dirty || m->hold_count || m->wire_count ||
3026 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3027 return (0);
3028 pmap_remove_all(m);
3029 if (m->dirty)
3030 return (0);
3031 vm_page_free(m);
3032 return (1);
3033 }
3034
3035 /*
3036 * vm_page_advise
3037 *
3038 * Deactivate or do nothing, as appropriate.
3039 *
3040 * The object and page must be locked.
3041 */
3042 void
3043 vm_page_advise(vm_page_t m, int advice)
3044 {
3045
3046 vm_page_assert_locked(m);
3047 VM_OBJECT_ASSERT_WLOCKED(m->object);
3048 if (advice == MADV_FREE)
3049 /*
3050 * Mark the page clean. This will allow the page to be freed
3051 * without first paging it out. MADV_FREE pages are often
3052 * quickly reused by malloc(3), so we do not do anything that
3053 * would result in a page fault on a later access.
3054 */
3055 vm_page_undirty(m);
3056 else if (advice != MADV_DONTNEED)
3057 return;
3058
3059 /*
3060 * Clear any references to the page. Otherwise, the page daemon will
3061 * immediately reactivate the page.
3062 */
3063 vm_page_aflag_clear(m, PGA_REFERENCED);
3064
3065 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3066 vm_page_dirty(m);
3067
3068 /*
3069 * Place clean pages near the head of the inactive queue rather than
3070 * the tail, thus defeating the queue's LRU operation and ensuring that
3071 * the page will be reused quickly. Dirty pages not already in the
3072 * laundry are moved there.
3073 */
3074 if (m->dirty == 0)
3075 vm_page_deactivate_noreuse(m);
3076 else
3077 vm_page_launder(m);
3078 }
3079
3080 /*
3081 * Grab a page, waiting until we are waken up due to the page
3082 * changing state. We keep on waiting, if the page continues
3083 * to be in the object. If the page doesn't exist, first allocate it
3084 * and then conditionally zero it.
3085 *
3086 * This routine may sleep.
3087 *
3088 * The object must be locked on entry. The lock will, however, be released
3089 * and reacquired if the routine sleeps.
3090 */
3091 vm_page_t
3092 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3093 {
3094 vm_page_t m;
3095 int sleep;
3096
3097 VM_OBJECT_ASSERT_WLOCKED(object);
3098 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3099 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3100 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3101 retrylookup:
3102 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3103 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3104 vm_page_xbusied(m) : vm_page_busied(m);
3105 if (sleep) {
3106 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3107 return (NULL);
3108 /*
3109 * Reference the page before unlocking and
3110 * sleeping so that the page daemon is less
3111 * likely to reclaim it.
3112 */
3113 vm_page_aflag_set(m, PGA_REFERENCED);
3114 vm_page_lock(m);
3115 VM_OBJECT_WUNLOCK(object);
3116 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3117 VM_ALLOC_IGN_SBUSY) != 0);
3118 VM_OBJECT_WLOCK(object);
3119 goto retrylookup;
3120 } else {
3121 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3122 vm_page_lock(m);
3123 vm_page_wire(m);
3124 vm_page_unlock(m);
3125 }
3126 if ((allocflags &
3127 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3128 vm_page_xbusy(m);
3129 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3130 vm_page_sbusy(m);
3131 return (m);
3132 }
3133 }
3134 m = vm_page_alloc(object, pindex, allocflags);
3135 if (m == NULL) {
3136 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3137 return (NULL);
3138 VM_OBJECT_WUNLOCK(object);
3139 VM_WAIT;
3140 VM_OBJECT_WLOCK(object);
3141 goto retrylookup;
3142 }
3143 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3144 pmap_zero_page(m);
3145 return (m);
3146 }
3147
3148 /*
3149 * Mapping function for valid or dirty bits in a page.
3150 *
3151 * Inputs are required to range within a page.
3152 */
3153 vm_page_bits_t
3154 vm_page_bits(int base, int size)
3155 {
3156 int first_bit;
3157 int last_bit;
3158
3159 KASSERT(
3160 base + size <= PAGE_SIZE,
3161 ("vm_page_bits: illegal base/size %d/%d", base, size)
3162 );
3163
3164 if (size == 0) /* handle degenerate case */
3165 return (0);
3166
3167 first_bit = base >> DEV_BSHIFT;
3168 last_bit = (base + size - 1) >> DEV_BSHIFT;
3169
3170 return (((vm_page_bits_t)2 << last_bit) -
3171 ((vm_page_bits_t)1 << first_bit));
3172 }
3173
3174 /*
3175 * vm_page_set_valid_range:
3176 *
3177 * Sets portions of a page valid. The arguments are expected
3178 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3179 * of any partial chunks touched by the range. The invalid portion of
3180 * such chunks will be zeroed.
3181 *
3182 * (base + size) must be less then or equal to PAGE_SIZE.
3183 */
3184 void
3185 vm_page_set_valid_range(vm_page_t m, int base, int size)
3186 {
3187 int endoff, frag;
3188
3189 VM_OBJECT_ASSERT_WLOCKED(m->object);
3190 if (size == 0) /* handle degenerate case */
3191 return;
3192
3193 /*
3194 * If the base is not DEV_BSIZE aligned and the valid
3195 * bit is clear, we have to zero out a portion of the
3196 * first block.
3197 */
3198 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3199 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3200 pmap_zero_page_area(m, frag, base - frag);
3201
3202 /*
3203 * If the ending offset is not DEV_BSIZE aligned and the
3204 * valid bit is clear, we have to zero out a portion of
3205 * the last block.
3206 */
3207 endoff = base + size;
3208 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3209 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3210 pmap_zero_page_area(m, endoff,
3211 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3212
3213 /*
3214 * Assert that no previously invalid block that is now being validated
3215 * is already dirty.
3216 */
3217 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3218 ("vm_page_set_valid_range: page %p is dirty", m));
3219
3220 /*
3221 * Set valid bits inclusive of any overlap.
3222 */
3223 m->valid |= vm_page_bits(base, size);
3224 }
3225
3226 /*
3227 * Clear the given bits from the specified page's dirty field.
3228 */
3229 static __inline void
3230 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3231 {
3232 uintptr_t addr;
3233 #if PAGE_SIZE < 16384
3234 int shift;
3235 #endif
3236
3237 /*
3238 * If the object is locked and the page is neither exclusive busy nor
3239 * write mapped, then the page's dirty field cannot possibly be
3240 * set by a concurrent pmap operation.
3241 */
3242 VM_OBJECT_ASSERT_WLOCKED(m->object);
3243 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3244 m->dirty &= ~pagebits;
3245 else {
3246 /*
3247 * The pmap layer can call vm_page_dirty() without
3248 * holding a distinguished lock. The combination of
3249 * the object's lock and an atomic operation suffice
3250 * to guarantee consistency of the page dirty field.
3251 *
3252 * For PAGE_SIZE == 32768 case, compiler already
3253 * properly aligns the dirty field, so no forcible
3254 * alignment is needed. Only require existence of
3255 * atomic_clear_64 when page size is 32768.
3256 */
3257 addr = (uintptr_t)&m->dirty;
3258 #if PAGE_SIZE == 32768
3259 atomic_clear_64((uint64_t *)addr, pagebits);
3260 #elif PAGE_SIZE == 16384
3261 atomic_clear_32((uint32_t *)addr, pagebits);
3262 #else /* PAGE_SIZE <= 8192 */
3263 /*
3264 * Use a trick to perform a 32-bit atomic on the
3265 * containing aligned word, to not depend on the existence
3266 * of atomic_clear_{8, 16}.
3267 */
3268 shift = addr & (sizeof(uint32_t) - 1);
3269 #if BYTE_ORDER == BIG_ENDIAN
3270 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3271 #else
3272 shift *= NBBY;
3273 #endif
3274 addr &= ~(sizeof(uint32_t) - 1);
3275 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3276 #endif /* PAGE_SIZE */
3277 }
3278 }
3279
3280 /*
3281 * vm_page_set_validclean:
3282 *
3283 * Sets portions of a page valid and clean. The arguments are expected
3284 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3285 * of any partial chunks touched by the range. The invalid portion of
3286 * such chunks will be zero'd.
3287 *
3288 * (base + size) must be less then or equal to PAGE_SIZE.
3289 */
3290 void
3291 vm_page_set_validclean(vm_page_t m, int base, int size)
3292 {
3293 vm_page_bits_t oldvalid, pagebits;
3294 int endoff, frag;
3295
3296 VM_OBJECT_ASSERT_WLOCKED(m->object);
3297 if (size == 0) /* handle degenerate case */
3298 return;
3299
3300 /*
3301 * If the base is not DEV_BSIZE aligned and the valid
3302 * bit is clear, we have to zero out a portion of the
3303 * first block.
3304 */
3305 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3306 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3307 pmap_zero_page_area(m, frag, base - frag);
3308
3309 /*
3310 * If the ending offset is not DEV_BSIZE aligned and the
3311 * valid bit is clear, we have to zero out a portion of
3312 * the last block.
3313 */
3314 endoff = base + size;
3315 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3316 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3317 pmap_zero_page_area(m, endoff,
3318 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3319
3320 /*
3321 * Set valid, clear dirty bits. If validating the entire
3322 * page we can safely clear the pmap modify bit. We also
3323 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3324 * takes a write fault on a MAP_NOSYNC memory area the flag will
3325 * be set again.
3326 *
3327 * We set valid bits inclusive of any overlap, but we can only
3328 * clear dirty bits for DEV_BSIZE chunks that are fully within
3329 * the range.
3330 */
3331 oldvalid = m->valid;
3332 pagebits = vm_page_bits(base, size);
3333 m->valid |= pagebits;
3334 #if 0 /* NOT YET */
3335 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3336 frag = DEV_BSIZE - frag;
3337 base += frag;
3338 size -= frag;
3339 if (size < 0)
3340 size = 0;
3341 }
3342 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3343 #endif
3344 if (base == 0 && size == PAGE_SIZE) {
3345 /*
3346 * The page can only be modified within the pmap if it is
3347 * mapped, and it can only be mapped if it was previously
3348 * fully valid.
3349 */
3350 if (oldvalid == VM_PAGE_BITS_ALL)
3351 /*
3352 * Perform the pmap_clear_modify() first. Otherwise,
3353 * a concurrent pmap operation, such as
3354 * pmap_protect(), could clear a modification in the
3355 * pmap and set the dirty field on the page before
3356 * pmap_clear_modify() had begun and after the dirty
3357 * field was cleared here.
3358 */
3359 pmap_clear_modify(m);
3360 m->dirty = 0;
3361 m->oflags &= ~VPO_NOSYNC;
3362 } else if (oldvalid != VM_PAGE_BITS_ALL)
3363 m->dirty &= ~pagebits;
3364 else
3365 vm_page_clear_dirty_mask(m, pagebits);
3366 }
3367
3368 void
3369 vm_page_clear_dirty(vm_page_t m, int base, int size)
3370 {
3371
3372 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3373 }
3374
3375 /*
3376 * vm_page_set_invalid:
3377 *
3378 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3379 * valid and dirty bits for the effected areas are cleared.
3380 */
3381 void
3382 vm_page_set_invalid(vm_page_t m, int base, int size)
3383 {
3384 vm_page_bits_t bits;
3385 vm_object_t object;
3386
3387 object = m->object;
3388 VM_OBJECT_ASSERT_WLOCKED(object);
3389 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3390 size >= object->un_pager.vnp.vnp_size)
3391 bits = VM_PAGE_BITS_ALL;
3392 else
3393 bits = vm_page_bits(base, size);
3394 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3395 bits != 0)
3396 pmap_remove_all(m);
3397 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3398 !pmap_page_is_mapped(m),
3399 ("vm_page_set_invalid: page %p is mapped", m));
3400 m->valid &= ~bits;
3401 m->dirty &= ~bits;
3402 }
3403
3404 /*
3405 * vm_page_zero_invalid()
3406 *
3407 * The kernel assumes that the invalid portions of a page contain
3408 * garbage, but such pages can be mapped into memory by user code.
3409 * When this occurs, we must zero out the non-valid portions of the
3410 * page so user code sees what it expects.
3411 *
3412 * Pages are most often semi-valid when the end of a file is mapped
3413 * into memory and the file's size is not page aligned.
3414 */
3415 void
3416 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3417 {
3418 int b;
3419 int i;
3420
3421 VM_OBJECT_ASSERT_WLOCKED(m->object);
3422 /*
3423 * Scan the valid bits looking for invalid sections that
3424 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3425 * valid bit may be set ) have already been zeroed by
3426 * vm_page_set_validclean().
3427 */
3428 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3429 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3430 (m->valid & ((vm_page_bits_t)1 << i))) {
3431 if (i > b) {
3432 pmap_zero_page_area(m,
3433 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3434 }
3435 b = i + 1;
3436 }
3437 }
3438
3439 /*
3440 * setvalid is TRUE when we can safely set the zero'd areas
3441 * as being valid. We can do this if there are no cache consistancy
3442 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3443 */
3444 if (setvalid)
3445 m->valid = VM_PAGE_BITS_ALL;
3446 }
3447
3448 /*
3449 * vm_page_is_valid:
3450 *
3451 * Is (partial) page valid? Note that the case where size == 0
3452 * will return FALSE in the degenerate case where the page is
3453 * entirely invalid, and TRUE otherwise.
3454 */
3455 int
3456 vm_page_is_valid(vm_page_t m, int base, int size)
3457 {
3458 vm_page_bits_t bits;
3459
3460 VM_OBJECT_ASSERT_LOCKED(m->object);
3461 bits = vm_page_bits(base, size);
3462 return (m->valid != 0 && (m->valid & bits) == bits);
3463 }
3464
3465 /*
3466 * vm_page_ps_is_valid:
3467 *
3468 * Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3469 */
3470 boolean_t
3471 vm_page_ps_is_valid(vm_page_t m)
3472 {
3473 int i, npages;
3474
3475 VM_OBJECT_ASSERT_LOCKED(m->object);
3476 npages = atop(pagesizes[m->psind]);
3477
3478 /*
3479 * The physically contiguous pages that make up a superpage, i.e., a
3480 * page with a page size index ("psind") greater than zero, will
3481 * occupy adjacent entries in vm_page_array[].
3482 */
3483 for (i = 0; i < npages; i++) {
3484 if (m[i].valid != VM_PAGE_BITS_ALL)
3485 return (FALSE);
3486 }
3487 return (TRUE);
3488 }
3489
3490 /*
3491 * Set the page's dirty bits if the page is modified.
3492 */
3493 void
3494 vm_page_test_dirty(vm_page_t m)
3495 {
3496
3497 VM_OBJECT_ASSERT_WLOCKED(m->object);
3498 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3499 vm_page_dirty(m);
3500 }
3501
3502 void
3503 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3504 {
3505
3506 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3507 }
3508
3509 void
3510 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3511 {
3512
3513 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3514 }
3515
3516 int
3517 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3518 {
3519
3520 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3521 }
3522
3523 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3524 void
3525 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3526 {
3527
3528 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3529 }
3530
3531 void
3532 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3533 {
3534
3535 mtx_assert_(vm_page_lockptr(m), a, file, line);
3536 }
3537 #endif
3538
3539 #ifdef INVARIANTS
3540 void
3541 vm_page_object_lock_assert(vm_page_t m)
3542 {
3543
3544 /*
3545 * Certain of the page's fields may only be modified by the
3546 * holder of the containing object's lock or the exclusive busy.
3547 * holder. Unfortunately, the holder of the write busy is
3548 * not recorded, and thus cannot be checked here.
3549 */
3550 if (m->object != NULL && !vm_page_xbusied(m))
3551 VM_OBJECT_ASSERT_WLOCKED(m->object);
3552 }
3553
3554 void
3555 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3556 {
3557
3558 if ((bits & PGA_WRITEABLE) == 0)
3559 return;
3560
3561 /*
3562 * The PGA_WRITEABLE flag can only be set if the page is
3563 * managed, is exclusively busied or the object is locked.
3564 * Currently, this flag is only set by pmap_enter().
3565 */
3566 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3567 ("PGA_WRITEABLE on unmanaged page"));
3568 if (!vm_page_xbusied(m))
3569 VM_OBJECT_ASSERT_LOCKED(m->object);
3570 }
3571 #endif
3572
3573 #include "opt_ddb.h"
3574 #ifdef DDB
3575 #include <sys/kernel.h>
3576
3577 #include <ddb/ddb.h>
3578
3579 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3580 {
3581
3582 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3583 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3584 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3585 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3586 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3587 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3588 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3589 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3590 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3591 }
3592
3593 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3594 {
3595 int dom;
3596
3597 db_printf("pq_free %d\n", vm_cnt.v_free_count);
3598 for (dom = 0; dom < vm_ndomains; dom++) {
3599 db_printf(
3600 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d\n",
3601 dom,
3602 vm_dom[dom].vmd_page_count,
3603 vm_dom[dom].vmd_free_count,
3604 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3605 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3606 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt);
3607 }
3608 }
3609
3610 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3611 {
3612 vm_page_t m;
3613 boolean_t phys;
3614
3615 if (!have_addr) {
3616 db_printf("show pginfo addr\n");
3617 return;
3618 }
3619
3620 phys = strchr(modif, 'p') != NULL;
3621 if (phys)
3622 m = PHYS_TO_VM_PAGE(addr);
3623 else
3624 m = (vm_page_t)addr;
3625 db_printf(
3626 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3627 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3628 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3629 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3630 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3631 }
3632 #endif /* DDB */
Cache object: 4bbc3e0a4ea638976999fd56ff299a3c
|