FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_page.c
1 /*-
2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3 *
4 * Copyright (c) 1991 Regents of the University of California.
5 * All rights reserved.
6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
7 *
8 * This code is derived from software contributed to Berkeley by
9 * The Mach Operating System project at Carnegie-Mellon University.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. Neither the name of the University nor the names of its contributors
20 * may be used to endorse or promote products derived from this software
21 * without specific prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * SUCH DAMAGE.
34 *
35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
36 */
37
38 /*-
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
41 *
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43 *
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
49 *
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53 *
54 * Carnegie Mellon requests users of this software to return to
55 *
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
60 *
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
63 */
64
65 /*
66 * Resident memory management module.
67 */
68
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71
72 #include "opt_vm.h"
73
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/lock.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/malloc.h>
82 #include <sys/mman.h>
83 #include <sys/msgbuf.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/rwlock.h>
87 #include <sys/sbuf.h>
88 #include <sys/sched.h>
89 #include <sys/smp.h>
90 #include <sys/sysctl.h>
91 #include <sys/vmmeter.h>
92 #include <sys/vnode.h>
93
94 #include <vm/vm.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_param.h>
97 #include <vm/vm_domainset.h>
98 #include <vm/vm_kern.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/vm_pager.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_reserv.h>
108 #include <vm/vm_extern.h>
109 #include <vm/uma.h>
110 #include <vm/uma_int.h>
111
112 #include <machine/md_var.h>
113
114 extern int uma_startup_count(int);
115 extern void uma_startup(void *, int);
116 extern int vmem_startup_count(void);
117
118 struct vm_domain vm_dom[MAXMEMDOM];
119
120 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
121
122 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
123
124 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
125 /* The following fields are protected by the domainset lock. */
126 domainset_t __exclusive_cache_line vm_min_domains;
127 domainset_t __exclusive_cache_line vm_severe_domains;
128 static int vm_min_waiters;
129 static int vm_severe_waiters;
130 static int vm_pageproc_waiters;
131
132 /*
133 * bogus page -- for I/O to/from partially complete buffers,
134 * or for paging into sparsely invalid regions.
135 */
136 vm_page_t bogus_page;
137
138 vm_page_t vm_page_array;
139 long vm_page_array_size;
140 long first_page;
141
142 static int boot_pages;
143 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
144 &boot_pages, 0,
145 "number of pages allocated for bootstrapping the VM system");
146
147 static int pa_tryrelock_restart;
148 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
149 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
150
151 static TAILQ_HEAD(, vm_page) blacklist_head;
152 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
153 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
154 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
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_dequeue_complete(vm_page_t m);
161 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
162 static void vm_page_init(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, int domain, u_long npages,
168 vm_page_t m_run, vm_paddr_t high);
169 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
170 int req);
171 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
172 int flags);
173 static void vm_page_zone_release(void *arg, void **store, int cnt);
174
175 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
176
177 static void
178 vm_page_init(void *dummy)
179 {
180
181 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
182 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
183 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
184 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
185 }
186
187 /*
188 * The cache page zone is initialized later since we need to be able to allocate
189 * pages before UMA is fully initialized.
190 */
191 static void
192 vm_page_init_cache_zones(void *dummy __unused)
193 {
194 struct vm_domain *vmd;
195 struct vm_pgcache *pgcache;
196 int domain, pool;
197
198 for (domain = 0; domain < vm_ndomains; domain++) {
199 vmd = VM_DOMAIN(domain);
200
201 /*
202 * Don't allow the page caches to take up more than .1875% of
203 * memory. A UMA bucket contains at most 256 free pages, and we
204 * have two buckets per CPU per free pool.
205 */
206 if (vmd->vmd_page_count / 600 < 2 * 256 * mp_ncpus *
207 VM_NFREEPOOL)
208 continue;
209 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
210 pgcache = &vmd->vmd_pgcache[pool];
211 pgcache->domain = domain;
212 pgcache->pool = pool;
213 pgcache->zone = uma_zcache_create("vm pgcache",
214 sizeof(struct vm_page), NULL, NULL, NULL, NULL,
215 vm_page_zone_import, vm_page_zone_release, pgcache,
216 UMA_ZONE_NOBUCKETCACHE | UMA_ZONE_MAXBUCKET |
217 UMA_ZONE_VM);
218 }
219 }
220 }
221 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
222
223 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
224 #if PAGE_SIZE == 32768
225 #ifdef CTASSERT
226 CTASSERT(sizeof(u_long) >= 8);
227 #endif
228 #endif
229
230 /*
231 * Try to acquire a physical address lock while a pmap is locked. If we
232 * fail to trylock we unlock and lock the pmap directly and cache the
233 * locked pa in *locked. The caller should then restart their loop in case
234 * the virtual to physical mapping has changed.
235 */
236 int
237 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
238 {
239 vm_paddr_t lockpa;
240
241 lockpa = *locked;
242 *locked = pa;
243 if (lockpa) {
244 PA_LOCK_ASSERT(lockpa, MA_OWNED);
245 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
246 return (0);
247 PA_UNLOCK(lockpa);
248 }
249 if (PA_TRYLOCK(pa))
250 return (0);
251 PMAP_UNLOCK(pmap);
252 atomic_add_int(&pa_tryrelock_restart, 1);
253 PA_LOCK(pa);
254 PMAP_LOCK(pmap);
255 return (EAGAIN);
256 }
257
258 /*
259 * vm_set_page_size:
260 *
261 * Sets the page size, perhaps based upon the memory
262 * size. Must be called before any use of page-size
263 * dependent functions.
264 */
265 void
266 vm_set_page_size(void)
267 {
268 if (vm_cnt.v_page_size == 0)
269 vm_cnt.v_page_size = PAGE_SIZE;
270 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
271 panic("vm_set_page_size: page size not a power of two");
272 }
273
274 /*
275 * vm_page_blacklist_next:
276 *
277 * Find the next entry in the provided string of blacklist
278 * addresses. Entries are separated by space, comma, or newline.
279 * If an invalid integer is encountered then the rest of the
280 * string is skipped. Updates the list pointer to the next
281 * character, or NULL if the string is exhausted or invalid.
282 */
283 static vm_paddr_t
284 vm_page_blacklist_next(char **list, char *end)
285 {
286 vm_paddr_t bad;
287 char *cp, *pos;
288
289 if (list == NULL || *list == NULL)
290 return (0);
291 if (**list =='\0') {
292 *list = NULL;
293 return (0);
294 }
295
296 /*
297 * If there's no end pointer then the buffer is coming from
298 * the kenv and we know it's null-terminated.
299 */
300 if (end == NULL)
301 end = *list + strlen(*list);
302
303 /* Ensure that strtoq() won't walk off the end */
304 if (*end != '\0') {
305 if (*end == '\n' || *end == ' ' || *end == ',')
306 *end = '\0';
307 else {
308 printf("Blacklist not terminated, skipping\n");
309 *list = NULL;
310 return (0);
311 }
312 }
313
314 for (pos = *list; *pos != '\0'; pos = cp) {
315 bad = strtoq(pos, &cp, 0);
316 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
317 if (bad == 0) {
318 if (++cp < end)
319 continue;
320 else
321 break;
322 }
323 } else
324 break;
325 if (*cp == '\0' || ++cp >= end)
326 *list = NULL;
327 else
328 *list = cp;
329 return (trunc_page(bad));
330 }
331 printf("Garbage in RAM blacklist, skipping\n");
332 *list = NULL;
333 return (0);
334 }
335
336 bool
337 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
338 {
339 struct vm_domain *vmd;
340 vm_page_t m;
341 int ret;
342
343 m = vm_phys_paddr_to_vm_page(pa);
344 if (m == NULL)
345 return (true); /* page does not exist, no failure */
346
347 vmd = vm_pagequeue_domain(m);
348 vm_domain_free_lock(vmd);
349 ret = vm_phys_unfree_page(m);
350 vm_domain_free_unlock(vmd);
351 if (ret != 0) {
352 vm_domain_freecnt_inc(vmd, -1);
353 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
354 if (verbose)
355 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
356 }
357 return (ret);
358 }
359
360 /*
361 * vm_page_blacklist_check:
362 *
363 * Iterate through the provided string of blacklist addresses, pulling
364 * each entry out of the physical allocator free list and putting it
365 * onto a list for reporting via the vm.page_blacklist sysctl.
366 */
367 static void
368 vm_page_blacklist_check(char *list, char *end)
369 {
370 vm_paddr_t pa;
371 char *next;
372
373 next = list;
374 while (next != NULL) {
375 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
376 continue;
377 vm_page_blacklist_add(pa, bootverbose);
378 }
379 }
380
381 /*
382 * vm_page_blacklist_load:
383 *
384 * Search for a special module named "ram_blacklist". It'll be a
385 * plain text file provided by the user via the loader directive
386 * of the same name.
387 */
388 static void
389 vm_page_blacklist_load(char **list, char **end)
390 {
391 void *mod;
392 u_char *ptr;
393 u_int len;
394
395 mod = NULL;
396 ptr = NULL;
397
398 mod = preload_search_by_type("ram_blacklist");
399 if (mod != NULL) {
400 ptr = preload_fetch_addr(mod);
401 len = preload_fetch_size(mod);
402 }
403 *list = ptr;
404 if (ptr != NULL)
405 *end = ptr + len;
406 else
407 *end = NULL;
408 return;
409 }
410
411 static int
412 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
413 {
414 vm_page_t m;
415 struct sbuf sbuf;
416 int error, first;
417
418 first = 1;
419 error = sysctl_wire_old_buffer(req, 0);
420 if (error != 0)
421 return (error);
422 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
423 TAILQ_FOREACH(m, &blacklist_head, listq) {
424 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
425 (uintmax_t)m->phys_addr);
426 first = 0;
427 }
428 error = sbuf_finish(&sbuf);
429 sbuf_delete(&sbuf);
430 return (error);
431 }
432
433 /*
434 * Initialize a dummy page for use in scans of the specified paging queue.
435 * In principle, this function only needs to set the flag PG_MARKER.
436 * Nonetheless, it write busies and initializes the hold count to one as
437 * safety precautions.
438 */
439 static void
440 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags)
441 {
442
443 bzero(marker, sizeof(*marker));
444 marker->flags = PG_MARKER;
445 marker->aflags = aflags;
446 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
447 marker->queue = queue;
448 marker->hold_count = 1;
449 }
450
451 static void
452 vm_page_domain_init(int domain)
453 {
454 struct vm_domain *vmd;
455 struct vm_pagequeue *pq;
456 int i;
457
458 vmd = VM_DOMAIN(domain);
459 bzero(vmd, sizeof(*vmd));
460 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
461 "vm inactive pagequeue";
462 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
463 "vm active pagequeue";
464 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
465 "vm laundry pagequeue";
466 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
467 "vm unswappable pagequeue";
468 vmd->vmd_domain = domain;
469 vmd->vmd_page_count = 0;
470 vmd->vmd_free_count = 0;
471 vmd->vmd_segs = 0;
472 vmd->vmd_oom = FALSE;
473 for (i = 0; i < PQ_COUNT; i++) {
474 pq = &vmd->vmd_pagequeues[i];
475 TAILQ_INIT(&pq->pq_pl);
476 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
477 MTX_DEF | MTX_DUPOK);
478 pq->pq_pdpages = 0;
479 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
480 }
481 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
482 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
483 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
484
485 /*
486 * inacthead is used to provide FIFO ordering for LRU-bypassing
487 * insertions.
488 */
489 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
490 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
491 &vmd->vmd_inacthead, plinks.q);
492
493 /*
494 * The clock pages are used to implement active queue scanning without
495 * requeues. Scans start at clock[0], which is advanced after the scan
496 * ends. When the two clock hands meet, they are reset and scanning
497 * resumes from the head of the queue.
498 */
499 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
500 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
501 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
502 &vmd->vmd_clock[0], plinks.q);
503 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
504 &vmd->vmd_clock[1], plinks.q);
505 }
506
507 /*
508 * Initialize a physical page in preparation for adding it to the free
509 * lists.
510 */
511 static void
512 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
513 {
514
515 m->object = NULL;
516 m->wire_count = 0;
517 m->busy_lock = VPB_UNBUSIED;
518 m->hold_count = 0;
519 m->flags = m->aflags = 0;
520 m->phys_addr = pa;
521 m->queue = PQ_NONE;
522 m->psind = 0;
523 m->segind = segind;
524 m->order = VM_NFREEORDER;
525 m->pool = VM_FREEPOOL_DEFAULT;
526 m->valid = m->dirty = 0;
527 pmap_page_init(m);
528 }
529
530 /*
531 * vm_page_startup:
532 *
533 * Initializes the resident memory module. Allocates physical memory for
534 * bootstrapping UMA and some data structures that are used to manage
535 * physical pages. Initializes these structures, and populates the free
536 * page queues.
537 */
538 vm_offset_t
539 vm_page_startup(vm_offset_t vaddr)
540 {
541 struct vm_phys_seg *seg;
542 vm_page_t m;
543 char *list, *listend;
544 vm_offset_t mapped;
545 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
546 vm_paddr_t biggestsize, last_pa, pa;
547 u_long pagecount;
548 int biggestone, i, segind;
549 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
550 long ii;
551 #endif
552
553 biggestsize = 0;
554 biggestone = 0;
555 vaddr = round_page(vaddr);
556
557 for (i = 0; phys_avail[i + 1]; i += 2) {
558 phys_avail[i] = round_page(phys_avail[i]);
559 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
560 }
561 for (i = 0; phys_avail[i + 1]; i += 2) {
562 size = phys_avail[i + 1] - phys_avail[i];
563 if (size > biggestsize) {
564 biggestone = i;
565 biggestsize = size;
566 }
567 }
568
569 end = phys_avail[biggestone+1];
570
571 /*
572 * Initialize the page and queue locks.
573 */
574 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
575 for (i = 0; i < PA_LOCK_COUNT; i++)
576 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
577 for (i = 0; i < vm_ndomains; i++)
578 vm_page_domain_init(i);
579
580 /*
581 * Allocate memory for use when boot strapping the kernel memory
582 * allocator. Tell UMA how many zones we are going to create
583 * before going fully functional. UMA will add its zones.
584 *
585 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
586 * KMAP ENTRY, MAP ENTRY, VMSPACE.
587 */
588 boot_pages = uma_startup_count(8);
589
590 #ifndef UMA_MD_SMALL_ALLOC
591 /* vmem_startup() calls uma_prealloc(). */
592 boot_pages += vmem_startup_count();
593 /* vm_map_startup() calls uma_prealloc(). */
594 boot_pages += howmany(MAX_KMAP,
595 UMA_SLAB_SPACE / sizeof(struct vm_map));
596
597 /*
598 * Before going fully functional kmem_init() does allocation
599 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
600 */
601 boot_pages += 2;
602 #endif
603 /*
604 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
605 * manually fetch the value.
606 */
607 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
608 new_end = end - (boot_pages * UMA_SLAB_SIZE);
609 new_end = trunc_page(new_end);
610 mapped = pmap_map(&vaddr, new_end, end,
611 VM_PROT_READ | VM_PROT_WRITE);
612 bzero((void *)mapped, end - new_end);
613 uma_startup((void *)mapped, boot_pages);
614
615 #ifdef WITNESS
616 end = new_end;
617 new_end = end - round_page(witness_startup_count());
618 mapped = pmap_map(&vaddr, new_end, end,
619 VM_PROT_READ | VM_PROT_WRITE);
620 bzero((void *)mapped, end - new_end);
621 witness_startup((void *)mapped);
622 #endif
623
624 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
625 defined(__i386__) || defined(__mips__) || defined(__riscv)
626 /*
627 * Allocate a bitmap to indicate that a random physical page
628 * needs to be included in a minidump.
629 *
630 * The amd64 port needs this to indicate which direct map pages
631 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
632 *
633 * However, i386 still needs this workspace internally within the
634 * minidump code. In theory, they are not needed on i386, but are
635 * included should the sf_buf code decide to use them.
636 */
637 last_pa = 0;
638 for (i = 0; dump_avail[i + 1] != 0; i += 2)
639 if (dump_avail[i + 1] > last_pa)
640 last_pa = dump_avail[i + 1];
641 page_range = last_pa / PAGE_SIZE;
642 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
643 new_end -= vm_page_dump_size;
644 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
645 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
646 bzero((void *)vm_page_dump, vm_page_dump_size);
647 #else
648 (void)last_pa;
649 #endif
650 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
651 defined(__riscv)
652 /*
653 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
654 * When pmap_map() uses the direct map, they are not automatically
655 * included.
656 */
657 for (pa = new_end; pa < end; pa += PAGE_SIZE)
658 dump_add_page(pa);
659 #endif
660 phys_avail[biggestone + 1] = new_end;
661 #ifdef __amd64__
662 /*
663 * Request that the physical pages underlying the message buffer be
664 * included in a crash dump. Since the message buffer is accessed
665 * through the direct map, they are not automatically included.
666 */
667 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
668 last_pa = pa + round_page(msgbufsize);
669 while (pa < last_pa) {
670 dump_add_page(pa);
671 pa += PAGE_SIZE;
672 }
673 #endif
674 /*
675 * Compute the number of pages of memory that will be available for
676 * use, taking into account the overhead of a page structure per page.
677 * In other words, solve
678 * "available physical memory" - round_page(page_range *
679 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
680 * for page_range.
681 */
682 low_avail = phys_avail[0];
683 high_avail = phys_avail[1];
684 for (i = 0; i < vm_phys_nsegs; i++) {
685 if (vm_phys_segs[i].start < low_avail)
686 low_avail = vm_phys_segs[i].start;
687 if (vm_phys_segs[i].end > high_avail)
688 high_avail = vm_phys_segs[i].end;
689 }
690 /* Skip the first chunk. It is already accounted for. */
691 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
692 if (phys_avail[i] < low_avail)
693 low_avail = phys_avail[i];
694 if (phys_avail[i + 1] > high_avail)
695 high_avail = phys_avail[i + 1];
696 }
697 first_page = low_avail / PAGE_SIZE;
698 #ifdef VM_PHYSSEG_SPARSE
699 size = 0;
700 for (i = 0; i < vm_phys_nsegs; i++)
701 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
702 for (i = 0; phys_avail[i + 1] != 0; i += 2)
703 size += phys_avail[i + 1] - phys_avail[i];
704 #elif defined(VM_PHYSSEG_DENSE)
705 size = high_avail - low_avail;
706 #else
707 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
708 #endif
709
710 #ifdef VM_PHYSSEG_DENSE
711 /*
712 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
713 * the overhead of a page structure per page only if vm_page_array is
714 * allocated from the last physical memory chunk. Otherwise, we must
715 * allocate page structures representing the physical memory
716 * underlying vm_page_array, even though they will not be used.
717 */
718 if (new_end != high_avail)
719 page_range = size / PAGE_SIZE;
720 else
721 #endif
722 {
723 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
724
725 /*
726 * If the partial bytes remaining are large enough for
727 * a page (PAGE_SIZE) without a corresponding
728 * 'struct vm_page', then new_end will contain an
729 * extra page after subtracting the length of the VM
730 * page array. Compensate by subtracting an extra
731 * page from new_end.
732 */
733 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
734 if (new_end == high_avail)
735 high_avail -= PAGE_SIZE;
736 new_end -= PAGE_SIZE;
737 }
738 }
739 end = new_end;
740
741 /*
742 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
743 * However, because this page is allocated from KVM, out-of-bounds
744 * accesses using the direct map will not be trapped.
745 */
746 vaddr += PAGE_SIZE;
747
748 /*
749 * Allocate physical memory for the page structures, and map it.
750 */
751 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
752 mapped = pmap_map(&vaddr, new_end, end,
753 VM_PROT_READ | VM_PROT_WRITE);
754 vm_page_array = (vm_page_t)mapped;
755 vm_page_array_size = page_range;
756
757 #if VM_NRESERVLEVEL > 0
758 /*
759 * Allocate physical memory for the reservation management system's
760 * data structures, and map it.
761 */
762 if (high_avail == end)
763 high_avail = new_end;
764 new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
765 #endif
766 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
767 defined(__riscv)
768 /*
769 * Include vm_page_array and vm_reserv_array in a crash dump.
770 */
771 for (pa = new_end; pa < end; pa += PAGE_SIZE)
772 dump_add_page(pa);
773 #endif
774 phys_avail[biggestone + 1] = new_end;
775
776 /*
777 * Add physical memory segments corresponding to the available
778 * physical pages.
779 */
780 for (i = 0; phys_avail[i + 1] != 0; i += 2)
781 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
782
783 /*
784 * Initialize the physical memory allocator.
785 */
786 vm_phys_init();
787
788 /*
789 * Initialize the page structures and add every available page to the
790 * physical memory allocator's free lists.
791 */
792 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
793 for (ii = 0; ii < vm_page_array_size; ii++) {
794 m = &vm_page_array[ii];
795 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
796 m->flags = PG_FICTITIOUS;
797 }
798 #endif
799 vm_cnt.v_page_count = 0;
800 for (segind = 0; segind < vm_phys_nsegs; segind++) {
801 seg = &vm_phys_segs[segind];
802 for (m = seg->first_page, pa = seg->start; pa < seg->end;
803 m++, pa += PAGE_SIZE)
804 vm_page_init_page(m, pa, segind);
805
806 /*
807 * Add the segment to the free lists only if it is covered by
808 * one of the ranges in phys_avail. Because we've added the
809 * ranges to the vm_phys_segs array, we can assume that each
810 * segment is either entirely contained in one of the ranges,
811 * or doesn't overlap any of them.
812 */
813 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
814 struct vm_domain *vmd;
815
816 if (seg->start < phys_avail[i] ||
817 seg->end > phys_avail[i + 1])
818 continue;
819
820 m = seg->first_page;
821 pagecount = (u_long)atop(seg->end - seg->start);
822
823 vmd = VM_DOMAIN(seg->domain);
824 vm_domain_free_lock(vmd);
825 vm_phys_free_contig(m, pagecount);
826 vm_domain_free_unlock(vmd);
827 vm_domain_freecnt_inc(vmd, pagecount);
828 vm_cnt.v_page_count += (u_int)pagecount;
829
830 vmd = VM_DOMAIN(seg->domain);
831 vmd->vmd_page_count += (u_int)pagecount;
832 vmd->vmd_segs |= 1UL << m->segind;
833 break;
834 }
835 }
836
837 /*
838 * Remove blacklisted pages from the physical memory allocator.
839 */
840 TAILQ_INIT(&blacklist_head);
841 vm_page_blacklist_load(&list, &listend);
842 vm_page_blacklist_check(list, listend);
843
844 list = kern_getenv("vm.blacklist");
845 vm_page_blacklist_check(list, NULL);
846
847 freeenv(list);
848 #if VM_NRESERVLEVEL > 0
849 /*
850 * Initialize the reservation management system.
851 */
852 vm_reserv_init();
853 #endif
854
855 return (vaddr);
856 }
857
858 void
859 vm_page_reference(vm_page_t m)
860 {
861
862 vm_page_aflag_set(m, PGA_REFERENCED);
863 }
864
865 /*
866 * vm_page_busy_downgrade:
867 *
868 * Downgrade an exclusive busy page into a single shared busy page.
869 */
870 void
871 vm_page_busy_downgrade(vm_page_t m)
872 {
873 u_int x;
874 bool locked;
875
876 vm_page_assert_xbusied(m);
877 locked = mtx_owned(vm_page_lockptr(m));
878
879 for (;;) {
880 x = m->busy_lock;
881 x &= VPB_BIT_WAITERS;
882 if (x != 0 && !locked)
883 vm_page_lock(m);
884 if (atomic_cmpset_rel_int(&m->busy_lock,
885 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
886 break;
887 if (x != 0 && !locked)
888 vm_page_unlock(m);
889 }
890 if (x != 0) {
891 wakeup(m);
892 if (!locked)
893 vm_page_unlock(m);
894 }
895 }
896
897 /*
898 * vm_page_sbusied:
899 *
900 * Return a positive value if the page is shared busied, 0 otherwise.
901 */
902 int
903 vm_page_sbusied(vm_page_t m)
904 {
905 u_int x;
906
907 x = m->busy_lock;
908 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
909 }
910
911 /*
912 * vm_page_sunbusy:
913 *
914 * Shared unbusy a page.
915 */
916 void
917 vm_page_sunbusy(vm_page_t m)
918 {
919 u_int x;
920
921 vm_page_lock_assert(m, MA_NOTOWNED);
922 vm_page_assert_sbusied(m);
923
924 for (;;) {
925 x = m->busy_lock;
926 if (VPB_SHARERS(x) > 1) {
927 if (atomic_cmpset_int(&m->busy_lock, x,
928 x - VPB_ONE_SHARER))
929 break;
930 continue;
931 }
932 if ((x & VPB_BIT_WAITERS) == 0) {
933 KASSERT(x == VPB_SHARERS_WORD(1),
934 ("vm_page_sunbusy: invalid lock state"));
935 if (atomic_cmpset_int(&m->busy_lock,
936 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
937 break;
938 continue;
939 }
940 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
941 ("vm_page_sunbusy: invalid lock state for waiters"));
942
943 vm_page_lock(m);
944 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
945 vm_page_unlock(m);
946 continue;
947 }
948 wakeup(m);
949 vm_page_unlock(m);
950 break;
951 }
952 }
953
954 /*
955 * vm_page_busy_sleep:
956 *
957 * Sleep and release the page lock, using the page pointer as wchan.
958 * This is used to implement the hard-path of busying mechanism.
959 *
960 * The given page must be locked.
961 *
962 * If nonshared is true, sleep only if the page is xbusy.
963 */
964 void
965 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
966 {
967 u_int x;
968
969 vm_page_assert_locked(m);
970
971 x = m->busy_lock;
972 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
973 ((x & VPB_BIT_WAITERS) == 0 &&
974 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
975 vm_page_unlock(m);
976 return;
977 }
978 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
979 }
980
981 /*
982 * vm_page_trysbusy:
983 *
984 * Try to shared busy a page.
985 * If the operation succeeds 1 is returned otherwise 0.
986 * The operation never sleeps.
987 */
988 int
989 vm_page_trysbusy(vm_page_t m)
990 {
991 u_int x;
992
993 for (;;) {
994 x = m->busy_lock;
995 if ((x & VPB_BIT_SHARED) == 0)
996 return (0);
997 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
998 return (1);
999 }
1000 }
1001
1002 static void
1003 vm_page_xunbusy_locked(vm_page_t m)
1004 {
1005
1006 vm_page_assert_xbusied(m);
1007 vm_page_assert_locked(m);
1008
1009 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1010 /* There is a waiter, do wakeup() instead of vm_page_flash(). */
1011 wakeup(m);
1012 }
1013
1014 void
1015 vm_page_xunbusy_maybelocked(vm_page_t m)
1016 {
1017 bool lockacq;
1018
1019 vm_page_assert_xbusied(m);
1020
1021 /*
1022 * Fast path for unbusy. If it succeeds, we know that there
1023 * are no waiters, so we do not need a wakeup.
1024 */
1025 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
1026 VPB_UNBUSIED))
1027 return;
1028
1029 lockacq = !mtx_owned(vm_page_lockptr(m));
1030 if (lockacq)
1031 vm_page_lock(m);
1032 vm_page_xunbusy_locked(m);
1033 if (lockacq)
1034 vm_page_unlock(m);
1035 }
1036
1037 /*
1038 * vm_page_xunbusy_hard:
1039 *
1040 * Called after the first try the exclusive unbusy of a page failed.
1041 * It is assumed that the waiters bit is on.
1042 */
1043 void
1044 vm_page_xunbusy_hard(vm_page_t m)
1045 {
1046
1047 vm_page_assert_xbusied(m);
1048
1049 vm_page_lock(m);
1050 vm_page_xunbusy_locked(m);
1051 vm_page_unlock(m);
1052 }
1053
1054 /*
1055 * vm_page_flash:
1056 *
1057 * Wakeup anyone waiting for the page.
1058 * The ownership bits do not change.
1059 *
1060 * The given page must be locked.
1061 */
1062 void
1063 vm_page_flash(vm_page_t m)
1064 {
1065 u_int x;
1066
1067 vm_page_lock_assert(m, MA_OWNED);
1068
1069 for (;;) {
1070 x = m->busy_lock;
1071 if ((x & VPB_BIT_WAITERS) == 0)
1072 return;
1073 if (atomic_cmpset_int(&m->busy_lock, x,
1074 x & (~VPB_BIT_WAITERS)))
1075 break;
1076 }
1077 wakeup(m);
1078 }
1079
1080 /*
1081 * Avoid releasing and reacquiring the same page lock.
1082 */
1083 void
1084 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1085 {
1086 struct mtx *mtx1;
1087
1088 mtx1 = vm_page_lockptr(m);
1089 if (*mtx == mtx1)
1090 return;
1091 if (*mtx != NULL)
1092 mtx_unlock(*mtx);
1093 *mtx = mtx1;
1094 mtx_lock(mtx1);
1095 }
1096
1097 /*
1098 * Keep page from being freed by the page daemon
1099 * much of the same effect as wiring, except much lower
1100 * overhead and should be used only for *very* temporary
1101 * holding ("wiring").
1102 */
1103 void
1104 vm_page_hold(vm_page_t mem)
1105 {
1106
1107 vm_page_lock_assert(mem, MA_OWNED);
1108 mem->hold_count++;
1109 }
1110
1111 void
1112 vm_page_unhold(vm_page_t mem)
1113 {
1114
1115 vm_page_lock_assert(mem, MA_OWNED);
1116 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1117 --mem->hold_count;
1118 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1119 vm_page_free_toq(mem);
1120 }
1121
1122 /*
1123 * vm_page_unhold_pages:
1124 *
1125 * Unhold each of the pages that is referenced by the given array.
1126 */
1127 void
1128 vm_page_unhold_pages(vm_page_t *ma, int count)
1129 {
1130 struct mtx *mtx;
1131
1132 mtx = NULL;
1133 for (; count != 0; count--) {
1134 vm_page_change_lock(*ma, &mtx);
1135 vm_page_unhold(*ma);
1136 ma++;
1137 }
1138 if (mtx != NULL)
1139 mtx_unlock(mtx);
1140 }
1141
1142 vm_page_t
1143 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1144 {
1145 vm_page_t m;
1146
1147 #ifdef VM_PHYSSEG_SPARSE
1148 m = vm_phys_paddr_to_vm_page(pa);
1149 if (m == NULL)
1150 m = vm_phys_fictitious_to_vm_page(pa);
1151 return (m);
1152 #elif defined(VM_PHYSSEG_DENSE)
1153 long pi;
1154
1155 pi = atop(pa);
1156 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1157 m = &vm_page_array[pi - first_page];
1158 return (m);
1159 }
1160 return (vm_phys_fictitious_to_vm_page(pa));
1161 #else
1162 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1163 #endif
1164 }
1165
1166 /*
1167 * vm_page_getfake:
1168 *
1169 * Create a fictitious page with the specified physical address and
1170 * memory attribute. The memory attribute is the only the machine-
1171 * dependent aspect of a fictitious page that must be initialized.
1172 */
1173 vm_page_t
1174 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1175 {
1176 vm_page_t m;
1177
1178 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1179 vm_page_initfake(m, paddr, memattr);
1180 return (m);
1181 }
1182
1183 void
1184 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1185 {
1186
1187 if ((m->flags & PG_FICTITIOUS) != 0) {
1188 /*
1189 * The page's memattr might have changed since the
1190 * previous initialization. Update the pmap to the
1191 * new memattr.
1192 */
1193 goto memattr;
1194 }
1195 m->phys_addr = paddr;
1196 m->queue = PQ_NONE;
1197 /* Fictitious pages don't use "segind". */
1198 m->flags = PG_FICTITIOUS;
1199 /* Fictitious pages don't use "order" or "pool". */
1200 m->oflags = VPO_UNMANAGED;
1201 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1202 m->wire_count = 1;
1203 pmap_page_init(m);
1204 memattr:
1205 pmap_page_set_memattr(m, memattr);
1206 }
1207
1208 /*
1209 * vm_page_putfake:
1210 *
1211 * Release a fictitious page.
1212 */
1213 void
1214 vm_page_putfake(vm_page_t m)
1215 {
1216
1217 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1218 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1219 ("vm_page_putfake: bad page %p", m));
1220 uma_zfree(fakepg_zone, m);
1221 }
1222
1223 /*
1224 * vm_page_updatefake:
1225 *
1226 * Update the given fictitious page to the specified physical address and
1227 * memory attribute.
1228 */
1229 void
1230 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1231 {
1232
1233 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1234 ("vm_page_updatefake: bad page %p", m));
1235 m->phys_addr = paddr;
1236 pmap_page_set_memattr(m, memattr);
1237 }
1238
1239 /*
1240 * vm_page_free:
1241 *
1242 * Free a page.
1243 */
1244 void
1245 vm_page_free(vm_page_t m)
1246 {
1247
1248 m->flags &= ~PG_ZERO;
1249 vm_page_free_toq(m);
1250 }
1251
1252 /*
1253 * vm_page_free_zero:
1254 *
1255 * Free a page to the zerod-pages queue
1256 */
1257 void
1258 vm_page_free_zero(vm_page_t m)
1259 {
1260
1261 m->flags |= PG_ZERO;
1262 vm_page_free_toq(m);
1263 }
1264
1265 /*
1266 * Unbusy and handle the page queueing for a page from a getpages request that
1267 * was optionally read ahead or behind.
1268 */
1269 void
1270 vm_page_readahead_finish(vm_page_t m)
1271 {
1272
1273 /* We shouldn't put invalid pages on queues. */
1274 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1275
1276 /*
1277 * Since the page is not the actually needed one, whether it should
1278 * be activated or deactivated is not obvious. Empirical results
1279 * have shown that deactivating the page is usually the best choice,
1280 * unless the page is wanted by another thread.
1281 */
1282 vm_page_lock(m);
1283 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1284 vm_page_activate(m);
1285 else
1286 vm_page_deactivate(m);
1287 vm_page_unlock(m);
1288 vm_page_xunbusy(m);
1289 }
1290
1291 /*
1292 * vm_page_sleep_if_busy:
1293 *
1294 * Sleep and release the page queues lock if the page is busied.
1295 * Returns TRUE if the thread slept.
1296 *
1297 * The given page must be unlocked and object containing it must
1298 * be locked.
1299 */
1300 int
1301 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1302 {
1303 vm_object_t obj;
1304
1305 vm_page_lock_assert(m, MA_NOTOWNED);
1306 VM_OBJECT_ASSERT_WLOCKED(m->object);
1307
1308 if (vm_page_busied(m)) {
1309 /*
1310 * The page-specific object must be cached because page
1311 * identity can change during the sleep, causing the
1312 * re-lock of a different object.
1313 * It is assumed that a reference to the object is already
1314 * held by the callers.
1315 */
1316 obj = m->object;
1317 vm_page_lock(m);
1318 VM_OBJECT_WUNLOCK(obj);
1319 vm_page_busy_sleep(m, msg, false);
1320 VM_OBJECT_WLOCK(obj);
1321 return (TRUE);
1322 }
1323 return (FALSE);
1324 }
1325
1326 /*
1327 * vm_page_dirty_KBI: [ internal use only ]
1328 *
1329 * Set all bits in the page's dirty field.
1330 *
1331 * The object containing the specified page must be locked if the
1332 * call is made from the machine-independent layer.
1333 *
1334 * See vm_page_clear_dirty_mask().
1335 *
1336 * This function should only be called by vm_page_dirty().
1337 */
1338 void
1339 vm_page_dirty_KBI(vm_page_t m)
1340 {
1341
1342 /* Refer to this operation by its public name. */
1343 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1344 ("vm_page_dirty: page is invalid!"));
1345 m->dirty = VM_PAGE_BITS_ALL;
1346 }
1347
1348 /*
1349 * vm_page_insert: [ internal use only ]
1350 *
1351 * Inserts the given mem entry into the object and object list.
1352 *
1353 * The object must be locked.
1354 */
1355 int
1356 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1357 {
1358 vm_page_t mpred;
1359
1360 VM_OBJECT_ASSERT_WLOCKED(object);
1361 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1362 return (vm_page_insert_after(m, object, pindex, mpred));
1363 }
1364
1365 /*
1366 * vm_page_insert_after:
1367 *
1368 * Inserts the page "m" into the specified object at offset "pindex".
1369 *
1370 * The page "mpred" must immediately precede the offset "pindex" within
1371 * the specified object.
1372 *
1373 * The object must be locked.
1374 */
1375 static int
1376 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1377 vm_page_t mpred)
1378 {
1379 vm_page_t msucc;
1380
1381 VM_OBJECT_ASSERT_WLOCKED(object);
1382 KASSERT(m->object == NULL,
1383 ("vm_page_insert_after: page already inserted"));
1384 if (mpred != NULL) {
1385 KASSERT(mpred->object == object,
1386 ("vm_page_insert_after: object doesn't contain mpred"));
1387 KASSERT(mpred->pindex < pindex,
1388 ("vm_page_insert_after: mpred doesn't precede pindex"));
1389 msucc = TAILQ_NEXT(mpred, listq);
1390 } else
1391 msucc = TAILQ_FIRST(&object->memq);
1392 if (msucc != NULL)
1393 KASSERT(msucc->pindex > pindex,
1394 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1395
1396 /*
1397 * Record the object/offset pair in this page
1398 */
1399 m->object = object;
1400 m->pindex = pindex;
1401
1402 /*
1403 * Now link into the object's ordered list of backed pages.
1404 */
1405 if (vm_radix_insert(&object->rtree, m)) {
1406 m->object = NULL;
1407 m->pindex = 0;
1408 return (1);
1409 }
1410 vm_page_insert_radixdone(m, object, mpred);
1411 return (0);
1412 }
1413
1414 /*
1415 * vm_page_insert_radixdone:
1416 *
1417 * Complete page "m" insertion into the specified object after the
1418 * radix trie hooking.
1419 *
1420 * The page "mpred" must precede the offset "m->pindex" within the
1421 * specified object.
1422 *
1423 * The object must be locked.
1424 */
1425 static void
1426 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1427 {
1428
1429 VM_OBJECT_ASSERT_WLOCKED(object);
1430 KASSERT(object != NULL && m->object == object,
1431 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1432 if (mpred != NULL) {
1433 KASSERT(mpred->object == object,
1434 ("vm_page_insert_after: object doesn't contain mpred"));
1435 KASSERT(mpred->pindex < m->pindex,
1436 ("vm_page_insert_after: mpred doesn't precede pindex"));
1437 }
1438
1439 if (mpred != NULL)
1440 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1441 else
1442 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1443
1444 /*
1445 * Show that the object has one more resident page.
1446 */
1447 object->resident_page_count++;
1448
1449 /*
1450 * Hold the vnode until the last page is released.
1451 */
1452 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1453 vhold(object->handle);
1454
1455 /*
1456 * Since we are inserting a new and possibly dirty page,
1457 * update the object's OBJ_MIGHTBEDIRTY flag.
1458 */
1459 if (pmap_page_is_write_mapped(m))
1460 vm_object_set_writeable_dirty(object);
1461 }
1462
1463 /*
1464 * vm_page_remove:
1465 *
1466 * Removes the specified page from its containing object, but does not
1467 * invalidate any backing storage. Return true if the page may be safely
1468 * freed and false otherwise.
1469 *
1470 * The object must be locked. The page must be locked if it is managed.
1471 */
1472 bool
1473 vm_page_remove(vm_page_t m)
1474 {
1475 vm_object_t object;
1476 vm_page_t mrem;
1477
1478 object = m->object;
1479
1480 if ((m->oflags & VPO_UNMANAGED) == 0)
1481 vm_page_assert_locked(m);
1482 VM_OBJECT_ASSERT_WLOCKED(object);
1483 if (vm_page_xbusied(m))
1484 vm_page_xunbusy_maybelocked(m);
1485 mrem = vm_radix_remove(&object->rtree, m->pindex);
1486 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1487
1488 /*
1489 * Now remove from the object's list of backed pages.
1490 */
1491 TAILQ_REMOVE(&object->memq, m, listq);
1492
1493 /*
1494 * And show that the object has one fewer resident page.
1495 */
1496 object->resident_page_count--;
1497
1498 /*
1499 * The vnode may now be recycled.
1500 */
1501 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1502 vdrop(object->handle);
1503
1504 m->object = NULL;
1505 return (!vm_page_wired(m));
1506 }
1507
1508 /*
1509 * vm_page_lookup:
1510 *
1511 * Returns the page associated with the object/offset
1512 * pair specified; if none is found, NULL is returned.
1513 *
1514 * The object must be locked.
1515 */
1516 vm_page_t
1517 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1518 {
1519
1520 VM_OBJECT_ASSERT_LOCKED(object);
1521 return (vm_radix_lookup(&object->rtree, pindex));
1522 }
1523
1524 /*
1525 * vm_page_find_least:
1526 *
1527 * Returns the page associated with the object with least pindex
1528 * greater than or equal to the parameter pindex, or NULL.
1529 *
1530 * The object must be locked.
1531 */
1532 vm_page_t
1533 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1534 {
1535 vm_page_t m;
1536
1537 VM_OBJECT_ASSERT_LOCKED(object);
1538 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1539 m = vm_radix_lookup_ge(&object->rtree, pindex);
1540 return (m);
1541 }
1542
1543 /*
1544 * Returns the given page's successor (by pindex) within the object if it is
1545 * resident; if none is found, NULL is returned.
1546 *
1547 * The object must be locked.
1548 */
1549 vm_page_t
1550 vm_page_next(vm_page_t m)
1551 {
1552 vm_page_t next;
1553
1554 VM_OBJECT_ASSERT_LOCKED(m->object);
1555 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1556 MPASS(next->object == m->object);
1557 if (next->pindex != m->pindex + 1)
1558 next = NULL;
1559 }
1560 return (next);
1561 }
1562
1563 /*
1564 * Returns the given page's predecessor (by pindex) within the object if it is
1565 * resident; if none is found, NULL is returned.
1566 *
1567 * The object must be locked.
1568 */
1569 vm_page_t
1570 vm_page_prev(vm_page_t m)
1571 {
1572 vm_page_t prev;
1573
1574 VM_OBJECT_ASSERT_LOCKED(m->object);
1575 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1576 MPASS(prev->object == m->object);
1577 if (prev->pindex != m->pindex - 1)
1578 prev = NULL;
1579 }
1580 return (prev);
1581 }
1582
1583 /*
1584 * Uses the page mnew as a replacement for an existing page at index
1585 * pindex which must be already present in the object.
1586 *
1587 * The existing page must not be on a paging queue.
1588 */
1589 vm_page_t
1590 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1591 {
1592 vm_page_t mold;
1593
1594 VM_OBJECT_ASSERT_WLOCKED(object);
1595 KASSERT(mnew->object == NULL,
1596 ("vm_page_replace: page %p already in object", mnew));
1597 KASSERT(mnew->queue == PQ_NONE,
1598 ("vm_page_replace: new page %p is on a paging queue", mnew));
1599
1600 /*
1601 * This function mostly follows vm_page_insert() and
1602 * vm_page_remove() without the radix, object count and vnode
1603 * dance. Double check such functions for more comments.
1604 */
1605
1606 mnew->object = object;
1607 mnew->pindex = pindex;
1608 mold = vm_radix_replace(&object->rtree, mnew);
1609 KASSERT(mold->queue == PQ_NONE,
1610 ("vm_page_replace: old page %p is on a paging queue", mold));
1611
1612 /* Keep the resident page list in sorted order. */
1613 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1614 TAILQ_REMOVE(&object->memq, mold, listq);
1615
1616 mold->object = NULL;
1617 vm_page_xunbusy_maybelocked(mold);
1618
1619 /*
1620 * The object's resident_page_count does not change because we have
1621 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1622 */
1623 if (pmap_page_is_write_mapped(mnew))
1624 vm_object_set_writeable_dirty(object);
1625 return (mold);
1626 }
1627
1628 /*
1629 * vm_page_rename:
1630 *
1631 * Move the given memory entry from its
1632 * current object to the specified target object/offset.
1633 *
1634 * Note: swap associated with the page must be invalidated by the move. We
1635 * have to do this for several reasons: (1) we aren't freeing the
1636 * page, (2) we are dirtying the page, (3) the VM system is probably
1637 * moving the page from object A to B, and will then later move
1638 * the backing store from A to B and we can't have a conflict.
1639 *
1640 * Note: we *always* dirty the page. It is necessary both for the
1641 * fact that we moved it, and because we may be invalidating
1642 * swap.
1643 *
1644 * The objects must be locked.
1645 */
1646 int
1647 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1648 {
1649 vm_page_t mpred;
1650 vm_pindex_t opidx;
1651
1652 VM_OBJECT_ASSERT_WLOCKED(new_object);
1653
1654 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1655 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1656 ("vm_page_rename: pindex already renamed"));
1657
1658 /*
1659 * Create a custom version of vm_page_insert() which does not depend
1660 * by m_prev and can cheat on the implementation aspects of the
1661 * function.
1662 */
1663 opidx = m->pindex;
1664 m->pindex = new_pindex;
1665 if (vm_radix_insert(&new_object->rtree, m)) {
1666 m->pindex = opidx;
1667 return (1);
1668 }
1669
1670 /*
1671 * The operation cannot fail anymore. The removal must happen before
1672 * the listq iterator is tainted.
1673 */
1674 m->pindex = opidx;
1675 vm_page_lock(m);
1676 (void)vm_page_remove(m);
1677
1678 /* Return back to the new pindex to complete vm_page_insert(). */
1679 m->pindex = new_pindex;
1680 m->object = new_object;
1681 vm_page_unlock(m);
1682 vm_page_insert_radixdone(m, new_object, mpred);
1683 vm_page_dirty(m);
1684 return (0);
1685 }
1686
1687 /*
1688 * vm_page_alloc:
1689 *
1690 * Allocate and return a page that is associated with the specified
1691 * object and offset pair. By default, this page is exclusive busied.
1692 *
1693 * The caller must always specify an allocation class.
1694 *
1695 * allocation classes:
1696 * VM_ALLOC_NORMAL normal process request
1697 * VM_ALLOC_SYSTEM system *really* needs a page
1698 * VM_ALLOC_INTERRUPT interrupt time request
1699 *
1700 * optional allocation flags:
1701 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1702 * intends to allocate
1703 * VM_ALLOC_NOBUSY do not exclusive busy the page
1704 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1705 * VM_ALLOC_NOOBJ page is not associated with an object and
1706 * should not be exclusive busy
1707 * VM_ALLOC_SBUSY shared busy the allocated page
1708 * VM_ALLOC_WIRED wire the allocated page
1709 * VM_ALLOC_ZERO prefer a zeroed page
1710 */
1711 vm_page_t
1712 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1713 {
1714
1715 return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1716 vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1717 }
1718
1719 vm_page_t
1720 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1721 int req)
1722 {
1723
1724 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1725 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1726 NULL));
1727 }
1728
1729 /*
1730 * Allocate a page in the specified object with the given page index. To
1731 * optimize insertion of the page into the object, the caller must also specifiy
1732 * the resident page in the object with largest index smaller than the given
1733 * page index, or NULL if no such page exists.
1734 */
1735 vm_page_t
1736 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1737 int req, vm_page_t mpred)
1738 {
1739 struct vm_domainset_iter di;
1740 vm_page_t m;
1741 int domain;
1742
1743 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1744 do {
1745 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1746 mpred);
1747 if (m != NULL)
1748 break;
1749 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1750
1751 return (m);
1752 }
1753
1754 /*
1755 * Returns true if the number of free pages exceeds the minimum
1756 * for the request class and false otherwise.
1757 */
1758 static int
1759 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1760 {
1761 u_int limit, old, new;
1762
1763 if (req_class == VM_ALLOC_INTERRUPT)
1764 limit = 0;
1765 else if (req_class == VM_ALLOC_SYSTEM)
1766 limit = vmd->vmd_interrupt_free_min;
1767 else
1768 limit = vmd->vmd_free_reserved;
1769
1770 /*
1771 * Attempt to reserve the pages. Fail if we're below the limit.
1772 */
1773 limit += npages;
1774 old = vmd->vmd_free_count;
1775 do {
1776 if (old < limit)
1777 return (0);
1778 new = old - npages;
1779 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1780
1781 /* Wake the page daemon if we've crossed the threshold. */
1782 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1783 pagedaemon_wakeup(vmd->vmd_domain);
1784
1785 /* Only update bitsets on transitions. */
1786 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1787 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1788 vm_domain_set(vmd);
1789
1790 return (1);
1791 }
1792
1793 int
1794 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1795 {
1796 int req_class;
1797
1798 /*
1799 * The page daemon is allowed to dig deeper into the free page list.
1800 */
1801 req_class = req & VM_ALLOC_CLASS_MASK;
1802 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1803 req_class = VM_ALLOC_SYSTEM;
1804 return (_vm_domain_allocate(vmd, req_class, npages));
1805 }
1806
1807 vm_page_t
1808 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1809 int req, vm_page_t mpred)
1810 {
1811 struct vm_domain *vmd;
1812 vm_page_t m;
1813 int flags, pool;
1814
1815 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1816 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1817 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1818 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1819 ("inconsistent object(%p)/req(%x)", object, req));
1820 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1821 ("Can't sleep and retry object insertion."));
1822 KASSERT(mpred == NULL || mpred->pindex < pindex,
1823 ("mpred %p doesn't precede pindex 0x%jx", mpred,
1824 (uintmax_t)pindex));
1825 if (object != NULL)
1826 VM_OBJECT_ASSERT_WLOCKED(object);
1827
1828 flags = 0;
1829 m = NULL;
1830 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1831 again:
1832 #if VM_NRESERVLEVEL > 0
1833 /*
1834 * Can we allocate the page from a reservation?
1835 */
1836 if (vm_object_reserv(object) &&
1837 ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL ||
1838 (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) {
1839 domain = vm_phys_domain(m);
1840 vmd = VM_DOMAIN(domain);
1841 goto found;
1842 }
1843 #endif
1844 vmd = VM_DOMAIN(domain);
1845 if (vmd->vmd_pgcache[pool].zone != NULL) {
1846 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1847 if (m != NULL) {
1848 flags |= PG_PCPU_CACHE;
1849 goto found;
1850 }
1851 }
1852 if (vm_domain_allocate(vmd, req, 1)) {
1853 /*
1854 * If not, allocate it from the free page queues.
1855 */
1856 vm_domain_free_lock(vmd);
1857 m = vm_phys_alloc_pages(domain, pool, 0);
1858 vm_domain_free_unlock(vmd);
1859 if (m == NULL) {
1860 vm_domain_freecnt_inc(vmd, 1);
1861 #if VM_NRESERVLEVEL > 0
1862 if (vm_reserv_reclaim_inactive(domain))
1863 goto again;
1864 #endif
1865 }
1866 }
1867 if (m == NULL) {
1868 /*
1869 * Not allocatable, give up.
1870 */
1871 if (vm_domain_alloc_fail(vmd, object, req))
1872 goto again;
1873 return (NULL);
1874 }
1875
1876 /*
1877 * At this point we had better have found a good page.
1878 */
1879 found:
1880 vm_page_dequeue(m);
1881 vm_page_alloc_check(m);
1882
1883 /*
1884 * Initialize the page. Only the PG_ZERO flag is inherited.
1885 */
1886 if ((req & VM_ALLOC_ZERO) != 0)
1887 flags |= (m->flags & PG_ZERO);
1888 if ((req & VM_ALLOC_NODUMP) != 0)
1889 flags |= PG_NODUMP;
1890 m->flags = flags;
1891 m->aflags = 0;
1892 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1893 VPO_UNMANAGED : 0;
1894 m->busy_lock = VPB_UNBUSIED;
1895 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1896 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1897 if ((req & VM_ALLOC_SBUSY) != 0)
1898 m->busy_lock = VPB_SHARERS_WORD(1);
1899 if (req & VM_ALLOC_WIRED) {
1900 /*
1901 * The page lock is not required for wiring a page until that
1902 * page is inserted into the object.
1903 */
1904 vm_wire_add(1);
1905 m->wire_count = 1;
1906 }
1907 m->act_count = 0;
1908
1909 if (object != NULL) {
1910 if (vm_page_insert_after(m, object, pindex, mpred)) {
1911 if (req & VM_ALLOC_WIRED) {
1912 vm_wire_sub(1);
1913 m->wire_count = 0;
1914 }
1915 KASSERT(m->object == NULL, ("page %p has object", m));
1916 m->oflags = VPO_UNMANAGED;
1917 m->busy_lock = VPB_UNBUSIED;
1918 /* Don't change PG_ZERO. */
1919 vm_page_free_toq(m);
1920 if (req & VM_ALLOC_WAITFAIL) {
1921 VM_OBJECT_WUNLOCK(object);
1922 vm_radix_wait();
1923 VM_OBJECT_WLOCK(object);
1924 }
1925 return (NULL);
1926 }
1927
1928 /* Ignore device objects; the pager sets "memattr" for them. */
1929 if (object->memattr != VM_MEMATTR_DEFAULT &&
1930 (object->flags & OBJ_FICTITIOUS) == 0)
1931 pmap_page_set_memattr(m, object->memattr);
1932 } else
1933 m->pindex = pindex;
1934
1935 return (m);
1936 }
1937
1938 /*
1939 * vm_page_alloc_contig:
1940 *
1941 * Allocate a contiguous set of physical pages of the given size "npages"
1942 * from the free lists. All of the physical pages must be at or above
1943 * the given physical address "low" and below the given physical address
1944 * "high". The given value "alignment" determines the alignment of the
1945 * first physical page in the set. If the given value "boundary" is
1946 * non-zero, then the set of physical pages cannot cross any physical
1947 * address boundary that is a multiple of that value. Both "alignment"
1948 * and "boundary" must be a power of two.
1949 *
1950 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1951 * then the memory attribute setting for the physical pages is configured
1952 * to the object's memory attribute setting. Otherwise, the memory
1953 * attribute setting for the physical pages is configured to "memattr",
1954 * overriding the object's memory attribute setting. However, if the
1955 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1956 * memory attribute setting for the physical pages cannot be configured
1957 * to VM_MEMATTR_DEFAULT.
1958 *
1959 * The specified object may not contain fictitious pages.
1960 *
1961 * The caller must always specify an allocation class.
1962 *
1963 * allocation classes:
1964 * VM_ALLOC_NORMAL normal process request
1965 * VM_ALLOC_SYSTEM system *really* needs a page
1966 * VM_ALLOC_INTERRUPT interrupt time request
1967 *
1968 * optional allocation flags:
1969 * VM_ALLOC_NOBUSY do not exclusive busy the page
1970 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1971 * VM_ALLOC_NOOBJ page is not associated with an object and
1972 * should not be exclusive busy
1973 * VM_ALLOC_SBUSY shared busy the allocated page
1974 * VM_ALLOC_WIRED wire the allocated page
1975 * VM_ALLOC_ZERO prefer a zeroed page
1976 */
1977 vm_page_t
1978 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1979 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1980 vm_paddr_t boundary, vm_memattr_t memattr)
1981 {
1982 struct vm_domainset_iter di;
1983 vm_page_t m;
1984 int domain;
1985
1986 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1987 do {
1988 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1989 npages, low, high, alignment, boundary, memattr);
1990 if (m != NULL)
1991 break;
1992 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1993
1994 return (m);
1995 }
1996
1997 vm_page_t
1998 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1999 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2000 vm_paddr_t boundary, vm_memattr_t memattr)
2001 {
2002 struct vm_domain *vmd;
2003 vm_page_t m, m_ret, mpred;
2004 u_int busy_lock, flags, oflags;
2005
2006 mpred = NULL; /* XXX: pacify gcc */
2007 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2008 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2009 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2010 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2011 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2012 req));
2013 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2014 ("Can't sleep and retry object insertion."));
2015 if (object != NULL) {
2016 VM_OBJECT_ASSERT_WLOCKED(object);
2017 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2018 ("vm_page_alloc_contig: object %p has fictitious pages",
2019 object));
2020 }
2021 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2022
2023 if (object != NULL) {
2024 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2025 KASSERT(mpred == NULL || mpred->pindex != pindex,
2026 ("vm_page_alloc_contig: pindex already allocated"));
2027 }
2028
2029 /*
2030 * Can we allocate the pages without the number of free pages falling
2031 * below the lower bound for the allocation class?
2032 */
2033 m_ret = NULL;
2034 again:
2035 #if VM_NRESERVLEVEL > 0
2036 /*
2037 * Can we allocate the pages from a reservation?
2038 */
2039 if (vm_object_reserv(object) &&
2040 ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
2041 npages, low, high, alignment, boundary, mpred)) != NULL ||
2042 (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain,
2043 npages, low, high, alignment, boundary, mpred)) != NULL)) {
2044 domain = vm_phys_domain(m_ret);
2045 vmd = VM_DOMAIN(domain);
2046 goto found;
2047 }
2048 #endif
2049 vmd = VM_DOMAIN(domain);
2050 if (vm_domain_allocate(vmd, req, npages)) {
2051 /*
2052 * allocate them from the free page queues.
2053 */
2054 vm_domain_free_lock(vmd);
2055 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2056 alignment, boundary);
2057 vm_domain_free_unlock(vmd);
2058 if (m_ret == NULL) {
2059 vm_domain_freecnt_inc(vmd, npages);
2060 #if VM_NRESERVLEVEL > 0
2061 if (vm_reserv_reclaim_contig(domain, npages, low,
2062 high, alignment, boundary))
2063 goto again;
2064 #endif
2065 }
2066 }
2067 if (m_ret == NULL) {
2068 if (vm_domain_alloc_fail(vmd, object, req))
2069 goto again;
2070 return (NULL);
2071 }
2072 #if VM_NRESERVLEVEL > 0
2073 found:
2074 #endif
2075 for (m = m_ret; m < &m_ret[npages]; m++) {
2076 vm_page_dequeue(m);
2077 vm_page_alloc_check(m);
2078 }
2079
2080 /*
2081 * Initialize the pages. Only the PG_ZERO flag is inherited.
2082 */
2083 flags = 0;
2084 if ((req & VM_ALLOC_ZERO) != 0)
2085 flags = PG_ZERO;
2086 if ((req & VM_ALLOC_NODUMP) != 0)
2087 flags |= PG_NODUMP;
2088 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2089 VPO_UNMANAGED : 0;
2090 busy_lock = VPB_UNBUSIED;
2091 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2092 busy_lock = VPB_SINGLE_EXCLUSIVER;
2093 if ((req & VM_ALLOC_SBUSY) != 0)
2094 busy_lock = VPB_SHARERS_WORD(1);
2095 if ((req & VM_ALLOC_WIRED) != 0)
2096 vm_wire_add(npages);
2097 if (object != NULL) {
2098 if (object->memattr != VM_MEMATTR_DEFAULT &&
2099 memattr == VM_MEMATTR_DEFAULT)
2100 memattr = object->memattr;
2101 }
2102 for (m = m_ret; m < &m_ret[npages]; m++) {
2103 m->aflags = 0;
2104 m->flags = (m->flags | PG_NODUMP) & flags;
2105 m->busy_lock = busy_lock;
2106 if ((req & VM_ALLOC_WIRED) != 0)
2107 m->wire_count = 1;
2108 m->act_count = 0;
2109 m->oflags = oflags;
2110 if (object != NULL) {
2111 if (vm_page_insert_after(m, object, pindex, mpred)) {
2112 if ((req & VM_ALLOC_WIRED) != 0)
2113 vm_wire_sub(npages);
2114 KASSERT(m->object == NULL,
2115 ("page %p has object", m));
2116 mpred = m;
2117 for (m = m_ret; m < &m_ret[npages]; m++) {
2118 if (m <= mpred &&
2119 (req & VM_ALLOC_WIRED) != 0)
2120 m->wire_count = 0;
2121 m->oflags = VPO_UNMANAGED;
2122 m->busy_lock = VPB_UNBUSIED;
2123 /* Don't change PG_ZERO. */
2124 vm_page_free_toq(m);
2125 }
2126 if (req & VM_ALLOC_WAITFAIL) {
2127 VM_OBJECT_WUNLOCK(object);
2128 vm_radix_wait();
2129 VM_OBJECT_WLOCK(object);
2130 }
2131 return (NULL);
2132 }
2133 mpred = m;
2134 } else
2135 m->pindex = pindex;
2136 if (memattr != VM_MEMATTR_DEFAULT)
2137 pmap_page_set_memattr(m, memattr);
2138 pindex++;
2139 }
2140 return (m_ret);
2141 }
2142
2143 /*
2144 * Check a page that has been freshly dequeued from a freelist.
2145 */
2146 static void
2147 vm_page_alloc_check(vm_page_t m)
2148 {
2149
2150 KASSERT(m->object == NULL, ("page %p has object", m));
2151 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
2152 ("page %p has unexpected queue %d, flags %#x",
2153 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK)));
2154 KASSERT(!vm_page_held(m), ("page %p is held", m));
2155 KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2156 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2157 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2158 ("page %p has unexpected memattr %d",
2159 m, pmap_page_get_memattr(m)));
2160 KASSERT(m->valid == 0, ("free page %p is valid", m));
2161 }
2162
2163 /*
2164 * vm_page_alloc_freelist:
2165 *
2166 * Allocate a physical page from the specified free page list.
2167 *
2168 * The caller must always specify an allocation class.
2169 *
2170 * allocation classes:
2171 * VM_ALLOC_NORMAL normal process request
2172 * VM_ALLOC_SYSTEM system *really* needs a page
2173 * VM_ALLOC_INTERRUPT interrupt time request
2174 *
2175 * optional allocation flags:
2176 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2177 * intends to allocate
2178 * VM_ALLOC_WIRED wire the allocated page
2179 * VM_ALLOC_ZERO prefer a zeroed page
2180 */
2181 vm_page_t
2182 vm_page_alloc_freelist(int freelist, int req)
2183 {
2184 struct vm_domainset_iter di;
2185 vm_page_t m;
2186 int domain;
2187
2188 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2189 do {
2190 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2191 if (m != NULL)
2192 break;
2193 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2194
2195 return (m);
2196 }
2197
2198 vm_page_t
2199 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2200 {
2201 struct vm_domain *vmd;
2202 vm_page_t m;
2203 u_int flags;
2204
2205 m = NULL;
2206 vmd = VM_DOMAIN(domain);
2207 again:
2208 if (vm_domain_allocate(vmd, req, 1)) {
2209 vm_domain_free_lock(vmd);
2210 m = vm_phys_alloc_freelist_pages(domain, freelist,
2211 VM_FREEPOOL_DIRECT, 0);
2212 vm_domain_free_unlock(vmd);
2213 if (m == NULL)
2214 vm_domain_freecnt_inc(vmd, 1);
2215 }
2216 if (m == NULL) {
2217 if (vm_domain_alloc_fail(vmd, NULL, req))
2218 goto again;
2219 return (NULL);
2220 }
2221 vm_page_dequeue(m);
2222 vm_page_alloc_check(m);
2223
2224 /*
2225 * Initialize the page. Only the PG_ZERO flag is inherited.
2226 */
2227 m->aflags = 0;
2228 flags = 0;
2229 if ((req & VM_ALLOC_ZERO) != 0)
2230 flags = PG_ZERO;
2231 m->flags &= flags;
2232 if ((req & VM_ALLOC_WIRED) != 0) {
2233 /*
2234 * The page lock is not required for wiring a page that does
2235 * not belong to an object.
2236 */
2237 vm_wire_add(1);
2238 m->wire_count = 1;
2239 }
2240 /* Unmanaged pages don't use "act_count". */
2241 m->oflags = VPO_UNMANAGED;
2242 return (m);
2243 }
2244
2245 static int
2246 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2247 {
2248 struct vm_domain *vmd;
2249 struct vm_pgcache *pgcache;
2250 int i;
2251
2252 pgcache = arg;
2253 vmd = VM_DOMAIN(pgcache->domain);
2254
2255 /*
2256 * The page daemon should avoid creating extra memory pressure since its
2257 * main purpose is to replenish the store of free pages.
2258 */
2259 if (vmd->vmd_severeset || curproc == pageproc ||
2260 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2261 return (0);
2262 domain = vmd->vmd_domain;
2263 vm_domain_free_lock(vmd);
2264 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2265 (vm_page_t *)store);
2266 vm_domain_free_unlock(vmd);
2267 if (cnt != i)
2268 vm_domain_freecnt_inc(vmd, cnt - i);
2269
2270 return (i);
2271 }
2272
2273 static void
2274 vm_page_zone_release(void *arg, void **store, int cnt)
2275 {
2276 struct vm_domain *vmd;
2277 struct vm_pgcache *pgcache;
2278 vm_page_t m;
2279 int i;
2280
2281 pgcache = arg;
2282 vmd = VM_DOMAIN(pgcache->domain);
2283 vm_domain_free_lock(vmd);
2284 for (i = 0; i < cnt; i++) {
2285 m = (vm_page_t)store[i];
2286 vm_phys_free_pages(m, 0);
2287 }
2288 vm_domain_free_unlock(vmd);
2289 vm_domain_freecnt_inc(vmd, cnt);
2290 }
2291
2292 #define VPSC_ANY 0 /* No restrictions. */
2293 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2294 #define VPSC_NOSUPER 2 /* Skip superpages. */
2295
2296 /*
2297 * vm_page_scan_contig:
2298 *
2299 * Scan vm_page_array[] between the specified entries "m_start" and
2300 * "m_end" for a run of contiguous physical pages that satisfy the
2301 * specified conditions, and return the lowest page in the run. The
2302 * specified "alignment" determines the alignment of the lowest physical
2303 * page in the run. If the specified "boundary" is non-zero, then the
2304 * run of physical pages cannot span a physical address that is a
2305 * multiple of "boundary".
2306 *
2307 * "m_end" is never dereferenced, so it need not point to a vm_page
2308 * structure within vm_page_array[].
2309 *
2310 * "npages" must be greater than zero. "m_start" and "m_end" must not
2311 * span a hole (or discontiguity) in the physical address space. Both
2312 * "alignment" and "boundary" must be a power of two.
2313 */
2314 vm_page_t
2315 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2316 u_long alignment, vm_paddr_t boundary, int options)
2317 {
2318 struct mtx *m_mtx;
2319 vm_object_t object;
2320 vm_paddr_t pa;
2321 vm_page_t m, m_run;
2322 #if VM_NRESERVLEVEL > 0
2323 int level;
2324 #endif
2325 int m_inc, order, run_ext, run_len;
2326
2327 KASSERT(npages > 0, ("npages is 0"));
2328 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2329 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2330 m_run = NULL;
2331 run_len = 0;
2332 m_mtx = NULL;
2333 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2334 KASSERT((m->flags & PG_MARKER) == 0,
2335 ("page %p is PG_MARKER", m));
2336 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2337 ("fictitious page %p has invalid wire count", m));
2338
2339 /*
2340 * If the current page would be the start of a run, check its
2341 * physical address against the end, alignment, and boundary
2342 * conditions. If it doesn't satisfy these conditions, either
2343 * terminate the scan or advance to the next page that
2344 * satisfies the failed condition.
2345 */
2346 if (run_len == 0) {
2347 KASSERT(m_run == NULL, ("m_run != NULL"));
2348 if (m + npages > m_end)
2349 break;
2350 pa = VM_PAGE_TO_PHYS(m);
2351 if ((pa & (alignment - 1)) != 0) {
2352 m_inc = atop(roundup2(pa, alignment) - pa);
2353 continue;
2354 }
2355 if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2356 boundary) != 0) {
2357 m_inc = atop(roundup2(pa, boundary) - pa);
2358 continue;
2359 }
2360 } else
2361 KASSERT(m_run != NULL, ("m_run == NULL"));
2362
2363 vm_page_change_lock(m, &m_mtx);
2364 m_inc = 1;
2365 retry:
2366 if (vm_page_held(m))
2367 run_ext = 0;
2368 #if VM_NRESERVLEVEL > 0
2369 else if ((level = vm_reserv_level(m)) >= 0 &&
2370 (options & VPSC_NORESERV) != 0) {
2371 run_ext = 0;
2372 /* Advance to the end of the reservation. */
2373 pa = VM_PAGE_TO_PHYS(m);
2374 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2375 pa);
2376 }
2377 #endif
2378 else if ((object = m->object) != NULL) {
2379 /*
2380 * The page is considered eligible for relocation if
2381 * and only if it could be laundered or reclaimed by
2382 * the page daemon.
2383 */
2384 if (!VM_OBJECT_TRYRLOCK(object)) {
2385 mtx_unlock(m_mtx);
2386 VM_OBJECT_RLOCK(object);
2387 mtx_lock(m_mtx);
2388 if (m->object != object) {
2389 /*
2390 * The page may have been freed.
2391 */
2392 VM_OBJECT_RUNLOCK(object);
2393 goto retry;
2394 } else if (vm_page_held(m)) {
2395 run_ext = 0;
2396 goto unlock;
2397 }
2398 }
2399 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2400 ("page %p is PG_UNHOLDFREE", m));
2401 /* Don't care: PG_NODUMP, PG_ZERO. */
2402 if (object->type != OBJT_DEFAULT &&
2403 object->type != OBJT_SWAP &&
2404 object->type != OBJT_VNODE) {
2405 run_ext = 0;
2406 #if VM_NRESERVLEVEL > 0
2407 } else if ((options & VPSC_NOSUPER) != 0 &&
2408 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2409 run_ext = 0;
2410 /* Advance to the end of the superpage. */
2411 pa = VM_PAGE_TO_PHYS(m);
2412 m_inc = atop(roundup2(pa + 1,
2413 vm_reserv_size(level)) - pa);
2414 #endif
2415 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2416 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2417 /*
2418 * The page is allocated but eligible for
2419 * relocation. Extend the current run by one
2420 * page.
2421 */
2422 KASSERT(pmap_page_get_memattr(m) ==
2423 VM_MEMATTR_DEFAULT,
2424 ("page %p has an unexpected memattr", m));
2425 KASSERT((m->oflags & (VPO_SWAPINPROG |
2426 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2427 ("page %p has unexpected oflags", m));
2428 /* Don't care: VPO_NOSYNC. */
2429 run_ext = 1;
2430 } else
2431 run_ext = 0;
2432 unlock:
2433 VM_OBJECT_RUNLOCK(object);
2434 #if VM_NRESERVLEVEL > 0
2435 } else if (level >= 0) {
2436 /*
2437 * The page is reserved but not yet allocated. In
2438 * other words, it is still free. Extend the current
2439 * run by one page.
2440 */
2441 run_ext = 1;
2442 #endif
2443 } else if ((order = m->order) < VM_NFREEORDER) {
2444 /*
2445 * The page is enqueued in the physical memory
2446 * allocator's free page queues. Moreover, it is the
2447 * first page in a power-of-two-sized run of
2448 * contiguous free pages. Add these pages to the end
2449 * of the current run, and jump ahead.
2450 */
2451 run_ext = 1 << order;
2452 m_inc = 1 << order;
2453 } else {
2454 /*
2455 * Skip the page for one of the following reasons: (1)
2456 * It is enqueued in the physical memory allocator's
2457 * free page queues. However, it is not the first
2458 * page in a run of contiguous free pages. (This case
2459 * rarely occurs because the scan is performed in
2460 * ascending order.) (2) It is not reserved, and it is
2461 * transitioning from free to allocated. (Conversely,
2462 * the transition from allocated to free for managed
2463 * pages is blocked by the page lock.) (3) It is
2464 * allocated but not contained by an object and not
2465 * wired, e.g., allocated by Xen's balloon driver.
2466 */
2467 run_ext = 0;
2468 }
2469
2470 /*
2471 * Extend or reset the current run of pages.
2472 */
2473 if (run_ext > 0) {
2474 if (run_len == 0)
2475 m_run = m;
2476 run_len += run_ext;
2477 } else {
2478 if (run_len > 0) {
2479 m_run = NULL;
2480 run_len = 0;
2481 }
2482 }
2483 }
2484 if (m_mtx != NULL)
2485 mtx_unlock(m_mtx);
2486 if (run_len >= npages)
2487 return (m_run);
2488 return (NULL);
2489 }
2490
2491 /*
2492 * vm_page_reclaim_run:
2493 *
2494 * Try to relocate each of the allocated virtual pages within the
2495 * specified run of physical pages to a new physical address. Free the
2496 * physical pages underlying the relocated virtual pages. A virtual page
2497 * is relocatable if and only if it could be laundered or reclaimed by
2498 * the page daemon. Whenever possible, a virtual page is relocated to a
2499 * physical address above "high".
2500 *
2501 * Returns 0 if every physical page within the run was already free or
2502 * just freed by a successful relocation. Otherwise, returns a non-zero
2503 * value indicating why the last attempt to relocate a virtual page was
2504 * unsuccessful.
2505 *
2506 * "req_class" must be an allocation class.
2507 */
2508 static int
2509 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2510 vm_paddr_t high)
2511 {
2512 struct vm_domain *vmd;
2513 struct mtx *m_mtx;
2514 struct spglist free;
2515 vm_object_t object;
2516 vm_paddr_t pa;
2517 vm_page_t m, m_end, m_new;
2518 int error, order, req;
2519
2520 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2521 ("req_class is not an allocation class"));
2522 SLIST_INIT(&free);
2523 error = 0;
2524 m = m_run;
2525 m_end = m_run + npages;
2526 m_mtx = NULL;
2527 for (; error == 0 && m < m_end; m++) {
2528 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2529 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2530
2531 /*
2532 * Avoid releasing and reacquiring the same page lock.
2533 */
2534 vm_page_change_lock(m, &m_mtx);
2535 retry:
2536 if (vm_page_held(m))
2537 error = EBUSY;
2538 else if ((object = m->object) != NULL) {
2539 /*
2540 * The page is relocated if and only if it could be
2541 * laundered or reclaimed by the page daemon.
2542 */
2543 if (!VM_OBJECT_TRYWLOCK(object)) {
2544 mtx_unlock(m_mtx);
2545 VM_OBJECT_WLOCK(object);
2546 mtx_lock(m_mtx);
2547 if (m->object != object) {
2548 /*
2549 * The page may have been freed.
2550 */
2551 VM_OBJECT_WUNLOCK(object);
2552 goto retry;
2553 } else if (vm_page_held(m)) {
2554 error = EBUSY;
2555 goto unlock;
2556 }
2557 }
2558 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2559 ("page %p is PG_UNHOLDFREE", m));
2560 /* Don't care: PG_NODUMP, PG_ZERO. */
2561 if (object->type != OBJT_DEFAULT &&
2562 object->type != OBJT_SWAP &&
2563 object->type != OBJT_VNODE)
2564 error = EINVAL;
2565 else if (object->memattr != VM_MEMATTR_DEFAULT)
2566 error = EINVAL;
2567 else if (vm_page_queue(m) != PQ_NONE &&
2568 !vm_page_busied(m)) {
2569 KASSERT(pmap_page_get_memattr(m) ==
2570 VM_MEMATTR_DEFAULT,
2571 ("page %p has an unexpected memattr", m));
2572 KASSERT((m->oflags & (VPO_SWAPINPROG |
2573 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2574 ("page %p has unexpected oflags", m));
2575 /* Don't care: VPO_NOSYNC. */
2576 if (m->valid != 0) {
2577 /*
2578 * First, try to allocate a new page
2579 * that is above "high". Failing
2580 * that, try to allocate a new page
2581 * that is below "m_run". Allocate
2582 * the new page between the end of
2583 * "m_run" and "high" only as a last
2584 * resort.
2585 */
2586 req = req_class | VM_ALLOC_NOOBJ;
2587 if ((m->flags & PG_NODUMP) != 0)
2588 req |= VM_ALLOC_NODUMP;
2589 if (trunc_page(high) !=
2590 ~(vm_paddr_t)PAGE_MASK) {
2591 m_new = vm_page_alloc_contig(
2592 NULL, 0, req, 1,
2593 round_page(high),
2594 ~(vm_paddr_t)0,
2595 PAGE_SIZE, 0,
2596 VM_MEMATTR_DEFAULT);
2597 } else
2598 m_new = NULL;
2599 if (m_new == NULL) {
2600 pa = VM_PAGE_TO_PHYS(m_run);
2601 m_new = vm_page_alloc_contig(
2602 NULL, 0, req, 1,
2603 0, pa - 1, PAGE_SIZE, 0,
2604 VM_MEMATTR_DEFAULT);
2605 }
2606 if (m_new == NULL) {
2607 pa += ptoa(npages);
2608 m_new = vm_page_alloc_contig(
2609 NULL, 0, req, 1,
2610 pa, high, PAGE_SIZE, 0,
2611 VM_MEMATTR_DEFAULT);
2612 }
2613 if (m_new == NULL) {
2614 error = ENOMEM;
2615 goto unlock;
2616 }
2617 KASSERT(!vm_page_wired(m_new),
2618 ("page %p is wired", m_new));
2619
2620 /*
2621 * Replace "m" with the new page. For
2622 * vm_page_replace(), "m" must be busy
2623 * and dequeued. Finally, change "m"
2624 * as if vm_page_free() was called.
2625 */
2626 if (object->ref_count != 0)
2627 pmap_remove_all(m);
2628 m_new->aflags = m->aflags &
2629 ~PGA_QUEUE_STATE_MASK;
2630 KASSERT(m_new->oflags == VPO_UNMANAGED,
2631 ("page %p is managed", m_new));
2632 m_new->oflags = m->oflags & VPO_NOSYNC;
2633 pmap_copy_page(m, m_new);
2634 m_new->valid = m->valid;
2635 m_new->dirty = m->dirty;
2636 m->flags &= ~PG_ZERO;
2637 vm_page_xbusy(m);
2638 vm_page_dequeue(m);
2639 vm_page_replace_checked(m_new, object,
2640 m->pindex, m);
2641 if (vm_page_free_prep(m))
2642 SLIST_INSERT_HEAD(&free, m,
2643 plinks.s.ss);
2644
2645 /*
2646 * The new page must be deactivated
2647 * before the object is unlocked.
2648 */
2649 vm_page_change_lock(m_new, &m_mtx);
2650 vm_page_deactivate(m_new);
2651 } else {
2652 m->flags &= ~PG_ZERO;
2653 vm_page_dequeue(m);
2654 if (vm_page_free_prep(m))
2655 SLIST_INSERT_HEAD(&free, m,
2656 plinks.s.ss);
2657 KASSERT(m->dirty == 0,
2658 ("page %p is dirty", m));
2659 }
2660 } else
2661 error = EBUSY;
2662 unlock:
2663 VM_OBJECT_WUNLOCK(object);
2664 } else {
2665 MPASS(vm_phys_domain(m) == domain);
2666 vmd = VM_DOMAIN(domain);
2667 vm_domain_free_lock(vmd);
2668 order = m->order;
2669 if (order < VM_NFREEORDER) {
2670 /*
2671 * The page is enqueued in the physical memory
2672 * allocator's free page queues. Moreover, it
2673 * is the first page in a power-of-two-sized
2674 * run of contiguous free pages. Jump ahead
2675 * to the last page within that run, and
2676 * continue from there.
2677 */
2678 m += (1 << order) - 1;
2679 }
2680 #if VM_NRESERVLEVEL > 0
2681 else if (vm_reserv_is_page_free(m))
2682 order = 0;
2683 #endif
2684 vm_domain_free_unlock(vmd);
2685 if (order == VM_NFREEORDER)
2686 error = EINVAL;
2687 }
2688 }
2689 if (m_mtx != NULL)
2690 mtx_unlock(m_mtx);
2691 if ((m = SLIST_FIRST(&free)) != NULL) {
2692 int cnt;
2693
2694 vmd = VM_DOMAIN(domain);
2695 cnt = 0;
2696 vm_domain_free_lock(vmd);
2697 do {
2698 MPASS(vm_phys_domain(m) == domain);
2699 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2700 vm_phys_free_pages(m, 0);
2701 cnt++;
2702 } while ((m = SLIST_FIRST(&free)) != NULL);
2703 vm_domain_free_unlock(vmd);
2704 vm_domain_freecnt_inc(vmd, cnt);
2705 }
2706 return (error);
2707 }
2708
2709 #define NRUNS 16
2710
2711 CTASSERT(powerof2(NRUNS));
2712
2713 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2714
2715 #define MIN_RECLAIM 8
2716
2717 /*
2718 * vm_page_reclaim_contig:
2719 *
2720 * Reclaim allocated, contiguous physical memory satisfying the specified
2721 * conditions by relocating the virtual pages using that physical memory.
2722 * Returns true if reclamation is successful and false otherwise. Since
2723 * relocation requires the allocation of physical pages, reclamation may
2724 * fail due to a shortage of free pages. When reclamation fails, callers
2725 * are expected to perform vm_wait() before retrying a failed allocation
2726 * operation, e.g., vm_page_alloc_contig().
2727 *
2728 * The caller must always specify an allocation class through "req".
2729 *
2730 * allocation classes:
2731 * VM_ALLOC_NORMAL normal process request
2732 * VM_ALLOC_SYSTEM system *really* needs a page
2733 * VM_ALLOC_INTERRUPT interrupt time request
2734 *
2735 * The optional allocation flags are ignored.
2736 *
2737 * "npages" must be greater than zero. Both "alignment" and "boundary"
2738 * must be a power of two.
2739 */
2740 bool
2741 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2742 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2743 {
2744 struct vm_domain *vmd;
2745 vm_paddr_t curr_low;
2746 vm_page_t m_run, m_runs[NRUNS];
2747 u_long count, reclaimed;
2748 int error, i, options, req_class;
2749
2750 KASSERT(npages > 0, ("npages is 0"));
2751 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2752 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2753 req_class = req & VM_ALLOC_CLASS_MASK;
2754
2755 /*
2756 * The page daemon is allowed to dig deeper into the free page list.
2757 */
2758 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2759 req_class = VM_ALLOC_SYSTEM;
2760
2761 /*
2762 * Return if the number of free pages cannot satisfy the requested
2763 * allocation.
2764 */
2765 vmd = VM_DOMAIN(domain);
2766 count = vmd->vmd_free_count;
2767 if (count < npages + vmd->vmd_free_reserved || (count < npages +
2768 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2769 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2770 return (false);
2771
2772 /*
2773 * Scan up to three times, relaxing the restrictions ("options") on
2774 * the reclamation of reservations and superpages each time.
2775 */
2776 for (options = VPSC_NORESERV;;) {
2777 /*
2778 * Find the highest runs that satisfy the given constraints
2779 * and restrictions, and record them in "m_runs".
2780 */
2781 curr_low = low;
2782 count = 0;
2783 for (;;) {
2784 m_run = vm_phys_scan_contig(domain, npages, curr_low,
2785 high, alignment, boundary, options);
2786 if (m_run == NULL)
2787 break;
2788 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2789 m_runs[RUN_INDEX(count)] = m_run;
2790 count++;
2791 }
2792
2793 /*
2794 * Reclaim the highest runs in LIFO (descending) order until
2795 * the number of reclaimed pages, "reclaimed", is at least
2796 * MIN_RECLAIM. Reset "reclaimed" each time because each
2797 * reclamation is idempotent, and runs will (likely) recur
2798 * from one scan to the next as restrictions are relaxed.
2799 */
2800 reclaimed = 0;
2801 for (i = 0; count > 0 && i < NRUNS; i++) {
2802 count--;
2803 m_run = m_runs[RUN_INDEX(count)];
2804 error = vm_page_reclaim_run(req_class, domain, npages,
2805 m_run, high);
2806 if (error == 0) {
2807 reclaimed += npages;
2808 if (reclaimed >= MIN_RECLAIM)
2809 return (true);
2810 }
2811 }
2812
2813 /*
2814 * Either relax the restrictions on the next scan or return if
2815 * the last scan had no restrictions.
2816 */
2817 if (options == VPSC_NORESERV)
2818 options = VPSC_NOSUPER;
2819 else if (options == VPSC_NOSUPER)
2820 options = VPSC_ANY;
2821 else if (options == VPSC_ANY)
2822 return (reclaimed != 0);
2823 }
2824 }
2825
2826 bool
2827 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2828 u_long alignment, vm_paddr_t boundary)
2829 {
2830 struct vm_domainset_iter di;
2831 int domain;
2832 bool ret;
2833
2834 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2835 do {
2836 ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2837 high, alignment, boundary);
2838 if (ret)
2839 break;
2840 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2841
2842 return (ret);
2843 }
2844
2845 /*
2846 * Set the domain in the appropriate page level domainset.
2847 */
2848 void
2849 vm_domain_set(struct vm_domain *vmd)
2850 {
2851
2852 mtx_lock(&vm_domainset_lock);
2853 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2854 vmd->vmd_minset = 1;
2855 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2856 }
2857 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2858 vmd->vmd_severeset = 1;
2859 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2860 }
2861 mtx_unlock(&vm_domainset_lock);
2862 }
2863
2864 /*
2865 * Clear the domain from the appropriate page level domainset.
2866 */
2867 void
2868 vm_domain_clear(struct vm_domain *vmd)
2869 {
2870
2871 mtx_lock(&vm_domainset_lock);
2872 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2873 vmd->vmd_minset = 0;
2874 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2875 if (vm_min_waiters != 0) {
2876 vm_min_waiters = 0;
2877 wakeup(&vm_min_domains);
2878 }
2879 }
2880 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2881 vmd->vmd_severeset = 0;
2882 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2883 if (vm_severe_waiters != 0) {
2884 vm_severe_waiters = 0;
2885 wakeup(&vm_severe_domains);
2886 }
2887 }
2888
2889 /*
2890 * If pageout daemon needs pages, then tell it that there are
2891 * some free.
2892 */
2893 if (vmd->vmd_pageout_pages_needed &&
2894 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2895 wakeup(&vmd->vmd_pageout_pages_needed);
2896 vmd->vmd_pageout_pages_needed = 0;
2897 }
2898
2899 /* See comments in vm_wait_doms(). */
2900 if (vm_pageproc_waiters) {
2901 vm_pageproc_waiters = 0;
2902 wakeup(&vm_pageproc_waiters);
2903 }
2904 mtx_unlock(&vm_domainset_lock);
2905 }
2906
2907 /*
2908 * Wait for free pages to exceed the min threshold globally.
2909 */
2910 void
2911 vm_wait_min(void)
2912 {
2913
2914 mtx_lock(&vm_domainset_lock);
2915 while (vm_page_count_min()) {
2916 vm_min_waiters++;
2917 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2918 }
2919 mtx_unlock(&vm_domainset_lock);
2920 }
2921
2922 /*
2923 * Wait for free pages to exceed the severe threshold globally.
2924 */
2925 void
2926 vm_wait_severe(void)
2927 {
2928
2929 mtx_lock(&vm_domainset_lock);
2930 while (vm_page_count_severe()) {
2931 vm_severe_waiters++;
2932 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2933 "vmwait", 0);
2934 }
2935 mtx_unlock(&vm_domainset_lock);
2936 }
2937
2938 u_int
2939 vm_wait_count(void)
2940 {
2941
2942 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2943 }
2944
2945 int
2946 vm_wait_doms(const domainset_t *wdoms, int mflags)
2947 {
2948 int error;
2949
2950 error = 0;
2951
2952 /*
2953 * We use racey wakeup synchronization to avoid expensive global
2954 * locking for the pageproc when sleeping with a non-specific vm_wait.
2955 * To handle this, we only sleep for one tick in this instance. It
2956 * is expected that most allocations for the pageproc will come from
2957 * kmem or vm_page_grab* which will use the more specific and
2958 * race-free vm_wait_domain().
2959 */
2960 if (curproc == pageproc) {
2961 mtx_lock(&vm_domainset_lock);
2962 vm_pageproc_waiters++;
2963 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
2964 PVM | PDROP | mflags, "pageprocwait", 1);
2965 } else {
2966 /*
2967 * XXX Ideally we would wait only until the allocation could
2968 * be satisfied. This condition can cause new allocators to
2969 * consume all freed pages while old allocators wait.
2970 */
2971 mtx_lock(&vm_domainset_lock);
2972 if (vm_page_count_min_set(wdoms)) {
2973 vm_min_waiters++;
2974 error = msleep(&vm_min_domains, &vm_domainset_lock,
2975 PVM | PDROP | mflags, "vmwait", 0);
2976 } else
2977 mtx_unlock(&vm_domainset_lock);
2978 }
2979 return (error);
2980 }
2981
2982 /*
2983 * vm_wait_domain:
2984 *
2985 * Sleep until free pages are available for allocation.
2986 * - Called in various places after failed memory allocations.
2987 */
2988 void
2989 vm_wait_domain(int domain)
2990 {
2991 struct vm_domain *vmd;
2992 domainset_t wdom;
2993
2994 vmd = VM_DOMAIN(domain);
2995 vm_domain_free_assert_unlocked(vmd);
2996
2997 if (curproc == pageproc) {
2998 mtx_lock(&vm_domainset_lock);
2999 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3000 vmd->vmd_pageout_pages_needed = 1;
3001 msleep(&vmd->vmd_pageout_pages_needed,
3002 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3003 } else
3004 mtx_unlock(&vm_domainset_lock);
3005 } else {
3006 if (pageproc == NULL)
3007 panic("vm_wait in early boot");
3008 DOMAINSET_ZERO(&wdom);
3009 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3010 vm_wait_doms(&wdom, 0);
3011 }
3012 }
3013
3014 static int
3015 vm_wait_flags(vm_object_t obj, int mflags)
3016 {
3017 struct domainset *d;
3018
3019 d = NULL;
3020
3021 /*
3022 * Carefully fetch pointers only once: the struct domainset
3023 * itself is ummutable but the pointer might change.
3024 */
3025 if (obj != NULL)
3026 d = obj->domain.dr_policy;
3027 if (d == NULL)
3028 d = curthread->td_domain.dr_policy;
3029
3030 return (vm_wait_doms(&d->ds_mask, mflags));
3031 }
3032
3033 /*
3034 * vm_wait:
3035 *
3036 * Sleep until free pages are available for allocation in the
3037 * affinity domains of the obj. If obj is NULL, the domain set
3038 * for the calling thread is used.
3039 * Called in various places after failed memory allocations.
3040 */
3041 void
3042 vm_wait(vm_object_t obj)
3043 {
3044 (void)vm_wait_flags(obj, 0);
3045 }
3046
3047 int
3048 vm_wait_intr(vm_object_t obj)
3049 {
3050 return (vm_wait_flags(obj, PCATCH));
3051 }
3052
3053 /*
3054 * vm_domain_alloc_fail:
3055 *
3056 * Called when a page allocation function fails. Informs the
3057 * pagedaemon and performs the requested wait. Requires the
3058 * domain_free and object lock on entry. Returns with the
3059 * object lock held and free lock released. Returns an error when
3060 * retry is necessary.
3061 *
3062 */
3063 static int
3064 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3065 {
3066
3067 vm_domain_free_assert_unlocked(vmd);
3068
3069 atomic_add_int(&vmd->vmd_pageout_deficit,
3070 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3071 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3072 if (object != NULL)
3073 VM_OBJECT_WUNLOCK(object);
3074 vm_wait_domain(vmd->vmd_domain);
3075 if (object != NULL)
3076 VM_OBJECT_WLOCK(object);
3077 if (req & VM_ALLOC_WAITOK)
3078 return (EAGAIN);
3079 }
3080
3081 return (0);
3082 }
3083
3084 /*
3085 * vm_waitpfault:
3086 *
3087 * Sleep until free pages are available for allocation.
3088 * - Called only in vm_fault so that processes page faulting
3089 * can be easily tracked.
3090 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3091 * processes will be able to grab memory first. Do not change
3092 * this balance without careful testing first.
3093 */
3094 void
3095 vm_waitpfault(struct domainset *dset, int timo)
3096 {
3097
3098 /*
3099 * XXX Ideally we would wait only until the allocation could
3100 * be satisfied. This condition can cause new allocators to
3101 * consume all freed pages while old allocators wait.
3102 */
3103 mtx_lock(&vm_domainset_lock);
3104 if (vm_page_count_min_set(&dset->ds_mask)) {
3105 vm_min_waiters++;
3106 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3107 "pfault", timo);
3108 } else
3109 mtx_unlock(&vm_domainset_lock);
3110 }
3111
3112 static struct vm_pagequeue *
3113 vm_page_pagequeue(vm_page_t m)
3114 {
3115
3116 uint8_t queue;
3117
3118 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE)
3119 return (NULL);
3120 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3121 }
3122
3123 static inline void
3124 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3125 {
3126 struct vm_domain *vmd;
3127 uint8_t qflags;
3128
3129 CRITICAL_ASSERT(curthread);
3130 vm_pagequeue_assert_locked(pq);
3131
3132 /*
3133 * The page daemon is allowed to set m->queue = PQ_NONE without
3134 * the page queue lock held. In this case it is about to free the page,
3135 * which must not have any queue state.
3136 */
3137 qflags = atomic_load_8(&m->aflags);
3138 KASSERT(pq == vm_page_pagequeue(m) ||
3139 (qflags & PGA_QUEUE_STATE_MASK) == 0,
3140 ("page %p doesn't belong to queue %p but has aflags %#x",
3141 m, pq, qflags));
3142
3143 if ((qflags & PGA_DEQUEUE) != 0) {
3144 if (__predict_true((qflags & PGA_ENQUEUED) != 0))
3145 vm_pagequeue_remove(pq, m);
3146 vm_page_dequeue_complete(m);
3147 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3148 if ((qflags & PGA_ENQUEUED) != 0)
3149 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3150 else {
3151 vm_pagequeue_cnt_inc(pq);
3152 vm_page_aflag_set(m, PGA_ENQUEUED);
3153 }
3154
3155 /*
3156 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.
3157 * In particular, if both flags are set in close succession,
3158 * only PGA_REQUEUE_HEAD will be applied, even if it was set
3159 * first.
3160 */
3161 if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3162 KASSERT(m->queue == PQ_INACTIVE,
3163 ("head enqueue not supported for page %p", m));
3164 vmd = vm_pagequeue_domain(m);
3165 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3166 } else
3167 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3168
3169 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE |
3170 PGA_REQUEUE_HEAD));
3171 }
3172 }
3173
3174 static void
3175 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3176 uint8_t queue)
3177 {
3178 vm_page_t m;
3179 int i;
3180
3181 for (i = 0; i < bq->bq_cnt; i++) {
3182 m = bq->bq_pa[i];
3183 if (__predict_false(m->queue != queue))
3184 continue;
3185 vm_pqbatch_process_page(pq, m);
3186 }
3187 vm_batchqueue_init(bq);
3188 }
3189
3190 static void
3191 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue)
3192 {
3193 struct vm_batchqueue *bq;
3194 struct vm_pagequeue *pq;
3195 int domain;
3196
3197 vm_page_assert_locked(m);
3198 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3199
3200 domain = vm_phys_domain(m);
3201 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3202
3203 critical_enter();
3204 bq = DPCPU_PTR(pqbatch[domain][queue]);
3205 if (vm_batchqueue_insert(bq, m)) {
3206 critical_exit();
3207 return;
3208 }
3209 if (!vm_pagequeue_trylock(pq)) {
3210 critical_exit();
3211 vm_pagequeue_lock(pq);
3212 critical_enter();
3213 bq = DPCPU_PTR(pqbatch[domain][queue]);
3214 }
3215 vm_pqbatch_process(pq, bq, queue);
3216
3217 /*
3218 * The page may have been logically dequeued before we acquired the
3219 * page queue lock. In this case, the page lock prevents the page
3220 * from being logically enqueued elsewhere.
3221 */
3222 if (__predict_true(m->queue == queue))
3223 vm_pqbatch_process_page(pq, m);
3224 else {
3225 KASSERT(m->queue == PQ_NONE,
3226 ("invalid queue transition for page %p", m));
3227 KASSERT((m->aflags & PGA_ENQUEUED) == 0,
3228 ("page %p is enqueued with invalid queue index", m));
3229 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3230 }
3231 vm_pagequeue_unlock(pq);
3232 critical_exit();
3233 }
3234
3235 /*
3236 * vm_page_drain_pqbatch: [ internal use only ]
3237 *
3238 * Force all per-CPU page queue batch queues to be drained. This is
3239 * intended for use in severe memory shortages, to ensure that pages
3240 * do not remain stuck in the batch queues.
3241 */
3242 void
3243 vm_page_drain_pqbatch(void)
3244 {
3245 struct thread *td;
3246 struct vm_domain *vmd;
3247 struct vm_pagequeue *pq;
3248 int cpu, domain, queue;
3249
3250 td = curthread;
3251 CPU_FOREACH(cpu) {
3252 thread_lock(td);
3253 sched_bind(td, cpu);
3254 thread_unlock(td);
3255
3256 for (domain = 0; domain < vm_ndomains; domain++) {
3257 vmd = VM_DOMAIN(domain);
3258 for (queue = 0; queue < PQ_COUNT; queue++) {
3259 pq = &vmd->vmd_pagequeues[queue];
3260 vm_pagequeue_lock(pq);
3261 critical_enter();
3262 vm_pqbatch_process(pq,
3263 DPCPU_PTR(pqbatch[domain][queue]), queue);
3264 critical_exit();
3265 vm_pagequeue_unlock(pq);
3266 }
3267 }
3268 }
3269 thread_lock(td);
3270 sched_unbind(td);
3271 thread_unlock(td);
3272 }
3273
3274 /*
3275 * Complete the logical removal of a page from a page queue. We must be
3276 * careful to synchronize with the page daemon, which may be concurrently
3277 * examining the page with only the page lock held. The page must not be
3278 * in a state where it appears to be logically enqueued.
3279 */
3280 static void
3281 vm_page_dequeue_complete(vm_page_t m)
3282 {
3283
3284 m->queue = PQ_NONE;
3285 atomic_thread_fence_rel();
3286 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3287 }
3288
3289 /*
3290 * vm_page_dequeue_deferred: [ internal use only ]
3291 *
3292 * Request removal of the given page from its current page
3293 * queue. Physical removal from the queue may be deferred
3294 * indefinitely.
3295 *
3296 * The page must be locked.
3297 */
3298 void
3299 vm_page_dequeue_deferred(vm_page_t m)
3300 {
3301 uint8_t queue;
3302
3303 vm_page_assert_locked(m);
3304
3305 if ((queue = vm_page_queue(m)) == PQ_NONE)
3306 return;
3307 vm_page_aflag_set(m, PGA_DEQUEUE);
3308 vm_pqbatch_submit_page(m, queue);
3309 }
3310
3311 /*
3312 * vm_page_dequeue:
3313 *
3314 * Remove the page from whichever page queue it's in, if any.
3315 * The page must either be locked or unallocated. This constraint
3316 * ensures that the queue state of the page will remain consistent
3317 * after this function returns.
3318 */
3319 void
3320 vm_page_dequeue(vm_page_t m)
3321 {
3322 struct vm_pagequeue *pq, *pq1;
3323 uint8_t aflags;
3324
3325 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3326 ("page %p is allocated and unlocked", m));
3327
3328 for (pq = vm_page_pagequeue(m);; pq = pq1) {
3329 if (pq == NULL) {
3330 /*
3331 * A thread may be concurrently executing
3332 * vm_page_dequeue_complete(). Ensure that all queue
3333 * state is cleared before we return.
3334 */
3335 aflags = atomic_load_8(&m->aflags);
3336 if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3337 return;
3338 KASSERT((aflags & PGA_DEQUEUE) != 0,
3339 ("page %p has unexpected queue state flags %#x",
3340 m, aflags));
3341
3342 /*
3343 * Busy wait until the thread updating queue state is
3344 * finished. Such a thread must be executing in a
3345 * critical section.
3346 */
3347 cpu_spinwait();
3348 pq1 = vm_page_pagequeue(m);
3349 continue;
3350 }
3351 vm_pagequeue_lock(pq);
3352 if ((pq1 = vm_page_pagequeue(m)) == pq)
3353 break;
3354 vm_pagequeue_unlock(pq);
3355 }
3356 KASSERT(pq == vm_page_pagequeue(m),
3357 ("%s: page %p migrated directly between queues", __func__, m));
3358 KASSERT((m->aflags & PGA_DEQUEUE) != 0 ||
3359 mtx_owned(vm_page_lockptr(m)),
3360 ("%s: queued unlocked page %p", __func__, m));
3361
3362 if ((m->aflags & PGA_ENQUEUED) != 0)
3363 vm_pagequeue_remove(pq, m);
3364 vm_page_dequeue_complete(m);
3365 vm_pagequeue_unlock(pq);
3366 }
3367
3368 /*
3369 * Schedule the given page for insertion into the specified page queue.
3370 * Physical insertion of the page may be deferred indefinitely.
3371 */
3372 static void
3373 vm_page_enqueue(vm_page_t m, uint8_t queue)
3374 {
3375
3376 vm_page_assert_locked(m);
3377 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0,
3378 ("%s: page %p is already enqueued", __func__, m));
3379
3380 m->queue = queue;
3381 if ((m->aflags & PGA_REQUEUE) == 0)
3382 vm_page_aflag_set(m, PGA_REQUEUE);
3383 vm_pqbatch_submit_page(m, queue);
3384 }
3385
3386 /*
3387 * vm_page_requeue: [ internal use only ]
3388 *
3389 * Schedule a requeue of the given page.
3390 *
3391 * The page must be locked.
3392 */
3393 void
3394 vm_page_requeue(vm_page_t m)
3395 {
3396
3397 vm_page_assert_locked(m);
3398 KASSERT(vm_page_queue(m) != PQ_NONE,
3399 ("%s: page %p is not logically enqueued", __func__, m));
3400
3401 if ((m->aflags & PGA_REQUEUE) == 0)
3402 vm_page_aflag_set(m, PGA_REQUEUE);
3403 vm_pqbatch_submit_page(m, atomic_load_8(&m->queue));
3404 }
3405
3406 /*
3407 * vm_page_free_prep:
3408 *
3409 * Prepares the given page to be put on the free list,
3410 * disassociating it from any VM object. The caller may return
3411 * the page to the free list only if this function returns true.
3412 *
3413 * The object must be locked. The page must be locked if it is
3414 * managed.
3415 */
3416 bool
3417 vm_page_free_prep(vm_page_t m)
3418 {
3419
3420 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3421 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3422 uint64_t *p;
3423 int i;
3424 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3425 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3426 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3427 m, i, (uintmax_t)*p));
3428 }
3429 #endif
3430 if ((m->oflags & VPO_UNMANAGED) == 0) {
3431 vm_page_lock_assert(m, MA_OWNED);
3432 KASSERT(!pmap_page_is_mapped(m),
3433 ("vm_page_free_prep: freeing mapped page %p", m));
3434 KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3435 ("vm_page_free_prep: mapping flags set in page %p", m));
3436 } else {
3437 KASSERT(m->queue == PQ_NONE,
3438 ("vm_page_free_prep: unmanaged page %p is queued", m));
3439 }
3440 VM_CNT_INC(v_tfree);
3441
3442 if (vm_page_sbusied(m))
3443 panic("vm_page_free_prep: freeing busy page %p", m);
3444
3445 if (m->object != NULL)
3446 (void)vm_page_remove(m);
3447
3448 /*
3449 * If fictitious remove object association and
3450 * return.
3451 */
3452 if ((m->flags & PG_FICTITIOUS) != 0) {
3453 KASSERT(m->wire_count == 1,
3454 ("fictitious page %p is not wired", m));
3455 KASSERT(m->queue == PQ_NONE,
3456 ("fictitious page %p is queued", m));
3457 return (false);
3458 }
3459
3460 /*
3461 * Pages need not be dequeued before they are returned to the physical
3462 * memory allocator, but they must at least be marked for a deferred
3463 * dequeue.
3464 */
3465 if ((m->oflags & VPO_UNMANAGED) == 0)
3466 vm_page_dequeue_deferred(m);
3467
3468 m->valid = 0;
3469 vm_page_undirty(m);
3470
3471 if (vm_page_wired(m) != 0)
3472 panic("vm_page_free_prep: freeing wired page %p", m);
3473 if (m->hold_count != 0) {
3474 m->flags &= ~PG_ZERO;
3475 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3476 ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m));
3477 m->flags |= PG_UNHOLDFREE;
3478 return (false);
3479 }
3480
3481 /*
3482 * Restore the default memory attribute to the page.
3483 */
3484 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3485 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3486
3487 #if VM_NRESERVLEVEL > 0
3488 /*
3489 * Determine whether the page belongs to a reservation. If the page was
3490 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3491 * as an optimization, we avoid the check in that case.
3492 */
3493 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3494 return (false);
3495 #endif
3496
3497 return (true);
3498 }
3499
3500 /*
3501 * vm_page_free_toq:
3502 *
3503 * Returns the given page to the free list, disassociating it
3504 * from any VM object.
3505 *
3506 * The object must be locked. The page must be locked if it is
3507 * managed.
3508 */
3509 void
3510 vm_page_free_toq(vm_page_t m)
3511 {
3512 struct vm_domain *vmd;
3513 uma_zone_t zone;
3514
3515 if (!vm_page_free_prep(m))
3516 return;
3517
3518 vmd = vm_pagequeue_domain(m);
3519 zone = vmd->vmd_pgcache[m->pool].zone;
3520 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3521 uma_zfree(zone, m);
3522 return;
3523 }
3524 vm_domain_free_lock(vmd);
3525 vm_phys_free_pages(m, 0);
3526 vm_domain_free_unlock(vmd);
3527 vm_domain_freecnt_inc(vmd, 1);
3528 }
3529
3530 /*
3531 * vm_page_free_pages_toq:
3532 *
3533 * Returns a list of pages to the free list, disassociating it
3534 * from any VM object. In other words, this is equivalent to
3535 * calling vm_page_free_toq() for each page of a list of VM objects.
3536 *
3537 * The objects must be locked. The pages must be locked if it is
3538 * managed.
3539 */
3540 void
3541 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3542 {
3543 vm_page_t m;
3544 int count;
3545
3546 if (SLIST_EMPTY(free))
3547 return;
3548
3549 count = 0;
3550 while ((m = SLIST_FIRST(free)) != NULL) {
3551 count++;
3552 SLIST_REMOVE_HEAD(free, plinks.s.ss);
3553 vm_page_free_toq(m);
3554 }
3555
3556 if (update_wire_count)
3557 vm_wire_sub(count);
3558 }
3559
3560 /*
3561 * vm_page_wire:
3562 *
3563 * Mark this page as wired down. If the page is fictitious, then
3564 * its wire count must remain one.
3565 *
3566 * The page must be locked.
3567 */
3568 void
3569 vm_page_wire(vm_page_t m)
3570 {
3571
3572 vm_page_assert_locked(m);
3573 if ((m->flags & PG_FICTITIOUS) != 0) {
3574 KASSERT(m->wire_count == 1,
3575 ("vm_page_wire: fictitious page %p's wire count isn't one",
3576 m));
3577 return;
3578 }
3579 if (!vm_page_wired(m)) {
3580 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3581 m->queue == PQ_NONE,
3582 ("vm_page_wire: unmanaged page %p is queued", m));
3583 vm_wire_add(1);
3584 }
3585 m->wire_count++;
3586 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3587 }
3588
3589 /*
3590 * vm_page_unwire:
3591 *
3592 * Release one wiring of the specified page, potentially allowing it to be
3593 * paged out. Returns TRUE if the number of wirings transitions to zero and
3594 * FALSE otherwise.
3595 *
3596 * Only managed pages belonging to an object can be paged out. If the number
3597 * of wirings transitions to zero and the page is eligible for page out, then
3598 * the page is added to the specified paging queue (unless PQ_NONE is
3599 * specified, in which case the page is dequeued if it belongs to a paging
3600 * queue).
3601 *
3602 * If a page is fictitious, then its wire count must always be one.
3603 *
3604 * A managed page must be locked.
3605 */
3606 bool
3607 vm_page_unwire(vm_page_t m, uint8_t queue)
3608 {
3609 bool unwired;
3610
3611 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3612 ("vm_page_unwire: invalid queue %u request for page %p",
3613 queue, m));
3614 if ((m->oflags & VPO_UNMANAGED) == 0)
3615 vm_page_assert_locked(m);
3616
3617 unwired = vm_page_unwire_noq(m);
3618 if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL)
3619 return (unwired);
3620
3621 if (vm_page_queue(m) == queue) {
3622 if (queue == PQ_ACTIVE)
3623 vm_page_reference(m);
3624 else if (queue != PQ_NONE)
3625 vm_page_requeue(m);
3626 } else {
3627 vm_page_dequeue(m);
3628 if (queue != PQ_NONE) {
3629 vm_page_enqueue(m, queue);
3630 if (queue == PQ_ACTIVE)
3631 /* Initialize act_count. */
3632 vm_page_activate(m);
3633 }
3634 }
3635 return (unwired);
3636 }
3637
3638 /*
3639 *
3640 * vm_page_unwire_noq:
3641 *
3642 * Unwire a page without (re-)inserting it into a page queue. It is up
3643 * to the caller to enqueue, requeue, or free the page as appropriate.
3644 * In most cases, vm_page_unwire() should be used instead.
3645 */
3646 bool
3647 vm_page_unwire_noq(vm_page_t m)
3648 {
3649
3650 if ((m->oflags & VPO_UNMANAGED) == 0)
3651 vm_page_assert_locked(m);
3652 if ((m->flags & PG_FICTITIOUS) != 0) {
3653 KASSERT(m->wire_count == 1,
3654 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3655 return (false);
3656 }
3657 if (!vm_page_wired(m))
3658 panic("vm_page_unwire: page %p's wire count is zero", m);
3659 m->wire_count--;
3660 if (m->wire_count == 0) {
3661 vm_wire_sub(1);
3662 return (true);
3663 } else
3664 return (false);
3665 }
3666
3667 /*
3668 * vm_page_activate:
3669 *
3670 * Put the specified page on the active list (if appropriate).
3671 * Ensure that act_count is at least ACT_INIT but do not otherwise
3672 * mess with it.
3673 *
3674 * The page must be locked.
3675 */
3676 void
3677 vm_page_activate(vm_page_t m)
3678 {
3679
3680 vm_page_assert_locked(m);
3681
3682 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3683 return;
3684 if (vm_page_queue(m) == PQ_ACTIVE) {
3685 if (m->act_count < ACT_INIT)
3686 m->act_count = ACT_INIT;
3687 return;
3688 }
3689
3690 vm_page_dequeue(m);
3691 if (m->act_count < ACT_INIT)
3692 m->act_count = ACT_INIT;
3693 vm_page_enqueue(m, PQ_ACTIVE);
3694 }
3695
3696 /*
3697 * Move the specified page to the tail of the inactive queue, or requeue
3698 * the page if it is already in the inactive queue.
3699 *
3700 * The page must be locked.
3701 */
3702 void
3703 vm_page_deactivate(vm_page_t m)
3704 {
3705
3706 vm_page_assert_locked(m);
3707
3708 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3709 return;
3710
3711 if (!vm_page_inactive(m)) {
3712 vm_page_dequeue(m);
3713 vm_page_enqueue(m, PQ_INACTIVE);
3714 } else
3715 vm_page_requeue(m);
3716 }
3717
3718 /*
3719 * Move the specified page close to the head of the inactive queue,
3720 * bypassing LRU. A marker page is used to maintain FIFO ordering.
3721 * As with regular enqueues, we use a per-CPU batch queue to reduce
3722 * contention on the page queue lock.
3723 *
3724 * The page must be locked.
3725 */
3726 void
3727 vm_page_deactivate_noreuse(vm_page_t m)
3728 {
3729
3730 vm_page_assert_locked(m);
3731
3732 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3733 return;
3734
3735 if (!vm_page_inactive(m)) {
3736 vm_page_dequeue(m);
3737 m->queue = PQ_INACTIVE;
3738 }
3739 if ((m->aflags & PGA_REQUEUE_HEAD) == 0)
3740 vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
3741 vm_pqbatch_submit_page(m, PQ_INACTIVE);
3742 }
3743
3744 /*
3745 * vm_page_launder
3746 *
3747 * Put a page in the laundry, or requeue it if it is already there.
3748 */
3749 void
3750 vm_page_launder(vm_page_t m)
3751 {
3752
3753 vm_page_assert_locked(m);
3754 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0)
3755 return;
3756
3757 if (vm_page_in_laundry(m))
3758 vm_page_requeue(m);
3759 else {
3760 vm_page_dequeue(m);
3761 vm_page_enqueue(m, PQ_LAUNDRY);
3762 }
3763 }
3764
3765 /*
3766 * vm_page_unswappable
3767 *
3768 * Put a page in the PQ_UNSWAPPABLE holding queue.
3769 */
3770 void
3771 vm_page_unswappable(vm_page_t m)
3772 {
3773
3774 vm_page_assert_locked(m);
3775 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
3776 ("page %p already unswappable", m));
3777
3778 vm_page_dequeue(m);
3779 vm_page_enqueue(m, PQ_UNSWAPPABLE);
3780 }
3781
3782 static void
3783 vm_page_release_toq(vm_page_t m, int flags)
3784 {
3785
3786 /*
3787 * Use a check of the valid bits to determine whether we should
3788 * accelerate reclamation of the page. The object lock might not be
3789 * held here, in which case the check is racy. At worst we will either
3790 * accelerate reclamation of a valid page and violate LRU, or
3791 * unnecessarily defer reclamation of an invalid page.
3792 *
3793 * If we were asked to not cache the page, place it near the head of the
3794 * inactive queue so that is reclaimed sooner.
3795 */
3796 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
3797 vm_page_deactivate_noreuse(m);
3798 else if (vm_page_active(m))
3799 vm_page_reference(m);
3800 else
3801 vm_page_deactivate(m);
3802 }
3803
3804 /*
3805 * Unwire a page and either attempt to free it or re-add it to the page queues.
3806 */
3807 void
3808 vm_page_release(vm_page_t m, int flags)
3809 {
3810 vm_object_t object;
3811 bool freed;
3812
3813 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3814 ("vm_page_release: page %p is unmanaged", m));
3815
3816 vm_page_lock(m);
3817 if (m->object != NULL)
3818 VM_OBJECT_ASSERT_UNLOCKED(m->object);
3819 if (vm_page_unwire_noq(m)) {
3820 if ((object = m->object) == NULL) {
3821 vm_page_free(m);
3822 } else {
3823 freed = false;
3824 if ((flags & VPR_TRYFREE) != 0 && !vm_page_busied(m) &&
3825 /* Depends on type stability. */
3826 VM_OBJECT_TRYWLOCK(object)) {
3827 /*
3828 * Only free unmapped pages. The busy test from
3829 * before the object was locked cannot be relied
3830 * upon.
3831 */
3832 if ((object->ref_count == 0 ||
3833 !pmap_page_is_mapped(m)) && m->dirty == 0 &&
3834 !vm_page_busied(m)) {
3835 vm_page_free(m);
3836 freed = true;
3837 }
3838 VM_OBJECT_WUNLOCK(object);
3839 }
3840
3841 if (!freed)
3842 vm_page_release_toq(m, flags);
3843 }
3844 }
3845 vm_page_unlock(m);
3846 }
3847
3848 /* See vm_page_release(). */
3849 void
3850 vm_page_release_locked(vm_page_t m, int flags)
3851 {
3852
3853 VM_OBJECT_ASSERT_WLOCKED(m->object);
3854 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3855 ("vm_page_release_locked: page %p is unmanaged", m));
3856
3857 vm_page_lock(m);
3858 if (vm_page_unwire_noq(m)) {
3859 if ((flags & VPR_TRYFREE) != 0 &&
3860 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
3861 m->dirty == 0 && !vm_page_busied(m)) {
3862 vm_page_free(m);
3863 } else {
3864 vm_page_release_toq(m, flags);
3865 }
3866 }
3867 vm_page_unlock(m);
3868 }
3869
3870 /*
3871 * vm_page_advise
3872 *
3873 * Apply the specified advice to the given page.
3874 *
3875 * The object and page must be locked.
3876 */
3877 void
3878 vm_page_advise(vm_page_t m, int advice)
3879 {
3880
3881 vm_page_assert_locked(m);
3882 VM_OBJECT_ASSERT_WLOCKED(m->object);
3883 if (advice == MADV_FREE)
3884 /*
3885 * Mark the page clean. This will allow the page to be freed
3886 * without first paging it out. MADV_FREE pages are often
3887 * quickly reused by malloc(3), so we do not do anything that
3888 * would result in a page fault on a later access.
3889 */
3890 vm_page_undirty(m);
3891 else if (advice != MADV_DONTNEED) {
3892 if (advice == MADV_WILLNEED)
3893 vm_page_activate(m);
3894 return;
3895 }
3896
3897 /*
3898 * Clear any references to the page. Otherwise, the page daemon will
3899 * immediately reactivate the page.
3900 */
3901 vm_page_aflag_clear(m, PGA_REFERENCED);
3902
3903 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3904 vm_page_dirty(m);
3905
3906 /*
3907 * Place clean pages near the head of the inactive queue rather than
3908 * the tail, thus defeating the queue's LRU operation and ensuring that
3909 * the page will be reused quickly. Dirty pages not already in the
3910 * laundry are moved there.
3911 */
3912 if (m->dirty == 0)
3913 vm_page_deactivate_noreuse(m);
3914 else if (!vm_page_in_laundry(m))
3915 vm_page_launder(m);
3916 }
3917
3918 /*
3919 * Grab a page, waiting until we are waken up due to the page
3920 * changing state. We keep on waiting, if the page continues
3921 * to be in the object. If the page doesn't exist, first allocate it
3922 * and then conditionally zero it.
3923 *
3924 * This routine may sleep.
3925 *
3926 * The object must be locked on entry. The lock will, however, be released
3927 * and reacquired if the routine sleeps.
3928 */
3929 vm_page_t
3930 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3931 {
3932 vm_page_t m;
3933 int sleep;
3934 int pflags;
3935
3936 VM_OBJECT_ASSERT_WLOCKED(object);
3937 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3938 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3939 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3940 pflags = allocflags &
3941 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3942 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3943 pflags |= VM_ALLOC_WAITFAIL;
3944 retrylookup:
3945 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3946 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3947 vm_page_xbusied(m) : vm_page_busied(m);
3948 if (sleep) {
3949 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3950 return (NULL);
3951 /*
3952 * Reference the page before unlocking and
3953 * sleeping so that the page daemon is less
3954 * likely to reclaim it.
3955 */
3956 vm_page_aflag_set(m, PGA_REFERENCED);
3957 vm_page_lock(m);
3958 VM_OBJECT_WUNLOCK(object);
3959 vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3960 VM_ALLOC_IGN_SBUSY) != 0);
3961 VM_OBJECT_WLOCK(object);
3962 goto retrylookup;
3963 } else {
3964 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3965 vm_page_lock(m);
3966 vm_page_wire(m);
3967 vm_page_unlock(m);
3968 }
3969 if ((allocflags &
3970 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3971 vm_page_xbusy(m);
3972 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3973 vm_page_sbusy(m);
3974 return (m);
3975 }
3976 }
3977 m = vm_page_alloc(object, pindex, pflags);
3978 if (m == NULL) {
3979 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3980 return (NULL);
3981 goto retrylookup;
3982 }
3983 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3984 pmap_zero_page(m);
3985 return (m);
3986 }
3987
3988 /*
3989 * Return the specified range of pages from the given object. For each
3990 * page offset within the range, if a page already exists within the object
3991 * at that offset and it is busy, then wait for it to change state. If,
3992 * instead, the page doesn't exist, then allocate it.
3993 *
3994 * The caller must always specify an allocation class.
3995 *
3996 * allocation classes:
3997 * VM_ALLOC_NORMAL normal process request
3998 * VM_ALLOC_SYSTEM system *really* needs the pages
3999 *
4000 * The caller must always specify that the pages are to be busied and/or
4001 * wired.
4002 *
4003 * optional allocation flags:
4004 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4005 * VM_ALLOC_NOBUSY do not exclusive busy the page
4006 * VM_ALLOC_NOWAIT do not sleep
4007 * VM_ALLOC_SBUSY set page to sbusy state
4008 * VM_ALLOC_WIRED wire the pages
4009 * VM_ALLOC_ZERO zero and validate any invalid pages
4010 *
4011 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4012 * may return a partial prefix of the requested range.
4013 */
4014 int
4015 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4016 vm_page_t *ma, int count)
4017 {
4018 vm_page_t m, mpred;
4019 int pflags;
4020 int i;
4021 bool sleep;
4022
4023 VM_OBJECT_ASSERT_WLOCKED(object);
4024 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4025 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4026 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4027 (allocflags & VM_ALLOC_WIRED) != 0,
4028 ("vm_page_grab_pages: the pages must be busied or wired"));
4029 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4030 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4031 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
4032 if (count == 0)
4033 return (0);
4034 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
4035 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
4036 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4037 pflags |= VM_ALLOC_WAITFAIL;
4038 i = 0;
4039 retrylookup:
4040 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4041 if (m == NULL || m->pindex != pindex + i) {
4042 mpred = m;
4043 m = NULL;
4044 } else
4045 mpred = TAILQ_PREV(m, pglist, listq);
4046 for (; i < count; i++) {
4047 if (m != NULL) {
4048 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
4049 vm_page_xbusied(m) : vm_page_busied(m);
4050 if (sleep) {
4051 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4052 break;
4053 /*
4054 * Reference the page before unlocking and
4055 * sleeping so that the page daemon is less
4056 * likely to reclaim it.
4057 */
4058 vm_page_aflag_set(m, PGA_REFERENCED);
4059 vm_page_lock(m);
4060 VM_OBJECT_WUNLOCK(object);
4061 vm_page_busy_sleep(m, "grbmaw", (allocflags &
4062 VM_ALLOC_IGN_SBUSY) != 0);
4063 VM_OBJECT_WLOCK(object);
4064 goto retrylookup;
4065 }
4066 if ((allocflags & VM_ALLOC_WIRED) != 0) {
4067 vm_page_lock(m);
4068 vm_page_wire(m);
4069 vm_page_unlock(m);
4070 }
4071 if ((allocflags & (VM_ALLOC_NOBUSY |
4072 VM_ALLOC_SBUSY)) == 0)
4073 vm_page_xbusy(m);
4074 if ((allocflags & VM_ALLOC_SBUSY) != 0)
4075 vm_page_sbusy(m);
4076 } else {
4077 m = vm_page_alloc_after(object, pindex + i,
4078 pflags | VM_ALLOC_COUNT(count - i), mpred);
4079 if (m == NULL) {
4080 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4081 break;
4082 goto retrylookup;
4083 }
4084 }
4085 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
4086 if ((m->flags & PG_ZERO) == 0)
4087 pmap_zero_page(m);
4088 m->valid = VM_PAGE_BITS_ALL;
4089 }
4090 ma[i] = mpred = m;
4091 m = vm_page_next(m);
4092 }
4093 return (i);
4094 }
4095
4096 /*
4097 * Mapping function for valid or dirty bits in a page.
4098 *
4099 * Inputs are required to range within a page.
4100 */
4101 vm_page_bits_t
4102 vm_page_bits(int base, int size)
4103 {
4104 int first_bit;
4105 int last_bit;
4106
4107 KASSERT(
4108 base + size <= PAGE_SIZE,
4109 ("vm_page_bits: illegal base/size %d/%d", base, size)
4110 );
4111
4112 if (size == 0) /* handle degenerate case */
4113 return (0);
4114
4115 first_bit = base >> DEV_BSHIFT;
4116 last_bit = (base + size - 1) >> DEV_BSHIFT;
4117
4118 return (((vm_page_bits_t)2 << last_bit) -
4119 ((vm_page_bits_t)1 << first_bit));
4120 }
4121
4122 /*
4123 * vm_page_set_valid_range:
4124 *
4125 * Sets portions of a page valid. The arguments are expected
4126 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4127 * of any partial chunks touched by the range. The invalid portion of
4128 * such chunks will be zeroed.
4129 *
4130 * (base + size) must be less then or equal to PAGE_SIZE.
4131 */
4132 void
4133 vm_page_set_valid_range(vm_page_t m, int base, int size)
4134 {
4135 int endoff, frag;
4136
4137 VM_OBJECT_ASSERT_WLOCKED(m->object);
4138 if (size == 0) /* handle degenerate case */
4139 return;
4140
4141 /*
4142 * If the base is not DEV_BSIZE aligned and the valid
4143 * bit is clear, we have to zero out a portion of the
4144 * first block.
4145 */
4146 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4147 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4148 pmap_zero_page_area(m, frag, base - frag);
4149
4150 /*
4151 * If the ending offset is not DEV_BSIZE aligned and the
4152 * valid bit is clear, we have to zero out a portion of
4153 * the last block.
4154 */
4155 endoff = base + size;
4156 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4157 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4158 pmap_zero_page_area(m, endoff,
4159 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4160
4161 /*
4162 * Assert that no previously invalid block that is now being validated
4163 * is already dirty.
4164 */
4165 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4166 ("vm_page_set_valid_range: page %p is dirty", m));
4167
4168 /*
4169 * Set valid bits inclusive of any overlap.
4170 */
4171 m->valid |= vm_page_bits(base, size);
4172 }
4173
4174 /*
4175 * Clear the given bits from the specified page's dirty field.
4176 */
4177 static __inline void
4178 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4179 {
4180 uintptr_t addr;
4181 #if PAGE_SIZE < 16384
4182 int shift;
4183 #endif
4184
4185 /*
4186 * If the object is locked and the page is neither exclusive busy nor
4187 * write mapped, then the page's dirty field cannot possibly be
4188 * set by a concurrent pmap operation.
4189 */
4190 VM_OBJECT_ASSERT_WLOCKED(m->object);
4191 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4192 m->dirty &= ~pagebits;
4193 else {
4194 /*
4195 * The pmap layer can call vm_page_dirty() without
4196 * holding a distinguished lock. The combination of
4197 * the object's lock and an atomic operation suffice
4198 * to guarantee consistency of the page dirty field.
4199 *
4200 * For PAGE_SIZE == 32768 case, compiler already
4201 * properly aligns the dirty field, so no forcible
4202 * alignment is needed. Only require existence of
4203 * atomic_clear_64 when page size is 32768.
4204 */
4205 addr = (uintptr_t)&m->dirty;
4206 #if PAGE_SIZE == 32768
4207 atomic_clear_64((uint64_t *)addr, pagebits);
4208 #elif PAGE_SIZE == 16384
4209 atomic_clear_32((uint32_t *)addr, pagebits);
4210 #else /* PAGE_SIZE <= 8192 */
4211 /*
4212 * Use a trick to perform a 32-bit atomic on the
4213 * containing aligned word, to not depend on the existence
4214 * of atomic_clear_{8, 16}.
4215 */
4216 shift = addr & (sizeof(uint32_t) - 1);
4217 #if BYTE_ORDER == BIG_ENDIAN
4218 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
4219 #else
4220 shift *= NBBY;
4221 #endif
4222 addr &= ~(sizeof(uint32_t) - 1);
4223 atomic_clear_32((uint32_t *)addr, pagebits << shift);
4224 #endif /* PAGE_SIZE */
4225 }
4226 }
4227
4228 /*
4229 * vm_page_set_validclean:
4230 *
4231 * Sets portions of a page valid and clean. The arguments are expected
4232 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4233 * of any partial chunks touched by the range. The invalid portion of
4234 * such chunks will be zero'd.
4235 *
4236 * (base + size) must be less then or equal to PAGE_SIZE.
4237 */
4238 void
4239 vm_page_set_validclean(vm_page_t m, int base, int size)
4240 {
4241 vm_page_bits_t oldvalid, pagebits;
4242 int endoff, frag;
4243
4244 VM_OBJECT_ASSERT_WLOCKED(m->object);
4245 if (size == 0) /* handle degenerate case */
4246 return;
4247
4248 /*
4249 * If the base is not DEV_BSIZE aligned and the valid
4250 * bit is clear, we have to zero out a portion of the
4251 * first block.
4252 */
4253 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4254 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4255 pmap_zero_page_area(m, frag, base - frag);
4256
4257 /*
4258 * If the ending offset is not DEV_BSIZE aligned and the
4259 * valid bit is clear, we have to zero out a portion of
4260 * the last block.
4261 */
4262 endoff = base + size;
4263 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4264 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4265 pmap_zero_page_area(m, endoff,
4266 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4267
4268 /*
4269 * Set valid, clear dirty bits. If validating the entire
4270 * page we can safely clear the pmap modify bit. We also
4271 * use this opportunity to clear the VPO_NOSYNC flag. If a process
4272 * takes a write fault on a MAP_NOSYNC memory area the flag will
4273 * be set again.
4274 *
4275 * We set valid bits inclusive of any overlap, but we can only
4276 * clear dirty bits for DEV_BSIZE chunks that are fully within
4277 * the range.
4278 */
4279 oldvalid = m->valid;
4280 pagebits = vm_page_bits(base, size);
4281 m->valid |= pagebits;
4282 #if 0 /* NOT YET */
4283 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4284 frag = DEV_BSIZE - frag;
4285 base += frag;
4286 size -= frag;
4287 if (size < 0)
4288 size = 0;
4289 }
4290 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4291 #endif
4292 if (base == 0 && size == PAGE_SIZE) {
4293 /*
4294 * The page can only be modified within the pmap if it is
4295 * mapped, and it can only be mapped if it was previously
4296 * fully valid.
4297 */
4298 if (oldvalid == VM_PAGE_BITS_ALL)
4299 /*
4300 * Perform the pmap_clear_modify() first. Otherwise,
4301 * a concurrent pmap operation, such as
4302 * pmap_protect(), could clear a modification in the
4303 * pmap and set the dirty field on the page before
4304 * pmap_clear_modify() had begun and after the dirty
4305 * field was cleared here.
4306 */
4307 pmap_clear_modify(m);
4308 m->dirty = 0;
4309 m->oflags &= ~VPO_NOSYNC;
4310 } else if (oldvalid != VM_PAGE_BITS_ALL)
4311 m->dirty &= ~pagebits;
4312 else
4313 vm_page_clear_dirty_mask(m, pagebits);
4314 }
4315
4316 void
4317 vm_page_clear_dirty(vm_page_t m, int base, int size)
4318 {
4319
4320 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4321 }
4322
4323 /*
4324 * vm_page_set_invalid:
4325 *
4326 * Invalidates DEV_BSIZE'd chunks within a page. Both the
4327 * valid and dirty bits for the effected areas are cleared.
4328 */
4329 void
4330 vm_page_set_invalid(vm_page_t m, int base, int size)
4331 {
4332 vm_page_bits_t bits;
4333 vm_object_t object;
4334
4335 object = m->object;
4336 VM_OBJECT_ASSERT_WLOCKED(object);
4337 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4338 size >= object->un_pager.vnp.vnp_size)
4339 bits = VM_PAGE_BITS_ALL;
4340 else
4341 bits = vm_page_bits(base, size);
4342 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
4343 bits != 0)
4344 pmap_remove_all(m);
4345 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
4346 !pmap_page_is_mapped(m),
4347 ("vm_page_set_invalid: page %p is mapped", m));
4348 m->valid &= ~bits;
4349 m->dirty &= ~bits;
4350 }
4351
4352 /*
4353 * vm_page_zero_invalid()
4354 *
4355 * The kernel assumes that the invalid portions of a page contain
4356 * garbage, but such pages can be mapped into memory by user code.
4357 * When this occurs, we must zero out the non-valid portions of the
4358 * page so user code sees what it expects.
4359 *
4360 * Pages are most often semi-valid when the end of a file is mapped
4361 * into memory and the file's size is not page aligned.
4362 */
4363 void
4364 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4365 {
4366 int b;
4367 int i;
4368
4369 VM_OBJECT_ASSERT_WLOCKED(m->object);
4370 /*
4371 * Scan the valid bits looking for invalid sections that
4372 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
4373 * valid bit may be set ) have already been zeroed by
4374 * vm_page_set_validclean().
4375 */
4376 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4377 if (i == (PAGE_SIZE / DEV_BSIZE) ||
4378 (m->valid & ((vm_page_bits_t)1 << i))) {
4379 if (i > b) {
4380 pmap_zero_page_area(m,
4381 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4382 }
4383 b = i + 1;
4384 }
4385 }
4386
4387 /*
4388 * setvalid is TRUE when we can safely set the zero'd areas
4389 * as being valid. We can do this if there are no cache consistency
4390 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
4391 */
4392 if (setvalid)
4393 m->valid = VM_PAGE_BITS_ALL;
4394 }
4395
4396 /*
4397 * vm_page_is_valid:
4398 *
4399 * Is (partial) page valid? Note that the case where size == 0
4400 * will return FALSE in the degenerate case where the page is
4401 * entirely invalid, and TRUE otherwise.
4402 */
4403 int
4404 vm_page_is_valid(vm_page_t m, int base, int size)
4405 {
4406 vm_page_bits_t bits;
4407
4408 VM_OBJECT_ASSERT_LOCKED(m->object);
4409 bits = vm_page_bits(base, size);
4410 return (m->valid != 0 && (m->valid & bits) == bits);
4411 }
4412
4413 /*
4414 * Returns true if all of the specified predicates are true for the entire
4415 * (super)page and false otherwise.
4416 */
4417 bool
4418 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4419 {
4420 vm_object_t object;
4421 int i, npages;
4422
4423 object = m->object;
4424 if (skip_m != NULL && skip_m->object != object)
4425 return (false);
4426 VM_OBJECT_ASSERT_LOCKED(object);
4427 npages = atop(pagesizes[m->psind]);
4428
4429 /*
4430 * The physically contiguous pages that make up a superpage, i.e., a
4431 * page with a page size index ("psind") greater than zero, will
4432 * occupy adjacent entries in vm_page_array[].
4433 */
4434 for (i = 0; i < npages; i++) {
4435 /* Always test object consistency, including "skip_m". */
4436 if (m[i].object != object)
4437 return (false);
4438 if (&m[i] == skip_m)
4439 continue;
4440 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4441 return (false);
4442 if ((flags & PS_ALL_DIRTY) != 0) {
4443 /*
4444 * Calling vm_page_test_dirty() or pmap_is_modified()
4445 * might stop this case from spuriously returning
4446 * "false". However, that would require a write lock
4447 * on the object containing "m[i]".
4448 */
4449 if (m[i].dirty != VM_PAGE_BITS_ALL)
4450 return (false);
4451 }
4452 if ((flags & PS_ALL_VALID) != 0 &&
4453 m[i].valid != VM_PAGE_BITS_ALL)
4454 return (false);
4455 }
4456 return (true);
4457 }
4458
4459 /*
4460 * Set the page's dirty bits if the page is modified.
4461 */
4462 void
4463 vm_page_test_dirty(vm_page_t m)
4464 {
4465
4466 VM_OBJECT_ASSERT_WLOCKED(m->object);
4467 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4468 vm_page_dirty(m);
4469 }
4470
4471 void
4472 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4473 {
4474
4475 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4476 }
4477
4478 void
4479 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4480 {
4481
4482 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4483 }
4484
4485 int
4486 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4487 {
4488
4489 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4490 }
4491
4492 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4493 void
4494 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4495 {
4496
4497 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4498 }
4499
4500 void
4501 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4502 {
4503
4504 mtx_assert_(vm_page_lockptr(m), a, file, line);
4505 }
4506 #endif
4507
4508 #ifdef INVARIANTS
4509 void
4510 vm_page_object_lock_assert(vm_page_t m)
4511 {
4512
4513 /*
4514 * Certain of the page's fields may only be modified by the
4515 * holder of the containing object's lock or the exclusive busy.
4516 * holder. Unfortunately, the holder of the write busy is
4517 * not recorded, and thus cannot be checked here.
4518 */
4519 if (m->object != NULL && !vm_page_xbusied(m))
4520 VM_OBJECT_ASSERT_WLOCKED(m->object);
4521 }
4522
4523 void
4524 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4525 {
4526
4527 if ((bits & PGA_WRITEABLE) == 0)
4528 return;
4529
4530 /*
4531 * The PGA_WRITEABLE flag can only be set if the page is
4532 * managed, is exclusively busied or the object is locked.
4533 * Currently, this flag is only set by pmap_enter().
4534 */
4535 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4536 ("PGA_WRITEABLE on unmanaged page"));
4537 if (!vm_page_xbusied(m))
4538 VM_OBJECT_ASSERT_LOCKED(m->object);
4539 }
4540 #endif
4541
4542 #include "opt_ddb.h"
4543 #ifdef DDB
4544 #include <sys/kernel.h>
4545
4546 #include <ddb/ddb.h>
4547
4548 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4549 {
4550
4551 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4552 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4553 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4554 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4555 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4556 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4557 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4558 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4559 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4560 }
4561
4562 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4563 {
4564 int dom;
4565
4566 db_printf("pq_free %d\n", vm_free_count());
4567 for (dom = 0; dom < vm_ndomains; dom++) {
4568 db_printf(
4569 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4570 dom,
4571 vm_dom[dom].vmd_page_count,
4572 vm_dom[dom].vmd_free_count,
4573 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4574 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4575 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4576 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4577 }
4578 }
4579
4580 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4581 {
4582 vm_page_t m;
4583 boolean_t phys, virt;
4584
4585 if (!have_addr) {
4586 db_printf("show pginfo addr\n");
4587 return;
4588 }
4589
4590 phys = strchr(modif, 'p') != NULL;
4591 virt = strchr(modif, 'v') != NULL;
4592 if (virt)
4593 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
4594 else if (phys)
4595 m = PHYS_TO_VM_PAGE(addr);
4596 else
4597 m = (vm_page_t)addr;
4598 db_printf(
4599 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4600 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4601 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4602 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4603 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4604 }
4605 #endif /* DDB */
Cache object: 55ad9211e3a70adfe0cc0cb0257568a1
|