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
36 /*-
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
39 *
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41 *
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
47 *
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51 *
52 * Carnegie Mellon requests users of this software to return to
53 *
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
58 *
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
61 */
62
63 /*
64 * Resident memory management module.
65 */
66
67 #include <sys/cdefs.h>
68 #include "opt_vm.h"
69
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/counter.h>
73 #include <sys/domainset.h>
74 #include <sys/kernel.h>
75 #include <sys/limits.h>
76 #include <sys/linker.h>
77 #include <sys/lock.h>
78 #include <sys/malloc.h>
79 #include <sys/mman.h>
80 #include <sys/msgbuf.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/rwlock.h>
84 #include <sys/sleepqueue.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/smp.h>
88 #include <sys/sysctl.h>
89 #include <sys/vmmeter.h>
90 #include <sys/vnode.h>
91
92 #include <vm/vm.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_param.h>
95 #include <vm/vm_domainset.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_dumpset.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110
111 #include <machine/md_var.h>
112
113 struct vm_domain vm_dom[MAXMEMDOM];
114
115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
116
117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
118
119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
120 /* The following fields are protected by the domainset lock. */
121 domainset_t __exclusive_cache_line vm_min_domains;
122 domainset_t __exclusive_cache_line vm_severe_domains;
123 static int vm_min_waiters;
124 static int vm_severe_waiters;
125 static int vm_pageproc_waiters;
126
127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
128 "VM page statistics");
129
130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
132 CTLFLAG_RD, &pqstate_commit_retries,
133 "Number of failed per-page atomic queue state updates");
134
135 static COUNTER_U64_DEFINE_EARLY(queue_ops);
136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
137 CTLFLAG_RD, &queue_ops,
138 "Number of batched queue operations");
139
140 static COUNTER_U64_DEFINE_EARLY(queue_nops);
141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
142 CTLFLAG_RD, &queue_nops,
143 "Number of batched queue operations with no effects");
144
145 /*
146 * bogus page -- for I/O to/from partially complete buffers,
147 * or for paging into sparsely invalid regions.
148 */
149 vm_page_t bogus_page;
150
151 vm_page_t vm_page_array;
152 long vm_page_array_size;
153 long first_page;
154
155 struct bitset *vm_page_dump;
156 long vm_page_dump_pages;
157
158 static TAILQ_HEAD(, vm_page) blacklist_head;
159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
161 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
162
163 static uma_zone_t fakepg_zone;
164
165 static void vm_page_alloc_check(vm_page_t m);
166 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
167 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
168 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
169 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
170 static bool vm_page_free_prep(vm_page_t m);
171 static void vm_page_free_toq(vm_page_t m);
172 static void vm_page_init(void *dummy);
173 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
174 vm_pindex_t pindex, vm_page_t mpred);
175 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
176 vm_page_t mpred);
177 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
178 const uint16_t nflag);
179 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
180 vm_page_t m_run, vm_paddr_t high);
181 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
182 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
183 int req);
184 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
185 int flags);
186 static void vm_page_zone_release(void *arg, void **store, int cnt);
187
188 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
189
190 static void
vm_page_init(void * dummy)191 vm_page_init(void *dummy)
192 {
193
194 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
195 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
196 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED);
197 }
198
199 static int pgcache_zone_max_pcpu;
200 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
201 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
202 "Per-CPU page cache size");
203
204 /*
205 * The cache page zone is initialized later since we need to be able to allocate
206 * pages before UMA is fully initialized.
207 */
208 static void
vm_page_init_cache_zones(void * dummy __unused)209 vm_page_init_cache_zones(void *dummy __unused)
210 {
211 struct vm_domain *vmd;
212 struct vm_pgcache *pgcache;
213 int cache, domain, maxcache, pool;
214
215 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
216 maxcache = pgcache_zone_max_pcpu * mp_ncpus;
217 for (domain = 0; domain < vm_ndomains; domain++) {
218 vmd = VM_DOMAIN(domain);
219 for (pool = 0; pool < VM_NFREEPOOL; pool++) {
220 pgcache = &vmd->vmd_pgcache[pool];
221 pgcache->domain = domain;
222 pgcache->pool = pool;
223 pgcache->zone = uma_zcache_create("vm pgcache",
224 PAGE_SIZE, NULL, NULL, NULL, NULL,
225 vm_page_zone_import, vm_page_zone_release, pgcache,
226 UMA_ZONE_VM);
227
228 /*
229 * Limit each pool's zone to 0.1% of the pages in the
230 * domain.
231 */
232 cache = maxcache != 0 ? maxcache :
233 vmd->vmd_page_count / 1000;
234 uma_zone_set_maxcache(pgcache->zone, cache);
235 }
236 }
237 }
238 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
239
240 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
241 #if PAGE_SIZE == 32768
242 #ifdef CTASSERT
243 CTASSERT(sizeof(u_long) >= 8);
244 #endif
245 #endif
246
247 /*
248 * vm_set_page_size:
249 *
250 * Sets the page size, perhaps based upon the memory
251 * size. Must be called before any use of page-size
252 * dependent functions.
253 */
254 void
vm_set_page_size(void)255 vm_set_page_size(void)
256 {
257 if (vm_cnt.v_page_size == 0)
258 vm_cnt.v_page_size = PAGE_SIZE;
259 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
260 panic("vm_set_page_size: page size not a power of two");
261 }
262
263 /*
264 * vm_page_blacklist_next:
265 *
266 * Find the next entry in the provided string of blacklist
267 * addresses. Entries are separated by space, comma, or newline.
268 * If an invalid integer is encountered then the rest of the
269 * string is skipped. Updates the list pointer to the next
270 * character, or NULL if the string is exhausted or invalid.
271 */
272 static vm_paddr_t
vm_page_blacklist_next(char ** list,char * end)273 vm_page_blacklist_next(char **list, char *end)
274 {
275 vm_paddr_t bad;
276 char *cp, *pos;
277
278 if (list == NULL || *list == NULL)
279 return (0);
280 if (**list =='\0') {
281 *list = NULL;
282 return (0);
283 }
284
285 /*
286 * If there's no end pointer then the buffer is coming from
287 * the kenv and we know it's null-terminated.
288 */
289 if (end == NULL)
290 end = *list + strlen(*list);
291
292 /* Ensure that strtoq() won't walk off the end */
293 if (*end != '\0') {
294 if (*end == '\n' || *end == ' ' || *end == ',')
295 *end = '\0';
296 else {
297 printf("Blacklist not terminated, skipping\n");
298 *list = NULL;
299 return (0);
300 }
301 }
302
303 for (pos = *list; *pos != '\0'; pos = cp) {
304 bad = strtoq(pos, &cp, 0);
305 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
306 if (bad == 0) {
307 if (++cp < end)
308 continue;
309 else
310 break;
311 }
312 } else
313 break;
314 if (*cp == '\0' || ++cp >= end)
315 *list = NULL;
316 else
317 *list = cp;
318 return (trunc_page(bad));
319 }
320 printf("Garbage in RAM blacklist, skipping\n");
321 *list = NULL;
322 return (0);
323 }
324
325 bool
vm_page_blacklist_add(vm_paddr_t pa,bool verbose)326 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
327 {
328 struct vm_domain *vmd;
329 vm_page_t m;
330 bool found;
331
332 m = vm_phys_paddr_to_vm_page(pa);
333 if (m == NULL)
334 return (true); /* page does not exist, no failure */
335
336 vmd = vm_pagequeue_domain(m);
337 vm_domain_free_lock(vmd);
338 found = vm_phys_unfree_page(m);
339 vm_domain_free_unlock(vmd);
340 if (found) {
341 vm_domain_freecnt_inc(vmd, -1);
342 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
343 if (verbose)
344 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
345 }
346 return (found);
347 }
348
349 /*
350 * vm_page_blacklist_check:
351 *
352 * Iterate through the provided string of blacklist addresses, pulling
353 * each entry out of the physical allocator free list and putting it
354 * onto a list for reporting via the vm.page_blacklist sysctl.
355 */
356 static void
vm_page_blacklist_check(char * list,char * end)357 vm_page_blacklist_check(char *list, char *end)
358 {
359 vm_paddr_t pa;
360 char *next;
361
362 next = list;
363 while (next != NULL) {
364 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
365 continue;
366 vm_page_blacklist_add(pa, bootverbose);
367 }
368 }
369
370 /*
371 * vm_page_blacklist_load:
372 *
373 * Search for a special module named "ram_blacklist". It'll be a
374 * plain text file provided by the user via the loader directive
375 * of the same name.
376 */
377 static void
vm_page_blacklist_load(char ** list,char ** end)378 vm_page_blacklist_load(char **list, char **end)
379 {
380 void *mod;
381 u_char *ptr;
382 u_int len;
383
384 mod = NULL;
385 ptr = NULL;
386
387 mod = preload_search_by_type("ram_blacklist");
388 if (mod != NULL) {
389 ptr = preload_fetch_addr(mod);
390 len = preload_fetch_size(mod);
391 }
392 *list = ptr;
393 if (ptr != NULL)
394 *end = ptr + len;
395 else
396 *end = NULL;
397 return;
398 }
399
400 static int
sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)401 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
402 {
403 vm_page_t m;
404 struct sbuf sbuf;
405 int error, first;
406
407 first = 1;
408 error = sysctl_wire_old_buffer(req, 0);
409 if (error != 0)
410 return (error);
411 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
412 TAILQ_FOREACH(m, &blacklist_head, listq) {
413 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
414 (uintmax_t)m->phys_addr);
415 first = 0;
416 }
417 error = sbuf_finish(&sbuf);
418 sbuf_delete(&sbuf);
419 return (error);
420 }
421
422 /*
423 * Initialize a dummy page for use in scans of the specified paging queue.
424 * In principle, this function only needs to set the flag PG_MARKER.
425 * Nonetheless, it write busies the page as a safety precaution.
426 */
427 void
vm_page_init_marker(vm_page_t marker,int queue,uint16_t aflags)428 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
429 {
430
431 bzero(marker, sizeof(*marker));
432 marker->flags = PG_MARKER;
433 marker->a.flags = aflags;
434 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
435 marker->a.queue = queue;
436 }
437
438 static void
vm_page_domain_init(int domain)439 vm_page_domain_init(int domain)
440 {
441 struct vm_domain *vmd;
442 struct vm_pagequeue *pq;
443 int i;
444
445 vmd = VM_DOMAIN(domain);
446 bzero(vmd, sizeof(*vmd));
447 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
448 "vm inactive pagequeue";
449 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
450 "vm active pagequeue";
451 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
452 "vm laundry pagequeue";
453 *__DECONST(const char **,
454 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
455 "vm unswappable pagequeue";
456 vmd->vmd_domain = domain;
457 vmd->vmd_page_count = 0;
458 vmd->vmd_free_count = 0;
459 vmd->vmd_segs = 0;
460 vmd->vmd_oom = FALSE;
461 for (i = 0; i < PQ_COUNT; i++) {
462 pq = &vmd->vmd_pagequeues[i];
463 TAILQ_INIT(&pq->pq_pl);
464 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
465 MTX_DEF | MTX_DUPOK);
466 pq->pq_pdpages = 0;
467 vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
468 }
469 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
470 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
471 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
472
473 /*
474 * inacthead is used to provide FIFO ordering for LRU-bypassing
475 * insertions.
476 */
477 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
478 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
479 &vmd->vmd_inacthead, plinks.q);
480
481 /*
482 * The clock pages are used to implement active queue scanning without
483 * requeues. Scans start at clock[0], which is advanced after the scan
484 * ends. When the two clock hands meet, they are reset and scanning
485 * resumes from the head of the queue.
486 */
487 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
488 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
489 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
490 &vmd->vmd_clock[0], plinks.q);
491 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
492 &vmd->vmd_clock[1], plinks.q);
493 }
494
495 /*
496 * Initialize a physical page in preparation for adding it to the free
497 * lists.
498 */
499 void
vm_page_init_page(vm_page_t m,vm_paddr_t pa,int segind)500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
501 {
502
503 m->object = NULL;
504 m->ref_count = 0;
505 m->busy_lock = VPB_FREED;
506 m->flags = m->a.flags = 0;
507 m->phys_addr = pa;
508 m->a.queue = PQ_NONE;
509 m->psind = 0;
510 m->segind = segind;
511 m->order = VM_NFREEORDER;
512 m->pool = VM_FREEPOOL_DEFAULT;
513 m->valid = m->dirty = 0;
514 pmap_page_init(m);
515 }
516
517 #ifndef PMAP_HAS_PAGE_ARRAY
518 static vm_paddr_t
vm_page_array_alloc(vm_offset_t * vaddr,vm_paddr_t end,vm_paddr_t page_range)519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
520 {
521 vm_paddr_t new_end;
522
523 /*
524 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
525 * However, because this page is allocated from KVM, out-of-bounds
526 * accesses using the direct map will not be trapped.
527 */
528 *vaddr += PAGE_SIZE;
529
530 /*
531 * Allocate physical memory for the page structures, and map it.
532 */
533 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
534 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
535 VM_PROT_READ | VM_PROT_WRITE);
536 vm_page_array_size = page_range;
537
538 return (new_end);
539 }
540 #endif
541
542 /*
543 * vm_page_startup:
544 *
545 * Initializes the resident memory module. Allocates physical memory for
546 * bootstrapping UMA and some data structures that are used to manage
547 * physical pages. Initializes these structures, and populates the free
548 * page queues.
549 */
550 vm_offset_t
vm_page_startup(vm_offset_t vaddr)551 vm_page_startup(vm_offset_t vaddr)
552 {
553 struct vm_phys_seg *seg;
554 struct vm_domain *vmd;
555 vm_page_t m;
556 char *list, *listend;
557 vm_paddr_t end, high_avail, low_avail, new_end, size;
558 vm_paddr_t page_range __unused;
559 vm_paddr_t last_pa, pa, startp, endp;
560 u_long pagecount;
561 #if MINIDUMP_PAGE_TRACKING
562 u_long vm_page_dump_size;
563 #endif
564 int biggestone, i, segind;
565 #ifdef WITNESS
566 vm_offset_t mapped;
567 int witness_size;
568 #endif
569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
570 long ii;
571 #endif
572
573 vaddr = round_page(vaddr);
574
575 vm_phys_early_startup();
576 biggestone = vm_phys_avail_largest();
577 end = phys_avail[biggestone+1];
578
579 /*
580 * Initialize the page and queue locks.
581 */
582 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
583 for (i = 0; i < PA_LOCK_COUNT; i++)
584 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
585 for (i = 0; i < vm_ndomains; i++)
586 vm_page_domain_init(i);
587
588 new_end = end;
589 #ifdef WITNESS
590 witness_size = round_page(witness_startup_count());
591 new_end -= witness_size;
592 mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
593 VM_PROT_READ | VM_PROT_WRITE);
594 bzero((void *)mapped, witness_size);
595 witness_startup((void *)mapped);
596 #endif
597
598 #if MINIDUMP_PAGE_TRACKING
599 /*
600 * Allocate a bitmap to indicate that a random physical page
601 * needs to be included in a minidump.
602 *
603 * The amd64 port needs this to indicate which direct map pages
604 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
605 *
606 * However, i386 still needs this workspace internally within the
607 * minidump code. In theory, they are not needed on i386, but are
608 * included should the sf_buf code decide to use them.
609 */
610 last_pa = 0;
611 vm_page_dump_pages = 0;
612 for (i = 0; dump_avail[i + 1] != 0; i += 2) {
613 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
614 dump_avail[i] / PAGE_SIZE;
615 if (dump_avail[i + 1] > last_pa)
616 last_pa = dump_avail[i + 1];
617 }
618 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
619 new_end -= vm_page_dump_size;
620 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
621 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
622 bzero((void *)vm_page_dump, vm_page_dump_size);
623 #else
624 (void)last_pa;
625 #endif
626 #if defined(__aarch64__) || defined(__amd64__) || \
627 defined(__riscv) || defined(__powerpc64__)
628 /*
629 * Include the UMA bootstrap pages, witness pages and vm_page_dump
630 * in a crash dump. When pmap_map() uses the direct map, they are
631 * not automatically included.
632 */
633 for (pa = new_end; pa < end; pa += PAGE_SIZE)
634 dump_add_page(pa);
635 #endif
636 phys_avail[biggestone + 1] = new_end;
637 #ifdef __amd64__
638 /*
639 * Request that the physical pages underlying the message buffer be
640 * included in a crash dump. Since the message buffer is accessed
641 * through the direct map, they are not automatically included.
642 */
643 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
644 last_pa = pa + round_page(msgbufsize);
645 while (pa < last_pa) {
646 dump_add_page(pa);
647 pa += PAGE_SIZE;
648 }
649 #endif
650 /*
651 * Compute the number of pages of memory that will be available for
652 * use, taking into account the overhead of a page structure per page.
653 * In other words, solve
654 * "available physical memory" - round_page(page_range *
655 * sizeof(struct vm_page)) = page_range * PAGE_SIZE
656 * for page_range.
657 */
658 low_avail = phys_avail[0];
659 high_avail = phys_avail[1];
660 for (i = 0; i < vm_phys_nsegs; i++) {
661 if (vm_phys_segs[i].start < low_avail)
662 low_avail = vm_phys_segs[i].start;
663 if (vm_phys_segs[i].end > high_avail)
664 high_avail = vm_phys_segs[i].end;
665 }
666 /* Skip the first chunk. It is already accounted for. */
667 for (i = 2; phys_avail[i + 1] != 0; i += 2) {
668 if (phys_avail[i] < low_avail)
669 low_avail = phys_avail[i];
670 if (phys_avail[i + 1] > high_avail)
671 high_avail = phys_avail[i + 1];
672 }
673 first_page = low_avail / PAGE_SIZE;
674 #ifdef VM_PHYSSEG_SPARSE
675 size = 0;
676 for (i = 0; i < vm_phys_nsegs; i++)
677 size += vm_phys_segs[i].end - vm_phys_segs[i].start;
678 for (i = 0; phys_avail[i + 1] != 0; i += 2)
679 size += phys_avail[i + 1] - phys_avail[i];
680 #elif defined(VM_PHYSSEG_DENSE)
681 size = high_avail - low_avail;
682 #else
683 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
684 #endif
685
686 #ifdef PMAP_HAS_PAGE_ARRAY
687 pmap_page_array_startup(size / PAGE_SIZE);
688 biggestone = vm_phys_avail_largest();
689 end = new_end = phys_avail[biggestone + 1];
690 #else
691 #ifdef VM_PHYSSEG_DENSE
692 /*
693 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
694 * the overhead of a page structure per page only if vm_page_array is
695 * allocated from the last physical memory chunk. Otherwise, we must
696 * allocate page structures representing the physical memory
697 * underlying vm_page_array, even though they will not be used.
698 */
699 if (new_end != high_avail)
700 page_range = size / PAGE_SIZE;
701 else
702 #endif
703 {
704 page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
705
706 /*
707 * If the partial bytes remaining are large enough for
708 * a page (PAGE_SIZE) without a corresponding
709 * 'struct vm_page', then new_end will contain an
710 * extra page after subtracting the length of the VM
711 * page array. Compensate by subtracting an extra
712 * page from new_end.
713 */
714 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
715 if (new_end == high_avail)
716 high_avail -= PAGE_SIZE;
717 new_end -= PAGE_SIZE;
718 }
719 }
720 end = new_end;
721 new_end = vm_page_array_alloc(&vaddr, end, page_range);
722 #endif
723
724 #if VM_NRESERVLEVEL > 0
725 /*
726 * Allocate physical memory for the reservation management system's
727 * data structures, and map it.
728 */
729 new_end = vm_reserv_startup(&vaddr, new_end);
730 #endif
731 #if defined(__aarch64__) || defined(__amd64__) || \
732 defined(__riscv) || defined(__powerpc64__)
733 /*
734 * Include vm_page_array and vm_reserv_array in a crash dump.
735 */
736 for (pa = new_end; pa < end; pa += PAGE_SIZE)
737 dump_add_page(pa);
738 #endif
739 phys_avail[biggestone + 1] = new_end;
740
741 /*
742 * Add physical memory segments corresponding to the available
743 * physical pages.
744 */
745 for (i = 0; phys_avail[i + 1] != 0; i += 2)
746 if (vm_phys_avail_size(i) != 0)
747 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
748
749 /*
750 * Initialize the physical memory allocator.
751 */
752 vm_phys_init();
753
754 /*
755 * Initialize the page structures and add every available page to the
756 * physical memory allocator's free lists.
757 */
758 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
759 for (ii = 0; ii < vm_page_array_size; ii++) {
760 m = &vm_page_array[ii];
761 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
762 m->flags = PG_FICTITIOUS;
763 }
764 #endif
765 vm_cnt.v_page_count = 0;
766 for (segind = 0; segind < vm_phys_nsegs; segind++) {
767 seg = &vm_phys_segs[segind];
768 for (m = seg->first_page, pa = seg->start; pa < seg->end;
769 m++, pa += PAGE_SIZE)
770 vm_page_init_page(m, pa, segind);
771
772 /*
773 * Add the segment's pages that are covered by one of
774 * phys_avail's ranges to the free lists.
775 */
776 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
777 if (seg->end <= phys_avail[i] ||
778 seg->start >= phys_avail[i + 1])
779 continue;
780
781 startp = MAX(seg->start, phys_avail[i]);
782 endp = MIN(seg->end, phys_avail[i + 1]);
783 pagecount = (u_long)atop(endp - startp);
784 if (pagecount == 0)
785 continue;
786
787 m = seg->first_page + atop(startp - seg->start);
788 vmd = VM_DOMAIN(seg->domain);
789 vm_domain_free_lock(vmd);
790 vm_phys_enqueue_contig(m, pagecount);
791 vm_domain_free_unlock(vmd);
792 vm_domain_freecnt_inc(vmd, pagecount);
793 vm_cnt.v_page_count += (u_int)pagecount;
794 vmd->vmd_page_count += (u_int)pagecount;
795 vmd->vmd_segs |= 1UL << segind;
796 }
797 }
798
799 /*
800 * Remove blacklisted pages from the physical memory allocator.
801 */
802 TAILQ_INIT(&blacklist_head);
803 vm_page_blacklist_load(&list, &listend);
804 vm_page_blacklist_check(list, listend);
805
806 list = kern_getenv("vm.blacklist");
807 vm_page_blacklist_check(list, NULL);
808
809 freeenv(list);
810 #if VM_NRESERVLEVEL > 0
811 /*
812 * Initialize the reservation management system.
813 */
814 vm_reserv_init();
815 #endif
816
817 return (vaddr);
818 }
819
820 void
vm_page_reference(vm_page_t m)821 vm_page_reference(vm_page_t m)
822 {
823
824 vm_page_aflag_set(m, PGA_REFERENCED);
825 }
826
827 /*
828 * vm_page_trybusy
829 *
830 * Helper routine for grab functions to trylock busy.
831 *
832 * Returns true on success and false on failure.
833 */
834 static bool
vm_page_trybusy(vm_page_t m,int allocflags)835 vm_page_trybusy(vm_page_t m, int allocflags)
836 {
837
838 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
839 return (vm_page_trysbusy(m));
840 else
841 return (vm_page_tryxbusy(m));
842 }
843
844 /*
845 * vm_page_tryacquire
846 *
847 * Helper routine for grab functions to trylock busy and wire.
848 *
849 * Returns true on success and false on failure.
850 */
851 static inline bool
vm_page_tryacquire(vm_page_t m,int allocflags)852 vm_page_tryacquire(vm_page_t m, int allocflags)
853 {
854 bool locked;
855
856 locked = vm_page_trybusy(m, allocflags);
857 if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
858 vm_page_wire(m);
859 return (locked);
860 }
861
862 /*
863 * vm_page_busy_acquire:
864 *
865 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
866 * and drop the object lock if necessary.
867 */
868 bool
vm_page_busy_acquire(vm_page_t m,int allocflags)869 vm_page_busy_acquire(vm_page_t m, int allocflags)
870 {
871 vm_object_t obj;
872 bool locked;
873
874 /*
875 * The page-specific object must be cached because page
876 * identity can change during the sleep, causing the
877 * re-lock of a different object.
878 * It is assumed that a reference to the object is already
879 * held by the callers.
880 */
881 obj = atomic_load_ptr(&m->object);
882 for (;;) {
883 if (vm_page_tryacquire(m, allocflags))
884 return (true);
885 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
886 return (false);
887 if (obj != NULL)
888 locked = VM_OBJECT_WOWNED(obj);
889 else
890 locked = false;
891 MPASS(locked || vm_page_wired(m));
892 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
893 locked) && locked)
894 VM_OBJECT_WLOCK(obj);
895 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
896 return (false);
897 KASSERT(m->object == obj || m->object == NULL,
898 ("vm_page_busy_acquire: page %p does not belong to %p",
899 m, obj));
900 }
901 }
902
903 /*
904 * vm_page_busy_downgrade:
905 *
906 * Downgrade an exclusive busy page into a single shared busy page.
907 */
908 void
vm_page_busy_downgrade(vm_page_t m)909 vm_page_busy_downgrade(vm_page_t m)
910 {
911 u_int x;
912
913 vm_page_assert_xbusied(m);
914
915 x = vm_page_busy_fetch(m);
916 for (;;) {
917 if (atomic_fcmpset_rel_int(&m->busy_lock,
918 &x, VPB_SHARERS_WORD(1)))
919 break;
920 }
921 if ((x & VPB_BIT_WAITERS) != 0)
922 wakeup(m);
923 }
924
925 /*
926 *
927 * vm_page_busy_tryupgrade:
928 *
929 * Attempt to upgrade a single shared busy into an exclusive busy.
930 */
931 int
vm_page_busy_tryupgrade(vm_page_t m)932 vm_page_busy_tryupgrade(vm_page_t m)
933 {
934 u_int ce, x;
935
936 vm_page_assert_sbusied(m);
937
938 x = vm_page_busy_fetch(m);
939 ce = VPB_CURTHREAD_EXCLUSIVE;
940 for (;;) {
941 if (VPB_SHARERS(x) > 1)
942 return (0);
943 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
944 ("vm_page_busy_tryupgrade: invalid lock state"));
945 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
946 ce | (x & VPB_BIT_WAITERS)))
947 continue;
948 return (1);
949 }
950 }
951
952 /*
953 * vm_page_sbusied:
954 *
955 * Return a positive value if the page is shared busied, 0 otherwise.
956 */
957 int
vm_page_sbusied(vm_page_t m)958 vm_page_sbusied(vm_page_t m)
959 {
960 u_int x;
961
962 x = vm_page_busy_fetch(m);
963 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
964 }
965
966 /*
967 * vm_page_sunbusy:
968 *
969 * Shared unbusy a page.
970 */
971 void
vm_page_sunbusy(vm_page_t m)972 vm_page_sunbusy(vm_page_t m)
973 {
974 u_int x;
975
976 vm_page_assert_sbusied(m);
977
978 x = vm_page_busy_fetch(m);
979 for (;;) {
980 KASSERT(x != VPB_FREED,
981 ("vm_page_sunbusy: Unlocking freed page."));
982 if (VPB_SHARERS(x) > 1) {
983 if (atomic_fcmpset_int(&m->busy_lock, &x,
984 x - VPB_ONE_SHARER))
985 break;
986 continue;
987 }
988 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
989 ("vm_page_sunbusy: invalid lock state"));
990 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
991 continue;
992 if ((x & VPB_BIT_WAITERS) == 0)
993 break;
994 wakeup(m);
995 break;
996 }
997 }
998
999 /*
1000 * vm_page_busy_sleep:
1001 *
1002 * Sleep if the page is busy, using the page pointer as wchan.
1003 * This is used to implement the hard-path of the busying mechanism.
1004 *
1005 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1006 * will not sleep if the page is shared-busy.
1007 *
1008 * The object lock must be held on entry.
1009 *
1010 * Returns true if it slept and dropped the object lock, or false
1011 * if there was no sleep and the lock is still held.
1012 */
1013 bool
vm_page_busy_sleep(vm_page_t m,const char * wmesg,int allocflags)1014 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1015 {
1016 vm_object_t obj;
1017
1018 obj = m->object;
1019 VM_OBJECT_ASSERT_LOCKED(obj);
1020
1021 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1022 true));
1023 }
1024
1025 /*
1026 * vm_page_busy_sleep_unlocked:
1027 *
1028 * Sleep if the page is busy, using the page pointer as wchan.
1029 * This is used to implement the hard-path of busying mechanism.
1030 *
1031 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1032 * will not sleep if the page is shared-busy.
1033 *
1034 * The object lock must not be held on entry. The operation will
1035 * return if the page changes identity.
1036 */
1037 void
vm_page_busy_sleep_unlocked(vm_object_t obj,vm_page_t m,vm_pindex_t pindex,const char * wmesg,int allocflags)1038 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1039 const char *wmesg, int allocflags)
1040 {
1041 VM_OBJECT_ASSERT_UNLOCKED(obj);
1042
1043 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1044 }
1045
1046 /*
1047 * _vm_page_busy_sleep:
1048 *
1049 * Internal busy sleep function. Verifies the page identity and
1050 * lockstate against parameters. Returns true if it sleeps and
1051 * false otherwise.
1052 *
1053 * allocflags uses VM_ALLOC_* flags to specify the lock required.
1054 *
1055 * If locked is true the lock will be dropped for any true returns
1056 * and held for any false returns.
1057 */
1058 static bool
_vm_page_busy_sleep(vm_object_t obj,vm_page_t m,vm_pindex_t pindex,const char * wmesg,int allocflags,bool locked)1059 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1060 const char *wmesg, int allocflags, bool locked)
1061 {
1062 bool xsleep;
1063 u_int x;
1064
1065 /*
1066 * If the object is busy we must wait for that to drain to zero
1067 * before trying the page again.
1068 */
1069 if (obj != NULL && vm_object_busied(obj)) {
1070 if (locked)
1071 VM_OBJECT_DROP(obj);
1072 vm_object_busy_wait(obj, wmesg);
1073 return (true);
1074 }
1075
1076 if (!vm_page_busied(m))
1077 return (false);
1078
1079 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1080 sleepq_lock(m);
1081 x = vm_page_busy_fetch(m);
1082 do {
1083 /*
1084 * If the page changes objects or becomes unlocked we can
1085 * simply return.
1086 */
1087 if (x == VPB_UNBUSIED ||
1088 (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1089 m->object != obj || m->pindex != pindex) {
1090 sleepq_release(m);
1091 return (false);
1092 }
1093 if ((x & VPB_BIT_WAITERS) != 0)
1094 break;
1095 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1096 if (locked)
1097 VM_OBJECT_DROP(obj);
1098 DROP_GIANT();
1099 sleepq_add(m, NULL, wmesg, 0, 0);
1100 sleepq_wait(m, PVM);
1101 PICKUP_GIANT();
1102 return (true);
1103 }
1104
1105 /*
1106 * vm_page_trysbusy:
1107 *
1108 * Try to shared busy a page.
1109 * If the operation succeeds 1 is returned otherwise 0.
1110 * The operation never sleeps.
1111 */
1112 int
vm_page_trysbusy(vm_page_t m)1113 vm_page_trysbusy(vm_page_t m)
1114 {
1115 vm_object_t obj;
1116 u_int x;
1117
1118 obj = m->object;
1119 x = vm_page_busy_fetch(m);
1120 for (;;) {
1121 if ((x & VPB_BIT_SHARED) == 0)
1122 return (0);
1123 /*
1124 * Reduce the window for transient busies that will trigger
1125 * false negatives in vm_page_ps_test().
1126 */
1127 if (obj != NULL && vm_object_busied(obj))
1128 return (0);
1129 if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1130 x + VPB_ONE_SHARER))
1131 break;
1132 }
1133
1134 /* Refetch the object now that we're guaranteed that it is stable. */
1135 obj = m->object;
1136 if (obj != NULL && vm_object_busied(obj)) {
1137 vm_page_sunbusy(m);
1138 return (0);
1139 }
1140 return (1);
1141 }
1142
1143 /*
1144 * vm_page_tryxbusy:
1145 *
1146 * Try to exclusive busy a page.
1147 * If the operation succeeds 1 is returned otherwise 0.
1148 * The operation never sleeps.
1149 */
1150 int
vm_page_tryxbusy(vm_page_t m)1151 vm_page_tryxbusy(vm_page_t m)
1152 {
1153 vm_object_t obj;
1154
1155 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1156 VPB_CURTHREAD_EXCLUSIVE) == 0)
1157 return (0);
1158
1159 obj = m->object;
1160 if (obj != NULL && vm_object_busied(obj)) {
1161 vm_page_xunbusy(m);
1162 return (0);
1163 }
1164 return (1);
1165 }
1166
1167 static void
vm_page_xunbusy_hard_tail(vm_page_t m)1168 vm_page_xunbusy_hard_tail(vm_page_t m)
1169 {
1170 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1171 /* Wake the waiter. */
1172 wakeup(m);
1173 }
1174
1175 /*
1176 * vm_page_xunbusy_hard:
1177 *
1178 * Called when unbusy has failed because there is a waiter.
1179 */
1180 void
vm_page_xunbusy_hard(vm_page_t m)1181 vm_page_xunbusy_hard(vm_page_t m)
1182 {
1183 vm_page_assert_xbusied(m);
1184 vm_page_xunbusy_hard_tail(m);
1185 }
1186
1187 void
vm_page_xunbusy_hard_unchecked(vm_page_t m)1188 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1189 {
1190 vm_page_assert_xbusied_unchecked(m);
1191 vm_page_xunbusy_hard_tail(m);
1192 }
1193
1194 static void
vm_page_busy_free(vm_page_t m)1195 vm_page_busy_free(vm_page_t m)
1196 {
1197 u_int x;
1198
1199 atomic_thread_fence_rel();
1200 x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1201 if ((x & VPB_BIT_WAITERS) != 0)
1202 wakeup(m);
1203 }
1204
1205 /*
1206 * vm_page_unhold_pages:
1207 *
1208 * Unhold each of the pages that is referenced by the given array.
1209 */
1210 void
vm_page_unhold_pages(vm_page_t * ma,int count)1211 vm_page_unhold_pages(vm_page_t *ma, int count)
1212 {
1213
1214 for (; count != 0; count--) {
1215 vm_page_unwire(*ma, PQ_ACTIVE);
1216 ma++;
1217 }
1218 }
1219
1220 vm_page_t
PHYS_TO_VM_PAGE(vm_paddr_t pa)1221 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1222 {
1223 vm_page_t m;
1224
1225 #ifdef VM_PHYSSEG_SPARSE
1226 m = vm_phys_paddr_to_vm_page(pa);
1227 if (m == NULL)
1228 m = vm_phys_fictitious_to_vm_page(pa);
1229 return (m);
1230 #elif defined(VM_PHYSSEG_DENSE)
1231 long pi;
1232
1233 pi = atop(pa);
1234 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1235 m = &vm_page_array[pi - first_page];
1236 return (m);
1237 }
1238 return (vm_phys_fictitious_to_vm_page(pa));
1239 #else
1240 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1241 #endif
1242 }
1243
1244 /*
1245 * vm_page_getfake:
1246 *
1247 * Create a fictitious page with the specified physical address and
1248 * memory attribute. The memory attribute is the only the machine-
1249 * dependent aspect of a fictitious page that must be initialized.
1250 */
1251 vm_page_t
vm_page_getfake(vm_paddr_t paddr,vm_memattr_t memattr)1252 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1253 {
1254 vm_page_t m;
1255
1256 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1257 vm_page_initfake(m, paddr, memattr);
1258 return (m);
1259 }
1260
1261 void
vm_page_initfake(vm_page_t m,vm_paddr_t paddr,vm_memattr_t memattr)1262 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1263 {
1264
1265 if ((m->flags & PG_FICTITIOUS) != 0) {
1266 /*
1267 * The page's memattr might have changed since the
1268 * previous initialization. Update the pmap to the
1269 * new memattr.
1270 */
1271 goto memattr;
1272 }
1273 m->phys_addr = paddr;
1274 m->a.queue = PQ_NONE;
1275 /* Fictitious pages don't use "segind". */
1276 m->flags = PG_FICTITIOUS;
1277 /* Fictitious pages don't use "order" or "pool". */
1278 m->oflags = VPO_UNMANAGED;
1279 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1280 /* Fictitious pages are unevictable. */
1281 m->ref_count = 1;
1282 pmap_page_init(m);
1283 memattr:
1284 pmap_page_set_memattr(m, memattr);
1285 }
1286
1287 /*
1288 * vm_page_putfake:
1289 *
1290 * Release a fictitious page.
1291 */
1292 void
vm_page_putfake(vm_page_t m)1293 vm_page_putfake(vm_page_t m)
1294 {
1295
1296 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1297 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1298 ("vm_page_putfake: bad page %p", m));
1299 vm_page_assert_xbusied(m);
1300 vm_page_busy_free(m);
1301 uma_zfree(fakepg_zone, m);
1302 }
1303
1304 /*
1305 * vm_page_updatefake:
1306 *
1307 * Update the given fictitious page to the specified physical address and
1308 * memory attribute.
1309 */
1310 void
vm_page_updatefake(vm_page_t m,vm_paddr_t paddr,vm_memattr_t memattr)1311 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1312 {
1313
1314 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1315 ("vm_page_updatefake: bad page %p", m));
1316 m->phys_addr = paddr;
1317 pmap_page_set_memattr(m, memattr);
1318 }
1319
1320 /*
1321 * vm_page_free:
1322 *
1323 * Free a page.
1324 */
1325 void
vm_page_free(vm_page_t m)1326 vm_page_free(vm_page_t m)
1327 {
1328
1329 m->flags &= ~PG_ZERO;
1330 vm_page_free_toq(m);
1331 }
1332
1333 /*
1334 * vm_page_free_zero:
1335 *
1336 * Free a page to the zerod-pages queue
1337 */
1338 void
vm_page_free_zero(vm_page_t m)1339 vm_page_free_zero(vm_page_t m)
1340 {
1341
1342 m->flags |= PG_ZERO;
1343 vm_page_free_toq(m);
1344 }
1345
1346 /*
1347 * Unbusy and handle the page queueing for a page from a getpages request that
1348 * was optionally read ahead or behind.
1349 */
1350 void
vm_page_readahead_finish(vm_page_t m)1351 vm_page_readahead_finish(vm_page_t m)
1352 {
1353
1354 /* We shouldn't put invalid pages on queues. */
1355 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1356
1357 /*
1358 * Since the page is not the actually needed one, whether it should
1359 * be activated or deactivated is not obvious. Empirical results
1360 * have shown that deactivating the page is usually the best choice,
1361 * unless the page is wanted by another thread.
1362 */
1363 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1364 vm_page_activate(m);
1365 else
1366 vm_page_deactivate(m);
1367 vm_page_xunbusy_unchecked(m);
1368 }
1369
1370 /*
1371 * Destroy the identity of an invalid page and free it if possible.
1372 * This is intended to be used when reading a page from backing store fails.
1373 */
1374 void
vm_page_free_invalid(vm_page_t m)1375 vm_page_free_invalid(vm_page_t m)
1376 {
1377
1378 KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1379 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1380 KASSERT(m->object != NULL, ("page %p has no object", m));
1381 VM_OBJECT_ASSERT_WLOCKED(m->object);
1382
1383 /*
1384 * We may be attempting to free the page as part of the handling for an
1385 * I/O error, in which case the page was xbusied by a different thread.
1386 */
1387 vm_page_xbusy_claim(m);
1388
1389 /*
1390 * If someone has wired this page while the object lock
1391 * was not held, then the thread that unwires is responsible
1392 * for freeing the page. Otherwise just free the page now.
1393 * The wire count of this unmapped page cannot change while
1394 * we have the page xbusy and the page's object wlocked.
1395 */
1396 if (vm_page_remove(m))
1397 vm_page_free(m);
1398 }
1399
1400 /*
1401 * vm_page_dirty_KBI: [ internal use only ]
1402 *
1403 * Set all bits in the page's dirty field.
1404 *
1405 * The object containing the specified page must be locked if the
1406 * call is made from the machine-independent layer.
1407 *
1408 * See vm_page_clear_dirty_mask().
1409 *
1410 * This function should only be called by vm_page_dirty().
1411 */
1412 void
vm_page_dirty_KBI(vm_page_t m)1413 vm_page_dirty_KBI(vm_page_t m)
1414 {
1415
1416 /* Refer to this operation by its public name. */
1417 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1418 m->dirty = VM_PAGE_BITS_ALL;
1419 }
1420
1421 /*
1422 * vm_page_insert: [ internal use only ]
1423 *
1424 * Inserts the given mem entry into the object and object list.
1425 *
1426 * The object must be locked.
1427 */
1428 int
vm_page_insert(vm_page_t m,vm_object_t object,vm_pindex_t pindex)1429 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1430 {
1431 vm_page_t mpred;
1432
1433 VM_OBJECT_ASSERT_WLOCKED(object);
1434 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1435 return (vm_page_insert_after(m, object, pindex, mpred));
1436 }
1437
1438 /*
1439 * vm_page_insert_after:
1440 *
1441 * Inserts the page "m" into the specified object at offset "pindex".
1442 *
1443 * The page "mpred" must immediately precede the offset "pindex" within
1444 * the specified object.
1445 *
1446 * The object must be locked.
1447 */
1448 static int
vm_page_insert_after(vm_page_t m,vm_object_t object,vm_pindex_t pindex,vm_page_t mpred)1449 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1450 vm_page_t mpred)
1451 {
1452 vm_page_t msucc;
1453
1454 VM_OBJECT_ASSERT_WLOCKED(object);
1455 KASSERT(m->object == NULL,
1456 ("vm_page_insert_after: page already inserted"));
1457 if (mpred != NULL) {
1458 KASSERT(mpred->object == object,
1459 ("vm_page_insert_after: object doesn't contain mpred"));
1460 KASSERT(mpred->pindex < pindex,
1461 ("vm_page_insert_after: mpred doesn't precede pindex"));
1462 msucc = TAILQ_NEXT(mpred, listq);
1463 } else
1464 msucc = TAILQ_FIRST(&object->memq);
1465 if (msucc != NULL)
1466 KASSERT(msucc->pindex > pindex,
1467 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1468
1469 /*
1470 * Record the object/offset pair in this page.
1471 */
1472 m->object = object;
1473 m->pindex = pindex;
1474 m->ref_count |= VPRC_OBJREF;
1475
1476 /*
1477 * Now link into the object's ordered list of backed pages.
1478 */
1479 if (vm_radix_insert(&object->rtree, m)) {
1480 m->object = NULL;
1481 m->pindex = 0;
1482 m->ref_count &= ~VPRC_OBJREF;
1483 return (1);
1484 }
1485 vm_page_insert_radixdone(m, object, mpred);
1486 vm_pager_page_inserted(object, m);
1487 return (0);
1488 }
1489
1490 /*
1491 * vm_page_insert_radixdone:
1492 *
1493 * Complete page "m" insertion into the specified object after the
1494 * radix trie hooking.
1495 *
1496 * The page "mpred" must precede the offset "m->pindex" within the
1497 * specified object.
1498 *
1499 * The object must be locked.
1500 */
1501 static void
vm_page_insert_radixdone(vm_page_t m,vm_object_t object,vm_page_t mpred)1502 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1503 {
1504
1505 VM_OBJECT_ASSERT_WLOCKED(object);
1506 KASSERT(object != NULL && m->object == object,
1507 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1508 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1509 ("vm_page_insert_radixdone: page %p is missing object ref", m));
1510 if (mpred != NULL) {
1511 KASSERT(mpred->object == object,
1512 ("vm_page_insert_radixdone: object doesn't contain mpred"));
1513 KASSERT(mpred->pindex < m->pindex,
1514 ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1515 }
1516
1517 if (mpred != NULL)
1518 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1519 else
1520 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1521
1522 /*
1523 * Show that the object has one more resident page.
1524 */
1525 object->resident_page_count++;
1526
1527 /*
1528 * Hold the vnode until the last page is released.
1529 */
1530 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1531 vhold(object->handle);
1532
1533 /*
1534 * Since we are inserting a new and possibly dirty page,
1535 * update the object's generation count.
1536 */
1537 if (pmap_page_is_write_mapped(m))
1538 vm_object_set_writeable_dirty(object);
1539 }
1540
1541 /*
1542 * Do the work to remove a page from its object. The caller is responsible for
1543 * updating the page's fields to reflect this removal.
1544 */
1545 static void
vm_page_object_remove(vm_page_t m)1546 vm_page_object_remove(vm_page_t m)
1547 {
1548 vm_object_t object;
1549 vm_page_t mrem __diagused;
1550
1551 vm_page_assert_xbusied(m);
1552 object = m->object;
1553 VM_OBJECT_ASSERT_WLOCKED(object);
1554 KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1555 ("page %p is missing its object ref", m));
1556
1557 /* Deferred free of swap space. */
1558 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1559 vm_pager_page_unswapped(m);
1560
1561 vm_pager_page_removed(object, m);
1562
1563 m->object = NULL;
1564 mrem = vm_radix_remove(&object->rtree, m->pindex);
1565 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1566
1567 /*
1568 * Now remove from the object's list of backed pages.
1569 */
1570 TAILQ_REMOVE(&object->memq, m, listq);
1571
1572 /*
1573 * And show that the object has one fewer resident page.
1574 */
1575 object->resident_page_count--;
1576
1577 /*
1578 * The vnode may now be recycled.
1579 */
1580 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1581 vdrop(object->handle);
1582 }
1583
1584 /*
1585 * vm_page_remove:
1586 *
1587 * Removes the specified page from its containing object, but does not
1588 * invalidate any backing storage. Returns true if the object's reference
1589 * was the last reference to the page, and false otherwise.
1590 *
1591 * The object must be locked and the page must be exclusively busied.
1592 * The exclusive busy will be released on return. If this is not the
1593 * final ref and the caller does not hold a wire reference it may not
1594 * continue to access the page.
1595 */
1596 bool
vm_page_remove(vm_page_t m)1597 vm_page_remove(vm_page_t m)
1598 {
1599 bool dropped;
1600
1601 dropped = vm_page_remove_xbusy(m);
1602 vm_page_xunbusy(m);
1603
1604 return (dropped);
1605 }
1606
1607 /*
1608 * vm_page_remove_xbusy
1609 *
1610 * Removes the page but leaves the xbusy held. Returns true if this
1611 * removed the final ref and false otherwise.
1612 */
1613 bool
vm_page_remove_xbusy(vm_page_t m)1614 vm_page_remove_xbusy(vm_page_t m)
1615 {
1616
1617 vm_page_object_remove(m);
1618 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1619 }
1620
1621 /*
1622 * vm_page_lookup:
1623 *
1624 * Returns the page associated with the object/offset
1625 * pair specified; if none is found, NULL is returned.
1626 *
1627 * The object must be locked.
1628 */
1629 vm_page_t
vm_page_lookup(vm_object_t object,vm_pindex_t pindex)1630 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1631 {
1632
1633 VM_OBJECT_ASSERT_LOCKED(object);
1634 return (vm_radix_lookup(&object->rtree, pindex));
1635 }
1636
1637 /*
1638 * vm_page_lookup_unlocked:
1639 *
1640 * Returns the page associated with the object/offset pair specified;
1641 * if none is found, NULL is returned. The page may be no longer be
1642 * present in the object at the time that this function returns. Only
1643 * useful for opportunistic checks such as inmem().
1644 */
1645 vm_page_t
vm_page_lookup_unlocked(vm_object_t object,vm_pindex_t pindex)1646 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1647 {
1648
1649 return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1650 }
1651
1652 /*
1653 * vm_page_relookup:
1654 *
1655 * Returns a page that must already have been busied by
1656 * the caller. Used for bogus page replacement.
1657 */
1658 vm_page_t
vm_page_relookup(vm_object_t object,vm_pindex_t pindex)1659 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1660 {
1661 vm_page_t m;
1662
1663 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1664 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1665 m->object == object && m->pindex == pindex,
1666 ("vm_page_relookup: Invalid page %p", m));
1667 return (m);
1668 }
1669
1670 /*
1671 * This should only be used by lockless functions for releasing transient
1672 * incorrect acquires. The page may have been freed after we acquired a
1673 * busy lock. In this case busy_lock == VPB_FREED and we have nothing
1674 * further to do.
1675 */
1676 static void
vm_page_busy_release(vm_page_t m)1677 vm_page_busy_release(vm_page_t m)
1678 {
1679 u_int x;
1680
1681 x = vm_page_busy_fetch(m);
1682 for (;;) {
1683 if (x == VPB_FREED)
1684 break;
1685 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1686 if (atomic_fcmpset_int(&m->busy_lock, &x,
1687 x - VPB_ONE_SHARER))
1688 break;
1689 continue;
1690 }
1691 KASSERT((x & VPB_BIT_SHARED) != 0 ||
1692 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1693 ("vm_page_busy_release: %p xbusy not owned.", m));
1694 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1695 continue;
1696 if ((x & VPB_BIT_WAITERS) != 0)
1697 wakeup(m);
1698 break;
1699 }
1700 }
1701
1702 /*
1703 * vm_page_find_least:
1704 *
1705 * Returns the page associated with the object with least pindex
1706 * greater than or equal to the parameter pindex, or NULL.
1707 *
1708 * The object must be locked.
1709 */
1710 vm_page_t
vm_page_find_least(vm_object_t object,vm_pindex_t pindex)1711 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1712 {
1713 vm_page_t m;
1714
1715 VM_OBJECT_ASSERT_LOCKED(object);
1716 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1717 m = vm_radix_lookup_ge(&object->rtree, pindex);
1718 return (m);
1719 }
1720
1721 /*
1722 * Returns the given page's successor (by pindex) within the object if it is
1723 * resident; if none is found, NULL is returned.
1724 *
1725 * The object must be locked.
1726 */
1727 vm_page_t
vm_page_next(vm_page_t m)1728 vm_page_next(vm_page_t m)
1729 {
1730 vm_page_t next;
1731
1732 VM_OBJECT_ASSERT_LOCKED(m->object);
1733 if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1734 MPASS(next->object == m->object);
1735 if (next->pindex != m->pindex + 1)
1736 next = NULL;
1737 }
1738 return (next);
1739 }
1740
1741 /*
1742 * Returns the given page's predecessor (by pindex) within the object if it is
1743 * resident; if none is found, NULL is returned.
1744 *
1745 * The object must be locked.
1746 */
1747 vm_page_t
vm_page_prev(vm_page_t m)1748 vm_page_prev(vm_page_t m)
1749 {
1750 vm_page_t prev;
1751
1752 VM_OBJECT_ASSERT_LOCKED(m->object);
1753 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1754 MPASS(prev->object == m->object);
1755 if (prev->pindex != m->pindex - 1)
1756 prev = NULL;
1757 }
1758 return (prev);
1759 }
1760
1761 /*
1762 * Uses the page mnew as a replacement for an existing page at index
1763 * pindex which must be already present in the object.
1764 *
1765 * Both pages must be exclusively busied on enter. The old page is
1766 * unbusied on exit.
1767 *
1768 * A return value of true means mold is now free. If this is not the
1769 * final ref and the caller does not hold a wire reference it may not
1770 * continue to access the page.
1771 */
1772 static bool
vm_page_replace_hold(vm_page_t mnew,vm_object_t object,vm_pindex_t pindex,vm_page_t mold)1773 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1774 vm_page_t mold)
1775 {
1776 vm_page_t mret __diagused;
1777 bool dropped;
1778
1779 VM_OBJECT_ASSERT_WLOCKED(object);
1780 vm_page_assert_xbusied(mold);
1781 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1782 ("vm_page_replace: page %p already in object", mnew));
1783
1784 /*
1785 * This function mostly follows vm_page_insert() and
1786 * vm_page_remove() without the radix, object count and vnode
1787 * dance. Double check such functions for more comments.
1788 */
1789
1790 mnew->object = object;
1791 mnew->pindex = pindex;
1792 atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1793 mret = vm_radix_replace(&object->rtree, mnew);
1794 KASSERT(mret == mold,
1795 ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1796 KASSERT((mold->oflags & VPO_UNMANAGED) ==
1797 (mnew->oflags & VPO_UNMANAGED),
1798 ("vm_page_replace: mismatched VPO_UNMANAGED"));
1799
1800 /* Keep the resident page list in sorted order. */
1801 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1802 TAILQ_REMOVE(&object->memq, mold, listq);
1803 mold->object = NULL;
1804
1805 /*
1806 * The object's resident_page_count does not change because we have
1807 * swapped one page for another, but the generation count should
1808 * change if the page is dirty.
1809 */
1810 if (pmap_page_is_write_mapped(mnew))
1811 vm_object_set_writeable_dirty(object);
1812 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1813 vm_page_xunbusy(mold);
1814
1815 return (dropped);
1816 }
1817
1818 void
vm_page_replace(vm_page_t mnew,vm_object_t object,vm_pindex_t pindex,vm_page_t mold)1819 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1820 vm_page_t mold)
1821 {
1822
1823 vm_page_assert_xbusied(mnew);
1824
1825 if (vm_page_replace_hold(mnew, object, pindex, mold))
1826 vm_page_free(mold);
1827 }
1828
1829 /*
1830 * vm_page_rename:
1831 *
1832 * Move the given memory entry from its
1833 * current object to the specified target object/offset.
1834 *
1835 * Note: swap associated with the page must be invalidated by the move. We
1836 * have to do this for several reasons: (1) we aren't freeing the
1837 * page, (2) we are dirtying the page, (3) the VM system is probably
1838 * moving the page from object A to B, and will then later move
1839 * the backing store from A to B and we can't have a conflict.
1840 *
1841 * Note: we *always* dirty the page. It is necessary both for the
1842 * fact that we moved it, and because we may be invalidating
1843 * swap.
1844 *
1845 * The objects must be locked.
1846 */
1847 int
vm_page_rename(vm_page_t m,vm_object_t new_object,vm_pindex_t new_pindex)1848 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1849 {
1850 vm_page_t mpred;
1851 vm_pindex_t opidx;
1852
1853 VM_OBJECT_ASSERT_WLOCKED(new_object);
1854
1855 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1856 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1857 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1858 ("vm_page_rename: pindex already renamed"));
1859
1860 /*
1861 * Create a custom version of vm_page_insert() which does not depend
1862 * by m_prev and can cheat on the implementation aspects of the
1863 * function.
1864 */
1865 opidx = m->pindex;
1866 m->pindex = new_pindex;
1867 if (vm_radix_insert(&new_object->rtree, m)) {
1868 m->pindex = opidx;
1869 return (1);
1870 }
1871
1872 /*
1873 * The operation cannot fail anymore. The removal must happen before
1874 * the listq iterator is tainted.
1875 */
1876 m->pindex = opidx;
1877 vm_page_object_remove(m);
1878
1879 /* Return back to the new pindex to complete vm_page_insert(). */
1880 m->pindex = new_pindex;
1881 m->object = new_object;
1882
1883 vm_page_insert_radixdone(m, new_object, mpred);
1884 vm_page_dirty(m);
1885 vm_pager_page_inserted(new_object, m);
1886 return (0);
1887 }
1888
1889 /*
1890 * vm_page_alloc:
1891 *
1892 * Allocate and return a page that is associated with the specified
1893 * object and offset pair. By default, this page is exclusive busied.
1894 *
1895 * The caller must always specify an allocation class.
1896 *
1897 * allocation classes:
1898 * VM_ALLOC_NORMAL normal process request
1899 * VM_ALLOC_SYSTEM system *really* needs a page
1900 * VM_ALLOC_INTERRUPT interrupt time request
1901 *
1902 * optional allocation flags:
1903 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1904 * intends to allocate
1905 * VM_ALLOC_NOBUSY do not exclusive busy the page
1906 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1907 * VM_ALLOC_SBUSY shared busy the allocated page
1908 * VM_ALLOC_WIRED wire the allocated page
1909 * VM_ALLOC_ZERO prefer a zeroed page
1910 */
1911 vm_page_t
vm_page_alloc(vm_object_t object,vm_pindex_t pindex,int req)1912 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1913 {
1914
1915 return (vm_page_alloc_after(object, pindex, req,
1916 vm_radix_lookup_le(&object->rtree, pindex)));
1917 }
1918
1919 vm_page_t
vm_page_alloc_domain(vm_object_t object,vm_pindex_t pindex,int domain,int req)1920 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1921 int req)
1922 {
1923
1924 return (vm_page_alloc_domain_after(object, pindex, domain, req,
1925 vm_radix_lookup_le(&object->rtree, pindex)));
1926 }
1927
1928 /*
1929 * Allocate a page in the specified object with the given page index. To
1930 * optimize insertion of the page into the object, the caller must also specifiy
1931 * the resident page in the object with largest index smaller than the given
1932 * page index, or NULL if no such page exists.
1933 */
1934 vm_page_t
vm_page_alloc_after(vm_object_t object,vm_pindex_t pindex,int req,vm_page_t mpred)1935 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1936 int req, vm_page_t mpred)
1937 {
1938 struct vm_domainset_iter di;
1939 vm_page_t m;
1940 int domain;
1941
1942 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1943 do {
1944 m = vm_page_alloc_domain_after(object, pindex, domain, req,
1945 mpred);
1946 if (m != NULL)
1947 break;
1948 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
1949
1950 return (m);
1951 }
1952
1953 /*
1954 * Returns true if the number of free pages exceeds the minimum
1955 * for the request class and false otherwise.
1956 */
1957 static int
_vm_domain_allocate(struct vm_domain * vmd,int req_class,int npages)1958 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1959 {
1960 u_int limit, old, new;
1961
1962 if (req_class == VM_ALLOC_INTERRUPT)
1963 limit = 0;
1964 else if (req_class == VM_ALLOC_SYSTEM)
1965 limit = vmd->vmd_interrupt_free_min;
1966 else
1967 limit = vmd->vmd_free_reserved;
1968
1969 /*
1970 * Attempt to reserve the pages. Fail if we're below the limit.
1971 */
1972 limit += npages;
1973 old = vmd->vmd_free_count;
1974 do {
1975 if (old < limit)
1976 return (0);
1977 new = old - npages;
1978 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1979
1980 /* Wake the page daemon if we've crossed the threshold. */
1981 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1982 pagedaemon_wakeup(vmd->vmd_domain);
1983
1984 /* Only update bitsets on transitions. */
1985 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1986 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1987 vm_domain_set(vmd);
1988
1989 return (1);
1990 }
1991
1992 int
vm_domain_allocate(struct vm_domain * vmd,int req,int npages)1993 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1994 {
1995 int req_class;
1996
1997 /*
1998 * The page daemon is allowed to dig deeper into the free page list.
1999 */
2000 req_class = req & VM_ALLOC_CLASS_MASK;
2001 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2002 req_class = VM_ALLOC_SYSTEM;
2003 return (_vm_domain_allocate(vmd, req_class, npages));
2004 }
2005
2006 vm_page_t
vm_page_alloc_domain_after(vm_object_t object,vm_pindex_t pindex,int domain,int req,vm_page_t mpred)2007 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2008 int req, vm_page_t mpred)
2009 {
2010 struct vm_domain *vmd;
2011 vm_page_t m;
2012 int flags;
2013
2014 #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \
2015 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \
2016 VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \
2017 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2018 KASSERT((req & ~VPA_FLAGS) == 0,
2019 ("invalid request %#x", req));
2020 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2021 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2022 ("invalid request %#x", req));
2023 KASSERT(mpred == NULL || mpred->pindex < pindex,
2024 ("mpred %p doesn't precede pindex 0x%jx", mpred,
2025 (uintmax_t)pindex));
2026 VM_OBJECT_ASSERT_WLOCKED(object);
2027
2028 flags = 0;
2029 m = NULL;
2030 if (!vm_pager_can_alloc_page(object, pindex))
2031 return (NULL);
2032 again:
2033 #if VM_NRESERVLEVEL > 0
2034 /*
2035 * Can we allocate the page from a reservation?
2036 */
2037 if (vm_object_reserv(object) &&
2038 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2039 NULL) {
2040 goto found;
2041 }
2042 #endif
2043 vmd = VM_DOMAIN(domain);
2044 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2045 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2046 M_NOWAIT | M_NOVM);
2047 if (m != NULL) {
2048 flags |= PG_PCPU_CACHE;
2049 goto found;
2050 }
2051 }
2052 if (vm_domain_allocate(vmd, req, 1)) {
2053 /*
2054 * If not, allocate it from the free page queues.
2055 */
2056 vm_domain_free_lock(vmd);
2057 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2058 vm_domain_free_unlock(vmd);
2059 if (m == NULL) {
2060 vm_domain_freecnt_inc(vmd, 1);
2061 #if VM_NRESERVLEVEL > 0
2062 if (vm_reserv_reclaim_inactive(domain))
2063 goto again;
2064 #endif
2065 }
2066 }
2067 if (m == NULL) {
2068 /*
2069 * Not allocatable, give up.
2070 */
2071 if (vm_domain_alloc_fail(vmd, object, req))
2072 goto again;
2073 return (NULL);
2074 }
2075
2076 /*
2077 * At this point we had better have found a good page.
2078 */
2079 found:
2080 vm_page_dequeue(m);
2081 vm_page_alloc_check(m);
2082
2083 /*
2084 * Initialize the page. Only the PG_ZERO flag is inherited.
2085 */
2086 flags |= m->flags & PG_ZERO;
2087 if ((req & VM_ALLOC_NODUMP) != 0)
2088 flags |= PG_NODUMP;
2089 m->flags = flags;
2090 m->a.flags = 0;
2091 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2092 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2093 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2094 else if ((req & VM_ALLOC_SBUSY) != 0)
2095 m->busy_lock = VPB_SHARERS_WORD(1);
2096 else
2097 m->busy_lock = VPB_UNBUSIED;
2098 if (req & VM_ALLOC_WIRED) {
2099 vm_wire_add(1);
2100 m->ref_count = 1;
2101 }
2102 m->a.act_count = 0;
2103
2104 if (vm_page_insert_after(m, object, pindex, mpred)) {
2105 if (req & VM_ALLOC_WIRED) {
2106 vm_wire_sub(1);
2107 m->ref_count = 0;
2108 }
2109 KASSERT(m->object == NULL, ("page %p has object", m));
2110 m->oflags = VPO_UNMANAGED;
2111 m->busy_lock = VPB_UNBUSIED;
2112 /* Don't change PG_ZERO. */
2113 vm_page_free_toq(m);
2114 if (req & VM_ALLOC_WAITFAIL) {
2115 VM_OBJECT_WUNLOCK(object);
2116 vm_radix_wait();
2117 VM_OBJECT_WLOCK(object);
2118 }
2119 return (NULL);
2120 }
2121
2122 /* Ignore device objects; the pager sets "memattr" for them. */
2123 if (object->memattr != VM_MEMATTR_DEFAULT &&
2124 (object->flags & OBJ_FICTITIOUS) == 0)
2125 pmap_page_set_memattr(m, object->memattr);
2126
2127 return (m);
2128 }
2129
2130 /*
2131 * vm_page_alloc_contig:
2132 *
2133 * Allocate a contiguous set of physical pages of the given size "npages"
2134 * from the free lists. All of the physical pages must be at or above
2135 * the given physical address "low" and below the given physical address
2136 * "high". The given value "alignment" determines the alignment of the
2137 * first physical page in the set. If the given value "boundary" is
2138 * non-zero, then the set of physical pages cannot cross any physical
2139 * address boundary that is a multiple of that value. Both "alignment"
2140 * and "boundary" must be a power of two.
2141 *
2142 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2143 * then the memory attribute setting for the physical pages is configured
2144 * to the object's memory attribute setting. Otherwise, the memory
2145 * attribute setting for the physical pages is configured to "memattr",
2146 * overriding the object's memory attribute setting. However, if the
2147 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2148 * memory attribute setting for the physical pages cannot be configured
2149 * to VM_MEMATTR_DEFAULT.
2150 *
2151 * The specified object may not contain fictitious pages.
2152 *
2153 * The caller must always specify an allocation class.
2154 *
2155 * allocation classes:
2156 * VM_ALLOC_NORMAL normal process request
2157 * VM_ALLOC_SYSTEM system *really* needs a page
2158 * VM_ALLOC_INTERRUPT interrupt time request
2159 *
2160 * optional allocation flags:
2161 * VM_ALLOC_NOBUSY do not exclusive busy the page
2162 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
2163 * VM_ALLOC_SBUSY shared busy the allocated page
2164 * VM_ALLOC_WIRED wire the allocated page
2165 * VM_ALLOC_ZERO prefer a zeroed page
2166 */
2167 vm_page_t
vm_page_alloc_contig(vm_object_t object,vm_pindex_t pindex,int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)2168 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2169 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2170 vm_paddr_t boundary, vm_memattr_t memattr)
2171 {
2172 struct vm_domainset_iter di;
2173 vm_page_t bounds[2];
2174 vm_page_t m;
2175 int domain;
2176 int start_segind;
2177
2178 start_segind = -1;
2179
2180 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2181 do {
2182 m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2183 npages, low, high, alignment, boundary, memattr);
2184 if (m != NULL)
2185 break;
2186 if (start_segind == -1)
2187 start_segind = vm_phys_lookup_segind(low);
2188 if (vm_phys_find_range(bounds, start_segind, domain,
2189 npages, low, high) == -1) {
2190 vm_domainset_iter_ignore(&di, domain);
2191 }
2192 } while (vm_domainset_iter_page(&di, object, &domain) == 0);
2193
2194 return (m);
2195 }
2196
2197 static vm_page_t
vm_page_find_contig_domain(int domain,int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary)2198 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2199 vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2200 {
2201 struct vm_domain *vmd;
2202 vm_page_t m_ret;
2203
2204 /*
2205 * Can we allocate the pages without the number of free pages falling
2206 * below the lower bound for the allocation class?
2207 */
2208 vmd = VM_DOMAIN(domain);
2209 if (!vm_domain_allocate(vmd, req, npages))
2210 return (NULL);
2211 /*
2212 * Try to allocate the pages from the free page queues.
2213 */
2214 vm_domain_free_lock(vmd);
2215 m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2216 alignment, boundary);
2217 vm_domain_free_unlock(vmd);
2218 if (m_ret != NULL)
2219 return (m_ret);
2220 #if VM_NRESERVLEVEL > 0
2221 /*
2222 * Try to break a reservation to allocate the pages.
2223 */
2224 if ((req & VM_ALLOC_NORECLAIM) == 0) {
2225 m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2226 high, alignment, boundary);
2227 if (m_ret != NULL)
2228 return (m_ret);
2229 }
2230 #endif
2231 vm_domain_freecnt_inc(vmd, npages);
2232 return (NULL);
2233 }
2234
2235 vm_page_t
vm_page_alloc_contig_domain(vm_object_t object,vm_pindex_t pindex,int domain,int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)2236 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2237 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2238 vm_paddr_t boundary, vm_memattr_t memattr)
2239 {
2240 vm_page_t m, m_ret, mpred;
2241 u_int busy_lock, flags, oflags;
2242
2243 #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM)
2244 KASSERT((req & ~VPAC_FLAGS) == 0,
2245 ("invalid request %#x", req));
2246 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2247 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2248 ("invalid request %#x", req));
2249 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2250 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2251 ("invalid request %#x", req));
2252 VM_OBJECT_ASSERT_WLOCKED(object);
2253 KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2254 ("vm_page_alloc_contig: object %p has fictitious pages",
2255 object));
2256 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2257
2258 mpred = vm_radix_lookup_le(&object->rtree, pindex);
2259 KASSERT(mpred == NULL || mpred->pindex != pindex,
2260 ("vm_page_alloc_contig: pindex already allocated"));
2261 for (;;) {
2262 #if VM_NRESERVLEVEL > 0
2263 /*
2264 * Can we allocate the pages from a reservation?
2265 */
2266 if (vm_object_reserv(object) &&
2267 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2268 mpred, npages, low, high, alignment, boundary)) != NULL) {
2269 break;
2270 }
2271 #endif
2272 if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2273 low, high, alignment, boundary)) != NULL)
2274 break;
2275 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2276 return (NULL);
2277 }
2278 for (m = m_ret; m < &m_ret[npages]; m++) {
2279 vm_page_dequeue(m);
2280 vm_page_alloc_check(m);
2281 }
2282
2283 /*
2284 * Initialize the pages. Only the PG_ZERO flag is inherited.
2285 */
2286 flags = PG_ZERO;
2287 if ((req & VM_ALLOC_NODUMP) != 0)
2288 flags |= PG_NODUMP;
2289 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2290 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2291 busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2292 else if ((req & VM_ALLOC_SBUSY) != 0)
2293 busy_lock = VPB_SHARERS_WORD(1);
2294 else
2295 busy_lock = VPB_UNBUSIED;
2296 if ((req & VM_ALLOC_WIRED) != 0)
2297 vm_wire_add(npages);
2298 if (object->memattr != VM_MEMATTR_DEFAULT &&
2299 memattr == VM_MEMATTR_DEFAULT)
2300 memattr = object->memattr;
2301 for (m = m_ret; m < &m_ret[npages]; m++) {
2302 m->a.flags = 0;
2303 m->flags = (m->flags | PG_NODUMP) & flags;
2304 m->busy_lock = busy_lock;
2305 if ((req & VM_ALLOC_WIRED) != 0)
2306 m->ref_count = 1;
2307 m->a.act_count = 0;
2308 m->oflags = oflags;
2309 if (vm_page_insert_after(m, object, pindex, mpred)) {
2310 if ((req & VM_ALLOC_WIRED) != 0)
2311 vm_wire_sub(npages);
2312 KASSERT(m->object == NULL,
2313 ("page %p has object", m));
2314 mpred = m;
2315 for (m = m_ret; m < &m_ret[npages]; m++) {
2316 if (m <= mpred &&
2317 (req & VM_ALLOC_WIRED) != 0)
2318 m->ref_count = 0;
2319 m->oflags = VPO_UNMANAGED;
2320 m->busy_lock = VPB_UNBUSIED;
2321 /* Don't change PG_ZERO. */
2322 vm_page_free_toq(m);
2323 }
2324 if (req & VM_ALLOC_WAITFAIL) {
2325 VM_OBJECT_WUNLOCK(object);
2326 vm_radix_wait();
2327 VM_OBJECT_WLOCK(object);
2328 }
2329 return (NULL);
2330 }
2331 mpred = m;
2332 if (memattr != VM_MEMATTR_DEFAULT)
2333 pmap_page_set_memattr(m, memattr);
2334 pindex++;
2335 }
2336 return (m_ret);
2337 }
2338
2339 /*
2340 * Allocate a physical page that is not intended to be inserted into a VM
2341 * object. If the "freelist" parameter is not equal to VM_NFREELIST, then only
2342 * pages from the specified vm_phys freelist will be returned.
2343 */
2344 static __always_inline vm_page_t
_vm_page_alloc_noobj_domain(int domain,const int freelist,int req)2345 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req)
2346 {
2347 struct vm_domain *vmd;
2348 vm_page_t m;
2349 int flags;
2350
2351 #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \
2352 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \
2353 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \
2354 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2355 KASSERT((req & ~VPAN_FLAGS) == 0,
2356 ("invalid request %#x", req));
2357
2358 flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0;
2359 vmd = VM_DOMAIN(domain);
2360 again:
2361 if (freelist == VM_NFREELIST &&
2362 vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2363 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2364 M_NOWAIT | M_NOVM);
2365 if (m != NULL) {
2366 flags |= PG_PCPU_CACHE;
2367 goto found;
2368 }
2369 }
2370
2371 if (vm_domain_allocate(vmd, req, 1)) {
2372 vm_domain_free_lock(vmd);
2373 if (freelist == VM_NFREELIST)
2374 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2375 else
2376 m = vm_phys_alloc_freelist_pages(domain, freelist,
2377 VM_FREEPOOL_DIRECT, 0);
2378 vm_domain_free_unlock(vmd);
2379 if (m == NULL) {
2380 vm_domain_freecnt_inc(vmd, 1);
2381 #if VM_NRESERVLEVEL > 0
2382 if (freelist == VM_NFREELIST &&
2383 vm_reserv_reclaim_inactive(domain))
2384 goto again;
2385 #endif
2386 }
2387 }
2388 if (m == NULL) {
2389 if (vm_domain_alloc_fail(vmd, NULL, req))
2390 goto again;
2391 return (NULL);
2392 }
2393
2394 found:
2395 vm_page_dequeue(m);
2396 vm_page_alloc_check(m);
2397
2398 /*
2399 * Consumers should not rely on a useful default pindex value.
2400 */
2401 m->pindex = 0xdeadc0dedeadc0de;
2402 m->flags = (m->flags & PG_ZERO) | flags;
2403 m->a.flags = 0;
2404 m->oflags = VPO_UNMANAGED;
2405 m->busy_lock = VPB_UNBUSIED;
2406 if ((req & VM_ALLOC_WIRED) != 0) {
2407 vm_wire_add(1);
2408 m->ref_count = 1;
2409 }
2410
2411 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2412 pmap_zero_page(m);
2413
2414 return (m);
2415 }
2416
2417 vm_page_t
vm_page_alloc_freelist(int freelist,int req)2418 vm_page_alloc_freelist(int freelist, int req)
2419 {
2420 struct vm_domainset_iter di;
2421 vm_page_t m;
2422 int domain;
2423
2424 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2425 do {
2426 m = vm_page_alloc_freelist_domain(domain, freelist, req);
2427 if (m != NULL)
2428 break;
2429 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2430
2431 return (m);
2432 }
2433
2434 vm_page_t
vm_page_alloc_freelist_domain(int domain,int freelist,int req)2435 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2436 {
2437 KASSERT(freelist >= 0 && freelist < VM_NFREELIST,
2438 ("%s: invalid freelist %d", __func__, freelist));
2439
2440 return (_vm_page_alloc_noobj_domain(domain, freelist, req));
2441 }
2442
2443 vm_page_t
vm_page_alloc_noobj(int req)2444 vm_page_alloc_noobj(int req)
2445 {
2446 struct vm_domainset_iter di;
2447 vm_page_t m;
2448 int domain;
2449
2450 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2451 do {
2452 m = vm_page_alloc_noobj_domain(domain, req);
2453 if (m != NULL)
2454 break;
2455 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2456
2457 return (m);
2458 }
2459
2460 vm_page_t
vm_page_alloc_noobj_domain(int domain,int req)2461 vm_page_alloc_noobj_domain(int domain, int req)
2462 {
2463 return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req));
2464 }
2465
2466 vm_page_t
vm_page_alloc_noobj_contig(int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)2467 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2468 vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2469 vm_memattr_t memattr)
2470 {
2471 struct vm_domainset_iter di;
2472 vm_page_t m;
2473 int domain;
2474
2475 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2476 do {
2477 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2478 high, alignment, boundary, memattr);
2479 if (m != NULL)
2480 break;
2481 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2482
2483 return (m);
2484 }
2485
2486 vm_page_t
vm_page_alloc_noobj_contig_domain(int domain,int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)2487 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2488 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2489 vm_memattr_t memattr)
2490 {
2491 vm_page_t m, m_ret;
2492 u_int flags;
2493
2494 #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2495 KASSERT((req & ~VPANC_FLAGS) == 0,
2496 ("invalid request %#x", req));
2497 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2498 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2499 ("invalid request %#x", req));
2500 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2501 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2502 ("invalid request %#x", req));
2503 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2504
2505 while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2506 low, high, alignment, boundary)) == NULL) {
2507 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2508 return (NULL);
2509 }
2510
2511 /*
2512 * Initialize the pages. Only the PG_ZERO flag is inherited.
2513 */
2514 flags = PG_ZERO;
2515 if ((req & VM_ALLOC_NODUMP) != 0)
2516 flags |= PG_NODUMP;
2517 if ((req & VM_ALLOC_WIRED) != 0)
2518 vm_wire_add(npages);
2519 for (m = m_ret; m < &m_ret[npages]; m++) {
2520 vm_page_dequeue(m);
2521 vm_page_alloc_check(m);
2522
2523 /*
2524 * Consumers should not rely on a useful default pindex value.
2525 */
2526 m->pindex = 0xdeadc0dedeadc0de;
2527 m->a.flags = 0;
2528 m->flags = (m->flags | PG_NODUMP) & flags;
2529 m->busy_lock = VPB_UNBUSIED;
2530 if ((req & VM_ALLOC_WIRED) != 0)
2531 m->ref_count = 1;
2532 m->a.act_count = 0;
2533 m->oflags = VPO_UNMANAGED;
2534
2535 /*
2536 * Zero the page before updating any mappings since the page is
2537 * not yet shared with any devices which might require the
2538 * non-default memory attribute. pmap_page_set_memattr()
2539 * flushes data caches before returning.
2540 */
2541 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2542 pmap_zero_page(m);
2543 if (memattr != VM_MEMATTR_DEFAULT)
2544 pmap_page_set_memattr(m, memattr);
2545 }
2546 return (m_ret);
2547 }
2548
2549 /*
2550 * Check a page that has been freshly dequeued from a freelist.
2551 */
2552 static void
vm_page_alloc_check(vm_page_t m)2553 vm_page_alloc_check(vm_page_t m)
2554 {
2555
2556 KASSERT(m->object == NULL, ("page %p has object", m));
2557 KASSERT(m->a.queue == PQ_NONE &&
2558 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2559 ("page %p has unexpected queue %d, flags %#x",
2560 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2561 KASSERT(m->ref_count == 0, ("page %p has references", m));
2562 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2563 KASSERT(m->dirty == 0, ("page %p is dirty", m));
2564 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2565 ("page %p has unexpected memattr %d",
2566 m, pmap_page_get_memattr(m)));
2567 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2568 pmap_vm_page_alloc_check(m);
2569 }
2570
2571 static int
vm_page_zone_import(void * arg,void ** store,int cnt,int domain,int flags)2572 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2573 {
2574 struct vm_domain *vmd;
2575 struct vm_pgcache *pgcache;
2576 int i;
2577
2578 pgcache = arg;
2579 vmd = VM_DOMAIN(pgcache->domain);
2580
2581 /*
2582 * The page daemon should avoid creating extra memory pressure since its
2583 * main purpose is to replenish the store of free pages.
2584 */
2585 if (vmd->vmd_severeset || curproc == pageproc ||
2586 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2587 return (0);
2588 domain = vmd->vmd_domain;
2589 vm_domain_free_lock(vmd);
2590 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2591 (vm_page_t *)store);
2592 vm_domain_free_unlock(vmd);
2593 if (cnt != i)
2594 vm_domain_freecnt_inc(vmd, cnt - i);
2595
2596 return (i);
2597 }
2598
2599 static void
vm_page_zone_release(void * arg,void ** store,int cnt)2600 vm_page_zone_release(void *arg, void **store, int cnt)
2601 {
2602 struct vm_domain *vmd;
2603 struct vm_pgcache *pgcache;
2604 vm_page_t m;
2605 int i;
2606
2607 pgcache = arg;
2608 vmd = VM_DOMAIN(pgcache->domain);
2609 vm_domain_free_lock(vmd);
2610 for (i = 0; i < cnt; i++) {
2611 m = (vm_page_t)store[i];
2612 vm_phys_free_pages(m, 0);
2613 }
2614 vm_domain_free_unlock(vmd);
2615 vm_domain_freecnt_inc(vmd, cnt);
2616 }
2617
2618 #define VPSC_ANY 0 /* No restrictions. */
2619 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2620 #define VPSC_NOSUPER 2 /* Skip superpages. */
2621
2622 /*
2623 * vm_page_scan_contig:
2624 *
2625 * Scan vm_page_array[] between the specified entries "m_start" and
2626 * "m_end" for a run of contiguous physical pages that satisfy the
2627 * specified conditions, and return the lowest page in the run. The
2628 * specified "alignment" determines the alignment of the lowest physical
2629 * page in the run. If the specified "boundary" is non-zero, then the
2630 * run of physical pages cannot span a physical address that is a
2631 * multiple of "boundary".
2632 *
2633 * "m_end" is never dereferenced, so it need not point to a vm_page
2634 * structure within vm_page_array[].
2635 *
2636 * "npages" must be greater than zero. "m_start" and "m_end" must not
2637 * span a hole (or discontiguity) in the physical address space. Both
2638 * "alignment" and "boundary" must be a power of two.
2639 */
2640 static vm_page_t
vm_page_scan_contig(u_long npages,vm_page_t m_start,vm_page_t m_end,u_long alignment,vm_paddr_t boundary,int options)2641 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2642 u_long alignment, vm_paddr_t boundary, int options)
2643 {
2644 vm_object_t object;
2645 vm_paddr_t pa;
2646 vm_page_t m, m_run;
2647 #if VM_NRESERVLEVEL > 0
2648 int level;
2649 #endif
2650 int m_inc, order, run_ext, run_len;
2651
2652 KASSERT(npages > 0, ("npages is 0"));
2653 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2654 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2655 m_run = NULL;
2656 run_len = 0;
2657 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2658 KASSERT((m->flags & PG_MARKER) == 0,
2659 ("page %p is PG_MARKER", m));
2660 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2661 ("fictitious page %p has invalid ref count", m));
2662
2663 /*
2664 * If the current page would be the start of a run, check its
2665 * physical address against the end, alignment, and boundary
2666 * conditions. If it doesn't satisfy these conditions, either
2667 * terminate the scan or advance to the next page that
2668 * satisfies the failed condition.
2669 */
2670 if (run_len == 0) {
2671 KASSERT(m_run == NULL, ("m_run != NULL"));
2672 if (m + npages > m_end)
2673 break;
2674 pa = VM_PAGE_TO_PHYS(m);
2675 if (!vm_addr_align_ok(pa, alignment)) {
2676 m_inc = atop(roundup2(pa, alignment) - pa);
2677 continue;
2678 }
2679 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2680 m_inc = atop(roundup2(pa, boundary) - pa);
2681 continue;
2682 }
2683 } else
2684 KASSERT(m_run != NULL, ("m_run == NULL"));
2685
2686 retry:
2687 m_inc = 1;
2688 if (vm_page_wired(m))
2689 run_ext = 0;
2690 #if VM_NRESERVLEVEL > 0
2691 else if ((level = vm_reserv_level(m)) >= 0 &&
2692 (options & VPSC_NORESERV) != 0) {
2693 run_ext = 0;
2694 /* Advance to the end of the reservation. */
2695 pa = VM_PAGE_TO_PHYS(m);
2696 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2697 pa);
2698 }
2699 #endif
2700 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2701 /*
2702 * The page is considered eligible for relocation if
2703 * and only if it could be laundered or reclaimed by
2704 * the page daemon.
2705 */
2706 VM_OBJECT_RLOCK(object);
2707 if (object != m->object) {
2708 VM_OBJECT_RUNLOCK(object);
2709 goto retry;
2710 }
2711 /* Don't care: PG_NODUMP, PG_ZERO. */
2712 if ((object->flags & OBJ_SWAP) == 0 &&
2713 object->type != OBJT_VNODE) {
2714 run_ext = 0;
2715 #if VM_NRESERVLEVEL > 0
2716 } else if ((options & VPSC_NOSUPER) != 0 &&
2717 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2718 run_ext = 0;
2719 /* Advance to the end of the superpage. */
2720 pa = VM_PAGE_TO_PHYS(m);
2721 m_inc = atop(roundup2(pa + 1,
2722 vm_reserv_size(level)) - pa);
2723 #endif
2724 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2725 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2726 /*
2727 * The page is allocated but eligible for
2728 * relocation. Extend the current run by one
2729 * page.
2730 */
2731 KASSERT(pmap_page_get_memattr(m) ==
2732 VM_MEMATTR_DEFAULT,
2733 ("page %p has an unexpected memattr", m));
2734 KASSERT((m->oflags & (VPO_SWAPINPROG |
2735 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2736 ("page %p has unexpected oflags", m));
2737 /* Don't care: PGA_NOSYNC. */
2738 run_ext = 1;
2739 } else
2740 run_ext = 0;
2741 VM_OBJECT_RUNLOCK(object);
2742 #if VM_NRESERVLEVEL > 0
2743 } else if (level >= 0) {
2744 /*
2745 * The page is reserved but not yet allocated. In
2746 * other words, it is still free. Extend the current
2747 * run by one page.
2748 */
2749 run_ext = 1;
2750 #endif
2751 } else if ((order = m->order) < VM_NFREEORDER) {
2752 /*
2753 * The page is enqueued in the physical memory
2754 * allocator's free page queues. Moreover, it is the
2755 * first page in a power-of-two-sized run of
2756 * contiguous free pages. Add these pages to the end
2757 * of the current run, and jump ahead.
2758 */
2759 run_ext = 1 << order;
2760 m_inc = 1 << order;
2761 } else {
2762 /*
2763 * Skip the page for one of the following reasons: (1)
2764 * It is enqueued in the physical memory allocator's
2765 * free page queues. However, it is not the first
2766 * page in a run of contiguous free pages. (This case
2767 * rarely occurs because the scan is performed in
2768 * ascending order.) (2) It is not reserved, and it is
2769 * transitioning from free to allocated. (Conversely,
2770 * the transition from allocated to free for managed
2771 * pages is blocked by the page busy lock.) (3) It is
2772 * allocated but not contained by an object and not
2773 * wired, e.g., allocated by Xen's balloon driver.
2774 */
2775 run_ext = 0;
2776 }
2777
2778 /*
2779 * Extend or reset the current run of pages.
2780 */
2781 if (run_ext > 0) {
2782 if (run_len == 0)
2783 m_run = m;
2784 run_len += run_ext;
2785 } else {
2786 if (run_len > 0) {
2787 m_run = NULL;
2788 run_len = 0;
2789 }
2790 }
2791 }
2792 if (run_len >= npages)
2793 return (m_run);
2794 return (NULL);
2795 }
2796
2797 /*
2798 * vm_page_reclaim_run:
2799 *
2800 * Try to relocate each of the allocated virtual pages within the
2801 * specified run of physical pages to a new physical address. Free the
2802 * physical pages underlying the relocated virtual pages. A virtual page
2803 * is relocatable if and only if it could be laundered or reclaimed by
2804 * the page daemon. Whenever possible, a virtual page is relocated to a
2805 * physical address above "high".
2806 *
2807 * Returns 0 if every physical page within the run was already free or
2808 * just freed by a successful relocation. Otherwise, returns a non-zero
2809 * value indicating why the last attempt to relocate a virtual page was
2810 * unsuccessful.
2811 *
2812 * "req_class" must be an allocation class.
2813 */
2814 static int
vm_page_reclaim_run(int req_class,int domain,u_long npages,vm_page_t m_run,vm_paddr_t high)2815 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2816 vm_paddr_t high)
2817 {
2818 struct vm_domain *vmd;
2819 struct spglist free;
2820 vm_object_t object;
2821 vm_paddr_t pa;
2822 vm_page_t m, m_end, m_new;
2823 int error, order, req;
2824
2825 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2826 ("req_class is not an allocation class"));
2827 SLIST_INIT(&free);
2828 error = 0;
2829 m = m_run;
2830 m_end = m_run + npages;
2831 for (; error == 0 && m < m_end; m++) {
2832 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2833 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2834
2835 /*
2836 * Racily check for wirings. Races are handled once the object
2837 * lock is held and the page is unmapped.
2838 */
2839 if (vm_page_wired(m))
2840 error = EBUSY;
2841 else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2842 /*
2843 * The page is relocated if and only if it could be
2844 * laundered or reclaimed by the page daemon.
2845 */
2846 VM_OBJECT_WLOCK(object);
2847 /* Don't care: PG_NODUMP, PG_ZERO. */
2848 if (m->object != object ||
2849 ((object->flags & OBJ_SWAP) == 0 &&
2850 object->type != OBJT_VNODE))
2851 error = EINVAL;
2852 else if (object->memattr != VM_MEMATTR_DEFAULT)
2853 error = EINVAL;
2854 else if (vm_page_queue(m) != PQ_NONE &&
2855 vm_page_tryxbusy(m) != 0) {
2856 if (vm_page_wired(m)) {
2857 vm_page_xunbusy(m);
2858 error = EBUSY;
2859 goto unlock;
2860 }
2861 KASSERT(pmap_page_get_memattr(m) ==
2862 VM_MEMATTR_DEFAULT,
2863 ("page %p has an unexpected memattr", m));
2864 KASSERT(m->oflags == 0,
2865 ("page %p has unexpected oflags", m));
2866 /* Don't care: PGA_NOSYNC. */
2867 if (!vm_page_none_valid(m)) {
2868 /*
2869 * First, try to allocate a new page
2870 * that is above "high". Failing
2871 * that, try to allocate a new page
2872 * that is below "m_run". Allocate
2873 * the new page between the end of
2874 * "m_run" and "high" only as a last
2875 * resort.
2876 */
2877 req = req_class;
2878 if ((m->flags & PG_NODUMP) != 0)
2879 req |= VM_ALLOC_NODUMP;
2880 if (trunc_page(high) !=
2881 ~(vm_paddr_t)PAGE_MASK) {
2882 m_new =
2883 vm_page_alloc_noobj_contig(
2884 req, 1, round_page(high),
2885 ~(vm_paddr_t)0, PAGE_SIZE,
2886 0, VM_MEMATTR_DEFAULT);
2887 } else
2888 m_new = NULL;
2889 if (m_new == NULL) {
2890 pa = VM_PAGE_TO_PHYS(m_run);
2891 m_new =
2892 vm_page_alloc_noobj_contig(
2893 req, 1, 0, pa - 1,
2894 PAGE_SIZE, 0,
2895 VM_MEMATTR_DEFAULT);
2896 }
2897 if (m_new == NULL) {
2898 pa += ptoa(npages);
2899 m_new =
2900 vm_page_alloc_noobj_contig(
2901 req, 1, pa, high, PAGE_SIZE,
2902 0, VM_MEMATTR_DEFAULT);
2903 }
2904 if (m_new == NULL) {
2905 vm_page_xunbusy(m);
2906 error = ENOMEM;
2907 goto unlock;
2908 }
2909
2910 /*
2911 * Unmap the page and check for new
2912 * wirings that may have been acquired
2913 * through a pmap lookup.
2914 */
2915 if (object->ref_count != 0 &&
2916 !vm_page_try_remove_all(m)) {
2917 vm_page_xunbusy(m);
2918 vm_page_free(m_new);
2919 error = EBUSY;
2920 goto unlock;
2921 }
2922
2923 /*
2924 * Replace "m" with the new page. For
2925 * vm_page_replace(), "m" must be busy
2926 * and dequeued. Finally, change "m"
2927 * as if vm_page_free() was called.
2928 */
2929 m_new->a.flags = m->a.flags &
2930 ~PGA_QUEUE_STATE_MASK;
2931 KASSERT(m_new->oflags == VPO_UNMANAGED,
2932 ("page %p is managed", m_new));
2933 m_new->oflags = 0;
2934 pmap_copy_page(m, m_new);
2935 m_new->valid = m->valid;
2936 m_new->dirty = m->dirty;
2937 m->flags &= ~PG_ZERO;
2938 vm_page_dequeue(m);
2939 if (vm_page_replace_hold(m_new, object,
2940 m->pindex, m) &&
2941 vm_page_free_prep(m))
2942 SLIST_INSERT_HEAD(&free, m,
2943 plinks.s.ss);
2944
2945 /*
2946 * The new page must be deactivated
2947 * before the object is unlocked.
2948 */
2949 vm_page_deactivate(m_new);
2950 } else {
2951 m->flags &= ~PG_ZERO;
2952 vm_page_dequeue(m);
2953 if (vm_page_free_prep(m))
2954 SLIST_INSERT_HEAD(&free, m,
2955 plinks.s.ss);
2956 KASSERT(m->dirty == 0,
2957 ("page %p is dirty", m));
2958 }
2959 } else
2960 error = EBUSY;
2961 unlock:
2962 VM_OBJECT_WUNLOCK(object);
2963 } else {
2964 MPASS(vm_page_domain(m) == domain);
2965 vmd = VM_DOMAIN(domain);
2966 vm_domain_free_lock(vmd);
2967 order = m->order;
2968 if (order < VM_NFREEORDER) {
2969 /*
2970 * The page is enqueued in the physical memory
2971 * allocator's free page queues. Moreover, it
2972 * is the first page in a power-of-two-sized
2973 * run of contiguous free pages. Jump ahead
2974 * to the last page within that run, and
2975 * continue from there.
2976 */
2977 m += (1 << order) - 1;
2978 }
2979 #if VM_NRESERVLEVEL > 0
2980 else if (vm_reserv_is_page_free(m))
2981 order = 0;
2982 #endif
2983 vm_domain_free_unlock(vmd);
2984 if (order == VM_NFREEORDER)
2985 error = EINVAL;
2986 }
2987 }
2988 if ((m = SLIST_FIRST(&free)) != NULL) {
2989 int cnt;
2990
2991 vmd = VM_DOMAIN(domain);
2992 cnt = 0;
2993 vm_domain_free_lock(vmd);
2994 do {
2995 MPASS(vm_page_domain(m) == domain);
2996 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2997 vm_phys_free_pages(m, 0);
2998 cnt++;
2999 } while ((m = SLIST_FIRST(&free)) != NULL);
3000 vm_domain_free_unlock(vmd);
3001 vm_domain_freecnt_inc(vmd, cnt);
3002 }
3003 return (error);
3004 }
3005
3006 #define NRUNS 16
3007
3008 #define RUN_INDEX(count, nruns) ((count) % (nruns))
3009
3010 #define MIN_RECLAIM 8
3011
3012 /*
3013 * vm_page_reclaim_contig:
3014 *
3015 * Reclaim allocated, contiguous physical memory satisfying the specified
3016 * conditions by relocating the virtual pages using that physical memory.
3017 * Returns 0 if reclamation is successful, ERANGE if the specified domain
3018 * can't possibly satisfy the reclamation request, or ENOMEM if not
3019 * currently able to reclaim the requested number of pages. Since
3020 * relocation requires the allocation of physical pages, reclamation may
3021 * fail with ENOMEM due to a shortage of free pages. When reclamation
3022 * fails in this manner, callers are expected to perform vm_wait() before
3023 * retrying a failed allocation operation, e.g., vm_page_alloc_contig().
3024 *
3025 * The caller must always specify an allocation class through "req".
3026 *
3027 * allocation classes:
3028 * VM_ALLOC_NORMAL normal process request
3029 * VM_ALLOC_SYSTEM system *really* needs a page
3030 * VM_ALLOC_INTERRUPT interrupt time request
3031 *
3032 * The optional allocation flags are ignored.
3033 *
3034 * "npages" must be greater than zero. Both "alignment" and "boundary"
3035 * must be a power of two.
3036 */
3037 int
vm_page_reclaim_contig_domain_ext(int domain,int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,int desired_runs)3038 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3039 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3040 int desired_runs)
3041 {
3042 struct vm_domain *vmd;
3043 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3044 u_long count, minalign, reclaimed;
3045 int error, i, min_reclaim, nruns, options, req_class;
3046 int segind, start_segind;
3047 int ret;
3048
3049 KASSERT(npages > 0, ("npages is 0"));
3050 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3051 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3052
3053 ret = ENOMEM;
3054
3055 /*
3056 * If the caller wants to reclaim multiple runs, try to allocate
3057 * space to store the runs. If that fails, fall back to the old
3058 * behavior of just reclaiming MIN_RECLAIM pages.
3059 */
3060 if (desired_runs > 1)
3061 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3062 M_TEMP, M_NOWAIT);
3063 else
3064 m_runs = NULL;
3065
3066 if (m_runs == NULL) {
3067 m_runs = _m_runs;
3068 nruns = NRUNS;
3069 } else {
3070 nruns = NRUNS + desired_runs - 1;
3071 }
3072 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3073
3074 /*
3075 * The caller will attempt an allocation after some runs have been
3076 * reclaimed and added to the vm_phys buddy lists. Due to limitations
3077 * of vm_phys_alloc_contig(), round up the requested length to the next
3078 * power of two or maximum chunk size, and ensure that each run is
3079 * suitably aligned.
3080 */
3081 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3082 npages = roundup2(npages, minalign);
3083 if (alignment < ptoa(minalign))
3084 alignment = ptoa(minalign);
3085
3086 /*
3087 * The page daemon is allowed to dig deeper into the free page list.
3088 */
3089 req_class = req & VM_ALLOC_CLASS_MASK;
3090 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3091 req_class = VM_ALLOC_SYSTEM;
3092
3093 start_segind = vm_phys_lookup_segind(low);
3094
3095 /*
3096 * Return if the number of free pages cannot satisfy the requested
3097 * allocation.
3098 */
3099 vmd = VM_DOMAIN(domain);
3100 count = vmd->vmd_free_count;
3101 if (count < npages + vmd->vmd_free_reserved || (count < npages +
3102 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3103 (count < npages && req_class == VM_ALLOC_INTERRUPT))
3104 goto done;
3105
3106 /*
3107 * Scan up to three times, relaxing the restrictions ("options") on
3108 * the reclamation of reservations and superpages each time.
3109 */
3110 for (options = VPSC_NORESERV;;) {
3111 bool phys_range_exists = false;
3112
3113 /*
3114 * Find the highest runs that satisfy the given constraints
3115 * and restrictions, and record them in "m_runs".
3116 */
3117 count = 0;
3118 segind = start_segind;
3119 while ((segind = vm_phys_find_range(bounds, segind, domain,
3120 npages, low, high)) != -1) {
3121 phys_range_exists = true;
3122 while ((m_run = vm_page_scan_contig(npages, bounds[0],
3123 bounds[1], alignment, boundary, options))) {
3124 bounds[0] = m_run + npages;
3125 m_runs[RUN_INDEX(count, nruns)] = m_run;
3126 count++;
3127 }
3128 segind++;
3129 }
3130
3131 if (!phys_range_exists) {
3132 ret = ERANGE;
3133 goto done;
3134 }
3135
3136 /*
3137 * Reclaim the highest runs in LIFO (descending) order until
3138 * the number of reclaimed pages, "reclaimed", is at least
3139 * "min_reclaim". Reset "reclaimed" each time because each
3140 * reclamation is idempotent, and runs will (likely) recur
3141 * from one scan to the next as restrictions are relaxed.
3142 */
3143 reclaimed = 0;
3144 for (i = 0; count > 0 && i < nruns; i++) {
3145 count--;
3146 m_run = m_runs[RUN_INDEX(count, nruns)];
3147 error = vm_page_reclaim_run(req_class, domain, npages,
3148 m_run, high);
3149 if (error == 0) {
3150 reclaimed += npages;
3151 if (reclaimed >= min_reclaim) {
3152 ret = 0;
3153 goto done;
3154 }
3155 }
3156 }
3157
3158 /*
3159 * Either relax the restrictions on the next scan or return if
3160 * the last scan had no restrictions.
3161 */
3162 if (options == VPSC_NORESERV)
3163 options = VPSC_NOSUPER;
3164 else if (options == VPSC_NOSUPER)
3165 options = VPSC_ANY;
3166 else if (options == VPSC_ANY) {
3167 if (reclaimed != 0)
3168 ret = 0;
3169 goto done;
3170 }
3171 }
3172 done:
3173 if (m_runs != _m_runs)
3174 free(m_runs, M_TEMP);
3175 return (ret);
3176 }
3177
3178 int
vm_page_reclaim_contig_domain(int domain,int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary)3179 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3180 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3181 {
3182 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3183 alignment, boundary, 1));
3184 }
3185
3186 int
vm_page_reclaim_contig(int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary)3187 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3188 u_long alignment, vm_paddr_t boundary)
3189 {
3190 struct vm_domainset_iter di;
3191 int domain, ret, status;
3192
3193 ret = ERANGE;
3194
3195 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3196 do {
3197 status = vm_page_reclaim_contig_domain(domain, req, npages, low,
3198 high, alignment, boundary);
3199 if (status == 0)
3200 return (0);
3201 else if (status == ERANGE)
3202 vm_domainset_iter_ignore(&di, domain);
3203 else {
3204 KASSERT(status == ENOMEM, ("Unrecognized error %d "
3205 "from vm_page_reclaim_contig_domain()", status));
3206 ret = ENOMEM;
3207 }
3208 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3209
3210 return (ret);
3211 }
3212
3213 /*
3214 * Set the domain in the appropriate page level domainset.
3215 */
3216 void
vm_domain_set(struct vm_domain * vmd)3217 vm_domain_set(struct vm_domain *vmd)
3218 {
3219
3220 mtx_lock(&vm_domainset_lock);
3221 if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3222 vmd->vmd_minset = 1;
3223 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3224 }
3225 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3226 vmd->vmd_severeset = 1;
3227 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3228 }
3229 mtx_unlock(&vm_domainset_lock);
3230 }
3231
3232 /*
3233 * Clear the domain from the appropriate page level domainset.
3234 */
3235 void
vm_domain_clear(struct vm_domain * vmd)3236 vm_domain_clear(struct vm_domain *vmd)
3237 {
3238
3239 mtx_lock(&vm_domainset_lock);
3240 if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3241 vmd->vmd_minset = 0;
3242 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3243 if (vm_min_waiters != 0) {
3244 vm_min_waiters = 0;
3245 wakeup(&vm_min_domains);
3246 }
3247 }
3248 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3249 vmd->vmd_severeset = 0;
3250 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3251 if (vm_severe_waiters != 0) {
3252 vm_severe_waiters = 0;
3253 wakeup(&vm_severe_domains);
3254 }
3255 }
3256
3257 /*
3258 * If pageout daemon needs pages, then tell it that there are
3259 * some free.
3260 */
3261 if (vmd->vmd_pageout_pages_needed &&
3262 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3263 wakeup(&vmd->vmd_pageout_pages_needed);
3264 vmd->vmd_pageout_pages_needed = 0;
3265 }
3266
3267 /* See comments in vm_wait_doms(). */
3268 if (vm_pageproc_waiters) {
3269 vm_pageproc_waiters = 0;
3270 wakeup(&vm_pageproc_waiters);
3271 }
3272 mtx_unlock(&vm_domainset_lock);
3273 }
3274
3275 /*
3276 * Wait for free pages to exceed the min threshold globally.
3277 */
3278 void
vm_wait_min(void)3279 vm_wait_min(void)
3280 {
3281
3282 mtx_lock(&vm_domainset_lock);
3283 while (vm_page_count_min()) {
3284 vm_min_waiters++;
3285 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3286 }
3287 mtx_unlock(&vm_domainset_lock);
3288 }
3289
3290 /*
3291 * Wait for free pages to exceed the severe threshold globally.
3292 */
3293 void
vm_wait_severe(void)3294 vm_wait_severe(void)
3295 {
3296
3297 mtx_lock(&vm_domainset_lock);
3298 while (vm_page_count_severe()) {
3299 vm_severe_waiters++;
3300 msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3301 "vmwait", 0);
3302 }
3303 mtx_unlock(&vm_domainset_lock);
3304 }
3305
3306 u_int
vm_wait_count(void)3307 vm_wait_count(void)
3308 {
3309
3310 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3311 }
3312
3313 int
vm_wait_doms(const domainset_t * wdoms,int mflags)3314 vm_wait_doms(const domainset_t *wdoms, int mflags)
3315 {
3316 int error;
3317
3318 error = 0;
3319
3320 /*
3321 * We use racey wakeup synchronization to avoid expensive global
3322 * locking for the pageproc when sleeping with a non-specific vm_wait.
3323 * To handle this, we only sleep for one tick in this instance. It
3324 * is expected that most allocations for the pageproc will come from
3325 * kmem or vm_page_grab* which will use the more specific and
3326 * race-free vm_wait_domain().
3327 */
3328 if (curproc == pageproc) {
3329 mtx_lock(&vm_domainset_lock);
3330 vm_pageproc_waiters++;
3331 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3332 PVM | PDROP | mflags, "pageprocwait", 1);
3333 } else {
3334 /*
3335 * XXX Ideally we would wait only until the allocation could
3336 * be satisfied. This condition can cause new allocators to
3337 * consume all freed pages while old allocators wait.
3338 */
3339 mtx_lock(&vm_domainset_lock);
3340 if (vm_page_count_min_set(wdoms)) {
3341 if (pageproc == NULL)
3342 panic("vm_wait in early boot");
3343 vm_min_waiters++;
3344 error = msleep(&vm_min_domains, &vm_domainset_lock,
3345 PVM | PDROP | mflags, "vmwait", 0);
3346 } else
3347 mtx_unlock(&vm_domainset_lock);
3348 }
3349 return (error);
3350 }
3351
3352 /*
3353 * vm_wait_domain:
3354 *
3355 * Sleep until free pages are available for allocation.
3356 * - Called in various places after failed memory allocations.
3357 */
3358 void
vm_wait_domain(int domain)3359 vm_wait_domain(int domain)
3360 {
3361 struct vm_domain *vmd;
3362 domainset_t wdom;
3363
3364 vmd = VM_DOMAIN(domain);
3365 vm_domain_free_assert_unlocked(vmd);
3366
3367 if (curproc == pageproc) {
3368 mtx_lock(&vm_domainset_lock);
3369 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3370 vmd->vmd_pageout_pages_needed = 1;
3371 msleep(&vmd->vmd_pageout_pages_needed,
3372 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3373 } else
3374 mtx_unlock(&vm_domainset_lock);
3375 } else {
3376 DOMAINSET_ZERO(&wdom);
3377 DOMAINSET_SET(vmd->vmd_domain, &wdom);
3378 vm_wait_doms(&wdom, 0);
3379 }
3380 }
3381
3382 static int
vm_wait_flags(vm_object_t obj,int mflags)3383 vm_wait_flags(vm_object_t obj, int mflags)
3384 {
3385 struct domainset *d;
3386
3387 d = NULL;
3388
3389 /*
3390 * Carefully fetch pointers only once: the struct domainset
3391 * itself is ummutable but the pointer might change.
3392 */
3393 if (obj != NULL)
3394 d = obj->domain.dr_policy;
3395 if (d == NULL)
3396 d = curthread->td_domain.dr_policy;
3397
3398 return (vm_wait_doms(&d->ds_mask, mflags));
3399 }
3400
3401 /*
3402 * vm_wait:
3403 *
3404 * Sleep until free pages are available for allocation in the
3405 * affinity domains of the obj. If obj is NULL, the domain set
3406 * for the calling thread is used.
3407 * Called in various places after failed memory allocations.
3408 */
3409 void
vm_wait(vm_object_t obj)3410 vm_wait(vm_object_t obj)
3411 {
3412 (void)vm_wait_flags(obj, 0);
3413 }
3414
3415 int
vm_wait_intr(vm_object_t obj)3416 vm_wait_intr(vm_object_t obj)
3417 {
3418 return (vm_wait_flags(obj, PCATCH));
3419 }
3420
3421 /*
3422 * vm_domain_alloc_fail:
3423 *
3424 * Called when a page allocation function fails. Informs the
3425 * pagedaemon and performs the requested wait. Requires the
3426 * domain_free and object lock on entry. Returns with the
3427 * object lock held and free lock released. Returns an error when
3428 * retry is necessary.
3429 *
3430 */
3431 static int
vm_domain_alloc_fail(struct vm_domain * vmd,vm_object_t object,int req)3432 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3433 {
3434
3435 vm_domain_free_assert_unlocked(vmd);
3436
3437 atomic_add_int(&vmd->vmd_pageout_deficit,
3438 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3439 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3440 if (object != NULL)
3441 VM_OBJECT_WUNLOCK(object);
3442 vm_wait_domain(vmd->vmd_domain);
3443 if (object != NULL)
3444 VM_OBJECT_WLOCK(object);
3445 if (req & VM_ALLOC_WAITOK)
3446 return (EAGAIN);
3447 }
3448
3449 return (0);
3450 }
3451
3452 /*
3453 * vm_waitpfault:
3454 *
3455 * Sleep until free pages are available for allocation.
3456 * - Called only in vm_fault so that processes page faulting
3457 * can be easily tracked.
3458 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3459 * processes will be able to grab memory first. Do not change
3460 * this balance without careful testing first.
3461 */
3462 void
vm_waitpfault(struct domainset * dset,int timo)3463 vm_waitpfault(struct domainset *dset, int timo)
3464 {
3465
3466 /*
3467 * XXX Ideally we would wait only until the allocation could
3468 * be satisfied. This condition can cause new allocators to
3469 * consume all freed pages while old allocators wait.
3470 */
3471 mtx_lock(&vm_domainset_lock);
3472 if (vm_page_count_min_set(&dset->ds_mask)) {
3473 vm_min_waiters++;
3474 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3475 "pfault", timo);
3476 } else
3477 mtx_unlock(&vm_domainset_lock);
3478 }
3479
3480 static struct vm_pagequeue *
_vm_page_pagequeue(vm_page_t m,uint8_t queue)3481 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3482 {
3483
3484 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3485 }
3486
3487 #ifdef INVARIANTS
3488 static struct vm_pagequeue *
vm_page_pagequeue(vm_page_t m)3489 vm_page_pagequeue(vm_page_t m)
3490 {
3491
3492 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3493 }
3494 #endif
3495
3496 static __always_inline bool
vm_page_pqstate_fcmpset(vm_page_t m,vm_page_astate_t * old,vm_page_astate_t new)3497 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3498 {
3499 vm_page_astate_t tmp;
3500
3501 tmp = *old;
3502 do {
3503 if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3504 return (true);
3505 counter_u64_add(pqstate_commit_retries, 1);
3506 } while (old->_bits == tmp._bits);
3507
3508 return (false);
3509 }
3510
3511 /*
3512 * Do the work of committing a queue state update that moves the page out of
3513 * its current queue.
3514 */
3515 static bool
_vm_page_pqstate_commit_dequeue(struct vm_pagequeue * pq,vm_page_t m,vm_page_astate_t * old,vm_page_astate_t new)3516 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3517 vm_page_astate_t *old, vm_page_astate_t new)
3518 {
3519 vm_page_t next;
3520
3521 vm_pagequeue_assert_locked(pq);
3522 KASSERT(vm_page_pagequeue(m) == pq,
3523 ("%s: queue %p does not match page %p", __func__, pq, m));
3524 KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3525 ("%s: invalid queue indices %d %d",
3526 __func__, old->queue, new.queue));
3527
3528 /*
3529 * Once the queue index of the page changes there is nothing
3530 * synchronizing with further updates to the page's physical
3531 * queue state. Therefore we must speculatively remove the page
3532 * from the queue now and be prepared to roll back if the queue
3533 * state update fails. If the page is not physically enqueued then
3534 * we just update its queue index.
3535 */
3536 if ((old->flags & PGA_ENQUEUED) != 0) {
3537 new.flags &= ~PGA_ENQUEUED;
3538 next = TAILQ_NEXT(m, plinks.q);
3539 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3540 vm_pagequeue_cnt_dec(pq);
3541 if (!vm_page_pqstate_fcmpset(m, old, new)) {
3542 if (next == NULL)
3543 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3544 else
3545 TAILQ_INSERT_BEFORE(next, m, plinks.q);
3546 vm_pagequeue_cnt_inc(pq);
3547 return (false);
3548 } else {
3549 return (true);
3550 }
3551 } else {
3552 return (vm_page_pqstate_fcmpset(m, old, new));
3553 }
3554 }
3555
3556 static bool
vm_page_pqstate_commit_dequeue(vm_page_t m,vm_page_astate_t * old,vm_page_astate_t new)3557 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3558 vm_page_astate_t new)
3559 {
3560 struct vm_pagequeue *pq;
3561 vm_page_astate_t as;
3562 bool ret;
3563
3564 pq = _vm_page_pagequeue(m, old->queue);
3565
3566 /*
3567 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3568 * corresponding page queue lock is held.
3569 */
3570 vm_pagequeue_lock(pq);
3571 as = vm_page_astate_load(m);
3572 if (__predict_false(as._bits != old->_bits)) {
3573 *old = as;
3574 ret = false;
3575 } else {
3576 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3577 }
3578 vm_pagequeue_unlock(pq);
3579 return (ret);
3580 }
3581
3582 /*
3583 * Commit a queue state update that enqueues or requeues a page.
3584 */
3585 static bool
_vm_page_pqstate_commit_requeue(struct vm_pagequeue * pq,vm_page_t m,vm_page_astate_t * old,vm_page_astate_t new)3586 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3587 vm_page_astate_t *old, vm_page_astate_t new)
3588 {
3589 struct vm_domain *vmd;
3590
3591 vm_pagequeue_assert_locked(pq);
3592 KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3593 ("%s: invalid queue indices %d %d",
3594 __func__, old->queue, new.queue));
3595
3596 new.flags |= PGA_ENQUEUED;
3597 if (!vm_page_pqstate_fcmpset(m, old, new))
3598 return (false);
3599
3600 if ((old->flags & PGA_ENQUEUED) != 0)
3601 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3602 else
3603 vm_pagequeue_cnt_inc(pq);
3604
3605 /*
3606 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if
3607 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3608 * applied, even if it was set first.
3609 */
3610 if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3611 vmd = vm_pagequeue_domain(m);
3612 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3613 ("%s: invalid page queue for page %p", __func__, m));
3614 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3615 } else {
3616 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3617 }
3618 return (true);
3619 }
3620
3621 /*
3622 * Commit a queue state update that encodes a request for a deferred queue
3623 * operation.
3624 */
3625 static bool
vm_page_pqstate_commit_request(vm_page_t m,vm_page_astate_t * old,vm_page_astate_t new)3626 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3627 vm_page_astate_t new)
3628 {
3629
3630 KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3631 ("%s: invalid state, queue %d flags %x",
3632 __func__, new.queue, new.flags));
3633
3634 if (old->_bits != new._bits &&
3635 !vm_page_pqstate_fcmpset(m, old, new))
3636 return (false);
3637 vm_page_pqbatch_submit(m, new.queue);
3638 return (true);
3639 }
3640
3641 /*
3642 * A generic queue state update function. This handles more cases than the
3643 * specialized functions above.
3644 */
3645 bool
vm_page_pqstate_commit(vm_page_t m,vm_page_astate_t * old,vm_page_astate_t new)3646 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3647 {
3648
3649 if (old->_bits == new._bits)
3650 return (true);
3651
3652 if (old->queue != PQ_NONE && new.queue != old->queue) {
3653 if (!vm_page_pqstate_commit_dequeue(m, old, new))
3654 return (false);
3655 if (new.queue != PQ_NONE)
3656 vm_page_pqbatch_submit(m, new.queue);
3657 } else {
3658 if (!vm_page_pqstate_fcmpset(m, old, new))
3659 return (false);
3660 if (new.queue != PQ_NONE &&
3661 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3662 vm_page_pqbatch_submit(m, new.queue);
3663 }
3664 return (true);
3665 }
3666
3667 /*
3668 * Apply deferred queue state updates to a page.
3669 */
3670 static inline void
vm_pqbatch_process_page(struct vm_pagequeue * pq,vm_page_t m,uint8_t queue)3671 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3672 {
3673 vm_page_astate_t new, old;
3674
3675 CRITICAL_ASSERT(curthread);
3676 vm_pagequeue_assert_locked(pq);
3677 KASSERT(queue < PQ_COUNT,
3678 ("%s: invalid queue index %d", __func__, queue));
3679 KASSERT(pq == _vm_page_pagequeue(m, queue),
3680 ("%s: page %p does not belong to queue %p", __func__, m, pq));
3681
3682 for (old = vm_page_astate_load(m);;) {
3683 if (__predict_false(old.queue != queue ||
3684 (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3685 counter_u64_add(queue_nops, 1);
3686 break;
3687 }
3688 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3689 ("%s: page %p is unmanaged", __func__, m));
3690
3691 new = old;
3692 if ((old.flags & PGA_DEQUEUE) != 0) {
3693 new.flags &= ~PGA_QUEUE_OP_MASK;
3694 new.queue = PQ_NONE;
3695 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3696 m, &old, new))) {
3697 counter_u64_add(queue_ops, 1);
3698 break;
3699 }
3700 } else {
3701 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3702 if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3703 m, &old, new))) {
3704 counter_u64_add(queue_ops, 1);
3705 break;
3706 }
3707 }
3708 }
3709 }
3710
3711 static void
vm_pqbatch_process(struct vm_pagequeue * pq,struct vm_batchqueue * bq,uint8_t queue)3712 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3713 uint8_t queue)
3714 {
3715 int i;
3716
3717 for (i = 0; i < bq->bq_cnt; i++)
3718 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3719 vm_batchqueue_init(bq);
3720 }
3721
3722 /*
3723 * vm_page_pqbatch_submit: [ internal use only ]
3724 *
3725 * Enqueue a page in the specified page queue's batched work queue.
3726 * The caller must have encoded the requested operation in the page
3727 * structure's a.flags field.
3728 */
3729 void
vm_page_pqbatch_submit(vm_page_t m,uint8_t queue)3730 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3731 {
3732 struct vm_batchqueue *bq;
3733 struct vm_pagequeue *pq;
3734 int domain, slots_remaining;
3735
3736 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3737
3738 domain = vm_page_domain(m);
3739 critical_enter();
3740 bq = DPCPU_PTR(pqbatch[domain][queue]);
3741 slots_remaining = vm_batchqueue_insert(bq, m);
3742 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3743 /* keep building the bq */
3744 critical_exit();
3745 return;
3746 } else if (slots_remaining > 0 ) {
3747 /* Try to process the bq if we can get the lock */
3748 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3749 if (vm_pagequeue_trylock(pq)) {
3750 vm_pqbatch_process(pq, bq, queue);
3751 vm_pagequeue_unlock(pq);
3752 }
3753 critical_exit();
3754 return;
3755 }
3756 critical_exit();
3757
3758 /* if we make it here, the bq is full so wait for the lock */
3759
3760 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3761 vm_pagequeue_lock(pq);
3762 critical_enter();
3763 bq = DPCPU_PTR(pqbatch[domain][queue]);
3764 vm_pqbatch_process(pq, bq, queue);
3765 vm_pqbatch_process_page(pq, m, queue);
3766 vm_pagequeue_unlock(pq);
3767 critical_exit();
3768 }
3769
3770 /*
3771 * vm_page_pqbatch_drain: [ internal use only ]
3772 *
3773 * Force all per-CPU page queue batch queues to be drained. This is
3774 * intended for use in severe memory shortages, to ensure that pages
3775 * do not remain stuck in the batch queues.
3776 */
3777 void
vm_page_pqbatch_drain(void)3778 vm_page_pqbatch_drain(void)
3779 {
3780 struct thread *td;
3781 struct vm_domain *vmd;
3782 struct vm_pagequeue *pq;
3783 int cpu, domain, queue;
3784
3785 td = curthread;
3786 CPU_FOREACH(cpu) {
3787 thread_lock(td);
3788 sched_bind(td, cpu);
3789 thread_unlock(td);
3790
3791 for (domain = 0; domain < vm_ndomains; domain++) {
3792 vmd = VM_DOMAIN(domain);
3793 for (queue = 0; queue < PQ_COUNT; queue++) {
3794 pq = &vmd->vmd_pagequeues[queue];
3795 vm_pagequeue_lock(pq);
3796 critical_enter();
3797 vm_pqbatch_process(pq,
3798 DPCPU_PTR(pqbatch[domain][queue]), queue);
3799 critical_exit();
3800 vm_pagequeue_unlock(pq);
3801 }
3802 }
3803 }
3804 thread_lock(td);
3805 sched_unbind(td);
3806 thread_unlock(td);
3807 }
3808
3809 /*
3810 * vm_page_dequeue_deferred: [ internal use only ]
3811 *
3812 * Request removal of the given page from its current page
3813 * queue. Physical removal from the queue may be deferred
3814 * indefinitely.
3815 */
3816 void
vm_page_dequeue_deferred(vm_page_t m)3817 vm_page_dequeue_deferred(vm_page_t m)
3818 {
3819 vm_page_astate_t new, old;
3820
3821 old = vm_page_astate_load(m);
3822 do {
3823 if (old.queue == PQ_NONE) {
3824 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3825 ("%s: page %p has unexpected queue state",
3826 __func__, m));
3827 break;
3828 }
3829 new = old;
3830 new.flags |= PGA_DEQUEUE;
3831 } while (!vm_page_pqstate_commit_request(m, &old, new));
3832 }
3833
3834 /*
3835 * vm_page_dequeue:
3836 *
3837 * Remove the page from whichever page queue it's in, if any, before
3838 * returning.
3839 */
3840 void
vm_page_dequeue(vm_page_t m)3841 vm_page_dequeue(vm_page_t m)
3842 {
3843 vm_page_astate_t new, old;
3844
3845 old = vm_page_astate_load(m);
3846 do {
3847 if (old.queue == PQ_NONE) {
3848 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3849 ("%s: page %p has unexpected queue state",
3850 __func__, m));
3851 break;
3852 }
3853 new = old;
3854 new.flags &= ~PGA_QUEUE_OP_MASK;
3855 new.queue = PQ_NONE;
3856 } while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3857
3858 }
3859
3860 /*
3861 * Schedule the given page for insertion into the specified page queue.
3862 * Physical insertion of the page may be deferred indefinitely.
3863 */
3864 static void
vm_page_enqueue(vm_page_t m,uint8_t queue)3865 vm_page_enqueue(vm_page_t m, uint8_t queue)
3866 {
3867
3868 KASSERT(m->a.queue == PQ_NONE &&
3869 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3870 ("%s: page %p is already enqueued", __func__, m));
3871 KASSERT(m->ref_count > 0,
3872 ("%s: page %p does not carry any references", __func__, m));
3873
3874 m->a.queue = queue;
3875 if ((m->a.flags & PGA_REQUEUE) == 0)
3876 vm_page_aflag_set(m, PGA_REQUEUE);
3877 vm_page_pqbatch_submit(m, queue);
3878 }
3879
3880 /*
3881 * vm_page_free_prep:
3882 *
3883 * Prepares the given page to be put on the free list,
3884 * disassociating it from any VM object. The caller may return
3885 * the page to the free list only if this function returns true.
3886 *
3887 * The object, if it exists, must be locked, and then the page must
3888 * be xbusy. Otherwise the page must be not busied. A managed
3889 * page must be unmapped.
3890 */
3891 static bool
vm_page_free_prep(vm_page_t m)3892 vm_page_free_prep(vm_page_t m)
3893 {
3894
3895 /*
3896 * Synchronize with threads that have dropped a reference to this
3897 * page.
3898 */
3899 atomic_thread_fence_acq();
3900
3901 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3902 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3903 uint64_t *p;
3904 int i;
3905 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3906 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3907 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3908 m, i, (uintmax_t)*p));
3909 }
3910 #endif
3911 if ((m->oflags & VPO_UNMANAGED) == 0) {
3912 KASSERT(!pmap_page_is_mapped(m),
3913 ("vm_page_free_prep: freeing mapped page %p", m));
3914 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3915 ("vm_page_free_prep: mapping flags set in page %p", m));
3916 } else {
3917 KASSERT(m->a.queue == PQ_NONE,
3918 ("vm_page_free_prep: unmanaged page %p is queued", m));
3919 }
3920 VM_CNT_INC(v_tfree);
3921
3922 if (m->object != NULL) {
3923 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3924 ((m->object->flags & OBJ_UNMANAGED) != 0),
3925 ("vm_page_free_prep: managed flag mismatch for page %p",
3926 m));
3927 vm_page_assert_xbusied(m);
3928
3929 /*
3930 * The object reference can be released without an atomic
3931 * operation.
3932 */
3933 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3934 m->ref_count == VPRC_OBJREF,
3935 ("vm_page_free_prep: page %p has unexpected ref_count %u",
3936 m, m->ref_count));
3937 vm_page_object_remove(m);
3938 m->ref_count -= VPRC_OBJREF;
3939 } else
3940 vm_page_assert_unbusied(m);
3941
3942 vm_page_busy_free(m);
3943
3944 /*
3945 * If fictitious remove object association and
3946 * return.
3947 */
3948 if ((m->flags & PG_FICTITIOUS) != 0) {
3949 KASSERT(m->ref_count == 1,
3950 ("fictitious page %p is referenced", m));
3951 KASSERT(m->a.queue == PQ_NONE,
3952 ("fictitious page %p is queued", m));
3953 return (false);
3954 }
3955
3956 /*
3957 * Pages need not be dequeued before they are returned to the physical
3958 * memory allocator, but they must at least be marked for a deferred
3959 * dequeue.
3960 */
3961 if ((m->oflags & VPO_UNMANAGED) == 0)
3962 vm_page_dequeue_deferred(m);
3963
3964 m->valid = 0;
3965 vm_page_undirty(m);
3966
3967 if (m->ref_count != 0)
3968 panic("vm_page_free_prep: page %p has references", m);
3969
3970 /*
3971 * Restore the default memory attribute to the page.
3972 */
3973 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3974 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3975
3976 #if VM_NRESERVLEVEL > 0
3977 /*
3978 * Determine whether the page belongs to a reservation. If the page was
3979 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3980 * as an optimization, we avoid the check in that case.
3981 */
3982 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3983 return (false);
3984 #endif
3985
3986 return (true);
3987 }
3988
3989 /*
3990 * vm_page_free_toq:
3991 *
3992 * Returns the given page to the free list, disassociating it
3993 * from any VM object.
3994 *
3995 * The object must be locked. The page must be exclusively busied if it
3996 * belongs to an object.
3997 */
3998 static void
vm_page_free_toq(vm_page_t m)3999 vm_page_free_toq(vm_page_t m)
4000 {
4001 struct vm_domain *vmd;
4002 uma_zone_t zone;
4003
4004 if (!vm_page_free_prep(m))
4005 return;
4006
4007 vmd = vm_pagequeue_domain(m);
4008 zone = vmd->vmd_pgcache[m->pool].zone;
4009 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
4010 uma_zfree(zone, m);
4011 return;
4012 }
4013 vm_domain_free_lock(vmd);
4014 vm_phys_free_pages(m, 0);
4015 vm_domain_free_unlock(vmd);
4016 vm_domain_freecnt_inc(vmd, 1);
4017 }
4018
4019 /*
4020 * vm_page_free_pages_toq:
4021 *
4022 * Returns a list of pages to the free list, disassociating it
4023 * from any VM object. In other words, this is equivalent to
4024 * calling vm_page_free_toq() for each page of a list of VM objects.
4025 */
4026 void
vm_page_free_pages_toq(struct spglist * free,bool update_wire_count)4027 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
4028 {
4029 vm_page_t m;
4030 int count;
4031
4032 if (SLIST_EMPTY(free))
4033 return;
4034
4035 count = 0;
4036 while ((m = SLIST_FIRST(free)) != NULL) {
4037 count++;
4038 SLIST_REMOVE_HEAD(free, plinks.s.ss);
4039 vm_page_free_toq(m);
4040 }
4041
4042 if (update_wire_count)
4043 vm_wire_sub(count);
4044 }
4045
4046 /*
4047 * Mark this page as wired down. For managed pages, this prevents reclamation
4048 * by the page daemon, or when the containing object, if any, is destroyed.
4049 */
4050 void
vm_page_wire(vm_page_t m)4051 vm_page_wire(vm_page_t m)
4052 {
4053 u_int old;
4054
4055 #ifdef INVARIANTS
4056 if (m->object != NULL && !vm_page_busied(m) &&
4057 !vm_object_busied(m->object))
4058 VM_OBJECT_ASSERT_LOCKED(m->object);
4059 #endif
4060 KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4061 VPRC_WIRE_COUNT(m->ref_count) >= 1,
4062 ("vm_page_wire: fictitious page %p has zero wirings", m));
4063
4064 old = atomic_fetchadd_int(&m->ref_count, 1);
4065 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4066 ("vm_page_wire: counter overflow for page %p", m));
4067 if (VPRC_WIRE_COUNT(old) == 0) {
4068 if ((m->oflags & VPO_UNMANAGED) == 0)
4069 vm_page_aflag_set(m, PGA_DEQUEUE);
4070 vm_wire_add(1);
4071 }
4072 }
4073
4074 /*
4075 * Attempt to wire a mapped page following a pmap lookup of that page.
4076 * This may fail if a thread is concurrently tearing down mappings of the page.
4077 * The transient failure is acceptable because it translates to the
4078 * failure of the caller pmap_extract_and_hold(), which should be then
4079 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4080 */
4081 bool
vm_page_wire_mapped(vm_page_t m)4082 vm_page_wire_mapped(vm_page_t m)
4083 {
4084 u_int old;
4085
4086 old = m->ref_count;
4087 do {
4088 KASSERT(old > 0,
4089 ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4090 if ((old & VPRC_BLOCKED) != 0)
4091 return (false);
4092 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4093
4094 if (VPRC_WIRE_COUNT(old) == 0) {
4095 if ((m->oflags & VPO_UNMANAGED) == 0)
4096 vm_page_aflag_set(m, PGA_DEQUEUE);
4097 vm_wire_add(1);
4098 }
4099 return (true);
4100 }
4101
4102 /*
4103 * Release a wiring reference to a managed page. If the page still belongs to
4104 * an object, update its position in the page queues to reflect the reference.
4105 * If the wiring was the last reference to the page, free the page.
4106 */
4107 static void
vm_page_unwire_managed(vm_page_t m,uint8_t nqueue,bool noreuse)4108 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4109 {
4110 u_int old;
4111
4112 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4113 ("%s: page %p is unmanaged", __func__, m));
4114
4115 /*
4116 * Update LRU state before releasing the wiring reference.
4117 * Use a release store when updating the reference count to
4118 * synchronize with vm_page_free_prep().
4119 */
4120 old = m->ref_count;
4121 do {
4122 KASSERT(VPRC_WIRE_COUNT(old) > 0,
4123 ("vm_page_unwire: wire count underflow for page %p", m));
4124
4125 if (old > VPRC_OBJREF + 1) {
4126 /*
4127 * The page has at least one other wiring reference. An
4128 * earlier iteration of this loop may have called
4129 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4130 * re-set it if necessary.
4131 */
4132 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4133 vm_page_aflag_set(m, PGA_DEQUEUE);
4134 } else if (old == VPRC_OBJREF + 1) {
4135 /*
4136 * This is the last wiring. Clear PGA_DEQUEUE and
4137 * update the page's queue state to reflect the
4138 * reference. If the page does not belong to an object
4139 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4140 * clear leftover queue state.
4141 */
4142 vm_page_release_toq(m, nqueue, noreuse);
4143 } else if (old == 1) {
4144 vm_page_aflag_clear(m, PGA_DEQUEUE);
4145 }
4146 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4147
4148 if (VPRC_WIRE_COUNT(old) == 1) {
4149 vm_wire_sub(1);
4150 if (old == 1)
4151 vm_page_free(m);
4152 }
4153 }
4154
4155 /*
4156 * Release one wiring of the specified page, potentially allowing it to be
4157 * paged out.
4158 *
4159 * Only managed pages belonging to an object can be paged out. If the number
4160 * of wirings transitions to zero and the page is eligible for page out, then
4161 * the page is added to the specified paging queue. If the released wiring
4162 * represented the last reference to the page, the page is freed.
4163 */
4164 void
vm_page_unwire(vm_page_t m,uint8_t nqueue)4165 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4166 {
4167
4168 KASSERT(nqueue < PQ_COUNT,
4169 ("vm_page_unwire: invalid queue %u request for page %p",
4170 nqueue, m));
4171
4172 if ((m->oflags & VPO_UNMANAGED) != 0) {
4173 if (vm_page_unwire_noq(m) && m->ref_count == 0)
4174 vm_page_free(m);
4175 return;
4176 }
4177 vm_page_unwire_managed(m, nqueue, false);
4178 }
4179
4180 /*
4181 * Unwire a page without (re-)inserting it into a page queue. It is up
4182 * to the caller to enqueue, requeue, or free the page as appropriate.
4183 * In most cases involving managed pages, vm_page_unwire() should be used
4184 * instead.
4185 */
4186 bool
vm_page_unwire_noq(vm_page_t m)4187 vm_page_unwire_noq(vm_page_t m)
4188 {
4189 u_int old;
4190
4191 old = vm_page_drop(m, 1);
4192 KASSERT(VPRC_WIRE_COUNT(old) != 0,
4193 ("%s: counter underflow for page %p", __func__, m));
4194 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4195 ("%s: missing ref on fictitious page %p", __func__, m));
4196
4197 if (VPRC_WIRE_COUNT(old) > 1)
4198 return (false);
4199 if ((m->oflags & VPO_UNMANAGED) == 0)
4200 vm_page_aflag_clear(m, PGA_DEQUEUE);
4201 vm_wire_sub(1);
4202 return (true);
4203 }
4204
4205 /*
4206 * Ensure that the page ends up in the specified page queue. If the page is
4207 * active or being moved to the active queue, ensure that its act_count is
4208 * at least ACT_INIT but do not otherwise mess with it.
4209 */
4210 static __always_inline void
vm_page_mvqueue(vm_page_t m,const uint8_t nqueue,const uint16_t nflag)4211 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4212 {
4213 vm_page_astate_t old, new;
4214
4215 KASSERT(m->ref_count > 0,
4216 ("%s: page %p does not carry any references", __func__, m));
4217 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4218 ("%s: invalid flags %x", __func__, nflag));
4219
4220 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4221 return;
4222
4223 old = vm_page_astate_load(m);
4224 do {
4225 if ((old.flags & PGA_DEQUEUE) != 0)
4226 break;
4227 new = old;
4228 new.flags &= ~PGA_QUEUE_OP_MASK;
4229 if (nqueue == PQ_ACTIVE)
4230 new.act_count = max(old.act_count, ACT_INIT);
4231 if (old.queue == nqueue) {
4232 /*
4233 * There is no need to requeue pages already in the
4234 * active queue.
4235 */
4236 if (nqueue != PQ_ACTIVE ||
4237 (old.flags & PGA_ENQUEUED) == 0)
4238 new.flags |= nflag;
4239 } else {
4240 new.flags |= nflag;
4241 new.queue = nqueue;
4242 }
4243 } while (!vm_page_pqstate_commit(m, &old, new));
4244 }
4245
4246 /*
4247 * Put the specified page on the active list (if appropriate).
4248 */
4249 void
vm_page_activate(vm_page_t m)4250 vm_page_activate(vm_page_t m)
4251 {
4252
4253 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4254 }
4255
4256 /*
4257 * Move the specified page to the tail of the inactive queue, or requeue
4258 * the page if it is already in the inactive queue.
4259 */
4260 void
vm_page_deactivate(vm_page_t m)4261 vm_page_deactivate(vm_page_t m)
4262 {
4263
4264 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4265 }
4266
4267 void
vm_page_deactivate_noreuse(vm_page_t m)4268 vm_page_deactivate_noreuse(vm_page_t m)
4269 {
4270
4271 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4272 }
4273
4274 /*
4275 * Put a page in the laundry, or requeue it if it is already there.
4276 */
4277 void
vm_page_launder(vm_page_t m)4278 vm_page_launder(vm_page_t m)
4279 {
4280
4281 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4282 }
4283
4284 /*
4285 * Put a page in the PQ_UNSWAPPABLE holding queue.
4286 */
4287 void
vm_page_unswappable(vm_page_t m)4288 vm_page_unswappable(vm_page_t m)
4289 {
4290
4291 VM_OBJECT_ASSERT_LOCKED(m->object);
4292 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4293 ("page %p already unswappable", m));
4294
4295 vm_page_dequeue(m);
4296 vm_page_enqueue(m, PQ_UNSWAPPABLE);
4297 }
4298
4299 /*
4300 * Release a page back to the page queues in preparation for unwiring.
4301 */
4302 static void
vm_page_release_toq(vm_page_t m,uint8_t nqueue,const bool noreuse)4303 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4304 {
4305 vm_page_astate_t old, new;
4306 uint16_t nflag;
4307
4308 /*
4309 * Use a check of the valid bits to determine whether we should
4310 * accelerate reclamation of the page. The object lock might not be
4311 * held here, in which case the check is racy. At worst we will either
4312 * accelerate reclamation of a valid page and violate LRU, or
4313 * unnecessarily defer reclamation of an invalid page.
4314 *
4315 * If we were asked to not cache the page, place it near the head of the
4316 * inactive queue so that is reclaimed sooner.
4317 */
4318 if (noreuse || vm_page_none_valid(m)) {
4319 nqueue = PQ_INACTIVE;
4320 nflag = PGA_REQUEUE_HEAD;
4321 } else {
4322 nflag = PGA_REQUEUE;
4323 }
4324
4325 old = vm_page_astate_load(m);
4326 do {
4327 new = old;
4328
4329 /*
4330 * If the page is already in the active queue and we are not
4331 * trying to accelerate reclamation, simply mark it as
4332 * referenced and avoid any queue operations.
4333 */
4334 new.flags &= ~PGA_QUEUE_OP_MASK;
4335 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4336 (old.flags & PGA_ENQUEUED) != 0)
4337 new.flags |= PGA_REFERENCED;
4338 else {
4339 new.flags |= nflag;
4340 new.queue = nqueue;
4341 }
4342 } while (!vm_page_pqstate_commit(m, &old, new));
4343 }
4344
4345 /*
4346 * Unwire a page and either attempt to free it or re-add it to the page queues.
4347 */
4348 void
vm_page_release(vm_page_t m,int flags)4349 vm_page_release(vm_page_t m, int flags)
4350 {
4351 vm_object_t object;
4352
4353 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4354 ("vm_page_release: page %p is unmanaged", m));
4355
4356 if ((flags & VPR_TRYFREE) != 0) {
4357 for (;;) {
4358 object = atomic_load_ptr(&m->object);
4359 if (object == NULL)
4360 break;
4361 /* Depends on type-stability. */
4362 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4363 break;
4364 if (object == m->object) {
4365 vm_page_release_locked(m, flags);
4366 VM_OBJECT_WUNLOCK(object);
4367 return;
4368 }
4369 VM_OBJECT_WUNLOCK(object);
4370 }
4371 }
4372 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4373 }
4374
4375 /* See vm_page_release(). */
4376 void
vm_page_release_locked(vm_page_t m,int flags)4377 vm_page_release_locked(vm_page_t m, int flags)
4378 {
4379
4380 VM_OBJECT_ASSERT_WLOCKED(m->object);
4381 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4382 ("vm_page_release_locked: page %p is unmanaged", m));
4383
4384 if (vm_page_unwire_noq(m)) {
4385 if ((flags & VPR_TRYFREE) != 0 &&
4386 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4387 m->dirty == 0 && vm_page_tryxbusy(m)) {
4388 /*
4389 * An unlocked lookup may have wired the page before the
4390 * busy lock was acquired, in which case the page must
4391 * not be freed.
4392 */
4393 if (__predict_true(!vm_page_wired(m))) {
4394 vm_page_free(m);
4395 return;
4396 }
4397 vm_page_xunbusy(m);
4398 } else {
4399 vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4400 }
4401 }
4402 }
4403
4404 static bool
vm_page_try_blocked_op(vm_page_t m,void (* op)(vm_page_t))4405 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4406 {
4407 u_int old;
4408
4409 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4410 ("vm_page_try_blocked_op: page %p has no object", m));
4411 KASSERT(vm_page_busied(m),
4412 ("vm_page_try_blocked_op: page %p is not busy", m));
4413 VM_OBJECT_ASSERT_LOCKED(m->object);
4414
4415 old = m->ref_count;
4416 do {
4417 KASSERT(old != 0,
4418 ("vm_page_try_blocked_op: page %p has no references", m));
4419 if (VPRC_WIRE_COUNT(old) != 0)
4420 return (false);
4421 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4422
4423 (op)(m);
4424
4425 /*
4426 * If the object is read-locked, new wirings may be created via an
4427 * object lookup.
4428 */
4429 old = vm_page_drop(m, VPRC_BLOCKED);
4430 KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4431 old == (VPRC_BLOCKED | VPRC_OBJREF),
4432 ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4433 old, m));
4434 return (true);
4435 }
4436
4437 /*
4438 * Atomically check for wirings and remove all mappings of the page.
4439 */
4440 bool
vm_page_try_remove_all(vm_page_t m)4441 vm_page_try_remove_all(vm_page_t m)
4442 {
4443
4444 return (vm_page_try_blocked_op(m, pmap_remove_all));
4445 }
4446
4447 /*
4448 * Atomically check for wirings and remove all writeable mappings of the page.
4449 */
4450 bool
vm_page_try_remove_write(vm_page_t m)4451 vm_page_try_remove_write(vm_page_t m)
4452 {
4453
4454 return (vm_page_try_blocked_op(m, pmap_remove_write));
4455 }
4456
4457 /*
4458 * vm_page_advise
4459 *
4460 * Apply the specified advice to the given page.
4461 */
4462 void
vm_page_advise(vm_page_t m,int advice)4463 vm_page_advise(vm_page_t m, int advice)
4464 {
4465
4466 VM_OBJECT_ASSERT_WLOCKED(m->object);
4467 vm_page_assert_xbusied(m);
4468
4469 if (advice == MADV_FREE)
4470 /*
4471 * Mark the page clean. This will allow the page to be freed
4472 * without first paging it out. MADV_FREE pages are often
4473 * quickly reused by malloc(3), so we do not do anything that
4474 * would result in a page fault on a later access.
4475 */
4476 vm_page_undirty(m);
4477 else if (advice != MADV_DONTNEED) {
4478 if (advice == MADV_WILLNEED)
4479 vm_page_activate(m);
4480 return;
4481 }
4482
4483 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4484 vm_page_dirty(m);
4485
4486 /*
4487 * Clear any references to the page. Otherwise, the page daemon will
4488 * immediately reactivate the page.
4489 */
4490 vm_page_aflag_clear(m, PGA_REFERENCED);
4491
4492 /*
4493 * Place clean pages near the head of the inactive queue rather than
4494 * the tail, thus defeating the queue's LRU operation and ensuring that
4495 * the page will be reused quickly. Dirty pages not already in the
4496 * laundry are moved there.
4497 */
4498 if (m->dirty == 0)
4499 vm_page_deactivate_noreuse(m);
4500 else if (!vm_page_in_laundry(m))
4501 vm_page_launder(m);
4502 }
4503
4504 /*
4505 * vm_page_grab_release
4506 *
4507 * Helper routine for grab functions to release busy on return.
4508 */
4509 static inline void
vm_page_grab_release(vm_page_t m,int allocflags)4510 vm_page_grab_release(vm_page_t m, int allocflags)
4511 {
4512
4513 if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4514 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4515 vm_page_sunbusy(m);
4516 else
4517 vm_page_xunbusy(m);
4518 }
4519 }
4520
4521 /*
4522 * vm_page_grab_sleep
4523 *
4524 * Sleep for busy according to VM_ALLOC_ parameters. Returns true
4525 * if the caller should retry and false otherwise.
4526 *
4527 * If the object is locked on entry the object will be unlocked with
4528 * false returns and still locked but possibly having been dropped
4529 * with true returns.
4530 */
4531 static bool
vm_page_grab_sleep(vm_object_t object,vm_page_t m,vm_pindex_t pindex,const char * wmesg,int allocflags,bool locked)4532 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4533 const char *wmesg, int allocflags, bool locked)
4534 {
4535
4536 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4537 return (false);
4538
4539 /*
4540 * Reference the page before unlocking and sleeping so that
4541 * the page daemon is less likely to reclaim it.
4542 */
4543 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4544 vm_page_reference(m);
4545
4546 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4547 locked)
4548 VM_OBJECT_WLOCK(object);
4549 if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4550 return (false);
4551
4552 return (true);
4553 }
4554
4555 /*
4556 * Assert that the grab flags are valid.
4557 */
4558 static inline void
vm_page_grab_check(int allocflags)4559 vm_page_grab_check(int allocflags)
4560 {
4561
4562 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4563 (allocflags & VM_ALLOC_WIRED) != 0,
4564 ("vm_page_grab*: the pages must be busied or wired"));
4565
4566 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4567 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4568 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4569 }
4570
4571 /*
4572 * Calculate the page allocation flags for grab.
4573 */
4574 static inline int
vm_page_grab_pflags(int allocflags)4575 vm_page_grab_pflags(int allocflags)
4576 {
4577 int pflags;
4578
4579 pflags = allocflags &
4580 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4581 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4582 if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4583 pflags |= VM_ALLOC_WAITFAIL;
4584 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4585 pflags |= VM_ALLOC_SBUSY;
4586
4587 return (pflags);
4588 }
4589
4590 /*
4591 * Grab a page, waiting until we are waken up due to the page
4592 * changing state. We keep on waiting, if the page continues
4593 * to be in the object. If the page doesn't exist, first allocate it
4594 * and then conditionally zero it.
4595 *
4596 * This routine may sleep.
4597 *
4598 * The object must be locked on entry. The lock will, however, be released
4599 * and reacquired if the routine sleeps.
4600 */
4601 vm_page_t
vm_page_grab(vm_object_t object,vm_pindex_t pindex,int allocflags)4602 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4603 {
4604 vm_page_t m;
4605
4606 VM_OBJECT_ASSERT_WLOCKED(object);
4607 vm_page_grab_check(allocflags);
4608
4609 retrylookup:
4610 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4611 if (!vm_page_tryacquire(m, allocflags)) {
4612 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4613 allocflags, true))
4614 goto retrylookup;
4615 return (NULL);
4616 }
4617 goto out;
4618 }
4619 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4620 return (NULL);
4621 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4622 if (m == NULL) {
4623 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4624 return (NULL);
4625 goto retrylookup;
4626 }
4627 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4628 pmap_zero_page(m);
4629
4630 out:
4631 vm_page_grab_release(m, allocflags);
4632
4633 return (m);
4634 }
4635
4636 /*
4637 * Locklessly attempt to acquire a page given a (object, pindex) tuple
4638 * and an optional previous page to avoid the radix lookup. The resulting
4639 * page will be validated against the identity tuple and busied or wired
4640 * as requested. A NULL *mp return guarantees that the page was not in
4641 * radix at the time of the call but callers must perform higher level
4642 * synchronization or retry the operation under a lock if they require
4643 * an atomic answer. This is the only lock free validation routine,
4644 * other routines can depend on the resulting page state.
4645 *
4646 * The return value indicates whether the operation failed due to caller
4647 * flags. The return is tri-state with mp:
4648 *
4649 * (true, *mp != NULL) - The operation was successful.
4650 * (true, *mp == NULL) - The page was not found in tree.
4651 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4652 */
4653 static bool
vm_page_acquire_unlocked(vm_object_t object,vm_pindex_t pindex,vm_page_t prev,vm_page_t * mp,int allocflags)4654 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4655 vm_page_t prev, vm_page_t *mp, int allocflags)
4656 {
4657 vm_page_t m;
4658
4659 vm_page_grab_check(allocflags);
4660 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4661
4662 *mp = NULL;
4663 for (;;) {
4664 /*
4665 * We may see a false NULL here because the previous page
4666 * has been removed or just inserted and the list is loaded
4667 * without barriers. Switch to radix to verify.
4668 */
4669 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4670 QMD_IS_TRASHED(m) || m->pindex != pindex ||
4671 atomic_load_ptr(&m->object) != object) {
4672 prev = NULL;
4673 /*
4674 * This guarantees the result is instantaneously
4675 * correct.
4676 */
4677 m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4678 }
4679 if (m == NULL)
4680 return (true);
4681 if (vm_page_trybusy(m, allocflags)) {
4682 if (m->object == object && m->pindex == pindex)
4683 break;
4684 /* relookup. */
4685 vm_page_busy_release(m);
4686 cpu_spinwait();
4687 continue;
4688 }
4689 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4690 allocflags, false))
4691 return (false);
4692 }
4693 if ((allocflags & VM_ALLOC_WIRED) != 0)
4694 vm_page_wire(m);
4695 vm_page_grab_release(m, allocflags);
4696 *mp = m;
4697 return (true);
4698 }
4699
4700 /*
4701 * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4702 * is not set.
4703 */
4704 vm_page_t
vm_page_grab_unlocked(vm_object_t object,vm_pindex_t pindex,int allocflags)4705 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4706 {
4707 vm_page_t m;
4708
4709 vm_page_grab_check(allocflags);
4710
4711 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4712 return (NULL);
4713 if (m != NULL)
4714 return (m);
4715
4716 /*
4717 * The radix lockless lookup should never return a false negative
4718 * errors. If the user specifies NOCREAT they are guaranteed there
4719 * was no page present at the instant of the call. A NOCREAT caller
4720 * must handle create races gracefully.
4721 */
4722 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4723 return (NULL);
4724
4725 VM_OBJECT_WLOCK(object);
4726 m = vm_page_grab(object, pindex, allocflags);
4727 VM_OBJECT_WUNLOCK(object);
4728
4729 return (m);
4730 }
4731
4732 /*
4733 * Grab a page and make it valid, paging in if necessary. Pages missing from
4734 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
4735 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4736 * in simultaneously. Additional pages will be left on a paging queue but
4737 * will neither be wired nor busy regardless of allocflags.
4738 */
4739 int
vm_page_grab_valid(vm_page_t * mp,vm_object_t object,vm_pindex_t pindex,int allocflags)4740 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4741 {
4742 vm_page_t m;
4743 vm_page_t ma[VM_INITIAL_PAGEIN];
4744 int after, i, pflags, rv;
4745
4746 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4747 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4748 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4749 KASSERT((allocflags &
4750 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4751 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4752 VM_OBJECT_ASSERT_WLOCKED(object);
4753 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4754 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4755 pflags |= VM_ALLOC_WAITFAIL;
4756
4757 retrylookup:
4758 if ((m = vm_page_lookup(object, pindex)) != NULL) {
4759 /*
4760 * If the page is fully valid it can only become invalid
4761 * with the object lock held. If it is not valid it can
4762 * become valid with the busy lock held. Therefore, we
4763 * may unnecessarily lock the exclusive busy here if we
4764 * race with I/O completion not using the object lock.
4765 * However, we will not end up with an invalid page and a
4766 * shared lock.
4767 */
4768 if (!vm_page_trybusy(m,
4769 vm_page_all_valid(m) ? allocflags : 0)) {
4770 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4771 allocflags, true);
4772 goto retrylookup;
4773 }
4774 if (vm_page_all_valid(m))
4775 goto out;
4776 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4777 vm_page_busy_release(m);
4778 *mp = NULL;
4779 return (VM_PAGER_FAIL);
4780 }
4781 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4782 *mp = NULL;
4783 return (VM_PAGER_FAIL);
4784 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4785 if (!vm_pager_can_alloc_page(object, pindex)) {
4786 *mp = NULL;
4787 return (VM_PAGER_AGAIN);
4788 }
4789 goto retrylookup;
4790 }
4791
4792 vm_page_assert_xbusied(m);
4793 if (vm_pager_has_page(object, pindex, NULL, &after)) {
4794 after = MIN(after, VM_INITIAL_PAGEIN);
4795 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4796 after = MAX(after, 1);
4797 ma[0] = m;
4798 for (i = 1; i < after; i++) {
4799 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4800 if (vm_page_any_valid(ma[i]) ||
4801 !vm_page_tryxbusy(ma[i]))
4802 break;
4803 } else {
4804 ma[i] = vm_page_alloc(object, m->pindex + i,
4805 VM_ALLOC_NORMAL);
4806 if (ma[i] == NULL)
4807 break;
4808 }
4809 }
4810 after = i;
4811 vm_object_pip_add(object, after);
4812 VM_OBJECT_WUNLOCK(object);
4813 rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4814 VM_OBJECT_WLOCK(object);
4815 vm_object_pip_wakeupn(object, after);
4816 /* Pager may have replaced a page. */
4817 m = ma[0];
4818 if (rv != VM_PAGER_OK) {
4819 for (i = 0; i < after; i++) {
4820 if (!vm_page_wired(ma[i]))
4821 vm_page_free(ma[i]);
4822 else
4823 vm_page_xunbusy(ma[i]);
4824 }
4825 *mp = NULL;
4826 return (rv);
4827 }
4828 for (i = 1; i < after; i++)
4829 vm_page_readahead_finish(ma[i]);
4830 MPASS(vm_page_all_valid(m));
4831 } else {
4832 vm_page_zero_invalid(m, TRUE);
4833 }
4834 out:
4835 if ((allocflags & VM_ALLOC_WIRED) != 0)
4836 vm_page_wire(m);
4837 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4838 vm_page_busy_downgrade(m);
4839 else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4840 vm_page_busy_release(m);
4841 *mp = m;
4842 return (VM_PAGER_OK);
4843 }
4844
4845 /*
4846 * Locklessly grab a valid page. If the page is not valid or not yet
4847 * allocated this will fall back to the object lock method.
4848 */
4849 int
vm_page_grab_valid_unlocked(vm_page_t * mp,vm_object_t object,vm_pindex_t pindex,int allocflags)4850 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4851 vm_pindex_t pindex, int allocflags)
4852 {
4853 vm_page_t m;
4854 int flags;
4855 int error;
4856
4857 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4858 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4859 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4860 "mismatch"));
4861 KASSERT((allocflags &
4862 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4863 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4864
4865 /*
4866 * Attempt a lockless lookup and busy. We need at least an sbusy
4867 * before we can inspect the valid field and return a wired page.
4868 */
4869 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4870 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4871 return (VM_PAGER_FAIL);
4872 if ((m = *mp) != NULL) {
4873 if (vm_page_all_valid(m)) {
4874 if ((allocflags & VM_ALLOC_WIRED) != 0)
4875 vm_page_wire(m);
4876 vm_page_grab_release(m, allocflags);
4877 return (VM_PAGER_OK);
4878 }
4879 vm_page_busy_release(m);
4880 }
4881 if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4882 *mp = NULL;
4883 return (VM_PAGER_FAIL);
4884 }
4885 VM_OBJECT_WLOCK(object);
4886 error = vm_page_grab_valid(mp, object, pindex, allocflags);
4887 VM_OBJECT_WUNLOCK(object);
4888
4889 return (error);
4890 }
4891
4892 /*
4893 * Return the specified range of pages from the given object. For each
4894 * page offset within the range, if a page already exists within the object
4895 * at that offset and it is busy, then wait for it to change state. If,
4896 * instead, the page doesn't exist, then allocate it.
4897 *
4898 * The caller must always specify an allocation class.
4899 *
4900 * allocation classes:
4901 * VM_ALLOC_NORMAL normal process request
4902 * VM_ALLOC_SYSTEM system *really* needs the pages
4903 *
4904 * The caller must always specify that the pages are to be busied and/or
4905 * wired.
4906 *
4907 * optional allocation flags:
4908 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages
4909 * VM_ALLOC_NOBUSY do not exclusive busy the page
4910 * VM_ALLOC_NOWAIT do not sleep
4911 * VM_ALLOC_SBUSY set page to sbusy state
4912 * VM_ALLOC_WIRED wire the pages
4913 * VM_ALLOC_ZERO zero and validate any invalid pages
4914 *
4915 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it
4916 * may return a partial prefix of the requested range.
4917 */
4918 int
vm_page_grab_pages(vm_object_t object,vm_pindex_t pindex,int allocflags,vm_page_t * ma,int count)4919 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4920 vm_page_t *ma, int count)
4921 {
4922 vm_page_t m, mpred;
4923 int pflags;
4924 int i;
4925
4926 VM_OBJECT_ASSERT_WLOCKED(object);
4927 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4928 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4929 KASSERT(count > 0,
4930 ("vm_page_grab_pages: invalid page count %d", count));
4931 vm_page_grab_check(allocflags);
4932
4933 pflags = vm_page_grab_pflags(allocflags);
4934 i = 0;
4935 retrylookup:
4936 m = vm_radix_lookup_le(&object->rtree, pindex + i);
4937 if (m == NULL || m->pindex != pindex + i) {
4938 mpred = m;
4939 m = NULL;
4940 } else
4941 mpred = TAILQ_PREV(m, pglist, listq);
4942 for (; i < count; i++) {
4943 if (m != NULL) {
4944 if (!vm_page_tryacquire(m, allocflags)) {
4945 if (vm_page_grab_sleep(object, m, pindex + i,
4946 "grbmaw", allocflags, true))
4947 goto retrylookup;
4948 break;
4949 }
4950 } else {
4951 if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4952 break;
4953 m = vm_page_alloc_after(object, pindex + i,
4954 pflags | VM_ALLOC_COUNT(count - i), mpred);
4955 if (m == NULL) {
4956 if ((allocflags & (VM_ALLOC_NOWAIT |
4957 VM_ALLOC_WAITFAIL)) != 0)
4958 break;
4959 goto retrylookup;
4960 }
4961 }
4962 if (vm_page_none_valid(m) &&
4963 (allocflags & VM_ALLOC_ZERO) != 0) {
4964 if ((m->flags & PG_ZERO) == 0)
4965 pmap_zero_page(m);
4966 vm_page_valid(m);
4967 }
4968 vm_page_grab_release(m, allocflags);
4969 ma[i] = mpred = m;
4970 m = vm_page_next(m);
4971 }
4972 return (i);
4973 }
4974
4975 /*
4976 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags
4977 * and will fall back to the locked variant to handle allocation.
4978 */
4979 int
vm_page_grab_pages_unlocked(vm_object_t object,vm_pindex_t pindex,int allocflags,vm_page_t * ma,int count)4980 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4981 int allocflags, vm_page_t *ma, int count)
4982 {
4983 vm_page_t m, pred;
4984 int flags;
4985 int i;
4986
4987 KASSERT(count > 0,
4988 ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4989 vm_page_grab_check(allocflags);
4990
4991 /*
4992 * Modify flags for lockless acquire to hold the page until we
4993 * set it valid if necessary.
4994 */
4995 flags = allocflags & ~VM_ALLOC_NOBUSY;
4996 pred = NULL;
4997 for (i = 0; i < count; i++, pindex++) {
4998 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4999 return (i);
5000 if (m == NULL)
5001 break;
5002 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
5003 if ((m->flags & PG_ZERO) == 0)
5004 pmap_zero_page(m);
5005 vm_page_valid(m);
5006 }
5007 /* m will still be wired or busy according to flags. */
5008 vm_page_grab_release(m, allocflags);
5009 pred = ma[i] = m;
5010 }
5011 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
5012 return (i);
5013 count -= i;
5014 VM_OBJECT_WLOCK(object);
5015 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
5016 VM_OBJECT_WUNLOCK(object);
5017
5018 return (i);
5019 }
5020
5021 /*
5022 * Mapping function for valid or dirty bits in a page.
5023 *
5024 * Inputs are required to range within a page.
5025 */
5026 vm_page_bits_t
vm_page_bits(int base,int size)5027 vm_page_bits(int base, int size)
5028 {
5029 int first_bit;
5030 int last_bit;
5031
5032 KASSERT(
5033 base + size <= PAGE_SIZE,
5034 ("vm_page_bits: illegal base/size %d/%d", base, size)
5035 );
5036
5037 if (size == 0) /* handle degenerate case */
5038 return (0);
5039
5040 first_bit = base >> DEV_BSHIFT;
5041 last_bit = (base + size - 1) >> DEV_BSHIFT;
5042
5043 return (((vm_page_bits_t)2 << last_bit) -
5044 ((vm_page_bits_t)1 << first_bit));
5045 }
5046
5047 void
vm_page_bits_set(vm_page_t m,vm_page_bits_t * bits,vm_page_bits_t set)5048 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5049 {
5050
5051 #if PAGE_SIZE == 32768
5052 atomic_set_64((uint64_t *)bits, set);
5053 #elif PAGE_SIZE == 16384
5054 atomic_set_32((uint32_t *)bits, set);
5055 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5056 atomic_set_16((uint16_t *)bits, set);
5057 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5058 atomic_set_8((uint8_t *)bits, set);
5059 #else /* PAGE_SIZE <= 8192 */
5060 uintptr_t addr;
5061 int shift;
5062
5063 addr = (uintptr_t)bits;
5064 /*
5065 * Use a trick to perform a 32-bit atomic on the
5066 * containing aligned word, to not depend on the existence
5067 * of atomic_{set, clear}_{8, 16}.
5068 */
5069 shift = addr & (sizeof(uint32_t) - 1);
5070 #if BYTE_ORDER == BIG_ENDIAN
5071 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5072 #else
5073 shift *= NBBY;
5074 #endif
5075 addr &= ~(sizeof(uint32_t) - 1);
5076 atomic_set_32((uint32_t *)addr, set << shift);
5077 #endif /* PAGE_SIZE */
5078 }
5079
5080 static inline void
vm_page_bits_clear(vm_page_t m,vm_page_bits_t * bits,vm_page_bits_t clear)5081 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5082 {
5083
5084 #if PAGE_SIZE == 32768
5085 atomic_clear_64((uint64_t *)bits, clear);
5086 #elif PAGE_SIZE == 16384
5087 atomic_clear_32((uint32_t *)bits, clear);
5088 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5089 atomic_clear_16((uint16_t *)bits, clear);
5090 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5091 atomic_clear_8((uint8_t *)bits, clear);
5092 #else /* PAGE_SIZE <= 8192 */
5093 uintptr_t addr;
5094 int shift;
5095
5096 addr = (uintptr_t)bits;
5097 /*
5098 * Use a trick to perform a 32-bit atomic on the
5099 * containing aligned word, to not depend on the existence
5100 * of atomic_{set, clear}_{8, 16}.
5101 */
5102 shift = addr & (sizeof(uint32_t) - 1);
5103 #if BYTE_ORDER == BIG_ENDIAN
5104 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5105 #else
5106 shift *= NBBY;
5107 #endif
5108 addr &= ~(sizeof(uint32_t) - 1);
5109 atomic_clear_32((uint32_t *)addr, clear << shift);
5110 #endif /* PAGE_SIZE */
5111 }
5112
5113 static inline vm_page_bits_t
vm_page_bits_swap(vm_page_t m,vm_page_bits_t * bits,vm_page_bits_t newbits)5114 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5115 {
5116 #if PAGE_SIZE == 32768
5117 uint64_t old;
5118
5119 old = *bits;
5120 while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5121 return (old);
5122 #elif PAGE_SIZE == 16384
5123 uint32_t old;
5124
5125 old = *bits;
5126 while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5127 return (old);
5128 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5129 uint16_t old;
5130
5131 old = *bits;
5132 while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5133 return (old);
5134 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5135 uint8_t old;
5136
5137 old = *bits;
5138 while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5139 return (old);
5140 #else /* PAGE_SIZE <= 4096*/
5141 uintptr_t addr;
5142 uint32_t old, new, mask;
5143 int shift;
5144
5145 addr = (uintptr_t)bits;
5146 /*
5147 * Use a trick to perform a 32-bit atomic on the
5148 * containing aligned word, to not depend on the existence
5149 * of atomic_{set, swap, clear}_{8, 16}.
5150 */
5151 shift = addr & (sizeof(uint32_t) - 1);
5152 #if BYTE_ORDER == BIG_ENDIAN
5153 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5154 #else
5155 shift *= NBBY;
5156 #endif
5157 addr &= ~(sizeof(uint32_t) - 1);
5158 mask = VM_PAGE_BITS_ALL << shift;
5159
5160 old = *bits;
5161 do {
5162 new = old & ~mask;
5163 new |= newbits << shift;
5164 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5165 return (old >> shift);
5166 #endif /* PAGE_SIZE */
5167 }
5168
5169 /*
5170 * vm_page_set_valid_range:
5171 *
5172 * Sets portions of a page valid. The arguments are expected
5173 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5174 * of any partial chunks touched by the range. The invalid portion of
5175 * such chunks will be zeroed.
5176 *
5177 * (base + size) must be less then or equal to PAGE_SIZE.
5178 */
5179 void
vm_page_set_valid_range(vm_page_t m,int base,int size)5180 vm_page_set_valid_range(vm_page_t m, int base, int size)
5181 {
5182 int endoff, frag;
5183 vm_page_bits_t pagebits;
5184
5185 vm_page_assert_busied(m);
5186 if (size == 0) /* handle degenerate case */
5187 return;
5188
5189 /*
5190 * If the base is not DEV_BSIZE aligned and the valid
5191 * bit is clear, we have to zero out a portion of the
5192 * first block.
5193 */
5194 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5195 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5196 pmap_zero_page_area(m, frag, base - frag);
5197
5198 /*
5199 * If the ending offset is not DEV_BSIZE aligned and the
5200 * valid bit is clear, we have to zero out a portion of
5201 * the last block.
5202 */
5203 endoff = base + size;
5204 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5205 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5206 pmap_zero_page_area(m, endoff,
5207 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5208
5209 /*
5210 * Assert that no previously invalid block that is now being validated
5211 * is already dirty.
5212 */
5213 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5214 ("vm_page_set_valid_range: page %p is dirty", m));
5215
5216 /*
5217 * Set valid bits inclusive of any overlap.
5218 */
5219 pagebits = vm_page_bits(base, size);
5220 if (vm_page_xbusied(m))
5221 m->valid |= pagebits;
5222 else
5223 vm_page_bits_set(m, &m->valid, pagebits);
5224 }
5225
5226 /*
5227 * Set the page dirty bits and free the invalid swap space if
5228 * present. Returns the previous dirty bits.
5229 */
5230 vm_page_bits_t
vm_page_set_dirty(vm_page_t m)5231 vm_page_set_dirty(vm_page_t m)
5232 {
5233 vm_page_bits_t old;
5234
5235 VM_PAGE_OBJECT_BUSY_ASSERT(m);
5236
5237 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5238 old = m->dirty;
5239 m->dirty = VM_PAGE_BITS_ALL;
5240 } else
5241 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5242 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5243 vm_pager_page_unswapped(m);
5244
5245 return (old);
5246 }
5247
5248 /*
5249 * Clear the given bits from the specified page's dirty field.
5250 */
5251 static __inline void
vm_page_clear_dirty_mask(vm_page_t m,vm_page_bits_t pagebits)5252 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5253 {
5254
5255 vm_page_assert_busied(m);
5256
5257 /*
5258 * If the page is xbusied and not write mapped we are the
5259 * only thread that can modify dirty bits. Otherwise, The pmap
5260 * layer can call vm_page_dirty() without holding a distinguished
5261 * lock. The combination of page busy and atomic operations
5262 * suffice to guarantee consistency of the page dirty field.
5263 */
5264 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5265 m->dirty &= ~pagebits;
5266 else
5267 vm_page_bits_clear(m, &m->dirty, pagebits);
5268 }
5269
5270 /*
5271 * vm_page_set_validclean:
5272 *
5273 * Sets portions of a page valid and clean. The arguments are expected
5274 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5275 * of any partial chunks touched by the range. The invalid portion of
5276 * such chunks will be zero'd.
5277 *
5278 * (base + size) must be less then or equal to PAGE_SIZE.
5279 */
5280 void
vm_page_set_validclean(vm_page_t m,int base,int size)5281 vm_page_set_validclean(vm_page_t m, int base, int size)
5282 {
5283 vm_page_bits_t oldvalid, pagebits;
5284 int endoff, frag;
5285
5286 vm_page_assert_busied(m);
5287 if (size == 0) /* handle degenerate case */
5288 return;
5289
5290 /*
5291 * If the base is not DEV_BSIZE aligned and the valid
5292 * bit is clear, we have to zero out a portion of the
5293 * first block.
5294 */
5295 if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5296 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5297 pmap_zero_page_area(m, frag, base - frag);
5298
5299 /*
5300 * If the ending offset is not DEV_BSIZE aligned and the
5301 * valid bit is clear, we have to zero out a portion of
5302 * the last block.
5303 */
5304 endoff = base + size;
5305 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5306 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5307 pmap_zero_page_area(m, endoff,
5308 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5309
5310 /*
5311 * Set valid, clear dirty bits. If validating the entire
5312 * page we can safely clear the pmap modify bit. We also
5313 * use this opportunity to clear the PGA_NOSYNC flag. If a process
5314 * takes a write fault on a MAP_NOSYNC memory area the flag will
5315 * be set again.
5316 *
5317 * We set valid bits inclusive of any overlap, but we can only
5318 * clear dirty bits for DEV_BSIZE chunks that are fully within
5319 * the range.
5320 */
5321 oldvalid = m->valid;
5322 pagebits = vm_page_bits(base, size);
5323 if (vm_page_xbusied(m))
5324 m->valid |= pagebits;
5325 else
5326 vm_page_bits_set(m, &m->valid, pagebits);
5327 #if 0 /* NOT YET */
5328 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5329 frag = DEV_BSIZE - frag;
5330 base += frag;
5331 size -= frag;
5332 if (size < 0)
5333 size = 0;
5334 }
5335 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5336 #endif
5337 if (base == 0 && size == PAGE_SIZE) {
5338 /*
5339 * The page can only be modified within the pmap if it is
5340 * mapped, and it can only be mapped if it was previously
5341 * fully valid.
5342 */
5343 if (oldvalid == VM_PAGE_BITS_ALL)
5344 /*
5345 * Perform the pmap_clear_modify() first. Otherwise,
5346 * a concurrent pmap operation, such as
5347 * pmap_protect(), could clear a modification in the
5348 * pmap and set the dirty field on the page before
5349 * pmap_clear_modify() had begun and after the dirty
5350 * field was cleared here.
5351 */
5352 pmap_clear_modify(m);
5353 m->dirty = 0;
5354 vm_page_aflag_clear(m, PGA_NOSYNC);
5355 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5356 m->dirty &= ~pagebits;
5357 else
5358 vm_page_clear_dirty_mask(m, pagebits);
5359 }
5360
5361 void
vm_page_clear_dirty(vm_page_t m,int base,int size)5362 vm_page_clear_dirty(vm_page_t m, int base, int size)
5363 {
5364
5365 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5366 }
5367
5368 /*
5369 * vm_page_set_invalid:
5370 *
5371 * Invalidates DEV_BSIZE'd chunks within a page. Both the
5372 * valid and dirty bits for the effected areas are cleared.
5373 */
5374 void
vm_page_set_invalid(vm_page_t m,int base,int size)5375 vm_page_set_invalid(vm_page_t m, int base, int size)
5376 {
5377 vm_page_bits_t bits;
5378 vm_object_t object;
5379
5380 /*
5381 * The object lock is required so that pages can't be mapped
5382 * read-only while we're in the process of invalidating them.
5383 */
5384 object = m->object;
5385 VM_OBJECT_ASSERT_WLOCKED(object);
5386 vm_page_assert_busied(m);
5387
5388 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5389 size >= object->un_pager.vnp.vnp_size)
5390 bits = VM_PAGE_BITS_ALL;
5391 else
5392 bits = vm_page_bits(base, size);
5393 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5394 pmap_remove_all(m);
5395 KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5396 !pmap_page_is_mapped(m),
5397 ("vm_page_set_invalid: page %p is mapped", m));
5398 if (vm_page_xbusied(m)) {
5399 m->valid &= ~bits;
5400 m->dirty &= ~bits;
5401 } else {
5402 vm_page_bits_clear(m, &m->valid, bits);
5403 vm_page_bits_clear(m, &m->dirty, bits);
5404 }
5405 }
5406
5407 /*
5408 * vm_page_invalid:
5409 *
5410 * Invalidates the entire page. The page must be busy, unmapped, and
5411 * the enclosing object must be locked. The object locks protects
5412 * against concurrent read-only pmap enter which is done without
5413 * busy.
5414 */
5415 void
vm_page_invalid(vm_page_t m)5416 vm_page_invalid(vm_page_t m)
5417 {
5418
5419 vm_page_assert_busied(m);
5420 VM_OBJECT_ASSERT_WLOCKED(m->object);
5421 MPASS(!pmap_page_is_mapped(m));
5422
5423 if (vm_page_xbusied(m))
5424 m->valid = 0;
5425 else
5426 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5427 }
5428
5429 /*
5430 * vm_page_zero_invalid()
5431 *
5432 * The kernel assumes that the invalid portions of a page contain
5433 * garbage, but such pages can be mapped into memory by user code.
5434 * When this occurs, we must zero out the non-valid portions of the
5435 * page so user code sees what it expects.
5436 *
5437 * Pages are most often semi-valid when the end of a file is mapped
5438 * into memory and the file's size is not page aligned.
5439 */
5440 void
vm_page_zero_invalid(vm_page_t m,boolean_t setvalid)5441 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5442 {
5443 int b;
5444 int i;
5445
5446 /*
5447 * Scan the valid bits looking for invalid sections that
5448 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
5449 * valid bit may be set ) have already been zeroed by
5450 * vm_page_set_validclean().
5451 */
5452 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5453 if (i == (PAGE_SIZE / DEV_BSIZE) ||
5454 (m->valid & ((vm_page_bits_t)1 << i))) {
5455 if (i > b) {
5456 pmap_zero_page_area(m,
5457 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5458 }
5459 b = i + 1;
5460 }
5461 }
5462
5463 /*
5464 * setvalid is TRUE when we can safely set the zero'd areas
5465 * as being valid. We can do this if there are no cache consistency
5466 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
5467 */
5468 if (setvalid)
5469 vm_page_valid(m);
5470 }
5471
5472 /*
5473 * vm_page_is_valid:
5474 *
5475 * Is (partial) page valid? Note that the case where size == 0
5476 * will return FALSE in the degenerate case where the page is
5477 * entirely invalid, and TRUE otherwise.
5478 *
5479 * Some callers envoke this routine without the busy lock held and
5480 * handle races via higher level locks. Typical callers should
5481 * hold a busy lock to prevent invalidation.
5482 */
5483 int
vm_page_is_valid(vm_page_t m,int base,int size)5484 vm_page_is_valid(vm_page_t m, int base, int size)
5485 {
5486 vm_page_bits_t bits;
5487
5488 bits = vm_page_bits(base, size);
5489 return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5490 }
5491
5492 /*
5493 * Returns true if all of the specified predicates are true for the entire
5494 * (super)page and false otherwise.
5495 */
5496 bool
vm_page_ps_test(vm_page_t m,int flags,vm_page_t skip_m)5497 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5498 {
5499 vm_object_t object;
5500 int i, npages;
5501
5502 object = m->object;
5503 if (skip_m != NULL && skip_m->object != object)
5504 return (false);
5505 VM_OBJECT_ASSERT_LOCKED(object);
5506 npages = atop(pagesizes[m->psind]);
5507
5508 /*
5509 * The physically contiguous pages that make up a superpage, i.e., a
5510 * page with a page size index ("psind") greater than zero, will
5511 * occupy adjacent entries in vm_page_array[].
5512 */
5513 for (i = 0; i < npages; i++) {
5514 /* Always test object consistency, including "skip_m". */
5515 if (m[i].object != object)
5516 return (false);
5517 if (&m[i] == skip_m)
5518 continue;
5519 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5520 return (false);
5521 if ((flags & PS_ALL_DIRTY) != 0) {
5522 /*
5523 * Calling vm_page_test_dirty() or pmap_is_modified()
5524 * might stop this case from spuriously returning
5525 * "false". However, that would require a write lock
5526 * on the object containing "m[i]".
5527 */
5528 if (m[i].dirty != VM_PAGE_BITS_ALL)
5529 return (false);
5530 }
5531 if ((flags & PS_ALL_VALID) != 0 &&
5532 m[i].valid != VM_PAGE_BITS_ALL)
5533 return (false);
5534 }
5535 return (true);
5536 }
5537
5538 /*
5539 * Set the page's dirty bits if the page is modified.
5540 */
5541 void
vm_page_test_dirty(vm_page_t m)5542 vm_page_test_dirty(vm_page_t m)
5543 {
5544
5545 vm_page_assert_busied(m);
5546 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5547 vm_page_dirty(m);
5548 }
5549
5550 void
vm_page_valid(vm_page_t m)5551 vm_page_valid(vm_page_t m)
5552 {
5553
5554 vm_page_assert_busied(m);
5555 if (vm_page_xbusied(m))
5556 m->valid = VM_PAGE_BITS_ALL;
5557 else
5558 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5559 }
5560
5561 void
vm_page_lock_KBI(vm_page_t m,const char * file,int line)5562 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5563 {
5564
5565 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5566 }
5567
5568 void
vm_page_unlock_KBI(vm_page_t m,const char * file,int line)5569 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5570 {
5571
5572 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5573 }
5574
5575 int
vm_page_trylock_KBI(vm_page_t m,const char * file,int line)5576 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5577 {
5578
5579 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5580 }
5581
5582 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5583 void
vm_page_assert_locked_KBI(vm_page_t m,const char * file,int line)5584 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5585 {
5586
5587 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5588 }
5589
5590 void
vm_page_lock_assert_KBI(vm_page_t m,int a,const char * file,int line)5591 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5592 {
5593
5594 mtx_assert_(vm_page_lockptr(m), a, file, line);
5595 }
5596 #endif
5597
5598 #ifdef INVARIANTS
5599 void
vm_page_object_busy_assert(vm_page_t m)5600 vm_page_object_busy_assert(vm_page_t m)
5601 {
5602
5603 /*
5604 * Certain of the page's fields may only be modified by the
5605 * holder of a page or object busy.
5606 */
5607 if (m->object != NULL && !vm_page_busied(m))
5608 VM_OBJECT_ASSERT_BUSY(m->object);
5609 }
5610
5611 void
vm_page_assert_pga_writeable(vm_page_t m,uint16_t bits)5612 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5613 {
5614
5615 if ((bits & PGA_WRITEABLE) == 0)
5616 return;
5617
5618 /*
5619 * The PGA_WRITEABLE flag can only be set if the page is
5620 * managed, is exclusively busied or the object is locked.
5621 * Currently, this flag is only set by pmap_enter().
5622 */
5623 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5624 ("PGA_WRITEABLE on unmanaged page"));
5625 if (!vm_page_xbusied(m))
5626 VM_OBJECT_ASSERT_BUSY(m->object);
5627 }
5628 #endif
5629
5630 #include "opt_ddb.h"
5631 #ifdef DDB
5632 #include <sys/kernel.h>
5633
5634 #include <ddb/ddb.h>
5635
DB_SHOW_COMMAND_FLAGS(page,vm_page_print_page_info,DB_CMD_MEMSAFE)5636 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5637 {
5638
5639 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5640 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5641 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5642 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5643 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5644 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5645 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5646 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5647 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5648 }
5649
DB_SHOW_COMMAND_FLAGS(pageq,vm_page_print_pageq_info,DB_CMD_MEMSAFE)5650 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5651 {
5652 int dom;
5653
5654 db_printf("pq_free %d\n", vm_free_count());
5655 for (dom = 0; dom < vm_ndomains; dom++) {
5656 db_printf(
5657 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5658 dom,
5659 vm_dom[dom].vmd_page_count,
5660 vm_dom[dom].vmd_free_count,
5661 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5662 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5663 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5664 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5665 }
5666 }
5667
DB_SHOW_COMMAND(pginfo,vm_page_print_pginfo)5668 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5669 {
5670 vm_page_t m;
5671 boolean_t phys, virt;
5672
5673 if (!have_addr) {
5674 db_printf("show pginfo addr\n");
5675 return;
5676 }
5677
5678 phys = strchr(modif, 'p') != NULL;
5679 virt = strchr(modif, 'v') != NULL;
5680 if (virt)
5681 m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5682 else if (phys)
5683 m = PHYS_TO_VM_PAGE(addr);
5684 else
5685 m = (vm_page_t)addr;
5686 db_printf(
5687 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5688 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5689 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5690 m->a.queue, m->ref_count, m->a.flags, m->oflags,
5691 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5692 }
5693 #endif /* DDB */
5694