1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * The Mach Operating System project at Carnegie-Mellon University. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 * 35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 36 */ 37 38 /*- 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 */ 64 65 /* 66 * Resident memory management module. 67 */ 68 69 #include <sys/cdefs.h> 70 __FBSDID("$FreeBSD$"); 71 72 #include "opt_vm.h" 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/counter.h> 77 #include <sys/domainset.h> 78 #include <sys/kernel.h> 79 #include <sys/limits.h> 80 #include <sys/linker.h> 81 #include <sys/lock.h> 82 #include <sys/malloc.h> 83 #include <sys/mman.h> 84 #include <sys/msgbuf.h> 85 #include <sys/mutex.h> 86 #include <sys/proc.h> 87 #include <sys/rwlock.h> 88 #include <sys/sleepqueue.h> 89 #include <sys/sbuf.h> 90 #include <sys/sched.h> 91 #include <sys/smp.h> 92 #include <sys/sysctl.h> 93 #include <sys/vmmeter.h> 94 #include <sys/vnode.h> 95 96 #include <vm/vm.h> 97 #include <vm/pmap.h> 98 #include <vm/vm_param.h> 99 #include <vm/vm_domainset.h> 100 #include <vm/vm_kern.h> 101 #include <vm/vm_map.h> 102 #include <vm/vm_object.h> 103 #include <vm/vm_page.h> 104 #include <vm/vm_pageout.h> 105 #include <vm/vm_phys.h> 106 #include <vm/vm_pagequeue.h> 107 #include <vm/vm_pager.h> 108 #include <vm/vm_radix.h> 109 #include <vm/vm_reserv.h> 110 #include <vm/vm_extern.h> 111 #include <vm/uma.h> 112 #include <vm/uma_int.h> 113 114 #include <machine/md_var.h> 115 116 struct vm_domain vm_dom[MAXMEMDOM]; 117 118 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 119 120 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 121 122 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 123 /* The following fields are protected by the domainset lock. */ 124 domainset_t __exclusive_cache_line vm_min_domains; 125 domainset_t __exclusive_cache_line vm_severe_domains; 126 static int vm_min_waiters; 127 static int vm_severe_waiters; 128 static int vm_pageproc_waiters; 129 130 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 131 "VM page statistics"); 132 133 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 134 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 135 CTLFLAG_RD, &pqstate_commit_retries, 136 "Number of failed per-page atomic queue state updates"); 137 138 static COUNTER_U64_DEFINE_EARLY(queue_ops); 139 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 140 CTLFLAG_RD, &queue_ops, 141 "Number of batched queue operations"); 142 143 static COUNTER_U64_DEFINE_EARLY(queue_nops); 144 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 145 CTLFLAG_RD, &queue_nops, 146 "Number of batched queue operations with no effects"); 147 148 /* 149 * bogus page -- for I/O to/from partially complete buffers, 150 * or for paging into sparsely invalid regions. 151 */ 152 vm_page_t bogus_page; 153 154 vm_page_t vm_page_array; 155 long vm_page_array_size; 156 long first_page; 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 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(NULL, 0, VM_ALLOC_NOOBJ | 197 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 198 } 199 200 /* 201 * The cache page zone is initialized later since we need to be able to allocate 202 * pages before UMA is fully initialized. 203 */ 204 static void 205 vm_page_init_cache_zones(void *dummy __unused) 206 { 207 struct vm_domain *vmd; 208 struct vm_pgcache *pgcache; 209 int cache, domain, maxcache, pool; 210 211 maxcache = 0; 212 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache); 213 maxcache *= mp_ncpus; 214 for (domain = 0; domain < vm_ndomains; domain++) { 215 vmd = VM_DOMAIN(domain); 216 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 217 pgcache = &vmd->vmd_pgcache[pool]; 218 pgcache->domain = domain; 219 pgcache->pool = pool; 220 pgcache->zone = uma_zcache_create("vm pgcache", 221 PAGE_SIZE, NULL, NULL, NULL, NULL, 222 vm_page_zone_import, vm_page_zone_release, pgcache, 223 UMA_ZONE_VM); 224 225 /* 226 * Limit each pool's zone to 0.1% of the pages in the 227 * domain. 228 */ 229 cache = maxcache != 0 ? maxcache : 230 vmd->vmd_page_count / 1000; 231 uma_zone_set_maxcache(pgcache->zone, cache); 232 } 233 } 234 } 235 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 236 237 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 238 #if PAGE_SIZE == 32768 239 #ifdef CTASSERT 240 CTASSERT(sizeof(u_long) >= 8); 241 #endif 242 #endif 243 244 /* 245 * vm_set_page_size: 246 * 247 * Sets the page size, perhaps based upon the memory 248 * size. Must be called before any use of page-size 249 * dependent functions. 250 */ 251 void 252 vm_set_page_size(void) 253 { 254 if (vm_cnt.v_page_size == 0) 255 vm_cnt.v_page_size = PAGE_SIZE; 256 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 257 panic("vm_set_page_size: page size not a power of two"); 258 } 259 260 /* 261 * vm_page_blacklist_next: 262 * 263 * Find the next entry in the provided string of blacklist 264 * addresses. Entries are separated by space, comma, or newline. 265 * If an invalid integer is encountered then the rest of the 266 * string is skipped. Updates the list pointer to the next 267 * character, or NULL if the string is exhausted or invalid. 268 */ 269 static vm_paddr_t 270 vm_page_blacklist_next(char **list, char *end) 271 { 272 vm_paddr_t bad; 273 char *cp, *pos; 274 275 if (list == NULL || *list == NULL) 276 return (0); 277 if (**list =='\0') { 278 *list = NULL; 279 return (0); 280 } 281 282 /* 283 * If there's no end pointer then the buffer is coming from 284 * the kenv and we know it's null-terminated. 285 */ 286 if (end == NULL) 287 end = *list + strlen(*list); 288 289 /* Ensure that strtoq() won't walk off the end */ 290 if (*end != '\0') { 291 if (*end == '\n' || *end == ' ' || *end == ',') 292 *end = '\0'; 293 else { 294 printf("Blacklist not terminated, skipping\n"); 295 *list = NULL; 296 return (0); 297 } 298 } 299 300 for (pos = *list; *pos != '\0'; pos = cp) { 301 bad = strtoq(pos, &cp, 0); 302 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 303 if (bad == 0) { 304 if (++cp < end) 305 continue; 306 else 307 break; 308 } 309 } else 310 break; 311 if (*cp == '\0' || ++cp >= end) 312 *list = NULL; 313 else 314 *list = cp; 315 return (trunc_page(bad)); 316 } 317 printf("Garbage in RAM blacklist, skipping\n"); 318 *list = NULL; 319 return (0); 320 } 321 322 bool 323 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 324 { 325 struct vm_domain *vmd; 326 vm_page_t m; 327 int ret; 328 329 m = vm_phys_paddr_to_vm_page(pa); 330 if (m == NULL) 331 return (true); /* page does not exist, no failure */ 332 333 vmd = vm_pagequeue_domain(m); 334 vm_domain_free_lock(vmd); 335 ret = vm_phys_unfree_page(m); 336 vm_domain_free_unlock(vmd); 337 if (ret != 0) { 338 vm_domain_freecnt_inc(vmd, -1); 339 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 340 if (verbose) 341 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 342 } 343 return (ret); 344 } 345 346 /* 347 * vm_page_blacklist_check: 348 * 349 * Iterate through the provided string of blacklist addresses, pulling 350 * each entry out of the physical allocator free list and putting it 351 * onto a list for reporting via the vm.page_blacklist sysctl. 352 */ 353 static void 354 vm_page_blacklist_check(char *list, char *end) 355 { 356 vm_paddr_t pa; 357 char *next; 358 359 next = list; 360 while (next != NULL) { 361 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 362 continue; 363 vm_page_blacklist_add(pa, bootverbose); 364 } 365 } 366 367 /* 368 * vm_page_blacklist_load: 369 * 370 * Search for a special module named "ram_blacklist". It'll be a 371 * plain text file provided by the user via the loader directive 372 * of the same name. 373 */ 374 static void 375 vm_page_blacklist_load(char **list, char **end) 376 { 377 void *mod; 378 u_char *ptr; 379 u_int len; 380 381 mod = NULL; 382 ptr = NULL; 383 384 mod = preload_search_by_type("ram_blacklist"); 385 if (mod != NULL) { 386 ptr = preload_fetch_addr(mod); 387 len = preload_fetch_size(mod); 388 } 389 *list = ptr; 390 if (ptr != NULL) 391 *end = ptr + len; 392 else 393 *end = NULL; 394 return; 395 } 396 397 static int 398 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 399 { 400 vm_page_t m; 401 struct sbuf sbuf; 402 int error, first; 403 404 first = 1; 405 error = sysctl_wire_old_buffer(req, 0); 406 if (error != 0) 407 return (error); 408 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 409 TAILQ_FOREACH(m, &blacklist_head, listq) { 410 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 411 (uintmax_t)m->phys_addr); 412 first = 0; 413 } 414 error = sbuf_finish(&sbuf); 415 sbuf_delete(&sbuf); 416 return (error); 417 } 418 419 /* 420 * Initialize a dummy page for use in scans of the specified paging queue. 421 * In principle, this function only needs to set the flag PG_MARKER. 422 * Nonetheless, it write busies the page as a safety precaution. 423 */ 424 static void 425 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 426 { 427 428 bzero(marker, sizeof(*marker)); 429 marker->flags = PG_MARKER; 430 marker->a.flags = aflags; 431 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 432 marker->a.queue = queue; 433 } 434 435 static void 436 vm_page_domain_init(int domain) 437 { 438 struct vm_domain *vmd; 439 struct vm_pagequeue *pq; 440 int i; 441 442 vmd = VM_DOMAIN(domain); 443 bzero(vmd, sizeof(*vmd)); 444 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 445 "vm inactive pagequeue"; 446 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 447 "vm active pagequeue"; 448 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 449 "vm laundry pagequeue"; 450 *__DECONST(const char **, 451 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 452 "vm unswappable pagequeue"; 453 vmd->vmd_domain = domain; 454 vmd->vmd_page_count = 0; 455 vmd->vmd_free_count = 0; 456 vmd->vmd_segs = 0; 457 vmd->vmd_oom = FALSE; 458 for (i = 0; i < PQ_COUNT; i++) { 459 pq = &vmd->vmd_pagequeues[i]; 460 TAILQ_INIT(&pq->pq_pl); 461 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 462 MTX_DEF | MTX_DUPOK); 463 pq->pq_pdpages = 0; 464 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 465 } 466 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 467 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 468 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 469 470 /* 471 * inacthead is used to provide FIFO ordering for LRU-bypassing 472 * insertions. 473 */ 474 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 475 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 476 &vmd->vmd_inacthead, plinks.q); 477 478 /* 479 * The clock pages are used to implement active queue scanning without 480 * requeues. Scans start at clock[0], which is advanced after the scan 481 * ends. When the two clock hands meet, they are reset and scanning 482 * resumes from the head of the queue. 483 */ 484 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 485 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 486 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 487 &vmd->vmd_clock[0], plinks.q); 488 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 489 &vmd->vmd_clock[1], plinks.q); 490 } 491 492 /* 493 * Initialize a physical page in preparation for adding it to the free 494 * lists. 495 */ 496 static void 497 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 498 { 499 500 m->object = NULL; 501 m->ref_count = 0; 502 m->busy_lock = VPB_FREED; 503 m->flags = m->a.flags = 0; 504 m->phys_addr = pa; 505 m->a.queue = PQ_NONE; 506 m->psind = 0; 507 m->segind = segind; 508 m->order = VM_NFREEORDER; 509 m->pool = VM_FREEPOOL_DEFAULT; 510 m->valid = m->dirty = 0; 511 pmap_page_init(m); 512 } 513 514 #ifndef PMAP_HAS_PAGE_ARRAY 515 static vm_paddr_t 516 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 517 { 518 vm_paddr_t new_end; 519 520 /* 521 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 522 * However, because this page is allocated from KVM, out-of-bounds 523 * accesses using the direct map will not be trapped. 524 */ 525 *vaddr += PAGE_SIZE; 526 527 /* 528 * Allocate physical memory for the page structures, and map it. 529 */ 530 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 531 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 532 VM_PROT_READ | VM_PROT_WRITE); 533 vm_page_array_size = page_range; 534 535 return (new_end); 536 } 537 #endif 538 539 /* 540 * vm_page_startup: 541 * 542 * Initializes the resident memory module. Allocates physical memory for 543 * bootstrapping UMA and some data structures that are used to manage 544 * physical pages. Initializes these structures, and populates the free 545 * page queues. 546 */ 547 vm_offset_t 548 vm_page_startup(vm_offset_t vaddr) 549 { 550 struct vm_phys_seg *seg; 551 vm_page_t m; 552 char *list, *listend; 553 vm_paddr_t end, high_avail, low_avail, new_end, size; 554 vm_paddr_t page_range __unused; 555 vm_paddr_t last_pa, pa; 556 u_long pagecount; 557 int biggestone, i, segind; 558 #ifdef WITNESS 559 vm_offset_t mapped; 560 int witness_size; 561 #endif 562 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 563 long ii; 564 #endif 565 566 vaddr = round_page(vaddr); 567 568 vm_phys_early_startup(); 569 biggestone = vm_phys_avail_largest(); 570 end = phys_avail[biggestone+1]; 571 572 /* 573 * Initialize the page and queue locks. 574 */ 575 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 576 for (i = 0; i < PA_LOCK_COUNT; i++) 577 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 578 for (i = 0; i < vm_ndomains; i++) 579 vm_page_domain_init(i); 580 581 new_end = end; 582 #ifdef WITNESS 583 witness_size = round_page(witness_startup_count()); 584 new_end -= witness_size; 585 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 586 VM_PROT_READ | VM_PROT_WRITE); 587 bzero((void *)mapped, witness_size); 588 witness_startup((void *)mapped); 589 #endif 590 591 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 592 defined(__i386__) || defined(__mips__) || defined(__riscv) || \ 593 defined(__powerpc64__) 594 /* 595 * Allocate a bitmap to indicate that a random physical page 596 * needs to be included in a minidump. 597 * 598 * The amd64 port needs this to indicate which direct map pages 599 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 600 * 601 * However, i386 still needs this workspace internally within the 602 * minidump code. In theory, they are not needed on i386, but are 603 * included should the sf_buf code decide to use them. 604 */ 605 last_pa = 0; 606 for (i = 0; dump_avail[i + 1] != 0; i += 2) 607 if (dump_avail[i + 1] > last_pa) 608 last_pa = dump_avail[i + 1]; 609 page_range = last_pa / PAGE_SIZE; 610 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 611 new_end -= vm_page_dump_size; 612 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 613 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 614 bzero((void *)vm_page_dump, vm_page_dump_size); 615 #else 616 (void)last_pa; 617 #endif 618 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 619 defined(__riscv) || defined(__powerpc64__) 620 /* 621 * Include the UMA bootstrap pages, witness pages and vm_page_dump 622 * in a crash dump. When pmap_map() uses the direct map, they are 623 * not automatically included. 624 */ 625 for (pa = new_end; pa < end; pa += PAGE_SIZE) 626 dump_add_page(pa); 627 #endif 628 phys_avail[biggestone + 1] = new_end; 629 #ifdef __amd64__ 630 /* 631 * Request that the physical pages underlying the message buffer be 632 * included in a crash dump. Since the message buffer is accessed 633 * through the direct map, they are not automatically included. 634 */ 635 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 636 last_pa = pa + round_page(msgbufsize); 637 while (pa < last_pa) { 638 dump_add_page(pa); 639 pa += PAGE_SIZE; 640 } 641 #endif 642 /* 643 * Compute the number of pages of memory that will be available for 644 * use, taking into account the overhead of a page structure per page. 645 * In other words, solve 646 * "available physical memory" - round_page(page_range * 647 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 648 * for page_range. 649 */ 650 low_avail = phys_avail[0]; 651 high_avail = phys_avail[1]; 652 for (i = 0; i < vm_phys_nsegs; i++) { 653 if (vm_phys_segs[i].start < low_avail) 654 low_avail = vm_phys_segs[i].start; 655 if (vm_phys_segs[i].end > high_avail) 656 high_avail = vm_phys_segs[i].end; 657 } 658 /* Skip the first chunk. It is already accounted for. */ 659 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 660 if (phys_avail[i] < low_avail) 661 low_avail = phys_avail[i]; 662 if (phys_avail[i + 1] > high_avail) 663 high_avail = phys_avail[i + 1]; 664 } 665 first_page = low_avail / PAGE_SIZE; 666 #ifdef VM_PHYSSEG_SPARSE 667 size = 0; 668 for (i = 0; i < vm_phys_nsegs; i++) 669 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 670 for (i = 0; phys_avail[i + 1] != 0; i += 2) 671 size += phys_avail[i + 1] - phys_avail[i]; 672 #elif defined(VM_PHYSSEG_DENSE) 673 size = high_avail - low_avail; 674 #else 675 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 676 #endif 677 678 #ifdef PMAP_HAS_PAGE_ARRAY 679 pmap_page_array_startup(size / PAGE_SIZE); 680 biggestone = vm_phys_avail_largest(); 681 end = new_end = phys_avail[biggestone + 1]; 682 #else 683 #ifdef VM_PHYSSEG_DENSE 684 /* 685 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 686 * the overhead of a page structure per page only if vm_page_array is 687 * allocated from the last physical memory chunk. Otherwise, we must 688 * allocate page structures representing the physical memory 689 * underlying vm_page_array, even though they will not be used. 690 */ 691 if (new_end != high_avail) 692 page_range = size / PAGE_SIZE; 693 else 694 #endif 695 { 696 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 697 698 /* 699 * If the partial bytes remaining are large enough for 700 * a page (PAGE_SIZE) without a corresponding 701 * 'struct vm_page', then new_end will contain an 702 * extra page after subtracting the length of the VM 703 * page array. Compensate by subtracting an extra 704 * page from new_end. 705 */ 706 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 707 if (new_end == high_avail) 708 high_avail -= PAGE_SIZE; 709 new_end -= PAGE_SIZE; 710 } 711 } 712 end = new_end; 713 new_end = vm_page_array_alloc(&vaddr, end, page_range); 714 #endif 715 716 #if VM_NRESERVLEVEL > 0 717 /* 718 * Allocate physical memory for the reservation management system's 719 * data structures, and map it. 720 */ 721 new_end = vm_reserv_startup(&vaddr, new_end); 722 #endif 723 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 724 defined(__riscv) || defined(__powerpc64__) 725 /* 726 * Include vm_page_array and vm_reserv_array in a crash dump. 727 */ 728 for (pa = new_end; pa < end; pa += PAGE_SIZE) 729 dump_add_page(pa); 730 #endif 731 phys_avail[biggestone + 1] = new_end; 732 733 /* 734 * Add physical memory segments corresponding to the available 735 * physical pages. 736 */ 737 for (i = 0; phys_avail[i + 1] != 0; i += 2) 738 if (vm_phys_avail_size(i) != 0) 739 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 740 741 /* 742 * Initialize the physical memory allocator. 743 */ 744 vm_phys_init(); 745 746 /* 747 * Initialize the page structures and add every available page to the 748 * physical memory allocator's free lists. 749 */ 750 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 751 for (ii = 0; ii < vm_page_array_size; ii++) { 752 m = &vm_page_array[ii]; 753 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 754 m->flags = PG_FICTITIOUS; 755 } 756 #endif 757 vm_cnt.v_page_count = 0; 758 for (segind = 0; segind < vm_phys_nsegs; segind++) { 759 seg = &vm_phys_segs[segind]; 760 for (m = seg->first_page, pa = seg->start; pa < seg->end; 761 m++, pa += PAGE_SIZE) 762 vm_page_init_page(m, pa, segind); 763 764 /* 765 * Add the segment to the free lists only if it is covered by 766 * one of the ranges in phys_avail. Because we've added the 767 * ranges to the vm_phys_segs array, we can assume that each 768 * segment is either entirely contained in one of the ranges, 769 * or doesn't overlap any of them. 770 */ 771 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 772 struct vm_domain *vmd; 773 774 if (seg->start < phys_avail[i] || 775 seg->end > phys_avail[i + 1]) 776 continue; 777 778 m = seg->first_page; 779 pagecount = (u_long)atop(seg->end - seg->start); 780 781 vmd = VM_DOMAIN(seg->domain); 782 vm_domain_free_lock(vmd); 783 vm_phys_enqueue_contig(m, pagecount); 784 vm_domain_free_unlock(vmd); 785 vm_domain_freecnt_inc(vmd, pagecount); 786 vm_cnt.v_page_count += (u_int)pagecount; 787 788 vmd = VM_DOMAIN(seg->domain); 789 vmd->vmd_page_count += (u_int)pagecount; 790 vmd->vmd_segs |= 1UL << m->segind; 791 break; 792 } 793 } 794 795 /* 796 * Remove blacklisted pages from the physical memory allocator. 797 */ 798 TAILQ_INIT(&blacklist_head); 799 vm_page_blacklist_load(&list, &listend); 800 vm_page_blacklist_check(list, listend); 801 802 list = kern_getenv("vm.blacklist"); 803 vm_page_blacklist_check(list, NULL); 804 805 freeenv(list); 806 #if VM_NRESERVLEVEL > 0 807 /* 808 * Initialize the reservation management system. 809 */ 810 vm_reserv_init(); 811 #endif 812 813 return (vaddr); 814 } 815 816 void 817 vm_page_reference(vm_page_t m) 818 { 819 820 vm_page_aflag_set(m, PGA_REFERENCED); 821 } 822 823 /* 824 * vm_page_trybusy 825 * 826 * Helper routine for grab functions to trylock busy. 827 * 828 * Returns true on success and false on failure. 829 */ 830 static bool 831 vm_page_trybusy(vm_page_t m, int allocflags) 832 { 833 834 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 835 return (vm_page_trysbusy(m)); 836 else 837 return (vm_page_tryxbusy(m)); 838 } 839 840 /* 841 * vm_page_tryacquire 842 * 843 * Helper routine for grab functions to trylock busy and wire. 844 * 845 * Returns true on success and false on failure. 846 */ 847 static inline bool 848 vm_page_tryacquire(vm_page_t m, int allocflags) 849 { 850 bool locked; 851 852 locked = vm_page_trybusy(m, allocflags); 853 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 854 vm_page_wire(m); 855 return (locked); 856 } 857 858 /* 859 * vm_page_busy_acquire: 860 * 861 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 862 * and drop the object lock if necessary. 863 */ 864 bool 865 vm_page_busy_acquire(vm_page_t m, int allocflags) 866 { 867 vm_object_t obj; 868 bool locked; 869 870 /* 871 * The page-specific object must be cached because page 872 * identity can change during the sleep, causing the 873 * re-lock of a different object. 874 * It is assumed that a reference to the object is already 875 * held by the callers. 876 */ 877 obj = m->object; 878 for (;;) { 879 if (vm_page_tryacquire(m, allocflags)) 880 return (true); 881 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 882 return (false); 883 if (obj != NULL) 884 locked = VM_OBJECT_WOWNED(obj); 885 else 886 locked = false; 887 MPASS(locked || vm_page_wired(m)); 888 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 889 locked) && locked) 890 VM_OBJECT_WLOCK(obj); 891 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 892 return (false); 893 KASSERT(m->object == obj || m->object == NULL, 894 ("vm_page_busy_acquire: page %p does not belong to %p", 895 m, obj)); 896 } 897 } 898 899 /* 900 * vm_page_busy_downgrade: 901 * 902 * Downgrade an exclusive busy page into a single shared busy page. 903 */ 904 void 905 vm_page_busy_downgrade(vm_page_t m) 906 { 907 u_int x; 908 909 vm_page_assert_xbusied(m); 910 911 x = m->busy_lock; 912 for (;;) { 913 if (atomic_fcmpset_rel_int(&m->busy_lock, 914 &x, VPB_SHARERS_WORD(1))) 915 break; 916 } 917 if ((x & VPB_BIT_WAITERS) != 0) 918 wakeup(m); 919 } 920 921 /* 922 * 923 * vm_page_busy_tryupgrade: 924 * 925 * Attempt to upgrade a single shared busy into an exclusive busy. 926 */ 927 int 928 vm_page_busy_tryupgrade(vm_page_t m) 929 { 930 u_int ce, x; 931 932 vm_page_assert_sbusied(m); 933 934 x = m->busy_lock; 935 ce = VPB_CURTHREAD_EXCLUSIVE; 936 for (;;) { 937 if (VPB_SHARERS(x) > 1) 938 return (0); 939 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 940 ("vm_page_busy_tryupgrade: invalid lock state")); 941 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 942 ce | (x & VPB_BIT_WAITERS))) 943 continue; 944 return (1); 945 } 946 } 947 948 /* 949 * vm_page_sbusied: 950 * 951 * Return a positive value if the page is shared busied, 0 otherwise. 952 */ 953 int 954 vm_page_sbusied(vm_page_t m) 955 { 956 u_int x; 957 958 x = m->busy_lock; 959 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 960 } 961 962 /* 963 * vm_page_sunbusy: 964 * 965 * Shared unbusy a page. 966 */ 967 void 968 vm_page_sunbusy(vm_page_t m) 969 { 970 u_int x; 971 972 vm_page_assert_sbusied(m); 973 974 x = m->busy_lock; 975 for (;;) { 976 KASSERT(x != VPB_FREED, 977 ("vm_page_sunbusy: Unlocking freed page.")); 978 if (VPB_SHARERS(x) > 1) { 979 if (atomic_fcmpset_int(&m->busy_lock, &x, 980 x - VPB_ONE_SHARER)) 981 break; 982 continue; 983 } 984 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 985 ("vm_page_sunbusy: invalid lock state")); 986 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 987 continue; 988 if ((x & VPB_BIT_WAITERS) == 0) 989 break; 990 wakeup(m); 991 break; 992 } 993 } 994 995 /* 996 * vm_page_busy_sleep: 997 * 998 * Sleep if the page is busy, using the page pointer as wchan. 999 * This is used to implement the hard-path of busying mechanism. 1000 * 1001 * If nonshared is true, sleep only if the page is xbusy. 1002 * 1003 * The object lock must be held on entry and will be released on exit. 1004 */ 1005 void 1006 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 1007 { 1008 vm_object_t obj; 1009 1010 obj = m->object; 1011 VM_OBJECT_ASSERT_LOCKED(obj); 1012 vm_page_lock_assert(m, MA_NOTOWNED); 1013 1014 if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 1015 nonshared ? VM_ALLOC_SBUSY : 0 , true)) 1016 VM_OBJECT_DROP(obj); 1017 } 1018 1019 /* 1020 * vm_page_busy_sleep_unlocked: 1021 * 1022 * Sleep if the page is busy, using the page pointer as wchan. 1023 * This is used to implement the hard-path of busying mechanism. 1024 * 1025 * If nonshared is true, sleep only if the page is xbusy. 1026 * 1027 * The object lock must not be held on entry. The operation will 1028 * return if the page changes identity. 1029 */ 1030 void 1031 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1032 const char *wmesg, bool nonshared) 1033 { 1034 1035 VM_OBJECT_ASSERT_UNLOCKED(obj); 1036 vm_page_lock_assert(m, MA_NOTOWNED); 1037 1038 _vm_page_busy_sleep(obj, m, pindex, wmesg, 1039 nonshared ? VM_ALLOC_SBUSY : 0, false); 1040 } 1041 1042 /* 1043 * _vm_page_busy_sleep: 1044 * 1045 * Internal busy sleep function. Verifies the page identity and 1046 * lockstate against parameters. Returns true if it sleeps and 1047 * false otherwise. 1048 * 1049 * If locked is true the lock will be dropped for any true returns 1050 * and held for any false returns. 1051 */ 1052 static bool 1053 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1054 const char *wmesg, int allocflags, bool locked) 1055 { 1056 bool xsleep; 1057 u_int x; 1058 1059 /* 1060 * If the object is busy we must wait for that to drain to zero 1061 * before trying the page again. 1062 */ 1063 if (obj != NULL && vm_object_busied(obj)) { 1064 if (locked) 1065 VM_OBJECT_DROP(obj); 1066 vm_object_busy_wait(obj, wmesg); 1067 return (true); 1068 } 1069 1070 if (!vm_page_busied(m)) 1071 return (false); 1072 1073 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1074 sleepq_lock(m); 1075 x = atomic_load_int(&m->busy_lock); 1076 do { 1077 /* 1078 * If the page changes objects or becomes unlocked we can 1079 * simply return. 1080 */ 1081 if (x == VPB_UNBUSIED || 1082 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1083 m->object != obj || m->pindex != pindex) { 1084 sleepq_release(m); 1085 return (false); 1086 } 1087 if ((x & VPB_BIT_WAITERS) != 0) 1088 break; 1089 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1090 if (locked) 1091 VM_OBJECT_DROP(obj); 1092 DROP_GIANT(); 1093 sleepq_add(m, NULL, wmesg, 0, 0); 1094 sleepq_wait(m, PVM); 1095 PICKUP_GIANT(); 1096 return (true); 1097 } 1098 1099 /* 1100 * vm_page_trysbusy: 1101 * 1102 * Try to shared busy a page. 1103 * If the operation succeeds 1 is returned otherwise 0. 1104 * The operation never sleeps. 1105 */ 1106 int 1107 vm_page_trysbusy(vm_page_t m) 1108 { 1109 vm_object_t obj; 1110 u_int x; 1111 1112 obj = m->object; 1113 x = m->busy_lock; 1114 for (;;) { 1115 if ((x & VPB_BIT_SHARED) == 0) 1116 return (0); 1117 /* 1118 * Reduce the window for transient busies that will trigger 1119 * false negatives in vm_page_ps_test(). 1120 */ 1121 if (obj != NULL && vm_object_busied(obj)) 1122 return (0); 1123 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1124 x + VPB_ONE_SHARER)) 1125 break; 1126 } 1127 1128 /* Refetch the object now that we're guaranteed that it is stable. */ 1129 obj = m->object; 1130 if (obj != NULL && vm_object_busied(obj)) { 1131 vm_page_sunbusy(m); 1132 return (0); 1133 } 1134 return (1); 1135 } 1136 1137 /* 1138 * vm_page_tryxbusy: 1139 * 1140 * Try to exclusive busy a page. 1141 * If the operation succeeds 1 is returned otherwise 0. 1142 * The operation never sleeps. 1143 */ 1144 int 1145 vm_page_tryxbusy(vm_page_t m) 1146 { 1147 vm_object_t obj; 1148 1149 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, 1150 VPB_CURTHREAD_EXCLUSIVE) == 0) 1151 return (0); 1152 1153 obj = m->object; 1154 if (obj != NULL && vm_object_busied(obj)) { 1155 vm_page_xunbusy(m); 1156 return (0); 1157 } 1158 return (1); 1159 } 1160 1161 static void 1162 vm_page_xunbusy_hard_tail(vm_page_t m) 1163 { 1164 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1165 /* Wake the waiter. */ 1166 wakeup(m); 1167 } 1168 1169 /* 1170 * vm_page_xunbusy_hard: 1171 * 1172 * Called when unbusy has failed because there is a waiter. 1173 */ 1174 void 1175 vm_page_xunbusy_hard(vm_page_t m) 1176 { 1177 vm_page_assert_xbusied(m); 1178 vm_page_xunbusy_hard_tail(m); 1179 } 1180 1181 void 1182 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1183 { 1184 vm_page_assert_xbusied_unchecked(m); 1185 vm_page_xunbusy_hard_tail(m); 1186 } 1187 1188 static void 1189 vm_page_busy_free(vm_page_t m) 1190 { 1191 u_int x; 1192 1193 atomic_thread_fence_rel(); 1194 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1195 if ((x & VPB_BIT_WAITERS) != 0) 1196 wakeup(m); 1197 } 1198 1199 /* 1200 * vm_page_unhold_pages: 1201 * 1202 * Unhold each of the pages that is referenced by the given array. 1203 */ 1204 void 1205 vm_page_unhold_pages(vm_page_t *ma, int count) 1206 { 1207 1208 for (; count != 0; count--) { 1209 vm_page_unwire(*ma, PQ_ACTIVE); 1210 ma++; 1211 } 1212 } 1213 1214 vm_page_t 1215 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1216 { 1217 vm_page_t m; 1218 1219 #ifdef VM_PHYSSEG_SPARSE 1220 m = vm_phys_paddr_to_vm_page(pa); 1221 if (m == NULL) 1222 m = vm_phys_fictitious_to_vm_page(pa); 1223 return (m); 1224 #elif defined(VM_PHYSSEG_DENSE) 1225 long pi; 1226 1227 pi = atop(pa); 1228 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1229 m = &vm_page_array[pi - first_page]; 1230 return (m); 1231 } 1232 return (vm_phys_fictitious_to_vm_page(pa)); 1233 #else 1234 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1235 #endif 1236 } 1237 1238 /* 1239 * vm_page_getfake: 1240 * 1241 * Create a fictitious page with the specified physical address and 1242 * memory attribute. The memory attribute is the only the machine- 1243 * dependent aspect of a fictitious page that must be initialized. 1244 */ 1245 vm_page_t 1246 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1247 { 1248 vm_page_t m; 1249 1250 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1251 vm_page_initfake(m, paddr, memattr); 1252 return (m); 1253 } 1254 1255 void 1256 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1257 { 1258 1259 if ((m->flags & PG_FICTITIOUS) != 0) { 1260 /* 1261 * The page's memattr might have changed since the 1262 * previous initialization. Update the pmap to the 1263 * new memattr. 1264 */ 1265 goto memattr; 1266 } 1267 m->phys_addr = paddr; 1268 m->a.queue = PQ_NONE; 1269 /* Fictitious pages don't use "segind". */ 1270 m->flags = PG_FICTITIOUS; 1271 /* Fictitious pages don't use "order" or "pool". */ 1272 m->oflags = VPO_UNMANAGED; 1273 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1274 /* Fictitious pages are unevictable. */ 1275 m->ref_count = 1; 1276 pmap_page_init(m); 1277 memattr: 1278 pmap_page_set_memattr(m, memattr); 1279 } 1280 1281 /* 1282 * vm_page_putfake: 1283 * 1284 * Release a fictitious page. 1285 */ 1286 void 1287 vm_page_putfake(vm_page_t m) 1288 { 1289 1290 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1291 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1292 ("vm_page_putfake: bad page %p", m)); 1293 vm_page_assert_xbusied(m); 1294 vm_page_busy_free(m); 1295 uma_zfree(fakepg_zone, m); 1296 } 1297 1298 /* 1299 * vm_page_updatefake: 1300 * 1301 * Update the given fictitious page to the specified physical address and 1302 * memory attribute. 1303 */ 1304 void 1305 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1306 { 1307 1308 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1309 ("vm_page_updatefake: bad page %p", m)); 1310 m->phys_addr = paddr; 1311 pmap_page_set_memattr(m, memattr); 1312 } 1313 1314 /* 1315 * vm_page_free: 1316 * 1317 * Free a page. 1318 */ 1319 void 1320 vm_page_free(vm_page_t m) 1321 { 1322 1323 m->flags &= ~PG_ZERO; 1324 vm_page_free_toq(m); 1325 } 1326 1327 /* 1328 * vm_page_free_zero: 1329 * 1330 * Free a page to the zerod-pages queue 1331 */ 1332 void 1333 vm_page_free_zero(vm_page_t m) 1334 { 1335 1336 m->flags |= PG_ZERO; 1337 vm_page_free_toq(m); 1338 } 1339 1340 /* 1341 * Unbusy and handle the page queueing for a page from a getpages request that 1342 * was optionally read ahead or behind. 1343 */ 1344 void 1345 vm_page_readahead_finish(vm_page_t m) 1346 { 1347 1348 /* We shouldn't put invalid pages on queues. */ 1349 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1350 1351 /* 1352 * Since the page is not the actually needed one, whether it should 1353 * be activated or deactivated is not obvious. Empirical results 1354 * have shown that deactivating the page is usually the best choice, 1355 * unless the page is wanted by another thread. 1356 */ 1357 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1358 vm_page_activate(m); 1359 else 1360 vm_page_deactivate(m); 1361 vm_page_xunbusy_unchecked(m); 1362 } 1363 1364 /* 1365 * Destroy the identity of an invalid page and free it if possible. 1366 * This is intended to be used when reading a page from backing store fails. 1367 */ 1368 void 1369 vm_page_free_invalid(vm_page_t m) 1370 { 1371 1372 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1373 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1374 vm_page_assert_xbusied(m); 1375 KASSERT(m->object != NULL, ("page %p has no object", m)); 1376 VM_OBJECT_ASSERT_WLOCKED(m->object); 1377 1378 /* 1379 * If someone has wired this page while the object lock 1380 * was not held, then the thread that unwires is responsible 1381 * for freeing the page. Otherwise just free the page now. 1382 * The wire count of this unmapped page cannot change while 1383 * we have the page xbusy and the page's object wlocked. 1384 */ 1385 if (vm_page_remove(m)) 1386 vm_page_free(m); 1387 } 1388 1389 /* 1390 * vm_page_sleep_if_busy: 1391 * 1392 * Sleep and release the object lock if the page is busied. 1393 * Returns TRUE if the thread slept. 1394 * 1395 * The given page must be unlocked and object containing it must 1396 * be locked. 1397 */ 1398 int 1399 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg) 1400 { 1401 vm_object_t obj; 1402 1403 vm_page_lock_assert(m, MA_NOTOWNED); 1404 VM_OBJECT_ASSERT_WLOCKED(m->object); 1405 1406 /* 1407 * The page-specific object must be cached because page 1408 * identity can change during the sleep, causing the 1409 * re-lock of a different object. 1410 * It is assumed that a reference to the object is already 1411 * held by the callers. 1412 */ 1413 obj = m->object; 1414 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) { 1415 VM_OBJECT_WLOCK(obj); 1416 return (TRUE); 1417 } 1418 return (FALSE); 1419 } 1420 1421 /* 1422 * vm_page_sleep_if_xbusy: 1423 * 1424 * Sleep and release the object lock if the page is xbusied. 1425 * Returns TRUE if the thread slept. 1426 * 1427 * The given page must be unlocked and object containing it must 1428 * be locked. 1429 */ 1430 int 1431 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg) 1432 { 1433 vm_object_t obj; 1434 1435 vm_page_lock_assert(m, MA_NOTOWNED); 1436 VM_OBJECT_ASSERT_WLOCKED(m->object); 1437 1438 /* 1439 * The page-specific object must be cached because page 1440 * identity can change during the sleep, causing the 1441 * re-lock of a different object. 1442 * It is assumed that a reference to the object is already 1443 * held by the callers. 1444 */ 1445 obj = m->object; 1446 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY, 1447 true)) { 1448 VM_OBJECT_WLOCK(obj); 1449 return (TRUE); 1450 } 1451 return (FALSE); 1452 } 1453 1454 /* 1455 * vm_page_dirty_KBI: [ internal use only ] 1456 * 1457 * Set all bits in the page's dirty field. 1458 * 1459 * The object containing the specified page must be locked if the 1460 * call is made from the machine-independent layer. 1461 * 1462 * See vm_page_clear_dirty_mask(). 1463 * 1464 * This function should only be called by vm_page_dirty(). 1465 */ 1466 void 1467 vm_page_dirty_KBI(vm_page_t m) 1468 { 1469 1470 /* Refer to this operation by its public name. */ 1471 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1472 m->dirty = VM_PAGE_BITS_ALL; 1473 } 1474 1475 /* 1476 * vm_page_insert: [ internal use only ] 1477 * 1478 * Inserts the given mem entry into the object and object list. 1479 * 1480 * The object must be locked. 1481 */ 1482 int 1483 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1484 { 1485 vm_page_t mpred; 1486 1487 VM_OBJECT_ASSERT_WLOCKED(object); 1488 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1489 return (vm_page_insert_after(m, object, pindex, mpred)); 1490 } 1491 1492 /* 1493 * vm_page_insert_after: 1494 * 1495 * Inserts the page "m" into the specified object at offset "pindex". 1496 * 1497 * The page "mpred" must immediately precede the offset "pindex" within 1498 * the specified object. 1499 * 1500 * The object must be locked. 1501 */ 1502 static int 1503 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1504 vm_page_t mpred) 1505 { 1506 vm_page_t msucc; 1507 1508 VM_OBJECT_ASSERT_WLOCKED(object); 1509 KASSERT(m->object == NULL, 1510 ("vm_page_insert_after: page already inserted")); 1511 if (mpred != NULL) { 1512 KASSERT(mpred->object == object, 1513 ("vm_page_insert_after: object doesn't contain mpred")); 1514 KASSERT(mpred->pindex < pindex, 1515 ("vm_page_insert_after: mpred doesn't precede pindex")); 1516 msucc = TAILQ_NEXT(mpred, listq); 1517 } else 1518 msucc = TAILQ_FIRST(&object->memq); 1519 if (msucc != NULL) 1520 KASSERT(msucc->pindex > pindex, 1521 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1522 1523 /* 1524 * Record the object/offset pair in this page. 1525 */ 1526 m->object = object; 1527 m->pindex = pindex; 1528 m->ref_count |= VPRC_OBJREF; 1529 1530 /* 1531 * Now link into the object's ordered list of backed pages. 1532 */ 1533 if (vm_radix_insert(&object->rtree, m)) { 1534 m->object = NULL; 1535 m->pindex = 0; 1536 m->ref_count &= ~VPRC_OBJREF; 1537 return (1); 1538 } 1539 vm_page_insert_radixdone(m, object, mpred); 1540 return (0); 1541 } 1542 1543 /* 1544 * vm_page_insert_radixdone: 1545 * 1546 * Complete page "m" insertion into the specified object after the 1547 * radix trie hooking. 1548 * 1549 * The page "mpred" must precede the offset "m->pindex" within the 1550 * specified object. 1551 * 1552 * The object must be locked. 1553 */ 1554 static void 1555 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1556 { 1557 1558 VM_OBJECT_ASSERT_WLOCKED(object); 1559 KASSERT(object != NULL && m->object == object, 1560 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1561 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1562 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1563 if (mpred != NULL) { 1564 KASSERT(mpred->object == object, 1565 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1566 KASSERT(mpred->pindex < m->pindex, 1567 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1568 } 1569 1570 if (mpred != NULL) 1571 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1572 else 1573 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1574 1575 /* 1576 * Show that the object has one more resident page. 1577 */ 1578 object->resident_page_count++; 1579 1580 /* 1581 * Hold the vnode until the last page is released. 1582 */ 1583 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1584 vhold(object->handle); 1585 1586 /* 1587 * Since we are inserting a new and possibly dirty page, 1588 * update the object's generation count. 1589 */ 1590 if (pmap_page_is_write_mapped(m)) 1591 vm_object_set_writeable_dirty(object); 1592 } 1593 1594 /* 1595 * Do the work to remove a page from its object. The caller is responsible for 1596 * updating the page's fields to reflect this removal. 1597 */ 1598 static void 1599 vm_page_object_remove(vm_page_t m) 1600 { 1601 vm_object_t object; 1602 vm_page_t mrem; 1603 1604 vm_page_assert_xbusied(m); 1605 object = m->object; 1606 VM_OBJECT_ASSERT_WLOCKED(object); 1607 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1608 ("page %p is missing its object ref", m)); 1609 1610 /* Deferred free of swap space. */ 1611 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1612 vm_pager_page_unswapped(m); 1613 1614 m->object = NULL; 1615 mrem = vm_radix_remove(&object->rtree, m->pindex); 1616 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1617 1618 /* 1619 * Now remove from the object's list of backed pages. 1620 */ 1621 TAILQ_REMOVE(&object->memq, m, listq); 1622 1623 /* 1624 * And show that the object has one fewer resident page. 1625 */ 1626 object->resident_page_count--; 1627 1628 /* 1629 * The vnode may now be recycled. 1630 */ 1631 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1632 vdrop(object->handle); 1633 } 1634 1635 /* 1636 * vm_page_remove: 1637 * 1638 * Removes the specified page from its containing object, but does not 1639 * invalidate any backing storage. Returns true if the object's reference 1640 * was the last reference to the page, and false otherwise. 1641 * 1642 * The object must be locked and the page must be exclusively busied. 1643 * The exclusive busy will be released on return. If this is not the 1644 * final ref and the caller does not hold a wire reference it may not 1645 * continue to access the page. 1646 */ 1647 bool 1648 vm_page_remove(vm_page_t m) 1649 { 1650 bool dropped; 1651 1652 dropped = vm_page_remove_xbusy(m); 1653 vm_page_xunbusy(m); 1654 1655 return (dropped); 1656 } 1657 1658 /* 1659 * vm_page_remove_xbusy 1660 * 1661 * Removes the page but leaves the xbusy held. Returns true if this 1662 * removed the final ref and false otherwise. 1663 */ 1664 bool 1665 vm_page_remove_xbusy(vm_page_t m) 1666 { 1667 1668 vm_page_object_remove(m); 1669 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1670 } 1671 1672 /* 1673 * vm_page_lookup: 1674 * 1675 * Returns the page associated with the object/offset 1676 * pair specified; if none is found, NULL is returned. 1677 * 1678 * The object must be locked. 1679 */ 1680 vm_page_t 1681 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1682 { 1683 1684 VM_OBJECT_ASSERT_LOCKED(object); 1685 return (vm_radix_lookup(&object->rtree, pindex)); 1686 } 1687 1688 /* 1689 * vm_page_relookup: 1690 * 1691 * Returns a page that must already have been busied by 1692 * the caller. Used for bogus page replacement. 1693 */ 1694 vm_page_t 1695 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1696 { 1697 vm_page_t m; 1698 1699 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 1700 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1701 m->object == object && m->pindex == pindex, 1702 ("vm_page_relookup: Invalid page %p", m)); 1703 return (m); 1704 } 1705 1706 /* 1707 * This should only be used by lockless functions for releasing transient 1708 * incorrect acquires. The page may have been freed after we acquired a 1709 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1710 * further to do. 1711 */ 1712 static void 1713 vm_page_busy_release(vm_page_t m) 1714 { 1715 u_int x; 1716 1717 x = atomic_load_int(&m->busy_lock); 1718 for (;;) { 1719 if (x == VPB_FREED) 1720 break; 1721 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1722 if (atomic_fcmpset_int(&m->busy_lock, &x, 1723 x - VPB_ONE_SHARER)) 1724 break; 1725 continue; 1726 } 1727 KASSERT((x & VPB_BIT_SHARED) != 0 || 1728 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1729 ("vm_page_busy_release: %p xbusy not owned.", m)); 1730 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1731 continue; 1732 if ((x & VPB_BIT_WAITERS) != 0) 1733 wakeup(m); 1734 break; 1735 } 1736 } 1737 1738 /* 1739 * vm_page_find_least: 1740 * 1741 * Returns the page associated with the object with least pindex 1742 * greater than or equal to the parameter pindex, or NULL. 1743 * 1744 * The object must be locked. 1745 */ 1746 vm_page_t 1747 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1748 { 1749 vm_page_t m; 1750 1751 VM_OBJECT_ASSERT_LOCKED(object); 1752 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1753 m = vm_radix_lookup_ge(&object->rtree, pindex); 1754 return (m); 1755 } 1756 1757 /* 1758 * Returns the given page's successor (by pindex) within the object if it is 1759 * resident; if none is found, NULL is returned. 1760 * 1761 * The object must be locked. 1762 */ 1763 vm_page_t 1764 vm_page_next(vm_page_t m) 1765 { 1766 vm_page_t next; 1767 1768 VM_OBJECT_ASSERT_LOCKED(m->object); 1769 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1770 MPASS(next->object == m->object); 1771 if (next->pindex != m->pindex + 1) 1772 next = NULL; 1773 } 1774 return (next); 1775 } 1776 1777 /* 1778 * Returns the given page's predecessor (by pindex) within the object if it is 1779 * resident; if none is found, NULL is returned. 1780 * 1781 * The object must be locked. 1782 */ 1783 vm_page_t 1784 vm_page_prev(vm_page_t m) 1785 { 1786 vm_page_t prev; 1787 1788 VM_OBJECT_ASSERT_LOCKED(m->object); 1789 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1790 MPASS(prev->object == m->object); 1791 if (prev->pindex != m->pindex - 1) 1792 prev = NULL; 1793 } 1794 return (prev); 1795 } 1796 1797 /* 1798 * Uses the page mnew as a replacement for an existing page at index 1799 * pindex which must be already present in the object. 1800 * 1801 * Both pages must be exclusively busied on enter. The old page is 1802 * unbusied on exit. 1803 * 1804 * A return value of true means mold is now free. If this is not the 1805 * final ref and the caller does not hold a wire reference it may not 1806 * continue to access the page. 1807 */ 1808 static bool 1809 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1810 vm_page_t mold) 1811 { 1812 vm_page_t mret; 1813 bool dropped; 1814 1815 VM_OBJECT_ASSERT_WLOCKED(object); 1816 vm_page_assert_xbusied(mold); 1817 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1818 ("vm_page_replace: page %p already in object", mnew)); 1819 1820 /* 1821 * This function mostly follows vm_page_insert() and 1822 * vm_page_remove() without the radix, object count and vnode 1823 * dance. Double check such functions for more comments. 1824 */ 1825 1826 mnew->object = object; 1827 mnew->pindex = pindex; 1828 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1829 mret = vm_radix_replace(&object->rtree, mnew); 1830 KASSERT(mret == mold, 1831 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1832 KASSERT((mold->oflags & VPO_UNMANAGED) == 1833 (mnew->oflags & VPO_UNMANAGED), 1834 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1835 1836 /* Keep the resident page list in sorted order. */ 1837 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1838 TAILQ_REMOVE(&object->memq, mold, listq); 1839 mold->object = NULL; 1840 1841 /* 1842 * The object's resident_page_count does not change because we have 1843 * swapped one page for another, but the generation count should 1844 * change if the page is dirty. 1845 */ 1846 if (pmap_page_is_write_mapped(mnew)) 1847 vm_object_set_writeable_dirty(object); 1848 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1849 vm_page_xunbusy(mold); 1850 1851 return (dropped); 1852 } 1853 1854 void 1855 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1856 vm_page_t mold) 1857 { 1858 1859 vm_page_assert_xbusied(mnew); 1860 1861 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1862 vm_page_free(mold); 1863 } 1864 1865 /* 1866 * vm_page_rename: 1867 * 1868 * Move the given memory entry from its 1869 * current object to the specified target object/offset. 1870 * 1871 * Note: swap associated with the page must be invalidated by the move. We 1872 * have to do this for several reasons: (1) we aren't freeing the 1873 * page, (2) we are dirtying the page, (3) the VM system is probably 1874 * moving the page from object A to B, and will then later move 1875 * the backing store from A to B and we can't have a conflict. 1876 * 1877 * Note: we *always* dirty the page. It is necessary both for the 1878 * fact that we moved it, and because we may be invalidating 1879 * swap. 1880 * 1881 * The objects must be locked. 1882 */ 1883 int 1884 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1885 { 1886 vm_page_t mpred; 1887 vm_pindex_t opidx; 1888 1889 VM_OBJECT_ASSERT_WLOCKED(new_object); 1890 1891 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1892 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1893 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1894 ("vm_page_rename: pindex already renamed")); 1895 1896 /* 1897 * Create a custom version of vm_page_insert() which does not depend 1898 * by m_prev and can cheat on the implementation aspects of the 1899 * function. 1900 */ 1901 opidx = m->pindex; 1902 m->pindex = new_pindex; 1903 if (vm_radix_insert(&new_object->rtree, m)) { 1904 m->pindex = opidx; 1905 return (1); 1906 } 1907 1908 /* 1909 * The operation cannot fail anymore. The removal must happen before 1910 * the listq iterator is tainted. 1911 */ 1912 m->pindex = opidx; 1913 vm_page_object_remove(m); 1914 1915 /* Return back to the new pindex to complete vm_page_insert(). */ 1916 m->pindex = new_pindex; 1917 m->object = new_object; 1918 1919 vm_page_insert_radixdone(m, new_object, mpred); 1920 vm_page_dirty(m); 1921 return (0); 1922 } 1923 1924 /* 1925 * vm_page_alloc: 1926 * 1927 * Allocate and return a page that is associated with the specified 1928 * object and offset pair. By default, this page is exclusive busied. 1929 * 1930 * The caller must always specify an allocation class. 1931 * 1932 * allocation classes: 1933 * VM_ALLOC_NORMAL normal process request 1934 * VM_ALLOC_SYSTEM system *really* needs a page 1935 * VM_ALLOC_INTERRUPT interrupt time request 1936 * 1937 * optional allocation flags: 1938 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1939 * intends to allocate 1940 * VM_ALLOC_NOBUSY do not exclusive busy the page 1941 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1942 * VM_ALLOC_NOOBJ page is not associated with an object and 1943 * should not be exclusive busy 1944 * VM_ALLOC_SBUSY shared busy the allocated page 1945 * VM_ALLOC_WIRED wire the allocated page 1946 * VM_ALLOC_ZERO prefer a zeroed page 1947 */ 1948 vm_page_t 1949 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1950 { 1951 1952 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1953 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1954 } 1955 1956 vm_page_t 1957 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1958 int req) 1959 { 1960 1961 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1962 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1963 NULL)); 1964 } 1965 1966 /* 1967 * Allocate a page in the specified object with the given page index. To 1968 * optimize insertion of the page into the object, the caller must also specifiy 1969 * the resident page in the object with largest index smaller than the given 1970 * page index, or NULL if no such page exists. 1971 */ 1972 vm_page_t 1973 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1974 int req, vm_page_t mpred) 1975 { 1976 struct vm_domainset_iter di; 1977 vm_page_t m; 1978 int domain; 1979 1980 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1981 do { 1982 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1983 mpred); 1984 if (m != NULL) 1985 break; 1986 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1987 1988 return (m); 1989 } 1990 1991 /* 1992 * Returns true if the number of free pages exceeds the minimum 1993 * for the request class and false otherwise. 1994 */ 1995 static int 1996 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 1997 { 1998 u_int limit, old, new; 1999 2000 if (req_class == VM_ALLOC_INTERRUPT) 2001 limit = 0; 2002 else if (req_class == VM_ALLOC_SYSTEM) 2003 limit = vmd->vmd_interrupt_free_min; 2004 else 2005 limit = vmd->vmd_free_reserved; 2006 2007 /* 2008 * Attempt to reserve the pages. Fail if we're below the limit. 2009 */ 2010 limit += npages; 2011 old = vmd->vmd_free_count; 2012 do { 2013 if (old < limit) 2014 return (0); 2015 new = old - npages; 2016 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2017 2018 /* Wake the page daemon if we've crossed the threshold. */ 2019 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2020 pagedaemon_wakeup(vmd->vmd_domain); 2021 2022 /* Only update bitsets on transitions. */ 2023 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2024 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2025 vm_domain_set(vmd); 2026 2027 return (1); 2028 } 2029 2030 int 2031 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2032 { 2033 int req_class; 2034 2035 /* 2036 * The page daemon is allowed to dig deeper into the free page list. 2037 */ 2038 req_class = req & VM_ALLOC_CLASS_MASK; 2039 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2040 req_class = VM_ALLOC_SYSTEM; 2041 return (_vm_domain_allocate(vmd, req_class, npages)); 2042 } 2043 2044 vm_page_t 2045 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 2046 int req, vm_page_t mpred) 2047 { 2048 struct vm_domain *vmd; 2049 vm_page_t m; 2050 int flags, pool; 2051 2052 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2053 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2054 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2055 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2056 ("inconsistent object(%p)/req(%x)", object, req)); 2057 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2058 ("Can't sleep and retry object insertion.")); 2059 KASSERT(mpred == NULL || mpred->pindex < pindex, 2060 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2061 (uintmax_t)pindex)); 2062 if (object != NULL) 2063 VM_OBJECT_ASSERT_WLOCKED(object); 2064 2065 flags = 0; 2066 m = NULL; 2067 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; 2068 again: 2069 #if VM_NRESERVLEVEL > 0 2070 /* 2071 * Can we allocate the page from a reservation? 2072 */ 2073 if (vm_object_reserv(object) && 2074 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 2075 NULL) { 2076 goto found; 2077 } 2078 #endif 2079 vmd = VM_DOMAIN(domain); 2080 if (vmd->vmd_pgcache[pool].zone != NULL) { 2081 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM); 2082 if (m != NULL) { 2083 flags |= PG_PCPU_CACHE; 2084 goto found; 2085 } 2086 } 2087 if (vm_domain_allocate(vmd, req, 1)) { 2088 /* 2089 * If not, allocate it from the free page queues. 2090 */ 2091 vm_domain_free_lock(vmd); 2092 m = vm_phys_alloc_pages(domain, pool, 0); 2093 vm_domain_free_unlock(vmd); 2094 if (m == NULL) { 2095 vm_domain_freecnt_inc(vmd, 1); 2096 #if VM_NRESERVLEVEL > 0 2097 if (vm_reserv_reclaim_inactive(domain)) 2098 goto again; 2099 #endif 2100 } 2101 } 2102 if (m == NULL) { 2103 /* 2104 * Not allocatable, give up. 2105 */ 2106 if (vm_domain_alloc_fail(vmd, object, req)) 2107 goto again; 2108 return (NULL); 2109 } 2110 2111 /* 2112 * At this point we had better have found a good page. 2113 */ 2114 found: 2115 vm_page_dequeue(m); 2116 vm_page_alloc_check(m); 2117 2118 /* 2119 * Initialize the page. Only the PG_ZERO flag is inherited. 2120 */ 2121 if ((req & VM_ALLOC_ZERO) != 0) 2122 flags |= (m->flags & PG_ZERO); 2123 if ((req & VM_ALLOC_NODUMP) != 0) 2124 flags |= PG_NODUMP; 2125 m->flags = flags; 2126 m->a.flags = 0; 2127 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2128 VPO_UNMANAGED : 0; 2129 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2130 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2131 else if ((req & VM_ALLOC_SBUSY) != 0) 2132 m->busy_lock = VPB_SHARERS_WORD(1); 2133 else 2134 m->busy_lock = VPB_UNBUSIED; 2135 if (req & VM_ALLOC_WIRED) { 2136 vm_wire_add(1); 2137 m->ref_count = 1; 2138 } 2139 m->a.act_count = 0; 2140 2141 if (object != NULL) { 2142 if (vm_page_insert_after(m, object, pindex, mpred)) { 2143 if (req & VM_ALLOC_WIRED) { 2144 vm_wire_sub(1); 2145 m->ref_count = 0; 2146 } 2147 KASSERT(m->object == NULL, ("page %p has object", m)); 2148 m->oflags = VPO_UNMANAGED; 2149 m->busy_lock = VPB_UNBUSIED; 2150 /* Don't change PG_ZERO. */ 2151 vm_page_free_toq(m); 2152 if (req & VM_ALLOC_WAITFAIL) { 2153 VM_OBJECT_WUNLOCK(object); 2154 vm_radix_wait(); 2155 VM_OBJECT_WLOCK(object); 2156 } 2157 return (NULL); 2158 } 2159 2160 /* Ignore device objects; the pager sets "memattr" for them. */ 2161 if (object->memattr != VM_MEMATTR_DEFAULT && 2162 (object->flags & OBJ_FICTITIOUS) == 0) 2163 pmap_page_set_memattr(m, object->memattr); 2164 } else 2165 m->pindex = pindex; 2166 2167 return (m); 2168 } 2169 2170 /* 2171 * vm_page_alloc_contig: 2172 * 2173 * Allocate a contiguous set of physical pages of the given size "npages" 2174 * from the free lists. All of the physical pages must be at or above 2175 * the given physical address "low" and below the given physical address 2176 * "high". The given value "alignment" determines the alignment of the 2177 * first physical page in the set. If the given value "boundary" is 2178 * non-zero, then the set of physical pages cannot cross any physical 2179 * address boundary that is a multiple of that value. Both "alignment" 2180 * and "boundary" must be a power of two. 2181 * 2182 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2183 * then the memory attribute setting for the physical pages is configured 2184 * to the object's memory attribute setting. Otherwise, the memory 2185 * attribute setting for the physical pages is configured to "memattr", 2186 * overriding the object's memory attribute setting. However, if the 2187 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2188 * memory attribute setting for the physical pages cannot be configured 2189 * to VM_MEMATTR_DEFAULT. 2190 * 2191 * The specified object may not contain fictitious pages. 2192 * 2193 * The caller must always specify an allocation class. 2194 * 2195 * allocation classes: 2196 * VM_ALLOC_NORMAL normal process request 2197 * VM_ALLOC_SYSTEM system *really* needs a page 2198 * VM_ALLOC_INTERRUPT interrupt time request 2199 * 2200 * optional allocation flags: 2201 * VM_ALLOC_NOBUSY do not exclusive busy the page 2202 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2203 * VM_ALLOC_NOOBJ page is not associated with an object and 2204 * should not be exclusive busy 2205 * VM_ALLOC_SBUSY shared busy the allocated page 2206 * VM_ALLOC_WIRED wire the allocated page 2207 * VM_ALLOC_ZERO prefer a zeroed page 2208 */ 2209 vm_page_t 2210 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2211 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2212 vm_paddr_t boundary, vm_memattr_t memattr) 2213 { 2214 struct vm_domainset_iter di; 2215 vm_page_t m; 2216 int domain; 2217 2218 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2219 do { 2220 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2221 npages, low, high, alignment, boundary, memattr); 2222 if (m != NULL) 2223 break; 2224 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2225 2226 return (m); 2227 } 2228 2229 vm_page_t 2230 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2231 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2232 vm_paddr_t boundary, vm_memattr_t memattr) 2233 { 2234 struct vm_domain *vmd; 2235 vm_page_t m, m_ret, mpred; 2236 u_int busy_lock, flags, oflags; 2237 2238 mpred = NULL; /* XXX: pacify gcc */ 2239 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2240 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2241 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2242 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2243 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 2244 req)); 2245 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2246 ("Can't sleep and retry object insertion.")); 2247 if (object != NULL) { 2248 VM_OBJECT_ASSERT_WLOCKED(object); 2249 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2250 ("vm_page_alloc_contig: object %p has fictitious pages", 2251 object)); 2252 } 2253 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2254 2255 if (object != NULL) { 2256 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2257 KASSERT(mpred == NULL || mpred->pindex != pindex, 2258 ("vm_page_alloc_contig: pindex already allocated")); 2259 } 2260 2261 /* 2262 * Can we allocate the pages without the number of free pages falling 2263 * below the lower bound for the allocation class? 2264 */ 2265 m_ret = NULL; 2266 again: 2267 #if VM_NRESERVLEVEL > 0 2268 /* 2269 * Can we allocate the pages from a reservation? 2270 */ 2271 if (vm_object_reserv(object) && 2272 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2273 mpred, npages, low, high, alignment, boundary)) != NULL) { 2274 goto found; 2275 } 2276 #endif 2277 vmd = VM_DOMAIN(domain); 2278 if (vm_domain_allocate(vmd, req, npages)) { 2279 /* 2280 * allocate them from the free page queues. 2281 */ 2282 vm_domain_free_lock(vmd); 2283 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2284 alignment, boundary); 2285 vm_domain_free_unlock(vmd); 2286 if (m_ret == NULL) { 2287 vm_domain_freecnt_inc(vmd, npages); 2288 #if VM_NRESERVLEVEL > 0 2289 if (vm_reserv_reclaim_contig(domain, npages, low, 2290 high, alignment, boundary)) 2291 goto again; 2292 #endif 2293 } 2294 } 2295 if (m_ret == NULL) { 2296 if (vm_domain_alloc_fail(vmd, object, req)) 2297 goto again; 2298 return (NULL); 2299 } 2300 #if VM_NRESERVLEVEL > 0 2301 found: 2302 #endif 2303 for (m = m_ret; m < &m_ret[npages]; m++) { 2304 vm_page_dequeue(m); 2305 vm_page_alloc_check(m); 2306 } 2307 2308 /* 2309 * Initialize the pages. Only the PG_ZERO flag is inherited. 2310 */ 2311 flags = 0; 2312 if ((req & VM_ALLOC_ZERO) != 0) 2313 flags = PG_ZERO; 2314 if ((req & VM_ALLOC_NODUMP) != 0) 2315 flags |= PG_NODUMP; 2316 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2317 VPO_UNMANAGED : 0; 2318 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2319 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2320 else if ((req & VM_ALLOC_SBUSY) != 0) 2321 busy_lock = VPB_SHARERS_WORD(1); 2322 else 2323 busy_lock = VPB_UNBUSIED; 2324 if ((req & VM_ALLOC_WIRED) != 0) 2325 vm_wire_add(npages); 2326 if (object != NULL) { 2327 if (object->memattr != VM_MEMATTR_DEFAULT && 2328 memattr == VM_MEMATTR_DEFAULT) 2329 memattr = object->memattr; 2330 } 2331 for (m = m_ret; m < &m_ret[npages]; m++) { 2332 m->a.flags = 0; 2333 m->flags = (m->flags | PG_NODUMP) & flags; 2334 m->busy_lock = busy_lock; 2335 if ((req & VM_ALLOC_WIRED) != 0) 2336 m->ref_count = 1; 2337 m->a.act_count = 0; 2338 m->oflags = oflags; 2339 if (object != NULL) { 2340 if (vm_page_insert_after(m, object, pindex, mpred)) { 2341 if ((req & VM_ALLOC_WIRED) != 0) 2342 vm_wire_sub(npages); 2343 KASSERT(m->object == NULL, 2344 ("page %p has object", m)); 2345 mpred = m; 2346 for (m = m_ret; m < &m_ret[npages]; m++) { 2347 if (m <= mpred && 2348 (req & VM_ALLOC_WIRED) != 0) 2349 m->ref_count = 0; 2350 m->oflags = VPO_UNMANAGED; 2351 m->busy_lock = VPB_UNBUSIED; 2352 /* Don't change PG_ZERO. */ 2353 vm_page_free_toq(m); 2354 } 2355 if (req & VM_ALLOC_WAITFAIL) { 2356 VM_OBJECT_WUNLOCK(object); 2357 vm_radix_wait(); 2358 VM_OBJECT_WLOCK(object); 2359 } 2360 return (NULL); 2361 } 2362 mpred = m; 2363 } else 2364 m->pindex = pindex; 2365 if (memattr != VM_MEMATTR_DEFAULT) 2366 pmap_page_set_memattr(m, memattr); 2367 pindex++; 2368 } 2369 return (m_ret); 2370 } 2371 2372 /* 2373 * Check a page that has been freshly dequeued from a freelist. 2374 */ 2375 static void 2376 vm_page_alloc_check(vm_page_t m) 2377 { 2378 2379 KASSERT(m->object == NULL, ("page %p has object", m)); 2380 KASSERT(m->a.queue == PQ_NONE && 2381 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2382 ("page %p has unexpected queue %d, flags %#x", 2383 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2384 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2385 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2386 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2387 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2388 ("page %p has unexpected memattr %d", 2389 m, pmap_page_get_memattr(m))); 2390 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2391 } 2392 2393 /* 2394 * vm_page_alloc_freelist: 2395 * 2396 * Allocate a physical page from the specified free page list. 2397 * 2398 * The caller must always specify an allocation class. 2399 * 2400 * allocation classes: 2401 * VM_ALLOC_NORMAL normal process request 2402 * VM_ALLOC_SYSTEM system *really* needs a page 2403 * VM_ALLOC_INTERRUPT interrupt time request 2404 * 2405 * optional allocation flags: 2406 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2407 * intends to allocate 2408 * VM_ALLOC_WIRED wire the allocated page 2409 * VM_ALLOC_ZERO prefer a zeroed page 2410 */ 2411 vm_page_t 2412 vm_page_alloc_freelist(int freelist, int req) 2413 { 2414 struct vm_domainset_iter di; 2415 vm_page_t m; 2416 int domain; 2417 2418 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2419 do { 2420 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2421 if (m != NULL) 2422 break; 2423 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2424 2425 return (m); 2426 } 2427 2428 vm_page_t 2429 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2430 { 2431 struct vm_domain *vmd; 2432 vm_page_t m; 2433 u_int flags; 2434 2435 m = NULL; 2436 vmd = VM_DOMAIN(domain); 2437 again: 2438 if (vm_domain_allocate(vmd, req, 1)) { 2439 vm_domain_free_lock(vmd); 2440 m = vm_phys_alloc_freelist_pages(domain, freelist, 2441 VM_FREEPOOL_DIRECT, 0); 2442 vm_domain_free_unlock(vmd); 2443 if (m == NULL) 2444 vm_domain_freecnt_inc(vmd, 1); 2445 } 2446 if (m == NULL) { 2447 if (vm_domain_alloc_fail(vmd, NULL, req)) 2448 goto again; 2449 return (NULL); 2450 } 2451 vm_page_dequeue(m); 2452 vm_page_alloc_check(m); 2453 2454 /* 2455 * Initialize the page. Only the PG_ZERO flag is inherited. 2456 */ 2457 m->a.flags = 0; 2458 flags = 0; 2459 if ((req & VM_ALLOC_ZERO) != 0) 2460 flags = PG_ZERO; 2461 m->flags &= flags; 2462 if ((req & VM_ALLOC_WIRED) != 0) { 2463 vm_wire_add(1); 2464 m->ref_count = 1; 2465 } 2466 /* Unmanaged pages don't use "act_count". */ 2467 m->oflags = VPO_UNMANAGED; 2468 return (m); 2469 } 2470 2471 static int 2472 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2473 { 2474 struct vm_domain *vmd; 2475 struct vm_pgcache *pgcache; 2476 int i; 2477 2478 pgcache = arg; 2479 vmd = VM_DOMAIN(pgcache->domain); 2480 2481 /* 2482 * The page daemon should avoid creating extra memory pressure since its 2483 * main purpose is to replenish the store of free pages. 2484 */ 2485 if (vmd->vmd_severeset || curproc == pageproc || 2486 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2487 return (0); 2488 domain = vmd->vmd_domain; 2489 vm_domain_free_lock(vmd); 2490 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2491 (vm_page_t *)store); 2492 vm_domain_free_unlock(vmd); 2493 if (cnt != i) 2494 vm_domain_freecnt_inc(vmd, cnt - i); 2495 2496 return (i); 2497 } 2498 2499 static void 2500 vm_page_zone_release(void *arg, void **store, int cnt) 2501 { 2502 struct vm_domain *vmd; 2503 struct vm_pgcache *pgcache; 2504 vm_page_t m; 2505 int i; 2506 2507 pgcache = arg; 2508 vmd = VM_DOMAIN(pgcache->domain); 2509 vm_domain_free_lock(vmd); 2510 for (i = 0; i < cnt; i++) { 2511 m = (vm_page_t)store[i]; 2512 vm_phys_free_pages(m, 0); 2513 } 2514 vm_domain_free_unlock(vmd); 2515 vm_domain_freecnt_inc(vmd, cnt); 2516 } 2517 2518 #define VPSC_ANY 0 /* No restrictions. */ 2519 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2520 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2521 2522 /* 2523 * vm_page_scan_contig: 2524 * 2525 * Scan vm_page_array[] between the specified entries "m_start" and 2526 * "m_end" for a run of contiguous physical pages that satisfy the 2527 * specified conditions, and return the lowest page in the run. The 2528 * specified "alignment" determines the alignment of the lowest physical 2529 * page in the run. If the specified "boundary" is non-zero, then the 2530 * run of physical pages cannot span a physical address that is a 2531 * multiple of "boundary". 2532 * 2533 * "m_end" is never dereferenced, so it need not point to a vm_page 2534 * structure within vm_page_array[]. 2535 * 2536 * "npages" must be greater than zero. "m_start" and "m_end" must not 2537 * span a hole (or discontiguity) in the physical address space. Both 2538 * "alignment" and "boundary" must be a power of two. 2539 */ 2540 vm_page_t 2541 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2542 u_long alignment, vm_paddr_t boundary, int options) 2543 { 2544 vm_object_t object; 2545 vm_paddr_t pa; 2546 vm_page_t m, m_run; 2547 #if VM_NRESERVLEVEL > 0 2548 int level; 2549 #endif 2550 int m_inc, order, run_ext, run_len; 2551 2552 KASSERT(npages > 0, ("npages is 0")); 2553 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2554 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2555 m_run = NULL; 2556 run_len = 0; 2557 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2558 KASSERT((m->flags & PG_MARKER) == 0, 2559 ("page %p is PG_MARKER", m)); 2560 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2561 ("fictitious page %p has invalid ref count", m)); 2562 2563 /* 2564 * If the current page would be the start of a run, check its 2565 * physical address against the end, alignment, and boundary 2566 * conditions. If it doesn't satisfy these conditions, either 2567 * terminate the scan or advance to the next page that 2568 * satisfies the failed condition. 2569 */ 2570 if (run_len == 0) { 2571 KASSERT(m_run == NULL, ("m_run != NULL")); 2572 if (m + npages > m_end) 2573 break; 2574 pa = VM_PAGE_TO_PHYS(m); 2575 if ((pa & (alignment - 1)) != 0) { 2576 m_inc = atop(roundup2(pa, alignment) - pa); 2577 continue; 2578 } 2579 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2580 boundary) != 0) { 2581 m_inc = atop(roundup2(pa, boundary) - pa); 2582 continue; 2583 } 2584 } else 2585 KASSERT(m_run != NULL, ("m_run == NULL")); 2586 2587 retry: 2588 m_inc = 1; 2589 if (vm_page_wired(m)) 2590 run_ext = 0; 2591 #if VM_NRESERVLEVEL > 0 2592 else if ((level = vm_reserv_level(m)) >= 0 && 2593 (options & VPSC_NORESERV) != 0) { 2594 run_ext = 0; 2595 /* Advance to the end of the reservation. */ 2596 pa = VM_PAGE_TO_PHYS(m); 2597 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2598 pa); 2599 } 2600 #endif 2601 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2602 /* 2603 * The page is considered eligible for relocation if 2604 * and only if it could be laundered or reclaimed by 2605 * the page daemon. 2606 */ 2607 VM_OBJECT_RLOCK(object); 2608 if (object != m->object) { 2609 VM_OBJECT_RUNLOCK(object); 2610 goto retry; 2611 } 2612 /* Don't care: PG_NODUMP, PG_ZERO. */ 2613 if (object->type != OBJT_DEFAULT && 2614 object->type != OBJT_SWAP && 2615 object->type != OBJT_VNODE) { 2616 run_ext = 0; 2617 #if VM_NRESERVLEVEL > 0 2618 } else if ((options & VPSC_NOSUPER) != 0 && 2619 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2620 run_ext = 0; 2621 /* Advance to the end of the superpage. */ 2622 pa = VM_PAGE_TO_PHYS(m); 2623 m_inc = atop(roundup2(pa + 1, 2624 vm_reserv_size(level)) - pa); 2625 #endif 2626 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2627 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2628 /* 2629 * The page is allocated but eligible for 2630 * relocation. Extend the current run by one 2631 * page. 2632 */ 2633 KASSERT(pmap_page_get_memattr(m) == 2634 VM_MEMATTR_DEFAULT, 2635 ("page %p has an unexpected memattr", m)); 2636 KASSERT((m->oflags & (VPO_SWAPINPROG | 2637 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2638 ("page %p has unexpected oflags", m)); 2639 /* Don't care: PGA_NOSYNC. */ 2640 run_ext = 1; 2641 } else 2642 run_ext = 0; 2643 VM_OBJECT_RUNLOCK(object); 2644 #if VM_NRESERVLEVEL > 0 2645 } else if (level >= 0) { 2646 /* 2647 * The page is reserved but not yet allocated. In 2648 * other words, it is still free. Extend the current 2649 * run by one page. 2650 */ 2651 run_ext = 1; 2652 #endif 2653 } else if ((order = m->order) < VM_NFREEORDER) { 2654 /* 2655 * The page is enqueued in the physical memory 2656 * allocator's free page queues. Moreover, it is the 2657 * first page in a power-of-two-sized run of 2658 * contiguous free pages. Add these pages to the end 2659 * of the current run, and jump ahead. 2660 */ 2661 run_ext = 1 << order; 2662 m_inc = 1 << order; 2663 } else { 2664 /* 2665 * Skip the page for one of the following reasons: (1) 2666 * It is enqueued in the physical memory allocator's 2667 * free page queues. However, it is not the first 2668 * page in a run of contiguous free pages. (This case 2669 * rarely occurs because the scan is performed in 2670 * ascending order.) (2) It is not reserved, and it is 2671 * transitioning from free to allocated. (Conversely, 2672 * the transition from allocated to free for managed 2673 * pages is blocked by the page lock.) (3) It is 2674 * allocated but not contained by an object and not 2675 * wired, e.g., allocated by Xen's balloon driver. 2676 */ 2677 run_ext = 0; 2678 } 2679 2680 /* 2681 * Extend or reset the current run of pages. 2682 */ 2683 if (run_ext > 0) { 2684 if (run_len == 0) 2685 m_run = m; 2686 run_len += run_ext; 2687 } else { 2688 if (run_len > 0) { 2689 m_run = NULL; 2690 run_len = 0; 2691 } 2692 } 2693 } 2694 if (run_len >= npages) 2695 return (m_run); 2696 return (NULL); 2697 } 2698 2699 /* 2700 * vm_page_reclaim_run: 2701 * 2702 * Try to relocate each of the allocated virtual pages within the 2703 * specified run of physical pages to a new physical address. Free the 2704 * physical pages underlying the relocated virtual pages. A virtual page 2705 * is relocatable if and only if it could be laundered or reclaimed by 2706 * the page daemon. Whenever possible, a virtual page is relocated to a 2707 * physical address above "high". 2708 * 2709 * Returns 0 if every physical page within the run was already free or 2710 * just freed by a successful relocation. Otherwise, returns a non-zero 2711 * value indicating why the last attempt to relocate a virtual page was 2712 * unsuccessful. 2713 * 2714 * "req_class" must be an allocation class. 2715 */ 2716 static int 2717 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2718 vm_paddr_t high) 2719 { 2720 struct vm_domain *vmd; 2721 struct spglist free; 2722 vm_object_t object; 2723 vm_paddr_t pa; 2724 vm_page_t m, m_end, m_new; 2725 int error, order, req; 2726 2727 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2728 ("req_class is not an allocation class")); 2729 SLIST_INIT(&free); 2730 error = 0; 2731 m = m_run; 2732 m_end = m_run + npages; 2733 for (; error == 0 && m < m_end; m++) { 2734 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2735 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2736 2737 /* 2738 * Racily check for wirings. Races are handled once the object 2739 * lock is held and the page is unmapped. 2740 */ 2741 if (vm_page_wired(m)) 2742 error = EBUSY; 2743 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2744 /* 2745 * The page is relocated if and only if it could be 2746 * laundered or reclaimed by the page daemon. 2747 */ 2748 VM_OBJECT_WLOCK(object); 2749 /* Don't care: PG_NODUMP, PG_ZERO. */ 2750 if (m->object != object || 2751 (object->type != OBJT_DEFAULT && 2752 object->type != OBJT_SWAP && 2753 object->type != OBJT_VNODE)) 2754 error = EINVAL; 2755 else if (object->memattr != VM_MEMATTR_DEFAULT) 2756 error = EINVAL; 2757 else if (vm_page_queue(m) != PQ_NONE && 2758 vm_page_tryxbusy(m) != 0) { 2759 if (vm_page_wired(m)) { 2760 vm_page_xunbusy(m); 2761 error = EBUSY; 2762 goto unlock; 2763 } 2764 KASSERT(pmap_page_get_memattr(m) == 2765 VM_MEMATTR_DEFAULT, 2766 ("page %p has an unexpected memattr", m)); 2767 KASSERT(m->oflags == 0, 2768 ("page %p has unexpected oflags", m)); 2769 /* Don't care: PGA_NOSYNC. */ 2770 if (!vm_page_none_valid(m)) { 2771 /* 2772 * First, try to allocate a new page 2773 * that is above "high". Failing 2774 * that, try to allocate a new page 2775 * that is below "m_run". Allocate 2776 * the new page between the end of 2777 * "m_run" and "high" only as a last 2778 * resort. 2779 */ 2780 req = req_class | VM_ALLOC_NOOBJ; 2781 if ((m->flags & PG_NODUMP) != 0) 2782 req |= VM_ALLOC_NODUMP; 2783 if (trunc_page(high) != 2784 ~(vm_paddr_t)PAGE_MASK) { 2785 m_new = vm_page_alloc_contig( 2786 NULL, 0, req, 1, 2787 round_page(high), 2788 ~(vm_paddr_t)0, 2789 PAGE_SIZE, 0, 2790 VM_MEMATTR_DEFAULT); 2791 } else 2792 m_new = NULL; 2793 if (m_new == NULL) { 2794 pa = VM_PAGE_TO_PHYS(m_run); 2795 m_new = vm_page_alloc_contig( 2796 NULL, 0, req, 1, 2797 0, pa - 1, PAGE_SIZE, 0, 2798 VM_MEMATTR_DEFAULT); 2799 } 2800 if (m_new == NULL) { 2801 pa += ptoa(npages); 2802 m_new = vm_page_alloc_contig( 2803 NULL, 0, req, 1, 2804 pa, high, PAGE_SIZE, 0, 2805 VM_MEMATTR_DEFAULT); 2806 } 2807 if (m_new == NULL) { 2808 vm_page_xunbusy(m); 2809 error = ENOMEM; 2810 goto unlock; 2811 } 2812 2813 /* 2814 * Unmap the page and check for new 2815 * wirings that may have been acquired 2816 * through a pmap lookup. 2817 */ 2818 if (object->ref_count != 0 && 2819 !vm_page_try_remove_all(m)) { 2820 vm_page_xunbusy(m); 2821 vm_page_free(m_new); 2822 error = EBUSY; 2823 goto unlock; 2824 } 2825 2826 /* 2827 * Replace "m" with the new page. For 2828 * vm_page_replace(), "m" must be busy 2829 * and dequeued. Finally, change "m" 2830 * as if vm_page_free() was called. 2831 */ 2832 m_new->a.flags = m->a.flags & 2833 ~PGA_QUEUE_STATE_MASK; 2834 KASSERT(m_new->oflags == VPO_UNMANAGED, 2835 ("page %p is managed", m_new)); 2836 m_new->oflags = 0; 2837 pmap_copy_page(m, m_new); 2838 m_new->valid = m->valid; 2839 m_new->dirty = m->dirty; 2840 m->flags &= ~PG_ZERO; 2841 vm_page_dequeue(m); 2842 if (vm_page_replace_hold(m_new, object, 2843 m->pindex, m) && 2844 vm_page_free_prep(m)) 2845 SLIST_INSERT_HEAD(&free, m, 2846 plinks.s.ss); 2847 2848 /* 2849 * The new page must be deactivated 2850 * before the object is unlocked. 2851 */ 2852 vm_page_deactivate(m_new); 2853 } else { 2854 m->flags &= ~PG_ZERO; 2855 vm_page_dequeue(m); 2856 if (vm_page_free_prep(m)) 2857 SLIST_INSERT_HEAD(&free, m, 2858 plinks.s.ss); 2859 KASSERT(m->dirty == 0, 2860 ("page %p is dirty", m)); 2861 } 2862 } else 2863 error = EBUSY; 2864 unlock: 2865 VM_OBJECT_WUNLOCK(object); 2866 } else { 2867 MPASS(vm_phys_domain(m) == domain); 2868 vmd = VM_DOMAIN(domain); 2869 vm_domain_free_lock(vmd); 2870 order = m->order; 2871 if (order < VM_NFREEORDER) { 2872 /* 2873 * The page is enqueued in the physical memory 2874 * allocator's free page queues. Moreover, it 2875 * is the first page in a power-of-two-sized 2876 * run of contiguous free pages. Jump ahead 2877 * to the last page within that run, and 2878 * continue from there. 2879 */ 2880 m += (1 << order) - 1; 2881 } 2882 #if VM_NRESERVLEVEL > 0 2883 else if (vm_reserv_is_page_free(m)) 2884 order = 0; 2885 #endif 2886 vm_domain_free_unlock(vmd); 2887 if (order == VM_NFREEORDER) 2888 error = EINVAL; 2889 } 2890 } 2891 if ((m = SLIST_FIRST(&free)) != NULL) { 2892 int cnt; 2893 2894 vmd = VM_DOMAIN(domain); 2895 cnt = 0; 2896 vm_domain_free_lock(vmd); 2897 do { 2898 MPASS(vm_phys_domain(m) == domain); 2899 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2900 vm_phys_free_pages(m, 0); 2901 cnt++; 2902 } while ((m = SLIST_FIRST(&free)) != NULL); 2903 vm_domain_free_unlock(vmd); 2904 vm_domain_freecnt_inc(vmd, cnt); 2905 } 2906 return (error); 2907 } 2908 2909 #define NRUNS 16 2910 2911 CTASSERT(powerof2(NRUNS)); 2912 2913 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2914 2915 #define MIN_RECLAIM 8 2916 2917 /* 2918 * vm_page_reclaim_contig: 2919 * 2920 * Reclaim allocated, contiguous physical memory satisfying the specified 2921 * conditions by relocating the virtual pages using that physical memory. 2922 * Returns true if reclamation is successful and false otherwise. Since 2923 * relocation requires the allocation of physical pages, reclamation may 2924 * fail due to a shortage of free pages. When reclamation fails, callers 2925 * are expected to perform vm_wait() before retrying a failed allocation 2926 * operation, e.g., vm_page_alloc_contig(). 2927 * 2928 * The caller must always specify an allocation class through "req". 2929 * 2930 * allocation classes: 2931 * VM_ALLOC_NORMAL normal process request 2932 * VM_ALLOC_SYSTEM system *really* needs a page 2933 * VM_ALLOC_INTERRUPT interrupt time request 2934 * 2935 * The optional allocation flags are ignored. 2936 * 2937 * "npages" must be greater than zero. Both "alignment" and "boundary" 2938 * must be a power of two. 2939 */ 2940 bool 2941 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2942 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2943 { 2944 struct vm_domain *vmd; 2945 vm_paddr_t curr_low; 2946 vm_page_t m_run, m_runs[NRUNS]; 2947 u_long count, reclaimed; 2948 int error, i, options, req_class; 2949 2950 KASSERT(npages > 0, ("npages is 0")); 2951 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2952 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2953 req_class = req & VM_ALLOC_CLASS_MASK; 2954 2955 /* 2956 * The page daemon is allowed to dig deeper into the free page list. 2957 */ 2958 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2959 req_class = VM_ALLOC_SYSTEM; 2960 2961 /* 2962 * Return if the number of free pages cannot satisfy the requested 2963 * allocation. 2964 */ 2965 vmd = VM_DOMAIN(domain); 2966 count = vmd->vmd_free_count; 2967 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2968 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2969 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2970 return (false); 2971 2972 /* 2973 * Scan up to three times, relaxing the restrictions ("options") on 2974 * the reclamation of reservations and superpages each time. 2975 */ 2976 for (options = VPSC_NORESERV;;) { 2977 /* 2978 * Find the highest runs that satisfy the given constraints 2979 * and restrictions, and record them in "m_runs". 2980 */ 2981 curr_low = low; 2982 count = 0; 2983 for (;;) { 2984 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2985 high, alignment, boundary, options); 2986 if (m_run == NULL) 2987 break; 2988 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2989 m_runs[RUN_INDEX(count)] = m_run; 2990 count++; 2991 } 2992 2993 /* 2994 * Reclaim the highest runs in LIFO (descending) order until 2995 * the number of reclaimed pages, "reclaimed", is at least 2996 * MIN_RECLAIM. Reset "reclaimed" each time because each 2997 * reclamation is idempotent, and runs will (likely) recur 2998 * from one scan to the next as restrictions are relaxed. 2999 */ 3000 reclaimed = 0; 3001 for (i = 0; count > 0 && i < NRUNS; i++) { 3002 count--; 3003 m_run = m_runs[RUN_INDEX(count)]; 3004 error = vm_page_reclaim_run(req_class, domain, npages, 3005 m_run, high); 3006 if (error == 0) { 3007 reclaimed += npages; 3008 if (reclaimed >= MIN_RECLAIM) 3009 return (true); 3010 } 3011 } 3012 3013 /* 3014 * Either relax the restrictions on the next scan or return if 3015 * the last scan had no restrictions. 3016 */ 3017 if (options == VPSC_NORESERV) 3018 options = VPSC_NOSUPER; 3019 else if (options == VPSC_NOSUPER) 3020 options = VPSC_ANY; 3021 else if (options == VPSC_ANY) 3022 return (reclaimed != 0); 3023 } 3024 } 3025 3026 bool 3027 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3028 u_long alignment, vm_paddr_t boundary) 3029 { 3030 struct vm_domainset_iter di; 3031 int domain; 3032 bool ret; 3033 3034 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3035 do { 3036 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 3037 high, alignment, boundary); 3038 if (ret) 3039 break; 3040 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3041 3042 return (ret); 3043 } 3044 3045 /* 3046 * Set the domain in the appropriate page level domainset. 3047 */ 3048 void 3049 vm_domain_set(struct vm_domain *vmd) 3050 { 3051 3052 mtx_lock(&vm_domainset_lock); 3053 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3054 vmd->vmd_minset = 1; 3055 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3056 } 3057 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3058 vmd->vmd_severeset = 1; 3059 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3060 } 3061 mtx_unlock(&vm_domainset_lock); 3062 } 3063 3064 /* 3065 * Clear the domain from the appropriate page level domainset. 3066 */ 3067 void 3068 vm_domain_clear(struct vm_domain *vmd) 3069 { 3070 3071 mtx_lock(&vm_domainset_lock); 3072 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3073 vmd->vmd_minset = 0; 3074 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3075 if (vm_min_waiters != 0) { 3076 vm_min_waiters = 0; 3077 wakeup(&vm_min_domains); 3078 } 3079 } 3080 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3081 vmd->vmd_severeset = 0; 3082 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3083 if (vm_severe_waiters != 0) { 3084 vm_severe_waiters = 0; 3085 wakeup(&vm_severe_domains); 3086 } 3087 } 3088 3089 /* 3090 * If pageout daemon needs pages, then tell it that there are 3091 * some free. 3092 */ 3093 if (vmd->vmd_pageout_pages_needed && 3094 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3095 wakeup(&vmd->vmd_pageout_pages_needed); 3096 vmd->vmd_pageout_pages_needed = 0; 3097 } 3098 3099 /* See comments in vm_wait_doms(). */ 3100 if (vm_pageproc_waiters) { 3101 vm_pageproc_waiters = 0; 3102 wakeup(&vm_pageproc_waiters); 3103 } 3104 mtx_unlock(&vm_domainset_lock); 3105 } 3106 3107 /* 3108 * Wait for free pages to exceed the min threshold globally. 3109 */ 3110 void 3111 vm_wait_min(void) 3112 { 3113 3114 mtx_lock(&vm_domainset_lock); 3115 while (vm_page_count_min()) { 3116 vm_min_waiters++; 3117 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3118 } 3119 mtx_unlock(&vm_domainset_lock); 3120 } 3121 3122 /* 3123 * Wait for free pages to exceed the severe threshold globally. 3124 */ 3125 void 3126 vm_wait_severe(void) 3127 { 3128 3129 mtx_lock(&vm_domainset_lock); 3130 while (vm_page_count_severe()) { 3131 vm_severe_waiters++; 3132 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3133 "vmwait", 0); 3134 } 3135 mtx_unlock(&vm_domainset_lock); 3136 } 3137 3138 u_int 3139 vm_wait_count(void) 3140 { 3141 3142 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3143 } 3144 3145 void 3146 vm_wait_doms(const domainset_t *wdoms) 3147 { 3148 3149 /* 3150 * We use racey wakeup synchronization to avoid expensive global 3151 * locking for the pageproc when sleeping with a non-specific vm_wait. 3152 * To handle this, we only sleep for one tick in this instance. It 3153 * is expected that most allocations for the pageproc will come from 3154 * kmem or vm_page_grab* which will use the more specific and 3155 * race-free vm_wait_domain(). 3156 */ 3157 if (curproc == pageproc) { 3158 mtx_lock(&vm_domainset_lock); 3159 vm_pageproc_waiters++; 3160 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, 3161 "pageprocwait", 1); 3162 } else { 3163 /* 3164 * XXX Ideally we would wait only until the allocation could 3165 * be satisfied. This condition can cause new allocators to 3166 * consume all freed pages while old allocators wait. 3167 */ 3168 mtx_lock(&vm_domainset_lock); 3169 if (vm_page_count_min_set(wdoms)) { 3170 vm_min_waiters++; 3171 msleep(&vm_min_domains, &vm_domainset_lock, 3172 PVM | PDROP, "vmwait", 0); 3173 } else 3174 mtx_unlock(&vm_domainset_lock); 3175 } 3176 } 3177 3178 /* 3179 * vm_wait_domain: 3180 * 3181 * Sleep until free pages are available for allocation. 3182 * - Called in various places after failed memory allocations. 3183 */ 3184 void 3185 vm_wait_domain(int domain) 3186 { 3187 struct vm_domain *vmd; 3188 domainset_t wdom; 3189 3190 vmd = VM_DOMAIN(domain); 3191 vm_domain_free_assert_unlocked(vmd); 3192 3193 if (curproc == pageproc) { 3194 mtx_lock(&vm_domainset_lock); 3195 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3196 vmd->vmd_pageout_pages_needed = 1; 3197 msleep(&vmd->vmd_pageout_pages_needed, 3198 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3199 } else 3200 mtx_unlock(&vm_domainset_lock); 3201 } else { 3202 if (pageproc == NULL) 3203 panic("vm_wait in early boot"); 3204 DOMAINSET_ZERO(&wdom); 3205 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3206 vm_wait_doms(&wdom); 3207 } 3208 } 3209 3210 /* 3211 * vm_wait: 3212 * 3213 * Sleep until free pages are available for allocation in the 3214 * affinity domains of the obj. If obj is NULL, the domain set 3215 * for the calling thread is used. 3216 * Called in various places after failed memory allocations. 3217 */ 3218 void 3219 vm_wait(vm_object_t obj) 3220 { 3221 struct domainset *d; 3222 3223 d = NULL; 3224 3225 /* 3226 * Carefully fetch pointers only once: the struct domainset 3227 * itself is ummutable but the pointer might change. 3228 */ 3229 if (obj != NULL) 3230 d = obj->domain.dr_policy; 3231 if (d == NULL) 3232 d = curthread->td_domain.dr_policy; 3233 3234 vm_wait_doms(&d->ds_mask); 3235 } 3236 3237 /* 3238 * vm_domain_alloc_fail: 3239 * 3240 * Called when a page allocation function fails. Informs the 3241 * pagedaemon and performs the requested wait. Requires the 3242 * domain_free and object lock on entry. Returns with the 3243 * object lock held and free lock released. Returns an error when 3244 * retry is necessary. 3245 * 3246 */ 3247 static int 3248 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3249 { 3250 3251 vm_domain_free_assert_unlocked(vmd); 3252 3253 atomic_add_int(&vmd->vmd_pageout_deficit, 3254 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3255 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3256 if (object != NULL) 3257 VM_OBJECT_WUNLOCK(object); 3258 vm_wait_domain(vmd->vmd_domain); 3259 if (object != NULL) 3260 VM_OBJECT_WLOCK(object); 3261 if (req & VM_ALLOC_WAITOK) 3262 return (EAGAIN); 3263 } 3264 3265 return (0); 3266 } 3267 3268 /* 3269 * vm_waitpfault: 3270 * 3271 * Sleep until free pages are available for allocation. 3272 * - Called only in vm_fault so that processes page faulting 3273 * can be easily tracked. 3274 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3275 * processes will be able to grab memory first. Do not change 3276 * this balance without careful testing first. 3277 */ 3278 void 3279 vm_waitpfault(struct domainset *dset, int timo) 3280 { 3281 3282 /* 3283 * XXX Ideally we would wait only until the allocation could 3284 * be satisfied. This condition can cause new allocators to 3285 * consume all freed pages while old allocators wait. 3286 */ 3287 mtx_lock(&vm_domainset_lock); 3288 if (vm_page_count_min_set(&dset->ds_mask)) { 3289 vm_min_waiters++; 3290 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3291 "pfault", timo); 3292 } else 3293 mtx_unlock(&vm_domainset_lock); 3294 } 3295 3296 static struct vm_pagequeue * 3297 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3298 { 3299 3300 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3301 } 3302 3303 #ifdef INVARIANTS 3304 static struct vm_pagequeue * 3305 vm_page_pagequeue(vm_page_t m) 3306 { 3307 3308 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3309 } 3310 #endif 3311 3312 static __always_inline bool 3313 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3314 { 3315 vm_page_astate_t tmp; 3316 3317 tmp = *old; 3318 do { 3319 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3320 return (true); 3321 counter_u64_add(pqstate_commit_retries, 1); 3322 } while (old->_bits == tmp._bits); 3323 3324 return (false); 3325 } 3326 3327 /* 3328 * Do the work of committing a queue state update that moves the page out of 3329 * its current queue. 3330 */ 3331 static bool 3332 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3333 vm_page_astate_t *old, vm_page_astate_t new) 3334 { 3335 vm_page_t next; 3336 3337 vm_pagequeue_assert_locked(pq); 3338 KASSERT(vm_page_pagequeue(m) == pq, 3339 ("%s: queue %p does not match page %p", __func__, pq, m)); 3340 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3341 ("%s: invalid queue indices %d %d", 3342 __func__, old->queue, new.queue)); 3343 3344 /* 3345 * Once the queue index of the page changes there is nothing 3346 * synchronizing with further updates to the page's physical 3347 * queue state. Therefore we must speculatively remove the page 3348 * from the queue now and be prepared to roll back if the queue 3349 * state update fails. If the page is not physically enqueued then 3350 * we just update its queue index. 3351 */ 3352 if ((old->flags & PGA_ENQUEUED) != 0) { 3353 new.flags &= ~PGA_ENQUEUED; 3354 next = TAILQ_NEXT(m, plinks.q); 3355 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3356 vm_pagequeue_cnt_dec(pq); 3357 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3358 if (next == NULL) 3359 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3360 else 3361 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3362 vm_pagequeue_cnt_inc(pq); 3363 return (false); 3364 } else { 3365 return (true); 3366 } 3367 } else { 3368 return (vm_page_pqstate_fcmpset(m, old, new)); 3369 } 3370 } 3371 3372 static bool 3373 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3374 vm_page_astate_t new) 3375 { 3376 struct vm_pagequeue *pq; 3377 vm_page_astate_t as; 3378 bool ret; 3379 3380 pq = _vm_page_pagequeue(m, old->queue); 3381 3382 /* 3383 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3384 * corresponding page queue lock is held. 3385 */ 3386 vm_pagequeue_lock(pq); 3387 as = vm_page_astate_load(m); 3388 if (__predict_false(as._bits != old->_bits)) { 3389 *old = as; 3390 ret = false; 3391 } else { 3392 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3393 } 3394 vm_pagequeue_unlock(pq); 3395 return (ret); 3396 } 3397 3398 /* 3399 * Commit a queue state update that enqueues or requeues a page. 3400 */ 3401 static bool 3402 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3403 vm_page_astate_t *old, vm_page_astate_t new) 3404 { 3405 struct vm_domain *vmd; 3406 3407 vm_pagequeue_assert_locked(pq); 3408 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3409 ("%s: invalid queue indices %d %d", 3410 __func__, old->queue, new.queue)); 3411 3412 new.flags |= PGA_ENQUEUED; 3413 if (!vm_page_pqstate_fcmpset(m, old, new)) 3414 return (false); 3415 3416 if ((old->flags & PGA_ENQUEUED) != 0) 3417 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3418 else 3419 vm_pagequeue_cnt_inc(pq); 3420 3421 /* 3422 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3423 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3424 * applied, even if it was set first. 3425 */ 3426 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3427 vmd = vm_pagequeue_domain(m); 3428 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3429 ("%s: invalid page queue for page %p", __func__, m)); 3430 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3431 } else { 3432 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3433 } 3434 return (true); 3435 } 3436 3437 /* 3438 * Commit a queue state update that encodes a request for a deferred queue 3439 * operation. 3440 */ 3441 static bool 3442 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3443 vm_page_astate_t new) 3444 { 3445 3446 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3447 ("%s: invalid state, queue %d flags %x", 3448 __func__, new.queue, new.flags)); 3449 3450 if (old->_bits != new._bits && 3451 !vm_page_pqstate_fcmpset(m, old, new)) 3452 return (false); 3453 vm_page_pqbatch_submit(m, new.queue); 3454 return (true); 3455 } 3456 3457 /* 3458 * A generic queue state update function. This handles more cases than the 3459 * specialized functions above. 3460 */ 3461 bool 3462 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3463 { 3464 3465 if (old->_bits == new._bits) 3466 return (true); 3467 3468 if (old->queue != PQ_NONE && new.queue != old->queue) { 3469 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3470 return (false); 3471 if (new.queue != PQ_NONE) 3472 vm_page_pqbatch_submit(m, new.queue); 3473 } else { 3474 if (!vm_page_pqstate_fcmpset(m, old, new)) 3475 return (false); 3476 if (new.queue != PQ_NONE && 3477 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3478 vm_page_pqbatch_submit(m, new.queue); 3479 } 3480 return (true); 3481 } 3482 3483 /* 3484 * Apply deferred queue state updates to a page. 3485 */ 3486 static inline void 3487 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3488 { 3489 vm_page_astate_t new, old; 3490 3491 CRITICAL_ASSERT(curthread); 3492 vm_pagequeue_assert_locked(pq); 3493 KASSERT(queue < PQ_COUNT, 3494 ("%s: invalid queue index %d", __func__, queue)); 3495 KASSERT(pq == _vm_page_pagequeue(m, queue), 3496 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3497 3498 for (old = vm_page_astate_load(m);;) { 3499 if (__predict_false(old.queue != queue || 3500 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3501 counter_u64_add(queue_nops, 1); 3502 break; 3503 } 3504 KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0, 3505 ("%s: page %p has unexpected queue state", __func__, m)); 3506 3507 new = old; 3508 if ((old.flags & PGA_DEQUEUE) != 0) { 3509 new.flags &= ~PGA_QUEUE_OP_MASK; 3510 new.queue = PQ_NONE; 3511 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3512 m, &old, new))) { 3513 counter_u64_add(queue_ops, 1); 3514 break; 3515 } 3516 } else { 3517 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3518 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3519 m, &old, new))) { 3520 counter_u64_add(queue_ops, 1); 3521 break; 3522 } 3523 } 3524 } 3525 } 3526 3527 static void 3528 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3529 uint8_t queue) 3530 { 3531 int i; 3532 3533 for (i = 0; i < bq->bq_cnt; i++) 3534 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3535 vm_batchqueue_init(bq); 3536 } 3537 3538 /* 3539 * vm_page_pqbatch_submit: [ internal use only ] 3540 * 3541 * Enqueue a page in the specified page queue's batched work queue. 3542 * The caller must have encoded the requested operation in the page 3543 * structure's a.flags field. 3544 */ 3545 void 3546 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3547 { 3548 struct vm_batchqueue *bq; 3549 struct vm_pagequeue *pq; 3550 int domain; 3551 3552 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3553 ("page %p is unmanaged", m)); 3554 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3555 3556 domain = vm_phys_domain(m); 3557 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 3558 3559 critical_enter(); 3560 bq = DPCPU_PTR(pqbatch[domain][queue]); 3561 if (vm_batchqueue_insert(bq, m)) { 3562 critical_exit(); 3563 return; 3564 } 3565 critical_exit(); 3566 vm_pagequeue_lock(pq); 3567 critical_enter(); 3568 bq = DPCPU_PTR(pqbatch[domain][queue]); 3569 vm_pqbatch_process(pq, bq, queue); 3570 vm_pqbatch_process_page(pq, m, queue); 3571 vm_pagequeue_unlock(pq); 3572 critical_exit(); 3573 } 3574 3575 /* 3576 * vm_page_pqbatch_drain: [ internal use only ] 3577 * 3578 * Force all per-CPU page queue batch queues to be drained. This is 3579 * intended for use in severe memory shortages, to ensure that pages 3580 * do not remain stuck in the batch queues. 3581 */ 3582 void 3583 vm_page_pqbatch_drain(void) 3584 { 3585 struct thread *td; 3586 struct vm_domain *vmd; 3587 struct vm_pagequeue *pq; 3588 int cpu, domain, queue; 3589 3590 td = curthread; 3591 CPU_FOREACH(cpu) { 3592 thread_lock(td); 3593 sched_bind(td, cpu); 3594 thread_unlock(td); 3595 3596 for (domain = 0; domain < vm_ndomains; domain++) { 3597 vmd = VM_DOMAIN(domain); 3598 for (queue = 0; queue < PQ_COUNT; queue++) { 3599 pq = &vmd->vmd_pagequeues[queue]; 3600 vm_pagequeue_lock(pq); 3601 critical_enter(); 3602 vm_pqbatch_process(pq, 3603 DPCPU_PTR(pqbatch[domain][queue]), queue); 3604 critical_exit(); 3605 vm_pagequeue_unlock(pq); 3606 } 3607 } 3608 } 3609 thread_lock(td); 3610 sched_unbind(td); 3611 thread_unlock(td); 3612 } 3613 3614 /* 3615 * vm_page_dequeue_deferred: [ internal use only ] 3616 * 3617 * Request removal of the given page from its current page 3618 * queue. Physical removal from the queue may be deferred 3619 * indefinitely. 3620 * 3621 * The page must be locked. 3622 */ 3623 void 3624 vm_page_dequeue_deferred(vm_page_t m) 3625 { 3626 vm_page_astate_t new, old; 3627 3628 old = vm_page_astate_load(m); 3629 do { 3630 if (old.queue == PQ_NONE) { 3631 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3632 ("%s: page %p has unexpected queue state", 3633 __func__, m)); 3634 break; 3635 } 3636 new = old; 3637 new.flags |= PGA_DEQUEUE; 3638 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3639 } 3640 3641 /* 3642 * vm_page_dequeue: 3643 * 3644 * Remove the page from whichever page queue it's in, if any, before 3645 * returning. 3646 */ 3647 void 3648 vm_page_dequeue(vm_page_t m) 3649 { 3650 vm_page_astate_t new, old; 3651 3652 old = vm_page_astate_load(m); 3653 do { 3654 if (old.queue == PQ_NONE) { 3655 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3656 ("%s: page %p has unexpected queue state", 3657 __func__, m)); 3658 break; 3659 } 3660 new = old; 3661 new.flags &= ~PGA_QUEUE_OP_MASK; 3662 new.queue = PQ_NONE; 3663 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 3664 3665 } 3666 3667 /* 3668 * Schedule the given page for insertion into the specified page queue. 3669 * Physical insertion of the page may be deferred indefinitely. 3670 */ 3671 static void 3672 vm_page_enqueue(vm_page_t m, uint8_t queue) 3673 { 3674 3675 KASSERT(m->a.queue == PQ_NONE && 3676 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 3677 ("%s: page %p is already enqueued", __func__, m)); 3678 KASSERT(m->ref_count > 0, 3679 ("%s: page %p does not carry any references", __func__, m)); 3680 3681 m->a.queue = queue; 3682 if ((m->a.flags & PGA_REQUEUE) == 0) 3683 vm_page_aflag_set(m, PGA_REQUEUE); 3684 vm_page_pqbatch_submit(m, queue); 3685 } 3686 3687 /* 3688 * vm_page_free_prep: 3689 * 3690 * Prepares the given page to be put on the free list, 3691 * disassociating it from any VM object. The caller may return 3692 * the page to the free list only if this function returns true. 3693 * 3694 * The object, if it exists, must be locked, and then the page must 3695 * be xbusy. Otherwise the page must be not busied. A managed 3696 * page must be unmapped. 3697 */ 3698 static bool 3699 vm_page_free_prep(vm_page_t m) 3700 { 3701 3702 /* 3703 * Synchronize with threads that have dropped a reference to this 3704 * page. 3705 */ 3706 atomic_thread_fence_acq(); 3707 3708 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3709 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3710 uint64_t *p; 3711 int i; 3712 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3713 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3714 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3715 m, i, (uintmax_t)*p)); 3716 } 3717 #endif 3718 if ((m->oflags & VPO_UNMANAGED) == 0) { 3719 KASSERT(!pmap_page_is_mapped(m), 3720 ("vm_page_free_prep: freeing mapped page %p", m)); 3721 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3722 ("vm_page_free_prep: mapping flags set in page %p", m)); 3723 } else { 3724 KASSERT(m->a.queue == PQ_NONE, 3725 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3726 } 3727 VM_CNT_INC(v_tfree); 3728 3729 if (m->object != NULL) { 3730 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 3731 ((m->object->flags & OBJ_UNMANAGED) != 0), 3732 ("vm_page_free_prep: managed flag mismatch for page %p", 3733 m)); 3734 vm_page_assert_xbusied(m); 3735 3736 /* 3737 * The object reference can be released without an atomic 3738 * operation. 3739 */ 3740 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3741 m->ref_count == VPRC_OBJREF, 3742 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3743 m, m->ref_count)); 3744 vm_page_object_remove(m); 3745 m->ref_count -= VPRC_OBJREF; 3746 } else 3747 vm_page_assert_unbusied(m); 3748 3749 vm_page_busy_free(m); 3750 3751 /* 3752 * If fictitious remove object association and 3753 * return. 3754 */ 3755 if ((m->flags & PG_FICTITIOUS) != 0) { 3756 KASSERT(m->ref_count == 1, 3757 ("fictitious page %p is referenced", m)); 3758 KASSERT(m->a.queue == PQ_NONE, 3759 ("fictitious page %p is queued", m)); 3760 return (false); 3761 } 3762 3763 /* 3764 * Pages need not be dequeued before they are returned to the physical 3765 * memory allocator, but they must at least be marked for a deferred 3766 * dequeue. 3767 */ 3768 if ((m->oflags & VPO_UNMANAGED) == 0) 3769 vm_page_dequeue_deferred(m); 3770 3771 m->valid = 0; 3772 vm_page_undirty(m); 3773 3774 if (m->ref_count != 0) 3775 panic("vm_page_free_prep: page %p has references", m); 3776 3777 /* 3778 * Restore the default memory attribute to the page. 3779 */ 3780 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3781 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3782 3783 #if VM_NRESERVLEVEL > 0 3784 /* 3785 * Determine whether the page belongs to a reservation. If the page was 3786 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3787 * as an optimization, we avoid the check in that case. 3788 */ 3789 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3790 return (false); 3791 #endif 3792 3793 return (true); 3794 } 3795 3796 /* 3797 * vm_page_free_toq: 3798 * 3799 * Returns the given page to the free list, disassociating it 3800 * from any VM object. 3801 * 3802 * The object must be locked. The page must be locked if it is 3803 * managed. 3804 */ 3805 static void 3806 vm_page_free_toq(vm_page_t m) 3807 { 3808 struct vm_domain *vmd; 3809 uma_zone_t zone; 3810 3811 if (!vm_page_free_prep(m)) 3812 return; 3813 3814 vmd = vm_pagequeue_domain(m); 3815 zone = vmd->vmd_pgcache[m->pool].zone; 3816 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 3817 uma_zfree(zone, m); 3818 return; 3819 } 3820 vm_domain_free_lock(vmd); 3821 vm_phys_free_pages(m, 0); 3822 vm_domain_free_unlock(vmd); 3823 vm_domain_freecnt_inc(vmd, 1); 3824 } 3825 3826 /* 3827 * vm_page_free_pages_toq: 3828 * 3829 * Returns a list of pages to the free list, disassociating it 3830 * from any VM object. In other words, this is equivalent to 3831 * calling vm_page_free_toq() for each page of a list of VM objects. 3832 * 3833 * The objects must be locked. The pages must be locked if it is 3834 * managed. 3835 */ 3836 void 3837 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3838 { 3839 vm_page_t m; 3840 int count; 3841 3842 if (SLIST_EMPTY(free)) 3843 return; 3844 3845 count = 0; 3846 while ((m = SLIST_FIRST(free)) != NULL) { 3847 count++; 3848 SLIST_REMOVE_HEAD(free, plinks.s.ss); 3849 vm_page_free_toq(m); 3850 } 3851 3852 if (update_wire_count) 3853 vm_wire_sub(count); 3854 } 3855 3856 /* 3857 * Mark this page as wired down, preventing reclamation by the page daemon 3858 * or when the containing object is destroyed. 3859 */ 3860 void 3861 vm_page_wire(vm_page_t m) 3862 { 3863 u_int old; 3864 3865 KASSERT(m->object != NULL, 3866 ("vm_page_wire: page %p does not belong to an object", m)); 3867 if (!vm_page_busied(m) && !vm_object_busied(m->object)) 3868 VM_OBJECT_ASSERT_LOCKED(m->object); 3869 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 3870 VPRC_WIRE_COUNT(m->ref_count) >= 1, 3871 ("vm_page_wire: fictitious page %p has zero wirings", m)); 3872 3873 old = atomic_fetchadd_int(&m->ref_count, 1); 3874 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 3875 ("vm_page_wire: counter overflow for page %p", m)); 3876 if (VPRC_WIRE_COUNT(old) == 0) { 3877 if ((m->oflags & VPO_UNMANAGED) == 0) 3878 vm_page_aflag_set(m, PGA_DEQUEUE); 3879 vm_wire_add(1); 3880 } 3881 } 3882 3883 /* 3884 * Attempt to wire a mapped page following a pmap lookup of that page. 3885 * This may fail if a thread is concurrently tearing down mappings of the page. 3886 * The transient failure is acceptable because it translates to the 3887 * failure of the caller pmap_extract_and_hold(), which should be then 3888 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 3889 */ 3890 bool 3891 vm_page_wire_mapped(vm_page_t m) 3892 { 3893 u_int old; 3894 3895 old = m->ref_count; 3896 do { 3897 KASSERT(old > 0, 3898 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 3899 if ((old & VPRC_BLOCKED) != 0) 3900 return (false); 3901 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 3902 3903 if (VPRC_WIRE_COUNT(old) == 0) { 3904 if ((m->oflags & VPO_UNMANAGED) == 0) 3905 vm_page_aflag_set(m, PGA_DEQUEUE); 3906 vm_wire_add(1); 3907 } 3908 return (true); 3909 } 3910 3911 /* 3912 * Release a wiring reference to a managed page. If the page still belongs to 3913 * an object, update its position in the page queues to reflect the reference. 3914 * If the wiring was the last reference to the page, free the page. 3915 */ 3916 static void 3917 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 3918 { 3919 u_int old; 3920 3921 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3922 ("%s: page %p is unmanaged", __func__, m)); 3923 3924 /* 3925 * Update LRU state before releasing the wiring reference. 3926 * Use a release store when updating the reference count to 3927 * synchronize with vm_page_free_prep(). 3928 */ 3929 old = m->ref_count; 3930 do { 3931 KASSERT(VPRC_WIRE_COUNT(old) > 0, 3932 ("vm_page_unwire: wire count underflow for page %p", m)); 3933 3934 if (old > VPRC_OBJREF + 1) { 3935 /* 3936 * The page has at least one other wiring reference. An 3937 * earlier iteration of this loop may have called 3938 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 3939 * re-set it if necessary. 3940 */ 3941 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 3942 vm_page_aflag_set(m, PGA_DEQUEUE); 3943 } else if (old == VPRC_OBJREF + 1) { 3944 /* 3945 * This is the last wiring. Clear PGA_DEQUEUE and 3946 * update the page's queue state to reflect the 3947 * reference. If the page does not belong to an object 3948 * (i.e., the VPRC_OBJREF bit is clear), we only need to 3949 * clear leftover queue state. 3950 */ 3951 vm_page_release_toq(m, nqueue, false); 3952 } else if (old == 1) { 3953 vm_page_aflag_clear(m, PGA_DEQUEUE); 3954 } 3955 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 3956 3957 if (VPRC_WIRE_COUNT(old) == 1) { 3958 vm_wire_sub(1); 3959 if (old == 1) 3960 vm_page_free(m); 3961 } 3962 } 3963 3964 /* 3965 * Release one wiring of the specified page, potentially allowing it to be 3966 * paged out. 3967 * 3968 * Only managed pages belonging to an object can be paged out. If the number 3969 * of wirings transitions to zero and the page is eligible for page out, then 3970 * the page is added to the specified paging queue. If the released wiring 3971 * represented the last reference to the page, the page is freed. 3972 * 3973 * A managed page must be locked. 3974 */ 3975 void 3976 vm_page_unwire(vm_page_t m, uint8_t nqueue) 3977 { 3978 3979 KASSERT(nqueue < PQ_COUNT, 3980 ("vm_page_unwire: invalid queue %u request for page %p", 3981 nqueue, m)); 3982 3983 if ((m->oflags & VPO_UNMANAGED) != 0) { 3984 if (vm_page_unwire_noq(m) && m->ref_count == 0) 3985 vm_page_free(m); 3986 return; 3987 } 3988 vm_page_unwire_managed(m, nqueue, false); 3989 } 3990 3991 /* 3992 * Unwire a page without (re-)inserting it into a page queue. It is up 3993 * to the caller to enqueue, requeue, or free the page as appropriate. 3994 * In most cases involving managed pages, vm_page_unwire() should be used 3995 * instead. 3996 */ 3997 bool 3998 vm_page_unwire_noq(vm_page_t m) 3999 { 4000 u_int old; 4001 4002 old = vm_page_drop(m, 1); 4003 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4004 ("vm_page_unref: counter underflow for page %p", m)); 4005 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4006 ("vm_page_unref: missing ref on fictitious page %p", m)); 4007 4008 if (VPRC_WIRE_COUNT(old) > 1) 4009 return (false); 4010 if ((m->oflags & VPO_UNMANAGED) == 0) 4011 vm_page_aflag_clear(m, PGA_DEQUEUE); 4012 vm_wire_sub(1); 4013 return (true); 4014 } 4015 4016 /* 4017 * Ensure that the page ends up in the specified page queue. If the page is 4018 * active or being moved to the active queue, ensure that its act_count is 4019 * at least ACT_INIT but do not otherwise mess with it. 4020 * 4021 * A managed page must be locked. 4022 */ 4023 static __always_inline void 4024 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4025 { 4026 vm_page_astate_t old, new; 4027 4028 KASSERT(m->ref_count > 0, 4029 ("%s: page %p does not carry any references", __func__, m)); 4030 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4031 ("%s: invalid flags %x", __func__, nflag)); 4032 4033 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4034 return; 4035 4036 old = vm_page_astate_load(m); 4037 do { 4038 if ((old.flags & PGA_DEQUEUE) != 0) 4039 break; 4040 new = old; 4041 new.flags &= ~PGA_QUEUE_OP_MASK; 4042 if (nqueue == PQ_ACTIVE) 4043 new.act_count = max(old.act_count, ACT_INIT); 4044 if (old.queue == nqueue) { 4045 if (nqueue != PQ_ACTIVE) 4046 new.flags |= nflag; 4047 } else { 4048 new.flags |= nflag; 4049 new.queue = nqueue; 4050 } 4051 } while (!vm_page_pqstate_commit(m, &old, new)); 4052 } 4053 4054 /* 4055 * Put the specified page on the active list (if appropriate). 4056 */ 4057 void 4058 vm_page_activate(vm_page_t m) 4059 { 4060 4061 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4062 } 4063 4064 /* 4065 * Move the specified page to the tail of the inactive queue, or requeue 4066 * the page if it is already in the inactive queue. 4067 */ 4068 void 4069 vm_page_deactivate(vm_page_t m) 4070 { 4071 4072 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4073 } 4074 4075 void 4076 vm_page_deactivate_noreuse(vm_page_t m) 4077 { 4078 4079 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4080 } 4081 4082 /* 4083 * Put a page in the laundry, or requeue it if it is already there. 4084 */ 4085 void 4086 vm_page_launder(vm_page_t m) 4087 { 4088 4089 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4090 } 4091 4092 /* 4093 * Put a page in the PQ_UNSWAPPABLE holding queue. 4094 */ 4095 void 4096 vm_page_unswappable(vm_page_t m) 4097 { 4098 4099 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 4100 ("page %p already unswappable", m)); 4101 4102 vm_page_dequeue(m); 4103 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4104 } 4105 4106 /* 4107 * Release a page back to the page queues in preparation for unwiring. 4108 */ 4109 static void 4110 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4111 { 4112 vm_page_astate_t old, new; 4113 uint16_t nflag; 4114 4115 /* 4116 * Use a check of the valid bits to determine whether we should 4117 * accelerate reclamation of the page. The object lock might not be 4118 * held here, in which case the check is racy. At worst we will either 4119 * accelerate reclamation of a valid page and violate LRU, or 4120 * unnecessarily defer reclamation of an invalid page. 4121 * 4122 * If we were asked to not cache the page, place it near the head of the 4123 * inactive queue so that is reclaimed sooner. 4124 */ 4125 if (noreuse || m->valid == 0) { 4126 nqueue = PQ_INACTIVE; 4127 nflag = PGA_REQUEUE_HEAD; 4128 } else { 4129 nflag = PGA_REQUEUE; 4130 } 4131 4132 old = vm_page_astate_load(m); 4133 do { 4134 new = old; 4135 4136 /* 4137 * If the page is already in the active queue and we are not 4138 * trying to accelerate reclamation, simply mark it as 4139 * referenced and avoid any queue operations. 4140 */ 4141 new.flags &= ~PGA_QUEUE_OP_MASK; 4142 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE) 4143 new.flags |= PGA_REFERENCED; 4144 else { 4145 new.flags |= nflag; 4146 new.queue = nqueue; 4147 } 4148 } while (!vm_page_pqstate_commit(m, &old, new)); 4149 } 4150 4151 /* 4152 * Unwire a page and either attempt to free it or re-add it to the page queues. 4153 */ 4154 void 4155 vm_page_release(vm_page_t m, int flags) 4156 { 4157 vm_object_t object; 4158 4159 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4160 ("vm_page_release: page %p is unmanaged", m)); 4161 4162 if ((flags & VPR_TRYFREE) != 0) { 4163 for (;;) { 4164 object = atomic_load_ptr(&m->object); 4165 if (object == NULL) 4166 break; 4167 /* Depends on type-stability. */ 4168 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4169 break; 4170 if (object == m->object) { 4171 vm_page_release_locked(m, flags); 4172 VM_OBJECT_WUNLOCK(object); 4173 return; 4174 } 4175 VM_OBJECT_WUNLOCK(object); 4176 } 4177 } 4178 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4179 } 4180 4181 /* See vm_page_release(). */ 4182 void 4183 vm_page_release_locked(vm_page_t m, int flags) 4184 { 4185 4186 VM_OBJECT_ASSERT_WLOCKED(m->object); 4187 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4188 ("vm_page_release_locked: page %p is unmanaged", m)); 4189 4190 if (vm_page_unwire_noq(m)) { 4191 if ((flags & VPR_TRYFREE) != 0 && 4192 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4193 m->dirty == 0 && vm_page_tryxbusy(m)) { 4194 /* 4195 * An unlocked lookup may have wired the page before the 4196 * busy lock was acquired, in which case the page must 4197 * not be freed. 4198 */ 4199 if (__predict_true(!vm_page_wired(m))) { 4200 vm_page_free(m); 4201 return; 4202 } 4203 vm_page_xunbusy(m); 4204 } else { 4205 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4206 } 4207 } 4208 } 4209 4210 static bool 4211 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4212 { 4213 u_int old; 4214 4215 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4216 ("vm_page_try_blocked_op: page %p has no object", m)); 4217 KASSERT(vm_page_busied(m), 4218 ("vm_page_try_blocked_op: page %p is not busy", m)); 4219 VM_OBJECT_ASSERT_LOCKED(m->object); 4220 4221 old = m->ref_count; 4222 do { 4223 KASSERT(old != 0, 4224 ("vm_page_try_blocked_op: page %p has no references", m)); 4225 if (VPRC_WIRE_COUNT(old) != 0) 4226 return (false); 4227 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4228 4229 (op)(m); 4230 4231 /* 4232 * If the object is read-locked, new wirings may be created via an 4233 * object lookup. 4234 */ 4235 old = vm_page_drop(m, VPRC_BLOCKED); 4236 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4237 old == (VPRC_BLOCKED | VPRC_OBJREF), 4238 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4239 old, m)); 4240 return (true); 4241 } 4242 4243 /* 4244 * Atomically check for wirings and remove all mappings of the page. 4245 */ 4246 bool 4247 vm_page_try_remove_all(vm_page_t m) 4248 { 4249 4250 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4251 } 4252 4253 /* 4254 * Atomically check for wirings and remove all writeable mappings of the page. 4255 */ 4256 bool 4257 vm_page_try_remove_write(vm_page_t m) 4258 { 4259 4260 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4261 } 4262 4263 /* 4264 * vm_page_advise 4265 * 4266 * Apply the specified advice to the given page. 4267 * 4268 * The object and page must be locked. 4269 */ 4270 void 4271 vm_page_advise(vm_page_t m, int advice) 4272 { 4273 4274 VM_OBJECT_ASSERT_WLOCKED(m->object); 4275 if (advice == MADV_FREE) 4276 /* 4277 * Mark the page clean. This will allow the page to be freed 4278 * without first paging it out. MADV_FREE pages are often 4279 * quickly reused by malloc(3), so we do not do anything that 4280 * would result in a page fault on a later access. 4281 */ 4282 vm_page_undirty(m); 4283 else if (advice != MADV_DONTNEED) { 4284 if (advice == MADV_WILLNEED) 4285 vm_page_activate(m); 4286 return; 4287 } 4288 4289 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4290 vm_page_dirty(m); 4291 4292 /* 4293 * Clear any references to the page. Otherwise, the page daemon will 4294 * immediately reactivate the page. 4295 */ 4296 vm_page_aflag_clear(m, PGA_REFERENCED); 4297 4298 /* 4299 * Place clean pages near the head of the inactive queue rather than 4300 * the tail, thus defeating the queue's LRU operation and ensuring that 4301 * the page will be reused quickly. Dirty pages not already in the 4302 * laundry are moved there. 4303 */ 4304 if (m->dirty == 0) 4305 vm_page_deactivate_noreuse(m); 4306 else if (!vm_page_in_laundry(m)) 4307 vm_page_launder(m); 4308 } 4309 4310 /* 4311 * vm_page_grab_release 4312 * 4313 * Helper routine for grab functions to release busy on return. 4314 */ 4315 static inline void 4316 vm_page_grab_release(vm_page_t m, int allocflags) 4317 { 4318 4319 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4320 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4321 vm_page_sunbusy(m); 4322 else 4323 vm_page_xunbusy(m); 4324 } 4325 } 4326 4327 /* 4328 * vm_page_grab_sleep 4329 * 4330 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4331 * if the caller should retry and false otherwise. 4332 * 4333 * If the object is locked on entry the object will be unlocked with 4334 * false returns and still locked but possibly having been dropped 4335 * with true returns. 4336 */ 4337 static bool 4338 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4339 const char *wmesg, int allocflags, bool locked) 4340 { 4341 4342 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4343 return (false); 4344 4345 /* 4346 * Reference the page before unlocking and sleeping so that 4347 * the page daemon is less likely to reclaim it. 4348 */ 4349 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4350 vm_page_reference(m); 4351 4352 if (_vm_page_busy_sleep(object, m, m->pindex, wmesg, allocflags, 4353 locked) && locked) 4354 VM_OBJECT_WLOCK(object); 4355 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4356 return (false); 4357 4358 return (true); 4359 } 4360 4361 /* 4362 * Assert that the grab flags are valid. 4363 */ 4364 static inline void 4365 vm_page_grab_check(int allocflags) 4366 { 4367 4368 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4369 (allocflags & VM_ALLOC_WIRED) != 0, 4370 ("vm_page_grab*: the pages must be busied or wired")); 4371 4372 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4373 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4374 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4375 } 4376 4377 /* 4378 * Calculate the page allocation flags for grab. 4379 */ 4380 static inline int 4381 vm_page_grab_pflags(int allocflags) 4382 { 4383 int pflags; 4384 4385 pflags = allocflags & 4386 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4387 VM_ALLOC_NOBUSY); 4388 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4389 pflags |= VM_ALLOC_WAITFAIL; 4390 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4391 pflags |= VM_ALLOC_SBUSY; 4392 4393 return (pflags); 4394 } 4395 4396 /* 4397 * Grab a page, waiting until we are waken up due to the page 4398 * changing state. We keep on waiting, if the page continues 4399 * to be in the object. If the page doesn't exist, first allocate it 4400 * and then conditionally zero it. 4401 * 4402 * This routine may sleep. 4403 * 4404 * The object must be locked on entry. The lock will, however, be released 4405 * and reacquired if the routine sleeps. 4406 */ 4407 vm_page_t 4408 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4409 { 4410 vm_page_t m; 4411 4412 VM_OBJECT_ASSERT_WLOCKED(object); 4413 vm_page_grab_check(allocflags); 4414 4415 retrylookup: 4416 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4417 if (!vm_page_tryacquire(m, allocflags)) { 4418 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4419 allocflags, true)) 4420 goto retrylookup; 4421 return (NULL); 4422 } 4423 goto out; 4424 } 4425 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4426 return (NULL); 4427 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); 4428 if (m == NULL) { 4429 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4430 return (NULL); 4431 goto retrylookup; 4432 } 4433 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4434 pmap_zero_page(m); 4435 4436 out: 4437 vm_page_grab_release(m, allocflags); 4438 4439 return (m); 4440 } 4441 4442 /* 4443 * Locklessly attempt to acquire a page given a (object, pindex) tuple 4444 * and an optional previous page to avoid the radix lookup. The resulting 4445 * page will be validated against the identity tuple and busied or wired 4446 * as requested. A NULL *mp return guarantees that the page was not in 4447 * radix at the time of the call but callers must perform higher level 4448 * synchronization or retry the operation under a lock if they require 4449 * an atomic answer. This is the only lock free validation routine, 4450 * other routines can depend on the resulting page state. 4451 * 4452 * The return value indicates whether the operation failed due to caller 4453 * flags. The return is tri-state with mp: 4454 * 4455 * (true, *mp != NULL) - The operation was successful. 4456 * (true, *mp == NULL) - The page was not found in tree. 4457 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition. 4458 */ 4459 static bool 4460 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, 4461 vm_page_t prev, vm_page_t *mp, int allocflags) 4462 { 4463 vm_page_t m; 4464 4465 vm_page_grab_check(allocflags); 4466 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev)); 4467 4468 *mp = NULL; 4469 for (;;) { 4470 /* 4471 * We may see a false NULL here because the previous page 4472 * has been removed or just inserted and the list is loaded 4473 * without barriers. Switch to radix to verify. 4474 */ 4475 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL || 4476 QMD_IS_TRASHED(m) || m->pindex != pindex || 4477 atomic_load_ptr(&m->object) != object) { 4478 prev = NULL; 4479 /* 4480 * This guarantees the result is instantaneously 4481 * correct. 4482 */ 4483 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 4484 } 4485 if (m == NULL) 4486 return (true); 4487 if (vm_page_trybusy(m, allocflags)) { 4488 if (m->object == object && m->pindex == pindex) 4489 break; 4490 /* relookup. */ 4491 vm_page_busy_release(m); 4492 cpu_spinwait(); 4493 continue; 4494 } 4495 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4496 allocflags, false)) 4497 return (false); 4498 } 4499 if ((allocflags & VM_ALLOC_WIRED) != 0) 4500 vm_page_wire(m); 4501 vm_page_grab_release(m, allocflags); 4502 *mp = m; 4503 return (true); 4504 } 4505 4506 /* 4507 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4508 * is not set. 4509 */ 4510 vm_page_t 4511 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4512 { 4513 vm_page_t m; 4514 4515 vm_page_grab_check(allocflags); 4516 4517 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags)) 4518 return (NULL); 4519 if (m != NULL) 4520 return (m); 4521 4522 /* 4523 * The radix lockless lookup should never return a false negative 4524 * errors. If the user specifies NOCREAT they are guaranteed there 4525 * was no page present at the instant of the call. A NOCREAT caller 4526 * must handle create races gracefully. 4527 */ 4528 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4529 return (NULL); 4530 4531 VM_OBJECT_WLOCK(object); 4532 m = vm_page_grab(object, pindex, allocflags); 4533 VM_OBJECT_WUNLOCK(object); 4534 4535 return (m); 4536 } 4537 4538 /* 4539 * Grab a page and make it valid, paging in if necessary. Pages missing from 4540 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4541 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4542 * in simultaneously. Additional pages will be left on a paging queue but 4543 * will neither be wired nor busy regardless of allocflags. 4544 */ 4545 int 4546 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4547 { 4548 vm_page_t m; 4549 vm_page_t ma[VM_INITIAL_PAGEIN]; 4550 int after, i, pflags, rv; 4551 4552 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4553 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4554 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4555 KASSERT((allocflags & 4556 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4557 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4558 VM_OBJECT_ASSERT_WLOCKED(object); 4559 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4560 VM_ALLOC_WIRED); 4561 pflags |= VM_ALLOC_WAITFAIL; 4562 4563 retrylookup: 4564 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4565 /* 4566 * If the page is fully valid it can only become invalid 4567 * with the object lock held. If it is not valid it can 4568 * become valid with the busy lock held. Therefore, we 4569 * may unnecessarily lock the exclusive busy here if we 4570 * race with I/O completion not using the object lock. 4571 * However, we will not end up with an invalid page and a 4572 * shared lock. 4573 */ 4574 if (!vm_page_trybusy(m, 4575 vm_page_all_valid(m) ? allocflags : 0)) { 4576 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4577 allocflags, true); 4578 goto retrylookup; 4579 } 4580 if (vm_page_all_valid(m)) 4581 goto out; 4582 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4583 vm_page_busy_release(m); 4584 *mp = NULL; 4585 return (VM_PAGER_FAIL); 4586 } 4587 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4588 *mp = NULL; 4589 return (VM_PAGER_FAIL); 4590 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { 4591 goto retrylookup; 4592 } 4593 4594 vm_page_assert_xbusied(m); 4595 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4596 after = MIN(after, VM_INITIAL_PAGEIN); 4597 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4598 after = MAX(after, 1); 4599 ma[0] = m; 4600 for (i = 1; i < after; i++) { 4601 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 4602 if (ma[i]->valid || !vm_page_tryxbusy(ma[i])) 4603 break; 4604 } else { 4605 ma[i] = vm_page_alloc(object, m->pindex + i, 4606 VM_ALLOC_NORMAL); 4607 if (ma[i] == NULL) 4608 break; 4609 } 4610 } 4611 after = i; 4612 vm_object_pip_add(object, after); 4613 VM_OBJECT_WUNLOCK(object); 4614 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 4615 VM_OBJECT_WLOCK(object); 4616 vm_object_pip_wakeupn(object, after); 4617 /* Pager may have replaced a page. */ 4618 m = ma[0]; 4619 if (rv != VM_PAGER_OK) { 4620 for (i = 0; i < after; i++) { 4621 if (!vm_page_wired(ma[i])) 4622 vm_page_free(ma[i]); 4623 else 4624 vm_page_xunbusy(ma[i]); 4625 } 4626 *mp = NULL; 4627 return (rv); 4628 } 4629 for (i = 1; i < after; i++) 4630 vm_page_readahead_finish(ma[i]); 4631 MPASS(vm_page_all_valid(m)); 4632 } else { 4633 vm_page_zero_invalid(m, TRUE); 4634 } 4635 out: 4636 if ((allocflags & VM_ALLOC_WIRED) != 0) 4637 vm_page_wire(m); 4638 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 4639 vm_page_busy_downgrade(m); 4640 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 4641 vm_page_busy_release(m); 4642 *mp = m; 4643 return (VM_PAGER_OK); 4644 } 4645 4646 /* 4647 * Locklessly grab a valid page. If the page is not valid or not yet 4648 * allocated this will fall back to the object lock method. 4649 */ 4650 int 4651 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 4652 vm_pindex_t pindex, int allocflags) 4653 { 4654 vm_page_t m; 4655 int flags; 4656 int error; 4657 4658 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4659 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4660 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 4661 "mismatch")); 4662 KASSERT((allocflags & 4663 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4664 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 4665 4666 /* 4667 * Attempt a lockless lookup and busy. We need at least an sbusy 4668 * before we can inspect the valid field and return a wired page. 4669 */ 4670 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 4671 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags)) 4672 return (VM_PAGER_FAIL); 4673 if ((m = *mp) != NULL) { 4674 if (vm_page_all_valid(m)) { 4675 if ((allocflags & VM_ALLOC_WIRED) != 0) 4676 vm_page_wire(m); 4677 vm_page_grab_release(m, allocflags); 4678 return (VM_PAGER_OK); 4679 } 4680 vm_page_busy_release(m); 4681 } 4682 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4683 *mp = NULL; 4684 return (VM_PAGER_FAIL); 4685 } 4686 VM_OBJECT_WLOCK(object); 4687 error = vm_page_grab_valid(mp, object, pindex, allocflags); 4688 VM_OBJECT_WUNLOCK(object); 4689 4690 return (error); 4691 } 4692 4693 /* 4694 * Return the specified range of pages from the given object. For each 4695 * page offset within the range, if a page already exists within the object 4696 * at that offset and it is busy, then wait for it to change state. If, 4697 * instead, the page doesn't exist, then allocate it. 4698 * 4699 * The caller must always specify an allocation class. 4700 * 4701 * allocation classes: 4702 * VM_ALLOC_NORMAL normal process request 4703 * VM_ALLOC_SYSTEM system *really* needs the pages 4704 * 4705 * The caller must always specify that the pages are to be busied and/or 4706 * wired. 4707 * 4708 * optional allocation flags: 4709 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4710 * VM_ALLOC_NOBUSY do not exclusive busy the page 4711 * VM_ALLOC_NOWAIT do not sleep 4712 * VM_ALLOC_SBUSY set page to sbusy state 4713 * VM_ALLOC_WIRED wire the pages 4714 * VM_ALLOC_ZERO zero and validate any invalid pages 4715 * 4716 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4717 * may return a partial prefix of the requested range. 4718 */ 4719 int 4720 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4721 vm_page_t *ma, int count) 4722 { 4723 vm_page_t m, mpred; 4724 int pflags; 4725 int i; 4726 4727 VM_OBJECT_ASSERT_WLOCKED(object); 4728 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4729 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4730 vm_page_grab_check(allocflags); 4731 4732 pflags = vm_page_grab_pflags(allocflags); 4733 if (count == 0) 4734 return (0); 4735 4736 i = 0; 4737 retrylookup: 4738 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4739 if (m == NULL || m->pindex != pindex + i) { 4740 mpred = m; 4741 m = NULL; 4742 } else 4743 mpred = TAILQ_PREV(m, pglist, listq); 4744 for (; i < count; i++) { 4745 if (m != NULL) { 4746 if (!vm_page_tryacquire(m, allocflags)) { 4747 if (vm_page_grab_sleep(object, m, pindex, 4748 "grbmaw", allocflags, true)) 4749 goto retrylookup; 4750 break; 4751 } 4752 } else { 4753 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4754 break; 4755 m = vm_page_alloc_after(object, pindex + i, 4756 pflags | VM_ALLOC_COUNT(count - i), mpred); 4757 if (m == NULL) { 4758 if ((allocflags & (VM_ALLOC_NOWAIT | 4759 VM_ALLOC_WAITFAIL)) != 0) 4760 break; 4761 goto retrylookup; 4762 } 4763 } 4764 if (vm_page_none_valid(m) && 4765 (allocflags & VM_ALLOC_ZERO) != 0) { 4766 if ((m->flags & PG_ZERO) == 0) 4767 pmap_zero_page(m); 4768 vm_page_valid(m); 4769 } 4770 vm_page_grab_release(m, allocflags); 4771 ma[i] = mpred = m; 4772 m = vm_page_next(m); 4773 } 4774 return (i); 4775 } 4776 4777 /* 4778 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 4779 * and will fall back to the locked variant to handle allocation. 4780 */ 4781 int 4782 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 4783 int allocflags, vm_page_t *ma, int count) 4784 { 4785 vm_page_t m, pred; 4786 int flags; 4787 int i; 4788 4789 vm_page_grab_check(allocflags); 4790 4791 /* 4792 * Modify flags for lockless acquire to hold the page until we 4793 * set it valid if necessary. 4794 */ 4795 flags = allocflags & ~VM_ALLOC_NOBUSY; 4796 pred = NULL; 4797 for (i = 0; i < count; i++, pindex++) { 4798 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags)) 4799 return (i); 4800 if (m == NULL) 4801 break; 4802 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 4803 if ((m->flags & PG_ZERO) == 0) 4804 pmap_zero_page(m); 4805 vm_page_valid(m); 4806 } 4807 /* m will still be wired or busy according to flags. */ 4808 vm_page_grab_release(m, allocflags); 4809 pred = ma[i] = m; 4810 } 4811 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4812 return (i); 4813 count -= i; 4814 VM_OBJECT_WLOCK(object); 4815 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 4816 VM_OBJECT_WUNLOCK(object); 4817 4818 return (i); 4819 } 4820 4821 /* 4822 * Mapping function for valid or dirty bits in a page. 4823 * 4824 * Inputs are required to range within a page. 4825 */ 4826 vm_page_bits_t 4827 vm_page_bits(int base, int size) 4828 { 4829 int first_bit; 4830 int last_bit; 4831 4832 KASSERT( 4833 base + size <= PAGE_SIZE, 4834 ("vm_page_bits: illegal base/size %d/%d", base, size) 4835 ); 4836 4837 if (size == 0) /* handle degenerate case */ 4838 return (0); 4839 4840 first_bit = base >> DEV_BSHIFT; 4841 last_bit = (base + size - 1) >> DEV_BSHIFT; 4842 4843 return (((vm_page_bits_t)2 << last_bit) - 4844 ((vm_page_bits_t)1 << first_bit)); 4845 } 4846 4847 void 4848 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 4849 { 4850 4851 #if PAGE_SIZE == 32768 4852 atomic_set_64((uint64_t *)bits, set); 4853 #elif PAGE_SIZE == 16384 4854 atomic_set_32((uint32_t *)bits, set); 4855 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 4856 atomic_set_16((uint16_t *)bits, set); 4857 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 4858 atomic_set_8((uint8_t *)bits, set); 4859 #else /* PAGE_SIZE <= 8192 */ 4860 uintptr_t addr; 4861 int shift; 4862 4863 addr = (uintptr_t)bits; 4864 /* 4865 * Use a trick to perform a 32-bit atomic on the 4866 * containing aligned word, to not depend on the existence 4867 * of atomic_{set, clear}_{8, 16}. 4868 */ 4869 shift = addr & (sizeof(uint32_t) - 1); 4870 #if BYTE_ORDER == BIG_ENDIAN 4871 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4872 #else 4873 shift *= NBBY; 4874 #endif 4875 addr &= ~(sizeof(uint32_t) - 1); 4876 atomic_set_32((uint32_t *)addr, set << shift); 4877 #endif /* PAGE_SIZE */ 4878 } 4879 4880 static inline void 4881 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 4882 { 4883 4884 #if PAGE_SIZE == 32768 4885 atomic_clear_64((uint64_t *)bits, clear); 4886 #elif PAGE_SIZE == 16384 4887 atomic_clear_32((uint32_t *)bits, clear); 4888 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 4889 atomic_clear_16((uint16_t *)bits, clear); 4890 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 4891 atomic_clear_8((uint8_t *)bits, clear); 4892 #else /* PAGE_SIZE <= 8192 */ 4893 uintptr_t addr; 4894 int shift; 4895 4896 addr = (uintptr_t)bits; 4897 /* 4898 * Use a trick to perform a 32-bit atomic on the 4899 * containing aligned word, to not depend on the existence 4900 * of atomic_{set, clear}_{8, 16}. 4901 */ 4902 shift = addr & (sizeof(uint32_t) - 1); 4903 #if BYTE_ORDER == BIG_ENDIAN 4904 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4905 #else 4906 shift *= NBBY; 4907 #endif 4908 addr &= ~(sizeof(uint32_t) - 1); 4909 atomic_clear_32((uint32_t *)addr, clear << shift); 4910 #endif /* PAGE_SIZE */ 4911 } 4912 4913 static inline vm_page_bits_t 4914 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 4915 { 4916 #if PAGE_SIZE == 32768 4917 uint64_t old; 4918 4919 old = *bits; 4920 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 4921 return (old); 4922 #elif PAGE_SIZE == 16384 4923 uint32_t old; 4924 4925 old = *bits; 4926 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 4927 return (old); 4928 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 4929 uint16_t old; 4930 4931 old = *bits; 4932 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 4933 return (old); 4934 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 4935 uint8_t old; 4936 4937 old = *bits; 4938 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 4939 return (old); 4940 #else /* PAGE_SIZE <= 4096*/ 4941 uintptr_t addr; 4942 uint32_t old, new, mask; 4943 int shift; 4944 4945 addr = (uintptr_t)bits; 4946 /* 4947 * Use a trick to perform a 32-bit atomic on the 4948 * containing aligned word, to not depend on the existence 4949 * of atomic_{set, swap, clear}_{8, 16}. 4950 */ 4951 shift = addr & (sizeof(uint32_t) - 1); 4952 #if BYTE_ORDER == BIG_ENDIAN 4953 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4954 #else 4955 shift *= NBBY; 4956 #endif 4957 addr &= ~(sizeof(uint32_t) - 1); 4958 mask = VM_PAGE_BITS_ALL << shift; 4959 4960 old = *bits; 4961 do { 4962 new = old & ~mask; 4963 new |= newbits << shift; 4964 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 4965 return (old >> shift); 4966 #endif /* PAGE_SIZE */ 4967 } 4968 4969 /* 4970 * vm_page_set_valid_range: 4971 * 4972 * Sets portions of a page valid. The arguments are expected 4973 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4974 * of any partial chunks touched by the range. The invalid portion of 4975 * such chunks will be zeroed. 4976 * 4977 * (base + size) must be less then or equal to PAGE_SIZE. 4978 */ 4979 void 4980 vm_page_set_valid_range(vm_page_t m, int base, int size) 4981 { 4982 int endoff, frag; 4983 vm_page_bits_t pagebits; 4984 4985 vm_page_assert_busied(m); 4986 if (size == 0) /* handle degenerate case */ 4987 return; 4988 4989 /* 4990 * If the base is not DEV_BSIZE aligned and the valid 4991 * bit is clear, we have to zero out a portion of the 4992 * first block. 4993 */ 4994 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4995 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 4996 pmap_zero_page_area(m, frag, base - frag); 4997 4998 /* 4999 * If the ending offset is not DEV_BSIZE aligned and the 5000 * valid bit is clear, we have to zero out a portion of 5001 * the last block. 5002 */ 5003 endoff = base + size; 5004 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5005 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5006 pmap_zero_page_area(m, endoff, 5007 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5008 5009 /* 5010 * Assert that no previously invalid block that is now being validated 5011 * is already dirty. 5012 */ 5013 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5014 ("vm_page_set_valid_range: page %p is dirty", m)); 5015 5016 /* 5017 * Set valid bits inclusive of any overlap. 5018 */ 5019 pagebits = vm_page_bits(base, size); 5020 if (vm_page_xbusied(m)) 5021 m->valid |= pagebits; 5022 else 5023 vm_page_bits_set(m, &m->valid, pagebits); 5024 } 5025 5026 /* 5027 * Set the page dirty bits and free the invalid swap space if 5028 * present. Returns the previous dirty bits. 5029 */ 5030 vm_page_bits_t 5031 vm_page_set_dirty(vm_page_t m) 5032 { 5033 vm_page_bits_t old; 5034 5035 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5036 5037 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5038 old = m->dirty; 5039 m->dirty = VM_PAGE_BITS_ALL; 5040 } else 5041 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5042 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5043 vm_pager_page_unswapped(m); 5044 5045 return (old); 5046 } 5047 5048 /* 5049 * Clear the given bits from the specified page's dirty field. 5050 */ 5051 static __inline void 5052 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5053 { 5054 5055 vm_page_assert_busied(m); 5056 5057 /* 5058 * If the page is xbusied and not write mapped we are the 5059 * only thread that can modify dirty bits. Otherwise, The pmap 5060 * layer can call vm_page_dirty() without holding a distinguished 5061 * lock. The combination of page busy and atomic operations 5062 * suffice to guarantee consistency of the page dirty field. 5063 */ 5064 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5065 m->dirty &= ~pagebits; 5066 else 5067 vm_page_bits_clear(m, &m->dirty, pagebits); 5068 } 5069 5070 /* 5071 * vm_page_set_validclean: 5072 * 5073 * Sets portions of a page valid and clean. The arguments are expected 5074 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5075 * of any partial chunks touched by the range. The invalid portion of 5076 * such chunks will be zero'd. 5077 * 5078 * (base + size) must be less then or equal to PAGE_SIZE. 5079 */ 5080 void 5081 vm_page_set_validclean(vm_page_t m, int base, int size) 5082 { 5083 vm_page_bits_t oldvalid, pagebits; 5084 int endoff, frag; 5085 5086 vm_page_assert_busied(m); 5087 if (size == 0) /* handle degenerate case */ 5088 return; 5089 5090 /* 5091 * If the base is not DEV_BSIZE aligned and the valid 5092 * bit is clear, we have to zero out a portion of the 5093 * first block. 5094 */ 5095 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5096 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5097 pmap_zero_page_area(m, frag, base - frag); 5098 5099 /* 5100 * If the ending offset is not DEV_BSIZE aligned and the 5101 * valid bit is clear, we have to zero out a portion of 5102 * the last block. 5103 */ 5104 endoff = base + size; 5105 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5106 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5107 pmap_zero_page_area(m, endoff, 5108 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5109 5110 /* 5111 * Set valid, clear dirty bits. If validating the entire 5112 * page we can safely clear the pmap modify bit. We also 5113 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5114 * takes a write fault on a MAP_NOSYNC memory area the flag will 5115 * be set again. 5116 * 5117 * We set valid bits inclusive of any overlap, but we can only 5118 * clear dirty bits for DEV_BSIZE chunks that are fully within 5119 * the range. 5120 */ 5121 oldvalid = m->valid; 5122 pagebits = vm_page_bits(base, size); 5123 if (vm_page_xbusied(m)) 5124 m->valid |= pagebits; 5125 else 5126 vm_page_bits_set(m, &m->valid, pagebits); 5127 #if 0 /* NOT YET */ 5128 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5129 frag = DEV_BSIZE - frag; 5130 base += frag; 5131 size -= frag; 5132 if (size < 0) 5133 size = 0; 5134 } 5135 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5136 #endif 5137 if (base == 0 && size == PAGE_SIZE) { 5138 /* 5139 * The page can only be modified within the pmap if it is 5140 * mapped, and it can only be mapped if it was previously 5141 * fully valid. 5142 */ 5143 if (oldvalid == VM_PAGE_BITS_ALL) 5144 /* 5145 * Perform the pmap_clear_modify() first. Otherwise, 5146 * a concurrent pmap operation, such as 5147 * pmap_protect(), could clear a modification in the 5148 * pmap and set the dirty field on the page before 5149 * pmap_clear_modify() had begun and after the dirty 5150 * field was cleared here. 5151 */ 5152 pmap_clear_modify(m); 5153 m->dirty = 0; 5154 vm_page_aflag_clear(m, PGA_NOSYNC); 5155 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5156 m->dirty &= ~pagebits; 5157 else 5158 vm_page_clear_dirty_mask(m, pagebits); 5159 } 5160 5161 void 5162 vm_page_clear_dirty(vm_page_t m, int base, int size) 5163 { 5164 5165 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5166 } 5167 5168 /* 5169 * vm_page_set_invalid: 5170 * 5171 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5172 * valid and dirty bits for the effected areas are cleared. 5173 */ 5174 void 5175 vm_page_set_invalid(vm_page_t m, int base, int size) 5176 { 5177 vm_page_bits_t bits; 5178 vm_object_t object; 5179 5180 /* 5181 * The object lock is required so that pages can't be mapped 5182 * read-only while we're in the process of invalidating them. 5183 */ 5184 object = m->object; 5185 VM_OBJECT_ASSERT_WLOCKED(object); 5186 vm_page_assert_busied(m); 5187 5188 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5189 size >= object->un_pager.vnp.vnp_size) 5190 bits = VM_PAGE_BITS_ALL; 5191 else 5192 bits = vm_page_bits(base, size); 5193 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5194 pmap_remove_all(m); 5195 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5196 !pmap_page_is_mapped(m), 5197 ("vm_page_set_invalid: page %p is mapped", m)); 5198 if (vm_page_xbusied(m)) { 5199 m->valid &= ~bits; 5200 m->dirty &= ~bits; 5201 } else { 5202 vm_page_bits_clear(m, &m->valid, bits); 5203 vm_page_bits_clear(m, &m->dirty, bits); 5204 } 5205 } 5206 5207 /* 5208 * vm_page_invalid: 5209 * 5210 * Invalidates the entire page. The page must be busy, unmapped, and 5211 * the enclosing object must be locked. The object locks protects 5212 * against concurrent read-only pmap enter which is done without 5213 * busy. 5214 */ 5215 void 5216 vm_page_invalid(vm_page_t m) 5217 { 5218 5219 vm_page_assert_busied(m); 5220 VM_OBJECT_ASSERT_LOCKED(m->object); 5221 MPASS(!pmap_page_is_mapped(m)); 5222 5223 if (vm_page_xbusied(m)) 5224 m->valid = 0; 5225 else 5226 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5227 } 5228 5229 /* 5230 * vm_page_zero_invalid() 5231 * 5232 * The kernel assumes that the invalid portions of a page contain 5233 * garbage, but such pages can be mapped into memory by user code. 5234 * When this occurs, we must zero out the non-valid portions of the 5235 * page so user code sees what it expects. 5236 * 5237 * Pages are most often semi-valid when the end of a file is mapped 5238 * into memory and the file's size is not page aligned. 5239 */ 5240 void 5241 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5242 { 5243 int b; 5244 int i; 5245 5246 /* 5247 * Scan the valid bits looking for invalid sections that 5248 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5249 * valid bit may be set ) have already been zeroed by 5250 * vm_page_set_validclean(). 5251 */ 5252 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5253 if (i == (PAGE_SIZE / DEV_BSIZE) || 5254 (m->valid & ((vm_page_bits_t)1 << i))) { 5255 if (i > b) { 5256 pmap_zero_page_area(m, 5257 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5258 } 5259 b = i + 1; 5260 } 5261 } 5262 5263 /* 5264 * setvalid is TRUE when we can safely set the zero'd areas 5265 * as being valid. We can do this if there are no cache consistancy 5266 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5267 */ 5268 if (setvalid) 5269 vm_page_valid(m); 5270 } 5271 5272 /* 5273 * vm_page_is_valid: 5274 * 5275 * Is (partial) page valid? Note that the case where size == 0 5276 * will return FALSE in the degenerate case where the page is 5277 * entirely invalid, and TRUE otherwise. 5278 * 5279 * Some callers envoke this routine without the busy lock held and 5280 * handle races via higher level locks. Typical callers should 5281 * hold a busy lock to prevent invalidation. 5282 */ 5283 int 5284 vm_page_is_valid(vm_page_t m, int base, int size) 5285 { 5286 vm_page_bits_t bits; 5287 5288 bits = vm_page_bits(base, size); 5289 return (m->valid != 0 && (m->valid & bits) == bits); 5290 } 5291 5292 /* 5293 * Returns true if all of the specified predicates are true for the entire 5294 * (super)page and false otherwise. 5295 */ 5296 bool 5297 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 5298 { 5299 vm_object_t object; 5300 int i, npages; 5301 5302 object = m->object; 5303 if (skip_m != NULL && skip_m->object != object) 5304 return (false); 5305 VM_OBJECT_ASSERT_LOCKED(object); 5306 npages = atop(pagesizes[m->psind]); 5307 5308 /* 5309 * The physically contiguous pages that make up a superpage, i.e., a 5310 * page with a page size index ("psind") greater than zero, will 5311 * occupy adjacent entries in vm_page_array[]. 5312 */ 5313 for (i = 0; i < npages; i++) { 5314 /* Always test object consistency, including "skip_m". */ 5315 if (m[i].object != object) 5316 return (false); 5317 if (&m[i] == skip_m) 5318 continue; 5319 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5320 return (false); 5321 if ((flags & PS_ALL_DIRTY) != 0) { 5322 /* 5323 * Calling vm_page_test_dirty() or pmap_is_modified() 5324 * might stop this case from spuriously returning 5325 * "false". However, that would require a write lock 5326 * on the object containing "m[i]". 5327 */ 5328 if (m[i].dirty != VM_PAGE_BITS_ALL) 5329 return (false); 5330 } 5331 if ((flags & PS_ALL_VALID) != 0 && 5332 m[i].valid != VM_PAGE_BITS_ALL) 5333 return (false); 5334 } 5335 return (true); 5336 } 5337 5338 /* 5339 * Set the page's dirty bits if the page is modified. 5340 */ 5341 void 5342 vm_page_test_dirty(vm_page_t m) 5343 { 5344 5345 vm_page_assert_busied(m); 5346 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5347 vm_page_dirty(m); 5348 } 5349 5350 void 5351 vm_page_valid(vm_page_t m) 5352 { 5353 5354 vm_page_assert_busied(m); 5355 if (vm_page_xbusied(m)) 5356 m->valid = VM_PAGE_BITS_ALL; 5357 else 5358 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5359 } 5360 5361 void 5362 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5363 { 5364 5365 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5366 } 5367 5368 void 5369 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5370 { 5371 5372 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5373 } 5374 5375 int 5376 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5377 { 5378 5379 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5380 } 5381 5382 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5383 void 5384 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5385 { 5386 5387 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5388 } 5389 5390 void 5391 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5392 { 5393 5394 mtx_assert_(vm_page_lockptr(m), a, file, line); 5395 } 5396 #endif 5397 5398 #ifdef INVARIANTS 5399 void 5400 vm_page_object_busy_assert(vm_page_t m) 5401 { 5402 5403 /* 5404 * Certain of the page's fields may only be modified by the 5405 * holder of a page or object busy. 5406 */ 5407 if (m->object != NULL && !vm_page_busied(m)) 5408 VM_OBJECT_ASSERT_BUSY(m->object); 5409 } 5410 5411 void 5412 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5413 { 5414 5415 if ((bits & PGA_WRITEABLE) == 0) 5416 return; 5417 5418 /* 5419 * The PGA_WRITEABLE flag can only be set if the page is 5420 * managed, is exclusively busied or the object is locked. 5421 * Currently, this flag is only set by pmap_enter(). 5422 */ 5423 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5424 ("PGA_WRITEABLE on unmanaged page")); 5425 if (!vm_page_xbusied(m)) 5426 VM_OBJECT_ASSERT_BUSY(m->object); 5427 } 5428 #endif 5429 5430 #include "opt_ddb.h" 5431 #ifdef DDB 5432 #include <sys/kernel.h> 5433 5434 #include <ddb/ddb.h> 5435 5436 DB_SHOW_COMMAND(page, vm_page_print_page_info) 5437 { 5438 5439 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5440 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5441 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5442 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5443 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5444 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5445 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5446 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5447 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5448 } 5449 5450 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 5451 { 5452 int dom; 5453 5454 db_printf("pq_free %d\n", vm_free_count()); 5455 for (dom = 0; dom < vm_ndomains; dom++) { 5456 db_printf( 5457 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5458 dom, 5459 vm_dom[dom].vmd_page_count, 5460 vm_dom[dom].vmd_free_count, 5461 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5462 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5463 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5464 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5465 } 5466 } 5467 5468 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5469 { 5470 vm_page_t m; 5471 boolean_t phys, virt; 5472 5473 if (!have_addr) { 5474 db_printf("show pginfo addr\n"); 5475 return; 5476 } 5477 5478 phys = strchr(modif, 'p') != NULL; 5479 virt = strchr(modif, 'v') != NULL; 5480 if (virt) 5481 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5482 else if (phys) 5483 m = PHYS_TO_VM_PAGE(addr); 5484 else 5485 m = (vm_page_t)addr; 5486 db_printf( 5487 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5488 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5489 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5490 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5491 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5492 } 5493 #endif /* DDB */ 5494