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 * vm_page_sleep_if_busy: 1366 * 1367 * Sleep and release the object lock if the page is busied. 1368 * Returns TRUE if the thread slept. 1369 * 1370 * The given page must be unlocked and object containing it must 1371 * be locked. 1372 */ 1373 int 1374 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg) 1375 { 1376 vm_object_t obj; 1377 1378 vm_page_lock_assert(m, MA_NOTOWNED); 1379 VM_OBJECT_ASSERT_WLOCKED(m->object); 1380 1381 /* 1382 * The page-specific object must be cached because page 1383 * identity can change during the sleep, causing the 1384 * re-lock of a different object. 1385 * It is assumed that a reference to the object is already 1386 * held by the callers. 1387 */ 1388 obj = m->object; 1389 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) { 1390 VM_OBJECT_WLOCK(obj); 1391 return (TRUE); 1392 } 1393 return (FALSE); 1394 } 1395 1396 /* 1397 * vm_page_sleep_if_xbusy: 1398 * 1399 * Sleep and release the object lock if the page is xbusied. 1400 * Returns TRUE if the thread slept. 1401 * 1402 * The given page must be unlocked and object containing it must 1403 * be locked. 1404 */ 1405 int 1406 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg) 1407 { 1408 vm_object_t obj; 1409 1410 vm_page_lock_assert(m, MA_NOTOWNED); 1411 VM_OBJECT_ASSERT_WLOCKED(m->object); 1412 1413 /* 1414 * The page-specific object must be cached because page 1415 * identity can change during the sleep, causing the 1416 * re-lock of a different object. 1417 * It is assumed that a reference to the object is already 1418 * held by the callers. 1419 */ 1420 obj = m->object; 1421 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY, 1422 true)) { 1423 VM_OBJECT_WLOCK(obj); 1424 return (TRUE); 1425 } 1426 return (FALSE); 1427 } 1428 1429 /* 1430 * vm_page_dirty_KBI: [ internal use only ] 1431 * 1432 * Set all bits in the page's dirty field. 1433 * 1434 * The object containing the specified page must be locked if the 1435 * call is made from the machine-independent layer. 1436 * 1437 * See vm_page_clear_dirty_mask(). 1438 * 1439 * This function should only be called by vm_page_dirty(). 1440 */ 1441 void 1442 vm_page_dirty_KBI(vm_page_t m) 1443 { 1444 1445 /* Refer to this operation by its public name. */ 1446 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1447 m->dirty = VM_PAGE_BITS_ALL; 1448 } 1449 1450 /* 1451 * vm_page_insert: [ internal use only ] 1452 * 1453 * Inserts the given mem entry into the object and object list. 1454 * 1455 * The object must be locked. 1456 */ 1457 int 1458 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1459 { 1460 vm_page_t mpred; 1461 1462 VM_OBJECT_ASSERT_WLOCKED(object); 1463 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1464 return (vm_page_insert_after(m, object, pindex, mpred)); 1465 } 1466 1467 /* 1468 * vm_page_insert_after: 1469 * 1470 * Inserts the page "m" into the specified object at offset "pindex". 1471 * 1472 * The page "mpred" must immediately precede the offset "pindex" within 1473 * the specified object. 1474 * 1475 * The object must be locked. 1476 */ 1477 static int 1478 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1479 vm_page_t mpred) 1480 { 1481 vm_page_t msucc; 1482 1483 VM_OBJECT_ASSERT_WLOCKED(object); 1484 KASSERT(m->object == NULL, 1485 ("vm_page_insert_after: page already inserted")); 1486 if (mpred != NULL) { 1487 KASSERT(mpred->object == object, 1488 ("vm_page_insert_after: object doesn't contain mpred")); 1489 KASSERT(mpred->pindex < pindex, 1490 ("vm_page_insert_after: mpred doesn't precede pindex")); 1491 msucc = TAILQ_NEXT(mpred, listq); 1492 } else 1493 msucc = TAILQ_FIRST(&object->memq); 1494 if (msucc != NULL) 1495 KASSERT(msucc->pindex > pindex, 1496 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1497 1498 /* 1499 * Record the object/offset pair in this page. 1500 */ 1501 m->object = object; 1502 m->pindex = pindex; 1503 m->ref_count |= VPRC_OBJREF; 1504 1505 /* 1506 * Now link into the object's ordered list of backed pages. 1507 */ 1508 if (vm_radix_insert(&object->rtree, m)) { 1509 m->object = NULL; 1510 m->pindex = 0; 1511 m->ref_count &= ~VPRC_OBJREF; 1512 return (1); 1513 } 1514 vm_page_insert_radixdone(m, object, mpred); 1515 return (0); 1516 } 1517 1518 /* 1519 * vm_page_insert_radixdone: 1520 * 1521 * Complete page "m" insertion into the specified object after the 1522 * radix trie hooking. 1523 * 1524 * The page "mpred" must precede the offset "m->pindex" within the 1525 * specified object. 1526 * 1527 * The object must be locked. 1528 */ 1529 static void 1530 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1531 { 1532 1533 VM_OBJECT_ASSERT_WLOCKED(object); 1534 KASSERT(object != NULL && m->object == object, 1535 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1536 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1537 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1538 if (mpred != NULL) { 1539 KASSERT(mpred->object == object, 1540 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1541 KASSERT(mpred->pindex < m->pindex, 1542 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1543 } 1544 1545 if (mpred != NULL) 1546 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1547 else 1548 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1549 1550 /* 1551 * Show that the object has one more resident page. 1552 */ 1553 object->resident_page_count++; 1554 1555 /* 1556 * Hold the vnode until the last page is released. 1557 */ 1558 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1559 vhold(object->handle); 1560 1561 /* 1562 * Since we are inserting a new and possibly dirty page, 1563 * update the object's generation count. 1564 */ 1565 if (pmap_page_is_write_mapped(m)) 1566 vm_object_set_writeable_dirty(object); 1567 } 1568 1569 /* 1570 * Do the work to remove a page from its object. The caller is responsible for 1571 * updating the page's fields to reflect this removal. 1572 */ 1573 static void 1574 vm_page_object_remove(vm_page_t m) 1575 { 1576 vm_object_t object; 1577 vm_page_t mrem; 1578 1579 vm_page_assert_xbusied(m); 1580 object = m->object; 1581 VM_OBJECT_ASSERT_WLOCKED(object); 1582 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1583 ("page %p is missing its object ref", m)); 1584 1585 /* Deferred free of swap space. */ 1586 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1587 vm_pager_page_unswapped(m); 1588 1589 m->object = NULL; 1590 mrem = vm_radix_remove(&object->rtree, m->pindex); 1591 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1592 1593 /* 1594 * Now remove from the object's list of backed pages. 1595 */ 1596 TAILQ_REMOVE(&object->memq, m, listq); 1597 1598 /* 1599 * And show that the object has one fewer resident page. 1600 */ 1601 object->resident_page_count--; 1602 1603 /* 1604 * The vnode may now be recycled. 1605 */ 1606 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1607 vdrop(object->handle); 1608 } 1609 1610 /* 1611 * vm_page_remove: 1612 * 1613 * Removes the specified page from its containing object, but does not 1614 * invalidate any backing storage. Returns true if the object's reference 1615 * was the last reference to the page, and false otherwise. 1616 * 1617 * The object must be locked and the page must be exclusively busied. 1618 * The exclusive busy will be released on return. If this is not the 1619 * final ref and the caller does not hold a wire reference it may not 1620 * continue to access the page. 1621 */ 1622 bool 1623 vm_page_remove(vm_page_t m) 1624 { 1625 bool dropped; 1626 1627 dropped = vm_page_remove_xbusy(m); 1628 vm_page_xunbusy(m); 1629 1630 return (dropped); 1631 } 1632 1633 /* 1634 * vm_page_remove_xbusy 1635 * 1636 * Removes the page but leaves the xbusy held. Returns true if this 1637 * removed the final ref and false otherwise. 1638 */ 1639 bool 1640 vm_page_remove_xbusy(vm_page_t m) 1641 { 1642 1643 vm_page_object_remove(m); 1644 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1645 } 1646 1647 /* 1648 * vm_page_lookup: 1649 * 1650 * Returns the page associated with the object/offset 1651 * pair specified; if none is found, NULL is returned. 1652 * 1653 * The object must be locked. 1654 */ 1655 vm_page_t 1656 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1657 { 1658 1659 VM_OBJECT_ASSERT_LOCKED(object); 1660 return (vm_radix_lookup(&object->rtree, pindex)); 1661 } 1662 1663 /* 1664 * vm_page_relookup: 1665 * 1666 * Returns a page that must already have been busied by 1667 * the caller. Used for bogus page replacement. 1668 */ 1669 vm_page_t 1670 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1671 { 1672 vm_page_t m; 1673 1674 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 1675 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1676 m->object == object && m->pindex == pindex, 1677 ("vm_page_relookup: Invalid page %p", m)); 1678 return (m); 1679 } 1680 1681 /* 1682 * This should only be used by lockless functions for releasing transient 1683 * incorrect acquires. The page may have been freed after we acquired a 1684 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1685 * further to do. 1686 */ 1687 static void 1688 vm_page_busy_release(vm_page_t m) 1689 { 1690 u_int x; 1691 1692 x = atomic_load_int(&m->busy_lock); 1693 for (;;) { 1694 if (x == VPB_FREED) 1695 break; 1696 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1697 if (atomic_fcmpset_int(&m->busy_lock, &x, 1698 x - VPB_ONE_SHARER)) 1699 break; 1700 continue; 1701 } 1702 KASSERT((x & VPB_BIT_SHARED) != 0 || 1703 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1704 ("vm_page_busy_release: %p xbusy not owned.", m)); 1705 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1706 continue; 1707 if ((x & VPB_BIT_WAITERS) != 0) 1708 wakeup(m); 1709 break; 1710 } 1711 } 1712 1713 /* 1714 * vm_page_find_least: 1715 * 1716 * Returns the page associated with the object with least pindex 1717 * greater than or equal to the parameter pindex, or NULL. 1718 * 1719 * The object must be locked. 1720 */ 1721 vm_page_t 1722 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1723 { 1724 vm_page_t m; 1725 1726 VM_OBJECT_ASSERT_LOCKED(object); 1727 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1728 m = vm_radix_lookup_ge(&object->rtree, pindex); 1729 return (m); 1730 } 1731 1732 /* 1733 * Returns the given page's successor (by pindex) within the object if it is 1734 * resident; if none is found, NULL is returned. 1735 * 1736 * The object must be locked. 1737 */ 1738 vm_page_t 1739 vm_page_next(vm_page_t m) 1740 { 1741 vm_page_t next; 1742 1743 VM_OBJECT_ASSERT_LOCKED(m->object); 1744 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1745 MPASS(next->object == m->object); 1746 if (next->pindex != m->pindex + 1) 1747 next = NULL; 1748 } 1749 return (next); 1750 } 1751 1752 /* 1753 * Returns the given page's predecessor (by pindex) within the object if it is 1754 * resident; if none is found, NULL is returned. 1755 * 1756 * The object must be locked. 1757 */ 1758 vm_page_t 1759 vm_page_prev(vm_page_t m) 1760 { 1761 vm_page_t prev; 1762 1763 VM_OBJECT_ASSERT_LOCKED(m->object); 1764 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1765 MPASS(prev->object == m->object); 1766 if (prev->pindex != m->pindex - 1) 1767 prev = NULL; 1768 } 1769 return (prev); 1770 } 1771 1772 /* 1773 * Uses the page mnew as a replacement for an existing page at index 1774 * pindex which must be already present in the object. 1775 * 1776 * Both pages must be exclusively busied on enter. The old page is 1777 * unbusied on exit. 1778 * 1779 * A return value of true means mold is now free. If this is not the 1780 * final ref and the caller does not hold a wire reference it may not 1781 * continue to access the page. 1782 */ 1783 static bool 1784 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1785 vm_page_t mold) 1786 { 1787 vm_page_t mret; 1788 bool dropped; 1789 1790 VM_OBJECT_ASSERT_WLOCKED(object); 1791 vm_page_assert_xbusied(mold); 1792 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1793 ("vm_page_replace: page %p already in object", mnew)); 1794 1795 /* 1796 * This function mostly follows vm_page_insert() and 1797 * vm_page_remove() without the radix, object count and vnode 1798 * dance. Double check such functions for more comments. 1799 */ 1800 1801 mnew->object = object; 1802 mnew->pindex = pindex; 1803 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1804 mret = vm_radix_replace(&object->rtree, mnew); 1805 KASSERT(mret == mold, 1806 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1807 KASSERT((mold->oflags & VPO_UNMANAGED) == 1808 (mnew->oflags & VPO_UNMANAGED), 1809 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1810 1811 /* Keep the resident page list in sorted order. */ 1812 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1813 TAILQ_REMOVE(&object->memq, mold, listq); 1814 mold->object = NULL; 1815 1816 /* 1817 * The object's resident_page_count does not change because we have 1818 * swapped one page for another, but the generation count should 1819 * change if the page is dirty. 1820 */ 1821 if (pmap_page_is_write_mapped(mnew)) 1822 vm_object_set_writeable_dirty(object); 1823 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1824 vm_page_xunbusy(mold); 1825 1826 return (dropped); 1827 } 1828 1829 void 1830 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1831 vm_page_t mold) 1832 { 1833 1834 vm_page_assert_xbusied(mnew); 1835 1836 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1837 vm_page_free(mold); 1838 } 1839 1840 /* 1841 * vm_page_rename: 1842 * 1843 * Move the given memory entry from its 1844 * current object to the specified target object/offset. 1845 * 1846 * Note: swap associated with the page must be invalidated by the move. We 1847 * have to do this for several reasons: (1) we aren't freeing the 1848 * page, (2) we are dirtying the page, (3) the VM system is probably 1849 * moving the page from object A to B, and will then later move 1850 * the backing store from A to B and we can't have a conflict. 1851 * 1852 * Note: we *always* dirty the page. It is necessary both for the 1853 * fact that we moved it, and because we may be invalidating 1854 * swap. 1855 * 1856 * The objects must be locked. 1857 */ 1858 int 1859 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1860 { 1861 vm_page_t mpred; 1862 vm_pindex_t opidx; 1863 1864 VM_OBJECT_ASSERT_WLOCKED(new_object); 1865 1866 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1867 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1868 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1869 ("vm_page_rename: pindex already renamed")); 1870 1871 /* 1872 * Create a custom version of vm_page_insert() which does not depend 1873 * by m_prev and can cheat on the implementation aspects of the 1874 * function. 1875 */ 1876 opidx = m->pindex; 1877 m->pindex = new_pindex; 1878 if (vm_radix_insert(&new_object->rtree, m)) { 1879 m->pindex = opidx; 1880 return (1); 1881 } 1882 1883 /* 1884 * The operation cannot fail anymore. The removal must happen before 1885 * the listq iterator is tainted. 1886 */ 1887 m->pindex = opidx; 1888 vm_page_object_remove(m); 1889 1890 /* Return back to the new pindex to complete vm_page_insert(). */ 1891 m->pindex = new_pindex; 1892 m->object = new_object; 1893 1894 vm_page_insert_radixdone(m, new_object, mpred); 1895 vm_page_dirty(m); 1896 return (0); 1897 } 1898 1899 /* 1900 * vm_page_alloc: 1901 * 1902 * Allocate and return a page that is associated with the specified 1903 * object and offset pair. By default, this page is exclusive busied. 1904 * 1905 * The caller must always specify an allocation class. 1906 * 1907 * allocation classes: 1908 * VM_ALLOC_NORMAL normal process request 1909 * VM_ALLOC_SYSTEM system *really* needs a page 1910 * VM_ALLOC_INTERRUPT interrupt time request 1911 * 1912 * optional allocation flags: 1913 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1914 * intends to allocate 1915 * VM_ALLOC_NOBUSY do not exclusive busy the page 1916 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1917 * VM_ALLOC_NOOBJ page is not associated with an object and 1918 * should not be exclusive busy 1919 * VM_ALLOC_SBUSY shared busy the allocated page 1920 * VM_ALLOC_WIRED wire the allocated page 1921 * VM_ALLOC_ZERO prefer a zeroed page 1922 */ 1923 vm_page_t 1924 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1925 { 1926 1927 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1928 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1929 } 1930 1931 vm_page_t 1932 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1933 int req) 1934 { 1935 1936 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1937 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1938 NULL)); 1939 } 1940 1941 /* 1942 * Allocate a page in the specified object with the given page index. To 1943 * optimize insertion of the page into the object, the caller must also specifiy 1944 * the resident page in the object with largest index smaller than the given 1945 * page index, or NULL if no such page exists. 1946 */ 1947 vm_page_t 1948 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1949 int req, vm_page_t mpred) 1950 { 1951 struct vm_domainset_iter di; 1952 vm_page_t m; 1953 int domain; 1954 1955 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1956 do { 1957 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1958 mpred); 1959 if (m != NULL) 1960 break; 1961 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1962 1963 return (m); 1964 } 1965 1966 /* 1967 * Returns true if the number of free pages exceeds the minimum 1968 * for the request class and false otherwise. 1969 */ 1970 static int 1971 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 1972 { 1973 u_int limit, old, new; 1974 1975 if (req_class == VM_ALLOC_INTERRUPT) 1976 limit = 0; 1977 else if (req_class == VM_ALLOC_SYSTEM) 1978 limit = vmd->vmd_interrupt_free_min; 1979 else 1980 limit = vmd->vmd_free_reserved; 1981 1982 /* 1983 * Attempt to reserve the pages. Fail if we're below the limit. 1984 */ 1985 limit += npages; 1986 old = vmd->vmd_free_count; 1987 do { 1988 if (old < limit) 1989 return (0); 1990 new = old - npages; 1991 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 1992 1993 /* Wake the page daemon if we've crossed the threshold. */ 1994 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 1995 pagedaemon_wakeup(vmd->vmd_domain); 1996 1997 /* Only update bitsets on transitions. */ 1998 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 1999 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2000 vm_domain_set(vmd); 2001 2002 return (1); 2003 } 2004 2005 int 2006 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2007 { 2008 int req_class; 2009 2010 /* 2011 * The page daemon is allowed to dig deeper into the free page list. 2012 */ 2013 req_class = req & VM_ALLOC_CLASS_MASK; 2014 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2015 req_class = VM_ALLOC_SYSTEM; 2016 return (_vm_domain_allocate(vmd, req_class, npages)); 2017 } 2018 2019 vm_page_t 2020 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 2021 int req, vm_page_t mpred) 2022 { 2023 struct vm_domain *vmd; 2024 vm_page_t m; 2025 int flags, pool; 2026 2027 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2028 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2029 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2030 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2031 ("inconsistent object(%p)/req(%x)", object, req)); 2032 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2033 ("Can't sleep and retry object insertion.")); 2034 KASSERT(mpred == NULL || mpred->pindex < pindex, 2035 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2036 (uintmax_t)pindex)); 2037 if (object != NULL) 2038 VM_OBJECT_ASSERT_WLOCKED(object); 2039 2040 flags = 0; 2041 m = NULL; 2042 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; 2043 again: 2044 #if VM_NRESERVLEVEL > 0 2045 /* 2046 * Can we allocate the page from a reservation? 2047 */ 2048 if (vm_object_reserv(object) && 2049 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 2050 NULL) { 2051 domain = vm_phys_domain(m); 2052 vmd = VM_DOMAIN(domain); 2053 goto found; 2054 } 2055 #endif 2056 vmd = VM_DOMAIN(domain); 2057 if (vmd->vmd_pgcache[pool].zone != NULL) { 2058 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM); 2059 if (m != NULL) { 2060 flags |= PG_PCPU_CACHE; 2061 goto found; 2062 } 2063 } 2064 if (vm_domain_allocate(vmd, req, 1)) { 2065 /* 2066 * If not, allocate it from the free page queues. 2067 */ 2068 vm_domain_free_lock(vmd); 2069 m = vm_phys_alloc_pages(domain, pool, 0); 2070 vm_domain_free_unlock(vmd); 2071 if (m == NULL) { 2072 vm_domain_freecnt_inc(vmd, 1); 2073 #if VM_NRESERVLEVEL > 0 2074 if (vm_reserv_reclaim_inactive(domain)) 2075 goto again; 2076 #endif 2077 } 2078 } 2079 if (m == NULL) { 2080 /* 2081 * Not allocatable, give up. 2082 */ 2083 if (vm_domain_alloc_fail(vmd, object, req)) 2084 goto again; 2085 return (NULL); 2086 } 2087 2088 /* 2089 * At this point we had better have found a good page. 2090 */ 2091 found: 2092 vm_page_dequeue(m); 2093 vm_page_alloc_check(m); 2094 2095 /* 2096 * Initialize the page. Only the PG_ZERO flag is inherited. 2097 */ 2098 if ((req & VM_ALLOC_ZERO) != 0) 2099 flags |= (m->flags & PG_ZERO); 2100 if ((req & VM_ALLOC_NODUMP) != 0) 2101 flags |= PG_NODUMP; 2102 m->flags = flags; 2103 m->a.flags = 0; 2104 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2105 VPO_UNMANAGED : 0; 2106 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2107 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2108 else if ((req & VM_ALLOC_SBUSY) != 0) 2109 m->busy_lock = VPB_SHARERS_WORD(1); 2110 else 2111 m->busy_lock = VPB_UNBUSIED; 2112 if (req & VM_ALLOC_WIRED) { 2113 vm_wire_add(1); 2114 m->ref_count = 1; 2115 } 2116 m->a.act_count = 0; 2117 2118 if (object != NULL) { 2119 if (vm_page_insert_after(m, object, pindex, mpred)) { 2120 if (req & VM_ALLOC_WIRED) { 2121 vm_wire_sub(1); 2122 m->ref_count = 0; 2123 } 2124 KASSERT(m->object == NULL, ("page %p has object", m)); 2125 m->oflags = VPO_UNMANAGED; 2126 m->busy_lock = VPB_UNBUSIED; 2127 /* Don't change PG_ZERO. */ 2128 vm_page_free_toq(m); 2129 if (req & VM_ALLOC_WAITFAIL) { 2130 VM_OBJECT_WUNLOCK(object); 2131 vm_radix_wait(); 2132 VM_OBJECT_WLOCK(object); 2133 } 2134 return (NULL); 2135 } 2136 2137 /* Ignore device objects; the pager sets "memattr" for them. */ 2138 if (object->memattr != VM_MEMATTR_DEFAULT && 2139 (object->flags & OBJ_FICTITIOUS) == 0) 2140 pmap_page_set_memattr(m, object->memattr); 2141 } else 2142 m->pindex = pindex; 2143 2144 return (m); 2145 } 2146 2147 /* 2148 * vm_page_alloc_contig: 2149 * 2150 * Allocate a contiguous set of physical pages of the given size "npages" 2151 * from the free lists. All of the physical pages must be at or above 2152 * the given physical address "low" and below the given physical address 2153 * "high". The given value "alignment" determines the alignment of the 2154 * first physical page in the set. If the given value "boundary" is 2155 * non-zero, then the set of physical pages cannot cross any physical 2156 * address boundary that is a multiple of that value. Both "alignment" 2157 * and "boundary" must be a power of two. 2158 * 2159 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2160 * then the memory attribute setting for the physical pages is configured 2161 * to the object's memory attribute setting. Otherwise, the memory 2162 * attribute setting for the physical pages is configured to "memattr", 2163 * overriding the object's memory attribute setting. However, if the 2164 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2165 * memory attribute setting for the physical pages cannot be configured 2166 * to VM_MEMATTR_DEFAULT. 2167 * 2168 * The specified object may not contain fictitious pages. 2169 * 2170 * The caller must always specify an allocation class. 2171 * 2172 * allocation classes: 2173 * VM_ALLOC_NORMAL normal process request 2174 * VM_ALLOC_SYSTEM system *really* needs a page 2175 * VM_ALLOC_INTERRUPT interrupt time request 2176 * 2177 * optional allocation flags: 2178 * VM_ALLOC_NOBUSY do not exclusive busy the page 2179 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2180 * VM_ALLOC_NOOBJ page is not associated with an object and 2181 * should not be exclusive busy 2182 * VM_ALLOC_SBUSY shared busy the allocated page 2183 * VM_ALLOC_WIRED wire the allocated page 2184 * VM_ALLOC_ZERO prefer a zeroed page 2185 */ 2186 vm_page_t 2187 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2188 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2189 vm_paddr_t boundary, vm_memattr_t memattr) 2190 { 2191 struct vm_domainset_iter di; 2192 vm_page_t m; 2193 int domain; 2194 2195 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2196 do { 2197 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2198 npages, low, high, alignment, boundary, memattr); 2199 if (m != NULL) 2200 break; 2201 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2202 2203 return (m); 2204 } 2205 2206 vm_page_t 2207 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2208 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2209 vm_paddr_t boundary, vm_memattr_t memattr) 2210 { 2211 struct vm_domain *vmd; 2212 vm_page_t m, m_ret, mpred; 2213 u_int busy_lock, flags, oflags; 2214 2215 mpred = NULL; /* XXX: pacify gcc */ 2216 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2217 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2218 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2219 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2220 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 2221 req)); 2222 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2223 ("Can't sleep and retry object insertion.")); 2224 if (object != NULL) { 2225 VM_OBJECT_ASSERT_WLOCKED(object); 2226 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2227 ("vm_page_alloc_contig: object %p has fictitious pages", 2228 object)); 2229 } 2230 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2231 2232 if (object != NULL) { 2233 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2234 KASSERT(mpred == NULL || mpred->pindex != pindex, 2235 ("vm_page_alloc_contig: pindex already allocated")); 2236 } 2237 2238 /* 2239 * Can we allocate the pages without the number of free pages falling 2240 * below the lower bound for the allocation class? 2241 */ 2242 m_ret = NULL; 2243 again: 2244 #if VM_NRESERVLEVEL > 0 2245 /* 2246 * Can we allocate the pages from a reservation? 2247 */ 2248 if (vm_object_reserv(object) && 2249 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2250 mpred, npages, low, high, alignment, boundary)) != NULL) { 2251 domain = vm_phys_domain(m_ret); 2252 vmd = VM_DOMAIN(domain); 2253 goto found; 2254 } 2255 #endif 2256 vmd = VM_DOMAIN(domain); 2257 if (vm_domain_allocate(vmd, req, npages)) { 2258 /* 2259 * allocate them from the free page queues. 2260 */ 2261 vm_domain_free_lock(vmd); 2262 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2263 alignment, boundary); 2264 vm_domain_free_unlock(vmd); 2265 if (m_ret == NULL) { 2266 vm_domain_freecnt_inc(vmd, npages); 2267 #if VM_NRESERVLEVEL > 0 2268 if (vm_reserv_reclaim_contig(domain, npages, low, 2269 high, alignment, boundary)) 2270 goto again; 2271 #endif 2272 } 2273 } 2274 if (m_ret == NULL) { 2275 if (vm_domain_alloc_fail(vmd, object, req)) 2276 goto again; 2277 return (NULL); 2278 } 2279 #if VM_NRESERVLEVEL > 0 2280 found: 2281 #endif 2282 for (m = m_ret; m < &m_ret[npages]; m++) { 2283 vm_page_dequeue(m); 2284 vm_page_alloc_check(m); 2285 } 2286 2287 /* 2288 * Initialize the pages. Only the PG_ZERO flag is inherited. 2289 */ 2290 flags = 0; 2291 if ((req & VM_ALLOC_ZERO) != 0) 2292 flags = PG_ZERO; 2293 if ((req & VM_ALLOC_NODUMP) != 0) 2294 flags |= PG_NODUMP; 2295 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2296 VPO_UNMANAGED : 0; 2297 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2298 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2299 else if ((req & VM_ALLOC_SBUSY) != 0) 2300 busy_lock = VPB_SHARERS_WORD(1); 2301 else 2302 busy_lock = VPB_UNBUSIED; 2303 if ((req & VM_ALLOC_WIRED) != 0) 2304 vm_wire_add(npages); 2305 if (object != NULL) { 2306 if (object->memattr != VM_MEMATTR_DEFAULT && 2307 memattr == VM_MEMATTR_DEFAULT) 2308 memattr = object->memattr; 2309 } 2310 for (m = m_ret; m < &m_ret[npages]; m++) { 2311 m->a.flags = 0; 2312 m->flags = (m->flags | PG_NODUMP) & flags; 2313 m->busy_lock = busy_lock; 2314 if ((req & VM_ALLOC_WIRED) != 0) 2315 m->ref_count = 1; 2316 m->a.act_count = 0; 2317 m->oflags = oflags; 2318 if (object != NULL) { 2319 if (vm_page_insert_after(m, object, pindex, mpred)) { 2320 if ((req & VM_ALLOC_WIRED) != 0) 2321 vm_wire_sub(npages); 2322 KASSERT(m->object == NULL, 2323 ("page %p has object", m)); 2324 mpred = m; 2325 for (m = m_ret; m < &m_ret[npages]; m++) { 2326 if (m <= mpred && 2327 (req & VM_ALLOC_WIRED) != 0) 2328 m->ref_count = 0; 2329 m->oflags = VPO_UNMANAGED; 2330 m->busy_lock = VPB_UNBUSIED; 2331 /* Don't change PG_ZERO. */ 2332 vm_page_free_toq(m); 2333 } 2334 if (req & VM_ALLOC_WAITFAIL) { 2335 VM_OBJECT_WUNLOCK(object); 2336 vm_radix_wait(); 2337 VM_OBJECT_WLOCK(object); 2338 } 2339 return (NULL); 2340 } 2341 mpred = m; 2342 } else 2343 m->pindex = pindex; 2344 if (memattr != VM_MEMATTR_DEFAULT) 2345 pmap_page_set_memattr(m, memattr); 2346 pindex++; 2347 } 2348 return (m_ret); 2349 } 2350 2351 /* 2352 * Check a page that has been freshly dequeued from a freelist. 2353 */ 2354 static void 2355 vm_page_alloc_check(vm_page_t m) 2356 { 2357 2358 KASSERT(m->object == NULL, ("page %p has object", m)); 2359 KASSERT(m->a.queue == PQ_NONE && 2360 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2361 ("page %p has unexpected queue %d, flags %#x", 2362 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2363 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2364 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2365 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2366 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2367 ("page %p has unexpected memattr %d", 2368 m, pmap_page_get_memattr(m))); 2369 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2370 } 2371 2372 /* 2373 * vm_page_alloc_freelist: 2374 * 2375 * Allocate a physical page from the specified free page list. 2376 * 2377 * The caller must always specify an allocation class. 2378 * 2379 * allocation classes: 2380 * VM_ALLOC_NORMAL normal process request 2381 * VM_ALLOC_SYSTEM system *really* needs a page 2382 * VM_ALLOC_INTERRUPT interrupt time request 2383 * 2384 * optional allocation flags: 2385 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2386 * intends to allocate 2387 * VM_ALLOC_WIRED wire the allocated page 2388 * VM_ALLOC_ZERO prefer a zeroed page 2389 */ 2390 vm_page_t 2391 vm_page_alloc_freelist(int freelist, int req) 2392 { 2393 struct vm_domainset_iter di; 2394 vm_page_t m; 2395 int domain; 2396 2397 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2398 do { 2399 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2400 if (m != NULL) 2401 break; 2402 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2403 2404 return (m); 2405 } 2406 2407 vm_page_t 2408 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2409 { 2410 struct vm_domain *vmd; 2411 vm_page_t m; 2412 u_int flags; 2413 2414 m = NULL; 2415 vmd = VM_DOMAIN(domain); 2416 again: 2417 if (vm_domain_allocate(vmd, req, 1)) { 2418 vm_domain_free_lock(vmd); 2419 m = vm_phys_alloc_freelist_pages(domain, freelist, 2420 VM_FREEPOOL_DIRECT, 0); 2421 vm_domain_free_unlock(vmd); 2422 if (m == NULL) 2423 vm_domain_freecnt_inc(vmd, 1); 2424 } 2425 if (m == NULL) { 2426 if (vm_domain_alloc_fail(vmd, NULL, req)) 2427 goto again; 2428 return (NULL); 2429 } 2430 vm_page_dequeue(m); 2431 vm_page_alloc_check(m); 2432 2433 /* 2434 * Initialize the page. Only the PG_ZERO flag is inherited. 2435 */ 2436 m->a.flags = 0; 2437 flags = 0; 2438 if ((req & VM_ALLOC_ZERO) != 0) 2439 flags = PG_ZERO; 2440 m->flags &= flags; 2441 if ((req & VM_ALLOC_WIRED) != 0) { 2442 vm_wire_add(1); 2443 m->ref_count = 1; 2444 } 2445 /* Unmanaged pages don't use "act_count". */ 2446 m->oflags = VPO_UNMANAGED; 2447 return (m); 2448 } 2449 2450 static int 2451 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2452 { 2453 struct vm_domain *vmd; 2454 struct vm_pgcache *pgcache; 2455 int i; 2456 2457 pgcache = arg; 2458 vmd = VM_DOMAIN(pgcache->domain); 2459 2460 /* 2461 * The page daemon should avoid creating extra memory pressure since its 2462 * main purpose is to replenish the store of free pages. 2463 */ 2464 if (vmd->vmd_severeset || curproc == pageproc || 2465 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2466 return (0); 2467 domain = vmd->vmd_domain; 2468 vm_domain_free_lock(vmd); 2469 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2470 (vm_page_t *)store); 2471 vm_domain_free_unlock(vmd); 2472 if (cnt != i) 2473 vm_domain_freecnt_inc(vmd, cnt - i); 2474 2475 return (i); 2476 } 2477 2478 static void 2479 vm_page_zone_release(void *arg, void **store, int cnt) 2480 { 2481 struct vm_domain *vmd; 2482 struct vm_pgcache *pgcache; 2483 vm_page_t m; 2484 int i; 2485 2486 pgcache = arg; 2487 vmd = VM_DOMAIN(pgcache->domain); 2488 vm_domain_free_lock(vmd); 2489 for (i = 0; i < cnt; i++) { 2490 m = (vm_page_t)store[i]; 2491 vm_phys_free_pages(m, 0); 2492 } 2493 vm_domain_free_unlock(vmd); 2494 vm_domain_freecnt_inc(vmd, cnt); 2495 } 2496 2497 #define VPSC_ANY 0 /* No restrictions. */ 2498 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2499 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2500 2501 /* 2502 * vm_page_scan_contig: 2503 * 2504 * Scan vm_page_array[] between the specified entries "m_start" and 2505 * "m_end" for a run of contiguous physical pages that satisfy the 2506 * specified conditions, and return the lowest page in the run. The 2507 * specified "alignment" determines the alignment of the lowest physical 2508 * page in the run. If the specified "boundary" is non-zero, then the 2509 * run of physical pages cannot span a physical address that is a 2510 * multiple of "boundary". 2511 * 2512 * "m_end" is never dereferenced, so it need not point to a vm_page 2513 * structure within vm_page_array[]. 2514 * 2515 * "npages" must be greater than zero. "m_start" and "m_end" must not 2516 * span a hole (or discontiguity) in the physical address space. Both 2517 * "alignment" and "boundary" must be a power of two. 2518 */ 2519 vm_page_t 2520 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2521 u_long alignment, vm_paddr_t boundary, int options) 2522 { 2523 vm_object_t object; 2524 vm_paddr_t pa; 2525 vm_page_t m, m_run; 2526 #if VM_NRESERVLEVEL > 0 2527 int level; 2528 #endif 2529 int m_inc, order, run_ext, run_len; 2530 2531 KASSERT(npages > 0, ("npages is 0")); 2532 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2533 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2534 m_run = NULL; 2535 run_len = 0; 2536 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2537 KASSERT((m->flags & PG_MARKER) == 0, 2538 ("page %p is PG_MARKER", m)); 2539 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2540 ("fictitious page %p has invalid ref count", m)); 2541 2542 /* 2543 * If the current page would be the start of a run, check its 2544 * physical address against the end, alignment, and boundary 2545 * conditions. If it doesn't satisfy these conditions, either 2546 * terminate the scan or advance to the next page that 2547 * satisfies the failed condition. 2548 */ 2549 if (run_len == 0) { 2550 KASSERT(m_run == NULL, ("m_run != NULL")); 2551 if (m + npages > m_end) 2552 break; 2553 pa = VM_PAGE_TO_PHYS(m); 2554 if ((pa & (alignment - 1)) != 0) { 2555 m_inc = atop(roundup2(pa, alignment) - pa); 2556 continue; 2557 } 2558 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2559 boundary) != 0) { 2560 m_inc = atop(roundup2(pa, boundary) - pa); 2561 continue; 2562 } 2563 } else 2564 KASSERT(m_run != NULL, ("m_run == NULL")); 2565 2566 retry: 2567 m_inc = 1; 2568 if (vm_page_wired(m)) 2569 run_ext = 0; 2570 #if VM_NRESERVLEVEL > 0 2571 else if ((level = vm_reserv_level(m)) >= 0 && 2572 (options & VPSC_NORESERV) != 0) { 2573 run_ext = 0; 2574 /* Advance to the end of the reservation. */ 2575 pa = VM_PAGE_TO_PHYS(m); 2576 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2577 pa); 2578 } 2579 #endif 2580 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2581 /* 2582 * The page is considered eligible for relocation if 2583 * and only if it could be laundered or reclaimed by 2584 * the page daemon. 2585 */ 2586 VM_OBJECT_RLOCK(object); 2587 if (object != m->object) { 2588 VM_OBJECT_RUNLOCK(object); 2589 goto retry; 2590 } 2591 /* Don't care: PG_NODUMP, PG_ZERO. */ 2592 if (object->type != OBJT_DEFAULT && 2593 object->type != OBJT_SWAP && 2594 object->type != OBJT_VNODE) { 2595 run_ext = 0; 2596 #if VM_NRESERVLEVEL > 0 2597 } else if ((options & VPSC_NOSUPER) != 0 && 2598 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2599 run_ext = 0; 2600 /* Advance to the end of the superpage. */ 2601 pa = VM_PAGE_TO_PHYS(m); 2602 m_inc = atop(roundup2(pa + 1, 2603 vm_reserv_size(level)) - pa); 2604 #endif 2605 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2606 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2607 /* 2608 * The page is allocated but eligible for 2609 * relocation. Extend the current run by one 2610 * page. 2611 */ 2612 KASSERT(pmap_page_get_memattr(m) == 2613 VM_MEMATTR_DEFAULT, 2614 ("page %p has an unexpected memattr", m)); 2615 KASSERT((m->oflags & (VPO_SWAPINPROG | 2616 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2617 ("page %p has unexpected oflags", m)); 2618 /* Don't care: PGA_NOSYNC. */ 2619 run_ext = 1; 2620 } else 2621 run_ext = 0; 2622 VM_OBJECT_RUNLOCK(object); 2623 #if VM_NRESERVLEVEL > 0 2624 } else if (level >= 0) { 2625 /* 2626 * The page is reserved but not yet allocated. In 2627 * other words, it is still free. Extend the current 2628 * run by one page. 2629 */ 2630 run_ext = 1; 2631 #endif 2632 } else if ((order = m->order) < VM_NFREEORDER) { 2633 /* 2634 * The page is enqueued in the physical memory 2635 * allocator's free page queues. Moreover, it is the 2636 * first page in a power-of-two-sized run of 2637 * contiguous free pages. Add these pages to the end 2638 * of the current run, and jump ahead. 2639 */ 2640 run_ext = 1 << order; 2641 m_inc = 1 << order; 2642 } else { 2643 /* 2644 * Skip the page for one of the following reasons: (1) 2645 * It is enqueued in the physical memory allocator's 2646 * free page queues. However, it is not the first 2647 * page in a run of contiguous free pages. (This case 2648 * rarely occurs because the scan is performed in 2649 * ascending order.) (2) It is not reserved, and it is 2650 * transitioning from free to allocated. (Conversely, 2651 * the transition from allocated to free for managed 2652 * pages is blocked by the page lock.) (3) It is 2653 * allocated but not contained by an object and not 2654 * wired, e.g., allocated by Xen's balloon driver. 2655 */ 2656 run_ext = 0; 2657 } 2658 2659 /* 2660 * Extend or reset the current run of pages. 2661 */ 2662 if (run_ext > 0) { 2663 if (run_len == 0) 2664 m_run = m; 2665 run_len += run_ext; 2666 } else { 2667 if (run_len > 0) { 2668 m_run = NULL; 2669 run_len = 0; 2670 } 2671 } 2672 } 2673 if (run_len >= npages) 2674 return (m_run); 2675 return (NULL); 2676 } 2677 2678 /* 2679 * vm_page_reclaim_run: 2680 * 2681 * Try to relocate each of the allocated virtual pages within the 2682 * specified run of physical pages to a new physical address. Free the 2683 * physical pages underlying the relocated virtual pages. A virtual page 2684 * is relocatable if and only if it could be laundered or reclaimed by 2685 * the page daemon. Whenever possible, a virtual page is relocated to a 2686 * physical address above "high". 2687 * 2688 * Returns 0 if every physical page within the run was already free or 2689 * just freed by a successful relocation. Otherwise, returns a non-zero 2690 * value indicating why the last attempt to relocate a virtual page was 2691 * unsuccessful. 2692 * 2693 * "req_class" must be an allocation class. 2694 */ 2695 static int 2696 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2697 vm_paddr_t high) 2698 { 2699 struct vm_domain *vmd; 2700 struct spglist free; 2701 vm_object_t object; 2702 vm_paddr_t pa; 2703 vm_page_t m, m_end, m_new; 2704 int error, order, req; 2705 2706 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2707 ("req_class is not an allocation class")); 2708 SLIST_INIT(&free); 2709 error = 0; 2710 m = m_run; 2711 m_end = m_run + npages; 2712 for (; error == 0 && m < m_end; m++) { 2713 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2714 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2715 2716 /* 2717 * Racily check for wirings. Races are handled once the object 2718 * lock is held and the page is unmapped. 2719 */ 2720 if (vm_page_wired(m)) 2721 error = EBUSY; 2722 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2723 /* 2724 * The page is relocated if and only if it could be 2725 * laundered or reclaimed by the page daemon. 2726 */ 2727 VM_OBJECT_WLOCK(object); 2728 /* Don't care: PG_NODUMP, PG_ZERO. */ 2729 if (m->object != object || 2730 (object->type != OBJT_DEFAULT && 2731 object->type != OBJT_SWAP && 2732 object->type != OBJT_VNODE)) 2733 error = EINVAL; 2734 else if (object->memattr != VM_MEMATTR_DEFAULT) 2735 error = EINVAL; 2736 else if (vm_page_queue(m) != PQ_NONE && 2737 vm_page_tryxbusy(m) != 0) { 2738 if (vm_page_wired(m)) { 2739 vm_page_xunbusy(m); 2740 error = EBUSY; 2741 goto unlock; 2742 } 2743 KASSERT(pmap_page_get_memattr(m) == 2744 VM_MEMATTR_DEFAULT, 2745 ("page %p has an unexpected memattr", m)); 2746 KASSERT(m->oflags == 0, 2747 ("page %p has unexpected oflags", m)); 2748 /* Don't care: PGA_NOSYNC. */ 2749 if (!vm_page_none_valid(m)) { 2750 /* 2751 * First, try to allocate a new page 2752 * that is above "high". Failing 2753 * that, try to allocate a new page 2754 * that is below "m_run". Allocate 2755 * the new page between the end of 2756 * "m_run" and "high" only as a last 2757 * resort. 2758 */ 2759 req = req_class | VM_ALLOC_NOOBJ; 2760 if ((m->flags & PG_NODUMP) != 0) 2761 req |= VM_ALLOC_NODUMP; 2762 if (trunc_page(high) != 2763 ~(vm_paddr_t)PAGE_MASK) { 2764 m_new = vm_page_alloc_contig( 2765 NULL, 0, req, 1, 2766 round_page(high), 2767 ~(vm_paddr_t)0, 2768 PAGE_SIZE, 0, 2769 VM_MEMATTR_DEFAULT); 2770 } else 2771 m_new = NULL; 2772 if (m_new == NULL) { 2773 pa = VM_PAGE_TO_PHYS(m_run); 2774 m_new = vm_page_alloc_contig( 2775 NULL, 0, req, 1, 2776 0, pa - 1, PAGE_SIZE, 0, 2777 VM_MEMATTR_DEFAULT); 2778 } 2779 if (m_new == NULL) { 2780 pa += ptoa(npages); 2781 m_new = vm_page_alloc_contig( 2782 NULL, 0, req, 1, 2783 pa, high, PAGE_SIZE, 0, 2784 VM_MEMATTR_DEFAULT); 2785 } 2786 if (m_new == NULL) { 2787 vm_page_xunbusy(m); 2788 error = ENOMEM; 2789 goto unlock; 2790 } 2791 2792 /* 2793 * Unmap the page and check for new 2794 * wirings that may have been acquired 2795 * through a pmap lookup. 2796 */ 2797 if (object->ref_count != 0 && 2798 !vm_page_try_remove_all(m)) { 2799 vm_page_xunbusy(m); 2800 vm_page_free(m_new); 2801 error = EBUSY; 2802 goto unlock; 2803 } 2804 2805 /* 2806 * Replace "m" with the new page. For 2807 * vm_page_replace(), "m" must be busy 2808 * and dequeued. Finally, change "m" 2809 * as if vm_page_free() was called. 2810 */ 2811 m_new->a.flags = m->a.flags & 2812 ~PGA_QUEUE_STATE_MASK; 2813 KASSERT(m_new->oflags == VPO_UNMANAGED, 2814 ("page %p is managed", m_new)); 2815 m_new->oflags = 0; 2816 pmap_copy_page(m, m_new); 2817 m_new->valid = m->valid; 2818 m_new->dirty = m->dirty; 2819 m->flags &= ~PG_ZERO; 2820 vm_page_dequeue(m); 2821 if (vm_page_replace_hold(m_new, object, 2822 m->pindex, m) && 2823 vm_page_free_prep(m)) 2824 SLIST_INSERT_HEAD(&free, m, 2825 plinks.s.ss); 2826 2827 /* 2828 * The new page must be deactivated 2829 * before the object is unlocked. 2830 */ 2831 vm_page_deactivate(m_new); 2832 } else { 2833 m->flags &= ~PG_ZERO; 2834 vm_page_dequeue(m); 2835 if (vm_page_free_prep(m)) 2836 SLIST_INSERT_HEAD(&free, m, 2837 plinks.s.ss); 2838 KASSERT(m->dirty == 0, 2839 ("page %p is dirty", m)); 2840 } 2841 } else 2842 error = EBUSY; 2843 unlock: 2844 VM_OBJECT_WUNLOCK(object); 2845 } else { 2846 MPASS(vm_phys_domain(m) == domain); 2847 vmd = VM_DOMAIN(domain); 2848 vm_domain_free_lock(vmd); 2849 order = m->order; 2850 if (order < VM_NFREEORDER) { 2851 /* 2852 * The page is enqueued in the physical memory 2853 * allocator's free page queues. Moreover, it 2854 * is the first page in a power-of-two-sized 2855 * run of contiguous free pages. Jump ahead 2856 * to the last page within that run, and 2857 * continue from there. 2858 */ 2859 m += (1 << order) - 1; 2860 } 2861 #if VM_NRESERVLEVEL > 0 2862 else if (vm_reserv_is_page_free(m)) 2863 order = 0; 2864 #endif 2865 vm_domain_free_unlock(vmd); 2866 if (order == VM_NFREEORDER) 2867 error = EINVAL; 2868 } 2869 } 2870 if ((m = SLIST_FIRST(&free)) != NULL) { 2871 int cnt; 2872 2873 vmd = VM_DOMAIN(domain); 2874 cnt = 0; 2875 vm_domain_free_lock(vmd); 2876 do { 2877 MPASS(vm_phys_domain(m) == domain); 2878 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2879 vm_phys_free_pages(m, 0); 2880 cnt++; 2881 } while ((m = SLIST_FIRST(&free)) != NULL); 2882 vm_domain_free_unlock(vmd); 2883 vm_domain_freecnt_inc(vmd, cnt); 2884 } 2885 return (error); 2886 } 2887 2888 #define NRUNS 16 2889 2890 CTASSERT(powerof2(NRUNS)); 2891 2892 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2893 2894 #define MIN_RECLAIM 8 2895 2896 /* 2897 * vm_page_reclaim_contig: 2898 * 2899 * Reclaim allocated, contiguous physical memory satisfying the specified 2900 * conditions by relocating the virtual pages using that physical memory. 2901 * Returns true if reclamation is successful and false otherwise. Since 2902 * relocation requires the allocation of physical pages, reclamation may 2903 * fail due to a shortage of free pages. When reclamation fails, callers 2904 * are expected to perform vm_wait() before retrying a failed allocation 2905 * operation, e.g., vm_page_alloc_contig(). 2906 * 2907 * The caller must always specify an allocation class through "req". 2908 * 2909 * allocation classes: 2910 * VM_ALLOC_NORMAL normal process request 2911 * VM_ALLOC_SYSTEM system *really* needs a page 2912 * VM_ALLOC_INTERRUPT interrupt time request 2913 * 2914 * The optional allocation flags are ignored. 2915 * 2916 * "npages" must be greater than zero. Both "alignment" and "boundary" 2917 * must be a power of two. 2918 */ 2919 bool 2920 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2921 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2922 { 2923 struct vm_domain *vmd; 2924 vm_paddr_t curr_low; 2925 vm_page_t m_run, m_runs[NRUNS]; 2926 u_long count, reclaimed; 2927 int error, i, options, req_class; 2928 2929 KASSERT(npages > 0, ("npages is 0")); 2930 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2931 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2932 req_class = req & VM_ALLOC_CLASS_MASK; 2933 2934 /* 2935 * The page daemon is allowed to dig deeper into the free page list. 2936 */ 2937 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2938 req_class = VM_ALLOC_SYSTEM; 2939 2940 /* 2941 * Return if the number of free pages cannot satisfy the requested 2942 * allocation. 2943 */ 2944 vmd = VM_DOMAIN(domain); 2945 count = vmd->vmd_free_count; 2946 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2947 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2948 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2949 return (false); 2950 2951 /* 2952 * Scan up to three times, relaxing the restrictions ("options") on 2953 * the reclamation of reservations and superpages each time. 2954 */ 2955 for (options = VPSC_NORESERV;;) { 2956 /* 2957 * Find the highest runs that satisfy the given constraints 2958 * and restrictions, and record them in "m_runs". 2959 */ 2960 curr_low = low; 2961 count = 0; 2962 for (;;) { 2963 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2964 high, alignment, boundary, options); 2965 if (m_run == NULL) 2966 break; 2967 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2968 m_runs[RUN_INDEX(count)] = m_run; 2969 count++; 2970 } 2971 2972 /* 2973 * Reclaim the highest runs in LIFO (descending) order until 2974 * the number of reclaimed pages, "reclaimed", is at least 2975 * MIN_RECLAIM. Reset "reclaimed" each time because each 2976 * reclamation is idempotent, and runs will (likely) recur 2977 * from one scan to the next as restrictions are relaxed. 2978 */ 2979 reclaimed = 0; 2980 for (i = 0; count > 0 && i < NRUNS; i++) { 2981 count--; 2982 m_run = m_runs[RUN_INDEX(count)]; 2983 error = vm_page_reclaim_run(req_class, domain, npages, 2984 m_run, high); 2985 if (error == 0) { 2986 reclaimed += npages; 2987 if (reclaimed >= MIN_RECLAIM) 2988 return (true); 2989 } 2990 } 2991 2992 /* 2993 * Either relax the restrictions on the next scan or return if 2994 * the last scan had no restrictions. 2995 */ 2996 if (options == VPSC_NORESERV) 2997 options = VPSC_NOSUPER; 2998 else if (options == VPSC_NOSUPER) 2999 options = VPSC_ANY; 3000 else if (options == VPSC_ANY) 3001 return (reclaimed != 0); 3002 } 3003 } 3004 3005 bool 3006 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3007 u_long alignment, vm_paddr_t boundary) 3008 { 3009 struct vm_domainset_iter di; 3010 int domain; 3011 bool ret; 3012 3013 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3014 do { 3015 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 3016 high, alignment, boundary); 3017 if (ret) 3018 break; 3019 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3020 3021 return (ret); 3022 } 3023 3024 /* 3025 * Set the domain in the appropriate page level domainset. 3026 */ 3027 void 3028 vm_domain_set(struct vm_domain *vmd) 3029 { 3030 3031 mtx_lock(&vm_domainset_lock); 3032 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3033 vmd->vmd_minset = 1; 3034 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3035 } 3036 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3037 vmd->vmd_severeset = 1; 3038 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3039 } 3040 mtx_unlock(&vm_domainset_lock); 3041 } 3042 3043 /* 3044 * Clear the domain from the appropriate page level domainset. 3045 */ 3046 void 3047 vm_domain_clear(struct vm_domain *vmd) 3048 { 3049 3050 mtx_lock(&vm_domainset_lock); 3051 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3052 vmd->vmd_minset = 0; 3053 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3054 if (vm_min_waiters != 0) { 3055 vm_min_waiters = 0; 3056 wakeup(&vm_min_domains); 3057 } 3058 } 3059 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3060 vmd->vmd_severeset = 0; 3061 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3062 if (vm_severe_waiters != 0) { 3063 vm_severe_waiters = 0; 3064 wakeup(&vm_severe_domains); 3065 } 3066 } 3067 3068 /* 3069 * If pageout daemon needs pages, then tell it that there are 3070 * some free. 3071 */ 3072 if (vmd->vmd_pageout_pages_needed && 3073 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3074 wakeup(&vmd->vmd_pageout_pages_needed); 3075 vmd->vmd_pageout_pages_needed = 0; 3076 } 3077 3078 /* See comments in vm_wait_doms(). */ 3079 if (vm_pageproc_waiters) { 3080 vm_pageproc_waiters = 0; 3081 wakeup(&vm_pageproc_waiters); 3082 } 3083 mtx_unlock(&vm_domainset_lock); 3084 } 3085 3086 /* 3087 * Wait for free pages to exceed the min threshold globally. 3088 */ 3089 void 3090 vm_wait_min(void) 3091 { 3092 3093 mtx_lock(&vm_domainset_lock); 3094 while (vm_page_count_min()) { 3095 vm_min_waiters++; 3096 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3097 } 3098 mtx_unlock(&vm_domainset_lock); 3099 } 3100 3101 /* 3102 * Wait for free pages to exceed the severe threshold globally. 3103 */ 3104 void 3105 vm_wait_severe(void) 3106 { 3107 3108 mtx_lock(&vm_domainset_lock); 3109 while (vm_page_count_severe()) { 3110 vm_severe_waiters++; 3111 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3112 "vmwait", 0); 3113 } 3114 mtx_unlock(&vm_domainset_lock); 3115 } 3116 3117 u_int 3118 vm_wait_count(void) 3119 { 3120 3121 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3122 } 3123 3124 void 3125 vm_wait_doms(const domainset_t *wdoms) 3126 { 3127 3128 /* 3129 * We use racey wakeup synchronization to avoid expensive global 3130 * locking for the pageproc when sleeping with a non-specific vm_wait. 3131 * To handle this, we only sleep for one tick in this instance. It 3132 * is expected that most allocations for the pageproc will come from 3133 * kmem or vm_page_grab* which will use the more specific and 3134 * race-free vm_wait_domain(). 3135 */ 3136 if (curproc == pageproc) { 3137 mtx_lock(&vm_domainset_lock); 3138 vm_pageproc_waiters++; 3139 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, 3140 "pageprocwait", 1); 3141 } else { 3142 /* 3143 * XXX Ideally we would wait only until the allocation could 3144 * be satisfied. This condition can cause new allocators to 3145 * consume all freed pages while old allocators wait. 3146 */ 3147 mtx_lock(&vm_domainset_lock); 3148 if (vm_page_count_min_set(wdoms)) { 3149 vm_min_waiters++; 3150 msleep(&vm_min_domains, &vm_domainset_lock, 3151 PVM | PDROP, "vmwait", 0); 3152 } else 3153 mtx_unlock(&vm_domainset_lock); 3154 } 3155 } 3156 3157 /* 3158 * vm_wait_domain: 3159 * 3160 * Sleep until free pages are available for allocation. 3161 * - Called in various places after failed memory allocations. 3162 */ 3163 void 3164 vm_wait_domain(int domain) 3165 { 3166 struct vm_domain *vmd; 3167 domainset_t wdom; 3168 3169 vmd = VM_DOMAIN(domain); 3170 vm_domain_free_assert_unlocked(vmd); 3171 3172 if (curproc == pageproc) { 3173 mtx_lock(&vm_domainset_lock); 3174 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3175 vmd->vmd_pageout_pages_needed = 1; 3176 msleep(&vmd->vmd_pageout_pages_needed, 3177 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3178 } else 3179 mtx_unlock(&vm_domainset_lock); 3180 } else { 3181 if (pageproc == NULL) 3182 panic("vm_wait in early boot"); 3183 DOMAINSET_ZERO(&wdom); 3184 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3185 vm_wait_doms(&wdom); 3186 } 3187 } 3188 3189 /* 3190 * vm_wait: 3191 * 3192 * Sleep until free pages are available for allocation in the 3193 * affinity domains of the obj. If obj is NULL, the domain set 3194 * for the calling thread is used. 3195 * Called in various places after failed memory allocations. 3196 */ 3197 void 3198 vm_wait(vm_object_t obj) 3199 { 3200 struct domainset *d; 3201 3202 d = NULL; 3203 3204 /* 3205 * Carefully fetch pointers only once: the struct domainset 3206 * itself is ummutable but the pointer might change. 3207 */ 3208 if (obj != NULL) 3209 d = obj->domain.dr_policy; 3210 if (d == NULL) 3211 d = curthread->td_domain.dr_policy; 3212 3213 vm_wait_doms(&d->ds_mask); 3214 } 3215 3216 /* 3217 * vm_domain_alloc_fail: 3218 * 3219 * Called when a page allocation function fails. Informs the 3220 * pagedaemon and performs the requested wait. Requires the 3221 * domain_free and object lock on entry. Returns with the 3222 * object lock held and free lock released. Returns an error when 3223 * retry is necessary. 3224 * 3225 */ 3226 static int 3227 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3228 { 3229 3230 vm_domain_free_assert_unlocked(vmd); 3231 3232 atomic_add_int(&vmd->vmd_pageout_deficit, 3233 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3234 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3235 if (object != NULL) 3236 VM_OBJECT_WUNLOCK(object); 3237 vm_wait_domain(vmd->vmd_domain); 3238 if (object != NULL) 3239 VM_OBJECT_WLOCK(object); 3240 if (req & VM_ALLOC_WAITOK) 3241 return (EAGAIN); 3242 } 3243 3244 return (0); 3245 } 3246 3247 /* 3248 * vm_waitpfault: 3249 * 3250 * Sleep until free pages are available for allocation. 3251 * - Called only in vm_fault so that processes page faulting 3252 * can be easily tracked. 3253 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3254 * processes will be able to grab memory first. Do not change 3255 * this balance without careful testing first. 3256 */ 3257 void 3258 vm_waitpfault(struct domainset *dset, int timo) 3259 { 3260 3261 /* 3262 * XXX Ideally we would wait only until the allocation could 3263 * be satisfied. This condition can cause new allocators to 3264 * consume all freed pages while old allocators wait. 3265 */ 3266 mtx_lock(&vm_domainset_lock); 3267 if (vm_page_count_min_set(&dset->ds_mask)) { 3268 vm_min_waiters++; 3269 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3270 "pfault", timo); 3271 } else 3272 mtx_unlock(&vm_domainset_lock); 3273 } 3274 3275 static struct vm_pagequeue * 3276 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3277 { 3278 3279 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3280 } 3281 3282 #ifdef INVARIANTS 3283 static struct vm_pagequeue * 3284 vm_page_pagequeue(vm_page_t m) 3285 { 3286 3287 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3288 } 3289 #endif 3290 3291 static __always_inline bool 3292 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3293 { 3294 vm_page_astate_t tmp; 3295 3296 tmp = *old; 3297 do { 3298 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3299 return (true); 3300 counter_u64_add(pqstate_commit_retries, 1); 3301 } while (old->_bits == tmp._bits); 3302 3303 return (false); 3304 } 3305 3306 /* 3307 * Do the work of committing a queue state update that moves the page out of 3308 * its current queue. 3309 */ 3310 static bool 3311 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3312 vm_page_astate_t *old, vm_page_astate_t new) 3313 { 3314 vm_page_t next; 3315 3316 vm_pagequeue_assert_locked(pq); 3317 KASSERT(vm_page_pagequeue(m) == pq, 3318 ("%s: queue %p does not match page %p", __func__, pq, m)); 3319 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3320 ("%s: invalid queue indices %d %d", 3321 __func__, old->queue, new.queue)); 3322 3323 /* 3324 * Once the queue index of the page changes there is nothing 3325 * synchronizing with further updates to the page's physical 3326 * queue state. Therefore we must speculatively remove the page 3327 * from the queue now and be prepared to roll back if the queue 3328 * state update fails. If the page is not physically enqueued then 3329 * we just update its queue index. 3330 */ 3331 if ((old->flags & PGA_ENQUEUED) != 0) { 3332 new.flags &= ~PGA_ENQUEUED; 3333 next = TAILQ_NEXT(m, plinks.q); 3334 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3335 vm_pagequeue_cnt_dec(pq); 3336 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3337 if (next == NULL) 3338 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3339 else 3340 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3341 vm_pagequeue_cnt_inc(pq); 3342 return (false); 3343 } else { 3344 return (true); 3345 } 3346 } else { 3347 return (vm_page_pqstate_fcmpset(m, old, new)); 3348 } 3349 } 3350 3351 static bool 3352 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3353 vm_page_astate_t new) 3354 { 3355 struct vm_pagequeue *pq; 3356 vm_page_astate_t as; 3357 bool ret; 3358 3359 pq = _vm_page_pagequeue(m, old->queue); 3360 3361 /* 3362 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3363 * corresponding page queue lock is held. 3364 */ 3365 vm_pagequeue_lock(pq); 3366 as = vm_page_astate_load(m); 3367 if (__predict_false(as._bits != old->_bits)) { 3368 *old = as; 3369 ret = false; 3370 } else { 3371 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3372 } 3373 vm_pagequeue_unlock(pq); 3374 return (ret); 3375 } 3376 3377 /* 3378 * Commit a queue state update that enqueues or requeues a page. 3379 */ 3380 static bool 3381 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3382 vm_page_astate_t *old, vm_page_astate_t new) 3383 { 3384 struct vm_domain *vmd; 3385 3386 vm_pagequeue_assert_locked(pq); 3387 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3388 ("%s: invalid queue indices %d %d", 3389 __func__, old->queue, new.queue)); 3390 3391 new.flags |= PGA_ENQUEUED; 3392 if (!vm_page_pqstate_fcmpset(m, old, new)) 3393 return (false); 3394 3395 if ((old->flags & PGA_ENQUEUED) != 0) 3396 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3397 else 3398 vm_pagequeue_cnt_inc(pq); 3399 3400 /* 3401 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3402 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3403 * applied, even if it was set first. 3404 */ 3405 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3406 vmd = vm_pagequeue_domain(m); 3407 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3408 ("%s: invalid page queue for page %p", __func__, m)); 3409 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3410 } else { 3411 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3412 } 3413 return (true); 3414 } 3415 3416 /* 3417 * Commit a queue state update that encodes a request for a deferred queue 3418 * operation. 3419 */ 3420 static bool 3421 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3422 vm_page_astate_t new) 3423 { 3424 3425 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3426 ("%s: invalid state, queue %d flags %x", 3427 __func__, new.queue, new.flags)); 3428 3429 if (old->_bits != new._bits && 3430 !vm_page_pqstate_fcmpset(m, old, new)) 3431 return (false); 3432 vm_page_pqbatch_submit(m, new.queue); 3433 return (true); 3434 } 3435 3436 /* 3437 * A generic queue state update function. This handles more cases than the 3438 * specialized functions above. 3439 */ 3440 bool 3441 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3442 { 3443 3444 if (old->_bits == new._bits) 3445 return (true); 3446 3447 if (old->queue != PQ_NONE && new.queue != old->queue) { 3448 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3449 return (false); 3450 if (new.queue != PQ_NONE) 3451 vm_page_pqbatch_submit(m, new.queue); 3452 } else { 3453 if (!vm_page_pqstate_fcmpset(m, old, new)) 3454 return (false); 3455 if (new.queue != PQ_NONE && 3456 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3457 vm_page_pqbatch_submit(m, new.queue); 3458 } 3459 return (true); 3460 } 3461 3462 /* 3463 * Apply deferred queue state updates to a page. 3464 */ 3465 static inline void 3466 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3467 { 3468 vm_page_astate_t new, old; 3469 3470 CRITICAL_ASSERT(curthread); 3471 vm_pagequeue_assert_locked(pq); 3472 KASSERT(queue < PQ_COUNT, 3473 ("%s: invalid queue index %d", __func__, queue)); 3474 KASSERT(pq == _vm_page_pagequeue(m, queue), 3475 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3476 3477 for (old = vm_page_astate_load(m);;) { 3478 if (__predict_false(old.queue != queue || 3479 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3480 counter_u64_add(queue_nops, 1); 3481 break; 3482 } 3483 KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0, 3484 ("%s: page %p has unexpected queue state", __func__, m)); 3485 3486 new = old; 3487 if ((old.flags & PGA_DEQUEUE) != 0) { 3488 new.flags &= ~PGA_QUEUE_OP_MASK; 3489 new.queue = PQ_NONE; 3490 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3491 m, &old, new))) { 3492 counter_u64_add(queue_ops, 1); 3493 break; 3494 } 3495 } else { 3496 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3497 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3498 m, &old, new))) { 3499 counter_u64_add(queue_ops, 1); 3500 break; 3501 } 3502 } 3503 } 3504 } 3505 3506 static void 3507 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3508 uint8_t queue) 3509 { 3510 int i; 3511 3512 for (i = 0; i < bq->bq_cnt; i++) 3513 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3514 vm_batchqueue_init(bq); 3515 } 3516 3517 /* 3518 * vm_page_pqbatch_submit: [ internal use only ] 3519 * 3520 * Enqueue a page in the specified page queue's batched work queue. 3521 * The caller must have encoded the requested operation in the page 3522 * structure's a.flags field. 3523 */ 3524 void 3525 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3526 { 3527 struct vm_batchqueue *bq; 3528 struct vm_pagequeue *pq; 3529 int domain; 3530 3531 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3532 ("page %p is unmanaged", m)); 3533 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3534 3535 domain = vm_phys_domain(m); 3536 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 3537 3538 critical_enter(); 3539 bq = DPCPU_PTR(pqbatch[domain][queue]); 3540 if (vm_batchqueue_insert(bq, m)) { 3541 critical_exit(); 3542 return; 3543 } 3544 critical_exit(); 3545 vm_pagequeue_lock(pq); 3546 critical_enter(); 3547 bq = DPCPU_PTR(pqbatch[domain][queue]); 3548 vm_pqbatch_process(pq, bq, queue); 3549 vm_pqbatch_process_page(pq, m, queue); 3550 vm_pagequeue_unlock(pq); 3551 critical_exit(); 3552 } 3553 3554 /* 3555 * vm_page_pqbatch_drain: [ internal use only ] 3556 * 3557 * Force all per-CPU page queue batch queues to be drained. This is 3558 * intended for use in severe memory shortages, to ensure that pages 3559 * do not remain stuck in the batch queues. 3560 */ 3561 void 3562 vm_page_pqbatch_drain(void) 3563 { 3564 struct thread *td; 3565 struct vm_domain *vmd; 3566 struct vm_pagequeue *pq; 3567 int cpu, domain, queue; 3568 3569 td = curthread; 3570 CPU_FOREACH(cpu) { 3571 thread_lock(td); 3572 sched_bind(td, cpu); 3573 thread_unlock(td); 3574 3575 for (domain = 0; domain < vm_ndomains; domain++) { 3576 vmd = VM_DOMAIN(domain); 3577 for (queue = 0; queue < PQ_COUNT; queue++) { 3578 pq = &vmd->vmd_pagequeues[queue]; 3579 vm_pagequeue_lock(pq); 3580 critical_enter(); 3581 vm_pqbatch_process(pq, 3582 DPCPU_PTR(pqbatch[domain][queue]), queue); 3583 critical_exit(); 3584 vm_pagequeue_unlock(pq); 3585 } 3586 } 3587 } 3588 thread_lock(td); 3589 sched_unbind(td); 3590 thread_unlock(td); 3591 } 3592 3593 /* 3594 * vm_page_dequeue_deferred: [ internal use only ] 3595 * 3596 * Request removal of the given page from its current page 3597 * queue. Physical removal from the queue may be deferred 3598 * indefinitely. 3599 * 3600 * The page must be locked. 3601 */ 3602 void 3603 vm_page_dequeue_deferred(vm_page_t m) 3604 { 3605 vm_page_astate_t new, old; 3606 3607 old = vm_page_astate_load(m); 3608 do { 3609 if (old.queue == PQ_NONE) { 3610 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3611 ("%s: page %p has unexpected queue state", 3612 __func__, m)); 3613 break; 3614 } 3615 new = old; 3616 new.flags |= PGA_DEQUEUE; 3617 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3618 } 3619 3620 /* 3621 * vm_page_dequeue: 3622 * 3623 * Remove the page from whichever page queue it's in, if any, before 3624 * returning. 3625 */ 3626 void 3627 vm_page_dequeue(vm_page_t m) 3628 { 3629 vm_page_astate_t new, old; 3630 3631 old = vm_page_astate_load(m); 3632 do { 3633 if (old.queue == PQ_NONE) { 3634 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3635 ("%s: page %p has unexpected queue state", 3636 __func__, m)); 3637 break; 3638 } 3639 new = old; 3640 new.flags &= ~PGA_QUEUE_OP_MASK; 3641 new.queue = PQ_NONE; 3642 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 3643 3644 } 3645 3646 /* 3647 * Schedule the given page for insertion into the specified page queue. 3648 * Physical insertion of the page may be deferred indefinitely. 3649 */ 3650 static void 3651 vm_page_enqueue(vm_page_t m, uint8_t queue) 3652 { 3653 3654 KASSERT(m->a.queue == PQ_NONE && 3655 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 3656 ("%s: page %p is already enqueued", __func__, m)); 3657 KASSERT(m->ref_count > 0, 3658 ("%s: page %p does not carry any references", __func__, m)); 3659 3660 m->a.queue = queue; 3661 if ((m->a.flags & PGA_REQUEUE) == 0) 3662 vm_page_aflag_set(m, PGA_REQUEUE); 3663 vm_page_pqbatch_submit(m, queue); 3664 } 3665 3666 /* 3667 * vm_page_free_prep: 3668 * 3669 * Prepares the given page to be put on the free list, 3670 * disassociating it from any VM object. The caller may return 3671 * the page to the free list only if this function returns true. 3672 * 3673 * The object must be locked. The page must be locked if it is 3674 * managed. 3675 */ 3676 static bool 3677 vm_page_free_prep(vm_page_t m) 3678 { 3679 3680 /* 3681 * Synchronize with threads that have dropped a reference to this 3682 * page. 3683 */ 3684 atomic_thread_fence_acq(); 3685 3686 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3687 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3688 uint64_t *p; 3689 int i; 3690 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3691 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3692 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3693 m, i, (uintmax_t)*p)); 3694 } 3695 #endif 3696 if ((m->oflags & VPO_UNMANAGED) == 0) { 3697 KASSERT(!pmap_page_is_mapped(m), 3698 ("vm_page_free_prep: freeing mapped page %p", m)); 3699 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3700 ("vm_page_free_prep: mapping flags set in page %p", m)); 3701 } else { 3702 KASSERT(m->a.queue == PQ_NONE, 3703 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3704 } 3705 VM_CNT_INC(v_tfree); 3706 3707 if (m->object != NULL) { 3708 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 3709 ((m->object->flags & OBJ_UNMANAGED) != 0), 3710 ("vm_page_free_prep: managed flag mismatch for page %p", 3711 m)); 3712 vm_page_assert_xbusied(m); 3713 3714 /* 3715 * The object reference can be released without an atomic 3716 * operation. 3717 */ 3718 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3719 m->ref_count == VPRC_OBJREF, 3720 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3721 m, m->ref_count)); 3722 vm_page_object_remove(m); 3723 m->ref_count -= VPRC_OBJREF; 3724 } else 3725 vm_page_assert_unbusied(m); 3726 3727 vm_page_busy_free(m); 3728 3729 /* 3730 * If fictitious remove object association and 3731 * return. 3732 */ 3733 if ((m->flags & PG_FICTITIOUS) != 0) { 3734 KASSERT(m->ref_count == 1, 3735 ("fictitious page %p is referenced", m)); 3736 KASSERT(m->a.queue == PQ_NONE, 3737 ("fictitious page %p is queued", m)); 3738 return (false); 3739 } 3740 3741 /* 3742 * Pages need not be dequeued before they are returned to the physical 3743 * memory allocator, but they must at least be marked for a deferred 3744 * dequeue. 3745 */ 3746 if ((m->oflags & VPO_UNMANAGED) == 0) 3747 vm_page_dequeue_deferred(m); 3748 3749 m->valid = 0; 3750 vm_page_undirty(m); 3751 3752 if (m->ref_count != 0) 3753 panic("vm_page_free_prep: page %p has references", m); 3754 3755 /* 3756 * Restore the default memory attribute to the page. 3757 */ 3758 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3759 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3760 3761 #if VM_NRESERVLEVEL > 0 3762 /* 3763 * Determine whether the page belongs to a reservation. If the page was 3764 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3765 * as an optimization, we avoid the check in that case. 3766 */ 3767 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3768 return (false); 3769 #endif 3770 3771 return (true); 3772 } 3773 3774 /* 3775 * vm_page_free_toq: 3776 * 3777 * Returns the given page to the free list, disassociating it 3778 * from any VM object. 3779 * 3780 * The object must be locked. The page must be locked if it is 3781 * managed. 3782 */ 3783 static void 3784 vm_page_free_toq(vm_page_t m) 3785 { 3786 struct vm_domain *vmd; 3787 uma_zone_t zone; 3788 3789 if (!vm_page_free_prep(m)) 3790 return; 3791 3792 vmd = vm_pagequeue_domain(m); 3793 zone = vmd->vmd_pgcache[m->pool].zone; 3794 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 3795 uma_zfree(zone, m); 3796 return; 3797 } 3798 vm_domain_free_lock(vmd); 3799 vm_phys_free_pages(m, 0); 3800 vm_domain_free_unlock(vmd); 3801 vm_domain_freecnt_inc(vmd, 1); 3802 } 3803 3804 /* 3805 * vm_page_free_pages_toq: 3806 * 3807 * Returns a list of pages to the free list, disassociating it 3808 * from any VM object. In other words, this is equivalent to 3809 * calling vm_page_free_toq() for each page of a list of VM objects. 3810 * 3811 * The objects must be locked. The pages must be locked if it is 3812 * managed. 3813 */ 3814 void 3815 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3816 { 3817 vm_page_t m; 3818 int count; 3819 3820 if (SLIST_EMPTY(free)) 3821 return; 3822 3823 count = 0; 3824 while ((m = SLIST_FIRST(free)) != NULL) { 3825 count++; 3826 SLIST_REMOVE_HEAD(free, plinks.s.ss); 3827 vm_page_free_toq(m); 3828 } 3829 3830 if (update_wire_count) 3831 vm_wire_sub(count); 3832 } 3833 3834 /* 3835 * Mark this page as wired down, preventing reclamation by the page daemon 3836 * or when the containing object is destroyed. 3837 */ 3838 void 3839 vm_page_wire(vm_page_t m) 3840 { 3841 u_int old; 3842 3843 KASSERT(m->object != NULL, 3844 ("vm_page_wire: page %p does not belong to an object", m)); 3845 if (!vm_page_busied(m) && !vm_object_busied(m->object)) 3846 VM_OBJECT_ASSERT_LOCKED(m->object); 3847 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 3848 VPRC_WIRE_COUNT(m->ref_count) >= 1, 3849 ("vm_page_wire: fictitious page %p has zero wirings", m)); 3850 3851 old = atomic_fetchadd_int(&m->ref_count, 1); 3852 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 3853 ("vm_page_wire: counter overflow for page %p", m)); 3854 if (VPRC_WIRE_COUNT(old) == 0) { 3855 if ((m->oflags & VPO_UNMANAGED) == 0) 3856 vm_page_aflag_set(m, PGA_DEQUEUE); 3857 vm_wire_add(1); 3858 } 3859 } 3860 3861 /* 3862 * Attempt to wire a mapped page following a pmap lookup of that page. 3863 * This may fail if a thread is concurrently tearing down mappings of the page. 3864 * The transient failure is acceptable because it translates to the 3865 * failure of the caller pmap_extract_and_hold(), which should be then 3866 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 3867 */ 3868 bool 3869 vm_page_wire_mapped(vm_page_t m) 3870 { 3871 u_int old; 3872 3873 old = m->ref_count; 3874 do { 3875 KASSERT(old > 0, 3876 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 3877 if ((old & VPRC_BLOCKED) != 0) 3878 return (false); 3879 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 3880 3881 if (VPRC_WIRE_COUNT(old) == 0) { 3882 if ((m->oflags & VPO_UNMANAGED) == 0) 3883 vm_page_aflag_set(m, PGA_DEQUEUE); 3884 vm_wire_add(1); 3885 } 3886 return (true); 3887 } 3888 3889 /* 3890 * Release a wiring reference to a managed page. If the page still belongs to 3891 * an object, update its position in the page queues to reflect the reference. 3892 * If the wiring was the last reference to the page, free the page. 3893 */ 3894 static void 3895 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 3896 { 3897 u_int old; 3898 3899 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3900 ("%s: page %p is unmanaged", __func__, m)); 3901 3902 /* 3903 * Update LRU state before releasing the wiring reference. 3904 * Use a release store when updating the reference count to 3905 * synchronize with vm_page_free_prep(). 3906 */ 3907 old = m->ref_count; 3908 do { 3909 KASSERT(VPRC_WIRE_COUNT(old) > 0, 3910 ("vm_page_unwire: wire count underflow for page %p", m)); 3911 3912 if (old > VPRC_OBJREF + 1) { 3913 /* 3914 * The page has at least one other wiring reference. An 3915 * earlier iteration of this loop may have called 3916 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 3917 * re-set it if necessary. 3918 */ 3919 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 3920 vm_page_aflag_set(m, PGA_DEQUEUE); 3921 } else if (old == VPRC_OBJREF + 1) { 3922 /* 3923 * This is the last wiring. Clear PGA_DEQUEUE and 3924 * update the page's queue state to reflect the 3925 * reference. If the page does not belong to an object 3926 * (i.e., the VPRC_OBJREF bit is clear), we only need to 3927 * clear leftover queue state. 3928 */ 3929 vm_page_release_toq(m, nqueue, false); 3930 } else if (old == 1) { 3931 vm_page_aflag_clear(m, PGA_DEQUEUE); 3932 } 3933 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 3934 3935 if (VPRC_WIRE_COUNT(old) == 1) { 3936 vm_wire_sub(1); 3937 if (old == 1) 3938 vm_page_free(m); 3939 } 3940 } 3941 3942 /* 3943 * Release one wiring of the specified page, potentially allowing it to be 3944 * paged out. 3945 * 3946 * Only managed pages belonging to an object can be paged out. If the number 3947 * of wirings transitions to zero and the page is eligible for page out, then 3948 * the page is added to the specified paging queue. If the released wiring 3949 * represented the last reference to the page, the page is freed. 3950 * 3951 * A managed page must be locked. 3952 */ 3953 void 3954 vm_page_unwire(vm_page_t m, uint8_t nqueue) 3955 { 3956 3957 KASSERT(nqueue < PQ_COUNT, 3958 ("vm_page_unwire: invalid queue %u request for page %p", 3959 nqueue, m)); 3960 3961 if ((m->oflags & VPO_UNMANAGED) != 0) { 3962 if (vm_page_unwire_noq(m) && m->ref_count == 0) 3963 vm_page_free(m); 3964 return; 3965 } 3966 vm_page_unwire_managed(m, nqueue, false); 3967 } 3968 3969 /* 3970 * Unwire a page without (re-)inserting it into a page queue. It is up 3971 * to the caller to enqueue, requeue, or free the page as appropriate. 3972 * In most cases involving managed pages, vm_page_unwire() should be used 3973 * instead. 3974 */ 3975 bool 3976 vm_page_unwire_noq(vm_page_t m) 3977 { 3978 u_int old; 3979 3980 old = vm_page_drop(m, 1); 3981 KASSERT(VPRC_WIRE_COUNT(old) != 0, 3982 ("vm_page_unref: counter underflow for page %p", m)); 3983 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 3984 ("vm_page_unref: missing ref on fictitious page %p", m)); 3985 3986 if (VPRC_WIRE_COUNT(old) > 1) 3987 return (false); 3988 if ((m->oflags & VPO_UNMANAGED) == 0) 3989 vm_page_aflag_clear(m, PGA_DEQUEUE); 3990 vm_wire_sub(1); 3991 return (true); 3992 } 3993 3994 /* 3995 * Ensure that the page ends up in the specified page queue. If the page is 3996 * active or being moved to the active queue, ensure that its act_count is 3997 * at least ACT_INIT but do not otherwise mess with it. 3998 * 3999 * A managed page must be locked. 4000 */ 4001 static __always_inline void 4002 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4003 { 4004 vm_page_astate_t old, new; 4005 4006 KASSERT(m->ref_count > 0, 4007 ("%s: page %p does not carry any references", __func__, m)); 4008 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4009 ("%s: invalid flags %x", __func__, nflag)); 4010 4011 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4012 return; 4013 4014 old = vm_page_astate_load(m); 4015 do { 4016 if ((old.flags & PGA_DEQUEUE) != 0) 4017 break; 4018 new = old; 4019 new.flags &= ~PGA_QUEUE_OP_MASK; 4020 if (nqueue == PQ_ACTIVE) 4021 new.act_count = max(old.act_count, ACT_INIT); 4022 if (old.queue == nqueue) { 4023 if (nqueue != PQ_ACTIVE) 4024 new.flags |= nflag; 4025 } else { 4026 new.flags |= nflag; 4027 new.queue = nqueue; 4028 } 4029 } while (!vm_page_pqstate_commit(m, &old, new)); 4030 } 4031 4032 /* 4033 * Put the specified page on the active list (if appropriate). 4034 */ 4035 void 4036 vm_page_activate(vm_page_t m) 4037 { 4038 4039 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4040 } 4041 4042 /* 4043 * Move the specified page to the tail of the inactive queue, or requeue 4044 * the page if it is already in the inactive queue. 4045 */ 4046 void 4047 vm_page_deactivate(vm_page_t m) 4048 { 4049 4050 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4051 } 4052 4053 void 4054 vm_page_deactivate_noreuse(vm_page_t m) 4055 { 4056 4057 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4058 } 4059 4060 /* 4061 * Put a page in the laundry, or requeue it if it is already there. 4062 */ 4063 void 4064 vm_page_launder(vm_page_t m) 4065 { 4066 4067 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4068 } 4069 4070 /* 4071 * Put a page in the PQ_UNSWAPPABLE holding queue. 4072 */ 4073 void 4074 vm_page_unswappable(vm_page_t m) 4075 { 4076 4077 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 4078 ("page %p already unswappable", m)); 4079 4080 vm_page_dequeue(m); 4081 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4082 } 4083 4084 /* 4085 * Release a page back to the page queues in preparation for unwiring. 4086 */ 4087 static void 4088 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4089 { 4090 vm_page_astate_t old, new; 4091 uint16_t nflag; 4092 4093 /* 4094 * Use a check of the valid bits to determine whether we should 4095 * accelerate reclamation of the page. The object lock might not be 4096 * held here, in which case the check is racy. At worst we will either 4097 * accelerate reclamation of a valid page and violate LRU, or 4098 * unnecessarily defer reclamation of an invalid page. 4099 * 4100 * If we were asked to not cache the page, place it near the head of the 4101 * inactive queue so that is reclaimed sooner. 4102 */ 4103 if (noreuse || m->valid == 0) { 4104 nqueue = PQ_INACTIVE; 4105 nflag = PGA_REQUEUE_HEAD; 4106 } else { 4107 nflag = PGA_REQUEUE; 4108 } 4109 4110 old = vm_page_astate_load(m); 4111 do { 4112 new = old; 4113 4114 /* 4115 * If the page is already in the active queue and we are not 4116 * trying to accelerate reclamation, simply mark it as 4117 * referenced and avoid any queue operations. 4118 */ 4119 new.flags &= ~PGA_QUEUE_OP_MASK; 4120 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE) 4121 new.flags |= PGA_REFERENCED; 4122 else { 4123 new.flags |= nflag; 4124 new.queue = nqueue; 4125 } 4126 } while (!vm_page_pqstate_commit(m, &old, new)); 4127 } 4128 4129 /* 4130 * Unwire a page and either attempt to free it or re-add it to the page queues. 4131 */ 4132 void 4133 vm_page_release(vm_page_t m, int flags) 4134 { 4135 vm_object_t object; 4136 4137 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4138 ("vm_page_release: page %p is unmanaged", m)); 4139 4140 if ((flags & VPR_TRYFREE) != 0) { 4141 for (;;) { 4142 object = atomic_load_ptr(&m->object); 4143 if (object == NULL) 4144 break; 4145 /* Depends on type-stability. */ 4146 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4147 break; 4148 if (object == m->object) { 4149 vm_page_release_locked(m, flags); 4150 VM_OBJECT_WUNLOCK(object); 4151 return; 4152 } 4153 VM_OBJECT_WUNLOCK(object); 4154 } 4155 } 4156 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4157 } 4158 4159 /* See vm_page_release(). */ 4160 void 4161 vm_page_release_locked(vm_page_t m, int flags) 4162 { 4163 4164 VM_OBJECT_ASSERT_WLOCKED(m->object); 4165 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4166 ("vm_page_release_locked: page %p is unmanaged", m)); 4167 4168 if (vm_page_unwire_noq(m)) { 4169 if ((flags & VPR_TRYFREE) != 0 && 4170 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4171 m->dirty == 0 && vm_page_tryxbusy(m)) { 4172 vm_page_free(m); 4173 } else { 4174 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4175 } 4176 } 4177 } 4178 4179 static bool 4180 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4181 { 4182 u_int old; 4183 4184 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4185 ("vm_page_try_blocked_op: page %p has no object", m)); 4186 KASSERT(vm_page_busied(m), 4187 ("vm_page_try_blocked_op: page %p is not busy", m)); 4188 VM_OBJECT_ASSERT_LOCKED(m->object); 4189 4190 old = m->ref_count; 4191 do { 4192 KASSERT(old != 0, 4193 ("vm_page_try_blocked_op: page %p has no references", m)); 4194 if (VPRC_WIRE_COUNT(old) != 0) 4195 return (false); 4196 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4197 4198 (op)(m); 4199 4200 /* 4201 * If the object is read-locked, new wirings may be created via an 4202 * object lookup. 4203 */ 4204 old = vm_page_drop(m, VPRC_BLOCKED); 4205 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4206 old == (VPRC_BLOCKED | VPRC_OBJREF), 4207 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4208 old, m)); 4209 return (true); 4210 } 4211 4212 /* 4213 * Atomically check for wirings and remove all mappings of the page. 4214 */ 4215 bool 4216 vm_page_try_remove_all(vm_page_t m) 4217 { 4218 4219 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4220 } 4221 4222 /* 4223 * Atomically check for wirings and remove all writeable mappings of the page. 4224 */ 4225 bool 4226 vm_page_try_remove_write(vm_page_t m) 4227 { 4228 4229 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4230 } 4231 4232 /* 4233 * vm_page_advise 4234 * 4235 * Apply the specified advice to the given page. 4236 * 4237 * The object and page must be locked. 4238 */ 4239 void 4240 vm_page_advise(vm_page_t m, int advice) 4241 { 4242 4243 VM_OBJECT_ASSERT_WLOCKED(m->object); 4244 if (advice == MADV_FREE) 4245 /* 4246 * Mark the page clean. This will allow the page to be freed 4247 * without first paging it out. MADV_FREE pages are often 4248 * quickly reused by malloc(3), so we do not do anything that 4249 * would result in a page fault on a later access. 4250 */ 4251 vm_page_undirty(m); 4252 else if (advice != MADV_DONTNEED) { 4253 if (advice == MADV_WILLNEED) 4254 vm_page_activate(m); 4255 return; 4256 } 4257 4258 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4259 vm_page_dirty(m); 4260 4261 /* 4262 * Clear any references to the page. Otherwise, the page daemon will 4263 * immediately reactivate the page. 4264 */ 4265 vm_page_aflag_clear(m, PGA_REFERENCED); 4266 4267 /* 4268 * Place clean pages near the head of the inactive queue rather than 4269 * the tail, thus defeating the queue's LRU operation and ensuring that 4270 * the page will be reused quickly. Dirty pages not already in the 4271 * laundry are moved there. 4272 */ 4273 if (m->dirty == 0) 4274 vm_page_deactivate_noreuse(m); 4275 else if (!vm_page_in_laundry(m)) 4276 vm_page_launder(m); 4277 } 4278 4279 /* 4280 * vm_page_grab_release 4281 * 4282 * Helper routine for grab functions to release busy on return. 4283 */ 4284 static inline void 4285 vm_page_grab_release(vm_page_t m, int allocflags) 4286 { 4287 4288 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4289 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4290 vm_page_sunbusy(m); 4291 else 4292 vm_page_xunbusy(m); 4293 } 4294 } 4295 4296 /* 4297 * vm_page_grab_sleep 4298 * 4299 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4300 * if the caller should retry and false otherwise. 4301 * 4302 * If the object is locked on entry the object will be unlocked with 4303 * false returns and still locked but possibly having been dropped 4304 * with true returns. 4305 */ 4306 static bool 4307 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4308 const char *wmesg, int allocflags, bool locked) 4309 { 4310 4311 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4312 return (false); 4313 4314 /* 4315 * Reference the page before unlocking and sleeping so that 4316 * the page daemon is less likely to reclaim it. 4317 */ 4318 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4319 vm_page_reference(m); 4320 4321 if (_vm_page_busy_sleep(object, m, m->pindex, wmesg, allocflags, 4322 locked) && locked) 4323 VM_OBJECT_WLOCK(object); 4324 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4325 return (false); 4326 4327 return (true); 4328 } 4329 4330 /* 4331 * Assert that the grab flags are valid. 4332 */ 4333 static inline void 4334 vm_page_grab_check(int allocflags) 4335 { 4336 4337 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4338 (allocflags & VM_ALLOC_WIRED) != 0, 4339 ("vm_page_grab*: the pages must be busied or wired")); 4340 4341 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4342 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4343 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4344 } 4345 4346 /* 4347 * Calculate the page allocation flags for grab. 4348 */ 4349 static inline int 4350 vm_page_grab_pflags(int allocflags) 4351 { 4352 int pflags; 4353 4354 pflags = allocflags & 4355 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4356 VM_ALLOC_NOBUSY); 4357 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4358 pflags |= VM_ALLOC_WAITFAIL; 4359 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4360 pflags |= VM_ALLOC_SBUSY; 4361 4362 return (pflags); 4363 } 4364 4365 /* 4366 * Grab a page, waiting until we are waken up due to the page 4367 * changing state. We keep on waiting, if the page continues 4368 * to be in the object. If the page doesn't exist, first allocate it 4369 * and then conditionally zero it. 4370 * 4371 * This routine may sleep. 4372 * 4373 * The object must be locked on entry. The lock will, however, be released 4374 * and reacquired if the routine sleeps. 4375 */ 4376 vm_page_t 4377 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4378 { 4379 vm_page_t m; 4380 4381 VM_OBJECT_ASSERT_WLOCKED(object); 4382 vm_page_grab_check(allocflags); 4383 4384 retrylookup: 4385 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4386 if (!vm_page_tryacquire(m, allocflags)) { 4387 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4388 allocflags, true)) 4389 goto retrylookup; 4390 return (NULL); 4391 } 4392 goto out; 4393 } 4394 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4395 return (NULL); 4396 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); 4397 if (m == NULL) { 4398 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4399 return (NULL); 4400 goto retrylookup; 4401 } 4402 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4403 pmap_zero_page(m); 4404 4405 out: 4406 vm_page_grab_release(m, allocflags); 4407 4408 return (m); 4409 } 4410 4411 /* 4412 * Locklessly attempt to acquire a page given a (object, pindex) tuple 4413 * and an optional previous page to avoid the radix lookup. The resulting 4414 * page will be validated against the identity tuple and busied or wired 4415 * as requested. A NULL *mp return guarantees that the page was not in 4416 * radix at the time of the call but callers must perform higher level 4417 * synchronization or retry the operation under a lock if they require 4418 * an atomic answer. This is the only lock free validation routine, 4419 * other routines can depend on the resulting page state. 4420 * 4421 * The return value indicates whether the operation failed due to caller 4422 * flags. The return is tri-state with mp: 4423 * 4424 * (true, *mp != NULL) - The operation was successful. 4425 * (true, *mp == NULL) - The page was not found in tree. 4426 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition. 4427 */ 4428 static bool 4429 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, 4430 vm_page_t prev, vm_page_t *mp, int allocflags) 4431 { 4432 vm_page_t m; 4433 4434 vm_page_grab_check(allocflags); 4435 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev)); 4436 4437 *mp = NULL; 4438 for (;;) { 4439 /* 4440 * We may see a false NULL here because the previous page 4441 * has been removed or just inserted and the list is loaded 4442 * without barriers. Switch to radix to verify. 4443 */ 4444 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL || 4445 m->pindex != pindex || 4446 atomic_load_ptr(&m->object) != object) { 4447 prev = NULL; 4448 /* 4449 * This guarantees the result is instantaneously 4450 * correct. 4451 */ 4452 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 4453 } 4454 if (m == NULL) 4455 return (true); 4456 if (vm_page_trybusy(m, allocflags)) { 4457 if (m->object == object && m->pindex == pindex) 4458 break; 4459 /* relookup. */ 4460 vm_page_busy_release(m); 4461 cpu_spinwait(); 4462 continue; 4463 } 4464 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4465 allocflags, false)) 4466 return (false); 4467 } 4468 if ((allocflags & VM_ALLOC_WIRED) != 0) 4469 vm_page_wire(m); 4470 vm_page_grab_release(m, allocflags); 4471 *mp = m; 4472 return (true); 4473 } 4474 4475 /* 4476 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4477 * is not set. 4478 */ 4479 vm_page_t 4480 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4481 { 4482 vm_page_t m; 4483 4484 vm_page_grab_check(allocflags); 4485 4486 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags)) 4487 return (NULL); 4488 if (m != NULL) 4489 return (m); 4490 4491 /* 4492 * The radix lockless lookup should never return a false negative 4493 * errors. If the user specifies NOCREAT they are guaranteed there 4494 * was no page present at the instant of the call. A NOCREAT caller 4495 * must handle create races gracefully. 4496 */ 4497 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4498 return (NULL); 4499 4500 VM_OBJECT_WLOCK(object); 4501 m = vm_page_grab(object, pindex, allocflags); 4502 VM_OBJECT_WUNLOCK(object); 4503 4504 return (m); 4505 } 4506 4507 /* 4508 * Grab a page and make it valid, paging in if necessary. Pages missing from 4509 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4510 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4511 * in simultaneously. Additional pages will be left on a paging queue but 4512 * will neither be wired nor busy regardless of allocflags. 4513 */ 4514 int 4515 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4516 { 4517 vm_page_t m; 4518 vm_page_t ma[VM_INITIAL_PAGEIN]; 4519 int after, i, pflags, rv; 4520 4521 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4522 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4523 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4524 KASSERT((allocflags & 4525 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4526 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4527 VM_OBJECT_ASSERT_WLOCKED(object); 4528 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4529 VM_ALLOC_WIRED); 4530 pflags |= VM_ALLOC_WAITFAIL; 4531 4532 retrylookup: 4533 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4534 /* 4535 * If the page is fully valid it can only become invalid 4536 * with the object lock held. If it is not valid it can 4537 * become valid with the busy lock held. Therefore, we 4538 * may unnecessarily lock the exclusive busy here if we 4539 * race with I/O completion not using the object lock. 4540 * However, we will not end up with an invalid page and a 4541 * shared lock. 4542 */ 4543 if (!vm_page_trybusy(m, 4544 vm_page_all_valid(m) ? allocflags : 0)) { 4545 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4546 allocflags, true); 4547 goto retrylookup; 4548 } 4549 if (vm_page_all_valid(m)) 4550 goto out; 4551 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4552 vm_page_busy_release(m); 4553 *mp = NULL; 4554 return (VM_PAGER_FAIL); 4555 } 4556 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4557 *mp = NULL; 4558 return (VM_PAGER_FAIL); 4559 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { 4560 goto retrylookup; 4561 } 4562 4563 vm_page_assert_xbusied(m); 4564 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4565 after = MIN(after, VM_INITIAL_PAGEIN); 4566 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4567 after = MAX(after, 1); 4568 ma[0] = m; 4569 for (i = 1; i < after; i++) { 4570 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 4571 if (ma[i]->valid || !vm_page_tryxbusy(ma[i])) 4572 break; 4573 } else { 4574 ma[i] = vm_page_alloc(object, m->pindex + i, 4575 VM_ALLOC_NORMAL); 4576 if (ma[i] == NULL) 4577 break; 4578 } 4579 } 4580 after = i; 4581 vm_object_pip_add(object, after); 4582 VM_OBJECT_WUNLOCK(object); 4583 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 4584 VM_OBJECT_WLOCK(object); 4585 vm_object_pip_wakeupn(object, after); 4586 /* Pager may have replaced a page. */ 4587 m = ma[0]; 4588 if (rv != VM_PAGER_OK) { 4589 for (i = 0; i < after; i++) { 4590 if (!vm_page_wired(ma[i])) 4591 vm_page_free(ma[i]); 4592 else 4593 vm_page_xunbusy(ma[i]); 4594 } 4595 *mp = NULL; 4596 return (rv); 4597 } 4598 for (i = 1; i < after; i++) 4599 vm_page_readahead_finish(ma[i]); 4600 MPASS(vm_page_all_valid(m)); 4601 } else { 4602 vm_page_zero_invalid(m, TRUE); 4603 } 4604 out: 4605 if ((allocflags & VM_ALLOC_WIRED) != 0) 4606 vm_page_wire(m); 4607 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 4608 vm_page_busy_downgrade(m); 4609 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 4610 vm_page_busy_release(m); 4611 *mp = m; 4612 return (VM_PAGER_OK); 4613 } 4614 4615 /* 4616 * Locklessly grab a valid page. If the page is not valid or not yet 4617 * allocated this will fall back to the object lock method. 4618 */ 4619 int 4620 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 4621 vm_pindex_t pindex, int allocflags) 4622 { 4623 vm_page_t m; 4624 int flags; 4625 int error; 4626 4627 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4628 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4629 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 4630 "mismatch")); 4631 KASSERT((allocflags & 4632 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4633 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 4634 4635 /* 4636 * Attempt a lockless lookup and busy. We need at least an sbusy 4637 * before we can inspect the valid field and return a wired page. 4638 */ 4639 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 4640 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags)) 4641 return (VM_PAGER_FAIL); 4642 if ((m = *mp) != NULL) { 4643 if (vm_page_all_valid(m)) { 4644 if ((allocflags & VM_ALLOC_WIRED) != 0) 4645 vm_page_wire(m); 4646 vm_page_grab_release(m, allocflags); 4647 return (VM_PAGER_OK); 4648 } 4649 vm_page_busy_release(m); 4650 } 4651 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4652 *mp = NULL; 4653 return (VM_PAGER_FAIL); 4654 } 4655 VM_OBJECT_WLOCK(object); 4656 error = vm_page_grab_valid(mp, object, pindex, allocflags); 4657 VM_OBJECT_WUNLOCK(object); 4658 4659 return (error); 4660 } 4661 4662 /* 4663 * Return the specified range of pages from the given object. For each 4664 * page offset within the range, if a page already exists within the object 4665 * at that offset and it is busy, then wait for it to change state. If, 4666 * instead, the page doesn't exist, then allocate it. 4667 * 4668 * The caller must always specify an allocation class. 4669 * 4670 * allocation classes: 4671 * VM_ALLOC_NORMAL normal process request 4672 * VM_ALLOC_SYSTEM system *really* needs the pages 4673 * 4674 * The caller must always specify that the pages are to be busied and/or 4675 * wired. 4676 * 4677 * optional allocation flags: 4678 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4679 * VM_ALLOC_NOBUSY do not exclusive busy the page 4680 * VM_ALLOC_NOWAIT do not sleep 4681 * VM_ALLOC_SBUSY set page to sbusy state 4682 * VM_ALLOC_WIRED wire the pages 4683 * VM_ALLOC_ZERO zero and validate any invalid pages 4684 * 4685 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4686 * may return a partial prefix of the requested range. 4687 */ 4688 int 4689 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4690 vm_page_t *ma, int count) 4691 { 4692 vm_page_t m, mpred; 4693 int pflags; 4694 int i; 4695 4696 VM_OBJECT_ASSERT_WLOCKED(object); 4697 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4698 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4699 vm_page_grab_check(allocflags); 4700 4701 pflags = vm_page_grab_pflags(allocflags); 4702 if (count == 0) 4703 return (0); 4704 4705 i = 0; 4706 retrylookup: 4707 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4708 if (m == NULL || m->pindex != pindex + i) { 4709 mpred = m; 4710 m = NULL; 4711 } else 4712 mpred = TAILQ_PREV(m, pglist, listq); 4713 for (; i < count; i++) { 4714 if (m != NULL) { 4715 if (!vm_page_tryacquire(m, allocflags)) { 4716 if (vm_page_grab_sleep(object, m, pindex, 4717 "grbmaw", allocflags, true)) 4718 goto retrylookup; 4719 break; 4720 } 4721 } else { 4722 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4723 break; 4724 m = vm_page_alloc_after(object, pindex + i, 4725 pflags | VM_ALLOC_COUNT(count - i), mpred); 4726 if (m == NULL) { 4727 if ((allocflags & (VM_ALLOC_NOWAIT | 4728 VM_ALLOC_WAITFAIL)) != 0) 4729 break; 4730 goto retrylookup; 4731 } 4732 } 4733 if (vm_page_none_valid(m) && 4734 (allocflags & VM_ALLOC_ZERO) != 0) { 4735 if ((m->flags & PG_ZERO) == 0) 4736 pmap_zero_page(m); 4737 vm_page_valid(m); 4738 } 4739 vm_page_grab_release(m, allocflags); 4740 ma[i] = mpred = m; 4741 m = vm_page_next(m); 4742 } 4743 return (i); 4744 } 4745 4746 /* 4747 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 4748 * and will fall back to the locked variant to handle allocation. 4749 */ 4750 int 4751 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 4752 int allocflags, vm_page_t *ma, int count) 4753 { 4754 vm_page_t m, pred; 4755 int flags; 4756 int i; 4757 4758 vm_page_grab_check(allocflags); 4759 4760 /* 4761 * Modify flags for lockless acquire to hold the page until we 4762 * set it valid if necessary. 4763 */ 4764 flags = allocflags & ~VM_ALLOC_NOBUSY; 4765 pred = NULL; 4766 for (i = 0; i < count; i++, pindex++) { 4767 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags)) 4768 return (i); 4769 if (m == NULL) 4770 break; 4771 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 4772 if ((m->flags & PG_ZERO) == 0) 4773 pmap_zero_page(m); 4774 vm_page_valid(m); 4775 } 4776 /* m will still be wired or busy according to flags. */ 4777 vm_page_grab_release(m, allocflags); 4778 pred = ma[i] = m; 4779 } 4780 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4781 return (i); 4782 count -= i; 4783 VM_OBJECT_WLOCK(object); 4784 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 4785 VM_OBJECT_WUNLOCK(object); 4786 4787 return (i); 4788 } 4789 4790 /* 4791 * Mapping function for valid or dirty bits in a page. 4792 * 4793 * Inputs are required to range within a page. 4794 */ 4795 vm_page_bits_t 4796 vm_page_bits(int base, int size) 4797 { 4798 int first_bit; 4799 int last_bit; 4800 4801 KASSERT( 4802 base + size <= PAGE_SIZE, 4803 ("vm_page_bits: illegal base/size %d/%d", base, size) 4804 ); 4805 4806 if (size == 0) /* handle degenerate case */ 4807 return (0); 4808 4809 first_bit = base >> DEV_BSHIFT; 4810 last_bit = (base + size - 1) >> DEV_BSHIFT; 4811 4812 return (((vm_page_bits_t)2 << last_bit) - 4813 ((vm_page_bits_t)1 << first_bit)); 4814 } 4815 4816 void 4817 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 4818 { 4819 4820 #if PAGE_SIZE == 32768 4821 atomic_set_64((uint64_t *)bits, set); 4822 #elif PAGE_SIZE == 16384 4823 atomic_set_32((uint32_t *)bits, set); 4824 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 4825 atomic_set_16((uint16_t *)bits, set); 4826 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 4827 atomic_set_8((uint8_t *)bits, set); 4828 #else /* PAGE_SIZE <= 8192 */ 4829 uintptr_t addr; 4830 int shift; 4831 4832 addr = (uintptr_t)bits; 4833 /* 4834 * Use a trick to perform a 32-bit atomic on the 4835 * containing aligned word, to not depend on the existence 4836 * of atomic_{set, clear}_{8, 16}. 4837 */ 4838 shift = addr & (sizeof(uint32_t) - 1); 4839 #if BYTE_ORDER == BIG_ENDIAN 4840 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4841 #else 4842 shift *= NBBY; 4843 #endif 4844 addr &= ~(sizeof(uint32_t) - 1); 4845 atomic_set_32((uint32_t *)addr, set << shift); 4846 #endif /* PAGE_SIZE */ 4847 } 4848 4849 static inline void 4850 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 4851 { 4852 4853 #if PAGE_SIZE == 32768 4854 atomic_clear_64((uint64_t *)bits, clear); 4855 #elif PAGE_SIZE == 16384 4856 atomic_clear_32((uint32_t *)bits, clear); 4857 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 4858 atomic_clear_16((uint16_t *)bits, clear); 4859 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 4860 atomic_clear_8((uint8_t *)bits, clear); 4861 #else /* PAGE_SIZE <= 8192 */ 4862 uintptr_t addr; 4863 int shift; 4864 4865 addr = (uintptr_t)bits; 4866 /* 4867 * Use a trick to perform a 32-bit atomic on the 4868 * containing aligned word, to not depend on the existence 4869 * of atomic_{set, clear}_{8, 16}. 4870 */ 4871 shift = addr & (sizeof(uint32_t) - 1); 4872 #if BYTE_ORDER == BIG_ENDIAN 4873 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4874 #else 4875 shift *= NBBY; 4876 #endif 4877 addr &= ~(sizeof(uint32_t) - 1); 4878 atomic_clear_32((uint32_t *)addr, clear << shift); 4879 #endif /* PAGE_SIZE */ 4880 } 4881 4882 static inline vm_page_bits_t 4883 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 4884 { 4885 #if PAGE_SIZE == 32768 4886 uint64_t old; 4887 4888 old = *bits; 4889 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 4890 return (old); 4891 #elif PAGE_SIZE == 16384 4892 uint32_t old; 4893 4894 old = *bits; 4895 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 4896 return (old); 4897 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 4898 uint16_t old; 4899 4900 old = *bits; 4901 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 4902 return (old); 4903 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 4904 uint8_t old; 4905 4906 old = *bits; 4907 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 4908 return (old); 4909 #else /* PAGE_SIZE <= 4096*/ 4910 uintptr_t addr; 4911 uint32_t old, new, mask; 4912 int shift; 4913 4914 addr = (uintptr_t)bits; 4915 /* 4916 * Use a trick to perform a 32-bit atomic on the 4917 * containing aligned word, to not depend on the existence 4918 * of atomic_{set, swap, clear}_{8, 16}. 4919 */ 4920 shift = addr & (sizeof(uint32_t) - 1); 4921 #if BYTE_ORDER == BIG_ENDIAN 4922 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4923 #else 4924 shift *= NBBY; 4925 #endif 4926 addr &= ~(sizeof(uint32_t) - 1); 4927 mask = VM_PAGE_BITS_ALL << shift; 4928 4929 old = *bits; 4930 do { 4931 new = old & ~mask; 4932 new |= newbits << shift; 4933 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 4934 return (old >> shift); 4935 #endif /* PAGE_SIZE */ 4936 } 4937 4938 /* 4939 * vm_page_set_valid_range: 4940 * 4941 * Sets portions of a page valid. The arguments are expected 4942 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4943 * of any partial chunks touched by the range. The invalid portion of 4944 * such chunks will be zeroed. 4945 * 4946 * (base + size) must be less then or equal to PAGE_SIZE. 4947 */ 4948 void 4949 vm_page_set_valid_range(vm_page_t m, int base, int size) 4950 { 4951 int endoff, frag; 4952 vm_page_bits_t pagebits; 4953 4954 vm_page_assert_busied(m); 4955 if (size == 0) /* handle degenerate case */ 4956 return; 4957 4958 /* 4959 * If the base is not DEV_BSIZE aligned and the valid 4960 * bit is clear, we have to zero out a portion of the 4961 * first block. 4962 */ 4963 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4964 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 4965 pmap_zero_page_area(m, frag, base - frag); 4966 4967 /* 4968 * If the ending offset is not DEV_BSIZE aligned and the 4969 * valid bit is clear, we have to zero out a portion of 4970 * the last block. 4971 */ 4972 endoff = base + size; 4973 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4974 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 4975 pmap_zero_page_area(m, endoff, 4976 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4977 4978 /* 4979 * Assert that no previously invalid block that is now being validated 4980 * is already dirty. 4981 */ 4982 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 4983 ("vm_page_set_valid_range: page %p is dirty", m)); 4984 4985 /* 4986 * Set valid bits inclusive of any overlap. 4987 */ 4988 pagebits = vm_page_bits(base, size); 4989 if (vm_page_xbusied(m)) 4990 m->valid |= pagebits; 4991 else 4992 vm_page_bits_set(m, &m->valid, pagebits); 4993 } 4994 4995 /* 4996 * Set the page dirty bits and free the invalid swap space if 4997 * present. Returns the previous dirty bits. 4998 */ 4999 vm_page_bits_t 5000 vm_page_set_dirty(vm_page_t m) 5001 { 5002 vm_page_bits_t old; 5003 5004 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5005 5006 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5007 old = m->dirty; 5008 m->dirty = VM_PAGE_BITS_ALL; 5009 } else 5010 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5011 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5012 vm_pager_page_unswapped(m); 5013 5014 return (old); 5015 } 5016 5017 /* 5018 * Clear the given bits from the specified page's dirty field. 5019 */ 5020 static __inline void 5021 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5022 { 5023 5024 vm_page_assert_busied(m); 5025 5026 /* 5027 * If the page is xbusied and not write mapped we are the 5028 * only thread that can modify dirty bits. Otherwise, The pmap 5029 * layer can call vm_page_dirty() without holding a distinguished 5030 * lock. The combination of page busy and atomic operations 5031 * suffice to guarantee consistency of the page dirty field. 5032 */ 5033 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5034 m->dirty &= ~pagebits; 5035 else 5036 vm_page_bits_clear(m, &m->dirty, pagebits); 5037 } 5038 5039 /* 5040 * vm_page_set_validclean: 5041 * 5042 * Sets portions of a page valid and clean. The arguments are expected 5043 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5044 * of any partial chunks touched by the range. The invalid portion of 5045 * such chunks will be zero'd. 5046 * 5047 * (base + size) must be less then or equal to PAGE_SIZE. 5048 */ 5049 void 5050 vm_page_set_validclean(vm_page_t m, int base, int size) 5051 { 5052 vm_page_bits_t oldvalid, pagebits; 5053 int endoff, frag; 5054 5055 vm_page_assert_busied(m); 5056 if (size == 0) /* handle degenerate case */ 5057 return; 5058 5059 /* 5060 * If the base is not DEV_BSIZE aligned and the valid 5061 * bit is clear, we have to zero out a portion of the 5062 * first block. 5063 */ 5064 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5065 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5066 pmap_zero_page_area(m, frag, base - frag); 5067 5068 /* 5069 * If the ending offset is not DEV_BSIZE aligned and the 5070 * valid bit is clear, we have to zero out a portion of 5071 * the last block. 5072 */ 5073 endoff = base + size; 5074 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5075 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5076 pmap_zero_page_area(m, endoff, 5077 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5078 5079 /* 5080 * Set valid, clear dirty bits. If validating the entire 5081 * page we can safely clear the pmap modify bit. We also 5082 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5083 * takes a write fault on a MAP_NOSYNC memory area the flag will 5084 * be set again. 5085 * 5086 * We set valid bits inclusive of any overlap, but we can only 5087 * clear dirty bits for DEV_BSIZE chunks that are fully within 5088 * the range. 5089 */ 5090 oldvalid = m->valid; 5091 pagebits = vm_page_bits(base, size); 5092 if (vm_page_xbusied(m)) 5093 m->valid |= pagebits; 5094 else 5095 vm_page_bits_set(m, &m->valid, pagebits); 5096 #if 0 /* NOT YET */ 5097 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5098 frag = DEV_BSIZE - frag; 5099 base += frag; 5100 size -= frag; 5101 if (size < 0) 5102 size = 0; 5103 } 5104 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5105 #endif 5106 if (base == 0 && size == PAGE_SIZE) { 5107 /* 5108 * The page can only be modified within the pmap if it is 5109 * mapped, and it can only be mapped if it was previously 5110 * fully valid. 5111 */ 5112 if (oldvalid == VM_PAGE_BITS_ALL) 5113 /* 5114 * Perform the pmap_clear_modify() first. Otherwise, 5115 * a concurrent pmap operation, such as 5116 * pmap_protect(), could clear a modification in the 5117 * pmap and set the dirty field on the page before 5118 * pmap_clear_modify() had begun and after the dirty 5119 * field was cleared here. 5120 */ 5121 pmap_clear_modify(m); 5122 m->dirty = 0; 5123 vm_page_aflag_clear(m, PGA_NOSYNC); 5124 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5125 m->dirty &= ~pagebits; 5126 else 5127 vm_page_clear_dirty_mask(m, pagebits); 5128 } 5129 5130 void 5131 vm_page_clear_dirty(vm_page_t m, int base, int size) 5132 { 5133 5134 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5135 } 5136 5137 /* 5138 * vm_page_set_invalid: 5139 * 5140 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5141 * valid and dirty bits for the effected areas are cleared. 5142 */ 5143 void 5144 vm_page_set_invalid(vm_page_t m, int base, int size) 5145 { 5146 vm_page_bits_t bits; 5147 vm_object_t object; 5148 5149 /* 5150 * The object lock is required so that pages can't be mapped 5151 * read-only while we're in the process of invalidating them. 5152 */ 5153 object = m->object; 5154 VM_OBJECT_ASSERT_WLOCKED(object); 5155 vm_page_assert_busied(m); 5156 5157 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5158 size >= object->un_pager.vnp.vnp_size) 5159 bits = VM_PAGE_BITS_ALL; 5160 else 5161 bits = vm_page_bits(base, size); 5162 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5163 pmap_remove_all(m); 5164 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5165 !pmap_page_is_mapped(m), 5166 ("vm_page_set_invalid: page %p is mapped", m)); 5167 if (vm_page_xbusied(m)) { 5168 m->valid &= ~bits; 5169 m->dirty &= ~bits; 5170 } else { 5171 vm_page_bits_clear(m, &m->valid, bits); 5172 vm_page_bits_clear(m, &m->dirty, bits); 5173 } 5174 } 5175 5176 /* 5177 * vm_page_invalid: 5178 * 5179 * Invalidates the entire page. The page must be busy, unmapped, and 5180 * the enclosing object must be locked. The object locks protects 5181 * against concurrent read-only pmap enter which is done without 5182 * busy. 5183 */ 5184 void 5185 vm_page_invalid(vm_page_t m) 5186 { 5187 5188 vm_page_assert_busied(m); 5189 VM_OBJECT_ASSERT_LOCKED(m->object); 5190 MPASS(!pmap_page_is_mapped(m)); 5191 5192 if (vm_page_xbusied(m)) 5193 m->valid = 0; 5194 else 5195 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5196 } 5197 5198 /* 5199 * vm_page_zero_invalid() 5200 * 5201 * The kernel assumes that the invalid portions of a page contain 5202 * garbage, but such pages can be mapped into memory by user code. 5203 * When this occurs, we must zero out the non-valid portions of the 5204 * page so user code sees what it expects. 5205 * 5206 * Pages are most often semi-valid when the end of a file is mapped 5207 * into memory and the file's size is not page aligned. 5208 */ 5209 void 5210 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5211 { 5212 int b; 5213 int i; 5214 5215 /* 5216 * Scan the valid bits looking for invalid sections that 5217 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5218 * valid bit may be set ) have already been zeroed by 5219 * vm_page_set_validclean(). 5220 */ 5221 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5222 if (i == (PAGE_SIZE / DEV_BSIZE) || 5223 (m->valid & ((vm_page_bits_t)1 << i))) { 5224 if (i > b) { 5225 pmap_zero_page_area(m, 5226 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5227 } 5228 b = i + 1; 5229 } 5230 } 5231 5232 /* 5233 * setvalid is TRUE when we can safely set the zero'd areas 5234 * as being valid. We can do this if there are no cache consistancy 5235 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5236 */ 5237 if (setvalid) 5238 vm_page_valid(m); 5239 } 5240 5241 /* 5242 * vm_page_is_valid: 5243 * 5244 * Is (partial) page valid? Note that the case where size == 0 5245 * will return FALSE in the degenerate case where the page is 5246 * entirely invalid, and TRUE otherwise. 5247 * 5248 * Some callers envoke this routine without the busy lock held and 5249 * handle races via higher level locks. Typical callers should 5250 * hold a busy lock to prevent invalidation. 5251 */ 5252 int 5253 vm_page_is_valid(vm_page_t m, int base, int size) 5254 { 5255 vm_page_bits_t bits; 5256 5257 bits = vm_page_bits(base, size); 5258 return (m->valid != 0 && (m->valid & bits) == bits); 5259 } 5260 5261 /* 5262 * Returns true if all of the specified predicates are true for the entire 5263 * (super)page and false otherwise. 5264 */ 5265 bool 5266 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 5267 { 5268 vm_object_t object; 5269 int i, npages; 5270 5271 object = m->object; 5272 if (skip_m != NULL && skip_m->object != object) 5273 return (false); 5274 VM_OBJECT_ASSERT_LOCKED(object); 5275 npages = atop(pagesizes[m->psind]); 5276 5277 /* 5278 * The physically contiguous pages that make up a superpage, i.e., a 5279 * page with a page size index ("psind") greater than zero, will 5280 * occupy adjacent entries in vm_page_array[]. 5281 */ 5282 for (i = 0; i < npages; i++) { 5283 /* Always test object consistency, including "skip_m". */ 5284 if (m[i].object != object) 5285 return (false); 5286 if (&m[i] == skip_m) 5287 continue; 5288 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5289 return (false); 5290 if ((flags & PS_ALL_DIRTY) != 0) { 5291 /* 5292 * Calling vm_page_test_dirty() or pmap_is_modified() 5293 * might stop this case from spuriously returning 5294 * "false". However, that would require a write lock 5295 * on the object containing "m[i]". 5296 */ 5297 if (m[i].dirty != VM_PAGE_BITS_ALL) 5298 return (false); 5299 } 5300 if ((flags & PS_ALL_VALID) != 0 && 5301 m[i].valid != VM_PAGE_BITS_ALL) 5302 return (false); 5303 } 5304 return (true); 5305 } 5306 5307 /* 5308 * Set the page's dirty bits if the page is modified. 5309 */ 5310 void 5311 vm_page_test_dirty(vm_page_t m) 5312 { 5313 5314 vm_page_assert_busied(m); 5315 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5316 vm_page_dirty(m); 5317 } 5318 5319 void 5320 vm_page_valid(vm_page_t m) 5321 { 5322 5323 vm_page_assert_busied(m); 5324 if (vm_page_xbusied(m)) 5325 m->valid = VM_PAGE_BITS_ALL; 5326 else 5327 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5328 } 5329 5330 void 5331 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5332 { 5333 5334 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5335 } 5336 5337 void 5338 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5339 { 5340 5341 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5342 } 5343 5344 int 5345 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5346 { 5347 5348 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5349 } 5350 5351 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5352 void 5353 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5354 { 5355 5356 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5357 } 5358 5359 void 5360 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5361 { 5362 5363 mtx_assert_(vm_page_lockptr(m), a, file, line); 5364 } 5365 #endif 5366 5367 #ifdef INVARIANTS 5368 void 5369 vm_page_object_busy_assert(vm_page_t m) 5370 { 5371 5372 /* 5373 * Certain of the page's fields may only be modified by the 5374 * holder of a page or object busy. 5375 */ 5376 if (m->object != NULL && !vm_page_busied(m)) 5377 VM_OBJECT_ASSERT_BUSY(m->object); 5378 } 5379 5380 void 5381 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5382 { 5383 5384 if ((bits & PGA_WRITEABLE) == 0) 5385 return; 5386 5387 /* 5388 * The PGA_WRITEABLE flag can only be set if the page is 5389 * managed, is exclusively busied or the object is locked. 5390 * Currently, this flag is only set by pmap_enter(). 5391 */ 5392 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5393 ("PGA_WRITEABLE on unmanaged page")); 5394 if (!vm_page_xbusied(m)) 5395 VM_OBJECT_ASSERT_BUSY(m->object); 5396 } 5397 #endif 5398 5399 #include "opt_ddb.h" 5400 #ifdef DDB 5401 #include <sys/kernel.h> 5402 5403 #include <ddb/ddb.h> 5404 5405 DB_SHOW_COMMAND(page, vm_page_print_page_info) 5406 { 5407 5408 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5409 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5410 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5411 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5412 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5413 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5414 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5415 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5416 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5417 } 5418 5419 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 5420 { 5421 int dom; 5422 5423 db_printf("pq_free %d\n", vm_free_count()); 5424 for (dom = 0; dom < vm_ndomains; dom++) { 5425 db_printf( 5426 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5427 dom, 5428 vm_dom[dom].vmd_page_count, 5429 vm_dom[dom].vmd_free_count, 5430 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5431 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5432 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5433 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5434 } 5435 } 5436 5437 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5438 { 5439 vm_page_t m; 5440 boolean_t phys, virt; 5441 5442 if (!have_addr) { 5443 db_printf("show pginfo addr\n"); 5444 return; 5445 } 5446 5447 phys = strchr(modif, 'p') != NULL; 5448 virt = strchr(modif, 'v') != NULL; 5449 if (virt) 5450 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5451 else if (phys) 5452 m = PHYS_TO_VM_PAGE(addr); 5453 else 5454 m = (vm_page_t)addr; 5455 db_printf( 5456 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" 5457 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5458 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5459 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5460 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5461 } 5462 #endif /* DDB */ 5463