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, 0, 131 "VM page statistics"); 132 133 static counter_u64_t pqstate_commit_retries = EARLY_COUNTER; 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_t queue_ops = EARLY_COUNTER; 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_t queue_nops = EARLY_COUNTER; 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 static void 149 counter_startup(void) 150 { 151 152 pqstate_commit_retries = counter_u64_alloc(M_WAITOK); 153 queue_ops = counter_u64_alloc(M_WAITOK); 154 queue_nops = counter_u64_alloc(M_WAITOK); 155 } 156 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); 157 158 /* 159 * bogus page -- for I/O to/from partially complete buffers, 160 * or for paging into sparsely invalid regions. 161 */ 162 vm_page_t bogus_page; 163 164 vm_page_t vm_page_array; 165 long vm_page_array_size; 166 long first_page; 167 168 static TAILQ_HEAD(, vm_page) blacklist_head; 169 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 170 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 171 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 172 173 static uma_zone_t fakepg_zone; 174 175 static void vm_page_alloc_check(vm_page_t m); 176 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 177 const char *wmesg, bool nonshared, bool locked); 178 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 179 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 180 static bool vm_page_free_prep(vm_page_t m); 181 static void vm_page_free_toq(vm_page_t m); 182 static void vm_page_init(void *dummy); 183 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 184 vm_pindex_t pindex, vm_page_t mpred); 185 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 186 vm_page_t mpred); 187 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 188 const uint16_t nflag); 189 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 190 vm_page_t m_run, vm_paddr_t high); 191 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 192 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 193 int req); 194 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 195 int flags); 196 static void vm_page_zone_release(void *arg, void **store, int cnt); 197 198 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 199 200 static void 201 vm_page_init(void *dummy) 202 { 203 204 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 205 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 206 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 207 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 208 } 209 210 /* 211 * The cache page zone is initialized later since we need to be able to allocate 212 * pages before UMA is fully initialized. 213 */ 214 static void 215 vm_page_init_cache_zones(void *dummy __unused) 216 { 217 struct vm_domain *vmd; 218 struct vm_pgcache *pgcache; 219 int cache, domain, maxcache, pool; 220 221 maxcache = 0; 222 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache); 223 maxcache *= mp_ncpus; 224 for (domain = 0; domain < vm_ndomains; domain++) { 225 vmd = VM_DOMAIN(domain); 226 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 227 pgcache = &vmd->vmd_pgcache[pool]; 228 pgcache->domain = domain; 229 pgcache->pool = pool; 230 pgcache->zone = uma_zcache_create("vm pgcache", 231 PAGE_SIZE, NULL, NULL, NULL, NULL, 232 vm_page_zone_import, vm_page_zone_release, pgcache, 233 UMA_ZONE_VM); 234 235 /* 236 * Limit each pool's zone to 0.1% of the pages in the 237 * domain. 238 */ 239 cache = maxcache != 0 ? maxcache : 240 vmd->vmd_page_count / 1000; 241 uma_zone_set_maxcache(pgcache->zone, cache); 242 } 243 } 244 } 245 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 246 247 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 248 #if PAGE_SIZE == 32768 249 #ifdef CTASSERT 250 CTASSERT(sizeof(u_long) >= 8); 251 #endif 252 #endif 253 254 /* 255 * vm_set_page_size: 256 * 257 * Sets the page size, perhaps based upon the memory 258 * size. Must be called before any use of page-size 259 * dependent functions. 260 */ 261 void 262 vm_set_page_size(void) 263 { 264 if (vm_cnt.v_page_size == 0) 265 vm_cnt.v_page_size = PAGE_SIZE; 266 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 267 panic("vm_set_page_size: page size not a power of two"); 268 } 269 270 /* 271 * vm_page_blacklist_next: 272 * 273 * Find the next entry in the provided string of blacklist 274 * addresses. Entries are separated by space, comma, or newline. 275 * If an invalid integer is encountered then the rest of the 276 * string is skipped. Updates the list pointer to the next 277 * character, or NULL if the string is exhausted or invalid. 278 */ 279 static vm_paddr_t 280 vm_page_blacklist_next(char **list, char *end) 281 { 282 vm_paddr_t bad; 283 char *cp, *pos; 284 285 if (list == NULL || *list == NULL) 286 return (0); 287 if (**list =='\0') { 288 *list = NULL; 289 return (0); 290 } 291 292 /* 293 * If there's no end pointer then the buffer is coming from 294 * the kenv and we know it's null-terminated. 295 */ 296 if (end == NULL) 297 end = *list + strlen(*list); 298 299 /* Ensure that strtoq() won't walk off the end */ 300 if (*end != '\0') { 301 if (*end == '\n' || *end == ' ' || *end == ',') 302 *end = '\0'; 303 else { 304 printf("Blacklist not terminated, skipping\n"); 305 *list = NULL; 306 return (0); 307 } 308 } 309 310 for (pos = *list; *pos != '\0'; pos = cp) { 311 bad = strtoq(pos, &cp, 0); 312 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 313 if (bad == 0) { 314 if (++cp < end) 315 continue; 316 else 317 break; 318 } 319 } else 320 break; 321 if (*cp == '\0' || ++cp >= end) 322 *list = NULL; 323 else 324 *list = cp; 325 return (trunc_page(bad)); 326 } 327 printf("Garbage in RAM blacklist, skipping\n"); 328 *list = NULL; 329 return (0); 330 } 331 332 bool 333 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 334 { 335 struct vm_domain *vmd; 336 vm_page_t m; 337 int ret; 338 339 m = vm_phys_paddr_to_vm_page(pa); 340 if (m == NULL) 341 return (true); /* page does not exist, no failure */ 342 343 vmd = vm_pagequeue_domain(m); 344 vm_domain_free_lock(vmd); 345 ret = vm_phys_unfree_page(m); 346 vm_domain_free_unlock(vmd); 347 if (ret != 0) { 348 vm_domain_freecnt_inc(vmd, -1); 349 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 350 if (verbose) 351 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 352 } 353 return (ret); 354 } 355 356 /* 357 * vm_page_blacklist_check: 358 * 359 * Iterate through the provided string of blacklist addresses, pulling 360 * each entry out of the physical allocator free list and putting it 361 * onto a list for reporting via the vm.page_blacklist sysctl. 362 */ 363 static void 364 vm_page_blacklist_check(char *list, char *end) 365 { 366 vm_paddr_t pa; 367 char *next; 368 369 next = list; 370 while (next != NULL) { 371 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 372 continue; 373 vm_page_blacklist_add(pa, bootverbose); 374 } 375 } 376 377 /* 378 * vm_page_blacklist_load: 379 * 380 * Search for a special module named "ram_blacklist". It'll be a 381 * plain text file provided by the user via the loader directive 382 * of the same name. 383 */ 384 static void 385 vm_page_blacklist_load(char **list, char **end) 386 { 387 void *mod; 388 u_char *ptr; 389 u_int len; 390 391 mod = NULL; 392 ptr = NULL; 393 394 mod = preload_search_by_type("ram_blacklist"); 395 if (mod != NULL) { 396 ptr = preload_fetch_addr(mod); 397 len = preload_fetch_size(mod); 398 } 399 *list = ptr; 400 if (ptr != NULL) 401 *end = ptr + len; 402 else 403 *end = NULL; 404 return; 405 } 406 407 static int 408 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 409 { 410 vm_page_t m; 411 struct sbuf sbuf; 412 int error, first; 413 414 first = 1; 415 error = sysctl_wire_old_buffer(req, 0); 416 if (error != 0) 417 return (error); 418 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 419 TAILQ_FOREACH(m, &blacklist_head, listq) { 420 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 421 (uintmax_t)m->phys_addr); 422 first = 0; 423 } 424 error = sbuf_finish(&sbuf); 425 sbuf_delete(&sbuf); 426 return (error); 427 } 428 429 /* 430 * Initialize a dummy page for use in scans of the specified paging queue. 431 * In principle, this function only needs to set the flag PG_MARKER. 432 * Nonetheless, it write busies the page as a safety precaution. 433 */ 434 static void 435 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 436 { 437 438 bzero(marker, sizeof(*marker)); 439 marker->flags = PG_MARKER; 440 marker->a.flags = aflags; 441 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 442 marker->a.queue = queue; 443 } 444 445 static void 446 vm_page_domain_init(int domain) 447 { 448 struct vm_domain *vmd; 449 struct vm_pagequeue *pq; 450 int i; 451 452 vmd = VM_DOMAIN(domain); 453 bzero(vmd, sizeof(*vmd)); 454 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 455 "vm inactive pagequeue"; 456 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 457 "vm active pagequeue"; 458 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 459 "vm laundry pagequeue"; 460 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 461 "vm unswappable pagequeue"; 462 vmd->vmd_domain = domain; 463 vmd->vmd_page_count = 0; 464 vmd->vmd_free_count = 0; 465 vmd->vmd_segs = 0; 466 vmd->vmd_oom = FALSE; 467 for (i = 0; i < PQ_COUNT; i++) { 468 pq = &vmd->vmd_pagequeues[i]; 469 TAILQ_INIT(&pq->pq_pl); 470 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 471 MTX_DEF | MTX_DUPOK); 472 pq->pq_pdpages = 0; 473 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 474 } 475 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 476 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 477 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 478 479 /* 480 * inacthead is used to provide FIFO ordering for LRU-bypassing 481 * insertions. 482 */ 483 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 484 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 485 &vmd->vmd_inacthead, plinks.q); 486 487 /* 488 * The clock pages are used to implement active queue scanning without 489 * requeues. Scans start at clock[0], which is advanced after the scan 490 * ends. When the two clock hands meet, they are reset and scanning 491 * resumes from the head of the queue. 492 */ 493 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 494 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 495 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 496 &vmd->vmd_clock[0], plinks.q); 497 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 498 &vmd->vmd_clock[1], plinks.q); 499 } 500 501 /* 502 * Initialize a physical page in preparation for adding it to the free 503 * lists. 504 */ 505 static void 506 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 507 { 508 509 m->object = NULL; 510 m->ref_count = 0; 511 m->busy_lock = VPB_FREED; 512 m->flags = m->a.flags = 0; 513 m->phys_addr = pa; 514 m->a.queue = PQ_NONE; 515 m->psind = 0; 516 m->segind = segind; 517 m->order = VM_NFREEORDER; 518 m->pool = VM_FREEPOOL_DEFAULT; 519 m->valid = m->dirty = 0; 520 pmap_page_init(m); 521 } 522 523 #ifndef PMAP_HAS_PAGE_ARRAY 524 static vm_paddr_t 525 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 526 { 527 vm_paddr_t new_end; 528 529 /* 530 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 531 * However, because this page is allocated from KVM, out-of-bounds 532 * accesses using the direct map will not be trapped. 533 */ 534 *vaddr += PAGE_SIZE; 535 536 /* 537 * Allocate physical memory for the page structures, and map it. 538 */ 539 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 540 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 541 VM_PROT_READ | VM_PROT_WRITE); 542 vm_page_array_size = page_range; 543 544 return (new_end); 545 } 546 #endif 547 548 /* 549 * vm_page_startup: 550 * 551 * Initializes the resident memory module. Allocates physical memory for 552 * bootstrapping UMA and some data structures that are used to manage 553 * physical pages. Initializes these structures, and populates the free 554 * page queues. 555 */ 556 vm_offset_t 557 vm_page_startup(vm_offset_t vaddr) 558 { 559 struct vm_phys_seg *seg; 560 vm_page_t m; 561 char *list, *listend; 562 vm_paddr_t end, high_avail, low_avail, new_end, size; 563 vm_paddr_t page_range __unused; 564 vm_paddr_t last_pa, pa; 565 u_long pagecount; 566 int biggestone, i, segind; 567 #ifdef WITNESS 568 vm_offset_t mapped; 569 int witness_size; 570 #endif 571 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 572 long ii; 573 #endif 574 575 vaddr = round_page(vaddr); 576 577 vm_phys_early_startup(); 578 biggestone = vm_phys_avail_largest(); 579 end = phys_avail[biggestone+1]; 580 581 /* 582 * Initialize the page and queue locks. 583 */ 584 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 585 for (i = 0; i < PA_LOCK_COUNT; i++) 586 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 587 for (i = 0; i < vm_ndomains; i++) 588 vm_page_domain_init(i); 589 590 new_end = end; 591 #ifdef WITNESS 592 witness_size = round_page(witness_startup_count()); 593 new_end -= witness_size; 594 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 595 VM_PROT_READ | VM_PROT_WRITE); 596 bzero((void *)mapped, witness_size); 597 witness_startup((void *)mapped); 598 #endif 599 600 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 601 defined(__i386__) || defined(__mips__) || defined(__riscv) || \ 602 defined(__powerpc64__) 603 /* 604 * Allocate a bitmap to indicate that a random physical page 605 * needs to be included in a minidump. 606 * 607 * The amd64 port needs this to indicate which direct map pages 608 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 609 * 610 * However, i386 still needs this workspace internally within the 611 * minidump code. In theory, they are not needed on i386, but are 612 * included should the sf_buf code decide to use them. 613 */ 614 last_pa = 0; 615 for (i = 0; dump_avail[i + 1] != 0; i += 2) 616 if (dump_avail[i + 1] > last_pa) 617 last_pa = dump_avail[i + 1]; 618 page_range = last_pa / PAGE_SIZE; 619 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 620 new_end -= vm_page_dump_size; 621 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 622 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 623 bzero((void *)vm_page_dump, vm_page_dump_size); 624 #else 625 (void)last_pa; 626 #endif 627 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 628 defined(__riscv) || defined(__powerpc64__) 629 /* 630 * Include the UMA bootstrap pages, witness pages and vm_page_dump 631 * in a crash dump. When pmap_map() uses the direct map, they are 632 * not automatically included. 633 */ 634 for (pa = new_end; pa < end; pa += PAGE_SIZE) 635 dump_add_page(pa); 636 #endif 637 phys_avail[biggestone + 1] = new_end; 638 #ifdef __amd64__ 639 /* 640 * Request that the physical pages underlying the message buffer be 641 * included in a crash dump. Since the message buffer is accessed 642 * through the direct map, they are not automatically included. 643 */ 644 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 645 last_pa = pa + round_page(msgbufsize); 646 while (pa < last_pa) { 647 dump_add_page(pa); 648 pa += PAGE_SIZE; 649 } 650 #endif 651 /* 652 * Compute the number of pages of memory that will be available for 653 * use, taking into account the overhead of a page structure per page. 654 * In other words, solve 655 * "available physical memory" - round_page(page_range * 656 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 657 * for page_range. 658 */ 659 low_avail = phys_avail[0]; 660 high_avail = phys_avail[1]; 661 for (i = 0; i < vm_phys_nsegs; i++) { 662 if (vm_phys_segs[i].start < low_avail) 663 low_avail = vm_phys_segs[i].start; 664 if (vm_phys_segs[i].end > high_avail) 665 high_avail = vm_phys_segs[i].end; 666 } 667 /* Skip the first chunk. It is already accounted for. */ 668 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 669 if (phys_avail[i] < low_avail) 670 low_avail = phys_avail[i]; 671 if (phys_avail[i + 1] > high_avail) 672 high_avail = phys_avail[i + 1]; 673 } 674 first_page = low_avail / PAGE_SIZE; 675 #ifdef VM_PHYSSEG_SPARSE 676 size = 0; 677 for (i = 0; i < vm_phys_nsegs; i++) 678 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 679 for (i = 0; phys_avail[i + 1] != 0; i += 2) 680 size += phys_avail[i + 1] - phys_avail[i]; 681 #elif defined(VM_PHYSSEG_DENSE) 682 size = high_avail - low_avail; 683 #else 684 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 685 #endif 686 687 #ifdef PMAP_HAS_PAGE_ARRAY 688 pmap_page_array_startup(size / PAGE_SIZE); 689 biggestone = vm_phys_avail_largest(); 690 end = new_end = phys_avail[biggestone + 1]; 691 #else 692 #ifdef VM_PHYSSEG_DENSE 693 /* 694 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 695 * the overhead of a page structure per page only if vm_page_array is 696 * allocated from the last physical memory chunk. Otherwise, we must 697 * allocate page structures representing the physical memory 698 * underlying vm_page_array, even though they will not be used. 699 */ 700 if (new_end != high_avail) 701 page_range = size / PAGE_SIZE; 702 else 703 #endif 704 { 705 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 706 707 /* 708 * If the partial bytes remaining are large enough for 709 * a page (PAGE_SIZE) without a corresponding 710 * 'struct vm_page', then new_end will contain an 711 * extra page after subtracting the length of the VM 712 * page array. Compensate by subtracting an extra 713 * page from new_end. 714 */ 715 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 716 if (new_end == high_avail) 717 high_avail -= PAGE_SIZE; 718 new_end -= PAGE_SIZE; 719 } 720 } 721 end = new_end; 722 new_end = vm_page_array_alloc(&vaddr, end, page_range); 723 #endif 724 725 #if VM_NRESERVLEVEL > 0 726 /* 727 * Allocate physical memory for the reservation management system's 728 * data structures, and map it. 729 */ 730 new_end = vm_reserv_startup(&vaddr, new_end); 731 #endif 732 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 733 defined(__riscv) || defined(__powerpc64__) 734 /* 735 * Include vm_page_array and vm_reserv_array in a crash dump. 736 */ 737 for (pa = new_end; pa < end; pa += PAGE_SIZE) 738 dump_add_page(pa); 739 #endif 740 phys_avail[biggestone + 1] = new_end; 741 742 /* 743 * Add physical memory segments corresponding to the available 744 * physical pages. 745 */ 746 for (i = 0; phys_avail[i + 1] != 0; i += 2) 747 if (vm_phys_avail_size(i) != 0) 748 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 749 750 /* 751 * Initialize the physical memory allocator. 752 */ 753 vm_phys_init(); 754 755 /* 756 * Initialize the page structures and add every available page to the 757 * physical memory allocator's free lists. 758 */ 759 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 760 for (ii = 0; ii < vm_page_array_size; ii++) { 761 m = &vm_page_array[ii]; 762 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 763 m->flags = PG_FICTITIOUS; 764 } 765 #endif 766 vm_cnt.v_page_count = 0; 767 for (segind = 0; segind < vm_phys_nsegs; segind++) { 768 seg = &vm_phys_segs[segind]; 769 for (m = seg->first_page, pa = seg->start; pa < seg->end; 770 m++, pa += PAGE_SIZE) 771 vm_page_init_page(m, pa, segind); 772 773 /* 774 * Add the segment to the free lists only if it is covered by 775 * one of the ranges in phys_avail. Because we've added the 776 * ranges to the vm_phys_segs array, we can assume that each 777 * segment is either entirely contained in one of the ranges, 778 * or doesn't overlap any of them. 779 */ 780 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 781 struct vm_domain *vmd; 782 783 if (seg->start < phys_avail[i] || 784 seg->end > phys_avail[i + 1]) 785 continue; 786 787 m = seg->first_page; 788 pagecount = (u_long)atop(seg->end - seg->start); 789 790 vmd = VM_DOMAIN(seg->domain); 791 vm_domain_free_lock(vmd); 792 vm_phys_enqueue_contig(m, pagecount); 793 vm_domain_free_unlock(vmd); 794 vm_domain_freecnt_inc(vmd, pagecount); 795 vm_cnt.v_page_count += (u_int)pagecount; 796 797 vmd = VM_DOMAIN(seg->domain); 798 vmd->vmd_page_count += (u_int)pagecount; 799 vmd->vmd_segs |= 1UL << m->segind; 800 break; 801 } 802 } 803 804 /* 805 * Remove blacklisted pages from the physical memory allocator. 806 */ 807 TAILQ_INIT(&blacklist_head); 808 vm_page_blacklist_load(&list, &listend); 809 vm_page_blacklist_check(list, listend); 810 811 list = kern_getenv("vm.blacklist"); 812 vm_page_blacklist_check(list, NULL); 813 814 freeenv(list); 815 #if VM_NRESERVLEVEL > 0 816 /* 817 * Initialize the reservation management system. 818 */ 819 vm_reserv_init(); 820 #endif 821 822 return (vaddr); 823 } 824 825 void 826 vm_page_reference(vm_page_t m) 827 { 828 829 vm_page_aflag_set(m, PGA_REFERENCED); 830 } 831 832 static bool 833 vm_page_acquire_flags(vm_page_t m, int allocflags) 834 { 835 bool locked; 836 837 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 838 locked = vm_page_trysbusy(m); 839 else 840 locked = vm_page_tryxbusy(m); 841 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 842 vm_page_wire(m); 843 return (locked); 844 } 845 846 /* 847 * vm_page_busy_sleep_flags 848 * 849 * Sleep for busy according to VM_ALLOC_ parameters. 850 */ 851 static bool 852 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg, 853 int allocflags) 854 { 855 856 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 857 return (false); 858 /* 859 * Reference the page before unlocking and 860 * sleeping so that the page daemon is less 861 * likely to reclaim it. 862 */ 863 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 864 vm_page_aflag_set(m, PGA_REFERENCED); 865 if (_vm_page_busy_sleep(object, m, wmesg, (allocflags & 866 VM_ALLOC_IGN_SBUSY) != 0, true)) 867 VM_OBJECT_WLOCK(object); 868 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 869 return (false); 870 return (true); 871 } 872 873 /* 874 * vm_page_busy_acquire: 875 * 876 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 877 * and drop the object lock if necessary. 878 */ 879 bool 880 vm_page_busy_acquire(vm_page_t m, int allocflags) 881 { 882 vm_object_t obj; 883 bool locked; 884 885 /* 886 * The page-specific object must be cached because page 887 * identity can change during the sleep, causing the 888 * re-lock of a different object. 889 * It is assumed that a reference to the object is already 890 * held by the callers. 891 */ 892 obj = m->object; 893 for (;;) { 894 if (vm_page_acquire_flags(m, allocflags)) 895 return (true); 896 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 897 return (false); 898 if (obj != NULL) 899 locked = VM_OBJECT_WOWNED(obj); 900 else 901 locked = false; 902 MPASS(locked || vm_page_wired(m)); 903 if (_vm_page_busy_sleep(obj, m, "vmpba", 904 (allocflags & VM_ALLOC_SBUSY) != 0, locked)) 905 VM_OBJECT_WLOCK(obj); 906 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 907 return (false); 908 KASSERT(m->object == obj || m->object == NULL, 909 ("vm_page_busy_acquire: page %p does not belong to %p", 910 m, obj)); 911 } 912 } 913 914 /* 915 * vm_page_busy_downgrade: 916 * 917 * Downgrade an exclusive busy page into a single shared busy page. 918 */ 919 void 920 vm_page_busy_downgrade(vm_page_t m) 921 { 922 u_int x; 923 924 vm_page_assert_xbusied(m); 925 926 x = m->busy_lock; 927 for (;;) { 928 if (atomic_fcmpset_rel_int(&m->busy_lock, 929 &x, VPB_SHARERS_WORD(1))) 930 break; 931 } 932 if ((x & VPB_BIT_WAITERS) != 0) 933 wakeup(m); 934 } 935 936 /* 937 * 938 * vm_page_busy_tryupgrade: 939 * 940 * Attempt to upgrade a single shared busy into an exclusive busy. 941 */ 942 int 943 vm_page_busy_tryupgrade(vm_page_t m) 944 { 945 u_int ce, x; 946 947 vm_page_assert_sbusied(m); 948 949 x = m->busy_lock; 950 ce = VPB_CURTHREAD_EXCLUSIVE; 951 for (;;) { 952 if (VPB_SHARERS(x) > 1) 953 return (0); 954 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 955 ("vm_page_busy_tryupgrade: invalid lock state")); 956 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 957 ce | (x & VPB_BIT_WAITERS))) 958 continue; 959 return (1); 960 } 961 } 962 963 /* 964 * vm_page_sbusied: 965 * 966 * Return a positive value if the page is shared busied, 0 otherwise. 967 */ 968 int 969 vm_page_sbusied(vm_page_t m) 970 { 971 u_int x; 972 973 x = m->busy_lock; 974 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 975 } 976 977 /* 978 * vm_page_sunbusy: 979 * 980 * Shared unbusy a page. 981 */ 982 void 983 vm_page_sunbusy(vm_page_t m) 984 { 985 u_int x; 986 987 vm_page_assert_sbusied(m); 988 989 x = m->busy_lock; 990 for (;;) { 991 KASSERT(x != VPB_FREED, 992 ("vm_page_sunbusy: Unlocking freed page.")); 993 if (VPB_SHARERS(x) > 1) { 994 if (atomic_fcmpset_int(&m->busy_lock, &x, 995 x - VPB_ONE_SHARER)) 996 break; 997 continue; 998 } 999 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1000 ("vm_page_sunbusy: invalid lock state")); 1001 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1002 continue; 1003 if ((x & VPB_BIT_WAITERS) == 0) 1004 break; 1005 wakeup(m); 1006 break; 1007 } 1008 } 1009 1010 /* 1011 * vm_page_busy_sleep: 1012 * 1013 * Sleep if the page is busy, using the page pointer as wchan. 1014 * This is used to implement the hard-path of busying mechanism. 1015 * 1016 * If nonshared is true, sleep only if the page is xbusy. 1017 * 1018 * The object lock must be held on entry and will be released on exit. 1019 */ 1020 void 1021 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 1022 { 1023 vm_object_t obj; 1024 1025 obj = m->object; 1026 VM_OBJECT_ASSERT_LOCKED(obj); 1027 vm_page_lock_assert(m, MA_NOTOWNED); 1028 1029 if (!_vm_page_busy_sleep(obj, m, wmesg, nonshared, true)) 1030 VM_OBJECT_DROP(obj); 1031 } 1032 1033 /* 1034 * _vm_page_busy_sleep: 1035 * 1036 * Internal busy sleep function. 1037 */ 1038 static bool 1039 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, 1040 bool nonshared, bool locked) 1041 { 1042 u_int x; 1043 1044 /* 1045 * If the object is busy we must wait for that to drain to zero 1046 * before trying the page again. 1047 */ 1048 if (obj != NULL && vm_object_busied(obj)) { 1049 if (locked) 1050 VM_OBJECT_DROP(obj); 1051 vm_object_busy_wait(obj, wmesg); 1052 return (locked); 1053 } 1054 sleepq_lock(m); 1055 x = m->busy_lock; 1056 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 1057 ((x & VPB_BIT_WAITERS) == 0 && 1058 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 1059 sleepq_release(m); 1060 return (false); 1061 } 1062 if (locked) 1063 VM_OBJECT_DROP(obj); 1064 DROP_GIANT(); 1065 sleepq_add(m, NULL, wmesg, 0, 0); 1066 sleepq_wait(m, PVM); 1067 PICKUP_GIANT(); 1068 return (locked); 1069 } 1070 1071 /* 1072 * vm_page_trysbusy: 1073 * 1074 * Try to shared busy a page. 1075 * If the operation succeeds 1 is returned otherwise 0. 1076 * The operation never sleeps. 1077 */ 1078 int 1079 vm_page_trysbusy(vm_page_t m) 1080 { 1081 vm_object_t obj; 1082 u_int x; 1083 1084 obj = m->object; 1085 x = m->busy_lock; 1086 for (;;) { 1087 if ((x & VPB_BIT_SHARED) == 0) 1088 return (0); 1089 /* 1090 * Reduce the window for transient busies that will trigger 1091 * false negatives in vm_page_ps_test(). 1092 */ 1093 if (obj != NULL && vm_object_busied(obj)) 1094 return (0); 1095 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1096 x + VPB_ONE_SHARER)) 1097 break; 1098 } 1099 1100 /* Refetch the object now that we're guaranteed that it is stable. */ 1101 obj = m->object; 1102 if (obj != NULL && vm_object_busied(obj)) { 1103 vm_page_sunbusy(m); 1104 return (0); 1105 } 1106 return (1); 1107 } 1108 1109 /* 1110 * vm_page_tryxbusy: 1111 * 1112 * Try to exclusive busy a page. 1113 * If the operation succeeds 1 is returned otherwise 0. 1114 * The operation never sleeps. 1115 */ 1116 int 1117 vm_page_tryxbusy(vm_page_t m) 1118 { 1119 vm_object_t obj; 1120 1121 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, 1122 VPB_CURTHREAD_EXCLUSIVE) == 0) 1123 return (0); 1124 1125 obj = m->object; 1126 if (obj != NULL && vm_object_busied(obj)) { 1127 vm_page_xunbusy(m); 1128 return (0); 1129 } 1130 return (1); 1131 } 1132 1133 static void 1134 vm_page_xunbusy_hard_tail(vm_page_t m) 1135 { 1136 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1137 /* Wake the waiter. */ 1138 wakeup(m); 1139 } 1140 1141 /* 1142 * vm_page_xunbusy_hard: 1143 * 1144 * Called when unbusy has failed because there is a waiter. 1145 */ 1146 void 1147 vm_page_xunbusy_hard(vm_page_t m) 1148 { 1149 vm_page_assert_xbusied(m); 1150 vm_page_xunbusy_hard_tail(m); 1151 } 1152 1153 void 1154 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1155 { 1156 vm_page_assert_xbusied_unchecked(m); 1157 vm_page_xunbusy_hard_tail(m); 1158 } 1159 1160 static void 1161 vm_page_busy_free(vm_page_t m) 1162 { 1163 u_int x; 1164 1165 atomic_thread_fence_rel(); 1166 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1167 if ((x & VPB_BIT_WAITERS) != 0) 1168 wakeup(m); 1169 } 1170 1171 /* 1172 * vm_page_unhold_pages: 1173 * 1174 * Unhold each of the pages that is referenced by the given array. 1175 */ 1176 void 1177 vm_page_unhold_pages(vm_page_t *ma, int count) 1178 { 1179 1180 for (; count != 0; count--) { 1181 vm_page_unwire(*ma, PQ_ACTIVE); 1182 ma++; 1183 } 1184 } 1185 1186 vm_page_t 1187 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1188 { 1189 vm_page_t m; 1190 1191 #ifdef VM_PHYSSEG_SPARSE 1192 m = vm_phys_paddr_to_vm_page(pa); 1193 if (m == NULL) 1194 m = vm_phys_fictitious_to_vm_page(pa); 1195 return (m); 1196 #elif defined(VM_PHYSSEG_DENSE) 1197 long pi; 1198 1199 pi = atop(pa); 1200 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1201 m = &vm_page_array[pi - first_page]; 1202 return (m); 1203 } 1204 return (vm_phys_fictitious_to_vm_page(pa)); 1205 #else 1206 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1207 #endif 1208 } 1209 1210 /* 1211 * vm_page_getfake: 1212 * 1213 * Create a fictitious page with the specified physical address and 1214 * memory attribute. The memory attribute is the only the machine- 1215 * dependent aspect of a fictitious page that must be initialized. 1216 */ 1217 vm_page_t 1218 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1219 { 1220 vm_page_t m; 1221 1222 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1223 vm_page_initfake(m, paddr, memattr); 1224 return (m); 1225 } 1226 1227 void 1228 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1229 { 1230 1231 if ((m->flags & PG_FICTITIOUS) != 0) { 1232 /* 1233 * The page's memattr might have changed since the 1234 * previous initialization. Update the pmap to the 1235 * new memattr. 1236 */ 1237 goto memattr; 1238 } 1239 m->phys_addr = paddr; 1240 m->a.queue = PQ_NONE; 1241 /* Fictitious pages don't use "segind". */ 1242 m->flags = PG_FICTITIOUS; 1243 /* Fictitious pages don't use "order" or "pool". */ 1244 m->oflags = VPO_UNMANAGED; 1245 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1246 /* Fictitious pages are unevictable. */ 1247 m->ref_count = 1; 1248 pmap_page_init(m); 1249 memattr: 1250 pmap_page_set_memattr(m, memattr); 1251 } 1252 1253 /* 1254 * vm_page_putfake: 1255 * 1256 * Release a fictitious page. 1257 */ 1258 void 1259 vm_page_putfake(vm_page_t m) 1260 { 1261 1262 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1263 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1264 ("vm_page_putfake: bad page %p", m)); 1265 vm_page_assert_xbusied(m); 1266 vm_page_busy_free(m); 1267 uma_zfree(fakepg_zone, m); 1268 } 1269 1270 /* 1271 * vm_page_updatefake: 1272 * 1273 * Update the given fictitious page to the specified physical address and 1274 * memory attribute. 1275 */ 1276 void 1277 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1278 { 1279 1280 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1281 ("vm_page_updatefake: bad page %p", m)); 1282 m->phys_addr = paddr; 1283 pmap_page_set_memattr(m, memattr); 1284 } 1285 1286 /* 1287 * vm_page_free: 1288 * 1289 * Free a page. 1290 */ 1291 void 1292 vm_page_free(vm_page_t m) 1293 { 1294 1295 m->flags &= ~PG_ZERO; 1296 vm_page_free_toq(m); 1297 } 1298 1299 /* 1300 * vm_page_free_zero: 1301 * 1302 * Free a page to the zerod-pages queue 1303 */ 1304 void 1305 vm_page_free_zero(vm_page_t m) 1306 { 1307 1308 m->flags |= PG_ZERO; 1309 vm_page_free_toq(m); 1310 } 1311 1312 /* 1313 * Unbusy and handle the page queueing for a page from a getpages request that 1314 * was optionally read ahead or behind. 1315 */ 1316 void 1317 vm_page_readahead_finish(vm_page_t m) 1318 { 1319 1320 /* We shouldn't put invalid pages on queues. */ 1321 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1322 1323 /* 1324 * Since the page is not the actually needed one, whether it should 1325 * be activated or deactivated is not obvious. Empirical results 1326 * have shown that deactivating the page is usually the best choice, 1327 * unless the page is wanted by another thread. 1328 */ 1329 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1330 vm_page_activate(m); 1331 else 1332 vm_page_deactivate(m); 1333 vm_page_xunbusy_unchecked(m); 1334 } 1335 1336 /* 1337 * vm_page_sleep_if_busy: 1338 * 1339 * Sleep and release the object lock if the page is busied. 1340 * Returns TRUE if the thread slept. 1341 * 1342 * The given page must be unlocked and object containing it must 1343 * be locked. 1344 */ 1345 int 1346 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1347 { 1348 vm_object_t obj; 1349 1350 vm_page_lock_assert(m, MA_NOTOWNED); 1351 VM_OBJECT_ASSERT_WLOCKED(m->object); 1352 1353 /* 1354 * The page-specific object must be cached because page 1355 * identity can change during the sleep, causing the 1356 * re-lock of a different object. 1357 * It is assumed that a reference to the object is already 1358 * held by the callers. 1359 */ 1360 obj = m->object; 1361 if (vm_page_busied(m) || (obj != NULL && obj->busy)) { 1362 vm_page_busy_sleep(m, msg, false); 1363 VM_OBJECT_WLOCK(obj); 1364 return (TRUE); 1365 } 1366 return (FALSE); 1367 } 1368 1369 /* 1370 * vm_page_sleep_if_xbusy: 1371 * 1372 * Sleep and release the object lock if the page is xbusied. 1373 * Returns TRUE if the thread slept. 1374 * 1375 * The given page must be unlocked and object containing it must 1376 * be locked. 1377 */ 1378 int 1379 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg) 1380 { 1381 vm_object_t obj; 1382 1383 vm_page_lock_assert(m, MA_NOTOWNED); 1384 VM_OBJECT_ASSERT_WLOCKED(m->object); 1385 1386 /* 1387 * The page-specific object must be cached because page 1388 * identity can change during the sleep, causing the 1389 * re-lock of a different object. 1390 * It is assumed that a reference to the object is already 1391 * held by the callers. 1392 */ 1393 obj = m->object; 1394 if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) { 1395 vm_page_busy_sleep(m, msg, true); 1396 VM_OBJECT_WLOCK(obj); 1397 return (TRUE); 1398 } 1399 return (FALSE); 1400 } 1401 1402 /* 1403 * vm_page_dirty_KBI: [ internal use only ] 1404 * 1405 * Set all bits in the page's dirty field. 1406 * 1407 * The object containing the specified page must be locked if the 1408 * call is made from the machine-independent layer. 1409 * 1410 * See vm_page_clear_dirty_mask(). 1411 * 1412 * This function should only be called by vm_page_dirty(). 1413 */ 1414 void 1415 vm_page_dirty_KBI(vm_page_t m) 1416 { 1417 1418 /* Refer to this operation by its public name. */ 1419 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1420 m->dirty = VM_PAGE_BITS_ALL; 1421 } 1422 1423 /* 1424 * vm_page_insert: [ internal use only ] 1425 * 1426 * Inserts the given mem entry into the object and object list. 1427 * 1428 * The object must be locked. 1429 */ 1430 int 1431 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1432 { 1433 vm_page_t mpred; 1434 1435 VM_OBJECT_ASSERT_WLOCKED(object); 1436 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1437 return (vm_page_insert_after(m, object, pindex, mpred)); 1438 } 1439 1440 /* 1441 * vm_page_insert_after: 1442 * 1443 * Inserts the page "m" into the specified object at offset "pindex". 1444 * 1445 * The page "mpred" must immediately precede the offset "pindex" within 1446 * the specified object. 1447 * 1448 * The object must be locked. 1449 */ 1450 static int 1451 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1452 vm_page_t mpred) 1453 { 1454 vm_page_t msucc; 1455 1456 VM_OBJECT_ASSERT_WLOCKED(object); 1457 KASSERT(m->object == NULL, 1458 ("vm_page_insert_after: page already inserted")); 1459 if (mpred != NULL) { 1460 KASSERT(mpred->object == object, 1461 ("vm_page_insert_after: object doesn't contain mpred")); 1462 KASSERT(mpred->pindex < pindex, 1463 ("vm_page_insert_after: mpred doesn't precede pindex")); 1464 msucc = TAILQ_NEXT(mpred, listq); 1465 } else 1466 msucc = TAILQ_FIRST(&object->memq); 1467 if (msucc != NULL) 1468 KASSERT(msucc->pindex > pindex, 1469 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1470 1471 /* 1472 * Record the object/offset pair in this page. 1473 */ 1474 m->object = object; 1475 m->pindex = pindex; 1476 m->ref_count |= VPRC_OBJREF; 1477 1478 /* 1479 * Now link into the object's ordered list of backed pages. 1480 */ 1481 if (vm_radix_insert(&object->rtree, m)) { 1482 m->object = NULL; 1483 m->pindex = 0; 1484 m->ref_count &= ~VPRC_OBJREF; 1485 return (1); 1486 } 1487 vm_page_insert_radixdone(m, object, mpred); 1488 return (0); 1489 } 1490 1491 /* 1492 * vm_page_insert_radixdone: 1493 * 1494 * Complete page "m" insertion into the specified object after the 1495 * radix trie hooking. 1496 * 1497 * The page "mpred" must precede the offset "m->pindex" within the 1498 * specified object. 1499 * 1500 * The object must be locked. 1501 */ 1502 static void 1503 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1504 { 1505 1506 VM_OBJECT_ASSERT_WLOCKED(object); 1507 KASSERT(object != NULL && m->object == object, 1508 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1509 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1510 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1511 if (mpred != NULL) { 1512 KASSERT(mpred->object == object, 1513 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1514 KASSERT(mpred->pindex < m->pindex, 1515 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1516 } 1517 1518 if (mpred != NULL) 1519 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1520 else 1521 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1522 1523 /* 1524 * Show that the object has one more resident page. 1525 */ 1526 object->resident_page_count++; 1527 1528 /* 1529 * Hold the vnode until the last page is released. 1530 */ 1531 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1532 vhold(object->handle); 1533 1534 /* 1535 * Since we are inserting a new and possibly dirty page, 1536 * update the object's generation count. 1537 */ 1538 if (pmap_page_is_write_mapped(m)) 1539 vm_object_set_writeable_dirty(object); 1540 } 1541 1542 /* 1543 * Do the work to remove a page from its object. The caller is responsible for 1544 * updating the page's fields to reflect this removal. 1545 */ 1546 static void 1547 vm_page_object_remove(vm_page_t m) 1548 { 1549 vm_object_t object; 1550 vm_page_t mrem; 1551 1552 vm_page_assert_xbusied(m); 1553 object = m->object; 1554 VM_OBJECT_ASSERT_WLOCKED(object); 1555 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1556 ("page %p is missing its object ref", m)); 1557 1558 /* Deferred free of swap space. */ 1559 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1560 vm_pager_page_unswapped(m); 1561 1562 mrem = vm_radix_remove(&object->rtree, m->pindex); 1563 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1564 1565 /* 1566 * Now remove from the object's list of backed pages. 1567 */ 1568 TAILQ_REMOVE(&object->memq, m, listq); 1569 1570 /* 1571 * And show that the object has one fewer resident page. 1572 */ 1573 object->resident_page_count--; 1574 1575 /* 1576 * The vnode may now be recycled. 1577 */ 1578 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1579 vdrop(object->handle); 1580 } 1581 1582 /* 1583 * vm_page_remove: 1584 * 1585 * Removes the specified page from its containing object, but does not 1586 * invalidate any backing storage. Returns true if the object's reference 1587 * was the last reference to the page, and false otherwise. 1588 * 1589 * The object must be locked and the page must be exclusively busied. 1590 * The exclusive busy will be released on return. If this is not the 1591 * final ref and the caller does not hold a wire reference it may not 1592 * continue to access the page. 1593 */ 1594 bool 1595 vm_page_remove(vm_page_t m) 1596 { 1597 bool dropped; 1598 1599 dropped = vm_page_remove_xbusy(m); 1600 vm_page_xunbusy(m); 1601 1602 return (dropped); 1603 } 1604 1605 /* 1606 * vm_page_remove_xbusy 1607 * 1608 * Removes the page but leaves the xbusy held. Returns true if this 1609 * removed the final ref and false otherwise. 1610 */ 1611 bool 1612 vm_page_remove_xbusy(vm_page_t m) 1613 { 1614 1615 vm_page_object_remove(m); 1616 m->object = NULL; 1617 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1618 } 1619 1620 /* 1621 * vm_page_lookup: 1622 * 1623 * Returns the page associated with the object/offset 1624 * pair specified; if none is found, NULL is returned. 1625 * 1626 * The object must be locked. 1627 */ 1628 vm_page_t 1629 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1630 { 1631 1632 VM_OBJECT_ASSERT_LOCKED(object); 1633 return (vm_radix_lookup(&object->rtree, pindex)); 1634 } 1635 1636 /* 1637 * vm_page_find_least: 1638 * 1639 * Returns the page associated with the object with least pindex 1640 * greater than or equal to the parameter pindex, or NULL. 1641 * 1642 * The object must be locked. 1643 */ 1644 vm_page_t 1645 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1646 { 1647 vm_page_t m; 1648 1649 VM_OBJECT_ASSERT_LOCKED(object); 1650 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1651 m = vm_radix_lookup_ge(&object->rtree, pindex); 1652 return (m); 1653 } 1654 1655 /* 1656 * Returns the given page's successor (by pindex) within the object if it is 1657 * resident; if none is found, NULL is returned. 1658 * 1659 * The object must be locked. 1660 */ 1661 vm_page_t 1662 vm_page_next(vm_page_t m) 1663 { 1664 vm_page_t next; 1665 1666 VM_OBJECT_ASSERT_LOCKED(m->object); 1667 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1668 MPASS(next->object == m->object); 1669 if (next->pindex != m->pindex + 1) 1670 next = NULL; 1671 } 1672 return (next); 1673 } 1674 1675 /* 1676 * Returns the given page's predecessor (by pindex) within the object if it is 1677 * resident; if none is found, NULL is returned. 1678 * 1679 * The object must be locked. 1680 */ 1681 vm_page_t 1682 vm_page_prev(vm_page_t m) 1683 { 1684 vm_page_t prev; 1685 1686 VM_OBJECT_ASSERT_LOCKED(m->object); 1687 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1688 MPASS(prev->object == m->object); 1689 if (prev->pindex != m->pindex - 1) 1690 prev = NULL; 1691 } 1692 return (prev); 1693 } 1694 1695 /* 1696 * Uses the page mnew as a replacement for an existing page at index 1697 * pindex which must be already present in the object. 1698 * 1699 * Both pages must be exclusively busied on enter. The old page is 1700 * unbusied on exit. 1701 * 1702 * A return value of true means mold is now free. If this is not the 1703 * final ref and the caller does not hold a wire reference it may not 1704 * continue to access the page. 1705 */ 1706 static bool 1707 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1708 vm_page_t mold) 1709 { 1710 vm_page_t mret; 1711 bool dropped; 1712 1713 VM_OBJECT_ASSERT_WLOCKED(object); 1714 vm_page_assert_xbusied(mold); 1715 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1716 ("vm_page_replace: page %p already in object", mnew)); 1717 1718 /* 1719 * This function mostly follows vm_page_insert() and 1720 * vm_page_remove() without the radix, object count and vnode 1721 * dance. Double check such functions for more comments. 1722 */ 1723 1724 mnew->object = object; 1725 mnew->pindex = pindex; 1726 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1727 mret = vm_radix_replace(&object->rtree, mnew); 1728 KASSERT(mret == mold, 1729 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1730 KASSERT((mold->oflags & VPO_UNMANAGED) == 1731 (mnew->oflags & VPO_UNMANAGED), 1732 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1733 1734 /* Keep the resident page list in sorted order. */ 1735 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1736 TAILQ_REMOVE(&object->memq, mold, listq); 1737 mold->object = NULL; 1738 1739 /* 1740 * The object's resident_page_count does not change because we have 1741 * swapped one page for another, but the generation count should 1742 * change if the page is dirty. 1743 */ 1744 if (pmap_page_is_write_mapped(mnew)) 1745 vm_object_set_writeable_dirty(object); 1746 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1747 vm_page_xunbusy(mold); 1748 1749 return (dropped); 1750 } 1751 1752 void 1753 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1754 vm_page_t mold) 1755 { 1756 1757 vm_page_assert_xbusied(mnew); 1758 1759 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1760 vm_page_free(mold); 1761 } 1762 1763 /* 1764 * vm_page_rename: 1765 * 1766 * Move the given memory entry from its 1767 * current object to the specified target object/offset. 1768 * 1769 * Note: swap associated with the page must be invalidated by the move. We 1770 * have to do this for several reasons: (1) we aren't freeing the 1771 * page, (2) we are dirtying the page, (3) the VM system is probably 1772 * moving the page from object A to B, and will then later move 1773 * the backing store from A to B and we can't have a conflict. 1774 * 1775 * Note: we *always* dirty the page. It is necessary both for the 1776 * fact that we moved it, and because we may be invalidating 1777 * swap. 1778 * 1779 * The objects must be locked. 1780 */ 1781 int 1782 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1783 { 1784 vm_page_t mpred; 1785 vm_pindex_t opidx; 1786 1787 VM_OBJECT_ASSERT_WLOCKED(new_object); 1788 1789 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1790 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1791 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1792 ("vm_page_rename: pindex already renamed")); 1793 1794 /* 1795 * Create a custom version of vm_page_insert() which does not depend 1796 * by m_prev and can cheat on the implementation aspects of the 1797 * function. 1798 */ 1799 opidx = m->pindex; 1800 m->pindex = new_pindex; 1801 if (vm_radix_insert(&new_object->rtree, m)) { 1802 m->pindex = opidx; 1803 return (1); 1804 } 1805 1806 /* 1807 * The operation cannot fail anymore. The removal must happen before 1808 * the listq iterator is tainted. 1809 */ 1810 m->pindex = opidx; 1811 vm_page_object_remove(m); 1812 1813 /* Return back to the new pindex to complete vm_page_insert(). */ 1814 m->pindex = new_pindex; 1815 m->object = new_object; 1816 1817 vm_page_insert_radixdone(m, new_object, mpred); 1818 vm_page_dirty(m); 1819 return (0); 1820 } 1821 1822 /* 1823 * vm_page_alloc: 1824 * 1825 * Allocate and return a page that is associated with the specified 1826 * object and offset pair. By default, this page is exclusive busied. 1827 * 1828 * The caller must always specify an allocation class. 1829 * 1830 * allocation classes: 1831 * VM_ALLOC_NORMAL normal process request 1832 * VM_ALLOC_SYSTEM system *really* needs a page 1833 * VM_ALLOC_INTERRUPT interrupt time request 1834 * 1835 * optional allocation flags: 1836 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1837 * intends to allocate 1838 * VM_ALLOC_NOBUSY do not exclusive busy the page 1839 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1840 * VM_ALLOC_NOOBJ page is not associated with an object and 1841 * should not be exclusive busy 1842 * VM_ALLOC_SBUSY shared busy the allocated page 1843 * VM_ALLOC_WIRED wire the allocated page 1844 * VM_ALLOC_ZERO prefer a zeroed page 1845 */ 1846 vm_page_t 1847 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1848 { 1849 1850 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1851 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1852 } 1853 1854 vm_page_t 1855 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1856 int req) 1857 { 1858 1859 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1860 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1861 NULL)); 1862 } 1863 1864 /* 1865 * Allocate a page in the specified object with the given page index. To 1866 * optimize insertion of the page into the object, the caller must also specifiy 1867 * the resident page in the object with largest index smaller than the given 1868 * page index, or NULL if no such page exists. 1869 */ 1870 vm_page_t 1871 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1872 int req, vm_page_t mpred) 1873 { 1874 struct vm_domainset_iter di; 1875 vm_page_t m; 1876 int domain; 1877 1878 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1879 do { 1880 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1881 mpred); 1882 if (m != NULL) 1883 break; 1884 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1885 1886 return (m); 1887 } 1888 1889 /* 1890 * Returns true if the number of free pages exceeds the minimum 1891 * for the request class and false otherwise. 1892 */ 1893 static int 1894 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 1895 { 1896 u_int limit, old, new; 1897 1898 if (req_class == VM_ALLOC_INTERRUPT) 1899 limit = 0; 1900 else if (req_class == VM_ALLOC_SYSTEM) 1901 limit = vmd->vmd_interrupt_free_min; 1902 else 1903 limit = vmd->vmd_free_reserved; 1904 1905 /* 1906 * Attempt to reserve the pages. Fail if we're below the limit. 1907 */ 1908 limit += npages; 1909 old = vmd->vmd_free_count; 1910 do { 1911 if (old < limit) 1912 return (0); 1913 new = old - npages; 1914 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 1915 1916 /* Wake the page daemon if we've crossed the threshold. */ 1917 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 1918 pagedaemon_wakeup(vmd->vmd_domain); 1919 1920 /* Only update bitsets on transitions. */ 1921 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 1922 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 1923 vm_domain_set(vmd); 1924 1925 return (1); 1926 } 1927 1928 int 1929 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 1930 { 1931 int req_class; 1932 1933 /* 1934 * The page daemon is allowed to dig deeper into the free page list. 1935 */ 1936 req_class = req & VM_ALLOC_CLASS_MASK; 1937 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1938 req_class = VM_ALLOC_SYSTEM; 1939 return (_vm_domain_allocate(vmd, req_class, npages)); 1940 } 1941 1942 vm_page_t 1943 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 1944 int req, vm_page_t mpred) 1945 { 1946 struct vm_domain *vmd; 1947 vm_page_t m; 1948 int flags, pool; 1949 1950 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1951 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1952 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1953 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1954 ("inconsistent object(%p)/req(%x)", object, req)); 1955 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1956 ("Can't sleep and retry object insertion.")); 1957 KASSERT(mpred == NULL || mpred->pindex < pindex, 1958 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1959 (uintmax_t)pindex)); 1960 if (object != NULL) 1961 VM_OBJECT_ASSERT_WLOCKED(object); 1962 1963 flags = 0; 1964 m = NULL; 1965 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; 1966 again: 1967 #if VM_NRESERVLEVEL > 0 1968 /* 1969 * Can we allocate the page from a reservation? 1970 */ 1971 if (vm_object_reserv(object) && 1972 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 1973 NULL) { 1974 domain = vm_phys_domain(m); 1975 vmd = VM_DOMAIN(domain); 1976 goto found; 1977 } 1978 #endif 1979 vmd = VM_DOMAIN(domain); 1980 if (vmd->vmd_pgcache[pool].zone != NULL) { 1981 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT); 1982 if (m != NULL) { 1983 flags |= PG_PCPU_CACHE; 1984 goto found; 1985 } 1986 } 1987 if (vm_domain_allocate(vmd, req, 1)) { 1988 /* 1989 * If not, allocate it from the free page queues. 1990 */ 1991 vm_domain_free_lock(vmd); 1992 m = vm_phys_alloc_pages(domain, pool, 0); 1993 vm_domain_free_unlock(vmd); 1994 if (m == NULL) { 1995 vm_domain_freecnt_inc(vmd, 1); 1996 #if VM_NRESERVLEVEL > 0 1997 if (vm_reserv_reclaim_inactive(domain)) 1998 goto again; 1999 #endif 2000 } 2001 } 2002 if (m == NULL) { 2003 /* 2004 * Not allocatable, give up. 2005 */ 2006 if (vm_domain_alloc_fail(vmd, object, req)) 2007 goto again; 2008 return (NULL); 2009 } 2010 2011 /* 2012 * At this point we had better have found a good page. 2013 */ 2014 found: 2015 vm_page_dequeue(m); 2016 vm_page_alloc_check(m); 2017 2018 /* 2019 * Initialize the page. Only the PG_ZERO flag is inherited. 2020 */ 2021 if ((req & VM_ALLOC_ZERO) != 0) 2022 flags |= (m->flags & PG_ZERO); 2023 if ((req & VM_ALLOC_NODUMP) != 0) 2024 flags |= PG_NODUMP; 2025 m->flags = flags; 2026 m->a.flags = 0; 2027 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2028 VPO_UNMANAGED : 0; 2029 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2030 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2031 else if ((req & VM_ALLOC_SBUSY) != 0) 2032 m->busy_lock = VPB_SHARERS_WORD(1); 2033 else 2034 m->busy_lock = VPB_UNBUSIED; 2035 if (req & VM_ALLOC_WIRED) { 2036 vm_wire_add(1); 2037 m->ref_count = 1; 2038 } 2039 m->a.act_count = 0; 2040 2041 if (object != NULL) { 2042 if (vm_page_insert_after(m, object, pindex, mpred)) { 2043 if (req & VM_ALLOC_WIRED) { 2044 vm_wire_sub(1); 2045 m->ref_count = 0; 2046 } 2047 KASSERT(m->object == NULL, ("page %p has object", m)); 2048 m->oflags = VPO_UNMANAGED; 2049 m->busy_lock = VPB_UNBUSIED; 2050 /* Don't change PG_ZERO. */ 2051 vm_page_free_toq(m); 2052 if (req & VM_ALLOC_WAITFAIL) { 2053 VM_OBJECT_WUNLOCK(object); 2054 vm_radix_wait(); 2055 VM_OBJECT_WLOCK(object); 2056 } 2057 return (NULL); 2058 } 2059 2060 /* Ignore device objects; the pager sets "memattr" for them. */ 2061 if (object->memattr != VM_MEMATTR_DEFAULT && 2062 (object->flags & OBJ_FICTITIOUS) == 0) 2063 pmap_page_set_memattr(m, object->memattr); 2064 } else 2065 m->pindex = pindex; 2066 2067 return (m); 2068 } 2069 2070 /* 2071 * vm_page_alloc_contig: 2072 * 2073 * Allocate a contiguous set of physical pages of the given size "npages" 2074 * from the free lists. All of the physical pages must be at or above 2075 * the given physical address "low" and below the given physical address 2076 * "high". The given value "alignment" determines the alignment of the 2077 * first physical page in the set. If the given value "boundary" is 2078 * non-zero, then the set of physical pages cannot cross any physical 2079 * address boundary that is a multiple of that value. Both "alignment" 2080 * and "boundary" must be a power of two. 2081 * 2082 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2083 * then the memory attribute setting for the physical pages is configured 2084 * to the object's memory attribute setting. Otherwise, the memory 2085 * attribute setting for the physical pages is configured to "memattr", 2086 * overriding the object's memory attribute setting. However, if the 2087 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2088 * memory attribute setting for the physical pages cannot be configured 2089 * to VM_MEMATTR_DEFAULT. 2090 * 2091 * The specified object may not contain fictitious pages. 2092 * 2093 * The caller must always specify an allocation class. 2094 * 2095 * allocation classes: 2096 * VM_ALLOC_NORMAL normal process request 2097 * VM_ALLOC_SYSTEM system *really* needs a page 2098 * VM_ALLOC_INTERRUPT interrupt time request 2099 * 2100 * optional allocation flags: 2101 * VM_ALLOC_NOBUSY do not exclusive busy the page 2102 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2103 * VM_ALLOC_NOOBJ page is not associated with an object and 2104 * should not be exclusive busy 2105 * VM_ALLOC_SBUSY shared busy the allocated page 2106 * VM_ALLOC_WIRED wire the allocated page 2107 * VM_ALLOC_ZERO prefer a zeroed page 2108 */ 2109 vm_page_t 2110 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2111 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2112 vm_paddr_t boundary, vm_memattr_t memattr) 2113 { 2114 struct vm_domainset_iter di; 2115 vm_page_t m; 2116 int domain; 2117 2118 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2119 do { 2120 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2121 npages, low, high, alignment, boundary, memattr); 2122 if (m != NULL) 2123 break; 2124 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2125 2126 return (m); 2127 } 2128 2129 vm_page_t 2130 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2131 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2132 vm_paddr_t boundary, vm_memattr_t memattr) 2133 { 2134 struct vm_domain *vmd; 2135 vm_page_t m, m_ret, mpred; 2136 u_int busy_lock, flags, oflags; 2137 2138 mpred = NULL; /* XXX: pacify gcc */ 2139 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2140 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2141 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2142 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2143 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 2144 req)); 2145 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2146 ("Can't sleep and retry object insertion.")); 2147 if (object != NULL) { 2148 VM_OBJECT_ASSERT_WLOCKED(object); 2149 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2150 ("vm_page_alloc_contig: object %p has fictitious pages", 2151 object)); 2152 } 2153 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2154 2155 if (object != NULL) { 2156 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2157 KASSERT(mpred == NULL || mpred->pindex != pindex, 2158 ("vm_page_alloc_contig: pindex already allocated")); 2159 } 2160 2161 /* 2162 * Can we allocate the pages without the number of free pages falling 2163 * below the lower bound for the allocation class? 2164 */ 2165 m_ret = NULL; 2166 again: 2167 #if VM_NRESERVLEVEL > 0 2168 /* 2169 * Can we allocate the pages from a reservation? 2170 */ 2171 if (vm_object_reserv(object) && 2172 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2173 mpred, npages, low, high, alignment, boundary)) != NULL) { 2174 domain = vm_phys_domain(m_ret); 2175 vmd = VM_DOMAIN(domain); 2176 goto found; 2177 } 2178 #endif 2179 vmd = VM_DOMAIN(domain); 2180 if (vm_domain_allocate(vmd, req, npages)) { 2181 /* 2182 * allocate them from the free page queues. 2183 */ 2184 vm_domain_free_lock(vmd); 2185 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2186 alignment, boundary); 2187 vm_domain_free_unlock(vmd); 2188 if (m_ret == NULL) { 2189 vm_domain_freecnt_inc(vmd, npages); 2190 #if VM_NRESERVLEVEL > 0 2191 if (vm_reserv_reclaim_contig(domain, npages, low, 2192 high, alignment, boundary)) 2193 goto again; 2194 #endif 2195 } 2196 } 2197 if (m_ret == NULL) { 2198 if (vm_domain_alloc_fail(vmd, object, req)) 2199 goto again; 2200 return (NULL); 2201 } 2202 #if VM_NRESERVLEVEL > 0 2203 found: 2204 #endif 2205 for (m = m_ret; m < &m_ret[npages]; m++) { 2206 vm_page_dequeue(m); 2207 vm_page_alloc_check(m); 2208 } 2209 2210 /* 2211 * Initialize the pages. Only the PG_ZERO flag is inherited. 2212 */ 2213 flags = 0; 2214 if ((req & VM_ALLOC_ZERO) != 0) 2215 flags = PG_ZERO; 2216 if ((req & VM_ALLOC_NODUMP) != 0) 2217 flags |= PG_NODUMP; 2218 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2219 VPO_UNMANAGED : 0; 2220 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2221 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2222 else if ((req & VM_ALLOC_SBUSY) != 0) 2223 busy_lock = VPB_SHARERS_WORD(1); 2224 else 2225 busy_lock = VPB_UNBUSIED; 2226 if ((req & VM_ALLOC_WIRED) != 0) 2227 vm_wire_add(npages); 2228 if (object != NULL) { 2229 if (object->memattr != VM_MEMATTR_DEFAULT && 2230 memattr == VM_MEMATTR_DEFAULT) 2231 memattr = object->memattr; 2232 } 2233 for (m = m_ret; m < &m_ret[npages]; m++) { 2234 m->a.flags = 0; 2235 m->flags = (m->flags | PG_NODUMP) & flags; 2236 m->busy_lock = busy_lock; 2237 if ((req & VM_ALLOC_WIRED) != 0) 2238 m->ref_count = 1; 2239 m->a.act_count = 0; 2240 m->oflags = oflags; 2241 if (object != NULL) { 2242 if (vm_page_insert_after(m, object, pindex, mpred)) { 2243 if ((req & VM_ALLOC_WIRED) != 0) 2244 vm_wire_sub(npages); 2245 KASSERT(m->object == NULL, 2246 ("page %p has object", m)); 2247 mpred = m; 2248 for (m = m_ret; m < &m_ret[npages]; m++) { 2249 if (m <= mpred && 2250 (req & VM_ALLOC_WIRED) != 0) 2251 m->ref_count = 0; 2252 m->oflags = VPO_UNMANAGED; 2253 m->busy_lock = VPB_UNBUSIED; 2254 /* Don't change PG_ZERO. */ 2255 vm_page_free_toq(m); 2256 } 2257 if (req & VM_ALLOC_WAITFAIL) { 2258 VM_OBJECT_WUNLOCK(object); 2259 vm_radix_wait(); 2260 VM_OBJECT_WLOCK(object); 2261 } 2262 return (NULL); 2263 } 2264 mpred = m; 2265 } else 2266 m->pindex = pindex; 2267 if (memattr != VM_MEMATTR_DEFAULT) 2268 pmap_page_set_memattr(m, memattr); 2269 pindex++; 2270 } 2271 return (m_ret); 2272 } 2273 2274 /* 2275 * Check a page that has been freshly dequeued from a freelist. 2276 */ 2277 static void 2278 vm_page_alloc_check(vm_page_t m) 2279 { 2280 2281 KASSERT(m->object == NULL, ("page %p has object", m)); 2282 KASSERT(m->a.queue == PQ_NONE && 2283 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2284 ("page %p has unexpected queue %d, flags %#x", 2285 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2286 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2287 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2288 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2289 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2290 ("page %p has unexpected memattr %d", 2291 m, pmap_page_get_memattr(m))); 2292 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2293 } 2294 2295 /* 2296 * vm_page_alloc_freelist: 2297 * 2298 * Allocate a physical page from the specified free page list. 2299 * 2300 * The caller must always specify an allocation class. 2301 * 2302 * allocation classes: 2303 * VM_ALLOC_NORMAL normal process request 2304 * VM_ALLOC_SYSTEM system *really* needs a page 2305 * VM_ALLOC_INTERRUPT interrupt time request 2306 * 2307 * optional allocation flags: 2308 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2309 * intends to allocate 2310 * VM_ALLOC_WIRED wire the allocated page 2311 * VM_ALLOC_ZERO prefer a zeroed page 2312 */ 2313 vm_page_t 2314 vm_page_alloc_freelist(int freelist, int req) 2315 { 2316 struct vm_domainset_iter di; 2317 vm_page_t m; 2318 int domain; 2319 2320 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2321 do { 2322 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2323 if (m != NULL) 2324 break; 2325 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2326 2327 return (m); 2328 } 2329 2330 vm_page_t 2331 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2332 { 2333 struct vm_domain *vmd; 2334 vm_page_t m; 2335 u_int flags; 2336 2337 m = NULL; 2338 vmd = VM_DOMAIN(domain); 2339 again: 2340 if (vm_domain_allocate(vmd, req, 1)) { 2341 vm_domain_free_lock(vmd); 2342 m = vm_phys_alloc_freelist_pages(domain, freelist, 2343 VM_FREEPOOL_DIRECT, 0); 2344 vm_domain_free_unlock(vmd); 2345 if (m == NULL) 2346 vm_domain_freecnt_inc(vmd, 1); 2347 } 2348 if (m == NULL) { 2349 if (vm_domain_alloc_fail(vmd, NULL, req)) 2350 goto again; 2351 return (NULL); 2352 } 2353 vm_page_dequeue(m); 2354 vm_page_alloc_check(m); 2355 2356 /* 2357 * Initialize the page. Only the PG_ZERO flag is inherited. 2358 */ 2359 m->a.flags = 0; 2360 flags = 0; 2361 if ((req & VM_ALLOC_ZERO) != 0) 2362 flags = PG_ZERO; 2363 m->flags &= flags; 2364 if ((req & VM_ALLOC_WIRED) != 0) { 2365 vm_wire_add(1); 2366 m->ref_count = 1; 2367 } 2368 /* Unmanaged pages don't use "act_count". */ 2369 m->oflags = VPO_UNMANAGED; 2370 return (m); 2371 } 2372 2373 static int 2374 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2375 { 2376 struct vm_domain *vmd; 2377 struct vm_pgcache *pgcache; 2378 int i; 2379 2380 pgcache = arg; 2381 vmd = VM_DOMAIN(pgcache->domain); 2382 2383 /* 2384 * The page daemon should avoid creating extra memory pressure since its 2385 * main purpose is to replenish the store of free pages. 2386 */ 2387 if (vmd->vmd_severeset || curproc == pageproc || 2388 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2389 return (0); 2390 domain = vmd->vmd_domain; 2391 vm_domain_free_lock(vmd); 2392 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2393 (vm_page_t *)store); 2394 vm_domain_free_unlock(vmd); 2395 if (cnt != i) 2396 vm_domain_freecnt_inc(vmd, cnt - i); 2397 2398 return (i); 2399 } 2400 2401 static void 2402 vm_page_zone_release(void *arg, void **store, int cnt) 2403 { 2404 struct vm_domain *vmd; 2405 struct vm_pgcache *pgcache; 2406 vm_page_t m; 2407 int i; 2408 2409 pgcache = arg; 2410 vmd = VM_DOMAIN(pgcache->domain); 2411 vm_domain_free_lock(vmd); 2412 for (i = 0; i < cnt; i++) { 2413 m = (vm_page_t)store[i]; 2414 vm_phys_free_pages(m, 0); 2415 } 2416 vm_domain_free_unlock(vmd); 2417 vm_domain_freecnt_inc(vmd, cnt); 2418 } 2419 2420 #define VPSC_ANY 0 /* No restrictions. */ 2421 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2422 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2423 2424 /* 2425 * vm_page_scan_contig: 2426 * 2427 * Scan vm_page_array[] between the specified entries "m_start" and 2428 * "m_end" for a run of contiguous physical pages that satisfy the 2429 * specified conditions, and return the lowest page in the run. The 2430 * specified "alignment" determines the alignment of the lowest physical 2431 * page in the run. If the specified "boundary" is non-zero, then the 2432 * run of physical pages cannot span a physical address that is a 2433 * multiple of "boundary". 2434 * 2435 * "m_end" is never dereferenced, so it need not point to a vm_page 2436 * structure within vm_page_array[]. 2437 * 2438 * "npages" must be greater than zero. "m_start" and "m_end" must not 2439 * span a hole (or discontiguity) in the physical address space. Both 2440 * "alignment" and "boundary" must be a power of two. 2441 */ 2442 vm_page_t 2443 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2444 u_long alignment, vm_paddr_t boundary, int options) 2445 { 2446 vm_object_t object; 2447 vm_paddr_t pa; 2448 vm_page_t m, m_run; 2449 #if VM_NRESERVLEVEL > 0 2450 int level; 2451 #endif 2452 int m_inc, order, run_ext, run_len; 2453 2454 KASSERT(npages > 0, ("npages is 0")); 2455 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2456 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2457 m_run = NULL; 2458 run_len = 0; 2459 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2460 KASSERT((m->flags & PG_MARKER) == 0, 2461 ("page %p is PG_MARKER", m)); 2462 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2463 ("fictitious page %p has invalid ref count", m)); 2464 2465 /* 2466 * If the current page would be the start of a run, check its 2467 * physical address against the end, alignment, and boundary 2468 * conditions. If it doesn't satisfy these conditions, either 2469 * terminate the scan or advance to the next page that 2470 * satisfies the failed condition. 2471 */ 2472 if (run_len == 0) { 2473 KASSERT(m_run == NULL, ("m_run != NULL")); 2474 if (m + npages > m_end) 2475 break; 2476 pa = VM_PAGE_TO_PHYS(m); 2477 if ((pa & (alignment - 1)) != 0) { 2478 m_inc = atop(roundup2(pa, alignment) - pa); 2479 continue; 2480 } 2481 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2482 boundary) != 0) { 2483 m_inc = atop(roundup2(pa, boundary) - pa); 2484 continue; 2485 } 2486 } else 2487 KASSERT(m_run != NULL, ("m_run == NULL")); 2488 2489 retry: 2490 m_inc = 1; 2491 if (vm_page_wired(m)) 2492 run_ext = 0; 2493 #if VM_NRESERVLEVEL > 0 2494 else if ((level = vm_reserv_level(m)) >= 0 && 2495 (options & VPSC_NORESERV) != 0) { 2496 run_ext = 0; 2497 /* Advance to the end of the reservation. */ 2498 pa = VM_PAGE_TO_PHYS(m); 2499 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2500 pa); 2501 } 2502 #endif 2503 else if ((object = 2504 (vm_object_t)atomic_load_ptr(&m->object)) != NULL) { 2505 /* 2506 * The page is considered eligible for relocation if 2507 * and only if it could be laundered or reclaimed by 2508 * the page daemon. 2509 */ 2510 VM_OBJECT_RLOCK(object); 2511 if (object != m->object) { 2512 VM_OBJECT_RUNLOCK(object); 2513 goto retry; 2514 } 2515 /* Don't care: PG_NODUMP, PG_ZERO. */ 2516 if (object->type != OBJT_DEFAULT && 2517 object->type != OBJT_SWAP && 2518 object->type != OBJT_VNODE) { 2519 run_ext = 0; 2520 #if VM_NRESERVLEVEL > 0 2521 } else if ((options & VPSC_NOSUPER) != 0 && 2522 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2523 run_ext = 0; 2524 /* Advance to the end of the superpage. */ 2525 pa = VM_PAGE_TO_PHYS(m); 2526 m_inc = atop(roundup2(pa + 1, 2527 vm_reserv_size(level)) - pa); 2528 #endif 2529 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2530 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2531 /* 2532 * The page is allocated but eligible for 2533 * relocation. Extend the current run by one 2534 * page. 2535 */ 2536 KASSERT(pmap_page_get_memattr(m) == 2537 VM_MEMATTR_DEFAULT, 2538 ("page %p has an unexpected memattr", m)); 2539 KASSERT((m->oflags & (VPO_SWAPINPROG | 2540 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2541 ("page %p has unexpected oflags", m)); 2542 /* Don't care: PGA_NOSYNC. */ 2543 run_ext = 1; 2544 } else 2545 run_ext = 0; 2546 VM_OBJECT_RUNLOCK(object); 2547 #if VM_NRESERVLEVEL > 0 2548 } else if (level >= 0) { 2549 /* 2550 * The page is reserved but not yet allocated. In 2551 * other words, it is still free. Extend the current 2552 * run by one page. 2553 */ 2554 run_ext = 1; 2555 #endif 2556 } else if ((order = m->order) < VM_NFREEORDER) { 2557 /* 2558 * The page is enqueued in the physical memory 2559 * allocator's free page queues. Moreover, it is the 2560 * first page in a power-of-two-sized run of 2561 * contiguous free pages. Add these pages to the end 2562 * of the current run, and jump ahead. 2563 */ 2564 run_ext = 1 << order; 2565 m_inc = 1 << order; 2566 } else { 2567 /* 2568 * Skip the page for one of the following reasons: (1) 2569 * It is enqueued in the physical memory allocator's 2570 * free page queues. However, it is not the first 2571 * page in a run of contiguous free pages. (This case 2572 * rarely occurs because the scan is performed in 2573 * ascending order.) (2) It is not reserved, and it is 2574 * transitioning from free to allocated. (Conversely, 2575 * the transition from allocated to free for managed 2576 * pages is blocked by the page lock.) (3) It is 2577 * allocated but not contained by an object and not 2578 * wired, e.g., allocated by Xen's balloon driver. 2579 */ 2580 run_ext = 0; 2581 } 2582 2583 /* 2584 * Extend or reset the current run of pages. 2585 */ 2586 if (run_ext > 0) { 2587 if (run_len == 0) 2588 m_run = m; 2589 run_len += run_ext; 2590 } else { 2591 if (run_len > 0) { 2592 m_run = NULL; 2593 run_len = 0; 2594 } 2595 } 2596 } 2597 if (run_len >= npages) 2598 return (m_run); 2599 return (NULL); 2600 } 2601 2602 /* 2603 * vm_page_reclaim_run: 2604 * 2605 * Try to relocate each of the allocated virtual pages within the 2606 * specified run of physical pages to a new physical address. Free the 2607 * physical pages underlying the relocated virtual pages. A virtual page 2608 * is relocatable if and only if it could be laundered or reclaimed by 2609 * the page daemon. Whenever possible, a virtual page is relocated to a 2610 * physical address above "high". 2611 * 2612 * Returns 0 if every physical page within the run was already free or 2613 * just freed by a successful relocation. Otherwise, returns a non-zero 2614 * value indicating why the last attempt to relocate a virtual page was 2615 * unsuccessful. 2616 * 2617 * "req_class" must be an allocation class. 2618 */ 2619 static int 2620 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2621 vm_paddr_t high) 2622 { 2623 struct vm_domain *vmd; 2624 struct spglist free; 2625 vm_object_t object; 2626 vm_paddr_t pa; 2627 vm_page_t m, m_end, m_new; 2628 int error, order, req; 2629 2630 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2631 ("req_class is not an allocation class")); 2632 SLIST_INIT(&free); 2633 error = 0; 2634 m = m_run; 2635 m_end = m_run + npages; 2636 for (; error == 0 && m < m_end; m++) { 2637 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2638 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2639 2640 /* 2641 * Racily check for wirings. Races are handled once the object 2642 * lock is held and the page is unmapped. 2643 */ 2644 if (vm_page_wired(m)) 2645 error = EBUSY; 2646 else if ((object = 2647 (vm_object_t)atomic_load_ptr(&m->object)) != NULL) { 2648 /* 2649 * The page is relocated if and only if it could be 2650 * laundered or reclaimed by the page daemon. 2651 */ 2652 VM_OBJECT_WLOCK(object); 2653 /* Don't care: PG_NODUMP, PG_ZERO. */ 2654 if (m->object != object || 2655 (object->type != OBJT_DEFAULT && 2656 object->type != OBJT_SWAP && 2657 object->type != OBJT_VNODE)) 2658 error = EINVAL; 2659 else if (object->memattr != VM_MEMATTR_DEFAULT) 2660 error = EINVAL; 2661 else if (vm_page_queue(m) != PQ_NONE && 2662 vm_page_tryxbusy(m) != 0) { 2663 if (vm_page_wired(m)) { 2664 vm_page_xunbusy(m); 2665 error = EBUSY; 2666 goto unlock; 2667 } 2668 KASSERT(pmap_page_get_memattr(m) == 2669 VM_MEMATTR_DEFAULT, 2670 ("page %p has an unexpected memattr", m)); 2671 KASSERT(m->oflags == 0, 2672 ("page %p has unexpected oflags", m)); 2673 /* Don't care: PGA_NOSYNC. */ 2674 if (!vm_page_none_valid(m)) { 2675 /* 2676 * First, try to allocate a new page 2677 * that is above "high". Failing 2678 * that, try to allocate a new page 2679 * that is below "m_run". Allocate 2680 * the new page between the end of 2681 * "m_run" and "high" only as a last 2682 * resort. 2683 */ 2684 req = req_class | VM_ALLOC_NOOBJ; 2685 if ((m->flags & PG_NODUMP) != 0) 2686 req |= VM_ALLOC_NODUMP; 2687 if (trunc_page(high) != 2688 ~(vm_paddr_t)PAGE_MASK) { 2689 m_new = vm_page_alloc_contig( 2690 NULL, 0, req, 1, 2691 round_page(high), 2692 ~(vm_paddr_t)0, 2693 PAGE_SIZE, 0, 2694 VM_MEMATTR_DEFAULT); 2695 } else 2696 m_new = NULL; 2697 if (m_new == NULL) { 2698 pa = VM_PAGE_TO_PHYS(m_run); 2699 m_new = vm_page_alloc_contig( 2700 NULL, 0, req, 1, 2701 0, pa - 1, PAGE_SIZE, 0, 2702 VM_MEMATTR_DEFAULT); 2703 } 2704 if (m_new == NULL) { 2705 pa += ptoa(npages); 2706 m_new = vm_page_alloc_contig( 2707 NULL, 0, req, 1, 2708 pa, high, PAGE_SIZE, 0, 2709 VM_MEMATTR_DEFAULT); 2710 } 2711 if (m_new == NULL) { 2712 vm_page_xunbusy(m); 2713 error = ENOMEM; 2714 goto unlock; 2715 } 2716 2717 /* 2718 * Unmap the page and check for new 2719 * wirings that may have been acquired 2720 * through a pmap lookup. 2721 */ 2722 if (object->ref_count != 0 && 2723 !vm_page_try_remove_all(m)) { 2724 vm_page_xunbusy(m); 2725 vm_page_free(m_new); 2726 error = EBUSY; 2727 goto unlock; 2728 } 2729 2730 /* 2731 * Replace "m" with the new page. For 2732 * vm_page_replace(), "m" must be busy 2733 * and dequeued. Finally, change "m" 2734 * as if vm_page_free() was called. 2735 */ 2736 m_new->a.flags = m->a.flags & 2737 ~PGA_QUEUE_STATE_MASK; 2738 KASSERT(m_new->oflags == VPO_UNMANAGED, 2739 ("page %p is managed", m_new)); 2740 m_new->oflags = 0; 2741 pmap_copy_page(m, m_new); 2742 m_new->valid = m->valid; 2743 m_new->dirty = m->dirty; 2744 m->flags &= ~PG_ZERO; 2745 vm_page_dequeue(m); 2746 if (vm_page_replace_hold(m_new, object, 2747 m->pindex, m) && 2748 vm_page_free_prep(m)) 2749 SLIST_INSERT_HEAD(&free, m, 2750 plinks.s.ss); 2751 2752 /* 2753 * The new page must be deactivated 2754 * before the object is unlocked. 2755 */ 2756 vm_page_deactivate(m_new); 2757 } else { 2758 m->flags &= ~PG_ZERO; 2759 vm_page_dequeue(m); 2760 if (vm_page_free_prep(m)) 2761 SLIST_INSERT_HEAD(&free, m, 2762 plinks.s.ss); 2763 KASSERT(m->dirty == 0, 2764 ("page %p is dirty", m)); 2765 } 2766 } else 2767 error = EBUSY; 2768 unlock: 2769 VM_OBJECT_WUNLOCK(object); 2770 } else { 2771 MPASS(vm_phys_domain(m) == domain); 2772 vmd = VM_DOMAIN(domain); 2773 vm_domain_free_lock(vmd); 2774 order = m->order; 2775 if (order < VM_NFREEORDER) { 2776 /* 2777 * The page is enqueued in the physical memory 2778 * allocator's free page queues. Moreover, it 2779 * is the first page in a power-of-two-sized 2780 * run of contiguous free pages. Jump ahead 2781 * to the last page within that run, and 2782 * continue from there. 2783 */ 2784 m += (1 << order) - 1; 2785 } 2786 #if VM_NRESERVLEVEL > 0 2787 else if (vm_reserv_is_page_free(m)) 2788 order = 0; 2789 #endif 2790 vm_domain_free_unlock(vmd); 2791 if (order == VM_NFREEORDER) 2792 error = EINVAL; 2793 } 2794 } 2795 if ((m = SLIST_FIRST(&free)) != NULL) { 2796 int cnt; 2797 2798 vmd = VM_DOMAIN(domain); 2799 cnt = 0; 2800 vm_domain_free_lock(vmd); 2801 do { 2802 MPASS(vm_phys_domain(m) == domain); 2803 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2804 vm_phys_free_pages(m, 0); 2805 cnt++; 2806 } while ((m = SLIST_FIRST(&free)) != NULL); 2807 vm_domain_free_unlock(vmd); 2808 vm_domain_freecnt_inc(vmd, cnt); 2809 } 2810 return (error); 2811 } 2812 2813 #define NRUNS 16 2814 2815 CTASSERT(powerof2(NRUNS)); 2816 2817 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2818 2819 #define MIN_RECLAIM 8 2820 2821 /* 2822 * vm_page_reclaim_contig: 2823 * 2824 * Reclaim allocated, contiguous physical memory satisfying the specified 2825 * conditions by relocating the virtual pages using that physical memory. 2826 * Returns true if reclamation is successful and false otherwise. Since 2827 * relocation requires the allocation of physical pages, reclamation may 2828 * fail due to a shortage of free pages. When reclamation fails, callers 2829 * are expected to perform vm_wait() before retrying a failed allocation 2830 * operation, e.g., vm_page_alloc_contig(). 2831 * 2832 * The caller must always specify an allocation class through "req". 2833 * 2834 * allocation classes: 2835 * VM_ALLOC_NORMAL normal process request 2836 * VM_ALLOC_SYSTEM system *really* needs a page 2837 * VM_ALLOC_INTERRUPT interrupt time request 2838 * 2839 * The optional allocation flags are ignored. 2840 * 2841 * "npages" must be greater than zero. Both "alignment" and "boundary" 2842 * must be a power of two. 2843 */ 2844 bool 2845 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2846 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2847 { 2848 struct vm_domain *vmd; 2849 vm_paddr_t curr_low; 2850 vm_page_t m_run, m_runs[NRUNS]; 2851 u_long count, reclaimed; 2852 int error, i, options, req_class; 2853 2854 KASSERT(npages > 0, ("npages is 0")); 2855 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2856 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2857 req_class = req & VM_ALLOC_CLASS_MASK; 2858 2859 /* 2860 * The page daemon is allowed to dig deeper into the free page list. 2861 */ 2862 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2863 req_class = VM_ALLOC_SYSTEM; 2864 2865 /* 2866 * Return if the number of free pages cannot satisfy the requested 2867 * allocation. 2868 */ 2869 vmd = VM_DOMAIN(domain); 2870 count = vmd->vmd_free_count; 2871 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2872 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2873 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2874 return (false); 2875 2876 /* 2877 * Scan up to three times, relaxing the restrictions ("options") on 2878 * the reclamation of reservations and superpages each time. 2879 */ 2880 for (options = VPSC_NORESERV;;) { 2881 /* 2882 * Find the highest runs that satisfy the given constraints 2883 * and restrictions, and record them in "m_runs". 2884 */ 2885 curr_low = low; 2886 count = 0; 2887 for (;;) { 2888 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2889 high, alignment, boundary, options); 2890 if (m_run == NULL) 2891 break; 2892 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2893 m_runs[RUN_INDEX(count)] = m_run; 2894 count++; 2895 } 2896 2897 /* 2898 * Reclaim the highest runs in LIFO (descending) order until 2899 * the number of reclaimed pages, "reclaimed", is at least 2900 * MIN_RECLAIM. Reset "reclaimed" each time because each 2901 * reclamation is idempotent, and runs will (likely) recur 2902 * from one scan to the next as restrictions are relaxed. 2903 */ 2904 reclaimed = 0; 2905 for (i = 0; count > 0 && i < NRUNS; i++) { 2906 count--; 2907 m_run = m_runs[RUN_INDEX(count)]; 2908 error = vm_page_reclaim_run(req_class, domain, npages, 2909 m_run, high); 2910 if (error == 0) { 2911 reclaimed += npages; 2912 if (reclaimed >= MIN_RECLAIM) 2913 return (true); 2914 } 2915 } 2916 2917 /* 2918 * Either relax the restrictions on the next scan or return if 2919 * the last scan had no restrictions. 2920 */ 2921 if (options == VPSC_NORESERV) 2922 options = VPSC_NOSUPER; 2923 else if (options == VPSC_NOSUPER) 2924 options = VPSC_ANY; 2925 else if (options == VPSC_ANY) 2926 return (reclaimed != 0); 2927 } 2928 } 2929 2930 bool 2931 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2932 u_long alignment, vm_paddr_t boundary) 2933 { 2934 struct vm_domainset_iter di; 2935 int domain; 2936 bool ret; 2937 2938 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2939 do { 2940 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 2941 high, alignment, boundary); 2942 if (ret) 2943 break; 2944 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2945 2946 return (ret); 2947 } 2948 2949 /* 2950 * Set the domain in the appropriate page level domainset. 2951 */ 2952 void 2953 vm_domain_set(struct vm_domain *vmd) 2954 { 2955 2956 mtx_lock(&vm_domainset_lock); 2957 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 2958 vmd->vmd_minset = 1; 2959 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 2960 } 2961 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 2962 vmd->vmd_severeset = 1; 2963 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 2964 } 2965 mtx_unlock(&vm_domainset_lock); 2966 } 2967 2968 /* 2969 * Clear the domain from the appropriate page level domainset. 2970 */ 2971 void 2972 vm_domain_clear(struct vm_domain *vmd) 2973 { 2974 2975 mtx_lock(&vm_domainset_lock); 2976 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 2977 vmd->vmd_minset = 0; 2978 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 2979 if (vm_min_waiters != 0) { 2980 vm_min_waiters = 0; 2981 wakeup(&vm_min_domains); 2982 } 2983 } 2984 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 2985 vmd->vmd_severeset = 0; 2986 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 2987 if (vm_severe_waiters != 0) { 2988 vm_severe_waiters = 0; 2989 wakeup(&vm_severe_domains); 2990 } 2991 } 2992 2993 /* 2994 * If pageout daemon needs pages, then tell it that there are 2995 * some free. 2996 */ 2997 if (vmd->vmd_pageout_pages_needed && 2998 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 2999 wakeup(&vmd->vmd_pageout_pages_needed); 3000 vmd->vmd_pageout_pages_needed = 0; 3001 } 3002 3003 /* See comments in vm_wait_doms(). */ 3004 if (vm_pageproc_waiters) { 3005 vm_pageproc_waiters = 0; 3006 wakeup(&vm_pageproc_waiters); 3007 } 3008 mtx_unlock(&vm_domainset_lock); 3009 } 3010 3011 /* 3012 * Wait for free pages to exceed the min threshold globally. 3013 */ 3014 void 3015 vm_wait_min(void) 3016 { 3017 3018 mtx_lock(&vm_domainset_lock); 3019 while (vm_page_count_min()) { 3020 vm_min_waiters++; 3021 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3022 } 3023 mtx_unlock(&vm_domainset_lock); 3024 } 3025 3026 /* 3027 * Wait for free pages to exceed the severe threshold globally. 3028 */ 3029 void 3030 vm_wait_severe(void) 3031 { 3032 3033 mtx_lock(&vm_domainset_lock); 3034 while (vm_page_count_severe()) { 3035 vm_severe_waiters++; 3036 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3037 "vmwait", 0); 3038 } 3039 mtx_unlock(&vm_domainset_lock); 3040 } 3041 3042 u_int 3043 vm_wait_count(void) 3044 { 3045 3046 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3047 } 3048 3049 void 3050 vm_wait_doms(const domainset_t *wdoms) 3051 { 3052 3053 /* 3054 * We use racey wakeup synchronization to avoid expensive global 3055 * locking for the pageproc when sleeping with a non-specific vm_wait. 3056 * To handle this, we only sleep for one tick in this instance. It 3057 * is expected that most allocations for the pageproc will come from 3058 * kmem or vm_page_grab* which will use the more specific and 3059 * race-free vm_wait_domain(). 3060 */ 3061 if (curproc == pageproc) { 3062 mtx_lock(&vm_domainset_lock); 3063 vm_pageproc_waiters++; 3064 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, 3065 "pageprocwait", 1); 3066 } else { 3067 /* 3068 * XXX Ideally we would wait only until the allocation could 3069 * be satisfied. This condition can cause new allocators to 3070 * consume all freed pages while old allocators wait. 3071 */ 3072 mtx_lock(&vm_domainset_lock); 3073 if (vm_page_count_min_set(wdoms)) { 3074 vm_min_waiters++; 3075 msleep(&vm_min_domains, &vm_domainset_lock, 3076 PVM | PDROP, "vmwait", 0); 3077 } else 3078 mtx_unlock(&vm_domainset_lock); 3079 } 3080 } 3081 3082 /* 3083 * vm_wait_domain: 3084 * 3085 * Sleep until free pages are available for allocation. 3086 * - Called in various places after failed memory allocations. 3087 */ 3088 void 3089 vm_wait_domain(int domain) 3090 { 3091 struct vm_domain *vmd; 3092 domainset_t wdom; 3093 3094 vmd = VM_DOMAIN(domain); 3095 vm_domain_free_assert_unlocked(vmd); 3096 3097 if (curproc == pageproc) { 3098 mtx_lock(&vm_domainset_lock); 3099 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3100 vmd->vmd_pageout_pages_needed = 1; 3101 msleep(&vmd->vmd_pageout_pages_needed, 3102 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3103 } else 3104 mtx_unlock(&vm_domainset_lock); 3105 } else { 3106 if (pageproc == NULL) 3107 panic("vm_wait in early boot"); 3108 DOMAINSET_ZERO(&wdom); 3109 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3110 vm_wait_doms(&wdom); 3111 } 3112 } 3113 3114 /* 3115 * vm_wait: 3116 * 3117 * Sleep until free pages are available for allocation in the 3118 * affinity domains of the obj. If obj is NULL, the domain set 3119 * for the calling thread is used. 3120 * Called in various places after failed memory allocations. 3121 */ 3122 void 3123 vm_wait(vm_object_t obj) 3124 { 3125 struct domainset *d; 3126 3127 d = NULL; 3128 3129 /* 3130 * Carefully fetch pointers only once: the struct domainset 3131 * itself is ummutable but the pointer might change. 3132 */ 3133 if (obj != NULL) 3134 d = obj->domain.dr_policy; 3135 if (d == NULL) 3136 d = curthread->td_domain.dr_policy; 3137 3138 vm_wait_doms(&d->ds_mask); 3139 } 3140 3141 /* 3142 * vm_domain_alloc_fail: 3143 * 3144 * Called when a page allocation function fails. Informs the 3145 * pagedaemon and performs the requested wait. Requires the 3146 * domain_free and object lock on entry. Returns with the 3147 * object lock held and free lock released. Returns an error when 3148 * retry is necessary. 3149 * 3150 */ 3151 static int 3152 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3153 { 3154 3155 vm_domain_free_assert_unlocked(vmd); 3156 3157 atomic_add_int(&vmd->vmd_pageout_deficit, 3158 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3159 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3160 if (object != NULL) 3161 VM_OBJECT_WUNLOCK(object); 3162 vm_wait_domain(vmd->vmd_domain); 3163 if (object != NULL) 3164 VM_OBJECT_WLOCK(object); 3165 if (req & VM_ALLOC_WAITOK) 3166 return (EAGAIN); 3167 } 3168 3169 return (0); 3170 } 3171 3172 /* 3173 * vm_waitpfault: 3174 * 3175 * Sleep until free pages are available for allocation. 3176 * - Called only in vm_fault so that processes page faulting 3177 * can be easily tracked. 3178 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3179 * processes will be able to grab memory first. Do not change 3180 * this balance without careful testing first. 3181 */ 3182 void 3183 vm_waitpfault(struct domainset *dset, int timo) 3184 { 3185 3186 /* 3187 * XXX Ideally we would wait only until the allocation could 3188 * be satisfied. This condition can cause new allocators to 3189 * consume all freed pages while old allocators wait. 3190 */ 3191 mtx_lock(&vm_domainset_lock); 3192 if (vm_page_count_min_set(&dset->ds_mask)) { 3193 vm_min_waiters++; 3194 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3195 "pfault", timo); 3196 } else 3197 mtx_unlock(&vm_domainset_lock); 3198 } 3199 3200 static struct vm_pagequeue * 3201 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3202 { 3203 3204 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3205 } 3206 3207 #ifdef INVARIANTS 3208 static struct vm_pagequeue * 3209 vm_page_pagequeue(vm_page_t m) 3210 { 3211 3212 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3213 } 3214 #endif 3215 3216 static __always_inline bool 3217 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3218 { 3219 vm_page_astate_t tmp; 3220 3221 tmp = *old; 3222 do { 3223 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3224 return (true); 3225 counter_u64_add(pqstate_commit_retries, 1); 3226 } while (old->_bits == tmp._bits); 3227 3228 return (false); 3229 } 3230 3231 /* 3232 * Do the work of committing a queue state update that moves the page out of 3233 * its current queue. 3234 */ 3235 static bool 3236 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3237 vm_page_astate_t *old, vm_page_astate_t new) 3238 { 3239 vm_page_t next; 3240 3241 vm_pagequeue_assert_locked(pq); 3242 KASSERT(vm_page_pagequeue(m) == pq, 3243 ("%s: queue %p does not match page %p", __func__, pq, m)); 3244 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3245 ("%s: invalid queue indices %d %d", 3246 __func__, old->queue, new.queue)); 3247 3248 /* 3249 * Once the queue index of the page changes there is nothing 3250 * synchronizing with further updates to the page's physical 3251 * queue state. Therefore we must speculatively remove the page 3252 * from the queue now and be prepared to roll back if the queue 3253 * state update fails. If the page is not physically enqueued then 3254 * we just update its queue index. 3255 */ 3256 if ((old->flags & PGA_ENQUEUED) != 0) { 3257 new.flags &= ~PGA_ENQUEUED; 3258 next = TAILQ_NEXT(m, plinks.q); 3259 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3260 vm_pagequeue_cnt_dec(pq); 3261 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3262 if (next == NULL) 3263 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3264 else 3265 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3266 vm_pagequeue_cnt_inc(pq); 3267 return (false); 3268 } else { 3269 return (true); 3270 } 3271 } else { 3272 return (vm_page_pqstate_fcmpset(m, old, new)); 3273 } 3274 } 3275 3276 static bool 3277 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3278 vm_page_astate_t new) 3279 { 3280 struct vm_pagequeue *pq; 3281 vm_page_astate_t as; 3282 bool ret; 3283 3284 pq = _vm_page_pagequeue(m, old->queue); 3285 3286 /* 3287 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3288 * corresponding page queue lock is held. 3289 */ 3290 vm_pagequeue_lock(pq); 3291 as = vm_page_astate_load(m); 3292 if (__predict_false(as._bits != old->_bits)) { 3293 *old = as; 3294 ret = false; 3295 } else { 3296 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3297 } 3298 vm_pagequeue_unlock(pq); 3299 return (ret); 3300 } 3301 3302 /* 3303 * Commit a queue state update that enqueues or requeues a page. 3304 */ 3305 static bool 3306 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3307 vm_page_astate_t *old, vm_page_astate_t new) 3308 { 3309 struct vm_domain *vmd; 3310 3311 vm_pagequeue_assert_locked(pq); 3312 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3313 ("%s: invalid queue indices %d %d", 3314 __func__, old->queue, new.queue)); 3315 3316 new.flags |= PGA_ENQUEUED; 3317 if (!vm_page_pqstate_fcmpset(m, old, new)) 3318 return (false); 3319 3320 if ((old->flags & PGA_ENQUEUED) != 0) 3321 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3322 else 3323 vm_pagequeue_cnt_inc(pq); 3324 3325 /* 3326 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3327 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3328 * applied, even if it was set first. 3329 */ 3330 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3331 vmd = vm_pagequeue_domain(m); 3332 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3333 ("%s: invalid page queue for page %p", __func__, m)); 3334 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3335 } else { 3336 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3337 } 3338 return (true); 3339 } 3340 3341 /* 3342 * Commit a queue state update that encodes a request for a deferred queue 3343 * operation. 3344 */ 3345 static bool 3346 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3347 vm_page_astate_t new) 3348 { 3349 3350 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3351 ("%s: invalid state, queue %d flags %x", 3352 __func__, new.queue, new.flags)); 3353 3354 if (old->_bits != new._bits && 3355 !vm_page_pqstate_fcmpset(m, old, new)) 3356 return (false); 3357 vm_page_pqbatch_submit(m, new.queue); 3358 return (true); 3359 } 3360 3361 /* 3362 * A generic queue state update function. This handles more cases than the 3363 * specialized functions above. 3364 */ 3365 bool 3366 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3367 { 3368 3369 if (old->_bits == new._bits) 3370 return (true); 3371 3372 if (old->queue != PQ_NONE && new.queue != old->queue) { 3373 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3374 return (false); 3375 if (new.queue != PQ_NONE) 3376 vm_page_pqbatch_submit(m, new.queue); 3377 } else { 3378 if (!vm_page_pqstate_fcmpset(m, old, new)) 3379 return (false); 3380 if (new.queue != PQ_NONE && 3381 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3382 vm_page_pqbatch_submit(m, new.queue); 3383 } 3384 return (true); 3385 } 3386 3387 /* 3388 * Apply deferred queue state updates to a page. 3389 */ 3390 static inline void 3391 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3392 { 3393 vm_page_astate_t new, old; 3394 3395 CRITICAL_ASSERT(curthread); 3396 vm_pagequeue_assert_locked(pq); 3397 KASSERT(queue < PQ_COUNT, 3398 ("%s: invalid queue index %d", __func__, queue)); 3399 KASSERT(pq == _vm_page_pagequeue(m, queue), 3400 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3401 3402 for (old = vm_page_astate_load(m);;) { 3403 if (__predict_false(old.queue != queue || 3404 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3405 counter_u64_add(queue_nops, 1); 3406 break; 3407 } 3408 KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0, 3409 ("%s: page %p has unexpected queue state", __func__, m)); 3410 3411 new = old; 3412 if ((old.flags & PGA_DEQUEUE) != 0) { 3413 new.flags &= ~PGA_QUEUE_OP_MASK; 3414 new.queue = PQ_NONE; 3415 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3416 m, &old, new))) { 3417 counter_u64_add(queue_ops, 1); 3418 break; 3419 } 3420 } else { 3421 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3422 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3423 m, &old, new))) { 3424 counter_u64_add(queue_ops, 1); 3425 break; 3426 } 3427 } 3428 } 3429 } 3430 3431 static void 3432 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3433 uint8_t queue) 3434 { 3435 int i; 3436 3437 for (i = 0; i < bq->bq_cnt; i++) 3438 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3439 vm_batchqueue_init(bq); 3440 } 3441 3442 /* 3443 * vm_page_pqbatch_submit: [ internal use only ] 3444 * 3445 * Enqueue a page in the specified page queue's batched work queue. 3446 * The caller must have encoded the requested operation in the page 3447 * structure's a.flags field. 3448 */ 3449 void 3450 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3451 { 3452 struct vm_batchqueue *bq; 3453 struct vm_pagequeue *pq; 3454 int domain; 3455 3456 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3457 ("page %p is unmanaged", m)); 3458 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3459 3460 domain = vm_phys_domain(m); 3461 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 3462 3463 critical_enter(); 3464 bq = DPCPU_PTR(pqbatch[domain][queue]); 3465 if (vm_batchqueue_insert(bq, m)) { 3466 critical_exit(); 3467 return; 3468 } 3469 critical_exit(); 3470 vm_pagequeue_lock(pq); 3471 critical_enter(); 3472 bq = DPCPU_PTR(pqbatch[domain][queue]); 3473 vm_pqbatch_process(pq, bq, queue); 3474 vm_pqbatch_process_page(pq, m, queue); 3475 vm_pagequeue_unlock(pq); 3476 critical_exit(); 3477 } 3478 3479 /* 3480 * vm_page_pqbatch_drain: [ internal use only ] 3481 * 3482 * Force all per-CPU page queue batch queues to be drained. This is 3483 * intended for use in severe memory shortages, to ensure that pages 3484 * do not remain stuck in the batch queues. 3485 */ 3486 void 3487 vm_page_pqbatch_drain(void) 3488 { 3489 struct thread *td; 3490 struct vm_domain *vmd; 3491 struct vm_pagequeue *pq; 3492 int cpu, domain, queue; 3493 3494 td = curthread; 3495 CPU_FOREACH(cpu) { 3496 thread_lock(td); 3497 sched_bind(td, cpu); 3498 thread_unlock(td); 3499 3500 for (domain = 0; domain < vm_ndomains; domain++) { 3501 vmd = VM_DOMAIN(domain); 3502 for (queue = 0; queue < PQ_COUNT; queue++) { 3503 pq = &vmd->vmd_pagequeues[queue]; 3504 vm_pagequeue_lock(pq); 3505 critical_enter(); 3506 vm_pqbatch_process(pq, 3507 DPCPU_PTR(pqbatch[domain][queue]), queue); 3508 critical_exit(); 3509 vm_pagequeue_unlock(pq); 3510 } 3511 } 3512 } 3513 thread_lock(td); 3514 sched_unbind(td); 3515 thread_unlock(td); 3516 } 3517 3518 /* 3519 * vm_page_dequeue_deferred: [ internal use only ] 3520 * 3521 * Request removal of the given page from its current page 3522 * queue. Physical removal from the queue may be deferred 3523 * indefinitely. 3524 * 3525 * The page must be locked. 3526 */ 3527 void 3528 vm_page_dequeue_deferred(vm_page_t m) 3529 { 3530 vm_page_astate_t new, old; 3531 3532 old = vm_page_astate_load(m); 3533 do { 3534 if (old.queue == PQ_NONE) { 3535 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3536 ("%s: page %p has unexpected queue state", 3537 __func__, m)); 3538 break; 3539 } 3540 new = old; 3541 new.flags |= PGA_DEQUEUE; 3542 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3543 } 3544 3545 /* 3546 * vm_page_dequeue: 3547 * 3548 * Remove the page from whichever page queue it's in, if any, before 3549 * returning. 3550 */ 3551 void 3552 vm_page_dequeue(vm_page_t m) 3553 { 3554 vm_page_astate_t new, old; 3555 3556 old = vm_page_astate_load(m); 3557 do { 3558 if (old.queue == PQ_NONE) { 3559 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3560 ("%s: page %p has unexpected queue state", 3561 __func__, m)); 3562 break; 3563 } 3564 new = old; 3565 new.flags &= ~PGA_QUEUE_OP_MASK; 3566 new.queue = PQ_NONE; 3567 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 3568 3569 } 3570 3571 /* 3572 * Schedule the given page for insertion into the specified page queue. 3573 * Physical insertion of the page may be deferred indefinitely. 3574 */ 3575 static void 3576 vm_page_enqueue(vm_page_t m, uint8_t queue) 3577 { 3578 3579 KASSERT(m->a.queue == PQ_NONE && 3580 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 3581 ("%s: page %p is already enqueued", __func__, m)); 3582 KASSERT(m->ref_count > 0, 3583 ("%s: page %p does not carry any references", __func__, m)); 3584 3585 m->a.queue = queue; 3586 if ((m->a.flags & PGA_REQUEUE) == 0) 3587 vm_page_aflag_set(m, PGA_REQUEUE); 3588 vm_page_pqbatch_submit(m, queue); 3589 } 3590 3591 /* 3592 * vm_page_free_prep: 3593 * 3594 * Prepares the given page to be put on the free list, 3595 * disassociating it from any VM object. The caller may return 3596 * the page to the free list only if this function returns true. 3597 * 3598 * The object must be locked. The page must be locked if it is 3599 * managed. 3600 */ 3601 static bool 3602 vm_page_free_prep(vm_page_t m) 3603 { 3604 3605 /* 3606 * Synchronize with threads that have dropped a reference to this 3607 * page. 3608 */ 3609 atomic_thread_fence_acq(); 3610 3611 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3612 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3613 uint64_t *p; 3614 int i; 3615 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3616 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3617 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3618 m, i, (uintmax_t)*p)); 3619 } 3620 #endif 3621 if ((m->oflags & VPO_UNMANAGED) == 0) { 3622 KASSERT(!pmap_page_is_mapped(m), 3623 ("vm_page_free_prep: freeing mapped page %p", m)); 3624 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3625 ("vm_page_free_prep: mapping flags set in page %p", m)); 3626 } else { 3627 KASSERT(m->a.queue == PQ_NONE, 3628 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3629 } 3630 VM_CNT_INC(v_tfree); 3631 3632 if (m->object != NULL) { 3633 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 3634 ((m->object->flags & OBJ_UNMANAGED) != 0), 3635 ("vm_page_free_prep: managed flag mismatch for page %p", 3636 m)); 3637 vm_page_assert_xbusied(m); 3638 3639 /* 3640 * The object reference can be released without an atomic 3641 * operation. 3642 */ 3643 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3644 m->ref_count == VPRC_OBJREF, 3645 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3646 m, m->ref_count)); 3647 vm_page_object_remove(m); 3648 m->object = NULL; 3649 m->ref_count -= VPRC_OBJREF; 3650 } else 3651 vm_page_assert_unbusied(m); 3652 3653 vm_page_busy_free(m); 3654 3655 /* 3656 * If fictitious remove object association and 3657 * return. 3658 */ 3659 if ((m->flags & PG_FICTITIOUS) != 0) { 3660 KASSERT(m->ref_count == 1, 3661 ("fictitious page %p is referenced", m)); 3662 KASSERT(m->a.queue == PQ_NONE, 3663 ("fictitious page %p is queued", m)); 3664 return (false); 3665 } 3666 3667 /* 3668 * Pages need not be dequeued before they are returned to the physical 3669 * memory allocator, but they must at least be marked for a deferred 3670 * dequeue. 3671 */ 3672 if ((m->oflags & VPO_UNMANAGED) == 0) 3673 vm_page_dequeue_deferred(m); 3674 3675 m->valid = 0; 3676 vm_page_undirty(m); 3677 3678 if (m->ref_count != 0) 3679 panic("vm_page_free_prep: page %p has references", m); 3680 3681 /* 3682 * Restore the default memory attribute to the page. 3683 */ 3684 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3685 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3686 3687 #if VM_NRESERVLEVEL > 0 3688 /* 3689 * Determine whether the page belongs to a reservation. If the page was 3690 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3691 * as an optimization, we avoid the check in that case. 3692 */ 3693 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3694 return (false); 3695 #endif 3696 3697 return (true); 3698 } 3699 3700 /* 3701 * vm_page_free_toq: 3702 * 3703 * Returns the given page to the free list, disassociating it 3704 * from any VM object. 3705 * 3706 * The object must be locked. The page must be locked if it is 3707 * managed. 3708 */ 3709 static void 3710 vm_page_free_toq(vm_page_t m) 3711 { 3712 struct vm_domain *vmd; 3713 uma_zone_t zone; 3714 3715 if (!vm_page_free_prep(m)) 3716 return; 3717 3718 vmd = vm_pagequeue_domain(m); 3719 zone = vmd->vmd_pgcache[m->pool].zone; 3720 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 3721 uma_zfree(zone, m); 3722 return; 3723 } 3724 vm_domain_free_lock(vmd); 3725 vm_phys_free_pages(m, 0); 3726 vm_domain_free_unlock(vmd); 3727 vm_domain_freecnt_inc(vmd, 1); 3728 } 3729 3730 /* 3731 * vm_page_free_pages_toq: 3732 * 3733 * Returns a list of pages to the free list, disassociating it 3734 * from any VM object. In other words, this is equivalent to 3735 * calling vm_page_free_toq() for each page of a list of VM objects. 3736 * 3737 * The objects must be locked. The pages must be locked if it is 3738 * managed. 3739 */ 3740 void 3741 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3742 { 3743 vm_page_t m; 3744 int count; 3745 3746 if (SLIST_EMPTY(free)) 3747 return; 3748 3749 count = 0; 3750 while ((m = SLIST_FIRST(free)) != NULL) { 3751 count++; 3752 SLIST_REMOVE_HEAD(free, plinks.s.ss); 3753 vm_page_free_toq(m); 3754 } 3755 3756 if (update_wire_count) 3757 vm_wire_sub(count); 3758 } 3759 3760 /* 3761 * Mark this page as wired down, preventing reclamation by the page daemon 3762 * or when the containing object is destroyed. 3763 */ 3764 void 3765 vm_page_wire(vm_page_t m) 3766 { 3767 u_int old; 3768 3769 KASSERT(m->object != NULL, 3770 ("vm_page_wire: page %p does not belong to an object", m)); 3771 if (!vm_page_busied(m) && !vm_object_busied(m->object)) 3772 VM_OBJECT_ASSERT_LOCKED(m->object); 3773 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 3774 VPRC_WIRE_COUNT(m->ref_count) >= 1, 3775 ("vm_page_wire: fictitious page %p has zero wirings", m)); 3776 3777 old = atomic_fetchadd_int(&m->ref_count, 1); 3778 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 3779 ("vm_page_wire: counter overflow for page %p", m)); 3780 if (VPRC_WIRE_COUNT(old) == 0) { 3781 if ((m->oflags & VPO_UNMANAGED) == 0) 3782 vm_page_aflag_set(m, PGA_DEQUEUE); 3783 vm_wire_add(1); 3784 } 3785 } 3786 3787 /* 3788 * Attempt to wire a mapped page following a pmap lookup of that page. 3789 * This may fail if a thread is concurrently tearing down mappings of the page. 3790 * The transient failure is acceptable because it translates to the 3791 * failure of the caller pmap_extract_and_hold(), which should be then 3792 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 3793 */ 3794 bool 3795 vm_page_wire_mapped(vm_page_t m) 3796 { 3797 u_int old; 3798 3799 old = m->ref_count; 3800 do { 3801 KASSERT(old > 0, 3802 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 3803 if ((old & VPRC_BLOCKED) != 0) 3804 return (false); 3805 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 3806 3807 if (VPRC_WIRE_COUNT(old) == 0) { 3808 if ((m->oflags & VPO_UNMANAGED) == 0) 3809 vm_page_aflag_set(m, PGA_DEQUEUE); 3810 vm_wire_add(1); 3811 } 3812 return (true); 3813 } 3814 3815 /* 3816 * Release a wiring reference to a managed page. If the page still belongs to 3817 * an object, update its position in the page queues to reflect the reference. 3818 * If the wiring was the last reference to the page, free the page. 3819 */ 3820 static void 3821 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 3822 { 3823 u_int old; 3824 3825 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3826 ("%s: page %p is unmanaged", __func__, m)); 3827 3828 /* 3829 * Update LRU state before releasing the wiring reference. 3830 * Use a release store when updating the reference count to 3831 * synchronize with vm_page_free_prep(). 3832 */ 3833 old = m->ref_count; 3834 do { 3835 KASSERT(VPRC_WIRE_COUNT(old) > 0, 3836 ("vm_page_unwire: wire count underflow for page %p", m)); 3837 3838 if (old > VPRC_OBJREF + 1) { 3839 /* 3840 * The page has at least one other wiring reference. An 3841 * earlier iteration of this loop may have called 3842 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 3843 * re-set it if necessary. 3844 */ 3845 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 3846 vm_page_aflag_set(m, PGA_DEQUEUE); 3847 } else if (old == VPRC_OBJREF + 1) { 3848 /* 3849 * This is the last wiring. Clear PGA_DEQUEUE and 3850 * update the page's queue state to reflect the 3851 * reference. If the page does not belong to an object 3852 * (i.e., the VPRC_OBJREF bit is clear), we only need to 3853 * clear leftover queue state. 3854 */ 3855 vm_page_release_toq(m, nqueue, false); 3856 } else if (old == 1) { 3857 vm_page_aflag_clear(m, PGA_DEQUEUE); 3858 } 3859 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 3860 3861 if (VPRC_WIRE_COUNT(old) == 1) { 3862 vm_wire_sub(1); 3863 if (old == 1) 3864 vm_page_free(m); 3865 } 3866 } 3867 3868 /* 3869 * Release one wiring of the specified page, potentially allowing it to be 3870 * paged out. 3871 * 3872 * Only managed pages belonging to an object can be paged out. If the number 3873 * of wirings transitions to zero and the page is eligible for page out, then 3874 * the page is added to the specified paging queue. If the released wiring 3875 * represented the last reference to the page, the page is freed. 3876 * 3877 * A managed page must be locked. 3878 */ 3879 void 3880 vm_page_unwire(vm_page_t m, uint8_t nqueue) 3881 { 3882 3883 KASSERT(nqueue < PQ_COUNT, 3884 ("vm_page_unwire: invalid queue %u request for page %p", 3885 nqueue, m)); 3886 3887 if ((m->oflags & VPO_UNMANAGED) != 0) { 3888 if (vm_page_unwire_noq(m) && m->ref_count == 0) 3889 vm_page_free(m); 3890 return; 3891 } 3892 vm_page_unwire_managed(m, nqueue, false); 3893 } 3894 3895 /* 3896 * Unwire a page without (re-)inserting it into a page queue. It is up 3897 * to the caller to enqueue, requeue, or free the page as appropriate. 3898 * In most cases involving managed pages, vm_page_unwire() should be used 3899 * instead. 3900 */ 3901 bool 3902 vm_page_unwire_noq(vm_page_t m) 3903 { 3904 u_int old; 3905 3906 old = vm_page_drop(m, 1); 3907 KASSERT(VPRC_WIRE_COUNT(old) != 0, 3908 ("vm_page_unref: counter underflow for page %p", m)); 3909 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 3910 ("vm_page_unref: missing ref on fictitious page %p", m)); 3911 3912 if (VPRC_WIRE_COUNT(old) > 1) 3913 return (false); 3914 if ((m->oflags & VPO_UNMANAGED) == 0) 3915 vm_page_aflag_clear(m, PGA_DEQUEUE); 3916 vm_wire_sub(1); 3917 return (true); 3918 } 3919 3920 /* 3921 * Ensure that the page ends up in the specified page queue. If the page is 3922 * active or being moved to the active queue, ensure that its act_count is 3923 * at least ACT_INIT but do not otherwise mess with it. 3924 * 3925 * A managed page must be locked. 3926 */ 3927 static __always_inline void 3928 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 3929 { 3930 vm_page_astate_t old, new; 3931 3932 KASSERT(m->ref_count > 0, 3933 ("%s: page %p does not carry any references", __func__, m)); 3934 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 3935 ("%s: invalid flags %x", __func__, nflag)); 3936 3937 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3938 return; 3939 3940 old = vm_page_astate_load(m); 3941 do { 3942 if ((old.flags & PGA_DEQUEUE) != 0) 3943 break; 3944 new = old; 3945 new.flags &= ~PGA_QUEUE_OP_MASK; 3946 if (nqueue == PQ_ACTIVE) 3947 new.act_count = max(old.act_count, ACT_INIT); 3948 if (old.queue == nqueue) { 3949 if (nqueue != PQ_ACTIVE) 3950 new.flags |= nflag; 3951 } else { 3952 new.flags |= nflag; 3953 new.queue = nqueue; 3954 } 3955 } while (!vm_page_pqstate_commit(m, &old, new)); 3956 } 3957 3958 /* 3959 * Put the specified page on the active list (if appropriate). 3960 */ 3961 void 3962 vm_page_activate(vm_page_t m) 3963 { 3964 3965 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 3966 } 3967 3968 /* 3969 * Move the specified page to the tail of the inactive queue, or requeue 3970 * the page if it is already in the inactive queue. 3971 */ 3972 void 3973 vm_page_deactivate(vm_page_t m) 3974 { 3975 3976 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 3977 } 3978 3979 void 3980 vm_page_deactivate_noreuse(vm_page_t m) 3981 { 3982 3983 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 3984 } 3985 3986 /* 3987 * Put a page in the laundry, or requeue it if it is already there. 3988 */ 3989 void 3990 vm_page_launder(vm_page_t m) 3991 { 3992 3993 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 3994 } 3995 3996 /* 3997 * Put a page in the PQ_UNSWAPPABLE holding queue. 3998 */ 3999 void 4000 vm_page_unswappable(vm_page_t m) 4001 { 4002 4003 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 4004 ("page %p already unswappable", m)); 4005 4006 vm_page_dequeue(m); 4007 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4008 } 4009 4010 /* 4011 * Release a page back to the page queues in preparation for unwiring. 4012 */ 4013 static void 4014 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4015 { 4016 vm_page_astate_t old, new; 4017 uint16_t nflag; 4018 4019 /* 4020 * Use a check of the valid bits to determine whether we should 4021 * accelerate reclamation of the page. The object lock might not be 4022 * held here, in which case the check is racy. At worst we will either 4023 * accelerate reclamation of a valid page and violate LRU, or 4024 * unnecessarily defer reclamation of an invalid page. 4025 * 4026 * If we were asked to not cache the page, place it near the head of the 4027 * inactive queue so that is reclaimed sooner. 4028 */ 4029 if (noreuse || m->valid == 0) { 4030 nqueue = PQ_INACTIVE; 4031 nflag = PGA_REQUEUE_HEAD; 4032 } else { 4033 nflag = PGA_REQUEUE; 4034 } 4035 4036 old = vm_page_astate_load(m); 4037 do { 4038 new = old; 4039 4040 /* 4041 * If the page is already in the active queue and we are not 4042 * trying to accelerate reclamation, simply mark it as 4043 * referenced and avoid any queue operations. 4044 */ 4045 new.flags &= ~PGA_QUEUE_OP_MASK; 4046 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE) 4047 new.flags |= PGA_REFERENCED; 4048 else { 4049 new.flags |= nflag; 4050 new.queue = nqueue; 4051 } 4052 } while (!vm_page_pqstate_commit(m, &old, new)); 4053 } 4054 4055 /* 4056 * Unwire a page and either attempt to free it or re-add it to the page queues. 4057 */ 4058 void 4059 vm_page_release(vm_page_t m, int flags) 4060 { 4061 vm_object_t object; 4062 4063 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4064 ("vm_page_release: page %p is unmanaged", m)); 4065 4066 if ((flags & VPR_TRYFREE) != 0) { 4067 for (;;) { 4068 object = (vm_object_t)atomic_load_ptr(&m->object); 4069 if (object == NULL) 4070 break; 4071 /* Depends on type-stability. */ 4072 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4073 break; 4074 if (object == m->object) { 4075 vm_page_release_locked(m, flags); 4076 VM_OBJECT_WUNLOCK(object); 4077 return; 4078 } 4079 VM_OBJECT_WUNLOCK(object); 4080 } 4081 } 4082 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4083 } 4084 4085 /* See vm_page_release(). */ 4086 void 4087 vm_page_release_locked(vm_page_t m, int flags) 4088 { 4089 4090 VM_OBJECT_ASSERT_WLOCKED(m->object); 4091 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4092 ("vm_page_release_locked: page %p is unmanaged", m)); 4093 4094 if (vm_page_unwire_noq(m)) { 4095 if ((flags & VPR_TRYFREE) != 0 && 4096 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4097 m->dirty == 0 && vm_page_tryxbusy(m)) { 4098 vm_page_free(m); 4099 } else { 4100 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4101 } 4102 } 4103 } 4104 4105 static bool 4106 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4107 { 4108 u_int old; 4109 4110 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4111 ("vm_page_try_blocked_op: page %p has no object", m)); 4112 KASSERT(vm_page_busied(m), 4113 ("vm_page_try_blocked_op: page %p is not busy", m)); 4114 VM_OBJECT_ASSERT_LOCKED(m->object); 4115 4116 old = m->ref_count; 4117 do { 4118 KASSERT(old != 0, 4119 ("vm_page_try_blocked_op: page %p has no references", m)); 4120 if (VPRC_WIRE_COUNT(old) != 0) 4121 return (false); 4122 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4123 4124 (op)(m); 4125 4126 /* 4127 * If the object is read-locked, new wirings may be created via an 4128 * object lookup. 4129 */ 4130 old = vm_page_drop(m, VPRC_BLOCKED); 4131 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4132 old == (VPRC_BLOCKED | VPRC_OBJREF), 4133 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4134 old, m)); 4135 return (true); 4136 } 4137 4138 /* 4139 * Atomically check for wirings and remove all mappings of the page. 4140 */ 4141 bool 4142 vm_page_try_remove_all(vm_page_t m) 4143 { 4144 4145 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4146 } 4147 4148 /* 4149 * Atomically check for wirings and remove all writeable mappings of the page. 4150 */ 4151 bool 4152 vm_page_try_remove_write(vm_page_t m) 4153 { 4154 4155 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4156 } 4157 4158 /* 4159 * vm_page_advise 4160 * 4161 * Apply the specified advice to the given page. 4162 * 4163 * The object and page must be locked. 4164 */ 4165 void 4166 vm_page_advise(vm_page_t m, int advice) 4167 { 4168 4169 VM_OBJECT_ASSERT_WLOCKED(m->object); 4170 if (advice == MADV_FREE) 4171 /* 4172 * Mark the page clean. This will allow the page to be freed 4173 * without first paging it out. MADV_FREE pages are often 4174 * quickly reused by malloc(3), so we do not do anything that 4175 * would result in a page fault on a later access. 4176 */ 4177 vm_page_undirty(m); 4178 else if (advice != MADV_DONTNEED) { 4179 if (advice == MADV_WILLNEED) 4180 vm_page_activate(m); 4181 return; 4182 } 4183 4184 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4185 vm_page_dirty(m); 4186 4187 /* 4188 * Clear any references to the page. Otherwise, the page daemon will 4189 * immediately reactivate the page. 4190 */ 4191 vm_page_aflag_clear(m, PGA_REFERENCED); 4192 4193 /* 4194 * Place clean pages near the head of the inactive queue rather than 4195 * the tail, thus defeating the queue's LRU operation and ensuring that 4196 * the page will be reused quickly. Dirty pages not already in the 4197 * laundry are moved there. 4198 */ 4199 if (m->dirty == 0) 4200 vm_page_deactivate_noreuse(m); 4201 else if (!vm_page_in_laundry(m)) 4202 vm_page_launder(m); 4203 } 4204 4205 static inline int 4206 vm_page_grab_pflags(int allocflags) 4207 { 4208 int pflags; 4209 4210 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4211 (allocflags & VM_ALLOC_WIRED) != 0, 4212 ("vm_page_grab_pflags: the pages must be busied or wired")); 4213 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4214 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4215 ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 4216 "mismatch")); 4217 pflags = allocflags & 4218 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4219 VM_ALLOC_NOBUSY); 4220 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4221 pflags |= VM_ALLOC_WAITFAIL; 4222 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4223 pflags |= VM_ALLOC_SBUSY; 4224 4225 return (pflags); 4226 } 4227 4228 /* 4229 * Grab a page, waiting until we are waken up due to the page 4230 * changing state. We keep on waiting, if the page continues 4231 * to be in the object. If the page doesn't exist, first allocate it 4232 * and then conditionally zero it. 4233 * 4234 * This routine may sleep. 4235 * 4236 * The object must be locked on entry. The lock will, however, be released 4237 * and reacquired if the routine sleeps. 4238 */ 4239 vm_page_t 4240 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4241 { 4242 vm_page_t m; 4243 int pflags; 4244 4245 VM_OBJECT_ASSERT_WLOCKED(object); 4246 pflags = vm_page_grab_pflags(allocflags); 4247 retrylookup: 4248 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4249 if (!vm_page_acquire_flags(m, allocflags)) { 4250 if (vm_page_busy_sleep_flags(object, m, "pgrbwt", 4251 allocflags)) 4252 goto retrylookup; 4253 return (NULL); 4254 } 4255 goto out; 4256 } 4257 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4258 return (NULL); 4259 m = vm_page_alloc(object, pindex, pflags); 4260 if (m == NULL) { 4261 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4262 return (NULL); 4263 goto retrylookup; 4264 } 4265 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4266 pmap_zero_page(m); 4267 4268 out: 4269 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4270 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4271 vm_page_sunbusy(m); 4272 else 4273 vm_page_xunbusy(m); 4274 } 4275 return (m); 4276 } 4277 4278 /* 4279 * Grab a page and make it valid, paging in if necessary. Pages missing from 4280 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4281 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4282 * in simultaneously. Additional pages will be left on a paging queue but 4283 * will neither be wired nor busy regardless of allocflags. 4284 */ 4285 int 4286 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4287 { 4288 vm_page_t m; 4289 vm_page_t ma[VM_INITIAL_PAGEIN]; 4290 bool sleep, xbusy; 4291 int after, i, pflags, rv; 4292 4293 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4294 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4295 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4296 KASSERT((allocflags & 4297 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4298 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4299 VM_OBJECT_ASSERT_WLOCKED(object); 4300 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY); 4301 pflags |= VM_ALLOC_WAITFAIL; 4302 4303 retrylookup: 4304 xbusy = false; 4305 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4306 /* 4307 * If the page is fully valid it can only become invalid 4308 * with the object lock held. If it is not valid it can 4309 * become valid with the busy lock held. Therefore, we 4310 * may unnecessarily lock the exclusive busy here if we 4311 * race with I/O completion not using the object lock. 4312 * However, we will not end up with an invalid page and a 4313 * shared lock. 4314 */ 4315 if (!vm_page_all_valid(m) || 4316 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) { 4317 sleep = !vm_page_tryxbusy(m); 4318 xbusy = true; 4319 } else 4320 sleep = !vm_page_trysbusy(m); 4321 if (sleep) { 4322 (void)vm_page_busy_sleep_flags(object, m, "pgrbwt", 4323 allocflags); 4324 goto retrylookup; 4325 } 4326 if ((allocflags & VM_ALLOC_NOCREAT) != 0 && 4327 !vm_page_all_valid(m)) { 4328 if (xbusy) 4329 vm_page_xunbusy(m); 4330 else 4331 vm_page_sunbusy(m); 4332 *mp = NULL; 4333 return (VM_PAGER_FAIL); 4334 } 4335 if ((allocflags & VM_ALLOC_WIRED) != 0) 4336 vm_page_wire(m); 4337 if (vm_page_all_valid(m)) 4338 goto out; 4339 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4340 *mp = NULL; 4341 return (VM_PAGER_FAIL); 4342 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) { 4343 xbusy = true; 4344 } else { 4345 goto retrylookup; 4346 } 4347 4348 vm_page_assert_xbusied(m); 4349 MPASS(xbusy); 4350 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4351 after = MIN(after, VM_INITIAL_PAGEIN); 4352 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4353 after = MAX(after, 1); 4354 ma[0] = m; 4355 for (i = 1; i < after; i++) { 4356 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 4357 if (ma[i]->valid || !vm_page_tryxbusy(ma[i])) 4358 break; 4359 } else { 4360 ma[i] = vm_page_alloc(object, m->pindex + i, 4361 VM_ALLOC_NORMAL); 4362 if (ma[i] == NULL) 4363 break; 4364 } 4365 } 4366 after = i; 4367 vm_object_pip_add(object, after); 4368 VM_OBJECT_WUNLOCK(object); 4369 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 4370 VM_OBJECT_WLOCK(object); 4371 vm_object_pip_wakeupn(object, after); 4372 /* Pager may have replaced a page. */ 4373 m = ma[0]; 4374 if (rv != VM_PAGER_OK) { 4375 if ((allocflags & VM_ALLOC_WIRED) != 0) 4376 vm_page_unwire_noq(m); 4377 for (i = 0; i < after; i++) { 4378 if (!vm_page_wired(ma[i])) 4379 vm_page_free(ma[i]); 4380 else 4381 vm_page_xunbusy(ma[i]); 4382 } 4383 *mp = NULL; 4384 return (rv); 4385 } 4386 for (i = 1; i < after; i++) 4387 vm_page_readahead_finish(ma[i]); 4388 MPASS(vm_page_all_valid(m)); 4389 } else { 4390 vm_page_zero_invalid(m, TRUE); 4391 } 4392 out: 4393 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4394 if (xbusy) 4395 vm_page_xunbusy(m); 4396 else 4397 vm_page_sunbusy(m); 4398 } 4399 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy) 4400 vm_page_busy_downgrade(m); 4401 *mp = m; 4402 return (VM_PAGER_OK); 4403 } 4404 4405 /* 4406 * Return the specified range of pages from the given object. For each 4407 * page offset within the range, if a page already exists within the object 4408 * at that offset and it is busy, then wait for it to change state. If, 4409 * instead, the page doesn't exist, then allocate it. 4410 * 4411 * The caller must always specify an allocation class. 4412 * 4413 * allocation classes: 4414 * VM_ALLOC_NORMAL normal process request 4415 * VM_ALLOC_SYSTEM system *really* needs the pages 4416 * 4417 * The caller must always specify that the pages are to be busied and/or 4418 * wired. 4419 * 4420 * optional allocation flags: 4421 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4422 * VM_ALLOC_NOBUSY do not exclusive busy the page 4423 * VM_ALLOC_NOWAIT do not sleep 4424 * VM_ALLOC_SBUSY set page to sbusy state 4425 * VM_ALLOC_WIRED wire the pages 4426 * VM_ALLOC_ZERO zero and validate any invalid pages 4427 * 4428 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4429 * may return a partial prefix of the requested range. 4430 */ 4431 int 4432 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4433 vm_page_t *ma, int count) 4434 { 4435 vm_page_t m, mpred; 4436 int pflags; 4437 int i; 4438 4439 VM_OBJECT_ASSERT_WLOCKED(object); 4440 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4441 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4442 4443 pflags = vm_page_grab_pflags(allocflags); 4444 if (count == 0) 4445 return (0); 4446 4447 i = 0; 4448 retrylookup: 4449 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4450 if (m == NULL || m->pindex != pindex + i) { 4451 mpred = m; 4452 m = NULL; 4453 } else 4454 mpred = TAILQ_PREV(m, pglist, listq); 4455 for (; i < count; i++) { 4456 if (m != NULL) { 4457 if (!vm_page_acquire_flags(m, allocflags)) { 4458 if (vm_page_busy_sleep_flags(object, m, 4459 "grbmaw", allocflags)) 4460 goto retrylookup; 4461 break; 4462 } 4463 } else { 4464 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4465 break; 4466 m = vm_page_alloc_after(object, pindex + i, 4467 pflags | VM_ALLOC_COUNT(count - i), mpred); 4468 if (m == NULL) { 4469 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4470 break; 4471 goto retrylookup; 4472 } 4473 } 4474 if (vm_page_none_valid(m) && 4475 (allocflags & VM_ALLOC_ZERO) != 0) { 4476 if ((m->flags & PG_ZERO) == 0) 4477 pmap_zero_page(m); 4478 vm_page_valid(m); 4479 } 4480 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4481 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4482 vm_page_sunbusy(m); 4483 else 4484 vm_page_xunbusy(m); 4485 } 4486 ma[i] = mpred = m; 4487 m = vm_page_next(m); 4488 } 4489 return (i); 4490 } 4491 4492 /* 4493 * Mapping function for valid or dirty bits in a page. 4494 * 4495 * Inputs are required to range within a page. 4496 */ 4497 vm_page_bits_t 4498 vm_page_bits(int base, int size) 4499 { 4500 int first_bit; 4501 int last_bit; 4502 4503 KASSERT( 4504 base + size <= PAGE_SIZE, 4505 ("vm_page_bits: illegal base/size %d/%d", base, size) 4506 ); 4507 4508 if (size == 0) /* handle degenerate case */ 4509 return (0); 4510 4511 first_bit = base >> DEV_BSHIFT; 4512 last_bit = (base + size - 1) >> DEV_BSHIFT; 4513 4514 return (((vm_page_bits_t)2 << last_bit) - 4515 ((vm_page_bits_t)1 << first_bit)); 4516 } 4517 4518 void 4519 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 4520 { 4521 4522 #if PAGE_SIZE == 32768 4523 atomic_set_64((uint64_t *)bits, set); 4524 #elif PAGE_SIZE == 16384 4525 atomic_set_32((uint32_t *)bits, set); 4526 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 4527 atomic_set_16((uint16_t *)bits, set); 4528 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 4529 atomic_set_8((uint8_t *)bits, set); 4530 #else /* PAGE_SIZE <= 8192 */ 4531 uintptr_t addr; 4532 int shift; 4533 4534 addr = (uintptr_t)bits; 4535 /* 4536 * Use a trick to perform a 32-bit atomic on the 4537 * containing aligned word, to not depend on the existence 4538 * of atomic_{set, clear}_{8, 16}. 4539 */ 4540 shift = addr & (sizeof(uint32_t) - 1); 4541 #if BYTE_ORDER == BIG_ENDIAN 4542 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4543 #else 4544 shift *= NBBY; 4545 #endif 4546 addr &= ~(sizeof(uint32_t) - 1); 4547 atomic_set_32((uint32_t *)addr, set << shift); 4548 #endif /* PAGE_SIZE */ 4549 } 4550 4551 static inline void 4552 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 4553 { 4554 4555 #if PAGE_SIZE == 32768 4556 atomic_clear_64((uint64_t *)bits, clear); 4557 #elif PAGE_SIZE == 16384 4558 atomic_clear_32((uint32_t *)bits, clear); 4559 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 4560 atomic_clear_16((uint16_t *)bits, clear); 4561 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 4562 atomic_clear_8((uint8_t *)bits, clear); 4563 #else /* PAGE_SIZE <= 8192 */ 4564 uintptr_t addr; 4565 int shift; 4566 4567 addr = (uintptr_t)bits; 4568 /* 4569 * Use a trick to perform a 32-bit atomic on the 4570 * containing aligned word, to not depend on the existence 4571 * of atomic_{set, clear}_{8, 16}. 4572 */ 4573 shift = addr & (sizeof(uint32_t) - 1); 4574 #if BYTE_ORDER == BIG_ENDIAN 4575 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4576 #else 4577 shift *= NBBY; 4578 #endif 4579 addr &= ~(sizeof(uint32_t) - 1); 4580 atomic_clear_32((uint32_t *)addr, clear << shift); 4581 #endif /* PAGE_SIZE */ 4582 } 4583 4584 static inline vm_page_bits_t 4585 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 4586 { 4587 #if PAGE_SIZE == 32768 4588 uint64_t old; 4589 4590 old = *bits; 4591 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 4592 return (old); 4593 #elif PAGE_SIZE == 16384 4594 uint32_t old; 4595 4596 old = *bits; 4597 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 4598 return (old); 4599 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 4600 uint16_t old; 4601 4602 old = *bits; 4603 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 4604 return (old); 4605 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 4606 uint8_t old; 4607 4608 old = *bits; 4609 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 4610 return (old); 4611 #else /* PAGE_SIZE <= 4096*/ 4612 uintptr_t addr; 4613 uint32_t old, new, mask; 4614 int shift; 4615 4616 addr = (uintptr_t)bits; 4617 /* 4618 * Use a trick to perform a 32-bit atomic on the 4619 * containing aligned word, to not depend on the existence 4620 * of atomic_{set, swap, clear}_{8, 16}. 4621 */ 4622 shift = addr & (sizeof(uint32_t) - 1); 4623 #if BYTE_ORDER == BIG_ENDIAN 4624 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4625 #else 4626 shift *= NBBY; 4627 #endif 4628 addr &= ~(sizeof(uint32_t) - 1); 4629 mask = VM_PAGE_BITS_ALL << shift; 4630 4631 old = *bits; 4632 do { 4633 new = old & ~mask; 4634 new |= newbits << shift; 4635 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 4636 return (old >> shift); 4637 #endif /* PAGE_SIZE */ 4638 } 4639 4640 /* 4641 * vm_page_set_valid_range: 4642 * 4643 * Sets portions of a page valid. The arguments are expected 4644 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4645 * of any partial chunks touched by the range. The invalid portion of 4646 * such chunks will be zeroed. 4647 * 4648 * (base + size) must be less then or equal to PAGE_SIZE. 4649 */ 4650 void 4651 vm_page_set_valid_range(vm_page_t m, int base, int size) 4652 { 4653 int endoff, frag; 4654 vm_page_bits_t pagebits; 4655 4656 vm_page_assert_busied(m); 4657 if (size == 0) /* handle degenerate case */ 4658 return; 4659 4660 /* 4661 * If the base is not DEV_BSIZE aligned and the valid 4662 * bit is clear, we have to zero out a portion of the 4663 * first block. 4664 */ 4665 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4666 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 4667 pmap_zero_page_area(m, frag, base - frag); 4668 4669 /* 4670 * If the ending offset is not DEV_BSIZE aligned and the 4671 * valid bit is clear, we have to zero out a portion of 4672 * the last block. 4673 */ 4674 endoff = base + size; 4675 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4676 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 4677 pmap_zero_page_area(m, endoff, 4678 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4679 4680 /* 4681 * Assert that no previously invalid block that is now being validated 4682 * is already dirty. 4683 */ 4684 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 4685 ("vm_page_set_valid_range: page %p is dirty", m)); 4686 4687 /* 4688 * Set valid bits inclusive of any overlap. 4689 */ 4690 pagebits = vm_page_bits(base, size); 4691 if (vm_page_xbusied(m)) 4692 m->valid |= pagebits; 4693 else 4694 vm_page_bits_set(m, &m->valid, pagebits); 4695 } 4696 4697 /* 4698 * Set the page dirty bits and free the invalid swap space if 4699 * present. Returns the previous dirty bits. 4700 */ 4701 vm_page_bits_t 4702 vm_page_set_dirty(vm_page_t m) 4703 { 4704 vm_page_bits_t old; 4705 4706 VM_PAGE_OBJECT_BUSY_ASSERT(m); 4707 4708 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 4709 old = m->dirty; 4710 m->dirty = VM_PAGE_BITS_ALL; 4711 } else 4712 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 4713 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 4714 vm_pager_page_unswapped(m); 4715 4716 return (old); 4717 } 4718 4719 /* 4720 * Clear the given bits from the specified page's dirty field. 4721 */ 4722 static __inline void 4723 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 4724 { 4725 4726 vm_page_assert_busied(m); 4727 4728 /* 4729 * If the page is xbusied and not write mapped we are the 4730 * only thread that can modify dirty bits. Otherwise, The pmap 4731 * layer can call vm_page_dirty() without holding a distinguished 4732 * lock. The combination of page busy and atomic operations 4733 * suffice to guarantee consistency of the page dirty field. 4734 */ 4735 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 4736 m->dirty &= ~pagebits; 4737 else 4738 vm_page_bits_clear(m, &m->dirty, pagebits); 4739 } 4740 4741 /* 4742 * vm_page_set_validclean: 4743 * 4744 * Sets portions of a page valid and clean. The arguments are expected 4745 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4746 * of any partial chunks touched by the range. The invalid portion of 4747 * such chunks will be zero'd. 4748 * 4749 * (base + size) must be less then or equal to PAGE_SIZE. 4750 */ 4751 void 4752 vm_page_set_validclean(vm_page_t m, int base, int size) 4753 { 4754 vm_page_bits_t oldvalid, pagebits; 4755 int endoff, frag; 4756 4757 vm_page_assert_busied(m); 4758 if (size == 0) /* handle degenerate case */ 4759 return; 4760 4761 /* 4762 * If the base is not DEV_BSIZE aligned and the valid 4763 * bit is clear, we have to zero out a portion of the 4764 * first block. 4765 */ 4766 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4767 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 4768 pmap_zero_page_area(m, frag, base - frag); 4769 4770 /* 4771 * If the ending offset is not DEV_BSIZE aligned and the 4772 * valid bit is clear, we have to zero out a portion of 4773 * the last block. 4774 */ 4775 endoff = base + size; 4776 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4777 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 4778 pmap_zero_page_area(m, endoff, 4779 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4780 4781 /* 4782 * Set valid, clear dirty bits. If validating the entire 4783 * page we can safely clear the pmap modify bit. We also 4784 * use this opportunity to clear the PGA_NOSYNC flag. If a process 4785 * takes a write fault on a MAP_NOSYNC memory area the flag will 4786 * be set again. 4787 * 4788 * We set valid bits inclusive of any overlap, but we can only 4789 * clear dirty bits for DEV_BSIZE chunks that are fully within 4790 * the range. 4791 */ 4792 oldvalid = m->valid; 4793 pagebits = vm_page_bits(base, size); 4794 if (vm_page_xbusied(m)) 4795 m->valid |= pagebits; 4796 else 4797 vm_page_bits_set(m, &m->valid, pagebits); 4798 #if 0 /* NOT YET */ 4799 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 4800 frag = DEV_BSIZE - frag; 4801 base += frag; 4802 size -= frag; 4803 if (size < 0) 4804 size = 0; 4805 } 4806 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 4807 #endif 4808 if (base == 0 && size == PAGE_SIZE) { 4809 /* 4810 * The page can only be modified within the pmap if it is 4811 * mapped, and it can only be mapped if it was previously 4812 * fully valid. 4813 */ 4814 if (oldvalid == VM_PAGE_BITS_ALL) 4815 /* 4816 * Perform the pmap_clear_modify() first. Otherwise, 4817 * a concurrent pmap operation, such as 4818 * pmap_protect(), could clear a modification in the 4819 * pmap and set the dirty field on the page before 4820 * pmap_clear_modify() had begun and after the dirty 4821 * field was cleared here. 4822 */ 4823 pmap_clear_modify(m); 4824 m->dirty = 0; 4825 vm_page_aflag_clear(m, PGA_NOSYNC); 4826 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 4827 m->dirty &= ~pagebits; 4828 else 4829 vm_page_clear_dirty_mask(m, pagebits); 4830 } 4831 4832 void 4833 vm_page_clear_dirty(vm_page_t m, int base, int size) 4834 { 4835 4836 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 4837 } 4838 4839 /* 4840 * vm_page_set_invalid: 4841 * 4842 * Invalidates DEV_BSIZE'd chunks within a page. Both the 4843 * valid and dirty bits for the effected areas are cleared. 4844 */ 4845 void 4846 vm_page_set_invalid(vm_page_t m, int base, int size) 4847 { 4848 vm_page_bits_t bits; 4849 vm_object_t object; 4850 4851 /* 4852 * The object lock is required so that pages can't be mapped 4853 * read-only while we're in the process of invalidating them. 4854 */ 4855 object = m->object; 4856 VM_OBJECT_ASSERT_WLOCKED(object); 4857 vm_page_assert_busied(m); 4858 4859 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 4860 size >= object->un_pager.vnp.vnp_size) 4861 bits = VM_PAGE_BITS_ALL; 4862 else 4863 bits = vm_page_bits(base, size); 4864 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 4865 pmap_remove_all(m); 4866 KASSERT((bits == 0 && vm_page_all_valid(m)) || 4867 !pmap_page_is_mapped(m), 4868 ("vm_page_set_invalid: page %p is mapped", m)); 4869 if (vm_page_xbusied(m)) { 4870 m->valid &= ~bits; 4871 m->dirty &= ~bits; 4872 } else { 4873 vm_page_bits_clear(m, &m->valid, bits); 4874 vm_page_bits_clear(m, &m->dirty, bits); 4875 } 4876 } 4877 4878 /* 4879 * vm_page_invalid: 4880 * 4881 * Invalidates the entire page. The page must be busy, unmapped, and 4882 * the enclosing object must be locked. The object locks protects 4883 * against concurrent read-only pmap enter which is done without 4884 * busy. 4885 */ 4886 void 4887 vm_page_invalid(vm_page_t m) 4888 { 4889 4890 vm_page_assert_busied(m); 4891 VM_OBJECT_ASSERT_LOCKED(m->object); 4892 MPASS(!pmap_page_is_mapped(m)); 4893 4894 if (vm_page_xbusied(m)) 4895 m->valid = 0; 4896 else 4897 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 4898 } 4899 4900 /* 4901 * vm_page_zero_invalid() 4902 * 4903 * The kernel assumes that the invalid portions of a page contain 4904 * garbage, but such pages can be mapped into memory by user code. 4905 * When this occurs, we must zero out the non-valid portions of the 4906 * page so user code sees what it expects. 4907 * 4908 * Pages are most often semi-valid when the end of a file is mapped 4909 * into memory and the file's size is not page aligned. 4910 */ 4911 void 4912 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 4913 { 4914 int b; 4915 int i; 4916 4917 /* 4918 * Scan the valid bits looking for invalid sections that 4919 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 4920 * valid bit may be set ) have already been zeroed by 4921 * vm_page_set_validclean(). 4922 */ 4923 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 4924 if (i == (PAGE_SIZE / DEV_BSIZE) || 4925 (m->valid & ((vm_page_bits_t)1 << i))) { 4926 if (i > b) { 4927 pmap_zero_page_area(m, 4928 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 4929 } 4930 b = i + 1; 4931 } 4932 } 4933 4934 /* 4935 * setvalid is TRUE when we can safely set the zero'd areas 4936 * as being valid. We can do this if there are no cache consistancy 4937 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 4938 */ 4939 if (setvalid) 4940 vm_page_valid(m); 4941 } 4942 4943 /* 4944 * vm_page_is_valid: 4945 * 4946 * Is (partial) page valid? Note that the case where size == 0 4947 * will return FALSE in the degenerate case where the page is 4948 * entirely invalid, and TRUE otherwise. 4949 * 4950 * Some callers envoke this routine without the busy lock held and 4951 * handle races via higher level locks. Typical callers should 4952 * hold a busy lock to prevent invalidation. 4953 */ 4954 int 4955 vm_page_is_valid(vm_page_t m, int base, int size) 4956 { 4957 vm_page_bits_t bits; 4958 4959 bits = vm_page_bits(base, size); 4960 return (m->valid != 0 && (m->valid & bits) == bits); 4961 } 4962 4963 /* 4964 * Returns true if all of the specified predicates are true for the entire 4965 * (super)page and false otherwise. 4966 */ 4967 bool 4968 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 4969 { 4970 vm_object_t object; 4971 int i, npages; 4972 4973 object = m->object; 4974 if (skip_m != NULL && skip_m->object != object) 4975 return (false); 4976 VM_OBJECT_ASSERT_LOCKED(object); 4977 npages = atop(pagesizes[m->psind]); 4978 4979 /* 4980 * The physically contiguous pages that make up a superpage, i.e., a 4981 * page with a page size index ("psind") greater than zero, will 4982 * occupy adjacent entries in vm_page_array[]. 4983 */ 4984 for (i = 0; i < npages; i++) { 4985 /* Always test object consistency, including "skip_m". */ 4986 if (m[i].object != object) 4987 return (false); 4988 if (&m[i] == skip_m) 4989 continue; 4990 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 4991 return (false); 4992 if ((flags & PS_ALL_DIRTY) != 0) { 4993 /* 4994 * Calling vm_page_test_dirty() or pmap_is_modified() 4995 * might stop this case from spuriously returning 4996 * "false". However, that would require a write lock 4997 * on the object containing "m[i]". 4998 */ 4999 if (m[i].dirty != VM_PAGE_BITS_ALL) 5000 return (false); 5001 } 5002 if ((flags & PS_ALL_VALID) != 0 && 5003 m[i].valid != VM_PAGE_BITS_ALL) 5004 return (false); 5005 } 5006 return (true); 5007 } 5008 5009 /* 5010 * Set the page's dirty bits if the page is modified. 5011 */ 5012 void 5013 vm_page_test_dirty(vm_page_t m) 5014 { 5015 5016 vm_page_assert_busied(m); 5017 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5018 vm_page_dirty(m); 5019 } 5020 5021 void 5022 vm_page_valid(vm_page_t m) 5023 { 5024 5025 vm_page_assert_busied(m); 5026 if (vm_page_xbusied(m)) 5027 m->valid = VM_PAGE_BITS_ALL; 5028 else 5029 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5030 } 5031 5032 void 5033 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5034 { 5035 5036 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5037 } 5038 5039 void 5040 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5041 { 5042 5043 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5044 } 5045 5046 int 5047 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5048 { 5049 5050 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5051 } 5052 5053 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5054 void 5055 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5056 { 5057 5058 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5059 } 5060 5061 void 5062 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5063 { 5064 5065 mtx_assert_(vm_page_lockptr(m), a, file, line); 5066 } 5067 #endif 5068 5069 #ifdef INVARIANTS 5070 void 5071 vm_page_object_busy_assert(vm_page_t m) 5072 { 5073 5074 /* 5075 * Certain of the page's fields may only be modified by the 5076 * holder of a page or object busy. 5077 */ 5078 if (m->object != NULL && !vm_page_busied(m)) 5079 VM_OBJECT_ASSERT_BUSY(m->object); 5080 } 5081 5082 void 5083 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5084 { 5085 5086 if ((bits & PGA_WRITEABLE) == 0) 5087 return; 5088 5089 /* 5090 * The PGA_WRITEABLE flag can only be set if the page is 5091 * managed, is exclusively busied or the object is locked. 5092 * Currently, this flag is only set by pmap_enter(). 5093 */ 5094 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5095 ("PGA_WRITEABLE on unmanaged page")); 5096 if (!vm_page_xbusied(m)) 5097 VM_OBJECT_ASSERT_BUSY(m->object); 5098 } 5099 #endif 5100 5101 #include "opt_ddb.h" 5102 #ifdef DDB 5103 #include <sys/kernel.h> 5104 5105 #include <ddb/ddb.h> 5106 5107 DB_SHOW_COMMAND(page, vm_page_print_page_info) 5108 { 5109 5110 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5111 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5112 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5113 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5114 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5115 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5116 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5117 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5118 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5119 } 5120 5121 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 5122 { 5123 int dom; 5124 5125 db_printf("pq_free %d\n", vm_free_count()); 5126 for (dom = 0; dom < vm_ndomains; dom++) { 5127 db_printf( 5128 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5129 dom, 5130 vm_dom[dom].vmd_page_count, 5131 vm_dom[dom].vmd_free_count, 5132 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5133 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5134 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5135 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5136 } 5137 } 5138 5139 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5140 { 5141 vm_page_t m; 5142 boolean_t phys, virt; 5143 5144 if (!have_addr) { 5145 db_printf("show pginfo addr\n"); 5146 return; 5147 } 5148 5149 phys = strchr(modif, 'p') != NULL; 5150 virt = strchr(modif, 'v') != NULL; 5151 if (virt) 5152 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5153 else if (phys) 5154 m = PHYS_TO_VM_PAGE(addr); 5155 else 5156 m = (vm_page_t)addr; 5157 db_printf( 5158 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" 5159 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5160 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5161 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5162 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5163 } 5164 #endif /* DDB */ 5165