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