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 bool found; 331 332 m = vm_phys_paddr_to_vm_page(pa); 333 if (m == NULL) 334 return (true); /* page does not exist, no failure */ 335 336 vmd = vm_pagequeue_domain(m); 337 vm_domain_free_lock(vmd); 338 found = vm_phys_unfree_page(m); 339 vm_domain_free_unlock(vmd); 340 if (found) { 341 vm_domain_freecnt_inc(vmd, -1); 342 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 343 if (verbose) 344 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 345 } 346 return (found); 347 } 348 349 /* 350 * vm_page_blacklist_check: 351 * 352 * Iterate through the provided string of blacklist addresses, pulling 353 * each entry out of the physical allocator free list and putting it 354 * onto a list for reporting via the vm.page_blacklist sysctl. 355 */ 356 static void 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__) || \ 627 defined(__riscv) || defined(__powerpc64__) 628 /* 629 * Include the UMA bootstrap pages, witness pages and vm_page_dump 630 * in a crash dump. When pmap_map() uses the direct map, they are 631 * not automatically included. 632 */ 633 for (pa = new_end; pa < end; pa += PAGE_SIZE) 634 dump_add_page(pa); 635 #endif 636 phys_avail[biggestone + 1] = new_end; 637 #ifdef __amd64__ 638 /* 639 * Request that the physical pages underlying the message buffer be 640 * included in a crash dump. Since the message buffer is accessed 641 * through the direct map, they are not automatically included. 642 */ 643 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 644 last_pa = pa + round_page(msgbufsize); 645 while (pa < last_pa) { 646 dump_add_page(pa); 647 pa += PAGE_SIZE; 648 } 649 #endif 650 /* 651 * Compute the number of pages of memory that will be available for 652 * use, taking into account the overhead of a page structure per page. 653 * In other words, solve 654 * "available physical memory" - round_page(page_range * 655 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 656 * for page_range. 657 */ 658 low_avail = phys_avail[0]; 659 high_avail = phys_avail[1]; 660 for (i = 0; i < vm_phys_nsegs; i++) { 661 if (vm_phys_segs[i].start < low_avail) 662 low_avail = vm_phys_segs[i].start; 663 if (vm_phys_segs[i].end > high_avail) 664 high_avail = vm_phys_segs[i].end; 665 } 666 /* Skip the first chunk. It is already accounted for. */ 667 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 668 if (phys_avail[i] < low_avail) 669 low_avail = phys_avail[i]; 670 if (phys_avail[i + 1] > high_avail) 671 high_avail = phys_avail[i + 1]; 672 } 673 first_page = low_avail / PAGE_SIZE; 674 #ifdef VM_PHYSSEG_SPARSE 675 size = 0; 676 for (i = 0; i < vm_phys_nsegs; i++) 677 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 678 for (i = 0; phys_avail[i + 1] != 0; i += 2) 679 size += phys_avail[i + 1] - phys_avail[i]; 680 #elif defined(VM_PHYSSEG_DENSE) 681 size = high_avail - low_avail; 682 #else 683 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 684 #endif 685 686 #ifdef PMAP_HAS_PAGE_ARRAY 687 pmap_page_array_startup(size / PAGE_SIZE); 688 biggestone = vm_phys_avail_largest(); 689 end = new_end = phys_avail[biggestone + 1]; 690 #else 691 #ifdef VM_PHYSSEG_DENSE 692 /* 693 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 694 * the overhead of a page structure per page only if vm_page_array is 695 * allocated from the last physical memory chunk. Otherwise, we must 696 * allocate page structures representing the physical memory 697 * underlying vm_page_array, even though they will not be used. 698 */ 699 if (new_end != high_avail) 700 page_range = size / PAGE_SIZE; 701 else 702 #endif 703 { 704 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 705 706 /* 707 * If the partial bytes remaining are large enough for 708 * a page (PAGE_SIZE) without a corresponding 709 * 'struct vm_page', then new_end will contain an 710 * extra page after subtracting the length of the VM 711 * page array. Compensate by subtracting an extra 712 * page from new_end. 713 */ 714 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 715 if (new_end == high_avail) 716 high_avail -= PAGE_SIZE; 717 new_end -= PAGE_SIZE; 718 } 719 } 720 end = new_end; 721 new_end = vm_page_array_alloc(&vaddr, end, page_range); 722 #endif 723 724 #if VM_NRESERVLEVEL > 0 725 /* 726 * Allocate physical memory for the reservation management system's 727 * data structures, and map it. 728 */ 729 new_end = vm_reserv_startup(&vaddr, new_end); 730 #endif 731 #if defined(__aarch64__) || defined(__amd64__) || \ 732 defined(__riscv) || defined(__powerpc64__) 733 /* 734 * Include vm_page_array and vm_reserv_array in a crash dump. 735 */ 736 for (pa = new_end; pa < end; pa += PAGE_SIZE) 737 dump_add_page(pa); 738 #endif 739 phys_avail[biggestone + 1] = new_end; 740 741 /* 742 * Add physical memory segments corresponding to the available 743 * physical pages. 744 */ 745 for (i = 0; phys_avail[i + 1] != 0; i += 2) 746 if (vm_phys_avail_size(i) != 0) 747 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 748 749 /* 750 * Initialize the physical memory allocator. 751 */ 752 vm_phys_init(); 753 754 /* 755 * Initialize the page structures and add every available page to the 756 * physical memory allocator's free lists. 757 */ 758 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 759 for (ii = 0; ii < vm_page_array_size; ii++) { 760 m = &vm_page_array[ii]; 761 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 762 m->flags = PG_FICTITIOUS; 763 } 764 #endif 765 vm_cnt.v_page_count = 0; 766 for (segind = 0; segind < vm_phys_nsegs; segind++) { 767 seg = &vm_phys_segs[segind]; 768 for (m = seg->first_page, pa = seg->start; pa < seg->end; 769 m++, pa += PAGE_SIZE) 770 vm_page_init_page(m, pa, segind); 771 772 /* 773 * Add the segment's pages that are covered by one of 774 * phys_avail's ranges to the free lists. 775 */ 776 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 777 if (seg->end <= phys_avail[i] || 778 seg->start >= phys_avail[i + 1]) 779 continue; 780 781 startp = MAX(seg->start, phys_avail[i]); 782 endp = MIN(seg->end, phys_avail[i + 1]); 783 pagecount = (u_long)atop(endp - startp); 784 if (pagecount == 0) 785 continue; 786 787 m = seg->first_page + atop(startp - seg->start); 788 vmd = VM_DOMAIN(seg->domain); 789 vm_domain_free_lock(vmd); 790 vm_phys_enqueue_contig(m, pagecount); 791 vm_domain_free_unlock(vmd); 792 vm_domain_freecnt_inc(vmd, pagecount); 793 vm_cnt.v_page_count += (u_int)pagecount; 794 vmd->vmd_page_count += (u_int)pagecount; 795 vmd->vmd_segs |= 1UL << segind; 796 } 797 } 798 799 /* 800 * Remove blacklisted pages from the physical memory allocator. 801 */ 802 TAILQ_INIT(&blacklist_head); 803 vm_page_blacklist_load(&list, &listend); 804 vm_page_blacklist_check(list, listend); 805 806 list = kern_getenv("vm.blacklist"); 807 vm_page_blacklist_check(list, NULL); 808 809 freeenv(list); 810 #if VM_NRESERVLEVEL > 0 811 /* 812 * Initialize the reservation management system. 813 */ 814 vm_reserv_init(); 815 #endif 816 817 return (vaddr); 818 } 819 820 void 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 vm_pager_page_inserted(object, m); 1487 return (0); 1488 } 1489 1490 /* 1491 * vm_page_insert_radixdone: 1492 * 1493 * Complete page "m" insertion into the specified object after the 1494 * radix trie hooking. 1495 * 1496 * The page "mpred" must precede the offset "m->pindex" within the 1497 * specified object. 1498 * 1499 * The object must be locked. 1500 */ 1501 static void 1502 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1503 { 1504 1505 VM_OBJECT_ASSERT_WLOCKED(object); 1506 KASSERT(object != NULL && m->object == object, 1507 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1508 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1509 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1510 if (mpred != NULL) { 1511 KASSERT(mpred->object == object, 1512 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1513 KASSERT(mpred->pindex < m->pindex, 1514 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1515 } 1516 1517 if (mpred != NULL) 1518 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1519 else 1520 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1521 1522 /* 1523 * Show that the object has one more resident page. 1524 */ 1525 object->resident_page_count++; 1526 1527 /* 1528 * Hold the vnode until the last page is released. 1529 */ 1530 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1531 vhold(object->handle); 1532 1533 /* 1534 * Since we are inserting a new and possibly dirty page, 1535 * update the object's generation count. 1536 */ 1537 if (pmap_page_is_write_mapped(m)) 1538 vm_object_set_writeable_dirty(object); 1539 } 1540 1541 /* 1542 * Do the work to remove a page from its object. The caller is responsible for 1543 * updating the page's fields to reflect this removal. 1544 */ 1545 static void 1546 vm_page_object_remove(vm_page_t m) 1547 { 1548 vm_object_t object; 1549 vm_page_t mrem __diagused; 1550 1551 vm_page_assert_xbusied(m); 1552 object = m->object; 1553 VM_OBJECT_ASSERT_WLOCKED(object); 1554 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1555 ("page %p is missing its object ref", m)); 1556 1557 /* Deferred free of swap space. */ 1558 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1559 vm_pager_page_unswapped(m); 1560 1561 vm_pager_page_removed(object, m); 1562 1563 m->object = NULL; 1564 mrem = vm_radix_remove(&object->rtree, m->pindex); 1565 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1566 1567 /* 1568 * Now remove from the object's list of backed pages. 1569 */ 1570 TAILQ_REMOVE(&object->memq, m, listq); 1571 1572 /* 1573 * And show that the object has one fewer resident page. 1574 */ 1575 object->resident_page_count--; 1576 1577 /* 1578 * The vnode may now be recycled. 1579 */ 1580 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1581 vdrop(object->handle); 1582 } 1583 1584 /* 1585 * vm_page_remove: 1586 * 1587 * Removes the specified page from its containing object, but does not 1588 * invalidate any backing storage. Returns true if the object's reference 1589 * was the last reference to the page, and false otherwise. 1590 * 1591 * The object must be locked and the page must be exclusively busied. 1592 * The exclusive busy will be released on return. If this is not the 1593 * final ref and the caller does not hold a wire reference it may not 1594 * continue to access the page. 1595 */ 1596 bool 1597 vm_page_remove(vm_page_t m) 1598 { 1599 bool dropped; 1600 1601 dropped = vm_page_remove_xbusy(m); 1602 vm_page_xunbusy(m); 1603 1604 return (dropped); 1605 } 1606 1607 /* 1608 * vm_page_remove_xbusy 1609 * 1610 * Removes the page but leaves the xbusy held. Returns true if this 1611 * removed the final ref and false otherwise. 1612 */ 1613 bool 1614 vm_page_remove_xbusy(vm_page_t m) 1615 { 1616 1617 vm_page_object_remove(m); 1618 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1619 } 1620 1621 /* 1622 * vm_page_lookup: 1623 * 1624 * Returns the page associated with the object/offset 1625 * pair specified; if none is found, NULL is returned. 1626 * 1627 * The object must be locked. 1628 */ 1629 vm_page_t 1630 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1631 { 1632 1633 VM_OBJECT_ASSERT_LOCKED(object); 1634 return (vm_radix_lookup(&object->rtree, pindex)); 1635 } 1636 1637 /* 1638 * vm_page_lookup_unlocked: 1639 * 1640 * Returns the page associated with the object/offset pair specified; 1641 * if none is found, NULL is returned. The page may be no longer be 1642 * present in the object at the time that this function returns. Only 1643 * useful for opportunistic checks such as inmem(). 1644 */ 1645 vm_page_t 1646 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1647 { 1648 1649 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1650 } 1651 1652 /* 1653 * vm_page_relookup: 1654 * 1655 * Returns a page that must already have been busied by 1656 * the caller. Used for bogus page replacement. 1657 */ 1658 vm_page_t 1659 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1660 { 1661 vm_page_t m; 1662 1663 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 1664 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1665 m->object == object && m->pindex == pindex, 1666 ("vm_page_relookup: Invalid page %p", m)); 1667 return (m); 1668 } 1669 1670 /* 1671 * This should only be used by lockless functions for releasing transient 1672 * incorrect acquires. The page may have been freed after we acquired a 1673 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1674 * further to do. 1675 */ 1676 static void 1677 vm_page_busy_release(vm_page_t m) 1678 { 1679 u_int x; 1680 1681 x = vm_page_busy_fetch(m); 1682 for (;;) { 1683 if (x == VPB_FREED) 1684 break; 1685 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1686 if (atomic_fcmpset_int(&m->busy_lock, &x, 1687 x - VPB_ONE_SHARER)) 1688 break; 1689 continue; 1690 } 1691 KASSERT((x & VPB_BIT_SHARED) != 0 || 1692 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1693 ("vm_page_busy_release: %p xbusy not owned.", m)); 1694 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1695 continue; 1696 if ((x & VPB_BIT_WAITERS) != 0) 1697 wakeup(m); 1698 break; 1699 } 1700 } 1701 1702 /* 1703 * vm_page_find_least: 1704 * 1705 * Returns the page associated with the object with least pindex 1706 * greater than or equal to the parameter pindex, or NULL. 1707 * 1708 * The object must be locked. 1709 */ 1710 vm_page_t 1711 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1712 { 1713 vm_page_t m; 1714 1715 VM_OBJECT_ASSERT_LOCKED(object); 1716 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1717 m = vm_radix_lookup_ge(&object->rtree, pindex); 1718 return (m); 1719 } 1720 1721 /* 1722 * Returns the given page's successor (by pindex) within the object if it is 1723 * resident; if none is found, NULL is returned. 1724 * 1725 * The object must be locked. 1726 */ 1727 vm_page_t 1728 vm_page_next(vm_page_t m) 1729 { 1730 vm_page_t next; 1731 1732 VM_OBJECT_ASSERT_LOCKED(m->object); 1733 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1734 MPASS(next->object == m->object); 1735 if (next->pindex != m->pindex + 1) 1736 next = NULL; 1737 } 1738 return (next); 1739 } 1740 1741 /* 1742 * Returns the given page's predecessor (by pindex) within the object if it is 1743 * resident; if none is found, NULL is returned. 1744 * 1745 * The object must be locked. 1746 */ 1747 vm_page_t 1748 vm_page_prev(vm_page_t m) 1749 { 1750 vm_page_t prev; 1751 1752 VM_OBJECT_ASSERT_LOCKED(m->object); 1753 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1754 MPASS(prev->object == m->object); 1755 if (prev->pindex != m->pindex - 1) 1756 prev = NULL; 1757 } 1758 return (prev); 1759 } 1760 1761 /* 1762 * Uses the page mnew as a replacement for an existing page at index 1763 * pindex which must be already present in the object. 1764 * 1765 * Both pages must be exclusively busied on enter. The old page is 1766 * unbusied on exit. 1767 * 1768 * A return value of true means mold is now free. If this is not the 1769 * final ref and the caller does not hold a wire reference it may not 1770 * continue to access the page. 1771 */ 1772 static bool 1773 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1774 vm_page_t mold) 1775 { 1776 vm_page_t mret __diagused; 1777 bool dropped; 1778 1779 VM_OBJECT_ASSERT_WLOCKED(object); 1780 vm_page_assert_xbusied(mold); 1781 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1782 ("vm_page_replace: page %p already in object", mnew)); 1783 1784 /* 1785 * This function mostly follows vm_page_insert() and 1786 * vm_page_remove() without the radix, object count and vnode 1787 * dance. Double check such functions for more comments. 1788 */ 1789 1790 mnew->object = object; 1791 mnew->pindex = pindex; 1792 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1793 mret = vm_radix_replace(&object->rtree, mnew); 1794 KASSERT(mret == mold, 1795 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1796 KASSERT((mold->oflags & VPO_UNMANAGED) == 1797 (mnew->oflags & VPO_UNMANAGED), 1798 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1799 1800 /* Keep the resident page list in sorted order. */ 1801 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1802 TAILQ_REMOVE(&object->memq, mold, listq); 1803 mold->object = NULL; 1804 1805 /* 1806 * The object's resident_page_count does not change because we have 1807 * swapped one page for another, but the generation count should 1808 * change if the page is dirty. 1809 */ 1810 if (pmap_page_is_write_mapped(mnew)) 1811 vm_object_set_writeable_dirty(object); 1812 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1813 vm_page_xunbusy(mold); 1814 1815 return (dropped); 1816 } 1817 1818 void 1819 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1820 vm_page_t mold) 1821 { 1822 1823 vm_page_assert_xbusied(mnew); 1824 1825 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1826 vm_page_free(mold); 1827 } 1828 1829 /* 1830 * vm_page_rename: 1831 * 1832 * Move the given memory entry from its 1833 * current object to the specified target object/offset. 1834 * 1835 * Note: swap associated with the page must be invalidated by the move. We 1836 * have to do this for several reasons: (1) we aren't freeing the 1837 * page, (2) we are dirtying the page, (3) the VM system is probably 1838 * moving the page from object A to B, and will then later move 1839 * the backing store from A to B and we can't have a conflict. 1840 * 1841 * Note: we *always* dirty the page. It is necessary both for the 1842 * fact that we moved it, and because we may be invalidating 1843 * swap. 1844 * 1845 * The objects must be locked. 1846 */ 1847 int 1848 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1849 { 1850 vm_page_t mpred; 1851 vm_pindex_t opidx; 1852 1853 VM_OBJECT_ASSERT_WLOCKED(new_object); 1854 1855 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1856 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1857 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1858 ("vm_page_rename: pindex already renamed")); 1859 1860 /* 1861 * Create a custom version of vm_page_insert() which does not depend 1862 * by m_prev and can cheat on the implementation aspects of the 1863 * function. 1864 */ 1865 opidx = m->pindex; 1866 m->pindex = new_pindex; 1867 if (vm_radix_insert(&new_object->rtree, m)) { 1868 m->pindex = opidx; 1869 return (1); 1870 } 1871 1872 /* 1873 * The operation cannot fail anymore. The removal must happen before 1874 * the listq iterator is tainted. 1875 */ 1876 m->pindex = opidx; 1877 vm_page_object_remove(m); 1878 1879 /* Return back to the new pindex to complete vm_page_insert(). */ 1880 m->pindex = new_pindex; 1881 m->object = new_object; 1882 1883 vm_page_insert_radixdone(m, new_object, mpred); 1884 vm_page_dirty(m); 1885 vm_pager_page_inserted(new_object, m); 1886 return (0); 1887 } 1888 1889 /* 1890 * vm_page_alloc: 1891 * 1892 * Allocate and return a page that is associated with the specified 1893 * object and offset pair. By default, this page is exclusive busied. 1894 * 1895 * The caller must always specify an allocation class. 1896 * 1897 * allocation classes: 1898 * VM_ALLOC_NORMAL normal process request 1899 * VM_ALLOC_SYSTEM system *really* needs a page 1900 * VM_ALLOC_INTERRUPT interrupt time request 1901 * 1902 * optional allocation flags: 1903 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1904 * intends to allocate 1905 * VM_ALLOC_NOBUSY do not exclusive busy the page 1906 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1907 * VM_ALLOC_SBUSY shared busy the allocated page 1908 * VM_ALLOC_WIRED wire the allocated page 1909 * VM_ALLOC_ZERO prefer a zeroed page 1910 */ 1911 vm_page_t 1912 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1913 { 1914 1915 return (vm_page_alloc_after(object, pindex, req, 1916 vm_radix_lookup_le(&object->rtree, pindex))); 1917 } 1918 1919 vm_page_t 1920 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1921 int req) 1922 { 1923 1924 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1925 vm_radix_lookup_le(&object->rtree, pindex))); 1926 } 1927 1928 /* 1929 * Allocate a page in the specified object with the given page index. To 1930 * optimize insertion of the page into the object, the caller must also specifiy 1931 * the resident page in the object with largest index smaller than the given 1932 * page index, or NULL if no such page exists. 1933 */ 1934 vm_page_t 1935 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1936 int req, vm_page_t mpred) 1937 { 1938 struct vm_domainset_iter di; 1939 vm_page_t m; 1940 int domain; 1941 1942 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1943 do { 1944 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1945 mpred); 1946 if (m != NULL) 1947 break; 1948 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1949 1950 return (m); 1951 } 1952 1953 /* 1954 * Returns true if the number of free pages exceeds the minimum 1955 * for the request class and false otherwise. 1956 */ 1957 static int 1958 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 1959 { 1960 u_int limit, old, new; 1961 1962 if (req_class == VM_ALLOC_INTERRUPT) 1963 limit = 0; 1964 else if (req_class == VM_ALLOC_SYSTEM) 1965 limit = vmd->vmd_interrupt_free_min; 1966 else 1967 limit = vmd->vmd_free_reserved; 1968 1969 /* 1970 * Attempt to reserve the pages. Fail if we're below the limit. 1971 */ 1972 limit += npages; 1973 old = vmd->vmd_free_count; 1974 do { 1975 if (old < limit) 1976 return (0); 1977 new = old - npages; 1978 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 1979 1980 /* Wake the page daemon if we've crossed the threshold. */ 1981 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 1982 pagedaemon_wakeup(vmd->vmd_domain); 1983 1984 /* Only update bitsets on transitions. */ 1985 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 1986 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 1987 vm_domain_set(vmd); 1988 1989 return (1); 1990 } 1991 1992 int 1993 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 1994 { 1995 int req_class; 1996 1997 /* 1998 * The page daemon is allowed to dig deeper into the free page list. 1999 */ 2000 req_class = req & VM_ALLOC_CLASS_MASK; 2001 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2002 req_class = VM_ALLOC_SYSTEM; 2003 return (_vm_domain_allocate(vmd, req_class, npages)); 2004 } 2005 2006 vm_page_t 2007 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 2008 int req, vm_page_t mpred) 2009 { 2010 struct vm_domain *vmd; 2011 vm_page_t m; 2012 int flags; 2013 2014 #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2015 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ 2016 VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ 2017 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) 2018 KASSERT((req & ~VPA_FLAGS) == 0, 2019 ("invalid request %#x", req)); 2020 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2021 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2022 ("invalid request %#x", req)); 2023 KASSERT(mpred == NULL || mpred->pindex < pindex, 2024 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2025 (uintmax_t)pindex)); 2026 VM_OBJECT_ASSERT_WLOCKED(object); 2027 2028 flags = 0; 2029 m = NULL; 2030 if (!vm_pager_can_alloc_page(object, pindex)) 2031 return (NULL); 2032 again: 2033 #if VM_NRESERVLEVEL > 0 2034 /* 2035 * Can we allocate the page from a reservation? 2036 */ 2037 if (vm_object_reserv(object) && 2038 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 2039 NULL) { 2040 goto found; 2041 } 2042 #endif 2043 vmd = VM_DOMAIN(domain); 2044 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2045 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2046 M_NOWAIT | M_NOVM); 2047 if (m != NULL) { 2048 flags |= PG_PCPU_CACHE; 2049 goto found; 2050 } 2051 } 2052 if (vm_domain_allocate(vmd, req, 1)) { 2053 /* 2054 * If not, allocate it from the free page queues. 2055 */ 2056 vm_domain_free_lock(vmd); 2057 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2058 vm_domain_free_unlock(vmd); 2059 if (m == NULL) { 2060 vm_domain_freecnt_inc(vmd, 1); 2061 #if VM_NRESERVLEVEL > 0 2062 if (vm_reserv_reclaim_inactive(domain)) 2063 goto again; 2064 #endif 2065 } 2066 } 2067 if (m == NULL) { 2068 /* 2069 * Not allocatable, give up. 2070 */ 2071 if (vm_domain_alloc_fail(vmd, object, req)) 2072 goto again; 2073 return (NULL); 2074 } 2075 2076 /* 2077 * At this point we had better have found a good page. 2078 */ 2079 found: 2080 vm_page_dequeue(m); 2081 vm_page_alloc_check(m); 2082 2083 /* 2084 * Initialize the page. Only the PG_ZERO flag is inherited. 2085 */ 2086 flags |= m->flags & PG_ZERO; 2087 if ((req & VM_ALLOC_NODUMP) != 0) 2088 flags |= PG_NODUMP; 2089 m->flags = flags; 2090 m->a.flags = 0; 2091 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2092 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2093 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2094 else if ((req & VM_ALLOC_SBUSY) != 0) 2095 m->busy_lock = VPB_SHARERS_WORD(1); 2096 else 2097 m->busy_lock = VPB_UNBUSIED; 2098 if (req & VM_ALLOC_WIRED) { 2099 vm_wire_add(1); 2100 m->ref_count = 1; 2101 } 2102 m->a.act_count = 0; 2103 2104 if (vm_page_insert_after(m, object, pindex, mpred)) { 2105 if (req & VM_ALLOC_WIRED) { 2106 vm_wire_sub(1); 2107 m->ref_count = 0; 2108 } 2109 KASSERT(m->object == NULL, ("page %p has object", m)); 2110 m->oflags = VPO_UNMANAGED; 2111 m->busy_lock = VPB_UNBUSIED; 2112 /* Don't change PG_ZERO. */ 2113 vm_page_free_toq(m); 2114 if (req & VM_ALLOC_WAITFAIL) { 2115 VM_OBJECT_WUNLOCK(object); 2116 vm_radix_wait(); 2117 VM_OBJECT_WLOCK(object); 2118 } 2119 return (NULL); 2120 } 2121 2122 /* Ignore device objects; the pager sets "memattr" for them. */ 2123 if (object->memattr != VM_MEMATTR_DEFAULT && 2124 (object->flags & OBJ_FICTITIOUS) == 0) 2125 pmap_page_set_memattr(m, object->memattr); 2126 2127 return (m); 2128 } 2129 2130 /* 2131 * vm_page_alloc_contig: 2132 * 2133 * Allocate a contiguous set of physical pages of the given size "npages" 2134 * from the free lists. All of the physical pages must be at or above 2135 * the given physical address "low" and below the given physical address 2136 * "high". The given value "alignment" determines the alignment of the 2137 * first physical page in the set. If the given value "boundary" is 2138 * non-zero, then the set of physical pages cannot cross any physical 2139 * address boundary that is a multiple of that value. Both "alignment" 2140 * and "boundary" must be a power of two. 2141 * 2142 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2143 * then the memory attribute setting for the physical pages is configured 2144 * to the object's memory attribute setting. Otherwise, the memory 2145 * attribute setting for the physical pages is configured to "memattr", 2146 * overriding the object's memory attribute setting. However, if the 2147 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2148 * memory attribute setting for the physical pages cannot be configured 2149 * to VM_MEMATTR_DEFAULT. 2150 * 2151 * The specified object may not contain fictitious pages. 2152 * 2153 * The caller must always specify an allocation class. 2154 * 2155 * allocation classes: 2156 * VM_ALLOC_NORMAL normal process request 2157 * VM_ALLOC_SYSTEM system *really* needs a page 2158 * VM_ALLOC_INTERRUPT interrupt time request 2159 * 2160 * optional allocation flags: 2161 * VM_ALLOC_NOBUSY do not exclusive busy the page 2162 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2163 * VM_ALLOC_SBUSY shared busy the allocated page 2164 * VM_ALLOC_WIRED wire the allocated page 2165 * VM_ALLOC_ZERO prefer a zeroed page 2166 */ 2167 vm_page_t 2168 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2169 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2170 vm_paddr_t boundary, vm_memattr_t memattr) 2171 { 2172 struct vm_domainset_iter di; 2173 vm_page_t m; 2174 int domain; 2175 2176 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2177 do { 2178 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2179 npages, low, high, alignment, boundary, memattr); 2180 if (m != NULL) 2181 break; 2182 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2183 2184 return (m); 2185 } 2186 2187 static vm_page_t 2188 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2189 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2190 { 2191 struct vm_domain *vmd; 2192 vm_page_t m_ret; 2193 2194 /* 2195 * Can we allocate the pages without the number of free pages falling 2196 * below the lower bound for the allocation class? 2197 */ 2198 vmd = VM_DOMAIN(domain); 2199 if (!vm_domain_allocate(vmd, req, npages)) 2200 return (NULL); 2201 /* 2202 * Try to allocate the pages from the free page queues. 2203 */ 2204 vm_domain_free_lock(vmd); 2205 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2206 alignment, boundary); 2207 vm_domain_free_unlock(vmd); 2208 if (m_ret != NULL) 2209 return (m_ret); 2210 #if VM_NRESERVLEVEL > 0 2211 /* 2212 * Try to break a reservation to allocate the pages. 2213 */ 2214 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2215 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2216 high, alignment, boundary); 2217 if (m_ret != NULL) 2218 return (m_ret); 2219 } 2220 #endif 2221 vm_domain_freecnt_inc(vmd, npages); 2222 return (NULL); 2223 } 2224 2225 vm_page_t 2226 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2227 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2228 vm_paddr_t boundary, vm_memattr_t memattr) 2229 { 2230 vm_page_t m, m_ret, mpred; 2231 u_int busy_lock, flags, oflags; 2232 2233 #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) 2234 KASSERT((req & ~VPAC_FLAGS) == 0, 2235 ("invalid request %#x", req)); 2236 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2237 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2238 ("invalid request %#x", req)); 2239 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2240 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2241 ("invalid request %#x", req)); 2242 VM_OBJECT_ASSERT_WLOCKED(object); 2243 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2244 ("vm_page_alloc_contig: object %p has fictitious pages", 2245 object)); 2246 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2247 2248 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2249 KASSERT(mpred == NULL || mpred->pindex != pindex, 2250 ("vm_page_alloc_contig: pindex already allocated")); 2251 for (;;) { 2252 #if VM_NRESERVLEVEL > 0 2253 /* 2254 * Can we allocate the pages from a reservation? 2255 */ 2256 if (vm_object_reserv(object) && 2257 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2258 mpred, npages, low, high, alignment, boundary)) != NULL) { 2259 break; 2260 } 2261 #endif 2262 if ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2263 low, high, alignment, boundary)) != NULL) 2264 break; 2265 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req)) 2266 return (NULL); 2267 } 2268 for (m = m_ret; m < &m_ret[npages]; m++) { 2269 vm_page_dequeue(m); 2270 vm_page_alloc_check(m); 2271 } 2272 2273 /* 2274 * Initialize the pages. Only the PG_ZERO flag is inherited. 2275 */ 2276 flags = PG_ZERO; 2277 if ((req & VM_ALLOC_NODUMP) != 0) 2278 flags |= PG_NODUMP; 2279 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2280 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2281 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2282 else if ((req & VM_ALLOC_SBUSY) != 0) 2283 busy_lock = VPB_SHARERS_WORD(1); 2284 else 2285 busy_lock = VPB_UNBUSIED; 2286 if ((req & VM_ALLOC_WIRED) != 0) 2287 vm_wire_add(npages); 2288 if (object->memattr != VM_MEMATTR_DEFAULT && 2289 memattr == VM_MEMATTR_DEFAULT) 2290 memattr = object->memattr; 2291 for (m = m_ret; m < &m_ret[npages]; m++) { 2292 m->a.flags = 0; 2293 m->flags = (m->flags | PG_NODUMP) & flags; 2294 m->busy_lock = busy_lock; 2295 if ((req & VM_ALLOC_WIRED) != 0) 2296 m->ref_count = 1; 2297 m->a.act_count = 0; 2298 m->oflags = oflags; 2299 if (vm_page_insert_after(m, object, pindex, mpred)) { 2300 if ((req & VM_ALLOC_WIRED) != 0) 2301 vm_wire_sub(npages); 2302 KASSERT(m->object == NULL, 2303 ("page %p has object", m)); 2304 mpred = m; 2305 for (m = m_ret; m < &m_ret[npages]; m++) { 2306 if (m <= mpred && 2307 (req & VM_ALLOC_WIRED) != 0) 2308 m->ref_count = 0; 2309 m->oflags = VPO_UNMANAGED; 2310 m->busy_lock = VPB_UNBUSIED; 2311 /* Don't change PG_ZERO. */ 2312 vm_page_free_toq(m); 2313 } 2314 if (req & VM_ALLOC_WAITFAIL) { 2315 VM_OBJECT_WUNLOCK(object); 2316 vm_radix_wait(); 2317 VM_OBJECT_WLOCK(object); 2318 } 2319 return (NULL); 2320 } 2321 mpred = m; 2322 if (memattr != VM_MEMATTR_DEFAULT) 2323 pmap_page_set_memattr(m, memattr); 2324 pindex++; 2325 } 2326 return (m_ret); 2327 } 2328 2329 /* 2330 * Allocate a physical page that is not intended to be inserted into a VM 2331 * object. If the "freelist" parameter is not equal to VM_NFREELIST, then only 2332 * pages from the specified vm_phys freelist will be returned. 2333 */ 2334 static __always_inline vm_page_t 2335 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req) 2336 { 2337 struct vm_domain *vmd; 2338 vm_page_t m; 2339 int flags; 2340 2341 #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2342 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ 2343 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ 2344 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) 2345 KASSERT((req & ~VPAN_FLAGS) == 0, 2346 ("invalid request %#x", req)); 2347 2348 flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0; 2349 vmd = VM_DOMAIN(domain); 2350 again: 2351 if (freelist == VM_NFREELIST && 2352 vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2353 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2354 M_NOWAIT | M_NOVM); 2355 if (m != NULL) { 2356 flags |= PG_PCPU_CACHE; 2357 goto found; 2358 } 2359 } 2360 2361 if (vm_domain_allocate(vmd, req, 1)) { 2362 vm_domain_free_lock(vmd); 2363 if (freelist == VM_NFREELIST) 2364 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2365 else 2366 m = vm_phys_alloc_freelist_pages(domain, freelist, 2367 VM_FREEPOOL_DIRECT, 0); 2368 vm_domain_free_unlock(vmd); 2369 if (m == NULL) { 2370 vm_domain_freecnt_inc(vmd, 1); 2371 #if VM_NRESERVLEVEL > 0 2372 if (freelist == VM_NFREELIST && 2373 vm_reserv_reclaim_inactive(domain)) 2374 goto again; 2375 #endif 2376 } 2377 } 2378 if (m == NULL) { 2379 if (vm_domain_alloc_fail(vmd, NULL, req)) 2380 goto again; 2381 return (NULL); 2382 } 2383 2384 found: 2385 vm_page_dequeue(m); 2386 vm_page_alloc_check(m); 2387 2388 /* 2389 * Consumers should not rely on a useful default pindex value. 2390 */ 2391 m->pindex = 0xdeadc0dedeadc0de; 2392 m->flags = (m->flags & PG_ZERO) | flags; 2393 m->a.flags = 0; 2394 m->oflags = VPO_UNMANAGED; 2395 m->busy_lock = VPB_UNBUSIED; 2396 if ((req & VM_ALLOC_WIRED) != 0) { 2397 vm_wire_add(1); 2398 m->ref_count = 1; 2399 } 2400 2401 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2402 pmap_zero_page(m); 2403 2404 return (m); 2405 } 2406 2407 vm_page_t 2408 vm_page_alloc_freelist(int freelist, int req) 2409 { 2410 struct vm_domainset_iter di; 2411 vm_page_t m; 2412 int domain; 2413 2414 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2415 do { 2416 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2417 if (m != NULL) 2418 break; 2419 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2420 2421 return (m); 2422 } 2423 2424 vm_page_t 2425 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2426 { 2427 KASSERT(freelist >= 0 && freelist < VM_NFREELIST, 2428 ("%s: invalid freelist %d", __func__, freelist)); 2429 2430 return (_vm_page_alloc_noobj_domain(domain, freelist, req)); 2431 } 2432 2433 vm_page_t 2434 vm_page_alloc_noobj(int req) 2435 { 2436 struct vm_domainset_iter di; 2437 vm_page_t m; 2438 int domain; 2439 2440 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2441 do { 2442 m = vm_page_alloc_noobj_domain(domain, req); 2443 if (m != NULL) 2444 break; 2445 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2446 2447 return (m); 2448 } 2449 2450 vm_page_t 2451 vm_page_alloc_noobj_domain(int domain, int req) 2452 { 2453 return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req)); 2454 } 2455 2456 vm_page_t 2457 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2458 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2459 vm_memattr_t memattr) 2460 { 2461 struct vm_domainset_iter di; 2462 vm_page_t m; 2463 int domain; 2464 2465 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2466 do { 2467 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2468 high, alignment, boundary, memattr); 2469 if (m != NULL) 2470 break; 2471 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2472 2473 return (m); 2474 } 2475 2476 vm_page_t 2477 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2478 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2479 vm_memattr_t memattr) 2480 { 2481 vm_page_t m, m_ret; 2482 u_int flags; 2483 2484 #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) 2485 KASSERT((req & ~VPANC_FLAGS) == 0, 2486 ("invalid request %#x", req)); 2487 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2488 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2489 ("invalid request %#x", req)); 2490 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2491 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2492 ("invalid request %#x", req)); 2493 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2494 2495 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2496 low, high, alignment, boundary)) == NULL) { 2497 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2498 return (NULL); 2499 } 2500 2501 /* 2502 * Initialize the pages. Only the PG_ZERO flag is inherited. 2503 */ 2504 flags = PG_ZERO; 2505 if ((req & VM_ALLOC_NODUMP) != 0) 2506 flags |= PG_NODUMP; 2507 if ((req & VM_ALLOC_WIRED) != 0) 2508 vm_wire_add(npages); 2509 for (m = m_ret; m < &m_ret[npages]; m++) { 2510 vm_page_dequeue(m); 2511 vm_page_alloc_check(m); 2512 2513 /* 2514 * Consumers should not rely on a useful default pindex value. 2515 */ 2516 m->pindex = 0xdeadc0dedeadc0de; 2517 m->a.flags = 0; 2518 m->flags = (m->flags | PG_NODUMP) & flags; 2519 m->busy_lock = VPB_UNBUSIED; 2520 if ((req & VM_ALLOC_WIRED) != 0) 2521 m->ref_count = 1; 2522 m->a.act_count = 0; 2523 m->oflags = VPO_UNMANAGED; 2524 2525 /* 2526 * Zero the page before updating any mappings since the page is 2527 * not yet shared with any devices which might require the 2528 * non-default memory attribute. pmap_page_set_memattr() 2529 * flushes data caches before returning. 2530 */ 2531 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2532 pmap_zero_page(m); 2533 if (memattr != VM_MEMATTR_DEFAULT) 2534 pmap_page_set_memattr(m, memattr); 2535 } 2536 return (m_ret); 2537 } 2538 2539 /* 2540 * Check a page that has been freshly dequeued from a freelist. 2541 */ 2542 static void 2543 vm_page_alloc_check(vm_page_t m) 2544 { 2545 2546 KASSERT(m->object == NULL, ("page %p has object", m)); 2547 KASSERT(m->a.queue == PQ_NONE && 2548 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2549 ("page %p has unexpected queue %d, flags %#x", 2550 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2551 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2552 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2553 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2554 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2555 ("page %p has unexpected memattr %d", 2556 m, pmap_page_get_memattr(m))); 2557 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2558 pmap_vm_page_alloc_check(m); 2559 } 2560 2561 static int 2562 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2563 { 2564 struct vm_domain *vmd; 2565 struct vm_pgcache *pgcache; 2566 int i; 2567 2568 pgcache = arg; 2569 vmd = VM_DOMAIN(pgcache->domain); 2570 2571 /* 2572 * The page daemon should avoid creating extra memory pressure since its 2573 * main purpose is to replenish the store of free pages. 2574 */ 2575 if (vmd->vmd_severeset || curproc == pageproc || 2576 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2577 return (0); 2578 domain = vmd->vmd_domain; 2579 vm_domain_free_lock(vmd); 2580 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2581 (vm_page_t *)store); 2582 vm_domain_free_unlock(vmd); 2583 if (cnt != i) 2584 vm_domain_freecnt_inc(vmd, cnt - i); 2585 2586 return (i); 2587 } 2588 2589 static void 2590 vm_page_zone_release(void *arg, void **store, int cnt) 2591 { 2592 struct vm_domain *vmd; 2593 struct vm_pgcache *pgcache; 2594 vm_page_t m; 2595 int i; 2596 2597 pgcache = arg; 2598 vmd = VM_DOMAIN(pgcache->domain); 2599 vm_domain_free_lock(vmd); 2600 for (i = 0; i < cnt; i++) { 2601 m = (vm_page_t)store[i]; 2602 vm_phys_free_pages(m, 0); 2603 } 2604 vm_domain_free_unlock(vmd); 2605 vm_domain_freecnt_inc(vmd, cnt); 2606 } 2607 2608 #define VPSC_ANY 0 /* No restrictions. */ 2609 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2610 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2611 2612 /* 2613 * vm_page_scan_contig: 2614 * 2615 * Scan vm_page_array[] between the specified entries "m_start" and 2616 * "m_end" for a run of contiguous physical pages that satisfy the 2617 * specified conditions, and return the lowest page in the run. The 2618 * specified "alignment" determines the alignment of the lowest physical 2619 * page in the run. If the specified "boundary" is non-zero, then the 2620 * run of physical pages cannot span a physical address that is a 2621 * multiple of "boundary". 2622 * 2623 * "m_end" is never dereferenced, so it need not point to a vm_page 2624 * structure within vm_page_array[]. 2625 * 2626 * "npages" must be greater than zero. "m_start" and "m_end" must not 2627 * span a hole (or discontiguity) in the physical address space. Both 2628 * "alignment" and "boundary" must be a power of two. 2629 */ 2630 vm_page_t 2631 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2632 u_long alignment, vm_paddr_t boundary, int options) 2633 { 2634 vm_object_t object; 2635 vm_paddr_t pa; 2636 vm_page_t m, m_run; 2637 #if VM_NRESERVLEVEL > 0 2638 int level; 2639 #endif 2640 int m_inc, order, run_ext, run_len; 2641 2642 KASSERT(npages > 0, ("npages is 0")); 2643 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2644 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2645 m_run = NULL; 2646 run_len = 0; 2647 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2648 KASSERT((m->flags & PG_MARKER) == 0, 2649 ("page %p is PG_MARKER", m)); 2650 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2651 ("fictitious page %p has invalid ref count", m)); 2652 2653 /* 2654 * If the current page would be the start of a run, check its 2655 * physical address against the end, alignment, and boundary 2656 * conditions. If it doesn't satisfy these conditions, either 2657 * terminate the scan or advance to the next page that 2658 * satisfies the failed condition. 2659 */ 2660 if (run_len == 0) { 2661 KASSERT(m_run == NULL, ("m_run != NULL")); 2662 if (m + npages > m_end) 2663 break; 2664 pa = VM_PAGE_TO_PHYS(m); 2665 if (!vm_addr_align_ok(pa, alignment)) { 2666 m_inc = atop(roundup2(pa, alignment) - pa); 2667 continue; 2668 } 2669 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2670 m_inc = atop(roundup2(pa, boundary) - pa); 2671 continue; 2672 } 2673 } else 2674 KASSERT(m_run != NULL, ("m_run == NULL")); 2675 2676 retry: 2677 m_inc = 1; 2678 if (vm_page_wired(m)) 2679 run_ext = 0; 2680 #if VM_NRESERVLEVEL > 0 2681 else if ((level = vm_reserv_level(m)) >= 0 && 2682 (options & VPSC_NORESERV) != 0) { 2683 run_ext = 0; 2684 /* Advance to the end of the reservation. */ 2685 pa = VM_PAGE_TO_PHYS(m); 2686 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2687 pa); 2688 } 2689 #endif 2690 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2691 /* 2692 * The page is considered eligible for relocation if 2693 * and only if it could be laundered or reclaimed by 2694 * the page daemon. 2695 */ 2696 VM_OBJECT_RLOCK(object); 2697 if (object != m->object) { 2698 VM_OBJECT_RUNLOCK(object); 2699 goto retry; 2700 } 2701 /* Don't care: PG_NODUMP, PG_ZERO. */ 2702 if ((object->flags & OBJ_SWAP) == 0 && 2703 object->type != OBJT_VNODE) { 2704 run_ext = 0; 2705 #if VM_NRESERVLEVEL > 0 2706 } else if ((options & VPSC_NOSUPER) != 0 && 2707 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2708 run_ext = 0; 2709 /* Advance to the end of the superpage. */ 2710 pa = VM_PAGE_TO_PHYS(m); 2711 m_inc = atop(roundup2(pa + 1, 2712 vm_reserv_size(level)) - pa); 2713 #endif 2714 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2715 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2716 /* 2717 * The page is allocated but eligible for 2718 * relocation. Extend the current run by one 2719 * page. 2720 */ 2721 KASSERT(pmap_page_get_memattr(m) == 2722 VM_MEMATTR_DEFAULT, 2723 ("page %p has an unexpected memattr", m)); 2724 KASSERT((m->oflags & (VPO_SWAPINPROG | 2725 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2726 ("page %p has unexpected oflags", m)); 2727 /* Don't care: PGA_NOSYNC. */ 2728 run_ext = 1; 2729 } else 2730 run_ext = 0; 2731 VM_OBJECT_RUNLOCK(object); 2732 #if VM_NRESERVLEVEL > 0 2733 } else if (level >= 0) { 2734 /* 2735 * The page is reserved but not yet allocated. In 2736 * other words, it is still free. Extend the current 2737 * run by one page. 2738 */ 2739 run_ext = 1; 2740 #endif 2741 } else if ((order = m->order) < VM_NFREEORDER) { 2742 /* 2743 * The page is enqueued in the physical memory 2744 * allocator's free page queues. Moreover, it is the 2745 * first page in a power-of-two-sized run of 2746 * contiguous free pages. Add these pages to the end 2747 * of the current run, and jump ahead. 2748 */ 2749 run_ext = 1 << order; 2750 m_inc = 1 << order; 2751 } else { 2752 /* 2753 * Skip the page for one of the following reasons: (1) 2754 * It is enqueued in the physical memory allocator's 2755 * free page queues. However, it is not the first 2756 * page in a run of contiguous free pages. (This case 2757 * rarely occurs because the scan is performed in 2758 * ascending order.) (2) It is not reserved, and it is 2759 * transitioning from free to allocated. (Conversely, 2760 * the transition from allocated to free for managed 2761 * pages is blocked by the page busy lock.) (3) It is 2762 * allocated but not contained by an object and not 2763 * wired, e.g., allocated by Xen's balloon driver. 2764 */ 2765 run_ext = 0; 2766 } 2767 2768 /* 2769 * Extend or reset the current run of pages. 2770 */ 2771 if (run_ext > 0) { 2772 if (run_len == 0) 2773 m_run = m; 2774 run_len += run_ext; 2775 } else { 2776 if (run_len > 0) { 2777 m_run = NULL; 2778 run_len = 0; 2779 } 2780 } 2781 } 2782 if (run_len >= npages) 2783 return (m_run); 2784 return (NULL); 2785 } 2786 2787 /* 2788 * vm_page_reclaim_run: 2789 * 2790 * Try to relocate each of the allocated virtual pages within the 2791 * specified run of physical pages to a new physical address. Free the 2792 * physical pages underlying the relocated virtual pages. A virtual page 2793 * is relocatable if and only if it could be laundered or reclaimed by 2794 * the page daemon. Whenever possible, a virtual page is relocated to a 2795 * physical address above "high". 2796 * 2797 * Returns 0 if every physical page within the run was already free or 2798 * just freed by a successful relocation. Otherwise, returns a non-zero 2799 * value indicating why the last attempt to relocate a virtual page was 2800 * unsuccessful. 2801 * 2802 * "req_class" must be an allocation class. 2803 */ 2804 static int 2805 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2806 vm_paddr_t high) 2807 { 2808 struct vm_domain *vmd; 2809 struct spglist free; 2810 vm_object_t object; 2811 vm_paddr_t pa; 2812 vm_page_t m, m_end, m_new; 2813 int error, order, req; 2814 2815 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2816 ("req_class is not an allocation class")); 2817 SLIST_INIT(&free); 2818 error = 0; 2819 m = m_run; 2820 m_end = m_run + npages; 2821 for (; error == 0 && m < m_end; m++) { 2822 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2823 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2824 2825 /* 2826 * Racily check for wirings. Races are handled once the object 2827 * lock is held and the page is unmapped. 2828 */ 2829 if (vm_page_wired(m)) 2830 error = EBUSY; 2831 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2832 /* 2833 * The page is relocated if and only if it could be 2834 * laundered or reclaimed by the page daemon. 2835 */ 2836 VM_OBJECT_WLOCK(object); 2837 /* Don't care: PG_NODUMP, PG_ZERO. */ 2838 if (m->object != object || 2839 ((object->flags & OBJ_SWAP) == 0 && 2840 object->type != OBJT_VNODE)) 2841 error = EINVAL; 2842 else if (object->memattr != VM_MEMATTR_DEFAULT) 2843 error = EINVAL; 2844 else if (vm_page_queue(m) != PQ_NONE && 2845 vm_page_tryxbusy(m) != 0) { 2846 if (vm_page_wired(m)) { 2847 vm_page_xunbusy(m); 2848 error = EBUSY; 2849 goto unlock; 2850 } 2851 KASSERT(pmap_page_get_memattr(m) == 2852 VM_MEMATTR_DEFAULT, 2853 ("page %p has an unexpected memattr", m)); 2854 KASSERT(m->oflags == 0, 2855 ("page %p has unexpected oflags", m)); 2856 /* Don't care: PGA_NOSYNC. */ 2857 if (!vm_page_none_valid(m)) { 2858 /* 2859 * First, try to allocate a new page 2860 * that is above "high". Failing 2861 * that, try to allocate a new page 2862 * that is below "m_run". Allocate 2863 * the new page between the end of 2864 * "m_run" and "high" only as a last 2865 * resort. 2866 */ 2867 req = req_class; 2868 if ((m->flags & PG_NODUMP) != 0) 2869 req |= VM_ALLOC_NODUMP; 2870 if (trunc_page(high) != 2871 ~(vm_paddr_t)PAGE_MASK) { 2872 m_new = 2873 vm_page_alloc_noobj_contig( 2874 req, 1, round_page(high), 2875 ~(vm_paddr_t)0, PAGE_SIZE, 2876 0, VM_MEMATTR_DEFAULT); 2877 } else 2878 m_new = NULL; 2879 if (m_new == NULL) { 2880 pa = VM_PAGE_TO_PHYS(m_run); 2881 m_new = 2882 vm_page_alloc_noobj_contig( 2883 req, 1, 0, pa - 1, 2884 PAGE_SIZE, 0, 2885 VM_MEMATTR_DEFAULT); 2886 } 2887 if (m_new == NULL) { 2888 pa += ptoa(npages); 2889 m_new = 2890 vm_page_alloc_noobj_contig( 2891 req, 1, pa, high, PAGE_SIZE, 2892 0, VM_MEMATTR_DEFAULT); 2893 } 2894 if (m_new == NULL) { 2895 vm_page_xunbusy(m); 2896 error = ENOMEM; 2897 goto unlock; 2898 } 2899 2900 /* 2901 * Unmap the page and check for new 2902 * wirings that may have been acquired 2903 * through a pmap lookup. 2904 */ 2905 if (object->ref_count != 0 && 2906 !vm_page_try_remove_all(m)) { 2907 vm_page_xunbusy(m); 2908 vm_page_free(m_new); 2909 error = EBUSY; 2910 goto unlock; 2911 } 2912 2913 /* 2914 * Replace "m" with the new page. For 2915 * vm_page_replace(), "m" must be busy 2916 * and dequeued. Finally, change "m" 2917 * as if vm_page_free() was called. 2918 */ 2919 m_new->a.flags = m->a.flags & 2920 ~PGA_QUEUE_STATE_MASK; 2921 KASSERT(m_new->oflags == VPO_UNMANAGED, 2922 ("page %p is managed", m_new)); 2923 m_new->oflags = 0; 2924 pmap_copy_page(m, m_new); 2925 m_new->valid = m->valid; 2926 m_new->dirty = m->dirty; 2927 m->flags &= ~PG_ZERO; 2928 vm_page_dequeue(m); 2929 if (vm_page_replace_hold(m_new, object, 2930 m->pindex, m) && 2931 vm_page_free_prep(m)) 2932 SLIST_INSERT_HEAD(&free, m, 2933 plinks.s.ss); 2934 2935 /* 2936 * The new page must be deactivated 2937 * before the object is unlocked. 2938 */ 2939 vm_page_deactivate(m_new); 2940 } else { 2941 m->flags &= ~PG_ZERO; 2942 vm_page_dequeue(m); 2943 if (vm_page_free_prep(m)) 2944 SLIST_INSERT_HEAD(&free, m, 2945 plinks.s.ss); 2946 KASSERT(m->dirty == 0, 2947 ("page %p is dirty", m)); 2948 } 2949 } else 2950 error = EBUSY; 2951 unlock: 2952 VM_OBJECT_WUNLOCK(object); 2953 } else { 2954 MPASS(vm_page_domain(m) == domain); 2955 vmd = VM_DOMAIN(domain); 2956 vm_domain_free_lock(vmd); 2957 order = m->order; 2958 if (order < VM_NFREEORDER) { 2959 /* 2960 * The page is enqueued in the physical memory 2961 * allocator's free page queues. Moreover, it 2962 * is the first page in a power-of-two-sized 2963 * run of contiguous free pages. Jump ahead 2964 * to the last page within that run, and 2965 * continue from there. 2966 */ 2967 m += (1 << order) - 1; 2968 } 2969 #if VM_NRESERVLEVEL > 0 2970 else if (vm_reserv_is_page_free(m)) 2971 order = 0; 2972 #endif 2973 vm_domain_free_unlock(vmd); 2974 if (order == VM_NFREEORDER) 2975 error = EINVAL; 2976 } 2977 } 2978 if ((m = SLIST_FIRST(&free)) != NULL) { 2979 int cnt; 2980 2981 vmd = VM_DOMAIN(domain); 2982 cnt = 0; 2983 vm_domain_free_lock(vmd); 2984 do { 2985 MPASS(vm_page_domain(m) == domain); 2986 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2987 vm_phys_free_pages(m, 0); 2988 cnt++; 2989 } while ((m = SLIST_FIRST(&free)) != NULL); 2990 vm_domain_free_unlock(vmd); 2991 vm_domain_freecnt_inc(vmd, cnt); 2992 } 2993 return (error); 2994 } 2995 2996 #define NRUNS 16 2997 2998 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 2999 3000 #define MIN_RECLAIM 8 3001 3002 /* 3003 * vm_page_reclaim_contig: 3004 * 3005 * Reclaim allocated, contiguous physical memory satisfying the specified 3006 * conditions by relocating the virtual pages using that physical memory. 3007 * Returns true if reclamation is successful and false otherwise. Since 3008 * relocation requires the allocation of physical pages, reclamation may 3009 * fail due to a shortage of free pages. When reclamation fails, callers 3010 * are expected to perform vm_wait() before retrying a failed allocation 3011 * operation, e.g., vm_page_alloc_contig(). 3012 * 3013 * The caller must always specify an allocation class through "req". 3014 * 3015 * allocation classes: 3016 * VM_ALLOC_NORMAL normal process request 3017 * VM_ALLOC_SYSTEM system *really* needs a page 3018 * VM_ALLOC_INTERRUPT interrupt time request 3019 * 3020 * The optional allocation flags are ignored. 3021 * 3022 * "npages" must be greater than zero. Both "alignment" and "boundary" 3023 * must be a power of two. 3024 */ 3025 bool 3026 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3027 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3028 int desired_runs) 3029 { 3030 struct vm_domain *vmd; 3031 vm_paddr_t curr_low; 3032 vm_page_t m_run, _m_runs[NRUNS], *m_runs; 3033 u_long count, minalign, reclaimed; 3034 int error, i, min_reclaim, nruns, options, req_class; 3035 bool ret; 3036 3037 KASSERT(npages > 0, ("npages is 0")); 3038 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3039 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3040 3041 ret = false; 3042 3043 /* 3044 * If the caller wants to reclaim multiple runs, try to allocate 3045 * space to store the runs. If that fails, fall back to the old 3046 * behavior of just reclaiming MIN_RECLAIM pages. 3047 */ 3048 if (desired_runs > 1) 3049 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3050 M_TEMP, M_NOWAIT); 3051 else 3052 m_runs = NULL; 3053 3054 if (m_runs == NULL) { 3055 m_runs = _m_runs; 3056 nruns = NRUNS; 3057 } else { 3058 nruns = NRUNS + desired_runs - 1; 3059 } 3060 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3061 3062 /* 3063 * The caller will attempt an allocation after some runs have been 3064 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3065 * of vm_phys_alloc_contig(), round up the requested length to the next 3066 * power of two or maximum chunk size, and ensure that each run is 3067 * suitably aligned. 3068 */ 3069 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3070 npages = roundup2(npages, minalign); 3071 if (alignment < ptoa(minalign)) 3072 alignment = ptoa(minalign); 3073 3074 /* 3075 * The page daemon is allowed to dig deeper into the free page list. 3076 */ 3077 req_class = req & VM_ALLOC_CLASS_MASK; 3078 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3079 req_class = VM_ALLOC_SYSTEM; 3080 3081 /* 3082 * Return if the number of free pages cannot satisfy the requested 3083 * allocation. 3084 */ 3085 vmd = VM_DOMAIN(domain); 3086 count = vmd->vmd_free_count; 3087 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3088 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3089 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3090 goto done; 3091 3092 /* 3093 * Scan up to three times, relaxing the restrictions ("options") on 3094 * the reclamation of reservations and superpages each time. 3095 */ 3096 for (options = VPSC_NORESERV;;) { 3097 /* 3098 * Find the highest runs that satisfy the given constraints 3099 * and restrictions, and record them in "m_runs". 3100 */ 3101 curr_low = low; 3102 count = 0; 3103 for (;;) { 3104 m_run = vm_phys_scan_contig(domain, npages, curr_low, 3105 high, alignment, boundary, options); 3106 if (m_run == NULL) 3107 break; 3108 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 3109 m_runs[RUN_INDEX(count, nruns)] = m_run; 3110 count++; 3111 } 3112 3113 /* 3114 * Reclaim the highest runs in LIFO (descending) order until 3115 * the number of reclaimed pages, "reclaimed", is at least 3116 * "min_reclaim". Reset "reclaimed" each time because each 3117 * reclamation is idempotent, and runs will (likely) recur 3118 * from one scan to the next as restrictions are relaxed. 3119 */ 3120 reclaimed = 0; 3121 for (i = 0; count > 0 && i < nruns; i++) { 3122 count--; 3123 m_run = m_runs[RUN_INDEX(count, nruns)]; 3124 error = vm_page_reclaim_run(req_class, domain, npages, 3125 m_run, high); 3126 if (error == 0) { 3127 reclaimed += npages; 3128 if (reclaimed >= min_reclaim) { 3129 ret = true; 3130 goto done; 3131 } 3132 } 3133 } 3134 3135 /* 3136 * Either relax the restrictions on the next scan or return if 3137 * the last scan had no restrictions. 3138 */ 3139 if (options == VPSC_NORESERV) 3140 options = VPSC_NOSUPER; 3141 else if (options == VPSC_NOSUPER) 3142 options = VPSC_ANY; 3143 else if (options == VPSC_ANY) { 3144 ret = reclaimed != 0; 3145 goto done; 3146 } 3147 } 3148 done: 3149 if (m_runs != _m_runs) 3150 free(m_runs, M_TEMP); 3151 return (ret); 3152 } 3153 3154 bool 3155 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3156 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3157 { 3158 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high, 3159 alignment, boundary, 1)); 3160 } 3161 3162 bool 3163 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3164 u_long alignment, vm_paddr_t boundary) 3165 { 3166 struct vm_domainset_iter di; 3167 int domain; 3168 bool ret; 3169 3170 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3171 do { 3172 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 3173 high, alignment, boundary); 3174 if (ret) 3175 break; 3176 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3177 3178 return (ret); 3179 } 3180 3181 /* 3182 * Set the domain in the appropriate page level domainset. 3183 */ 3184 void 3185 vm_domain_set(struct vm_domain *vmd) 3186 { 3187 3188 mtx_lock(&vm_domainset_lock); 3189 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3190 vmd->vmd_minset = 1; 3191 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3192 } 3193 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3194 vmd->vmd_severeset = 1; 3195 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3196 } 3197 mtx_unlock(&vm_domainset_lock); 3198 } 3199 3200 /* 3201 * Clear the domain from the appropriate page level domainset. 3202 */ 3203 void 3204 vm_domain_clear(struct vm_domain *vmd) 3205 { 3206 3207 mtx_lock(&vm_domainset_lock); 3208 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3209 vmd->vmd_minset = 0; 3210 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3211 if (vm_min_waiters != 0) { 3212 vm_min_waiters = 0; 3213 wakeup(&vm_min_domains); 3214 } 3215 } 3216 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3217 vmd->vmd_severeset = 0; 3218 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3219 if (vm_severe_waiters != 0) { 3220 vm_severe_waiters = 0; 3221 wakeup(&vm_severe_domains); 3222 } 3223 } 3224 3225 /* 3226 * If pageout daemon needs pages, then tell it that there are 3227 * some free. 3228 */ 3229 if (vmd->vmd_pageout_pages_needed && 3230 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3231 wakeup(&vmd->vmd_pageout_pages_needed); 3232 vmd->vmd_pageout_pages_needed = 0; 3233 } 3234 3235 /* See comments in vm_wait_doms(). */ 3236 if (vm_pageproc_waiters) { 3237 vm_pageproc_waiters = 0; 3238 wakeup(&vm_pageproc_waiters); 3239 } 3240 mtx_unlock(&vm_domainset_lock); 3241 } 3242 3243 /* 3244 * Wait for free pages to exceed the min threshold globally. 3245 */ 3246 void 3247 vm_wait_min(void) 3248 { 3249 3250 mtx_lock(&vm_domainset_lock); 3251 while (vm_page_count_min()) { 3252 vm_min_waiters++; 3253 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3254 } 3255 mtx_unlock(&vm_domainset_lock); 3256 } 3257 3258 /* 3259 * Wait for free pages to exceed the severe threshold globally. 3260 */ 3261 void 3262 vm_wait_severe(void) 3263 { 3264 3265 mtx_lock(&vm_domainset_lock); 3266 while (vm_page_count_severe()) { 3267 vm_severe_waiters++; 3268 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3269 "vmwait", 0); 3270 } 3271 mtx_unlock(&vm_domainset_lock); 3272 } 3273 3274 u_int 3275 vm_wait_count(void) 3276 { 3277 3278 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3279 } 3280 3281 int 3282 vm_wait_doms(const domainset_t *wdoms, int mflags) 3283 { 3284 int error; 3285 3286 error = 0; 3287 3288 /* 3289 * We use racey wakeup synchronization to avoid expensive global 3290 * locking for the pageproc when sleeping with a non-specific vm_wait. 3291 * To handle this, we only sleep for one tick in this instance. It 3292 * is expected that most allocations for the pageproc will come from 3293 * kmem or vm_page_grab* which will use the more specific and 3294 * race-free vm_wait_domain(). 3295 */ 3296 if (curproc == pageproc) { 3297 mtx_lock(&vm_domainset_lock); 3298 vm_pageproc_waiters++; 3299 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3300 PVM | PDROP | mflags, "pageprocwait", 1); 3301 } else { 3302 /* 3303 * XXX Ideally we would wait only until the allocation could 3304 * be satisfied. This condition can cause new allocators to 3305 * consume all freed pages while old allocators wait. 3306 */ 3307 mtx_lock(&vm_domainset_lock); 3308 if (vm_page_count_min_set(wdoms)) { 3309 if (pageproc == NULL) 3310 panic("vm_wait in early boot"); 3311 vm_min_waiters++; 3312 error = msleep(&vm_min_domains, &vm_domainset_lock, 3313 PVM | PDROP | mflags, "vmwait", 0); 3314 } else 3315 mtx_unlock(&vm_domainset_lock); 3316 } 3317 return (error); 3318 } 3319 3320 /* 3321 * vm_wait_domain: 3322 * 3323 * Sleep until free pages are available for allocation. 3324 * - Called in various places after failed memory allocations. 3325 */ 3326 void 3327 vm_wait_domain(int domain) 3328 { 3329 struct vm_domain *vmd; 3330 domainset_t wdom; 3331 3332 vmd = VM_DOMAIN(domain); 3333 vm_domain_free_assert_unlocked(vmd); 3334 3335 if (curproc == pageproc) { 3336 mtx_lock(&vm_domainset_lock); 3337 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3338 vmd->vmd_pageout_pages_needed = 1; 3339 msleep(&vmd->vmd_pageout_pages_needed, 3340 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3341 } else 3342 mtx_unlock(&vm_domainset_lock); 3343 } else { 3344 DOMAINSET_ZERO(&wdom); 3345 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3346 vm_wait_doms(&wdom, 0); 3347 } 3348 } 3349 3350 static int 3351 vm_wait_flags(vm_object_t obj, int mflags) 3352 { 3353 struct domainset *d; 3354 3355 d = NULL; 3356 3357 /* 3358 * Carefully fetch pointers only once: the struct domainset 3359 * itself is ummutable but the pointer might change. 3360 */ 3361 if (obj != NULL) 3362 d = obj->domain.dr_policy; 3363 if (d == NULL) 3364 d = curthread->td_domain.dr_policy; 3365 3366 return (vm_wait_doms(&d->ds_mask, mflags)); 3367 } 3368 3369 /* 3370 * vm_wait: 3371 * 3372 * Sleep until free pages are available for allocation in the 3373 * affinity domains of the obj. If obj is NULL, the domain set 3374 * for the calling thread is used. 3375 * Called in various places after failed memory allocations. 3376 */ 3377 void 3378 vm_wait(vm_object_t obj) 3379 { 3380 (void)vm_wait_flags(obj, 0); 3381 } 3382 3383 int 3384 vm_wait_intr(vm_object_t obj) 3385 { 3386 return (vm_wait_flags(obj, PCATCH)); 3387 } 3388 3389 /* 3390 * vm_domain_alloc_fail: 3391 * 3392 * Called when a page allocation function fails. Informs the 3393 * pagedaemon and performs the requested wait. Requires the 3394 * domain_free and object lock on entry. Returns with the 3395 * object lock held and free lock released. Returns an error when 3396 * retry is necessary. 3397 * 3398 */ 3399 static int 3400 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3401 { 3402 3403 vm_domain_free_assert_unlocked(vmd); 3404 3405 atomic_add_int(&vmd->vmd_pageout_deficit, 3406 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3407 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3408 if (object != NULL) 3409 VM_OBJECT_WUNLOCK(object); 3410 vm_wait_domain(vmd->vmd_domain); 3411 if (object != NULL) 3412 VM_OBJECT_WLOCK(object); 3413 if (req & VM_ALLOC_WAITOK) 3414 return (EAGAIN); 3415 } 3416 3417 return (0); 3418 } 3419 3420 /* 3421 * vm_waitpfault: 3422 * 3423 * Sleep until free pages are available for allocation. 3424 * - Called only in vm_fault so that processes page faulting 3425 * can be easily tracked. 3426 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3427 * processes will be able to grab memory first. Do not change 3428 * this balance without careful testing first. 3429 */ 3430 void 3431 vm_waitpfault(struct domainset *dset, int timo) 3432 { 3433 3434 /* 3435 * XXX Ideally we would wait only until the allocation could 3436 * be satisfied. This condition can cause new allocators to 3437 * consume all freed pages while old allocators wait. 3438 */ 3439 mtx_lock(&vm_domainset_lock); 3440 if (vm_page_count_min_set(&dset->ds_mask)) { 3441 vm_min_waiters++; 3442 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3443 "pfault", timo); 3444 } else 3445 mtx_unlock(&vm_domainset_lock); 3446 } 3447 3448 static struct vm_pagequeue * 3449 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3450 { 3451 3452 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3453 } 3454 3455 #ifdef INVARIANTS 3456 static struct vm_pagequeue * 3457 vm_page_pagequeue(vm_page_t m) 3458 { 3459 3460 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3461 } 3462 #endif 3463 3464 static __always_inline bool 3465 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3466 { 3467 vm_page_astate_t tmp; 3468 3469 tmp = *old; 3470 do { 3471 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3472 return (true); 3473 counter_u64_add(pqstate_commit_retries, 1); 3474 } while (old->_bits == tmp._bits); 3475 3476 return (false); 3477 } 3478 3479 /* 3480 * Do the work of committing a queue state update that moves the page out of 3481 * its current queue. 3482 */ 3483 static bool 3484 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3485 vm_page_astate_t *old, vm_page_astate_t new) 3486 { 3487 vm_page_t next; 3488 3489 vm_pagequeue_assert_locked(pq); 3490 KASSERT(vm_page_pagequeue(m) == pq, 3491 ("%s: queue %p does not match page %p", __func__, pq, m)); 3492 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3493 ("%s: invalid queue indices %d %d", 3494 __func__, old->queue, new.queue)); 3495 3496 /* 3497 * Once the queue index of the page changes there is nothing 3498 * synchronizing with further updates to the page's physical 3499 * queue state. Therefore we must speculatively remove the page 3500 * from the queue now and be prepared to roll back if the queue 3501 * state update fails. If the page is not physically enqueued then 3502 * we just update its queue index. 3503 */ 3504 if ((old->flags & PGA_ENQUEUED) != 0) { 3505 new.flags &= ~PGA_ENQUEUED; 3506 next = TAILQ_NEXT(m, plinks.q); 3507 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3508 vm_pagequeue_cnt_dec(pq); 3509 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3510 if (next == NULL) 3511 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3512 else 3513 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3514 vm_pagequeue_cnt_inc(pq); 3515 return (false); 3516 } else { 3517 return (true); 3518 } 3519 } else { 3520 return (vm_page_pqstate_fcmpset(m, old, new)); 3521 } 3522 } 3523 3524 static bool 3525 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3526 vm_page_astate_t new) 3527 { 3528 struct vm_pagequeue *pq; 3529 vm_page_astate_t as; 3530 bool ret; 3531 3532 pq = _vm_page_pagequeue(m, old->queue); 3533 3534 /* 3535 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3536 * corresponding page queue lock is held. 3537 */ 3538 vm_pagequeue_lock(pq); 3539 as = vm_page_astate_load(m); 3540 if (__predict_false(as._bits != old->_bits)) { 3541 *old = as; 3542 ret = false; 3543 } else { 3544 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3545 } 3546 vm_pagequeue_unlock(pq); 3547 return (ret); 3548 } 3549 3550 /* 3551 * Commit a queue state update that enqueues or requeues a page. 3552 */ 3553 static bool 3554 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3555 vm_page_astate_t *old, vm_page_astate_t new) 3556 { 3557 struct vm_domain *vmd; 3558 3559 vm_pagequeue_assert_locked(pq); 3560 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3561 ("%s: invalid queue indices %d %d", 3562 __func__, old->queue, new.queue)); 3563 3564 new.flags |= PGA_ENQUEUED; 3565 if (!vm_page_pqstate_fcmpset(m, old, new)) 3566 return (false); 3567 3568 if ((old->flags & PGA_ENQUEUED) != 0) 3569 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3570 else 3571 vm_pagequeue_cnt_inc(pq); 3572 3573 /* 3574 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3575 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3576 * applied, even if it was set first. 3577 */ 3578 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3579 vmd = vm_pagequeue_domain(m); 3580 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3581 ("%s: invalid page queue for page %p", __func__, m)); 3582 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3583 } else { 3584 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3585 } 3586 return (true); 3587 } 3588 3589 /* 3590 * Commit a queue state update that encodes a request for a deferred queue 3591 * operation. 3592 */ 3593 static bool 3594 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3595 vm_page_astate_t new) 3596 { 3597 3598 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3599 ("%s: invalid state, queue %d flags %x", 3600 __func__, new.queue, new.flags)); 3601 3602 if (old->_bits != new._bits && 3603 !vm_page_pqstate_fcmpset(m, old, new)) 3604 return (false); 3605 vm_page_pqbatch_submit(m, new.queue); 3606 return (true); 3607 } 3608 3609 /* 3610 * A generic queue state update function. This handles more cases than the 3611 * specialized functions above. 3612 */ 3613 bool 3614 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3615 { 3616 3617 if (old->_bits == new._bits) 3618 return (true); 3619 3620 if (old->queue != PQ_NONE && new.queue != old->queue) { 3621 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3622 return (false); 3623 if (new.queue != PQ_NONE) 3624 vm_page_pqbatch_submit(m, new.queue); 3625 } else { 3626 if (!vm_page_pqstate_fcmpset(m, old, new)) 3627 return (false); 3628 if (new.queue != PQ_NONE && 3629 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3630 vm_page_pqbatch_submit(m, new.queue); 3631 } 3632 return (true); 3633 } 3634 3635 /* 3636 * Apply deferred queue state updates to a page. 3637 */ 3638 static inline void 3639 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3640 { 3641 vm_page_astate_t new, old; 3642 3643 CRITICAL_ASSERT(curthread); 3644 vm_pagequeue_assert_locked(pq); 3645 KASSERT(queue < PQ_COUNT, 3646 ("%s: invalid queue index %d", __func__, queue)); 3647 KASSERT(pq == _vm_page_pagequeue(m, queue), 3648 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3649 3650 for (old = vm_page_astate_load(m);;) { 3651 if (__predict_false(old.queue != queue || 3652 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3653 counter_u64_add(queue_nops, 1); 3654 break; 3655 } 3656 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3657 ("%s: page %p is unmanaged", __func__, m)); 3658 3659 new = old; 3660 if ((old.flags & PGA_DEQUEUE) != 0) { 3661 new.flags &= ~PGA_QUEUE_OP_MASK; 3662 new.queue = PQ_NONE; 3663 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3664 m, &old, new))) { 3665 counter_u64_add(queue_ops, 1); 3666 break; 3667 } 3668 } else { 3669 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3670 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3671 m, &old, new))) { 3672 counter_u64_add(queue_ops, 1); 3673 break; 3674 } 3675 } 3676 } 3677 } 3678 3679 static void 3680 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3681 uint8_t queue) 3682 { 3683 int i; 3684 3685 for (i = 0; i < bq->bq_cnt; i++) 3686 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3687 vm_batchqueue_init(bq); 3688 } 3689 3690 /* 3691 * vm_page_pqbatch_submit: [ internal use only ] 3692 * 3693 * Enqueue a page in the specified page queue's batched work queue. 3694 * The caller must have encoded the requested operation in the page 3695 * structure's a.flags field. 3696 */ 3697 void 3698 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3699 { 3700 struct vm_batchqueue *bq; 3701 struct vm_pagequeue *pq; 3702 int domain, slots_remaining; 3703 3704 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3705 3706 domain = vm_page_domain(m); 3707 critical_enter(); 3708 bq = DPCPU_PTR(pqbatch[domain][queue]); 3709 slots_remaining = vm_batchqueue_insert(bq, m); 3710 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 3711 /* keep building the bq */ 3712 critical_exit(); 3713 return; 3714 } else if (slots_remaining > 0 ) { 3715 /* Try to process the bq if we can get the lock */ 3716 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3717 if (vm_pagequeue_trylock(pq)) { 3718 vm_pqbatch_process(pq, bq, queue); 3719 vm_pagequeue_unlock(pq); 3720 } 3721 critical_exit(); 3722 return; 3723 } 3724 critical_exit(); 3725 3726 /* if we make it here, the bq is full so wait for the lock */ 3727 3728 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3729 vm_pagequeue_lock(pq); 3730 critical_enter(); 3731 bq = DPCPU_PTR(pqbatch[domain][queue]); 3732 vm_pqbatch_process(pq, bq, queue); 3733 vm_pqbatch_process_page(pq, m, queue); 3734 vm_pagequeue_unlock(pq); 3735 critical_exit(); 3736 } 3737 3738 /* 3739 * vm_page_pqbatch_drain: [ internal use only ] 3740 * 3741 * Force all per-CPU page queue batch queues to be drained. This is 3742 * intended for use in severe memory shortages, to ensure that pages 3743 * do not remain stuck in the batch queues. 3744 */ 3745 void 3746 vm_page_pqbatch_drain(void) 3747 { 3748 struct thread *td; 3749 struct vm_domain *vmd; 3750 struct vm_pagequeue *pq; 3751 int cpu, domain, queue; 3752 3753 td = curthread; 3754 CPU_FOREACH(cpu) { 3755 thread_lock(td); 3756 sched_bind(td, cpu); 3757 thread_unlock(td); 3758 3759 for (domain = 0; domain < vm_ndomains; domain++) { 3760 vmd = VM_DOMAIN(domain); 3761 for (queue = 0; queue < PQ_COUNT; queue++) { 3762 pq = &vmd->vmd_pagequeues[queue]; 3763 vm_pagequeue_lock(pq); 3764 critical_enter(); 3765 vm_pqbatch_process(pq, 3766 DPCPU_PTR(pqbatch[domain][queue]), queue); 3767 critical_exit(); 3768 vm_pagequeue_unlock(pq); 3769 } 3770 } 3771 } 3772 thread_lock(td); 3773 sched_unbind(td); 3774 thread_unlock(td); 3775 } 3776 3777 /* 3778 * vm_page_dequeue_deferred: [ internal use only ] 3779 * 3780 * Request removal of the given page from its current page 3781 * queue. Physical removal from the queue may be deferred 3782 * indefinitely. 3783 */ 3784 void 3785 vm_page_dequeue_deferred(vm_page_t m) 3786 { 3787 vm_page_astate_t new, old; 3788 3789 old = vm_page_astate_load(m); 3790 do { 3791 if (old.queue == PQ_NONE) { 3792 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3793 ("%s: page %p has unexpected queue state", 3794 __func__, m)); 3795 break; 3796 } 3797 new = old; 3798 new.flags |= PGA_DEQUEUE; 3799 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3800 } 3801 3802 /* 3803 * vm_page_dequeue: 3804 * 3805 * Remove the page from whichever page queue it's in, if any, before 3806 * returning. 3807 */ 3808 void 3809 vm_page_dequeue(vm_page_t m) 3810 { 3811 vm_page_astate_t new, old; 3812 3813 old = vm_page_astate_load(m); 3814 do { 3815 if (old.queue == PQ_NONE) { 3816 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3817 ("%s: page %p has unexpected queue state", 3818 __func__, m)); 3819 break; 3820 } 3821 new = old; 3822 new.flags &= ~PGA_QUEUE_OP_MASK; 3823 new.queue = PQ_NONE; 3824 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 3825 3826 } 3827 3828 /* 3829 * Schedule the given page for insertion into the specified page queue. 3830 * Physical insertion of the page may be deferred indefinitely. 3831 */ 3832 static void 3833 vm_page_enqueue(vm_page_t m, uint8_t queue) 3834 { 3835 3836 KASSERT(m->a.queue == PQ_NONE && 3837 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 3838 ("%s: page %p is already enqueued", __func__, m)); 3839 KASSERT(m->ref_count > 0, 3840 ("%s: page %p does not carry any references", __func__, m)); 3841 3842 m->a.queue = queue; 3843 if ((m->a.flags & PGA_REQUEUE) == 0) 3844 vm_page_aflag_set(m, PGA_REQUEUE); 3845 vm_page_pqbatch_submit(m, queue); 3846 } 3847 3848 /* 3849 * vm_page_free_prep: 3850 * 3851 * Prepares the given page to be put on the free list, 3852 * disassociating it from any VM object. The caller may return 3853 * the page to the free list only if this function returns true. 3854 * 3855 * The object, if it exists, must be locked, and then the page must 3856 * be xbusy. Otherwise the page must be not busied. A managed 3857 * page must be unmapped. 3858 */ 3859 static bool 3860 vm_page_free_prep(vm_page_t m) 3861 { 3862 3863 /* 3864 * Synchronize with threads that have dropped a reference to this 3865 * page. 3866 */ 3867 atomic_thread_fence_acq(); 3868 3869 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3870 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3871 uint64_t *p; 3872 int i; 3873 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3874 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3875 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3876 m, i, (uintmax_t)*p)); 3877 } 3878 #endif 3879 if ((m->oflags & VPO_UNMANAGED) == 0) { 3880 KASSERT(!pmap_page_is_mapped(m), 3881 ("vm_page_free_prep: freeing mapped page %p", m)); 3882 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3883 ("vm_page_free_prep: mapping flags set in page %p", m)); 3884 } else { 3885 KASSERT(m->a.queue == PQ_NONE, 3886 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3887 } 3888 VM_CNT_INC(v_tfree); 3889 3890 if (m->object != NULL) { 3891 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 3892 ((m->object->flags & OBJ_UNMANAGED) != 0), 3893 ("vm_page_free_prep: managed flag mismatch for page %p", 3894 m)); 3895 vm_page_assert_xbusied(m); 3896 3897 /* 3898 * The object reference can be released without an atomic 3899 * operation. 3900 */ 3901 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3902 m->ref_count == VPRC_OBJREF, 3903 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3904 m, m->ref_count)); 3905 vm_page_object_remove(m); 3906 m->ref_count -= VPRC_OBJREF; 3907 } else 3908 vm_page_assert_unbusied(m); 3909 3910 vm_page_busy_free(m); 3911 3912 /* 3913 * If fictitious remove object association and 3914 * return. 3915 */ 3916 if ((m->flags & PG_FICTITIOUS) != 0) { 3917 KASSERT(m->ref_count == 1, 3918 ("fictitious page %p is referenced", m)); 3919 KASSERT(m->a.queue == PQ_NONE, 3920 ("fictitious page %p is queued", m)); 3921 return (false); 3922 } 3923 3924 /* 3925 * Pages need not be dequeued before they are returned to the physical 3926 * memory allocator, but they must at least be marked for a deferred 3927 * dequeue. 3928 */ 3929 if ((m->oflags & VPO_UNMANAGED) == 0) 3930 vm_page_dequeue_deferred(m); 3931 3932 m->valid = 0; 3933 vm_page_undirty(m); 3934 3935 if (m->ref_count != 0) 3936 panic("vm_page_free_prep: page %p has references", m); 3937 3938 /* 3939 * Restore the default memory attribute to the page. 3940 */ 3941 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3942 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3943 3944 #if VM_NRESERVLEVEL > 0 3945 /* 3946 * Determine whether the page belongs to a reservation. If the page was 3947 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3948 * as an optimization, we avoid the check in that case. 3949 */ 3950 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3951 return (false); 3952 #endif 3953 3954 return (true); 3955 } 3956 3957 /* 3958 * vm_page_free_toq: 3959 * 3960 * Returns the given page to the free list, disassociating it 3961 * from any VM object. 3962 * 3963 * The object must be locked. The page must be exclusively busied if it 3964 * belongs to an object. 3965 */ 3966 static void 3967 vm_page_free_toq(vm_page_t m) 3968 { 3969 struct vm_domain *vmd; 3970 uma_zone_t zone; 3971 3972 if (!vm_page_free_prep(m)) 3973 return; 3974 3975 vmd = vm_pagequeue_domain(m); 3976 zone = vmd->vmd_pgcache[m->pool].zone; 3977 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 3978 uma_zfree(zone, m); 3979 return; 3980 } 3981 vm_domain_free_lock(vmd); 3982 vm_phys_free_pages(m, 0); 3983 vm_domain_free_unlock(vmd); 3984 vm_domain_freecnt_inc(vmd, 1); 3985 } 3986 3987 /* 3988 * vm_page_free_pages_toq: 3989 * 3990 * Returns a list of pages to the free list, disassociating it 3991 * from any VM object. In other words, this is equivalent to 3992 * calling vm_page_free_toq() for each page of a list of VM objects. 3993 */ 3994 void 3995 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3996 { 3997 vm_page_t m; 3998 int count; 3999 4000 if (SLIST_EMPTY(free)) 4001 return; 4002 4003 count = 0; 4004 while ((m = SLIST_FIRST(free)) != NULL) { 4005 count++; 4006 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4007 vm_page_free_toq(m); 4008 } 4009 4010 if (update_wire_count) 4011 vm_wire_sub(count); 4012 } 4013 4014 /* 4015 * Mark this page as wired down. For managed pages, this prevents reclamation 4016 * by the page daemon, or when the containing object, if any, is destroyed. 4017 */ 4018 void 4019 vm_page_wire(vm_page_t m) 4020 { 4021 u_int old; 4022 4023 #ifdef INVARIANTS 4024 if (m->object != NULL && !vm_page_busied(m) && 4025 !vm_object_busied(m->object)) 4026 VM_OBJECT_ASSERT_LOCKED(m->object); 4027 #endif 4028 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4029 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4030 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4031 4032 old = atomic_fetchadd_int(&m->ref_count, 1); 4033 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4034 ("vm_page_wire: counter overflow for page %p", m)); 4035 if (VPRC_WIRE_COUNT(old) == 0) { 4036 if ((m->oflags & VPO_UNMANAGED) == 0) 4037 vm_page_aflag_set(m, PGA_DEQUEUE); 4038 vm_wire_add(1); 4039 } 4040 } 4041 4042 /* 4043 * Attempt to wire a mapped page following a pmap lookup of that page. 4044 * This may fail if a thread is concurrently tearing down mappings of the page. 4045 * The transient failure is acceptable because it translates to the 4046 * failure of the caller pmap_extract_and_hold(), which should be then 4047 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4048 */ 4049 bool 4050 vm_page_wire_mapped(vm_page_t m) 4051 { 4052 u_int old; 4053 4054 old = m->ref_count; 4055 do { 4056 KASSERT(old > 0, 4057 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4058 if ((old & VPRC_BLOCKED) != 0) 4059 return (false); 4060 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4061 4062 if (VPRC_WIRE_COUNT(old) == 0) { 4063 if ((m->oflags & VPO_UNMANAGED) == 0) 4064 vm_page_aflag_set(m, PGA_DEQUEUE); 4065 vm_wire_add(1); 4066 } 4067 return (true); 4068 } 4069 4070 /* 4071 * Release a wiring reference to a managed page. If the page still belongs to 4072 * an object, update its position in the page queues to reflect the reference. 4073 * If the wiring was the last reference to the page, free the page. 4074 */ 4075 static void 4076 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4077 { 4078 u_int old; 4079 4080 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4081 ("%s: page %p is unmanaged", __func__, m)); 4082 4083 /* 4084 * Update LRU state before releasing the wiring reference. 4085 * Use a release store when updating the reference count to 4086 * synchronize with vm_page_free_prep(). 4087 */ 4088 old = m->ref_count; 4089 do { 4090 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4091 ("vm_page_unwire: wire count underflow for page %p", m)); 4092 4093 if (old > VPRC_OBJREF + 1) { 4094 /* 4095 * The page has at least one other wiring reference. An 4096 * earlier iteration of this loop may have called 4097 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4098 * re-set it if necessary. 4099 */ 4100 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4101 vm_page_aflag_set(m, PGA_DEQUEUE); 4102 } else if (old == VPRC_OBJREF + 1) { 4103 /* 4104 * This is the last wiring. Clear PGA_DEQUEUE and 4105 * update the page's queue state to reflect the 4106 * reference. If the page does not belong to an object 4107 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4108 * clear leftover queue state. 4109 */ 4110 vm_page_release_toq(m, nqueue, noreuse); 4111 } else if (old == 1) { 4112 vm_page_aflag_clear(m, PGA_DEQUEUE); 4113 } 4114 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4115 4116 if (VPRC_WIRE_COUNT(old) == 1) { 4117 vm_wire_sub(1); 4118 if (old == 1) 4119 vm_page_free(m); 4120 } 4121 } 4122 4123 /* 4124 * Release one wiring of the specified page, potentially allowing it to be 4125 * paged out. 4126 * 4127 * Only managed pages belonging to an object can be paged out. If the number 4128 * of wirings transitions to zero and the page is eligible for page out, then 4129 * the page is added to the specified paging queue. If the released wiring 4130 * represented the last reference to the page, the page is freed. 4131 */ 4132 void 4133 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4134 { 4135 4136 KASSERT(nqueue < PQ_COUNT, 4137 ("vm_page_unwire: invalid queue %u request for page %p", 4138 nqueue, m)); 4139 4140 if ((m->oflags & VPO_UNMANAGED) != 0) { 4141 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4142 vm_page_free(m); 4143 return; 4144 } 4145 vm_page_unwire_managed(m, nqueue, false); 4146 } 4147 4148 /* 4149 * Unwire a page without (re-)inserting it into a page queue. It is up 4150 * to the caller to enqueue, requeue, or free the page as appropriate. 4151 * In most cases involving managed pages, vm_page_unwire() should be used 4152 * instead. 4153 */ 4154 bool 4155 vm_page_unwire_noq(vm_page_t m) 4156 { 4157 u_int old; 4158 4159 old = vm_page_drop(m, 1); 4160 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4161 ("%s: counter underflow for page %p", __func__, m)); 4162 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4163 ("%s: missing ref on fictitious page %p", __func__, m)); 4164 4165 if (VPRC_WIRE_COUNT(old) > 1) 4166 return (false); 4167 if ((m->oflags & VPO_UNMANAGED) == 0) 4168 vm_page_aflag_clear(m, PGA_DEQUEUE); 4169 vm_wire_sub(1); 4170 return (true); 4171 } 4172 4173 /* 4174 * Ensure that the page ends up in the specified page queue. If the page is 4175 * active or being moved to the active queue, ensure that its act_count is 4176 * at least ACT_INIT but do not otherwise mess with it. 4177 */ 4178 static __always_inline void 4179 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4180 { 4181 vm_page_astate_t old, new; 4182 4183 KASSERT(m->ref_count > 0, 4184 ("%s: page %p does not carry any references", __func__, m)); 4185 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4186 ("%s: invalid flags %x", __func__, nflag)); 4187 4188 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4189 return; 4190 4191 old = vm_page_astate_load(m); 4192 do { 4193 if ((old.flags & PGA_DEQUEUE) != 0) 4194 break; 4195 new = old; 4196 new.flags &= ~PGA_QUEUE_OP_MASK; 4197 if (nqueue == PQ_ACTIVE) 4198 new.act_count = max(old.act_count, ACT_INIT); 4199 if (old.queue == nqueue) { 4200 /* 4201 * There is no need to requeue pages already in the 4202 * active queue. 4203 */ 4204 if (nqueue != PQ_ACTIVE || 4205 (old.flags & PGA_ENQUEUED) == 0) 4206 new.flags |= nflag; 4207 } else { 4208 new.flags |= nflag; 4209 new.queue = nqueue; 4210 } 4211 } while (!vm_page_pqstate_commit(m, &old, new)); 4212 } 4213 4214 /* 4215 * Put the specified page on the active list (if appropriate). 4216 */ 4217 void 4218 vm_page_activate(vm_page_t m) 4219 { 4220 4221 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4222 } 4223 4224 /* 4225 * Move the specified page to the tail of the inactive queue, or requeue 4226 * the page if it is already in the inactive queue. 4227 */ 4228 void 4229 vm_page_deactivate(vm_page_t m) 4230 { 4231 4232 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4233 } 4234 4235 void 4236 vm_page_deactivate_noreuse(vm_page_t m) 4237 { 4238 4239 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4240 } 4241 4242 /* 4243 * Put a page in the laundry, or requeue it if it is already there. 4244 */ 4245 void 4246 vm_page_launder(vm_page_t m) 4247 { 4248 4249 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4250 } 4251 4252 /* 4253 * Put a page in the PQ_UNSWAPPABLE holding queue. 4254 */ 4255 void 4256 vm_page_unswappable(vm_page_t m) 4257 { 4258 4259 VM_OBJECT_ASSERT_LOCKED(m->object); 4260 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4261 ("page %p already unswappable", m)); 4262 4263 vm_page_dequeue(m); 4264 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4265 } 4266 4267 /* 4268 * Release a page back to the page queues in preparation for unwiring. 4269 */ 4270 static void 4271 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4272 { 4273 vm_page_astate_t old, new; 4274 uint16_t nflag; 4275 4276 /* 4277 * Use a check of the valid bits to determine whether we should 4278 * accelerate reclamation of the page. The object lock might not be 4279 * held here, in which case the check is racy. At worst we will either 4280 * accelerate reclamation of a valid page and violate LRU, or 4281 * unnecessarily defer reclamation of an invalid page. 4282 * 4283 * If we were asked to not cache the page, place it near the head of the 4284 * inactive queue so that is reclaimed sooner. 4285 */ 4286 if (noreuse || vm_page_none_valid(m)) { 4287 nqueue = PQ_INACTIVE; 4288 nflag = PGA_REQUEUE_HEAD; 4289 } else { 4290 nflag = PGA_REQUEUE; 4291 } 4292 4293 old = vm_page_astate_load(m); 4294 do { 4295 new = old; 4296 4297 /* 4298 * If the page is already in the active queue and we are not 4299 * trying to accelerate reclamation, simply mark it as 4300 * referenced and avoid any queue operations. 4301 */ 4302 new.flags &= ~PGA_QUEUE_OP_MASK; 4303 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4304 (old.flags & PGA_ENQUEUED) != 0) 4305 new.flags |= PGA_REFERENCED; 4306 else { 4307 new.flags |= nflag; 4308 new.queue = nqueue; 4309 } 4310 } while (!vm_page_pqstate_commit(m, &old, new)); 4311 } 4312 4313 /* 4314 * Unwire a page and either attempt to free it or re-add it to the page queues. 4315 */ 4316 void 4317 vm_page_release(vm_page_t m, int flags) 4318 { 4319 vm_object_t object; 4320 4321 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4322 ("vm_page_release: page %p is unmanaged", m)); 4323 4324 if ((flags & VPR_TRYFREE) != 0) { 4325 for (;;) { 4326 object = atomic_load_ptr(&m->object); 4327 if (object == NULL) 4328 break; 4329 /* Depends on type-stability. */ 4330 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4331 break; 4332 if (object == m->object) { 4333 vm_page_release_locked(m, flags); 4334 VM_OBJECT_WUNLOCK(object); 4335 return; 4336 } 4337 VM_OBJECT_WUNLOCK(object); 4338 } 4339 } 4340 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4341 } 4342 4343 /* See vm_page_release(). */ 4344 void 4345 vm_page_release_locked(vm_page_t m, int flags) 4346 { 4347 4348 VM_OBJECT_ASSERT_WLOCKED(m->object); 4349 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4350 ("vm_page_release_locked: page %p is unmanaged", m)); 4351 4352 if (vm_page_unwire_noq(m)) { 4353 if ((flags & VPR_TRYFREE) != 0 && 4354 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4355 m->dirty == 0 && vm_page_tryxbusy(m)) { 4356 /* 4357 * An unlocked lookup may have wired the page before the 4358 * busy lock was acquired, in which case the page must 4359 * not be freed. 4360 */ 4361 if (__predict_true(!vm_page_wired(m))) { 4362 vm_page_free(m); 4363 return; 4364 } 4365 vm_page_xunbusy(m); 4366 } else { 4367 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4368 } 4369 } 4370 } 4371 4372 static bool 4373 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4374 { 4375 u_int old; 4376 4377 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4378 ("vm_page_try_blocked_op: page %p has no object", m)); 4379 KASSERT(vm_page_busied(m), 4380 ("vm_page_try_blocked_op: page %p is not busy", m)); 4381 VM_OBJECT_ASSERT_LOCKED(m->object); 4382 4383 old = m->ref_count; 4384 do { 4385 KASSERT(old != 0, 4386 ("vm_page_try_blocked_op: page %p has no references", m)); 4387 if (VPRC_WIRE_COUNT(old) != 0) 4388 return (false); 4389 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4390 4391 (op)(m); 4392 4393 /* 4394 * If the object is read-locked, new wirings may be created via an 4395 * object lookup. 4396 */ 4397 old = vm_page_drop(m, VPRC_BLOCKED); 4398 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4399 old == (VPRC_BLOCKED | VPRC_OBJREF), 4400 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4401 old, m)); 4402 return (true); 4403 } 4404 4405 /* 4406 * Atomically check for wirings and remove all mappings of the page. 4407 */ 4408 bool 4409 vm_page_try_remove_all(vm_page_t m) 4410 { 4411 4412 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4413 } 4414 4415 /* 4416 * Atomically check for wirings and remove all writeable mappings of the page. 4417 */ 4418 bool 4419 vm_page_try_remove_write(vm_page_t m) 4420 { 4421 4422 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4423 } 4424 4425 /* 4426 * vm_page_advise 4427 * 4428 * Apply the specified advice to the given page. 4429 */ 4430 void 4431 vm_page_advise(vm_page_t m, int advice) 4432 { 4433 4434 VM_OBJECT_ASSERT_WLOCKED(m->object); 4435 vm_page_assert_xbusied(m); 4436 4437 if (advice == MADV_FREE) 4438 /* 4439 * Mark the page clean. This will allow the page to be freed 4440 * without first paging it out. MADV_FREE pages are often 4441 * quickly reused by malloc(3), so we do not do anything that 4442 * would result in a page fault on a later access. 4443 */ 4444 vm_page_undirty(m); 4445 else if (advice != MADV_DONTNEED) { 4446 if (advice == MADV_WILLNEED) 4447 vm_page_activate(m); 4448 return; 4449 } 4450 4451 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4452 vm_page_dirty(m); 4453 4454 /* 4455 * Clear any references to the page. Otherwise, the page daemon will 4456 * immediately reactivate the page. 4457 */ 4458 vm_page_aflag_clear(m, PGA_REFERENCED); 4459 4460 /* 4461 * Place clean pages near the head of the inactive queue rather than 4462 * the tail, thus defeating the queue's LRU operation and ensuring that 4463 * the page will be reused quickly. Dirty pages not already in the 4464 * laundry are moved there. 4465 */ 4466 if (m->dirty == 0) 4467 vm_page_deactivate_noreuse(m); 4468 else if (!vm_page_in_laundry(m)) 4469 vm_page_launder(m); 4470 } 4471 4472 /* 4473 * vm_page_grab_release 4474 * 4475 * Helper routine for grab functions to release busy on return. 4476 */ 4477 static inline void 4478 vm_page_grab_release(vm_page_t m, int allocflags) 4479 { 4480 4481 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4482 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4483 vm_page_sunbusy(m); 4484 else 4485 vm_page_xunbusy(m); 4486 } 4487 } 4488 4489 /* 4490 * vm_page_grab_sleep 4491 * 4492 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4493 * if the caller should retry and false otherwise. 4494 * 4495 * If the object is locked on entry the object will be unlocked with 4496 * false returns and still locked but possibly having been dropped 4497 * with true returns. 4498 */ 4499 static bool 4500 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4501 const char *wmesg, int allocflags, bool locked) 4502 { 4503 4504 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4505 return (false); 4506 4507 /* 4508 * Reference the page before unlocking and sleeping so that 4509 * the page daemon is less likely to reclaim it. 4510 */ 4511 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4512 vm_page_reference(m); 4513 4514 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4515 locked) 4516 VM_OBJECT_WLOCK(object); 4517 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4518 return (false); 4519 4520 return (true); 4521 } 4522 4523 /* 4524 * Assert that the grab flags are valid. 4525 */ 4526 static inline void 4527 vm_page_grab_check(int allocflags) 4528 { 4529 4530 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4531 (allocflags & VM_ALLOC_WIRED) != 0, 4532 ("vm_page_grab*: the pages must be busied or wired")); 4533 4534 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4535 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4536 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4537 } 4538 4539 /* 4540 * Calculate the page allocation flags for grab. 4541 */ 4542 static inline int 4543 vm_page_grab_pflags(int allocflags) 4544 { 4545 int pflags; 4546 4547 pflags = allocflags & 4548 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4549 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY); 4550 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4551 pflags |= VM_ALLOC_WAITFAIL; 4552 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4553 pflags |= VM_ALLOC_SBUSY; 4554 4555 return (pflags); 4556 } 4557 4558 /* 4559 * Grab a page, waiting until we are waken up due to the page 4560 * changing state. We keep on waiting, if the page continues 4561 * to be in the object. If the page doesn't exist, first allocate it 4562 * and then conditionally zero it. 4563 * 4564 * This routine may sleep. 4565 * 4566 * The object must be locked on entry. The lock will, however, be released 4567 * and reacquired if the routine sleeps. 4568 */ 4569 vm_page_t 4570 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4571 { 4572 vm_page_t m; 4573 4574 VM_OBJECT_ASSERT_WLOCKED(object); 4575 vm_page_grab_check(allocflags); 4576 4577 retrylookup: 4578 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4579 if (!vm_page_tryacquire(m, allocflags)) { 4580 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4581 allocflags, true)) 4582 goto retrylookup; 4583 return (NULL); 4584 } 4585 goto out; 4586 } 4587 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4588 return (NULL); 4589 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); 4590 if (m == NULL) { 4591 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4592 return (NULL); 4593 goto retrylookup; 4594 } 4595 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4596 pmap_zero_page(m); 4597 4598 out: 4599 vm_page_grab_release(m, allocflags); 4600 4601 return (m); 4602 } 4603 4604 /* 4605 * Locklessly attempt to acquire a page given a (object, pindex) tuple 4606 * and an optional previous page to avoid the radix lookup. The resulting 4607 * page will be validated against the identity tuple and busied or wired 4608 * as requested. A NULL *mp return guarantees that the page was not in 4609 * radix at the time of the call but callers must perform higher level 4610 * synchronization or retry the operation under a lock if they require 4611 * an atomic answer. This is the only lock free validation routine, 4612 * other routines can depend on the resulting page state. 4613 * 4614 * The return value indicates whether the operation failed due to caller 4615 * flags. The return is tri-state with mp: 4616 * 4617 * (true, *mp != NULL) - The operation was successful. 4618 * (true, *mp == NULL) - The page was not found in tree. 4619 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition. 4620 */ 4621 static bool 4622 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, 4623 vm_page_t prev, vm_page_t *mp, int allocflags) 4624 { 4625 vm_page_t m; 4626 4627 vm_page_grab_check(allocflags); 4628 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev)); 4629 4630 *mp = NULL; 4631 for (;;) { 4632 /* 4633 * We may see a false NULL here because the previous page 4634 * has been removed or just inserted and the list is loaded 4635 * without barriers. Switch to radix to verify. 4636 */ 4637 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL || 4638 QMD_IS_TRASHED(m) || m->pindex != pindex || 4639 atomic_load_ptr(&m->object) != object) { 4640 prev = NULL; 4641 /* 4642 * This guarantees the result is instantaneously 4643 * correct. 4644 */ 4645 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 4646 } 4647 if (m == NULL) 4648 return (true); 4649 if (vm_page_trybusy(m, allocflags)) { 4650 if (m->object == object && m->pindex == pindex) 4651 break; 4652 /* relookup. */ 4653 vm_page_busy_release(m); 4654 cpu_spinwait(); 4655 continue; 4656 } 4657 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4658 allocflags, false)) 4659 return (false); 4660 } 4661 if ((allocflags & VM_ALLOC_WIRED) != 0) 4662 vm_page_wire(m); 4663 vm_page_grab_release(m, allocflags); 4664 *mp = m; 4665 return (true); 4666 } 4667 4668 /* 4669 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4670 * is not set. 4671 */ 4672 vm_page_t 4673 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4674 { 4675 vm_page_t m; 4676 4677 vm_page_grab_check(allocflags); 4678 4679 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags)) 4680 return (NULL); 4681 if (m != NULL) 4682 return (m); 4683 4684 /* 4685 * The radix lockless lookup should never return a false negative 4686 * errors. If the user specifies NOCREAT they are guaranteed there 4687 * was no page present at the instant of the call. A NOCREAT caller 4688 * must handle create races gracefully. 4689 */ 4690 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4691 return (NULL); 4692 4693 VM_OBJECT_WLOCK(object); 4694 m = vm_page_grab(object, pindex, allocflags); 4695 VM_OBJECT_WUNLOCK(object); 4696 4697 return (m); 4698 } 4699 4700 /* 4701 * Grab a page and make it valid, paging in if necessary. Pages missing from 4702 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4703 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4704 * in simultaneously. Additional pages will be left on a paging queue but 4705 * will neither be wired nor busy regardless of allocflags. 4706 */ 4707 int 4708 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4709 { 4710 vm_page_t m; 4711 vm_page_t ma[VM_INITIAL_PAGEIN]; 4712 int after, i, pflags, rv; 4713 4714 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4715 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4716 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4717 KASSERT((allocflags & 4718 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4719 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4720 VM_OBJECT_ASSERT_WLOCKED(object); 4721 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4722 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4723 pflags |= VM_ALLOC_WAITFAIL; 4724 4725 retrylookup: 4726 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4727 /* 4728 * If the page is fully valid it can only become invalid 4729 * with the object lock held. If it is not valid it can 4730 * become valid with the busy lock held. Therefore, we 4731 * may unnecessarily lock the exclusive busy here if we 4732 * race with I/O completion not using the object lock. 4733 * However, we will not end up with an invalid page and a 4734 * shared lock. 4735 */ 4736 if (!vm_page_trybusy(m, 4737 vm_page_all_valid(m) ? allocflags : 0)) { 4738 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4739 allocflags, true); 4740 goto retrylookup; 4741 } 4742 if (vm_page_all_valid(m)) 4743 goto out; 4744 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4745 vm_page_busy_release(m); 4746 *mp = NULL; 4747 return (VM_PAGER_FAIL); 4748 } 4749 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4750 *mp = NULL; 4751 return (VM_PAGER_FAIL); 4752 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { 4753 if (!vm_pager_can_alloc_page(object, pindex)) { 4754 *mp = NULL; 4755 return (VM_PAGER_AGAIN); 4756 } 4757 goto retrylookup; 4758 } 4759 4760 vm_page_assert_xbusied(m); 4761 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4762 after = MIN(after, VM_INITIAL_PAGEIN); 4763 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4764 after = MAX(after, 1); 4765 ma[0] = m; 4766 for (i = 1; i < after; i++) { 4767 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 4768 if (vm_page_any_valid(ma[i]) || 4769 !vm_page_tryxbusy(ma[i])) 4770 break; 4771 } else { 4772 ma[i] = vm_page_alloc(object, m->pindex + i, 4773 VM_ALLOC_NORMAL); 4774 if (ma[i] == NULL) 4775 break; 4776 } 4777 } 4778 after = i; 4779 vm_object_pip_add(object, after); 4780 VM_OBJECT_WUNLOCK(object); 4781 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 4782 VM_OBJECT_WLOCK(object); 4783 vm_object_pip_wakeupn(object, after); 4784 /* Pager may have replaced a page. */ 4785 m = ma[0]; 4786 if (rv != VM_PAGER_OK) { 4787 for (i = 0; i < after; i++) { 4788 if (!vm_page_wired(ma[i])) 4789 vm_page_free(ma[i]); 4790 else 4791 vm_page_xunbusy(ma[i]); 4792 } 4793 *mp = NULL; 4794 return (rv); 4795 } 4796 for (i = 1; i < after; i++) 4797 vm_page_readahead_finish(ma[i]); 4798 MPASS(vm_page_all_valid(m)); 4799 } else { 4800 vm_page_zero_invalid(m, TRUE); 4801 } 4802 out: 4803 if ((allocflags & VM_ALLOC_WIRED) != 0) 4804 vm_page_wire(m); 4805 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 4806 vm_page_busy_downgrade(m); 4807 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 4808 vm_page_busy_release(m); 4809 *mp = m; 4810 return (VM_PAGER_OK); 4811 } 4812 4813 /* 4814 * Locklessly grab a valid page. If the page is not valid or not yet 4815 * allocated this will fall back to the object lock method. 4816 */ 4817 int 4818 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 4819 vm_pindex_t pindex, int allocflags) 4820 { 4821 vm_page_t m; 4822 int flags; 4823 int error; 4824 4825 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4826 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4827 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 4828 "mismatch")); 4829 KASSERT((allocflags & 4830 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4831 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 4832 4833 /* 4834 * Attempt a lockless lookup and busy. We need at least an sbusy 4835 * before we can inspect the valid field and return a wired page. 4836 */ 4837 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 4838 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags)) 4839 return (VM_PAGER_FAIL); 4840 if ((m = *mp) != NULL) { 4841 if (vm_page_all_valid(m)) { 4842 if ((allocflags & VM_ALLOC_WIRED) != 0) 4843 vm_page_wire(m); 4844 vm_page_grab_release(m, allocflags); 4845 return (VM_PAGER_OK); 4846 } 4847 vm_page_busy_release(m); 4848 } 4849 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4850 *mp = NULL; 4851 return (VM_PAGER_FAIL); 4852 } 4853 VM_OBJECT_WLOCK(object); 4854 error = vm_page_grab_valid(mp, object, pindex, allocflags); 4855 VM_OBJECT_WUNLOCK(object); 4856 4857 return (error); 4858 } 4859 4860 /* 4861 * Return the specified range of pages from the given object. For each 4862 * page offset within the range, if a page already exists within the object 4863 * at that offset and it is busy, then wait for it to change state. If, 4864 * instead, the page doesn't exist, then allocate it. 4865 * 4866 * The caller must always specify an allocation class. 4867 * 4868 * allocation classes: 4869 * VM_ALLOC_NORMAL normal process request 4870 * VM_ALLOC_SYSTEM system *really* needs the pages 4871 * 4872 * The caller must always specify that the pages are to be busied and/or 4873 * wired. 4874 * 4875 * optional allocation flags: 4876 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4877 * VM_ALLOC_NOBUSY do not exclusive busy the page 4878 * VM_ALLOC_NOWAIT do not sleep 4879 * VM_ALLOC_SBUSY set page to sbusy state 4880 * VM_ALLOC_WIRED wire the pages 4881 * VM_ALLOC_ZERO zero and validate any invalid pages 4882 * 4883 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4884 * may return a partial prefix of the requested range. 4885 */ 4886 int 4887 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4888 vm_page_t *ma, int count) 4889 { 4890 vm_page_t m, mpred; 4891 int pflags; 4892 int i; 4893 4894 VM_OBJECT_ASSERT_WLOCKED(object); 4895 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4896 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4897 KASSERT(count > 0, 4898 ("vm_page_grab_pages: invalid page count %d", count)); 4899 vm_page_grab_check(allocflags); 4900 4901 pflags = vm_page_grab_pflags(allocflags); 4902 i = 0; 4903 retrylookup: 4904 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4905 if (m == NULL || m->pindex != pindex + i) { 4906 mpred = m; 4907 m = NULL; 4908 } else 4909 mpred = TAILQ_PREV(m, pglist, listq); 4910 for (; i < count; i++) { 4911 if (m != NULL) { 4912 if (!vm_page_tryacquire(m, allocflags)) { 4913 if (vm_page_grab_sleep(object, m, pindex + i, 4914 "grbmaw", allocflags, true)) 4915 goto retrylookup; 4916 break; 4917 } 4918 } else { 4919 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4920 break; 4921 m = vm_page_alloc_after(object, pindex + i, 4922 pflags | VM_ALLOC_COUNT(count - i), mpred); 4923 if (m == NULL) { 4924 if ((allocflags & (VM_ALLOC_NOWAIT | 4925 VM_ALLOC_WAITFAIL)) != 0) 4926 break; 4927 goto retrylookup; 4928 } 4929 } 4930 if (vm_page_none_valid(m) && 4931 (allocflags & VM_ALLOC_ZERO) != 0) { 4932 if ((m->flags & PG_ZERO) == 0) 4933 pmap_zero_page(m); 4934 vm_page_valid(m); 4935 } 4936 vm_page_grab_release(m, allocflags); 4937 ma[i] = mpred = m; 4938 m = vm_page_next(m); 4939 } 4940 return (i); 4941 } 4942 4943 /* 4944 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 4945 * and will fall back to the locked variant to handle allocation. 4946 */ 4947 int 4948 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 4949 int allocflags, vm_page_t *ma, int count) 4950 { 4951 vm_page_t m, pred; 4952 int flags; 4953 int i; 4954 4955 KASSERT(count > 0, 4956 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 4957 vm_page_grab_check(allocflags); 4958 4959 /* 4960 * Modify flags for lockless acquire to hold the page until we 4961 * set it valid if necessary. 4962 */ 4963 flags = allocflags & ~VM_ALLOC_NOBUSY; 4964 pred = NULL; 4965 for (i = 0; i < count; i++, pindex++) { 4966 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags)) 4967 return (i); 4968 if (m == NULL) 4969 break; 4970 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 4971 if ((m->flags & PG_ZERO) == 0) 4972 pmap_zero_page(m); 4973 vm_page_valid(m); 4974 } 4975 /* m will still be wired or busy according to flags. */ 4976 vm_page_grab_release(m, allocflags); 4977 pred = ma[i] = m; 4978 } 4979 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 4980 return (i); 4981 count -= i; 4982 VM_OBJECT_WLOCK(object); 4983 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 4984 VM_OBJECT_WUNLOCK(object); 4985 4986 return (i); 4987 } 4988 4989 /* 4990 * Mapping function for valid or dirty bits in a page. 4991 * 4992 * Inputs are required to range within a page. 4993 */ 4994 vm_page_bits_t 4995 vm_page_bits(int base, int size) 4996 { 4997 int first_bit; 4998 int last_bit; 4999 5000 KASSERT( 5001 base + size <= PAGE_SIZE, 5002 ("vm_page_bits: illegal base/size %d/%d", base, size) 5003 ); 5004 5005 if (size == 0) /* handle degenerate case */ 5006 return (0); 5007 5008 first_bit = base >> DEV_BSHIFT; 5009 last_bit = (base + size - 1) >> DEV_BSHIFT; 5010 5011 return (((vm_page_bits_t)2 << last_bit) - 5012 ((vm_page_bits_t)1 << first_bit)); 5013 } 5014 5015 void 5016 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5017 { 5018 5019 #if PAGE_SIZE == 32768 5020 atomic_set_64((uint64_t *)bits, set); 5021 #elif PAGE_SIZE == 16384 5022 atomic_set_32((uint32_t *)bits, set); 5023 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5024 atomic_set_16((uint16_t *)bits, set); 5025 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5026 atomic_set_8((uint8_t *)bits, set); 5027 #else /* PAGE_SIZE <= 8192 */ 5028 uintptr_t addr; 5029 int shift; 5030 5031 addr = (uintptr_t)bits; 5032 /* 5033 * Use a trick to perform a 32-bit atomic on the 5034 * containing aligned word, to not depend on the existence 5035 * of atomic_{set, clear}_{8, 16}. 5036 */ 5037 shift = addr & (sizeof(uint32_t) - 1); 5038 #if BYTE_ORDER == BIG_ENDIAN 5039 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5040 #else 5041 shift *= NBBY; 5042 #endif 5043 addr &= ~(sizeof(uint32_t) - 1); 5044 atomic_set_32((uint32_t *)addr, set << shift); 5045 #endif /* PAGE_SIZE */ 5046 } 5047 5048 static inline void 5049 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5050 { 5051 5052 #if PAGE_SIZE == 32768 5053 atomic_clear_64((uint64_t *)bits, clear); 5054 #elif PAGE_SIZE == 16384 5055 atomic_clear_32((uint32_t *)bits, clear); 5056 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5057 atomic_clear_16((uint16_t *)bits, clear); 5058 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5059 atomic_clear_8((uint8_t *)bits, clear); 5060 #else /* PAGE_SIZE <= 8192 */ 5061 uintptr_t addr; 5062 int shift; 5063 5064 addr = (uintptr_t)bits; 5065 /* 5066 * Use a trick to perform a 32-bit atomic on the 5067 * containing aligned word, to not depend on the existence 5068 * of atomic_{set, clear}_{8, 16}. 5069 */ 5070 shift = addr & (sizeof(uint32_t) - 1); 5071 #if BYTE_ORDER == BIG_ENDIAN 5072 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5073 #else 5074 shift *= NBBY; 5075 #endif 5076 addr &= ~(sizeof(uint32_t) - 1); 5077 atomic_clear_32((uint32_t *)addr, clear << shift); 5078 #endif /* PAGE_SIZE */ 5079 } 5080 5081 static inline vm_page_bits_t 5082 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5083 { 5084 #if PAGE_SIZE == 32768 5085 uint64_t old; 5086 5087 old = *bits; 5088 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5089 return (old); 5090 #elif PAGE_SIZE == 16384 5091 uint32_t old; 5092 5093 old = *bits; 5094 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5095 return (old); 5096 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5097 uint16_t old; 5098 5099 old = *bits; 5100 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5101 return (old); 5102 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5103 uint8_t old; 5104 5105 old = *bits; 5106 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5107 return (old); 5108 #else /* PAGE_SIZE <= 4096*/ 5109 uintptr_t addr; 5110 uint32_t old, new, mask; 5111 int shift; 5112 5113 addr = (uintptr_t)bits; 5114 /* 5115 * Use a trick to perform a 32-bit atomic on the 5116 * containing aligned word, to not depend on the existence 5117 * of atomic_{set, swap, clear}_{8, 16}. 5118 */ 5119 shift = addr & (sizeof(uint32_t) - 1); 5120 #if BYTE_ORDER == BIG_ENDIAN 5121 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5122 #else 5123 shift *= NBBY; 5124 #endif 5125 addr &= ~(sizeof(uint32_t) - 1); 5126 mask = VM_PAGE_BITS_ALL << shift; 5127 5128 old = *bits; 5129 do { 5130 new = old & ~mask; 5131 new |= newbits << shift; 5132 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5133 return (old >> shift); 5134 #endif /* PAGE_SIZE */ 5135 } 5136 5137 /* 5138 * vm_page_set_valid_range: 5139 * 5140 * Sets portions of a page valid. The arguments are expected 5141 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5142 * of any partial chunks touched by the range. The invalid portion of 5143 * such chunks will be zeroed. 5144 * 5145 * (base + size) must be less then or equal to PAGE_SIZE. 5146 */ 5147 void 5148 vm_page_set_valid_range(vm_page_t m, int base, int size) 5149 { 5150 int endoff, frag; 5151 vm_page_bits_t pagebits; 5152 5153 vm_page_assert_busied(m); 5154 if (size == 0) /* handle degenerate case */ 5155 return; 5156 5157 /* 5158 * If the base is not DEV_BSIZE aligned and the valid 5159 * bit is clear, we have to zero out a portion of the 5160 * first block. 5161 */ 5162 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5163 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5164 pmap_zero_page_area(m, frag, base - frag); 5165 5166 /* 5167 * If the ending offset is not DEV_BSIZE aligned and the 5168 * valid bit is clear, we have to zero out a portion of 5169 * the last block. 5170 */ 5171 endoff = base + size; 5172 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5173 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5174 pmap_zero_page_area(m, endoff, 5175 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5176 5177 /* 5178 * Assert that no previously invalid block that is now being validated 5179 * is already dirty. 5180 */ 5181 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5182 ("vm_page_set_valid_range: page %p is dirty", m)); 5183 5184 /* 5185 * Set valid bits inclusive of any overlap. 5186 */ 5187 pagebits = vm_page_bits(base, size); 5188 if (vm_page_xbusied(m)) 5189 m->valid |= pagebits; 5190 else 5191 vm_page_bits_set(m, &m->valid, pagebits); 5192 } 5193 5194 /* 5195 * Set the page dirty bits and free the invalid swap space if 5196 * present. Returns the previous dirty bits. 5197 */ 5198 vm_page_bits_t 5199 vm_page_set_dirty(vm_page_t m) 5200 { 5201 vm_page_bits_t old; 5202 5203 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5204 5205 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5206 old = m->dirty; 5207 m->dirty = VM_PAGE_BITS_ALL; 5208 } else 5209 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5210 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5211 vm_pager_page_unswapped(m); 5212 5213 return (old); 5214 } 5215 5216 /* 5217 * Clear the given bits from the specified page's dirty field. 5218 */ 5219 static __inline void 5220 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5221 { 5222 5223 vm_page_assert_busied(m); 5224 5225 /* 5226 * If the page is xbusied and not write mapped we are the 5227 * only thread that can modify dirty bits. Otherwise, The pmap 5228 * layer can call vm_page_dirty() without holding a distinguished 5229 * lock. The combination of page busy and atomic operations 5230 * suffice to guarantee consistency of the page dirty field. 5231 */ 5232 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5233 m->dirty &= ~pagebits; 5234 else 5235 vm_page_bits_clear(m, &m->dirty, pagebits); 5236 } 5237 5238 /* 5239 * vm_page_set_validclean: 5240 * 5241 * Sets portions of a page valid and clean. The arguments are expected 5242 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5243 * of any partial chunks touched by the range. The invalid portion of 5244 * such chunks will be zero'd. 5245 * 5246 * (base + size) must be less then or equal to PAGE_SIZE. 5247 */ 5248 void 5249 vm_page_set_validclean(vm_page_t m, int base, int size) 5250 { 5251 vm_page_bits_t oldvalid, pagebits; 5252 int endoff, frag; 5253 5254 vm_page_assert_busied(m); 5255 if (size == 0) /* handle degenerate case */ 5256 return; 5257 5258 /* 5259 * If the base is not DEV_BSIZE aligned and the valid 5260 * bit is clear, we have to zero out a portion of the 5261 * first block. 5262 */ 5263 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5264 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5265 pmap_zero_page_area(m, frag, base - frag); 5266 5267 /* 5268 * If the ending offset is not DEV_BSIZE aligned and the 5269 * valid bit is clear, we have to zero out a portion of 5270 * the last block. 5271 */ 5272 endoff = base + size; 5273 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5274 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5275 pmap_zero_page_area(m, endoff, 5276 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5277 5278 /* 5279 * Set valid, clear dirty bits. If validating the entire 5280 * page we can safely clear the pmap modify bit. We also 5281 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5282 * takes a write fault on a MAP_NOSYNC memory area the flag will 5283 * be set again. 5284 * 5285 * We set valid bits inclusive of any overlap, but we can only 5286 * clear dirty bits for DEV_BSIZE chunks that are fully within 5287 * the range. 5288 */ 5289 oldvalid = m->valid; 5290 pagebits = vm_page_bits(base, size); 5291 if (vm_page_xbusied(m)) 5292 m->valid |= pagebits; 5293 else 5294 vm_page_bits_set(m, &m->valid, pagebits); 5295 #if 0 /* NOT YET */ 5296 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5297 frag = DEV_BSIZE - frag; 5298 base += frag; 5299 size -= frag; 5300 if (size < 0) 5301 size = 0; 5302 } 5303 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5304 #endif 5305 if (base == 0 && size == PAGE_SIZE) { 5306 /* 5307 * The page can only be modified within the pmap if it is 5308 * mapped, and it can only be mapped if it was previously 5309 * fully valid. 5310 */ 5311 if (oldvalid == VM_PAGE_BITS_ALL) 5312 /* 5313 * Perform the pmap_clear_modify() first. Otherwise, 5314 * a concurrent pmap operation, such as 5315 * pmap_protect(), could clear a modification in the 5316 * pmap and set the dirty field on the page before 5317 * pmap_clear_modify() had begun and after the dirty 5318 * field was cleared here. 5319 */ 5320 pmap_clear_modify(m); 5321 m->dirty = 0; 5322 vm_page_aflag_clear(m, PGA_NOSYNC); 5323 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5324 m->dirty &= ~pagebits; 5325 else 5326 vm_page_clear_dirty_mask(m, pagebits); 5327 } 5328 5329 void 5330 vm_page_clear_dirty(vm_page_t m, int base, int size) 5331 { 5332 5333 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5334 } 5335 5336 /* 5337 * vm_page_set_invalid: 5338 * 5339 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5340 * valid and dirty bits for the effected areas are cleared. 5341 */ 5342 void 5343 vm_page_set_invalid(vm_page_t m, int base, int size) 5344 { 5345 vm_page_bits_t bits; 5346 vm_object_t object; 5347 5348 /* 5349 * The object lock is required so that pages can't be mapped 5350 * read-only while we're in the process of invalidating them. 5351 */ 5352 object = m->object; 5353 VM_OBJECT_ASSERT_WLOCKED(object); 5354 vm_page_assert_busied(m); 5355 5356 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5357 size >= object->un_pager.vnp.vnp_size) 5358 bits = VM_PAGE_BITS_ALL; 5359 else 5360 bits = vm_page_bits(base, size); 5361 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5362 pmap_remove_all(m); 5363 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5364 !pmap_page_is_mapped(m), 5365 ("vm_page_set_invalid: page %p is mapped", m)); 5366 if (vm_page_xbusied(m)) { 5367 m->valid &= ~bits; 5368 m->dirty &= ~bits; 5369 } else { 5370 vm_page_bits_clear(m, &m->valid, bits); 5371 vm_page_bits_clear(m, &m->dirty, bits); 5372 } 5373 } 5374 5375 /* 5376 * vm_page_invalid: 5377 * 5378 * Invalidates the entire page. The page must be busy, unmapped, and 5379 * the enclosing object must be locked. The object locks protects 5380 * against concurrent read-only pmap enter which is done without 5381 * busy. 5382 */ 5383 void 5384 vm_page_invalid(vm_page_t m) 5385 { 5386 5387 vm_page_assert_busied(m); 5388 VM_OBJECT_ASSERT_WLOCKED(m->object); 5389 MPASS(!pmap_page_is_mapped(m)); 5390 5391 if (vm_page_xbusied(m)) 5392 m->valid = 0; 5393 else 5394 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5395 } 5396 5397 /* 5398 * vm_page_zero_invalid() 5399 * 5400 * The kernel assumes that the invalid portions of a page contain 5401 * garbage, but such pages can be mapped into memory by user code. 5402 * When this occurs, we must zero out the non-valid portions of the 5403 * page so user code sees what it expects. 5404 * 5405 * Pages are most often semi-valid when the end of a file is mapped 5406 * into memory and the file's size is not page aligned. 5407 */ 5408 void 5409 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5410 { 5411 int b; 5412 int i; 5413 5414 /* 5415 * Scan the valid bits looking for invalid sections that 5416 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5417 * valid bit may be set ) have already been zeroed by 5418 * vm_page_set_validclean(). 5419 */ 5420 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5421 if (i == (PAGE_SIZE / DEV_BSIZE) || 5422 (m->valid & ((vm_page_bits_t)1 << i))) { 5423 if (i > b) { 5424 pmap_zero_page_area(m, 5425 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5426 } 5427 b = i + 1; 5428 } 5429 } 5430 5431 /* 5432 * setvalid is TRUE when we can safely set the zero'd areas 5433 * as being valid. We can do this if there are no cache consistency 5434 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5435 */ 5436 if (setvalid) 5437 vm_page_valid(m); 5438 } 5439 5440 /* 5441 * vm_page_is_valid: 5442 * 5443 * Is (partial) page valid? Note that the case where size == 0 5444 * will return FALSE in the degenerate case where the page is 5445 * entirely invalid, and TRUE otherwise. 5446 * 5447 * Some callers envoke this routine without the busy lock held and 5448 * handle races via higher level locks. Typical callers should 5449 * hold a busy lock to prevent invalidation. 5450 */ 5451 int 5452 vm_page_is_valid(vm_page_t m, int base, int size) 5453 { 5454 vm_page_bits_t bits; 5455 5456 bits = vm_page_bits(base, size); 5457 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5458 } 5459 5460 /* 5461 * Returns true if all of the specified predicates are true for the entire 5462 * (super)page and false otherwise. 5463 */ 5464 bool 5465 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 5466 { 5467 vm_object_t object; 5468 int i, npages; 5469 5470 object = m->object; 5471 if (skip_m != NULL && skip_m->object != object) 5472 return (false); 5473 VM_OBJECT_ASSERT_LOCKED(object); 5474 npages = atop(pagesizes[m->psind]); 5475 5476 /* 5477 * The physically contiguous pages that make up a superpage, i.e., a 5478 * page with a page size index ("psind") greater than zero, will 5479 * occupy adjacent entries in vm_page_array[]. 5480 */ 5481 for (i = 0; i < npages; i++) { 5482 /* Always test object consistency, including "skip_m". */ 5483 if (m[i].object != object) 5484 return (false); 5485 if (&m[i] == skip_m) 5486 continue; 5487 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5488 return (false); 5489 if ((flags & PS_ALL_DIRTY) != 0) { 5490 /* 5491 * Calling vm_page_test_dirty() or pmap_is_modified() 5492 * might stop this case from spuriously returning 5493 * "false". However, that would require a write lock 5494 * on the object containing "m[i]". 5495 */ 5496 if (m[i].dirty != VM_PAGE_BITS_ALL) 5497 return (false); 5498 } 5499 if ((flags & PS_ALL_VALID) != 0 && 5500 m[i].valid != VM_PAGE_BITS_ALL) 5501 return (false); 5502 } 5503 return (true); 5504 } 5505 5506 /* 5507 * Set the page's dirty bits if the page is modified. 5508 */ 5509 void 5510 vm_page_test_dirty(vm_page_t m) 5511 { 5512 5513 vm_page_assert_busied(m); 5514 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5515 vm_page_dirty(m); 5516 } 5517 5518 void 5519 vm_page_valid(vm_page_t m) 5520 { 5521 5522 vm_page_assert_busied(m); 5523 if (vm_page_xbusied(m)) 5524 m->valid = VM_PAGE_BITS_ALL; 5525 else 5526 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5527 } 5528 5529 void 5530 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5531 { 5532 5533 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5534 } 5535 5536 void 5537 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5538 { 5539 5540 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5541 } 5542 5543 int 5544 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5545 { 5546 5547 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5548 } 5549 5550 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5551 void 5552 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5553 { 5554 5555 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5556 } 5557 5558 void 5559 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5560 { 5561 5562 mtx_assert_(vm_page_lockptr(m), a, file, line); 5563 } 5564 #endif 5565 5566 #ifdef INVARIANTS 5567 void 5568 vm_page_object_busy_assert(vm_page_t m) 5569 { 5570 5571 /* 5572 * Certain of the page's fields may only be modified by the 5573 * holder of a page or object busy. 5574 */ 5575 if (m->object != NULL && !vm_page_busied(m)) 5576 VM_OBJECT_ASSERT_BUSY(m->object); 5577 } 5578 5579 void 5580 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5581 { 5582 5583 if ((bits & PGA_WRITEABLE) == 0) 5584 return; 5585 5586 /* 5587 * The PGA_WRITEABLE flag can only be set if the page is 5588 * managed, is exclusively busied or the object is locked. 5589 * Currently, this flag is only set by pmap_enter(). 5590 */ 5591 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5592 ("PGA_WRITEABLE on unmanaged page")); 5593 if (!vm_page_xbusied(m)) 5594 VM_OBJECT_ASSERT_BUSY(m->object); 5595 } 5596 #endif 5597 5598 #include "opt_ddb.h" 5599 #ifdef DDB 5600 #include <sys/kernel.h> 5601 5602 #include <ddb/ddb.h> 5603 5604 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5605 { 5606 5607 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5608 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5609 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5610 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5611 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5612 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5613 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5614 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5615 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5616 } 5617 5618 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5619 { 5620 int dom; 5621 5622 db_printf("pq_free %d\n", vm_free_count()); 5623 for (dom = 0; dom < vm_ndomains; dom++) { 5624 db_printf( 5625 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5626 dom, 5627 vm_dom[dom].vmd_page_count, 5628 vm_dom[dom].vmd_free_count, 5629 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5630 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5631 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5632 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5633 } 5634 } 5635 5636 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5637 { 5638 vm_page_t m; 5639 boolean_t phys, virt; 5640 5641 if (!have_addr) { 5642 db_printf("show pginfo addr\n"); 5643 return; 5644 } 5645 5646 phys = strchr(modif, 'p') != NULL; 5647 virt = strchr(modif, 'v') != NULL; 5648 if (virt) 5649 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5650 else if (phys) 5651 m = PHYS_TO_VM_PAGE(addr); 5652 else 5653 m = (vm_page_t)addr; 5654 db_printf( 5655 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5656 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5657 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5658 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5659 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5660 } 5661 #endif /* DDB */ 5662