1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * The Mach Operating System project at Carnegie-Mellon University. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 * 35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 36 */ 37 38 /*- 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 */ 64 65 /* 66 * Resident memory management module. 67 */ 68 69 #include <sys/cdefs.h> 70 __FBSDID("$FreeBSD$"); 71 72 #include "opt_vm.h" 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/counter.h> 77 #include <sys/domainset.h> 78 #include <sys/kernel.h> 79 #include <sys/limits.h> 80 #include <sys/linker.h> 81 #include <sys/lock.h> 82 #include <sys/malloc.h> 83 #include <sys/mman.h> 84 #include <sys/msgbuf.h> 85 #include <sys/mutex.h> 86 #include <sys/proc.h> 87 #include <sys/rwlock.h> 88 #include <sys/sleepqueue.h> 89 #include <sys/sbuf.h> 90 #include <sys/sched.h> 91 #include <sys/smp.h> 92 #include <sys/sysctl.h> 93 #include <sys/vmmeter.h> 94 #include <sys/vnode.h> 95 96 #include <vm/vm.h> 97 #include <vm/pmap.h> 98 #include <vm/vm_param.h> 99 #include <vm/vm_domainset.h> 100 #include <vm/vm_kern.h> 101 #include <vm/vm_map.h> 102 #include <vm/vm_object.h> 103 #include <vm/vm_page.h> 104 #include <vm/vm_pageout.h> 105 #include <vm/vm_phys.h> 106 #include <vm/vm_pagequeue.h> 107 #include <vm/vm_pager.h> 108 #include <vm/vm_radix.h> 109 #include <vm/vm_reserv.h> 110 #include <vm/vm_extern.h> 111 #include <vm/vm_dumpset.h> 112 #include <vm/uma.h> 113 #include <vm/uma_int.h> 114 115 #include <machine/md_var.h> 116 117 struct vm_domain vm_dom[MAXMEMDOM]; 118 119 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 120 121 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 122 123 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 124 /* The following fields are protected by the domainset lock. */ 125 domainset_t __exclusive_cache_line vm_min_domains; 126 domainset_t __exclusive_cache_line vm_severe_domains; 127 static int vm_min_waiters; 128 static int vm_severe_waiters; 129 static int vm_pageproc_waiters; 130 131 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 132 "VM page statistics"); 133 134 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 135 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 136 CTLFLAG_RD, &pqstate_commit_retries, 137 "Number of failed per-page atomic queue state updates"); 138 139 static COUNTER_U64_DEFINE_EARLY(queue_ops); 140 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 141 CTLFLAG_RD, &queue_ops, 142 "Number of batched queue operations"); 143 144 static COUNTER_U64_DEFINE_EARLY(queue_nops); 145 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 146 CTLFLAG_RD, &queue_nops, 147 "Number of batched queue operations with no effects"); 148 149 /* 150 * bogus page -- for I/O to/from partially complete buffers, 151 * or for paging into sparsely invalid regions. 152 */ 153 vm_page_t bogus_page; 154 155 vm_page_t vm_page_array; 156 long vm_page_array_size; 157 long first_page; 158 159 struct bitset *vm_page_dump; 160 long vm_page_dump_pages; 161 162 static TAILQ_HEAD(, vm_page) blacklist_head; 163 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 164 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 165 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 166 167 static uma_zone_t fakepg_zone; 168 169 static void vm_page_alloc_check(vm_page_t m); 170 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 171 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 172 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 173 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 174 static bool vm_page_free_prep(vm_page_t m); 175 static void vm_page_free_toq(vm_page_t m); 176 static void vm_page_init(void *dummy); 177 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 178 vm_pindex_t pindex, vm_page_t mpred); 179 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 180 vm_page_t mpred); 181 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 182 const uint16_t nflag); 183 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 184 vm_page_t m_run, vm_paddr_t high); 185 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 187 int req); 188 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 189 int flags); 190 static void vm_page_zone_release(void *arg, void **store, int cnt); 191 192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 193 194 static void 195 vm_page_init(void *dummy) 196 { 197 198 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 199 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 200 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED); 201 } 202 203 /* 204 * The cache page zone is initialized later since we need to be able to allocate 205 * pages before UMA is fully initialized. 206 */ 207 static void 208 vm_page_init_cache_zones(void *dummy __unused) 209 { 210 struct vm_domain *vmd; 211 struct vm_pgcache *pgcache; 212 int cache, domain, maxcache, pool; 213 214 maxcache = 0; 215 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache); 216 maxcache *= mp_ncpus; 217 for (domain = 0; domain < vm_ndomains; domain++) { 218 vmd = VM_DOMAIN(domain); 219 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 220 pgcache = &vmd->vmd_pgcache[pool]; 221 pgcache->domain = domain; 222 pgcache->pool = pool; 223 pgcache->zone = uma_zcache_create("vm pgcache", 224 PAGE_SIZE, NULL, NULL, NULL, NULL, 225 vm_page_zone_import, vm_page_zone_release, pgcache, 226 UMA_ZONE_VM); 227 228 /* 229 * Limit each pool's zone to 0.1% of the pages in the 230 * domain. 231 */ 232 cache = maxcache != 0 ? maxcache : 233 vmd->vmd_page_count / 1000; 234 uma_zone_set_maxcache(pgcache->zone, cache); 235 } 236 } 237 } 238 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 239 240 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 241 #if PAGE_SIZE == 32768 242 #ifdef CTASSERT 243 CTASSERT(sizeof(u_long) >= 8); 244 #endif 245 #endif 246 247 /* 248 * vm_set_page_size: 249 * 250 * Sets the page size, perhaps based upon the memory 251 * size. Must be called before any use of page-size 252 * dependent functions. 253 */ 254 void 255 vm_set_page_size(void) 256 { 257 if (vm_cnt.v_page_size == 0) 258 vm_cnt.v_page_size = PAGE_SIZE; 259 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 260 panic("vm_set_page_size: page size not a power of two"); 261 } 262 263 /* 264 * vm_page_blacklist_next: 265 * 266 * Find the next entry in the provided string of blacklist 267 * addresses. Entries are separated by space, comma, or newline. 268 * If an invalid integer is encountered then the rest of the 269 * string is skipped. Updates the list pointer to the next 270 * character, or NULL if the string is exhausted or invalid. 271 */ 272 static vm_paddr_t 273 vm_page_blacklist_next(char **list, char *end) 274 { 275 vm_paddr_t bad; 276 char *cp, *pos; 277 278 if (list == NULL || *list == NULL) 279 return (0); 280 if (**list =='\0') { 281 *list = NULL; 282 return (0); 283 } 284 285 /* 286 * If there's no end pointer then the buffer is coming from 287 * the kenv and we know it's null-terminated. 288 */ 289 if (end == NULL) 290 end = *list + strlen(*list); 291 292 /* Ensure that strtoq() won't walk off the end */ 293 if (*end != '\0') { 294 if (*end == '\n' || *end == ' ' || *end == ',') 295 *end = '\0'; 296 else { 297 printf("Blacklist not terminated, skipping\n"); 298 *list = NULL; 299 return (0); 300 } 301 } 302 303 for (pos = *list; *pos != '\0'; pos = cp) { 304 bad = strtoq(pos, &cp, 0); 305 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 306 if (bad == 0) { 307 if (++cp < end) 308 continue; 309 else 310 break; 311 } 312 } else 313 break; 314 if (*cp == '\0' || ++cp >= end) 315 *list = NULL; 316 else 317 *list = cp; 318 return (trunc_page(bad)); 319 } 320 printf("Garbage in RAM blacklist, skipping\n"); 321 *list = NULL; 322 return (0); 323 } 324 325 bool 326 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 327 { 328 struct vm_domain *vmd; 329 vm_page_t m; 330 int ret; 331 332 m = vm_phys_paddr_to_vm_page(pa); 333 if (m == NULL) 334 return (true); /* page does not exist, no failure */ 335 336 vmd = vm_pagequeue_domain(m); 337 vm_domain_free_lock(vmd); 338 ret = vm_phys_unfree_page(m); 339 vm_domain_free_unlock(vmd); 340 if (ret != 0) { 341 vm_domain_freecnt_inc(vmd, -1); 342 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 343 if (verbose) 344 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 345 } 346 return (ret); 347 } 348 349 /* 350 * vm_page_blacklist_check: 351 * 352 * Iterate through the provided string of blacklist addresses, pulling 353 * each entry out of the physical allocator free list and putting it 354 * onto a list for reporting via the vm.page_blacklist sysctl. 355 */ 356 static void 357 vm_page_blacklist_check(char *list, char *end) 358 { 359 vm_paddr_t pa; 360 char *next; 361 362 next = list; 363 while (next != NULL) { 364 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 365 continue; 366 vm_page_blacklist_add(pa, bootverbose); 367 } 368 } 369 370 /* 371 * vm_page_blacklist_load: 372 * 373 * Search for a special module named "ram_blacklist". It'll be a 374 * plain text file provided by the user via the loader directive 375 * of the same name. 376 */ 377 static void 378 vm_page_blacklist_load(char **list, char **end) 379 { 380 void *mod; 381 u_char *ptr; 382 u_int len; 383 384 mod = NULL; 385 ptr = NULL; 386 387 mod = preload_search_by_type("ram_blacklist"); 388 if (mod != NULL) { 389 ptr = preload_fetch_addr(mod); 390 len = preload_fetch_size(mod); 391 } 392 *list = ptr; 393 if (ptr != NULL) 394 *end = ptr + len; 395 else 396 *end = NULL; 397 return; 398 } 399 400 static int 401 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 402 { 403 vm_page_t m; 404 struct sbuf sbuf; 405 int error, first; 406 407 first = 1; 408 error = sysctl_wire_old_buffer(req, 0); 409 if (error != 0) 410 return (error); 411 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 412 TAILQ_FOREACH(m, &blacklist_head, listq) { 413 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 414 (uintmax_t)m->phys_addr); 415 first = 0; 416 } 417 error = sbuf_finish(&sbuf); 418 sbuf_delete(&sbuf); 419 return (error); 420 } 421 422 /* 423 * Initialize a dummy page for use in scans of the specified paging queue. 424 * In principle, this function only needs to set the flag PG_MARKER. 425 * Nonetheless, it write busies the page as a safety precaution. 426 */ 427 void 428 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 429 { 430 431 bzero(marker, sizeof(*marker)); 432 marker->flags = PG_MARKER; 433 marker->a.flags = aflags; 434 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 435 marker->a.queue = queue; 436 } 437 438 static void 439 vm_page_domain_init(int domain) 440 { 441 struct vm_domain *vmd; 442 struct vm_pagequeue *pq; 443 int i; 444 445 vmd = VM_DOMAIN(domain); 446 bzero(vmd, sizeof(*vmd)); 447 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 448 "vm inactive pagequeue"; 449 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 450 "vm active pagequeue"; 451 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 452 "vm laundry pagequeue"; 453 *__DECONST(const char **, 454 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 455 "vm unswappable pagequeue"; 456 vmd->vmd_domain = domain; 457 vmd->vmd_page_count = 0; 458 vmd->vmd_free_count = 0; 459 vmd->vmd_segs = 0; 460 vmd->vmd_oom = FALSE; 461 for (i = 0; i < PQ_COUNT; i++) { 462 pq = &vmd->vmd_pagequeues[i]; 463 TAILQ_INIT(&pq->pq_pl); 464 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 465 MTX_DEF | MTX_DUPOK); 466 pq->pq_pdpages = 0; 467 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 468 } 469 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 470 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 471 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 472 473 /* 474 * inacthead is used to provide FIFO ordering for LRU-bypassing 475 * insertions. 476 */ 477 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 478 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 479 &vmd->vmd_inacthead, plinks.q); 480 481 /* 482 * The clock pages are used to implement active queue scanning without 483 * requeues. Scans start at clock[0], which is advanced after the scan 484 * ends. When the two clock hands meet, they are reset and scanning 485 * resumes from the head of the queue. 486 */ 487 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 488 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 489 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 490 &vmd->vmd_clock[0], plinks.q); 491 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 492 &vmd->vmd_clock[1], plinks.q); 493 } 494 495 /* 496 * Initialize a physical page in preparation for adding it to the free 497 * lists. 498 */ 499 void 500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 501 { 502 503 m->object = NULL; 504 m->ref_count = 0; 505 m->busy_lock = VPB_FREED; 506 m->flags = m->a.flags = 0; 507 m->phys_addr = pa; 508 m->a.queue = PQ_NONE; 509 m->psind = 0; 510 m->segind = segind; 511 m->order = VM_NFREEORDER; 512 m->pool = VM_FREEPOOL_DEFAULT; 513 m->valid = m->dirty = 0; 514 pmap_page_init(m); 515 } 516 517 #ifndef PMAP_HAS_PAGE_ARRAY 518 static vm_paddr_t 519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 520 { 521 vm_paddr_t new_end; 522 523 /* 524 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 525 * However, because this page is allocated from KVM, out-of-bounds 526 * accesses using the direct map will not be trapped. 527 */ 528 *vaddr += PAGE_SIZE; 529 530 /* 531 * Allocate physical memory for the page structures, and map it. 532 */ 533 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 534 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 535 VM_PROT_READ | VM_PROT_WRITE); 536 vm_page_array_size = page_range; 537 538 return (new_end); 539 } 540 #endif 541 542 /* 543 * vm_page_startup: 544 * 545 * Initializes the resident memory module. Allocates physical memory for 546 * bootstrapping UMA and some data structures that are used to manage 547 * physical pages. Initializes these structures, and populates the free 548 * page queues. 549 */ 550 vm_offset_t 551 vm_page_startup(vm_offset_t vaddr) 552 { 553 struct vm_phys_seg *seg; 554 struct vm_domain *vmd; 555 vm_page_t m; 556 char *list, *listend; 557 vm_paddr_t end, high_avail, low_avail, new_end, size; 558 vm_paddr_t page_range __unused; 559 vm_paddr_t last_pa, pa, startp, endp; 560 u_long pagecount; 561 #if MINIDUMP_PAGE_TRACKING 562 u_long vm_page_dump_size; 563 #endif 564 int biggestone, i, segind; 565 #ifdef WITNESS 566 vm_offset_t mapped; 567 int witness_size; 568 #endif 569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 570 long ii; 571 #endif 572 573 vaddr = round_page(vaddr); 574 575 vm_phys_early_startup(); 576 biggestone = vm_phys_avail_largest(); 577 end = phys_avail[biggestone+1]; 578 579 /* 580 * Initialize the page and queue locks. 581 */ 582 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 583 for (i = 0; i < PA_LOCK_COUNT; i++) 584 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 585 for (i = 0; i < vm_ndomains; i++) 586 vm_page_domain_init(i); 587 588 new_end = end; 589 #ifdef WITNESS 590 witness_size = round_page(witness_startup_count()); 591 new_end -= witness_size; 592 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 593 VM_PROT_READ | VM_PROT_WRITE); 594 bzero((void *)mapped, witness_size); 595 witness_startup((void *)mapped); 596 #endif 597 598 #if MINIDUMP_PAGE_TRACKING 599 /* 600 * Allocate a bitmap to indicate that a random physical page 601 * needs to be included in a minidump. 602 * 603 * The amd64 port needs this to indicate which direct map pages 604 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 605 * 606 * However, i386 still needs this workspace internally within the 607 * minidump code. In theory, they are not needed on i386, but are 608 * included should the sf_buf code decide to use them. 609 */ 610 last_pa = 0; 611 vm_page_dump_pages = 0; 612 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 613 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 614 dump_avail[i] / PAGE_SIZE; 615 if (dump_avail[i + 1] > last_pa) 616 last_pa = dump_avail[i + 1]; 617 } 618 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 619 new_end -= vm_page_dump_size; 620 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 621 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 622 bzero((void *)vm_page_dump, vm_page_dump_size); 623 #else 624 (void)last_pa; 625 #endif 626 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 627 defined(__riscv) || defined(__powerpc64__) 628 /* 629 * Include the UMA bootstrap pages, witness pages and vm_page_dump 630 * in a crash dump. When pmap_map() uses the direct map, they are 631 * not automatically included. 632 */ 633 for (pa = new_end; pa < end; pa += PAGE_SIZE) 634 dump_add_page(pa); 635 #endif 636 phys_avail[biggestone + 1] = new_end; 637 #ifdef __amd64__ 638 /* 639 * Request that the physical pages underlying the message buffer be 640 * included in a crash dump. Since the message buffer is accessed 641 * through the direct map, they are not automatically included. 642 */ 643 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 644 last_pa = pa + round_page(msgbufsize); 645 while (pa < last_pa) { 646 dump_add_page(pa); 647 pa += PAGE_SIZE; 648 } 649 #endif 650 /* 651 * Compute the number of pages of memory that will be available for 652 * use, taking into account the overhead of a page structure per page. 653 * In other words, solve 654 * "available physical memory" - round_page(page_range * 655 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 656 * for page_range. 657 */ 658 low_avail = phys_avail[0]; 659 high_avail = phys_avail[1]; 660 for (i = 0; i < vm_phys_nsegs; i++) { 661 if (vm_phys_segs[i].start < low_avail) 662 low_avail = vm_phys_segs[i].start; 663 if (vm_phys_segs[i].end > high_avail) 664 high_avail = vm_phys_segs[i].end; 665 } 666 /* Skip the first chunk. It is already accounted for. */ 667 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 668 if (phys_avail[i] < low_avail) 669 low_avail = phys_avail[i]; 670 if (phys_avail[i + 1] > high_avail) 671 high_avail = phys_avail[i + 1]; 672 } 673 first_page = low_avail / PAGE_SIZE; 674 #ifdef VM_PHYSSEG_SPARSE 675 size = 0; 676 for (i = 0; i < vm_phys_nsegs; i++) 677 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 678 for (i = 0; phys_avail[i + 1] != 0; i += 2) 679 size += phys_avail[i + 1] - phys_avail[i]; 680 #elif defined(VM_PHYSSEG_DENSE) 681 size = high_avail - low_avail; 682 #else 683 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 684 #endif 685 686 #ifdef PMAP_HAS_PAGE_ARRAY 687 pmap_page_array_startup(size / PAGE_SIZE); 688 biggestone = vm_phys_avail_largest(); 689 end = new_end = phys_avail[biggestone + 1]; 690 #else 691 #ifdef VM_PHYSSEG_DENSE 692 /* 693 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 694 * the overhead of a page structure per page only if vm_page_array is 695 * allocated from the last physical memory chunk. Otherwise, we must 696 * allocate page structures representing the physical memory 697 * underlying vm_page_array, even though they will not be used. 698 */ 699 if (new_end != high_avail) 700 page_range = size / PAGE_SIZE; 701 else 702 #endif 703 { 704 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 705 706 /* 707 * If the partial bytes remaining are large enough for 708 * a page (PAGE_SIZE) without a corresponding 709 * 'struct vm_page', then new_end will contain an 710 * extra page after subtracting the length of the VM 711 * page array. Compensate by subtracting an extra 712 * page from new_end. 713 */ 714 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 715 if (new_end == high_avail) 716 high_avail -= PAGE_SIZE; 717 new_end -= PAGE_SIZE; 718 } 719 } 720 end = new_end; 721 new_end = vm_page_array_alloc(&vaddr, end, page_range); 722 #endif 723 724 #if VM_NRESERVLEVEL > 0 725 /* 726 * Allocate physical memory for the reservation management system's 727 * data structures, and map it. 728 */ 729 new_end = vm_reserv_startup(&vaddr, new_end); 730 #endif 731 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 732 defined(__riscv) || defined(__powerpc64__) 733 /* 734 * Include vm_page_array and vm_reserv_array in a crash dump. 735 */ 736 for (pa = new_end; pa < end; pa += PAGE_SIZE) 737 dump_add_page(pa); 738 #endif 739 phys_avail[biggestone + 1] = new_end; 740 741 /* 742 * Add physical memory segments corresponding to the available 743 * physical pages. 744 */ 745 for (i = 0; phys_avail[i + 1] != 0; i += 2) 746 if (vm_phys_avail_size(i) != 0) 747 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 748 749 /* 750 * Initialize the physical memory allocator. 751 */ 752 vm_phys_init(); 753 754 /* 755 * Initialize the page structures and add every available page to the 756 * physical memory allocator's free lists. 757 */ 758 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 759 for (ii = 0; ii < vm_page_array_size; ii++) { 760 m = &vm_page_array[ii]; 761 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 762 m->flags = PG_FICTITIOUS; 763 } 764 #endif 765 vm_cnt.v_page_count = 0; 766 for (segind = 0; segind < vm_phys_nsegs; segind++) { 767 seg = &vm_phys_segs[segind]; 768 for (m = seg->first_page, pa = seg->start; pa < seg->end; 769 m++, pa += PAGE_SIZE) 770 vm_page_init_page(m, pa, segind); 771 772 /* 773 * Add the segment's pages that are covered by one of 774 * phys_avail's ranges to the free lists. 775 */ 776 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 777 if (seg->end <= phys_avail[i] || 778 seg->start >= phys_avail[i + 1]) 779 continue; 780 781 startp = MAX(seg->start, phys_avail[i]); 782 endp = MIN(seg->end, phys_avail[i + 1]); 783 pagecount = (u_long)atop(endp - startp); 784 if (pagecount == 0) 785 continue; 786 787 m = seg->first_page + atop(startp - seg->start); 788 vmd = VM_DOMAIN(seg->domain); 789 vm_domain_free_lock(vmd); 790 vm_phys_enqueue_contig(m, pagecount); 791 vm_domain_free_unlock(vmd); 792 vm_domain_freecnt_inc(vmd, pagecount); 793 vm_cnt.v_page_count += (u_int)pagecount; 794 vmd->vmd_page_count += (u_int)pagecount; 795 vmd->vmd_segs |= 1UL << segind; 796 } 797 } 798 799 /* 800 * Remove blacklisted pages from the physical memory allocator. 801 */ 802 TAILQ_INIT(&blacklist_head); 803 vm_page_blacklist_load(&list, &listend); 804 vm_page_blacklist_check(list, listend); 805 806 list = kern_getenv("vm.blacklist"); 807 vm_page_blacklist_check(list, NULL); 808 809 freeenv(list); 810 #if VM_NRESERVLEVEL > 0 811 /* 812 * Initialize the reservation management system. 813 */ 814 vm_reserv_init(); 815 #endif 816 817 return (vaddr); 818 } 819 820 void 821 vm_page_reference(vm_page_t m) 822 { 823 824 vm_page_aflag_set(m, PGA_REFERENCED); 825 } 826 827 /* 828 * vm_page_trybusy 829 * 830 * Helper routine for grab functions to trylock busy. 831 * 832 * Returns true on success and false on failure. 833 */ 834 static bool 835 vm_page_trybusy(vm_page_t m, int allocflags) 836 { 837 838 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 839 return (vm_page_trysbusy(m)); 840 else 841 return (vm_page_tryxbusy(m)); 842 } 843 844 /* 845 * vm_page_tryacquire 846 * 847 * Helper routine for grab functions to trylock busy and wire. 848 * 849 * Returns true on success and false on failure. 850 */ 851 static inline bool 852 vm_page_tryacquire(vm_page_t m, int allocflags) 853 { 854 bool locked; 855 856 locked = vm_page_trybusy(m, allocflags); 857 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 858 vm_page_wire(m); 859 return (locked); 860 } 861 862 /* 863 * vm_page_busy_acquire: 864 * 865 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 866 * and drop the object lock if necessary. 867 */ 868 bool 869 vm_page_busy_acquire(vm_page_t m, int allocflags) 870 { 871 vm_object_t obj; 872 bool locked; 873 874 /* 875 * The page-specific object must be cached because page 876 * identity can change during the sleep, causing the 877 * re-lock of a different object. 878 * It is assumed that a reference to the object is already 879 * held by the callers. 880 */ 881 obj = atomic_load_ptr(&m->object); 882 for (;;) { 883 if (vm_page_tryacquire(m, allocflags)) 884 return (true); 885 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 886 return (false); 887 if (obj != NULL) 888 locked = VM_OBJECT_WOWNED(obj); 889 else 890 locked = false; 891 MPASS(locked || vm_page_wired(m)); 892 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 893 locked) && locked) 894 VM_OBJECT_WLOCK(obj); 895 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 896 return (false); 897 KASSERT(m->object == obj || m->object == NULL, 898 ("vm_page_busy_acquire: page %p does not belong to %p", 899 m, obj)); 900 } 901 } 902 903 /* 904 * vm_page_busy_downgrade: 905 * 906 * Downgrade an exclusive busy page into a single shared busy page. 907 */ 908 void 909 vm_page_busy_downgrade(vm_page_t m) 910 { 911 u_int x; 912 913 vm_page_assert_xbusied(m); 914 915 x = vm_page_busy_fetch(m); 916 for (;;) { 917 if (atomic_fcmpset_rel_int(&m->busy_lock, 918 &x, VPB_SHARERS_WORD(1))) 919 break; 920 } 921 if ((x & VPB_BIT_WAITERS) != 0) 922 wakeup(m); 923 } 924 925 /* 926 * 927 * vm_page_busy_tryupgrade: 928 * 929 * Attempt to upgrade a single shared busy into an exclusive busy. 930 */ 931 int 932 vm_page_busy_tryupgrade(vm_page_t m) 933 { 934 u_int ce, x; 935 936 vm_page_assert_sbusied(m); 937 938 x = vm_page_busy_fetch(m); 939 ce = VPB_CURTHREAD_EXCLUSIVE; 940 for (;;) { 941 if (VPB_SHARERS(x) > 1) 942 return (0); 943 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 944 ("vm_page_busy_tryupgrade: invalid lock state")); 945 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 946 ce | (x & VPB_BIT_WAITERS))) 947 continue; 948 return (1); 949 } 950 } 951 952 /* 953 * vm_page_sbusied: 954 * 955 * Return a positive value if the page is shared busied, 0 otherwise. 956 */ 957 int 958 vm_page_sbusied(vm_page_t m) 959 { 960 u_int x; 961 962 x = vm_page_busy_fetch(m); 963 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 964 } 965 966 /* 967 * vm_page_sunbusy: 968 * 969 * Shared unbusy a page. 970 */ 971 void 972 vm_page_sunbusy(vm_page_t m) 973 { 974 u_int x; 975 976 vm_page_assert_sbusied(m); 977 978 x = vm_page_busy_fetch(m); 979 for (;;) { 980 KASSERT(x != VPB_FREED, 981 ("vm_page_sunbusy: Unlocking freed page.")); 982 if (VPB_SHARERS(x) > 1) { 983 if (atomic_fcmpset_int(&m->busy_lock, &x, 984 x - VPB_ONE_SHARER)) 985 break; 986 continue; 987 } 988 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 989 ("vm_page_sunbusy: invalid lock state")); 990 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 991 continue; 992 if ((x & VPB_BIT_WAITERS) == 0) 993 break; 994 wakeup(m); 995 break; 996 } 997 } 998 999 /* 1000 * vm_page_busy_sleep: 1001 * 1002 * Sleep if the page is busy, using the page pointer as wchan. 1003 * This is used to implement the hard-path of busying mechanism. 1004 * 1005 * If nonshared is true, sleep only if the page is xbusy. 1006 * 1007 * The object lock must be held on entry and will be released on exit. 1008 */ 1009 void 1010 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 1011 { 1012 vm_object_t obj; 1013 1014 obj = m->object; 1015 VM_OBJECT_ASSERT_LOCKED(obj); 1016 vm_page_lock_assert(m, MA_NOTOWNED); 1017 1018 if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 1019 nonshared ? VM_ALLOC_SBUSY : 0 , true)) 1020 VM_OBJECT_DROP(obj); 1021 } 1022 1023 /* 1024 * vm_page_busy_sleep_unlocked: 1025 * 1026 * Sleep if the page is busy, using the page pointer as wchan. 1027 * This is used to implement the hard-path of busying mechanism. 1028 * 1029 * If nonshared is true, sleep only if the page is xbusy. 1030 * 1031 * The object lock must not be held on entry. The operation will 1032 * return if the page changes identity. 1033 */ 1034 void 1035 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1036 const char *wmesg, bool nonshared) 1037 { 1038 1039 VM_OBJECT_ASSERT_UNLOCKED(obj); 1040 vm_page_lock_assert(m, MA_NOTOWNED); 1041 1042 _vm_page_busy_sleep(obj, m, pindex, wmesg, 1043 nonshared ? VM_ALLOC_SBUSY : 0, 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 * If locked is true the lock will be dropped for any true returns 1054 * and held for any false returns. 1055 */ 1056 static bool 1057 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1058 const char *wmesg, int allocflags, bool locked) 1059 { 1060 bool xsleep; 1061 u_int x; 1062 1063 /* 1064 * If the object is busy we must wait for that to drain to zero 1065 * before trying the page again. 1066 */ 1067 if (obj != NULL && vm_object_busied(obj)) { 1068 if (locked) 1069 VM_OBJECT_DROP(obj); 1070 vm_object_busy_wait(obj, wmesg); 1071 return (true); 1072 } 1073 1074 if (!vm_page_busied(m)) 1075 return (false); 1076 1077 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1078 sleepq_lock(m); 1079 x = vm_page_busy_fetch(m); 1080 do { 1081 /* 1082 * If the page changes objects or becomes unlocked we can 1083 * simply return. 1084 */ 1085 if (x == VPB_UNBUSIED || 1086 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1087 m->object != obj || m->pindex != pindex) { 1088 sleepq_release(m); 1089 return (false); 1090 } 1091 if ((x & VPB_BIT_WAITERS) != 0) 1092 break; 1093 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1094 if (locked) 1095 VM_OBJECT_DROP(obj); 1096 DROP_GIANT(); 1097 sleepq_add(m, NULL, wmesg, 0, 0); 1098 sleepq_wait(m, PVM); 1099 PICKUP_GIANT(); 1100 return (true); 1101 } 1102 1103 /* 1104 * vm_page_trysbusy: 1105 * 1106 * Try to shared busy a page. 1107 * If the operation succeeds 1 is returned otherwise 0. 1108 * The operation never sleeps. 1109 */ 1110 int 1111 vm_page_trysbusy(vm_page_t m) 1112 { 1113 vm_object_t obj; 1114 u_int x; 1115 1116 obj = m->object; 1117 x = vm_page_busy_fetch(m); 1118 for (;;) { 1119 if ((x & VPB_BIT_SHARED) == 0) 1120 return (0); 1121 /* 1122 * Reduce the window for transient busies that will trigger 1123 * false negatives in vm_page_ps_test(). 1124 */ 1125 if (obj != NULL && vm_object_busied(obj)) 1126 return (0); 1127 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1128 x + VPB_ONE_SHARER)) 1129 break; 1130 } 1131 1132 /* Refetch the object now that we're guaranteed that it is stable. */ 1133 obj = m->object; 1134 if (obj != NULL && vm_object_busied(obj)) { 1135 vm_page_sunbusy(m); 1136 return (0); 1137 } 1138 return (1); 1139 } 1140 1141 /* 1142 * vm_page_tryxbusy: 1143 * 1144 * Try to exclusive busy a page. 1145 * If the operation succeeds 1 is returned otherwise 0. 1146 * The operation never sleeps. 1147 */ 1148 int 1149 vm_page_tryxbusy(vm_page_t m) 1150 { 1151 vm_object_t obj; 1152 1153 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1154 VPB_CURTHREAD_EXCLUSIVE) == 0) 1155 return (0); 1156 1157 obj = m->object; 1158 if (obj != NULL && vm_object_busied(obj)) { 1159 vm_page_xunbusy(m); 1160 return (0); 1161 } 1162 return (1); 1163 } 1164 1165 static void 1166 vm_page_xunbusy_hard_tail(vm_page_t m) 1167 { 1168 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1169 /* Wake the waiter. */ 1170 wakeup(m); 1171 } 1172 1173 /* 1174 * vm_page_xunbusy_hard: 1175 * 1176 * Called when unbusy has failed because there is a waiter. 1177 */ 1178 void 1179 vm_page_xunbusy_hard(vm_page_t m) 1180 { 1181 vm_page_assert_xbusied(m); 1182 vm_page_xunbusy_hard_tail(m); 1183 } 1184 1185 void 1186 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1187 { 1188 vm_page_assert_xbusied_unchecked(m); 1189 vm_page_xunbusy_hard_tail(m); 1190 } 1191 1192 static void 1193 vm_page_busy_free(vm_page_t m) 1194 { 1195 u_int x; 1196 1197 atomic_thread_fence_rel(); 1198 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1199 if ((x & VPB_BIT_WAITERS) != 0) 1200 wakeup(m); 1201 } 1202 1203 /* 1204 * vm_page_unhold_pages: 1205 * 1206 * Unhold each of the pages that is referenced by the given array. 1207 */ 1208 void 1209 vm_page_unhold_pages(vm_page_t *ma, int count) 1210 { 1211 1212 for (; count != 0; count--) { 1213 vm_page_unwire(*ma, PQ_ACTIVE); 1214 ma++; 1215 } 1216 } 1217 1218 vm_page_t 1219 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1220 { 1221 vm_page_t m; 1222 1223 #ifdef VM_PHYSSEG_SPARSE 1224 m = vm_phys_paddr_to_vm_page(pa); 1225 if (m == NULL) 1226 m = vm_phys_fictitious_to_vm_page(pa); 1227 return (m); 1228 #elif defined(VM_PHYSSEG_DENSE) 1229 long pi; 1230 1231 pi = atop(pa); 1232 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1233 m = &vm_page_array[pi - first_page]; 1234 return (m); 1235 } 1236 return (vm_phys_fictitious_to_vm_page(pa)); 1237 #else 1238 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1239 #endif 1240 } 1241 1242 /* 1243 * vm_page_getfake: 1244 * 1245 * Create a fictitious page with the specified physical address and 1246 * memory attribute. The memory attribute is the only the machine- 1247 * dependent aspect of a fictitious page that must be initialized. 1248 */ 1249 vm_page_t 1250 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1251 { 1252 vm_page_t m; 1253 1254 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1255 vm_page_initfake(m, paddr, memattr); 1256 return (m); 1257 } 1258 1259 void 1260 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1261 { 1262 1263 if ((m->flags & PG_FICTITIOUS) != 0) { 1264 /* 1265 * The page's memattr might have changed since the 1266 * previous initialization. Update the pmap to the 1267 * new memattr. 1268 */ 1269 goto memattr; 1270 } 1271 m->phys_addr = paddr; 1272 m->a.queue = PQ_NONE; 1273 /* Fictitious pages don't use "segind". */ 1274 m->flags = PG_FICTITIOUS; 1275 /* Fictitious pages don't use "order" or "pool". */ 1276 m->oflags = VPO_UNMANAGED; 1277 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1278 /* Fictitious pages are unevictable. */ 1279 m->ref_count = 1; 1280 pmap_page_init(m); 1281 memattr: 1282 pmap_page_set_memattr(m, memattr); 1283 } 1284 1285 /* 1286 * vm_page_putfake: 1287 * 1288 * Release a fictitious page. 1289 */ 1290 void 1291 vm_page_putfake(vm_page_t m) 1292 { 1293 1294 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1295 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1296 ("vm_page_putfake: bad page %p", m)); 1297 vm_page_assert_xbusied(m); 1298 vm_page_busy_free(m); 1299 uma_zfree(fakepg_zone, m); 1300 } 1301 1302 /* 1303 * vm_page_updatefake: 1304 * 1305 * Update the given fictitious page to the specified physical address and 1306 * memory attribute. 1307 */ 1308 void 1309 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1310 { 1311 1312 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1313 ("vm_page_updatefake: bad page %p", m)); 1314 m->phys_addr = paddr; 1315 pmap_page_set_memattr(m, memattr); 1316 } 1317 1318 /* 1319 * vm_page_free: 1320 * 1321 * Free a page. 1322 */ 1323 void 1324 vm_page_free(vm_page_t m) 1325 { 1326 1327 m->flags &= ~PG_ZERO; 1328 vm_page_free_toq(m); 1329 } 1330 1331 /* 1332 * vm_page_free_zero: 1333 * 1334 * Free a page to the zerod-pages queue 1335 */ 1336 void 1337 vm_page_free_zero(vm_page_t m) 1338 { 1339 1340 m->flags |= PG_ZERO; 1341 vm_page_free_toq(m); 1342 } 1343 1344 /* 1345 * Unbusy and handle the page queueing for a page from a getpages request that 1346 * was optionally read ahead or behind. 1347 */ 1348 void 1349 vm_page_readahead_finish(vm_page_t m) 1350 { 1351 1352 /* We shouldn't put invalid pages on queues. */ 1353 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1354 1355 /* 1356 * Since the page is not the actually needed one, whether it should 1357 * be activated or deactivated is not obvious. Empirical results 1358 * have shown that deactivating the page is usually the best choice, 1359 * unless the page is wanted by another thread. 1360 */ 1361 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1362 vm_page_activate(m); 1363 else 1364 vm_page_deactivate(m); 1365 vm_page_xunbusy_unchecked(m); 1366 } 1367 1368 /* 1369 * Destroy the identity of an invalid page and free it if possible. 1370 * This is intended to be used when reading a page from backing store fails. 1371 */ 1372 void 1373 vm_page_free_invalid(vm_page_t m) 1374 { 1375 1376 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1377 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1378 KASSERT(m->object != NULL, ("page %p has no object", m)); 1379 VM_OBJECT_ASSERT_WLOCKED(m->object); 1380 1381 /* 1382 * We may be attempting to free the page as part of the handling for an 1383 * I/O error, in which case the page was xbusied by a different thread. 1384 */ 1385 vm_page_xbusy_claim(m); 1386 1387 /* 1388 * If someone has wired this page while the object lock 1389 * was not held, then the thread that unwires is responsible 1390 * for freeing the page. Otherwise just free the page now. 1391 * The wire count of this unmapped page cannot change while 1392 * we have the page xbusy and the page's object wlocked. 1393 */ 1394 if (vm_page_remove(m)) 1395 vm_page_free(m); 1396 } 1397 1398 /* 1399 * vm_page_sleep_if_busy: 1400 * 1401 * Sleep and release the object lock if the page is busied. 1402 * Returns TRUE if the thread slept. 1403 * 1404 * The given page must be unlocked and object containing it must 1405 * be locked. 1406 */ 1407 int 1408 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg) 1409 { 1410 vm_object_t obj; 1411 1412 vm_page_lock_assert(m, MA_NOTOWNED); 1413 VM_OBJECT_ASSERT_WLOCKED(m->object); 1414 1415 /* 1416 * The page-specific object must be cached because page 1417 * identity can change during the sleep, causing the 1418 * re-lock of a different object. 1419 * It is assumed that a reference to the object is already 1420 * held by the callers. 1421 */ 1422 obj = m->object; 1423 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) { 1424 VM_OBJECT_WLOCK(obj); 1425 return (TRUE); 1426 } 1427 return (FALSE); 1428 } 1429 1430 /* 1431 * vm_page_sleep_if_xbusy: 1432 * 1433 * Sleep and release the object lock if the page is xbusied. 1434 * Returns TRUE if the thread slept. 1435 * 1436 * The given page must be unlocked and object containing it must 1437 * be locked. 1438 */ 1439 int 1440 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg) 1441 { 1442 vm_object_t obj; 1443 1444 vm_page_lock_assert(m, MA_NOTOWNED); 1445 VM_OBJECT_ASSERT_WLOCKED(m->object); 1446 1447 /* 1448 * The page-specific object must be cached because page 1449 * identity can change during the sleep, causing the 1450 * re-lock of a different object. 1451 * It is assumed that a reference to the object is already 1452 * held by the callers. 1453 */ 1454 obj = m->object; 1455 if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY, 1456 true)) { 1457 VM_OBJECT_WLOCK(obj); 1458 return (TRUE); 1459 } 1460 return (FALSE); 1461 } 1462 1463 /* 1464 * vm_page_dirty_KBI: [ internal use only ] 1465 * 1466 * Set all bits in the page's dirty field. 1467 * 1468 * The object containing the specified page must be locked if the 1469 * call is made from the machine-independent layer. 1470 * 1471 * See vm_page_clear_dirty_mask(). 1472 * 1473 * This function should only be called by vm_page_dirty(). 1474 */ 1475 void 1476 vm_page_dirty_KBI(vm_page_t m) 1477 { 1478 1479 /* Refer to this operation by its public name. */ 1480 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1481 m->dirty = VM_PAGE_BITS_ALL; 1482 } 1483 1484 /* 1485 * vm_page_insert: [ internal use only ] 1486 * 1487 * Inserts the given mem entry into the object and object list. 1488 * 1489 * The object must be locked. 1490 */ 1491 int 1492 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1493 { 1494 vm_page_t mpred; 1495 1496 VM_OBJECT_ASSERT_WLOCKED(object); 1497 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1498 return (vm_page_insert_after(m, object, pindex, mpred)); 1499 } 1500 1501 /* 1502 * vm_page_insert_after: 1503 * 1504 * Inserts the page "m" into the specified object at offset "pindex". 1505 * 1506 * The page "mpred" must immediately precede the offset "pindex" within 1507 * the specified object. 1508 * 1509 * The object must be locked. 1510 */ 1511 static int 1512 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1513 vm_page_t mpred) 1514 { 1515 vm_page_t msucc; 1516 1517 VM_OBJECT_ASSERT_WLOCKED(object); 1518 KASSERT(m->object == NULL, 1519 ("vm_page_insert_after: page already inserted")); 1520 if (mpred != NULL) { 1521 KASSERT(mpred->object == object, 1522 ("vm_page_insert_after: object doesn't contain mpred")); 1523 KASSERT(mpred->pindex < pindex, 1524 ("vm_page_insert_after: mpred doesn't precede pindex")); 1525 msucc = TAILQ_NEXT(mpred, listq); 1526 } else 1527 msucc = TAILQ_FIRST(&object->memq); 1528 if (msucc != NULL) 1529 KASSERT(msucc->pindex > pindex, 1530 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1531 1532 /* 1533 * Record the object/offset pair in this page. 1534 */ 1535 m->object = object; 1536 m->pindex = pindex; 1537 m->ref_count |= VPRC_OBJREF; 1538 1539 /* 1540 * Now link into the object's ordered list of backed pages. 1541 */ 1542 if (vm_radix_insert(&object->rtree, m)) { 1543 m->object = NULL; 1544 m->pindex = 0; 1545 m->ref_count &= ~VPRC_OBJREF; 1546 return (1); 1547 } 1548 vm_page_insert_radixdone(m, object, mpred); 1549 return (0); 1550 } 1551 1552 /* 1553 * vm_page_insert_radixdone: 1554 * 1555 * Complete page "m" insertion into the specified object after the 1556 * radix trie hooking. 1557 * 1558 * The page "mpred" must precede the offset "m->pindex" within the 1559 * specified object. 1560 * 1561 * The object must be locked. 1562 */ 1563 static void 1564 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1565 { 1566 1567 VM_OBJECT_ASSERT_WLOCKED(object); 1568 KASSERT(object != NULL && m->object == object, 1569 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1570 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1571 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1572 if (mpred != NULL) { 1573 KASSERT(mpred->object == object, 1574 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1575 KASSERT(mpred->pindex < m->pindex, 1576 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1577 } 1578 1579 if (mpred != NULL) 1580 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1581 else 1582 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1583 1584 /* 1585 * Show that the object has one more resident page. 1586 */ 1587 object->resident_page_count++; 1588 1589 /* 1590 * Hold the vnode until the last page is released. 1591 */ 1592 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1593 vhold(object->handle); 1594 1595 /* 1596 * Since we are inserting a new and possibly dirty page, 1597 * update the object's generation count. 1598 */ 1599 if (pmap_page_is_write_mapped(m)) 1600 vm_object_set_writeable_dirty(object); 1601 } 1602 1603 /* 1604 * Do the work to remove a page from its object. The caller is responsible for 1605 * updating the page's fields to reflect this removal. 1606 */ 1607 static void 1608 vm_page_object_remove(vm_page_t m) 1609 { 1610 vm_object_t object; 1611 vm_page_t mrem; 1612 1613 vm_page_assert_xbusied(m); 1614 object = m->object; 1615 VM_OBJECT_ASSERT_WLOCKED(object); 1616 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1617 ("page %p is missing its object ref", m)); 1618 1619 /* Deferred free of swap space. */ 1620 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1621 vm_pager_page_unswapped(m); 1622 1623 m->object = NULL; 1624 mrem = vm_radix_remove(&object->rtree, m->pindex); 1625 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1626 1627 /* 1628 * Now remove from the object's list of backed pages. 1629 */ 1630 TAILQ_REMOVE(&object->memq, m, listq); 1631 1632 /* 1633 * And show that the object has one fewer resident page. 1634 */ 1635 object->resident_page_count--; 1636 1637 /* 1638 * The vnode may now be recycled. 1639 */ 1640 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1641 vdrop(object->handle); 1642 } 1643 1644 /* 1645 * vm_page_remove: 1646 * 1647 * Removes the specified page from its containing object, but does not 1648 * invalidate any backing storage. Returns true if the object's reference 1649 * was the last reference to the page, and false otherwise. 1650 * 1651 * The object must be locked and the page must be exclusively busied. 1652 * The exclusive busy will be released on return. If this is not the 1653 * final ref and the caller does not hold a wire reference it may not 1654 * continue to access the page. 1655 */ 1656 bool 1657 vm_page_remove(vm_page_t m) 1658 { 1659 bool dropped; 1660 1661 dropped = vm_page_remove_xbusy(m); 1662 vm_page_xunbusy(m); 1663 1664 return (dropped); 1665 } 1666 1667 /* 1668 * vm_page_remove_xbusy 1669 * 1670 * Removes the page but leaves the xbusy held. Returns true if this 1671 * removed the final ref and false otherwise. 1672 */ 1673 bool 1674 vm_page_remove_xbusy(vm_page_t m) 1675 { 1676 1677 vm_page_object_remove(m); 1678 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1679 } 1680 1681 /* 1682 * vm_page_lookup: 1683 * 1684 * Returns the page associated with the object/offset 1685 * pair specified; if none is found, NULL is returned. 1686 * 1687 * The object must be locked. 1688 */ 1689 vm_page_t 1690 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1691 { 1692 1693 VM_OBJECT_ASSERT_LOCKED(object); 1694 return (vm_radix_lookup(&object->rtree, pindex)); 1695 } 1696 1697 /* 1698 * vm_page_lookup_unlocked: 1699 * 1700 * Returns the page associated with the object/offset pair specified; 1701 * if none is found, NULL is returned. The page may be no longer be 1702 * present in the object at the time that this function returns. Only 1703 * useful for opportunistic checks such as inmem(). 1704 */ 1705 vm_page_t 1706 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1707 { 1708 1709 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1710 } 1711 1712 /* 1713 * vm_page_relookup: 1714 * 1715 * Returns a page that must already have been busied by 1716 * the caller. Used for bogus page replacement. 1717 */ 1718 vm_page_t 1719 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1720 { 1721 vm_page_t m; 1722 1723 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 1724 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1725 m->object == object && m->pindex == pindex, 1726 ("vm_page_relookup: Invalid page %p", m)); 1727 return (m); 1728 } 1729 1730 /* 1731 * This should only be used by lockless functions for releasing transient 1732 * incorrect acquires. The page may have been freed after we acquired a 1733 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1734 * further to do. 1735 */ 1736 static void 1737 vm_page_busy_release(vm_page_t m) 1738 { 1739 u_int x; 1740 1741 x = vm_page_busy_fetch(m); 1742 for (;;) { 1743 if (x == VPB_FREED) 1744 break; 1745 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1746 if (atomic_fcmpset_int(&m->busy_lock, &x, 1747 x - VPB_ONE_SHARER)) 1748 break; 1749 continue; 1750 } 1751 KASSERT((x & VPB_BIT_SHARED) != 0 || 1752 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1753 ("vm_page_busy_release: %p xbusy not owned.", m)); 1754 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1755 continue; 1756 if ((x & VPB_BIT_WAITERS) != 0) 1757 wakeup(m); 1758 break; 1759 } 1760 } 1761 1762 /* 1763 * vm_page_find_least: 1764 * 1765 * Returns the page associated with the object with least pindex 1766 * greater than or equal to the parameter pindex, or NULL. 1767 * 1768 * The object must be locked. 1769 */ 1770 vm_page_t 1771 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1772 { 1773 vm_page_t m; 1774 1775 VM_OBJECT_ASSERT_LOCKED(object); 1776 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1777 m = vm_radix_lookup_ge(&object->rtree, pindex); 1778 return (m); 1779 } 1780 1781 /* 1782 * Returns the given page's successor (by pindex) within the object if it is 1783 * resident; if none is found, NULL is returned. 1784 * 1785 * The object must be locked. 1786 */ 1787 vm_page_t 1788 vm_page_next(vm_page_t m) 1789 { 1790 vm_page_t next; 1791 1792 VM_OBJECT_ASSERT_LOCKED(m->object); 1793 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1794 MPASS(next->object == m->object); 1795 if (next->pindex != m->pindex + 1) 1796 next = NULL; 1797 } 1798 return (next); 1799 } 1800 1801 /* 1802 * Returns the given page's predecessor (by pindex) within the object if it is 1803 * resident; if none is found, NULL is returned. 1804 * 1805 * The object must be locked. 1806 */ 1807 vm_page_t 1808 vm_page_prev(vm_page_t m) 1809 { 1810 vm_page_t prev; 1811 1812 VM_OBJECT_ASSERT_LOCKED(m->object); 1813 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1814 MPASS(prev->object == m->object); 1815 if (prev->pindex != m->pindex - 1) 1816 prev = NULL; 1817 } 1818 return (prev); 1819 } 1820 1821 /* 1822 * Uses the page mnew as a replacement for an existing page at index 1823 * pindex which must be already present in the object. 1824 * 1825 * Both pages must be exclusively busied on enter. The old page is 1826 * unbusied on exit. 1827 * 1828 * A return value of true means mold is now free. If this is not the 1829 * final ref and the caller does not hold a wire reference it may not 1830 * continue to access the page. 1831 */ 1832 static bool 1833 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1834 vm_page_t mold) 1835 { 1836 vm_page_t mret; 1837 bool dropped; 1838 1839 VM_OBJECT_ASSERT_WLOCKED(object); 1840 vm_page_assert_xbusied(mold); 1841 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1842 ("vm_page_replace: page %p already in object", mnew)); 1843 1844 /* 1845 * This function mostly follows vm_page_insert() and 1846 * vm_page_remove() without the radix, object count and vnode 1847 * dance. Double check such functions for more comments. 1848 */ 1849 1850 mnew->object = object; 1851 mnew->pindex = pindex; 1852 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1853 mret = vm_radix_replace(&object->rtree, mnew); 1854 KASSERT(mret == mold, 1855 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1856 KASSERT((mold->oflags & VPO_UNMANAGED) == 1857 (mnew->oflags & VPO_UNMANAGED), 1858 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1859 1860 /* Keep the resident page list in sorted order. */ 1861 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1862 TAILQ_REMOVE(&object->memq, mold, listq); 1863 mold->object = NULL; 1864 1865 /* 1866 * The object's resident_page_count does not change because we have 1867 * swapped one page for another, but the generation count should 1868 * change if the page is dirty. 1869 */ 1870 if (pmap_page_is_write_mapped(mnew)) 1871 vm_object_set_writeable_dirty(object); 1872 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1873 vm_page_xunbusy(mold); 1874 1875 return (dropped); 1876 } 1877 1878 void 1879 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1880 vm_page_t mold) 1881 { 1882 1883 vm_page_assert_xbusied(mnew); 1884 1885 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1886 vm_page_free(mold); 1887 } 1888 1889 /* 1890 * vm_page_rename: 1891 * 1892 * Move the given memory entry from its 1893 * current object to the specified target object/offset. 1894 * 1895 * Note: swap associated with the page must be invalidated by the move. We 1896 * have to do this for several reasons: (1) we aren't freeing the 1897 * page, (2) we are dirtying the page, (3) the VM system is probably 1898 * moving the page from object A to B, and will then later move 1899 * the backing store from A to B and we can't have a conflict. 1900 * 1901 * Note: we *always* dirty the page. It is necessary both for the 1902 * fact that we moved it, and because we may be invalidating 1903 * swap. 1904 * 1905 * The objects must be locked. 1906 */ 1907 int 1908 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1909 { 1910 vm_page_t mpred; 1911 vm_pindex_t opidx; 1912 1913 VM_OBJECT_ASSERT_WLOCKED(new_object); 1914 1915 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1916 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1917 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1918 ("vm_page_rename: pindex already renamed")); 1919 1920 /* 1921 * Create a custom version of vm_page_insert() which does not depend 1922 * by m_prev and can cheat on the implementation aspects of the 1923 * function. 1924 */ 1925 opidx = m->pindex; 1926 m->pindex = new_pindex; 1927 if (vm_radix_insert(&new_object->rtree, m)) { 1928 m->pindex = opidx; 1929 return (1); 1930 } 1931 1932 /* 1933 * The operation cannot fail anymore. The removal must happen before 1934 * the listq iterator is tainted. 1935 */ 1936 m->pindex = opidx; 1937 vm_page_object_remove(m); 1938 1939 /* Return back to the new pindex to complete vm_page_insert(). */ 1940 m->pindex = new_pindex; 1941 m->object = new_object; 1942 1943 vm_page_insert_radixdone(m, new_object, mpred); 1944 vm_page_dirty(m); 1945 return (0); 1946 } 1947 1948 /* 1949 * vm_page_alloc: 1950 * 1951 * Allocate and return a page that is associated with the specified 1952 * object and offset pair. By default, this page is exclusive busied. 1953 * 1954 * The caller must always specify an allocation class. 1955 * 1956 * allocation classes: 1957 * VM_ALLOC_NORMAL normal process request 1958 * VM_ALLOC_SYSTEM system *really* needs a page 1959 * VM_ALLOC_INTERRUPT interrupt time request 1960 * 1961 * optional allocation flags: 1962 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1963 * intends to allocate 1964 * VM_ALLOC_NOBUSY do not exclusive busy the page 1965 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1966 * VM_ALLOC_SBUSY shared busy the allocated page 1967 * VM_ALLOC_WIRED wire the allocated page 1968 * VM_ALLOC_ZERO prefer a zeroed page 1969 */ 1970 vm_page_t 1971 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1972 { 1973 1974 return (vm_page_alloc_after(object, pindex, req, 1975 vm_radix_lookup_le(&object->rtree, pindex))); 1976 } 1977 1978 vm_page_t 1979 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1980 int req) 1981 { 1982 1983 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1984 vm_radix_lookup_le(&object->rtree, pindex))); 1985 } 1986 1987 /* 1988 * Allocate a page in the specified object with the given page index. To 1989 * optimize insertion of the page into the object, the caller must also specifiy 1990 * the resident page in the object with largest index smaller than the given 1991 * page index, or NULL if no such page exists. 1992 */ 1993 vm_page_t 1994 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1995 int req, vm_page_t mpred) 1996 { 1997 struct vm_domainset_iter di; 1998 vm_page_t m; 1999 int domain; 2000 2001 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2002 do { 2003 m = vm_page_alloc_domain_after(object, pindex, domain, req, 2004 mpred); 2005 if (m != NULL) 2006 break; 2007 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2008 2009 return (m); 2010 } 2011 2012 /* 2013 * Returns true if the number of free pages exceeds the minimum 2014 * for the request class and false otherwise. 2015 */ 2016 static int 2017 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2018 { 2019 u_int limit, old, new; 2020 2021 if (req_class == VM_ALLOC_INTERRUPT) 2022 limit = 0; 2023 else if (req_class == VM_ALLOC_SYSTEM) 2024 limit = vmd->vmd_interrupt_free_min; 2025 else 2026 limit = vmd->vmd_free_reserved; 2027 2028 /* 2029 * Attempt to reserve the pages. Fail if we're below the limit. 2030 */ 2031 limit += npages; 2032 old = vmd->vmd_free_count; 2033 do { 2034 if (old < limit) 2035 return (0); 2036 new = old - npages; 2037 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2038 2039 /* Wake the page daemon if we've crossed the threshold. */ 2040 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2041 pagedaemon_wakeup(vmd->vmd_domain); 2042 2043 /* Only update bitsets on transitions. */ 2044 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2045 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2046 vm_domain_set(vmd); 2047 2048 return (1); 2049 } 2050 2051 int 2052 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2053 { 2054 int req_class; 2055 2056 /* 2057 * The page daemon is allowed to dig deeper into the free page list. 2058 */ 2059 req_class = req & VM_ALLOC_CLASS_MASK; 2060 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2061 req_class = VM_ALLOC_SYSTEM; 2062 return (_vm_domain_allocate(vmd, req_class, npages)); 2063 } 2064 2065 vm_page_t 2066 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 2067 int req, vm_page_t mpred) 2068 { 2069 struct vm_domain *vmd; 2070 vm_page_t m; 2071 int flags; 2072 2073 #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2074 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ 2075 VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ 2076 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) 2077 KASSERT((req & ~VPA_FLAGS) == 0, 2078 ("invalid request %#x", req)); 2079 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2080 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2081 ("invalid request %#x", req)); 2082 KASSERT(mpred == NULL || mpred->pindex < pindex, 2083 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2084 (uintmax_t)pindex)); 2085 VM_OBJECT_ASSERT_WLOCKED(object); 2086 2087 flags = 0; 2088 m = NULL; 2089 again: 2090 #if VM_NRESERVLEVEL > 0 2091 /* 2092 * Can we allocate the page from a reservation? 2093 */ 2094 if (vm_object_reserv(object) && 2095 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 2096 NULL) { 2097 goto found; 2098 } 2099 #endif 2100 vmd = VM_DOMAIN(domain); 2101 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2102 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2103 M_NOWAIT | M_NOVM); 2104 if (m != NULL) { 2105 flags |= PG_PCPU_CACHE; 2106 goto found; 2107 } 2108 } 2109 if (vm_domain_allocate(vmd, req, 1)) { 2110 /* 2111 * If not, allocate it from the free page queues. 2112 */ 2113 vm_domain_free_lock(vmd); 2114 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2115 vm_domain_free_unlock(vmd); 2116 if (m == NULL) { 2117 vm_domain_freecnt_inc(vmd, 1); 2118 #if VM_NRESERVLEVEL > 0 2119 if (vm_reserv_reclaim_inactive(domain)) 2120 goto again; 2121 #endif 2122 } 2123 } 2124 if (m == NULL) { 2125 /* 2126 * Not allocatable, give up. 2127 */ 2128 if (vm_domain_alloc_fail(vmd, object, req)) 2129 goto again; 2130 return (NULL); 2131 } 2132 2133 /* 2134 * At this point we had better have found a good page. 2135 */ 2136 found: 2137 vm_page_dequeue(m); 2138 vm_page_alloc_check(m); 2139 2140 /* 2141 * Initialize the page. Only the PG_ZERO flag is inherited. 2142 */ 2143 flags |= m->flags & PG_ZERO; 2144 if ((req & VM_ALLOC_NODUMP) != 0) 2145 flags |= PG_NODUMP; 2146 m->flags = flags; 2147 m->a.flags = 0; 2148 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2149 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2150 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2151 else if ((req & VM_ALLOC_SBUSY) != 0) 2152 m->busy_lock = VPB_SHARERS_WORD(1); 2153 else 2154 m->busy_lock = VPB_UNBUSIED; 2155 if (req & VM_ALLOC_WIRED) { 2156 vm_wire_add(1); 2157 m->ref_count = 1; 2158 } 2159 m->a.act_count = 0; 2160 2161 if (vm_page_insert_after(m, object, pindex, mpred)) { 2162 if (req & VM_ALLOC_WIRED) { 2163 vm_wire_sub(1); 2164 m->ref_count = 0; 2165 } 2166 KASSERT(m->object == NULL, ("page %p has object", m)); 2167 m->oflags = VPO_UNMANAGED; 2168 m->busy_lock = VPB_UNBUSIED; 2169 /* Don't change PG_ZERO. */ 2170 vm_page_free_toq(m); 2171 if (req & VM_ALLOC_WAITFAIL) { 2172 VM_OBJECT_WUNLOCK(object); 2173 vm_radix_wait(); 2174 VM_OBJECT_WLOCK(object); 2175 } 2176 return (NULL); 2177 } 2178 2179 /* Ignore device objects; the pager sets "memattr" for them. */ 2180 if (object->memattr != VM_MEMATTR_DEFAULT && 2181 (object->flags & OBJ_FICTITIOUS) == 0) 2182 pmap_page_set_memattr(m, object->memattr); 2183 2184 return (m); 2185 } 2186 2187 /* 2188 * vm_page_alloc_contig: 2189 * 2190 * Allocate a contiguous set of physical pages of the given size "npages" 2191 * from the free lists. All of the physical pages must be at or above 2192 * the given physical address "low" and below the given physical address 2193 * "high". The given value "alignment" determines the alignment of the 2194 * first physical page in the set. If the given value "boundary" is 2195 * non-zero, then the set of physical pages cannot cross any physical 2196 * address boundary that is a multiple of that value. Both "alignment" 2197 * and "boundary" must be a power of two. 2198 * 2199 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2200 * then the memory attribute setting for the physical pages is configured 2201 * to the object's memory attribute setting. Otherwise, the memory 2202 * attribute setting for the physical pages is configured to "memattr", 2203 * overriding the object's memory attribute setting. However, if the 2204 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2205 * memory attribute setting for the physical pages cannot be configured 2206 * to VM_MEMATTR_DEFAULT. 2207 * 2208 * The specified object may not contain fictitious pages. 2209 * 2210 * The caller must always specify an allocation class. 2211 * 2212 * allocation classes: 2213 * VM_ALLOC_NORMAL normal process request 2214 * VM_ALLOC_SYSTEM system *really* needs a page 2215 * VM_ALLOC_INTERRUPT interrupt time request 2216 * 2217 * optional allocation flags: 2218 * VM_ALLOC_NOBUSY do not exclusive busy the page 2219 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2220 * VM_ALLOC_SBUSY shared busy the allocated page 2221 * VM_ALLOC_WIRED wire the allocated page 2222 * VM_ALLOC_ZERO prefer a zeroed page 2223 */ 2224 vm_page_t 2225 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2226 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2227 vm_paddr_t boundary, vm_memattr_t memattr) 2228 { 2229 struct vm_domainset_iter di; 2230 vm_page_t m; 2231 int domain; 2232 2233 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2234 do { 2235 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2236 npages, low, high, alignment, boundary, memattr); 2237 if (m != NULL) 2238 break; 2239 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2240 2241 return (m); 2242 } 2243 2244 vm_page_t 2245 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2246 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2247 vm_paddr_t boundary, vm_memattr_t memattr) 2248 { 2249 struct vm_domain *vmd; 2250 vm_page_t m, m_ret, mpred; 2251 u_int busy_lock, flags, oflags; 2252 2253 #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) 2254 KASSERT((req & ~VPAC_FLAGS) == 0, 2255 ("invalid request %#x", req)); 2256 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2257 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2258 ("invalid request %#x", req)); 2259 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2260 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2261 ("invalid request %#x", req)); 2262 VM_OBJECT_ASSERT_WLOCKED(object); 2263 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2264 ("vm_page_alloc_contig: object %p has fictitious pages", 2265 object)); 2266 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2267 2268 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2269 KASSERT(mpred == NULL || mpred->pindex != pindex, 2270 ("vm_page_alloc_contig: pindex already allocated")); 2271 2272 /* 2273 * Can we allocate the pages without the number of free pages falling 2274 * below the lower bound for the allocation class? 2275 */ 2276 m_ret = NULL; 2277 again: 2278 #if VM_NRESERVLEVEL > 0 2279 /* 2280 * Can we allocate the pages from a reservation? 2281 */ 2282 if (vm_object_reserv(object) && 2283 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2284 mpred, npages, low, high, alignment, boundary)) != NULL) { 2285 goto found; 2286 } 2287 #endif 2288 vmd = VM_DOMAIN(domain); 2289 if (vm_domain_allocate(vmd, req, npages)) { 2290 /* 2291 * allocate them from the free page queues. 2292 */ 2293 vm_domain_free_lock(vmd); 2294 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2295 alignment, boundary); 2296 vm_domain_free_unlock(vmd); 2297 if (m_ret == NULL) { 2298 vm_domain_freecnt_inc(vmd, npages); 2299 #if VM_NRESERVLEVEL > 0 2300 if ((req & VM_ALLOC_NORECLAIM) == 0 && 2301 vm_reserv_reclaim_contig(domain, npages, low, 2302 high, alignment, boundary)) 2303 goto again; 2304 #endif 2305 } 2306 } 2307 if (m_ret == NULL) { 2308 if (vm_domain_alloc_fail(vmd, object, req)) 2309 goto again; 2310 return (NULL); 2311 } 2312 #if VM_NRESERVLEVEL > 0 2313 found: 2314 #endif 2315 for (m = m_ret; m < &m_ret[npages]; m++) { 2316 vm_page_dequeue(m); 2317 vm_page_alloc_check(m); 2318 } 2319 2320 /* 2321 * Initialize the pages. Only the PG_ZERO flag is inherited. 2322 */ 2323 flags = PG_ZERO; 2324 if ((req & VM_ALLOC_NODUMP) != 0) 2325 flags |= PG_NODUMP; 2326 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2327 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2328 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2329 else if ((req & VM_ALLOC_SBUSY) != 0) 2330 busy_lock = VPB_SHARERS_WORD(1); 2331 else 2332 busy_lock = VPB_UNBUSIED; 2333 if ((req & VM_ALLOC_WIRED) != 0) 2334 vm_wire_add(npages); 2335 if (object->memattr != VM_MEMATTR_DEFAULT && 2336 memattr == VM_MEMATTR_DEFAULT) 2337 memattr = object->memattr; 2338 for (m = m_ret; m < &m_ret[npages]; m++) { 2339 m->a.flags = 0; 2340 m->flags = (m->flags | PG_NODUMP) & flags; 2341 m->busy_lock = busy_lock; 2342 if ((req & VM_ALLOC_WIRED) != 0) 2343 m->ref_count = 1; 2344 m->a.act_count = 0; 2345 m->oflags = oflags; 2346 if (vm_page_insert_after(m, object, pindex, mpred)) { 2347 if ((req & VM_ALLOC_WIRED) != 0) 2348 vm_wire_sub(npages); 2349 KASSERT(m->object == NULL, 2350 ("page %p has object", m)); 2351 mpred = m; 2352 for (m = m_ret; m < &m_ret[npages]; m++) { 2353 if (m <= mpred && 2354 (req & VM_ALLOC_WIRED) != 0) 2355 m->ref_count = 0; 2356 m->oflags = VPO_UNMANAGED; 2357 m->busy_lock = VPB_UNBUSIED; 2358 /* Don't change PG_ZERO. */ 2359 vm_page_free_toq(m); 2360 } 2361 if (req & VM_ALLOC_WAITFAIL) { 2362 VM_OBJECT_WUNLOCK(object); 2363 vm_radix_wait(); 2364 VM_OBJECT_WLOCK(object); 2365 } 2366 return (NULL); 2367 } 2368 mpred = m; 2369 if (memattr != VM_MEMATTR_DEFAULT) 2370 pmap_page_set_memattr(m, memattr); 2371 pindex++; 2372 } 2373 return (m_ret); 2374 } 2375 2376 /* 2377 * Allocate a physical page that is not intended to be inserted into a VM 2378 * object. If the "freelist" parameter is not equal to VM_NFREELIST, then only 2379 * pages from the specified vm_phys freelist will be returned. 2380 */ 2381 static __always_inline vm_page_t 2382 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req) 2383 { 2384 struct vm_domain *vmd; 2385 vm_page_t m; 2386 int flags; 2387 2388 #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2389 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ 2390 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ 2391 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) 2392 KASSERT((req & ~VPAN_FLAGS) == 0, 2393 ("invalid request %#x", req)); 2394 2395 flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0; 2396 vmd = VM_DOMAIN(domain); 2397 again: 2398 if (freelist == VM_NFREELIST && 2399 vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2400 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2401 M_NOWAIT | M_NOVM); 2402 if (m != NULL) { 2403 flags |= PG_PCPU_CACHE; 2404 goto found; 2405 } 2406 } 2407 2408 if (vm_domain_allocate(vmd, req, 1)) { 2409 vm_domain_free_lock(vmd); 2410 if (freelist == VM_NFREELIST) 2411 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2412 else 2413 m = vm_phys_alloc_freelist_pages(domain, freelist, 2414 VM_FREEPOOL_DIRECT, 0); 2415 vm_domain_free_unlock(vmd); 2416 if (m == NULL) { 2417 vm_domain_freecnt_inc(vmd, 1); 2418 #if VM_NRESERVLEVEL > 0 2419 if (freelist == VM_NFREELIST && 2420 vm_reserv_reclaim_inactive(domain)) 2421 goto again; 2422 #endif 2423 } 2424 } 2425 if (m == NULL) { 2426 if (vm_domain_alloc_fail(vmd, NULL, req)) 2427 goto again; 2428 return (NULL); 2429 } 2430 2431 found: 2432 vm_page_dequeue(m); 2433 vm_page_alloc_check(m); 2434 2435 /* 2436 * Consumers should not rely on a useful default pindex value. 2437 */ 2438 m->pindex = 0xdeadc0dedeadc0de; 2439 m->flags = (m->flags & PG_ZERO) | flags; 2440 m->a.flags = 0; 2441 m->oflags = VPO_UNMANAGED; 2442 m->busy_lock = VPB_UNBUSIED; 2443 if ((req & VM_ALLOC_WIRED) != 0) { 2444 vm_wire_add(1); 2445 m->ref_count = 1; 2446 } 2447 2448 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2449 pmap_zero_page(m); 2450 2451 return (m); 2452 } 2453 2454 vm_page_t 2455 vm_page_alloc_freelist(int freelist, int req) 2456 { 2457 struct vm_domainset_iter di; 2458 vm_page_t m; 2459 int domain; 2460 2461 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2462 do { 2463 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2464 if (m != NULL) 2465 break; 2466 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2467 2468 return (m); 2469 } 2470 2471 vm_page_t 2472 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2473 { 2474 KASSERT(freelist >= 0 && freelist < VM_NFREELIST, 2475 ("%s: invalid freelist %d", __func__, freelist)); 2476 2477 return (_vm_page_alloc_noobj_domain(domain, freelist, req)); 2478 } 2479 2480 vm_page_t 2481 vm_page_alloc_noobj(int req) 2482 { 2483 struct vm_domainset_iter di; 2484 vm_page_t m; 2485 int domain; 2486 2487 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2488 do { 2489 m = vm_page_alloc_noobj_domain(domain, req); 2490 if (m != NULL) 2491 break; 2492 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2493 2494 return (m); 2495 } 2496 2497 vm_page_t 2498 vm_page_alloc_noobj_domain(int domain, int req) 2499 { 2500 return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req)); 2501 } 2502 2503 vm_page_t 2504 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2505 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2506 vm_memattr_t memattr) 2507 { 2508 struct vm_domainset_iter di; 2509 vm_page_t m; 2510 int domain; 2511 2512 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2513 do { 2514 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2515 high, alignment, boundary, memattr); 2516 if (m != NULL) 2517 break; 2518 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2519 2520 return (m); 2521 } 2522 2523 vm_page_t 2524 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2525 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2526 vm_memattr_t memattr) 2527 { 2528 struct vm_domain *vmd; 2529 vm_page_t m, m_ret; 2530 u_int flags; 2531 2532 #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) 2533 KASSERT((req & ~VPANC_FLAGS) == 0, 2534 ("invalid request %#x", req)); 2535 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2536 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2537 ("invalid request %#x", req)); 2538 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2539 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2540 ("invalid request %#x", req)); 2541 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2542 2543 m_ret = NULL; 2544 again: 2545 vmd = VM_DOMAIN(domain); 2546 if (vm_domain_allocate(vmd, req, npages)) { 2547 /* 2548 * allocate them from the free page queues. 2549 */ 2550 vm_domain_free_lock(vmd); 2551 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2552 alignment, boundary); 2553 vm_domain_free_unlock(vmd); 2554 if (m_ret == NULL) { 2555 vm_domain_freecnt_inc(vmd, npages); 2556 #if VM_NRESERVLEVEL > 0 2557 if ((req & VM_ALLOC_NORECLAIM) == 0 && 2558 vm_reserv_reclaim_contig(domain, npages, low, 2559 high, alignment, boundary)) 2560 goto again; 2561 #endif 2562 } 2563 } 2564 if (m_ret == NULL) { 2565 if (vm_domain_alloc_fail(vmd, NULL, req)) 2566 goto again; 2567 return (NULL); 2568 } 2569 2570 /* 2571 * Initialize the pages. Only the PG_ZERO flag is inherited. 2572 */ 2573 flags = PG_ZERO; 2574 if ((req & VM_ALLOC_NODUMP) != 0) 2575 flags |= PG_NODUMP; 2576 if ((req & VM_ALLOC_WIRED) != 0) 2577 vm_wire_add(npages); 2578 for (m = m_ret; m < &m_ret[npages]; m++) { 2579 vm_page_dequeue(m); 2580 vm_page_alloc_check(m); 2581 2582 /* 2583 * Consumers should not rely on a useful default pindex value. 2584 */ 2585 m->pindex = 0xdeadc0dedeadc0de; 2586 m->a.flags = 0; 2587 m->flags = (m->flags | PG_NODUMP) & flags; 2588 m->busy_lock = VPB_UNBUSIED; 2589 if ((req & VM_ALLOC_WIRED) != 0) 2590 m->ref_count = 1; 2591 m->a.act_count = 0; 2592 m->oflags = VPO_UNMANAGED; 2593 2594 /* 2595 * Zero the page before updating any mappings since the page is 2596 * not yet shared with any devices which might require the 2597 * non-default memory attribute. pmap_page_set_memattr() 2598 * flushes data caches before returning. 2599 */ 2600 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2601 pmap_zero_page(m); 2602 if (memattr != VM_MEMATTR_DEFAULT) 2603 pmap_page_set_memattr(m, memattr); 2604 } 2605 return (m_ret); 2606 } 2607 2608 /* 2609 * Check a page that has been freshly dequeued from a freelist. 2610 */ 2611 static void 2612 vm_page_alloc_check(vm_page_t m) 2613 { 2614 2615 KASSERT(m->object == NULL, ("page %p has object", m)); 2616 KASSERT(m->a.queue == PQ_NONE && 2617 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2618 ("page %p has unexpected queue %d, flags %#x", 2619 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2620 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2621 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2622 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2623 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2624 ("page %p has unexpected memattr %d", 2625 m, pmap_page_get_memattr(m))); 2626 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2627 pmap_vm_page_alloc_check(m); 2628 } 2629 2630 static int 2631 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2632 { 2633 struct vm_domain *vmd; 2634 struct vm_pgcache *pgcache; 2635 int i; 2636 2637 pgcache = arg; 2638 vmd = VM_DOMAIN(pgcache->domain); 2639 2640 /* 2641 * The page daemon should avoid creating extra memory pressure since its 2642 * main purpose is to replenish the store of free pages. 2643 */ 2644 if (vmd->vmd_severeset || curproc == pageproc || 2645 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2646 return (0); 2647 domain = vmd->vmd_domain; 2648 vm_domain_free_lock(vmd); 2649 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2650 (vm_page_t *)store); 2651 vm_domain_free_unlock(vmd); 2652 if (cnt != i) 2653 vm_domain_freecnt_inc(vmd, cnt - i); 2654 2655 return (i); 2656 } 2657 2658 static void 2659 vm_page_zone_release(void *arg, void **store, int cnt) 2660 { 2661 struct vm_domain *vmd; 2662 struct vm_pgcache *pgcache; 2663 vm_page_t m; 2664 int i; 2665 2666 pgcache = arg; 2667 vmd = VM_DOMAIN(pgcache->domain); 2668 vm_domain_free_lock(vmd); 2669 for (i = 0; i < cnt; i++) { 2670 m = (vm_page_t)store[i]; 2671 vm_phys_free_pages(m, 0); 2672 } 2673 vm_domain_free_unlock(vmd); 2674 vm_domain_freecnt_inc(vmd, cnt); 2675 } 2676 2677 #define VPSC_ANY 0 /* No restrictions. */ 2678 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2679 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2680 2681 /* 2682 * vm_page_scan_contig: 2683 * 2684 * Scan vm_page_array[] between the specified entries "m_start" and 2685 * "m_end" for a run of contiguous physical pages that satisfy the 2686 * specified conditions, and return the lowest page in the run. The 2687 * specified "alignment" determines the alignment of the lowest physical 2688 * page in the run. If the specified "boundary" is non-zero, then the 2689 * run of physical pages cannot span a physical address that is a 2690 * multiple of "boundary". 2691 * 2692 * "m_end" is never dereferenced, so it need not point to a vm_page 2693 * structure within vm_page_array[]. 2694 * 2695 * "npages" must be greater than zero. "m_start" and "m_end" must not 2696 * span a hole (or discontiguity) in the physical address space. Both 2697 * "alignment" and "boundary" must be a power of two. 2698 */ 2699 vm_page_t 2700 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2701 u_long alignment, vm_paddr_t boundary, int options) 2702 { 2703 vm_object_t object; 2704 vm_paddr_t pa; 2705 vm_page_t m, m_run; 2706 #if VM_NRESERVLEVEL > 0 2707 int level; 2708 #endif 2709 int m_inc, order, run_ext, run_len; 2710 2711 KASSERT(npages > 0, ("npages is 0")); 2712 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2713 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2714 m_run = NULL; 2715 run_len = 0; 2716 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2717 KASSERT((m->flags & PG_MARKER) == 0, 2718 ("page %p is PG_MARKER", m)); 2719 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2720 ("fictitious page %p has invalid ref count", m)); 2721 2722 /* 2723 * If the current page would be the start of a run, check its 2724 * physical address against the end, alignment, and boundary 2725 * conditions. If it doesn't satisfy these conditions, either 2726 * terminate the scan or advance to the next page that 2727 * satisfies the failed condition. 2728 */ 2729 if (run_len == 0) { 2730 KASSERT(m_run == NULL, ("m_run != NULL")); 2731 if (m + npages > m_end) 2732 break; 2733 pa = VM_PAGE_TO_PHYS(m); 2734 if ((pa & (alignment - 1)) != 0) { 2735 m_inc = atop(roundup2(pa, alignment) - pa); 2736 continue; 2737 } 2738 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2739 boundary) != 0) { 2740 m_inc = atop(roundup2(pa, boundary) - pa); 2741 continue; 2742 } 2743 } else 2744 KASSERT(m_run != NULL, ("m_run == NULL")); 2745 2746 retry: 2747 m_inc = 1; 2748 if (vm_page_wired(m)) 2749 run_ext = 0; 2750 #if VM_NRESERVLEVEL > 0 2751 else if ((level = vm_reserv_level(m)) >= 0 && 2752 (options & VPSC_NORESERV) != 0) { 2753 run_ext = 0; 2754 /* Advance to the end of the reservation. */ 2755 pa = VM_PAGE_TO_PHYS(m); 2756 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2757 pa); 2758 } 2759 #endif 2760 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2761 /* 2762 * The page is considered eligible for relocation if 2763 * and only if it could be laundered or reclaimed by 2764 * the page daemon. 2765 */ 2766 VM_OBJECT_RLOCK(object); 2767 if (object != m->object) { 2768 VM_OBJECT_RUNLOCK(object); 2769 goto retry; 2770 } 2771 /* Don't care: PG_NODUMP, PG_ZERO. */ 2772 if (object->type != OBJT_DEFAULT && 2773 (object->flags & OBJ_SWAP) == 0 && 2774 object->type != OBJT_VNODE) { 2775 run_ext = 0; 2776 #if VM_NRESERVLEVEL > 0 2777 } else if ((options & VPSC_NOSUPER) != 0 && 2778 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2779 run_ext = 0; 2780 /* Advance to the end of the superpage. */ 2781 pa = VM_PAGE_TO_PHYS(m); 2782 m_inc = atop(roundup2(pa + 1, 2783 vm_reserv_size(level)) - pa); 2784 #endif 2785 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2786 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2787 /* 2788 * The page is allocated but eligible for 2789 * relocation. Extend the current run by one 2790 * page. 2791 */ 2792 KASSERT(pmap_page_get_memattr(m) == 2793 VM_MEMATTR_DEFAULT, 2794 ("page %p has an unexpected memattr", m)); 2795 KASSERT((m->oflags & (VPO_SWAPINPROG | 2796 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2797 ("page %p has unexpected oflags", m)); 2798 /* Don't care: PGA_NOSYNC. */ 2799 run_ext = 1; 2800 } else 2801 run_ext = 0; 2802 VM_OBJECT_RUNLOCK(object); 2803 #if VM_NRESERVLEVEL > 0 2804 } else if (level >= 0) { 2805 /* 2806 * The page is reserved but not yet allocated. In 2807 * other words, it is still free. Extend the current 2808 * run by one page. 2809 */ 2810 run_ext = 1; 2811 #endif 2812 } else if ((order = m->order) < VM_NFREEORDER) { 2813 /* 2814 * The page is enqueued in the physical memory 2815 * allocator's free page queues. Moreover, it is the 2816 * first page in a power-of-two-sized run of 2817 * contiguous free pages. Add these pages to the end 2818 * of the current run, and jump ahead. 2819 */ 2820 run_ext = 1 << order; 2821 m_inc = 1 << order; 2822 } else { 2823 /* 2824 * Skip the page for one of the following reasons: (1) 2825 * It is enqueued in the physical memory allocator's 2826 * free page queues. However, it is not the first 2827 * page in a run of contiguous free pages. (This case 2828 * rarely occurs because the scan is performed in 2829 * ascending order.) (2) It is not reserved, and it is 2830 * transitioning from free to allocated. (Conversely, 2831 * the transition from allocated to free for managed 2832 * pages is blocked by the page busy lock.) (3) It is 2833 * allocated but not contained by an object and not 2834 * wired, e.g., allocated by Xen's balloon driver. 2835 */ 2836 run_ext = 0; 2837 } 2838 2839 /* 2840 * Extend or reset the current run of pages. 2841 */ 2842 if (run_ext > 0) { 2843 if (run_len == 0) 2844 m_run = m; 2845 run_len += run_ext; 2846 } else { 2847 if (run_len > 0) { 2848 m_run = NULL; 2849 run_len = 0; 2850 } 2851 } 2852 } 2853 if (run_len >= npages) 2854 return (m_run); 2855 return (NULL); 2856 } 2857 2858 /* 2859 * vm_page_reclaim_run: 2860 * 2861 * Try to relocate each of the allocated virtual pages within the 2862 * specified run of physical pages to a new physical address. Free the 2863 * physical pages underlying the relocated virtual pages. A virtual page 2864 * is relocatable if and only if it could be laundered or reclaimed by 2865 * the page daemon. Whenever possible, a virtual page is relocated to a 2866 * physical address above "high". 2867 * 2868 * Returns 0 if every physical page within the run was already free or 2869 * just freed by a successful relocation. Otherwise, returns a non-zero 2870 * value indicating why the last attempt to relocate a virtual page was 2871 * unsuccessful. 2872 * 2873 * "req_class" must be an allocation class. 2874 */ 2875 static int 2876 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2877 vm_paddr_t high) 2878 { 2879 struct vm_domain *vmd; 2880 struct spglist free; 2881 vm_object_t object; 2882 vm_paddr_t pa; 2883 vm_page_t m, m_end, m_new; 2884 int error, order, req; 2885 2886 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2887 ("req_class is not an allocation class")); 2888 SLIST_INIT(&free); 2889 error = 0; 2890 m = m_run; 2891 m_end = m_run + npages; 2892 for (; error == 0 && m < m_end; m++) { 2893 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2894 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2895 2896 /* 2897 * Racily check for wirings. Races are handled once the object 2898 * lock is held and the page is unmapped. 2899 */ 2900 if (vm_page_wired(m)) 2901 error = EBUSY; 2902 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2903 /* 2904 * The page is relocated if and only if it could be 2905 * laundered or reclaimed by the page daemon. 2906 */ 2907 VM_OBJECT_WLOCK(object); 2908 /* Don't care: PG_NODUMP, PG_ZERO. */ 2909 if (m->object != object || 2910 (object->type != OBJT_DEFAULT && 2911 (object->flags & OBJ_SWAP) == 0 && 2912 object->type != OBJT_VNODE)) 2913 error = EINVAL; 2914 else if (object->memattr != VM_MEMATTR_DEFAULT) 2915 error = EINVAL; 2916 else if (vm_page_queue(m) != PQ_NONE && 2917 vm_page_tryxbusy(m) != 0) { 2918 if (vm_page_wired(m)) { 2919 vm_page_xunbusy(m); 2920 error = EBUSY; 2921 goto unlock; 2922 } 2923 KASSERT(pmap_page_get_memattr(m) == 2924 VM_MEMATTR_DEFAULT, 2925 ("page %p has an unexpected memattr", m)); 2926 KASSERT(m->oflags == 0, 2927 ("page %p has unexpected oflags", m)); 2928 /* Don't care: PGA_NOSYNC. */ 2929 if (!vm_page_none_valid(m)) { 2930 /* 2931 * First, try to allocate a new page 2932 * that is above "high". Failing 2933 * that, try to allocate a new page 2934 * that is below "m_run". Allocate 2935 * the new page between the end of 2936 * "m_run" and "high" only as a last 2937 * resort. 2938 */ 2939 req = req_class; 2940 if ((m->flags & PG_NODUMP) != 0) 2941 req |= VM_ALLOC_NODUMP; 2942 if (trunc_page(high) != 2943 ~(vm_paddr_t)PAGE_MASK) { 2944 m_new = 2945 vm_page_alloc_noobj_contig( 2946 req, 1, round_page(high), 2947 ~(vm_paddr_t)0, PAGE_SIZE, 2948 0, VM_MEMATTR_DEFAULT); 2949 } else 2950 m_new = NULL; 2951 if (m_new == NULL) { 2952 pa = VM_PAGE_TO_PHYS(m_run); 2953 m_new = 2954 vm_page_alloc_noobj_contig( 2955 req, 1, 0, pa - 1, 2956 PAGE_SIZE, 0, 2957 VM_MEMATTR_DEFAULT); 2958 } 2959 if (m_new == NULL) { 2960 pa += ptoa(npages); 2961 m_new = 2962 vm_page_alloc_noobj_contig( 2963 req, 1, pa, high, PAGE_SIZE, 2964 0, VM_MEMATTR_DEFAULT); 2965 } 2966 if (m_new == NULL) { 2967 vm_page_xunbusy(m); 2968 error = ENOMEM; 2969 goto unlock; 2970 } 2971 2972 /* 2973 * Unmap the page and check for new 2974 * wirings that may have been acquired 2975 * through a pmap lookup. 2976 */ 2977 if (object->ref_count != 0 && 2978 !vm_page_try_remove_all(m)) { 2979 vm_page_xunbusy(m); 2980 vm_page_free(m_new); 2981 error = EBUSY; 2982 goto unlock; 2983 } 2984 2985 /* 2986 * Replace "m" with the new page. For 2987 * vm_page_replace(), "m" must be busy 2988 * and dequeued. Finally, change "m" 2989 * as if vm_page_free() was called. 2990 */ 2991 m_new->a.flags = m->a.flags & 2992 ~PGA_QUEUE_STATE_MASK; 2993 KASSERT(m_new->oflags == VPO_UNMANAGED, 2994 ("page %p is managed", m_new)); 2995 m_new->oflags = 0; 2996 pmap_copy_page(m, m_new); 2997 m_new->valid = m->valid; 2998 m_new->dirty = m->dirty; 2999 m->flags &= ~PG_ZERO; 3000 vm_page_dequeue(m); 3001 if (vm_page_replace_hold(m_new, object, 3002 m->pindex, m) && 3003 vm_page_free_prep(m)) 3004 SLIST_INSERT_HEAD(&free, m, 3005 plinks.s.ss); 3006 3007 /* 3008 * The new page must be deactivated 3009 * before the object is unlocked. 3010 */ 3011 vm_page_deactivate(m_new); 3012 } else { 3013 m->flags &= ~PG_ZERO; 3014 vm_page_dequeue(m); 3015 if (vm_page_free_prep(m)) 3016 SLIST_INSERT_HEAD(&free, m, 3017 plinks.s.ss); 3018 KASSERT(m->dirty == 0, 3019 ("page %p is dirty", m)); 3020 } 3021 } else 3022 error = EBUSY; 3023 unlock: 3024 VM_OBJECT_WUNLOCK(object); 3025 } else { 3026 MPASS(vm_page_domain(m) == domain); 3027 vmd = VM_DOMAIN(domain); 3028 vm_domain_free_lock(vmd); 3029 order = m->order; 3030 if (order < VM_NFREEORDER) { 3031 /* 3032 * The page is enqueued in the physical memory 3033 * allocator's free page queues. Moreover, it 3034 * is the first page in a power-of-two-sized 3035 * run of contiguous free pages. Jump ahead 3036 * to the last page within that run, and 3037 * continue from there. 3038 */ 3039 m += (1 << order) - 1; 3040 } 3041 #if VM_NRESERVLEVEL > 0 3042 else if (vm_reserv_is_page_free(m)) 3043 order = 0; 3044 #endif 3045 vm_domain_free_unlock(vmd); 3046 if (order == VM_NFREEORDER) 3047 error = EINVAL; 3048 } 3049 } 3050 if ((m = SLIST_FIRST(&free)) != NULL) { 3051 int cnt; 3052 3053 vmd = VM_DOMAIN(domain); 3054 cnt = 0; 3055 vm_domain_free_lock(vmd); 3056 do { 3057 MPASS(vm_page_domain(m) == domain); 3058 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3059 vm_phys_free_pages(m, 0); 3060 cnt++; 3061 } while ((m = SLIST_FIRST(&free)) != NULL); 3062 vm_domain_free_unlock(vmd); 3063 vm_domain_freecnt_inc(vmd, cnt); 3064 } 3065 return (error); 3066 } 3067 3068 #define NRUNS 16 3069 3070 CTASSERT(powerof2(NRUNS)); 3071 3072 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 3073 3074 #define MIN_RECLAIM 8 3075 3076 /* 3077 * vm_page_reclaim_contig: 3078 * 3079 * Reclaim allocated, contiguous physical memory satisfying the specified 3080 * conditions by relocating the virtual pages using that physical memory. 3081 * Returns true if reclamation is successful and false otherwise. Since 3082 * relocation requires the allocation of physical pages, reclamation may 3083 * fail due to a shortage of free pages. When reclamation fails, callers 3084 * are expected to perform vm_wait() before retrying a failed allocation 3085 * operation, e.g., vm_page_alloc_contig(). 3086 * 3087 * The caller must always specify an allocation class through "req". 3088 * 3089 * allocation classes: 3090 * VM_ALLOC_NORMAL normal process request 3091 * VM_ALLOC_SYSTEM system *really* needs a page 3092 * VM_ALLOC_INTERRUPT interrupt time request 3093 * 3094 * The optional allocation flags are ignored. 3095 * 3096 * "npages" must be greater than zero. Both "alignment" and "boundary" 3097 * must be a power of two. 3098 */ 3099 bool 3100 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3101 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3102 { 3103 struct vm_domain *vmd; 3104 vm_paddr_t curr_low; 3105 vm_page_t m_run, m_runs[NRUNS]; 3106 u_long count, minalign, reclaimed; 3107 int error, i, options, req_class; 3108 3109 KASSERT(npages > 0, ("npages is 0")); 3110 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3111 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3112 3113 /* 3114 * The caller will attempt an allocation after some runs have been 3115 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3116 * of vm_phys_alloc_contig(), round up the requested length to the next 3117 * power of two or maximum chunk size, and ensure that each run is 3118 * suitably aligned. 3119 */ 3120 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3121 npages = roundup2(npages, minalign); 3122 if (alignment < ptoa(minalign)) 3123 alignment = ptoa(minalign); 3124 3125 /* 3126 * The page daemon is allowed to dig deeper into the free page list. 3127 */ 3128 req_class = req & VM_ALLOC_CLASS_MASK; 3129 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3130 req_class = VM_ALLOC_SYSTEM; 3131 3132 /* 3133 * Return if the number of free pages cannot satisfy the requested 3134 * allocation. 3135 */ 3136 vmd = VM_DOMAIN(domain); 3137 count = vmd->vmd_free_count; 3138 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3139 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3140 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3141 return (false); 3142 3143 /* 3144 * Scan up to three times, relaxing the restrictions ("options") on 3145 * the reclamation of reservations and superpages each time. 3146 */ 3147 for (options = VPSC_NORESERV;;) { 3148 /* 3149 * Find the highest runs that satisfy the given constraints 3150 * and restrictions, and record them in "m_runs". 3151 */ 3152 curr_low = low; 3153 count = 0; 3154 for (;;) { 3155 m_run = vm_phys_scan_contig(domain, npages, curr_low, 3156 high, alignment, boundary, options); 3157 if (m_run == NULL) 3158 break; 3159 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 3160 m_runs[RUN_INDEX(count)] = m_run; 3161 count++; 3162 } 3163 3164 /* 3165 * Reclaim the highest runs in LIFO (descending) order until 3166 * the number of reclaimed pages, "reclaimed", is at least 3167 * MIN_RECLAIM. Reset "reclaimed" each time because each 3168 * reclamation is idempotent, and runs will (likely) recur 3169 * from one scan to the next as restrictions are relaxed. 3170 */ 3171 reclaimed = 0; 3172 for (i = 0; count > 0 && i < NRUNS; i++) { 3173 count--; 3174 m_run = m_runs[RUN_INDEX(count)]; 3175 error = vm_page_reclaim_run(req_class, domain, npages, 3176 m_run, high); 3177 if (error == 0) { 3178 reclaimed += npages; 3179 if (reclaimed >= MIN_RECLAIM) 3180 return (true); 3181 } 3182 } 3183 3184 /* 3185 * Either relax the restrictions on the next scan or return if 3186 * the last scan had no restrictions. 3187 */ 3188 if (options == VPSC_NORESERV) 3189 options = VPSC_NOSUPER; 3190 else if (options == VPSC_NOSUPER) 3191 options = VPSC_ANY; 3192 else if (options == VPSC_ANY) 3193 return (reclaimed != 0); 3194 } 3195 } 3196 3197 bool 3198 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3199 u_long alignment, vm_paddr_t boundary) 3200 { 3201 struct vm_domainset_iter di; 3202 int domain; 3203 bool ret; 3204 3205 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3206 do { 3207 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 3208 high, alignment, boundary); 3209 if (ret) 3210 break; 3211 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3212 3213 return (ret); 3214 } 3215 3216 /* 3217 * Set the domain in the appropriate page level domainset. 3218 */ 3219 void 3220 vm_domain_set(struct vm_domain *vmd) 3221 { 3222 3223 mtx_lock(&vm_domainset_lock); 3224 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3225 vmd->vmd_minset = 1; 3226 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3227 } 3228 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3229 vmd->vmd_severeset = 1; 3230 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3231 } 3232 mtx_unlock(&vm_domainset_lock); 3233 } 3234 3235 /* 3236 * Clear the domain from the appropriate page level domainset. 3237 */ 3238 void 3239 vm_domain_clear(struct vm_domain *vmd) 3240 { 3241 3242 mtx_lock(&vm_domainset_lock); 3243 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3244 vmd->vmd_minset = 0; 3245 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3246 if (vm_min_waiters != 0) { 3247 vm_min_waiters = 0; 3248 wakeup(&vm_min_domains); 3249 } 3250 } 3251 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3252 vmd->vmd_severeset = 0; 3253 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3254 if (vm_severe_waiters != 0) { 3255 vm_severe_waiters = 0; 3256 wakeup(&vm_severe_domains); 3257 } 3258 } 3259 3260 /* 3261 * If pageout daemon needs pages, then tell it that there are 3262 * some free. 3263 */ 3264 if (vmd->vmd_pageout_pages_needed && 3265 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3266 wakeup(&vmd->vmd_pageout_pages_needed); 3267 vmd->vmd_pageout_pages_needed = 0; 3268 } 3269 3270 /* See comments in vm_wait_doms(). */ 3271 if (vm_pageproc_waiters) { 3272 vm_pageproc_waiters = 0; 3273 wakeup(&vm_pageproc_waiters); 3274 } 3275 mtx_unlock(&vm_domainset_lock); 3276 } 3277 3278 /* 3279 * Wait for free pages to exceed the min threshold globally. 3280 */ 3281 void 3282 vm_wait_min(void) 3283 { 3284 3285 mtx_lock(&vm_domainset_lock); 3286 while (vm_page_count_min()) { 3287 vm_min_waiters++; 3288 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3289 } 3290 mtx_unlock(&vm_domainset_lock); 3291 } 3292 3293 /* 3294 * Wait for free pages to exceed the severe threshold globally. 3295 */ 3296 void 3297 vm_wait_severe(void) 3298 { 3299 3300 mtx_lock(&vm_domainset_lock); 3301 while (vm_page_count_severe()) { 3302 vm_severe_waiters++; 3303 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3304 "vmwait", 0); 3305 } 3306 mtx_unlock(&vm_domainset_lock); 3307 } 3308 3309 u_int 3310 vm_wait_count(void) 3311 { 3312 3313 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3314 } 3315 3316 int 3317 vm_wait_doms(const domainset_t *wdoms, int mflags) 3318 { 3319 int error; 3320 3321 error = 0; 3322 3323 /* 3324 * We use racey wakeup synchronization to avoid expensive global 3325 * locking for the pageproc when sleeping with a non-specific vm_wait. 3326 * To handle this, we only sleep for one tick in this instance. It 3327 * is expected that most allocations for the pageproc will come from 3328 * kmem or vm_page_grab* which will use the more specific and 3329 * race-free vm_wait_domain(). 3330 */ 3331 if (curproc == pageproc) { 3332 mtx_lock(&vm_domainset_lock); 3333 vm_pageproc_waiters++; 3334 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3335 PVM | PDROP | mflags, "pageprocwait", 1); 3336 } else { 3337 /* 3338 * XXX Ideally we would wait only until the allocation could 3339 * be satisfied. This condition can cause new allocators to 3340 * consume all freed pages while old allocators wait. 3341 */ 3342 mtx_lock(&vm_domainset_lock); 3343 if (vm_page_count_min_set(wdoms)) { 3344 vm_min_waiters++; 3345 error = msleep(&vm_min_domains, &vm_domainset_lock, 3346 PVM | PDROP | mflags, "vmwait", 0); 3347 } else 3348 mtx_unlock(&vm_domainset_lock); 3349 } 3350 return (error); 3351 } 3352 3353 /* 3354 * vm_wait_domain: 3355 * 3356 * Sleep until free pages are available for allocation. 3357 * - Called in various places after failed memory allocations. 3358 */ 3359 void 3360 vm_wait_domain(int domain) 3361 { 3362 struct vm_domain *vmd; 3363 domainset_t wdom; 3364 3365 vmd = VM_DOMAIN(domain); 3366 vm_domain_free_assert_unlocked(vmd); 3367 3368 if (curproc == pageproc) { 3369 mtx_lock(&vm_domainset_lock); 3370 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3371 vmd->vmd_pageout_pages_needed = 1; 3372 msleep(&vmd->vmd_pageout_pages_needed, 3373 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3374 } else 3375 mtx_unlock(&vm_domainset_lock); 3376 } else { 3377 if (pageproc == NULL) 3378 panic("vm_wait in early boot"); 3379 DOMAINSET_ZERO(&wdom); 3380 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3381 vm_wait_doms(&wdom, 0); 3382 } 3383 } 3384 3385 static int 3386 vm_wait_flags(vm_object_t obj, int mflags) 3387 { 3388 struct domainset *d; 3389 3390 d = NULL; 3391 3392 /* 3393 * Carefully fetch pointers only once: the struct domainset 3394 * itself is ummutable but the pointer might change. 3395 */ 3396 if (obj != NULL) 3397 d = obj->domain.dr_policy; 3398 if (d == NULL) 3399 d = curthread->td_domain.dr_policy; 3400 3401 return (vm_wait_doms(&d->ds_mask, mflags)); 3402 } 3403 3404 /* 3405 * vm_wait: 3406 * 3407 * Sleep until free pages are available for allocation in the 3408 * affinity domains of the obj. If obj is NULL, the domain set 3409 * for the calling thread is used. 3410 * Called in various places after failed memory allocations. 3411 */ 3412 void 3413 vm_wait(vm_object_t obj) 3414 { 3415 (void)vm_wait_flags(obj, 0); 3416 } 3417 3418 int 3419 vm_wait_intr(vm_object_t obj) 3420 { 3421 return (vm_wait_flags(obj, PCATCH)); 3422 } 3423 3424 /* 3425 * vm_domain_alloc_fail: 3426 * 3427 * Called when a page allocation function fails. Informs the 3428 * pagedaemon and performs the requested wait. Requires the 3429 * domain_free and object lock on entry. Returns with the 3430 * object lock held and free lock released. Returns an error when 3431 * retry is necessary. 3432 * 3433 */ 3434 static int 3435 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3436 { 3437 3438 vm_domain_free_assert_unlocked(vmd); 3439 3440 atomic_add_int(&vmd->vmd_pageout_deficit, 3441 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3442 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3443 if (object != NULL) 3444 VM_OBJECT_WUNLOCK(object); 3445 vm_wait_domain(vmd->vmd_domain); 3446 if (object != NULL) 3447 VM_OBJECT_WLOCK(object); 3448 if (req & VM_ALLOC_WAITOK) 3449 return (EAGAIN); 3450 } 3451 3452 return (0); 3453 } 3454 3455 /* 3456 * vm_waitpfault: 3457 * 3458 * Sleep until free pages are available for allocation. 3459 * - Called only in vm_fault so that processes page faulting 3460 * can be easily tracked. 3461 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3462 * processes will be able to grab memory first. Do not change 3463 * this balance without careful testing first. 3464 */ 3465 void 3466 vm_waitpfault(struct domainset *dset, int timo) 3467 { 3468 3469 /* 3470 * XXX Ideally we would wait only until the allocation could 3471 * be satisfied. This condition can cause new allocators to 3472 * consume all freed pages while old allocators wait. 3473 */ 3474 mtx_lock(&vm_domainset_lock); 3475 if (vm_page_count_min_set(&dset->ds_mask)) { 3476 vm_min_waiters++; 3477 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3478 "pfault", timo); 3479 } else 3480 mtx_unlock(&vm_domainset_lock); 3481 } 3482 3483 static struct vm_pagequeue * 3484 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3485 { 3486 3487 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3488 } 3489 3490 #ifdef INVARIANTS 3491 static struct vm_pagequeue * 3492 vm_page_pagequeue(vm_page_t m) 3493 { 3494 3495 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3496 } 3497 #endif 3498 3499 static __always_inline bool 3500 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3501 { 3502 vm_page_astate_t tmp; 3503 3504 tmp = *old; 3505 do { 3506 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3507 return (true); 3508 counter_u64_add(pqstate_commit_retries, 1); 3509 } while (old->_bits == tmp._bits); 3510 3511 return (false); 3512 } 3513 3514 /* 3515 * Do the work of committing a queue state update that moves the page out of 3516 * its current queue. 3517 */ 3518 static bool 3519 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3520 vm_page_astate_t *old, vm_page_astate_t new) 3521 { 3522 vm_page_t next; 3523 3524 vm_pagequeue_assert_locked(pq); 3525 KASSERT(vm_page_pagequeue(m) == pq, 3526 ("%s: queue %p does not match page %p", __func__, pq, m)); 3527 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3528 ("%s: invalid queue indices %d %d", 3529 __func__, old->queue, new.queue)); 3530 3531 /* 3532 * Once the queue index of the page changes there is nothing 3533 * synchronizing with further updates to the page's physical 3534 * queue state. Therefore we must speculatively remove the page 3535 * from the queue now and be prepared to roll back if the queue 3536 * state update fails. If the page is not physically enqueued then 3537 * we just update its queue index. 3538 */ 3539 if ((old->flags & PGA_ENQUEUED) != 0) { 3540 new.flags &= ~PGA_ENQUEUED; 3541 next = TAILQ_NEXT(m, plinks.q); 3542 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3543 vm_pagequeue_cnt_dec(pq); 3544 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3545 if (next == NULL) 3546 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3547 else 3548 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3549 vm_pagequeue_cnt_inc(pq); 3550 return (false); 3551 } else { 3552 return (true); 3553 } 3554 } else { 3555 return (vm_page_pqstate_fcmpset(m, old, new)); 3556 } 3557 } 3558 3559 static bool 3560 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3561 vm_page_astate_t new) 3562 { 3563 struct vm_pagequeue *pq; 3564 vm_page_astate_t as; 3565 bool ret; 3566 3567 pq = _vm_page_pagequeue(m, old->queue); 3568 3569 /* 3570 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3571 * corresponding page queue lock is held. 3572 */ 3573 vm_pagequeue_lock(pq); 3574 as = vm_page_astate_load(m); 3575 if (__predict_false(as._bits != old->_bits)) { 3576 *old = as; 3577 ret = false; 3578 } else { 3579 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3580 } 3581 vm_pagequeue_unlock(pq); 3582 return (ret); 3583 } 3584 3585 /* 3586 * Commit a queue state update that enqueues or requeues a page. 3587 */ 3588 static bool 3589 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3590 vm_page_astate_t *old, vm_page_astate_t new) 3591 { 3592 struct vm_domain *vmd; 3593 3594 vm_pagequeue_assert_locked(pq); 3595 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3596 ("%s: invalid queue indices %d %d", 3597 __func__, old->queue, new.queue)); 3598 3599 new.flags |= PGA_ENQUEUED; 3600 if (!vm_page_pqstate_fcmpset(m, old, new)) 3601 return (false); 3602 3603 if ((old->flags & PGA_ENQUEUED) != 0) 3604 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3605 else 3606 vm_pagequeue_cnt_inc(pq); 3607 3608 /* 3609 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3610 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3611 * applied, even if it was set first. 3612 */ 3613 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3614 vmd = vm_pagequeue_domain(m); 3615 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3616 ("%s: invalid page queue for page %p", __func__, m)); 3617 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3618 } else { 3619 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3620 } 3621 return (true); 3622 } 3623 3624 /* 3625 * Commit a queue state update that encodes a request for a deferred queue 3626 * operation. 3627 */ 3628 static bool 3629 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3630 vm_page_astate_t new) 3631 { 3632 3633 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3634 ("%s: invalid state, queue %d flags %x", 3635 __func__, new.queue, new.flags)); 3636 3637 if (old->_bits != new._bits && 3638 !vm_page_pqstate_fcmpset(m, old, new)) 3639 return (false); 3640 vm_page_pqbatch_submit(m, new.queue); 3641 return (true); 3642 } 3643 3644 /* 3645 * A generic queue state update function. This handles more cases than the 3646 * specialized functions above. 3647 */ 3648 bool 3649 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3650 { 3651 3652 if (old->_bits == new._bits) 3653 return (true); 3654 3655 if (old->queue != PQ_NONE && new.queue != old->queue) { 3656 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3657 return (false); 3658 if (new.queue != PQ_NONE) 3659 vm_page_pqbatch_submit(m, new.queue); 3660 } else { 3661 if (!vm_page_pqstate_fcmpset(m, old, new)) 3662 return (false); 3663 if (new.queue != PQ_NONE && 3664 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3665 vm_page_pqbatch_submit(m, new.queue); 3666 } 3667 return (true); 3668 } 3669 3670 /* 3671 * Apply deferred queue state updates to a page. 3672 */ 3673 static inline void 3674 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3675 { 3676 vm_page_astate_t new, old; 3677 3678 CRITICAL_ASSERT(curthread); 3679 vm_pagequeue_assert_locked(pq); 3680 KASSERT(queue < PQ_COUNT, 3681 ("%s: invalid queue index %d", __func__, queue)); 3682 KASSERT(pq == _vm_page_pagequeue(m, queue), 3683 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3684 3685 for (old = vm_page_astate_load(m);;) { 3686 if (__predict_false(old.queue != queue || 3687 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3688 counter_u64_add(queue_nops, 1); 3689 break; 3690 } 3691 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3692 ("%s: page %p is unmanaged", __func__, m)); 3693 3694 new = old; 3695 if ((old.flags & PGA_DEQUEUE) != 0) { 3696 new.flags &= ~PGA_QUEUE_OP_MASK; 3697 new.queue = PQ_NONE; 3698 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3699 m, &old, new))) { 3700 counter_u64_add(queue_ops, 1); 3701 break; 3702 } 3703 } else { 3704 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3705 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3706 m, &old, new))) { 3707 counter_u64_add(queue_ops, 1); 3708 break; 3709 } 3710 } 3711 } 3712 } 3713 3714 static void 3715 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3716 uint8_t queue) 3717 { 3718 int i; 3719 3720 for (i = 0; i < bq->bq_cnt; i++) 3721 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3722 vm_batchqueue_init(bq); 3723 } 3724 3725 /* 3726 * vm_page_pqbatch_submit: [ internal use only ] 3727 * 3728 * Enqueue a page in the specified page queue's batched work queue. 3729 * The caller must have encoded the requested operation in the page 3730 * structure's a.flags field. 3731 */ 3732 void 3733 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3734 { 3735 struct vm_batchqueue *bq; 3736 struct vm_pagequeue *pq; 3737 int domain; 3738 3739 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3740 3741 domain = vm_page_domain(m); 3742 critical_enter(); 3743 bq = DPCPU_PTR(pqbatch[domain][queue]); 3744 if (vm_batchqueue_insert(bq, m)) { 3745 critical_exit(); 3746 return; 3747 } 3748 critical_exit(); 3749 3750 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3751 vm_pagequeue_lock(pq); 3752 critical_enter(); 3753 bq = DPCPU_PTR(pqbatch[domain][queue]); 3754 vm_pqbatch_process(pq, bq, queue); 3755 vm_pqbatch_process_page(pq, m, queue); 3756 vm_pagequeue_unlock(pq); 3757 critical_exit(); 3758 } 3759 3760 /* 3761 * vm_page_pqbatch_drain: [ internal use only ] 3762 * 3763 * Force all per-CPU page queue batch queues to be drained. This is 3764 * intended for use in severe memory shortages, to ensure that pages 3765 * do not remain stuck in the batch queues. 3766 */ 3767 void 3768 vm_page_pqbatch_drain(void) 3769 { 3770 struct thread *td; 3771 struct vm_domain *vmd; 3772 struct vm_pagequeue *pq; 3773 int cpu, domain, queue; 3774 3775 td = curthread; 3776 CPU_FOREACH(cpu) { 3777 thread_lock(td); 3778 sched_bind(td, cpu); 3779 thread_unlock(td); 3780 3781 for (domain = 0; domain < vm_ndomains; domain++) { 3782 vmd = VM_DOMAIN(domain); 3783 for (queue = 0; queue < PQ_COUNT; queue++) { 3784 pq = &vmd->vmd_pagequeues[queue]; 3785 vm_pagequeue_lock(pq); 3786 critical_enter(); 3787 vm_pqbatch_process(pq, 3788 DPCPU_PTR(pqbatch[domain][queue]), queue); 3789 critical_exit(); 3790 vm_pagequeue_unlock(pq); 3791 } 3792 } 3793 } 3794 thread_lock(td); 3795 sched_unbind(td); 3796 thread_unlock(td); 3797 } 3798 3799 /* 3800 * vm_page_dequeue_deferred: [ internal use only ] 3801 * 3802 * Request removal of the given page from its current page 3803 * queue. Physical removal from the queue may be deferred 3804 * indefinitely. 3805 */ 3806 void 3807 vm_page_dequeue_deferred(vm_page_t m) 3808 { 3809 vm_page_astate_t new, old; 3810 3811 old = vm_page_astate_load(m); 3812 do { 3813 if (old.queue == PQ_NONE) { 3814 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3815 ("%s: page %p has unexpected queue state", 3816 __func__, m)); 3817 break; 3818 } 3819 new = old; 3820 new.flags |= PGA_DEQUEUE; 3821 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3822 } 3823 3824 /* 3825 * vm_page_dequeue: 3826 * 3827 * Remove the page from whichever page queue it's in, if any, before 3828 * returning. 3829 */ 3830 void 3831 vm_page_dequeue(vm_page_t m) 3832 { 3833 vm_page_astate_t new, old; 3834 3835 old = vm_page_astate_load(m); 3836 do { 3837 if (old.queue == PQ_NONE) { 3838 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3839 ("%s: page %p has unexpected queue state", 3840 __func__, m)); 3841 break; 3842 } 3843 new = old; 3844 new.flags &= ~PGA_QUEUE_OP_MASK; 3845 new.queue = PQ_NONE; 3846 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 3847 3848 } 3849 3850 /* 3851 * Schedule the given page for insertion into the specified page queue. 3852 * Physical insertion of the page may be deferred indefinitely. 3853 */ 3854 static void 3855 vm_page_enqueue(vm_page_t m, uint8_t queue) 3856 { 3857 3858 KASSERT(m->a.queue == PQ_NONE && 3859 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 3860 ("%s: page %p is already enqueued", __func__, m)); 3861 KASSERT(m->ref_count > 0, 3862 ("%s: page %p does not carry any references", __func__, m)); 3863 3864 m->a.queue = queue; 3865 if ((m->a.flags & PGA_REQUEUE) == 0) 3866 vm_page_aflag_set(m, PGA_REQUEUE); 3867 vm_page_pqbatch_submit(m, queue); 3868 } 3869 3870 /* 3871 * vm_page_free_prep: 3872 * 3873 * Prepares the given page to be put on the free list, 3874 * disassociating it from any VM object. The caller may return 3875 * the page to the free list only if this function returns true. 3876 * 3877 * The object, if it exists, must be locked, and then the page must 3878 * be xbusy. Otherwise the page must be not busied. A managed 3879 * page must be unmapped. 3880 */ 3881 static bool 3882 vm_page_free_prep(vm_page_t m) 3883 { 3884 3885 /* 3886 * Synchronize with threads that have dropped a reference to this 3887 * page. 3888 */ 3889 atomic_thread_fence_acq(); 3890 3891 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3892 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3893 uint64_t *p; 3894 int i; 3895 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3896 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3897 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3898 m, i, (uintmax_t)*p)); 3899 } 3900 #endif 3901 if ((m->oflags & VPO_UNMANAGED) == 0) { 3902 KASSERT(!pmap_page_is_mapped(m), 3903 ("vm_page_free_prep: freeing mapped page %p", m)); 3904 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3905 ("vm_page_free_prep: mapping flags set in page %p", m)); 3906 } else { 3907 KASSERT(m->a.queue == PQ_NONE, 3908 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3909 } 3910 VM_CNT_INC(v_tfree); 3911 3912 if (m->object != NULL) { 3913 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 3914 ((m->object->flags & OBJ_UNMANAGED) != 0), 3915 ("vm_page_free_prep: managed flag mismatch for page %p", 3916 m)); 3917 vm_page_assert_xbusied(m); 3918 3919 /* 3920 * The object reference can be released without an atomic 3921 * operation. 3922 */ 3923 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3924 m->ref_count == VPRC_OBJREF, 3925 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3926 m, m->ref_count)); 3927 vm_page_object_remove(m); 3928 m->ref_count -= VPRC_OBJREF; 3929 } else 3930 vm_page_assert_unbusied(m); 3931 3932 vm_page_busy_free(m); 3933 3934 /* 3935 * If fictitious remove object association and 3936 * return. 3937 */ 3938 if ((m->flags & PG_FICTITIOUS) != 0) { 3939 KASSERT(m->ref_count == 1, 3940 ("fictitious page %p is referenced", m)); 3941 KASSERT(m->a.queue == PQ_NONE, 3942 ("fictitious page %p is queued", m)); 3943 return (false); 3944 } 3945 3946 /* 3947 * Pages need not be dequeued before they are returned to the physical 3948 * memory allocator, but they must at least be marked for a deferred 3949 * dequeue. 3950 */ 3951 if ((m->oflags & VPO_UNMANAGED) == 0) 3952 vm_page_dequeue_deferred(m); 3953 3954 m->valid = 0; 3955 vm_page_undirty(m); 3956 3957 if (m->ref_count != 0) 3958 panic("vm_page_free_prep: page %p has references", m); 3959 3960 /* 3961 * Restore the default memory attribute to the page. 3962 */ 3963 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3964 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3965 3966 #if VM_NRESERVLEVEL > 0 3967 /* 3968 * Determine whether the page belongs to a reservation. If the page was 3969 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3970 * as an optimization, we avoid the check in that case. 3971 */ 3972 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3973 return (false); 3974 #endif 3975 3976 return (true); 3977 } 3978 3979 /* 3980 * vm_page_free_toq: 3981 * 3982 * Returns the given page to the free list, disassociating it 3983 * from any VM object. 3984 * 3985 * The object must be locked. The page must be exclusively busied if it 3986 * belongs to an object. 3987 */ 3988 static void 3989 vm_page_free_toq(vm_page_t m) 3990 { 3991 struct vm_domain *vmd; 3992 uma_zone_t zone; 3993 3994 if (!vm_page_free_prep(m)) 3995 return; 3996 3997 vmd = vm_pagequeue_domain(m); 3998 zone = vmd->vmd_pgcache[m->pool].zone; 3999 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4000 uma_zfree(zone, m); 4001 return; 4002 } 4003 vm_domain_free_lock(vmd); 4004 vm_phys_free_pages(m, 0); 4005 vm_domain_free_unlock(vmd); 4006 vm_domain_freecnt_inc(vmd, 1); 4007 } 4008 4009 /* 4010 * vm_page_free_pages_toq: 4011 * 4012 * Returns a list of pages to the free list, disassociating it 4013 * from any VM object. In other words, this is equivalent to 4014 * calling vm_page_free_toq() for each page of a list of VM objects. 4015 */ 4016 void 4017 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4018 { 4019 vm_page_t m; 4020 int count; 4021 4022 if (SLIST_EMPTY(free)) 4023 return; 4024 4025 count = 0; 4026 while ((m = SLIST_FIRST(free)) != NULL) { 4027 count++; 4028 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4029 vm_page_free_toq(m); 4030 } 4031 4032 if (update_wire_count) 4033 vm_wire_sub(count); 4034 } 4035 4036 /* 4037 * Mark this page as wired down. For managed pages, this prevents reclamation 4038 * by the page daemon, or when the containing object, if any, is destroyed. 4039 */ 4040 void 4041 vm_page_wire(vm_page_t m) 4042 { 4043 u_int old; 4044 4045 #ifdef INVARIANTS 4046 if (m->object != NULL && !vm_page_busied(m) && 4047 !vm_object_busied(m->object)) 4048 VM_OBJECT_ASSERT_LOCKED(m->object); 4049 #endif 4050 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4051 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4052 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4053 4054 old = atomic_fetchadd_int(&m->ref_count, 1); 4055 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4056 ("vm_page_wire: counter overflow for page %p", m)); 4057 if (VPRC_WIRE_COUNT(old) == 0) { 4058 if ((m->oflags & VPO_UNMANAGED) == 0) 4059 vm_page_aflag_set(m, PGA_DEQUEUE); 4060 vm_wire_add(1); 4061 } 4062 } 4063 4064 /* 4065 * Attempt to wire a mapped page following a pmap lookup of that page. 4066 * This may fail if a thread is concurrently tearing down mappings of the page. 4067 * The transient failure is acceptable because it translates to the 4068 * failure of the caller pmap_extract_and_hold(), which should be then 4069 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4070 */ 4071 bool 4072 vm_page_wire_mapped(vm_page_t m) 4073 { 4074 u_int old; 4075 4076 old = m->ref_count; 4077 do { 4078 KASSERT(old > 0, 4079 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4080 if ((old & VPRC_BLOCKED) != 0) 4081 return (false); 4082 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4083 4084 if (VPRC_WIRE_COUNT(old) == 0) { 4085 if ((m->oflags & VPO_UNMANAGED) == 0) 4086 vm_page_aflag_set(m, PGA_DEQUEUE); 4087 vm_wire_add(1); 4088 } 4089 return (true); 4090 } 4091 4092 /* 4093 * Release a wiring reference to a managed page. If the page still belongs to 4094 * an object, update its position in the page queues to reflect the reference. 4095 * If the wiring was the last reference to the page, free the page. 4096 */ 4097 static void 4098 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4099 { 4100 u_int old; 4101 4102 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4103 ("%s: page %p is unmanaged", __func__, m)); 4104 4105 /* 4106 * Update LRU state before releasing the wiring reference. 4107 * Use a release store when updating the reference count to 4108 * synchronize with vm_page_free_prep(). 4109 */ 4110 old = m->ref_count; 4111 do { 4112 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4113 ("vm_page_unwire: wire count underflow for page %p", m)); 4114 4115 if (old > VPRC_OBJREF + 1) { 4116 /* 4117 * The page has at least one other wiring reference. An 4118 * earlier iteration of this loop may have called 4119 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4120 * re-set it if necessary. 4121 */ 4122 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4123 vm_page_aflag_set(m, PGA_DEQUEUE); 4124 } else if (old == VPRC_OBJREF + 1) { 4125 /* 4126 * This is the last wiring. Clear PGA_DEQUEUE and 4127 * update the page's queue state to reflect the 4128 * reference. If the page does not belong to an object 4129 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4130 * clear leftover queue state. 4131 */ 4132 vm_page_release_toq(m, nqueue, noreuse); 4133 } else if (old == 1) { 4134 vm_page_aflag_clear(m, PGA_DEQUEUE); 4135 } 4136 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4137 4138 if (VPRC_WIRE_COUNT(old) == 1) { 4139 vm_wire_sub(1); 4140 if (old == 1) 4141 vm_page_free(m); 4142 } 4143 } 4144 4145 /* 4146 * Release one wiring of the specified page, potentially allowing it to be 4147 * paged out. 4148 * 4149 * Only managed pages belonging to an object can be paged out. If the number 4150 * of wirings transitions to zero and the page is eligible for page out, then 4151 * the page is added to the specified paging queue. If the released wiring 4152 * represented the last reference to the page, the page is freed. 4153 */ 4154 void 4155 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4156 { 4157 4158 KASSERT(nqueue < PQ_COUNT, 4159 ("vm_page_unwire: invalid queue %u request for page %p", 4160 nqueue, m)); 4161 4162 if ((m->oflags & VPO_UNMANAGED) != 0) { 4163 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4164 vm_page_free(m); 4165 return; 4166 } 4167 vm_page_unwire_managed(m, nqueue, false); 4168 } 4169 4170 /* 4171 * Unwire a page without (re-)inserting it into a page queue. It is up 4172 * to the caller to enqueue, requeue, or free the page as appropriate. 4173 * In most cases involving managed pages, vm_page_unwire() should be used 4174 * instead. 4175 */ 4176 bool 4177 vm_page_unwire_noq(vm_page_t m) 4178 { 4179 u_int old; 4180 4181 old = vm_page_drop(m, 1); 4182 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4183 ("%s: counter underflow for page %p", __func__, m)); 4184 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4185 ("%s: missing ref on fictitious page %p", __func__, m)); 4186 4187 if (VPRC_WIRE_COUNT(old) > 1) 4188 return (false); 4189 if ((m->oflags & VPO_UNMANAGED) == 0) 4190 vm_page_aflag_clear(m, PGA_DEQUEUE); 4191 vm_wire_sub(1); 4192 return (true); 4193 } 4194 4195 /* 4196 * Ensure that the page ends up in the specified page queue. If the page is 4197 * active or being moved to the active queue, ensure that its act_count is 4198 * at least ACT_INIT but do not otherwise mess with it. 4199 */ 4200 static __always_inline void 4201 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4202 { 4203 vm_page_astate_t old, new; 4204 4205 KASSERT(m->ref_count > 0, 4206 ("%s: page %p does not carry any references", __func__, m)); 4207 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4208 ("%s: invalid flags %x", __func__, nflag)); 4209 4210 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4211 return; 4212 4213 old = vm_page_astate_load(m); 4214 do { 4215 if ((old.flags & PGA_DEQUEUE) != 0) 4216 break; 4217 new = old; 4218 new.flags &= ~PGA_QUEUE_OP_MASK; 4219 if (nqueue == PQ_ACTIVE) 4220 new.act_count = max(old.act_count, ACT_INIT); 4221 if (old.queue == nqueue) { 4222 if (nqueue != PQ_ACTIVE) 4223 new.flags |= nflag; 4224 } else { 4225 new.flags |= nflag; 4226 new.queue = nqueue; 4227 } 4228 } while (!vm_page_pqstate_commit(m, &old, new)); 4229 } 4230 4231 /* 4232 * Put the specified page on the active list (if appropriate). 4233 */ 4234 void 4235 vm_page_activate(vm_page_t m) 4236 { 4237 4238 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4239 } 4240 4241 /* 4242 * Move the specified page to the tail of the inactive queue, or requeue 4243 * the page if it is already in the inactive queue. 4244 */ 4245 void 4246 vm_page_deactivate(vm_page_t m) 4247 { 4248 4249 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4250 } 4251 4252 void 4253 vm_page_deactivate_noreuse(vm_page_t m) 4254 { 4255 4256 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4257 } 4258 4259 /* 4260 * Put a page in the laundry, or requeue it if it is already there. 4261 */ 4262 void 4263 vm_page_launder(vm_page_t m) 4264 { 4265 4266 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4267 } 4268 4269 /* 4270 * Put a page in the PQ_UNSWAPPABLE holding queue. 4271 */ 4272 void 4273 vm_page_unswappable(vm_page_t m) 4274 { 4275 4276 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 4277 ("page %p already unswappable", m)); 4278 4279 vm_page_dequeue(m); 4280 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4281 } 4282 4283 /* 4284 * Release a page back to the page queues in preparation for unwiring. 4285 */ 4286 static void 4287 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4288 { 4289 vm_page_astate_t old, new; 4290 uint16_t nflag; 4291 4292 /* 4293 * Use a check of the valid bits to determine whether we should 4294 * accelerate reclamation of the page. The object lock might not be 4295 * held here, in which case the check is racy. At worst we will either 4296 * accelerate reclamation of a valid page and violate LRU, or 4297 * unnecessarily defer reclamation of an invalid page. 4298 * 4299 * If we were asked to not cache the page, place it near the head of the 4300 * inactive queue so that is reclaimed sooner. 4301 */ 4302 if (noreuse || m->valid == 0) { 4303 nqueue = PQ_INACTIVE; 4304 nflag = PGA_REQUEUE_HEAD; 4305 } else { 4306 nflag = PGA_REQUEUE; 4307 } 4308 4309 old = vm_page_astate_load(m); 4310 do { 4311 new = old; 4312 4313 /* 4314 * If the page is already in the active queue and we are not 4315 * trying to accelerate reclamation, simply mark it as 4316 * referenced and avoid any queue operations. 4317 */ 4318 new.flags &= ~PGA_QUEUE_OP_MASK; 4319 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE) 4320 new.flags |= PGA_REFERENCED; 4321 else { 4322 new.flags |= nflag; 4323 new.queue = nqueue; 4324 } 4325 } while (!vm_page_pqstate_commit(m, &old, new)); 4326 } 4327 4328 /* 4329 * Unwire a page and either attempt to free it or re-add it to the page queues. 4330 */ 4331 void 4332 vm_page_release(vm_page_t m, int flags) 4333 { 4334 vm_object_t object; 4335 4336 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4337 ("vm_page_release: page %p is unmanaged", m)); 4338 4339 if ((flags & VPR_TRYFREE) != 0) { 4340 for (;;) { 4341 object = atomic_load_ptr(&m->object); 4342 if (object == NULL) 4343 break; 4344 /* Depends on type-stability. */ 4345 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4346 break; 4347 if (object == m->object) { 4348 vm_page_release_locked(m, flags); 4349 VM_OBJECT_WUNLOCK(object); 4350 return; 4351 } 4352 VM_OBJECT_WUNLOCK(object); 4353 } 4354 } 4355 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4356 } 4357 4358 /* See vm_page_release(). */ 4359 void 4360 vm_page_release_locked(vm_page_t m, int flags) 4361 { 4362 4363 VM_OBJECT_ASSERT_WLOCKED(m->object); 4364 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4365 ("vm_page_release_locked: page %p is unmanaged", m)); 4366 4367 if (vm_page_unwire_noq(m)) { 4368 if ((flags & VPR_TRYFREE) != 0 && 4369 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4370 m->dirty == 0 && vm_page_tryxbusy(m)) { 4371 /* 4372 * An unlocked lookup may have wired the page before the 4373 * busy lock was acquired, in which case the page must 4374 * not be freed. 4375 */ 4376 if (__predict_true(!vm_page_wired(m))) { 4377 vm_page_free(m); 4378 return; 4379 } 4380 vm_page_xunbusy(m); 4381 } else { 4382 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4383 } 4384 } 4385 } 4386 4387 static bool 4388 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4389 { 4390 u_int old; 4391 4392 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4393 ("vm_page_try_blocked_op: page %p has no object", m)); 4394 KASSERT(vm_page_busied(m), 4395 ("vm_page_try_blocked_op: page %p is not busy", m)); 4396 VM_OBJECT_ASSERT_LOCKED(m->object); 4397 4398 old = m->ref_count; 4399 do { 4400 KASSERT(old != 0, 4401 ("vm_page_try_blocked_op: page %p has no references", m)); 4402 if (VPRC_WIRE_COUNT(old) != 0) 4403 return (false); 4404 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4405 4406 (op)(m); 4407 4408 /* 4409 * If the object is read-locked, new wirings may be created via an 4410 * object lookup. 4411 */ 4412 old = vm_page_drop(m, VPRC_BLOCKED); 4413 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4414 old == (VPRC_BLOCKED | VPRC_OBJREF), 4415 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4416 old, m)); 4417 return (true); 4418 } 4419 4420 /* 4421 * Atomically check for wirings and remove all mappings of the page. 4422 */ 4423 bool 4424 vm_page_try_remove_all(vm_page_t m) 4425 { 4426 4427 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4428 } 4429 4430 /* 4431 * Atomically check for wirings and remove all writeable mappings of the page. 4432 */ 4433 bool 4434 vm_page_try_remove_write(vm_page_t m) 4435 { 4436 4437 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4438 } 4439 4440 /* 4441 * vm_page_advise 4442 * 4443 * Apply the specified advice to the given page. 4444 */ 4445 void 4446 vm_page_advise(vm_page_t m, int advice) 4447 { 4448 4449 VM_OBJECT_ASSERT_WLOCKED(m->object); 4450 vm_page_assert_xbusied(m); 4451 4452 if (advice == MADV_FREE) 4453 /* 4454 * Mark the page clean. This will allow the page to be freed 4455 * without first paging it out. MADV_FREE pages are often 4456 * quickly reused by malloc(3), so we do not do anything that 4457 * would result in a page fault on a later access. 4458 */ 4459 vm_page_undirty(m); 4460 else if (advice != MADV_DONTNEED) { 4461 if (advice == MADV_WILLNEED) 4462 vm_page_activate(m); 4463 return; 4464 } 4465 4466 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4467 vm_page_dirty(m); 4468 4469 /* 4470 * Clear any references to the page. Otherwise, the page daemon will 4471 * immediately reactivate the page. 4472 */ 4473 vm_page_aflag_clear(m, PGA_REFERENCED); 4474 4475 /* 4476 * Place clean pages near the head of the inactive queue rather than 4477 * the tail, thus defeating the queue's LRU operation and ensuring that 4478 * the page will be reused quickly. Dirty pages not already in the 4479 * laundry are moved there. 4480 */ 4481 if (m->dirty == 0) 4482 vm_page_deactivate_noreuse(m); 4483 else if (!vm_page_in_laundry(m)) 4484 vm_page_launder(m); 4485 } 4486 4487 /* 4488 * vm_page_grab_release 4489 * 4490 * Helper routine for grab functions to release busy on return. 4491 */ 4492 static inline void 4493 vm_page_grab_release(vm_page_t m, int allocflags) 4494 { 4495 4496 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4497 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4498 vm_page_sunbusy(m); 4499 else 4500 vm_page_xunbusy(m); 4501 } 4502 } 4503 4504 /* 4505 * vm_page_grab_sleep 4506 * 4507 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4508 * if the caller should retry and false otherwise. 4509 * 4510 * If the object is locked on entry the object will be unlocked with 4511 * false returns and still locked but possibly having been dropped 4512 * with true returns. 4513 */ 4514 static bool 4515 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4516 const char *wmesg, int allocflags, bool locked) 4517 { 4518 4519 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4520 return (false); 4521 4522 /* 4523 * Reference the page before unlocking and sleeping so that 4524 * the page daemon is less likely to reclaim it. 4525 */ 4526 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4527 vm_page_reference(m); 4528 4529 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4530 locked) 4531 VM_OBJECT_WLOCK(object); 4532 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4533 return (false); 4534 4535 return (true); 4536 } 4537 4538 /* 4539 * Assert that the grab flags are valid. 4540 */ 4541 static inline void 4542 vm_page_grab_check(int allocflags) 4543 { 4544 4545 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4546 (allocflags & VM_ALLOC_WIRED) != 0, 4547 ("vm_page_grab*: the pages must be busied or wired")); 4548 4549 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4550 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4551 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4552 } 4553 4554 /* 4555 * Calculate the page allocation flags for grab. 4556 */ 4557 static inline int 4558 vm_page_grab_pflags(int allocflags) 4559 { 4560 int pflags; 4561 4562 pflags = allocflags & 4563 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4564 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY); 4565 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4566 pflags |= VM_ALLOC_WAITFAIL; 4567 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4568 pflags |= VM_ALLOC_SBUSY; 4569 4570 return (pflags); 4571 } 4572 4573 /* 4574 * Grab a page, waiting until we are waken up due to the page 4575 * changing state. We keep on waiting, if the page continues 4576 * to be in the object. If the page doesn't exist, first allocate it 4577 * and then conditionally zero it. 4578 * 4579 * This routine may sleep. 4580 * 4581 * The object must be locked on entry. The lock will, however, be released 4582 * and reacquired if the routine sleeps. 4583 */ 4584 vm_page_t 4585 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4586 { 4587 vm_page_t m; 4588 4589 VM_OBJECT_ASSERT_WLOCKED(object); 4590 vm_page_grab_check(allocflags); 4591 4592 retrylookup: 4593 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4594 if (!vm_page_tryacquire(m, allocflags)) { 4595 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4596 allocflags, true)) 4597 goto retrylookup; 4598 return (NULL); 4599 } 4600 goto out; 4601 } 4602 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4603 return (NULL); 4604 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); 4605 if (m == NULL) { 4606 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4607 return (NULL); 4608 goto retrylookup; 4609 } 4610 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4611 pmap_zero_page(m); 4612 4613 out: 4614 vm_page_grab_release(m, allocflags); 4615 4616 return (m); 4617 } 4618 4619 /* 4620 * Locklessly attempt to acquire a page given a (object, pindex) tuple 4621 * and an optional previous page to avoid the radix lookup. The resulting 4622 * page will be validated against the identity tuple and busied or wired 4623 * as requested. A NULL *mp return guarantees that the page was not in 4624 * radix at the time of the call but callers must perform higher level 4625 * synchronization or retry the operation under a lock if they require 4626 * an atomic answer. This is the only lock free validation routine, 4627 * other routines can depend on the resulting page state. 4628 * 4629 * The return value indicates whether the operation failed due to caller 4630 * flags. The return is tri-state with mp: 4631 * 4632 * (true, *mp != NULL) - The operation was successful. 4633 * (true, *mp == NULL) - The page was not found in tree. 4634 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition. 4635 */ 4636 static bool 4637 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, 4638 vm_page_t prev, vm_page_t *mp, int allocflags) 4639 { 4640 vm_page_t m; 4641 4642 vm_page_grab_check(allocflags); 4643 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev)); 4644 4645 *mp = NULL; 4646 for (;;) { 4647 /* 4648 * We may see a false NULL here because the previous page 4649 * has been removed or just inserted and the list is loaded 4650 * without barriers. Switch to radix to verify. 4651 */ 4652 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL || 4653 QMD_IS_TRASHED(m) || m->pindex != pindex || 4654 atomic_load_ptr(&m->object) != object) { 4655 prev = NULL; 4656 /* 4657 * This guarantees the result is instantaneously 4658 * correct. 4659 */ 4660 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 4661 } 4662 if (m == NULL) 4663 return (true); 4664 if (vm_page_trybusy(m, allocflags)) { 4665 if (m->object == object && m->pindex == pindex) 4666 break; 4667 /* relookup. */ 4668 vm_page_busy_release(m); 4669 cpu_spinwait(); 4670 continue; 4671 } 4672 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4673 allocflags, false)) 4674 return (false); 4675 } 4676 if ((allocflags & VM_ALLOC_WIRED) != 0) 4677 vm_page_wire(m); 4678 vm_page_grab_release(m, allocflags); 4679 *mp = m; 4680 return (true); 4681 } 4682 4683 /* 4684 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4685 * is not set. 4686 */ 4687 vm_page_t 4688 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4689 { 4690 vm_page_t m; 4691 4692 vm_page_grab_check(allocflags); 4693 4694 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags)) 4695 return (NULL); 4696 if (m != NULL) 4697 return (m); 4698 4699 /* 4700 * The radix lockless lookup should never return a false negative 4701 * errors. If the user specifies NOCREAT they are guaranteed there 4702 * was no page present at the instant of the call. A NOCREAT caller 4703 * must handle create races gracefully. 4704 */ 4705 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4706 return (NULL); 4707 4708 VM_OBJECT_WLOCK(object); 4709 m = vm_page_grab(object, pindex, allocflags); 4710 VM_OBJECT_WUNLOCK(object); 4711 4712 return (m); 4713 } 4714 4715 /* 4716 * Grab a page and make it valid, paging in if necessary. Pages missing from 4717 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4718 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4719 * in simultaneously. Additional pages will be left on a paging queue but 4720 * will neither be wired nor busy regardless of allocflags. 4721 */ 4722 int 4723 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4724 { 4725 vm_page_t m; 4726 vm_page_t ma[VM_INITIAL_PAGEIN]; 4727 int after, i, pflags, rv; 4728 4729 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4730 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4731 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4732 KASSERT((allocflags & 4733 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4734 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4735 VM_OBJECT_ASSERT_WLOCKED(object); 4736 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4737 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4738 pflags |= VM_ALLOC_WAITFAIL; 4739 4740 retrylookup: 4741 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4742 /* 4743 * If the page is fully valid it can only become invalid 4744 * with the object lock held. If it is not valid it can 4745 * become valid with the busy lock held. Therefore, we 4746 * may unnecessarily lock the exclusive busy here if we 4747 * race with I/O completion not using the object lock. 4748 * However, we will not end up with an invalid page and a 4749 * shared lock. 4750 */ 4751 if (!vm_page_trybusy(m, 4752 vm_page_all_valid(m) ? allocflags : 0)) { 4753 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4754 allocflags, true); 4755 goto retrylookup; 4756 } 4757 if (vm_page_all_valid(m)) 4758 goto out; 4759 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4760 vm_page_busy_release(m); 4761 *mp = NULL; 4762 return (VM_PAGER_FAIL); 4763 } 4764 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4765 *mp = NULL; 4766 return (VM_PAGER_FAIL); 4767 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { 4768 goto retrylookup; 4769 } 4770 4771 vm_page_assert_xbusied(m); 4772 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4773 after = MIN(after, VM_INITIAL_PAGEIN); 4774 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4775 after = MAX(after, 1); 4776 ma[0] = m; 4777 for (i = 1; i < after; i++) { 4778 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 4779 if (ma[i]->valid || !vm_page_tryxbusy(ma[i])) 4780 break; 4781 } else { 4782 ma[i] = vm_page_alloc(object, m->pindex + i, 4783 VM_ALLOC_NORMAL); 4784 if (ma[i] == NULL) 4785 break; 4786 } 4787 } 4788 after = i; 4789 vm_object_pip_add(object, after); 4790 VM_OBJECT_WUNLOCK(object); 4791 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 4792 VM_OBJECT_WLOCK(object); 4793 vm_object_pip_wakeupn(object, after); 4794 /* Pager may have replaced a page. */ 4795 m = ma[0]; 4796 if (rv != VM_PAGER_OK) { 4797 for (i = 0; i < after; i++) { 4798 if (!vm_page_wired(ma[i])) 4799 vm_page_free(ma[i]); 4800 else 4801 vm_page_xunbusy(ma[i]); 4802 } 4803 *mp = NULL; 4804 return (rv); 4805 } 4806 for (i = 1; i < after; i++) 4807 vm_page_readahead_finish(ma[i]); 4808 MPASS(vm_page_all_valid(m)); 4809 } else { 4810 vm_page_zero_invalid(m, TRUE); 4811 } 4812 out: 4813 if ((allocflags & VM_ALLOC_WIRED) != 0) 4814 vm_page_wire(m); 4815 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 4816 vm_page_busy_downgrade(m); 4817 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 4818 vm_page_busy_release(m); 4819 *mp = m; 4820 return (VM_PAGER_OK); 4821 } 4822 4823 /* 4824 * Locklessly grab a valid page. If the page is not valid or not yet 4825 * allocated this will fall back to the object lock method. 4826 */ 4827 int 4828 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 4829 vm_pindex_t pindex, int allocflags) 4830 { 4831 vm_page_t m; 4832 int flags; 4833 int error; 4834 4835 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4836 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4837 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 4838 "mismatch")); 4839 KASSERT((allocflags & 4840 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4841 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 4842 4843 /* 4844 * Attempt a lockless lookup and busy. We need at least an sbusy 4845 * before we can inspect the valid field and return a wired page. 4846 */ 4847 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 4848 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags)) 4849 return (VM_PAGER_FAIL); 4850 if ((m = *mp) != NULL) { 4851 if (vm_page_all_valid(m)) { 4852 if ((allocflags & VM_ALLOC_WIRED) != 0) 4853 vm_page_wire(m); 4854 vm_page_grab_release(m, allocflags); 4855 return (VM_PAGER_OK); 4856 } 4857 vm_page_busy_release(m); 4858 } 4859 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4860 *mp = NULL; 4861 return (VM_PAGER_FAIL); 4862 } 4863 VM_OBJECT_WLOCK(object); 4864 error = vm_page_grab_valid(mp, object, pindex, allocflags); 4865 VM_OBJECT_WUNLOCK(object); 4866 4867 return (error); 4868 } 4869 4870 /* 4871 * Return the specified range of pages from the given object. For each 4872 * page offset within the range, if a page already exists within the object 4873 * at that offset and it is busy, then wait for it to change state. If, 4874 * instead, the page doesn't exist, then allocate it. 4875 * 4876 * The caller must always specify an allocation class. 4877 * 4878 * allocation classes: 4879 * VM_ALLOC_NORMAL normal process request 4880 * VM_ALLOC_SYSTEM system *really* needs the pages 4881 * 4882 * The caller must always specify that the pages are to be busied and/or 4883 * wired. 4884 * 4885 * optional allocation flags: 4886 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4887 * VM_ALLOC_NOBUSY do not exclusive busy the page 4888 * VM_ALLOC_NOWAIT do not sleep 4889 * VM_ALLOC_SBUSY set page to sbusy state 4890 * VM_ALLOC_WIRED wire the pages 4891 * VM_ALLOC_ZERO zero and validate any invalid pages 4892 * 4893 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4894 * may return a partial prefix of the requested range. 4895 */ 4896 int 4897 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4898 vm_page_t *ma, int count) 4899 { 4900 vm_page_t m, mpred; 4901 int pflags; 4902 int i; 4903 4904 VM_OBJECT_ASSERT_WLOCKED(object); 4905 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4906 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4907 KASSERT(count > 0, 4908 ("vm_page_grab_pages: invalid page count %d", count)); 4909 vm_page_grab_check(allocflags); 4910 4911 pflags = vm_page_grab_pflags(allocflags); 4912 i = 0; 4913 retrylookup: 4914 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4915 if (m == NULL || m->pindex != pindex + i) { 4916 mpred = m; 4917 m = NULL; 4918 } else 4919 mpred = TAILQ_PREV(m, pglist, listq); 4920 for (; i < count; i++) { 4921 if (m != NULL) { 4922 if (!vm_page_tryacquire(m, allocflags)) { 4923 if (vm_page_grab_sleep(object, m, pindex + i, 4924 "grbmaw", allocflags, true)) 4925 goto retrylookup; 4926 break; 4927 } 4928 } else { 4929 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4930 break; 4931 m = vm_page_alloc_after(object, pindex + i, 4932 pflags | VM_ALLOC_COUNT(count - i), mpred); 4933 if (m == NULL) { 4934 if ((allocflags & (VM_ALLOC_NOWAIT | 4935 VM_ALLOC_WAITFAIL)) != 0) 4936 break; 4937 goto retrylookup; 4938 } 4939 } 4940 if (vm_page_none_valid(m) && 4941 (allocflags & VM_ALLOC_ZERO) != 0) { 4942 if ((m->flags & PG_ZERO) == 0) 4943 pmap_zero_page(m); 4944 vm_page_valid(m); 4945 } 4946 vm_page_grab_release(m, allocflags); 4947 ma[i] = mpred = m; 4948 m = vm_page_next(m); 4949 } 4950 return (i); 4951 } 4952 4953 /* 4954 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 4955 * and will fall back to the locked variant to handle allocation. 4956 */ 4957 int 4958 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 4959 int allocflags, vm_page_t *ma, int count) 4960 { 4961 vm_page_t m, pred; 4962 int flags; 4963 int i; 4964 4965 KASSERT(count > 0, 4966 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 4967 vm_page_grab_check(allocflags); 4968 4969 /* 4970 * Modify flags for lockless acquire to hold the page until we 4971 * set it valid if necessary. 4972 */ 4973 flags = allocflags & ~VM_ALLOC_NOBUSY; 4974 pred = NULL; 4975 for (i = 0; i < count; i++, pindex++) { 4976 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags)) 4977 return (i); 4978 if (m == NULL) 4979 break; 4980 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 4981 if ((m->flags & PG_ZERO) == 0) 4982 pmap_zero_page(m); 4983 vm_page_valid(m); 4984 } 4985 /* m will still be wired or busy according to flags. */ 4986 vm_page_grab_release(m, allocflags); 4987 pred = ma[i] = m; 4988 } 4989 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 4990 return (i); 4991 count -= i; 4992 VM_OBJECT_WLOCK(object); 4993 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 4994 VM_OBJECT_WUNLOCK(object); 4995 4996 return (i); 4997 } 4998 4999 /* 5000 * Mapping function for valid or dirty bits in a page. 5001 * 5002 * Inputs are required to range within a page. 5003 */ 5004 vm_page_bits_t 5005 vm_page_bits(int base, int size) 5006 { 5007 int first_bit; 5008 int last_bit; 5009 5010 KASSERT( 5011 base + size <= PAGE_SIZE, 5012 ("vm_page_bits: illegal base/size %d/%d", base, size) 5013 ); 5014 5015 if (size == 0) /* handle degenerate case */ 5016 return (0); 5017 5018 first_bit = base >> DEV_BSHIFT; 5019 last_bit = (base + size - 1) >> DEV_BSHIFT; 5020 5021 return (((vm_page_bits_t)2 << last_bit) - 5022 ((vm_page_bits_t)1 << first_bit)); 5023 } 5024 5025 void 5026 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5027 { 5028 5029 #if PAGE_SIZE == 32768 5030 atomic_set_64((uint64_t *)bits, set); 5031 #elif PAGE_SIZE == 16384 5032 atomic_set_32((uint32_t *)bits, set); 5033 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5034 atomic_set_16((uint16_t *)bits, set); 5035 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5036 atomic_set_8((uint8_t *)bits, set); 5037 #else /* PAGE_SIZE <= 8192 */ 5038 uintptr_t addr; 5039 int shift; 5040 5041 addr = (uintptr_t)bits; 5042 /* 5043 * Use a trick to perform a 32-bit atomic on the 5044 * containing aligned word, to not depend on the existence 5045 * of atomic_{set, clear}_{8, 16}. 5046 */ 5047 shift = addr & (sizeof(uint32_t) - 1); 5048 #if BYTE_ORDER == BIG_ENDIAN 5049 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5050 #else 5051 shift *= NBBY; 5052 #endif 5053 addr &= ~(sizeof(uint32_t) - 1); 5054 atomic_set_32((uint32_t *)addr, set << shift); 5055 #endif /* PAGE_SIZE */ 5056 } 5057 5058 static inline void 5059 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5060 { 5061 5062 #if PAGE_SIZE == 32768 5063 atomic_clear_64((uint64_t *)bits, clear); 5064 #elif PAGE_SIZE == 16384 5065 atomic_clear_32((uint32_t *)bits, clear); 5066 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5067 atomic_clear_16((uint16_t *)bits, clear); 5068 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5069 atomic_clear_8((uint8_t *)bits, clear); 5070 #else /* PAGE_SIZE <= 8192 */ 5071 uintptr_t addr; 5072 int shift; 5073 5074 addr = (uintptr_t)bits; 5075 /* 5076 * Use a trick to perform a 32-bit atomic on the 5077 * containing aligned word, to not depend on the existence 5078 * of atomic_{set, clear}_{8, 16}. 5079 */ 5080 shift = addr & (sizeof(uint32_t) - 1); 5081 #if BYTE_ORDER == BIG_ENDIAN 5082 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5083 #else 5084 shift *= NBBY; 5085 #endif 5086 addr &= ~(sizeof(uint32_t) - 1); 5087 atomic_clear_32((uint32_t *)addr, clear << shift); 5088 #endif /* PAGE_SIZE */ 5089 } 5090 5091 static inline vm_page_bits_t 5092 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5093 { 5094 #if PAGE_SIZE == 32768 5095 uint64_t old; 5096 5097 old = *bits; 5098 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5099 return (old); 5100 #elif PAGE_SIZE == 16384 5101 uint32_t old; 5102 5103 old = *bits; 5104 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5105 return (old); 5106 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5107 uint16_t old; 5108 5109 old = *bits; 5110 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5111 return (old); 5112 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5113 uint8_t old; 5114 5115 old = *bits; 5116 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5117 return (old); 5118 #else /* PAGE_SIZE <= 4096*/ 5119 uintptr_t addr; 5120 uint32_t old, new, mask; 5121 int shift; 5122 5123 addr = (uintptr_t)bits; 5124 /* 5125 * Use a trick to perform a 32-bit atomic on the 5126 * containing aligned word, to not depend on the existence 5127 * of atomic_{set, swap, clear}_{8, 16}. 5128 */ 5129 shift = addr & (sizeof(uint32_t) - 1); 5130 #if BYTE_ORDER == BIG_ENDIAN 5131 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5132 #else 5133 shift *= NBBY; 5134 #endif 5135 addr &= ~(sizeof(uint32_t) - 1); 5136 mask = VM_PAGE_BITS_ALL << shift; 5137 5138 old = *bits; 5139 do { 5140 new = old & ~mask; 5141 new |= newbits << shift; 5142 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5143 return (old >> shift); 5144 #endif /* PAGE_SIZE */ 5145 } 5146 5147 /* 5148 * vm_page_set_valid_range: 5149 * 5150 * Sets portions of a page valid. The arguments are expected 5151 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5152 * of any partial chunks touched by the range. The invalid portion of 5153 * such chunks will be zeroed. 5154 * 5155 * (base + size) must be less then or equal to PAGE_SIZE. 5156 */ 5157 void 5158 vm_page_set_valid_range(vm_page_t m, int base, int size) 5159 { 5160 int endoff, frag; 5161 vm_page_bits_t pagebits; 5162 5163 vm_page_assert_busied(m); 5164 if (size == 0) /* handle degenerate case */ 5165 return; 5166 5167 /* 5168 * If the base is not DEV_BSIZE aligned and the valid 5169 * bit is clear, we have to zero out a portion of the 5170 * first block. 5171 */ 5172 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5173 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5174 pmap_zero_page_area(m, frag, base - frag); 5175 5176 /* 5177 * If the ending offset is not DEV_BSIZE aligned and the 5178 * valid bit is clear, we have to zero out a portion of 5179 * the last block. 5180 */ 5181 endoff = base + size; 5182 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5183 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5184 pmap_zero_page_area(m, endoff, 5185 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5186 5187 /* 5188 * Assert that no previously invalid block that is now being validated 5189 * is already dirty. 5190 */ 5191 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5192 ("vm_page_set_valid_range: page %p is dirty", m)); 5193 5194 /* 5195 * Set valid bits inclusive of any overlap. 5196 */ 5197 pagebits = vm_page_bits(base, size); 5198 if (vm_page_xbusied(m)) 5199 m->valid |= pagebits; 5200 else 5201 vm_page_bits_set(m, &m->valid, pagebits); 5202 } 5203 5204 /* 5205 * Set the page dirty bits and free the invalid swap space if 5206 * present. Returns the previous dirty bits. 5207 */ 5208 vm_page_bits_t 5209 vm_page_set_dirty(vm_page_t m) 5210 { 5211 vm_page_bits_t old; 5212 5213 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5214 5215 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5216 old = m->dirty; 5217 m->dirty = VM_PAGE_BITS_ALL; 5218 } else 5219 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5220 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5221 vm_pager_page_unswapped(m); 5222 5223 return (old); 5224 } 5225 5226 /* 5227 * Clear the given bits from the specified page's dirty field. 5228 */ 5229 static __inline void 5230 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5231 { 5232 5233 vm_page_assert_busied(m); 5234 5235 /* 5236 * If the page is xbusied and not write mapped we are the 5237 * only thread that can modify dirty bits. Otherwise, The pmap 5238 * layer can call vm_page_dirty() without holding a distinguished 5239 * lock. The combination of page busy and atomic operations 5240 * suffice to guarantee consistency of the page dirty field. 5241 */ 5242 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5243 m->dirty &= ~pagebits; 5244 else 5245 vm_page_bits_clear(m, &m->dirty, pagebits); 5246 } 5247 5248 /* 5249 * vm_page_set_validclean: 5250 * 5251 * Sets portions of a page valid and clean. The arguments are expected 5252 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5253 * of any partial chunks touched by the range. The invalid portion of 5254 * such chunks will be zero'd. 5255 * 5256 * (base + size) must be less then or equal to PAGE_SIZE. 5257 */ 5258 void 5259 vm_page_set_validclean(vm_page_t m, int base, int size) 5260 { 5261 vm_page_bits_t oldvalid, pagebits; 5262 int endoff, frag; 5263 5264 vm_page_assert_busied(m); 5265 if (size == 0) /* handle degenerate case */ 5266 return; 5267 5268 /* 5269 * If the base is not DEV_BSIZE aligned and the valid 5270 * bit is clear, we have to zero out a portion of the 5271 * first block. 5272 */ 5273 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5274 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5275 pmap_zero_page_area(m, frag, base - frag); 5276 5277 /* 5278 * If the ending offset is not DEV_BSIZE aligned and the 5279 * valid bit is clear, we have to zero out a portion of 5280 * the last block. 5281 */ 5282 endoff = base + size; 5283 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5284 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5285 pmap_zero_page_area(m, endoff, 5286 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5287 5288 /* 5289 * Set valid, clear dirty bits. If validating the entire 5290 * page we can safely clear the pmap modify bit. We also 5291 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5292 * takes a write fault on a MAP_NOSYNC memory area the flag will 5293 * be set again. 5294 * 5295 * We set valid bits inclusive of any overlap, but we can only 5296 * clear dirty bits for DEV_BSIZE chunks that are fully within 5297 * the range. 5298 */ 5299 oldvalid = m->valid; 5300 pagebits = vm_page_bits(base, size); 5301 if (vm_page_xbusied(m)) 5302 m->valid |= pagebits; 5303 else 5304 vm_page_bits_set(m, &m->valid, pagebits); 5305 #if 0 /* NOT YET */ 5306 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5307 frag = DEV_BSIZE - frag; 5308 base += frag; 5309 size -= frag; 5310 if (size < 0) 5311 size = 0; 5312 } 5313 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5314 #endif 5315 if (base == 0 && size == PAGE_SIZE) { 5316 /* 5317 * The page can only be modified within the pmap if it is 5318 * mapped, and it can only be mapped if it was previously 5319 * fully valid. 5320 */ 5321 if (oldvalid == VM_PAGE_BITS_ALL) 5322 /* 5323 * Perform the pmap_clear_modify() first. Otherwise, 5324 * a concurrent pmap operation, such as 5325 * pmap_protect(), could clear a modification in the 5326 * pmap and set the dirty field on the page before 5327 * pmap_clear_modify() had begun and after the dirty 5328 * field was cleared here. 5329 */ 5330 pmap_clear_modify(m); 5331 m->dirty = 0; 5332 vm_page_aflag_clear(m, PGA_NOSYNC); 5333 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5334 m->dirty &= ~pagebits; 5335 else 5336 vm_page_clear_dirty_mask(m, pagebits); 5337 } 5338 5339 void 5340 vm_page_clear_dirty(vm_page_t m, int base, int size) 5341 { 5342 5343 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5344 } 5345 5346 /* 5347 * vm_page_set_invalid: 5348 * 5349 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5350 * valid and dirty bits for the effected areas are cleared. 5351 */ 5352 void 5353 vm_page_set_invalid(vm_page_t m, int base, int size) 5354 { 5355 vm_page_bits_t bits; 5356 vm_object_t object; 5357 5358 /* 5359 * The object lock is required so that pages can't be mapped 5360 * read-only while we're in the process of invalidating them. 5361 */ 5362 object = m->object; 5363 VM_OBJECT_ASSERT_WLOCKED(object); 5364 vm_page_assert_busied(m); 5365 5366 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5367 size >= object->un_pager.vnp.vnp_size) 5368 bits = VM_PAGE_BITS_ALL; 5369 else 5370 bits = vm_page_bits(base, size); 5371 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5372 pmap_remove_all(m); 5373 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5374 !pmap_page_is_mapped(m), 5375 ("vm_page_set_invalid: page %p is mapped", m)); 5376 if (vm_page_xbusied(m)) { 5377 m->valid &= ~bits; 5378 m->dirty &= ~bits; 5379 } else { 5380 vm_page_bits_clear(m, &m->valid, bits); 5381 vm_page_bits_clear(m, &m->dirty, bits); 5382 } 5383 } 5384 5385 /* 5386 * vm_page_invalid: 5387 * 5388 * Invalidates the entire page. The page must be busy, unmapped, and 5389 * the enclosing object must be locked. The object locks protects 5390 * against concurrent read-only pmap enter which is done without 5391 * busy. 5392 */ 5393 void 5394 vm_page_invalid(vm_page_t m) 5395 { 5396 5397 vm_page_assert_busied(m); 5398 VM_OBJECT_ASSERT_LOCKED(m->object); 5399 MPASS(!pmap_page_is_mapped(m)); 5400 5401 if (vm_page_xbusied(m)) 5402 m->valid = 0; 5403 else 5404 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5405 } 5406 5407 /* 5408 * vm_page_zero_invalid() 5409 * 5410 * The kernel assumes that the invalid portions of a page contain 5411 * garbage, but such pages can be mapped into memory by user code. 5412 * When this occurs, we must zero out the non-valid portions of the 5413 * page so user code sees what it expects. 5414 * 5415 * Pages are most often semi-valid when the end of a file is mapped 5416 * into memory and the file's size is not page aligned. 5417 */ 5418 void 5419 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5420 { 5421 int b; 5422 int i; 5423 5424 /* 5425 * Scan the valid bits looking for invalid sections that 5426 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5427 * valid bit may be set ) have already been zeroed by 5428 * vm_page_set_validclean(). 5429 */ 5430 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5431 if (i == (PAGE_SIZE / DEV_BSIZE) || 5432 (m->valid & ((vm_page_bits_t)1 << i))) { 5433 if (i > b) { 5434 pmap_zero_page_area(m, 5435 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5436 } 5437 b = i + 1; 5438 } 5439 } 5440 5441 /* 5442 * setvalid is TRUE when we can safely set the zero'd areas 5443 * as being valid. We can do this if there are no cache consistancy 5444 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5445 */ 5446 if (setvalid) 5447 vm_page_valid(m); 5448 } 5449 5450 /* 5451 * vm_page_is_valid: 5452 * 5453 * Is (partial) page valid? Note that the case where size == 0 5454 * will return FALSE in the degenerate case where the page is 5455 * entirely invalid, and TRUE otherwise. 5456 * 5457 * Some callers envoke this routine without the busy lock held and 5458 * handle races via higher level locks. Typical callers should 5459 * hold a busy lock to prevent invalidation. 5460 */ 5461 int 5462 vm_page_is_valid(vm_page_t m, int base, int size) 5463 { 5464 vm_page_bits_t bits; 5465 5466 bits = vm_page_bits(base, size); 5467 return (m->valid != 0 && (m->valid & bits) == bits); 5468 } 5469 5470 /* 5471 * Returns true if all of the specified predicates are true for the entire 5472 * (super)page and false otherwise. 5473 */ 5474 bool 5475 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 5476 { 5477 vm_object_t object; 5478 int i, npages; 5479 5480 object = m->object; 5481 if (skip_m != NULL && skip_m->object != object) 5482 return (false); 5483 VM_OBJECT_ASSERT_LOCKED(object); 5484 npages = atop(pagesizes[m->psind]); 5485 5486 /* 5487 * The physically contiguous pages that make up a superpage, i.e., a 5488 * page with a page size index ("psind") greater than zero, will 5489 * occupy adjacent entries in vm_page_array[]. 5490 */ 5491 for (i = 0; i < npages; i++) { 5492 /* Always test object consistency, including "skip_m". */ 5493 if (m[i].object != object) 5494 return (false); 5495 if (&m[i] == skip_m) 5496 continue; 5497 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5498 return (false); 5499 if ((flags & PS_ALL_DIRTY) != 0) { 5500 /* 5501 * Calling vm_page_test_dirty() or pmap_is_modified() 5502 * might stop this case from spuriously returning 5503 * "false". However, that would require a write lock 5504 * on the object containing "m[i]". 5505 */ 5506 if (m[i].dirty != VM_PAGE_BITS_ALL) 5507 return (false); 5508 } 5509 if ((flags & PS_ALL_VALID) != 0 && 5510 m[i].valid != VM_PAGE_BITS_ALL) 5511 return (false); 5512 } 5513 return (true); 5514 } 5515 5516 /* 5517 * Set the page's dirty bits if the page is modified. 5518 */ 5519 void 5520 vm_page_test_dirty(vm_page_t m) 5521 { 5522 5523 vm_page_assert_busied(m); 5524 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5525 vm_page_dirty(m); 5526 } 5527 5528 void 5529 vm_page_valid(vm_page_t m) 5530 { 5531 5532 vm_page_assert_busied(m); 5533 if (vm_page_xbusied(m)) 5534 m->valid = VM_PAGE_BITS_ALL; 5535 else 5536 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5537 } 5538 5539 void 5540 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5541 { 5542 5543 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5544 } 5545 5546 void 5547 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5548 { 5549 5550 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5551 } 5552 5553 int 5554 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5555 { 5556 5557 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5558 } 5559 5560 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5561 void 5562 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5563 { 5564 5565 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5566 } 5567 5568 void 5569 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5570 { 5571 5572 mtx_assert_(vm_page_lockptr(m), a, file, line); 5573 } 5574 #endif 5575 5576 #ifdef INVARIANTS 5577 void 5578 vm_page_object_busy_assert(vm_page_t m) 5579 { 5580 5581 /* 5582 * Certain of the page's fields may only be modified by the 5583 * holder of a page or object busy. 5584 */ 5585 if (m->object != NULL && !vm_page_busied(m)) 5586 VM_OBJECT_ASSERT_BUSY(m->object); 5587 } 5588 5589 void 5590 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5591 { 5592 5593 if ((bits & PGA_WRITEABLE) == 0) 5594 return; 5595 5596 /* 5597 * The PGA_WRITEABLE flag can only be set if the page is 5598 * managed, is exclusively busied or the object is locked. 5599 * Currently, this flag is only set by pmap_enter(). 5600 */ 5601 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5602 ("PGA_WRITEABLE on unmanaged page")); 5603 if (!vm_page_xbusied(m)) 5604 VM_OBJECT_ASSERT_BUSY(m->object); 5605 } 5606 #endif 5607 5608 #include "opt_ddb.h" 5609 #ifdef DDB 5610 #include <sys/kernel.h> 5611 5612 #include <ddb/ddb.h> 5613 5614 DB_SHOW_COMMAND(page, vm_page_print_page_info) 5615 { 5616 5617 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5618 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5619 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5620 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5621 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5622 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5623 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5624 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5625 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5626 } 5627 5628 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 5629 { 5630 int dom; 5631 5632 db_printf("pq_free %d\n", vm_free_count()); 5633 for (dom = 0; dom < vm_ndomains; dom++) { 5634 db_printf( 5635 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5636 dom, 5637 vm_dom[dom].vmd_page_count, 5638 vm_dom[dom].vmd_free_count, 5639 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5640 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5641 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5642 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5643 } 5644 } 5645 5646 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5647 { 5648 vm_page_t m; 5649 boolean_t phys, virt; 5650 5651 if (!have_addr) { 5652 db_printf("show pginfo addr\n"); 5653 return; 5654 } 5655 5656 phys = strchr(modif, 'p') != NULL; 5657 virt = strchr(modif, 'v') != NULL; 5658 if (virt) 5659 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5660 else if (phys) 5661 m = PHYS_TO_VM_PAGE(addr); 5662 else 5663 m = (vm_page_t)addr; 5664 db_printf( 5665 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5666 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5667 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5668 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5669 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5670 } 5671 #endif /* DDB */ 5672