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 36 /*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Resident memory management module. 65 */ 66 67 #include <sys/cdefs.h> 68 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/counter.h> 73 #include <sys/domainset.h> 74 #include <sys/kernel.h> 75 #include <sys/limits.h> 76 #include <sys/linker.h> 77 #include <sys/lock.h> 78 #include <sys/malloc.h> 79 #include <sys/mman.h> 80 #include <sys/msgbuf.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> 83 #include <sys/rwlock.h> 84 #include <sys/sleepqueue.h> 85 #include <sys/sbuf.h> 86 #include <sys/sched.h> 87 #include <sys/smp.h> 88 #include <sys/sysctl.h> 89 #include <sys/vmmeter.h> 90 #include <sys/vnode.h> 91 92 #include <vm/vm.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_param.h> 95 #include <vm/vm_domainset.h> 96 #include <vm/vm_kern.h> 97 #include <vm/vm_map.h> 98 #include <vm/vm_object.h> 99 #include <vm/vm_page.h> 100 #include <vm/vm_pageout.h> 101 #include <vm/vm_phys.h> 102 #include <vm/vm_pagequeue.h> 103 #include <vm/vm_pager.h> 104 #include <vm/vm_radix.h> 105 #include <vm/vm_reserv.h> 106 #include <vm/vm_extern.h> 107 #include <vm/vm_dumpset.h> 108 #include <vm/uma.h> 109 #include <vm/uma_int.h> 110 111 #include <machine/md_var.h> 112 113 struct vm_domain vm_dom[MAXMEMDOM]; 114 115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 116 117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 118 119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 120 /* The following fields are protected by the domainset lock. */ 121 domainset_t __exclusive_cache_line vm_min_domains; 122 domainset_t __exclusive_cache_line vm_severe_domains; 123 static int vm_min_waiters; 124 static int vm_severe_waiters; 125 static int vm_pageproc_waiters; 126 127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 128 "VM page statistics"); 129 130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 132 CTLFLAG_RD, &pqstate_commit_retries, 133 "Number of failed per-page atomic queue state updates"); 134 135 static COUNTER_U64_DEFINE_EARLY(queue_ops); 136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 137 CTLFLAG_RD, &queue_ops, 138 "Number of batched queue operations"); 139 140 static COUNTER_U64_DEFINE_EARLY(queue_nops); 141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 142 CTLFLAG_RD, &queue_nops, 143 "Number of batched queue operations with no effects"); 144 145 /* 146 * bogus page -- for I/O to/from partially complete buffers, 147 * or for paging into sparsely invalid regions. 148 */ 149 vm_page_t bogus_page; 150 151 vm_page_t vm_page_array; 152 long vm_page_array_size; 153 long first_page; 154 155 struct bitset *vm_page_dump; 156 long vm_page_dump_pages; 157 158 static TAILQ_HEAD(, vm_page) blacklist_head; 159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 161 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 162 163 static uma_zone_t fakepg_zone; 164 165 static void vm_page_alloc_check(vm_page_t m); 166 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 167 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 168 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 169 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 170 static bool vm_page_free_prep(vm_page_t m); 171 static void vm_page_free_toq(vm_page_t m); 172 static void vm_page_init(void *dummy); 173 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 174 vm_pindex_t pindex, vm_page_t mpred); 175 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 176 vm_page_t mpred); 177 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 178 const uint16_t nflag); 179 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 180 vm_page_t m_run, vm_paddr_t high); 181 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 182 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 183 int req); 184 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 185 int flags); 186 static void vm_page_zone_release(void *arg, void **store, int cnt); 187 188 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 189 190 static void 191 vm_page_init(void *dummy) 192 { 193 194 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 195 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 196 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED); 197 } 198 199 static int pgcache_zone_max_pcpu; 200 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu, 201 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0, 202 "Per-CPU page cache size"); 203 204 /* 205 * The cache page zone is initialized later since we need to be able to allocate 206 * pages before UMA is fully initialized. 207 */ 208 static void 209 vm_page_init_cache_zones(void *dummy __unused) 210 { 211 struct vm_domain *vmd; 212 struct vm_pgcache *pgcache; 213 int cache, domain, maxcache, pool; 214 215 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu); 216 maxcache = pgcache_zone_max_pcpu * mp_ncpus; 217 for (domain = 0; domain < vm_ndomains; domain++) { 218 vmd = VM_DOMAIN(domain); 219 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 220 pgcache = &vmd->vmd_pgcache[pool]; 221 pgcache->domain = domain; 222 pgcache->pool = pool; 223 pgcache->zone = uma_zcache_create("vm pgcache", 224 PAGE_SIZE, NULL, NULL, NULL, NULL, 225 vm_page_zone_import, vm_page_zone_release, pgcache, 226 UMA_ZONE_VM); 227 228 /* 229 * Limit each pool's zone to 0.1% of the pages in the 230 * domain. 231 */ 232 cache = maxcache != 0 ? maxcache : 233 vmd->vmd_page_count / 1000; 234 uma_zone_set_maxcache(pgcache->zone, cache); 235 } 236 } 237 } 238 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 239 240 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 241 #if PAGE_SIZE == 32768 242 #ifdef CTASSERT 243 CTASSERT(sizeof(u_long) >= 8); 244 #endif 245 #endif 246 247 /* 248 * vm_set_page_size: 249 * 250 * Sets the page size, perhaps based upon the memory 251 * size. Must be called before any use of page-size 252 * dependent functions. 253 */ 254 void 255 vm_set_page_size(void) 256 { 257 if (vm_cnt.v_page_size == 0) 258 vm_cnt.v_page_size = PAGE_SIZE; 259 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 260 panic("vm_set_page_size: page size not a power of two"); 261 } 262 263 /* 264 * vm_page_blacklist_next: 265 * 266 * Find the next entry in the provided string of blacklist 267 * addresses. Entries are separated by space, comma, or newline. 268 * If an invalid integer is encountered then the rest of the 269 * string is skipped. Updates the list pointer to the next 270 * character, or NULL if the string is exhausted or invalid. 271 */ 272 static vm_paddr_t 273 vm_page_blacklist_next(char **list, char *end) 274 { 275 vm_paddr_t bad; 276 char *cp, *pos; 277 278 if (list == NULL || *list == NULL) 279 return (0); 280 if (**list =='\0') { 281 *list = NULL; 282 return (0); 283 } 284 285 /* 286 * If there's no end pointer then the buffer is coming from 287 * the kenv and we know it's null-terminated. 288 */ 289 if (end == NULL) 290 end = *list + strlen(*list); 291 292 /* Ensure that strtoq() won't walk off the end */ 293 if (*end != '\0') { 294 if (*end == '\n' || *end == ' ' || *end == ',') 295 *end = '\0'; 296 else { 297 printf("Blacklist not terminated, skipping\n"); 298 *list = NULL; 299 return (0); 300 } 301 } 302 303 for (pos = *list; *pos != '\0'; pos = cp) { 304 bad = strtoq(pos, &cp, 0); 305 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 306 if (bad == 0) { 307 if (++cp < end) 308 continue; 309 else 310 break; 311 } 312 } else 313 break; 314 if (*cp == '\0' || ++cp >= end) 315 *list = NULL; 316 else 317 *list = cp; 318 return (trunc_page(bad)); 319 } 320 printf("Garbage in RAM blacklist, skipping\n"); 321 *list = NULL; 322 return (0); 323 } 324 325 bool 326 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 327 { 328 struct vm_domain *vmd; 329 vm_page_t m; 330 bool found; 331 332 m = vm_phys_paddr_to_vm_page(pa); 333 if (m == NULL) 334 return (true); /* page does not exist, no failure */ 335 336 vmd = VM_DOMAIN(vm_phys_domain(pa)); 337 vm_domain_free_lock(vmd); 338 found = vm_phys_unfree_page(pa); 339 vm_domain_free_unlock(vmd); 340 if (found) { 341 vm_domain_freecnt_inc(vmd, -1); 342 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 343 if (verbose) 344 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 345 } 346 return (found); 347 } 348 349 /* 350 * vm_page_blacklist_check: 351 * 352 * Iterate through the provided string of blacklist addresses, pulling 353 * each entry out of the physical allocator free list and putting it 354 * onto a list for reporting via the vm.page_blacklist sysctl. 355 */ 356 static void 357 vm_page_blacklist_check(char *list, char *end) 358 { 359 vm_paddr_t pa; 360 char *next; 361 362 next = list; 363 while (next != NULL) { 364 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 365 continue; 366 vm_page_blacklist_add(pa, bootverbose); 367 } 368 } 369 370 /* 371 * vm_page_blacklist_load: 372 * 373 * Search for a special module named "ram_blacklist". It'll be a 374 * plain text file provided by the user via the loader directive 375 * of the same name. 376 */ 377 static void 378 vm_page_blacklist_load(char **list, char **end) 379 { 380 void *mod; 381 u_char *ptr; 382 u_int len; 383 384 mod = NULL; 385 ptr = NULL; 386 387 mod = preload_search_by_type("ram_blacklist"); 388 if (mod != NULL) { 389 ptr = preload_fetch_addr(mod); 390 len = preload_fetch_size(mod); 391 } 392 *list = ptr; 393 if (ptr != NULL) 394 *end = ptr + len; 395 else 396 *end = NULL; 397 return; 398 } 399 400 static int 401 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 402 { 403 vm_page_t m; 404 struct sbuf sbuf; 405 int error, first; 406 407 first = 1; 408 error = sysctl_wire_old_buffer(req, 0); 409 if (error != 0) 410 return (error); 411 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 412 TAILQ_FOREACH(m, &blacklist_head, listq) { 413 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 414 (uintmax_t)m->phys_addr); 415 first = 0; 416 } 417 error = sbuf_finish(&sbuf); 418 sbuf_delete(&sbuf); 419 return (error); 420 } 421 422 /* 423 * Initialize a dummy page for use in scans of the specified paging queue. 424 * In principle, this function only needs to set the flag PG_MARKER. 425 * Nonetheless, it write busies the page as a safety precaution. 426 */ 427 void 428 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 429 { 430 431 bzero(marker, sizeof(*marker)); 432 marker->flags = PG_MARKER; 433 marker->a.flags = aflags; 434 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 435 marker->a.queue = queue; 436 } 437 438 static void 439 vm_page_domain_init(int domain) 440 { 441 struct vm_domain *vmd; 442 struct vm_pagequeue *pq; 443 int i; 444 445 vmd = VM_DOMAIN(domain); 446 bzero(vmd, sizeof(*vmd)); 447 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 448 "vm inactive pagequeue"; 449 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 450 "vm active pagequeue"; 451 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 452 "vm laundry pagequeue"; 453 *__DECONST(const char **, 454 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 455 "vm unswappable pagequeue"; 456 vmd->vmd_domain = domain; 457 vmd->vmd_page_count = 0; 458 vmd->vmd_free_count = 0; 459 vmd->vmd_segs = 0; 460 vmd->vmd_oom = FALSE; 461 for (i = 0; i < PQ_COUNT; i++) { 462 pq = &vmd->vmd_pagequeues[i]; 463 TAILQ_INIT(&pq->pq_pl); 464 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 465 MTX_DEF | MTX_DUPOK); 466 pq->pq_pdpages = 0; 467 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 468 } 469 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 470 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 471 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 472 473 /* 474 * inacthead is used to provide FIFO ordering for LRU-bypassing 475 * insertions. 476 */ 477 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 478 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 479 &vmd->vmd_inacthead, plinks.q); 480 481 /* 482 * The clock pages are used to implement active queue scanning without 483 * requeues. Scans start at clock[0], which is advanced after the scan 484 * ends. When the two clock hands meet, they are reset and scanning 485 * resumes from the head of the queue. 486 */ 487 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 488 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 489 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 490 &vmd->vmd_clock[0], plinks.q); 491 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 492 &vmd->vmd_clock[1], plinks.q); 493 } 494 495 /* 496 * Initialize a physical page in preparation for adding it to the free 497 * lists. 498 */ 499 void 500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool) 501 { 502 m->object = NULL; 503 m->ref_count = 0; 504 m->busy_lock = VPB_FREED; 505 m->flags = m->a.flags = 0; 506 m->phys_addr = pa; 507 m->a.queue = PQ_NONE; 508 m->psind = 0; 509 m->segind = segind; 510 m->order = VM_NFREEORDER; 511 m->pool = pool; 512 m->valid = m->dirty = 0; 513 pmap_page_init(m); 514 } 515 516 #ifndef PMAP_HAS_PAGE_ARRAY 517 static vm_paddr_t 518 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 519 { 520 vm_paddr_t new_end; 521 522 /* 523 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 524 * However, because this page is allocated from KVM, out-of-bounds 525 * accesses using the direct map will not be trapped. 526 */ 527 *vaddr += PAGE_SIZE; 528 529 /* 530 * Allocate physical memory for the page structures, and map it. 531 */ 532 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 533 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 534 VM_PROT_READ | VM_PROT_WRITE); 535 vm_page_array_size = page_range; 536 537 return (new_end); 538 } 539 #endif 540 541 /* 542 * vm_page_startup: 543 * 544 * Initializes the resident memory module. Allocates physical memory for 545 * bootstrapping UMA and some data structures that are used to manage 546 * physical pages. Initializes these structures, and populates the free 547 * page queues. 548 */ 549 vm_offset_t 550 vm_page_startup(vm_offset_t vaddr) 551 { 552 struct vm_phys_seg *seg; 553 struct vm_domain *vmd; 554 vm_page_t m; 555 char *list, *listend; 556 vm_paddr_t end, high_avail, low_avail, new_end, size; 557 vm_paddr_t page_range __unused; 558 vm_paddr_t last_pa, pa, startp, endp; 559 u_long pagecount; 560 #if MINIDUMP_PAGE_TRACKING 561 u_long vm_page_dump_size; 562 #endif 563 int biggestone, i, segind; 564 #ifdef WITNESS 565 vm_offset_t mapped; 566 int witness_size; 567 #endif 568 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 569 long ii; 570 #endif 571 #ifdef VM_FREEPOOL_LAZYINIT 572 int lazyinit; 573 #endif 574 575 vaddr = round_page(vaddr); 576 577 vm_phys_early_startup(); 578 biggestone = vm_phys_avail_largest(); 579 end = phys_avail[biggestone+1]; 580 581 /* 582 * Initialize the page and queue locks. 583 */ 584 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 585 for (i = 0; i < PA_LOCK_COUNT; i++) 586 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 587 for (i = 0; i < vm_ndomains; i++) 588 vm_page_domain_init(i); 589 590 new_end = end; 591 #ifdef WITNESS 592 witness_size = round_page(witness_startup_count()); 593 new_end -= witness_size; 594 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 595 VM_PROT_READ | VM_PROT_WRITE); 596 bzero((void *)mapped, witness_size); 597 witness_startup((void *)mapped); 598 #endif 599 600 #if MINIDUMP_PAGE_TRACKING 601 /* 602 * Allocate a bitmap to indicate that a random physical page 603 * needs to be included in a minidump. 604 * 605 * The amd64 port needs this to indicate which direct map pages 606 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 607 * 608 * However, i386 still needs this workspace internally within the 609 * minidump code. In theory, they are not needed on i386, but are 610 * included should the sf_buf code decide to use them. 611 */ 612 last_pa = 0; 613 vm_page_dump_pages = 0; 614 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 615 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 616 dump_avail[i] / PAGE_SIZE; 617 if (dump_avail[i + 1] > last_pa) 618 last_pa = dump_avail[i + 1]; 619 } 620 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 621 new_end -= vm_page_dump_size; 622 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 623 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 624 bzero((void *)vm_page_dump, vm_page_dump_size); 625 #if MINIDUMP_STARTUP_PAGE_TRACKING 626 /* 627 * Include the UMA bootstrap pages, witness pages and vm_page_dump 628 * in a crash dump. When pmap_map() uses the direct map, they are 629 * not automatically included. 630 */ 631 for (pa = new_end; pa < end; pa += PAGE_SIZE) 632 dump_add_page(pa); 633 #endif 634 #else 635 (void)last_pa; 636 #endif 637 phys_avail[biggestone + 1] = new_end; 638 #ifdef __amd64__ 639 /* 640 * Request that the physical pages underlying the message buffer be 641 * included in a crash dump. Since the message buffer is accessed 642 * through the direct map, they are not automatically included. 643 */ 644 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 645 last_pa = pa + round_page(msgbufsize); 646 while (pa < last_pa) { 647 dump_add_page(pa); 648 pa += PAGE_SIZE; 649 } 650 #endif 651 /* 652 * Compute the number of pages of memory that will be available for 653 * use, taking into account the overhead of a page structure per page. 654 * In other words, solve 655 * "available physical memory" - round_page(page_range * 656 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 657 * for page_range. 658 */ 659 low_avail = phys_avail[0]; 660 high_avail = phys_avail[1]; 661 for (i = 0; i < vm_phys_nsegs; i++) { 662 if (vm_phys_segs[i].start < low_avail) 663 low_avail = vm_phys_segs[i].start; 664 if (vm_phys_segs[i].end > high_avail) 665 high_avail = vm_phys_segs[i].end; 666 } 667 /* Skip the first chunk. It is already accounted for. */ 668 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 669 if (phys_avail[i] < low_avail) 670 low_avail = phys_avail[i]; 671 if (phys_avail[i + 1] > high_avail) 672 high_avail = phys_avail[i + 1]; 673 } 674 first_page = low_avail / PAGE_SIZE; 675 #ifdef VM_PHYSSEG_SPARSE 676 size = 0; 677 for (i = 0; i < vm_phys_nsegs; i++) 678 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 679 for (i = 0; phys_avail[i + 1] != 0; i += 2) 680 size += phys_avail[i + 1] - phys_avail[i]; 681 #elif defined(VM_PHYSSEG_DENSE) 682 size = high_avail - low_avail; 683 #else 684 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 685 #endif 686 687 #ifdef PMAP_HAS_PAGE_ARRAY 688 pmap_page_array_startup(size / PAGE_SIZE); 689 biggestone = vm_phys_avail_largest(); 690 end = new_end = phys_avail[biggestone + 1]; 691 #else 692 #ifdef VM_PHYSSEG_DENSE 693 /* 694 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 695 * the overhead of a page structure per page only if vm_page_array is 696 * allocated from the last physical memory chunk. Otherwise, we must 697 * allocate page structures representing the physical memory 698 * underlying vm_page_array, even though they will not be used. 699 */ 700 if (new_end != high_avail) 701 page_range = size / PAGE_SIZE; 702 else 703 #endif 704 { 705 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 706 707 /* 708 * If the partial bytes remaining are large enough for 709 * a page (PAGE_SIZE) without a corresponding 710 * 'struct vm_page', then new_end will contain an 711 * extra page after subtracting the length of the VM 712 * page array. Compensate by subtracting an extra 713 * page from new_end. 714 */ 715 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 716 if (new_end == high_avail) 717 high_avail -= PAGE_SIZE; 718 new_end -= PAGE_SIZE; 719 } 720 } 721 end = new_end; 722 new_end = vm_page_array_alloc(&vaddr, end, page_range); 723 #endif 724 725 #if VM_NRESERVLEVEL > 0 726 /* 727 * Allocate physical memory for the reservation management system's 728 * data structures, and map it. 729 */ 730 new_end = vm_reserv_startup(&vaddr, new_end); 731 #endif 732 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 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 #ifdef VM_FREEPOOL_LAZYINIT 755 lazyinit = 1; 756 TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit); 757 #endif 758 759 /* 760 * Initialize the page structures and add every available page to the 761 * physical memory allocator's free lists. 762 */ 763 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 764 for (ii = 0; ii < vm_page_array_size; ii++) { 765 m = &vm_page_array[ii]; 766 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0, 767 VM_FREEPOOL_DEFAULT); 768 m->flags = PG_FICTITIOUS; 769 } 770 #endif 771 vm_cnt.v_page_count = 0; 772 for (segind = 0; segind < vm_phys_nsegs; segind++) { 773 seg = &vm_phys_segs[segind]; 774 775 /* 776 * If lazy vm_page initialization is not enabled, simply 777 * initialize all of the pages in the segment. Otherwise, we 778 * only initialize: 779 * 1. Pages not covered by phys_avail[], since they might be 780 * freed to the allocator at some future point, e.g., by 781 * kmem_bootstrap_free(). 782 * 2. The first page of each run of free pages handed to the 783 * vm_phys allocator, which in turn defers initialization 784 * of pages until they are needed. 785 * This avoids blocking the boot process for long periods, which 786 * may be relevant for VMs (which ought to boot as quickly as 787 * possible) and/or systems with large amounts of physical 788 * memory. 789 */ 790 #ifdef VM_FREEPOOL_LAZYINIT 791 if (lazyinit) { 792 startp = seg->start; 793 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 794 if (startp >= seg->end) 795 break; 796 797 if (phys_avail[i + 1] < startp) 798 continue; 799 if (phys_avail[i] <= startp) { 800 startp = phys_avail[i + 1]; 801 continue; 802 } 803 804 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 805 for (endp = MIN(phys_avail[i], seg->end); 806 startp < endp; startp += PAGE_SIZE, m++) { 807 vm_page_init_page(m, startp, segind, 808 VM_FREEPOOL_DEFAULT); 809 } 810 } 811 } else 812 #endif 813 for (m = seg->first_page, pa = seg->start; 814 pa < seg->end; m++, pa += PAGE_SIZE) { 815 vm_page_init_page(m, pa, segind, 816 VM_FREEPOOL_DEFAULT); 817 } 818 819 /* 820 * Add the segment's pages that are covered by one of 821 * phys_avail's ranges to the free lists. 822 */ 823 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 824 if (seg->end <= phys_avail[i] || 825 seg->start >= phys_avail[i + 1]) 826 continue; 827 828 startp = MAX(seg->start, phys_avail[i]); 829 endp = MIN(seg->end, phys_avail[i + 1]); 830 pagecount = (u_long)atop(endp - startp); 831 if (pagecount == 0) 832 continue; 833 834 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 835 #ifdef VM_FREEPOOL_LAZYINIT 836 if (lazyinit) { 837 vm_page_init_page(m, startp, segind, 838 VM_FREEPOOL_LAZYINIT); 839 } 840 #endif 841 vmd = VM_DOMAIN(seg->domain); 842 vm_domain_free_lock(vmd); 843 vm_phys_enqueue_contig(m, pagecount); 844 vm_domain_free_unlock(vmd); 845 vm_domain_freecnt_inc(vmd, pagecount); 846 vm_cnt.v_page_count += (u_int)pagecount; 847 vmd->vmd_page_count += (u_int)pagecount; 848 vmd->vmd_segs |= 1UL << segind; 849 } 850 } 851 852 /* 853 * Remove blacklisted pages from the physical memory allocator. 854 */ 855 TAILQ_INIT(&blacklist_head); 856 vm_page_blacklist_load(&list, &listend); 857 vm_page_blacklist_check(list, listend); 858 859 list = kern_getenv("vm.blacklist"); 860 vm_page_blacklist_check(list, NULL); 861 862 freeenv(list); 863 #if VM_NRESERVLEVEL > 0 864 /* 865 * Initialize the reservation management system. 866 */ 867 vm_reserv_init(); 868 #endif 869 870 return (vaddr); 871 } 872 873 void 874 vm_page_reference(vm_page_t m) 875 { 876 877 vm_page_aflag_set(m, PGA_REFERENCED); 878 } 879 880 /* 881 * vm_page_trybusy 882 * 883 * Helper routine for grab functions to trylock busy. 884 * 885 * Returns true on success and false on failure. 886 */ 887 static bool 888 vm_page_trybusy(vm_page_t m, int allocflags) 889 { 890 891 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 892 return (vm_page_trysbusy(m)); 893 else 894 return (vm_page_tryxbusy(m)); 895 } 896 897 /* 898 * vm_page_tryacquire 899 * 900 * Helper routine for grab functions to trylock busy and wire. 901 * 902 * Returns true on success and false on failure. 903 */ 904 static inline bool 905 vm_page_tryacquire(vm_page_t m, int allocflags) 906 { 907 bool locked; 908 909 locked = vm_page_trybusy(m, allocflags); 910 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 911 vm_page_wire(m); 912 return (locked); 913 } 914 915 /* 916 * vm_page_busy_acquire: 917 * 918 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 919 * and drop the object lock if necessary. 920 */ 921 bool 922 vm_page_busy_acquire(vm_page_t m, int allocflags) 923 { 924 vm_object_t obj; 925 bool locked; 926 927 /* 928 * The page-specific object must be cached because page 929 * identity can change during the sleep, causing the 930 * re-lock of a different object. 931 * It is assumed that a reference to the object is already 932 * held by the callers. 933 */ 934 obj = atomic_load_ptr(&m->object); 935 for (;;) { 936 if (vm_page_tryacquire(m, allocflags)) 937 return (true); 938 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 939 return (false); 940 if (obj != NULL) 941 locked = VM_OBJECT_WOWNED(obj); 942 else 943 locked = false; 944 MPASS(locked || vm_page_wired(m)); 945 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 946 locked) && locked) 947 VM_OBJECT_WLOCK(obj); 948 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 949 return (false); 950 KASSERT(m->object == obj || m->object == NULL, 951 ("vm_page_busy_acquire: page %p does not belong to %p", 952 m, obj)); 953 } 954 } 955 956 /* 957 * vm_page_busy_downgrade: 958 * 959 * Downgrade an exclusive busy page into a single shared busy page. 960 */ 961 void 962 vm_page_busy_downgrade(vm_page_t m) 963 { 964 u_int x; 965 966 vm_page_assert_xbusied(m); 967 968 x = vm_page_busy_fetch(m); 969 for (;;) { 970 if (atomic_fcmpset_rel_int(&m->busy_lock, 971 &x, VPB_SHARERS_WORD(1))) 972 break; 973 } 974 if ((x & VPB_BIT_WAITERS) != 0) 975 wakeup(m); 976 } 977 978 /* 979 * 980 * vm_page_busy_tryupgrade: 981 * 982 * Attempt to upgrade a single shared busy into an exclusive busy. 983 */ 984 int 985 vm_page_busy_tryupgrade(vm_page_t m) 986 { 987 u_int ce, x; 988 989 vm_page_assert_sbusied(m); 990 991 x = vm_page_busy_fetch(m); 992 ce = VPB_CURTHREAD_EXCLUSIVE; 993 for (;;) { 994 if (VPB_SHARERS(x) > 1) 995 return (0); 996 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 997 ("vm_page_busy_tryupgrade: invalid lock state")); 998 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 999 ce | (x & VPB_BIT_WAITERS))) 1000 continue; 1001 return (1); 1002 } 1003 } 1004 1005 /* 1006 * vm_page_sbusied: 1007 * 1008 * Return a positive value if the page is shared busied, 0 otherwise. 1009 */ 1010 int 1011 vm_page_sbusied(vm_page_t m) 1012 { 1013 u_int x; 1014 1015 x = vm_page_busy_fetch(m); 1016 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1017 } 1018 1019 /* 1020 * vm_page_sunbusy: 1021 * 1022 * Shared unbusy a page. 1023 */ 1024 void 1025 vm_page_sunbusy(vm_page_t m) 1026 { 1027 u_int x; 1028 1029 vm_page_assert_sbusied(m); 1030 1031 x = vm_page_busy_fetch(m); 1032 for (;;) { 1033 KASSERT(x != VPB_FREED, 1034 ("vm_page_sunbusy: Unlocking freed page.")); 1035 if (VPB_SHARERS(x) > 1) { 1036 if (atomic_fcmpset_int(&m->busy_lock, &x, 1037 x - VPB_ONE_SHARER)) 1038 break; 1039 continue; 1040 } 1041 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1042 ("vm_page_sunbusy: invalid lock state")); 1043 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1044 continue; 1045 if ((x & VPB_BIT_WAITERS) == 0) 1046 break; 1047 wakeup(m); 1048 break; 1049 } 1050 } 1051 1052 /* 1053 * vm_page_busy_sleep: 1054 * 1055 * Sleep if the page is busy, using the page pointer as wchan. 1056 * This is used to implement the hard-path of the busying mechanism. 1057 * 1058 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1059 * will not sleep if the page is shared-busy. 1060 * 1061 * The object lock must be held on entry. 1062 * 1063 * Returns true if it slept and dropped the object lock, or false 1064 * if there was no sleep and the lock is still held. 1065 */ 1066 bool 1067 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) 1068 { 1069 vm_object_t obj; 1070 1071 obj = m->object; 1072 VM_OBJECT_ASSERT_LOCKED(obj); 1073 1074 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, 1075 true)); 1076 } 1077 1078 /* 1079 * vm_page_busy_sleep_unlocked: 1080 * 1081 * Sleep if the page is busy, using the page pointer as wchan. 1082 * This is used to implement the hard-path of busying mechanism. 1083 * 1084 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1085 * will not sleep if the page is shared-busy. 1086 * 1087 * The object lock must not be held on entry. The operation will 1088 * return if the page changes identity. 1089 */ 1090 void 1091 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1092 const char *wmesg, int allocflags) 1093 { 1094 VM_OBJECT_ASSERT_UNLOCKED(obj); 1095 1096 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); 1097 } 1098 1099 /* 1100 * _vm_page_busy_sleep: 1101 * 1102 * Internal busy sleep function. Verifies the page identity and 1103 * lockstate against parameters. Returns true if it sleeps and 1104 * false otherwise. 1105 * 1106 * allocflags uses VM_ALLOC_* flags to specify the lock required. 1107 * 1108 * If locked is true the lock will be dropped for any true returns 1109 * and held for any false returns. 1110 */ 1111 static bool 1112 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1113 const char *wmesg, int allocflags, bool locked) 1114 { 1115 bool xsleep; 1116 u_int x; 1117 1118 /* 1119 * If the object is busy we must wait for that to drain to zero 1120 * before trying the page again. 1121 */ 1122 if (obj != NULL && vm_object_busied(obj)) { 1123 if (locked) 1124 VM_OBJECT_DROP(obj); 1125 vm_object_busy_wait(obj, wmesg); 1126 return (true); 1127 } 1128 1129 if (!vm_page_busied(m)) 1130 return (false); 1131 1132 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1133 sleepq_lock(m); 1134 x = vm_page_busy_fetch(m); 1135 do { 1136 /* 1137 * If the page changes objects or becomes unlocked we can 1138 * simply return. 1139 */ 1140 if (x == VPB_UNBUSIED || 1141 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1142 m->object != obj || m->pindex != pindex) { 1143 sleepq_release(m); 1144 return (false); 1145 } 1146 if ((x & VPB_BIT_WAITERS) != 0) 1147 break; 1148 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1149 if (locked) 1150 VM_OBJECT_DROP(obj); 1151 DROP_GIANT(); 1152 sleepq_add(m, NULL, wmesg, 0, 0); 1153 sleepq_wait(m, PVM); 1154 PICKUP_GIANT(); 1155 return (true); 1156 } 1157 1158 /* 1159 * vm_page_trysbusy: 1160 * 1161 * Try to shared busy a page. 1162 * If the operation succeeds 1 is returned otherwise 0. 1163 * The operation never sleeps. 1164 */ 1165 int 1166 vm_page_trysbusy(vm_page_t m) 1167 { 1168 vm_object_t obj; 1169 u_int x; 1170 1171 obj = m->object; 1172 x = vm_page_busy_fetch(m); 1173 for (;;) { 1174 if ((x & VPB_BIT_SHARED) == 0) 1175 return (0); 1176 /* 1177 * Reduce the window for transient busies that will trigger 1178 * false negatives in vm_page_ps_test(). 1179 */ 1180 if (obj != NULL && vm_object_busied(obj)) 1181 return (0); 1182 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1183 x + VPB_ONE_SHARER)) 1184 break; 1185 } 1186 1187 /* Refetch the object now that we're guaranteed that it is stable. */ 1188 obj = m->object; 1189 if (obj != NULL && vm_object_busied(obj)) { 1190 vm_page_sunbusy(m); 1191 return (0); 1192 } 1193 return (1); 1194 } 1195 1196 /* 1197 * vm_page_tryxbusy: 1198 * 1199 * Try to exclusive busy a page. 1200 * If the operation succeeds 1 is returned otherwise 0. 1201 * The operation never sleeps. 1202 */ 1203 int 1204 vm_page_tryxbusy(vm_page_t m) 1205 { 1206 vm_object_t obj; 1207 1208 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1209 VPB_CURTHREAD_EXCLUSIVE) == 0) 1210 return (0); 1211 1212 obj = m->object; 1213 if (obj != NULL && vm_object_busied(obj)) { 1214 vm_page_xunbusy(m); 1215 return (0); 1216 } 1217 return (1); 1218 } 1219 1220 static void 1221 vm_page_xunbusy_hard_tail(vm_page_t m) 1222 { 1223 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1224 /* Wake the waiter. */ 1225 wakeup(m); 1226 } 1227 1228 /* 1229 * vm_page_xunbusy_hard: 1230 * 1231 * Called when unbusy has failed because there is a waiter. 1232 */ 1233 void 1234 vm_page_xunbusy_hard(vm_page_t m) 1235 { 1236 vm_page_assert_xbusied(m); 1237 vm_page_xunbusy_hard_tail(m); 1238 } 1239 1240 void 1241 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1242 { 1243 vm_page_assert_xbusied_unchecked(m); 1244 vm_page_xunbusy_hard_tail(m); 1245 } 1246 1247 static void 1248 vm_page_busy_free(vm_page_t m) 1249 { 1250 u_int x; 1251 1252 atomic_thread_fence_rel(); 1253 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1254 if ((x & VPB_BIT_WAITERS) != 0) 1255 wakeup(m); 1256 } 1257 1258 /* 1259 * vm_page_unhold_pages: 1260 * 1261 * Unhold each of the pages that is referenced by the given array. 1262 */ 1263 void 1264 vm_page_unhold_pages(vm_page_t *ma, int count) 1265 { 1266 1267 for (; count != 0; count--) { 1268 vm_page_unwire(*ma, PQ_ACTIVE); 1269 ma++; 1270 } 1271 } 1272 1273 vm_page_t 1274 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1275 { 1276 vm_page_t m; 1277 1278 #ifdef VM_PHYSSEG_SPARSE 1279 m = vm_phys_paddr_to_vm_page(pa); 1280 if (m == NULL) 1281 m = vm_phys_fictitious_to_vm_page(pa); 1282 return (m); 1283 #elif defined(VM_PHYSSEG_DENSE) 1284 long pi; 1285 1286 pi = atop(pa); 1287 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1288 m = &vm_page_array[pi - first_page]; 1289 return (m); 1290 } 1291 return (vm_phys_fictitious_to_vm_page(pa)); 1292 #else 1293 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1294 #endif 1295 } 1296 1297 /* 1298 * vm_page_getfake: 1299 * 1300 * Create a fictitious page with the specified physical address and 1301 * memory attribute. The memory attribute is the only the machine- 1302 * dependent aspect of a fictitious page that must be initialized. 1303 */ 1304 vm_page_t 1305 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1306 { 1307 vm_page_t m; 1308 1309 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1310 vm_page_initfake(m, paddr, memattr); 1311 return (m); 1312 } 1313 1314 void 1315 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1316 { 1317 1318 if ((m->flags & PG_FICTITIOUS) != 0) { 1319 /* 1320 * The page's memattr might have changed since the 1321 * previous initialization. Update the pmap to the 1322 * new memattr. 1323 */ 1324 goto memattr; 1325 } 1326 m->phys_addr = paddr; 1327 m->a.queue = PQ_NONE; 1328 /* Fictitious pages don't use "segind". */ 1329 m->flags = PG_FICTITIOUS; 1330 /* Fictitious pages don't use "order" or "pool". */ 1331 m->oflags = VPO_UNMANAGED; 1332 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1333 /* Fictitious pages are unevictable. */ 1334 m->ref_count = 1; 1335 pmap_page_init(m); 1336 memattr: 1337 pmap_page_set_memattr(m, memattr); 1338 } 1339 1340 /* 1341 * vm_page_putfake: 1342 * 1343 * Release a fictitious page. 1344 */ 1345 void 1346 vm_page_putfake(vm_page_t m) 1347 { 1348 1349 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1350 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1351 ("vm_page_putfake: bad page %p", m)); 1352 vm_page_assert_xbusied(m); 1353 vm_page_busy_free(m); 1354 uma_zfree(fakepg_zone, m); 1355 } 1356 1357 /* 1358 * vm_page_updatefake: 1359 * 1360 * Update the given fictitious page to the specified physical address and 1361 * memory attribute. 1362 */ 1363 void 1364 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1365 { 1366 1367 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1368 ("vm_page_updatefake: bad page %p", m)); 1369 m->phys_addr = paddr; 1370 pmap_page_set_memattr(m, memattr); 1371 } 1372 1373 /* 1374 * vm_page_free: 1375 * 1376 * Free a page. 1377 */ 1378 void 1379 vm_page_free(vm_page_t m) 1380 { 1381 1382 m->flags &= ~PG_ZERO; 1383 vm_page_free_toq(m); 1384 } 1385 1386 /* 1387 * vm_page_free_zero: 1388 * 1389 * Free a page to the zerod-pages queue 1390 */ 1391 void 1392 vm_page_free_zero(vm_page_t m) 1393 { 1394 1395 m->flags |= PG_ZERO; 1396 vm_page_free_toq(m); 1397 } 1398 1399 /* 1400 * Unbusy and handle the page queueing for a page from a getpages request that 1401 * was optionally read ahead or behind. 1402 */ 1403 void 1404 vm_page_readahead_finish(vm_page_t m) 1405 { 1406 1407 /* We shouldn't put invalid pages on queues. */ 1408 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1409 1410 /* 1411 * Since the page is not the actually needed one, whether it should 1412 * be activated or deactivated is not obvious. Empirical results 1413 * have shown that deactivating the page is usually the best choice, 1414 * unless the page is wanted by another thread. 1415 */ 1416 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1417 vm_page_activate(m); 1418 else 1419 vm_page_deactivate(m); 1420 vm_page_xunbusy_unchecked(m); 1421 } 1422 1423 /* 1424 * Destroy the identity of an invalid page and free it if possible. 1425 * This is intended to be used when reading a page from backing store fails. 1426 */ 1427 void 1428 vm_page_free_invalid(vm_page_t m) 1429 { 1430 1431 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1432 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1433 KASSERT(m->object != NULL, ("page %p has no object", m)); 1434 VM_OBJECT_ASSERT_WLOCKED(m->object); 1435 1436 /* 1437 * We may be attempting to free the page as part of the handling for an 1438 * I/O error, in which case the page was xbusied by a different thread. 1439 */ 1440 vm_page_xbusy_claim(m); 1441 1442 /* 1443 * If someone has wired this page while the object lock 1444 * was not held, then the thread that unwires is responsible 1445 * for freeing the page. Otherwise just free the page now. 1446 * The wire count of this unmapped page cannot change while 1447 * we have the page xbusy and the page's object wlocked. 1448 */ 1449 if (vm_page_remove(m)) 1450 vm_page_free(m); 1451 } 1452 1453 /* 1454 * vm_page_dirty_KBI: [ internal use only ] 1455 * 1456 * Set all bits in the page's dirty field. 1457 * 1458 * The object containing the specified page must be locked if the 1459 * call is made from the machine-independent layer. 1460 * 1461 * See vm_page_clear_dirty_mask(). 1462 * 1463 * This function should only be called by vm_page_dirty(). 1464 */ 1465 void 1466 vm_page_dirty_KBI(vm_page_t m) 1467 { 1468 1469 /* Refer to this operation by its public name. */ 1470 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1471 m->dirty = VM_PAGE_BITS_ALL; 1472 } 1473 1474 /* 1475 * Insert the given page into the given object at the given pindex. mpred is 1476 * used for memq linkage. From vm_page_insert, lookup is true, mpred is 1477 * initially NULL, and this procedure looks it up. From vm_page_insert_after, 1478 * lookup is false and mpred is known to the caller to be valid, and may be 1479 * NULL if this will be the page with the lowest pindex. 1480 * 1481 * The procedure is marked __always_inline to suggest to the compiler to 1482 * eliminate the lookup parameter and the associated alternate branch. 1483 */ 1484 static __always_inline int 1485 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1486 vm_page_t mpred, bool lookup) 1487 { 1488 int error; 1489 1490 VM_OBJECT_ASSERT_WLOCKED(object); 1491 KASSERT(m->object == NULL, 1492 ("vm_page_insert: page %p already inserted", m)); 1493 1494 /* 1495 * Record the object/offset pair in this page. 1496 */ 1497 m->object = object; 1498 m->pindex = pindex; 1499 m->ref_count |= VPRC_OBJREF; 1500 1501 /* 1502 * Add this page to the object's radix tree, and look up mpred if 1503 * needed. 1504 */ 1505 if (lookup) 1506 error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred); 1507 else 1508 error = vm_radix_insert(&object->rtree, m); 1509 if (__predict_false(error != 0)) { 1510 m->object = NULL; 1511 m->pindex = 0; 1512 m->ref_count &= ~VPRC_OBJREF; 1513 return (1); 1514 } 1515 1516 /* 1517 * Now link into the object's ordered list of backed pages. 1518 */ 1519 vm_page_insert_radixdone(m, object, mpred); 1520 vm_pager_page_inserted(object, m); 1521 return (0); 1522 } 1523 1524 /* 1525 * vm_page_insert: [ internal use only ] 1526 * 1527 * Inserts the given mem entry into the object and object list. 1528 * 1529 * The object must be locked. 1530 */ 1531 int 1532 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1533 { 1534 return (vm_page_insert_lookup(m, object, pindex, NULL, true)); 1535 } 1536 1537 /* 1538 * vm_page_insert_after: 1539 * 1540 * Inserts the page "m" into the specified object at offset "pindex". 1541 * 1542 * The page "mpred" must immediately precede the offset "pindex" within 1543 * the specified object. 1544 * 1545 * The object must be locked. 1546 */ 1547 static int 1548 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1549 vm_page_t mpred) 1550 { 1551 return (vm_page_insert_lookup(m, object, pindex, mpred, false)); 1552 } 1553 1554 /* 1555 * vm_page_insert_radixdone: 1556 * 1557 * Complete page "m" insertion into the specified object after the 1558 * radix trie hooking. 1559 * 1560 * The page "mpred" must precede the offset "m->pindex" within the 1561 * specified object. 1562 * 1563 * The object must be locked. 1564 */ 1565 static void 1566 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1567 { 1568 1569 VM_OBJECT_ASSERT_WLOCKED(object); 1570 KASSERT(object != NULL && m->object == object, 1571 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1572 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1573 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1574 if (mpred != NULL) { 1575 KASSERT(mpred->object == object, 1576 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1577 KASSERT(mpred->pindex < m->pindex, 1578 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1579 KASSERT(TAILQ_NEXT(mpred, listq) == NULL || 1580 m->pindex < TAILQ_NEXT(mpred, listq)->pindex, 1581 ("vm_page_insert_radixdone: pindex doesn't precede msucc")); 1582 } else { 1583 KASSERT(TAILQ_EMPTY(&object->memq) || 1584 m->pindex < TAILQ_FIRST(&object->memq)->pindex, 1585 ("vm_page_insert_radixdone: no mpred but not first page")); 1586 } 1587 1588 if (mpred != NULL) 1589 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1590 else 1591 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1592 1593 /* 1594 * Show that the object has one more resident page. 1595 */ 1596 object->resident_page_count++; 1597 1598 /* 1599 * Hold the vnode until the last page is released. 1600 */ 1601 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1602 vhold(object->handle); 1603 1604 /* 1605 * Since we are inserting a new and possibly dirty page, 1606 * update the object's generation count. 1607 */ 1608 if (pmap_page_is_write_mapped(m)) 1609 vm_object_set_writeable_dirty(object); 1610 } 1611 1612 /* 1613 * Do the work to remove a page from its object. The caller is responsible for 1614 * updating the page's fields to reflect this removal. 1615 */ 1616 static void 1617 vm_page_object_remove(vm_page_t m) 1618 { 1619 vm_object_t object; 1620 vm_page_t mrem __diagused; 1621 1622 vm_page_assert_xbusied(m); 1623 object = m->object; 1624 VM_OBJECT_ASSERT_WLOCKED(object); 1625 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1626 ("page %p is missing its object ref", m)); 1627 1628 /* Deferred free of swap space. */ 1629 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1630 vm_pager_page_unswapped(m); 1631 1632 vm_pager_page_removed(object, m); 1633 1634 m->object = NULL; 1635 mrem = vm_radix_remove(&object->rtree, m->pindex); 1636 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1637 1638 /* 1639 * Now remove from the object's list of backed pages. 1640 */ 1641 TAILQ_REMOVE(&object->memq, m, listq); 1642 1643 /* 1644 * And show that the object has one fewer resident page. 1645 */ 1646 object->resident_page_count--; 1647 1648 /* 1649 * The vnode may now be recycled. 1650 */ 1651 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1652 vdrop(object->handle); 1653 } 1654 1655 /* 1656 * vm_page_remove: 1657 * 1658 * Removes the specified page from its containing object, but does not 1659 * invalidate any backing storage. Returns true if the object's reference 1660 * was the last reference to the page, and false otherwise. 1661 * 1662 * The object must be locked and the page must be exclusively busied. 1663 * The exclusive busy will be released on return. If this is not the 1664 * final ref and the caller does not hold a wire reference it may not 1665 * continue to access the page. 1666 */ 1667 bool 1668 vm_page_remove(vm_page_t m) 1669 { 1670 bool dropped; 1671 1672 dropped = vm_page_remove_xbusy(m); 1673 vm_page_xunbusy(m); 1674 1675 return (dropped); 1676 } 1677 1678 /* 1679 * vm_page_remove_xbusy 1680 * 1681 * Removes the page but leaves the xbusy held. Returns true if this 1682 * removed the final ref and false otherwise. 1683 */ 1684 bool 1685 vm_page_remove_xbusy(vm_page_t m) 1686 { 1687 1688 vm_page_object_remove(m); 1689 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1690 } 1691 1692 /* 1693 * vm_page_lookup: 1694 * 1695 * Returns the page associated with the object/offset 1696 * pair specified; if none is found, NULL is returned. 1697 * 1698 * The object must be locked. 1699 */ 1700 vm_page_t 1701 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1702 { 1703 1704 VM_OBJECT_ASSERT_LOCKED(object); 1705 return (vm_radix_lookup(&object->rtree, pindex)); 1706 } 1707 1708 /* 1709 * vm_page_lookup_unlocked: 1710 * 1711 * Returns the page associated with the object/offset pair specified; 1712 * if none is found, NULL is returned. The page may be no longer be 1713 * present in the object at the time that this function returns. Only 1714 * useful for opportunistic checks such as inmem(). 1715 */ 1716 vm_page_t 1717 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1718 { 1719 1720 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1721 } 1722 1723 /* 1724 * vm_page_relookup: 1725 * 1726 * Returns a page that must already have been busied by 1727 * the caller. Used for bogus page replacement. 1728 */ 1729 vm_page_t 1730 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1731 { 1732 vm_page_t m; 1733 1734 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 1735 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1736 m->object == object && m->pindex == pindex, 1737 ("vm_page_relookup: Invalid page %p", m)); 1738 return (m); 1739 } 1740 1741 /* 1742 * This should only be used by lockless functions for releasing transient 1743 * incorrect acquires. The page may have been freed after we acquired a 1744 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1745 * further to do. 1746 */ 1747 static void 1748 vm_page_busy_release(vm_page_t m) 1749 { 1750 u_int x; 1751 1752 x = vm_page_busy_fetch(m); 1753 for (;;) { 1754 if (x == VPB_FREED) 1755 break; 1756 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1757 if (atomic_fcmpset_int(&m->busy_lock, &x, 1758 x - VPB_ONE_SHARER)) 1759 break; 1760 continue; 1761 } 1762 KASSERT((x & VPB_BIT_SHARED) != 0 || 1763 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1764 ("vm_page_busy_release: %p xbusy not owned.", m)); 1765 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1766 continue; 1767 if ((x & VPB_BIT_WAITERS) != 0) 1768 wakeup(m); 1769 break; 1770 } 1771 } 1772 1773 /* 1774 * vm_page_find_least: 1775 * 1776 * Returns the page associated with the object with least pindex 1777 * greater than or equal to the parameter pindex, or NULL. 1778 * 1779 * The object must be locked. 1780 */ 1781 vm_page_t 1782 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1783 { 1784 vm_page_t m; 1785 1786 VM_OBJECT_ASSERT_LOCKED(object); 1787 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1788 m = vm_radix_lookup_ge(&object->rtree, pindex); 1789 return (m); 1790 } 1791 1792 /* 1793 * Returns the given page's successor (by pindex) within the object if it is 1794 * resident; if none is found, NULL is returned. 1795 * 1796 * The object must be locked. 1797 */ 1798 vm_page_t 1799 vm_page_next(vm_page_t m) 1800 { 1801 vm_page_t next; 1802 1803 VM_OBJECT_ASSERT_LOCKED(m->object); 1804 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1805 MPASS(next->object == m->object); 1806 if (next->pindex != m->pindex + 1) 1807 next = NULL; 1808 } 1809 return (next); 1810 } 1811 1812 /* 1813 * Returns the given page's predecessor (by pindex) within the object if it is 1814 * resident; if none is found, NULL is returned. 1815 * 1816 * The object must be locked. 1817 */ 1818 vm_page_t 1819 vm_page_prev(vm_page_t m) 1820 { 1821 vm_page_t prev; 1822 1823 VM_OBJECT_ASSERT_LOCKED(m->object); 1824 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1825 MPASS(prev->object == m->object); 1826 if (prev->pindex != m->pindex - 1) 1827 prev = NULL; 1828 } 1829 return (prev); 1830 } 1831 1832 /* 1833 * Uses the page mnew as a replacement for an existing page at index 1834 * pindex which must be already present in the object. 1835 * 1836 * Both pages must be exclusively busied on enter. The old page is 1837 * unbusied on exit. 1838 * 1839 * A return value of true means mold is now free. If this is not the 1840 * final ref and the caller does not hold a wire reference it may not 1841 * continue to access the page. 1842 */ 1843 static bool 1844 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1845 vm_page_t mold) 1846 { 1847 vm_page_t mret __diagused; 1848 bool dropped; 1849 1850 VM_OBJECT_ASSERT_WLOCKED(object); 1851 vm_page_assert_xbusied(mold); 1852 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1853 ("vm_page_replace: page %p already in object", mnew)); 1854 1855 /* 1856 * This function mostly follows vm_page_insert() and 1857 * vm_page_remove() without the radix, object count and vnode 1858 * dance. Double check such functions for more comments. 1859 */ 1860 1861 mnew->object = object; 1862 mnew->pindex = pindex; 1863 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1864 mret = vm_radix_replace(&object->rtree, mnew); 1865 KASSERT(mret == mold, 1866 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1867 KASSERT((mold->oflags & VPO_UNMANAGED) == 1868 (mnew->oflags & VPO_UNMANAGED), 1869 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1870 1871 /* Keep the resident page list in sorted order. */ 1872 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1873 TAILQ_REMOVE(&object->memq, mold, listq); 1874 mold->object = NULL; 1875 1876 /* 1877 * The object's resident_page_count does not change because we have 1878 * swapped one page for another, but the generation count should 1879 * change if the page is dirty. 1880 */ 1881 if (pmap_page_is_write_mapped(mnew)) 1882 vm_object_set_writeable_dirty(object); 1883 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1884 vm_page_xunbusy(mold); 1885 1886 return (dropped); 1887 } 1888 1889 void 1890 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1891 vm_page_t mold) 1892 { 1893 1894 vm_page_assert_xbusied(mnew); 1895 1896 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1897 vm_page_free(mold); 1898 } 1899 1900 /* 1901 * vm_page_rename: 1902 * 1903 * Move the given memory entry from its 1904 * current object to the specified target object/offset. 1905 * 1906 * Note: swap associated with the page must be invalidated by the move. We 1907 * have to do this for several reasons: (1) we aren't freeing the 1908 * page, (2) we are dirtying the page, (3) the VM system is probably 1909 * moving the page from object A to B, and will then later move 1910 * the backing store from A to B and we can't have a conflict. 1911 * 1912 * Note: we *always* dirty the page. It is necessary both for the 1913 * fact that we moved it, and because we may be invalidating 1914 * swap. 1915 * 1916 * The objects must be locked. 1917 */ 1918 int 1919 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1920 { 1921 vm_page_t mpred; 1922 vm_pindex_t opidx; 1923 1924 VM_OBJECT_ASSERT_WLOCKED(new_object); 1925 1926 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1927 1928 /* 1929 * Create a custom version of vm_page_insert() which does not depend 1930 * by m_prev and can cheat on the implementation aspects of the 1931 * function. 1932 */ 1933 opidx = m->pindex; 1934 m->pindex = new_pindex; 1935 if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) { 1936 m->pindex = opidx; 1937 return (1); 1938 } 1939 1940 /* 1941 * The operation cannot fail anymore. The removal must happen before 1942 * the listq iterator is tainted. 1943 */ 1944 m->pindex = opidx; 1945 vm_page_object_remove(m); 1946 1947 /* Return back to the new pindex to complete vm_page_insert(). */ 1948 m->pindex = new_pindex; 1949 m->object = new_object; 1950 1951 vm_page_insert_radixdone(m, new_object, mpred); 1952 vm_page_dirty(m); 1953 vm_pager_page_inserted(new_object, m); 1954 return (0); 1955 } 1956 1957 /* 1958 * vm_page_alloc: 1959 * 1960 * Allocate and return a page that is associated with the specified 1961 * object and offset pair. By default, this page is exclusive busied. 1962 * 1963 * The caller must always specify an allocation class. 1964 * 1965 * allocation classes: 1966 * VM_ALLOC_NORMAL normal process request 1967 * VM_ALLOC_SYSTEM system *really* needs a page 1968 * VM_ALLOC_INTERRUPT interrupt time request 1969 * 1970 * optional allocation flags: 1971 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1972 * intends to allocate 1973 * VM_ALLOC_NOBUSY do not exclusive busy the page 1974 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1975 * VM_ALLOC_SBUSY shared busy the allocated page 1976 * VM_ALLOC_WIRED wire the allocated page 1977 * VM_ALLOC_ZERO prefer a zeroed page 1978 */ 1979 vm_page_t 1980 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1981 { 1982 1983 return (vm_page_alloc_after(object, pindex, req, 1984 vm_radix_lookup_le(&object->rtree, pindex))); 1985 } 1986 1987 vm_page_t 1988 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1989 int req) 1990 { 1991 1992 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1993 vm_radix_lookup_le(&object->rtree, pindex))); 1994 } 1995 1996 /* 1997 * Allocate a page in the specified object with the given page index. To 1998 * optimize insertion of the page into the object, the caller must also specify 1999 * the resident page in the object with largest index smaller than the given 2000 * page index, or NULL if no such page exists. 2001 */ 2002 vm_page_t 2003 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 2004 int req, vm_page_t mpred) 2005 { 2006 struct vm_domainset_iter di; 2007 vm_page_t m; 2008 int domain; 2009 2010 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2011 do { 2012 m = vm_page_alloc_domain_after(object, pindex, domain, req, 2013 mpred); 2014 if (m != NULL) 2015 break; 2016 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2017 2018 return (m); 2019 } 2020 2021 /* 2022 * Returns true if the number of free pages exceeds the minimum 2023 * for the request class and false otherwise. 2024 */ 2025 static int 2026 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2027 { 2028 u_int limit, old, new; 2029 2030 if (req_class == VM_ALLOC_INTERRUPT) 2031 limit = 0; 2032 else if (req_class == VM_ALLOC_SYSTEM) 2033 limit = vmd->vmd_interrupt_free_min; 2034 else 2035 limit = vmd->vmd_free_reserved; 2036 2037 /* 2038 * Attempt to reserve the pages. Fail if we're below the limit. 2039 */ 2040 limit += npages; 2041 old = vmd->vmd_free_count; 2042 do { 2043 if (old < limit) 2044 return (0); 2045 new = old - npages; 2046 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2047 2048 /* Wake the page daemon if we've crossed the threshold. */ 2049 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2050 pagedaemon_wakeup(vmd->vmd_domain); 2051 2052 /* Only update bitsets on transitions. */ 2053 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2054 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2055 vm_domain_set(vmd); 2056 2057 return (1); 2058 } 2059 2060 int 2061 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2062 { 2063 int req_class; 2064 2065 /* 2066 * The page daemon is allowed to dig deeper into the free page list. 2067 */ 2068 req_class = req & VM_ALLOC_CLASS_MASK; 2069 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2070 req_class = VM_ALLOC_SYSTEM; 2071 return (_vm_domain_allocate(vmd, req_class, npages)); 2072 } 2073 2074 vm_page_t 2075 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 2076 int req, vm_page_t mpred) 2077 { 2078 struct vm_domain *vmd; 2079 vm_page_t m; 2080 int flags; 2081 2082 #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2083 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ 2084 VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ 2085 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) 2086 KASSERT((req & ~VPA_FLAGS) == 0, 2087 ("invalid request %#x", req)); 2088 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2089 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2090 ("invalid request %#x", req)); 2091 KASSERT(mpred == NULL || mpred->pindex < pindex, 2092 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2093 (uintmax_t)pindex)); 2094 VM_OBJECT_ASSERT_WLOCKED(object); 2095 2096 flags = 0; 2097 m = NULL; 2098 if (!vm_pager_can_alloc_page(object, pindex)) 2099 return (NULL); 2100 again: 2101 #if VM_NRESERVLEVEL > 0 2102 /* 2103 * Can we allocate the page from a reservation? 2104 */ 2105 if (vm_object_reserv(object) && 2106 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 2107 NULL) { 2108 goto found; 2109 } 2110 #endif 2111 vmd = VM_DOMAIN(domain); 2112 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2113 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2114 M_NOWAIT | M_NOVM); 2115 if (m != NULL) { 2116 flags |= PG_PCPU_CACHE; 2117 goto found; 2118 } 2119 } 2120 if (vm_domain_allocate(vmd, req, 1)) { 2121 /* 2122 * If not, allocate it from the free page queues. 2123 */ 2124 vm_domain_free_lock(vmd); 2125 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2126 vm_domain_free_unlock(vmd); 2127 if (m == NULL) { 2128 vm_domain_freecnt_inc(vmd, 1); 2129 #if VM_NRESERVLEVEL > 0 2130 if (vm_reserv_reclaim_inactive(domain)) 2131 goto again; 2132 #endif 2133 } 2134 } 2135 if (m == NULL) { 2136 /* 2137 * Not allocatable, give up. 2138 */ 2139 if (vm_domain_alloc_fail(vmd, object, req)) 2140 goto again; 2141 return (NULL); 2142 } 2143 2144 /* 2145 * At this point we had better have found a good page. 2146 */ 2147 found: 2148 vm_page_dequeue(m); 2149 vm_page_alloc_check(m); 2150 2151 /* 2152 * Initialize the page. Only the PG_ZERO flag is inherited. 2153 */ 2154 flags |= m->flags & PG_ZERO; 2155 if ((req & VM_ALLOC_NODUMP) != 0) 2156 flags |= PG_NODUMP; 2157 m->flags = flags; 2158 m->a.flags = 0; 2159 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2160 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2161 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2162 else if ((req & VM_ALLOC_SBUSY) != 0) 2163 m->busy_lock = VPB_SHARERS_WORD(1); 2164 else 2165 m->busy_lock = VPB_UNBUSIED; 2166 if (req & VM_ALLOC_WIRED) { 2167 vm_wire_add(1); 2168 m->ref_count = 1; 2169 } 2170 m->a.act_count = 0; 2171 2172 if (vm_page_insert_after(m, object, pindex, mpred)) { 2173 if (req & VM_ALLOC_WIRED) { 2174 vm_wire_sub(1); 2175 m->ref_count = 0; 2176 } 2177 KASSERT(m->object == NULL, ("page %p has object", m)); 2178 m->oflags = VPO_UNMANAGED; 2179 m->busy_lock = VPB_UNBUSIED; 2180 /* Don't change PG_ZERO. */ 2181 vm_page_free_toq(m); 2182 if (req & VM_ALLOC_WAITFAIL) { 2183 VM_OBJECT_WUNLOCK(object); 2184 vm_radix_wait(); 2185 VM_OBJECT_WLOCK(object); 2186 } 2187 return (NULL); 2188 } 2189 2190 /* Ignore device objects; the pager sets "memattr" for them. */ 2191 if (object->memattr != VM_MEMATTR_DEFAULT && 2192 (object->flags & OBJ_FICTITIOUS) == 0) 2193 pmap_page_set_memattr(m, object->memattr); 2194 2195 return (m); 2196 } 2197 2198 /* 2199 * vm_page_alloc_contig: 2200 * 2201 * Allocate a contiguous set of physical pages of the given size "npages" 2202 * from the free lists. All of the physical pages must be at or above 2203 * the given physical address "low" and below the given physical address 2204 * "high". The given value "alignment" determines the alignment of the 2205 * first physical page in the set. If the given value "boundary" is 2206 * non-zero, then the set of physical pages cannot cross any physical 2207 * address boundary that is a multiple of that value. Both "alignment" 2208 * and "boundary" must be a power of two. 2209 * 2210 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2211 * then the memory attribute setting for the physical pages is configured 2212 * to the object's memory attribute setting. Otherwise, the memory 2213 * attribute setting for the physical pages is configured to "memattr", 2214 * overriding the object's memory attribute setting. However, if the 2215 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2216 * memory attribute setting for the physical pages cannot be configured 2217 * to VM_MEMATTR_DEFAULT. 2218 * 2219 * The specified object may not contain fictitious pages. 2220 * 2221 * The caller must always specify an allocation class. 2222 * 2223 * allocation classes: 2224 * VM_ALLOC_NORMAL normal process request 2225 * VM_ALLOC_SYSTEM system *really* needs a page 2226 * VM_ALLOC_INTERRUPT interrupt time request 2227 * 2228 * optional allocation flags: 2229 * VM_ALLOC_NOBUSY do not exclusive busy the page 2230 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2231 * VM_ALLOC_SBUSY shared busy the allocated page 2232 * VM_ALLOC_WIRED wire the allocated page 2233 * VM_ALLOC_ZERO prefer a zeroed page 2234 */ 2235 vm_page_t 2236 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2237 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2238 vm_paddr_t boundary, vm_memattr_t memattr) 2239 { 2240 struct vm_domainset_iter di; 2241 vm_page_t bounds[2]; 2242 vm_page_t m; 2243 int domain; 2244 int start_segind; 2245 2246 start_segind = -1; 2247 2248 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2249 do { 2250 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2251 npages, low, high, alignment, boundary, memattr); 2252 if (m != NULL) 2253 break; 2254 if (start_segind == -1) 2255 start_segind = vm_phys_lookup_segind(low); 2256 if (vm_phys_find_range(bounds, start_segind, domain, 2257 npages, low, high) == -1) { 2258 vm_domainset_iter_ignore(&di, domain); 2259 } 2260 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2261 2262 return (m); 2263 } 2264 2265 static vm_page_t 2266 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2267 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2268 { 2269 struct vm_domain *vmd; 2270 vm_page_t m_ret; 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 vmd = VM_DOMAIN(domain); 2277 if (!vm_domain_allocate(vmd, req, npages)) 2278 return (NULL); 2279 /* 2280 * Try to allocate the pages from the free page queues. 2281 */ 2282 vm_domain_free_lock(vmd); 2283 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2284 alignment, boundary); 2285 vm_domain_free_unlock(vmd); 2286 if (m_ret != NULL) 2287 return (m_ret); 2288 #if VM_NRESERVLEVEL > 0 2289 /* 2290 * Try to break a reservation to allocate the pages. 2291 */ 2292 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2293 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2294 high, alignment, boundary); 2295 if (m_ret != NULL) 2296 return (m_ret); 2297 } 2298 #endif 2299 vm_domain_freecnt_inc(vmd, npages); 2300 return (NULL); 2301 } 2302 2303 vm_page_t 2304 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2305 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2306 vm_paddr_t boundary, vm_memattr_t memattr) 2307 { 2308 vm_page_t m, m_ret, mpred; 2309 u_int busy_lock, flags, oflags; 2310 2311 #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) 2312 KASSERT((req & ~VPAC_FLAGS) == 0, 2313 ("invalid request %#x", req)); 2314 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2315 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2316 ("invalid request %#x", req)); 2317 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2318 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2319 ("invalid request %#x", req)); 2320 VM_OBJECT_ASSERT_WLOCKED(object); 2321 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2322 ("vm_page_alloc_contig: object %p has fictitious pages", 2323 object)); 2324 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2325 2326 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2327 KASSERT(mpred == NULL || mpred->pindex != pindex, 2328 ("vm_page_alloc_contig: pindex already allocated")); 2329 for (;;) { 2330 #if VM_NRESERVLEVEL > 0 2331 /* 2332 * Can we allocate the pages from a reservation? 2333 */ 2334 if (vm_object_reserv(object) && 2335 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2336 mpred, npages, low, high, alignment, boundary)) != NULL) { 2337 break; 2338 } 2339 #endif 2340 if ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2341 low, high, alignment, boundary)) != NULL) 2342 break; 2343 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req)) 2344 return (NULL); 2345 } 2346 for (m = m_ret; m < &m_ret[npages]; m++) { 2347 vm_page_dequeue(m); 2348 vm_page_alloc_check(m); 2349 } 2350 2351 /* 2352 * Initialize the pages. Only the PG_ZERO flag is inherited. 2353 */ 2354 flags = PG_ZERO; 2355 if ((req & VM_ALLOC_NODUMP) != 0) 2356 flags |= PG_NODUMP; 2357 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2358 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2359 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2360 else if ((req & VM_ALLOC_SBUSY) != 0) 2361 busy_lock = VPB_SHARERS_WORD(1); 2362 else 2363 busy_lock = VPB_UNBUSIED; 2364 if ((req & VM_ALLOC_WIRED) != 0) 2365 vm_wire_add(npages); 2366 if (object->memattr != VM_MEMATTR_DEFAULT && 2367 memattr == VM_MEMATTR_DEFAULT) 2368 memattr = object->memattr; 2369 for (m = m_ret; m < &m_ret[npages]; m++) { 2370 m->a.flags = 0; 2371 m->flags = (m->flags | PG_NODUMP) & flags; 2372 m->busy_lock = busy_lock; 2373 if ((req & VM_ALLOC_WIRED) != 0) 2374 m->ref_count = 1; 2375 m->a.act_count = 0; 2376 m->oflags = oflags; 2377 if (vm_page_insert_after(m, object, pindex, mpred)) { 2378 if ((req & VM_ALLOC_WIRED) != 0) 2379 vm_wire_sub(npages); 2380 KASSERT(m->object == NULL, 2381 ("page %p has object", m)); 2382 mpred = m; 2383 for (m = m_ret; m < &m_ret[npages]; m++) { 2384 if (m <= mpred && 2385 (req & VM_ALLOC_WIRED) != 0) 2386 m->ref_count = 0; 2387 m->oflags = VPO_UNMANAGED; 2388 m->busy_lock = VPB_UNBUSIED; 2389 /* Don't change PG_ZERO. */ 2390 vm_page_free_toq(m); 2391 } 2392 if (req & VM_ALLOC_WAITFAIL) { 2393 VM_OBJECT_WUNLOCK(object); 2394 vm_radix_wait(); 2395 VM_OBJECT_WLOCK(object); 2396 } 2397 return (NULL); 2398 } 2399 mpred = m; 2400 if (memattr != VM_MEMATTR_DEFAULT) 2401 pmap_page_set_memattr(m, memattr); 2402 pindex++; 2403 } 2404 return (m_ret); 2405 } 2406 2407 /* 2408 * Allocate a physical page that is not intended to be inserted into a VM 2409 * object. If the "freelist" parameter is not equal to VM_NFREELIST, then only 2410 * pages from the specified vm_phys freelist will be returned. 2411 */ 2412 static __always_inline vm_page_t 2413 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req) 2414 { 2415 struct vm_domain *vmd; 2416 vm_page_t m; 2417 int flags; 2418 2419 #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2420 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ 2421 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ 2422 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK) 2423 KASSERT((req & ~VPAN_FLAGS) == 0, 2424 ("invalid request %#x", req)); 2425 2426 flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0; 2427 vmd = VM_DOMAIN(domain); 2428 again: 2429 if (freelist == VM_NFREELIST && 2430 vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2431 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2432 M_NOWAIT | M_NOVM); 2433 if (m != NULL) { 2434 flags |= PG_PCPU_CACHE; 2435 goto found; 2436 } 2437 } 2438 2439 if (vm_domain_allocate(vmd, req, 1)) { 2440 vm_domain_free_lock(vmd); 2441 if (freelist == VM_NFREELIST) 2442 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2443 else 2444 m = vm_phys_alloc_freelist_pages(domain, freelist, 2445 VM_FREEPOOL_DIRECT, 0); 2446 vm_domain_free_unlock(vmd); 2447 if (m == NULL) { 2448 vm_domain_freecnt_inc(vmd, 1); 2449 #if VM_NRESERVLEVEL > 0 2450 if (freelist == VM_NFREELIST && 2451 vm_reserv_reclaim_inactive(domain)) 2452 goto again; 2453 #endif 2454 } 2455 } 2456 if (m == NULL) { 2457 if (vm_domain_alloc_fail(vmd, NULL, req)) 2458 goto again; 2459 return (NULL); 2460 } 2461 2462 found: 2463 vm_page_dequeue(m); 2464 vm_page_alloc_check(m); 2465 2466 /* 2467 * Consumers should not rely on a useful default pindex value. 2468 */ 2469 m->pindex = 0xdeadc0dedeadc0de; 2470 m->flags = (m->flags & PG_ZERO) | flags; 2471 m->a.flags = 0; 2472 m->oflags = VPO_UNMANAGED; 2473 m->busy_lock = VPB_UNBUSIED; 2474 if ((req & VM_ALLOC_WIRED) != 0) { 2475 vm_wire_add(1); 2476 m->ref_count = 1; 2477 } 2478 2479 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2480 pmap_zero_page(m); 2481 2482 return (m); 2483 } 2484 2485 vm_page_t 2486 vm_page_alloc_freelist(int freelist, int req) 2487 { 2488 struct vm_domainset_iter di; 2489 vm_page_t m; 2490 int domain; 2491 2492 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2493 do { 2494 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2495 if (m != NULL) 2496 break; 2497 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2498 2499 return (m); 2500 } 2501 2502 vm_page_t 2503 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2504 { 2505 KASSERT(freelist >= 0 && freelist < VM_NFREELIST, 2506 ("%s: invalid freelist %d", __func__, freelist)); 2507 2508 return (_vm_page_alloc_noobj_domain(domain, freelist, req)); 2509 } 2510 2511 vm_page_t 2512 vm_page_alloc_noobj(int req) 2513 { 2514 struct vm_domainset_iter di; 2515 vm_page_t m; 2516 int domain; 2517 2518 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2519 do { 2520 m = vm_page_alloc_noobj_domain(domain, req); 2521 if (m != NULL) 2522 break; 2523 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2524 2525 return (m); 2526 } 2527 2528 vm_page_t 2529 vm_page_alloc_noobj_domain(int domain, int req) 2530 { 2531 return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req)); 2532 } 2533 2534 vm_page_t 2535 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2536 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2537 vm_memattr_t memattr) 2538 { 2539 struct vm_domainset_iter di; 2540 vm_page_t m; 2541 int domain; 2542 2543 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2544 do { 2545 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2546 high, alignment, boundary, memattr); 2547 if (m != NULL) 2548 break; 2549 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2550 2551 return (m); 2552 } 2553 2554 vm_page_t 2555 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2556 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2557 vm_memattr_t memattr) 2558 { 2559 vm_page_t m, m_ret; 2560 u_int flags; 2561 2562 #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) 2563 KASSERT((req & ~VPANC_FLAGS) == 0, 2564 ("invalid request %#x", req)); 2565 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2566 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2567 ("invalid request %#x", req)); 2568 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2569 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2570 ("invalid request %#x", req)); 2571 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2572 2573 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2574 low, high, alignment, boundary)) == NULL) { 2575 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2576 return (NULL); 2577 } 2578 2579 /* 2580 * Initialize the pages. Only the PG_ZERO flag is inherited. 2581 */ 2582 flags = PG_ZERO; 2583 if ((req & VM_ALLOC_NODUMP) != 0) 2584 flags |= PG_NODUMP; 2585 if ((req & VM_ALLOC_WIRED) != 0) 2586 vm_wire_add(npages); 2587 for (m = m_ret; m < &m_ret[npages]; m++) { 2588 vm_page_dequeue(m); 2589 vm_page_alloc_check(m); 2590 2591 /* 2592 * Consumers should not rely on a useful default pindex value. 2593 */ 2594 m->pindex = 0xdeadc0dedeadc0de; 2595 m->a.flags = 0; 2596 m->flags = (m->flags | PG_NODUMP) & flags; 2597 m->busy_lock = VPB_UNBUSIED; 2598 if ((req & VM_ALLOC_WIRED) != 0) 2599 m->ref_count = 1; 2600 m->a.act_count = 0; 2601 m->oflags = VPO_UNMANAGED; 2602 2603 /* 2604 * Zero the page before updating any mappings since the page is 2605 * not yet shared with any devices which might require the 2606 * non-default memory attribute. pmap_page_set_memattr() 2607 * flushes data caches before returning. 2608 */ 2609 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2610 pmap_zero_page(m); 2611 if (memattr != VM_MEMATTR_DEFAULT) 2612 pmap_page_set_memattr(m, memattr); 2613 } 2614 return (m_ret); 2615 } 2616 2617 /* 2618 * Check a page that has been freshly dequeued from a freelist. 2619 */ 2620 static void 2621 vm_page_alloc_check(vm_page_t m) 2622 { 2623 2624 KASSERT(m->object == NULL, ("page %p has object", m)); 2625 KASSERT(m->a.queue == PQ_NONE && 2626 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2627 ("page %p has unexpected queue %d, flags %#x", 2628 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2629 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2630 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2631 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2632 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2633 ("page %p has unexpected memattr %d", 2634 m, pmap_page_get_memattr(m))); 2635 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2636 pmap_vm_page_alloc_check(m); 2637 } 2638 2639 static int 2640 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2641 { 2642 struct vm_domain *vmd; 2643 struct vm_pgcache *pgcache; 2644 int i; 2645 2646 pgcache = arg; 2647 vmd = VM_DOMAIN(pgcache->domain); 2648 2649 /* 2650 * The page daemon should avoid creating extra memory pressure since its 2651 * main purpose is to replenish the store of free pages. 2652 */ 2653 if (vmd->vmd_severeset || curproc == pageproc || 2654 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2655 return (0); 2656 domain = vmd->vmd_domain; 2657 vm_domain_free_lock(vmd); 2658 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2659 (vm_page_t *)store); 2660 vm_domain_free_unlock(vmd); 2661 if (cnt != i) 2662 vm_domain_freecnt_inc(vmd, cnt - i); 2663 2664 return (i); 2665 } 2666 2667 static void 2668 vm_page_zone_release(void *arg, void **store, int cnt) 2669 { 2670 struct vm_domain *vmd; 2671 struct vm_pgcache *pgcache; 2672 vm_page_t m; 2673 int i; 2674 2675 pgcache = arg; 2676 vmd = VM_DOMAIN(pgcache->domain); 2677 vm_domain_free_lock(vmd); 2678 for (i = 0; i < cnt; i++) { 2679 m = (vm_page_t)store[i]; 2680 vm_phys_free_pages(m, 0); 2681 } 2682 vm_domain_free_unlock(vmd); 2683 vm_domain_freecnt_inc(vmd, cnt); 2684 } 2685 2686 #define VPSC_ANY 0 /* No restrictions. */ 2687 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2688 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2689 2690 /* 2691 * vm_page_scan_contig: 2692 * 2693 * Scan vm_page_array[] between the specified entries "m_start" and 2694 * "m_end" for a run of contiguous physical pages that satisfy the 2695 * specified conditions, and return the lowest page in the run. The 2696 * specified "alignment" determines the alignment of the lowest physical 2697 * page in the run. If the specified "boundary" is non-zero, then the 2698 * run of physical pages cannot span a physical address that is a 2699 * multiple of "boundary". 2700 * 2701 * "m_end" is never dereferenced, so it need not point to a vm_page 2702 * structure within vm_page_array[]. 2703 * 2704 * "npages" must be greater than zero. "m_start" and "m_end" must not 2705 * span a hole (or discontiguity) in the physical address space. Both 2706 * "alignment" and "boundary" must be a power of two. 2707 */ 2708 static vm_page_t 2709 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2710 u_long alignment, vm_paddr_t boundary, int options) 2711 { 2712 vm_object_t object; 2713 vm_paddr_t pa; 2714 vm_page_t m, m_run; 2715 #if VM_NRESERVLEVEL > 0 2716 int level; 2717 #endif 2718 int m_inc, order, run_ext, run_len; 2719 2720 KASSERT(npages > 0, ("npages is 0")); 2721 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2722 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2723 m_run = NULL; 2724 run_len = 0; 2725 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2726 KASSERT((m->flags & PG_MARKER) == 0, 2727 ("page %p is PG_MARKER", m)); 2728 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2729 ("fictitious page %p has invalid ref count", m)); 2730 2731 /* 2732 * If the current page would be the start of a run, check its 2733 * physical address against the end, alignment, and boundary 2734 * conditions. If it doesn't satisfy these conditions, either 2735 * terminate the scan or advance to the next page that 2736 * satisfies the failed condition. 2737 */ 2738 if (run_len == 0) { 2739 KASSERT(m_run == NULL, ("m_run != NULL")); 2740 if (m + npages > m_end) 2741 break; 2742 pa = VM_PAGE_TO_PHYS(m); 2743 if (!vm_addr_align_ok(pa, alignment)) { 2744 m_inc = atop(roundup2(pa, alignment) - pa); 2745 continue; 2746 } 2747 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2748 m_inc = atop(roundup2(pa, boundary) - pa); 2749 continue; 2750 } 2751 } else 2752 KASSERT(m_run != NULL, ("m_run == NULL")); 2753 2754 retry: 2755 m_inc = 1; 2756 if (vm_page_wired(m)) 2757 run_ext = 0; 2758 #if VM_NRESERVLEVEL > 0 2759 else if ((level = vm_reserv_level(m)) >= 0 && 2760 (options & VPSC_NORESERV) != 0) { 2761 run_ext = 0; 2762 /* Advance to the end of the reservation. */ 2763 pa = VM_PAGE_TO_PHYS(m); 2764 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2765 pa); 2766 } 2767 #endif 2768 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2769 /* 2770 * The page is considered eligible for relocation if 2771 * and only if it could be laundered or reclaimed by 2772 * the page daemon. 2773 */ 2774 VM_OBJECT_RLOCK(object); 2775 if (object != m->object) { 2776 VM_OBJECT_RUNLOCK(object); 2777 goto retry; 2778 } 2779 /* Don't care: PG_NODUMP, PG_ZERO. */ 2780 if ((object->flags & OBJ_SWAP) == 0 && 2781 object->type != OBJT_VNODE) { 2782 run_ext = 0; 2783 #if VM_NRESERVLEVEL > 0 2784 } else if ((options & VPSC_NOSUPER) != 0 && 2785 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2786 run_ext = 0; 2787 /* Advance to the end of the superpage. */ 2788 pa = VM_PAGE_TO_PHYS(m); 2789 m_inc = atop(roundup2(pa + 1, 2790 vm_reserv_size(level)) - pa); 2791 #endif 2792 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2793 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2794 /* 2795 * The page is allocated but eligible for 2796 * relocation. Extend the current run by one 2797 * page. 2798 */ 2799 KASSERT(pmap_page_get_memattr(m) == 2800 VM_MEMATTR_DEFAULT, 2801 ("page %p has an unexpected memattr", m)); 2802 KASSERT((m->oflags & (VPO_SWAPINPROG | 2803 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2804 ("page %p has unexpected oflags", m)); 2805 /* Don't care: PGA_NOSYNC. */ 2806 run_ext = 1; 2807 } else 2808 run_ext = 0; 2809 VM_OBJECT_RUNLOCK(object); 2810 #if VM_NRESERVLEVEL > 0 2811 } else if (level >= 0) { 2812 /* 2813 * The page is reserved but not yet allocated. In 2814 * other words, it is still free. Extend the current 2815 * run by one page. 2816 */ 2817 run_ext = 1; 2818 #endif 2819 } else if ((order = m->order) < VM_NFREEORDER) { 2820 /* 2821 * The page is enqueued in the physical memory 2822 * allocator's free page queues. Moreover, it is the 2823 * first page in a power-of-two-sized run of 2824 * contiguous free pages. Add these pages to the end 2825 * of the current run, and jump ahead. 2826 */ 2827 run_ext = 1 << order; 2828 m_inc = 1 << order; 2829 } else { 2830 /* 2831 * Skip the page for one of the following reasons: (1) 2832 * It is enqueued in the physical memory allocator's 2833 * free page queues. However, it is not the first 2834 * page in a run of contiguous free pages. (This case 2835 * rarely occurs because the scan is performed in 2836 * ascending order.) (2) It is not reserved, and it is 2837 * transitioning from free to allocated. (Conversely, 2838 * the transition from allocated to free for managed 2839 * pages is blocked by the page busy lock.) (3) It is 2840 * allocated but not contained by an object and not 2841 * wired, e.g., allocated by Xen's balloon driver. 2842 */ 2843 run_ext = 0; 2844 } 2845 2846 /* 2847 * Extend or reset the current run of pages. 2848 */ 2849 if (run_ext > 0) { 2850 if (run_len == 0) 2851 m_run = m; 2852 run_len += run_ext; 2853 } else { 2854 if (run_len > 0) { 2855 m_run = NULL; 2856 run_len = 0; 2857 } 2858 } 2859 } 2860 if (run_len >= npages) 2861 return (m_run); 2862 return (NULL); 2863 } 2864 2865 /* 2866 * vm_page_reclaim_run: 2867 * 2868 * Try to relocate each of the allocated virtual pages within the 2869 * specified run of physical pages to a new physical address. Free the 2870 * physical pages underlying the relocated virtual pages. A virtual page 2871 * is relocatable if and only if it could be laundered or reclaimed by 2872 * the page daemon. Whenever possible, a virtual page is relocated to a 2873 * physical address above "high". 2874 * 2875 * Returns 0 if every physical page within the run was already free or 2876 * just freed by a successful relocation. Otherwise, returns a non-zero 2877 * value indicating why the last attempt to relocate a virtual page was 2878 * unsuccessful. 2879 * 2880 * "req_class" must be an allocation class. 2881 */ 2882 static int 2883 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2884 vm_paddr_t high) 2885 { 2886 struct vm_domain *vmd; 2887 struct spglist free; 2888 vm_object_t object; 2889 vm_paddr_t pa; 2890 vm_page_t m, m_end, m_new; 2891 int error, order, req; 2892 2893 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2894 ("req_class is not an allocation class")); 2895 SLIST_INIT(&free); 2896 error = 0; 2897 m = m_run; 2898 m_end = m_run + npages; 2899 for (; error == 0 && m < m_end; m++) { 2900 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2901 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2902 2903 /* 2904 * Racily check for wirings. Races are handled once the object 2905 * lock is held and the page is unmapped. 2906 */ 2907 if (vm_page_wired(m)) 2908 error = EBUSY; 2909 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2910 /* 2911 * The page is relocated if and only if it could be 2912 * laundered or reclaimed by the page daemon. 2913 */ 2914 VM_OBJECT_WLOCK(object); 2915 /* Don't care: PG_NODUMP, PG_ZERO. */ 2916 if (m->object != object || 2917 ((object->flags & OBJ_SWAP) == 0 && 2918 object->type != OBJT_VNODE)) 2919 error = EINVAL; 2920 else if (object->memattr != VM_MEMATTR_DEFAULT) 2921 error = EINVAL; 2922 else if (vm_page_queue(m) != PQ_NONE && 2923 vm_page_tryxbusy(m) != 0) { 2924 if (vm_page_wired(m)) { 2925 vm_page_xunbusy(m); 2926 error = EBUSY; 2927 goto unlock; 2928 } 2929 KASSERT(pmap_page_get_memattr(m) == 2930 VM_MEMATTR_DEFAULT, 2931 ("page %p has an unexpected memattr", m)); 2932 KASSERT(m->oflags == 0, 2933 ("page %p has unexpected oflags", m)); 2934 /* Don't care: PGA_NOSYNC. */ 2935 if (!vm_page_none_valid(m)) { 2936 /* 2937 * First, try to allocate a new page 2938 * that is above "high". Failing 2939 * that, try to allocate a new page 2940 * that is below "m_run". Allocate 2941 * the new page between the end of 2942 * "m_run" and "high" only as a last 2943 * resort. 2944 */ 2945 req = req_class; 2946 if ((m->flags & PG_NODUMP) != 0) 2947 req |= VM_ALLOC_NODUMP; 2948 if (trunc_page(high) != 2949 ~(vm_paddr_t)PAGE_MASK) { 2950 m_new = 2951 vm_page_alloc_noobj_contig( 2952 req, 1, round_page(high), 2953 ~(vm_paddr_t)0, PAGE_SIZE, 2954 0, VM_MEMATTR_DEFAULT); 2955 } else 2956 m_new = NULL; 2957 if (m_new == NULL) { 2958 pa = VM_PAGE_TO_PHYS(m_run); 2959 m_new = 2960 vm_page_alloc_noobj_contig( 2961 req, 1, 0, pa - 1, 2962 PAGE_SIZE, 0, 2963 VM_MEMATTR_DEFAULT); 2964 } 2965 if (m_new == NULL) { 2966 pa += ptoa(npages); 2967 m_new = 2968 vm_page_alloc_noobj_contig( 2969 req, 1, pa, high, PAGE_SIZE, 2970 0, VM_MEMATTR_DEFAULT); 2971 } 2972 if (m_new == NULL) { 2973 vm_page_xunbusy(m); 2974 error = ENOMEM; 2975 goto unlock; 2976 } 2977 2978 /* 2979 * Unmap the page and check for new 2980 * wirings that may have been acquired 2981 * through a pmap lookup. 2982 */ 2983 if (object->ref_count != 0 && 2984 !vm_page_try_remove_all(m)) { 2985 vm_page_xunbusy(m); 2986 vm_page_free(m_new); 2987 error = EBUSY; 2988 goto unlock; 2989 } 2990 2991 /* 2992 * Replace "m" with the new page. For 2993 * vm_page_replace(), "m" must be busy 2994 * and dequeued. Finally, change "m" 2995 * as if vm_page_free() was called. 2996 */ 2997 m_new->a.flags = m->a.flags & 2998 ~PGA_QUEUE_STATE_MASK; 2999 KASSERT(m_new->oflags == VPO_UNMANAGED, 3000 ("page %p is managed", m_new)); 3001 m_new->oflags = 0; 3002 pmap_copy_page(m, m_new); 3003 m_new->valid = m->valid; 3004 m_new->dirty = m->dirty; 3005 m->flags &= ~PG_ZERO; 3006 vm_page_dequeue(m); 3007 if (vm_page_replace_hold(m_new, object, 3008 m->pindex, m) && 3009 vm_page_free_prep(m)) 3010 SLIST_INSERT_HEAD(&free, m, 3011 plinks.s.ss); 3012 3013 /* 3014 * The new page must be deactivated 3015 * before the object is unlocked. 3016 */ 3017 vm_page_deactivate(m_new); 3018 } else { 3019 m->flags &= ~PG_ZERO; 3020 vm_page_dequeue(m); 3021 if (vm_page_free_prep(m)) 3022 SLIST_INSERT_HEAD(&free, m, 3023 plinks.s.ss); 3024 KASSERT(m->dirty == 0, 3025 ("page %p is dirty", m)); 3026 } 3027 } else 3028 error = EBUSY; 3029 unlock: 3030 VM_OBJECT_WUNLOCK(object); 3031 } else { 3032 MPASS(vm_page_domain(m) == domain); 3033 vmd = VM_DOMAIN(domain); 3034 vm_domain_free_lock(vmd); 3035 order = m->order; 3036 if (order < VM_NFREEORDER) { 3037 /* 3038 * The page is enqueued in the physical memory 3039 * allocator's free page queues. Moreover, it 3040 * is the first page in a power-of-two-sized 3041 * run of contiguous free pages. Jump ahead 3042 * to the last page within that run, and 3043 * continue from there. 3044 */ 3045 m += (1 << order) - 1; 3046 } 3047 #if VM_NRESERVLEVEL > 0 3048 else if (vm_reserv_is_page_free(m)) 3049 order = 0; 3050 #endif 3051 vm_domain_free_unlock(vmd); 3052 if (order == VM_NFREEORDER) 3053 error = EINVAL; 3054 } 3055 } 3056 if ((m = SLIST_FIRST(&free)) != NULL) { 3057 int cnt; 3058 3059 vmd = VM_DOMAIN(domain); 3060 cnt = 0; 3061 vm_domain_free_lock(vmd); 3062 do { 3063 MPASS(vm_page_domain(m) == domain); 3064 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3065 vm_phys_free_pages(m, 0); 3066 cnt++; 3067 } while ((m = SLIST_FIRST(&free)) != NULL); 3068 vm_domain_free_unlock(vmd); 3069 vm_domain_freecnt_inc(vmd, cnt); 3070 } 3071 return (error); 3072 } 3073 3074 #define NRUNS 16 3075 3076 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 3077 3078 #define MIN_RECLAIM 8 3079 3080 /* 3081 * vm_page_reclaim_contig: 3082 * 3083 * Reclaim allocated, contiguous physical memory satisfying the specified 3084 * conditions by relocating the virtual pages using that physical memory. 3085 * Returns 0 if reclamation is successful, ERANGE if the specified domain 3086 * can't possibly satisfy the reclamation request, or ENOMEM if not 3087 * currently able to reclaim the requested number of pages. Since 3088 * relocation requires the allocation of physical pages, reclamation may 3089 * fail with ENOMEM due to a shortage of free pages. When reclamation 3090 * fails in this manner, callers are expected to perform vm_wait() before 3091 * retrying a failed allocation operation, e.g., vm_page_alloc_contig(). 3092 * 3093 * The caller must always specify an allocation class through "req". 3094 * 3095 * allocation classes: 3096 * VM_ALLOC_NORMAL normal process request 3097 * VM_ALLOC_SYSTEM system *really* needs a page 3098 * VM_ALLOC_INTERRUPT interrupt time request 3099 * 3100 * The optional allocation flags are ignored. 3101 * 3102 * "npages" must be greater than zero. Both "alignment" and "boundary" 3103 * must be a power of two. 3104 */ 3105 int 3106 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3107 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3108 int desired_runs) 3109 { 3110 struct vm_domain *vmd; 3111 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; 3112 u_long count, minalign, reclaimed; 3113 int error, i, min_reclaim, nruns, options, req_class; 3114 int segind, start_segind; 3115 int ret; 3116 3117 KASSERT(npages > 0, ("npages is 0")); 3118 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3119 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3120 3121 ret = ENOMEM; 3122 3123 /* 3124 * If the caller wants to reclaim multiple runs, try to allocate 3125 * space to store the runs. If that fails, fall back to the old 3126 * behavior of just reclaiming MIN_RECLAIM pages. 3127 */ 3128 if (desired_runs > 1) 3129 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3130 M_TEMP, M_NOWAIT); 3131 else 3132 m_runs = NULL; 3133 3134 if (m_runs == NULL) { 3135 m_runs = _m_runs; 3136 nruns = NRUNS; 3137 } else { 3138 nruns = NRUNS + desired_runs - 1; 3139 } 3140 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3141 3142 /* 3143 * The caller will attempt an allocation after some runs have been 3144 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3145 * of vm_phys_alloc_contig(), round up the requested length to the next 3146 * power of two or maximum chunk size, and ensure that each run is 3147 * suitably aligned. 3148 */ 3149 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3150 npages = roundup2(npages, minalign); 3151 if (alignment < ptoa(minalign)) 3152 alignment = ptoa(minalign); 3153 3154 /* 3155 * The page daemon is allowed to dig deeper into the free page list. 3156 */ 3157 req_class = req & VM_ALLOC_CLASS_MASK; 3158 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3159 req_class = VM_ALLOC_SYSTEM; 3160 3161 start_segind = vm_phys_lookup_segind(low); 3162 3163 /* 3164 * Return if the number of free pages cannot satisfy the requested 3165 * allocation. 3166 */ 3167 vmd = VM_DOMAIN(domain); 3168 count = vmd->vmd_free_count; 3169 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3170 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3171 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3172 goto done; 3173 3174 /* 3175 * Scan up to three times, relaxing the restrictions ("options") on 3176 * the reclamation of reservations and superpages each time. 3177 */ 3178 for (options = VPSC_NORESERV;;) { 3179 bool phys_range_exists = false; 3180 3181 /* 3182 * Find the highest runs that satisfy the given constraints 3183 * and restrictions, and record them in "m_runs". 3184 */ 3185 count = 0; 3186 segind = start_segind; 3187 while ((segind = vm_phys_find_range(bounds, segind, domain, 3188 npages, low, high)) != -1) { 3189 phys_range_exists = true; 3190 while ((m_run = vm_page_scan_contig(npages, bounds[0], 3191 bounds[1], alignment, boundary, options))) { 3192 bounds[0] = m_run + npages; 3193 m_runs[RUN_INDEX(count, nruns)] = m_run; 3194 count++; 3195 } 3196 segind++; 3197 } 3198 3199 if (!phys_range_exists) { 3200 ret = ERANGE; 3201 goto done; 3202 } 3203 3204 /* 3205 * Reclaim the highest runs in LIFO (descending) order until 3206 * the number of reclaimed pages, "reclaimed", is at least 3207 * "min_reclaim". Reset "reclaimed" each time because each 3208 * reclamation is idempotent, and runs will (likely) recur 3209 * from one scan to the next as restrictions are relaxed. 3210 */ 3211 reclaimed = 0; 3212 for (i = 0; count > 0 && i < nruns; i++) { 3213 count--; 3214 m_run = m_runs[RUN_INDEX(count, nruns)]; 3215 error = vm_page_reclaim_run(req_class, domain, npages, 3216 m_run, high); 3217 if (error == 0) { 3218 reclaimed += npages; 3219 if (reclaimed >= min_reclaim) { 3220 ret = 0; 3221 goto done; 3222 } 3223 } 3224 } 3225 3226 /* 3227 * Either relax the restrictions on the next scan or return if 3228 * the last scan had no restrictions. 3229 */ 3230 if (options == VPSC_NORESERV) 3231 options = VPSC_NOSUPER; 3232 else if (options == VPSC_NOSUPER) 3233 options = VPSC_ANY; 3234 else if (options == VPSC_ANY) { 3235 if (reclaimed != 0) 3236 ret = 0; 3237 goto done; 3238 } 3239 } 3240 done: 3241 if (m_runs != _m_runs) 3242 free(m_runs, M_TEMP); 3243 return (ret); 3244 } 3245 3246 int 3247 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3248 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3249 { 3250 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high, 3251 alignment, boundary, 1)); 3252 } 3253 3254 int 3255 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3256 u_long alignment, vm_paddr_t boundary) 3257 { 3258 struct vm_domainset_iter di; 3259 int domain, ret, status; 3260 3261 ret = ERANGE; 3262 3263 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3264 do { 3265 status = vm_page_reclaim_contig_domain(domain, req, npages, low, 3266 high, alignment, boundary); 3267 if (status == 0) 3268 return (0); 3269 else if (status == ERANGE) 3270 vm_domainset_iter_ignore(&di, domain); 3271 else { 3272 KASSERT(status == ENOMEM, ("Unrecognized error %d " 3273 "from vm_page_reclaim_contig_domain()", status)); 3274 ret = ENOMEM; 3275 } 3276 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3277 3278 return (ret); 3279 } 3280 3281 /* 3282 * Set the domain in the appropriate page level domainset. 3283 */ 3284 void 3285 vm_domain_set(struct vm_domain *vmd) 3286 { 3287 3288 mtx_lock(&vm_domainset_lock); 3289 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3290 vmd->vmd_minset = 1; 3291 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3292 } 3293 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3294 vmd->vmd_severeset = 1; 3295 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3296 } 3297 mtx_unlock(&vm_domainset_lock); 3298 } 3299 3300 /* 3301 * Clear the domain from the appropriate page level domainset. 3302 */ 3303 void 3304 vm_domain_clear(struct vm_domain *vmd) 3305 { 3306 3307 mtx_lock(&vm_domainset_lock); 3308 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3309 vmd->vmd_minset = 0; 3310 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3311 if (vm_min_waiters != 0) { 3312 vm_min_waiters = 0; 3313 wakeup(&vm_min_domains); 3314 } 3315 } 3316 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3317 vmd->vmd_severeset = 0; 3318 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3319 if (vm_severe_waiters != 0) { 3320 vm_severe_waiters = 0; 3321 wakeup(&vm_severe_domains); 3322 } 3323 } 3324 3325 /* 3326 * If pageout daemon needs pages, then tell it that there are 3327 * some free. 3328 */ 3329 if (vmd->vmd_pageout_pages_needed && 3330 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3331 wakeup(&vmd->vmd_pageout_pages_needed); 3332 vmd->vmd_pageout_pages_needed = 0; 3333 } 3334 3335 /* See comments in vm_wait_doms(). */ 3336 if (vm_pageproc_waiters) { 3337 vm_pageproc_waiters = 0; 3338 wakeup(&vm_pageproc_waiters); 3339 } 3340 mtx_unlock(&vm_domainset_lock); 3341 } 3342 3343 /* 3344 * Wait for free pages to exceed the min threshold globally. 3345 */ 3346 void 3347 vm_wait_min(void) 3348 { 3349 3350 mtx_lock(&vm_domainset_lock); 3351 while (vm_page_count_min()) { 3352 vm_min_waiters++; 3353 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3354 } 3355 mtx_unlock(&vm_domainset_lock); 3356 } 3357 3358 /* 3359 * Wait for free pages to exceed the severe threshold globally. 3360 */ 3361 void 3362 vm_wait_severe(void) 3363 { 3364 3365 mtx_lock(&vm_domainset_lock); 3366 while (vm_page_count_severe()) { 3367 vm_severe_waiters++; 3368 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3369 "vmwait", 0); 3370 } 3371 mtx_unlock(&vm_domainset_lock); 3372 } 3373 3374 u_int 3375 vm_wait_count(void) 3376 { 3377 3378 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3379 } 3380 3381 int 3382 vm_wait_doms(const domainset_t *wdoms, int mflags) 3383 { 3384 int error; 3385 3386 error = 0; 3387 3388 /* 3389 * We use racey wakeup synchronization to avoid expensive global 3390 * locking for the pageproc when sleeping with a non-specific vm_wait. 3391 * To handle this, we only sleep for one tick in this instance. It 3392 * is expected that most allocations for the pageproc will come from 3393 * kmem or vm_page_grab* which will use the more specific and 3394 * race-free vm_wait_domain(). 3395 */ 3396 if (curproc == pageproc) { 3397 mtx_lock(&vm_domainset_lock); 3398 vm_pageproc_waiters++; 3399 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3400 PVM | PDROP | mflags, "pageprocwait", 1); 3401 } else { 3402 /* 3403 * XXX Ideally we would wait only until the allocation could 3404 * be satisfied. This condition can cause new allocators to 3405 * consume all freed pages while old allocators wait. 3406 */ 3407 mtx_lock(&vm_domainset_lock); 3408 if (vm_page_count_min_set(wdoms)) { 3409 if (pageproc == NULL) 3410 panic("vm_wait in early boot"); 3411 vm_min_waiters++; 3412 error = msleep(&vm_min_domains, &vm_domainset_lock, 3413 PVM | PDROP | mflags, "vmwait", 0); 3414 } else 3415 mtx_unlock(&vm_domainset_lock); 3416 } 3417 return (error); 3418 } 3419 3420 /* 3421 * vm_wait_domain: 3422 * 3423 * Sleep until free pages are available for allocation. 3424 * - Called in various places after failed memory allocations. 3425 */ 3426 void 3427 vm_wait_domain(int domain) 3428 { 3429 struct vm_domain *vmd; 3430 domainset_t wdom; 3431 3432 vmd = VM_DOMAIN(domain); 3433 vm_domain_free_assert_unlocked(vmd); 3434 3435 if (curproc == pageproc) { 3436 mtx_lock(&vm_domainset_lock); 3437 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3438 vmd->vmd_pageout_pages_needed = 1; 3439 msleep(&vmd->vmd_pageout_pages_needed, 3440 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3441 } else 3442 mtx_unlock(&vm_domainset_lock); 3443 } else { 3444 DOMAINSET_ZERO(&wdom); 3445 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3446 vm_wait_doms(&wdom, 0); 3447 } 3448 } 3449 3450 static int 3451 vm_wait_flags(vm_object_t obj, int mflags) 3452 { 3453 struct domainset *d; 3454 3455 d = NULL; 3456 3457 /* 3458 * Carefully fetch pointers only once: the struct domainset 3459 * itself is ummutable but the pointer might change. 3460 */ 3461 if (obj != NULL) 3462 d = obj->domain.dr_policy; 3463 if (d == NULL) 3464 d = curthread->td_domain.dr_policy; 3465 3466 return (vm_wait_doms(&d->ds_mask, mflags)); 3467 } 3468 3469 /* 3470 * vm_wait: 3471 * 3472 * Sleep until free pages are available for allocation in the 3473 * affinity domains of the obj. If obj is NULL, the domain set 3474 * for the calling thread is used. 3475 * Called in various places after failed memory allocations. 3476 */ 3477 void 3478 vm_wait(vm_object_t obj) 3479 { 3480 (void)vm_wait_flags(obj, 0); 3481 } 3482 3483 int 3484 vm_wait_intr(vm_object_t obj) 3485 { 3486 return (vm_wait_flags(obj, PCATCH)); 3487 } 3488 3489 /* 3490 * vm_domain_alloc_fail: 3491 * 3492 * Called when a page allocation function fails. Informs the 3493 * pagedaemon and performs the requested wait. Requires the 3494 * domain_free and object lock on entry. Returns with the 3495 * object lock held and free lock released. Returns an error when 3496 * retry is necessary. 3497 * 3498 */ 3499 static int 3500 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3501 { 3502 3503 vm_domain_free_assert_unlocked(vmd); 3504 3505 atomic_add_int(&vmd->vmd_pageout_deficit, 3506 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3507 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3508 if (object != NULL) 3509 VM_OBJECT_WUNLOCK(object); 3510 vm_wait_domain(vmd->vmd_domain); 3511 if (object != NULL) 3512 VM_OBJECT_WLOCK(object); 3513 if (req & VM_ALLOC_WAITOK) 3514 return (EAGAIN); 3515 } 3516 3517 return (0); 3518 } 3519 3520 /* 3521 * vm_waitpfault: 3522 * 3523 * Sleep until free pages are available for allocation. 3524 * - Called only in vm_fault so that processes page faulting 3525 * can be easily tracked. 3526 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3527 * processes will be able to grab memory first. Do not change 3528 * this balance without careful testing first. 3529 */ 3530 void 3531 vm_waitpfault(struct domainset *dset, int timo) 3532 { 3533 3534 /* 3535 * XXX Ideally we would wait only until the allocation could 3536 * be satisfied. This condition can cause new allocators to 3537 * consume all freed pages while old allocators wait. 3538 */ 3539 mtx_lock(&vm_domainset_lock); 3540 if (vm_page_count_min_set(&dset->ds_mask)) { 3541 vm_min_waiters++; 3542 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3543 "pfault", timo); 3544 } else 3545 mtx_unlock(&vm_domainset_lock); 3546 } 3547 3548 static struct vm_pagequeue * 3549 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3550 { 3551 3552 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3553 } 3554 3555 #ifdef INVARIANTS 3556 static struct vm_pagequeue * 3557 vm_page_pagequeue(vm_page_t m) 3558 { 3559 3560 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3561 } 3562 #endif 3563 3564 static __always_inline bool 3565 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3566 { 3567 vm_page_astate_t tmp; 3568 3569 tmp = *old; 3570 do { 3571 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3572 return (true); 3573 counter_u64_add(pqstate_commit_retries, 1); 3574 } while (old->_bits == tmp._bits); 3575 3576 return (false); 3577 } 3578 3579 /* 3580 * Do the work of committing a queue state update that moves the page out of 3581 * its current queue. 3582 */ 3583 static bool 3584 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3585 vm_page_astate_t *old, vm_page_astate_t new) 3586 { 3587 vm_page_t next; 3588 3589 vm_pagequeue_assert_locked(pq); 3590 KASSERT(vm_page_pagequeue(m) == pq, 3591 ("%s: queue %p does not match page %p", __func__, pq, m)); 3592 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3593 ("%s: invalid queue indices %d %d", 3594 __func__, old->queue, new.queue)); 3595 3596 /* 3597 * Once the queue index of the page changes there is nothing 3598 * synchronizing with further updates to the page's physical 3599 * queue state. Therefore we must speculatively remove the page 3600 * from the queue now and be prepared to roll back if the queue 3601 * state update fails. If the page is not physically enqueued then 3602 * we just update its queue index. 3603 */ 3604 if ((old->flags & PGA_ENQUEUED) != 0) { 3605 new.flags &= ~PGA_ENQUEUED; 3606 next = TAILQ_NEXT(m, plinks.q); 3607 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3608 vm_pagequeue_cnt_dec(pq); 3609 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3610 if (next == NULL) 3611 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3612 else 3613 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3614 vm_pagequeue_cnt_inc(pq); 3615 return (false); 3616 } else { 3617 return (true); 3618 } 3619 } else { 3620 return (vm_page_pqstate_fcmpset(m, old, new)); 3621 } 3622 } 3623 3624 static bool 3625 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3626 vm_page_astate_t new) 3627 { 3628 struct vm_pagequeue *pq; 3629 vm_page_astate_t as; 3630 bool ret; 3631 3632 pq = _vm_page_pagequeue(m, old->queue); 3633 3634 /* 3635 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3636 * corresponding page queue lock is held. 3637 */ 3638 vm_pagequeue_lock(pq); 3639 as = vm_page_astate_load(m); 3640 if (__predict_false(as._bits != old->_bits)) { 3641 *old = as; 3642 ret = false; 3643 } else { 3644 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3645 } 3646 vm_pagequeue_unlock(pq); 3647 return (ret); 3648 } 3649 3650 /* 3651 * Commit a queue state update that enqueues or requeues a page. 3652 */ 3653 static bool 3654 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3655 vm_page_astate_t *old, vm_page_astate_t new) 3656 { 3657 struct vm_domain *vmd; 3658 3659 vm_pagequeue_assert_locked(pq); 3660 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3661 ("%s: invalid queue indices %d %d", 3662 __func__, old->queue, new.queue)); 3663 3664 new.flags |= PGA_ENQUEUED; 3665 if (!vm_page_pqstate_fcmpset(m, old, new)) 3666 return (false); 3667 3668 if ((old->flags & PGA_ENQUEUED) != 0) 3669 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3670 else 3671 vm_pagequeue_cnt_inc(pq); 3672 3673 /* 3674 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3675 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3676 * applied, even if it was set first. 3677 */ 3678 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3679 vmd = vm_pagequeue_domain(m); 3680 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3681 ("%s: invalid page queue for page %p", __func__, m)); 3682 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3683 } else { 3684 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3685 } 3686 return (true); 3687 } 3688 3689 /* 3690 * Commit a queue state update that encodes a request for a deferred queue 3691 * operation. 3692 */ 3693 static bool 3694 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3695 vm_page_astate_t new) 3696 { 3697 3698 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3699 ("%s: invalid state, queue %d flags %x", 3700 __func__, new.queue, new.flags)); 3701 3702 if (old->_bits != new._bits && 3703 !vm_page_pqstate_fcmpset(m, old, new)) 3704 return (false); 3705 vm_page_pqbatch_submit(m, new.queue); 3706 return (true); 3707 } 3708 3709 /* 3710 * A generic queue state update function. This handles more cases than the 3711 * specialized functions above. 3712 */ 3713 bool 3714 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3715 { 3716 3717 if (old->_bits == new._bits) 3718 return (true); 3719 3720 if (old->queue != PQ_NONE && new.queue != old->queue) { 3721 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3722 return (false); 3723 if (new.queue != PQ_NONE) 3724 vm_page_pqbatch_submit(m, new.queue); 3725 } else { 3726 if (!vm_page_pqstate_fcmpset(m, old, new)) 3727 return (false); 3728 if (new.queue != PQ_NONE && 3729 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3730 vm_page_pqbatch_submit(m, new.queue); 3731 } 3732 return (true); 3733 } 3734 3735 /* 3736 * Apply deferred queue state updates to a page. 3737 */ 3738 static inline void 3739 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3740 { 3741 vm_page_astate_t new, old; 3742 3743 CRITICAL_ASSERT(curthread); 3744 vm_pagequeue_assert_locked(pq); 3745 KASSERT(queue < PQ_COUNT, 3746 ("%s: invalid queue index %d", __func__, queue)); 3747 KASSERT(pq == _vm_page_pagequeue(m, queue), 3748 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3749 3750 for (old = vm_page_astate_load(m);;) { 3751 if (__predict_false(old.queue != queue || 3752 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3753 counter_u64_add(queue_nops, 1); 3754 break; 3755 } 3756 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3757 ("%s: page %p is unmanaged", __func__, m)); 3758 3759 new = old; 3760 if ((old.flags & PGA_DEQUEUE) != 0) { 3761 new.flags &= ~PGA_QUEUE_OP_MASK; 3762 new.queue = PQ_NONE; 3763 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3764 m, &old, new))) { 3765 counter_u64_add(queue_ops, 1); 3766 break; 3767 } 3768 } else { 3769 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3770 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3771 m, &old, new))) { 3772 counter_u64_add(queue_ops, 1); 3773 break; 3774 } 3775 } 3776 } 3777 } 3778 3779 static void 3780 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3781 uint8_t queue) 3782 { 3783 int i; 3784 3785 for (i = 0; i < bq->bq_cnt; i++) 3786 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3787 vm_batchqueue_init(bq); 3788 } 3789 3790 /* 3791 * vm_page_pqbatch_submit: [ internal use only ] 3792 * 3793 * Enqueue a page in the specified page queue's batched work queue. 3794 * The caller must have encoded the requested operation in the page 3795 * structure's a.flags field. 3796 */ 3797 void 3798 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3799 { 3800 struct vm_batchqueue *bq; 3801 struct vm_pagequeue *pq; 3802 int domain, slots_remaining; 3803 3804 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3805 3806 domain = vm_page_domain(m); 3807 critical_enter(); 3808 bq = DPCPU_PTR(pqbatch[domain][queue]); 3809 slots_remaining = vm_batchqueue_insert(bq, m); 3810 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 3811 /* keep building the bq */ 3812 critical_exit(); 3813 return; 3814 } else if (slots_remaining > 0 ) { 3815 /* Try to process the bq if we can get the lock */ 3816 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3817 if (vm_pagequeue_trylock(pq)) { 3818 vm_pqbatch_process(pq, bq, queue); 3819 vm_pagequeue_unlock(pq); 3820 } 3821 critical_exit(); 3822 return; 3823 } 3824 critical_exit(); 3825 3826 /* if we make it here, the bq is full so wait for the lock */ 3827 3828 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3829 vm_pagequeue_lock(pq); 3830 critical_enter(); 3831 bq = DPCPU_PTR(pqbatch[domain][queue]); 3832 vm_pqbatch_process(pq, bq, queue); 3833 vm_pqbatch_process_page(pq, m, queue); 3834 vm_pagequeue_unlock(pq); 3835 critical_exit(); 3836 } 3837 3838 /* 3839 * vm_page_pqbatch_drain: [ internal use only ] 3840 * 3841 * Force all per-CPU page queue batch queues to be drained. This is 3842 * intended for use in severe memory shortages, to ensure that pages 3843 * do not remain stuck in the batch queues. 3844 */ 3845 void 3846 vm_page_pqbatch_drain(void) 3847 { 3848 struct thread *td; 3849 struct vm_domain *vmd; 3850 struct vm_pagequeue *pq; 3851 int cpu, domain, queue; 3852 3853 td = curthread; 3854 CPU_FOREACH(cpu) { 3855 thread_lock(td); 3856 sched_bind(td, cpu); 3857 thread_unlock(td); 3858 3859 for (domain = 0; domain < vm_ndomains; domain++) { 3860 vmd = VM_DOMAIN(domain); 3861 for (queue = 0; queue < PQ_COUNT; queue++) { 3862 pq = &vmd->vmd_pagequeues[queue]; 3863 vm_pagequeue_lock(pq); 3864 critical_enter(); 3865 vm_pqbatch_process(pq, 3866 DPCPU_PTR(pqbatch[domain][queue]), queue); 3867 critical_exit(); 3868 vm_pagequeue_unlock(pq); 3869 } 3870 } 3871 } 3872 thread_lock(td); 3873 sched_unbind(td); 3874 thread_unlock(td); 3875 } 3876 3877 /* 3878 * vm_page_dequeue_deferred: [ internal use only ] 3879 * 3880 * Request removal of the given page from its current page 3881 * queue. Physical removal from the queue may be deferred 3882 * indefinitely. 3883 */ 3884 void 3885 vm_page_dequeue_deferred(vm_page_t m) 3886 { 3887 vm_page_astate_t new, old; 3888 3889 old = vm_page_astate_load(m); 3890 do { 3891 if (old.queue == PQ_NONE) { 3892 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3893 ("%s: page %p has unexpected queue state", 3894 __func__, m)); 3895 break; 3896 } 3897 new = old; 3898 new.flags |= PGA_DEQUEUE; 3899 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3900 } 3901 3902 /* 3903 * vm_page_dequeue: 3904 * 3905 * Remove the page from whichever page queue it's in, if any, before 3906 * returning. 3907 */ 3908 void 3909 vm_page_dequeue(vm_page_t m) 3910 { 3911 vm_page_astate_t new, old; 3912 3913 old = vm_page_astate_load(m); 3914 do { 3915 if (old.queue == PQ_NONE) { 3916 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3917 ("%s: page %p has unexpected queue state", 3918 __func__, m)); 3919 break; 3920 } 3921 new = old; 3922 new.flags &= ~PGA_QUEUE_OP_MASK; 3923 new.queue = PQ_NONE; 3924 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 3925 3926 } 3927 3928 /* 3929 * Schedule the given page for insertion into the specified page queue. 3930 * Physical insertion of the page may be deferred indefinitely. 3931 */ 3932 static void 3933 vm_page_enqueue(vm_page_t m, uint8_t queue) 3934 { 3935 3936 KASSERT(m->a.queue == PQ_NONE && 3937 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 3938 ("%s: page %p is already enqueued", __func__, m)); 3939 KASSERT(m->ref_count > 0, 3940 ("%s: page %p does not carry any references", __func__, m)); 3941 3942 m->a.queue = queue; 3943 if ((m->a.flags & PGA_REQUEUE) == 0) 3944 vm_page_aflag_set(m, PGA_REQUEUE); 3945 vm_page_pqbatch_submit(m, queue); 3946 } 3947 3948 /* 3949 * vm_page_free_prep: 3950 * 3951 * Prepares the given page to be put on the free list, 3952 * disassociating it from any VM object. The caller may return 3953 * the page to the free list only if this function returns true. 3954 * 3955 * The object, if it exists, must be locked, and then the page must 3956 * be xbusy. Otherwise the page must be not busied. A managed 3957 * page must be unmapped. 3958 */ 3959 static bool 3960 vm_page_free_prep(vm_page_t m) 3961 { 3962 3963 /* 3964 * Synchronize with threads that have dropped a reference to this 3965 * page. 3966 */ 3967 atomic_thread_fence_acq(); 3968 3969 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3970 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3971 uint64_t *p; 3972 int i; 3973 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3974 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3975 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3976 m, i, (uintmax_t)*p)); 3977 } 3978 #endif 3979 if ((m->oflags & VPO_UNMANAGED) == 0) { 3980 KASSERT(!pmap_page_is_mapped(m), 3981 ("vm_page_free_prep: freeing mapped page %p", m)); 3982 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3983 ("vm_page_free_prep: mapping flags set in page %p", m)); 3984 } else { 3985 KASSERT(m->a.queue == PQ_NONE, 3986 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3987 } 3988 VM_CNT_INC(v_tfree); 3989 3990 if (m->object != NULL) { 3991 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 3992 ((m->object->flags & OBJ_UNMANAGED) != 0), 3993 ("vm_page_free_prep: managed flag mismatch for page %p", 3994 m)); 3995 vm_page_assert_xbusied(m); 3996 3997 /* 3998 * The object reference can be released without an atomic 3999 * operation. 4000 */ 4001 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 4002 m->ref_count == VPRC_OBJREF, 4003 ("vm_page_free_prep: page %p has unexpected ref_count %u", 4004 m, m->ref_count)); 4005 vm_page_object_remove(m); 4006 m->ref_count -= VPRC_OBJREF; 4007 } else 4008 vm_page_assert_unbusied(m); 4009 4010 vm_page_busy_free(m); 4011 4012 /* 4013 * If fictitious remove object association and 4014 * return. 4015 */ 4016 if ((m->flags & PG_FICTITIOUS) != 0) { 4017 KASSERT(m->ref_count == 1, 4018 ("fictitious page %p is referenced", m)); 4019 KASSERT(m->a.queue == PQ_NONE, 4020 ("fictitious page %p is queued", m)); 4021 return (false); 4022 } 4023 4024 /* 4025 * Pages need not be dequeued before they are returned to the physical 4026 * memory allocator, but they must at least be marked for a deferred 4027 * dequeue. 4028 */ 4029 if ((m->oflags & VPO_UNMANAGED) == 0) 4030 vm_page_dequeue_deferred(m); 4031 4032 m->valid = 0; 4033 vm_page_undirty(m); 4034 4035 if (m->ref_count != 0) 4036 panic("vm_page_free_prep: page %p has references", m); 4037 4038 /* 4039 * Restore the default memory attribute to the page. 4040 */ 4041 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 4042 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 4043 4044 #if VM_NRESERVLEVEL > 0 4045 /* 4046 * Determine whether the page belongs to a reservation. If the page was 4047 * allocated from a per-CPU cache, it cannot belong to a reservation, so 4048 * as an optimization, we avoid the check in that case. 4049 */ 4050 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 4051 return (false); 4052 #endif 4053 4054 return (true); 4055 } 4056 4057 /* 4058 * vm_page_free_toq: 4059 * 4060 * Returns the given page to the free list, disassociating it 4061 * from any VM object. 4062 * 4063 * The object must be locked. The page must be exclusively busied if it 4064 * belongs to an object. 4065 */ 4066 static void 4067 vm_page_free_toq(vm_page_t m) 4068 { 4069 struct vm_domain *vmd; 4070 uma_zone_t zone; 4071 4072 if (!vm_page_free_prep(m)) 4073 return; 4074 4075 vmd = vm_pagequeue_domain(m); 4076 zone = vmd->vmd_pgcache[m->pool].zone; 4077 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4078 uma_zfree(zone, m); 4079 return; 4080 } 4081 vm_domain_free_lock(vmd); 4082 vm_phys_free_pages(m, 0); 4083 vm_domain_free_unlock(vmd); 4084 vm_domain_freecnt_inc(vmd, 1); 4085 } 4086 4087 /* 4088 * vm_page_free_pages_toq: 4089 * 4090 * Returns a list of pages to the free list, disassociating it 4091 * from any VM object. In other words, this is equivalent to 4092 * calling vm_page_free_toq() for each page of a list of VM objects. 4093 */ 4094 void 4095 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4096 { 4097 vm_page_t m; 4098 int count; 4099 4100 if (SLIST_EMPTY(free)) 4101 return; 4102 4103 count = 0; 4104 while ((m = SLIST_FIRST(free)) != NULL) { 4105 count++; 4106 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4107 vm_page_free_toq(m); 4108 } 4109 4110 if (update_wire_count) 4111 vm_wire_sub(count); 4112 } 4113 4114 /* 4115 * Mark this page as wired down. For managed pages, this prevents reclamation 4116 * by the page daemon, or when the containing object, if any, is destroyed. 4117 */ 4118 void 4119 vm_page_wire(vm_page_t m) 4120 { 4121 u_int old; 4122 4123 #ifdef INVARIANTS 4124 if (m->object != NULL && !vm_page_busied(m) && 4125 !vm_object_busied(m->object)) 4126 VM_OBJECT_ASSERT_LOCKED(m->object); 4127 #endif 4128 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4129 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4130 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4131 4132 old = atomic_fetchadd_int(&m->ref_count, 1); 4133 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4134 ("vm_page_wire: counter overflow for page %p", m)); 4135 if (VPRC_WIRE_COUNT(old) == 0) { 4136 if ((m->oflags & VPO_UNMANAGED) == 0) 4137 vm_page_aflag_set(m, PGA_DEQUEUE); 4138 vm_wire_add(1); 4139 } 4140 } 4141 4142 /* 4143 * Attempt to wire a mapped page following a pmap lookup of that page. 4144 * This may fail if a thread is concurrently tearing down mappings of the page. 4145 * The transient failure is acceptable because it translates to the 4146 * failure of the caller pmap_extract_and_hold(), which should be then 4147 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4148 */ 4149 bool 4150 vm_page_wire_mapped(vm_page_t m) 4151 { 4152 u_int old; 4153 4154 old = m->ref_count; 4155 do { 4156 KASSERT(old > 0, 4157 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4158 if ((old & VPRC_BLOCKED) != 0) 4159 return (false); 4160 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4161 4162 if (VPRC_WIRE_COUNT(old) == 0) { 4163 if ((m->oflags & VPO_UNMANAGED) == 0) 4164 vm_page_aflag_set(m, PGA_DEQUEUE); 4165 vm_wire_add(1); 4166 } 4167 return (true); 4168 } 4169 4170 /* 4171 * Release a wiring reference to a managed page. If the page still belongs to 4172 * an object, update its position in the page queues to reflect the reference. 4173 * If the wiring was the last reference to the page, free the page. 4174 */ 4175 static void 4176 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4177 { 4178 u_int old; 4179 4180 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4181 ("%s: page %p is unmanaged", __func__, m)); 4182 4183 /* 4184 * Update LRU state before releasing the wiring reference. 4185 * Use a release store when updating the reference count to 4186 * synchronize with vm_page_free_prep(). 4187 */ 4188 old = m->ref_count; 4189 do { 4190 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4191 ("vm_page_unwire: wire count underflow for page %p", m)); 4192 4193 if (old > VPRC_OBJREF + 1) { 4194 /* 4195 * The page has at least one other wiring reference. An 4196 * earlier iteration of this loop may have called 4197 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4198 * re-set it if necessary. 4199 */ 4200 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4201 vm_page_aflag_set(m, PGA_DEQUEUE); 4202 } else if (old == VPRC_OBJREF + 1) { 4203 /* 4204 * This is the last wiring. Clear PGA_DEQUEUE and 4205 * update the page's queue state to reflect the 4206 * reference. If the page does not belong to an object 4207 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4208 * clear leftover queue state. 4209 */ 4210 vm_page_release_toq(m, nqueue, noreuse); 4211 } else if (old == 1) { 4212 vm_page_aflag_clear(m, PGA_DEQUEUE); 4213 } 4214 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4215 4216 if (VPRC_WIRE_COUNT(old) == 1) { 4217 vm_wire_sub(1); 4218 if (old == 1) 4219 vm_page_free(m); 4220 } 4221 } 4222 4223 /* 4224 * Release one wiring of the specified page, potentially allowing it to be 4225 * paged out. 4226 * 4227 * Only managed pages belonging to an object can be paged out. If the number 4228 * of wirings transitions to zero and the page is eligible for page out, then 4229 * the page is added to the specified paging queue. If the released wiring 4230 * represented the last reference to the page, the page is freed. 4231 */ 4232 void 4233 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4234 { 4235 4236 KASSERT(nqueue < PQ_COUNT, 4237 ("vm_page_unwire: invalid queue %u request for page %p", 4238 nqueue, m)); 4239 4240 if ((m->oflags & VPO_UNMANAGED) != 0) { 4241 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4242 vm_page_free(m); 4243 return; 4244 } 4245 vm_page_unwire_managed(m, nqueue, false); 4246 } 4247 4248 /* 4249 * Unwire a page without (re-)inserting it into a page queue. It is up 4250 * to the caller to enqueue, requeue, or free the page as appropriate. 4251 * In most cases involving managed pages, vm_page_unwire() should be used 4252 * instead. 4253 */ 4254 bool 4255 vm_page_unwire_noq(vm_page_t m) 4256 { 4257 u_int old; 4258 4259 old = vm_page_drop(m, 1); 4260 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4261 ("%s: counter underflow for page %p", __func__, m)); 4262 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4263 ("%s: missing ref on fictitious page %p", __func__, m)); 4264 4265 if (VPRC_WIRE_COUNT(old) > 1) 4266 return (false); 4267 if ((m->oflags & VPO_UNMANAGED) == 0) 4268 vm_page_aflag_clear(m, PGA_DEQUEUE); 4269 vm_wire_sub(1); 4270 return (true); 4271 } 4272 4273 /* 4274 * Ensure that the page ends up in the specified page queue. If the page is 4275 * active or being moved to the active queue, ensure that its act_count is 4276 * at least ACT_INIT but do not otherwise mess with it. 4277 */ 4278 static __always_inline void 4279 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4280 { 4281 vm_page_astate_t old, new; 4282 4283 KASSERT(m->ref_count > 0, 4284 ("%s: page %p does not carry any references", __func__, m)); 4285 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4286 ("%s: invalid flags %x", __func__, nflag)); 4287 4288 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4289 return; 4290 4291 old = vm_page_astate_load(m); 4292 do { 4293 if ((old.flags & PGA_DEQUEUE) != 0) 4294 break; 4295 new = old; 4296 new.flags &= ~PGA_QUEUE_OP_MASK; 4297 if (nqueue == PQ_ACTIVE) 4298 new.act_count = max(old.act_count, ACT_INIT); 4299 if (old.queue == nqueue) { 4300 /* 4301 * There is no need to requeue pages already in the 4302 * active queue. 4303 */ 4304 if (nqueue != PQ_ACTIVE || 4305 (old.flags & PGA_ENQUEUED) == 0) 4306 new.flags |= nflag; 4307 } else { 4308 new.flags |= nflag; 4309 new.queue = nqueue; 4310 } 4311 } while (!vm_page_pqstate_commit(m, &old, new)); 4312 } 4313 4314 /* 4315 * Put the specified page on the active list (if appropriate). 4316 */ 4317 void 4318 vm_page_activate(vm_page_t m) 4319 { 4320 4321 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4322 } 4323 4324 /* 4325 * Move the specified page to the tail of the inactive queue, or requeue 4326 * the page if it is already in the inactive queue. 4327 */ 4328 void 4329 vm_page_deactivate(vm_page_t m) 4330 { 4331 4332 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4333 } 4334 4335 void 4336 vm_page_deactivate_noreuse(vm_page_t m) 4337 { 4338 4339 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4340 } 4341 4342 /* 4343 * Put a page in the laundry, or requeue it if it is already there. 4344 */ 4345 void 4346 vm_page_launder(vm_page_t m) 4347 { 4348 4349 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4350 } 4351 4352 /* 4353 * Put a page in the PQ_UNSWAPPABLE holding queue. 4354 */ 4355 void 4356 vm_page_unswappable(vm_page_t m) 4357 { 4358 4359 VM_OBJECT_ASSERT_LOCKED(m->object); 4360 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4361 ("page %p already unswappable", m)); 4362 4363 vm_page_dequeue(m); 4364 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4365 } 4366 4367 /* 4368 * Release a page back to the page queues in preparation for unwiring. 4369 */ 4370 static void 4371 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4372 { 4373 vm_page_astate_t old, new; 4374 uint16_t nflag; 4375 4376 /* 4377 * Use a check of the valid bits to determine whether we should 4378 * accelerate reclamation of the page. The object lock might not be 4379 * held here, in which case the check is racy. At worst we will either 4380 * accelerate reclamation of a valid page and violate LRU, or 4381 * unnecessarily defer reclamation of an invalid page. 4382 * 4383 * If we were asked to not cache the page, place it near the head of the 4384 * inactive queue so that is reclaimed sooner. 4385 */ 4386 if (noreuse || vm_page_none_valid(m)) { 4387 nqueue = PQ_INACTIVE; 4388 nflag = PGA_REQUEUE_HEAD; 4389 } else { 4390 nflag = PGA_REQUEUE; 4391 } 4392 4393 old = vm_page_astate_load(m); 4394 do { 4395 new = old; 4396 4397 /* 4398 * If the page is already in the active queue and we are not 4399 * trying to accelerate reclamation, simply mark it as 4400 * referenced and avoid any queue operations. 4401 */ 4402 new.flags &= ~PGA_QUEUE_OP_MASK; 4403 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4404 (old.flags & PGA_ENQUEUED) != 0) 4405 new.flags |= PGA_REFERENCED; 4406 else { 4407 new.flags |= nflag; 4408 new.queue = nqueue; 4409 } 4410 } while (!vm_page_pqstate_commit(m, &old, new)); 4411 } 4412 4413 /* 4414 * Unwire a page and either attempt to free it or re-add it to the page queues. 4415 */ 4416 void 4417 vm_page_release(vm_page_t m, int flags) 4418 { 4419 vm_object_t object; 4420 4421 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4422 ("vm_page_release: page %p is unmanaged", m)); 4423 4424 if ((flags & VPR_TRYFREE) != 0) { 4425 for (;;) { 4426 object = atomic_load_ptr(&m->object); 4427 if (object == NULL) 4428 break; 4429 /* Depends on type-stability. */ 4430 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4431 break; 4432 if (object == m->object) { 4433 vm_page_release_locked(m, flags); 4434 VM_OBJECT_WUNLOCK(object); 4435 return; 4436 } 4437 VM_OBJECT_WUNLOCK(object); 4438 } 4439 } 4440 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4441 } 4442 4443 /* See vm_page_release(). */ 4444 void 4445 vm_page_release_locked(vm_page_t m, int flags) 4446 { 4447 4448 VM_OBJECT_ASSERT_WLOCKED(m->object); 4449 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4450 ("vm_page_release_locked: page %p is unmanaged", m)); 4451 4452 if (vm_page_unwire_noq(m)) { 4453 if ((flags & VPR_TRYFREE) != 0 && 4454 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4455 m->dirty == 0 && vm_page_tryxbusy(m)) { 4456 /* 4457 * An unlocked lookup may have wired the page before the 4458 * busy lock was acquired, in which case the page must 4459 * not be freed. 4460 */ 4461 if (__predict_true(!vm_page_wired(m))) { 4462 vm_page_free(m); 4463 return; 4464 } 4465 vm_page_xunbusy(m); 4466 } else { 4467 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4468 } 4469 } 4470 } 4471 4472 static bool 4473 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4474 { 4475 u_int old; 4476 4477 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4478 ("vm_page_try_blocked_op: page %p has no object", m)); 4479 KASSERT(vm_page_busied(m), 4480 ("vm_page_try_blocked_op: page %p is not busy", m)); 4481 VM_OBJECT_ASSERT_LOCKED(m->object); 4482 4483 old = m->ref_count; 4484 do { 4485 KASSERT(old != 0, 4486 ("vm_page_try_blocked_op: page %p has no references", m)); 4487 if (VPRC_WIRE_COUNT(old) != 0) 4488 return (false); 4489 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4490 4491 (op)(m); 4492 4493 /* 4494 * If the object is read-locked, new wirings may be created via an 4495 * object lookup. 4496 */ 4497 old = vm_page_drop(m, VPRC_BLOCKED); 4498 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4499 old == (VPRC_BLOCKED | VPRC_OBJREF), 4500 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4501 old, m)); 4502 return (true); 4503 } 4504 4505 /* 4506 * Atomically check for wirings and remove all mappings of the page. 4507 */ 4508 bool 4509 vm_page_try_remove_all(vm_page_t m) 4510 { 4511 4512 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4513 } 4514 4515 /* 4516 * Atomically check for wirings and remove all writeable mappings of the page. 4517 */ 4518 bool 4519 vm_page_try_remove_write(vm_page_t m) 4520 { 4521 4522 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4523 } 4524 4525 /* 4526 * vm_page_advise 4527 * 4528 * Apply the specified advice to the given page. 4529 */ 4530 void 4531 vm_page_advise(vm_page_t m, int advice) 4532 { 4533 4534 VM_OBJECT_ASSERT_WLOCKED(m->object); 4535 vm_page_assert_xbusied(m); 4536 4537 if (advice == MADV_FREE) 4538 /* 4539 * Mark the page clean. This will allow the page to be freed 4540 * without first paging it out. MADV_FREE pages are often 4541 * quickly reused by malloc(3), so we do not do anything that 4542 * would result in a page fault on a later access. 4543 */ 4544 vm_page_undirty(m); 4545 else if (advice != MADV_DONTNEED) { 4546 if (advice == MADV_WILLNEED) 4547 vm_page_activate(m); 4548 return; 4549 } 4550 4551 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4552 vm_page_dirty(m); 4553 4554 /* 4555 * Clear any references to the page. Otherwise, the page daemon will 4556 * immediately reactivate the page. 4557 */ 4558 vm_page_aflag_clear(m, PGA_REFERENCED); 4559 4560 /* 4561 * Place clean pages near the head of the inactive queue rather than 4562 * the tail, thus defeating the queue's LRU operation and ensuring that 4563 * the page will be reused quickly. Dirty pages not already in the 4564 * laundry are moved there. 4565 */ 4566 if (m->dirty == 0) 4567 vm_page_deactivate_noreuse(m); 4568 else if (!vm_page_in_laundry(m)) 4569 vm_page_launder(m); 4570 } 4571 4572 /* 4573 * vm_page_grab_release 4574 * 4575 * Helper routine for grab functions to release busy on return. 4576 */ 4577 static inline void 4578 vm_page_grab_release(vm_page_t m, int allocflags) 4579 { 4580 4581 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4582 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4583 vm_page_sunbusy(m); 4584 else 4585 vm_page_xunbusy(m); 4586 } 4587 } 4588 4589 /* 4590 * vm_page_grab_sleep 4591 * 4592 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4593 * if the caller should retry and false otherwise. 4594 * 4595 * If the object is locked on entry the object will be unlocked with 4596 * false returns and still locked but possibly having been dropped 4597 * with true returns. 4598 */ 4599 static bool 4600 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4601 const char *wmesg, int allocflags, bool locked) 4602 { 4603 4604 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4605 return (false); 4606 4607 /* 4608 * Reference the page before unlocking and sleeping so that 4609 * the page daemon is less likely to reclaim it. 4610 */ 4611 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4612 vm_page_reference(m); 4613 4614 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4615 locked) 4616 VM_OBJECT_WLOCK(object); 4617 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4618 return (false); 4619 4620 return (true); 4621 } 4622 4623 /* 4624 * Assert that the grab flags are valid. 4625 */ 4626 static inline void 4627 vm_page_grab_check(int allocflags) 4628 { 4629 4630 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4631 (allocflags & VM_ALLOC_WIRED) != 0, 4632 ("vm_page_grab*: the pages must be busied or wired")); 4633 4634 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4635 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4636 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4637 } 4638 4639 /* 4640 * Calculate the page allocation flags for grab. 4641 */ 4642 static inline int 4643 vm_page_grab_pflags(int allocflags) 4644 { 4645 int pflags; 4646 4647 pflags = allocflags & 4648 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4649 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY); 4650 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4651 pflags |= VM_ALLOC_WAITFAIL; 4652 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4653 pflags |= VM_ALLOC_SBUSY; 4654 4655 return (pflags); 4656 } 4657 4658 /* 4659 * Grab a page, waiting until we are waken up due to the page 4660 * changing state. We keep on waiting, if the page continues 4661 * to be in the object. If the page doesn't exist, first allocate it 4662 * and then conditionally zero it. 4663 * 4664 * This routine may sleep. 4665 * 4666 * The object must be locked on entry. The lock will, however, be released 4667 * and reacquired if the routine sleeps. 4668 */ 4669 vm_page_t 4670 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4671 { 4672 vm_page_t m; 4673 4674 VM_OBJECT_ASSERT_WLOCKED(object); 4675 vm_page_grab_check(allocflags); 4676 4677 retrylookup: 4678 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4679 if (!vm_page_tryacquire(m, allocflags)) { 4680 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4681 allocflags, true)) 4682 goto retrylookup; 4683 return (NULL); 4684 } 4685 goto out; 4686 } 4687 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4688 return (NULL); 4689 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); 4690 if (m == NULL) { 4691 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4692 return (NULL); 4693 goto retrylookup; 4694 } 4695 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4696 pmap_zero_page(m); 4697 4698 out: 4699 vm_page_grab_release(m, allocflags); 4700 4701 return (m); 4702 } 4703 4704 /* 4705 * Locklessly attempt to acquire a page given a (object, pindex) tuple 4706 * and an optional previous page to avoid the radix lookup. The resulting 4707 * page will be validated against the identity tuple and busied or wired 4708 * as requested. A NULL *mp return guarantees that the page was not in 4709 * radix at the time of the call but callers must perform higher level 4710 * synchronization or retry the operation under a lock if they require 4711 * an atomic answer. This is the only lock free validation routine, 4712 * other routines can depend on the resulting page state. 4713 * 4714 * The return value indicates whether the operation failed due to caller 4715 * flags. The return is tri-state with mp: 4716 * 4717 * (true, *mp != NULL) - The operation was successful. 4718 * (true, *mp == NULL) - The page was not found in tree. 4719 * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition. 4720 */ 4721 static bool 4722 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, 4723 vm_page_t prev, vm_page_t *mp, int allocflags) 4724 { 4725 vm_page_t m; 4726 4727 vm_page_grab_check(allocflags); 4728 MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev)); 4729 4730 *mp = NULL; 4731 for (;;) { 4732 /* 4733 * We may see a false NULL here because the previous page 4734 * has been removed or just inserted and the list is loaded 4735 * without barriers. Switch to radix to verify. 4736 */ 4737 if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL || 4738 QMD_IS_TRASHED(m) || m->pindex != pindex || 4739 atomic_load_ptr(&m->object) != object) { 4740 prev = NULL; 4741 /* 4742 * This guarantees the result is instantaneously 4743 * correct. 4744 */ 4745 m = vm_radix_lookup_unlocked(&object->rtree, pindex); 4746 } 4747 if (m == NULL) 4748 return (true); 4749 if (vm_page_trybusy(m, allocflags)) { 4750 if (m->object == object && m->pindex == pindex) 4751 break; 4752 /* relookup. */ 4753 vm_page_busy_release(m); 4754 cpu_spinwait(); 4755 continue; 4756 } 4757 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4758 allocflags, false)) 4759 return (false); 4760 } 4761 if ((allocflags & VM_ALLOC_WIRED) != 0) 4762 vm_page_wire(m); 4763 vm_page_grab_release(m, allocflags); 4764 *mp = m; 4765 return (true); 4766 } 4767 4768 /* 4769 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4770 * is not set. 4771 */ 4772 vm_page_t 4773 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4774 { 4775 vm_page_t m; 4776 4777 vm_page_grab_check(allocflags); 4778 4779 if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags)) 4780 return (NULL); 4781 if (m != NULL) 4782 return (m); 4783 4784 /* 4785 * The radix lockless lookup should never return a false negative 4786 * errors. If the user specifies NOCREAT they are guaranteed there 4787 * was no page present at the instant of the call. A NOCREAT caller 4788 * must handle create races gracefully. 4789 */ 4790 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4791 return (NULL); 4792 4793 VM_OBJECT_WLOCK(object); 4794 m = vm_page_grab(object, pindex, allocflags); 4795 VM_OBJECT_WUNLOCK(object); 4796 4797 return (m); 4798 } 4799 4800 /* 4801 * Grab a page and make it valid, paging in if necessary. Pages missing from 4802 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4803 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4804 * in simultaneously. Additional pages will be left on a paging queue but 4805 * will neither be wired nor busy regardless of allocflags. 4806 */ 4807 int 4808 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4809 { 4810 vm_page_t m; 4811 vm_page_t ma[VM_INITIAL_PAGEIN]; 4812 int after, i, pflags, rv; 4813 4814 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4815 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4816 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4817 KASSERT((allocflags & 4818 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4819 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4820 VM_OBJECT_ASSERT_WLOCKED(object); 4821 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4822 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4823 pflags |= VM_ALLOC_WAITFAIL; 4824 4825 retrylookup: 4826 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4827 /* 4828 * If the page is fully valid it can only become invalid 4829 * with the object lock held. If it is not valid it can 4830 * become valid with the busy lock held. Therefore, we 4831 * may unnecessarily lock the exclusive busy here if we 4832 * race with I/O completion not using the object lock. 4833 * However, we will not end up with an invalid page and a 4834 * shared lock. 4835 */ 4836 if (!vm_page_trybusy(m, 4837 vm_page_all_valid(m) ? allocflags : 0)) { 4838 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4839 allocflags, true); 4840 goto retrylookup; 4841 } 4842 if (vm_page_all_valid(m)) 4843 goto out; 4844 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4845 vm_page_busy_release(m); 4846 *mp = NULL; 4847 return (VM_PAGER_FAIL); 4848 } 4849 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4850 *mp = NULL; 4851 return (VM_PAGER_FAIL); 4852 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { 4853 if (!vm_pager_can_alloc_page(object, pindex)) { 4854 *mp = NULL; 4855 return (VM_PAGER_AGAIN); 4856 } 4857 goto retrylookup; 4858 } 4859 4860 vm_page_assert_xbusied(m); 4861 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4862 after = MIN(after, VM_INITIAL_PAGEIN); 4863 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4864 after = MAX(after, 1); 4865 ma[0] = m; 4866 for (i = 1; i < after; i++) { 4867 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 4868 if (vm_page_any_valid(ma[i]) || 4869 !vm_page_tryxbusy(ma[i])) 4870 break; 4871 } else { 4872 ma[i] = vm_page_alloc(object, m->pindex + i, 4873 VM_ALLOC_NORMAL); 4874 if (ma[i] == NULL) 4875 break; 4876 } 4877 } 4878 after = i; 4879 vm_object_pip_add(object, after); 4880 VM_OBJECT_WUNLOCK(object); 4881 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 4882 VM_OBJECT_WLOCK(object); 4883 vm_object_pip_wakeupn(object, after); 4884 /* Pager may have replaced a page. */ 4885 m = ma[0]; 4886 if (rv != VM_PAGER_OK) { 4887 for (i = 0; i < after; i++) { 4888 if (!vm_page_wired(ma[i])) 4889 vm_page_free(ma[i]); 4890 else 4891 vm_page_xunbusy(ma[i]); 4892 } 4893 *mp = NULL; 4894 return (rv); 4895 } 4896 for (i = 1; i < after; i++) 4897 vm_page_readahead_finish(ma[i]); 4898 MPASS(vm_page_all_valid(m)); 4899 } else { 4900 vm_page_zero_invalid(m, TRUE); 4901 } 4902 out: 4903 if ((allocflags & VM_ALLOC_WIRED) != 0) 4904 vm_page_wire(m); 4905 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 4906 vm_page_busy_downgrade(m); 4907 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 4908 vm_page_busy_release(m); 4909 *mp = m; 4910 return (VM_PAGER_OK); 4911 } 4912 4913 /* 4914 * Locklessly grab a valid page. If the page is not valid or not yet 4915 * allocated this will fall back to the object lock method. 4916 */ 4917 int 4918 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 4919 vm_pindex_t pindex, int allocflags) 4920 { 4921 vm_page_t m; 4922 int flags; 4923 int error; 4924 4925 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4926 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4927 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 4928 "mismatch")); 4929 KASSERT((allocflags & 4930 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4931 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 4932 4933 /* 4934 * Attempt a lockless lookup and busy. We need at least an sbusy 4935 * before we can inspect the valid field and return a wired page. 4936 */ 4937 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 4938 if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags)) 4939 return (VM_PAGER_FAIL); 4940 if ((m = *mp) != NULL) { 4941 if (vm_page_all_valid(m)) { 4942 if ((allocflags & VM_ALLOC_WIRED) != 0) 4943 vm_page_wire(m); 4944 vm_page_grab_release(m, allocflags); 4945 return (VM_PAGER_OK); 4946 } 4947 vm_page_busy_release(m); 4948 } 4949 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4950 *mp = NULL; 4951 return (VM_PAGER_FAIL); 4952 } 4953 VM_OBJECT_WLOCK(object); 4954 error = vm_page_grab_valid(mp, object, pindex, allocflags); 4955 VM_OBJECT_WUNLOCK(object); 4956 4957 return (error); 4958 } 4959 4960 /* 4961 * Return the specified range of pages from the given object. For each 4962 * page offset within the range, if a page already exists within the object 4963 * at that offset and it is busy, then wait for it to change state. If, 4964 * instead, the page doesn't exist, then allocate it. 4965 * 4966 * The caller must always specify an allocation class. 4967 * 4968 * allocation classes: 4969 * VM_ALLOC_NORMAL normal process request 4970 * VM_ALLOC_SYSTEM system *really* needs the pages 4971 * 4972 * The caller must always specify that the pages are to be busied and/or 4973 * wired. 4974 * 4975 * optional allocation flags: 4976 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4977 * VM_ALLOC_NOBUSY do not exclusive busy the page 4978 * VM_ALLOC_NOWAIT do not sleep 4979 * VM_ALLOC_SBUSY set page to sbusy state 4980 * VM_ALLOC_WIRED wire the pages 4981 * VM_ALLOC_ZERO zero and validate any invalid pages 4982 * 4983 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4984 * may return a partial prefix of the requested range. 4985 */ 4986 int 4987 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4988 vm_page_t *ma, int count) 4989 { 4990 vm_page_t m, mpred; 4991 int pflags; 4992 int i; 4993 4994 VM_OBJECT_ASSERT_WLOCKED(object); 4995 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4996 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4997 KASSERT(count > 0, 4998 ("vm_page_grab_pages: invalid page count %d", count)); 4999 vm_page_grab_check(allocflags); 5000 5001 pflags = vm_page_grab_pflags(allocflags); 5002 i = 0; 5003 retrylookup: 5004 m = vm_radix_lookup_le(&object->rtree, pindex + i); 5005 if (m == NULL || m->pindex != pindex + i) { 5006 mpred = m; 5007 m = NULL; 5008 } else 5009 mpred = TAILQ_PREV(m, pglist, listq); 5010 for (; i < count; i++) { 5011 if (m != NULL) { 5012 if (!vm_page_tryacquire(m, allocflags)) { 5013 if (vm_page_grab_sleep(object, m, pindex + i, 5014 "grbmaw", allocflags, true)) 5015 goto retrylookup; 5016 break; 5017 } 5018 } else { 5019 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 5020 break; 5021 m = vm_page_alloc_after(object, pindex + i, 5022 pflags | VM_ALLOC_COUNT(count - i), mpred); 5023 if (m == NULL) { 5024 if ((allocflags & (VM_ALLOC_NOWAIT | 5025 VM_ALLOC_WAITFAIL)) != 0) 5026 break; 5027 goto retrylookup; 5028 } 5029 } 5030 if (vm_page_none_valid(m) && 5031 (allocflags & VM_ALLOC_ZERO) != 0) { 5032 if ((m->flags & PG_ZERO) == 0) 5033 pmap_zero_page(m); 5034 vm_page_valid(m); 5035 } 5036 vm_page_grab_release(m, allocflags); 5037 ma[i] = mpred = m; 5038 m = vm_page_next(m); 5039 } 5040 return (i); 5041 } 5042 5043 /* 5044 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 5045 * and will fall back to the locked variant to handle allocation. 5046 */ 5047 int 5048 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 5049 int allocflags, vm_page_t *ma, int count) 5050 { 5051 vm_page_t m, pred; 5052 int flags; 5053 int i; 5054 5055 KASSERT(count > 0, 5056 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 5057 vm_page_grab_check(allocflags); 5058 5059 /* 5060 * Modify flags for lockless acquire to hold the page until we 5061 * set it valid if necessary. 5062 */ 5063 flags = allocflags & ~VM_ALLOC_NOBUSY; 5064 pred = NULL; 5065 for (i = 0; i < count; i++, pindex++) { 5066 if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags)) 5067 return (i); 5068 if (m == NULL) 5069 break; 5070 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 5071 if ((m->flags & PG_ZERO) == 0) 5072 pmap_zero_page(m); 5073 vm_page_valid(m); 5074 } 5075 /* m will still be wired or busy according to flags. */ 5076 vm_page_grab_release(m, allocflags); 5077 pred = ma[i] = m; 5078 } 5079 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 5080 return (i); 5081 count -= i; 5082 VM_OBJECT_WLOCK(object); 5083 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 5084 VM_OBJECT_WUNLOCK(object); 5085 5086 return (i); 5087 } 5088 5089 /* 5090 * Mapping function for valid or dirty bits in a page. 5091 * 5092 * Inputs are required to range within a page. 5093 */ 5094 vm_page_bits_t 5095 vm_page_bits(int base, int size) 5096 { 5097 int first_bit; 5098 int last_bit; 5099 5100 KASSERT( 5101 base + size <= PAGE_SIZE, 5102 ("vm_page_bits: illegal base/size %d/%d", base, size) 5103 ); 5104 5105 if (size == 0) /* handle degenerate case */ 5106 return (0); 5107 5108 first_bit = base >> DEV_BSHIFT; 5109 last_bit = (base + size - 1) >> DEV_BSHIFT; 5110 5111 return (((vm_page_bits_t)2 << last_bit) - 5112 ((vm_page_bits_t)1 << first_bit)); 5113 } 5114 5115 void 5116 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5117 { 5118 5119 #if PAGE_SIZE == 32768 5120 atomic_set_64((uint64_t *)bits, set); 5121 #elif PAGE_SIZE == 16384 5122 atomic_set_32((uint32_t *)bits, set); 5123 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5124 atomic_set_16((uint16_t *)bits, set); 5125 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5126 atomic_set_8((uint8_t *)bits, set); 5127 #else /* PAGE_SIZE <= 8192 */ 5128 uintptr_t addr; 5129 int shift; 5130 5131 addr = (uintptr_t)bits; 5132 /* 5133 * Use a trick to perform a 32-bit atomic on the 5134 * containing aligned word, to not depend on the existence 5135 * of atomic_{set, clear}_{8, 16}. 5136 */ 5137 shift = addr & (sizeof(uint32_t) - 1); 5138 #if BYTE_ORDER == BIG_ENDIAN 5139 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5140 #else 5141 shift *= NBBY; 5142 #endif 5143 addr &= ~(sizeof(uint32_t) - 1); 5144 atomic_set_32((uint32_t *)addr, set << shift); 5145 #endif /* PAGE_SIZE */ 5146 } 5147 5148 static inline void 5149 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5150 { 5151 5152 #if PAGE_SIZE == 32768 5153 atomic_clear_64((uint64_t *)bits, clear); 5154 #elif PAGE_SIZE == 16384 5155 atomic_clear_32((uint32_t *)bits, clear); 5156 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5157 atomic_clear_16((uint16_t *)bits, clear); 5158 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5159 atomic_clear_8((uint8_t *)bits, clear); 5160 #else /* PAGE_SIZE <= 8192 */ 5161 uintptr_t addr; 5162 int shift; 5163 5164 addr = (uintptr_t)bits; 5165 /* 5166 * Use a trick to perform a 32-bit atomic on the 5167 * containing aligned word, to not depend on the existence 5168 * of atomic_{set, clear}_{8, 16}. 5169 */ 5170 shift = addr & (sizeof(uint32_t) - 1); 5171 #if BYTE_ORDER == BIG_ENDIAN 5172 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5173 #else 5174 shift *= NBBY; 5175 #endif 5176 addr &= ~(sizeof(uint32_t) - 1); 5177 atomic_clear_32((uint32_t *)addr, clear << shift); 5178 #endif /* PAGE_SIZE */ 5179 } 5180 5181 static inline vm_page_bits_t 5182 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5183 { 5184 #if PAGE_SIZE == 32768 5185 uint64_t old; 5186 5187 old = *bits; 5188 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5189 return (old); 5190 #elif PAGE_SIZE == 16384 5191 uint32_t old; 5192 5193 old = *bits; 5194 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5195 return (old); 5196 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5197 uint16_t old; 5198 5199 old = *bits; 5200 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5201 return (old); 5202 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5203 uint8_t old; 5204 5205 old = *bits; 5206 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5207 return (old); 5208 #else /* PAGE_SIZE <= 4096*/ 5209 uintptr_t addr; 5210 uint32_t old, new, mask; 5211 int shift; 5212 5213 addr = (uintptr_t)bits; 5214 /* 5215 * Use a trick to perform a 32-bit atomic on the 5216 * containing aligned word, to not depend on the existence 5217 * of atomic_{set, swap, clear}_{8, 16}. 5218 */ 5219 shift = addr & (sizeof(uint32_t) - 1); 5220 #if BYTE_ORDER == BIG_ENDIAN 5221 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5222 #else 5223 shift *= NBBY; 5224 #endif 5225 addr &= ~(sizeof(uint32_t) - 1); 5226 mask = VM_PAGE_BITS_ALL << shift; 5227 5228 old = *bits; 5229 do { 5230 new = old & ~mask; 5231 new |= newbits << shift; 5232 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5233 return (old >> shift); 5234 #endif /* PAGE_SIZE */ 5235 } 5236 5237 /* 5238 * vm_page_set_valid_range: 5239 * 5240 * Sets portions of a page valid. The arguments are expected 5241 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5242 * of any partial chunks touched by the range. The invalid portion of 5243 * such chunks will be zeroed. 5244 * 5245 * (base + size) must be less then or equal to PAGE_SIZE. 5246 */ 5247 void 5248 vm_page_set_valid_range(vm_page_t m, int base, int size) 5249 { 5250 int endoff, frag; 5251 vm_page_bits_t pagebits; 5252 5253 vm_page_assert_busied(m); 5254 if (size == 0) /* handle degenerate case */ 5255 return; 5256 5257 /* 5258 * If the base is not DEV_BSIZE aligned and the valid 5259 * bit is clear, we have to zero out a portion of the 5260 * first block. 5261 */ 5262 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5263 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5264 pmap_zero_page_area(m, frag, base - frag); 5265 5266 /* 5267 * If the ending offset is not DEV_BSIZE aligned and the 5268 * valid bit is clear, we have to zero out a portion of 5269 * the last block. 5270 */ 5271 endoff = base + size; 5272 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5273 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5274 pmap_zero_page_area(m, endoff, 5275 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5276 5277 /* 5278 * Assert that no previously invalid block that is now being validated 5279 * is already dirty. 5280 */ 5281 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5282 ("vm_page_set_valid_range: page %p is dirty", m)); 5283 5284 /* 5285 * Set valid bits inclusive of any overlap. 5286 */ 5287 pagebits = vm_page_bits(base, size); 5288 if (vm_page_xbusied(m)) 5289 m->valid |= pagebits; 5290 else 5291 vm_page_bits_set(m, &m->valid, pagebits); 5292 } 5293 5294 /* 5295 * Set the page dirty bits and free the invalid swap space if 5296 * present. Returns the previous dirty bits. 5297 */ 5298 vm_page_bits_t 5299 vm_page_set_dirty(vm_page_t m) 5300 { 5301 vm_page_bits_t old; 5302 5303 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5304 5305 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5306 old = m->dirty; 5307 m->dirty = VM_PAGE_BITS_ALL; 5308 } else 5309 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5310 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5311 vm_pager_page_unswapped(m); 5312 5313 return (old); 5314 } 5315 5316 /* 5317 * Clear the given bits from the specified page's dirty field. 5318 */ 5319 static __inline void 5320 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5321 { 5322 5323 vm_page_assert_busied(m); 5324 5325 /* 5326 * If the page is xbusied and not write mapped we are the 5327 * only thread that can modify dirty bits. Otherwise, The pmap 5328 * layer can call vm_page_dirty() without holding a distinguished 5329 * lock. The combination of page busy and atomic operations 5330 * suffice to guarantee consistency of the page dirty field. 5331 */ 5332 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5333 m->dirty &= ~pagebits; 5334 else 5335 vm_page_bits_clear(m, &m->dirty, pagebits); 5336 } 5337 5338 /* 5339 * vm_page_set_validclean: 5340 * 5341 * Sets portions of a page valid and clean. The arguments are expected 5342 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5343 * of any partial chunks touched by the range. The invalid portion of 5344 * such chunks will be zero'd. 5345 * 5346 * (base + size) must be less then or equal to PAGE_SIZE. 5347 */ 5348 void 5349 vm_page_set_validclean(vm_page_t m, int base, int size) 5350 { 5351 vm_page_bits_t oldvalid, pagebits; 5352 int endoff, frag; 5353 5354 vm_page_assert_busied(m); 5355 if (size == 0) /* handle degenerate case */ 5356 return; 5357 5358 /* 5359 * If the base is not DEV_BSIZE aligned and the valid 5360 * bit is clear, we have to zero out a portion of the 5361 * first block. 5362 */ 5363 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5364 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5365 pmap_zero_page_area(m, frag, base - frag); 5366 5367 /* 5368 * If the ending offset is not DEV_BSIZE aligned and the 5369 * valid bit is clear, we have to zero out a portion of 5370 * the last block. 5371 */ 5372 endoff = base + size; 5373 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5374 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5375 pmap_zero_page_area(m, endoff, 5376 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5377 5378 /* 5379 * Set valid, clear dirty bits. If validating the entire 5380 * page we can safely clear the pmap modify bit. We also 5381 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5382 * takes a write fault on a MAP_NOSYNC memory area the flag will 5383 * be set again. 5384 * 5385 * We set valid bits inclusive of any overlap, but we can only 5386 * clear dirty bits for DEV_BSIZE chunks that are fully within 5387 * the range. 5388 */ 5389 oldvalid = m->valid; 5390 pagebits = vm_page_bits(base, size); 5391 if (vm_page_xbusied(m)) 5392 m->valid |= pagebits; 5393 else 5394 vm_page_bits_set(m, &m->valid, pagebits); 5395 #if 0 /* NOT YET */ 5396 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5397 frag = DEV_BSIZE - frag; 5398 base += frag; 5399 size -= frag; 5400 if (size < 0) 5401 size = 0; 5402 } 5403 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5404 #endif 5405 if (base == 0 && size == PAGE_SIZE) { 5406 /* 5407 * The page can only be modified within the pmap if it is 5408 * mapped, and it can only be mapped if it was previously 5409 * fully valid. 5410 */ 5411 if (oldvalid == VM_PAGE_BITS_ALL) 5412 /* 5413 * Perform the pmap_clear_modify() first. Otherwise, 5414 * a concurrent pmap operation, such as 5415 * pmap_protect(), could clear a modification in the 5416 * pmap and set the dirty field on the page before 5417 * pmap_clear_modify() had begun and after the dirty 5418 * field was cleared here. 5419 */ 5420 pmap_clear_modify(m); 5421 m->dirty = 0; 5422 vm_page_aflag_clear(m, PGA_NOSYNC); 5423 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5424 m->dirty &= ~pagebits; 5425 else 5426 vm_page_clear_dirty_mask(m, pagebits); 5427 } 5428 5429 void 5430 vm_page_clear_dirty(vm_page_t m, int base, int size) 5431 { 5432 5433 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5434 } 5435 5436 /* 5437 * vm_page_set_invalid: 5438 * 5439 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5440 * valid and dirty bits for the effected areas are cleared. 5441 */ 5442 void 5443 vm_page_set_invalid(vm_page_t m, int base, int size) 5444 { 5445 vm_page_bits_t bits; 5446 vm_object_t object; 5447 5448 /* 5449 * The object lock is required so that pages can't be mapped 5450 * read-only while we're in the process of invalidating them. 5451 */ 5452 object = m->object; 5453 VM_OBJECT_ASSERT_WLOCKED(object); 5454 vm_page_assert_busied(m); 5455 5456 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5457 size >= object->un_pager.vnp.vnp_size) 5458 bits = VM_PAGE_BITS_ALL; 5459 else 5460 bits = vm_page_bits(base, size); 5461 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5462 pmap_remove_all(m); 5463 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5464 !pmap_page_is_mapped(m), 5465 ("vm_page_set_invalid: page %p is mapped", m)); 5466 if (vm_page_xbusied(m)) { 5467 m->valid &= ~bits; 5468 m->dirty &= ~bits; 5469 } else { 5470 vm_page_bits_clear(m, &m->valid, bits); 5471 vm_page_bits_clear(m, &m->dirty, bits); 5472 } 5473 } 5474 5475 /* 5476 * vm_page_invalid: 5477 * 5478 * Invalidates the entire page. The page must be busy, unmapped, and 5479 * the enclosing object must be locked. The object locks protects 5480 * against concurrent read-only pmap enter which is done without 5481 * busy. 5482 */ 5483 void 5484 vm_page_invalid(vm_page_t m) 5485 { 5486 5487 vm_page_assert_busied(m); 5488 VM_OBJECT_ASSERT_WLOCKED(m->object); 5489 MPASS(!pmap_page_is_mapped(m)); 5490 5491 if (vm_page_xbusied(m)) 5492 m->valid = 0; 5493 else 5494 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5495 } 5496 5497 /* 5498 * vm_page_zero_invalid() 5499 * 5500 * The kernel assumes that the invalid portions of a page contain 5501 * garbage, but such pages can be mapped into memory by user code. 5502 * When this occurs, we must zero out the non-valid portions of the 5503 * page so user code sees what it expects. 5504 * 5505 * Pages are most often semi-valid when the end of a file is mapped 5506 * into memory and the file's size is not page aligned. 5507 */ 5508 void 5509 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5510 { 5511 int b; 5512 int i; 5513 5514 /* 5515 * Scan the valid bits looking for invalid sections that 5516 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5517 * valid bit may be set ) have already been zeroed by 5518 * vm_page_set_validclean(). 5519 */ 5520 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5521 if (i == (PAGE_SIZE / DEV_BSIZE) || 5522 (m->valid & ((vm_page_bits_t)1 << i))) { 5523 if (i > b) { 5524 pmap_zero_page_area(m, 5525 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5526 } 5527 b = i + 1; 5528 } 5529 } 5530 5531 /* 5532 * setvalid is TRUE when we can safely set the zero'd areas 5533 * as being valid. We can do this if there are no cache consistency 5534 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5535 */ 5536 if (setvalid) 5537 vm_page_valid(m); 5538 } 5539 5540 /* 5541 * vm_page_is_valid: 5542 * 5543 * Is (partial) page valid? Note that the case where size == 0 5544 * will return FALSE in the degenerate case where the page is 5545 * entirely invalid, and TRUE otherwise. 5546 * 5547 * Some callers envoke this routine without the busy lock held and 5548 * handle races via higher level locks. Typical callers should 5549 * hold a busy lock to prevent invalidation. 5550 */ 5551 int 5552 vm_page_is_valid(vm_page_t m, int base, int size) 5553 { 5554 vm_page_bits_t bits; 5555 5556 bits = vm_page_bits(base, size); 5557 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5558 } 5559 5560 /* 5561 * Returns true if all of the specified predicates are true for the entire 5562 * (super)page and false otherwise. 5563 */ 5564 bool 5565 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 5566 { 5567 vm_object_t object; 5568 int i, npages; 5569 5570 object = m->object; 5571 if (skip_m != NULL && skip_m->object != object) 5572 return (false); 5573 VM_OBJECT_ASSERT_LOCKED(object); 5574 npages = atop(pagesizes[m->psind]); 5575 5576 /* 5577 * The physically contiguous pages that make up a superpage, i.e., a 5578 * page with a page size index ("psind") greater than zero, will 5579 * occupy adjacent entries in vm_page_array[]. 5580 */ 5581 for (i = 0; i < npages; i++) { 5582 /* Always test object consistency, including "skip_m". */ 5583 if (m[i].object != object) 5584 return (false); 5585 if (&m[i] == skip_m) 5586 continue; 5587 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5588 return (false); 5589 if ((flags & PS_ALL_DIRTY) != 0) { 5590 /* 5591 * Calling vm_page_test_dirty() or pmap_is_modified() 5592 * might stop this case from spuriously returning 5593 * "false". However, that would require a write lock 5594 * on the object containing "m[i]". 5595 */ 5596 if (m[i].dirty != VM_PAGE_BITS_ALL) 5597 return (false); 5598 } 5599 if ((flags & PS_ALL_VALID) != 0 && 5600 m[i].valid != VM_PAGE_BITS_ALL) 5601 return (false); 5602 } 5603 return (true); 5604 } 5605 5606 /* 5607 * Set the page's dirty bits if the page is modified. 5608 */ 5609 void 5610 vm_page_test_dirty(vm_page_t m) 5611 { 5612 5613 vm_page_assert_busied(m); 5614 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5615 vm_page_dirty(m); 5616 } 5617 5618 void 5619 vm_page_valid(vm_page_t m) 5620 { 5621 5622 vm_page_assert_busied(m); 5623 if (vm_page_xbusied(m)) 5624 m->valid = VM_PAGE_BITS_ALL; 5625 else 5626 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5627 } 5628 5629 void 5630 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5631 { 5632 5633 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5634 } 5635 5636 void 5637 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5638 { 5639 5640 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5641 } 5642 5643 int 5644 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5645 { 5646 5647 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5648 } 5649 5650 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5651 void 5652 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5653 { 5654 5655 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5656 } 5657 5658 void 5659 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5660 { 5661 5662 mtx_assert_(vm_page_lockptr(m), a, file, line); 5663 } 5664 #endif 5665 5666 #ifdef INVARIANTS 5667 void 5668 vm_page_object_busy_assert(vm_page_t m) 5669 { 5670 5671 /* 5672 * Certain of the page's fields may only be modified by the 5673 * holder of a page or object busy. 5674 */ 5675 if (m->object != NULL && !vm_page_busied(m)) 5676 VM_OBJECT_ASSERT_BUSY(m->object); 5677 } 5678 5679 void 5680 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5681 { 5682 5683 if ((bits & PGA_WRITEABLE) == 0) 5684 return; 5685 5686 /* 5687 * The PGA_WRITEABLE flag can only be set if the page is 5688 * managed, is exclusively busied or the object is locked. 5689 * Currently, this flag is only set by pmap_enter(). 5690 */ 5691 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5692 ("PGA_WRITEABLE on unmanaged page")); 5693 if (!vm_page_xbusied(m)) 5694 VM_OBJECT_ASSERT_BUSY(m->object); 5695 } 5696 #endif 5697 5698 #include "opt_ddb.h" 5699 #ifdef DDB 5700 #include <sys/kernel.h> 5701 5702 #include <ddb/ddb.h> 5703 5704 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5705 { 5706 5707 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5708 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5709 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5710 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5711 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5712 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5713 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5714 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5715 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5716 } 5717 5718 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5719 { 5720 int dom; 5721 5722 db_printf("pq_free %d\n", vm_free_count()); 5723 for (dom = 0; dom < vm_ndomains; dom++) { 5724 db_printf( 5725 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5726 dom, 5727 vm_dom[dom].vmd_page_count, 5728 vm_dom[dom].vmd_free_count, 5729 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5730 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5731 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5732 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5733 } 5734 } 5735 5736 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5737 { 5738 vm_page_t m; 5739 boolean_t phys, virt; 5740 5741 if (!have_addr) { 5742 db_printf("show pginfo addr\n"); 5743 return; 5744 } 5745 5746 phys = strchr(modif, 'p') != NULL; 5747 virt = strchr(modif, 'v') != NULL; 5748 if (virt) 5749 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5750 else if (phys) 5751 m = PHYS_TO_VM_PAGE(addr); 5752 else 5753 m = (vm_page_t)addr; 5754 db_printf( 5755 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5756 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5757 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5758 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5759 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5760 } 5761 #endif /* DDB */ 5762