1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * The Mach Operating System project at Carnegie-Mellon University. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 * 35 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 36 */ 37 38 /*- 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 */ 64 65 /* 66 * Resident memory management module. 67 */ 68 69 #include <sys/cdefs.h> 70 __FBSDID("$FreeBSD$"); 71 72 #include "opt_vm.h" 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/counter.h> 77 #include <sys/domainset.h> 78 #include <sys/kernel.h> 79 #include <sys/limits.h> 80 #include <sys/linker.h> 81 #include <sys/lock.h> 82 #include <sys/malloc.h> 83 #include <sys/mman.h> 84 #include <sys/msgbuf.h> 85 #include <sys/mutex.h> 86 #include <sys/proc.h> 87 #include <sys/rwlock.h> 88 #include <sys/sleepqueue.h> 89 #include <sys/sbuf.h> 90 #include <sys/sched.h> 91 #include <sys/smp.h> 92 #include <sys/sysctl.h> 93 #include <sys/vmmeter.h> 94 #include <sys/vnode.h> 95 96 #include <vm/vm.h> 97 #include <vm/pmap.h> 98 #include <vm/vm_param.h> 99 #include <vm/vm_domainset.h> 100 #include <vm/vm_kern.h> 101 #include <vm/vm_map.h> 102 #include <vm/vm_object.h> 103 #include <vm/vm_page.h> 104 #include <vm/vm_pageout.h> 105 #include <vm/vm_phys.h> 106 #include <vm/vm_pagequeue.h> 107 #include <vm/vm_pager.h> 108 #include <vm/vm_radix.h> 109 #include <vm/vm_reserv.h> 110 #include <vm/vm_extern.h> 111 #include <vm/uma.h> 112 #include <vm/uma_int.h> 113 114 #include <machine/md_var.h> 115 116 extern int uma_startup_count(int); 117 extern void uma_startup(void *, int); 118 extern int vmem_startup_count(void); 119 120 struct vm_domain vm_dom[MAXMEMDOM]; 121 122 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 123 124 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 125 126 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 127 /* The following fields are protected by the domainset lock. */ 128 domainset_t __exclusive_cache_line vm_min_domains; 129 domainset_t __exclusive_cache_line vm_severe_domains; 130 static int vm_min_waiters; 131 static int vm_severe_waiters; 132 static int vm_pageproc_waiters; 133 134 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0, 135 "VM page statistics"); 136 137 static counter_u64_t queue_ops = EARLY_COUNTER; 138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 139 CTLFLAG_RD, &queue_ops, 140 "Number of batched queue operations"); 141 142 static counter_u64_t queue_nops = EARLY_COUNTER; 143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 144 CTLFLAG_RD, &queue_nops, 145 "Number of batched queue operations with no effects"); 146 147 static void 148 counter_startup(void) 149 { 150 151 queue_ops = counter_u64_alloc(M_WAITOK); 152 queue_nops = counter_u64_alloc(M_WAITOK); 153 } 154 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); 155 156 /* 157 * bogus page -- for I/O to/from partially complete buffers, 158 * or for paging into sparsely invalid regions. 159 */ 160 vm_page_t bogus_page; 161 162 vm_page_t vm_page_array; 163 long vm_page_array_size; 164 long first_page; 165 166 static int boot_pages; 167 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 168 &boot_pages, 0, 169 "number of pages allocated for bootstrapping the VM system"); 170 171 static TAILQ_HEAD(, vm_page) blacklist_head; 172 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 173 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 174 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 175 176 static uma_zone_t fakepg_zone; 177 178 static void vm_page_alloc_check(vm_page_t m); 179 static void _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 180 const char *wmesg, bool nonshared, bool locked); 181 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 182 static void vm_page_dequeue_complete(vm_page_t m); 183 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 184 static void vm_page_init(void *dummy); 185 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 186 vm_pindex_t pindex, vm_page_t mpred); 187 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 188 vm_page_t mpred); 189 static void vm_page_mvqueue(vm_page_t m, uint8_t queue); 190 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 191 vm_page_t m_run, vm_paddr_t high); 192 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 193 int req); 194 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 195 int flags); 196 static void vm_page_zone_release(void *arg, void **store, int cnt); 197 198 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 199 200 static void 201 vm_page_init(void *dummy) 202 { 203 204 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 205 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 206 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 207 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 208 } 209 210 /* 211 * The cache page zone is initialized later since we need to be able to allocate 212 * pages before UMA is fully initialized. 213 */ 214 static void 215 vm_page_init_cache_zones(void *dummy __unused) 216 { 217 struct vm_domain *vmd; 218 struct vm_pgcache *pgcache; 219 int domain, maxcache, pool; 220 221 maxcache = 0; 222 TUNABLE_INT_FETCH("vm.pgcache_zone_max", &maxcache); 223 for (domain = 0; domain < vm_ndomains; domain++) { 224 vmd = VM_DOMAIN(domain); 225 226 /* 227 * Don't allow the page caches to take up more than .1875% of 228 * memory. A UMA bucket contains at most 256 free pages, and we 229 * have two buckets per CPU per free pool. 230 */ 231 if (vmd->vmd_page_count / 600 < 2 * 256 * mp_ncpus * 232 VM_NFREEPOOL) 233 continue; 234 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 235 pgcache = &vmd->vmd_pgcache[pool]; 236 pgcache->domain = domain; 237 pgcache->pool = pool; 238 pgcache->zone = uma_zcache_create("vm pgcache", 239 sizeof(struct vm_page), NULL, NULL, NULL, NULL, 240 vm_page_zone_import, vm_page_zone_release, pgcache, 241 UMA_ZONE_MAXBUCKET | UMA_ZONE_VM); 242 (void)uma_zone_set_maxcache(pgcache->zone, maxcache); 243 } 244 } 245 } 246 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 247 248 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 249 #if PAGE_SIZE == 32768 250 #ifdef CTASSERT 251 CTASSERT(sizeof(u_long) >= 8); 252 #endif 253 #endif 254 255 /* 256 * vm_set_page_size: 257 * 258 * Sets the page size, perhaps based upon the memory 259 * size. Must be called before any use of page-size 260 * dependent functions. 261 */ 262 void 263 vm_set_page_size(void) 264 { 265 if (vm_cnt.v_page_size == 0) 266 vm_cnt.v_page_size = PAGE_SIZE; 267 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 268 panic("vm_set_page_size: page size not a power of two"); 269 } 270 271 /* 272 * vm_page_blacklist_next: 273 * 274 * Find the next entry in the provided string of blacklist 275 * addresses. Entries are separated by space, comma, or newline. 276 * If an invalid integer is encountered then the rest of the 277 * string is skipped. Updates the list pointer to the next 278 * character, or NULL if the string is exhausted or invalid. 279 */ 280 static vm_paddr_t 281 vm_page_blacklist_next(char **list, char *end) 282 { 283 vm_paddr_t bad; 284 char *cp, *pos; 285 286 if (list == NULL || *list == NULL) 287 return (0); 288 if (**list =='\0') { 289 *list = NULL; 290 return (0); 291 } 292 293 /* 294 * If there's no end pointer then the buffer is coming from 295 * the kenv and we know it's null-terminated. 296 */ 297 if (end == NULL) 298 end = *list + strlen(*list); 299 300 /* Ensure that strtoq() won't walk off the end */ 301 if (*end != '\0') { 302 if (*end == '\n' || *end == ' ' || *end == ',') 303 *end = '\0'; 304 else { 305 printf("Blacklist not terminated, skipping\n"); 306 *list = NULL; 307 return (0); 308 } 309 } 310 311 for (pos = *list; *pos != '\0'; pos = cp) { 312 bad = strtoq(pos, &cp, 0); 313 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 314 if (bad == 0) { 315 if (++cp < end) 316 continue; 317 else 318 break; 319 } 320 } else 321 break; 322 if (*cp == '\0' || ++cp >= end) 323 *list = NULL; 324 else 325 *list = cp; 326 return (trunc_page(bad)); 327 } 328 printf("Garbage in RAM blacklist, skipping\n"); 329 *list = NULL; 330 return (0); 331 } 332 333 bool 334 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 335 { 336 struct vm_domain *vmd; 337 vm_page_t m; 338 int ret; 339 340 m = vm_phys_paddr_to_vm_page(pa); 341 if (m == NULL) 342 return (true); /* page does not exist, no failure */ 343 344 vmd = vm_pagequeue_domain(m); 345 vm_domain_free_lock(vmd); 346 ret = vm_phys_unfree_page(m); 347 vm_domain_free_unlock(vmd); 348 if (ret != 0) { 349 vm_domain_freecnt_inc(vmd, -1); 350 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 351 if (verbose) 352 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 353 } 354 return (ret); 355 } 356 357 /* 358 * vm_page_blacklist_check: 359 * 360 * Iterate through the provided string of blacklist addresses, pulling 361 * each entry out of the physical allocator free list and putting it 362 * onto a list for reporting via the vm.page_blacklist sysctl. 363 */ 364 static void 365 vm_page_blacklist_check(char *list, char *end) 366 { 367 vm_paddr_t pa; 368 char *next; 369 370 next = list; 371 while (next != NULL) { 372 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 373 continue; 374 vm_page_blacklist_add(pa, bootverbose); 375 } 376 } 377 378 /* 379 * vm_page_blacklist_load: 380 * 381 * Search for a special module named "ram_blacklist". It'll be a 382 * plain text file provided by the user via the loader directive 383 * of the same name. 384 */ 385 static void 386 vm_page_blacklist_load(char **list, char **end) 387 { 388 void *mod; 389 u_char *ptr; 390 u_int len; 391 392 mod = NULL; 393 ptr = NULL; 394 395 mod = preload_search_by_type("ram_blacklist"); 396 if (mod != NULL) { 397 ptr = preload_fetch_addr(mod); 398 len = preload_fetch_size(mod); 399 } 400 *list = ptr; 401 if (ptr != NULL) 402 *end = ptr + len; 403 else 404 *end = NULL; 405 return; 406 } 407 408 static int 409 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 410 { 411 vm_page_t m; 412 struct sbuf sbuf; 413 int error, first; 414 415 first = 1; 416 error = sysctl_wire_old_buffer(req, 0); 417 if (error != 0) 418 return (error); 419 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 420 TAILQ_FOREACH(m, &blacklist_head, listq) { 421 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 422 (uintmax_t)m->phys_addr); 423 first = 0; 424 } 425 error = sbuf_finish(&sbuf); 426 sbuf_delete(&sbuf); 427 return (error); 428 } 429 430 /* 431 * Initialize a dummy page for use in scans of the specified paging queue. 432 * In principle, this function only needs to set the flag PG_MARKER. 433 * Nonetheless, it write busies the page as a safety precaution. 434 */ 435 static void 436 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags) 437 { 438 439 bzero(marker, sizeof(*marker)); 440 marker->flags = PG_MARKER; 441 marker->aflags = aflags; 442 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 443 marker->queue = queue; 444 } 445 446 static void 447 vm_page_domain_init(int domain) 448 { 449 struct vm_domain *vmd; 450 struct vm_pagequeue *pq; 451 int i; 452 453 vmd = VM_DOMAIN(domain); 454 bzero(vmd, sizeof(*vmd)); 455 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 456 "vm inactive pagequeue"; 457 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 458 "vm active pagequeue"; 459 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 460 "vm laundry pagequeue"; 461 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 462 "vm unswappable pagequeue"; 463 vmd->vmd_domain = domain; 464 vmd->vmd_page_count = 0; 465 vmd->vmd_free_count = 0; 466 vmd->vmd_segs = 0; 467 vmd->vmd_oom = FALSE; 468 for (i = 0; i < PQ_COUNT; i++) { 469 pq = &vmd->vmd_pagequeues[i]; 470 TAILQ_INIT(&pq->pq_pl); 471 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 472 MTX_DEF | MTX_DUPOK); 473 pq->pq_pdpages = 0; 474 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 475 } 476 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 477 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 478 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 479 480 /* 481 * inacthead is used to provide FIFO ordering for LRU-bypassing 482 * insertions. 483 */ 484 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 485 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 486 &vmd->vmd_inacthead, plinks.q); 487 488 /* 489 * The clock pages are used to implement active queue scanning without 490 * requeues. Scans start at clock[0], which is advanced after the scan 491 * ends. When the two clock hands meet, they are reset and scanning 492 * resumes from the head of the queue. 493 */ 494 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 495 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 496 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 497 &vmd->vmd_clock[0], plinks.q); 498 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 499 &vmd->vmd_clock[1], plinks.q); 500 } 501 502 /* 503 * Initialize a physical page in preparation for adding it to the free 504 * lists. 505 */ 506 static void 507 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 508 { 509 510 m->object = NULL; 511 m->ref_count = 0; 512 m->busy_lock = VPB_UNBUSIED; 513 m->flags = m->aflags = 0; 514 m->phys_addr = pa; 515 m->queue = PQ_NONE; 516 m->psind = 0; 517 m->segind = segind; 518 m->order = VM_NFREEORDER; 519 m->pool = VM_FREEPOOL_DEFAULT; 520 m->valid = m->dirty = 0; 521 pmap_page_init(m); 522 } 523 524 #ifndef PMAP_HAS_PAGE_ARRAY 525 static vm_paddr_t 526 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 527 { 528 vm_paddr_t new_end; 529 530 /* 531 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 532 * However, because this page is allocated from KVM, out-of-bounds 533 * accesses using the direct map will not be trapped. 534 */ 535 *vaddr += PAGE_SIZE; 536 537 /* 538 * Allocate physical memory for the page structures, and map it. 539 */ 540 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 541 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 542 VM_PROT_READ | VM_PROT_WRITE); 543 vm_page_array_size = page_range; 544 545 return (new_end); 546 } 547 #endif 548 549 /* 550 * vm_page_startup: 551 * 552 * Initializes the resident memory module. Allocates physical memory for 553 * bootstrapping UMA and some data structures that are used to manage 554 * physical pages. Initializes these structures, and populates the free 555 * page queues. 556 */ 557 vm_offset_t 558 vm_page_startup(vm_offset_t vaddr) 559 { 560 struct vm_phys_seg *seg; 561 vm_page_t m; 562 char *list, *listend; 563 vm_offset_t mapped; 564 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; 565 vm_paddr_t last_pa, pa; 566 u_long pagecount; 567 int biggestone, i, segind; 568 #ifdef WITNESS 569 int witness_size; 570 #endif 571 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 572 long ii; 573 #endif 574 575 vaddr = round_page(vaddr); 576 577 vm_phys_early_startup(); 578 biggestone = vm_phys_avail_largest(); 579 end = phys_avail[biggestone+1]; 580 581 /* 582 * Initialize the page and queue locks. 583 */ 584 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 585 for (i = 0; i < PA_LOCK_COUNT; i++) 586 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 587 for (i = 0; i < vm_ndomains; i++) 588 vm_page_domain_init(i); 589 590 /* 591 * Allocate memory for use when boot strapping the kernel memory 592 * allocator. Tell UMA how many zones we are going to create 593 * before going fully functional. UMA will add its zones. 594 * 595 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, 596 * KMAP ENTRY, MAP ENTRY, VMSPACE. 597 */ 598 boot_pages = uma_startup_count(8); 599 600 #ifndef UMA_MD_SMALL_ALLOC 601 /* vmem_startup() calls uma_prealloc(). */ 602 boot_pages += vmem_startup_count(); 603 /* vm_map_startup() calls uma_prealloc(). */ 604 boot_pages += howmany(MAX_KMAP, 605 UMA_SLAB_SPACE / sizeof(struct vm_map)); 606 607 /* 608 * Before going fully functional kmem_init() does allocation 609 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". 610 */ 611 boot_pages += 2; 612 #endif 613 /* 614 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 615 * manually fetch the value. 616 */ 617 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 618 new_end = end - (boot_pages * UMA_SLAB_SIZE); 619 new_end = trunc_page(new_end); 620 mapped = pmap_map(&vaddr, new_end, end, 621 VM_PROT_READ | VM_PROT_WRITE); 622 bzero((void *)mapped, end - new_end); 623 uma_startup((void *)mapped, boot_pages); 624 625 #ifdef WITNESS 626 witness_size = round_page(witness_startup_count()); 627 new_end -= witness_size; 628 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 629 VM_PROT_READ | VM_PROT_WRITE); 630 bzero((void *)mapped, witness_size); 631 witness_startup((void *)mapped); 632 #endif 633 634 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 635 defined(__i386__) || defined(__mips__) || defined(__riscv) || \ 636 defined(__powerpc64__) 637 /* 638 * Allocate a bitmap to indicate that a random physical page 639 * needs to be included in a minidump. 640 * 641 * The amd64 port needs this to indicate which direct map pages 642 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 643 * 644 * However, i386 still needs this workspace internally within the 645 * minidump code. In theory, they are not needed on i386, but are 646 * included should the sf_buf code decide to use them. 647 */ 648 last_pa = 0; 649 for (i = 0; dump_avail[i + 1] != 0; i += 2) 650 if (dump_avail[i + 1] > last_pa) 651 last_pa = dump_avail[i + 1]; 652 page_range = last_pa / PAGE_SIZE; 653 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 654 new_end -= vm_page_dump_size; 655 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 656 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 657 bzero((void *)vm_page_dump, vm_page_dump_size); 658 #else 659 (void)last_pa; 660 #endif 661 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 662 defined(__riscv) || defined(__powerpc64__) 663 /* 664 * Include the UMA bootstrap pages, witness pages and vm_page_dump 665 * in a crash dump. When pmap_map() uses the direct map, they are 666 * not automatically included. 667 */ 668 for (pa = new_end; pa < end; pa += PAGE_SIZE) 669 dump_add_page(pa); 670 #endif 671 phys_avail[biggestone + 1] = new_end; 672 #ifdef __amd64__ 673 /* 674 * Request that the physical pages underlying the message buffer be 675 * included in a crash dump. Since the message buffer is accessed 676 * through the direct map, they are not automatically included. 677 */ 678 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 679 last_pa = pa + round_page(msgbufsize); 680 while (pa < last_pa) { 681 dump_add_page(pa); 682 pa += PAGE_SIZE; 683 } 684 #endif 685 /* 686 * Compute the number of pages of memory that will be available for 687 * use, taking into account the overhead of a page structure per page. 688 * In other words, solve 689 * "available physical memory" - round_page(page_range * 690 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 691 * for page_range. 692 */ 693 low_avail = phys_avail[0]; 694 high_avail = phys_avail[1]; 695 for (i = 0; i < vm_phys_nsegs; i++) { 696 if (vm_phys_segs[i].start < low_avail) 697 low_avail = vm_phys_segs[i].start; 698 if (vm_phys_segs[i].end > high_avail) 699 high_avail = vm_phys_segs[i].end; 700 } 701 /* Skip the first chunk. It is already accounted for. */ 702 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 703 if (phys_avail[i] < low_avail) 704 low_avail = phys_avail[i]; 705 if (phys_avail[i + 1] > high_avail) 706 high_avail = phys_avail[i + 1]; 707 } 708 first_page = low_avail / PAGE_SIZE; 709 #ifdef VM_PHYSSEG_SPARSE 710 size = 0; 711 for (i = 0; i < vm_phys_nsegs; i++) 712 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 713 for (i = 0; phys_avail[i + 1] != 0; i += 2) 714 size += phys_avail[i + 1] - phys_avail[i]; 715 #elif defined(VM_PHYSSEG_DENSE) 716 size = high_avail - low_avail; 717 #else 718 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 719 #endif 720 721 #ifdef PMAP_HAS_PAGE_ARRAY 722 pmap_page_array_startup(size / PAGE_SIZE); 723 biggestone = vm_phys_avail_largest(); 724 end = new_end = phys_avail[biggestone + 1]; 725 #else 726 #ifdef VM_PHYSSEG_DENSE 727 /* 728 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 729 * the overhead of a page structure per page only if vm_page_array is 730 * allocated from the last physical memory chunk. Otherwise, we must 731 * allocate page structures representing the physical memory 732 * underlying vm_page_array, even though they will not be used. 733 */ 734 if (new_end != high_avail) 735 page_range = size / PAGE_SIZE; 736 else 737 #endif 738 { 739 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 740 741 /* 742 * If the partial bytes remaining are large enough for 743 * a page (PAGE_SIZE) without a corresponding 744 * 'struct vm_page', then new_end will contain an 745 * extra page after subtracting the length of the VM 746 * page array. Compensate by subtracting an extra 747 * page from new_end. 748 */ 749 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 750 if (new_end == high_avail) 751 high_avail -= PAGE_SIZE; 752 new_end -= PAGE_SIZE; 753 } 754 } 755 end = new_end; 756 new_end = vm_page_array_alloc(&vaddr, end, page_range); 757 #endif 758 759 #if VM_NRESERVLEVEL > 0 760 /* 761 * Allocate physical memory for the reservation management system's 762 * data structures, and map it. 763 */ 764 new_end = vm_reserv_startup(&vaddr, new_end); 765 #endif 766 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 767 defined(__riscv) || defined(__powerpc64__) 768 /* 769 * Include vm_page_array and vm_reserv_array in a crash dump. 770 */ 771 for (pa = new_end; pa < end; pa += PAGE_SIZE) 772 dump_add_page(pa); 773 #endif 774 phys_avail[biggestone + 1] = new_end; 775 776 /* 777 * Add physical memory segments corresponding to the available 778 * physical pages. 779 */ 780 for (i = 0; phys_avail[i + 1] != 0; i += 2) 781 if (vm_phys_avail_size(i) != 0) 782 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 783 784 /* 785 * Initialize the physical memory allocator. 786 */ 787 vm_phys_init(); 788 789 /* 790 * Initialize the page structures and add every available page to the 791 * physical memory allocator's free lists. 792 */ 793 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 794 for (ii = 0; ii < vm_page_array_size; ii++) { 795 m = &vm_page_array[ii]; 796 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 797 m->flags = PG_FICTITIOUS; 798 } 799 #endif 800 vm_cnt.v_page_count = 0; 801 for (segind = 0; segind < vm_phys_nsegs; segind++) { 802 seg = &vm_phys_segs[segind]; 803 for (m = seg->first_page, pa = seg->start; pa < seg->end; 804 m++, pa += PAGE_SIZE) 805 vm_page_init_page(m, pa, segind); 806 807 /* 808 * Add the segment to the free lists only if it is covered by 809 * one of the ranges in phys_avail. Because we've added the 810 * ranges to the vm_phys_segs array, we can assume that each 811 * segment is either entirely contained in one of the ranges, 812 * or doesn't overlap any of them. 813 */ 814 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 815 struct vm_domain *vmd; 816 817 if (seg->start < phys_avail[i] || 818 seg->end > phys_avail[i + 1]) 819 continue; 820 821 m = seg->first_page; 822 pagecount = (u_long)atop(seg->end - seg->start); 823 824 vmd = VM_DOMAIN(seg->domain); 825 vm_domain_free_lock(vmd); 826 vm_phys_enqueue_contig(m, pagecount); 827 vm_domain_free_unlock(vmd); 828 vm_domain_freecnt_inc(vmd, pagecount); 829 vm_cnt.v_page_count += (u_int)pagecount; 830 831 vmd = VM_DOMAIN(seg->domain); 832 vmd->vmd_page_count += (u_int)pagecount; 833 vmd->vmd_segs |= 1UL << m->segind; 834 break; 835 } 836 } 837 838 /* 839 * Remove blacklisted pages from the physical memory allocator. 840 */ 841 TAILQ_INIT(&blacklist_head); 842 vm_page_blacklist_load(&list, &listend); 843 vm_page_blacklist_check(list, listend); 844 845 list = kern_getenv("vm.blacklist"); 846 vm_page_blacklist_check(list, NULL); 847 848 freeenv(list); 849 #if VM_NRESERVLEVEL > 0 850 /* 851 * Initialize the reservation management system. 852 */ 853 vm_reserv_init(); 854 #endif 855 856 return (vaddr); 857 } 858 859 void 860 vm_page_reference(vm_page_t m) 861 { 862 863 vm_page_aflag_set(m, PGA_REFERENCED); 864 } 865 866 /* 867 * vm_page_busy_acquire: 868 * 869 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 870 * and drop the object lock if necessary. 871 */ 872 int 873 vm_page_busy_acquire(vm_page_t m, int allocflags) 874 { 875 vm_object_t obj; 876 bool locked; 877 878 /* 879 * The page-specific object must be cached because page 880 * identity can change during the sleep, causing the 881 * re-lock of a different object. 882 * It is assumed that a reference to the object is already 883 * held by the callers. 884 */ 885 obj = m->object; 886 for (;;) { 887 if ((allocflags & VM_ALLOC_SBUSY) == 0) { 888 if (vm_page_tryxbusy(m)) 889 return (TRUE); 890 } else { 891 if (vm_page_trysbusy(m)) 892 return (TRUE); 893 } 894 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 895 return (FALSE); 896 if (obj != NULL) 897 locked = VM_OBJECT_WOWNED(obj); 898 else 899 locked = FALSE; 900 MPASS(locked || vm_page_wired(m)); 901 _vm_page_busy_sleep(obj, m, "vmpba", 902 (allocflags & VM_ALLOC_SBUSY) != 0, locked); 903 if (locked) 904 VM_OBJECT_WLOCK(obj); 905 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 906 return (FALSE); 907 KASSERT(m->object == obj || m->object == NULL, 908 ("vm_page_busy_acquire: page %p does not belong to %p", 909 m, obj)); 910 } 911 } 912 913 /* 914 * vm_page_busy_downgrade: 915 * 916 * Downgrade an exclusive busy page into a single shared busy page. 917 */ 918 void 919 vm_page_busy_downgrade(vm_page_t m) 920 { 921 u_int x; 922 923 vm_page_assert_xbusied(m); 924 925 x = m->busy_lock; 926 for (;;) { 927 if (atomic_fcmpset_rel_int(&m->busy_lock, 928 &x, VPB_SHARERS_WORD(1))) 929 break; 930 } 931 if ((x & VPB_BIT_WAITERS) != 0) 932 wakeup(m); 933 } 934 935 /* 936 * 937 * vm_page_busy_tryupgrade: 938 * 939 * Attempt to upgrade a single shared busy into an exclusive busy. 940 */ 941 int 942 vm_page_busy_tryupgrade(vm_page_t m) 943 { 944 u_int x; 945 946 vm_page_assert_sbusied(m); 947 948 x = m->busy_lock; 949 for (;;) { 950 if (VPB_SHARERS(x) > 1) 951 return (0); 952 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 953 ("vm_page_busy_tryupgrade: invalid lock state")); 954 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 955 VPB_SINGLE_EXCLUSIVER | (x & VPB_BIT_WAITERS))) 956 continue; 957 return (1); 958 } 959 } 960 961 /* 962 * vm_page_sbusied: 963 * 964 * Return a positive value if the page is shared busied, 0 otherwise. 965 */ 966 int 967 vm_page_sbusied(vm_page_t m) 968 { 969 u_int x; 970 971 x = m->busy_lock; 972 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 973 } 974 975 /* 976 * vm_page_sunbusy: 977 * 978 * Shared unbusy a page. 979 */ 980 void 981 vm_page_sunbusy(vm_page_t m) 982 { 983 u_int x; 984 985 vm_page_assert_sbusied(m); 986 987 x = m->busy_lock; 988 for (;;) { 989 if (VPB_SHARERS(x) > 1) { 990 if (atomic_fcmpset_int(&m->busy_lock, &x, 991 x - VPB_ONE_SHARER)) 992 break; 993 continue; 994 } 995 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 996 ("vm_page_sunbusy: invalid lock state")); 997 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 998 continue; 999 if ((x & VPB_BIT_WAITERS) == 0) 1000 break; 1001 wakeup(m); 1002 break; 1003 } 1004 } 1005 1006 /* 1007 * vm_page_busy_sleep: 1008 * 1009 * Sleep if the page is busy, using the page pointer as wchan. 1010 * This is used to implement the hard-path of busying mechanism. 1011 * 1012 * If nonshared is true, sleep only if the page is xbusy. 1013 * 1014 * The object lock must be held on entry and will be released on exit. 1015 */ 1016 void 1017 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 1018 { 1019 vm_object_t obj; 1020 1021 obj = m->object; 1022 VM_OBJECT_ASSERT_LOCKED(obj); 1023 vm_page_lock_assert(m, MA_NOTOWNED); 1024 1025 _vm_page_busy_sleep(obj, m, wmesg, nonshared, true); 1026 } 1027 1028 static void 1029 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, 1030 bool nonshared, bool locked) 1031 { 1032 u_int x; 1033 1034 /* 1035 * If the object is busy we must wait for that to drain to zero 1036 * before trying the page again. 1037 */ 1038 if (obj != NULL && vm_object_busied(obj)) { 1039 if (locked) 1040 VM_OBJECT_DROP(obj); 1041 vm_object_busy_wait(obj, wmesg); 1042 return; 1043 } 1044 sleepq_lock(m); 1045 x = m->busy_lock; 1046 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 1047 ((x & VPB_BIT_WAITERS) == 0 && 1048 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 1049 if (locked) 1050 VM_OBJECT_DROP(obj); 1051 sleepq_release(m); 1052 return; 1053 } 1054 if (locked) 1055 VM_OBJECT_DROP(obj); 1056 sleepq_add(m, NULL, wmesg, 0, 0); 1057 sleepq_wait(m, PVM); 1058 } 1059 1060 /* 1061 * vm_page_trysbusy: 1062 * 1063 * Try to shared busy a page. 1064 * If the operation succeeds 1 is returned otherwise 0. 1065 * The operation never sleeps. 1066 */ 1067 int 1068 vm_page_trysbusy(vm_page_t m) 1069 { 1070 vm_object_t obj; 1071 u_int x; 1072 1073 obj = m->object; 1074 x = m->busy_lock; 1075 for (;;) { 1076 if ((x & VPB_BIT_SHARED) == 0) 1077 return (0); 1078 /* 1079 * Reduce the window for transient busies that will trigger 1080 * false negatives in vm_page_ps_test(). 1081 */ 1082 if (obj != NULL && vm_object_busied(obj)) 1083 return (0); 1084 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1085 x + VPB_ONE_SHARER)) 1086 break; 1087 } 1088 1089 /* Refetch the object now that we're guaranteed that it is stable. */ 1090 obj = m->object; 1091 if (obj != NULL && vm_object_busied(obj)) { 1092 vm_page_sunbusy(m); 1093 return (0); 1094 } 1095 return (1); 1096 } 1097 1098 /* 1099 * vm_page_tryxbusy: 1100 * 1101 * Try to exclusive busy a page. 1102 * If the operation succeeds 1 is returned otherwise 0. 1103 * The operation never sleeps. 1104 */ 1105 int 1106 vm_page_tryxbusy(vm_page_t m) 1107 { 1108 vm_object_t obj; 1109 1110 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, 1111 VPB_SINGLE_EXCLUSIVER) == 0) 1112 return (0); 1113 1114 obj = m->object; 1115 if (obj != NULL && vm_object_busied(obj)) { 1116 vm_page_xunbusy(m); 1117 return (0); 1118 } 1119 return (1); 1120 } 1121 1122 /* 1123 * vm_page_xunbusy_hard: 1124 * 1125 * Called when unbusy has failed because there is a waiter. 1126 */ 1127 void 1128 vm_page_xunbusy_hard(vm_page_t m) 1129 { 1130 1131 vm_page_assert_xbusied(m); 1132 1133 /* 1134 * Wake the waiter. 1135 */ 1136 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1137 wakeup(m); 1138 } 1139 1140 /* 1141 * Avoid releasing and reacquiring the same page lock. 1142 */ 1143 void 1144 vm_page_change_lock(vm_page_t m, struct mtx **mtx) 1145 { 1146 struct mtx *mtx1; 1147 1148 mtx1 = vm_page_lockptr(m); 1149 if (*mtx == mtx1) 1150 return; 1151 if (*mtx != NULL) 1152 mtx_unlock(*mtx); 1153 *mtx = mtx1; 1154 mtx_lock(mtx1); 1155 } 1156 1157 /* 1158 * vm_page_unhold_pages: 1159 * 1160 * Unhold each of the pages that is referenced by the given array. 1161 */ 1162 void 1163 vm_page_unhold_pages(vm_page_t *ma, int count) 1164 { 1165 1166 for (; count != 0; count--) { 1167 vm_page_unwire(*ma, PQ_ACTIVE); 1168 ma++; 1169 } 1170 } 1171 1172 vm_page_t 1173 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1174 { 1175 vm_page_t m; 1176 1177 #ifdef VM_PHYSSEG_SPARSE 1178 m = vm_phys_paddr_to_vm_page(pa); 1179 if (m == NULL) 1180 m = vm_phys_fictitious_to_vm_page(pa); 1181 return (m); 1182 #elif defined(VM_PHYSSEG_DENSE) 1183 long pi; 1184 1185 pi = atop(pa); 1186 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1187 m = &vm_page_array[pi - first_page]; 1188 return (m); 1189 } 1190 return (vm_phys_fictitious_to_vm_page(pa)); 1191 #else 1192 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1193 #endif 1194 } 1195 1196 /* 1197 * vm_page_getfake: 1198 * 1199 * Create a fictitious page with the specified physical address and 1200 * memory attribute. The memory attribute is the only the machine- 1201 * dependent aspect of a fictitious page that must be initialized. 1202 */ 1203 vm_page_t 1204 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1205 { 1206 vm_page_t m; 1207 1208 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1209 vm_page_initfake(m, paddr, memattr); 1210 return (m); 1211 } 1212 1213 void 1214 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1215 { 1216 1217 if ((m->flags & PG_FICTITIOUS) != 0) { 1218 /* 1219 * The page's memattr might have changed since the 1220 * previous initialization. Update the pmap to the 1221 * new memattr. 1222 */ 1223 goto memattr; 1224 } 1225 m->phys_addr = paddr; 1226 m->queue = PQ_NONE; 1227 /* Fictitious pages don't use "segind". */ 1228 m->flags = PG_FICTITIOUS; 1229 /* Fictitious pages don't use "order" or "pool". */ 1230 m->oflags = VPO_UNMANAGED; 1231 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1232 /* Fictitious pages are unevictable. */ 1233 m->ref_count = 1; 1234 pmap_page_init(m); 1235 memattr: 1236 pmap_page_set_memattr(m, memattr); 1237 } 1238 1239 /* 1240 * vm_page_putfake: 1241 * 1242 * Release a fictitious page. 1243 */ 1244 void 1245 vm_page_putfake(vm_page_t m) 1246 { 1247 1248 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1249 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1250 ("vm_page_putfake: bad page %p", m)); 1251 if (vm_page_xbusied(m)) 1252 vm_page_xunbusy(m); 1253 uma_zfree(fakepg_zone, m); 1254 } 1255 1256 /* 1257 * vm_page_updatefake: 1258 * 1259 * Update the given fictitious page to the specified physical address and 1260 * memory attribute. 1261 */ 1262 void 1263 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1264 { 1265 1266 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1267 ("vm_page_updatefake: bad page %p", m)); 1268 m->phys_addr = paddr; 1269 pmap_page_set_memattr(m, memattr); 1270 } 1271 1272 /* 1273 * vm_page_free: 1274 * 1275 * Free a page. 1276 */ 1277 void 1278 vm_page_free(vm_page_t m) 1279 { 1280 1281 m->flags &= ~PG_ZERO; 1282 vm_page_free_toq(m); 1283 } 1284 1285 /* 1286 * vm_page_free_zero: 1287 * 1288 * Free a page to the zerod-pages queue 1289 */ 1290 void 1291 vm_page_free_zero(vm_page_t m) 1292 { 1293 1294 m->flags |= PG_ZERO; 1295 vm_page_free_toq(m); 1296 } 1297 1298 /* 1299 * Unbusy and handle the page queueing for a page from a getpages request that 1300 * was optionally read ahead or behind. 1301 */ 1302 void 1303 vm_page_readahead_finish(vm_page_t m) 1304 { 1305 1306 /* We shouldn't put invalid pages on queues. */ 1307 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1308 1309 /* 1310 * Since the page is not the actually needed one, whether it should 1311 * be activated or deactivated is not obvious. Empirical results 1312 * have shown that deactivating the page is usually the best choice, 1313 * unless the page is wanted by another thread. 1314 */ 1315 vm_page_lock(m); 1316 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1317 vm_page_activate(m); 1318 else 1319 vm_page_deactivate(m); 1320 vm_page_unlock(m); 1321 vm_page_xunbusy(m); 1322 } 1323 1324 /* 1325 * vm_page_sleep_if_busy: 1326 * 1327 * Sleep and release the object lock if the page is busied. 1328 * Returns TRUE if the thread slept. 1329 * 1330 * The given page must be unlocked and object containing it must 1331 * be locked. 1332 */ 1333 int 1334 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1335 { 1336 vm_object_t obj; 1337 1338 vm_page_lock_assert(m, MA_NOTOWNED); 1339 VM_OBJECT_ASSERT_WLOCKED(m->object); 1340 1341 /* 1342 * The page-specific object must be cached because page 1343 * identity can change during the sleep, causing the 1344 * re-lock of a different object. 1345 * It is assumed that a reference to the object is already 1346 * held by the callers. 1347 */ 1348 obj = m->object; 1349 if (vm_page_busied(m) || (obj != NULL && obj->busy)) { 1350 vm_page_busy_sleep(m, msg, false); 1351 VM_OBJECT_WLOCK(obj); 1352 return (TRUE); 1353 } 1354 return (FALSE); 1355 } 1356 1357 /* 1358 * vm_page_sleep_if_xbusy: 1359 * 1360 * Sleep and release the object lock if the page is xbusied. 1361 * Returns TRUE if the thread slept. 1362 * 1363 * The given page must be unlocked and object containing it must 1364 * be locked. 1365 */ 1366 int 1367 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg) 1368 { 1369 vm_object_t obj; 1370 1371 vm_page_lock_assert(m, MA_NOTOWNED); 1372 VM_OBJECT_ASSERT_WLOCKED(m->object); 1373 1374 /* 1375 * The page-specific object must be cached because page 1376 * identity can change during the sleep, causing the 1377 * re-lock of a different object. 1378 * It is assumed that a reference to the object is already 1379 * held by the callers. 1380 */ 1381 obj = m->object; 1382 if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) { 1383 vm_page_busy_sleep(m, msg, true); 1384 VM_OBJECT_WLOCK(obj); 1385 return (TRUE); 1386 } 1387 return (FALSE); 1388 } 1389 1390 /* 1391 * vm_page_dirty_KBI: [ internal use only ] 1392 * 1393 * Set all bits in the page's dirty field. 1394 * 1395 * The object containing the specified page must be locked if the 1396 * call is made from the machine-independent layer. 1397 * 1398 * See vm_page_clear_dirty_mask(). 1399 * 1400 * This function should only be called by vm_page_dirty(). 1401 */ 1402 void 1403 vm_page_dirty_KBI(vm_page_t m) 1404 { 1405 1406 /* Refer to this operation by its public name. */ 1407 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1408 m->dirty = VM_PAGE_BITS_ALL; 1409 } 1410 1411 /* 1412 * vm_page_insert: [ internal use only ] 1413 * 1414 * Inserts the given mem entry into the object and object list. 1415 * 1416 * The object must be locked. 1417 */ 1418 int 1419 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1420 { 1421 vm_page_t mpred; 1422 1423 VM_OBJECT_ASSERT_WLOCKED(object); 1424 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1425 return (vm_page_insert_after(m, object, pindex, mpred)); 1426 } 1427 1428 /* 1429 * vm_page_insert_after: 1430 * 1431 * Inserts the page "m" into the specified object at offset "pindex". 1432 * 1433 * The page "mpred" must immediately precede the offset "pindex" within 1434 * the specified object. 1435 * 1436 * The object must be locked. 1437 */ 1438 static int 1439 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1440 vm_page_t mpred) 1441 { 1442 vm_page_t msucc; 1443 1444 VM_OBJECT_ASSERT_WLOCKED(object); 1445 KASSERT(m->object == NULL, 1446 ("vm_page_insert_after: page already inserted")); 1447 if (mpred != NULL) { 1448 KASSERT(mpred->object == object, 1449 ("vm_page_insert_after: object doesn't contain mpred")); 1450 KASSERT(mpred->pindex < pindex, 1451 ("vm_page_insert_after: mpred doesn't precede pindex")); 1452 msucc = TAILQ_NEXT(mpred, listq); 1453 } else 1454 msucc = TAILQ_FIRST(&object->memq); 1455 if (msucc != NULL) 1456 KASSERT(msucc->pindex > pindex, 1457 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1458 1459 /* 1460 * Record the object/offset pair in this page. 1461 */ 1462 m->object = object; 1463 m->pindex = pindex; 1464 m->ref_count |= VPRC_OBJREF; 1465 1466 /* 1467 * Now link into the object's ordered list of backed pages. 1468 */ 1469 if (vm_radix_insert(&object->rtree, m)) { 1470 m->object = NULL; 1471 m->pindex = 0; 1472 m->ref_count &= ~VPRC_OBJREF; 1473 return (1); 1474 } 1475 vm_page_insert_radixdone(m, object, mpred); 1476 return (0); 1477 } 1478 1479 /* 1480 * vm_page_insert_radixdone: 1481 * 1482 * Complete page "m" insertion into the specified object after the 1483 * radix trie hooking. 1484 * 1485 * The page "mpred" must precede the offset "m->pindex" within the 1486 * specified object. 1487 * 1488 * The object must be locked. 1489 */ 1490 static void 1491 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1492 { 1493 1494 VM_OBJECT_ASSERT_WLOCKED(object); 1495 KASSERT(object != NULL && m->object == object, 1496 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1497 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1498 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1499 if (mpred != NULL) { 1500 KASSERT(mpred->object == object, 1501 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1502 KASSERT(mpred->pindex < m->pindex, 1503 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1504 } 1505 1506 if (mpred != NULL) 1507 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1508 else 1509 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1510 1511 /* 1512 * Show that the object has one more resident page. 1513 */ 1514 object->resident_page_count++; 1515 1516 /* 1517 * Hold the vnode until the last page is released. 1518 */ 1519 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1520 vhold(object->handle); 1521 1522 /* 1523 * Since we are inserting a new and possibly dirty page, 1524 * update the object's generation count. 1525 */ 1526 if (pmap_page_is_write_mapped(m)) 1527 vm_object_set_writeable_dirty(object); 1528 } 1529 1530 /* 1531 * Do the work to remove a page from its object. The caller is responsible for 1532 * updating the page's fields to reflect this removal. 1533 */ 1534 static void 1535 vm_page_object_remove(vm_page_t m) 1536 { 1537 vm_object_t object; 1538 vm_page_t mrem; 1539 1540 object = m->object; 1541 VM_OBJECT_ASSERT_WLOCKED(object); 1542 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1543 ("page %p is missing its object ref", m)); 1544 1545 mrem = vm_radix_remove(&object->rtree, m->pindex); 1546 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1547 1548 /* 1549 * Now remove from the object's list of backed pages. 1550 */ 1551 TAILQ_REMOVE(&object->memq, m, listq); 1552 1553 /* 1554 * And show that the object has one fewer resident page. 1555 */ 1556 object->resident_page_count--; 1557 1558 /* 1559 * The vnode may now be recycled. 1560 */ 1561 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1562 vdrop(object->handle); 1563 } 1564 1565 /* 1566 * vm_page_remove: 1567 * 1568 * Removes the specified page from its containing object, but does not 1569 * invalidate any backing storage. Returns true if the object's reference 1570 * was the last reference to the page, and false otherwise. 1571 * 1572 * The object must be locked. 1573 */ 1574 bool 1575 vm_page_remove(vm_page_t m) 1576 { 1577 1578 vm_page_object_remove(m); 1579 m->object = NULL; 1580 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1581 } 1582 1583 /* 1584 * vm_page_lookup: 1585 * 1586 * Returns the page associated with the object/offset 1587 * pair specified; if none is found, NULL is returned. 1588 * 1589 * The object must be locked. 1590 */ 1591 vm_page_t 1592 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1593 { 1594 1595 VM_OBJECT_ASSERT_LOCKED(object); 1596 return (vm_radix_lookup(&object->rtree, pindex)); 1597 } 1598 1599 /* 1600 * vm_page_find_least: 1601 * 1602 * Returns the page associated with the object with least pindex 1603 * greater than or equal to the parameter pindex, or NULL. 1604 * 1605 * The object must be locked. 1606 */ 1607 vm_page_t 1608 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1609 { 1610 vm_page_t m; 1611 1612 VM_OBJECT_ASSERT_LOCKED(object); 1613 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1614 m = vm_radix_lookup_ge(&object->rtree, pindex); 1615 return (m); 1616 } 1617 1618 /* 1619 * Returns the given page's successor (by pindex) within the object if it is 1620 * resident; if none is found, NULL is returned. 1621 * 1622 * The object must be locked. 1623 */ 1624 vm_page_t 1625 vm_page_next(vm_page_t m) 1626 { 1627 vm_page_t next; 1628 1629 VM_OBJECT_ASSERT_LOCKED(m->object); 1630 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1631 MPASS(next->object == m->object); 1632 if (next->pindex != m->pindex + 1) 1633 next = NULL; 1634 } 1635 return (next); 1636 } 1637 1638 /* 1639 * Returns the given page's predecessor (by pindex) within the object if it is 1640 * resident; if none is found, NULL is returned. 1641 * 1642 * The object must be locked. 1643 */ 1644 vm_page_t 1645 vm_page_prev(vm_page_t m) 1646 { 1647 vm_page_t prev; 1648 1649 VM_OBJECT_ASSERT_LOCKED(m->object); 1650 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1651 MPASS(prev->object == m->object); 1652 if (prev->pindex != m->pindex - 1) 1653 prev = NULL; 1654 } 1655 return (prev); 1656 } 1657 1658 /* 1659 * Uses the page mnew as a replacement for an existing page at index 1660 * pindex which must be already present in the object. 1661 */ 1662 vm_page_t 1663 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1664 { 1665 vm_page_t mold; 1666 1667 VM_OBJECT_ASSERT_WLOCKED(object); 1668 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1669 ("vm_page_replace: page %p already in object", mnew)); 1670 1671 /* 1672 * This function mostly follows vm_page_insert() and 1673 * vm_page_remove() without the radix, object count and vnode 1674 * dance. Double check such functions for more comments. 1675 */ 1676 1677 mnew->object = object; 1678 mnew->pindex = pindex; 1679 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1680 mold = vm_radix_replace(&object->rtree, mnew); 1681 KASSERT(mold->queue == PQ_NONE, 1682 ("vm_page_replace: old page %p is on a paging queue", mold)); 1683 1684 /* Keep the resident page list in sorted order. */ 1685 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1686 TAILQ_REMOVE(&object->memq, mold, listq); 1687 1688 mold->object = NULL; 1689 atomic_clear_int(&mold->ref_count, VPRC_OBJREF); 1690 vm_page_xunbusy(mold); 1691 1692 /* 1693 * The object's resident_page_count does not change because we have 1694 * swapped one page for another, but the generation count should 1695 * change if the page is dirty. 1696 */ 1697 if (pmap_page_is_write_mapped(mnew)) 1698 vm_object_set_writeable_dirty(object); 1699 return (mold); 1700 } 1701 1702 /* 1703 * vm_page_rename: 1704 * 1705 * Move the given memory entry from its 1706 * current object to the specified target object/offset. 1707 * 1708 * Note: swap associated with the page must be invalidated by the move. We 1709 * have to do this for several reasons: (1) we aren't freeing the 1710 * page, (2) we are dirtying the page, (3) the VM system is probably 1711 * moving the page from object A to B, and will then later move 1712 * the backing store from A to B and we can't have a conflict. 1713 * 1714 * Note: we *always* dirty the page. It is necessary both for the 1715 * fact that we moved it, and because we may be invalidating 1716 * swap. 1717 * 1718 * The objects must be locked. 1719 */ 1720 int 1721 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1722 { 1723 vm_page_t mpred; 1724 vm_pindex_t opidx; 1725 1726 VM_OBJECT_ASSERT_WLOCKED(new_object); 1727 1728 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1729 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1730 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1731 ("vm_page_rename: pindex already renamed")); 1732 1733 /* 1734 * Create a custom version of vm_page_insert() which does not depend 1735 * by m_prev and can cheat on the implementation aspects of the 1736 * function. 1737 */ 1738 opidx = m->pindex; 1739 m->pindex = new_pindex; 1740 if (vm_radix_insert(&new_object->rtree, m)) { 1741 m->pindex = opidx; 1742 return (1); 1743 } 1744 1745 /* 1746 * The operation cannot fail anymore. The removal must happen before 1747 * the listq iterator is tainted. 1748 */ 1749 m->pindex = opidx; 1750 vm_page_object_remove(m); 1751 1752 /* Return back to the new pindex to complete vm_page_insert(). */ 1753 m->pindex = new_pindex; 1754 m->object = new_object; 1755 1756 vm_page_insert_radixdone(m, new_object, mpred); 1757 vm_page_dirty(m); 1758 return (0); 1759 } 1760 1761 /* 1762 * vm_page_alloc: 1763 * 1764 * Allocate and return a page that is associated with the specified 1765 * object and offset pair. By default, this page is exclusive busied. 1766 * 1767 * The caller must always specify an allocation class. 1768 * 1769 * allocation classes: 1770 * VM_ALLOC_NORMAL normal process request 1771 * VM_ALLOC_SYSTEM system *really* needs a page 1772 * VM_ALLOC_INTERRUPT interrupt time request 1773 * 1774 * optional allocation flags: 1775 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1776 * intends to allocate 1777 * VM_ALLOC_NOBUSY do not exclusive busy the page 1778 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1779 * VM_ALLOC_NOOBJ page is not associated with an object and 1780 * should not be exclusive busy 1781 * VM_ALLOC_SBUSY shared busy the allocated page 1782 * VM_ALLOC_WIRED wire the allocated page 1783 * VM_ALLOC_ZERO prefer a zeroed page 1784 */ 1785 vm_page_t 1786 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1787 { 1788 1789 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1790 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1791 } 1792 1793 vm_page_t 1794 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1795 int req) 1796 { 1797 1798 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1799 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1800 NULL)); 1801 } 1802 1803 /* 1804 * Allocate a page in the specified object with the given page index. To 1805 * optimize insertion of the page into the object, the caller must also specifiy 1806 * the resident page in the object with largest index smaller than the given 1807 * page index, or NULL if no such page exists. 1808 */ 1809 vm_page_t 1810 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1811 int req, vm_page_t mpred) 1812 { 1813 struct vm_domainset_iter di; 1814 vm_page_t m; 1815 int domain; 1816 1817 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1818 do { 1819 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1820 mpred); 1821 if (m != NULL) 1822 break; 1823 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1824 1825 return (m); 1826 } 1827 1828 /* 1829 * Returns true if the number of free pages exceeds the minimum 1830 * for the request class and false otherwise. 1831 */ 1832 int 1833 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 1834 { 1835 u_int limit, old, new; 1836 1837 req = req & VM_ALLOC_CLASS_MASK; 1838 1839 /* 1840 * The page daemon is allowed to dig deeper into the free page list. 1841 */ 1842 if (curproc == pageproc && req != VM_ALLOC_INTERRUPT) 1843 req = VM_ALLOC_SYSTEM; 1844 if (req == VM_ALLOC_INTERRUPT) 1845 limit = 0; 1846 else if (req == VM_ALLOC_SYSTEM) 1847 limit = vmd->vmd_interrupt_free_min; 1848 else 1849 limit = vmd->vmd_free_reserved; 1850 1851 /* 1852 * Attempt to reserve the pages. Fail if we're below the limit. 1853 */ 1854 limit += npages; 1855 old = vmd->vmd_free_count; 1856 do { 1857 if (old < limit) 1858 return (0); 1859 new = old - npages; 1860 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 1861 1862 /* Wake the page daemon if we've crossed the threshold. */ 1863 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 1864 pagedaemon_wakeup(vmd->vmd_domain); 1865 1866 /* Only update bitsets on transitions. */ 1867 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 1868 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 1869 vm_domain_set(vmd); 1870 1871 return (1); 1872 } 1873 1874 vm_page_t 1875 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 1876 int req, vm_page_t mpred) 1877 { 1878 struct vm_domain *vmd; 1879 vm_page_t m; 1880 int flags, pool; 1881 1882 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1883 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1884 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1885 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1886 ("inconsistent object(%p)/req(%x)", object, req)); 1887 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1888 ("Can't sleep and retry object insertion.")); 1889 KASSERT(mpred == NULL || mpred->pindex < pindex, 1890 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1891 (uintmax_t)pindex)); 1892 if (object != NULL) 1893 VM_OBJECT_ASSERT_WLOCKED(object); 1894 1895 flags = 0; 1896 m = NULL; 1897 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; 1898 again: 1899 #if VM_NRESERVLEVEL > 0 1900 /* 1901 * Can we allocate the page from a reservation? 1902 */ 1903 if (vm_object_reserv(object) && 1904 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 1905 NULL) { 1906 domain = vm_phys_domain(m); 1907 vmd = VM_DOMAIN(domain); 1908 goto found; 1909 } 1910 #endif 1911 vmd = VM_DOMAIN(domain); 1912 if (vmd->vmd_pgcache[pool].zone != NULL) { 1913 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT); 1914 if (m != NULL) { 1915 flags |= PG_PCPU_CACHE; 1916 goto found; 1917 } 1918 } 1919 if (vm_domain_allocate(vmd, req, 1)) { 1920 /* 1921 * If not, allocate it from the free page queues. 1922 */ 1923 vm_domain_free_lock(vmd); 1924 m = vm_phys_alloc_pages(domain, pool, 0); 1925 vm_domain_free_unlock(vmd); 1926 if (m == NULL) { 1927 vm_domain_freecnt_inc(vmd, 1); 1928 #if VM_NRESERVLEVEL > 0 1929 if (vm_reserv_reclaim_inactive(domain)) 1930 goto again; 1931 #endif 1932 } 1933 } 1934 if (m == NULL) { 1935 /* 1936 * Not allocatable, give up. 1937 */ 1938 if (vm_domain_alloc_fail(vmd, object, req)) 1939 goto again; 1940 return (NULL); 1941 } 1942 1943 /* 1944 * At this point we had better have found a good page. 1945 */ 1946 found: 1947 vm_page_dequeue(m); 1948 vm_page_alloc_check(m); 1949 1950 /* 1951 * Initialize the page. Only the PG_ZERO flag is inherited. 1952 */ 1953 if ((req & VM_ALLOC_ZERO) != 0) 1954 flags |= (m->flags & PG_ZERO); 1955 if ((req & VM_ALLOC_NODUMP) != 0) 1956 flags |= PG_NODUMP; 1957 m->flags = flags; 1958 m->aflags = 0; 1959 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1960 VPO_UNMANAGED : 0; 1961 m->busy_lock = VPB_UNBUSIED; 1962 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1963 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1964 if ((req & VM_ALLOC_SBUSY) != 0) 1965 m->busy_lock = VPB_SHARERS_WORD(1); 1966 if (req & VM_ALLOC_WIRED) { 1967 /* 1968 * The page lock is not required for wiring a page until that 1969 * page is inserted into the object. 1970 */ 1971 vm_wire_add(1); 1972 m->ref_count = 1; 1973 } 1974 m->act_count = 0; 1975 1976 if (object != NULL) { 1977 if (vm_page_insert_after(m, object, pindex, mpred)) { 1978 if (req & VM_ALLOC_WIRED) { 1979 vm_wire_sub(1); 1980 m->ref_count = 0; 1981 } 1982 KASSERT(m->object == NULL, ("page %p has object", m)); 1983 m->oflags = VPO_UNMANAGED; 1984 m->busy_lock = VPB_UNBUSIED; 1985 /* Don't change PG_ZERO. */ 1986 vm_page_free_toq(m); 1987 if (req & VM_ALLOC_WAITFAIL) { 1988 VM_OBJECT_WUNLOCK(object); 1989 vm_radix_wait(); 1990 VM_OBJECT_WLOCK(object); 1991 } 1992 return (NULL); 1993 } 1994 1995 /* Ignore device objects; the pager sets "memattr" for them. */ 1996 if (object->memattr != VM_MEMATTR_DEFAULT && 1997 (object->flags & OBJ_FICTITIOUS) == 0) 1998 pmap_page_set_memattr(m, object->memattr); 1999 } else 2000 m->pindex = pindex; 2001 2002 return (m); 2003 } 2004 2005 /* 2006 * vm_page_alloc_contig: 2007 * 2008 * Allocate a contiguous set of physical pages of the given size "npages" 2009 * from the free lists. All of the physical pages must be at or above 2010 * the given physical address "low" and below the given physical address 2011 * "high". The given value "alignment" determines the alignment of the 2012 * first physical page in the set. If the given value "boundary" is 2013 * non-zero, then the set of physical pages cannot cross any physical 2014 * address boundary that is a multiple of that value. Both "alignment" 2015 * and "boundary" must be a power of two. 2016 * 2017 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2018 * then the memory attribute setting for the physical pages is configured 2019 * to the object's memory attribute setting. Otherwise, the memory 2020 * attribute setting for the physical pages is configured to "memattr", 2021 * overriding the object's memory attribute setting. However, if the 2022 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2023 * memory attribute setting for the physical pages cannot be configured 2024 * to VM_MEMATTR_DEFAULT. 2025 * 2026 * The specified object may not contain fictitious pages. 2027 * 2028 * The caller must always specify an allocation class. 2029 * 2030 * allocation classes: 2031 * VM_ALLOC_NORMAL normal process request 2032 * VM_ALLOC_SYSTEM system *really* needs a page 2033 * VM_ALLOC_INTERRUPT interrupt time request 2034 * 2035 * optional allocation flags: 2036 * VM_ALLOC_NOBUSY do not exclusive busy the page 2037 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2038 * VM_ALLOC_NOOBJ page is not associated with an object and 2039 * should not be exclusive busy 2040 * VM_ALLOC_SBUSY shared busy the allocated page 2041 * VM_ALLOC_WIRED wire the allocated page 2042 * VM_ALLOC_ZERO prefer a zeroed page 2043 */ 2044 vm_page_t 2045 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2046 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2047 vm_paddr_t boundary, vm_memattr_t memattr) 2048 { 2049 struct vm_domainset_iter di; 2050 vm_page_t m; 2051 int domain; 2052 2053 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2054 do { 2055 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2056 npages, low, high, alignment, boundary, memattr); 2057 if (m != NULL) 2058 break; 2059 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2060 2061 return (m); 2062 } 2063 2064 vm_page_t 2065 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2066 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2067 vm_paddr_t boundary, vm_memattr_t memattr) 2068 { 2069 struct vm_domain *vmd; 2070 vm_page_t m, m_ret, mpred; 2071 u_int busy_lock, flags, oflags; 2072 2073 mpred = NULL; /* XXX: pacify gcc */ 2074 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2075 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2076 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2077 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2078 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 2079 req)); 2080 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2081 ("Can't sleep and retry object insertion.")); 2082 if (object != NULL) { 2083 VM_OBJECT_ASSERT_WLOCKED(object); 2084 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2085 ("vm_page_alloc_contig: object %p has fictitious pages", 2086 object)); 2087 } 2088 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2089 2090 if (object != NULL) { 2091 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2092 KASSERT(mpred == NULL || mpred->pindex != pindex, 2093 ("vm_page_alloc_contig: pindex already allocated")); 2094 } 2095 2096 /* 2097 * Can we allocate the pages without the number of free pages falling 2098 * below the lower bound for the allocation class? 2099 */ 2100 m_ret = NULL; 2101 again: 2102 #if VM_NRESERVLEVEL > 0 2103 /* 2104 * Can we allocate the pages from a reservation? 2105 */ 2106 if (vm_object_reserv(object) && 2107 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2108 mpred, npages, low, high, alignment, boundary)) != NULL) { 2109 domain = vm_phys_domain(m_ret); 2110 vmd = VM_DOMAIN(domain); 2111 goto found; 2112 } 2113 #endif 2114 vmd = VM_DOMAIN(domain); 2115 if (vm_domain_allocate(vmd, req, npages)) { 2116 /* 2117 * allocate them from the free page queues. 2118 */ 2119 vm_domain_free_lock(vmd); 2120 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2121 alignment, boundary); 2122 vm_domain_free_unlock(vmd); 2123 if (m_ret == NULL) { 2124 vm_domain_freecnt_inc(vmd, npages); 2125 #if VM_NRESERVLEVEL > 0 2126 if (vm_reserv_reclaim_contig(domain, npages, low, 2127 high, alignment, boundary)) 2128 goto again; 2129 #endif 2130 } 2131 } 2132 if (m_ret == NULL) { 2133 if (vm_domain_alloc_fail(vmd, object, req)) 2134 goto again; 2135 return (NULL); 2136 } 2137 #if VM_NRESERVLEVEL > 0 2138 found: 2139 #endif 2140 for (m = m_ret; m < &m_ret[npages]; m++) { 2141 vm_page_dequeue(m); 2142 vm_page_alloc_check(m); 2143 } 2144 2145 /* 2146 * Initialize the pages. Only the PG_ZERO flag is inherited. 2147 */ 2148 flags = 0; 2149 if ((req & VM_ALLOC_ZERO) != 0) 2150 flags = PG_ZERO; 2151 if ((req & VM_ALLOC_NODUMP) != 0) 2152 flags |= PG_NODUMP; 2153 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2154 VPO_UNMANAGED : 0; 2155 busy_lock = VPB_UNBUSIED; 2156 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2157 busy_lock = VPB_SINGLE_EXCLUSIVER; 2158 if ((req & VM_ALLOC_SBUSY) != 0) 2159 busy_lock = VPB_SHARERS_WORD(1); 2160 if ((req & VM_ALLOC_WIRED) != 0) 2161 vm_wire_add(npages); 2162 if (object != NULL) { 2163 if (object->memattr != VM_MEMATTR_DEFAULT && 2164 memattr == VM_MEMATTR_DEFAULT) 2165 memattr = object->memattr; 2166 } 2167 for (m = m_ret; m < &m_ret[npages]; m++) { 2168 m->aflags = 0; 2169 m->flags = (m->flags | PG_NODUMP) & flags; 2170 m->busy_lock = busy_lock; 2171 if ((req & VM_ALLOC_WIRED) != 0) 2172 m->ref_count = 1; 2173 m->act_count = 0; 2174 m->oflags = oflags; 2175 if (object != NULL) { 2176 if (vm_page_insert_after(m, object, pindex, mpred)) { 2177 if ((req & VM_ALLOC_WIRED) != 0) 2178 vm_wire_sub(npages); 2179 KASSERT(m->object == NULL, 2180 ("page %p has object", m)); 2181 mpred = m; 2182 for (m = m_ret; m < &m_ret[npages]; m++) { 2183 if (m <= mpred && 2184 (req & VM_ALLOC_WIRED) != 0) 2185 m->ref_count = 0; 2186 m->oflags = VPO_UNMANAGED; 2187 m->busy_lock = VPB_UNBUSIED; 2188 /* Don't change PG_ZERO. */ 2189 vm_page_free_toq(m); 2190 } 2191 if (req & VM_ALLOC_WAITFAIL) { 2192 VM_OBJECT_WUNLOCK(object); 2193 vm_radix_wait(); 2194 VM_OBJECT_WLOCK(object); 2195 } 2196 return (NULL); 2197 } 2198 mpred = m; 2199 } else 2200 m->pindex = pindex; 2201 if (memattr != VM_MEMATTR_DEFAULT) 2202 pmap_page_set_memattr(m, memattr); 2203 pindex++; 2204 } 2205 return (m_ret); 2206 } 2207 2208 /* 2209 * Check a page that has been freshly dequeued from a freelist. 2210 */ 2211 static void 2212 vm_page_alloc_check(vm_page_t m) 2213 { 2214 2215 KASSERT(m->object == NULL, ("page %p has object", m)); 2216 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 2217 ("page %p has unexpected queue %d, flags %#x", 2218 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK))); 2219 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2220 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 2221 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2222 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2223 ("page %p has unexpected memattr %d", 2224 m, pmap_page_get_memattr(m))); 2225 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2226 } 2227 2228 /* 2229 * vm_page_alloc_freelist: 2230 * 2231 * Allocate a physical page from the specified free page list. 2232 * 2233 * The caller must always specify an allocation class. 2234 * 2235 * allocation classes: 2236 * VM_ALLOC_NORMAL normal process request 2237 * VM_ALLOC_SYSTEM system *really* needs a page 2238 * VM_ALLOC_INTERRUPT interrupt time request 2239 * 2240 * optional allocation flags: 2241 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2242 * intends to allocate 2243 * VM_ALLOC_WIRED wire the allocated page 2244 * VM_ALLOC_ZERO prefer a zeroed page 2245 */ 2246 vm_page_t 2247 vm_page_alloc_freelist(int freelist, int req) 2248 { 2249 struct vm_domainset_iter di; 2250 vm_page_t m; 2251 int domain; 2252 2253 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2254 do { 2255 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2256 if (m != NULL) 2257 break; 2258 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2259 2260 return (m); 2261 } 2262 2263 vm_page_t 2264 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2265 { 2266 struct vm_domain *vmd; 2267 vm_page_t m; 2268 u_int flags; 2269 2270 m = NULL; 2271 vmd = VM_DOMAIN(domain); 2272 again: 2273 if (vm_domain_allocate(vmd, req, 1)) { 2274 vm_domain_free_lock(vmd); 2275 m = vm_phys_alloc_freelist_pages(domain, freelist, 2276 VM_FREEPOOL_DIRECT, 0); 2277 vm_domain_free_unlock(vmd); 2278 if (m == NULL) 2279 vm_domain_freecnt_inc(vmd, 1); 2280 } 2281 if (m == NULL) { 2282 if (vm_domain_alloc_fail(vmd, NULL, req)) 2283 goto again; 2284 return (NULL); 2285 } 2286 vm_page_dequeue(m); 2287 vm_page_alloc_check(m); 2288 2289 /* 2290 * Initialize the page. Only the PG_ZERO flag is inherited. 2291 */ 2292 m->aflags = 0; 2293 flags = 0; 2294 if ((req & VM_ALLOC_ZERO) != 0) 2295 flags = PG_ZERO; 2296 m->flags &= flags; 2297 if ((req & VM_ALLOC_WIRED) != 0) { 2298 /* 2299 * The page lock is not required for wiring a page that does 2300 * not belong to an object. 2301 */ 2302 vm_wire_add(1); 2303 m->ref_count = 1; 2304 } 2305 /* Unmanaged pages don't use "act_count". */ 2306 m->oflags = VPO_UNMANAGED; 2307 return (m); 2308 } 2309 2310 static int 2311 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2312 { 2313 struct vm_domain *vmd; 2314 struct vm_pgcache *pgcache; 2315 int i; 2316 2317 pgcache = arg; 2318 vmd = VM_DOMAIN(pgcache->domain); 2319 /* Only import if we can bring in a full bucket. */ 2320 if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2321 return (0); 2322 domain = vmd->vmd_domain; 2323 vm_domain_free_lock(vmd); 2324 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2325 (vm_page_t *)store); 2326 vm_domain_free_unlock(vmd); 2327 if (cnt != i) 2328 vm_domain_freecnt_inc(vmd, cnt - i); 2329 2330 return (i); 2331 } 2332 2333 static void 2334 vm_page_zone_release(void *arg, void **store, int cnt) 2335 { 2336 struct vm_domain *vmd; 2337 struct vm_pgcache *pgcache; 2338 vm_page_t m; 2339 int i; 2340 2341 pgcache = arg; 2342 vmd = VM_DOMAIN(pgcache->domain); 2343 vm_domain_free_lock(vmd); 2344 for (i = 0; i < cnt; i++) { 2345 m = (vm_page_t)store[i]; 2346 vm_phys_free_pages(m, 0); 2347 } 2348 vm_domain_free_unlock(vmd); 2349 vm_domain_freecnt_inc(vmd, cnt); 2350 } 2351 2352 #define VPSC_ANY 0 /* No restrictions. */ 2353 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2354 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2355 2356 /* 2357 * vm_page_scan_contig: 2358 * 2359 * Scan vm_page_array[] between the specified entries "m_start" and 2360 * "m_end" for a run of contiguous physical pages that satisfy the 2361 * specified conditions, and return the lowest page in the run. The 2362 * specified "alignment" determines the alignment of the lowest physical 2363 * page in the run. If the specified "boundary" is non-zero, then the 2364 * run of physical pages cannot span a physical address that is a 2365 * multiple of "boundary". 2366 * 2367 * "m_end" is never dereferenced, so it need not point to a vm_page 2368 * structure within vm_page_array[]. 2369 * 2370 * "npages" must be greater than zero. "m_start" and "m_end" must not 2371 * span a hole (or discontiguity) in the physical address space. Both 2372 * "alignment" and "boundary" must be a power of two. 2373 */ 2374 vm_page_t 2375 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2376 u_long alignment, vm_paddr_t boundary, int options) 2377 { 2378 struct mtx *m_mtx; 2379 vm_object_t object; 2380 vm_paddr_t pa; 2381 vm_page_t m, m_run; 2382 #if VM_NRESERVLEVEL > 0 2383 int level; 2384 #endif 2385 int m_inc, order, run_ext, run_len; 2386 2387 KASSERT(npages > 0, ("npages is 0")); 2388 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2389 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2390 m_run = NULL; 2391 run_len = 0; 2392 m_mtx = NULL; 2393 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2394 KASSERT((m->flags & PG_MARKER) == 0, 2395 ("page %p is PG_MARKER", m)); 2396 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2397 ("fictitious page %p has invalid ref count", m)); 2398 2399 /* 2400 * If the current page would be the start of a run, check its 2401 * physical address against the end, alignment, and boundary 2402 * conditions. If it doesn't satisfy these conditions, either 2403 * terminate the scan or advance to the next page that 2404 * satisfies the failed condition. 2405 */ 2406 if (run_len == 0) { 2407 KASSERT(m_run == NULL, ("m_run != NULL")); 2408 if (m + npages > m_end) 2409 break; 2410 pa = VM_PAGE_TO_PHYS(m); 2411 if ((pa & (alignment - 1)) != 0) { 2412 m_inc = atop(roundup2(pa, alignment) - pa); 2413 continue; 2414 } 2415 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2416 boundary) != 0) { 2417 m_inc = atop(roundup2(pa, boundary) - pa); 2418 continue; 2419 } 2420 } else 2421 KASSERT(m_run != NULL, ("m_run == NULL")); 2422 2423 vm_page_change_lock(m, &m_mtx); 2424 m_inc = 1; 2425 retry: 2426 if (vm_page_wired(m)) 2427 run_ext = 0; 2428 #if VM_NRESERVLEVEL > 0 2429 else if ((level = vm_reserv_level(m)) >= 0 && 2430 (options & VPSC_NORESERV) != 0) { 2431 run_ext = 0; 2432 /* Advance to the end of the reservation. */ 2433 pa = VM_PAGE_TO_PHYS(m); 2434 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2435 pa); 2436 } 2437 #endif 2438 else if ((object = m->object) != NULL) { 2439 /* 2440 * The page is considered eligible for relocation if 2441 * and only if it could be laundered or reclaimed by 2442 * the page daemon. 2443 */ 2444 if (!VM_OBJECT_TRYRLOCK(object)) { 2445 mtx_unlock(m_mtx); 2446 VM_OBJECT_RLOCK(object); 2447 mtx_lock(m_mtx); 2448 if (m->object != object) { 2449 /* 2450 * The page may have been freed. 2451 */ 2452 VM_OBJECT_RUNLOCK(object); 2453 goto retry; 2454 } 2455 } 2456 /* Don't care: PG_NODUMP, PG_ZERO. */ 2457 if (object->type != OBJT_DEFAULT && 2458 object->type != OBJT_SWAP && 2459 object->type != OBJT_VNODE) { 2460 run_ext = 0; 2461 #if VM_NRESERVLEVEL > 0 2462 } else if ((options & VPSC_NOSUPER) != 0 && 2463 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2464 run_ext = 0; 2465 /* Advance to the end of the superpage. */ 2466 pa = VM_PAGE_TO_PHYS(m); 2467 m_inc = atop(roundup2(pa + 1, 2468 vm_reserv_size(level)) - pa); 2469 #endif 2470 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2471 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) && 2472 !vm_page_wired(m)) { 2473 /* 2474 * The page is allocated but eligible for 2475 * relocation. Extend the current run by one 2476 * page. 2477 */ 2478 KASSERT(pmap_page_get_memattr(m) == 2479 VM_MEMATTR_DEFAULT, 2480 ("page %p has an unexpected memattr", m)); 2481 KASSERT((m->oflags & (VPO_SWAPINPROG | 2482 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2483 ("page %p has unexpected oflags", m)); 2484 /* Don't care: PGA_NOSYNC. */ 2485 run_ext = 1; 2486 } else 2487 run_ext = 0; 2488 VM_OBJECT_RUNLOCK(object); 2489 #if VM_NRESERVLEVEL > 0 2490 } else if (level >= 0) { 2491 /* 2492 * The page is reserved but not yet allocated. In 2493 * other words, it is still free. Extend the current 2494 * run by one page. 2495 */ 2496 run_ext = 1; 2497 #endif 2498 } else if ((order = m->order) < VM_NFREEORDER) { 2499 /* 2500 * The page is enqueued in the physical memory 2501 * allocator's free page queues. Moreover, it is the 2502 * first page in a power-of-two-sized run of 2503 * contiguous free pages. Add these pages to the end 2504 * of the current run, and jump ahead. 2505 */ 2506 run_ext = 1 << order; 2507 m_inc = 1 << order; 2508 } else { 2509 /* 2510 * Skip the page for one of the following reasons: (1) 2511 * It is enqueued in the physical memory allocator's 2512 * free page queues. However, it is not the first 2513 * page in a run of contiguous free pages. (This case 2514 * rarely occurs because the scan is performed in 2515 * ascending order.) (2) It is not reserved, and it is 2516 * transitioning from free to allocated. (Conversely, 2517 * the transition from allocated to free for managed 2518 * pages is blocked by the page lock.) (3) It is 2519 * allocated but not contained by an object and not 2520 * wired, e.g., allocated by Xen's balloon driver. 2521 */ 2522 run_ext = 0; 2523 } 2524 2525 /* 2526 * Extend or reset the current run of pages. 2527 */ 2528 if (run_ext > 0) { 2529 if (run_len == 0) 2530 m_run = m; 2531 run_len += run_ext; 2532 } else { 2533 if (run_len > 0) { 2534 m_run = NULL; 2535 run_len = 0; 2536 } 2537 } 2538 } 2539 if (m_mtx != NULL) 2540 mtx_unlock(m_mtx); 2541 if (run_len >= npages) 2542 return (m_run); 2543 return (NULL); 2544 } 2545 2546 /* 2547 * vm_page_reclaim_run: 2548 * 2549 * Try to relocate each of the allocated virtual pages within the 2550 * specified run of physical pages to a new physical address. Free the 2551 * physical pages underlying the relocated virtual pages. A virtual page 2552 * is relocatable if and only if it could be laundered or reclaimed by 2553 * the page daemon. Whenever possible, a virtual page is relocated to a 2554 * physical address above "high". 2555 * 2556 * Returns 0 if every physical page within the run was already free or 2557 * just freed by a successful relocation. Otherwise, returns a non-zero 2558 * value indicating why the last attempt to relocate a virtual page was 2559 * unsuccessful. 2560 * 2561 * "req_class" must be an allocation class. 2562 */ 2563 static int 2564 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2565 vm_paddr_t high) 2566 { 2567 struct vm_domain *vmd; 2568 struct mtx *m_mtx; 2569 struct spglist free; 2570 vm_object_t object; 2571 vm_paddr_t pa; 2572 vm_page_t m, m_end, m_new; 2573 int error, order, req; 2574 2575 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2576 ("req_class is not an allocation class")); 2577 SLIST_INIT(&free); 2578 error = 0; 2579 m = m_run; 2580 m_end = m_run + npages; 2581 m_mtx = NULL; 2582 for (; error == 0 && m < m_end; m++) { 2583 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2584 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2585 2586 /* 2587 * Avoid releasing and reacquiring the same page lock. 2588 */ 2589 vm_page_change_lock(m, &m_mtx); 2590 retry: 2591 /* 2592 * Racily check for wirings. Races are handled below. 2593 */ 2594 if (vm_page_wired(m)) 2595 error = EBUSY; 2596 else if ((object = m->object) != NULL) { 2597 /* 2598 * The page is relocated if and only if it could be 2599 * laundered or reclaimed by the page daemon. 2600 */ 2601 if (!VM_OBJECT_TRYWLOCK(object)) { 2602 mtx_unlock(m_mtx); 2603 VM_OBJECT_WLOCK(object); 2604 mtx_lock(m_mtx); 2605 if (m->object != object) { 2606 /* 2607 * The page may have been freed. 2608 */ 2609 VM_OBJECT_WUNLOCK(object); 2610 goto retry; 2611 } 2612 } 2613 /* Don't care: PG_NODUMP, PG_ZERO. */ 2614 if (object->type != OBJT_DEFAULT && 2615 object->type != OBJT_SWAP && 2616 object->type != OBJT_VNODE) 2617 error = EINVAL; 2618 else if (object->memattr != VM_MEMATTR_DEFAULT) 2619 error = EINVAL; 2620 else if (vm_page_queue(m) != PQ_NONE && 2621 vm_page_tryxbusy(m) != 0) { 2622 if (vm_page_wired(m)) { 2623 vm_page_xunbusy(m); 2624 error = EBUSY; 2625 goto unlock; 2626 } 2627 KASSERT(pmap_page_get_memattr(m) == 2628 VM_MEMATTR_DEFAULT, 2629 ("page %p has an unexpected memattr", m)); 2630 KASSERT((m->oflags & (VPO_SWAPINPROG | 2631 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2632 ("page %p has unexpected oflags", m)); 2633 /* Don't care: PGA_NOSYNC. */ 2634 if (!vm_page_none_valid(m)) { 2635 /* 2636 * First, try to allocate a new page 2637 * that is above "high". Failing 2638 * that, try to allocate a new page 2639 * that is below "m_run". Allocate 2640 * the new page between the end of 2641 * "m_run" and "high" only as a last 2642 * resort. 2643 */ 2644 req = req_class | VM_ALLOC_NOOBJ; 2645 if ((m->flags & PG_NODUMP) != 0) 2646 req |= VM_ALLOC_NODUMP; 2647 if (trunc_page(high) != 2648 ~(vm_paddr_t)PAGE_MASK) { 2649 m_new = vm_page_alloc_contig( 2650 NULL, 0, req, 1, 2651 round_page(high), 2652 ~(vm_paddr_t)0, 2653 PAGE_SIZE, 0, 2654 VM_MEMATTR_DEFAULT); 2655 } else 2656 m_new = NULL; 2657 if (m_new == NULL) { 2658 pa = VM_PAGE_TO_PHYS(m_run); 2659 m_new = vm_page_alloc_contig( 2660 NULL, 0, req, 1, 2661 0, pa - 1, PAGE_SIZE, 0, 2662 VM_MEMATTR_DEFAULT); 2663 } 2664 if (m_new == NULL) { 2665 pa += ptoa(npages); 2666 m_new = vm_page_alloc_contig( 2667 NULL, 0, req, 1, 2668 pa, high, PAGE_SIZE, 0, 2669 VM_MEMATTR_DEFAULT); 2670 } 2671 if (m_new == NULL) { 2672 vm_page_xunbusy(m); 2673 error = ENOMEM; 2674 goto unlock; 2675 } 2676 2677 /* 2678 * Unmap the page and check for new 2679 * wirings that may have been acquired 2680 * through a pmap lookup. 2681 */ 2682 if (object->ref_count != 0 && 2683 !vm_page_try_remove_all(m)) { 2684 vm_page_free(m_new); 2685 error = EBUSY; 2686 goto unlock; 2687 } 2688 2689 /* 2690 * Replace "m" with the new page. For 2691 * vm_page_replace(), "m" must be busy 2692 * and dequeued. Finally, change "m" 2693 * as if vm_page_free() was called. 2694 */ 2695 m_new->aflags = m->aflags & 2696 ~PGA_QUEUE_STATE_MASK; 2697 KASSERT(m_new->oflags == VPO_UNMANAGED, 2698 ("page %p is managed", m_new)); 2699 pmap_copy_page(m, m_new); 2700 m_new->valid = m->valid; 2701 m_new->dirty = m->dirty; 2702 m->flags &= ~PG_ZERO; 2703 vm_page_dequeue(m); 2704 vm_page_replace_checked(m_new, object, 2705 m->pindex, m); 2706 if (vm_page_free_prep(m)) 2707 SLIST_INSERT_HEAD(&free, m, 2708 plinks.s.ss); 2709 2710 /* 2711 * The new page must be deactivated 2712 * before the object is unlocked. 2713 */ 2714 vm_page_change_lock(m_new, &m_mtx); 2715 vm_page_deactivate(m_new); 2716 } else { 2717 m->flags &= ~PG_ZERO; 2718 vm_page_dequeue(m); 2719 if (vm_page_free_prep(m)) 2720 SLIST_INSERT_HEAD(&free, m, 2721 plinks.s.ss); 2722 KASSERT(m->dirty == 0, 2723 ("page %p is dirty", m)); 2724 } 2725 } else 2726 error = EBUSY; 2727 unlock: 2728 VM_OBJECT_WUNLOCK(object); 2729 } else { 2730 MPASS(vm_phys_domain(m) == domain); 2731 vmd = VM_DOMAIN(domain); 2732 vm_domain_free_lock(vmd); 2733 order = m->order; 2734 if (order < VM_NFREEORDER) { 2735 /* 2736 * The page is enqueued in the physical memory 2737 * allocator's free page queues. Moreover, it 2738 * is the first page in a power-of-two-sized 2739 * run of contiguous free pages. Jump ahead 2740 * to the last page within that run, and 2741 * continue from there. 2742 */ 2743 m += (1 << order) - 1; 2744 } 2745 #if VM_NRESERVLEVEL > 0 2746 else if (vm_reserv_is_page_free(m)) 2747 order = 0; 2748 #endif 2749 vm_domain_free_unlock(vmd); 2750 if (order == VM_NFREEORDER) 2751 error = EINVAL; 2752 } 2753 } 2754 if (m_mtx != NULL) 2755 mtx_unlock(m_mtx); 2756 if ((m = SLIST_FIRST(&free)) != NULL) { 2757 int cnt; 2758 2759 vmd = VM_DOMAIN(domain); 2760 cnt = 0; 2761 vm_domain_free_lock(vmd); 2762 do { 2763 MPASS(vm_phys_domain(m) == domain); 2764 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2765 vm_phys_free_pages(m, 0); 2766 cnt++; 2767 } while ((m = SLIST_FIRST(&free)) != NULL); 2768 vm_domain_free_unlock(vmd); 2769 vm_domain_freecnt_inc(vmd, cnt); 2770 } 2771 return (error); 2772 } 2773 2774 #define NRUNS 16 2775 2776 CTASSERT(powerof2(NRUNS)); 2777 2778 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2779 2780 #define MIN_RECLAIM 8 2781 2782 /* 2783 * vm_page_reclaim_contig: 2784 * 2785 * Reclaim allocated, contiguous physical memory satisfying the specified 2786 * conditions by relocating the virtual pages using that physical memory. 2787 * Returns true if reclamation is successful and false otherwise. Since 2788 * relocation requires the allocation of physical pages, reclamation may 2789 * fail due to a shortage of free pages. When reclamation fails, callers 2790 * are expected to perform vm_wait() before retrying a failed allocation 2791 * operation, e.g., vm_page_alloc_contig(). 2792 * 2793 * The caller must always specify an allocation class through "req". 2794 * 2795 * allocation classes: 2796 * VM_ALLOC_NORMAL normal process request 2797 * VM_ALLOC_SYSTEM system *really* needs a page 2798 * VM_ALLOC_INTERRUPT interrupt time request 2799 * 2800 * The optional allocation flags are ignored. 2801 * 2802 * "npages" must be greater than zero. Both "alignment" and "boundary" 2803 * must be a power of two. 2804 */ 2805 bool 2806 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2807 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2808 { 2809 struct vm_domain *vmd; 2810 vm_paddr_t curr_low; 2811 vm_page_t m_run, m_runs[NRUNS]; 2812 u_long count, reclaimed; 2813 int error, i, options, req_class; 2814 2815 KASSERT(npages > 0, ("npages is 0")); 2816 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2817 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2818 req_class = req & VM_ALLOC_CLASS_MASK; 2819 2820 /* 2821 * The page daemon is allowed to dig deeper into the free page list. 2822 */ 2823 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2824 req_class = VM_ALLOC_SYSTEM; 2825 2826 /* 2827 * Return if the number of free pages cannot satisfy the requested 2828 * allocation. 2829 */ 2830 vmd = VM_DOMAIN(domain); 2831 count = vmd->vmd_free_count; 2832 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2833 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2834 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2835 return (false); 2836 2837 /* 2838 * Scan up to three times, relaxing the restrictions ("options") on 2839 * the reclamation of reservations and superpages each time. 2840 */ 2841 for (options = VPSC_NORESERV;;) { 2842 /* 2843 * Find the highest runs that satisfy the given constraints 2844 * and restrictions, and record them in "m_runs". 2845 */ 2846 curr_low = low; 2847 count = 0; 2848 for (;;) { 2849 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2850 high, alignment, boundary, options); 2851 if (m_run == NULL) 2852 break; 2853 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2854 m_runs[RUN_INDEX(count)] = m_run; 2855 count++; 2856 } 2857 2858 /* 2859 * Reclaim the highest runs in LIFO (descending) order until 2860 * the number of reclaimed pages, "reclaimed", is at least 2861 * MIN_RECLAIM. Reset "reclaimed" each time because each 2862 * reclamation is idempotent, and runs will (likely) recur 2863 * from one scan to the next as restrictions are relaxed. 2864 */ 2865 reclaimed = 0; 2866 for (i = 0; count > 0 && i < NRUNS; i++) { 2867 count--; 2868 m_run = m_runs[RUN_INDEX(count)]; 2869 error = vm_page_reclaim_run(req_class, domain, npages, 2870 m_run, high); 2871 if (error == 0) { 2872 reclaimed += npages; 2873 if (reclaimed >= MIN_RECLAIM) 2874 return (true); 2875 } 2876 } 2877 2878 /* 2879 * Either relax the restrictions on the next scan or return if 2880 * the last scan had no restrictions. 2881 */ 2882 if (options == VPSC_NORESERV) 2883 options = VPSC_NOSUPER; 2884 else if (options == VPSC_NOSUPER) 2885 options = VPSC_ANY; 2886 else if (options == VPSC_ANY) 2887 return (reclaimed != 0); 2888 } 2889 } 2890 2891 bool 2892 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2893 u_long alignment, vm_paddr_t boundary) 2894 { 2895 struct vm_domainset_iter di; 2896 int domain; 2897 bool ret; 2898 2899 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2900 do { 2901 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 2902 high, alignment, boundary); 2903 if (ret) 2904 break; 2905 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2906 2907 return (ret); 2908 } 2909 2910 /* 2911 * Set the domain in the appropriate page level domainset. 2912 */ 2913 void 2914 vm_domain_set(struct vm_domain *vmd) 2915 { 2916 2917 mtx_lock(&vm_domainset_lock); 2918 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 2919 vmd->vmd_minset = 1; 2920 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 2921 } 2922 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 2923 vmd->vmd_severeset = 1; 2924 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 2925 } 2926 mtx_unlock(&vm_domainset_lock); 2927 } 2928 2929 /* 2930 * Clear the domain from the appropriate page level domainset. 2931 */ 2932 void 2933 vm_domain_clear(struct vm_domain *vmd) 2934 { 2935 2936 mtx_lock(&vm_domainset_lock); 2937 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 2938 vmd->vmd_minset = 0; 2939 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 2940 if (vm_min_waiters != 0) { 2941 vm_min_waiters = 0; 2942 wakeup(&vm_min_domains); 2943 } 2944 } 2945 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 2946 vmd->vmd_severeset = 0; 2947 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 2948 if (vm_severe_waiters != 0) { 2949 vm_severe_waiters = 0; 2950 wakeup(&vm_severe_domains); 2951 } 2952 } 2953 2954 /* 2955 * If pageout daemon needs pages, then tell it that there are 2956 * some free. 2957 */ 2958 if (vmd->vmd_pageout_pages_needed && 2959 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 2960 wakeup(&vmd->vmd_pageout_pages_needed); 2961 vmd->vmd_pageout_pages_needed = 0; 2962 } 2963 2964 /* See comments in vm_wait_doms(). */ 2965 if (vm_pageproc_waiters) { 2966 vm_pageproc_waiters = 0; 2967 wakeup(&vm_pageproc_waiters); 2968 } 2969 mtx_unlock(&vm_domainset_lock); 2970 } 2971 2972 /* 2973 * Wait for free pages to exceed the min threshold globally. 2974 */ 2975 void 2976 vm_wait_min(void) 2977 { 2978 2979 mtx_lock(&vm_domainset_lock); 2980 while (vm_page_count_min()) { 2981 vm_min_waiters++; 2982 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 2983 } 2984 mtx_unlock(&vm_domainset_lock); 2985 } 2986 2987 /* 2988 * Wait for free pages to exceed the severe threshold globally. 2989 */ 2990 void 2991 vm_wait_severe(void) 2992 { 2993 2994 mtx_lock(&vm_domainset_lock); 2995 while (vm_page_count_severe()) { 2996 vm_severe_waiters++; 2997 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 2998 "vmwait", 0); 2999 } 3000 mtx_unlock(&vm_domainset_lock); 3001 } 3002 3003 u_int 3004 vm_wait_count(void) 3005 { 3006 3007 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3008 } 3009 3010 void 3011 vm_wait_doms(const domainset_t *wdoms) 3012 { 3013 3014 /* 3015 * We use racey wakeup synchronization to avoid expensive global 3016 * locking for the pageproc when sleeping with a non-specific vm_wait. 3017 * To handle this, we only sleep for one tick in this instance. It 3018 * is expected that most allocations for the pageproc will come from 3019 * kmem or vm_page_grab* which will use the more specific and 3020 * race-free vm_wait_domain(). 3021 */ 3022 if (curproc == pageproc) { 3023 mtx_lock(&vm_domainset_lock); 3024 vm_pageproc_waiters++; 3025 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, 3026 "pageprocwait", 1); 3027 } else { 3028 /* 3029 * XXX Ideally we would wait only until the allocation could 3030 * be satisfied. This condition can cause new allocators to 3031 * consume all freed pages while old allocators wait. 3032 */ 3033 mtx_lock(&vm_domainset_lock); 3034 if (vm_page_count_min_set(wdoms)) { 3035 vm_min_waiters++; 3036 msleep(&vm_min_domains, &vm_domainset_lock, 3037 PVM | PDROP, "vmwait", 0); 3038 } else 3039 mtx_unlock(&vm_domainset_lock); 3040 } 3041 } 3042 3043 /* 3044 * vm_wait_domain: 3045 * 3046 * Sleep until free pages are available for allocation. 3047 * - Called in various places after failed memory allocations. 3048 */ 3049 void 3050 vm_wait_domain(int domain) 3051 { 3052 struct vm_domain *vmd; 3053 domainset_t wdom; 3054 3055 vmd = VM_DOMAIN(domain); 3056 vm_domain_free_assert_unlocked(vmd); 3057 3058 if (curproc == pageproc) { 3059 mtx_lock(&vm_domainset_lock); 3060 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3061 vmd->vmd_pageout_pages_needed = 1; 3062 msleep(&vmd->vmd_pageout_pages_needed, 3063 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3064 } else 3065 mtx_unlock(&vm_domainset_lock); 3066 } else { 3067 if (pageproc == NULL) 3068 panic("vm_wait in early boot"); 3069 DOMAINSET_ZERO(&wdom); 3070 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3071 vm_wait_doms(&wdom); 3072 } 3073 } 3074 3075 /* 3076 * vm_wait: 3077 * 3078 * Sleep until free pages are available for allocation in the 3079 * affinity domains of the obj. If obj is NULL, the domain set 3080 * for the calling thread is used. 3081 * Called in various places after failed memory allocations. 3082 */ 3083 void 3084 vm_wait(vm_object_t obj) 3085 { 3086 struct domainset *d; 3087 3088 d = NULL; 3089 3090 /* 3091 * Carefully fetch pointers only once: the struct domainset 3092 * itself is ummutable but the pointer might change. 3093 */ 3094 if (obj != NULL) 3095 d = obj->domain.dr_policy; 3096 if (d == NULL) 3097 d = curthread->td_domain.dr_policy; 3098 3099 vm_wait_doms(&d->ds_mask); 3100 } 3101 3102 /* 3103 * vm_domain_alloc_fail: 3104 * 3105 * Called when a page allocation function fails. Informs the 3106 * pagedaemon and performs the requested wait. Requires the 3107 * domain_free and object lock on entry. Returns with the 3108 * object lock held and free lock released. Returns an error when 3109 * retry is necessary. 3110 * 3111 */ 3112 static int 3113 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3114 { 3115 3116 vm_domain_free_assert_unlocked(vmd); 3117 3118 atomic_add_int(&vmd->vmd_pageout_deficit, 3119 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3120 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3121 if (object != NULL) 3122 VM_OBJECT_WUNLOCK(object); 3123 vm_wait_domain(vmd->vmd_domain); 3124 if (object != NULL) 3125 VM_OBJECT_WLOCK(object); 3126 if (req & VM_ALLOC_WAITOK) 3127 return (EAGAIN); 3128 } 3129 3130 return (0); 3131 } 3132 3133 /* 3134 * vm_waitpfault: 3135 * 3136 * Sleep until free pages are available for allocation. 3137 * - Called only in vm_fault so that processes page faulting 3138 * can be easily tracked. 3139 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3140 * processes will be able to grab memory first. Do not change 3141 * this balance without careful testing first. 3142 */ 3143 void 3144 vm_waitpfault(struct domainset *dset, int timo) 3145 { 3146 3147 /* 3148 * XXX Ideally we would wait only until the allocation could 3149 * be satisfied. This condition can cause new allocators to 3150 * consume all freed pages while old allocators wait. 3151 */ 3152 mtx_lock(&vm_domainset_lock); 3153 if (vm_page_count_min_set(&dset->ds_mask)) { 3154 vm_min_waiters++; 3155 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3156 "pfault", timo); 3157 } else 3158 mtx_unlock(&vm_domainset_lock); 3159 } 3160 3161 static struct vm_pagequeue * 3162 vm_page_pagequeue(vm_page_t m) 3163 { 3164 3165 uint8_t queue; 3166 3167 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) 3168 return (NULL); 3169 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3170 } 3171 3172 static inline void 3173 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m) 3174 { 3175 struct vm_domain *vmd; 3176 uint8_t qflags; 3177 3178 CRITICAL_ASSERT(curthread); 3179 vm_pagequeue_assert_locked(pq); 3180 3181 /* 3182 * The page daemon is allowed to set m->queue = PQ_NONE without 3183 * the page queue lock held. In this case it is about to free the page, 3184 * which must not have any queue state. 3185 */ 3186 qflags = atomic_load_8(&m->aflags); 3187 KASSERT(pq == vm_page_pagequeue(m) || 3188 (qflags & PGA_QUEUE_STATE_MASK) == 0, 3189 ("page %p doesn't belong to queue %p but has aflags %#x", 3190 m, pq, qflags)); 3191 3192 if ((qflags & PGA_DEQUEUE) != 0) { 3193 if (__predict_true((qflags & PGA_ENQUEUED) != 0)) 3194 vm_pagequeue_remove(pq, m); 3195 vm_page_dequeue_complete(m); 3196 counter_u64_add(queue_ops, 1); 3197 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) { 3198 if ((qflags & PGA_ENQUEUED) != 0) 3199 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3200 else { 3201 vm_pagequeue_cnt_inc(pq); 3202 vm_page_aflag_set(m, PGA_ENQUEUED); 3203 } 3204 3205 /* 3206 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. 3207 * In particular, if both flags are set in close succession, 3208 * only PGA_REQUEUE_HEAD will be applied, even if it was set 3209 * first. 3210 */ 3211 if ((qflags & PGA_REQUEUE_HEAD) != 0) { 3212 KASSERT(m->queue == PQ_INACTIVE, 3213 ("head enqueue not supported for page %p", m)); 3214 vmd = vm_pagequeue_domain(m); 3215 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3216 } else 3217 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3218 3219 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE | 3220 PGA_REQUEUE_HEAD)); 3221 counter_u64_add(queue_ops, 1); 3222 } else { 3223 counter_u64_add(queue_nops, 1); 3224 } 3225 } 3226 3227 static void 3228 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3229 uint8_t queue) 3230 { 3231 vm_page_t m; 3232 int i; 3233 3234 for (i = 0; i < bq->bq_cnt; i++) { 3235 m = bq->bq_pa[i]; 3236 if (__predict_false(m->queue != queue)) 3237 continue; 3238 vm_pqbatch_process_page(pq, m); 3239 } 3240 vm_batchqueue_init(bq); 3241 } 3242 3243 /* 3244 * vm_page_pqbatch_submit: [ internal use only ] 3245 * 3246 * Enqueue a page in the specified page queue's batched work queue. 3247 * The caller must have encoded the requested operation in the page 3248 * structure's aflags field. 3249 */ 3250 void 3251 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3252 { 3253 struct vm_batchqueue *bq; 3254 struct vm_pagequeue *pq; 3255 int domain; 3256 3257 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3258 ("page %p is unmanaged", m)); 3259 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, 3260 ("missing synchronization for page %p", m)); 3261 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3262 3263 domain = vm_phys_domain(m); 3264 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 3265 3266 critical_enter(); 3267 bq = DPCPU_PTR(pqbatch[domain][queue]); 3268 if (vm_batchqueue_insert(bq, m)) { 3269 critical_exit(); 3270 return; 3271 } 3272 critical_exit(); 3273 vm_pagequeue_lock(pq); 3274 critical_enter(); 3275 bq = DPCPU_PTR(pqbatch[domain][queue]); 3276 vm_pqbatch_process(pq, bq, queue); 3277 3278 /* 3279 * The page may have been logically dequeued before we acquired the 3280 * page queue lock. In this case, since we either hold the page lock 3281 * or the page is being freed, a different thread cannot be concurrently 3282 * enqueuing the page. 3283 */ 3284 if (__predict_true(m->queue == queue)) 3285 vm_pqbatch_process_page(pq, m); 3286 else { 3287 KASSERT(m->queue == PQ_NONE, 3288 ("invalid queue transition for page %p", m)); 3289 KASSERT((m->aflags & PGA_ENQUEUED) == 0, 3290 ("page %p is enqueued with invalid queue index", m)); 3291 } 3292 vm_pagequeue_unlock(pq); 3293 critical_exit(); 3294 } 3295 3296 /* 3297 * vm_page_pqbatch_drain: [ internal use only ] 3298 * 3299 * Force all per-CPU page queue batch queues to be drained. This is 3300 * intended for use in severe memory shortages, to ensure that pages 3301 * do not remain stuck in the batch queues. 3302 */ 3303 void 3304 vm_page_pqbatch_drain(void) 3305 { 3306 struct thread *td; 3307 struct vm_domain *vmd; 3308 struct vm_pagequeue *pq; 3309 int cpu, domain, queue; 3310 3311 td = curthread; 3312 CPU_FOREACH(cpu) { 3313 thread_lock(td); 3314 sched_bind(td, cpu); 3315 thread_unlock(td); 3316 3317 for (domain = 0; domain < vm_ndomains; domain++) { 3318 vmd = VM_DOMAIN(domain); 3319 for (queue = 0; queue < PQ_COUNT; queue++) { 3320 pq = &vmd->vmd_pagequeues[queue]; 3321 vm_pagequeue_lock(pq); 3322 critical_enter(); 3323 vm_pqbatch_process(pq, 3324 DPCPU_PTR(pqbatch[domain][queue]), queue); 3325 critical_exit(); 3326 vm_pagequeue_unlock(pq); 3327 } 3328 } 3329 } 3330 thread_lock(td); 3331 sched_unbind(td); 3332 thread_unlock(td); 3333 } 3334 3335 /* 3336 * Complete the logical removal of a page from a page queue. We must be 3337 * careful to synchronize with the page daemon, which may be concurrently 3338 * examining the page with only the page lock held. The page must not be 3339 * in a state where it appears to be logically enqueued. 3340 */ 3341 static void 3342 vm_page_dequeue_complete(vm_page_t m) 3343 { 3344 3345 m->queue = PQ_NONE; 3346 atomic_thread_fence_rel(); 3347 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); 3348 } 3349 3350 /* 3351 * vm_page_dequeue_deferred: [ internal use only ] 3352 * 3353 * Request removal of the given page from its current page 3354 * queue. Physical removal from the queue may be deferred 3355 * indefinitely. 3356 * 3357 * The page must be locked. 3358 */ 3359 void 3360 vm_page_dequeue_deferred(vm_page_t m) 3361 { 3362 uint8_t queue; 3363 3364 vm_page_assert_locked(m); 3365 3366 if ((queue = vm_page_queue(m)) == PQ_NONE) 3367 return; 3368 3369 /* 3370 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call 3371 * to vm_page_dequeue_deferred_free(). In particular, avoid modifying 3372 * the page's queue state once vm_page_dequeue_deferred_free() has been 3373 * called. In the event of a race, two batch queue entries for the page 3374 * will be created, but the second will have no effect. 3375 */ 3376 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE)) 3377 vm_page_pqbatch_submit(m, queue); 3378 } 3379 3380 /* 3381 * A variant of vm_page_dequeue_deferred() that does not assert the page 3382 * lock and is only to be called from vm_page_free_prep(). Because the 3383 * page is being freed, we can assume that nothing other than the page 3384 * daemon is scheduling queue operations on this page, so we get for 3385 * free the mutual exclusion that is otherwise provided by the page lock. 3386 * To handle races, the page daemon must take care to atomically check 3387 * for PGA_DEQUEUE when updating queue state. 3388 */ 3389 static void 3390 vm_page_dequeue_deferred_free(vm_page_t m) 3391 { 3392 uint8_t queue; 3393 3394 KASSERT(m->ref_count == 0, ("page %p has references", m)); 3395 3396 for (;;) { 3397 if ((m->aflags & PGA_DEQUEUE) != 0) 3398 return; 3399 atomic_thread_fence_acq(); 3400 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) 3401 return; 3402 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, 3403 PGA_DEQUEUE)) { 3404 vm_page_pqbatch_submit(m, queue); 3405 break; 3406 } 3407 } 3408 } 3409 3410 /* 3411 * vm_page_dequeue: 3412 * 3413 * Remove the page from whichever page queue it's in, if any. 3414 * The page must either be locked or unallocated. This constraint 3415 * ensures that the queue state of the page will remain consistent 3416 * after this function returns. 3417 */ 3418 void 3419 vm_page_dequeue(vm_page_t m) 3420 { 3421 struct vm_pagequeue *pq, *pq1; 3422 uint8_t aflags; 3423 3424 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, 3425 ("page %p is allocated and unlocked", m)); 3426 3427 for (pq = vm_page_pagequeue(m);; pq = pq1) { 3428 if (pq == NULL) { 3429 /* 3430 * A thread may be concurrently executing 3431 * vm_page_dequeue_complete(). Ensure that all queue 3432 * state is cleared before we return. 3433 */ 3434 aflags = atomic_load_8(&m->aflags); 3435 if ((aflags & PGA_QUEUE_STATE_MASK) == 0) 3436 return; 3437 KASSERT((aflags & PGA_DEQUEUE) != 0, 3438 ("page %p has unexpected queue state flags %#x", 3439 m, aflags)); 3440 3441 /* 3442 * Busy wait until the thread updating queue state is 3443 * finished. Such a thread must be executing in a 3444 * critical section. 3445 */ 3446 cpu_spinwait(); 3447 pq1 = vm_page_pagequeue(m); 3448 continue; 3449 } 3450 vm_pagequeue_lock(pq); 3451 if ((pq1 = vm_page_pagequeue(m)) == pq) 3452 break; 3453 vm_pagequeue_unlock(pq); 3454 } 3455 KASSERT(pq == vm_page_pagequeue(m), 3456 ("%s: page %p migrated directly between queues", __func__, m)); 3457 KASSERT((m->aflags & PGA_DEQUEUE) != 0 || 3458 mtx_owned(vm_page_lockptr(m)), 3459 ("%s: queued unlocked page %p", __func__, m)); 3460 3461 if ((m->aflags & PGA_ENQUEUED) != 0) 3462 vm_pagequeue_remove(pq, m); 3463 vm_page_dequeue_complete(m); 3464 vm_pagequeue_unlock(pq); 3465 } 3466 3467 /* 3468 * Schedule the given page for insertion into the specified page queue. 3469 * Physical insertion of the page may be deferred indefinitely. 3470 */ 3471 static void 3472 vm_page_enqueue(vm_page_t m, uint8_t queue) 3473 { 3474 3475 vm_page_assert_locked(m); 3476 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 3477 ("%s: page %p is already enqueued", __func__, m)); 3478 3479 m->queue = queue; 3480 if ((m->aflags & PGA_REQUEUE) == 0) 3481 vm_page_aflag_set(m, PGA_REQUEUE); 3482 vm_page_pqbatch_submit(m, queue); 3483 } 3484 3485 /* 3486 * vm_page_requeue: [ internal use only ] 3487 * 3488 * Schedule a requeue of the given page. 3489 * 3490 * The page must be locked. 3491 */ 3492 void 3493 vm_page_requeue(vm_page_t m) 3494 { 3495 3496 vm_page_assert_locked(m); 3497 KASSERT(vm_page_queue(m) != PQ_NONE, 3498 ("%s: page %p is not logically enqueued", __func__, m)); 3499 3500 if ((m->aflags & PGA_REQUEUE) == 0) 3501 vm_page_aflag_set(m, PGA_REQUEUE); 3502 vm_page_pqbatch_submit(m, atomic_load_8(&m->queue)); 3503 } 3504 3505 /* 3506 * vm_page_swapqueue: [ internal use only ] 3507 * 3508 * Move the page from one queue to another, or to the tail of its 3509 * current queue, in the face of a possible concurrent call to 3510 * vm_page_dequeue_deferred_free(). 3511 */ 3512 void 3513 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq) 3514 { 3515 struct vm_pagequeue *pq; 3516 vm_page_t next; 3517 bool queued; 3518 3519 KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq, 3520 ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq)); 3521 vm_page_assert_locked(m); 3522 3523 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq]; 3524 vm_pagequeue_lock(pq); 3525 3526 /* 3527 * The physical queue state might change at any point before the page 3528 * queue lock is acquired, so we must verify that we hold the correct 3529 * lock before proceeding. 3530 */ 3531 if (__predict_false(m->queue != oldq)) { 3532 vm_pagequeue_unlock(pq); 3533 return; 3534 } 3535 3536 /* 3537 * Once the queue index of the page changes, there is nothing 3538 * synchronizing with further updates to the physical queue state. 3539 * Therefore we must remove the page from the queue now in anticipation 3540 * of a successful commit, and be prepared to roll back. 3541 */ 3542 if (__predict_true((m->aflags & PGA_ENQUEUED) != 0)) { 3543 next = TAILQ_NEXT(m, plinks.q); 3544 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3545 vm_page_aflag_clear(m, PGA_ENQUEUED); 3546 queued = true; 3547 } else { 3548 queued = false; 3549 } 3550 3551 /* 3552 * Atomically update the queue field and set PGA_REQUEUE while 3553 * ensuring that PGA_DEQUEUE has not been set. 3554 */ 3555 if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE, 3556 PGA_REQUEUE))) { 3557 if (queued) { 3558 vm_page_aflag_set(m, PGA_ENQUEUED); 3559 if (next != NULL) 3560 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3561 else 3562 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3563 } 3564 vm_pagequeue_unlock(pq); 3565 return; 3566 } 3567 vm_pagequeue_cnt_dec(pq); 3568 vm_pagequeue_unlock(pq); 3569 vm_page_pqbatch_submit(m, newq); 3570 } 3571 3572 /* 3573 * vm_page_free_prep: 3574 * 3575 * Prepares the given page to be put on the free list, 3576 * disassociating it from any VM object. The caller may return 3577 * the page to the free list only if this function returns true. 3578 * 3579 * The object must be locked. The page must be locked if it is 3580 * managed. 3581 */ 3582 bool 3583 vm_page_free_prep(vm_page_t m) 3584 { 3585 3586 /* 3587 * Synchronize with threads that have dropped a reference to this 3588 * page. 3589 */ 3590 atomic_thread_fence_acq(); 3591 3592 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3593 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3594 uint64_t *p; 3595 int i; 3596 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3597 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3598 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3599 m, i, (uintmax_t)*p)); 3600 } 3601 #endif 3602 if ((m->oflags & VPO_UNMANAGED) == 0) { 3603 KASSERT(!pmap_page_is_mapped(m), 3604 ("vm_page_free_prep: freeing mapped page %p", m)); 3605 KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3606 ("vm_page_free_prep: mapping flags set in page %p", m)); 3607 } else { 3608 KASSERT(m->queue == PQ_NONE, 3609 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3610 } 3611 VM_CNT_INC(v_tfree); 3612 3613 if (vm_page_sbusied(m)) 3614 panic("vm_page_free_prep: freeing shared busy page %p", m); 3615 3616 if (m->object != NULL) { 3617 vm_page_object_remove(m); 3618 3619 /* 3620 * The object reference can be released without an atomic 3621 * operation. 3622 */ 3623 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3624 m->ref_count == VPRC_OBJREF, 3625 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3626 m, m->ref_count)); 3627 m->object = NULL; 3628 m->ref_count -= VPRC_OBJREF; 3629 } 3630 3631 if (vm_page_xbusied(m)) 3632 vm_page_xunbusy(m); 3633 3634 /* 3635 * If fictitious remove object association and 3636 * return. 3637 */ 3638 if ((m->flags & PG_FICTITIOUS) != 0) { 3639 KASSERT(m->ref_count == 1, 3640 ("fictitious page %p is referenced", m)); 3641 KASSERT(m->queue == PQ_NONE, 3642 ("fictitious page %p is queued", m)); 3643 return (false); 3644 } 3645 3646 /* 3647 * Pages need not be dequeued before they are returned to the physical 3648 * memory allocator, but they must at least be marked for a deferred 3649 * dequeue. 3650 */ 3651 if ((m->oflags & VPO_UNMANAGED) == 0) 3652 vm_page_dequeue_deferred_free(m); 3653 3654 m->valid = 0; 3655 vm_page_undirty(m); 3656 3657 if (m->ref_count != 0) 3658 panic("vm_page_free_prep: page %p has references", m); 3659 3660 /* 3661 * Restore the default memory attribute to the page. 3662 */ 3663 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3664 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3665 3666 #if VM_NRESERVLEVEL > 0 3667 /* 3668 * Determine whether the page belongs to a reservation. If the page was 3669 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3670 * as an optimization, we avoid the check in that case. 3671 */ 3672 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3673 return (false); 3674 #endif 3675 3676 return (true); 3677 } 3678 3679 /* 3680 * vm_page_free_toq: 3681 * 3682 * Returns the given page to the free list, disassociating it 3683 * from any VM object. 3684 * 3685 * The object must be locked. The page must be locked if it is 3686 * managed. 3687 */ 3688 void 3689 vm_page_free_toq(vm_page_t m) 3690 { 3691 struct vm_domain *vmd; 3692 uma_zone_t zone; 3693 3694 if (!vm_page_free_prep(m)) 3695 return; 3696 3697 vmd = vm_pagequeue_domain(m); 3698 zone = vmd->vmd_pgcache[m->pool].zone; 3699 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 3700 uma_zfree(zone, m); 3701 return; 3702 } 3703 vm_domain_free_lock(vmd); 3704 vm_phys_free_pages(m, 0); 3705 vm_domain_free_unlock(vmd); 3706 vm_domain_freecnt_inc(vmd, 1); 3707 } 3708 3709 /* 3710 * vm_page_free_pages_toq: 3711 * 3712 * Returns a list of pages to the free list, disassociating it 3713 * from any VM object. In other words, this is equivalent to 3714 * calling vm_page_free_toq() for each page of a list of VM objects. 3715 * 3716 * The objects must be locked. The pages must be locked if it is 3717 * managed. 3718 */ 3719 void 3720 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3721 { 3722 vm_page_t m; 3723 int count; 3724 3725 if (SLIST_EMPTY(free)) 3726 return; 3727 3728 count = 0; 3729 while ((m = SLIST_FIRST(free)) != NULL) { 3730 count++; 3731 SLIST_REMOVE_HEAD(free, plinks.s.ss); 3732 vm_page_free_toq(m); 3733 } 3734 3735 if (update_wire_count) 3736 vm_wire_sub(count); 3737 } 3738 3739 /* 3740 * Mark this page as wired down, preventing reclamation by the page daemon 3741 * or when the containing object is destroyed. 3742 */ 3743 void 3744 vm_page_wire(vm_page_t m) 3745 { 3746 u_int old; 3747 3748 KASSERT(m->object != NULL, 3749 ("vm_page_wire: page %p does not belong to an object", m)); 3750 if (!vm_page_busied(m)) 3751 VM_OBJECT_ASSERT_LOCKED(m->object); 3752 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 3753 VPRC_WIRE_COUNT(m->ref_count) >= 1, 3754 ("vm_page_wire: fictitious page %p has zero wirings", m)); 3755 3756 old = atomic_fetchadd_int(&m->ref_count, 1); 3757 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 3758 ("vm_page_wire: counter overflow for page %p", m)); 3759 if (VPRC_WIRE_COUNT(old) == 0) 3760 vm_wire_add(1); 3761 } 3762 3763 /* 3764 * Attempt to wire a mapped page following a pmap lookup of that page. 3765 * This may fail if a thread is concurrently tearing down mappings of the page. 3766 * The transient failure is acceptable because it translates to the 3767 * failure of the caller pmap_extract_and_hold(), which should be then 3768 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 3769 */ 3770 bool 3771 vm_page_wire_mapped(vm_page_t m) 3772 { 3773 u_int old; 3774 3775 old = m->ref_count; 3776 do { 3777 KASSERT(old > 0, 3778 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 3779 if ((old & VPRC_BLOCKED) != 0) 3780 return (false); 3781 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 3782 3783 if (VPRC_WIRE_COUNT(old) == 0) 3784 vm_wire_add(1); 3785 return (true); 3786 } 3787 3788 /* 3789 * Release one wiring of the specified page, potentially allowing it to be 3790 * paged out. 3791 * 3792 * Only managed pages belonging to an object can be paged out. If the number 3793 * of wirings transitions to zero and the page is eligible for page out, then 3794 * the page is added to the specified paging queue. If the released wiring 3795 * represented the last reference to the page, the page is freed. 3796 * 3797 * A managed page must be locked. 3798 */ 3799 void 3800 vm_page_unwire(vm_page_t m, uint8_t queue) 3801 { 3802 u_int old; 3803 bool locked; 3804 3805 KASSERT(queue < PQ_COUNT, 3806 ("vm_page_unwire: invalid queue %u request for page %p", queue, m)); 3807 3808 if ((m->oflags & VPO_UNMANAGED) != 0) { 3809 if (vm_page_unwire_noq(m) && m->ref_count == 0) 3810 vm_page_free(m); 3811 return; 3812 } 3813 3814 /* 3815 * Update LRU state before releasing the wiring reference. 3816 * We only need to do this once since we hold the page lock. 3817 * Use a release store when updating the reference count to 3818 * synchronize with vm_page_free_prep(). 3819 */ 3820 old = m->ref_count; 3821 locked = false; 3822 do { 3823 KASSERT(VPRC_WIRE_COUNT(old) > 0, 3824 ("vm_page_unwire: wire count underflow for page %p", m)); 3825 if (!locked && VPRC_WIRE_COUNT(old) == 1) { 3826 vm_page_lock(m); 3827 locked = true; 3828 if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE) 3829 vm_page_reference(m); 3830 else 3831 vm_page_mvqueue(m, queue); 3832 } 3833 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 3834 3835 /* 3836 * Release the lock only after the wiring is released, to ensure that 3837 * the page daemon does not encounter and dequeue the page while it is 3838 * still wired. 3839 */ 3840 if (locked) 3841 vm_page_unlock(m); 3842 3843 if (VPRC_WIRE_COUNT(old) == 1) { 3844 vm_wire_sub(1); 3845 if (old == 1) 3846 vm_page_free(m); 3847 } 3848 } 3849 3850 /* 3851 * Unwire a page without (re-)inserting it into a page queue. It is up 3852 * to the caller to enqueue, requeue, or free the page as appropriate. 3853 * In most cases involving managed pages, vm_page_unwire() should be used 3854 * instead. 3855 */ 3856 bool 3857 vm_page_unwire_noq(vm_page_t m) 3858 { 3859 u_int old; 3860 3861 old = vm_page_drop(m, 1); 3862 KASSERT(VPRC_WIRE_COUNT(old) != 0, 3863 ("vm_page_unref: counter underflow for page %p", m)); 3864 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 3865 ("vm_page_unref: missing ref on fictitious page %p", m)); 3866 3867 if (VPRC_WIRE_COUNT(old) > 1) 3868 return (false); 3869 vm_wire_sub(1); 3870 return (true); 3871 } 3872 3873 /* 3874 * Ensure that the page is in the specified page queue. If the page is 3875 * active or being moved to the active queue, ensure that its act_count is 3876 * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that 3877 * the page is at the tail of its page queue. 3878 * 3879 * The page may be wired. The caller should release its wiring reference 3880 * before releasing the page lock, otherwise the page daemon may immediately 3881 * dequeue the page. 3882 * 3883 * A managed page must be locked. 3884 */ 3885 static __always_inline void 3886 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue) 3887 { 3888 3889 vm_page_assert_locked(m); 3890 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3891 ("vm_page_mvqueue: page %p is unmanaged", m)); 3892 3893 if (vm_page_queue(m) != nqueue) { 3894 vm_page_dequeue(m); 3895 vm_page_enqueue(m, nqueue); 3896 } else if (nqueue != PQ_ACTIVE) { 3897 vm_page_requeue(m); 3898 } 3899 3900 if (nqueue == PQ_ACTIVE && m->act_count < ACT_INIT) 3901 m->act_count = ACT_INIT; 3902 } 3903 3904 /* 3905 * Put the specified page on the active list (if appropriate). 3906 */ 3907 void 3908 vm_page_activate(vm_page_t m) 3909 { 3910 3911 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3912 return; 3913 vm_page_mvqueue(m, PQ_ACTIVE); 3914 } 3915 3916 /* 3917 * Move the specified page to the tail of the inactive queue, or requeue 3918 * the page if it is already in the inactive queue. 3919 */ 3920 void 3921 vm_page_deactivate(vm_page_t m) 3922 { 3923 3924 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3925 return; 3926 vm_page_mvqueue(m, PQ_INACTIVE); 3927 } 3928 3929 /* 3930 * Move the specified page close to the head of the inactive queue, 3931 * bypassing LRU. A marker page is used to maintain FIFO ordering. 3932 * As with regular enqueues, we use a per-CPU batch queue to reduce 3933 * contention on the page queue lock. 3934 */ 3935 static void 3936 _vm_page_deactivate_noreuse(vm_page_t m) 3937 { 3938 3939 vm_page_assert_locked(m); 3940 3941 if (!vm_page_inactive(m)) { 3942 vm_page_dequeue(m); 3943 m->queue = PQ_INACTIVE; 3944 } 3945 if ((m->aflags & PGA_REQUEUE_HEAD) == 0) 3946 vm_page_aflag_set(m, PGA_REQUEUE_HEAD); 3947 vm_page_pqbatch_submit(m, PQ_INACTIVE); 3948 } 3949 3950 void 3951 vm_page_deactivate_noreuse(vm_page_t m) 3952 { 3953 3954 KASSERT(m->object != NULL, 3955 ("vm_page_deactivate_noreuse: page %p has no object", m)); 3956 3957 if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m)) 3958 _vm_page_deactivate_noreuse(m); 3959 } 3960 3961 /* 3962 * Put a page in the laundry, or requeue it if it is already there. 3963 */ 3964 void 3965 vm_page_launder(vm_page_t m) 3966 { 3967 3968 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3969 return; 3970 vm_page_mvqueue(m, PQ_LAUNDRY); 3971 } 3972 3973 /* 3974 * Put a page in the PQ_UNSWAPPABLE holding queue. 3975 */ 3976 void 3977 vm_page_unswappable(vm_page_t m) 3978 { 3979 3980 vm_page_assert_locked(m); 3981 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 3982 ("page %p already unswappable", m)); 3983 3984 vm_page_dequeue(m); 3985 vm_page_enqueue(m, PQ_UNSWAPPABLE); 3986 } 3987 3988 static void 3989 vm_page_release_toq(vm_page_t m, int flags) 3990 { 3991 3992 vm_page_assert_locked(m); 3993 3994 /* 3995 * Use a check of the valid bits to determine whether we should 3996 * accelerate reclamation of the page. The object lock might not be 3997 * held here, in which case the check is racy. At worst we will either 3998 * accelerate reclamation of a valid page and violate LRU, or 3999 * unnecessarily defer reclamation of an invalid page. 4000 * 4001 * If we were asked to not cache the page, place it near the head of the 4002 * inactive queue so that is reclaimed sooner. 4003 */ 4004 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0) 4005 _vm_page_deactivate_noreuse(m); 4006 else if (vm_page_active(m)) 4007 vm_page_reference(m); 4008 else 4009 vm_page_mvqueue(m, PQ_INACTIVE); 4010 } 4011 4012 /* 4013 * Unwire a page and either attempt to free it or re-add it to the page queues. 4014 */ 4015 void 4016 vm_page_release(vm_page_t m, int flags) 4017 { 4018 vm_object_t object; 4019 u_int old; 4020 bool locked; 4021 4022 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4023 ("vm_page_release: page %p is unmanaged", m)); 4024 4025 if ((flags & VPR_TRYFREE) != 0) { 4026 for (;;) { 4027 object = (vm_object_t)atomic_load_ptr(&m->object); 4028 if (object == NULL) 4029 break; 4030 /* Depends on type-stability. */ 4031 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) { 4032 object = NULL; 4033 break; 4034 } 4035 if (object == m->object) 4036 break; 4037 VM_OBJECT_WUNLOCK(object); 4038 } 4039 if (__predict_true(object != NULL)) { 4040 vm_page_release_locked(m, flags); 4041 VM_OBJECT_WUNLOCK(object); 4042 return; 4043 } 4044 } 4045 4046 /* 4047 * Update LRU state before releasing the wiring reference. 4048 * Use a release store when updating the reference count to 4049 * synchronize with vm_page_free_prep(). 4050 */ 4051 old = m->ref_count; 4052 locked = false; 4053 do { 4054 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4055 ("vm_page_unwire: wire count underflow for page %p", m)); 4056 if (!locked && VPRC_WIRE_COUNT(old) == 1) { 4057 vm_page_lock(m); 4058 locked = true; 4059 vm_page_release_toq(m, flags); 4060 } 4061 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4062 4063 /* 4064 * Release the lock only after the wiring is released, to ensure that 4065 * the page daemon does not encounter and dequeue the page while it is 4066 * still wired. 4067 */ 4068 if (locked) 4069 vm_page_unlock(m); 4070 4071 if (VPRC_WIRE_COUNT(old) == 1) { 4072 vm_wire_sub(1); 4073 if (old == 1) 4074 vm_page_free(m); 4075 } 4076 } 4077 4078 /* See vm_page_release(). */ 4079 void 4080 vm_page_release_locked(vm_page_t m, int flags) 4081 { 4082 4083 VM_OBJECT_ASSERT_WLOCKED(m->object); 4084 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4085 ("vm_page_release_locked: page %p is unmanaged", m)); 4086 4087 if (vm_page_unwire_noq(m)) { 4088 if ((flags & VPR_TRYFREE) != 0 && 4089 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4090 m->dirty == 0 && !vm_page_busied(m)) { 4091 vm_page_free(m); 4092 } else { 4093 vm_page_lock(m); 4094 vm_page_release_toq(m, flags); 4095 vm_page_unlock(m); 4096 } 4097 } 4098 } 4099 4100 static bool 4101 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4102 { 4103 u_int old; 4104 4105 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4106 ("vm_page_try_blocked_op: page %p has no object", m)); 4107 KASSERT(vm_page_busied(m), 4108 ("vm_page_try_blocked_op: page %p is not busy", m)); 4109 VM_OBJECT_ASSERT_LOCKED(m->object); 4110 4111 old = m->ref_count; 4112 do { 4113 KASSERT(old != 0, 4114 ("vm_page_try_blocked_op: page %p has no references", m)); 4115 if (VPRC_WIRE_COUNT(old) != 0) 4116 return (false); 4117 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4118 4119 (op)(m); 4120 4121 /* 4122 * If the object is read-locked, new wirings may be created via an 4123 * object lookup. 4124 */ 4125 old = vm_page_drop(m, VPRC_BLOCKED); 4126 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4127 old == (VPRC_BLOCKED | VPRC_OBJREF), 4128 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4129 old, m)); 4130 return (true); 4131 } 4132 4133 /* 4134 * Atomically check for wirings and remove all mappings of the page. 4135 */ 4136 bool 4137 vm_page_try_remove_all(vm_page_t m) 4138 { 4139 4140 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4141 } 4142 4143 /* 4144 * Atomically check for wirings and remove all writeable mappings of the page. 4145 */ 4146 bool 4147 vm_page_try_remove_write(vm_page_t m) 4148 { 4149 4150 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4151 } 4152 4153 /* 4154 * vm_page_advise 4155 * 4156 * Apply the specified advice to the given page. 4157 * 4158 * The object and page must be locked. 4159 */ 4160 void 4161 vm_page_advise(vm_page_t m, int advice) 4162 { 4163 4164 vm_page_assert_locked(m); 4165 VM_OBJECT_ASSERT_WLOCKED(m->object); 4166 if (advice == MADV_FREE) 4167 /* 4168 * Mark the page clean. This will allow the page to be freed 4169 * without first paging it out. MADV_FREE pages are often 4170 * quickly reused by malloc(3), so we do not do anything that 4171 * would result in a page fault on a later access. 4172 */ 4173 vm_page_undirty(m); 4174 else if (advice != MADV_DONTNEED) { 4175 if (advice == MADV_WILLNEED) 4176 vm_page_activate(m); 4177 return; 4178 } 4179 4180 /* 4181 * Clear any references to the page. Otherwise, the page daemon will 4182 * immediately reactivate the page. 4183 */ 4184 vm_page_aflag_clear(m, PGA_REFERENCED); 4185 4186 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4187 vm_page_dirty(m); 4188 4189 /* 4190 * Place clean pages near the head of the inactive queue rather than 4191 * the tail, thus defeating the queue's LRU operation and ensuring that 4192 * the page will be reused quickly. Dirty pages not already in the 4193 * laundry are moved there. 4194 */ 4195 if (m->dirty == 0) 4196 vm_page_deactivate_noreuse(m); 4197 else if (!vm_page_in_laundry(m)) 4198 vm_page_launder(m); 4199 } 4200 4201 /* 4202 * Grab a page, waiting until we are waken up due to the page 4203 * changing state. We keep on waiting, if the page continues 4204 * to be in the object. If the page doesn't exist, first allocate it 4205 * and then conditionally zero it. 4206 * 4207 * This routine may sleep. 4208 * 4209 * The object must be locked on entry. The lock will, however, be released 4210 * and reacquired if the routine sleeps. 4211 */ 4212 vm_page_t 4213 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4214 { 4215 vm_page_t m; 4216 int sleep; 4217 int pflags; 4218 4219 VM_OBJECT_ASSERT_WLOCKED(object); 4220 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4221 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4222 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4223 pflags = allocflags & 4224 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4225 VM_ALLOC_NOBUSY); 4226 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4227 pflags |= VM_ALLOC_WAITFAIL; 4228 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4229 pflags |= VM_ALLOC_SBUSY; 4230 retrylookup: 4231 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4232 if ((allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) != 0) 4233 sleep = !vm_page_trysbusy(m); 4234 else 4235 sleep = !vm_page_tryxbusy(m); 4236 if (sleep) { 4237 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4238 return (NULL); 4239 /* 4240 * Reference the page before unlocking and 4241 * sleeping so that the page daemon is less 4242 * likely to reclaim it. 4243 */ 4244 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 4245 vm_page_aflag_set(m, PGA_REFERENCED); 4246 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 4247 VM_ALLOC_IGN_SBUSY) != 0); 4248 VM_OBJECT_WLOCK(object); 4249 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4250 return (NULL); 4251 goto retrylookup; 4252 } else { 4253 if ((allocflags & VM_ALLOC_WIRED) != 0) 4254 vm_page_wire(m); 4255 goto out; 4256 } 4257 } 4258 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4259 return (NULL); 4260 m = vm_page_alloc(object, pindex, pflags); 4261 if (m == NULL) { 4262 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4263 return (NULL); 4264 goto retrylookup; 4265 } 4266 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4267 pmap_zero_page(m); 4268 4269 out: 4270 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4271 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4272 vm_page_sunbusy(m); 4273 else 4274 vm_page_xunbusy(m); 4275 } 4276 return (m); 4277 } 4278 4279 /* 4280 * Grab a page and make it valid, paging in if necessary. Pages missing from 4281 * their pager are zero filled and validated. 4282 */ 4283 int 4284 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4285 { 4286 vm_page_t m; 4287 bool sleep, xbusy; 4288 int pflags; 4289 int rv; 4290 4291 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4292 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4293 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4294 KASSERT((allocflags & 4295 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4296 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4297 VM_OBJECT_ASSERT_WLOCKED(object); 4298 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY); 4299 pflags |= VM_ALLOC_WAITFAIL; 4300 4301 retrylookup: 4302 xbusy = false; 4303 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4304 /* 4305 * If the page is fully valid it can only become invalid 4306 * with the object lock held. If it is not valid it can 4307 * become valid with the busy lock held. Therefore, we 4308 * may unnecessarily lock the exclusive busy here if we 4309 * race with I/O completion not using the object lock. 4310 * However, we will not end up with an invalid page and a 4311 * shared lock. 4312 */ 4313 if (!vm_page_all_valid(m) || 4314 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) { 4315 sleep = !vm_page_tryxbusy(m); 4316 xbusy = true; 4317 } else 4318 sleep = !vm_page_trysbusy(m); 4319 if (sleep) { 4320 /* 4321 * Reference the page before unlocking and 4322 * sleeping so that the page daemon is less 4323 * likely to reclaim it. 4324 */ 4325 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 4326 vm_page_aflag_set(m, PGA_REFERENCED); 4327 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 4328 VM_ALLOC_IGN_SBUSY) != 0); 4329 VM_OBJECT_WLOCK(object); 4330 goto retrylookup; 4331 } 4332 if ((allocflags & VM_ALLOC_NOCREAT) != 0 && 4333 !vm_page_all_valid(m)) { 4334 if (xbusy) 4335 vm_page_xunbusy(m); 4336 else 4337 vm_page_sunbusy(m); 4338 *mp = NULL; 4339 return (VM_PAGER_FAIL); 4340 } 4341 if ((allocflags & VM_ALLOC_WIRED) != 0) 4342 vm_page_wire(m); 4343 if (vm_page_all_valid(m)) 4344 goto out; 4345 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4346 *mp = NULL; 4347 return (VM_PAGER_FAIL); 4348 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) { 4349 xbusy = true; 4350 } else { 4351 goto retrylookup; 4352 } 4353 4354 vm_page_assert_xbusied(m); 4355 MPASS(xbusy); 4356 if (vm_pager_has_page(object, pindex, NULL, NULL)) { 4357 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 4358 if (rv != VM_PAGER_OK) { 4359 if (allocflags & VM_ALLOC_WIRED) 4360 vm_page_unwire_noq(m); 4361 vm_page_free(m); 4362 *mp = NULL; 4363 return (rv); 4364 } 4365 MPASS(vm_page_all_valid(m)); 4366 } else { 4367 vm_page_zero_invalid(m, TRUE); 4368 } 4369 out: 4370 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4371 if (xbusy) 4372 vm_page_xunbusy(m); 4373 else 4374 vm_page_sunbusy(m); 4375 } 4376 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy) 4377 vm_page_busy_downgrade(m); 4378 *mp = m; 4379 return (VM_PAGER_OK); 4380 } 4381 4382 /* 4383 * Return the specified range of pages from the given object. For each 4384 * page offset within the range, if a page already exists within the object 4385 * at that offset and it is busy, then wait for it to change state. If, 4386 * instead, the page doesn't exist, then allocate it. 4387 * 4388 * The caller must always specify an allocation class. 4389 * 4390 * allocation classes: 4391 * VM_ALLOC_NORMAL normal process request 4392 * VM_ALLOC_SYSTEM system *really* needs the pages 4393 * 4394 * The caller must always specify that the pages are to be busied and/or 4395 * wired. 4396 * 4397 * optional allocation flags: 4398 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4399 * VM_ALLOC_NOBUSY do not exclusive busy the page 4400 * VM_ALLOC_NOWAIT do not sleep 4401 * VM_ALLOC_SBUSY set page to sbusy state 4402 * VM_ALLOC_WIRED wire the pages 4403 * VM_ALLOC_ZERO zero and validate any invalid pages 4404 * 4405 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4406 * may return a partial prefix of the requested range. 4407 */ 4408 int 4409 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4410 vm_page_t *ma, int count) 4411 { 4412 vm_page_t m, mpred; 4413 int pflags; 4414 int i; 4415 bool sleep; 4416 4417 VM_OBJECT_ASSERT_WLOCKED(object); 4418 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4419 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4420 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4421 (allocflags & VM_ALLOC_WIRED) != 0, 4422 ("vm_page_grab_pages: the pages must be busied or wired")); 4423 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4424 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4425 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); 4426 if (count == 0) 4427 return (0); 4428 pflags = allocflags & 4429 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4430 VM_ALLOC_NOBUSY); 4431 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4432 pflags |= VM_ALLOC_WAITFAIL; 4433 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4434 pflags |= VM_ALLOC_SBUSY; 4435 i = 0; 4436 retrylookup: 4437 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4438 if (m == NULL || m->pindex != pindex + i) { 4439 mpred = m; 4440 m = NULL; 4441 } else 4442 mpred = TAILQ_PREV(m, pglist, listq); 4443 for (; i < count; i++) { 4444 if (m != NULL) { 4445 if ((allocflags & 4446 (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 4447 sleep = !vm_page_trysbusy(m); 4448 else 4449 sleep = !vm_page_tryxbusy(m); 4450 if (sleep) { 4451 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4452 break; 4453 /* 4454 * Reference the page before unlocking and 4455 * sleeping so that the page daemon is less 4456 * likely to reclaim it. 4457 */ 4458 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 4459 vm_page_aflag_set(m, PGA_REFERENCED); 4460 vm_page_busy_sleep(m, "grbmaw", (allocflags & 4461 VM_ALLOC_IGN_SBUSY) != 0); 4462 VM_OBJECT_WLOCK(object); 4463 goto retrylookup; 4464 } 4465 if ((allocflags & VM_ALLOC_WIRED) != 0) 4466 vm_page_wire(m); 4467 } else { 4468 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4469 break; 4470 m = vm_page_alloc_after(object, pindex + i, 4471 pflags | VM_ALLOC_COUNT(count - i), mpred); 4472 if (m == NULL) { 4473 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4474 break; 4475 goto retrylookup; 4476 } 4477 } 4478 if (vm_page_none_valid(m) && 4479 (allocflags & VM_ALLOC_ZERO) != 0) { 4480 if ((m->flags & PG_ZERO) == 0) 4481 pmap_zero_page(m); 4482 vm_page_valid(m); 4483 } 4484 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4485 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4486 vm_page_sunbusy(m); 4487 else 4488 vm_page_xunbusy(m); 4489 } 4490 ma[i] = mpred = m; 4491 m = vm_page_next(m); 4492 } 4493 return (i); 4494 } 4495 4496 /* 4497 * Mapping function for valid or dirty bits in a page. 4498 * 4499 * Inputs are required to range within a page. 4500 */ 4501 vm_page_bits_t 4502 vm_page_bits(int base, int size) 4503 { 4504 int first_bit; 4505 int last_bit; 4506 4507 KASSERT( 4508 base + size <= PAGE_SIZE, 4509 ("vm_page_bits: illegal base/size %d/%d", base, size) 4510 ); 4511 4512 if (size == 0) /* handle degenerate case */ 4513 return (0); 4514 4515 first_bit = base >> DEV_BSHIFT; 4516 last_bit = (base + size - 1) >> DEV_BSHIFT; 4517 4518 return (((vm_page_bits_t)2 << last_bit) - 4519 ((vm_page_bits_t)1 << first_bit)); 4520 } 4521 4522 static inline void 4523 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 4524 { 4525 4526 #if PAGE_SIZE == 32768 4527 atomic_set_64((uint64_t *)bits, set); 4528 #elif PAGE_SIZE == 16384 4529 atomic_set_32((uint32_t *)bits, set); 4530 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 4531 atomic_set_16((uint16_t *)bits, set); 4532 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 4533 atomic_set_8((uint8_t *)bits, set); 4534 #else /* PAGE_SIZE <= 8192 */ 4535 uintptr_t addr; 4536 int shift; 4537 4538 addr = (uintptr_t)bits; 4539 /* 4540 * Use a trick to perform a 32-bit atomic on the 4541 * containing aligned word, to not depend on the existence 4542 * of atomic_{set, clear}_{8, 16}. 4543 */ 4544 shift = addr & (sizeof(uint32_t) - 1); 4545 #if BYTE_ORDER == BIG_ENDIAN 4546 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4547 #else 4548 shift *= NBBY; 4549 #endif 4550 addr &= ~(sizeof(uint32_t) - 1); 4551 atomic_set_32((uint32_t *)addr, set << shift); 4552 #endif /* PAGE_SIZE */ 4553 } 4554 4555 static inline void 4556 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 4557 { 4558 4559 #if PAGE_SIZE == 32768 4560 atomic_clear_64((uint64_t *)bits, clear); 4561 #elif PAGE_SIZE == 16384 4562 atomic_clear_32((uint32_t *)bits, clear); 4563 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 4564 atomic_clear_16((uint16_t *)bits, clear); 4565 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 4566 atomic_clear_8((uint8_t *)bits, clear); 4567 #else /* PAGE_SIZE <= 8192 */ 4568 uintptr_t addr; 4569 int shift; 4570 4571 addr = (uintptr_t)bits; 4572 /* 4573 * Use a trick to perform a 32-bit atomic on the 4574 * containing aligned word, to not depend on the existence 4575 * of atomic_{set, clear}_{8, 16}. 4576 */ 4577 shift = addr & (sizeof(uint32_t) - 1); 4578 #if BYTE_ORDER == BIG_ENDIAN 4579 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4580 #else 4581 shift *= NBBY; 4582 #endif 4583 addr &= ~(sizeof(uint32_t) - 1); 4584 atomic_clear_32((uint32_t *)addr, clear << shift); 4585 #endif /* PAGE_SIZE */ 4586 } 4587 4588 /* 4589 * vm_page_set_valid_range: 4590 * 4591 * Sets portions of a page valid. The arguments are expected 4592 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4593 * of any partial chunks touched by the range. The invalid portion of 4594 * such chunks will be zeroed. 4595 * 4596 * (base + size) must be less then or equal to PAGE_SIZE. 4597 */ 4598 void 4599 vm_page_set_valid_range(vm_page_t m, int base, int size) 4600 { 4601 int endoff, frag; 4602 vm_page_bits_t pagebits; 4603 4604 vm_page_assert_busied(m); 4605 if (size == 0) /* handle degenerate case */ 4606 return; 4607 4608 /* 4609 * If the base is not DEV_BSIZE aligned and the valid 4610 * bit is clear, we have to zero out a portion of the 4611 * first block. 4612 */ 4613 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4614 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 4615 pmap_zero_page_area(m, frag, base - frag); 4616 4617 /* 4618 * If the ending offset is not DEV_BSIZE aligned and the 4619 * valid bit is clear, we have to zero out a portion of 4620 * the last block. 4621 */ 4622 endoff = base + size; 4623 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4624 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 4625 pmap_zero_page_area(m, endoff, 4626 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4627 4628 /* 4629 * Assert that no previously invalid block that is now being validated 4630 * is already dirty. 4631 */ 4632 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 4633 ("vm_page_set_valid_range: page %p is dirty", m)); 4634 4635 /* 4636 * Set valid bits inclusive of any overlap. 4637 */ 4638 pagebits = vm_page_bits(base, size); 4639 if (vm_page_xbusied(m)) 4640 m->valid |= pagebits; 4641 else 4642 vm_page_bits_set(m, &m->valid, pagebits); 4643 } 4644 4645 /* 4646 * Clear the given bits from the specified page's dirty field. 4647 */ 4648 static __inline void 4649 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 4650 { 4651 4652 vm_page_assert_busied(m); 4653 4654 /* 4655 * If the page is xbusied and not write mapped we are the 4656 * only thread that can modify dirty bits. Otherwise, The pmap 4657 * layer can call vm_page_dirty() without holding a distinguished 4658 * lock. The combination of page busy and atomic operations 4659 * suffice to guarantee consistency of the page dirty field. 4660 */ 4661 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 4662 m->dirty &= ~pagebits; 4663 else 4664 vm_page_bits_clear(m, &m->dirty, pagebits); 4665 } 4666 4667 /* 4668 * vm_page_set_validclean: 4669 * 4670 * Sets portions of a page valid and clean. The arguments are expected 4671 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4672 * of any partial chunks touched by the range. The invalid portion of 4673 * such chunks will be zero'd. 4674 * 4675 * (base + size) must be less then or equal to PAGE_SIZE. 4676 */ 4677 void 4678 vm_page_set_validclean(vm_page_t m, int base, int size) 4679 { 4680 vm_page_bits_t oldvalid, pagebits; 4681 int endoff, frag; 4682 4683 vm_page_assert_busied(m); 4684 if (size == 0) /* handle degenerate case */ 4685 return; 4686 4687 /* 4688 * If the base is not DEV_BSIZE aligned and the valid 4689 * bit is clear, we have to zero out a portion of the 4690 * first block. 4691 */ 4692 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4693 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 4694 pmap_zero_page_area(m, frag, base - frag); 4695 4696 /* 4697 * If the ending offset is not DEV_BSIZE aligned and the 4698 * valid bit is clear, we have to zero out a portion of 4699 * the last block. 4700 */ 4701 endoff = base + size; 4702 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4703 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 4704 pmap_zero_page_area(m, endoff, 4705 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4706 4707 /* 4708 * Set valid, clear dirty bits. If validating the entire 4709 * page we can safely clear the pmap modify bit. We also 4710 * use this opportunity to clear the PGA_NOSYNC flag. If a process 4711 * takes a write fault on a MAP_NOSYNC memory area the flag will 4712 * be set again. 4713 * 4714 * We set valid bits inclusive of any overlap, but we can only 4715 * clear dirty bits for DEV_BSIZE chunks that are fully within 4716 * the range. 4717 */ 4718 oldvalid = m->valid; 4719 pagebits = vm_page_bits(base, size); 4720 if (vm_page_xbusied(m)) 4721 m->valid |= pagebits; 4722 else 4723 vm_page_bits_set(m, &m->valid, pagebits); 4724 #if 0 /* NOT YET */ 4725 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 4726 frag = DEV_BSIZE - frag; 4727 base += frag; 4728 size -= frag; 4729 if (size < 0) 4730 size = 0; 4731 } 4732 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 4733 #endif 4734 if (base == 0 && size == PAGE_SIZE) { 4735 /* 4736 * The page can only be modified within the pmap if it is 4737 * mapped, and it can only be mapped if it was previously 4738 * fully valid. 4739 */ 4740 if (oldvalid == VM_PAGE_BITS_ALL) 4741 /* 4742 * Perform the pmap_clear_modify() first. Otherwise, 4743 * a concurrent pmap operation, such as 4744 * pmap_protect(), could clear a modification in the 4745 * pmap and set the dirty field on the page before 4746 * pmap_clear_modify() had begun and after the dirty 4747 * field was cleared here. 4748 */ 4749 pmap_clear_modify(m); 4750 m->dirty = 0; 4751 vm_page_aflag_clear(m, PGA_NOSYNC); 4752 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 4753 m->dirty &= ~pagebits; 4754 else 4755 vm_page_clear_dirty_mask(m, pagebits); 4756 } 4757 4758 void 4759 vm_page_clear_dirty(vm_page_t m, int base, int size) 4760 { 4761 4762 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 4763 } 4764 4765 /* 4766 * vm_page_set_invalid: 4767 * 4768 * Invalidates DEV_BSIZE'd chunks within a page. Both the 4769 * valid and dirty bits for the effected areas are cleared. 4770 */ 4771 void 4772 vm_page_set_invalid(vm_page_t m, int base, int size) 4773 { 4774 vm_page_bits_t bits; 4775 vm_object_t object; 4776 4777 /* 4778 * The object lock is required so that pages can't be mapped 4779 * read-only while we're in the process of invalidating them. 4780 */ 4781 object = m->object; 4782 VM_OBJECT_ASSERT_WLOCKED(object); 4783 vm_page_assert_busied(m); 4784 4785 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 4786 size >= object->un_pager.vnp.vnp_size) 4787 bits = VM_PAGE_BITS_ALL; 4788 else 4789 bits = vm_page_bits(base, size); 4790 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 4791 pmap_remove_all(m); 4792 KASSERT((bits == 0 && vm_page_all_valid(m)) || 4793 !pmap_page_is_mapped(m), 4794 ("vm_page_set_invalid: page %p is mapped", m)); 4795 if (vm_page_xbusied(m)) { 4796 m->valid &= ~bits; 4797 m->dirty &= ~bits; 4798 } else { 4799 vm_page_bits_clear(m, &m->valid, bits); 4800 vm_page_bits_clear(m, &m->dirty, bits); 4801 } 4802 } 4803 4804 /* 4805 * vm_page_invalid: 4806 * 4807 * Invalidates the entire page. The page must be busy, unmapped, and 4808 * the enclosing object must be locked. The object locks protects 4809 * against concurrent read-only pmap enter which is done without 4810 * busy. 4811 */ 4812 void 4813 vm_page_invalid(vm_page_t m) 4814 { 4815 4816 vm_page_assert_busied(m); 4817 VM_OBJECT_ASSERT_LOCKED(m->object); 4818 MPASS(!pmap_page_is_mapped(m)); 4819 4820 if (vm_page_xbusied(m)) 4821 m->valid = 0; 4822 else 4823 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 4824 } 4825 4826 /* 4827 * vm_page_zero_invalid() 4828 * 4829 * The kernel assumes that the invalid portions of a page contain 4830 * garbage, but such pages can be mapped into memory by user code. 4831 * When this occurs, we must zero out the non-valid portions of the 4832 * page so user code sees what it expects. 4833 * 4834 * Pages are most often semi-valid when the end of a file is mapped 4835 * into memory and the file's size is not page aligned. 4836 */ 4837 void 4838 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 4839 { 4840 int b; 4841 int i; 4842 4843 /* 4844 * Scan the valid bits looking for invalid sections that 4845 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 4846 * valid bit may be set ) have already been zeroed by 4847 * vm_page_set_validclean(). 4848 */ 4849 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 4850 if (i == (PAGE_SIZE / DEV_BSIZE) || 4851 (m->valid & ((vm_page_bits_t)1 << i))) { 4852 if (i > b) { 4853 pmap_zero_page_area(m, 4854 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 4855 } 4856 b = i + 1; 4857 } 4858 } 4859 4860 /* 4861 * setvalid is TRUE when we can safely set the zero'd areas 4862 * as being valid. We can do this if there are no cache consistancy 4863 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 4864 */ 4865 if (setvalid) 4866 vm_page_valid(m); 4867 } 4868 4869 /* 4870 * vm_page_is_valid: 4871 * 4872 * Is (partial) page valid? Note that the case where size == 0 4873 * will return FALSE in the degenerate case where the page is 4874 * entirely invalid, and TRUE otherwise. 4875 * 4876 * Some callers envoke this routine without the busy lock held and 4877 * handle races via higher level locks. Typical callers should 4878 * hold a busy lock to prevent invalidation. 4879 */ 4880 int 4881 vm_page_is_valid(vm_page_t m, int base, int size) 4882 { 4883 vm_page_bits_t bits; 4884 4885 bits = vm_page_bits(base, size); 4886 return (m->valid != 0 && (m->valid & bits) == bits); 4887 } 4888 4889 /* 4890 * Returns true if all of the specified predicates are true for the entire 4891 * (super)page and false otherwise. 4892 */ 4893 bool 4894 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 4895 { 4896 vm_object_t object; 4897 int i, npages; 4898 4899 object = m->object; 4900 if (skip_m != NULL && skip_m->object != object) 4901 return (false); 4902 VM_OBJECT_ASSERT_LOCKED(object); 4903 npages = atop(pagesizes[m->psind]); 4904 4905 /* 4906 * The physically contiguous pages that make up a superpage, i.e., a 4907 * page with a page size index ("psind") greater than zero, will 4908 * occupy adjacent entries in vm_page_array[]. 4909 */ 4910 for (i = 0; i < npages; i++) { 4911 /* Always test object consistency, including "skip_m". */ 4912 if (m[i].object != object) 4913 return (false); 4914 if (&m[i] == skip_m) 4915 continue; 4916 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 4917 return (false); 4918 if ((flags & PS_ALL_DIRTY) != 0) { 4919 /* 4920 * Calling vm_page_test_dirty() or pmap_is_modified() 4921 * might stop this case from spuriously returning 4922 * "false". However, that would require a write lock 4923 * on the object containing "m[i]". 4924 */ 4925 if (m[i].dirty != VM_PAGE_BITS_ALL) 4926 return (false); 4927 } 4928 if ((flags & PS_ALL_VALID) != 0 && 4929 m[i].valid != VM_PAGE_BITS_ALL) 4930 return (false); 4931 } 4932 return (true); 4933 } 4934 4935 /* 4936 * Set the page's dirty bits if the page is modified. 4937 */ 4938 void 4939 vm_page_test_dirty(vm_page_t m) 4940 { 4941 4942 vm_page_assert_busied(m); 4943 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 4944 vm_page_dirty(m); 4945 } 4946 4947 void 4948 vm_page_valid(vm_page_t m) 4949 { 4950 4951 vm_page_assert_busied(m); 4952 if (vm_page_xbusied(m)) 4953 m->valid = VM_PAGE_BITS_ALL; 4954 else 4955 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 4956 } 4957 4958 void 4959 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 4960 { 4961 4962 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 4963 } 4964 4965 void 4966 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 4967 { 4968 4969 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 4970 } 4971 4972 int 4973 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 4974 { 4975 4976 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 4977 } 4978 4979 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 4980 void 4981 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 4982 { 4983 4984 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 4985 } 4986 4987 void 4988 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 4989 { 4990 4991 mtx_assert_(vm_page_lockptr(m), a, file, line); 4992 } 4993 #endif 4994 4995 #ifdef INVARIANTS 4996 void 4997 vm_page_object_busy_assert(vm_page_t m) 4998 { 4999 5000 /* 5001 * Certain of the page's fields may only be modified by the 5002 * holder of a page or object busy. 5003 */ 5004 if (m->object != NULL && !vm_page_busied(m)) 5005 VM_OBJECT_ASSERT_BUSY(m->object); 5006 } 5007 5008 void 5009 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 5010 { 5011 5012 if ((bits & PGA_WRITEABLE) == 0) 5013 return; 5014 5015 /* 5016 * The PGA_WRITEABLE flag can only be set if the page is 5017 * managed, is exclusively busied or the object is locked. 5018 * Currently, this flag is only set by pmap_enter(). 5019 */ 5020 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5021 ("PGA_WRITEABLE on unmanaged page")); 5022 if (!vm_page_xbusied(m)) 5023 VM_OBJECT_ASSERT_BUSY(m->object); 5024 } 5025 #endif 5026 5027 #include "opt_ddb.h" 5028 #ifdef DDB 5029 #include <sys/kernel.h> 5030 5031 #include <ddb/ddb.h> 5032 5033 DB_SHOW_COMMAND(page, vm_page_print_page_info) 5034 { 5035 5036 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5037 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5038 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5039 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5040 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5041 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5042 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5043 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5044 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5045 } 5046 5047 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 5048 { 5049 int dom; 5050 5051 db_printf("pq_free %d\n", vm_free_count()); 5052 for (dom = 0; dom < vm_ndomains; dom++) { 5053 db_printf( 5054 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5055 dom, 5056 vm_dom[dom].vmd_page_count, 5057 vm_dom[dom].vmd_free_count, 5058 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5059 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5060 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5061 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5062 } 5063 } 5064 5065 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5066 { 5067 vm_page_t m; 5068 boolean_t phys, virt; 5069 5070 if (!have_addr) { 5071 db_printf("show pginfo addr\n"); 5072 return; 5073 } 5074 5075 phys = strchr(modif, 'p') != NULL; 5076 virt = strchr(modif, 'v') != NULL; 5077 if (virt) 5078 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5079 else if (phys) 5080 m = PHYS_TO_VM_PAGE(addr); 5081 else 5082 m = (vm_page_t)addr; 5083 db_printf( 5084 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" 5085 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5086 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5087 m->queue, m->ref_count, m->aflags, m->oflags, 5088 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 5089 } 5090 #endif /* DDB */ 5091