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 cache, 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 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 226 pgcache = &vmd->vmd_pgcache[pool]; 227 pgcache->domain = domain; 228 pgcache->pool = pool; 229 pgcache->zone = uma_zcache_create("vm pgcache", 230 PAGE_SIZE, NULL, NULL, NULL, NULL, 231 vm_page_zone_import, vm_page_zone_release, pgcache, 232 UMA_ZONE_VM); 233 234 /* 235 * Limit each pool's zone to 0.1% of the pages in the 236 * domain. 237 */ 238 cache = maxcache != 0 ? maxcache : 239 vmd->vmd_page_count / 1000; 240 uma_zone_set_maxcache(pgcache->zone, cache); 241 } 242 } 243 } 244 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 245 246 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 247 #if PAGE_SIZE == 32768 248 #ifdef CTASSERT 249 CTASSERT(sizeof(u_long) >= 8); 250 #endif 251 #endif 252 253 /* 254 * vm_set_page_size: 255 * 256 * Sets the page size, perhaps based upon the memory 257 * size. Must be called before any use of page-size 258 * dependent functions. 259 */ 260 void 261 vm_set_page_size(void) 262 { 263 if (vm_cnt.v_page_size == 0) 264 vm_cnt.v_page_size = PAGE_SIZE; 265 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 266 panic("vm_set_page_size: page size not a power of two"); 267 } 268 269 /* 270 * vm_page_blacklist_next: 271 * 272 * Find the next entry in the provided string of blacklist 273 * addresses. Entries are separated by space, comma, or newline. 274 * If an invalid integer is encountered then the rest of the 275 * string is skipped. Updates the list pointer to the next 276 * character, or NULL if the string is exhausted or invalid. 277 */ 278 static vm_paddr_t 279 vm_page_blacklist_next(char **list, char *end) 280 { 281 vm_paddr_t bad; 282 char *cp, *pos; 283 284 if (list == NULL || *list == NULL) 285 return (0); 286 if (**list =='\0') { 287 *list = NULL; 288 return (0); 289 } 290 291 /* 292 * If there's no end pointer then the buffer is coming from 293 * the kenv and we know it's null-terminated. 294 */ 295 if (end == NULL) 296 end = *list + strlen(*list); 297 298 /* Ensure that strtoq() won't walk off the end */ 299 if (*end != '\0') { 300 if (*end == '\n' || *end == ' ' || *end == ',') 301 *end = '\0'; 302 else { 303 printf("Blacklist not terminated, skipping\n"); 304 *list = NULL; 305 return (0); 306 } 307 } 308 309 for (pos = *list; *pos != '\0'; pos = cp) { 310 bad = strtoq(pos, &cp, 0); 311 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 312 if (bad == 0) { 313 if (++cp < end) 314 continue; 315 else 316 break; 317 } 318 } else 319 break; 320 if (*cp == '\0' || ++cp >= end) 321 *list = NULL; 322 else 323 *list = cp; 324 return (trunc_page(bad)); 325 } 326 printf("Garbage in RAM blacklist, skipping\n"); 327 *list = NULL; 328 return (0); 329 } 330 331 bool 332 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 333 { 334 struct vm_domain *vmd; 335 vm_page_t m; 336 int ret; 337 338 m = vm_phys_paddr_to_vm_page(pa); 339 if (m == NULL) 340 return (true); /* page does not exist, no failure */ 341 342 vmd = vm_pagequeue_domain(m); 343 vm_domain_free_lock(vmd); 344 ret = vm_phys_unfree_page(m); 345 vm_domain_free_unlock(vmd); 346 if (ret != 0) { 347 vm_domain_freecnt_inc(vmd, -1); 348 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 349 if (verbose) 350 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 351 } 352 return (ret); 353 } 354 355 /* 356 * vm_page_blacklist_check: 357 * 358 * Iterate through the provided string of blacklist addresses, pulling 359 * each entry out of the physical allocator free list and putting it 360 * onto a list for reporting via the vm.page_blacklist sysctl. 361 */ 362 static void 363 vm_page_blacklist_check(char *list, char *end) 364 { 365 vm_paddr_t pa; 366 char *next; 367 368 next = list; 369 while (next != NULL) { 370 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 371 continue; 372 vm_page_blacklist_add(pa, bootverbose); 373 } 374 } 375 376 /* 377 * vm_page_blacklist_load: 378 * 379 * Search for a special module named "ram_blacklist". It'll be a 380 * plain text file provided by the user via the loader directive 381 * of the same name. 382 */ 383 static void 384 vm_page_blacklist_load(char **list, char **end) 385 { 386 void *mod; 387 u_char *ptr; 388 u_int len; 389 390 mod = NULL; 391 ptr = NULL; 392 393 mod = preload_search_by_type("ram_blacklist"); 394 if (mod != NULL) { 395 ptr = preload_fetch_addr(mod); 396 len = preload_fetch_size(mod); 397 } 398 *list = ptr; 399 if (ptr != NULL) 400 *end = ptr + len; 401 else 402 *end = NULL; 403 return; 404 } 405 406 static int 407 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 408 { 409 vm_page_t m; 410 struct sbuf sbuf; 411 int error, first; 412 413 first = 1; 414 error = sysctl_wire_old_buffer(req, 0); 415 if (error != 0) 416 return (error); 417 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 418 TAILQ_FOREACH(m, &blacklist_head, listq) { 419 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 420 (uintmax_t)m->phys_addr); 421 first = 0; 422 } 423 error = sbuf_finish(&sbuf); 424 sbuf_delete(&sbuf); 425 return (error); 426 } 427 428 /* 429 * Initialize a dummy page for use in scans of the specified paging queue. 430 * In principle, this function only needs to set the flag PG_MARKER. 431 * Nonetheless, it write busies the page as a safety precaution. 432 */ 433 static void 434 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 435 { 436 437 bzero(marker, sizeof(*marker)); 438 marker->flags = PG_MARKER; 439 marker->aflags = aflags; 440 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 441 marker->queue = queue; 442 } 443 444 static void 445 vm_page_domain_init(int domain) 446 { 447 struct vm_domain *vmd; 448 struct vm_pagequeue *pq; 449 int i; 450 451 vmd = VM_DOMAIN(domain); 452 bzero(vmd, sizeof(*vmd)); 453 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 454 "vm inactive pagequeue"; 455 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 456 "vm active pagequeue"; 457 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 458 "vm laundry pagequeue"; 459 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 460 "vm unswappable pagequeue"; 461 vmd->vmd_domain = domain; 462 vmd->vmd_page_count = 0; 463 vmd->vmd_free_count = 0; 464 vmd->vmd_segs = 0; 465 vmd->vmd_oom = FALSE; 466 for (i = 0; i < PQ_COUNT; i++) { 467 pq = &vmd->vmd_pagequeues[i]; 468 TAILQ_INIT(&pq->pq_pl); 469 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 470 MTX_DEF | MTX_DUPOK); 471 pq->pq_pdpages = 0; 472 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 473 } 474 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 475 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 476 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 477 478 /* 479 * inacthead is used to provide FIFO ordering for LRU-bypassing 480 * insertions. 481 */ 482 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 483 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 484 &vmd->vmd_inacthead, plinks.q); 485 486 /* 487 * The clock pages are used to implement active queue scanning without 488 * requeues. Scans start at clock[0], which is advanced after the scan 489 * ends. When the two clock hands meet, they are reset and scanning 490 * resumes from the head of the queue. 491 */ 492 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 493 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 494 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 495 &vmd->vmd_clock[0], plinks.q); 496 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 497 &vmd->vmd_clock[1], plinks.q); 498 } 499 500 /* 501 * Initialize a physical page in preparation for adding it to the free 502 * lists. 503 */ 504 static void 505 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 506 { 507 508 m->object = NULL; 509 m->ref_count = 0; 510 m->busy_lock = VPB_UNBUSIED; 511 m->flags = m->aflags = 0; 512 m->phys_addr = pa; 513 m->queue = PQ_NONE; 514 m->psind = 0; 515 m->segind = segind; 516 m->order = VM_NFREEORDER; 517 m->pool = VM_FREEPOOL_DEFAULT; 518 m->valid = m->dirty = 0; 519 pmap_page_init(m); 520 } 521 522 #ifndef PMAP_HAS_PAGE_ARRAY 523 static vm_paddr_t 524 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 525 { 526 vm_paddr_t new_end; 527 528 /* 529 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 530 * However, because this page is allocated from KVM, out-of-bounds 531 * accesses using the direct map will not be trapped. 532 */ 533 *vaddr += PAGE_SIZE; 534 535 /* 536 * Allocate physical memory for the page structures, and map it. 537 */ 538 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 539 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 540 VM_PROT_READ | VM_PROT_WRITE); 541 vm_page_array_size = page_range; 542 543 return (new_end); 544 } 545 #endif 546 547 /* 548 * vm_page_startup: 549 * 550 * Initializes the resident memory module. Allocates physical memory for 551 * bootstrapping UMA and some data structures that are used to manage 552 * physical pages. Initializes these structures, and populates the free 553 * page queues. 554 */ 555 vm_offset_t 556 vm_page_startup(vm_offset_t vaddr) 557 { 558 struct vm_phys_seg *seg; 559 vm_page_t m; 560 char *list, *listend; 561 vm_offset_t mapped; 562 vm_paddr_t end, high_avail, low_avail, new_end, size; 563 vm_paddr_t page_range __unused; 564 vm_paddr_t last_pa, pa; 565 u_long pagecount; 566 int biggestone, i, segind; 567 #ifdef WITNESS 568 int witness_size; 569 #endif 570 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 571 long ii; 572 #endif 573 574 vaddr = round_page(vaddr); 575 576 vm_phys_early_startup(); 577 biggestone = vm_phys_avail_largest(); 578 end = phys_avail[biggestone+1]; 579 580 /* 581 * Initialize the page and queue locks. 582 */ 583 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 584 for (i = 0; i < PA_LOCK_COUNT; i++) 585 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 586 for (i = 0; i < vm_ndomains; i++) 587 vm_page_domain_init(i); 588 589 /* 590 * Allocate memory for use when boot strapping the kernel memory 591 * allocator. Tell UMA how many zones we are going to create 592 * before going fully functional. UMA will add its zones. 593 * 594 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, 595 * KMAP ENTRY, MAP ENTRY, VMSPACE. 596 */ 597 boot_pages = uma_startup_count(8); 598 599 #ifndef UMA_MD_SMALL_ALLOC 600 /* vmem_startup() calls uma_prealloc(). */ 601 boot_pages += vmem_startup_count(); 602 /* vm_map_startup() calls uma_prealloc(). */ 603 boot_pages += howmany(MAX_KMAP, 604 slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR)); 605 606 /* 607 * Before going fully functional kmem_init() does allocation 608 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". 609 */ 610 boot_pages += 2; 611 #endif 612 /* 613 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 614 * manually fetch the value. 615 */ 616 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 617 new_end = end - (boot_pages * UMA_SLAB_SIZE); 618 new_end = trunc_page(new_end); 619 mapped = pmap_map(&vaddr, new_end, end, 620 VM_PROT_READ | VM_PROT_WRITE); 621 bzero((void *)mapped, end - new_end); 622 uma_startup((void *)mapped, boot_pages); 623 624 #ifdef WITNESS 625 witness_size = round_page(witness_startup_count()); 626 new_end -= witness_size; 627 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 628 VM_PROT_READ | VM_PROT_WRITE); 629 bzero((void *)mapped, witness_size); 630 witness_startup((void *)mapped); 631 #endif 632 633 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 634 defined(__i386__) || defined(__mips__) || defined(__riscv) || \ 635 defined(__powerpc64__) 636 /* 637 * Allocate a bitmap to indicate that a random physical page 638 * needs to be included in a minidump. 639 * 640 * The amd64 port needs this to indicate which direct map pages 641 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 642 * 643 * However, i386 still needs this workspace internally within the 644 * minidump code. In theory, they are not needed on i386, but are 645 * included should the sf_buf code decide to use them. 646 */ 647 last_pa = 0; 648 for (i = 0; dump_avail[i + 1] != 0; i += 2) 649 if (dump_avail[i + 1] > last_pa) 650 last_pa = dump_avail[i + 1]; 651 page_range = last_pa / PAGE_SIZE; 652 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 653 new_end -= vm_page_dump_size; 654 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 655 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 656 bzero((void *)vm_page_dump, vm_page_dump_size); 657 #else 658 (void)last_pa; 659 #endif 660 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 661 defined(__riscv) || defined(__powerpc64__) 662 /* 663 * Include the UMA bootstrap pages, witness pages and vm_page_dump 664 * in a crash dump. When pmap_map() uses the direct map, they are 665 * not automatically included. 666 */ 667 for (pa = new_end; pa < end; pa += PAGE_SIZE) 668 dump_add_page(pa); 669 #endif 670 phys_avail[biggestone + 1] = new_end; 671 #ifdef __amd64__ 672 /* 673 * Request that the physical pages underlying the message buffer be 674 * included in a crash dump. Since the message buffer is accessed 675 * through the direct map, they are not automatically included. 676 */ 677 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 678 last_pa = pa + round_page(msgbufsize); 679 while (pa < last_pa) { 680 dump_add_page(pa); 681 pa += PAGE_SIZE; 682 } 683 #endif 684 /* 685 * Compute the number of pages of memory that will be available for 686 * use, taking into account the overhead of a page structure per page. 687 * In other words, solve 688 * "available physical memory" - round_page(page_range * 689 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 690 * for page_range. 691 */ 692 low_avail = phys_avail[0]; 693 high_avail = phys_avail[1]; 694 for (i = 0; i < vm_phys_nsegs; i++) { 695 if (vm_phys_segs[i].start < low_avail) 696 low_avail = vm_phys_segs[i].start; 697 if (vm_phys_segs[i].end > high_avail) 698 high_avail = vm_phys_segs[i].end; 699 } 700 /* Skip the first chunk. It is already accounted for. */ 701 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 702 if (phys_avail[i] < low_avail) 703 low_avail = phys_avail[i]; 704 if (phys_avail[i + 1] > high_avail) 705 high_avail = phys_avail[i + 1]; 706 } 707 first_page = low_avail / PAGE_SIZE; 708 #ifdef VM_PHYSSEG_SPARSE 709 size = 0; 710 for (i = 0; i < vm_phys_nsegs; i++) 711 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 712 for (i = 0; phys_avail[i + 1] != 0; i += 2) 713 size += phys_avail[i + 1] - phys_avail[i]; 714 #elif defined(VM_PHYSSEG_DENSE) 715 size = high_avail - low_avail; 716 #else 717 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 718 #endif 719 720 #ifdef PMAP_HAS_PAGE_ARRAY 721 pmap_page_array_startup(size / PAGE_SIZE); 722 biggestone = vm_phys_avail_largest(); 723 end = new_end = phys_avail[biggestone + 1]; 724 #else 725 #ifdef VM_PHYSSEG_DENSE 726 /* 727 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 728 * the overhead of a page structure per page only if vm_page_array is 729 * allocated from the last physical memory chunk. Otherwise, we must 730 * allocate page structures representing the physical memory 731 * underlying vm_page_array, even though they will not be used. 732 */ 733 if (new_end != high_avail) 734 page_range = size / PAGE_SIZE; 735 else 736 #endif 737 { 738 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 739 740 /* 741 * If the partial bytes remaining are large enough for 742 * a page (PAGE_SIZE) without a corresponding 743 * 'struct vm_page', then new_end will contain an 744 * extra page after subtracting the length of the VM 745 * page array. Compensate by subtracting an extra 746 * page from new_end. 747 */ 748 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 749 if (new_end == high_avail) 750 high_avail -= PAGE_SIZE; 751 new_end -= PAGE_SIZE; 752 } 753 } 754 end = new_end; 755 new_end = vm_page_array_alloc(&vaddr, end, page_range); 756 #endif 757 758 #if VM_NRESERVLEVEL > 0 759 /* 760 * Allocate physical memory for the reservation management system's 761 * data structures, and map it. 762 */ 763 new_end = vm_reserv_startup(&vaddr, new_end); 764 #endif 765 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 766 defined(__riscv) || defined(__powerpc64__) 767 /* 768 * Include vm_page_array and vm_reserv_array in a crash dump. 769 */ 770 for (pa = new_end; pa < end; pa += PAGE_SIZE) 771 dump_add_page(pa); 772 #endif 773 phys_avail[biggestone + 1] = new_end; 774 775 /* 776 * Add physical memory segments corresponding to the available 777 * physical pages. 778 */ 779 for (i = 0; phys_avail[i + 1] != 0; i += 2) 780 if (vm_phys_avail_size(i) != 0) 781 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 782 783 /* 784 * Initialize the physical memory allocator. 785 */ 786 vm_phys_init(); 787 788 /* 789 * Initialize the page structures and add every available page to the 790 * physical memory allocator's free lists. 791 */ 792 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 793 for (ii = 0; ii < vm_page_array_size; ii++) { 794 m = &vm_page_array[ii]; 795 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 796 m->flags = PG_FICTITIOUS; 797 } 798 #endif 799 vm_cnt.v_page_count = 0; 800 for (segind = 0; segind < vm_phys_nsegs; segind++) { 801 seg = &vm_phys_segs[segind]; 802 for (m = seg->first_page, pa = seg->start; pa < seg->end; 803 m++, pa += PAGE_SIZE) 804 vm_page_init_page(m, pa, segind); 805 806 /* 807 * Add the segment to the free lists only if it is covered by 808 * one of the ranges in phys_avail. Because we've added the 809 * ranges to the vm_phys_segs array, we can assume that each 810 * segment is either entirely contained in one of the ranges, 811 * or doesn't overlap any of them. 812 */ 813 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 814 struct vm_domain *vmd; 815 816 if (seg->start < phys_avail[i] || 817 seg->end > phys_avail[i + 1]) 818 continue; 819 820 m = seg->first_page; 821 pagecount = (u_long)atop(seg->end - seg->start); 822 823 vmd = VM_DOMAIN(seg->domain); 824 vm_domain_free_lock(vmd); 825 vm_phys_enqueue_contig(m, pagecount); 826 vm_domain_free_unlock(vmd); 827 vm_domain_freecnt_inc(vmd, pagecount); 828 vm_cnt.v_page_count += (u_int)pagecount; 829 830 vmd = VM_DOMAIN(seg->domain); 831 vmd->vmd_page_count += (u_int)pagecount; 832 vmd->vmd_segs |= 1UL << m->segind; 833 break; 834 } 835 } 836 837 /* 838 * Remove blacklisted pages from the physical memory allocator. 839 */ 840 TAILQ_INIT(&blacklist_head); 841 vm_page_blacklist_load(&list, &listend); 842 vm_page_blacklist_check(list, listend); 843 844 list = kern_getenv("vm.blacklist"); 845 vm_page_blacklist_check(list, NULL); 846 847 freeenv(list); 848 #if VM_NRESERVLEVEL > 0 849 /* 850 * Initialize the reservation management system. 851 */ 852 vm_reserv_init(); 853 #endif 854 855 return (vaddr); 856 } 857 858 void 859 vm_page_reference(vm_page_t m) 860 { 861 862 vm_page_aflag_set(m, PGA_REFERENCED); 863 } 864 865 static bool 866 vm_page_acquire_flags(vm_page_t m, int allocflags) 867 { 868 bool locked; 869 870 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 871 locked = vm_page_trysbusy(m); 872 else 873 locked = vm_page_tryxbusy(m); 874 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 875 vm_page_wire(m); 876 return (locked); 877 } 878 879 static bool 880 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wchan, 881 int allocflags) 882 { 883 884 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 885 return (false); 886 /* 887 * Reference the page before unlocking and 888 * sleeping so that the page daemon is less 889 * likely to reclaim it. 890 */ 891 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 892 vm_page_aflag_set(m, PGA_REFERENCED); 893 vm_page_busy_sleep(m, wchan, (allocflags & 894 VM_ALLOC_IGN_SBUSY) != 0); 895 VM_OBJECT_WLOCK(object); 896 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 897 return (false); 898 return (true); 899 } 900 901 /* 902 * vm_page_busy_acquire: 903 * 904 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 905 * and drop the object lock if necessary. 906 */ 907 bool 908 vm_page_busy_acquire(vm_page_t m, int allocflags) 909 { 910 vm_object_t obj; 911 bool locked; 912 913 /* 914 * The page-specific object must be cached because page 915 * identity can change during the sleep, causing the 916 * re-lock of a different object. 917 * It is assumed that a reference to the object is already 918 * held by the callers. 919 */ 920 obj = m->object; 921 for (;;) { 922 if (vm_page_acquire_flags(m, allocflags)) 923 return (true); 924 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 925 return (false); 926 if (obj != NULL) 927 locked = VM_OBJECT_WOWNED(obj); 928 else 929 locked = false; 930 MPASS(locked || vm_page_wired(m)); 931 _vm_page_busy_sleep(obj, m, "vmpba", 932 (allocflags & VM_ALLOC_SBUSY) != 0, locked); 933 if (locked) 934 VM_OBJECT_WLOCK(obj); 935 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 936 return (false); 937 KASSERT(m->object == obj || m->object == NULL, 938 ("vm_page_busy_acquire: page %p does not belong to %p", 939 m, obj)); 940 } 941 } 942 943 /* 944 * vm_page_busy_downgrade: 945 * 946 * Downgrade an exclusive busy page into a single shared busy page. 947 */ 948 void 949 vm_page_busy_downgrade(vm_page_t m) 950 { 951 u_int x; 952 953 vm_page_assert_xbusied(m); 954 955 x = m->busy_lock; 956 for (;;) { 957 if (atomic_fcmpset_rel_int(&m->busy_lock, 958 &x, VPB_SHARERS_WORD(1))) 959 break; 960 } 961 if ((x & VPB_BIT_WAITERS) != 0) 962 wakeup(m); 963 } 964 965 /* 966 * 967 * vm_page_busy_tryupgrade: 968 * 969 * Attempt to upgrade a single shared busy into an exclusive busy. 970 */ 971 int 972 vm_page_busy_tryupgrade(vm_page_t m) 973 { 974 u_int ce, x; 975 976 vm_page_assert_sbusied(m); 977 978 x = m->busy_lock; 979 ce = VPB_CURTHREAD_EXCLUSIVE; 980 for (;;) { 981 if (VPB_SHARERS(x) > 1) 982 return (0); 983 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 984 ("vm_page_busy_tryupgrade: invalid lock state")); 985 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 986 ce | (x & VPB_BIT_WAITERS))) 987 continue; 988 return (1); 989 } 990 } 991 992 /* 993 * vm_page_sbusied: 994 * 995 * Return a positive value if the page is shared busied, 0 otherwise. 996 */ 997 int 998 vm_page_sbusied(vm_page_t m) 999 { 1000 u_int x; 1001 1002 x = m->busy_lock; 1003 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1004 } 1005 1006 /* 1007 * vm_page_sunbusy: 1008 * 1009 * Shared unbusy a page. 1010 */ 1011 void 1012 vm_page_sunbusy(vm_page_t m) 1013 { 1014 u_int x; 1015 1016 vm_page_assert_sbusied(m); 1017 1018 x = m->busy_lock; 1019 for (;;) { 1020 if (VPB_SHARERS(x) > 1) { 1021 if (atomic_fcmpset_int(&m->busy_lock, &x, 1022 x - VPB_ONE_SHARER)) 1023 break; 1024 continue; 1025 } 1026 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1027 ("vm_page_sunbusy: invalid lock state")); 1028 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1029 continue; 1030 if ((x & VPB_BIT_WAITERS) == 0) 1031 break; 1032 wakeup(m); 1033 break; 1034 } 1035 } 1036 1037 /* 1038 * vm_page_busy_sleep: 1039 * 1040 * Sleep if the page is busy, using the page pointer as wchan. 1041 * This is used to implement the hard-path of busying mechanism. 1042 * 1043 * If nonshared is true, sleep only if the page is xbusy. 1044 * 1045 * The object lock must be held on entry and will be released on exit. 1046 */ 1047 void 1048 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 1049 { 1050 vm_object_t obj; 1051 1052 obj = m->object; 1053 VM_OBJECT_ASSERT_LOCKED(obj); 1054 vm_page_lock_assert(m, MA_NOTOWNED); 1055 1056 _vm_page_busy_sleep(obj, m, wmesg, nonshared, true); 1057 } 1058 1059 static void 1060 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, 1061 bool nonshared, bool locked) 1062 { 1063 u_int x; 1064 1065 /* 1066 * If the object is busy we must wait for that to drain to zero 1067 * before trying the page again. 1068 */ 1069 if (obj != NULL && vm_object_busied(obj)) { 1070 if (locked) 1071 VM_OBJECT_DROP(obj); 1072 vm_object_busy_wait(obj, wmesg); 1073 return; 1074 } 1075 sleepq_lock(m); 1076 x = m->busy_lock; 1077 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 1078 ((x & VPB_BIT_WAITERS) == 0 && 1079 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 1080 if (locked) 1081 VM_OBJECT_DROP(obj); 1082 sleepq_release(m); 1083 return; 1084 } 1085 if (locked) 1086 VM_OBJECT_DROP(obj); 1087 DROP_GIANT(); 1088 sleepq_add(m, NULL, wmesg, 0, 0); 1089 sleepq_wait(m, PVM); 1090 PICKUP_GIANT(); 1091 } 1092 1093 /* 1094 * vm_page_trysbusy: 1095 * 1096 * Try to shared busy a page. 1097 * If the operation succeeds 1 is returned otherwise 0. 1098 * The operation never sleeps. 1099 */ 1100 int 1101 vm_page_trysbusy(vm_page_t m) 1102 { 1103 vm_object_t obj; 1104 u_int x; 1105 1106 obj = m->object; 1107 x = m->busy_lock; 1108 for (;;) { 1109 if ((x & VPB_BIT_SHARED) == 0) 1110 return (0); 1111 /* 1112 * Reduce the window for transient busies that will trigger 1113 * false negatives in vm_page_ps_test(). 1114 */ 1115 if (obj != NULL && vm_object_busied(obj)) 1116 return (0); 1117 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1118 x + VPB_ONE_SHARER)) 1119 break; 1120 } 1121 1122 /* Refetch the object now that we're guaranteed that it is stable. */ 1123 obj = m->object; 1124 if (obj != NULL && vm_object_busied(obj)) { 1125 vm_page_sunbusy(m); 1126 return (0); 1127 } 1128 return (1); 1129 } 1130 1131 /* 1132 * vm_page_tryxbusy: 1133 * 1134 * Try to exclusive busy a page. 1135 * If the operation succeeds 1 is returned otherwise 0. 1136 * The operation never sleeps. 1137 */ 1138 int 1139 vm_page_tryxbusy(vm_page_t m) 1140 { 1141 vm_object_t obj; 1142 1143 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, 1144 VPB_CURTHREAD_EXCLUSIVE) == 0) 1145 return (0); 1146 1147 obj = m->object; 1148 if (obj != NULL && vm_object_busied(obj)) { 1149 vm_page_xunbusy(m); 1150 return (0); 1151 } 1152 return (1); 1153 } 1154 1155 static void 1156 vm_page_xunbusy_hard_tail(vm_page_t m) 1157 { 1158 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1159 /* Wake the waiter. */ 1160 wakeup(m); 1161 } 1162 1163 /* 1164 * vm_page_xunbusy_hard: 1165 * 1166 * Called when unbusy has failed because there is a waiter. 1167 */ 1168 void 1169 vm_page_xunbusy_hard(vm_page_t m) 1170 { 1171 vm_page_assert_xbusied(m); 1172 vm_page_xunbusy_hard_tail(m); 1173 } 1174 1175 void 1176 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1177 { 1178 vm_page_assert_xbusied_unchecked(m); 1179 vm_page_xunbusy_hard_tail(m); 1180 } 1181 1182 /* 1183 * Avoid releasing and reacquiring the same page lock. 1184 */ 1185 void 1186 vm_page_change_lock(vm_page_t m, struct mtx **mtx) 1187 { 1188 struct mtx *mtx1; 1189 1190 mtx1 = vm_page_lockptr(m); 1191 if (*mtx == mtx1) 1192 return; 1193 if (*mtx != NULL) 1194 mtx_unlock(*mtx); 1195 *mtx = mtx1; 1196 mtx_lock(mtx1); 1197 } 1198 1199 /* 1200 * vm_page_unhold_pages: 1201 * 1202 * Unhold each of the pages that is referenced by the given array. 1203 */ 1204 void 1205 vm_page_unhold_pages(vm_page_t *ma, int count) 1206 { 1207 1208 for (; count != 0; count--) { 1209 vm_page_unwire(*ma, PQ_ACTIVE); 1210 ma++; 1211 } 1212 } 1213 1214 vm_page_t 1215 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1216 { 1217 vm_page_t m; 1218 1219 #ifdef VM_PHYSSEG_SPARSE 1220 m = vm_phys_paddr_to_vm_page(pa); 1221 if (m == NULL) 1222 m = vm_phys_fictitious_to_vm_page(pa); 1223 return (m); 1224 #elif defined(VM_PHYSSEG_DENSE) 1225 long pi; 1226 1227 pi = atop(pa); 1228 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1229 m = &vm_page_array[pi - first_page]; 1230 return (m); 1231 } 1232 return (vm_phys_fictitious_to_vm_page(pa)); 1233 #else 1234 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1235 #endif 1236 } 1237 1238 /* 1239 * vm_page_getfake: 1240 * 1241 * Create a fictitious page with the specified physical address and 1242 * memory attribute. The memory attribute is the only the machine- 1243 * dependent aspect of a fictitious page that must be initialized. 1244 */ 1245 vm_page_t 1246 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1247 { 1248 vm_page_t m; 1249 1250 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1251 vm_page_initfake(m, paddr, memattr); 1252 return (m); 1253 } 1254 1255 void 1256 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1257 { 1258 1259 if ((m->flags & PG_FICTITIOUS) != 0) { 1260 /* 1261 * The page's memattr might have changed since the 1262 * previous initialization. Update the pmap to the 1263 * new memattr. 1264 */ 1265 goto memattr; 1266 } 1267 m->phys_addr = paddr; 1268 m->queue = PQ_NONE; 1269 /* Fictitious pages don't use "segind". */ 1270 m->flags = PG_FICTITIOUS; 1271 /* Fictitious pages don't use "order" or "pool". */ 1272 m->oflags = VPO_UNMANAGED; 1273 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1274 /* Fictitious pages are unevictable. */ 1275 m->ref_count = 1; 1276 pmap_page_init(m); 1277 memattr: 1278 pmap_page_set_memattr(m, memattr); 1279 } 1280 1281 /* 1282 * vm_page_putfake: 1283 * 1284 * Release a fictitious page. 1285 */ 1286 void 1287 vm_page_putfake(vm_page_t m) 1288 { 1289 1290 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1291 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1292 ("vm_page_putfake: bad page %p", m)); 1293 if (vm_page_xbusied(m)) 1294 vm_page_xunbusy(m); 1295 uma_zfree(fakepg_zone, m); 1296 } 1297 1298 /* 1299 * vm_page_updatefake: 1300 * 1301 * Update the given fictitious page to the specified physical address and 1302 * memory attribute. 1303 */ 1304 void 1305 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1306 { 1307 1308 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1309 ("vm_page_updatefake: bad page %p", m)); 1310 m->phys_addr = paddr; 1311 pmap_page_set_memattr(m, memattr); 1312 } 1313 1314 /* 1315 * vm_page_free: 1316 * 1317 * Free a page. 1318 */ 1319 void 1320 vm_page_free(vm_page_t m) 1321 { 1322 1323 m->flags &= ~PG_ZERO; 1324 vm_page_free_toq(m); 1325 } 1326 1327 /* 1328 * vm_page_free_zero: 1329 * 1330 * Free a page to the zerod-pages queue 1331 */ 1332 void 1333 vm_page_free_zero(vm_page_t m) 1334 { 1335 1336 m->flags |= PG_ZERO; 1337 vm_page_free_toq(m); 1338 } 1339 1340 /* 1341 * Unbusy and handle the page queueing for a page from a getpages request that 1342 * was optionally read ahead or behind. 1343 */ 1344 void 1345 vm_page_readahead_finish(vm_page_t m) 1346 { 1347 1348 /* We shouldn't put invalid pages on queues. */ 1349 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1350 1351 /* 1352 * Since the page is not the actually needed one, whether it should 1353 * be activated or deactivated is not obvious. Empirical results 1354 * have shown that deactivating the page is usually the best choice, 1355 * unless the page is wanted by another thread. 1356 */ 1357 vm_page_lock(m); 1358 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1359 vm_page_activate(m); 1360 else 1361 vm_page_deactivate(m); 1362 vm_page_unlock(m); 1363 vm_page_xunbusy_unchecked(m); 1364 } 1365 1366 /* 1367 * vm_page_sleep_if_busy: 1368 * 1369 * Sleep and release the object lock if the page is busied. 1370 * Returns TRUE if the thread slept. 1371 * 1372 * The given page must be unlocked and object containing it must 1373 * be locked. 1374 */ 1375 int 1376 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1377 { 1378 vm_object_t obj; 1379 1380 vm_page_lock_assert(m, MA_NOTOWNED); 1381 VM_OBJECT_ASSERT_WLOCKED(m->object); 1382 1383 /* 1384 * The page-specific object must be cached because page 1385 * identity can change during the sleep, causing the 1386 * re-lock of a different object. 1387 * It is assumed that a reference to the object is already 1388 * held by the callers. 1389 */ 1390 obj = m->object; 1391 if (vm_page_busied(m) || (obj != NULL && obj->busy)) { 1392 vm_page_busy_sleep(m, msg, false); 1393 VM_OBJECT_WLOCK(obj); 1394 return (TRUE); 1395 } 1396 return (FALSE); 1397 } 1398 1399 /* 1400 * vm_page_sleep_if_xbusy: 1401 * 1402 * Sleep and release the object lock if the page is xbusied. 1403 * Returns TRUE if the thread slept. 1404 * 1405 * The given page must be unlocked and object containing it must 1406 * be locked. 1407 */ 1408 int 1409 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg) 1410 { 1411 vm_object_t obj; 1412 1413 vm_page_lock_assert(m, MA_NOTOWNED); 1414 VM_OBJECT_ASSERT_WLOCKED(m->object); 1415 1416 /* 1417 * The page-specific object must be cached because page 1418 * identity can change during the sleep, causing the 1419 * re-lock of a different object. 1420 * It is assumed that a reference to the object is already 1421 * held by the callers. 1422 */ 1423 obj = m->object; 1424 if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) { 1425 vm_page_busy_sleep(m, msg, true); 1426 VM_OBJECT_WLOCK(obj); 1427 return (TRUE); 1428 } 1429 return (FALSE); 1430 } 1431 1432 /* 1433 * vm_page_dirty_KBI: [ internal use only ] 1434 * 1435 * Set all bits in the page's dirty field. 1436 * 1437 * The object containing the specified page must be locked if the 1438 * call is made from the machine-independent layer. 1439 * 1440 * See vm_page_clear_dirty_mask(). 1441 * 1442 * This function should only be called by vm_page_dirty(). 1443 */ 1444 void 1445 vm_page_dirty_KBI(vm_page_t m) 1446 { 1447 1448 /* Refer to this operation by its public name. */ 1449 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1450 m->dirty = VM_PAGE_BITS_ALL; 1451 } 1452 1453 /* 1454 * vm_page_insert: [ internal use only ] 1455 * 1456 * Inserts the given mem entry into the object and object list. 1457 * 1458 * The object must be locked. 1459 */ 1460 int 1461 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1462 { 1463 vm_page_t mpred; 1464 1465 VM_OBJECT_ASSERT_WLOCKED(object); 1466 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1467 return (vm_page_insert_after(m, object, pindex, mpred)); 1468 } 1469 1470 /* 1471 * vm_page_insert_after: 1472 * 1473 * Inserts the page "m" into the specified object at offset "pindex". 1474 * 1475 * The page "mpred" must immediately precede the offset "pindex" within 1476 * the specified object. 1477 * 1478 * The object must be locked. 1479 */ 1480 static int 1481 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1482 vm_page_t mpred) 1483 { 1484 vm_page_t msucc; 1485 1486 VM_OBJECT_ASSERT_WLOCKED(object); 1487 KASSERT(m->object == NULL, 1488 ("vm_page_insert_after: page already inserted")); 1489 if (mpred != NULL) { 1490 KASSERT(mpred->object == object, 1491 ("vm_page_insert_after: object doesn't contain mpred")); 1492 KASSERT(mpred->pindex < pindex, 1493 ("vm_page_insert_after: mpred doesn't precede pindex")); 1494 msucc = TAILQ_NEXT(mpred, listq); 1495 } else 1496 msucc = TAILQ_FIRST(&object->memq); 1497 if (msucc != NULL) 1498 KASSERT(msucc->pindex > pindex, 1499 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1500 1501 /* 1502 * Record the object/offset pair in this page. 1503 */ 1504 m->object = object; 1505 m->pindex = pindex; 1506 m->ref_count |= VPRC_OBJREF; 1507 1508 /* 1509 * Now link into the object's ordered list of backed pages. 1510 */ 1511 if (vm_radix_insert(&object->rtree, m)) { 1512 m->object = NULL; 1513 m->pindex = 0; 1514 m->ref_count &= ~VPRC_OBJREF; 1515 return (1); 1516 } 1517 vm_page_insert_radixdone(m, object, mpred); 1518 return (0); 1519 } 1520 1521 /* 1522 * vm_page_insert_radixdone: 1523 * 1524 * Complete page "m" insertion into the specified object after the 1525 * radix trie hooking. 1526 * 1527 * The page "mpred" must precede the offset "m->pindex" within the 1528 * specified object. 1529 * 1530 * The object must be locked. 1531 */ 1532 static void 1533 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1534 { 1535 1536 VM_OBJECT_ASSERT_WLOCKED(object); 1537 KASSERT(object != NULL && m->object == object, 1538 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1539 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1540 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1541 if (mpred != NULL) { 1542 KASSERT(mpred->object == object, 1543 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1544 KASSERT(mpred->pindex < m->pindex, 1545 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1546 } 1547 1548 if (mpred != NULL) 1549 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1550 else 1551 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1552 1553 /* 1554 * Show that the object has one more resident page. 1555 */ 1556 object->resident_page_count++; 1557 1558 /* 1559 * Hold the vnode until the last page is released. 1560 */ 1561 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1562 vhold(object->handle); 1563 1564 /* 1565 * Since we are inserting a new and possibly dirty page, 1566 * update the object's generation count. 1567 */ 1568 if (pmap_page_is_write_mapped(m)) 1569 vm_object_set_writeable_dirty(object); 1570 } 1571 1572 /* 1573 * Do the work to remove a page from its object. The caller is responsible for 1574 * updating the page's fields to reflect this removal. 1575 */ 1576 static void 1577 vm_page_object_remove(vm_page_t m) 1578 { 1579 vm_object_t object; 1580 vm_page_t mrem; 1581 1582 object = m->object; 1583 VM_OBJECT_ASSERT_WLOCKED(object); 1584 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1585 ("page %p is missing its object ref", m)); 1586 1587 mrem = vm_radix_remove(&object->rtree, m->pindex); 1588 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1589 1590 /* 1591 * Now remove from the object's list of backed pages. 1592 */ 1593 TAILQ_REMOVE(&object->memq, m, listq); 1594 1595 /* 1596 * And show that the object has one fewer resident page. 1597 */ 1598 object->resident_page_count--; 1599 1600 /* 1601 * The vnode may now be recycled. 1602 */ 1603 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1604 vdrop(object->handle); 1605 } 1606 1607 /* 1608 * vm_page_remove: 1609 * 1610 * Removes the specified page from its containing object, but does not 1611 * invalidate any backing storage. Returns true if the object's reference 1612 * was the last reference to the page, and false otherwise. 1613 * 1614 * The object must be locked. 1615 */ 1616 bool 1617 vm_page_remove(vm_page_t m) 1618 { 1619 1620 vm_page_object_remove(m); 1621 m->object = NULL; 1622 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1623 } 1624 1625 /* 1626 * vm_page_lookup: 1627 * 1628 * Returns the page associated with the object/offset 1629 * pair specified; if none is found, NULL is returned. 1630 * 1631 * The object must be locked. 1632 */ 1633 vm_page_t 1634 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1635 { 1636 1637 VM_OBJECT_ASSERT_LOCKED(object); 1638 return (vm_radix_lookup(&object->rtree, pindex)); 1639 } 1640 1641 /* 1642 * vm_page_find_least: 1643 * 1644 * Returns the page associated with the object with least pindex 1645 * greater than or equal to the parameter pindex, or NULL. 1646 * 1647 * The object must be locked. 1648 */ 1649 vm_page_t 1650 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1651 { 1652 vm_page_t m; 1653 1654 VM_OBJECT_ASSERT_LOCKED(object); 1655 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1656 m = vm_radix_lookup_ge(&object->rtree, pindex); 1657 return (m); 1658 } 1659 1660 /* 1661 * Returns the given page's successor (by pindex) within the object if it is 1662 * resident; if none is found, NULL is returned. 1663 * 1664 * The object must be locked. 1665 */ 1666 vm_page_t 1667 vm_page_next(vm_page_t m) 1668 { 1669 vm_page_t next; 1670 1671 VM_OBJECT_ASSERT_LOCKED(m->object); 1672 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1673 MPASS(next->object == m->object); 1674 if (next->pindex != m->pindex + 1) 1675 next = NULL; 1676 } 1677 return (next); 1678 } 1679 1680 /* 1681 * Returns the given page's predecessor (by pindex) within the object if it is 1682 * resident; if none is found, NULL is returned. 1683 * 1684 * The object must be locked. 1685 */ 1686 vm_page_t 1687 vm_page_prev(vm_page_t m) 1688 { 1689 vm_page_t prev; 1690 1691 VM_OBJECT_ASSERT_LOCKED(m->object); 1692 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1693 MPASS(prev->object == m->object); 1694 if (prev->pindex != m->pindex - 1) 1695 prev = NULL; 1696 } 1697 return (prev); 1698 } 1699 1700 /* 1701 * Uses the page mnew as a replacement for an existing page at index 1702 * pindex which must be already present in the object. 1703 */ 1704 vm_page_t 1705 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1706 { 1707 vm_page_t mold; 1708 1709 VM_OBJECT_ASSERT_WLOCKED(object); 1710 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1711 ("vm_page_replace: page %p already in object", mnew)); 1712 1713 /* 1714 * This function mostly follows vm_page_insert() and 1715 * vm_page_remove() without the radix, object count and vnode 1716 * dance. Double check such functions for more comments. 1717 */ 1718 1719 mnew->object = object; 1720 mnew->pindex = pindex; 1721 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1722 mold = vm_radix_replace(&object->rtree, mnew); 1723 1724 /* Keep the resident page list in sorted order. */ 1725 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1726 TAILQ_REMOVE(&object->memq, mold, listq); 1727 1728 mold->object = NULL; 1729 atomic_clear_int(&mold->ref_count, VPRC_OBJREF); 1730 vm_page_xunbusy(mold); 1731 1732 /* 1733 * The object's resident_page_count does not change because we have 1734 * swapped one page for another, but the generation count should 1735 * change if the page is dirty. 1736 */ 1737 if (pmap_page_is_write_mapped(mnew)) 1738 vm_object_set_writeable_dirty(object); 1739 return (mold); 1740 } 1741 1742 /* 1743 * vm_page_rename: 1744 * 1745 * Move the given memory entry from its 1746 * current object to the specified target object/offset. 1747 * 1748 * Note: swap associated with the page must be invalidated by the move. We 1749 * have to do this for several reasons: (1) we aren't freeing the 1750 * page, (2) we are dirtying the page, (3) the VM system is probably 1751 * moving the page from object A to B, and will then later move 1752 * the backing store from A to B and we can't have a conflict. 1753 * 1754 * Note: we *always* dirty the page. It is necessary both for the 1755 * fact that we moved it, and because we may be invalidating 1756 * swap. 1757 * 1758 * The objects must be locked. 1759 */ 1760 int 1761 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1762 { 1763 vm_page_t mpred; 1764 vm_pindex_t opidx; 1765 1766 VM_OBJECT_ASSERT_WLOCKED(new_object); 1767 1768 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1769 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1770 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1771 ("vm_page_rename: pindex already renamed")); 1772 1773 /* 1774 * Create a custom version of vm_page_insert() which does not depend 1775 * by m_prev and can cheat on the implementation aspects of the 1776 * function. 1777 */ 1778 opidx = m->pindex; 1779 m->pindex = new_pindex; 1780 if (vm_radix_insert(&new_object->rtree, m)) { 1781 m->pindex = opidx; 1782 return (1); 1783 } 1784 1785 /* 1786 * The operation cannot fail anymore. The removal must happen before 1787 * the listq iterator is tainted. 1788 */ 1789 m->pindex = opidx; 1790 vm_page_object_remove(m); 1791 1792 /* Return back to the new pindex to complete vm_page_insert(). */ 1793 m->pindex = new_pindex; 1794 m->object = new_object; 1795 1796 vm_page_insert_radixdone(m, new_object, mpred); 1797 vm_page_dirty(m); 1798 return (0); 1799 } 1800 1801 /* 1802 * vm_page_alloc: 1803 * 1804 * Allocate and return a page that is associated with the specified 1805 * object and offset pair. By default, this page is exclusive busied. 1806 * 1807 * The caller must always specify an allocation class. 1808 * 1809 * allocation classes: 1810 * VM_ALLOC_NORMAL normal process request 1811 * VM_ALLOC_SYSTEM system *really* needs a page 1812 * VM_ALLOC_INTERRUPT interrupt time request 1813 * 1814 * optional allocation flags: 1815 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1816 * intends to allocate 1817 * VM_ALLOC_NOBUSY do not exclusive busy the page 1818 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1819 * VM_ALLOC_NOOBJ page is not associated with an object and 1820 * should not be exclusive busy 1821 * VM_ALLOC_SBUSY shared busy the allocated page 1822 * VM_ALLOC_WIRED wire the allocated page 1823 * VM_ALLOC_ZERO prefer a zeroed page 1824 */ 1825 vm_page_t 1826 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1827 { 1828 1829 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1830 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1831 } 1832 1833 vm_page_t 1834 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1835 int req) 1836 { 1837 1838 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1839 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1840 NULL)); 1841 } 1842 1843 /* 1844 * Allocate a page in the specified object with the given page index. To 1845 * optimize insertion of the page into the object, the caller must also specifiy 1846 * the resident page in the object with largest index smaller than the given 1847 * page index, or NULL if no such page exists. 1848 */ 1849 vm_page_t 1850 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1851 int req, vm_page_t mpred) 1852 { 1853 struct vm_domainset_iter di; 1854 vm_page_t m; 1855 int domain; 1856 1857 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1858 do { 1859 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1860 mpred); 1861 if (m != NULL) 1862 break; 1863 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1864 1865 return (m); 1866 } 1867 1868 /* 1869 * Returns true if the number of free pages exceeds the minimum 1870 * for the request class and false otherwise. 1871 */ 1872 static int 1873 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 1874 { 1875 u_int limit, old, new; 1876 1877 if (req_class == VM_ALLOC_INTERRUPT) 1878 limit = 0; 1879 else if (req_class == VM_ALLOC_SYSTEM) 1880 limit = vmd->vmd_interrupt_free_min; 1881 else 1882 limit = vmd->vmd_free_reserved; 1883 1884 /* 1885 * Attempt to reserve the pages. Fail if we're below the limit. 1886 */ 1887 limit += npages; 1888 old = vmd->vmd_free_count; 1889 do { 1890 if (old < limit) 1891 return (0); 1892 new = old - npages; 1893 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 1894 1895 /* Wake the page daemon if we've crossed the threshold. */ 1896 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 1897 pagedaemon_wakeup(vmd->vmd_domain); 1898 1899 /* Only update bitsets on transitions. */ 1900 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 1901 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 1902 vm_domain_set(vmd); 1903 1904 return (1); 1905 } 1906 1907 int 1908 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 1909 { 1910 int req_class; 1911 1912 /* 1913 * The page daemon is allowed to dig deeper into the free page list. 1914 */ 1915 req_class = req & VM_ALLOC_CLASS_MASK; 1916 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1917 req_class = VM_ALLOC_SYSTEM; 1918 return (_vm_domain_allocate(vmd, req_class, npages)); 1919 } 1920 1921 vm_page_t 1922 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 1923 int req, vm_page_t mpred) 1924 { 1925 struct vm_domain *vmd; 1926 vm_page_t m; 1927 int flags, pool; 1928 1929 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1930 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1931 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1932 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1933 ("inconsistent object(%p)/req(%x)", object, req)); 1934 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1935 ("Can't sleep and retry object insertion.")); 1936 KASSERT(mpred == NULL || mpred->pindex < pindex, 1937 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1938 (uintmax_t)pindex)); 1939 if (object != NULL) 1940 VM_OBJECT_ASSERT_WLOCKED(object); 1941 1942 flags = 0; 1943 m = NULL; 1944 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; 1945 again: 1946 #if VM_NRESERVLEVEL > 0 1947 /* 1948 * Can we allocate the page from a reservation? 1949 */ 1950 if (vm_object_reserv(object) && 1951 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 1952 NULL) { 1953 domain = vm_phys_domain(m); 1954 vmd = VM_DOMAIN(domain); 1955 goto found; 1956 } 1957 #endif 1958 vmd = VM_DOMAIN(domain); 1959 if (vmd->vmd_pgcache[pool].zone != NULL) { 1960 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT); 1961 if (m != NULL) { 1962 flags |= PG_PCPU_CACHE; 1963 goto found; 1964 } 1965 } 1966 if (vm_domain_allocate(vmd, req, 1)) { 1967 /* 1968 * If not, allocate it from the free page queues. 1969 */ 1970 vm_domain_free_lock(vmd); 1971 m = vm_phys_alloc_pages(domain, pool, 0); 1972 vm_domain_free_unlock(vmd); 1973 if (m == NULL) { 1974 vm_domain_freecnt_inc(vmd, 1); 1975 #if VM_NRESERVLEVEL > 0 1976 if (vm_reserv_reclaim_inactive(domain)) 1977 goto again; 1978 #endif 1979 } 1980 } 1981 if (m == NULL) { 1982 /* 1983 * Not allocatable, give up. 1984 */ 1985 if (vm_domain_alloc_fail(vmd, object, req)) 1986 goto again; 1987 return (NULL); 1988 } 1989 1990 /* 1991 * At this point we had better have found a good page. 1992 */ 1993 found: 1994 vm_page_dequeue(m); 1995 vm_page_alloc_check(m); 1996 1997 /* 1998 * Initialize the page. Only the PG_ZERO flag is inherited. 1999 */ 2000 if ((req & VM_ALLOC_ZERO) != 0) 2001 flags |= (m->flags & PG_ZERO); 2002 if ((req & VM_ALLOC_NODUMP) != 0) 2003 flags |= PG_NODUMP; 2004 m->flags = flags; 2005 m->aflags = 0; 2006 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2007 VPO_UNMANAGED : 0; 2008 m->busy_lock = VPB_UNBUSIED; 2009 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2010 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2011 if ((req & VM_ALLOC_SBUSY) != 0) 2012 m->busy_lock = VPB_SHARERS_WORD(1); 2013 if (req & VM_ALLOC_WIRED) { 2014 /* 2015 * The page lock is not required for wiring a page until that 2016 * page is inserted into the object. 2017 */ 2018 vm_wire_add(1); 2019 m->ref_count = 1; 2020 } 2021 m->act_count = 0; 2022 2023 if (object != NULL) { 2024 if (vm_page_insert_after(m, object, pindex, mpred)) { 2025 if (req & VM_ALLOC_WIRED) { 2026 vm_wire_sub(1); 2027 m->ref_count = 0; 2028 } 2029 KASSERT(m->object == NULL, ("page %p has object", m)); 2030 m->oflags = VPO_UNMANAGED; 2031 m->busy_lock = VPB_UNBUSIED; 2032 /* Don't change PG_ZERO. */ 2033 vm_page_free_toq(m); 2034 if (req & VM_ALLOC_WAITFAIL) { 2035 VM_OBJECT_WUNLOCK(object); 2036 vm_radix_wait(); 2037 VM_OBJECT_WLOCK(object); 2038 } 2039 return (NULL); 2040 } 2041 2042 /* Ignore device objects; the pager sets "memattr" for them. */ 2043 if (object->memattr != VM_MEMATTR_DEFAULT && 2044 (object->flags & OBJ_FICTITIOUS) == 0) 2045 pmap_page_set_memattr(m, object->memattr); 2046 } else 2047 m->pindex = pindex; 2048 2049 return (m); 2050 } 2051 2052 /* 2053 * vm_page_alloc_contig: 2054 * 2055 * Allocate a contiguous set of physical pages of the given size "npages" 2056 * from the free lists. All of the physical pages must be at or above 2057 * the given physical address "low" and below the given physical address 2058 * "high". The given value "alignment" determines the alignment of the 2059 * first physical page in the set. If the given value "boundary" is 2060 * non-zero, then the set of physical pages cannot cross any physical 2061 * address boundary that is a multiple of that value. Both "alignment" 2062 * and "boundary" must be a power of two. 2063 * 2064 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2065 * then the memory attribute setting for the physical pages is configured 2066 * to the object's memory attribute setting. Otherwise, the memory 2067 * attribute setting for the physical pages is configured to "memattr", 2068 * overriding the object's memory attribute setting. However, if the 2069 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2070 * memory attribute setting for the physical pages cannot be configured 2071 * to VM_MEMATTR_DEFAULT. 2072 * 2073 * The specified object may not contain fictitious pages. 2074 * 2075 * The caller must always specify an allocation class. 2076 * 2077 * allocation classes: 2078 * VM_ALLOC_NORMAL normal process request 2079 * VM_ALLOC_SYSTEM system *really* needs a page 2080 * VM_ALLOC_INTERRUPT interrupt time request 2081 * 2082 * optional allocation flags: 2083 * VM_ALLOC_NOBUSY do not exclusive busy the page 2084 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2085 * VM_ALLOC_NOOBJ page is not associated with an object and 2086 * should not be exclusive busy 2087 * VM_ALLOC_SBUSY shared busy the allocated page 2088 * VM_ALLOC_WIRED wire the allocated page 2089 * VM_ALLOC_ZERO prefer a zeroed page 2090 */ 2091 vm_page_t 2092 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2093 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2094 vm_paddr_t boundary, vm_memattr_t memattr) 2095 { 2096 struct vm_domainset_iter di; 2097 vm_page_t m; 2098 int domain; 2099 2100 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2101 do { 2102 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2103 npages, low, high, alignment, boundary, memattr); 2104 if (m != NULL) 2105 break; 2106 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2107 2108 return (m); 2109 } 2110 2111 vm_page_t 2112 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2113 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2114 vm_paddr_t boundary, vm_memattr_t memattr) 2115 { 2116 struct vm_domain *vmd; 2117 vm_page_t m, m_ret, mpred; 2118 u_int busy_lock, flags, oflags; 2119 2120 mpred = NULL; /* XXX: pacify gcc */ 2121 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2122 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2123 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2124 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2125 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 2126 req)); 2127 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2128 ("Can't sleep and retry object insertion.")); 2129 if (object != NULL) { 2130 VM_OBJECT_ASSERT_WLOCKED(object); 2131 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2132 ("vm_page_alloc_contig: object %p has fictitious pages", 2133 object)); 2134 } 2135 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2136 2137 if (object != NULL) { 2138 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2139 KASSERT(mpred == NULL || mpred->pindex != pindex, 2140 ("vm_page_alloc_contig: pindex already allocated")); 2141 } 2142 2143 /* 2144 * Can we allocate the pages without the number of free pages falling 2145 * below the lower bound for the allocation class? 2146 */ 2147 m_ret = NULL; 2148 again: 2149 #if VM_NRESERVLEVEL > 0 2150 /* 2151 * Can we allocate the pages from a reservation? 2152 */ 2153 if (vm_object_reserv(object) && 2154 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2155 mpred, npages, low, high, alignment, boundary)) != NULL) { 2156 domain = vm_phys_domain(m_ret); 2157 vmd = VM_DOMAIN(domain); 2158 goto found; 2159 } 2160 #endif 2161 vmd = VM_DOMAIN(domain); 2162 if (vm_domain_allocate(vmd, req, npages)) { 2163 /* 2164 * allocate them from the free page queues. 2165 */ 2166 vm_domain_free_lock(vmd); 2167 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2168 alignment, boundary); 2169 vm_domain_free_unlock(vmd); 2170 if (m_ret == NULL) { 2171 vm_domain_freecnt_inc(vmd, npages); 2172 #if VM_NRESERVLEVEL > 0 2173 if (vm_reserv_reclaim_contig(domain, npages, low, 2174 high, alignment, boundary)) 2175 goto again; 2176 #endif 2177 } 2178 } 2179 if (m_ret == NULL) { 2180 if (vm_domain_alloc_fail(vmd, object, req)) 2181 goto again; 2182 return (NULL); 2183 } 2184 #if VM_NRESERVLEVEL > 0 2185 found: 2186 #endif 2187 for (m = m_ret; m < &m_ret[npages]; m++) { 2188 vm_page_dequeue(m); 2189 vm_page_alloc_check(m); 2190 } 2191 2192 /* 2193 * Initialize the pages. Only the PG_ZERO flag is inherited. 2194 */ 2195 flags = 0; 2196 if ((req & VM_ALLOC_ZERO) != 0) 2197 flags = PG_ZERO; 2198 if ((req & VM_ALLOC_NODUMP) != 0) 2199 flags |= PG_NODUMP; 2200 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2201 VPO_UNMANAGED : 0; 2202 busy_lock = VPB_UNBUSIED; 2203 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2204 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2205 if ((req & VM_ALLOC_SBUSY) != 0) 2206 busy_lock = VPB_SHARERS_WORD(1); 2207 if ((req & VM_ALLOC_WIRED) != 0) 2208 vm_wire_add(npages); 2209 if (object != NULL) { 2210 if (object->memattr != VM_MEMATTR_DEFAULT && 2211 memattr == VM_MEMATTR_DEFAULT) 2212 memattr = object->memattr; 2213 } 2214 for (m = m_ret; m < &m_ret[npages]; m++) { 2215 m->aflags = 0; 2216 m->flags = (m->flags | PG_NODUMP) & flags; 2217 m->busy_lock = busy_lock; 2218 if ((req & VM_ALLOC_WIRED) != 0) 2219 m->ref_count = 1; 2220 m->act_count = 0; 2221 m->oflags = oflags; 2222 if (object != NULL) { 2223 if (vm_page_insert_after(m, object, pindex, mpred)) { 2224 if ((req & VM_ALLOC_WIRED) != 0) 2225 vm_wire_sub(npages); 2226 KASSERT(m->object == NULL, 2227 ("page %p has object", m)); 2228 mpred = m; 2229 for (m = m_ret; m < &m_ret[npages]; m++) { 2230 if (m <= mpred && 2231 (req & VM_ALLOC_WIRED) != 0) 2232 m->ref_count = 0; 2233 m->oflags = VPO_UNMANAGED; 2234 m->busy_lock = VPB_UNBUSIED; 2235 /* Don't change PG_ZERO. */ 2236 vm_page_free_toq(m); 2237 } 2238 if (req & VM_ALLOC_WAITFAIL) { 2239 VM_OBJECT_WUNLOCK(object); 2240 vm_radix_wait(); 2241 VM_OBJECT_WLOCK(object); 2242 } 2243 return (NULL); 2244 } 2245 mpred = m; 2246 } else 2247 m->pindex = pindex; 2248 if (memattr != VM_MEMATTR_DEFAULT) 2249 pmap_page_set_memattr(m, memattr); 2250 pindex++; 2251 } 2252 return (m_ret); 2253 } 2254 2255 /* 2256 * Check a page that has been freshly dequeued from a freelist. 2257 */ 2258 static void 2259 vm_page_alloc_check(vm_page_t m) 2260 { 2261 2262 KASSERT(m->object == NULL, ("page %p has object", m)); 2263 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 2264 ("page %p has unexpected queue %d, flags %#x", 2265 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK))); 2266 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2267 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 2268 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2269 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2270 ("page %p has unexpected memattr %d", 2271 m, pmap_page_get_memattr(m))); 2272 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2273 } 2274 2275 /* 2276 * vm_page_alloc_freelist: 2277 * 2278 * Allocate a physical page from the specified free page list. 2279 * 2280 * The caller must always specify an allocation class. 2281 * 2282 * allocation classes: 2283 * VM_ALLOC_NORMAL normal process request 2284 * VM_ALLOC_SYSTEM system *really* needs a page 2285 * VM_ALLOC_INTERRUPT interrupt time request 2286 * 2287 * optional allocation flags: 2288 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2289 * intends to allocate 2290 * VM_ALLOC_WIRED wire the allocated page 2291 * VM_ALLOC_ZERO prefer a zeroed page 2292 */ 2293 vm_page_t 2294 vm_page_alloc_freelist(int freelist, int req) 2295 { 2296 struct vm_domainset_iter di; 2297 vm_page_t m; 2298 int domain; 2299 2300 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2301 do { 2302 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2303 if (m != NULL) 2304 break; 2305 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2306 2307 return (m); 2308 } 2309 2310 vm_page_t 2311 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2312 { 2313 struct vm_domain *vmd; 2314 vm_page_t m; 2315 u_int flags; 2316 2317 m = NULL; 2318 vmd = VM_DOMAIN(domain); 2319 again: 2320 if (vm_domain_allocate(vmd, req, 1)) { 2321 vm_domain_free_lock(vmd); 2322 m = vm_phys_alloc_freelist_pages(domain, freelist, 2323 VM_FREEPOOL_DIRECT, 0); 2324 vm_domain_free_unlock(vmd); 2325 if (m == NULL) 2326 vm_domain_freecnt_inc(vmd, 1); 2327 } 2328 if (m == NULL) { 2329 if (vm_domain_alloc_fail(vmd, NULL, req)) 2330 goto again; 2331 return (NULL); 2332 } 2333 vm_page_dequeue(m); 2334 vm_page_alloc_check(m); 2335 2336 /* 2337 * Initialize the page. Only the PG_ZERO flag is inherited. 2338 */ 2339 m->aflags = 0; 2340 flags = 0; 2341 if ((req & VM_ALLOC_ZERO) != 0) 2342 flags = PG_ZERO; 2343 m->flags &= flags; 2344 if ((req & VM_ALLOC_WIRED) != 0) { 2345 /* 2346 * The page lock is not required for wiring a page that does 2347 * not belong to an object. 2348 */ 2349 vm_wire_add(1); 2350 m->ref_count = 1; 2351 } 2352 /* Unmanaged pages don't use "act_count". */ 2353 m->oflags = VPO_UNMANAGED; 2354 return (m); 2355 } 2356 2357 static int 2358 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2359 { 2360 struct vm_domain *vmd; 2361 struct vm_pgcache *pgcache; 2362 int i; 2363 2364 pgcache = arg; 2365 vmd = VM_DOMAIN(pgcache->domain); 2366 2367 /* 2368 * The page daemon should avoid creating extra memory pressure since its 2369 * main purpose is to replenish the store of free pages. 2370 */ 2371 if (vmd->vmd_severeset || curproc == pageproc || 2372 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2373 return (0); 2374 domain = vmd->vmd_domain; 2375 vm_domain_free_lock(vmd); 2376 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2377 (vm_page_t *)store); 2378 vm_domain_free_unlock(vmd); 2379 if (cnt != i) 2380 vm_domain_freecnt_inc(vmd, cnt - i); 2381 2382 return (i); 2383 } 2384 2385 static void 2386 vm_page_zone_release(void *arg, void **store, int cnt) 2387 { 2388 struct vm_domain *vmd; 2389 struct vm_pgcache *pgcache; 2390 vm_page_t m; 2391 int i; 2392 2393 pgcache = arg; 2394 vmd = VM_DOMAIN(pgcache->domain); 2395 vm_domain_free_lock(vmd); 2396 for (i = 0; i < cnt; i++) { 2397 m = (vm_page_t)store[i]; 2398 vm_phys_free_pages(m, 0); 2399 } 2400 vm_domain_free_unlock(vmd); 2401 vm_domain_freecnt_inc(vmd, cnt); 2402 } 2403 2404 #define VPSC_ANY 0 /* No restrictions. */ 2405 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2406 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2407 2408 /* 2409 * vm_page_scan_contig: 2410 * 2411 * Scan vm_page_array[] between the specified entries "m_start" and 2412 * "m_end" for a run of contiguous physical pages that satisfy the 2413 * specified conditions, and return the lowest page in the run. The 2414 * specified "alignment" determines the alignment of the lowest physical 2415 * page in the run. If the specified "boundary" is non-zero, then the 2416 * run of physical pages cannot span a physical address that is a 2417 * multiple of "boundary". 2418 * 2419 * "m_end" is never dereferenced, so it need not point to a vm_page 2420 * structure within vm_page_array[]. 2421 * 2422 * "npages" must be greater than zero. "m_start" and "m_end" must not 2423 * span a hole (or discontiguity) in the physical address space. Both 2424 * "alignment" and "boundary" must be a power of two. 2425 */ 2426 vm_page_t 2427 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2428 u_long alignment, vm_paddr_t boundary, int options) 2429 { 2430 struct mtx *m_mtx; 2431 vm_object_t object; 2432 vm_paddr_t pa; 2433 vm_page_t m, m_run; 2434 #if VM_NRESERVLEVEL > 0 2435 int level; 2436 #endif 2437 int m_inc, order, run_ext, run_len; 2438 2439 KASSERT(npages > 0, ("npages is 0")); 2440 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2441 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2442 m_run = NULL; 2443 run_len = 0; 2444 m_mtx = NULL; 2445 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2446 KASSERT((m->flags & PG_MARKER) == 0, 2447 ("page %p is PG_MARKER", m)); 2448 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2449 ("fictitious page %p has invalid ref count", m)); 2450 2451 /* 2452 * If the current page would be the start of a run, check its 2453 * physical address against the end, alignment, and boundary 2454 * conditions. If it doesn't satisfy these conditions, either 2455 * terminate the scan or advance to the next page that 2456 * satisfies the failed condition. 2457 */ 2458 if (run_len == 0) { 2459 KASSERT(m_run == NULL, ("m_run != NULL")); 2460 if (m + npages > m_end) 2461 break; 2462 pa = VM_PAGE_TO_PHYS(m); 2463 if ((pa & (alignment - 1)) != 0) { 2464 m_inc = atop(roundup2(pa, alignment) - pa); 2465 continue; 2466 } 2467 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2468 boundary) != 0) { 2469 m_inc = atop(roundup2(pa, boundary) - pa); 2470 continue; 2471 } 2472 } else 2473 KASSERT(m_run != NULL, ("m_run == NULL")); 2474 2475 vm_page_change_lock(m, &m_mtx); 2476 m_inc = 1; 2477 retry: 2478 if (vm_page_wired(m)) 2479 run_ext = 0; 2480 #if VM_NRESERVLEVEL > 0 2481 else if ((level = vm_reserv_level(m)) >= 0 && 2482 (options & VPSC_NORESERV) != 0) { 2483 run_ext = 0; 2484 /* Advance to the end of the reservation. */ 2485 pa = VM_PAGE_TO_PHYS(m); 2486 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2487 pa); 2488 } 2489 #endif 2490 else if ((object = m->object) != NULL) { 2491 /* 2492 * The page is considered eligible for relocation if 2493 * and only if it could be laundered or reclaimed by 2494 * the page daemon. 2495 */ 2496 if (!VM_OBJECT_TRYRLOCK(object)) { 2497 mtx_unlock(m_mtx); 2498 VM_OBJECT_RLOCK(object); 2499 mtx_lock(m_mtx); 2500 if (m->object != object) { 2501 /* 2502 * The page may have been freed. 2503 */ 2504 VM_OBJECT_RUNLOCK(object); 2505 goto retry; 2506 } 2507 } 2508 /* Don't care: PG_NODUMP, PG_ZERO. */ 2509 if (object->type != OBJT_DEFAULT && 2510 object->type != OBJT_SWAP && 2511 object->type != OBJT_VNODE) { 2512 run_ext = 0; 2513 #if VM_NRESERVLEVEL > 0 2514 } else if ((options & VPSC_NOSUPER) != 0 && 2515 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2516 run_ext = 0; 2517 /* Advance to the end of the superpage. */ 2518 pa = VM_PAGE_TO_PHYS(m); 2519 m_inc = atop(roundup2(pa + 1, 2520 vm_reserv_size(level)) - pa); 2521 #endif 2522 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2523 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) && 2524 !vm_page_wired(m)) { 2525 /* 2526 * The page is allocated but eligible for 2527 * relocation. Extend the current run by one 2528 * page. 2529 */ 2530 KASSERT(pmap_page_get_memattr(m) == 2531 VM_MEMATTR_DEFAULT, 2532 ("page %p has an unexpected memattr", m)); 2533 KASSERT((m->oflags & (VPO_SWAPINPROG | 2534 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2535 ("page %p has unexpected oflags", m)); 2536 /* Don't care: PGA_NOSYNC. */ 2537 run_ext = 1; 2538 } else 2539 run_ext = 0; 2540 VM_OBJECT_RUNLOCK(object); 2541 #if VM_NRESERVLEVEL > 0 2542 } else if (level >= 0) { 2543 /* 2544 * The page is reserved but not yet allocated. In 2545 * other words, it is still free. Extend the current 2546 * run by one page. 2547 */ 2548 run_ext = 1; 2549 #endif 2550 } else if ((order = m->order) < VM_NFREEORDER) { 2551 /* 2552 * The page is enqueued in the physical memory 2553 * allocator's free page queues. Moreover, it is the 2554 * first page in a power-of-two-sized run of 2555 * contiguous free pages. Add these pages to the end 2556 * of the current run, and jump ahead. 2557 */ 2558 run_ext = 1 << order; 2559 m_inc = 1 << order; 2560 } else { 2561 /* 2562 * Skip the page for one of the following reasons: (1) 2563 * It is enqueued in the physical memory allocator's 2564 * free page queues. However, it is not the first 2565 * page in a run of contiguous free pages. (This case 2566 * rarely occurs because the scan is performed in 2567 * ascending order.) (2) It is not reserved, and it is 2568 * transitioning from free to allocated. (Conversely, 2569 * the transition from allocated to free for managed 2570 * pages is blocked by the page lock.) (3) It is 2571 * allocated but not contained by an object and not 2572 * wired, e.g., allocated by Xen's balloon driver. 2573 */ 2574 run_ext = 0; 2575 } 2576 2577 /* 2578 * Extend or reset the current run of pages. 2579 */ 2580 if (run_ext > 0) { 2581 if (run_len == 0) 2582 m_run = m; 2583 run_len += run_ext; 2584 } else { 2585 if (run_len > 0) { 2586 m_run = NULL; 2587 run_len = 0; 2588 } 2589 } 2590 } 2591 if (m_mtx != NULL) 2592 mtx_unlock(m_mtx); 2593 if (run_len >= npages) 2594 return (m_run); 2595 return (NULL); 2596 } 2597 2598 /* 2599 * vm_page_reclaim_run: 2600 * 2601 * Try to relocate each of the allocated virtual pages within the 2602 * specified run of physical pages to a new physical address. Free the 2603 * physical pages underlying the relocated virtual pages. A virtual page 2604 * is relocatable if and only if it could be laundered or reclaimed by 2605 * the page daemon. Whenever possible, a virtual page is relocated to a 2606 * physical address above "high". 2607 * 2608 * Returns 0 if every physical page within the run was already free or 2609 * just freed by a successful relocation. Otherwise, returns a non-zero 2610 * value indicating why the last attempt to relocate a virtual page was 2611 * unsuccessful. 2612 * 2613 * "req_class" must be an allocation class. 2614 */ 2615 static int 2616 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2617 vm_paddr_t high) 2618 { 2619 struct vm_domain *vmd; 2620 struct mtx *m_mtx; 2621 struct spglist free; 2622 vm_object_t object; 2623 vm_paddr_t pa; 2624 vm_page_t m, m_end, m_new; 2625 int error, order, req; 2626 2627 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2628 ("req_class is not an allocation class")); 2629 SLIST_INIT(&free); 2630 error = 0; 2631 m = m_run; 2632 m_end = m_run + npages; 2633 m_mtx = NULL; 2634 for (; error == 0 && m < m_end; m++) { 2635 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2636 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2637 2638 /* 2639 * Avoid releasing and reacquiring the same page lock. 2640 */ 2641 vm_page_change_lock(m, &m_mtx); 2642 retry: 2643 /* 2644 * Racily check for wirings. Races are handled below. 2645 */ 2646 if (vm_page_wired(m)) 2647 error = EBUSY; 2648 else if ((object = m->object) != NULL) { 2649 /* 2650 * The page is relocated if and only if it could be 2651 * laundered or reclaimed by the page daemon. 2652 */ 2653 if (!VM_OBJECT_TRYWLOCK(object)) { 2654 mtx_unlock(m_mtx); 2655 VM_OBJECT_WLOCK(object); 2656 mtx_lock(m_mtx); 2657 if (m->object != object) { 2658 /* 2659 * The page may have been freed. 2660 */ 2661 VM_OBJECT_WUNLOCK(object); 2662 goto retry; 2663 } 2664 } 2665 /* Don't care: PG_NODUMP, PG_ZERO. */ 2666 if (object->type != OBJT_DEFAULT && 2667 object->type != OBJT_SWAP && 2668 object->type != OBJT_VNODE) 2669 error = EINVAL; 2670 else if (object->memattr != VM_MEMATTR_DEFAULT) 2671 error = EINVAL; 2672 else if (vm_page_queue(m) != PQ_NONE && 2673 vm_page_tryxbusy(m) != 0) { 2674 if (vm_page_wired(m)) { 2675 vm_page_xunbusy(m); 2676 error = EBUSY; 2677 goto unlock; 2678 } 2679 KASSERT(pmap_page_get_memattr(m) == 2680 VM_MEMATTR_DEFAULT, 2681 ("page %p has an unexpected memattr", m)); 2682 KASSERT((m->oflags & (VPO_SWAPINPROG | 2683 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2684 ("page %p has unexpected oflags", m)); 2685 /* Don't care: PGA_NOSYNC. */ 2686 if (!vm_page_none_valid(m)) { 2687 /* 2688 * First, try to allocate a new page 2689 * that is above "high". Failing 2690 * that, try to allocate a new page 2691 * that is below "m_run". Allocate 2692 * the new page between the end of 2693 * "m_run" and "high" only as a last 2694 * resort. 2695 */ 2696 req = req_class | VM_ALLOC_NOOBJ; 2697 if ((m->flags & PG_NODUMP) != 0) 2698 req |= VM_ALLOC_NODUMP; 2699 if (trunc_page(high) != 2700 ~(vm_paddr_t)PAGE_MASK) { 2701 m_new = vm_page_alloc_contig( 2702 NULL, 0, req, 1, 2703 round_page(high), 2704 ~(vm_paddr_t)0, 2705 PAGE_SIZE, 0, 2706 VM_MEMATTR_DEFAULT); 2707 } else 2708 m_new = NULL; 2709 if (m_new == NULL) { 2710 pa = VM_PAGE_TO_PHYS(m_run); 2711 m_new = vm_page_alloc_contig( 2712 NULL, 0, req, 1, 2713 0, pa - 1, PAGE_SIZE, 0, 2714 VM_MEMATTR_DEFAULT); 2715 } 2716 if (m_new == NULL) { 2717 pa += ptoa(npages); 2718 m_new = vm_page_alloc_contig( 2719 NULL, 0, req, 1, 2720 pa, high, PAGE_SIZE, 0, 2721 VM_MEMATTR_DEFAULT); 2722 } 2723 if (m_new == NULL) { 2724 vm_page_xunbusy(m); 2725 error = ENOMEM; 2726 goto unlock; 2727 } 2728 2729 /* 2730 * Unmap the page and check for new 2731 * wirings that may have been acquired 2732 * through a pmap lookup. 2733 */ 2734 if (object->ref_count != 0 && 2735 !vm_page_try_remove_all(m)) { 2736 vm_page_free(m_new); 2737 error = EBUSY; 2738 goto unlock; 2739 } 2740 2741 /* 2742 * Replace "m" with the new page. For 2743 * vm_page_replace(), "m" must be busy 2744 * and dequeued. Finally, change "m" 2745 * as if vm_page_free() was called. 2746 */ 2747 m_new->aflags = m->aflags & 2748 ~PGA_QUEUE_STATE_MASK; 2749 KASSERT(m_new->oflags == VPO_UNMANAGED, 2750 ("page %p is managed", m_new)); 2751 pmap_copy_page(m, m_new); 2752 m_new->valid = m->valid; 2753 m_new->dirty = m->dirty; 2754 m->flags &= ~PG_ZERO; 2755 vm_page_dequeue(m); 2756 vm_page_replace_checked(m_new, object, 2757 m->pindex, m); 2758 if (vm_page_free_prep(m)) 2759 SLIST_INSERT_HEAD(&free, m, 2760 plinks.s.ss); 2761 2762 /* 2763 * The new page must be deactivated 2764 * before the object is unlocked. 2765 */ 2766 vm_page_change_lock(m_new, &m_mtx); 2767 vm_page_deactivate(m_new); 2768 } else { 2769 m->flags &= ~PG_ZERO; 2770 vm_page_dequeue(m); 2771 if (vm_page_free_prep(m)) 2772 SLIST_INSERT_HEAD(&free, m, 2773 plinks.s.ss); 2774 KASSERT(m->dirty == 0, 2775 ("page %p is dirty", m)); 2776 } 2777 } else 2778 error = EBUSY; 2779 unlock: 2780 VM_OBJECT_WUNLOCK(object); 2781 } else { 2782 MPASS(vm_phys_domain(m) == domain); 2783 vmd = VM_DOMAIN(domain); 2784 vm_domain_free_lock(vmd); 2785 order = m->order; 2786 if (order < VM_NFREEORDER) { 2787 /* 2788 * The page is enqueued in the physical memory 2789 * allocator's free page queues. Moreover, it 2790 * is the first page in a power-of-two-sized 2791 * run of contiguous free pages. Jump ahead 2792 * to the last page within that run, and 2793 * continue from there. 2794 */ 2795 m += (1 << order) - 1; 2796 } 2797 #if VM_NRESERVLEVEL > 0 2798 else if (vm_reserv_is_page_free(m)) 2799 order = 0; 2800 #endif 2801 vm_domain_free_unlock(vmd); 2802 if (order == VM_NFREEORDER) 2803 error = EINVAL; 2804 } 2805 } 2806 if (m_mtx != NULL) 2807 mtx_unlock(m_mtx); 2808 if ((m = SLIST_FIRST(&free)) != NULL) { 2809 int cnt; 2810 2811 vmd = VM_DOMAIN(domain); 2812 cnt = 0; 2813 vm_domain_free_lock(vmd); 2814 do { 2815 MPASS(vm_phys_domain(m) == domain); 2816 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2817 vm_phys_free_pages(m, 0); 2818 cnt++; 2819 } while ((m = SLIST_FIRST(&free)) != NULL); 2820 vm_domain_free_unlock(vmd); 2821 vm_domain_freecnt_inc(vmd, cnt); 2822 } 2823 return (error); 2824 } 2825 2826 #define NRUNS 16 2827 2828 CTASSERT(powerof2(NRUNS)); 2829 2830 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2831 2832 #define MIN_RECLAIM 8 2833 2834 /* 2835 * vm_page_reclaim_contig: 2836 * 2837 * Reclaim allocated, contiguous physical memory satisfying the specified 2838 * conditions by relocating the virtual pages using that physical memory. 2839 * Returns true if reclamation is successful and false otherwise. Since 2840 * relocation requires the allocation of physical pages, reclamation may 2841 * fail due to a shortage of free pages. When reclamation fails, callers 2842 * are expected to perform vm_wait() before retrying a failed allocation 2843 * operation, e.g., vm_page_alloc_contig(). 2844 * 2845 * The caller must always specify an allocation class through "req". 2846 * 2847 * allocation classes: 2848 * VM_ALLOC_NORMAL normal process request 2849 * VM_ALLOC_SYSTEM system *really* needs a page 2850 * VM_ALLOC_INTERRUPT interrupt time request 2851 * 2852 * The optional allocation flags are ignored. 2853 * 2854 * "npages" must be greater than zero. Both "alignment" and "boundary" 2855 * must be a power of two. 2856 */ 2857 bool 2858 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2859 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2860 { 2861 struct vm_domain *vmd; 2862 vm_paddr_t curr_low; 2863 vm_page_t m_run, m_runs[NRUNS]; 2864 u_long count, reclaimed; 2865 int error, i, options, req_class; 2866 2867 KASSERT(npages > 0, ("npages is 0")); 2868 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2869 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2870 req_class = req & VM_ALLOC_CLASS_MASK; 2871 2872 /* 2873 * The page daemon is allowed to dig deeper into the free page list. 2874 */ 2875 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2876 req_class = VM_ALLOC_SYSTEM; 2877 2878 /* 2879 * Return if the number of free pages cannot satisfy the requested 2880 * allocation. 2881 */ 2882 vmd = VM_DOMAIN(domain); 2883 count = vmd->vmd_free_count; 2884 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2885 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2886 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2887 return (false); 2888 2889 /* 2890 * Scan up to three times, relaxing the restrictions ("options") on 2891 * the reclamation of reservations and superpages each time. 2892 */ 2893 for (options = VPSC_NORESERV;;) { 2894 /* 2895 * Find the highest runs that satisfy the given constraints 2896 * and restrictions, and record them in "m_runs". 2897 */ 2898 curr_low = low; 2899 count = 0; 2900 for (;;) { 2901 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2902 high, alignment, boundary, options); 2903 if (m_run == NULL) 2904 break; 2905 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2906 m_runs[RUN_INDEX(count)] = m_run; 2907 count++; 2908 } 2909 2910 /* 2911 * Reclaim the highest runs in LIFO (descending) order until 2912 * the number of reclaimed pages, "reclaimed", is at least 2913 * MIN_RECLAIM. Reset "reclaimed" each time because each 2914 * reclamation is idempotent, and runs will (likely) recur 2915 * from one scan to the next as restrictions are relaxed. 2916 */ 2917 reclaimed = 0; 2918 for (i = 0; count > 0 && i < NRUNS; i++) { 2919 count--; 2920 m_run = m_runs[RUN_INDEX(count)]; 2921 error = vm_page_reclaim_run(req_class, domain, npages, 2922 m_run, high); 2923 if (error == 0) { 2924 reclaimed += npages; 2925 if (reclaimed >= MIN_RECLAIM) 2926 return (true); 2927 } 2928 } 2929 2930 /* 2931 * Either relax the restrictions on the next scan or return if 2932 * the last scan had no restrictions. 2933 */ 2934 if (options == VPSC_NORESERV) 2935 options = VPSC_NOSUPER; 2936 else if (options == VPSC_NOSUPER) 2937 options = VPSC_ANY; 2938 else if (options == VPSC_ANY) 2939 return (reclaimed != 0); 2940 } 2941 } 2942 2943 bool 2944 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2945 u_long alignment, vm_paddr_t boundary) 2946 { 2947 struct vm_domainset_iter di; 2948 int domain; 2949 bool ret; 2950 2951 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2952 do { 2953 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 2954 high, alignment, boundary); 2955 if (ret) 2956 break; 2957 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2958 2959 return (ret); 2960 } 2961 2962 /* 2963 * Set the domain in the appropriate page level domainset. 2964 */ 2965 void 2966 vm_domain_set(struct vm_domain *vmd) 2967 { 2968 2969 mtx_lock(&vm_domainset_lock); 2970 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 2971 vmd->vmd_minset = 1; 2972 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 2973 } 2974 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 2975 vmd->vmd_severeset = 1; 2976 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 2977 } 2978 mtx_unlock(&vm_domainset_lock); 2979 } 2980 2981 /* 2982 * Clear the domain from the appropriate page level domainset. 2983 */ 2984 void 2985 vm_domain_clear(struct vm_domain *vmd) 2986 { 2987 2988 mtx_lock(&vm_domainset_lock); 2989 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 2990 vmd->vmd_minset = 0; 2991 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 2992 if (vm_min_waiters != 0) { 2993 vm_min_waiters = 0; 2994 wakeup(&vm_min_domains); 2995 } 2996 } 2997 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 2998 vmd->vmd_severeset = 0; 2999 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3000 if (vm_severe_waiters != 0) { 3001 vm_severe_waiters = 0; 3002 wakeup(&vm_severe_domains); 3003 } 3004 } 3005 3006 /* 3007 * If pageout daemon needs pages, then tell it that there are 3008 * some free. 3009 */ 3010 if (vmd->vmd_pageout_pages_needed && 3011 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3012 wakeup(&vmd->vmd_pageout_pages_needed); 3013 vmd->vmd_pageout_pages_needed = 0; 3014 } 3015 3016 /* See comments in vm_wait_doms(). */ 3017 if (vm_pageproc_waiters) { 3018 vm_pageproc_waiters = 0; 3019 wakeup(&vm_pageproc_waiters); 3020 } 3021 mtx_unlock(&vm_domainset_lock); 3022 } 3023 3024 /* 3025 * Wait for free pages to exceed the min threshold globally. 3026 */ 3027 void 3028 vm_wait_min(void) 3029 { 3030 3031 mtx_lock(&vm_domainset_lock); 3032 while (vm_page_count_min()) { 3033 vm_min_waiters++; 3034 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3035 } 3036 mtx_unlock(&vm_domainset_lock); 3037 } 3038 3039 /* 3040 * Wait for free pages to exceed the severe threshold globally. 3041 */ 3042 void 3043 vm_wait_severe(void) 3044 { 3045 3046 mtx_lock(&vm_domainset_lock); 3047 while (vm_page_count_severe()) { 3048 vm_severe_waiters++; 3049 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3050 "vmwait", 0); 3051 } 3052 mtx_unlock(&vm_domainset_lock); 3053 } 3054 3055 u_int 3056 vm_wait_count(void) 3057 { 3058 3059 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3060 } 3061 3062 void 3063 vm_wait_doms(const domainset_t *wdoms) 3064 { 3065 3066 /* 3067 * We use racey wakeup synchronization to avoid expensive global 3068 * locking for the pageproc when sleeping with a non-specific vm_wait. 3069 * To handle this, we only sleep for one tick in this instance. It 3070 * is expected that most allocations for the pageproc will come from 3071 * kmem or vm_page_grab* which will use the more specific and 3072 * race-free vm_wait_domain(). 3073 */ 3074 if (curproc == pageproc) { 3075 mtx_lock(&vm_domainset_lock); 3076 vm_pageproc_waiters++; 3077 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, 3078 "pageprocwait", 1); 3079 } else { 3080 /* 3081 * XXX Ideally we would wait only until the allocation could 3082 * be satisfied. This condition can cause new allocators to 3083 * consume all freed pages while old allocators wait. 3084 */ 3085 mtx_lock(&vm_domainset_lock); 3086 if (vm_page_count_min_set(wdoms)) { 3087 vm_min_waiters++; 3088 msleep(&vm_min_domains, &vm_domainset_lock, 3089 PVM | PDROP, "vmwait", 0); 3090 } else 3091 mtx_unlock(&vm_domainset_lock); 3092 } 3093 } 3094 3095 /* 3096 * vm_wait_domain: 3097 * 3098 * Sleep until free pages are available for allocation. 3099 * - Called in various places after failed memory allocations. 3100 */ 3101 void 3102 vm_wait_domain(int domain) 3103 { 3104 struct vm_domain *vmd; 3105 domainset_t wdom; 3106 3107 vmd = VM_DOMAIN(domain); 3108 vm_domain_free_assert_unlocked(vmd); 3109 3110 if (curproc == pageproc) { 3111 mtx_lock(&vm_domainset_lock); 3112 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3113 vmd->vmd_pageout_pages_needed = 1; 3114 msleep(&vmd->vmd_pageout_pages_needed, 3115 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3116 } else 3117 mtx_unlock(&vm_domainset_lock); 3118 } else { 3119 if (pageproc == NULL) 3120 panic("vm_wait in early boot"); 3121 DOMAINSET_ZERO(&wdom); 3122 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3123 vm_wait_doms(&wdom); 3124 } 3125 } 3126 3127 /* 3128 * vm_wait: 3129 * 3130 * Sleep until free pages are available for allocation in the 3131 * affinity domains of the obj. If obj is NULL, the domain set 3132 * for the calling thread is used. 3133 * Called in various places after failed memory allocations. 3134 */ 3135 void 3136 vm_wait(vm_object_t obj) 3137 { 3138 struct domainset *d; 3139 3140 d = NULL; 3141 3142 /* 3143 * Carefully fetch pointers only once: the struct domainset 3144 * itself is ummutable but the pointer might change. 3145 */ 3146 if (obj != NULL) 3147 d = obj->domain.dr_policy; 3148 if (d == NULL) 3149 d = curthread->td_domain.dr_policy; 3150 3151 vm_wait_doms(&d->ds_mask); 3152 } 3153 3154 /* 3155 * vm_domain_alloc_fail: 3156 * 3157 * Called when a page allocation function fails. Informs the 3158 * pagedaemon and performs the requested wait. Requires the 3159 * domain_free and object lock on entry. Returns with the 3160 * object lock held and free lock released. Returns an error when 3161 * retry is necessary. 3162 * 3163 */ 3164 static int 3165 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3166 { 3167 3168 vm_domain_free_assert_unlocked(vmd); 3169 3170 atomic_add_int(&vmd->vmd_pageout_deficit, 3171 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3172 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3173 if (object != NULL) 3174 VM_OBJECT_WUNLOCK(object); 3175 vm_wait_domain(vmd->vmd_domain); 3176 if (object != NULL) 3177 VM_OBJECT_WLOCK(object); 3178 if (req & VM_ALLOC_WAITOK) 3179 return (EAGAIN); 3180 } 3181 3182 return (0); 3183 } 3184 3185 /* 3186 * vm_waitpfault: 3187 * 3188 * Sleep until free pages are available for allocation. 3189 * - Called only in vm_fault so that processes page faulting 3190 * can be easily tracked. 3191 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3192 * processes will be able to grab memory first. Do not change 3193 * this balance without careful testing first. 3194 */ 3195 void 3196 vm_waitpfault(struct domainset *dset, int timo) 3197 { 3198 3199 /* 3200 * XXX Ideally we would wait only until the allocation could 3201 * be satisfied. This condition can cause new allocators to 3202 * consume all freed pages while old allocators wait. 3203 */ 3204 mtx_lock(&vm_domainset_lock); 3205 if (vm_page_count_min_set(&dset->ds_mask)) { 3206 vm_min_waiters++; 3207 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3208 "pfault", timo); 3209 } else 3210 mtx_unlock(&vm_domainset_lock); 3211 } 3212 3213 static struct vm_pagequeue * 3214 vm_page_pagequeue(vm_page_t m) 3215 { 3216 3217 uint8_t queue; 3218 3219 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) 3220 return (NULL); 3221 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3222 } 3223 3224 static inline void 3225 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m) 3226 { 3227 struct vm_domain *vmd; 3228 uint16_t qflags; 3229 3230 CRITICAL_ASSERT(curthread); 3231 vm_pagequeue_assert_locked(pq); 3232 3233 /* 3234 * The page daemon is allowed to set m->queue = PQ_NONE without 3235 * the page queue lock held. In this case it is about to free the page, 3236 * which must not have any queue state. 3237 */ 3238 qflags = atomic_load_16(&m->aflags); 3239 KASSERT(pq == vm_page_pagequeue(m) || 3240 (qflags & PGA_QUEUE_STATE_MASK) == 0, 3241 ("page %p doesn't belong to queue %p but has aflags %#x", 3242 m, pq, qflags)); 3243 3244 if ((qflags & PGA_DEQUEUE) != 0) { 3245 if (__predict_true((qflags & PGA_ENQUEUED) != 0)) 3246 vm_pagequeue_remove(pq, m); 3247 vm_page_dequeue_complete(m); 3248 counter_u64_add(queue_ops, 1); 3249 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) { 3250 if ((qflags & PGA_ENQUEUED) != 0) 3251 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3252 else { 3253 vm_pagequeue_cnt_inc(pq); 3254 vm_page_aflag_set(m, PGA_ENQUEUED); 3255 } 3256 3257 /* 3258 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. 3259 * In particular, if both flags are set in close succession, 3260 * only PGA_REQUEUE_HEAD will be applied, even if it was set 3261 * first. 3262 */ 3263 if ((qflags & PGA_REQUEUE_HEAD) != 0) { 3264 KASSERT(m->queue == PQ_INACTIVE, 3265 ("head enqueue not supported for page %p", m)); 3266 vmd = vm_pagequeue_domain(m); 3267 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3268 } else 3269 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3270 3271 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE | 3272 PGA_REQUEUE_HEAD)); 3273 counter_u64_add(queue_ops, 1); 3274 } else { 3275 counter_u64_add(queue_nops, 1); 3276 } 3277 } 3278 3279 static void 3280 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3281 uint8_t queue) 3282 { 3283 vm_page_t m; 3284 int i; 3285 3286 for (i = 0; i < bq->bq_cnt; i++) { 3287 m = bq->bq_pa[i]; 3288 if (__predict_false(m->queue != queue)) 3289 continue; 3290 vm_pqbatch_process_page(pq, m); 3291 } 3292 vm_batchqueue_init(bq); 3293 } 3294 3295 /* 3296 * vm_page_pqbatch_submit: [ internal use only ] 3297 * 3298 * Enqueue a page in the specified page queue's batched work queue. 3299 * The caller must have encoded the requested operation in the page 3300 * structure's aflags field. 3301 */ 3302 void 3303 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3304 { 3305 struct vm_batchqueue *bq; 3306 struct vm_pagequeue *pq; 3307 int domain; 3308 3309 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3310 ("page %p is unmanaged", m)); 3311 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, 3312 ("missing synchronization for page %p", m)); 3313 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3314 3315 domain = vm_phys_domain(m); 3316 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 3317 3318 critical_enter(); 3319 bq = DPCPU_PTR(pqbatch[domain][queue]); 3320 if (vm_batchqueue_insert(bq, m)) { 3321 critical_exit(); 3322 return; 3323 } 3324 critical_exit(); 3325 vm_pagequeue_lock(pq); 3326 critical_enter(); 3327 bq = DPCPU_PTR(pqbatch[domain][queue]); 3328 vm_pqbatch_process(pq, bq, queue); 3329 3330 /* 3331 * The page may have been logically dequeued before we acquired the 3332 * page queue lock. In this case, since we either hold the page lock 3333 * or the page is being freed, a different thread cannot be concurrently 3334 * enqueuing the page. 3335 */ 3336 if (__predict_true(m->queue == queue)) 3337 vm_pqbatch_process_page(pq, m); 3338 else { 3339 KASSERT(m->queue == PQ_NONE, 3340 ("invalid queue transition for page %p", m)); 3341 KASSERT((m->aflags & PGA_ENQUEUED) == 0, 3342 ("page %p is enqueued with invalid queue index", m)); 3343 } 3344 vm_pagequeue_unlock(pq); 3345 critical_exit(); 3346 } 3347 3348 /* 3349 * vm_page_pqbatch_drain: [ internal use only ] 3350 * 3351 * Force all per-CPU page queue batch queues to be drained. This is 3352 * intended for use in severe memory shortages, to ensure that pages 3353 * do not remain stuck in the batch queues. 3354 */ 3355 void 3356 vm_page_pqbatch_drain(void) 3357 { 3358 struct thread *td; 3359 struct vm_domain *vmd; 3360 struct vm_pagequeue *pq; 3361 int cpu, domain, queue; 3362 3363 td = curthread; 3364 CPU_FOREACH(cpu) { 3365 thread_lock(td); 3366 sched_bind(td, cpu); 3367 thread_unlock(td); 3368 3369 for (domain = 0; domain < vm_ndomains; domain++) { 3370 vmd = VM_DOMAIN(domain); 3371 for (queue = 0; queue < PQ_COUNT; queue++) { 3372 pq = &vmd->vmd_pagequeues[queue]; 3373 vm_pagequeue_lock(pq); 3374 critical_enter(); 3375 vm_pqbatch_process(pq, 3376 DPCPU_PTR(pqbatch[domain][queue]), queue); 3377 critical_exit(); 3378 vm_pagequeue_unlock(pq); 3379 } 3380 } 3381 } 3382 thread_lock(td); 3383 sched_unbind(td); 3384 thread_unlock(td); 3385 } 3386 3387 /* 3388 * Complete the logical removal of a page from a page queue. We must be 3389 * careful to synchronize with the page daemon, which may be concurrently 3390 * examining the page with only the page lock held. The page must not be 3391 * in a state where it appears to be logically enqueued. 3392 */ 3393 static void 3394 vm_page_dequeue_complete(vm_page_t m) 3395 { 3396 3397 m->queue = PQ_NONE; 3398 atomic_thread_fence_rel(); 3399 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); 3400 } 3401 3402 /* 3403 * vm_page_dequeue_deferred: [ internal use only ] 3404 * 3405 * Request removal of the given page from its current page 3406 * queue. Physical removal from the queue may be deferred 3407 * indefinitely. 3408 * 3409 * The page must be locked. 3410 */ 3411 void 3412 vm_page_dequeue_deferred(vm_page_t m) 3413 { 3414 uint8_t queue; 3415 3416 vm_page_assert_locked(m); 3417 3418 if ((queue = vm_page_queue(m)) == PQ_NONE) 3419 return; 3420 3421 /* 3422 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call 3423 * to vm_page_dequeue_deferred_free(). In particular, avoid modifying 3424 * the page's queue state once vm_page_dequeue_deferred_free() has been 3425 * called. In the event of a race, two batch queue entries for the page 3426 * will be created, but the second will have no effect. 3427 */ 3428 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE)) 3429 vm_page_pqbatch_submit(m, queue); 3430 } 3431 3432 /* 3433 * A variant of vm_page_dequeue_deferred() that does not assert the page 3434 * lock and is only to be called from vm_page_free_prep(). Because the 3435 * page is being freed, we can assume that nothing other than the page 3436 * daemon is scheduling queue operations on this page, so we get for 3437 * free the mutual exclusion that is otherwise provided by the page lock. 3438 * To handle races, the page daemon must take care to atomically check 3439 * for PGA_DEQUEUE when updating queue state. 3440 */ 3441 static void 3442 vm_page_dequeue_deferred_free(vm_page_t m) 3443 { 3444 uint8_t queue; 3445 3446 KASSERT(m->ref_count == 0, ("page %p has references", m)); 3447 3448 for (;;) { 3449 if ((m->aflags & PGA_DEQUEUE) != 0) 3450 return; 3451 atomic_thread_fence_acq(); 3452 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) 3453 return; 3454 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, 3455 PGA_DEQUEUE)) { 3456 vm_page_pqbatch_submit(m, queue); 3457 break; 3458 } 3459 } 3460 } 3461 3462 /* 3463 * vm_page_dequeue: 3464 * 3465 * Remove the page from whichever page queue it's in, if any. 3466 * The page must either be locked or unallocated. This constraint 3467 * ensures that the queue state of the page will remain consistent 3468 * after this function returns. 3469 */ 3470 void 3471 vm_page_dequeue(vm_page_t m) 3472 { 3473 struct vm_pagequeue *pq, *pq1; 3474 uint16_t aflags; 3475 3476 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->ref_count == 0, 3477 ("page %p is allocated and unlocked", m)); 3478 3479 for (pq = vm_page_pagequeue(m);; pq = pq1) { 3480 if (pq == NULL) { 3481 /* 3482 * A thread may be concurrently executing 3483 * vm_page_dequeue_complete(). Ensure that all queue 3484 * state is cleared before we return. 3485 */ 3486 aflags = atomic_load_16(&m->aflags); 3487 if ((aflags & PGA_QUEUE_STATE_MASK) == 0) 3488 return; 3489 KASSERT((aflags & PGA_DEQUEUE) != 0, 3490 ("page %p has unexpected queue state flags %#x", 3491 m, aflags)); 3492 3493 /* 3494 * Busy wait until the thread updating queue state is 3495 * finished. Such a thread must be executing in a 3496 * critical section. 3497 */ 3498 cpu_spinwait(); 3499 pq1 = vm_page_pagequeue(m); 3500 continue; 3501 } 3502 vm_pagequeue_lock(pq); 3503 if ((pq1 = vm_page_pagequeue(m)) == pq) 3504 break; 3505 vm_pagequeue_unlock(pq); 3506 } 3507 KASSERT(pq == vm_page_pagequeue(m), 3508 ("%s: page %p migrated directly between queues", __func__, m)); 3509 KASSERT((m->aflags & PGA_DEQUEUE) != 0 || 3510 mtx_owned(vm_page_lockptr(m)), 3511 ("%s: queued unlocked page %p", __func__, m)); 3512 3513 if ((m->aflags & PGA_ENQUEUED) != 0) 3514 vm_pagequeue_remove(pq, m); 3515 vm_page_dequeue_complete(m); 3516 vm_pagequeue_unlock(pq); 3517 } 3518 3519 /* 3520 * Schedule the given page for insertion into the specified page queue. 3521 * Physical insertion of the page may be deferred indefinitely. 3522 */ 3523 static void 3524 vm_page_enqueue(vm_page_t m, uint8_t queue) 3525 { 3526 3527 vm_page_assert_locked(m); 3528 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 3529 ("%s: page %p is already enqueued", __func__, m)); 3530 KASSERT(m->ref_count > 0, 3531 ("%s: page %p does not carry any references", __func__, m)); 3532 3533 m->queue = queue; 3534 if ((m->aflags & PGA_REQUEUE) == 0) 3535 vm_page_aflag_set(m, PGA_REQUEUE); 3536 vm_page_pqbatch_submit(m, queue); 3537 } 3538 3539 /* 3540 * vm_page_requeue: [ internal use only ] 3541 * 3542 * Schedule a requeue of the given page. 3543 * 3544 * The page must be locked. 3545 */ 3546 void 3547 vm_page_requeue(vm_page_t m) 3548 { 3549 3550 vm_page_assert_locked(m); 3551 KASSERT(vm_page_queue(m) != PQ_NONE, 3552 ("%s: page %p is not logically enqueued", __func__, m)); 3553 KASSERT(m->ref_count > 0, 3554 ("%s: page %p does not carry any references", __func__, m)); 3555 3556 if ((m->aflags & PGA_REQUEUE) == 0) 3557 vm_page_aflag_set(m, PGA_REQUEUE); 3558 vm_page_pqbatch_submit(m, atomic_load_8(&m->queue)); 3559 } 3560 3561 /* 3562 * vm_page_swapqueue: [ internal use only ] 3563 * 3564 * Move the page from one queue to another, or to the tail of its 3565 * current queue, in the face of a possible concurrent call to 3566 * vm_page_dequeue_deferred_free(). 3567 */ 3568 void 3569 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq) 3570 { 3571 struct vm_pagequeue *pq; 3572 vm_page_t next; 3573 bool queued; 3574 3575 KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq, 3576 ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq)); 3577 vm_page_assert_locked(m); 3578 3579 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq]; 3580 vm_pagequeue_lock(pq); 3581 3582 /* 3583 * The physical queue state might change at any point before the page 3584 * queue lock is acquired, so we must verify that we hold the correct 3585 * lock before proceeding. 3586 */ 3587 if (__predict_false(m->queue != oldq)) { 3588 vm_pagequeue_unlock(pq); 3589 return; 3590 } 3591 3592 /* 3593 * Once the queue index of the page changes, there is nothing 3594 * synchronizing with further updates to the physical queue state. 3595 * Therefore we must remove the page from the queue now in anticipation 3596 * of a successful commit, and be prepared to roll back. 3597 */ 3598 if (__predict_true((m->aflags & PGA_ENQUEUED) != 0)) { 3599 next = TAILQ_NEXT(m, plinks.q); 3600 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3601 vm_page_aflag_clear(m, PGA_ENQUEUED); 3602 queued = true; 3603 } else { 3604 queued = false; 3605 } 3606 3607 /* 3608 * Atomically update the queue field and set PGA_REQUEUE while 3609 * ensuring that PGA_DEQUEUE has not been set. 3610 */ 3611 if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE, 3612 PGA_REQUEUE))) { 3613 if (queued) { 3614 vm_page_aflag_set(m, PGA_ENQUEUED); 3615 if (next != NULL) 3616 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3617 else 3618 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3619 } 3620 vm_pagequeue_unlock(pq); 3621 return; 3622 } 3623 vm_pagequeue_cnt_dec(pq); 3624 vm_pagequeue_unlock(pq); 3625 vm_page_pqbatch_submit(m, newq); 3626 } 3627 3628 /* 3629 * vm_page_free_prep: 3630 * 3631 * Prepares the given page to be put on the free list, 3632 * disassociating it from any VM object. The caller may return 3633 * the page to the free list only if this function returns true. 3634 * 3635 * The object must be locked. The page must be locked if it is 3636 * managed. 3637 */ 3638 bool 3639 vm_page_free_prep(vm_page_t m) 3640 { 3641 3642 /* 3643 * Synchronize with threads that have dropped a reference to this 3644 * page. 3645 */ 3646 atomic_thread_fence_acq(); 3647 3648 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3649 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3650 uint64_t *p; 3651 int i; 3652 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3653 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3654 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3655 m, i, (uintmax_t)*p)); 3656 } 3657 #endif 3658 if ((m->oflags & VPO_UNMANAGED) == 0) { 3659 KASSERT(!pmap_page_is_mapped(m), 3660 ("vm_page_free_prep: freeing mapped page %p", m)); 3661 KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3662 ("vm_page_free_prep: mapping flags set in page %p", m)); 3663 } else { 3664 KASSERT(m->queue == PQ_NONE, 3665 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3666 } 3667 VM_CNT_INC(v_tfree); 3668 3669 if (vm_page_sbusied(m)) 3670 panic("vm_page_free_prep: freeing shared busy page %p", m); 3671 3672 if (m->object != NULL) { 3673 vm_page_object_remove(m); 3674 3675 /* 3676 * The object reference can be released without an atomic 3677 * operation. 3678 */ 3679 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3680 m->ref_count == VPRC_OBJREF, 3681 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3682 m, m->ref_count)); 3683 m->object = NULL; 3684 m->ref_count -= VPRC_OBJREF; 3685 } 3686 3687 if (vm_page_xbusied(m)) 3688 vm_page_xunbusy(m); 3689 3690 /* 3691 * If fictitious remove object association and 3692 * return. 3693 */ 3694 if ((m->flags & PG_FICTITIOUS) != 0) { 3695 KASSERT(m->ref_count == 1, 3696 ("fictitious page %p is referenced", m)); 3697 KASSERT(m->queue == PQ_NONE, 3698 ("fictitious page %p is queued", m)); 3699 return (false); 3700 } 3701 3702 /* 3703 * Pages need not be dequeued before they are returned to the physical 3704 * memory allocator, but they must at least be marked for a deferred 3705 * dequeue. 3706 */ 3707 if ((m->oflags & VPO_UNMANAGED) == 0) 3708 vm_page_dequeue_deferred_free(m); 3709 3710 m->valid = 0; 3711 vm_page_undirty(m); 3712 3713 if (m->ref_count != 0) 3714 panic("vm_page_free_prep: page %p has references", m); 3715 3716 /* 3717 * Restore the default memory attribute to the page. 3718 */ 3719 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3720 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3721 3722 #if VM_NRESERVLEVEL > 0 3723 /* 3724 * Determine whether the page belongs to a reservation. If the page was 3725 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3726 * as an optimization, we avoid the check in that case. 3727 */ 3728 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3729 return (false); 3730 #endif 3731 3732 return (true); 3733 } 3734 3735 /* 3736 * vm_page_free_toq: 3737 * 3738 * Returns the given page to the free list, disassociating it 3739 * from any VM object. 3740 * 3741 * The object must be locked. The page must be locked if it is 3742 * managed. 3743 */ 3744 void 3745 vm_page_free_toq(vm_page_t m) 3746 { 3747 struct vm_domain *vmd; 3748 uma_zone_t zone; 3749 3750 if (!vm_page_free_prep(m)) 3751 return; 3752 3753 vmd = vm_pagequeue_domain(m); 3754 zone = vmd->vmd_pgcache[m->pool].zone; 3755 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 3756 uma_zfree(zone, m); 3757 return; 3758 } 3759 vm_domain_free_lock(vmd); 3760 vm_phys_free_pages(m, 0); 3761 vm_domain_free_unlock(vmd); 3762 vm_domain_freecnt_inc(vmd, 1); 3763 } 3764 3765 /* 3766 * vm_page_free_pages_toq: 3767 * 3768 * Returns a list of pages to the free list, disassociating it 3769 * from any VM object. In other words, this is equivalent to 3770 * calling vm_page_free_toq() for each page of a list of VM objects. 3771 * 3772 * The objects must be locked. The pages must be locked if it is 3773 * managed. 3774 */ 3775 void 3776 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3777 { 3778 vm_page_t m; 3779 int count; 3780 3781 if (SLIST_EMPTY(free)) 3782 return; 3783 3784 count = 0; 3785 while ((m = SLIST_FIRST(free)) != NULL) { 3786 count++; 3787 SLIST_REMOVE_HEAD(free, plinks.s.ss); 3788 vm_page_free_toq(m); 3789 } 3790 3791 if (update_wire_count) 3792 vm_wire_sub(count); 3793 } 3794 3795 /* 3796 * Mark this page as wired down, preventing reclamation by the page daemon 3797 * or when the containing object is destroyed. 3798 */ 3799 void 3800 vm_page_wire(vm_page_t m) 3801 { 3802 u_int old; 3803 3804 KASSERT(m->object != NULL, 3805 ("vm_page_wire: page %p does not belong to an object", m)); 3806 if (!vm_page_busied(m) && !vm_object_busied(m->object)) 3807 VM_OBJECT_ASSERT_LOCKED(m->object); 3808 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 3809 VPRC_WIRE_COUNT(m->ref_count) >= 1, 3810 ("vm_page_wire: fictitious page %p has zero wirings", m)); 3811 3812 old = atomic_fetchadd_int(&m->ref_count, 1); 3813 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 3814 ("vm_page_wire: counter overflow for page %p", m)); 3815 if (VPRC_WIRE_COUNT(old) == 0) 3816 vm_wire_add(1); 3817 } 3818 3819 /* 3820 * Attempt to wire a mapped page following a pmap lookup of that page. 3821 * This may fail if a thread is concurrently tearing down mappings of the page. 3822 * The transient failure is acceptable because it translates to the 3823 * failure of the caller pmap_extract_and_hold(), which should be then 3824 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 3825 */ 3826 bool 3827 vm_page_wire_mapped(vm_page_t m) 3828 { 3829 u_int old; 3830 3831 old = m->ref_count; 3832 do { 3833 KASSERT(old > 0, 3834 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 3835 if ((old & VPRC_BLOCKED) != 0) 3836 return (false); 3837 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 3838 3839 if (VPRC_WIRE_COUNT(old) == 0) 3840 vm_wire_add(1); 3841 return (true); 3842 } 3843 3844 /* 3845 * Release one wiring of the specified page, potentially allowing it to be 3846 * paged out. 3847 * 3848 * Only managed pages belonging to an object can be paged out. If the number 3849 * of wirings transitions to zero and the page is eligible for page out, then 3850 * the page is added to the specified paging queue. If the released wiring 3851 * represented the last reference to the page, the page is freed. 3852 * 3853 * A managed page must be locked. 3854 */ 3855 void 3856 vm_page_unwire(vm_page_t m, uint8_t queue) 3857 { 3858 u_int old; 3859 bool locked; 3860 3861 KASSERT(queue < PQ_COUNT, 3862 ("vm_page_unwire: invalid queue %u request for page %p", queue, m)); 3863 3864 if ((m->oflags & VPO_UNMANAGED) != 0) { 3865 if (vm_page_unwire_noq(m) && m->ref_count == 0) 3866 vm_page_free(m); 3867 return; 3868 } 3869 3870 /* 3871 * Update LRU state before releasing the wiring reference. 3872 * We only need to do this once since we hold the page lock. 3873 * Use a release store when updating the reference count to 3874 * synchronize with vm_page_free_prep(). 3875 */ 3876 old = m->ref_count; 3877 locked = false; 3878 do { 3879 KASSERT(VPRC_WIRE_COUNT(old) > 0, 3880 ("vm_page_unwire: wire count underflow for page %p", m)); 3881 if (!locked && VPRC_WIRE_COUNT(old) == 1) { 3882 vm_page_lock(m); 3883 locked = true; 3884 if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE) 3885 vm_page_reference(m); 3886 else 3887 vm_page_mvqueue(m, queue); 3888 } 3889 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 3890 3891 /* 3892 * Release the lock only after the wiring is released, to ensure that 3893 * the page daemon does not encounter and dequeue the page while it is 3894 * still wired. 3895 */ 3896 if (locked) 3897 vm_page_unlock(m); 3898 3899 if (VPRC_WIRE_COUNT(old) == 1) { 3900 vm_wire_sub(1); 3901 if (old == 1) 3902 vm_page_free(m); 3903 } 3904 } 3905 3906 /* 3907 * Unwire a page without (re-)inserting it into a page queue. It is up 3908 * to the caller to enqueue, requeue, or free the page as appropriate. 3909 * In most cases involving managed pages, vm_page_unwire() should be used 3910 * instead. 3911 */ 3912 bool 3913 vm_page_unwire_noq(vm_page_t m) 3914 { 3915 u_int old; 3916 3917 old = vm_page_drop(m, 1); 3918 KASSERT(VPRC_WIRE_COUNT(old) != 0, 3919 ("vm_page_unref: counter underflow for page %p", m)); 3920 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 3921 ("vm_page_unref: missing ref on fictitious page %p", m)); 3922 3923 if (VPRC_WIRE_COUNT(old) > 1) 3924 return (false); 3925 vm_wire_sub(1); 3926 return (true); 3927 } 3928 3929 /* 3930 * Ensure that the page is in the specified page queue. If the page is 3931 * active or being moved to the active queue, ensure that its act_count is 3932 * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that 3933 * the page is at the tail of its page queue. 3934 * 3935 * The page may be wired. The caller should release its wiring reference 3936 * before releasing the page lock, otherwise the page daemon may immediately 3937 * dequeue the page. 3938 * 3939 * A managed page must be locked. 3940 */ 3941 static __always_inline void 3942 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue) 3943 { 3944 3945 vm_page_assert_locked(m); 3946 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3947 ("vm_page_mvqueue: page %p is unmanaged", m)); 3948 KASSERT(m->ref_count > 0, 3949 ("%s: page %p does not carry any references", __func__, m)); 3950 3951 if (vm_page_queue(m) != nqueue) { 3952 vm_page_dequeue(m); 3953 vm_page_enqueue(m, nqueue); 3954 } else if (nqueue != PQ_ACTIVE) { 3955 vm_page_requeue(m); 3956 } 3957 3958 if (nqueue == PQ_ACTIVE && m->act_count < ACT_INIT) 3959 m->act_count = ACT_INIT; 3960 } 3961 3962 /* 3963 * Put the specified page on the active list (if appropriate). 3964 */ 3965 void 3966 vm_page_activate(vm_page_t m) 3967 { 3968 3969 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3970 return; 3971 vm_page_mvqueue(m, PQ_ACTIVE); 3972 } 3973 3974 /* 3975 * Move the specified page to the tail of the inactive queue, or requeue 3976 * the page if it is already in the inactive queue. 3977 */ 3978 void 3979 vm_page_deactivate(vm_page_t m) 3980 { 3981 3982 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3983 return; 3984 vm_page_mvqueue(m, PQ_INACTIVE); 3985 } 3986 3987 /* 3988 * Move the specified page close to the head of the inactive queue, 3989 * bypassing LRU. A marker page is used to maintain FIFO ordering. 3990 * As with regular enqueues, we use a per-CPU batch queue to reduce 3991 * contention on the page queue lock. 3992 */ 3993 static void 3994 _vm_page_deactivate_noreuse(vm_page_t m) 3995 { 3996 3997 vm_page_assert_locked(m); 3998 3999 if (!vm_page_inactive(m)) { 4000 vm_page_dequeue(m); 4001 m->queue = PQ_INACTIVE; 4002 } 4003 if ((m->aflags & PGA_REQUEUE_HEAD) == 0) 4004 vm_page_aflag_set(m, PGA_REQUEUE_HEAD); 4005 vm_page_pqbatch_submit(m, PQ_INACTIVE); 4006 } 4007 4008 void 4009 vm_page_deactivate_noreuse(vm_page_t m) 4010 { 4011 4012 KASSERT(m->object != NULL, 4013 ("vm_page_deactivate_noreuse: page %p has no object", m)); 4014 4015 if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m)) 4016 _vm_page_deactivate_noreuse(m); 4017 } 4018 4019 /* 4020 * Put a page in the laundry, or requeue it if it is already there. 4021 */ 4022 void 4023 vm_page_launder(vm_page_t m) 4024 { 4025 4026 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4027 return; 4028 vm_page_mvqueue(m, PQ_LAUNDRY); 4029 } 4030 4031 /* 4032 * Put a page in the PQ_UNSWAPPABLE holding queue. 4033 */ 4034 void 4035 vm_page_unswappable(vm_page_t m) 4036 { 4037 4038 vm_page_assert_locked(m); 4039 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 4040 ("page %p already unswappable", m)); 4041 4042 vm_page_dequeue(m); 4043 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4044 } 4045 4046 static void 4047 vm_page_release_toq(vm_page_t m, int flags) 4048 { 4049 4050 vm_page_assert_locked(m); 4051 4052 /* 4053 * Use a check of the valid bits to determine whether we should 4054 * accelerate reclamation of the page. The object lock might not be 4055 * held here, in which case the check is racy. At worst we will either 4056 * accelerate reclamation of a valid page and violate LRU, or 4057 * unnecessarily defer reclamation of an invalid page. 4058 * 4059 * If we were asked to not cache the page, place it near the head of the 4060 * inactive queue so that is reclaimed sooner. 4061 */ 4062 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0) 4063 _vm_page_deactivate_noreuse(m); 4064 else if (vm_page_active(m)) 4065 vm_page_reference(m); 4066 else 4067 vm_page_mvqueue(m, PQ_INACTIVE); 4068 } 4069 4070 /* 4071 * Unwire a page and either attempt to free it or re-add it to the page queues. 4072 */ 4073 void 4074 vm_page_release(vm_page_t m, int flags) 4075 { 4076 vm_object_t object; 4077 u_int old; 4078 bool locked; 4079 4080 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4081 ("vm_page_release: page %p is unmanaged", m)); 4082 4083 if ((flags & VPR_TRYFREE) != 0) { 4084 for (;;) { 4085 object = (vm_object_t)atomic_load_ptr(&m->object); 4086 if (object == NULL) 4087 break; 4088 /* Depends on type-stability. */ 4089 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) { 4090 object = NULL; 4091 break; 4092 } 4093 if (object == m->object) 4094 break; 4095 VM_OBJECT_WUNLOCK(object); 4096 } 4097 if (__predict_true(object != NULL)) { 4098 vm_page_release_locked(m, flags); 4099 VM_OBJECT_WUNLOCK(object); 4100 return; 4101 } 4102 } 4103 4104 /* 4105 * Update LRU state before releasing the wiring reference. 4106 * Use a release store when updating the reference count to 4107 * synchronize with vm_page_free_prep(). 4108 */ 4109 old = m->ref_count; 4110 locked = false; 4111 do { 4112 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4113 ("vm_page_unwire: wire count underflow for page %p", m)); 4114 if (!locked && VPRC_WIRE_COUNT(old) == 1) { 4115 vm_page_lock(m); 4116 locked = true; 4117 vm_page_release_toq(m, flags); 4118 } 4119 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4120 4121 /* 4122 * Release the lock only after the wiring is released, to ensure that 4123 * the page daemon does not encounter and dequeue the page while it is 4124 * still wired. 4125 */ 4126 if (locked) 4127 vm_page_unlock(m); 4128 4129 if (VPRC_WIRE_COUNT(old) == 1) { 4130 vm_wire_sub(1); 4131 if (old == 1) 4132 vm_page_free(m); 4133 } 4134 } 4135 4136 /* See vm_page_release(). */ 4137 void 4138 vm_page_release_locked(vm_page_t m, int flags) 4139 { 4140 4141 VM_OBJECT_ASSERT_WLOCKED(m->object); 4142 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4143 ("vm_page_release_locked: page %p is unmanaged", m)); 4144 4145 if (vm_page_unwire_noq(m)) { 4146 if ((flags & VPR_TRYFREE) != 0 && 4147 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4148 m->dirty == 0 && !vm_page_busied(m)) { 4149 vm_page_free(m); 4150 } else { 4151 vm_page_lock(m); 4152 vm_page_release_toq(m, flags); 4153 vm_page_unlock(m); 4154 } 4155 } 4156 } 4157 4158 static bool 4159 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4160 { 4161 u_int old; 4162 4163 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4164 ("vm_page_try_blocked_op: page %p has no object", m)); 4165 KASSERT(vm_page_busied(m), 4166 ("vm_page_try_blocked_op: page %p is not busy", m)); 4167 VM_OBJECT_ASSERT_LOCKED(m->object); 4168 4169 old = m->ref_count; 4170 do { 4171 KASSERT(old != 0, 4172 ("vm_page_try_blocked_op: page %p has no references", m)); 4173 if (VPRC_WIRE_COUNT(old) != 0) 4174 return (false); 4175 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4176 4177 (op)(m); 4178 4179 /* 4180 * If the object is read-locked, new wirings may be created via an 4181 * object lookup. 4182 */ 4183 old = vm_page_drop(m, VPRC_BLOCKED); 4184 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4185 old == (VPRC_BLOCKED | VPRC_OBJREF), 4186 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4187 old, m)); 4188 return (true); 4189 } 4190 4191 /* 4192 * Atomically check for wirings and remove all mappings of the page. 4193 */ 4194 bool 4195 vm_page_try_remove_all(vm_page_t m) 4196 { 4197 4198 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4199 } 4200 4201 /* 4202 * Atomically check for wirings and remove all writeable mappings of the page. 4203 */ 4204 bool 4205 vm_page_try_remove_write(vm_page_t m) 4206 { 4207 4208 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4209 } 4210 4211 /* 4212 * vm_page_advise 4213 * 4214 * Apply the specified advice to the given page. 4215 * 4216 * The object and page must be locked. 4217 */ 4218 void 4219 vm_page_advise(vm_page_t m, int advice) 4220 { 4221 4222 vm_page_assert_locked(m); 4223 VM_OBJECT_ASSERT_WLOCKED(m->object); 4224 if (advice == MADV_FREE) 4225 /* 4226 * Mark the page clean. This will allow the page to be freed 4227 * without first paging it out. MADV_FREE pages are often 4228 * quickly reused by malloc(3), so we do not do anything that 4229 * would result in a page fault on a later access. 4230 */ 4231 vm_page_undirty(m); 4232 else if (advice != MADV_DONTNEED) { 4233 if (advice == MADV_WILLNEED) 4234 vm_page_activate(m); 4235 return; 4236 } 4237 4238 /* 4239 * Clear any references to the page. Otherwise, the page daemon will 4240 * immediately reactivate the page. 4241 */ 4242 vm_page_aflag_clear(m, PGA_REFERENCED); 4243 4244 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4245 vm_page_dirty(m); 4246 4247 /* 4248 * Place clean pages near the head of the inactive queue rather than 4249 * the tail, thus defeating the queue's LRU operation and ensuring that 4250 * the page will be reused quickly. Dirty pages not already in the 4251 * laundry are moved there. 4252 */ 4253 if (m->dirty == 0) 4254 vm_page_deactivate_noreuse(m); 4255 else if (!vm_page_in_laundry(m)) 4256 vm_page_launder(m); 4257 } 4258 4259 static inline int 4260 vm_page_grab_pflags(int allocflags) 4261 { 4262 int pflags; 4263 4264 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4265 (allocflags & VM_ALLOC_WIRED) != 0, 4266 ("vm_page_grab_pflags: the pages must be busied or wired")); 4267 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4268 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4269 ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 4270 "mismatch")); 4271 pflags = allocflags & 4272 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4273 VM_ALLOC_NOBUSY); 4274 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4275 pflags |= VM_ALLOC_WAITFAIL; 4276 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4277 pflags |= VM_ALLOC_SBUSY; 4278 4279 return (pflags); 4280 } 4281 4282 /* 4283 * Grab a page, waiting until we are waken up due to the page 4284 * changing state. We keep on waiting, if the page continues 4285 * to be in the object. If the page doesn't exist, first allocate it 4286 * and then conditionally zero it. 4287 * 4288 * This routine may sleep. 4289 * 4290 * The object must be locked on entry. The lock will, however, be released 4291 * and reacquired if the routine sleeps. 4292 */ 4293 vm_page_t 4294 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4295 { 4296 vm_page_t m; 4297 int pflags; 4298 4299 VM_OBJECT_ASSERT_WLOCKED(object); 4300 pflags = vm_page_grab_pflags(allocflags); 4301 retrylookup: 4302 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4303 if (!vm_page_acquire_flags(m, allocflags)) { 4304 if (vm_page_busy_sleep_flags(object, m, "pgrbwt", 4305 allocflags)) 4306 goto retrylookup; 4307 return (NULL); 4308 } 4309 goto out; 4310 } 4311 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4312 return (NULL); 4313 m = vm_page_alloc(object, pindex, pflags); 4314 if (m == NULL) { 4315 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4316 return (NULL); 4317 goto retrylookup; 4318 } 4319 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4320 pmap_zero_page(m); 4321 4322 out: 4323 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4324 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4325 vm_page_sunbusy(m); 4326 else 4327 vm_page_xunbusy(m); 4328 } 4329 return (m); 4330 } 4331 4332 /* 4333 * Grab a page and make it valid, paging in if necessary. Pages missing from 4334 * their pager are zero filled and validated. 4335 */ 4336 int 4337 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4338 { 4339 vm_page_t m; 4340 bool sleep, xbusy; 4341 int pflags; 4342 int rv; 4343 4344 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4345 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4346 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4347 KASSERT((allocflags & 4348 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4349 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4350 VM_OBJECT_ASSERT_WLOCKED(object); 4351 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY); 4352 pflags |= VM_ALLOC_WAITFAIL; 4353 4354 retrylookup: 4355 xbusy = false; 4356 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4357 /* 4358 * If the page is fully valid it can only become invalid 4359 * with the object lock held. If it is not valid it can 4360 * become valid with the busy lock held. Therefore, we 4361 * may unnecessarily lock the exclusive busy here if we 4362 * race with I/O completion not using the object lock. 4363 * However, we will not end up with an invalid page and a 4364 * shared lock. 4365 */ 4366 if (!vm_page_all_valid(m) || 4367 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) { 4368 sleep = !vm_page_tryxbusy(m); 4369 xbusy = true; 4370 } else 4371 sleep = !vm_page_trysbusy(m); 4372 if (sleep) { 4373 (void)vm_page_busy_sleep_flags(object, m, "pgrbwt", 4374 allocflags); 4375 goto retrylookup; 4376 } 4377 if ((allocflags & VM_ALLOC_NOCREAT) != 0 && 4378 !vm_page_all_valid(m)) { 4379 if (xbusy) 4380 vm_page_xunbusy(m); 4381 else 4382 vm_page_sunbusy(m); 4383 *mp = NULL; 4384 return (VM_PAGER_FAIL); 4385 } 4386 if ((allocflags & VM_ALLOC_WIRED) != 0) 4387 vm_page_wire(m); 4388 if (vm_page_all_valid(m)) 4389 goto out; 4390 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4391 *mp = NULL; 4392 return (VM_PAGER_FAIL); 4393 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) { 4394 xbusy = true; 4395 } else { 4396 goto retrylookup; 4397 } 4398 4399 vm_page_assert_xbusied(m); 4400 MPASS(xbusy); 4401 if (vm_pager_has_page(object, pindex, NULL, NULL)) { 4402 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 4403 if (rv != VM_PAGER_OK) { 4404 if (allocflags & VM_ALLOC_WIRED) 4405 vm_page_unwire_noq(m); 4406 vm_page_free(m); 4407 *mp = NULL; 4408 return (rv); 4409 } 4410 MPASS(vm_page_all_valid(m)); 4411 } else { 4412 vm_page_zero_invalid(m, TRUE); 4413 } 4414 out: 4415 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4416 if (xbusy) 4417 vm_page_xunbusy(m); 4418 else 4419 vm_page_sunbusy(m); 4420 } 4421 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy) 4422 vm_page_busy_downgrade(m); 4423 *mp = m; 4424 return (VM_PAGER_OK); 4425 } 4426 4427 /* 4428 * Return the specified range of pages from the given object. For each 4429 * page offset within the range, if a page already exists within the object 4430 * at that offset and it is busy, then wait for it to change state. If, 4431 * instead, the page doesn't exist, then allocate it. 4432 * 4433 * The caller must always specify an allocation class. 4434 * 4435 * allocation classes: 4436 * VM_ALLOC_NORMAL normal process request 4437 * VM_ALLOC_SYSTEM system *really* needs the pages 4438 * 4439 * The caller must always specify that the pages are to be busied and/or 4440 * wired. 4441 * 4442 * optional allocation flags: 4443 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4444 * VM_ALLOC_NOBUSY do not exclusive busy the page 4445 * VM_ALLOC_NOWAIT do not sleep 4446 * VM_ALLOC_SBUSY set page to sbusy state 4447 * VM_ALLOC_WIRED wire the pages 4448 * VM_ALLOC_ZERO zero and validate any invalid pages 4449 * 4450 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4451 * may return a partial prefix of the requested range. 4452 */ 4453 int 4454 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4455 vm_page_t *ma, int count) 4456 { 4457 vm_page_t m, mpred; 4458 int pflags; 4459 int i; 4460 4461 VM_OBJECT_ASSERT_WLOCKED(object); 4462 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4463 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4464 4465 pflags = vm_page_grab_pflags(allocflags); 4466 if (count == 0) 4467 return (0); 4468 4469 i = 0; 4470 retrylookup: 4471 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4472 if (m == NULL || m->pindex != pindex + i) { 4473 mpred = m; 4474 m = NULL; 4475 } else 4476 mpred = TAILQ_PREV(m, pglist, listq); 4477 for (; i < count; i++) { 4478 if (m != NULL) { 4479 if (!vm_page_acquire_flags(m, allocflags)) { 4480 if (vm_page_busy_sleep_flags(object, m, 4481 "grbmaw", allocflags)) 4482 goto retrylookup; 4483 break; 4484 } 4485 } else { 4486 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4487 break; 4488 m = vm_page_alloc_after(object, pindex + i, 4489 pflags | VM_ALLOC_COUNT(count - i), mpred); 4490 if (m == NULL) { 4491 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4492 break; 4493 goto retrylookup; 4494 } 4495 } 4496 if (vm_page_none_valid(m) && 4497 (allocflags & VM_ALLOC_ZERO) != 0) { 4498 if ((m->flags & PG_ZERO) == 0) 4499 pmap_zero_page(m); 4500 vm_page_valid(m); 4501 } 4502 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4503 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4504 vm_page_sunbusy(m); 4505 else 4506 vm_page_xunbusy(m); 4507 } 4508 ma[i] = mpred = m; 4509 m = vm_page_next(m); 4510 } 4511 return (i); 4512 } 4513 4514 /* 4515 * Mapping function for valid or dirty bits in a page. 4516 * 4517 * Inputs are required to range within a page. 4518 */ 4519 vm_page_bits_t 4520 vm_page_bits(int base, int size) 4521 { 4522 int first_bit; 4523 int last_bit; 4524 4525 KASSERT( 4526 base + size <= PAGE_SIZE, 4527 ("vm_page_bits: illegal base/size %d/%d", base, size) 4528 ); 4529 4530 if (size == 0) /* handle degenerate case */ 4531 return (0); 4532 4533 first_bit = base >> DEV_BSHIFT; 4534 last_bit = (base + size - 1) >> DEV_BSHIFT; 4535 4536 return (((vm_page_bits_t)2 << last_bit) - 4537 ((vm_page_bits_t)1 << first_bit)); 4538 } 4539 4540 void 4541 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 4542 { 4543 4544 #if PAGE_SIZE == 32768 4545 atomic_set_64((uint64_t *)bits, set); 4546 #elif PAGE_SIZE == 16384 4547 atomic_set_32((uint32_t *)bits, set); 4548 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 4549 atomic_set_16((uint16_t *)bits, set); 4550 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 4551 atomic_set_8((uint8_t *)bits, set); 4552 #else /* PAGE_SIZE <= 8192 */ 4553 uintptr_t addr; 4554 int shift; 4555 4556 addr = (uintptr_t)bits; 4557 /* 4558 * Use a trick to perform a 32-bit atomic on the 4559 * containing aligned word, to not depend on the existence 4560 * of atomic_{set, clear}_{8, 16}. 4561 */ 4562 shift = addr & (sizeof(uint32_t) - 1); 4563 #if BYTE_ORDER == BIG_ENDIAN 4564 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4565 #else 4566 shift *= NBBY; 4567 #endif 4568 addr &= ~(sizeof(uint32_t) - 1); 4569 atomic_set_32((uint32_t *)addr, set << shift); 4570 #endif /* PAGE_SIZE */ 4571 } 4572 4573 static inline void 4574 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 4575 { 4576 4577 #if PAGE_SIZE == 32768 4578 atomic_clear_64((uint64_t *)bits, clear); 4579 #elif PAGE_SIZE == 16384 4580 atomic_clear_32((uint32_t *)bits, clear); 4581 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 4582 atomic_clear_16((uint16_t *)bits, clear); 4583 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 4584 atomic_clear_8((uint8_t *)bits, clear); 4585 #else /* PAGE_SIZE <= 8192 */ 4586 uintptr_t addr; 4587 int shift; 4588 4589 addr = (uintptr_t)bits; 4590 /* 4591 * Use a trick to perform a 32-bit atomic on the 4592 * containing aligned word, to not depend on the existence 4593 * of atomic_{set, clear}_{8, 16}. 4594 */ 4595 shift = addr & (sizeof(uint32_t) - 1); 4596 #if BYTE_ORDER == BIG_ENDIAN 4597 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4598 #else 4599 shift *= NBBY; 4600 #endif 4601 addr &= ~(sizeof(uint32_t) - 1); 4602 atomic_clear_32((uint32_t *)addr, clear << shift); 4603 #endif /* PAGE_SIZE */ 4604 } 4605 4606 /* 4607 * vm_page_set_valid_range: 4608 * 4609 * Sets portions of a page valid. The arguments are expected 4610 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4611 * of any partial chunks touched by the range. The invalid portion of 4612 * such chunks will be zeroed. 4613 * 4614 * (base + size) must be less then or equal to PAGE_SIZE. 4615 */ 4616 void 4617 vm_page_set_valid_range(vm_page_t m, int base, int size) 4618 { 4619 int endoff, frag; 4620 vm_page_bits_t pagebits; 4621 4622 vm_page_assert_busied(m); 4623 if (size == 0) /* handle degenerate case */ 4624 return; 4625 4626 /* 4627 * If the base is not DEV_BSIZE aligned and the valid 4628 * bit is clear, we have to zero out a portion of the 4629 * first block. 4630 */ 4631 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4632 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 4633 pmap_zero_page_area(m, frag, base - frag); 4634 4635 /* 4636 * If the ending offset is not DEV_BSIZE aligned and the 4637 * valid bit is clear, we have to zero out a portion of 4638 * the last block. 4639 */ 4640 endoff = base + size; 4641 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4642 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 4643 pmap_zero_page_area(m, endoff, 4644 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4645 4646 /* 4647 * Assert that no previously invalid block that is now being validated 4648 * is already dirty. 4649 */ 4650 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 4651 ("vm_page_set_valid_range: page %p is dirty", m)); 4652 4653 /* 4654 * Set valid bits inclusive of any overlap. 4655 */ 4656 pagebits = vm_page_bits(base, size); 4657 if (vm_page_xbusied(m)) 4658 m->valid |= pagebits; 4659 else 4660 vm_page_bits_set(m, &m->valid, pagebits); 4661 } 4662 4663 /* 4664 * Clear the given bits from the specified page's dirty field. 4665 */ 4666 static __inline void 4667 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 4668 { 4669 4670 vm_page_assert_busied(m); 4671 4672 /* 4673 * If the page is xbusied and not write mapped we are the 4674 * only thread that can modify dirty bits. Otherwise, The pmap 4675 * layer can call vm_page_dirty() without holding a distinguished 4676 * lock. The combination of page busy and atomic operations 4677 * suffice to guarantee consistency of the page dirty field. 4678 */ 4679 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 4680 m->dirty &= ~pagebits; 4681 else 4682 vm_page_bits_clear(m, &m->dirty, pagebits); 4683 } 4684 4685 /* 4686 * vm_page_set_validclean: 4687 * 4688 * Sets portions of a page valid and clean. The arguments are expected 4689 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4690 * of any partial chunks touched by the range. The invalid portion of 4691 * such chunks will be zero'd. 4692 * 4693 * (base + size) must be less then or equal to PAGE_SIZE. 4694 */ 4695 void 4696 vm_page_set_validclean(vm_page_t m, int base, int size) 4697 { 4698 vm_page_bits_t oldvalid, pagebits; 4699 int endoff, frag; 4700 4701 vm_page_assert_busied(m); 4702 if (size == 0) /* handle degenerate case */ 4703 return; 4704 4705 /* 4706 * If the base is not DEV_BSIZE aligned and the valid 4707 * bit is clear, we have to zero out a portion of the 4708 * first block. 4709 */ 4710 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4711 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 4712 pmap_zero_page_area(m, frag, base - frag); 4713 4714 /* 4715 * If the ending offset is not DEV_BSIZE aligned and the 4716 * valid bit is clear, we have to zero out a portion of 4717 * the last block. 4718 */ 4719 endoff = base + size; 4720 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4721 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 4722 pmap_zero_page_area(m, endoff, 4723 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4724 4725 /* 4726 * Set valid, clear dirty bits. If validating the entire 4727 * page we can safely clear the pmap modify bit. We also 4728 * use this opportunity to clear the PGA_NOSYNC flag. If a process 4729 * takes a write fault on a MAP_NOSYNC memory area the flag will 4730 * be set again. 4731 * 4732 * We set valid bits inclusive of any overlap, but we can only 4733 * clear dirty bits for DEV_BSIZE chunks that are fully within 4734 * the range. 4735 */ 4736 oldvalid = m->valid; 4737 pagebits = vm_page_bits(base, size); 4738 if (vm_page_xbusied(m)) 4739 m->valid |= pagebits; 4740 else 4741 vm_page_bits_set(m, &m->valid, pagebits); 4742 #if 0 /* NOT YET */ 4743 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 4744 frag = DEV_BSIZE - frag; 4745 base += frag; 4746 size -= frag; 4747 if (size < 0) 4748 size = 0; 4749 } 4750 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 4751 #endif 4752 if (base == 0 && size == PAGE_SIZE) { 4753 /* 4754 * The page can only be modified within the pmap if it is 4755 * mapped, and it can only be mapped if it was previously 4756 * fully valid. 4757 */ 4758 if (oldvalid == VM_PAGE_BITS_ALL) 4759 /* 4760 * Perform the pmap_clear_modify() first. Otherwise, 4761 * a concurrent pmap operation, such as 4762 * pmap_protect(), could clear a modification in the 4763 * pmap and set the dirty field on the page before 4764 * pmap_clear_modify() had begun and after the dirty 4765 * field was cleared here. 4766 */ 4767 pmap_clear_modify(m); 4768 m->dirty = 0; 4769 vm_page_aflag_clear(m, PGA_NOSYNC); 4770 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 4771 m->dirty &= ~pagebits; 4772 else 4773 vm_page_clear_dirty_mask(m, pagebits); 4774 } 4775 4776 void 4777 vm_page_clear_dirty(vm_page_t m, int base, int size) 4778 { 4779 4780 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 4781 } 4782 4783 /* 4784 * vm_page_set_invalid: 4785 * 4786 * Invalidates DEV_BSIZE'd chunks within a page. Both the 4787 * valid and dirty bits for the effected areas are cleared. 4788 */ 4789 void 4790 vm_page_set_invalid(vm_page_t m, int base, int size) 4791 { 4792 vm_page_bits_t bits; 4793 vm_object_t object; 4794 4795 /* 4796 * The object lock is required so that pages can't be mapped 4797 * read-only while we're in the process of invalidating them. 4798 */ 4799 object = m->object; 4800 VM_OBJECT_ASSERT_WLOCKED(object); 4801 vm_page_assert_busied(m); 4802 4803 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 4804 size >= object->un_pager.vnp.vnp_size) 4805 bits = VM_PAGE_BITS_ALL; 4806 else 4807 bits = vm_page_bits(base, size); 4808 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 4809 pmap_remove_all(m); 4810 KASSERT((bits == 0 && vm_page_all_valid(m)) || 4811 !pmap_page_is_mapped(m), 4812 ("vm_page_set_invalid: page %p is mapped", m)); 4813 if (vm_page_xbusied(m)) { 4814 m->valid &= ~bits; 4815 m->dirty &= ~bits; 4816 } else { 4817 vm_page_bits_clear(m, &m->valid, bits); 4818 vm_page_bits_clear(m, &m->dirty, bits); 4819 } 4820 } 4821 4822 /* 4823 * vm_page_invalid: 4824 * 4825 * Invalidates the entire page. The page must be busy, unmapped, and 4826 * the enclosing object must be locked. The object locks protects 4827 * against concurrent read-only pmap enter which is done without 4828 * busy. 4829 */ 4830 void 4831 vm_page_invalid(vm_page_t m) 4832 { 4833 4834 vm_page_assert_busied(m); 4835 VM_OBJECT_ASSERT_LOCKED(m->object); 4836 MPASS(!pmap_page_is_mapped(m)); 4837 4838 if (vm_page_xbusied(m)) 4839 m->valid = 0; 4840 else 4841 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 4842 } 4843 4844 /* 4845 * vm_page_zero_invalid() 4846 * 4847 * The kernel assumes that the invalid portions of a page contain 4848 * garbage, but such pages can be mapped into memory by user code. 4849 * When this occurs, we must zero out the non-valid portions of the 4850 * page so user code sees what it expects. 4851 * 4852 * Pages are most often semi-valid when the end of a file is mapped 4853 * into memory and the file's size is not page aligned. 4854 */ 4855 void 4856 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 4857 { 4858 int b; 4859 int i; 4860 4861 /* 4862 * Scan the valid bits looking for invalid sections that 4863 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 4864 * valid bit may be set ) have already been zeroed by 4865 * vm_page_set_validclean(). 4866 */ 4867 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 4868 if (i == (PAGE_SIZE / DEV_BSIZE) || 4869 (m->valid & ((vm_page_bits_t)1 << i))) { 4870 if (i > b) { 4871 pmap_zero_page_area(m, 4872 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 4873 } 4874 b = i + 1; 4875 } 4876 } 4877 4878 /* 4879 * setvalid is TRUE when we can safely set the zero'd areas 4880 * as being valid. We can do this if there are no cache consistancy 4881 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 4882 */ 4883 if (setvalid) 4884 vm_page_valid(m); 4885 } 4886 4887 /* 4888 * vm_page_is_valid: 4889 * 4890 * Is (partial) page valid? Note that the case where size == 0 4891 * will return FALSE in the degenerate case where the page is 4892 * entirely invalid, and TRUE otherwise. 4893 * 4894 * Some callers envoke this routine without the busy lock held and 4895 * handle races via higher level locks. Typical callers should 4896 * hold a busy lock to prevent invalidation. 4897 */ 4898 int 4899 vm_page_is_valid(vm_page_t m, int base, int size) 4900 { 4901 vm_page_bits_t bits; 4902 4903 bits = vm_page_bits(base, size); 4904 return (m->valid != 0 && (m->valid & bits) == bits); 4905 } 4906 4907 /* 4908 * Returns true if all of the specified predicates are true for the entire 4909 * (super)page and false otherwise. 4910 */ 4911 bool 4912 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 4913 { 4914 vm_object_t object; 4915 int i, npages; 4916 4917 object = m->object; 4918 if (skip_m != NULL && skip_m->object != object) 4919 return (false); 4920 VM_OBJECT_ASSERT_LOCKED(object); 4921 npages = atop(pagesizes[m->psind]); 4922 4923 /* 4924 * The physically contiguous pages that make up a superpage, i.e., a 4925 * page with a page size index ("psind") greater than zero, will 4926 * occupy adjacent entries in vm_page_array[]. 4927 */ 4928 for (i = 0; i < npages; i++) { 4929 /* Always test object consistency, including "skip_m". */ 4930 if (m[i].object != object) 4931 return (false); 4932 if (&m[i] == skip_m) 4933 continue; 4934 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 4935 return (false); 4936 if ((flags & PS_ALL_DIRTY) != 0) { 4937 /* 4938 * Calling vm_page_test_dirty() or pmap_is_modified() 4939 * might stop this case from spuriously returning 4940 * "false". However, that would require a write lock 4941 * on the object containing "m[i]". 4942 */ 4943 if (m[i].dirty != VM_PAGE_BITS_ALL) 4944 return (false); 4945 } 4946 if ((flags & PS_ALL_VALID) != 0 && 4947 m[i].valid != VM_PAGE_BITS_ALL) 4948 return (false); 4949 } 4950 return (true); 4951 } 4952 4953 /* 4954 * Set the page's dirty bits if the page is modified. 4955 */ 4956 void 4957 vm_page_test_dirty(vm_page_t m) 4958 { 4959 4960 vm_page_assert_busied(m); 4961 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 4962 vm_page_dirty(m); 4963 } 4964 4965 void 4966 vm_page_valid(vm_page_t m) 4967 { 4968 4969 vm_page_assert_busied(m); 4970 if (vm_page_xbusied(m)) 4971 m->valid = VM_PAGE_BITS_ALL; 4972 else 4973 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 4974 } 4975 4976 void 4977 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 4978 { 4979 4980 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 4981 } 4982 4983 void 4984 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 4985 { 4986 4987 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 4988 } 4989 4990 int 4991 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 4992 { 4993 4994 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 4995 } 4996 4997 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 4998 void 4999 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5000 { 5001 5002 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5003 } 5004 5005 void 5006 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5007 { 5008 5009 mtx_assert_(vm_page_lockptr(m), a, file, line); 5010 } 5011 #endif 5012 5013 #ifdef INVARIANTS 5014 void 5015 vm_page_object_busy_assert(vm_page_t m) 5016 { 5017 5018 /* 5019 * Certain of the page's fields may only be modified by the 5020 * holder of a page or object busy. 5021 */ 5022 if (m->object != NULL && !vm_page_busied(m)) 5023 VM_OBJECT_ASSERT_BUSY(m->object); 5024 } 5025 5026 void 5027 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5028 { 5029 5030 if ((bits & PGA_WRITEABLE) == 0) 5031 return; 5032 5033 /* 5034 * The PGA_WRITEABLE flag can only be set if the page is 5035 * managed, is exclusively busied or the object is locked. 5036 * Currently, this flag is only set by pmap_enter(). 5037 */ 5038 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5039 ("PGA_WRITEABLE on unmanaged page")); 5040 if (!vm_page_xbusied(m)) 5041 VM_OBJECT_ASSERT_BUSY(m->object); 5042 } 5043 #endif 5044 5045 #include "opt_ddb.h" 5046 #ifdef DDB 5047 #include <sys/kernel.h> 5048 5049 #include <ddb/ddb.h> 5050 5051 DB_SHOW_COMMAND(page, vm_page_print_page_info) 5052 { 5053 5054 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5055 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5056 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5057 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5058 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5059 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5060 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5061 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5062 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5063 } 5064 5065 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 5066 { 5067 int dom; 5068 5069 db_printf("pq_free %d\n", vm_free_count()); 5070 for (dom = 0; dom < vm_ndomains; dom++) { 5071 db_printf( 5072 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5073 dom, 5074 vm_dom[dom].vmd_page_count, 5075 vm_dom[dom].vmd_free_count, 5076 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5077 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5078 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5079 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5080 } 5081 } 5082 5083 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5084 { 5085 vm_page_t m; 5086 boolean_t phys, virt; 5087 5088 if (!have_addr) { 5089 db_printf("show pginfo addr\n"); 5090 return; 5091 } 5092 5093 phys = strchr(modif, 'p') != NULL; 5094 virt = strchr(modif, 'v') != NULL; 5095 if (virt) 5096 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5097 else if (phys) 5098 m = PHYS_TO_VM_PAGE(addr); 5099 else 5100 m = (vm_page_t)addr; 5101 db_printf( 5102 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" 5103 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5104 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5105 m->queue, m->ref_count, m->aflags, m->oflags, 5106 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 5107 } 5108 #endif /* DDB */ 5109