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, page_range, size; 563 vm_paddr_t last_pa, pa; 564 u_long pagecount; 565 int biggestone, i, segind; 566 #ifdef WITNESS 567 int witness_size; 568 #endif 569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 570 long ii; 571 #endif 572 573 vaddr = round_page(vaddr); 574 575 vm_phys_early_startup(); 576 biggestone = vm_phys_avail_largest(); 577 end = phys_avail[biggestone+1]; 578 579 /* 580 * Initialize the page and queue locks. 581 */ 582 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 583 for (i = 0; i < PA_LOCK_COUNT; i++) 584 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 585 for (i = 0; i < vm_ndomains; i++) 586 vm_page_domain_init(i); 587 588 /* 589 * Allocate memory for use when boot strapping the kernel memory 590 * allocator. Tell UMA how many zones we are going to create 591 * before going fully functional. UMA will add its zones. 592 * 593 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, 594 * KMAP ENTRY, MAP ENTRY, VMSPACE. 595 */ 596 boot_pages = uma_startup_count(8); 597 598 #ifndef UMA_MD_SMALL_ALLOC 599 /* vmem_startup() calls uma_prealloc(). */ 600 boot_pages += vmem_startup_count(); 601 /* vm_map_startup() calls uma_prealloc(). */ 602 boot_pages += howmany(MAX_KMAP, 603 slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR)); 604 605 /* 606 * Before going fully functional kmem_init() does allocation 607 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". 608 */ 609 boot_pages += 2; 610 #endif 611 /* 612 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 613 * manually fetch the value. 614 */ 615 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 616 new_end = end - (boot_pages * UMA_SLAB_SIZE); 617 new_end = trunc_page(new_end); 618 mapped = pmap_map(&vaddr, new_end, end, 619 VM_PROT_READ | VM_PROT_WRITE); 620 bzero((void *)mapped, end - new_end); 621 uma_startup((void *)mapped, boot_pages); 622 623 #ifdef WITNESS 624 witness_size = round_page(witness_startup_count()); 625 new_end -= witness_size; 626 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 627 VM_PROT_READ | VM_PROT_WRITE); 628 bzero((void *)mapped, witness_size); 629 witness_startup((void *)mapped); 630 #endif 631 632 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 633 defined(__i386__) || defined(__mips__) || defined(__riscv) || \ 634 defined(__powerpc64__) 635 /* 636 * Allocate a bitmap to indicate that a random physical page 637 * needs to be included in a minidump. 638 * 639 * The amd64 port needs this to indicate which direct map pages 640 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 641 * 642 * However, i386 still needs this workspace internally within the 643 * minidump code. In theory, they are not needed on i386, but are 644 * included should the sf_buf code decide to use them. 645 */ 646 last_pa = 0; 647 for (i = 0; dump_avail[i + 1] != 0; i += 2) 648 if (dump_avail[i + 1] > last_pa) 649 last_pa = dump_avail[i + 1]; 650 page_range = last_pa / PAGE_SIZE; 651 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 652 new_end -= vm_page_dump_size; 653 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 654 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 655 bzero((void *)vm_page_dump, vm_page_dump_size); 656 #else 657 (void)last_pa; 658 #endif 659 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 660 defined(__riscv) || defined(__powerpc64__) 661 /* 662 * Include the UMA bootstrap pages, witness pages and vm_page_dump 663 * in a crash dump. When pmap_map() uses the direct map, they are 664 * not automatically included. 665 */ 666 for (pa = new_end; pa < end; pa += PAGE_SIZE) 667 dump_add_page(pa); 668 #endif 669 phys_avail[biggestone + 1] = new_end; 670 #ifdef __amd64__ 671 /* 672 * Request that the physical pages underlying the message buffer be 673 * included in a crash dump. Since the message buffer is accessed 674 * through the direct map, they are not automatically included. 675 */ 676 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 677 last_pa = pa + round_page(msgbufsize); 678 while (pa < last_pa) { 679 dump_add_page(pa); 680 pa += PAGE_SIZE; 681 } 682 #endif 683 /* 684 * Compute the number of pages of memory that will be available for 685 * use, taking into account the overhead of a page structure per page. 686 * In other words, solve 687 * "available physical memory" - round_page(page_range * 688 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 689 * for page_range. 690 */ 691 low_avail = phys_avail[0]; 692 high_avail = phys_avail[1]; 693 for (i = 0; i < vm_phys_nsegs; i++) { 694 if (vm_phys_segs[i].start < low_avail) 695 low_avail = vm_phys_segs[i].start; 696 if (vm_phys_segs[i].end > high_avail) 697 high_avail = vm_phys_segs[i].end; 698 } 699 /* Skip the first chunk. It is already accounted for. */ 700 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 701 if (phys_avail[i] < low_avail) 702 low_avail = phys_avail[i]; 703 if (phys_avail[i + 1] > high_avail) 704 high_avail = phys_avail[i + 1]; 705 } 706 first_page = low_avail / PAGE_SIZE; 707 #ifdef VM_PHYSSEG_SPARSE 708 size = 0; 709 for (i = 0; i < vm_phys_nsegs; i++) 710 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 711 for (i = 0; phys_avail[i + 1] != 0; i += 2) 712 size += phys_avail[i + 1] - phys_avail[i]; 713 #elif defined(VM_PHYSSEG_DENSE) 714 size = high_avail - low_avail; 715 #else 716 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 717 #endif 718 719 #ifdef PMAP_HAS_PAGE_ARRAY 720 pmap_page_array_startup(size / PAGE_SIZE); 721 biggestone = vm_phys_avail_largest(); 722 end = new_end = phys_avail[biggestone + 1]; 723 #else 724 #ifdef VM_PHYSSEG_DENSE 725 /* 726 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 727 * the overhead of a page structure per page only if vm_page_array is 728 * allocated from the last physical memory chunk. Otherwise, we must 729 * allocate page structures representing the physical memory 730 * underlying vm_page_array, even though they will not be used. 731 */ 732 if (new_end != high_avail) 733 page_range = size / PAGE_SIZE; 734 else 735 #endif 736 { 737 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 738 739 /* 740 * If the partial bytes remaining are large enough for 741 * a page (PAGE_SIZE) without a corresponding 742 * 'struct vm_page', then new_end will contain an 743 * extra page after subtracting the length of the VM 744 * page array. Compensate by subtracting an extra 745 * page from new_end. 746 */ 747 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 748 if (new_end == high_avail) 749 high_avail -= PAGE_SIZE; 750 new_end -= PAGE_SIZE; 751 } 752 } 753 end = new_end; 754 new_end = vm_page_array_alloc(&vaddr, end, page_range); 755 #endif 756 757 #if VM_NRESERVLEVEL > 0 758 /* 759 * Allocate physical memory for the reservation management system's 760 * data structures, and map it. 761 */ 762 new_end = vm_reserv_startup(&vaddr, new_end); 763 #endif 764 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 765 defined(__riscv) || defined(__powerpc64__) 766 /* 767 * Include vm_page_array and vm_reserv_array in a crash dump. 768 */ 769 for (pa = new_end; pa < end; pa += PAGE_SIZE) 770 dump_add_page(pa); 771 #endif 772 phys_avail[biggestone + 1] = new_end; 773 774 /* 775 * Add physical memory segments corresponding to the available 776 * physical pages. 777 */ 778 for (i = 0; phys_avail[i + 1] != 0; i += 2) 779 if (vm_phys_avail_size(i) != 0) 780 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 781 782 /* 783 * Initialize the physical memory allocator. 784 */ 785 vm_phys_init(); 786 787 /* 788 * Initialize the page structures and add every available page to the 789 * physical memory allocator's free lists. 790 */ 791 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 792 for (ii = 0; ii < vm_page_array_size; ii++) { 793 m = &vm_page_array[ii]; 794 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 795 m->flags = PG_FICTITIOUS; 796 } 797 #endif 798 vm_cnt.v_page_count = 0; 799 for (segind = 0; segind < vm_phys_nsegs; segind++) { 800 seg = &vm_phys_segs[segind]; 801 for (m = seg->first_page, pa = seg->start; pa < seg->end; 802 m++, pa += PAGE_SIZE) 803 vm_page_init_page(m, pa, segind); 804 805 /* 806 * Add the segment to the free lists only if it is covered by 807 * one of the ranges in phys_avail. Because we've added the 808 * ranges to the vm_phys_segs array, we can assume that each 809 * segment is either entirely contained in one of the ranges, 810 * or doesn't overlap any of them. 811 */ 812 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 813 struct vm_domain *vmd; 814 815 if (seg->start < phys_avail[i] || 816 seg->end > phys_avail[i + 1]) 817 continue; 818 819 m = seg->first_page; 820 pagecount = (u_long)atop(seg->end - seg->start); 821 822 vmd = VM_DOMAIN(seg->domain); 823 vm_domain_free_lock(vmd); 824 vm_phys_enqueue_contig(m, pagecount); 825 vm_domain_free_unlock(vmd); 826 vm_domain_freecnt_inc(vmd, pagecount); 827 vm_cnt.v_page_count += (u_int)pagecount; 828 829 vmd = VM_DOMAIN(seg->domain); 830 vmd->vmd_page_count += (u_int)pagecount; 831 vmd->vmd_segs |= 1UL << m->segind; 832 break; 833 } 834 } 835 836 /* 837 * Remove blacklisted pages from the physical memory allocator. 838 */ 839 TAILQ_INIT(&blacklist_head); 840 vm_page_blacklist_load(&list, &listend); 841 vm_page_blacklist_check(list, listend); 842 843 list = kern_getenv("vm.blacklist"); 844 vm_page_blacklist_check(list, NULL); 845 846 freeenv(list); 847 #if VM_NRESERVLEVEL > 0 848 /* 849 * Initialize the reservation management system. 850 */ 851 vm_reserv_init(); 852 #endif 853 854 return (vaddr); 855 } 856 857 void 858 vm_page_reference(vm_page_t m) 859 { 860 861 vm_page_aflag_set(m, PGA_REFERENCED); 862 } 863 864 /* 865 * vm_page_busy_acquire: 866 * 867 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 868 * and drop the object lock if necessary. 869 */ 870 int 871 vm_page_busy_acquire(vm_page_t m, int allocflags) 872 { 873 vm_object_t obj; 874 bool locked; 875 876 /* 877 * The page-specific object must be cached because page 878 * identity can change during the sleep, causing the 879 * re-lock of a different object. 880 * It is assumed that a reference to the object is already 881 * held by the callers. 882 */ 883 obj = m->object; 884 for (;;) { 885 if ((allocflags & VM_ALLOC_SBUSY) == 0) { 886 if (vm_page_tryxbusy(m)) 887 return (TRUE); 888 } else { 889 if (vm_page_trysbusy(m)) 890 return (TRUE); 891 } 892 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 893 return (FALSE); 894 if (obj != NULL) 895 locked = VM_OBJECT_WOWNED(obj); 896 else 897 locked = FALSE; 898 MPASS(locked || vm_page_wired(m)); 899 _vm_page_busy_sleep(obj, m, "vmpba", 900 (allocflags & VM_ALLOC_SBUSY) != 0, locked); 901 if (locked) 902 VM_OBJECT_WLOCK(obj); 903 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 904 return (FALSE); 905 KASSERT(m->object == obj || m->object == NULL, 906 ("vm_page_busy_acquire: page %p does not belong to %p", 907 m, obj)); 908 } 909 } 910 911 /* 912 * vm_page_busy_downgrade: 913 * 914 * Downgrade an exclusive busy page into a single shared busy page. 915 */ 916 void 917 vm_page_busy_downgrade(vm_page_t m) 918 { 919 u_int x; 920 921 vm_page_assert_xbusied(m); 922 923 x = m->busy_lock; 924 for (;;) { 925 if (atomic_fcmpset_rel_int(&m->busy_lock, 926 &x, VPB_SHARERS_WORD(1))) 927 break; 928 } 929 if ((x & VPB_BIT_WAITERS) != 0) 930 wakeup(m); 931 } 932 933 /* 934 * 935 * vm_page_busy_tryupgrade: 936 * 937 * Attempt to upgrade a single shared busy into an exclusive busy. 938 */ 939 int 940 vm_page_busy_tryupgrade(vm_page_t m) 941 { 942 u_int ce, x; 943 944 vm_page_assert_sbusied(m); 945 946 x = m->busy_lock; 947 ce = VPB_CURTHREAD_EXCLUSIVE; 948 for (;;) { 949 if (VPB_SHARERS(x) > 1) 950 return (0); 951 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 952 ("vm_page_busy_tryupgrade: invalid lock state")); 953 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 954 ce | (x & VPB_BIT_WAITERS))) 955 continue; 956 return (1); 957 } 958 } 959 960 /* 961 * vm_page_sbusied: 962 * 963 * Return a positive value if the page is shared busied, 0 otherwise. 964 */ 965 int 966 vm_page_sbusied(vm_page_t m) 967 { 968 u_int x; 969 970 x = m->busy_lock; 971 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 972 } 973 974 /* 975 * vm_page_sunbusy: 976 * 977 * Shared unbusy a page. 978 */ 979 void 980 vm_page_sunbusy(vm_page_t m) 981 { 982 u_int x; 983 984 vm_page_assert_sbusied(m); 985 986 x = m->busy_lock; 987 for (;;) { 988 if (VPB_SHARERS(x) > 1) { 989 if (atomic_fcmpset_int(&m->busy_lock, &x, 990 x - VPB_ONE_SHARER)) 991 break; 992 continue; 993 } 994 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 995 ("vm_page_sunbusy: invalid lock state")); 996 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 997 continue; 998 if ((x & VPB_BIT_WAITERS) == 0) 999 break; 1000 wakeup(m); 1001 break; 1002 } 1003 } 1004 1005 /* 1006 * vm_page_busy_sleep: 1007 * 1008 * Sleep if the page is busy, using the page pointer as wchan. 1009 * This is used to implement the hard-path of busying mechanism. 1010 * 1011 * If nonshared is true, sleep only if the page is xbusy. 1012 * 1013 * The object lock must be held on entry and will be released on exit. 1014 */ 1015 void 1016 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 1017 { 1018 vm_object_t obj; 1019 1020 obj = m->object; 1021 VM_OBJECT_ASSERT_LOCKED(obj); 1022 vm_page_lock_assert(m, MA_NOTOWNED); 1023 1024 _vm_page_busy_sleep(obj, m, wmesg, nonshared, true); 1025 } 1026 1027 static void 1028 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, 1029 bool nonshared, bool locked) 1030 { 1031 u_int x; 1032 1033 /* 1034 * If the object is busy we must wait for that to drain to zero 1035 * before trying the page again. 1036 */ 1037 if (obj != NULL && vm_object_busied(obj)) { 1038 if (locked) 1039 VM_OBJECT_DROP(obj); 1040 vm_object_busy_wait(obj, wmesg); 1041 return; 1042 } 1043 sleepq_lock(m); 1044 x = m->busy_lock; 1045 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 1046 ((x & VPB_BIT_WAITERS) == 0 && 1047 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 1048 if (locked) 1049 VM_OBJECT_DROP(obj); 1050 sleepq_release(m); 1051 return; 1052 } 1053 if (locked) 1054 VM_OBJECT_DROP(obj); 1055 DROP_GIANT(); 1056 sleepq_add(m, NULL, wmesg, 0, 0); 1057 sleepq_wait(m, PVM); 1058 PICKUP_GIANT(); 1059 } 1060 1061 /* 1062 * vm_page_trysbusy: 1063 * 1064 * Try to shared busy a page. 1065 * If the operation succeeds 1 is returned otherwise 0. 1066 * The operation never sleeps. 1067 */ 1068 int 1069 vm_page_trysbusy(vm_page_t m) 1070 { 1071 vm_object_t obj; 1072 u_int x; 1073 1074 obj = m->object; 1075 x = m->busy_lock; 1076 for (;;) { 1077 if ((x & VPB_BIT_SHARED) == 0) 1078 return (0); 1079 /* 1080 * Reduce the window for transient busies that will trigger 1081 * false negatives in vm_page_ps_test(). 1082 */ 1083 if (obj != NULL && vm_object_busied(obj)) 1084 return (0); 1085 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1086 x + VPB_ONE_SHARER)) 1087 break; 1088 } 1089 1090 /* Refetch the object now that we're guaranteed that it is stable. */ 1091 obj = m->object; 1092 if (obj != NULL && vm_object_busied(obj)) { 1093 vm_page_sunbusy(m); 1094 return (0); 1095 } 1096 return (1); 1097 } 1098 1099 /* 1100 * vm_page_tryxbusy: 1101 * 1102 * Try to exclusive busy a page. 1103 * If the operation succeeds 1 is returned otherwise 0. 1104 * The operation never sleeps. 1105 */ 1106 int 1107 vm_page_tryxbusy(vm_page_t m) 1108 { 1109 vm_object_t obj; 1110 1111 if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, 1112 VPB_CURTHREAD_EXCLUSIVE) == 0) 1113 return (0); 1114 1115 obj = m->object; 1116 if (obj != NULL && vm_object_busied(obj)) { 1117 vm_page_xunbusy(m); 1118 return (0); 1119 } 1120 return (1); 1121 } 1122 1123 static void 1124 vm_page_xunbusy_hard_tail(vm_page_t m) 1125 { 1126 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1127 /* Wake the waiter. */ 1128 wakeup(m); 1129 } 1130 1131 /* 1132 * vm_page_xunbusy_hard: 1133 * 1134 * Called when unbusy has failed because there is a waiter. 1135 */ 1136 void 1137 vm_page_xunbusy_hard(vm_page_t m) 1138 { 1139 vm_page_assert_xbusied(m); 1140 vm_page_xunbusy_hard_tail(m); 1141 } 1142 1143 void 1144 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1145 { 1146 vm_page_assert_xbusied_unchecked(m); 1147 vm_page_xunbusy_hard_tail(m); 1148 } 1149 1150 /* 1151 * Avoid releasing and reacquiring the same page lock. 1152 */ 1153 void 1154 vm_page_change_lock(vm_page_t m, struct mtx **mtx) 1155 { 1156 struct mtx *mtx1; 1157 1158 mtx1 = vm_page_lockptr(m); 1159 if (*mtx == mtx1) 1160 return; 1161 if (*mtx != NULL) 1162 mtx_unlock(*mtx); 1163 *mtx = mtx1; 1164 mtx_lock(mtx1); 1165 } 1166 1167 /* 1168 * vm_page_unhold_pages: 1169 * 1170 * Unhold each of the pages that is referenced by the given array. 1171 */ 1172 void 1173 vm_page_unhold_pages(vm_page_t *ma, int count) 1174 { 1175 1176 for (; count != 0; count--) { 1177 vm_page_unwire(*ma, PQ_ACTIVE); 1178 ma++; 1179 } 1180 } 1181 1182 vm_page_t 1183 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1184 { 1185 vm_page_t m; 1186 1187 #ifdef VM_PHYSSEG_SPARSE 1188 m = vm_phys_paddr_to_vm_page(pa); 1189 if (m == NULL) 1190 m = vm_phys_fictitious_to_vm_page(pa); 1191 return (m); 1192 #elif defined(VM_PHYSSEG_DENSE) 1193 long pi; 1194 1195 pi = atop(pa); 1196 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1197 m = &vm_page_array[pi - first_page]; 1198 return (m); 1199 } 1200 return (vm_phys_fictitious_to_vm_page(pa)); 1201 #else 1202 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1203 #endif 1204 } 1205 1206 /* 1207 * vm_page_getfake: 1208 * 1209 * Create a fictitious page with the specified physical address and 1210 * memory attribute. The memory attribute is the only the machine- 1211 * dependent aspect of a fictitious page that must be initialized. 1212 */ 1213 vm_page_t 1214 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1215 { 1216 vm_page_t m; 1217 1218 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1219 vm_page_initfake(m, paddr, memattr); 1220 return (m); 1221 } 1222 1223 void 1224 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1225 { 1226 1227 if ((m->flags & PG_FICTITIOUS) != 0) { 1228 /* 1229 * The page's memattr might have changed since the 1230 * previous initialization. Update the pmap to the 1231 * new memattr. 1232 */ 1233 goto memattr; 1234 } 1235 m->phys_addr = paddr; 1236 m->queue = PQ_NONE; 1237 /* Fictitious pages don't use "segind". */ 1238 m->flags = PG_FICTITIOUS; 1239 /* Fictitious pages don't use "order" or "pool". */ 1240 m->oflags = VPO_UNMANAGED; 1241 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1242 /* Fictitious pages are unevictable. */ 1243 m->ref_count = 1; 1244 pmap_page_init(m); 1245 memattr: 1246 pmap_page_set_memattr(m, memattr); 1247 } 1248 1249 /* 1250 * vm_page_putfake: 1251 * 1252 * Release a fictitious page. 1253 */ 1254 void 1255 vm_page_putfake(vm_page_t m) 1256 { 1257 1258 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1259 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1260 ("vm_page_putfake: bad page %p", m)); 1261 if (vm_page_xbusied(m)) 1262 vm_page_xunbusy(m); 1263 uma_zfree(fakepg_zone, m); 1264 } 1265 1266 /* 1267 * vm_page_updatefake: 1268 * 1269 * Update the given fictitious page to the specified physical address and 1270 * memory attribute. 1271 */ 1272 void 1273 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1274 { 1275 1276 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1277 ("vm_page_updatefake: bad page %p", m)); 1278 m->phys_addr = paddr; 1279 pmap_page_set_memattr(m, memattr); 1280 } 1281 1282 /* 1283 * vm_page_free: 1284 * 1285 * Free a page. 1286 */ 1287 void 1288 vm_page_free(vm_page_t m) 1289 { 1290 1291 m->flags &= ~PG_ZERO; 1292 vm_page_free_toq(m); 1293 } 1294 1295 /* 1296 * vm_page_free_zero: 1297 * 1298 * Free a page to the zerod-pages queue 1299 */ 1300 void 1301 vm_page_free_zero(vm_page_t m) 1302 { 1303 1304 m->flags |= PG_ZERO; 1305 vm_page_free_toq(m); 1306 } 1307 1308 /* 1309 * Unbusy and handle the page queueing for a page from a getpages request that 1310 * was optionally read ahead or behind. 1311 */ 1312 void 1313 vm_page_readahead_finish(vm_page_t m) 1314 { 1315 1316 /* We shouldn't put invalid pages on queues. */ 1317 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1318 1319 /* 1320 * Since the page is not the actually needed one, whether it should 1321 * be activated or deactivated is not obvious. Empirical results 1322 * have shown that deactivating the page is usually the best choice, 1323 * unless the page is wanted by another thread. 1324 */ 1325 vm_page_lock(m); 1326 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1327 vm_page_activate(m); 1328 else 1329 vm_page_deactivate(m); 1330 vm_page_unlock(m); 1331 vm_page_xunbusy_unchecked(m); 1332 } 1333 1334 /* 1335 * vm_page_sleep_if_busy: 1336 * 1337 * Sleep and release the object lock if the page is busied. 1338 * Returns TRUE if the thread slept. 1339 * 1340 * The given page must be unlocked and object containing it must 1341 * be locked. 1342 */ 1343 int 1344 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1345 { 1346 vm_object_t obj; 1347 1348 vm_page_lock_assert(m, MA_NOTOWNED); 1349 VM_OBJECT_ASSERT_WLOCKED(m->object); 1350 1351 /* 1352 * The page-specific object must be cached because page 1353 * identity can change during the sleep, causing the 1354 * re-lock of a different object. 1355 * It is assumed that a reference to the object is already 1356 * held by the callers. 1357 */ 1358 obj = m->object; 1359 if (vm_page_busied(m) || (obj != NULL && obj->busy)) { 1360 vm_page_busy_sleep(m, msg, false); 1361 VM_OBJECT_WLOCK(obj); 1362 return (TRUE); 1363 } 1364 return (FALSE); 1365 } 1366 1367 /* 1368 * vm_page_sleep_if_xbusy: 1369 * 1370 * Sleep and release the object lock if the page is xbusied. 1371 * Returns TRUE if the thread slept. 1372 * 1373 * The given page must be unlocked and object containing it must 1374 * be locked. 1375 */ 1376 int 1377 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg) 1378 { 1379 vm_object_t obj; 1380 1381 vm_page_lock_assert(m, MA_NOTOWNED); 1382 VM_OBJECT_ASSERT_WLOCKED(m->object); 1383 1384 /* 1385 * The page-specific object must be cached because page 1386 * identity can change during the sleep, causing the 1387 * re-lock of a different object. 1388 * It is assumed that a reference to the object is already 1389 * held by the callers. 1390 */ 1391 obj = m->object; 1392 if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) { 1393 vm_page_busy_sleep(m, msg, true); 1394 VM_OBJECT_WLOCK(obj); 1395 return (TRUE); 1396 } 1397 return (FALSE); 1398 } 1399 1400 /* 1401 * vm_page_dirty_KBI: [ internal use only ] 1402 * 1403 * Set all bits in the page's dirty field. 1404 * 1405 * The object containing the specified page must be locked if the 1406 * call is made from the machine-independent layer. 1407 * 1408 * See vm_page_clear_dirty_mask(). 1409 * 1410 * This function should only be called by vm_page_dirty(). 1411 */ 1412 void 1413 vm_page_dirty_KBI(vm_page_t m) 1414 { 1415 1416 /* Refer to this operation by its public name. */ 1417 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1418 m->dirty = VM_PAGE_BITS_ALL; 1419 } 1420 1421 /* 1422 * vm_page_insert: [ internal use only ] 1423 * 1424 * Inserts the given mem entry into the object and object list. 1425 * 1426 * The object must be locked. 1427 */ 1428 int 1429 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1430 { 1431 vm_page_t mpred; 1432 1433 VM_OBJECT_ASSERT_WLOCKED(object); 1434 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1435 return (vm_page_insert_after(m, object, pindex, mpred)); 1436 } 1437 1438 /* 1439 * vm_page_insert_after: 1440 * 1441 * Inserts the page "m" into the specified object at offset "pindex". 1442 * 1443 * The page "mpred" must immediately precede the offset "pindex" within 1444 * the specified object. 1445 * 1446 * The object must be locked. 1447 */ 1448 static int 1449 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1450 vm_page_t mpred) 1451 { 1452 vm_page_t msucc; 1453 1454 VM_OBJECT_ASSERT_WLOCKED(object); 1455 KASSERT(m->object == NULL, 1456 ("vm_page_insert_after: page already inserted")); 1457 if (mpred != NULL) { 1458 KASSERT(mpred->object == object, 1459 ("vm_page_insert_after: object doesn't contain mpred")); 1460 KASSERT(mpred->pindex < pindex, 1461 ("vm_page_insert_after: mpred doesn't precede pindex")); 1462 msucc = TAILQ_NEXT(mpred, listq); 1463 } else 1464 msucc = TAILQ_FIRST(&object->memq); 1465 if (msucc != NULL) 1466 KASSERT(msucc->pindex > pindex, 1467 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1468 1469 /* 1470 * Record the object/offset pair in this page. 1471 */ 1472 m->object = object; 1473 m->pindex = pindex; 1474 m->ref_count |= VPRC_OBJREF; 1475 1476 /* 1477 * Now link into the object's ordered list of backed pages. 1478 */ 1479 if (vm_radix_insert(&object->rtree, m)) { 1480 m->object = NULL; 1481 m->pindex = 0; 1482 m->ref_count &= ~VPRC_OBJREF; 1483 return (1); 1484 } 1485 vm_page_insert_radixdone(m, object, mpred); 1486 return (0); 1487 } 1488 1489 /* 1490 * vm_page_insert_radixdone: 1491 * 1492 * Complete page "m" insertion into the specified object after the 1493 * radix trie hooking. 1494 * 1495 * The page "mpred" must precede the offset "m->pindex" within the 1496 * specified object. 1497 * 1498 * The object must be locked. 1499 */ 1500 static void 1501 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1502 { 1503 1504 VM_OBJECT_ASSERT_WLOCKED(object); 1505 KASSERT(object != NULL && m->object == object, 1506 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1507 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1508 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1509 if (mpred != NULL) { 1510 KASSERT(mpred->object == object, 1511 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1512 KASSERT(mpred->pindex < m->pindex, 1513 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1514 } 1515 1516 if (mpred != NULL) 1517 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1518 else 1519 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1520 1521 /* 1522 * Show that the object has one more resident page. 1523 */ 1524 object->resident_page_count++; 1525 1526 /* 1527 * Hold the vnode until the last page is released. 1528 */ 1529 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1530 vhold(object->handle); 1531 1532 /* 1533 * Since we are inserting a new and possibly dirty page, 1534 * update the object's generation count. 1535 */ 1536 if (pmap_page_is_write_mapped(m)) 1537 vm_object_set_writeable_dirty(object); 1538 } 1539 1540 /* 1541 * Do the work to remove a page from its object. The caller is responsible for 1542 * updating the page's fields to reflect this removal. 1543 */ 1544 static void 1545 vm_page_object_remove(vm_page_t m) 1546 { 1547 vm_object_t object; 1548 vm_page_t mrem; 1549 1550 object = m->object; 1551 VM_OBJECT_ASSERT_WLOCKED(object); 1552 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1553 ("page %p is missing its object ref", m)); 1554 1555 mrem = vm_radix_remove(&object->rtree, m->pindex); 1556 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1557 1558 /* 1559 * Now remove from the object's list of backed pages. 1560 */ 1561 TAILQ_REMOVE(&object->memq, m, listq); 1562 1563 /* 1564 * And show that the object has one fewer resident page. 1565 */ 1566 object->resident_page_count--; 1567 1568 /* 1569 * The vnode may now be recycled. 1570 */ 1571 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1572 vdrop(object->handle); 1573 } 1574 1575 /* 1576 * vm_page_remove: 1577 * 1578 * Removes the specified page from its containing object, but does not 1579 * invalidate any backing storage. Returns true if the object's reference 1580 * was the last reference to the page, and false otherwise. 1581 * 1582 * The object must be locked. 1583 */ 1584 bool 1585 vm_page_remove(vm_page_t m) 1586 { 1587 1588 vm_page_object_remove(m); 1589 m->object = NULL; 1590 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1591 } 1592 1593 /* 1594 * vm_page_lookup: 1595 * 1596 * Returns the page associated with the object/offset 1597 * pair specified; if none is found, NULL is returned. 1598 * 1599 * The object must be locked. 1600 */ 1601 vm_page_t 1602 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1603 { 1604 1605 VM_OBJECT_ASSERT_LOCKED(object); 1606 return (vm_radix_lookup(&object->rtree, pindex)); 1607 } 1608 1609 /* 1610 * vm_page_find_least: 1611 * 1612 * Returns the page associated with the object with least pindex 1613 * greater than or equal to the parameter pindex, or NULL. 1614 * 1615 * The object must be locked. 1616 */ 1617 vm_page_t 1618 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1619 { 1620 vm_page_t m; 1621 1622 VM_OBJECT_ASSERT_LOCKED(object); 1623 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1624 m = vm_radix_lookup_ge(&object->rtree, pindex); 1625 return (m); 1626 } 1627 1628 /* 1629 * Returns the given page's successor (by pindex) within the object if it is 1630 * resident; if none is found, NULL is returned. 1631 * 1632 * The object must be locked. 1633 */ 1634 vm_page_t 1635 vm_page_next(vm_page_t m) 1636 { 1637 vm_page_t next; 1638 1639 VM_OBJECT_ASSERT_LOCKED(m->object); 1640 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1641 MPASS(next->object == m->object); 1642 if (next->pindex != m->pindex + 1) 1643 next = NULL; 1644 } 1645 return (next); 1646 } 1647 1648 /* 1649 * Returns the given page's predecessor (by pindex) within the object if it is 1650 * resident; if none is found, NULL is returned. 1651 * 1652 * The object must be locked. 1653 */ 1654 vm_page_t 1655 vm_page_prev(vm_page_t m) 1656 { 1657 vm_page_t prev; 1658 1659 VM_OBJECT_ASSERT_LOCKED(m->object); 1660 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1661 MPASS(prev->object == m->object); 1662 if (prev->pindex != m->pindex - 1) 1663 prev = NULL; 1664 } 1665 return (prev); 1666 } 1667 1668 /* 1669 * Uses the page mnew as a replacement for an existing page at index 1670 * pindex which must be already present in the object. 1671 */ 1672 vm_page_t 1673 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1674 { 1675 vm_page_t mold; 1676 1677 VM_OBJECT_ASSERT_WLOCKED(object); 1678 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1679 ("vm_page_replace: page %p already in object", mnew)); 1680 1681 /* 1682 * This function mostly follows vm_page_insert() and 1683 * vm_page_remove() without the radix, object count and vnode 1684 * dance. Double check such functions for more comments. 1685 */ 1686 1687 mnew->object = object; 1688 mnew->pindex = pindex; 1689 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1690 mold = vm_radix_replace(&object->rtree, mnew); 1691 KASSERT(mold->queue == PQ_NONE, 1692 ("vm_page_replace: old page %p is on a paging queue", mold)); 1693 1694 /* Keep the resident page list in sorted order. */ 1695 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1696 TAILQ_REMOVE(&object->memq, mold, listq); 1697 1698 mold->object = NULL; 1699 atomic_clear_int(&mold->ref_count, VPRC_OBJREF); 1700 vm_page_xunbusy(mold); 1701 1702 /* 1703 * The object's resident_page_count does not change because we have 1704 * swapped one page for another, but the generation count should 1705 * change if the page is dirty. 1706 */ 1707 if (pmap_page_is_write_mapped(mnew)) 1708 vm_object_set_writeable_dirty(object); 1709 return (mold); 1710 } 1711 1712 /* 1713 * vm_page_rename: 1714 * 1715 * Move the given memory entry from its 1716 * current object to the specified target object/offset. 1717 * 1718 * Note: swap associated with the page must be invalidated by the move. We 1719 * have to do this for several reasons: (1) we aren't freeing the 1720 * page, (2) we are dirtying the page, (3) the VM system is probably 1721 * moving the page from object A to B, and will then later move 1722 * the backing store from A to B and we can't have a conflict. 1723 * 1724 * Note: we *always* dirty the page. It is necessary both for the 1725 * fact that we moved it, and because we may be invalidating 1726 * swap. 1727 * 1728 * The objects must be locked. 1729 */ 1730 int 1731 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1732 { 1733 vm_page_t mpred; 1734 vm_pindex_t opidx; 1735 1736 VM_OBJECT_ASSERT_WLOCKED(new_object); 1737 1738 KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); 1739 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1740 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1741 ("vm_page_rename: pindex already renamed")); 1742 1743 /* 1744 * Create a custom version of vm_page_insert() which does not depend 1745 * by m_prev and can cheat on the implementation aspects of the 1746 * function. 1747 */ 1748 opidx = m->pindex; 1749 m->pindex = new_pindex; 1750 if (vm_radix_insert(&new_object->rtree, m)) { 1751 m->pindex = opidx; 1752 return (1); 1753 } 1754 1755 /* 1756 * The operation cannot fail anymore. The removal must happen before 1757 * the listq iterator is tainted. 1758 */ 1759 m->pindex = opidx; 1760 vm_page_object_remove(m); 1761 1762 /* Return back to the new pindex to complete vm_page_insert(). */ 1763 m->pindex = new_pindex; 1764 m->object = new_object; 1765 1766 vm_page_insert_radixdone(m, new_object, mpred); 1767 vm_page_dirty(m); 1768 return (0); 1769 } 1770 1771 /* 1772 * vm_page_alloc: 1773 * 1774 * Allocate and return a page that is associated with the specified 1775 * object and offset pair. By default, this page is exclusive busied. 1776 * 1777 * The caller must always specify an allocation class. 1778 * 1779 * allocation classes: 1780 * VM_ALLOC_NORMAL normal process request 1781 * VM_ALLOC_SYSTEM system *really* needs a page 1782 * VM_ALLOC_INTERRUPT interrupt time request 1783 * 1784 * optional allocation flags: 1785 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1786 * intends to allocate 1787 * VM_ALLOC_NOBUSY do not exclusive busy the page 1788 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1789 * VM_ALLOC_NOOBJ page is not associated with an object and 1790 * should not be exclusive busy 1791 * VM_ALLOC_SBUSY shared busy the allocated page 1792 * VM_ALLOC_WIRED wire the allocated page 1793 * VM_ALLOC_ZERO prefer a zeroed page 1794 */ 1795 vm_page_t 1796 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1797 { 1798 1799 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1800 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1801 } 1802 1803 vm_page_t 1804 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1805 int req) 1806 { 1807 1808 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1809 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1810 NULL)); 1811 } 1812 1813 /* 1814 * Allocate a page in the specified object with the given page index. To 1815 * optimize insertion of the page into the object, the caller must also specifiy 1816 * the resident page in the object with largest index smaller than the given 1817 * page index, or NULL if no such page exists. 1818 */ 1819 vm_page_t 1820 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1821 int req, vm_page_t mpred) 1822 { 1823 struct vm_domainset_iter di; 1824 vm_page_t m; 1825 int domain; 1826 1827 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1828 do { 1829 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1830 mpred); 1831 if (m != NULL) 1832 break; 1833 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1834 1835 return (m); 1836 } 1837 1838 /* 1839 * Returns true if the number of free pages exceeds the minimum 1840 * for the request class and false otherwise. 1841 */ 1842 static int 1843 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 1844 { 1845 u_int limit, old, new; 1846 1847 if (req_class == VM_ALLOC_INTERRUPT) 1848 limit = 0; 1849 else if (req_class == VM_ALLOC_SYSTEM) 1850 limit = vmd->vmd_interrupt_free_min; 1851 else 1852 limit = vmd->vmd_free_reserved; 1853 1854 /* 1855 * Attempt to reserve the pages. Fail if we're below the limit. 1856 */ 1857 limit += npages; 1858 old = vmd->vmd_free_count; 1859 do { 1860 if (old < limit) 1861 return (0); 1862 new = old - npages; 1863 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 1864 1865 /* Wake the page daemon if we've crossed the threshold. */ 1866 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 1867 pagedaemon_wakeup(vmd->vmd_domain); 1868 1869 /* Only update bitsets on transitions. */ 1870 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 1871 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 1872 vm_domain_set(vmd); 1873 1874 return (1); 1875 } 1876 1877 int 1878 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 1879 { 1880 int req_class; 1881 1882 /* 1883 * The page daemon is allowed to dig deeper into the free page list. 1884 */ 1885 req_class = req & VM_ALLOC_CLASS_MASK; 1886 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1887 req_class = VM_ALLOC_SYSTEM; 1888 return (_vm_domain_allocate(vmd, req_class, npages)); 1889 } 1890 1891 vm_page_t 1892 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 1893 int req, vm_page_t mpred) 1894 { 1895 struct vm_domain *vmd; 1896 vm_page_t m; 1897 int flags, pool; 1898 1899 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1900 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1901 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1902 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1903 ("inconsistent object(%p)/req(%x)", object, req)); 1904 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1905 ("Can't sleep and retry object insertion.")); 1906 KASSERT(mpred == NULL || mpred->pindex < pindex, 1907 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1908 (uintmax_t)pindex)); 1909 if (object != NULL) 1910 VM_OBJECT_ASSERT_WLOCKED(object); 1911 1912 flags = 0; 1913 m = NULL; 1914 pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; 1915 again: 1916 #if VM_NRESERVLEVEL > 0 1917 /* 1918 * Can we allocate the page from a reservation? 1919 */ 1920 if (vm_object_reserv(object) && 1921 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 1922 NULL) { 1923 domain = vm_phys_domain(m); 1924 vmd = VM_DOMAIN(domain); 1925 goto found; 1926 } 1927 #endif 1928 vmd = VM_DOMAIN(domain); 1929 if (vmd->vmd_pgcache[pool].zone != NULL) { 1930 m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT); 1931 if (m != NULL) { 1932 flags |= PG_PCPU_CACHE; 1933 goto found; 1934 } 1935 } 1936 if (vm_domain_allocate(vmd, req, 1)) { 1937 /* 1938 * If not, allocate it from the free page queues. 1939 */ 1940 vm_domain_free_lock(vmd); 1941 m = vm_phys_alloc_pages(domain, pool, 0); 1942 vm_domain_free_unlock(vmd); 1943 if (m == NULL) { 1944 vm_domain_freecnt_inc(vmd, 1); 1945 #if VM_NRESERVLEVEL > 0 1946 if (vm_reserv_reclaim_inactive(domain)) 1947 goto again; 1948 #endif 1949 } 1950 } 1951 if (m == NULL) { 1952 /* 1953 * Not allocatable, give up. 1954 */ 1955 if (vm_domain_alloc_fail(vmd, object, req)) 1956 goto again; 1957 return (NULL); 1958 } 1959 1960 /* 1961 * At this point we had better have found a good page. 1962 */ 1963 found: 1964 vm_page_dequeue(m); 1965 vm_page_alloc_check(m); 1966 1967 /* 1968 * Initialize the page. Only the PG_ZERO flag is inherited. 1969 */ 1970 if ((req & VM_ALLOC_ZERO) != 0) 1971 flags |= (m->flags & PG_ZERO); 1972 if ((req & VM_ALLOC_NODUMP) != 0) 1973 flags |= PG_NODUMP; 1974 m->flags = flags; 1975 m->aflags = 0; 1976 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1977 VPO_UNMANAGED : 0; 1978 m->busy_lock = VPB_UNBUSIED; 1979 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1980 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1981 if ((req & VM_ALLOC_SBUSY) != 0) 1982 m->busy_lock = VPB_SHARERS_WORD(1); 1983 if (req & VM_ALLOC_WIRED) { 1984 /* 1985 * The page lock is not required for wiring a page until that 1986 * page is inserted into the object. 1987 */ 1988 vm_wire_add(1); 1989 m->ref_count = 1; 1990 } 1991 m->act_count = 0; 1992 1993 if (object != NULL) { 1994 if (vm_page_insert_after(m, object, pindex, mpred)) { 1995 if (req & VM_ALLOC_WIRED) { 1996 vm_wire_sub(1); 1997 m->ref_count = 0; 1998 } 1999 KASSERT(m->object == NULL, ("page %p has object", m)); 2000 m->oflags = VPO_UNMANAGED; 2001 m->busy_lock = VPB_UNBUSIED; 2002 /* Don't change PG_ZERO. */ 2003 vm_page_free_toq(m); 2004 if (req & VM_ALLOC_WAITFAIL) { 2005 VM_OBJECT_WUNLOCK(object); 2006 vm_radix_wait(); 2007 VM_OBJECT_WLOCK(object); 2008 } 2009 return (NULL); 2010 } 2011 2012 /* Ignore device objects; the pager sets "memattr" for them. */ 2013 if (object->memattr != VM_MEMATTR_DEFAULT && 2014 (object->flags & OBJ_FICTITIOUS) == 0) 2015 pmap_page_set_memattr(m, object->memattr); 2016 } else 2017 m->pindex = pindex; 2018 2019 return (m); 2020 } 2021 2022 /* 2023 * vm_page_alloc_contig: 2024 * 2025 * Allocate a contiguous set of physical pages of the given size "npages" 2026 * from the free lists. All of the physical pages must be at or above 2027 * the given physical address "low" and below the given physical address 2028 * "high". The given value "alignment" determines the alignment of the 2029 * first physical page in the set. If the given value "boundary" is 2030 * non-zero, then the set of physical pages cannot cross any physical 2031 * address boundary that is a multiple of that value. Both "alignment" 2032 * and "boundary" must be a power of two. 2033 * 2034 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2035 * then the memory attribute setting for the physical pages is configured 2036 * to the object's memory attribute setting. Otherwise, the memory 2037 * attribute setting for the physical pages is configured to "memattr", 2038 * overriding the object's memory attribute setting. However, if the 2039 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2040 * memory attribute setting for the physical pages cannot be configured 2041 * to VM_MEMATTR_DEFAULT. 2042 * 2043 * The specified object may not contain fictitious pages. 2044 * 2045 * The caller must always specify an allocation class. 2046 * 2047 * allocation classes: 2048 * VM_ALLOC_NORMAL normal process request 2049 * VM_ALLOC_SYSTEM system *really* needs a page 2050 * VM_ALLOC_INTERRUPT interrupt time request 2051 * 2052 * optional allocation flags: 2053 * VM_ALLOC_NOBUSY do not exclusive busy the page 2054 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2055 * VM_ALLOC_NOOBJ page is not associated with an object and 2056 * should not be exclusive busy 2057 * VM_ALLOC_SBUSY shared busy the allocated page 2058 * VM_ALLOC_WIRED wire the allocated page 2059 * VM_ALLOC_ZERO prefer a zeroed page 2060 */ 2061 vm_page_t 2062 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2063 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2064 vm_paddr_t boundary, vm_memattr_t memattr) 2065 { 2066 struct vm_domainset_iter di; 2067 vm_page_t m; 2068 int domain; 2069 2070 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2071 do { 2072 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2073 npages, low, high, alignment, boundary, memattr); 2074 if (m != NULL) 2075 break; 2076 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2077 2078 return (m); 2079 } 2080 2081 vm_page_t 2082 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2083 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2084 vm_paddr_t boundary, vm_memattr_t memattr) 2085 { 2086 struct vm_domain *vmd; 2087 vm_page_t m, m_ret, mpred; 2088 u_int busy_lock, flags, oflags; 2089 2090 mpred = NULL; /* XXX: pacify gcc */ 2091 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 2092 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 2093 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2094 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2095 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 2096 req)); 2097 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2098 ("Can't sleep and retry object insertion.")); 2099 if (object != NULL) { 2100 VM_OBJECT_ASSERT_WLOCKED(object); 2101 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2102 ("vm_page_alloc_contig: object %p has fictitious pages", 2103 object)); 2104 } 2105 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2106 2107 if (object != NULL) { 2108 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2109 KASSERT(mpred == NULL || mpred->pindex != pindex, 2110 ("vm_page_alloc_contig: pindex already allocated")); 2111 } 2112 2113 /* 2114 * Can we allocate the pages without the number of free pages falling 2115 * below the lower bound for the allocation class? 2116 */ 2117 m_ret = NULL; 2118 again: 2119 #if VM_NRESERVLEVEL > 0 2120 /* 2121 * Can we allocate the pages from a reservation? 2122 */ 2123 if (vm_object_reserv(object) && 2124 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2125 mpred, npages, low, high, alignment, boundary)) != NULL) { 2126 domain = vm_phys_domain(m_ret); 2127 vmd = VM_DOMAIN(domain); 2128 goto found; 2129 } 2130 #endif 2131 vmd = VM_DOMAIN(domain); 2132 if (vm_domain_allocate(vmd, req, npages)) { 2133 /* 2134 * allocate them from the free page queues. 2135 */ 2136 vm_domain_free_lock(vmd); 2137 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2138 alignment, boundary); 2139 vm_domain_free_unlock(vmd); 2140 if (m_ret == NULL) { 2141 vm_domain_freecnt_inc(vmd, npages); 2142 #if VM_NRESERVLEVEL > 0 2143 if (vm_reserv_reclaim_contig(domain, npages, low, 2144 high, alignment, boundary)) 2145 goto again; 2146 #endif 2147 } 2148 } 2149 if (m_ret == NULL) { 2150 if (vm_domain_alloc_fail(vmd, object, req)) 2151 goto again; 2152 return (NULL); 2153 } 2154 #if VM_NRESERVLEVEL > 0 2155 found: 2156 #endif 2157 for (m = m_ret; m < &m_ret[npages]; m++) { 2158 vm_page_dequeue(m); 2159 vm_page_alloc_check(m); 2160 } 2161 2162 /* 2163 * Initialize the pages. Only the PG_ZERO flag is inherited. 2164 */ 2165 flags = 0; 2166 if ((req & VM_ALLOC_ZERO) != 0) 2167 flags = PG_ZERO; 2168 if ((req & VM_ALLOC_NODUMP) != 0) 2169 flags |= PG_NODUMP; 2170 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2171 VPO_UNMANAGED : 0; 2172 busy_lock = VPB_UNBUSIED; 2173 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2174 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2175 if ((req & VM_ALLOC_SBUSY) != 0) 2176 busy_lock = VPB_SHARERS_WORD(1); 2177 if ((req & VM_ALLOC_WIRED) != 0) 2178 vm_wire_add(npages); 2179 if (object != NULL) { 2180 if (object->memattr != VM_MEMATTR_DEFAULT && 2181 memattr == VM_MEMATTR_DEFAULT) 2182 memattr = object->memattr; 2183 } 2184 for (m = m_ret; m < &m_ret[npages]; m++) { 2185 m->aflags = 0; 2186 m->flags = (m->flags | PG_NODUMP) & flags; 2187 m->busy_lock = busy_lock; 2188 if ((req & VM_ALLOC_WIRED) != 0) 2189 m->ref_count = 1; 2190 m->act_count = 0; 2191 m->oflags = oflags; 2192 if (object != NULL) { 2193 if (vm_page_insert_after(m, object, pindex, mpred)) { 2194 if ((req & VM_ALLOC_WIRED) != 0) 2195 vm_wire_sub(npages); 2196 KASSERT(m->object == NULL, 2197 ("page %p has object", m)); 2198 mpred = m; 2199 for (m = m_ret; m < &m_ret[npages]; m++) { 2200 if (m <= mpred && 2201 (req & VM_ALLOC_WIRED) != 0) 2202 m->ref_count = 0; 2203 m->oflags = VPO_UNMANAGED; 2204 m->busy_lock = VPB_UNBUSIED; 2205 /* Don't change PG_ZERO. */ 2206 vm_page_free_toq(m); 2207 } 2208 if (req & VM_ALLOC_WAITFAIL) { 2209 VM_OBJECT_WUNLOCK(object); 2210 vm_radix_wait(); 2211 VM_OBJECT_WLOCK(object); 2212 } 2213 return (NULL); 2214 } 2215 mpred = m; 2216 } else 2217 m->pindex = pindex; 2218 if (memattr != VM_MEMATTR_DEFAULT) 2219 pmap_page_set_memattr(m, memattr); 2220 pindex++; 2221 } 2222 return (m_ret); 2223 } 2224 2225 /* 2226 * Check a page that has been freshly dequeued from a freelist. 2227 */ 2228 static void 2229 vm_page_alloc_check(vm_page_t m) 2230 { 2231 2232 KASSERT(m->object == NULL, ("page %p has object", m)); 2233 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 2234 ("page %p has unexpected queue %d, flags %#x", 2235 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK))); 2236 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2237 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 2238 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2239 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2240 ("page %p has unexpected memattr %d", 2241 m, pmap_page_get_memattr(m))); 2242 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2243 } 2244 2245 /* 2246 * vm_page_alloc_freelist: 2247 * 2248 * Allocate a physical page from the specified free page list. 2249 * 2250 * The caller must always specify an allocation class. 2251 * 2252 * allocation classes: 2253 * VM_ALLOC_NORMAL normal process request 2254 * VM_ALLOC_SYSTEM system *really* needs a page 2255 * VM_ALLOC_INTERRUPT interrupt time request 2256 * 2257 * optional allocation flags: 2258 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2259 * intends to allocate 2260 * VM_ALLOC_WIRED wire the allocated page 2261 * VM_ALLOC_ZERO prefer a zeroed page 2262 */ 2263 vm_page_t 2264 vm_page_alloc_freelist(int freelist, int req) 2265 { 2266 struct vm_domainset_iter di; 2267 vm_page_t m; 2268 int domain; 2269 2270 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2271 do { 2272 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2273 if (m != NULL) 2274 break; 2275 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2276 2277 return (m); 2278 } 2279 2280 vm_page_t 2281 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2282 { 2283 struct vm_domain *vmd; 2284 vm_page_t m; 2285 u_int flags; 2286 2287 m = NULL; 2288 vmd = VM_DOMAIN(domain); 2289 again: 2290 if (vm_domain_allocate(vmd, req, 1)) { 2291 vm_domain_free_lock(vmd); 2292 m = vm_phys_alloc_freelist_pages(domain, freelist, 2293 VM_FREEPOOL_DIRECT, 0); 2294 vm_domain_free_unlock(vmd); 2295 if (m == NULL) 2296 vm_domain_freecnt_inc(vmd, 1); 2297 } 2298 if (m == NULL) { 2299 if (vm_domain_alloc_fail(vmd, NULL, req)) 2300 goto again; 2301 return (NULL); 2302 } 2303 vm_page_dequeue(m); 2304 vm_page_alloc_check(m); 2305 2306 /* 2307 * Initialize the page. Only the PG_ZERO flag is inherited. 2308 */ 2309 m->aflags = 0; 2310 flags = 0; 2311 if ((req & VM_ALLOC_ZERO) != 0) 2312 flags = PG_ZERO; 2313 m->flags &= flags; 2314 if ((req & VM_ALLOC_WIRED) != 0) { 2315 /* 2316 * The page lock is not required for wiring a page that does 2317 * not belong to an object. 2318 */ 2319 vm_wire_add(1); 2320 m->ref_count = 1; 2321 } 2322 /* Unmanaged pages don't use "act_count". */ 2323 m->oflags = VPO_UNMANAGED; 2324 return (m); 2325 } 2326 2327 static int 2328 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2329 { 2330 struct vm_domain *vmd; 2331 struct vm_pgcache *pgcache; 2332 int i; 2333 2334 pgcache = arg; 2335 vmd = VM_DOMAIN(pgcache->domain); 2336 2337 /* 2338 * The page daemon should avoid creating extra memory pressure since its 2339 * main purpose is to replenish the store of free pages. 2340 */ 2341 if (vmd->vmd_severeset || curproc == pageproc || 2342 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2343 return (0); 2344 domain = vmd->vmd_domain; 2345 vm_domain_free_lock(vmd); 2346 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2347 (vm_page_t *)store); 2348 vm_domain_free_unlock(vmd); 2349 if (cnt != i) 2350 vm_domain_freecnt_inc(vmd, cnt - i); 2351 2352 return (i); 2353 } 2354 2355 static void 2356 vm_page_zone_release(void *arg, void **store, int cnt) 2357 { 2358 struct vm_domain *vmd; 2359 struct vm_pgcache *pgcache; 2360 vm_page_t m; 2361 int i; 2362 2363 pgcache = arg; 2364 vmd = VM_DOMAIN(pgcache->domain); 2365 vm_domain_free_lock(vmd); 2366 for (i = 0; i < cnt; i++) { 2367 m = (vm_page_t)store[i]; 2368 vm_phys_free_pages(m, 0); 2369 } 2370 vm_domain_free_unlock(vmd); 2371 vm_domain_freecnt_inc(vmd, cnt); 2372 } 2373 2374 #define VPSC_ANY 0 /* No restrictions. */ 2375 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2376 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2377 2378 /* 2379 * vm_page_scan_contig: 2380 * 2381 * Scan vm_page_array[] between the specified entries "m_start" and 2382 * "m_end" for a run of contiguous physical pages that satisfy the 2383 * specified conditions, and return the lowest page in the run. The 2384 * specified "alignment" determines the alignment of the lowest physical 2385 * page in the run. If the specified "boundary" is non-zero, then the 2386 * run of physical pages cannot span a physical address that is a 2387 * multiple of "boundary". 2388 * 2389 * "m_end" is never dereferenced, so it need not point to a vm_page 2390 * structure within vm_page_array[]. 2391 * 2392 * "npages" must be greater than zero. "m_start" and "m_end" must not 2393 * span a hole (or discontiguity) in the physical address space. Both 2394 * "alignment" and "boundary" must be a power of two. 2395 */ 2396 vm_page_t 2397 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2398 u_long alignment, vm_paddr_t boundary, int options) 2399 { 2400 struct mtx *m_mtx; 2401 vm_object_t object; 2402 vm_paddr_t pa; 2403 vm_page_t m, m_run; 2404 #if VM_NRESERVLEVEL > 0 2405 int level; 2406 #endif 2407 int m_inc, order, run_ext, run_len; 2408 2409 KASSERT(npages > 0, ("npages is 0")); 2410 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2411 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2412 m_run = NULL; 2413 run_len = 0; 2414 m_mtx = NULL; 2415 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2416 KASSERT((m->flags & PG_MARKER) == 0, 2417 ("page %p is PG_MARKER", m)); 2418 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2419 ("fictitious page %p has invalid ref count", m)); 2420 2421 /* 2422 * If the current page would be the start of a run, check its 2423 * physical address against the end, alignment, and boundary 2424 * conditions. If it doesn't satisfy these conditions, either 2425 * terminate the scan or advance to the next page that 2426 * satisfies the failed condition. 2427 */ 2428 if (run_len == 0) { 2429 KASSERT(m_run == NULL, ("m_run != NULL")); 2430 if (m + npages > m_end) 2431 break; 2432 pa = VM_PAGE_TO_PHYS(m); 2433 if ((pa & (alignment - 1)) != 0) { 2434 m_inc = atop(roundup2(pa, alignment) - pa); 2435 continue; 2436 } 2437 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2438 boundary) != 0) { 2439 m_inc = atop(roundup2(pa, boundary) - pa); 2440 continue; 2441 } 2442 } else 2443 KASSERT(m_run != NULL, ("m_run == NULL")); 2444 2445 vm_page_change_lock(m, &m_mtx); 2446 m_inc = 1; 2447 retry: 2448 if (vm_page_wired(m)) 2449 run_ext = 0; 2450 #if VM_NRESERVLEVEL > 0 2451 else if ((level = vm_reserv_level(m)) >= 0 && 2452 (options & VPSC_NORESERV) != 0) { 2453 run_ext = 0; 2454 /* Advance to the end of the reservation. */ 2455 pa = VM_PAGE_TO_PHYS(m); 2456 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2457 pa); 2458 } 2459 #endif 2460 else if ((object = m->object) != NULL) { 2461 /* 2462 * The page is considered eligible for relocation if 2463 * and only if it could be laundered or reclaimed by 2464 * the page daemon. 2465 */ 2466 if (!VM_OBJECT_TRYRLOCK(object)) { 2467 mtx_unlock(m_mtx); 2468 VM_OBJECT_RLOCK(object); 2469 mtx_lock(m_mtx); 2470 if (m->object != object) { 2471 /* 2472 * The page may have been freed. 2473 */ 2474 VM_OBJECT_RUNLOCK(object); 2475 goto retry; 2476 } 2477 } 2478 /* Don't care: PG_NODUMP, PG_ZERO. */ 2479 if (object->type != OBJT_DEFAULT && 2480 object->type != OBJT_SWAP && 2481 object->type != OBJT_VNODE) { 2482 run_ext = 0; 2483 #if VM_NRESERVLEVEL > 0 2484 } else if ((options & VPSC_NOSUPER) != 0 && 2485 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2486 run_ext = 0; 2487 /* Advance to the end of the superpage. */ 2488 pa = VM_PAGE_TO_PHYS(m); 2489 m_inc = atop(roundup2(pa + 1, 2490 vm_reserv_size(level)) - pa); 2491 #endif 2492 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2493 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) && 2494 !vm_page_wired(m)) { 2495 /* 2496 * The page is allocated but eligible for 2497 * relocation. Extend the current run by one 2498 * page. 2499 */ 2500 KASSERT(pmap_page_get_memattr(m) == 2501 VM_MEMATTR_DEFAULT, 2502 ("page %p has an unexpected memattr", m)); 2503 KASSERT((m->oflags & (VPO_SWAPINPROG | 2504 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2505 ("page %p has unexpected oflags", m)); 2506 /* Don't care: PGA_NOSYNC. */ 2507 run_ext = 1; 2508 } else 2509 run_ext = 0; 2510 VM_OBJECT_RUNLOCK(object); 2511 #if VM_NRESERVLEVEL > 0 2512 } else if (level >= 0) { 2513 /* 2514 * The page is reserved but not yet allocated. In 2515 * other words, it is still free. Extend the current 2516 * run by one page. 2517 */ 2518 run_ext = 1; 2519 #endif 2520 } else if ((order = m->order) < VM_NFREEORDER) { 2521 /* 2522 * The page is enqueued in the physical memory 2523 * allocator's free page queues. Moreover, it is the 2524 * first page in a power-of-two-sized run of 2525 * contiguous free pages. Add these pages to the end 2526 * of the current run, and jump ahead. 2527 */ 2528 run_ext = 1 << order; 2529 m_inc = 1 << order; 2530 } else { 2531 /* 2532 * Skip the page for one of the following reasons: (1) 2533 * It is enqueued in the physical memory allocator's 2534 * free page queues. However, it is not the first 2535 * page in a run of contiguous free pages. (This case 2536 * rarely occurs because the scan is performed in 2537 * ascending order.) (2) It is not reserved, and it is 2538 * transitioning from free to allocated. (Conversely, 2539 * the transition from allocated to free for managed 2540 * pages is blocked by the page lock.) (3) It is 2541 * allocated but not contained by an object and not 2542 * wired, e.g., allocated by Xen's balloon driver. 2543 */ 2544 run_ext = 0; 2545 } 2546 2547 /* 2548 * Extend or reset the current run of pages. 2549 */ 2550 if (run_ext > 0) { 2551 if (run_len == 0) 2552 m_run = m; 2553 run_len += run_ext; 2554 } else { 2555 if (run_len > 0) { 2556 m_run = NULL; 2557 run_len = 0; 2558 } 2559 } 2560 } 2561 if (m_mtx != NULL) 2562 mtx_unlock(m_mtx); 2563 if (run_len >= npages) 2564 return (m_run); 2565 return (NULL); 2566 } 2567 2568 /* 2569 * vm_page_reclaim_run: 2570 * 2571 * Try to relocate each of the allocated virtual pages within the 2572 * specified run of physical pages to a new physical address. Free the 2573 * physical pages underlying the relocated virtual pages. A virtual page 2574 * is relocatable if and only if it could be laundered or reclaimed by 2575 * the page daemon. Whenever possible, a virtual page is relocated to a 2576 * physical address above "high". 2577 * 2578 * Returns 0 if every physical page within the run was already free or 2579 * just freed by a successful relocation. Otherwise, returns a non-zero 2580 * value indicating why the last attempt to relocate a virtual page was 2581 * unsuccessful. 2582 * 2583 * "req_class" must be an allocation class. 2584 */ 2585 static int 2586 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2587 vm_paddr_t high) 2588 { 2589 struct vm_domain *vmd; 2590 struct mtx *m_mtx; 2591 struct spglist free; 2592 vm_object_t object; 2593 vm_paddr_t pa; 2594 vm_page_t m, m_end, m_new; 2595 int error, order, req; 2596 2597 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2598 ("req_class is not an allocation class")); 2599 SLIST_INIT(&free); 2600 error = 0; 2601 m = m_run; 2602 m_end = m_run + npages; 2603 m_mtx = NULL; 2604 for (; error == 0 && m < m_end; m++) { 2605 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2606 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2607 2608 /* 2609 * Avoid releasing and reacquiring the same page lock. 2610 */ 2611 vm_page_change_lock(m, &m_mtx); 2612 retry: 2613 /* 2614 * Racily check for wirings. Races are handled below. 2615 */ 2616 if (vm_page_wired(m)) 2617 error = EBUSY; 2618 else if ((object = m->object) != NULL) { 2619 /* 2620 * The page is relocated if and only if it could be 2621 * laundered or reclaimed by the page daemon. 2622 */ 2623 if (!VM_OBJECT_TRYWLOCK(object)) { 2624 mtx_unlock(m_mtx); 2625 VM_OBJECT_WLOCK(object); 2626 mtx_lock(m_mtx); 2627 if (m->object != object) { 2628 /* 2629 * The page may have been freed. 2630 */ 2631 VM_OBJECT_WUNLOCK(object); 2632 goto retry; 2633 } 2634 } 2635 /* Don't care: PG_NODUMP, PG_ZERO. */ 2636 if (object->type != OBJT_DEFAULT && 2637 object->type != OBJT_SWAP && 2638 object->type != OBJT_VNODE) 2639 error = EINVAL; 2640 else if (object->memattr != VM_MEMATTR_DEFAULT) 2641 error = EINVAL; 2642 else if (vm_page_queue(m) != PQ_NONE && 2643 vm_page_tryxbusy(m) != 0) { 2644 if (vm_page_wired(m)) { 2645 vm_page_xunbusy(m); 2646 error = EBUSY; 2647 goto unlock; 2648 } 2649 KASSERT(pmap_page_get_memattr(m) == 2650 VM_MEMATTR_DEFAULT, 2651 ("page %p has an unexpected memattr", m)); 2652 KASSERT((m->oflags & (VPO_SWAPINPROG | 2653 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2654 ("page %p has unexpected oflags", m)); 2655 /* Don't care: PGA_NOSYNC. */ 2656 if (!vm_page_none_valid(m)) { 2657 /* 2658 * First, try to allocate a new page 2659 * that is above "high". Failing 2660 * that, try to allocate a new page 2661 * that is below "m_run". Allocate 2662 * the new page between the end of 2663 * "m_run" and "high" only as a last 2664 * resort. 2665 */ 2666 req = req_class | VM_ALLOC_NOOBJ; 2667 if ((m->flags & PG_NODUMP) != 0) 2668 req |= VM_ALLOC_NODUMP; 2669 if (trunc_page(high) != 2670 ~(vm_paddr_t)PAGE_MASK) { 2671 m_new = vm_page_alloc_contig( 2672 NULL, 0, req, 1, 2673 round_page(high), 2674 ~(vm_paddr_t)0, 2675 PAGE_SIZE, 0, 2676 VM_MEMATTR_DEFAULT); 2677 } else 2678 m_new = NULL; 2679 if (m_new == NULL) { 2680 pa = VM_PAGE_TO_PHYS(m_run); 2681 m_new = vm_page_alloc_contig( 2682 NULL, 0, req, 1, 2683 0, pa - 1, PAGE_SIZE, 0, 2684 VM_MEMATTR_DEFAULT); 2685 } 2686 if (m_new == NULL) { 2687 pa += ptoa(npages); 2688 m_new = vm_page_alloc_contig( 2689 NULL, 0, req, 1, 2690 pa, high, PAGE_SIZE, 0, 2691 VM_MEMATTR_DEFAULT); 2692 } 2693 if (m_new == NULL) { 2694 vm_page_xunbusy(m); 2695 error = ENOMEM; 2696 goto unlock; 2697 } 2698 2699 /* 2700 * Unmap the page and check for new 2701 * wirings that may have been acquired 2702 * through a pmap lookup. 2703 */ 2704 if (object->ref_count != 0 && 2705 !vm_page_try_remove_all(m)) { 2706 vm_page_free(m_new); 2707 error = EBUSY; 2708 goto unlock; 2709 } 2710 2711 /* 2712 * Replace "m" with the new page. For 2713 * vm_page_replace(), "m" must be busy 2714 * and dequeued. Finally, change "m" 2715 * as if vm_page_free() was called. 2716 */ 2717 m_new->aflags = m->aflags & 2718 ~PGA_QUEUE_STATE_MASK; 2719 KASSERT(m_new->oflags == VPO_UNMANAGED, 2720 ("page %p is managed", m_new)); 2721 pmap_copy_page(m, m_new); 2722 m_new->valid = m->valid; 2723 m_new->dirty = m->dirty; 2724 m->flags &= ~PG_ZERO; 2725 vm_page_dequeue(m); 2726 vm_page_replace_checked(m_new, object, 2727 m->pindex, m); 2728 if (vm_page_free_prep(m)) 2729 SLIST_INSERT_HEAD(&free, m, 2730 plinks.s.ss); 2731 2732 /* 2733 * The new page must be deactivated 2734 * before the object is unlocked. 2735 */ 2736 vm_page_change_lock(m_new, &m_mtx); 2737 vm_page_deactivate(m_new); 2738 } else { 2739 m->flags &= ~PG_ZERO; 2740 vm_page_dequeue(m); 2741 if (vm_page_free_prep(m)) 2742 SLIST_INSERT_HEAD(&free, m, 2743 plinks.s.ss); 2744 KASSERT(m->dirty == 0, 2745 ("page %p is dirty", m)); 2746 } 2747 } else 2748 error = EBUSY; 2749 unlock: 2750 VM_OBJECT_WUNLOCK(object); 2751 } else { 2752 MPASS(vm_phys_domain(m) == domain); 2753 vmd = VM_DOMAIN(domain); 2754 vm_domain_free_lock(vmd); 2755 order = m->order; 2756 if (order < VM_NFREEORDER) { 2757 /* 2758 * The page is enqueued in the physical memory 2759 * allocator's free page queues. Moreover, it 2760 * is the first page in a power-of-two-sized 2761 * run of contiguous free pages. Jump ahead 2762 * to the last page within that run, and 2763 * continue from there. 2764 */ 2765 m += (1 << order) - 1; 2766 } 2767 #if VM_NRESERVLEVEL > 0 2768 else if (vm_reserv_is_page_free(m)) 2769 order = 0; 2770 #endif 2771 vm_domain_free_unlock(vmd); 2772 if (order == VM_NFREEORDER) 2773 error = EINVAL; 2774 } 2775 } 2776 if (m_mtx != NULL) 2777 mtx_unlock(m_mtx); 2778 if ((m = SLIST_FIRST(&free)) != NULL) { 2779 int cnt; 2780 2781 vmd = VM_DOMAIN(domain); 2782 cnt = 0; 2783 vm_domain_free_lock(vmd); 2784 do { 2785 MPASS(vm_phys_domain(m) == domain); 2786 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2787 vm_phys_free_pages(m, 0); 2788 cnt++; 2789 } while ((m = SLIST_FIRST(&free)) != NULL); 2790 vm_domain_free_unlock(vmd); 2791 vm_domain_freecnt_inc(vmd, cnt); 2792 } 2793 return (error); 2794 } 2795 2796 #define NRUNS 16 2797 2798 CTASSERT(powerof2(NRUNS)); 2799 2800 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2801 2802 #define MIN_RECLAIM 8 2803 2804 /* 2805 * vm_page_reclaim_contig: 2806 * 2807 * Reclaim allocated, contiguous physical memory satisfying the specified 2808 * conditions by relocating the virtual pages using that physical memory. 2809 * Returns true if reclamation is successful and false otherwise. Since 2810 * relocation requires the allocation of physical pages, reclamation may 2811 * fail due to a shortage of free pages. When reclamation fails, callers 2812 * are expected to perform vm_wait() before retrying a failed allocation 2813 * operation, e.g., vm_page_alloc_contig(). 2814 * 2815 * The caller must always specify an allocation class through "req". 2816 * 2817 * allocation classes: 2818 * VM_ALLOC_NORMAL normal process request 2819 * VM_ALLOC_SYSTEM system *really* needs a page 2820 * VM_ALLOC_INTERRUPT interrupt time request 2821 * 2822 * The optional allocation flags are ignored. 2823 * 2824 * "npages" must be greater than zero. Both "alignment" and "boundary" 2825 * must be a power of two. 2826 */ 2827 bool 2828 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2829 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2830 { 2831 struct vm_domain *vmd; 2832 vm_paddr_t curr_low; 2833 vm_page_t m_run, m_runs[NRUNS]; 2834 u_long count, reclaimed; 2835 int error, i, options, req_class; 2836 2837 KASSERT(npages > 0, ("npages is 0")); 2838 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2839 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2840 req_class = req & VM_ALLOC_CLASS_MASK; 2841 2842 /* 2843 * The page daemon is allowed to dig deeper into the free page list. 2844 */ 2845 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2846 req_class = VM_ALLOC_SYSTEM; 2847 2848 /* 2849 * Return if the number of free pages cannot satisfy the requested 2850 * allocation. 2851 */ 2852 vmd = VM_DOMAIN(domain); 2853 count = vmd->vmd_free_count; 2854 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2855 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2856 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2857 return (false); 2858 2859 /* 2860 * Scan up to three times, relaxing the restrictions ("options") on 2861 * the reclamation of reservations and superpages each time. 2862 */ 2863 for (options = VPSC_NORESERV;;) { 2864 /* 2865 * Find the highest runs that satisfy the given constraints 2866 * and restrictions, and record them in "m_runs". 2867 */ 2868 curr_low = low; 2869 count = 0; 2870 for (;;) { 2871 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2872 high, alignment, boundary, options); 2873 if (m_run == NULL) 2874 break; 2875 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2876 m_runs[RUN_INDEX(count)] = m_run; 2877 count++; 2878 } 2879 2880 /* 2881 * Reclaim the highest runs in LIFO (descending) order until 2882 * the number of reclaimed pages, "reclaimed", is at least 2883 * MIN_RECLAIM. Reset "reclaimed" each time because each 2884 * reclamation is idempotent, and runs will (likely) recur 2885 * from one scan to the next as restrictions are relaxed. 2886 */ 2887 reclaimed = 0; 2888 for (i = 0; count > 0 && i < NRUNS; i++) { 2889 count--; 2890 m_run = m_runs[RUN_INDEX(count)]; 2891 error = vm_page_reclaim_run(req_class, domain, npages, 2892 m_run, high); 2893 if (error == 0) { 2894 reclaimed += npages; 2895 if (reclaimed >= MIN_RECLAIM) 2896 return (true); 2897 } 2898 } 2899 2900 /* 2901 * Either relax the restrictions on the next scan or return if 2902 * the last scan had no restrictions. 2903 */ 2904 if (options == VPSC_NORESERV) 2905 options = VPSC_NOSUPER; 2906 else if (options == VPSC_NOSUPER) 2907 options = VPSC_ANY; 2908 else if (options == VPSC_ANY) 2909 return (reclaimed != 0); 2910 } 2911 } 2912 2913 bool 2914 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2915 u_long alignment, vm_paddr_t boundary) 2916 { 2917 struct vm_domainset_iter di; 2918 int domain; 2919 bool ret; 2920 2921 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2922 do { 2923 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 2924 high, alignment, boundary); 2925 if (ret) 2926 break; 2927 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2928 2929 return (ret); 2930 } 2931 2932 /* 2933 * Set the domain in the appropriate page level domainset. 2934 */ 2935 void 2936 vm_domain_set(struct vm_domain *vmd) 2937 { 2938 2939 mtx_lock(&vm_domainset_lock); 2940 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 2941 vmd->vmd_minset = 1; 2942 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 2943 } 2944 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 2945 vmd->vmd_severeset = 1; 2946 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 2947 } 2948 mtx_unlock(&vm_domainset_lock); 2949 } 2950 2951 /* 2952 * Clear the domain from the appropriate page level domainset. 2953 */ 2954 void 2955 vm_domain_clear(struct vm_domain *vmd) 2956 { 2957 2958 mtx_lock(&vm_domainset_lock); 2959 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 2960 vmd->vmd_minset = 0; 2961 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 2962 if (vm_min_waiters != 0) { 2963 vm_min_waiters = 0; 2964 wakeup(&vm_min_domains); 2965 } 2966 } 2967 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 2968 vmd->vmd_severeset = 0; 2969 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 2970 if (vm_severe_waiters != 0) { 2971 vm_severe_waiters = 0; 2972 wakeup(&vm_severe_domains); 2973 } 2974 } 2975 2976 /* 2977 * If pageout daemon needs pages, then tell it that there are 2978 * some free. 2979 */ 2980 if (vmd->vmd_pageout_pages_needed && 2981 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 2982 wakeup(&vmd->vmd_pageout_pages_needed); 2983 vmd->vmd_pageout_pages_needed = 0; 2984 } 2985 2986 /* See comments in vm_wait_doms(). */ 2987 if (vm_pageproc_waiters) { 2988 vm_pageproc_waiters = 0; 2989 wakeup(&vm_pageproc_waiters); 2990 } 2991 mtx_unlock(&vm_domainset_lock); 2992 } 2993 2994 /* 2995 * Wait for free pages to exceed the min threshold globally. 2996 */ 2997 void 2998 vm_wait_min(void) 2999 { 3000 3001 mtx_lock(&vm_domainset_lock); 3002 while (vm_page_count_min()) { 3003 vm_min_waiters++; 3004 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3005 } 3006 mtx_unlock(&vm_domainset_lock); 3007 } 3008 3009 /* 3010 * Wait for free pages to exceed the severe threshold globally. 3011 */ 3012 void 3013 vm_wait_severe(void) 3014 { 3015 3016 mtx_lock(&vm_domainset_lock); 3017 while (vm_page_count_severe()) { 3018 vm_severe_waiters++; 3019 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3020 "vmwait", 0); 3021 } 3022 mtx_unlock(&vm_domainset_lock); 3023 } 3024 3025 u_int 3026 vm_wait_count(void) 3027 { 3028 3029 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3030 } 3031 3032 void 3033 vm_wait_doms(const domainset_t *wdoms) 3034 { 3035 3036 /* 3037 * We use racey wakeup synchronization to avoid expensive global 3038 * locking for the pageproc when sleeping with a non-specific vm_wait. 3039 * To handle this, we only sleep for one tick in this instance. It 3040 * is expected that most allocations for the pageproc will come from 3041 * kmem or vm_page_grab* which will use the more specific and 3042 * race-free vm_wait_domain(). 3043 */ 3044 if (curproc == pageproc) { 3045 mtx_lock(&vm_domainset_lock); 3046 vm_pageproc_waiters++; 3047 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, 3048 "pageprocwait", 1); 3049 } else { 3050 /* 3051 * XXX Ideally we would wait only until the allocation could 3052 * be satisfied. This condition can cause new allocators to 3053 * consume all freed pages while old allocators wait. 3054 */ 3055 mtx_lock(&vm_domainset_lock); 3056 if (vm_page_count_min_set(wdoms)) { 3057 vm_min_waiters++; 3058 msleep(&vm_min_domains, &vm_domainset_lock, 3059 PVM | PDROP, "vmwait", 0); 3060 } else 3061 mtx_unlock(&vm_domainset_lock); 3062 } 3063 } 3064 3065 /* 3066 * vm_wait_domain: 3067 * 3068 * Sleep until free pages are available for allocation. 3069 * - Called in various places after failed memory allocations. 3070 */ 3071 void 3072 vm_wait_domain(int domain) 3073 { 3074 struct vm_domain *vmd; 3075 domainset_t wdom; 3076 3077 vmd = VM_DOMAIN(domain); 3078 vm_domain_free_assert_unlocked(vmd); 3079 3080 if (curproc == pageproc) { 3081 mtx_lock(&vm_domainset_lock); 3082 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3083 vmd->vmd_pageout_pages_needed = 1; 3084 msleep(&vmd->vmd_pageout_pages_needed, 3085 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3086 } else 3087 mtx_unlock(&vm_domainset_lock); 3088 } else { 3089 if (pageproc == NULL) 3090 panic("vm_wait in early boot"); 3091 DOMAINSET_ZERO(&wdom); 3092 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3093 vm_wait_doms(&wdom); 3094 } 3095 } 3096 3097 /* 3098 * vm_wait: 3099 * 3100 * Sleep until free pages are available for allocation in the 3101 * affinity domains of the obj. If obj is NULL, the domain set 3102 * for the calling thread is used. 3103 * Called in various places after failed memory allocations. 3104 */ 3105 void 3106 vm_wait(vm_object_t obj) 3107 { 3108 struct domainset *d; 3109 3110 d = NULL; 3111 3112 /* 3113 * Carefully fetch pointers only once: the struct domainset 3114 * itself is ummutable but the pointer might change. 3115 */ 3116 if (obj != NULL) 3117 d = obj->domain.dr_policy; 3118 if (d == NULL) 3119 d = curthread->td_domain.dr_policy; 3120 3121 vm_wait_doms(&d->ds_mask); 3122 } 3123 3124 /* 3125 * vm_domain_alloc_fail: 3126 * 3127 * Called when a page allocation function fails. Informs the 3128 * pagedaemon and performs the requested wait. Requires the 3129 * domain_free and object lock on entry. Returns with the 3130 * object lock held and free lock released. Returns an error when 3131 * retry is necessary. 3132 * 3133 */ 3134 static int 3135 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3136 { 3137 3138 vm_domain_free_assert_unlocked(vmd); 3139 3140 atomic_add_int(&vmd->vmd_pageout_deficit, 3141 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3142 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3143 if (object != NULL) 3144 VM_OBJECT_WUNLOCK(object); 3145 vm_wait_domain(vmd->vmd_domain); 3146 if (object != NULL) 3147 VM_OBJECT_WLOCK(object); 3148 if (req & VM_ALLOC_WAITOK) 3149 return (EAGAIN); 3150 } 3151 3152 return (0); 3153 } 3154 3155 /* 3156 * vm_waitpfault: 3157 * 3158 * Sleep until free pages are available for allocation. 3159 * - Called only in vm_fault so that processes page faulting 3160 * can be easily tracked. 3161 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3162 * processes will be able to grab memory first. Do not change 3163 * this balance without careful testing first. 3164 */ 3165 void 3166 vm_waitpfault(struct domainset *dset, int timo) 3167 { 3168 3169 /* 3170 * XXX Ideally we would wait only until the allocation could 3171 * be satisfied. This condition can cause new allocators to 3172 * consume all freed pages while old allocators wait. 3173 */ 3174 mtx_lock(&vm_domainset_lock); 3175 if (vm_page_count_min_set(&dset->ds_mask)) { 3176 vm_min_waiters++; 3177 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3178 "pfault", timo); 3179 } else 3180 mtx_unlock(&vm_domainset_lock); 3181 } 3182 3183 static struct vm_pagequeue * 3184 vm_page_pagequeue(vm_page_t m) 3185 { 3186 3187 uint8_t queue; 3188 3189 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) 3190 return (NULL); 3191 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3192 } 3193 3194 static inline void 3195 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m) 3196 { 3197 struct vm_domain *vmd; 3198 uint16_t qflags; 3199 3200 CRITICAL_ASSERT(curthread); 3201 vm_pagequeue_assert_locked(pq); 3202 3203 /* 3204 * The page daemon is allowed to set m->queue = PQ_NONE without 3205 * the page queue lock held. In this case it is about to free the page, 3206 * which must not have any queue state. 3207 */ 3208 qflags = atomic_load_16(&m->aflags); 3209 KASSERT(pq == vm_page_pagequeue(m) || 3210 (qflags & PGA_QUEUE_STATE_MASK) == 0, 3211 ("page %p doesn't belong to queue %p but has aflags %#x", 3212 m, pq, qflags)); 3213 3214 if ((qflags & PGA_DEQUEUE) != 0) { 3215 if (__predict_true((qflags & PGA_ENQUEUED) != 0)) 3216 vm_pagequeue_remove(pq, m); 3217 vm_page_dequeue_complete(m); 3218 counter_u64_add(queue_ops, 1); 3219 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) { 3220 if ((qflags & PGA_ENQUEUED) != 0) 3221 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3222 else { 3223 vm_pagequeue_cnt_inc(pq); 3224 vm_page_aflag_set(m, PGA_ENQUEUED); 3225 } 3226 3227 /* 3228 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. 3229 * In particular, if both flags are set in close succession, 3230 * only PGA_REQUEUE_HEAD will be applied, even if it was set 3231 * first. 3232 */ 3233 if ((qflags & PGA_REQUEUE_HEAD) != 0) { 3234 KASSERT(m->queue == PQ_INACTIVE, 3235 ("head enqueue not supported for page %p", m)); 3236 vmd = vm_pagequeue_domain(m); 3237 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3238 } else 3239 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3240 3241 vm_page_aflag_clear(m, qflags & (PGA_REQUEUE | 3242 PGA_REQUEUE_HEAD)); 3243 counter_u64_add(queue_ops, 1); 3244 } else { 3245 counter_u64_add(queue_nops, 1); 3246 } 3247 } 3248 3249 static void 3250 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3251 uint8_t queue) 3252 { 3253 vm_page_t m; 3254 int i; 3255 3256 for (i = 0; i < bq->bq_cnt; i++) { 3257 m = bq->bq_pa[i]; 3258 if (__predict_false(m->queue != queue)) 3259 continue; 3260 vm_pqbatch_process_page(pq, m); 3261 } 3262 vm_batchqueue_init(bq); 3263 } 3264 3265 /* 3266 * vm_page_pqbatch_submit: [ internal use only ] 3267 * 3268 * Enqueue a page in the specified page queue's batched work queue. 3269 * The caller must have encoded the requested operation in the page 3270 * structure's aflags field. 3271 */ 3272 void 3273 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3274 { 3275 struct vm_batchqueue *bq; 3276 struct vm_pagequeue *pq; 3277 int domain; 3278 3279 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3280 ("page %p is unmanaged", m)); 3281 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL, 3282 ("missing synchronization for page %p", m)); 3283 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3284 3285 domain = vm_phys_domain(m); 3286 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 3287 3288 critical_enter(); 3289 bq = DPCPU_PTR(pqbatch[domain][queue]); 3290 if (vm_batchqueue_insert(bq, m)) { 3291 critical_exit(); 3292 return; 3293 } 3294 critical_exit(); 3295 vm_pagequeue_lock(pq); 3296 critical_enter(); 3297 bq = DPCPU_PTR(pqbatch[domain][queue]); 3298 vm_pqbatch_process(pq, bq, queue); 3299 3300 /* 3301 * The page may have been logically dequeued before we acquired the 3302 * page queue lock. In this case, since we either hold the page lock 3303 * or the page is being freed, a different thread cannot be concurrently 3304 * enqueuing the page. 3305 */ 3306 if (__predict_true(m->queue == queue)) 3307 vm_pqbatch_process_page(pq, m); 3308 else { 3309 KASSERT(m->queue == PQ_NONE, 3310 ("invalid queue transition for page %p", m)); 3311 KASSERT((m->aflags & PGA_ENQUEUED) == 0, 3312 ("page %p is enqueued with invalid queue index", m)); 3313 } 3314 vm_pagequeue_unlock(pq); 3315 critical_exit(); 3316 } 3317 3318 /* 3319 * vm_page_pqbatch_drain: [ internal use only ] 3320 * 3321 * Force all per-CPU page queue batch queues to be drained. This is 3322 * intended for use in severe memory shortages, to ensure that pages 3323 * do not remain stuck in the batch queues. 3324 */ 3325 void 3326 vm_page_pqbatch_drain(void) 3327 { 3328 struct thread *td; 3329 struct vm_domain *vmd; 3330 struct vm_pagequeue *pq; 3331 int cpu, domain, queue; 3332 3333 td = curthread; 3334 CPU_FOREACH(cpu) { 3335 thread_lock(td); 3336 sched_bind(td, cpu); 3337 thread_unlock(td); 3338 3339 for (domain = 0; domain < vm_ndomains; domain++) { 3340 vmd = VM_DOMAIN(domain); 3341 for (queue = 0; queue < PQ_COUNT; queue++) { 3342 pq = &vmd->vmd_pagequeues[queue]; 3343 vm_pagequeue_lock(pq); 3344 critical_enter(); 3345 vm_pqbatch_process(pq, 3346 DPCPU_PTR(pqbatch[domain][queue]), queue); 3347 critical_exit(); 3348 vm_pagequeue_unlock(pq); 3349 } 3350 } 3351 } 3352 thread_lock(td); 3353 sched_unbind(td); 3354 thread_unlock(td); 3355 } 3356 3357 /* 3358 * Complete the logical removal of a page from a page queue. We must be 3359 * careful to synchronize with the page daemon, which may be concurrently 3360 * examining the page with only the page lock held. The page must not be 3361 * in a state where it appears to be logically enqueued. 3362 */ 3363 static void 3364 vm_page_dequeue_complete(vm_page_t m) 3365 { 3366 3367 m->queue = PQ_NONE; 3368 atomic_thread_fence_rel(); 3369 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); 3370 } 3371 3372 /* 3373 * vm_page_dequeue_deferred: [ internal use only ] 3374 * 3375 * Request removal of the given page from its current page 3376 * queue. Physical removal from the queue may be deferred 3377 * indefinitely. 3378 * 3379 * The page must be locked. 3380 */ 3381 void 3382 vm_page_dequeue_deferred(vm_page_t m) 3383 { 3384 uint8_t queue; 3385 3386 vm_page_assert_locked(m); 3387 3388 if ((queue = vm_page_queue(m)) == PQ_NONE) 3389 return; 3390 3391 /* 3392 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call 3393 * to vm_page_dequeue_deferred_free(). In particular, avoid modifying 3394 * the page's queue state once vm_page_dequeue_deferred_free() has been 3395 * called. In the event of a race, two batch queue entries for the page 3396 * will be created, but the second will have no effect. 3397 */ 3398 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE)) 3399 vm_page_pqbatch_submit(m, queue); 3400 } 3401 3402 /* 3403 * A variant of vm_page_dequeue_deferred() that does not assert the page 3404 * lock and is only to be called from vm_page_free_prep(). Because the 3405 * page is being freed, we can assume that nothing other than the page 3406 * daemon is scheduling queue operations on this page, so we get for 3407 * free the mutual exclusion that is otherwise provided by the page lock. 3408 * To handle races, the page daemon must take care to atomically check 3409 * for PGA_DEQUEUE when updating queue state. 3410 */ 3411 static void 3412 vm_page_dequeue_deferred_free(vm_page_t m) 3413 { 3414 uint8_t queue; 3415 3416 KASSERT(m->ref_count == 0, ("page %p has references", m)); 3417 3418 for (;;) { 3419 if ((m->aflags & PGA_DEQUEUE) != 0) 3420 return; 3421 atomic_thread_fence_acq(); 3422 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) 3423 return; 3424 if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, 3425 PGA_DEQUEUE)) { 3426 vm_page_pqbatch_submit(m, queue); 3427 break; 3428 } 3429 } 3430 } 3431 3432 /* 3433 * vm_page_dequeue: 3434 * 3435 * Remove the page from whichever page queue it's in, if any. 3436 * The page must either be locked or unallocated. This constraint 3437 * ensures that the queue state of the page will remain consistent 3438 * after this function returns. 3439 */ 3440 void 3441 vm_page_dequeue(vm_page_t m) 3442 { 3443 struct vm_pagequeue *pq, *pq1; 3444 uint16_t aflags; 3445 3446 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->ref_count == 0, 3447 ("page %p is allocated and unlocked", m)); 3448 3449 for (pq = vm_page_pagequeue(m);; pq = pq1) { 3450 if (pq == NULL) { 3451 /* 3452 * A thread may be concurrently executing 3453 * vm_page_dequeue_complete(). Ensure that all queue 3454 * state is cleared before we return. 3455 */ 3456 aflags = atomic_load_16(&m->aflags); 3457 if ((aflags & PGA_QUEUE_STATE_MASK) == 0) 3458 return; 3459 KASSERT((aflags & PGA_DEQUEUE) != 0, 3460 ("page %p has unexpected queue state flags %#x", 3461 m, aflags)); 3462 3463 /* 3464 * Busy wait until the thread updating queue state is 3465 * finished. Such a thread must be executing in a 3466 * critical section. 3467 */ 3468 cpu_spinwait(); 3469 pq1 = vm_page_pagequeue(m); 3470 continue; 3471 } 3472 vm_pagequeue_lock(pq); 3473 if ((pq1 = vm_page_pagequeue(m)) == pq) 3474 break; 3475 vm_pagequeue_unlock(pq); 3476 } 3477 KASSERT(pq == vm_page_pagequeue(m), 3478 ("%s: page %p migrated directly between queues", __func__, m)); 3479 KASSERT((m->aflags & PGA_DEQUEUE) != 0 || 3480 mtx_owned(vm_page_lockptr(m)), 3481 ("%s: queued unlocked page %p", __func__, m)); 3482 3483 if ((m->aflags & PGA_ENQUEUED) != 0) 3484 vm_pagequeue_remove(pq, m); 3485 vm_page_dequeue_complete(m); 3486 vm_pagequeue_unlock(pq); 3487 } 3488 3489 /* 3490 * Schedule the given page for insertion into the specified page queue. 3491 * Physical insertion of the page may be deferred indefinitely. 3492 */ 3493 static void 3494 vm_page_enqueue(vm_page_t m, uint8_t queue) 3495 { 3496 3497 vm_page_assert_locked(m); 3498 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 3499 ("%s: page %p is already enqueued", __func__, m)); 3500 KASSERT(m->ref_count > 0, 3501 ("%s: page %p does not carry any references", __func__, m)); 3502 3503 m->queue = queue; 3504 if ((m->aflags & PGA_REQUEUE) == 0) 3505 vm_page_aflag_set(m, PGA_REQUEUE); 3506 vm_page_pqbatch_submit(m, queue); 3507 } 3508 3509 /* 3510 * vm_page_requeue: [ internal use only ] 3511 * 3512 * Schedule a requeue of the given page. 3513 * 3514 * The page must be locked. 3515 */ 3516 void 3517 vm_page_requeue(vm_page_t m) 3518 { 3519 3520 vm_page_assert_locked(m); 3521 KASSERT(vm_page_queue(m) != PQ_NONE, 3522 ("%s: page %p is not logically enqueued", __func__, m)); 3523 KASSERT(m->ref_count > 0, 3524 ("%s: page %p does not carry any references", __func__, m)); 3525 3526 if ((m->aflags & PGA_REQUEUE) == 0) 3527 vm_page_aflag_set(m, PGA_REQUEUE); 3528 vm_page_pqbatch_submit(m, atomic_load_8(&m->queue)); 3529 } 3530 3531 /* 3532 * vm_page_swapqueue: [ internal use only ] 3533 * 3534 * Move the page from one queue to another, or to the tail of its 3535 * current queue, in the face of a possible concurrent call to 3536 * vm_page_dequeue_deferred_free(). 3537 */ 3538 void 3539 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq) 3540 { 3541 struct vm_pagequeue *pq; 3542 vm_page_t next; 3543 bool queued; 3544 3545 KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq, 3546 ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq)); 3547 vm_page_assert_locked(m); 3548 3549 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq]; 3550 vm_pagequeue_lock(pq); 3551 3552 /* 3553 * The physical queue state might change at any point before the page 3554 * queue lock is acquired, so we must verify that we hold the correct 3555 * lock before proceeding. 3556 */ 3557 if (__predict_false(m->queue != oldq)) { 3558 vm_pagequeue_unlock(pq); 3559 return; 3560 } 3561 3562 /* 3563 * Once the queue index of the page changes, there is nothing 3564 * synchronizing with further updates to the physical queue state. 3565 * Therefore we must remove the page from the queue now in anticipation 3566 * of a successful commit, and be prepared to roll back. 3567 */ 3568 if (__predict_true((m->aflags & PGA_ENQUEUED) != 0)) { 3569 next = TAILQ_NEXT(m, plinks.q); 3570 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3571 vm_page_aflag_clear(m, PGA_ENQUEUED); 3572 queued = true; 3573 } else { 3574 queued = false; 3575 } 3576 3577 /* 3578 * Atomically update the queue field and set PGA_REQUEUE while 3579 * ensuring that PGA_DEQUEUE has not been set. 3580 */ 3581 if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE, 3582 PGA_REQUEUE))) { 3583 if (queued) { 3584 vm_page_aflag_set(m, PGA_ENQUEUED); 3585 if (next != NULL) 3586 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3587 else 3588 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3589 } 3590 vm_pagequeue_unlock(pq); 3591 return; 3592 } 3593 vm_pagequeue_cnt_dec(pq); 3594 vm_pagequeue_unlock(pq); 3595 vm_page_pqbatch_submit(m, newq); 3596 } 3597 3598 /* 3599 * vm_page_free_prep: 3600 * 3601 * Prepares the given page to be put on the free list, 3602 * disassociating it from any VM object. The caller may return 3603 * the page to the free list only if this function returns true. 3604 * 3605 * The object must be locked. The page must be locked if it is 3606 * managed. 3607 */ 3608 bool 3609 vm_page_free_prep(vm_page_t m) 3610 { 3611 3612 /* 3613 * Synchronize with threads that have dropped a reference to this 3614 * page. 3615 */ 3616 atomic_thread_fence_acq(); 3617 3618 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3619 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3620 uint64_t *p; 3621 int i; 3622 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3623 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3624 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3625 m, i, (uintmax_t)*p)); 3626 } 3627 #endif 3628 if ((m->oflags & VPO_UNMANAGED) == 0) { 3629 KASSERT(!pmap_page_is_mapped(m), 3630 ("vm_page_free_prep: freeing mapped page %p", m)); 3631 KASSERT((m->aflags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 3632 ("vm_page_free_prep: mapping flags set in page %p", m)); 3633 } else { 3634 KASSERT(m->queue == PQ_NONE, 3635 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3636 } 3637 VM_CNT_INC(v_tfree); 3638 3639 if (vm_page_sbusied(m)) 3640 panic("vm_page_free_prep: freeing shared busy page %p", m); 3641 3642 if (m->object != NULL) { 3643 vm_page_object_remove(m); 3644 3645 /* 3646 * The object reference can be released without an atomic 3647 * operation. 3648 */ 3649 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 3650 m->ref_count == VPRC_OBJREF, 3651 ("vm_page_free_prep: page %p has unexpected ref_count %u", 3652 m, m->ref_count)); 3653 m->object = NULL; 3654 m->ref_count -= VPRC_OBJREF; 3655 } 3656 3657 if (vm_page_xbusied(m)) 3658 vm_page_xunbusy(m); 3659 3660 /* 3661 * If fictitious remove object association and 3662 * return. 3663 */ 3664 if ((m->flags & PG_FICTITIOUS) != 0) { 3665 KASSERT(m->ref_count == 1, 3666 ("fictitious page %p is referenced", m)); 3667 KASSERT(m->queue == PQ_NONE, 3668 ("fictitious page %p is queued", m)); 3669 return (false); 3670 } 3671 3672 /* 3673 * Pages need not be dequeued before they are returned to the physical 3674 * memory allocator, but they must at least be marked for a deferred 3675 * dequeue. 3676 */ 3677 if ((m->oflags & VPO_UNMANAGED) == 0) 3678 vm_page_dequeue_deferred_free(m); 3679 3680 m->valid = 0; 3681 vm_page_undirty(m); 3682 3683 if (m->ref_count != 0) 3684 panic("vm_page_free_prep: page %p has references", m); 3685 3686 /* 3687 * Restore the default memory attribute to the page. 3688 */ 3689 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3690 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3691 3692 #if VM_NRESERVLEVEL > 0 3693 /* 3694 * Determine whether the page belongs to a reservation. If the page was 3695 * allocated from a per-CPU cache, it cannot belong to a reservation, so 3696 * as an optimization, we avoid the check in that case. 3697 */ 3698 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 3699 return (false); 3700 #endif 3701 3702 return (true); 3703 } 3704 3705 /* 3706 * vm_page_free_toq: 3707 * 3708 * Returns the given page to the free list, disassociating it 3709 * from any VM object. 3710 * 3711 * The object must be locked. The page must be locked if it is 3712 * managed. 3713 */ 3714 void 3715 vm_page_free_toq(vm_page_t m) 3716 { 3717 struct vm_domain *vmd; 3718 uma_zone_t zone; 3719 3720 if (!vm_page_free_prep(m)) 3721 return; 3722 3723 vmd = vm_pagequeue_domain(m); 3724 zone = vmd->vmd_pgcache[m->pool].zone; 3725 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 3726 uma_zfree(zone, m); 3727 return; 3728 } 3729 vm_domain_free_lock(vmd); 3730 vm_phys_free_pages(m, 0); 3731 vm_domain_free_unlock(vmd); 3732 vm_domain_freecnt_inc(vmd, 1); 3733 } 3734 3735 /* 3736 * vm_page_free_pages_toq: 3737 * 3738 * Returns a list of pages to the free list, disassociating it 3739 * from any VM object. In other words, this is equivalent to 3740 * calling vm_page_free_toq() for each page of a list of VM objects. 3741 * 3742 * The objects must be locked. The pages must be locked if it is 3743 * managed. 3744 */ 3745 void 3746 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3747 { 3748 vm_page_t m; 3749 int count; 3750 3751 if (SLIST_EMPTY(free)) 3752 return; 3753 3754 count = 0; 3755 while ((m = SLIST_FIRST(free)) != NULL) { 3756 count++; 3757 SLIST_REMOVE_HEAD(free, plinks.s.ss); 3758 vm_page_free_toq(m); 3759 } 3760 3761 if (update_wire_count) 3762 vm_wire_sub(count); 3763 } 3764 3765 /* 3766 * Mark this page as wired down, preventing reclamation by the page daemon 3767 * or when the containing object is destroyed. 3768 */ 3769 void 3770 vm_page_wire(vm_page_t m) 3771 { 3772 u_int old; 3773 3774 KASSERT(m->object != NULL, 3775 ("vm_page_wire: page %p does not belong to an object", m)); 3776 if (!vm_page_busied(m)) 3777 VM_OBJECT_ASSERT_LOCKED(m->object); 3778 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 3779 VPRC_WIRE_COUNT(m->ref_count) >= 1, 3780 ("vm_page_wire: fictitious page %p has zero wirings", m)); 3781 3782 old = atomic_fetchadd_int(&m->ref_count, 1); 3783 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 3784 ("vm_page_wire: counter overflow for page %p", m)); 3785 if (VPRC_WIRE_COUNT(old) == 0) 3786 vm_wire_add(1); 3787 } 3788 3789 /* 3790 * Attempt to wire a mapped page following a pmap lookup of that page. 3791 * This may fail if a thread is concurrently tearing down mappings of the page. 3792 * The transient failure is acceptable because it translates to the 3793 * failure of the caller pmap_extract_and_hold(), which should be then 3794 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 3795 */ 3796 bool 3797 vm_page_wire_mapped(vm_page_t m) 3798 { 3799 u_int old; 3800 3801 old = m->ref_count; 3802 do { 3803 KASSERT(old > 0, 3804 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 3805 if ((old & VPRC_BLOCKED) != 0) 3806 return (false); 3807 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 3808 3809 if (VPRC_WIRE_COUNT(old) == 0) 3810 vm_wire_add(1); 3811 return (true); 3812 } 3813 3814 /* 3815 * Release one wiring of the specified page, potentially allowing it to be 3816 * paged out. 3817 * 3818 * Only managed pages belonging to an object can be paged out. If the number 3819 * of wirings transitions to zero and the page is eligible for page out, then 3820 * the page is added to the specified paging queue. If the released wiring 3821 * represented the last reference to the page, the page is freed. 3822 * 3823 * A managed page must be locked. 3824 */ 3825 void 3826 vm_page_unwire(vm_page_t m, uint8_t queue) 3827 { 3828 u_int old; 3829 bool locked; 3830 3831 KASSERT(queue < PQ_COUNT, 3832 ("vm_page_unwire: invalid queue %u request for page %p", queue, m)); 3833 3834 if ((m->oflags & VPO_UNMANAGED) != 0) { 3835 if (vm_page_unwire_noq(m) && m->ref_count == 0) 3836 vm_page_free(m); 3837 return; 3838 } 3839 3840 /* 3841 * Update LRU state before releasing the wiring reference. 3842 * We only need to do this once since we hold the page lock. 3843 * Use a release store when updating the reference count to 3844 * synchronize with vm_page_free_prep(). 3845 */ 3846 old = m->ref_count; 3847 locked = false; 3848 do { 3849 KASSERT(VPRC_WIRE_COUNT(old) > 0, 3850 ("vm_page_unwire: wire count underflow for page %p", m)); 3851 if (!locked && VPRC_WIRE_COUNT(old) == 1) { 3852 vm_page_lock(m); 3853 locked = true; 3854 if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE) 3855 vm_page_reference(m); 3856 else 3857 vm_page_mvqueue(m, queue); 3858 } 3859 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 3860 3861 /* 3862 * Release the lock only after the wiring is released, to ensure that 3863 * the page daemon does not encounter and dequeue the page while it is 3864 * still wired. 3865 */ 3866 if (locked) 3867 vm_page_unlock(m); 3868 3869 if (VPRC_WIRE_COUNT(old) == 1) { 3870 vm_wire_sub(1); 3871 if (old == 1) 3872 vm_page_free(m); 3873 } 3874 } 3875 3876 /* 3877 * Unwire a page without (re-)inserting it into a page queue. It is up 3878 * to the caller to enqueue, requeue, or free the page as appropriate. 3879 * In most cases involving managed pages, vm_page_unwire() should be used 3880 * instead. 3881 */ 3882 bool 3883 vm_page_unwire_noq(vm_page_t m) 3884 { 3885 u_int old; 3886 3887 old = vm_page_drop(m, 1); 3888 KASSERT(VPRC_WIRE_COUNT(old) != 0, 3889 ("vm_page_unref: counter underflow for page %p", m)); 3890 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 3891 ("vm_page_unref: missing ref on fictitious page %p", m)); 3892 3893 if (VPRC_WIRE_COUNT(old) > 1) 3894 return (false); 3895 vm_wire_sub(1); 3896 return (true); 3897 } 3898 3899 /* 3900 * Ensure that the page is in the specified page queue. If the page is 3901 * active or being moved to the active queue, ensure that its act_count is 3902 * at least ACT_INIT but do not otherwise mess with it. Otherwise, ensure that 3903 * the page is at the tail of its page queue. 3904 * 3905 * The page may be wired. The caller should release its wiring reference 3906 * before releasing the page lock, otherwise the page daemon may immediately 3907 * dequeue the page. 3908 * 3909 * A managed page must be locked. 3910 */ 3911 static __always_inline void 3912 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue) 3913 { 3914 3915 vm_page_assert_locked(m); 3916 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3917 ("vm_page_mvqueue: page %p is unmanaged", m)); 3918 KASSERT(m->ref_count > 0, 3919 ("%s: page %p does not carry any references", __func__, m)); 3920 3921 if (vm_page_queue(m) != nqueue) { 3922 vm_page_dequeue(m); 3923 vm_page_enqueue(m, nqueue); 3924 } else if (nqueue != PQ_ACTIVE) { 3925 vm_page_requeue(m); 3926 } 3927 3928 if (nqueue == PQ_ACTIVE && m->act_count < ACT_INIT) 3929 m->act_count = ACT_INIT; 3930 } 3931 3932 /* 3933 * Put the specified page on the active list (if appropriate). 3934 */ 3935 void 3936 vm_page_activate(vm_page_t m) 3937 { 3938 3939 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3940 return; 3941 vm_page_mvqueue(m, PQ_ACTIVE); 3942 } 3943 3944 /* 3945 * Move the specified page to the tail of the inactive queue, or requeue 3946 * the page if it is already in the inactive queue. 3947 */ 3948 void 3949 vm_page_deactivate(vm_page_t m) 3950 { 3951 3952 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3953 return; 3954 vm_page_mvqueue(m, PQ_INACTIVE); 3955 } 3956 3957 /* 3958 * Move the specified page close to the head of the inactive queue, 3959 * bypassing LRU. A marker page is used to maintain FIFO ordering. 3960 * As with regular enqueues, we use a per-CPU batch queue to reduce 3961 * contention on the page queue lock. 3962 */ 3963 static void 3964 _vm_page_deactivate_noreuse(vm_page_t m) 3965 { 3966 3967 vm_page_assert_locked(m); 3968 3969 if (!vm_page_inactive(m)) { 3970 vm_page_dequeue(m); 3971 m->queue = PQ_INACTIVE; 3972 } 3973 if ((m->aflags & PGA_REQUEUE_HEAD) == 0) 3974 vm_page_aflag_set(m, PGA_REQUEUE_HEAD); 3975 vm_page_pqbatch_submit(m, PQ_INACTIVE); 3976 } 3977 3978 void 3979 vm_page_deactivate_noreuse(vm_page_t m) 3980 { 3981 3982 KASSERT(m->object != NULL, 3983 ("vm_page_deactivate_noreuse: page %p has no object", m)); 3984 3985 if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m)) 3986 _vm_page_deactivate_noreuse(m); 3987 } 3988 3989 /* 3990 * Put a page in the laundry, or requeue it if it is already there. 3991 */ 3992 void 3993 vm_page_launder(vm_page_t m) 3994 { 3995 3996 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 3997 return; 3998 vm_page_mvqueue(m, PQ_LAUNDRY); 3999 } 4000 4001 /* 4002 * Put a page in the PQ_UNSWAPPABLE holding queue. 4003 */ 4004 void 4005 vm_page_unswappable(vm_page_t m) 4006 { 4007 4008 vm_page_assert_locked(m); 4009 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 4010 ("page %p already unswappable", m)); 4011 4012 vm_page_dequeue(m); 4013 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4014 } 4015 4016 static void 4017 vm_page_release_toq(vm_page_t m, int flags) 4018 { 4019 4020 vm_page_assert_locked(m); 4021 4022 /* 4023 * Use a check of the valid bits to determine whether we should 4024 * accelerate reclamation of the page. The object lock might not be 4025 * held here, in which case the check is racy. At worst we will either 4026 * accelerate reclamation of a valid page and violate LRU, or 4027 * unnecessarily defer reclamation of an invalid page. 4028 * 4029 * If we were asked to not cache the page, place it near the head of the 4030 * inactive queue so that is reclaimed sooner. 4031 */ 4032 if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0) 4033 _vm_page_deactivate_noreuse(m); 4034 else if (vm_page_active(m)) 4035 vm_page_reference(m); 4036 else 4037 vm_page_mvqueue(m, PQ_INACTIVE); 4038 } 4039 4040 /* 4041 * Unwire a page and either attempt to free it or re-add it to the page queues. 4042 */ 4043 void 4044 vm_page_release(vm_page_t m, int flags) 4045 { 4046 vm_object_t object; 4047 u_int old; 4048 bool locked; 4049 4050 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4051 ("vm_page_release: page %p is unmanaged", m)); 4052 4053 if ((flags & VPR_TRYFREE) != 0) { 4054 for (;;) { 4055 object = (vm_object_t)atomic_load_ptr(&m->object); 4056 if (object == NULL) 4057 break; 4058 /* Depends on type-stability. */ 4059 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) { 4060 object = NULL; 4061 break; 4062 } 4063 if (object == m->object) 4064 break; 4065 VM_OBJECT_WUNLOCK(object); 4066 } 4067 if (__predict_true(object != NULL)) { 4068 vm_page_release_locked(m, flags); 4069 VM_OBJECT_WUNLOCK(object); 4070 return; 4071 } 4072 } 4073 4074 /* 4075 * Update LRU state before releasing the wiring reference. 4076 * Use a release store when updating the reference count to 4077 * synchronize with vm_page_free_prep(). 4078 */ 4079 old = m->ref_count; 4080 locked = false; 4081 do { 4082 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4083 ("vm_page_unwire: wire count underflow for page %p", m)); 4084 if (!locked && VPRC_WIRE_COUNT(old) == 1) { 4085 vm_page_lock(m); 4086 locked = true; 4087 vm_page_release_toq(m, flags); 4088 } 4089 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4090 4091 /* 4092 * Release the lock only after the wiring is released, to ensure that 4093 * the page daemon does not encounter and dequeue the page while it is 4094 * still wired. 4095 */ 4096 if (locked) 4097 vm_page_unlock(m); 4098 4099 if (VPRC_WIRE_COUNT(old) == 1) { 4100 vm_wire_sub(1); 4101 if (old == 1) 4102 vm_page_free(m); 4103 } 4104 } 4105 4106 /* See vm_page_release(). */ 4107 void 4108 vm_page_release_locked(vm_page_t m, int flags) 4109 { 4110 4111 VM_OBJECT_ASSERT_WLOCKED(m->object); 4112 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4113 ("vm_page_release_locked: page %p is unmanaged", m)); 4114 4115 if (vm_page_unwire_noq(m)) { 4116 if ((flags & VPR_TRYFREE) != 0 && 4117 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4118 m->dirty == 0 && !vm_page_busied(m)) { 4119 vm_page_free(m); 4120 } else { 4121 vm_page_lock(m); 4122 vm_page_release_toq(m, flags); 4123 vm_page_unlock(m); 4124 } 4125 } 4126 } 4127 4128 static bool 4129 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4130 { 4131 u_int old; 4132 4133 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4134 ("vm_page_try_blocked_op: page %p has no object", m)); 4135 KASSERT(vm_page_busied(m), 4136 ("vm_page_try_blocked_op: page %p is not busy", m)); 4137 VM_OBJECT_ASSERT_LOCKED(m->object); 4138 4139 old = m->ref_count; 4140 do { 4141 KASSERT(old != 0, 4142 ("vm_page_try_blocked_op: page %p has no references", m)); 4143 if (VPRC_WIRE_COUNT(old) != 0) 4144 return (false); 4145 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4146 4147 (op)(m); 4148 4149 /* 4150 * If the object is read-locked, new wirings may be created via an 4151 * object lookup. 4152 */ 4153 old = vm_page_drop(m, VPRC_BLOCKED); 4154 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4155 old == (VPRC_BLOCKED | VPRC_OBJREF), 4156 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4157 old, m)); 4158 return (true); 4159 } 4160 4161 /* 4162 * Atomically check for wirings and remove all mappings of the page. 4163 */ 4164 bool 4165 vm_page_try_remove_all(vm_page_t m) 4166 { 4167 4168 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4169 } 4170 4171 /* 4172 * Atomically check for wirings and remove all writeable mappings of the page. 4173 */ 4174 bool 4175 vm_page_try_remove_write(vm_page_t m) 4176 { 4177 4178 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4179 } 4180 4181 /* 4182 * vm_page_advise 4183 * 4184 * Apply the specified advice to the given page. 4185 * 4186 * The object and page must be locked. 4187 */ 4188 void 4189 vm_page_advise(vm_page_t m, int advice) 4190 { 4191 4192 vm_page_assert_locked(m); 4193 VM_OBJECT_ASSERT_WLOCKED(m->object); 4194 if (advice == MADV_FREE) 4195 /* 4196 * Mark the page clean. This will allow the page to be freed 4197 * without first paging it out. MADV_FREE pages are often 4198 * quickly reused by malloc(3), so we do not do anything that 4199 * would result in a page fault on a later access. 4200 */ 4201 vm_page_undirty(m); 4202 else if (advice != MADV_DONTNEED) { 4203 if (advice == MADV_WILLNEED) 4204 vm_page_activate(m); 4205 return; 4206 } 4207 4208 /* 4209 * Clear any references to the page. Otherwise, the page daemon will 4210 * immediately reactivate the page. 4211 */ 4212 vm_page_aflag_clear(m, PGA_REFERENCED); 4213 4214 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4215 vm_page_dirty(m); 4216 4217 /* 4218 * Place clean pages near the head of the inactive queue rather than 4219 * the tail, thus defeating the queue's LRU operation and ensuring that 4220 * the page will be reused quickly. Dirty pages not already in the 4221 * laundry are moved there. 4222 */ 4223 if (m->dirty == 0) 4224 vm_page_deactivate_noreuse(m); 4225 else if (!vm_page_in_laundry(m)) 4226 vm_page_launder(m); 4227 } 4228 4229 /* 4230 * Grab a page, waiting until we are waken up due to the page 4231 * changing state. We keep on waiting, if the page continues 4232 * to be in the object. If the page doesn't exist, first allocate it 4233 * and then conditionally zero it. 4234 * 4235 * This routine may sleep. 4236 * 4237 * The object must be locked on entry. The lock will, however, be released 4238 * and reacquired if the routine sleeps. 4239 */ 4240 vm_page_t 4241 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4242 { 4243 vm_page_t m; 4244 int sleep; 4245 int pflags; 4246 4247 VM_OBJECT_ASSERT_WLOCKED(object); 4248 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4249 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4250 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4251 pflags = allocflags & 4252 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4253 VM_ALLOC_NOBUSY); 4254 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4255 pflags |= VM_ALLOC_WAITFAIL; 4256 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4257 pflags |= VM_ALLOC_SBUSY; 4258 retrylookup: 4259 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4260 if ((allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) != 0) 4261 sleep = !vm_page_trysbusy(m); 4262 else 4263 sleep = !vm_page_tryxbusy(m); 4264 if (sleep) { 4265 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4266 return (NULL); 4267 /* 4268 * Reference the page before unlocking and 4269 * sleeping so that the page daemon is less 4270 * likely to reclaim it. 4271 */ 4272 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 4273 vm_page_aflag_set(m, PGA_REFERENCED); 4274 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 4275 VM_ALLOC_IGN_SBUSY) != 0); 4276 VM_OBJECT_WLOCK(object); 4277 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4278 return (NULL); 4279 goto retrylookup; 4280 } else { 4281 if ((allocflags & VM_ALLOC_WIRED) != 0) 4282 vm_page_wire(m); 4283 goto out; 4284 } 4285 } 4286 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4287 return (NULL); 4288 m = vm_page_alloc(object, pindex, pflags); 4289 if (m == NULL) { 4290 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4291 return (NULL); 4292 goto retrylookup; 4293 } 4294 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4295 pmap_zero_page(m); 4296 4297 out: 4298 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4299 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4300 vm_page_sunbusy(m); 4301 else 4302 vm_page_xunbusy(m); 4303 } 4304 return (m); 4305 } 4306 4307 /* 4308 * Grab a page and make it valid, paging in if necessary. Pages missing from 4309 * their pager are zero filled and validated. 4310 */ 4311 int 4312 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4313 { 4314 vm_page_t m; 4315 bool sleep, xbusy; 4316 int pflags; 4317 int rv; 4318 4319 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4320 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4321 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4322 KASSERT((allocflags & 4323 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4324 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4325 VM_OBJECT_ASSERT_WLOCKED(object); 4326 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY); 4327 pflags |= VM_ALLOC_WAITFAIL; 4328 4329 retrylookup: 4330 xbusy = false; 4331 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4332 /* 4333 * If the page is fully valid it can only become invalid 4334 * with the object lock held. If it is not valid it can 4335 * become valid with the busy lock held. Therefore, we 4336 * may unnecessarily lock the exclusive busy here if we 4337 * race with I/O completion not using the object lock. 4338 * However, we will not end up with an invalid page and a 4339 * shared lock. 4340 */ 4341 if (!vm_page_all_valid(m) || 4342 (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) { 4343 sleep = !vm_page_tryxbusy(m); 4344 xbusy = true; 4345 } else 4346 sleep = !vm_page_trysbusy(m); 4347 if (sleep) { 4348 /* 4349 * Reference the page before unlocking and 4350 * sleeping so that the page daemon is less 4351 * likely to reclaim it. 4352 */ 4353 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 4354 vm_page_aflag_set(m, PGA_REFERENCED); 4355 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 4356 VM_ALLOC_IGN_SBUSY) != 0); 4357 VM_OBJECT_WLOCK(object); 4358 goto retrylookup; 4359 } 4360 if ((allocflags & VM_ALLOC_NOCREAT) != 0 && 4361 !vm_page_all_valid(m)) { 4362 if (xbusy) 4363 vm_page_xunbusy(m); 4364 else 4365 vm_page_sunbusy(m); 4366 *mp = NULL; 4367 return (VM_PAGER_FAIL); 4368 } 4369 if ((allocflags & VM_ALLOC_WIRED) != 0) 4370 vm_page_wire(m); 4371 if (vm_page_all_valid(m)) 4372 goto out; 4373 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4374 *mp = NULL; 4375 return (VM_PAGER_FAIL); 4376 } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) { 4377 xbusy = true; 4378 } else { 4379 goto retrylookup; 4380 } 4381 4382 vm_page_assert_xbusied(m); 4383 MPASS(xbusy); 4384 if (vm_pager_has_page(object, pindex, NULL, NULL)) { 4385 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 4386 if (rv != VM_PAGER_OK) { 4387 if (allocflags & VM_ALLOC_WIRED) 4388 vm_page_unwire_noq(m); 4389 vm_page_free(m); 4390 *mp = NULL; 4391 return (rv); 4392 } 4393 MPASS(vm_page_all_valid(m)); 4394 } else { 4395 vm_page_zero_invalid(m, TRUE); 4396 } 4397 out: 4398 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4399 if (xbusy) 4400 vm_page_xunbusy(m); 4401 else 4402 vm_page_sunbusy(m); 4403 } 4404 if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy) 4405 vm_page_busy_downgrade(m); 4406 *mp = m; 4407 return (VM_PAGER_OK); 4408 } 4409 4410 /* 4411 * Return the specified range of pages from the given object. For each 4412 * page offset within the range, if a page already exists within the object 4413 * at that offset and it is busy, then wait for it to change state. If, 4414 * instead, the page doesn't exist, then allocate it. 4415 * 4416 * The caller must always specify an allocation class. 4417 * 4418 * allocation classes: 4419 * VM_ALLOC_NORMAL normal process request 4420 * VM_ALLOC_SYSTEM system *really* needs the pages 4421 * 4422 * The caller must always specify that the pages are to be busied and/or 4423 * wired. 4424 * 4425 * optional allocation flags: 4426 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 4427 * VM_ALLOC_NOBUSY do not exclusive busy the page 4428 * VM_ALLOC_NOWAIT do not sleep 4429 * VM_ALLOC_SBUSY set page to sbusy state 4430 * VM_ALLOC_WIRED wire the pages 4431 * VM_ALLOC_ZERO zero and validate any invalid pages 4432 * 4433 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 4434 * may return a partial prefix of the requested range. 4435 */ 4436 int 4437 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 4438 vm_page_t *ma, int count) 4439 { 4440 vm_page_t m, mpred; 4441 int pflags; 4442 int i; 4443 bool sleep; 4444 4445 VM_OBJECT_ASSERT_WLOCKED(object); 4446 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 4447 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 4448 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4449 (allocflags & VM_ALLOC_WIRED) != 0, 4450 ("vm_page_grab_pages: the pages must be busied or wired")); 4451 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4452 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4453 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); 4454 if (count == 0) 4455 return (0); 4456 pflags = allocflags & 4457 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4458 VM_ALLOC_NOBUSY); 4459 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4460 pflags |= VM_ALLOC_WAITFAIL; 4461 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4462 pflags |= VM_ALLOC_SBUSY; 4463 i = 0; 4464 retrylookup: 4465 m = vm_radix_lookup_le(&object->rtree, pindex + i); 4466 if (m == NULL || m->pindex != pindex + i) { 4467 mpred = m; 4468 m = NULL; 4469 } else 4470 mpred = TAILQ_PREV(m, pglist, listq); 4471 for (; i < count; i++) { 4472 if (m != NULL) { 4473 if ((allocflags & 4474 (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 4475 sleep = !vm_page_trysbusy(m); 4476 else 4477 sleep = !vm_page_tryxbusy(m); 4478 if (sleep) { 4479 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4480 break; 4481 /* 4482 * Reference the page before unlocking and 4483 * sleeping so that the page daemon is less 4484 * likely to reclaim it. 4485 */ 4486 if ((allocflags & VM_ALLOC_NOCREAT) == 0) 4487 vm_page_aflag_set(m, PGA_REFERENCED); 4488 vm_page_busy_sleep(m, "grbmaw", (allocflags & 4489 VM_ALLOC_IGN_SBUSY) != 0); 4490 VM_OBJECT_WLOCK(object); 4491 goto retrylookup; 4492 } 4493 if ((allocflags & VM_ALLOC_WIRED) != 0) 4494 vm_page_wire(m); 4495 } else { 4496 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4497 break; 4498 m = vm_page_alloc_after(object, pindex + i, 4499 pflags | VM_ALLOC_COUNT(count - i), mpred); 4500 if (m == NULL) { 4501 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4502 break; 4503 goto retrylookup; 4504 } 4505 } 4506 if (vm_page_none_valid(m) && 4507 (allocflags & VM_ALLOC_ZERO) != 0) { 4508 if ((m->flags & PG_ZERO) == 0) 4509 pmap_zero_page(m); 4510 vm_page_valid(m); 4511 } 4512 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4513 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4514 vm_page_sunbusy(m); 4515 else 4516 vm_page_xunbusy(m); 4517 } 4518 ma[i] = mpred = m; 4519 m = vm_page_next(m); 4520 } 4521 return (i); 4522 } 4523 4524 /* 4525 * Mapping function for valid or dirty bits in a page. 4526 * 4527 * Inputs are required to range within a page. 4528 */ 4529 vm_page_bits_t 4530 vm_page_bits(int base, int size) 4531 { 4532 int first_bit; 4533 int last_bit; 4534 4535 KASSERT( 4536 base + size <= PAGE_SIZE, 4537 ("vm_page_bits: illegal base/size %d/%d", base, size) 4538 ); 4539 4540 if (size == 0) /* handle degenerate case */ 4541 return (0); 4542 4543 first_bit = base >> DEV_BSHIFT; 4544 last_bit = (base + size - 1) >> DEV_BSHIFT; 4545 4546 return (((vm_page_bits_t)2 << last_bit) - 4547 ((vm_page_bits_t)1 << first_bit)); 4548 } 4549 4550 void 4551 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 4552 { 4553 4554 #if PAGE_SIZE == 32768 4555 atomic_set_64((uint64_t *)bits, set); 4556 #elif PAGE_SIZE == 16384 4557 atomic_set_32((uint32_t *)bits, set); 4558 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 4559 atomic_set_16((uint16_t *)bits, set); 4560 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 4561 atomic_set_8((uint8_t *)bits, set); 4562 #else /* PAGE_SIZE <= 8192 */ 4563 uintptr_t addr; 4564 int shift; 4565 4566 addr = (uintptr_t)bits; 4567 /* 4568 * Use a trick to perform a 32-bit atomic on the 4569 * containing aligned word, to not depend on the existence 4570 * of atomic_{set, clear}_{8, 16}. 4571 */ 4572 shift = addr & (sizeof(uint32_t) - 1); 4573 #if BYTE_ORDER == BIG_ENDIAN 4574 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4575 #else 4576 shift *= NBBY; 4577 #endif 4578 addr &= ~(sizeof(uint32_t) - 1); 4579 atomic_set_32((uint32_t *)addr, set << shift); 4580 #endif /* PAGE_SIZE */ 4581 } 4582 4583 static inline void 4584 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 4585 { 4586 4587 #if PAGE_SIZE == 32768 4588 atomic_clear_64((uint64_t *)bits, clear); 4589 #elif PAGE_SIZE == 16384 4590 atomic_clear_32((uint32_t *)bits, clear); 4591 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 4592 atomic_clear_16((uint16_t *)bits, clear); 4593 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 4594 atomic_clear_8((uint8_t *)bits, clear); 4595 #else /* PAGE_SIZE <= 8192 */ 4596 uintptr_t addr; 4597 int shift; 4598 4599 addr = (uintptr_t)bits; 4600 /* 4601 * Use a trick to perform a 32-bit atomic on the 4602 * containing aligned word, to not depend on the existence 4603 * of atomic_{set, clear}_{8, 16}. 4604 */ 4605 shift = addr & (sizeof(uint32_t) - 1); 4606 #if BYTE_ORDER == BIG_ENDIAN 4607 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 4608 #else 4609 shift *= NBBY; 4610 #endif 4611 addr &= ~(sizeof(uint32_t) - 1); 4612 atomic_clear_32((uint32_t *)addr, clear << shift); 4613 #endif /* PAGE_SIZE */ 4614 } 4615 4616 /* 4617 * vm_page_set_valid_range: 4618 * 4619 * Sets portions of a page valid. The arguments are expected 4620 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4621 * of any partial chunks touched by the range. The invalid portion of 4622 * such chunks will be zeroed. 4623 * 4624 * (base + size) must be less then or equal to PAGE_SIZE. 4625 */ 4626 void 4627 vm_page_set_valid_range(vm_page_t m, int base, int size) 4628 { 4629 int endoff, frag; 4630 vm_page_bits_t pagebits; 4631 4632 vm_page_assert_busied(m); 4633 if (size == 0) /* handle degenerate case */ 4634 return; 4635 4636 /* 4637 * If the base is not DEV_BSIZE aligned and the valid 4638 * bit is clear, we have to zero out a portion of the 4639 * first block. 4640 */ 4641 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4642 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 4643 pmap_zero_page_area(m, frag, base - frag); 4644 4645 /* 4646 * If the ending offset is not DEV_BSIZE aligned and the 4647 * valid bit is clear, we have to zero out a portion of 4648 * the last block. 4649 */ 4650 endoff = base + size; 4651 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4652 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 4653 pmap_zero_page_area(m, endoff, 4654 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4655 4656 /* 4657 * Assert that no previously invalid block that is now being validated 4658 * is already dirty. 4659 */ 4660 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 4661 ("vm_page_set_valid_range: page %p is dirty", m)); 4662 4663 /* 4664 * Set valid bits inclusive of any overlap. 4665 */ 4666 pagebits = vm_page_bits(base, size); 4667 if (vm_page_xbusied(m)) 4668 m->valid |= pagebits; 4669 else 4670 vm_page_bits_set(m, &m->valid, pagebits); 4671 } 4672 4673 /* 4674 * Clear the given bits from the specified page's dirty field. 4675 */ 4676 static __inline void 4677 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 4678 { 4679 4680 vm_page_assert_busied(m); 4681 4682 /* 4683 * If the page is xbusied and not write mapped we are the 4684 * only thread that can modify dirty bits. Otherwise, The pmap 4685 * layer can call vm_page_dirty() without holding a distinguished 4686 * lock. The combination of page busy and atomic operations 4687 * suffice to guarantee consistency of the page dirty field. 4688 */ 4689 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 4690 m->dirty &= ~pagebits; 4691 else 4692 vm_page_bits_clear(m, &m->dirty, pagebits); 4693 } 4694 4695 /* 4696 * vm_page_set_validclean: 4697 * 4698 * Sets portions of a page valid and clean. The arguments are expected 4699 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4700 * of any partial chunks touched by the range. The invalid portion of 4701 * such chunks will be zero'd. 4702 * 4703 * (base + size) must be less then or equal to PAGE_SIZE. 4704 */ 4705 void 4706 vm_page_set_validclean(vm_page_t m, int base, int size) 4707 { 4708 vm_page_bits_t oldvalid, pagebits; 4709 int endoff, frag; 4710 4711 vm_page_assert_busied(m); 4712 if (size == 0) /* handle degenerate case */ 4713 return; 4714 4715 /* 4716 * If the base is not DEV_BSIZE aligned and the valid 4717 * bit is clear, we have to zero out a portion of the 4718 * first block. 4719 */ 4720 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4721 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 4722 pmap_zero_page_area(m, frag, base - frag); 4723 4724 /* 4725 * If the ending offset is not DEV_BSIZE aligned and the 4726 * valid bit is clear, we have to zero out a portion of 4727 * the last block. 4728 */ 4729 endoff = base + size; 4730 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4731 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 4732 pmap_zero_page_area(m, endoff, 4733 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4734 4735 /* 4736 * Set valid, clear dirty bits. If validating the entire 4737 * page we can safely clear the pmap modify bit. We also 4738 * use this opportunity to clear the PGA_NOSYNC flag. If a process 4739 * takes a write fault on a MAP_NOSYNC memory area the flag will 4740 * be set again. 4741 * 4742 * We set valid bits inclusive of any overlap, but we can only 4743 * clear dirty bits for DEV_BSIZE chunks that are fully within 4744 * the range. 4745 */ 4746 oldvalid = m->valid; 4747 pagebits = vm_page_bits(base, size); 4748 if (vm_page_xbusied(m)) 4749 m->valid |= pagebits; 4750 else 4751 vm_page_bits_set(m, &m->valid, pagebits); 4752 #if 0 /* NOT YET */ 4753 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 4754 frag = DEV_BSIZE - frag; 4755 base += frag; 4756 size -= frag; 4757 if (size < 0) 4758 size = 0; 4759 } 4760 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 4761 #endif 4762 if (base == 0 && size == PAGE_SIZE) { 4763 /* 4764 * The page can only be modified within the pmap if it is 4765 * mapped, and it can only be mapped if it was previously 4766 * fully valid. 4767 */ 4768 if (oldvalid == VM_PAGE_BITS_ALL) 4769 /* 4770 * Perform the pmap_clear_modify() first. Otherwise, 4771 * a concurrent pmap operation, such as 4772 * pmap_protect(), could clear a modification in the 4773 * pmap and set the dirty field on the page before 4774 * pmap_clear_modify() had begun and after the dirty 4775 * field was cleared here. 4776 */ 4777 pmap_clear_modify(m); 4778 m->dirty = 0; 4779 vm_page_aflag_clear(m, PGA_NOSYNC); 4780 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 4781 m->dirty &= ~pagebits; 4782 else 4783 vm_page_clear_dirty_mask(m, pagebits); 4784 } 4785 4786 void 4787 vm_page_clear_dirty(vm_page_t m, int base, int size) 4788 { 4789 4790 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 4791 } 4792 4793 /* 4794 * vm_page_set_invalid: 4795 * 4796 * Invalidates DEV_BSIZE'd chunks within a page. Both the 4797 * valid and dirty bits for the effected areas are cleared. 4798 */ 4799 void 4800 vm_page_set_invalid(vm_page_t m, int base, int size) 4801 { 4802 vm_page_bits_t bits; 4803 vm_object_t object; 4804 4805 /* 4806 * The object lock is required so that pages can't be mapped 4807 * read-only while we're in the process of invalidating them. 4808 */ 4809 object = m->object; 4810 VM_OBJECT_ASSERT_WLOCKED(object); 4811 vm_page_assert_busied(m); 4812 4813 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 4814 size >= object->un_pager.vnp.vnp_size) 4815 bits = VM_PAGE_BITS_ALL; 4816 else 4817 bits = vm_page_bits(base, size); 4818 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 4819 pmap_remove_all(m); 4820 KASSERT((bits == 0 && vm_page_all_valid(m)) || 4821 !pmap_page_is_mapped(m), 4822 ("vm_page_set_invalid: page %p is mapped", m)); 4823 if (vm_page_xbusied(m)) { 4824 m->valid &= ~bits; 4825 m->dirty &= ~bits; 4826 } else { 4827 vm_page_bits_clear(m, &m->valid, bits); 4828 vm_page_bits_clear(m, &m->dirty, bits); 4829 } 4830 } 4831 4832 /* 4833 * vm_page_invalid: 4834 * 4835 * Invalidates the entire page. The page must be busy, unmapped, and 4836 * the enclosing object must be locked. The object locks protects 4837 * against concurrent read-only pmap enter which is done without 4838 * busy. 4839 */ 4840 void 4841 vm_page_invalid(vm_page_t m) 4842 { 4843 4844 vm_page_assert_busied(m); 4845 VM_OBJECT_ASSERT_LOCKED(m->object); 4846 MPASS(!pmap_page_is_mapped(m)); 4847 4848 if (vm_page_xbusied(m)) 4849 m->valid = 0; 4850 else 4851 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 4852 } 4853 4854 /* 4855 * vm_page_zero_invalid() 4856 * 4857 * The kernel assumes that the invalid portions of a page contain 4858 * garbage, but such pages can be mapped into memory by user code. 4859 * When this occurs, we must zero out the non-valid portions of the 4860 * page so user code sees what it expects. 4861 * 4862 * Pages are most often semi-valid when the end of a file is mapped 4863 * into memory and the file's size is not page aligned. 4864 */ 4865 void 4866 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 4867 { 4868 int b; 4869 int i; 4870 4871 /* 4872 * Scan the valid bits looking for invalid sections that 4873 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 4874 * valid bit may be set ) have already been zeroed by 4875 * vm_page_set_validclean(). 4876 */ 4877 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 4878 if (i == (PAGE_SIZE / DEV_BSIZE) || 4879 (m->valid & ((vm_page_bits_t)1 << i))) { 4880 if (i > b) { 4881 pmap_zero_page_area(m, 4882 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 4883 } 4884 b = i + 1; 4885 } 4886 } 4887 4888 /* 4889 * setvalid is TRUE when we can safely set the zero'd areas 4890 * as being valid. We can do this if there are no cache consistancy 4891 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 4892 */ 4893 if (setvalid) 4894 vm_page_valid(m); 4895 } 4896 4897 /* 4898 * vm_page_is_valid: 4899 * 4900 * Is (partial) page valid? Note that the case where size == 0 4901 * will return FALSE in the degenerate case where the page is 4902 * entirely invalid, and TRUE otherwise. 4903 * 4904 * Some callers envoke this routine without the busy lock held and 4905 * handle races via higher level locks. Typical callers should 4906 * hold a busy lock to prevent invalidation. 4907 */ 4908 int 4909 vm_page_is_valid(vm_page_t m, int base, int size) 4910 { 4911 vm_page_bits_t bits; 4912 4913 bits = vm_page_bits(base, size); 4914 return (m->valid != 0 && (m->valid & bits) == bits); 4915 } 4916 4917 /* 4918 * Returns true if all of the specified predicates are true for the entire 4919 * (super)page and false otherwise. 4920 */ 4921 bool 4922 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 4923 { 4924 vm_object_t object; 4925 int i, npages; 4926 4927 object = m->object; 4928 if (skip_m != NULL && skip_m->object != object) 4929 return (false); 4930 VM_OBJECT_ASSERT_LOCKED(object); 4931 npages = atop(pagesizes[m->psind]); 4932 4933 /* 4934 * The physically contiguous pages that make up a superpage, i.e., a 4935 * page with a page size index ("psind") greater than zero, will 4936 * occupy adjacent entries in vm_page_array[]. 4937 */ 4938 for (i = 0; i < npages; i++) { 4939 /* Always test object consistency, including "skip_m". */ 4940 if (m[i].object != object) 4941 return (false); 4942 if (&m[i] == skip_m) 4943 continue; 4944 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 4945 return (false); 4946 if ((flags & PS_ALL_DIRTY) != 0) { 4947 /* 4948 * Calling vm_page_test_dirty() or pmap_is_modified() 4949 * might stop this case from spuriously returning 4950 * "false". However, that would require a write lock 4951 * on the object containing "m[i]". 4952 */ 4953 if (m[i].dirty != VM_PAGE_BITS_ALL) 4954 return (false); 4955 } 4956 if ((flags & PS_ALL_VALID) != 0 && 4957 m[i].valid != VM_PAGE_BITS_ALL) 4958 return (false); 4959 } 4960 return (true); 4961 } 4962 4963 /* 4964 * Set the page's dirty bits if the page is modified. 4965 */ 4966 void 4967 vm_page_test_dirty(vm_page_t m) 4968 { 4969 4970 vm_page_assert_busied(m); 4971 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 4972 vm_page_dirty(m); 4973 } 4974 4975 void 4976 vm_page_valid(vm_page_t m) 4977 { 4978 4979 vm_page_assert_busied(m); 4980 if (vm_page_xbusied(m)) 4981 m->valid = VM_PAGE_BITS_ALL; 4982 else 4983 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 4984 } 4985 4986 void 4987 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 4988 { 4989 4990 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 4991 } 4992 4993 void 4994 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 4995 { 4996 4997 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 4998 } 4999 5000 int 5001 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5002 { 5003 5004 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5005 } 5006 5007 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5008 void 5009 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5010 { 5011 5012 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5013 } 5014 5015 void 5016 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5017 { 5018 5019 mtx_assert_(vm_page_lockptr(m), a, file, line); 5020 } 5021 #endif 5022 5023 #ifdef INVARIANTS 5024 void 5025 vm_page_object_busy_assert(vm_page_t m) 5026 { 5027 5028 /* 5029 * Certain of the page's fields may only be modified by the 5030 * holder of a page or object busy. 5031 */ 5032 if (m->object != NULL && !vm_page_busied(m)) 5033 VM_OBJECT_ASSERT_BUSY(m->object); 5034 } 5035 5036 void 5037 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5038 { 5039 5040 if ((bits & PGA_WRITEABLE) == 0) 5041 return; 5042 5043 /* 5044 * The PGA_WRITEABLE flag can only be set if the page is 5045 * managed, is exclusively busied or the object is locked. 5046 * Currently, this flag is only set by pmap_enter(). 5047 */ 5048 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5049 ("PGA_WRITEABLE on unmanaged page")); 5050 if (!vm_page_xbusied(m)) 5051 VM_OBJECT_ASSERT_BUSY(m->object); 5052 } 5053 #endif 5054 5055 #include "opt_ddb.h" 5056 #ifdef DDB 5057 #include <sys/kernel.h> 5058 5059 #include <ddb/ddb.h> 5060 5061 DB_SHOW_COMMAND(page, vm_page_print_page_info) 5062 { 5063 5064 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5065 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5066 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5067 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5068 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5069 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5070 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5071 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5072 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5073 } 5074 5075 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 5076 { 5077 int dom; 5078 5079 db_printf("pq_free %d\n", vm_free_count()); 5080 for (dom = 0; dom < vm_ndomains; dom++) { 5081 db_printf( 5082 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5083 dom, 5084 vm_dom[dom].vmd_page_count, 5085 vm_dom[dom].vmd_free_count, 5086 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5087 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5088 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5089 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5090 } 5091 } 5092 5093 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5094 { 5095 vm_page_t m; 5096 boolean_t phys, virt; 5097 5098 if (!have_addr) { 5099 db_printf("show pginfo addr\n"); 5100 return; 5101 } 5102 5103 phys = strchr(modif, 'p') != NULL; 5104 virt = strchr(modif, 'v') != NULL; 5105 if (virt) 5106 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5107 else if (phys) 5108 m = PHYS_TO_VM_PAGE(addr); 5109 else 5110 m = (vm_page_t)addr; 5111 db_printf( 5112 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" 5113 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5114 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5115 m->queue, m->ref_count, m->aflags, m->oflags, 5116 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 5117 } 5118 #endif /* DDB */ 5119