1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * The Mach Operating System project at Carnegie-Mellon University. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 36 /*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Resident memory management module. 65 */ 66 67 #include <sys/cdefs.h> 68 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/counter.h> 73 #include <sys/domainset.h> 74 #include <sys/kernel.h> 75 #include <sys/limits.h> 76 #include <sys/linker.h> 77 #include <sys/lock.h> 78 #include <sys/malloc.h> 79 #include <sys/mman.h> 80 #include <sys/msgbuf.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> 83 #include <sys/rwlock.h> 84 #include <sys/sleepqueue.h> 85 #include <sys/sbuf.h> 86 #include <sys/sched.h> 87 #include <sys/sf_buf.h> 88 #include <sys/smp.h> 89 #include <sys/sysctl.h> 90 #include <sys/vmmeter.h> 91 #include <sys/vnode.h> 92 93 #include <vm/vm.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_param.h> 96 #include <vm/vm_domainset.h> 97 #include <vm/vm_kern.h> 98 #include <vm/vm_map.h> 99 #include <vm/vm_object.h> 100 #include <vm/vm_page.h> 101 #include <vm/vm_pageout.h> 102 #include <vm/vm_phys.h> 103 #include <vm/vm_pagequeue.h> 104 #include <vm/vm_pager.h> 105 #include <vm/vm_radix.h> 106 #include <vm/vm_reserv.h> 107 #include <vm/vm_extern.h> 108 #include <vm/vm_dumpset.h> 109 #include <vm/uma.h> 110 #include <vm/uma_int.h> 111 112 #include <machine/md_var.h> 113 114 struct vm_domain vm_dom[MAXMEMDOM]; 115 116 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 117 118 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 119 /* The following fields are protected by the domainset lock. */ 120 domainset_t __exclusive_cache_line vm_min_domains; 121 domainset_t __exclusive_cache_line vm_severe_domains; 122 static int vm_min_waiters; 123 static int vm_severe_waiters; 124 static int vm_pageproc_waiters; 125 126 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 127 "VM page statistics"); 128 129 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 130 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 131 CTLFLAG_RD, &pqstate_commit_retries, 132 "Number of failed per-page atomic queue state updates"); 133 134 static COUNTER_U64_DEFINE_EARLY(queue_ops); 135 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 136 CTLFLAG_RD, &queue_ops, 137 "Number of batched queue operations"); 138 139 static COUNTER_U64_DEFINE_EARLY(queue_nops); 140 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 141 CTLFLAG_RD, &queue_nops, 142 "Number of batched queue operations with no effects"); 143 144 static unsigned long nofreeq_size; 145 SYSCTL_ULONG(_vm_stats_page, OID_AUTO, nofreeq_size, CTLFLAG_RD, 146 &nofreeq_size, 0, 147 "Size of the nofree queue"); 148 149 #ifdef INVARIANTS 150 bool vm_check_pg_zero = false; 151 SYSCTL_BOOL(_debug, OID_AUTO, vm_check_pg_zero, CTLFLAG_RWTUN, 152 &vm_check_pg_zero, 0, 153 "verify content of freed zero-filled pages"); 154 #endif 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 struct bitset *vm_page_dump; 167 long vm_page_dump_pages; 168 169 static TAILQ_HEAD(, vm_page) blacklist_head; 170 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 171 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 172 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 173 174 static uma_zone_t fakepg_zone; 175 176 static void vm_page_alloc_check(vm_page_t m); 177 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req); 178 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 179 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 180 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 181 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 182 static bool vm_page_free_prep(vm_page_t m); 183 static void vm_page_free_toq(vm_page_t m); 184 static void vm_page_init(void *dummy); 185 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object); 186 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 187 const uint16_t nflag); 188 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 189 vm_page_t m_run, vm_paddr_t high); 190 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 191 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 192 int req); 193 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 194 int flags); 195 static void vm_page_zone_release(void *arg, void **store, int cnt); 196 197 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 198 199 static void 200 vm_page_init(void *dummy) 201 { 202 203 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 204 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 205 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE); 206 } 207 208 static int pgcache_zone_max_pcpu; 209 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu, 210 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0, 211 "Per-CPU page cache size"); 212 213 /* 214 * The cache page zone is initialized later since we need to be able to allocate 215 * pages before UMA is fully initialized. 216 */ 217 static void 218 vm_page_init_cache_zones(void *dummy __unused) 219 { 220 struct vm_domain *vmd; 221 struct vm_pgcache *pgcache; 222 int cache, domain, maxcache, pool; 223 224 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu); 225 maxcache = pgcache_zone_max_pcpu * mp_ncpus; 226 for (domain = 0; domain < vm_ndomains; domain++) { 227 vmd = VM_DOMAIN(domain); 228 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 229 #ifdef VM_FREEPOOL_LAZYINIT 230 if (pool == VM_FREEPOOL_LAZYINIT) 231 continue; 232 #endif 233 pgcache = &vmd->vmd_pgcache[pool]; 234 pgcache->domain = domain; 235 pgcache->pool = pool; 236 pgcache->zone = uma_zcache_create("vm pgcache", 237 PAGE_SIZE, NULL, NULL, NULL, NULL, 238 vm_page_zone_import, vm_page_zone_release, pgcache, 239 UMA_ZONE_VM); 240 241 /* 242 * Limit each pool's zone to 0.1% of the pages in the 243 * domain. 244 */ 245 cache = maxcache != 0 ? maxcache : 246 vmd->vmd_page_count / 1000; 247 uma_zone_set_maxcache(pgcache->zone, cache); 248 } 249 } 250 } 251 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 252 253 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 254 #if PAGE_SIZE == 32768 255 #ifdef CTASSERT 256 CTASSERT(sizeof(u_long) >= 8); 257 #endif 258 #endif 259 260 /* 261 * vm_set_page_size: 262 * 263 * Sets the page size, perhaps based upon the memory 264 * size. Must be called before any use of page-size 265 * dependent functions. 266 */ 267 void 268 vm_set_page_size(void) 269 { 270 if (vm_cnt.v_page_size == 0) 271 vm_cnt.v_page_size = PAGE_SIZE; 272 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 273 panic("vm_set_page_size: page size not a power of two"); 274 } 275 276 /* 277 * vm_page_blacklist_next: 278 * 279 * Find the next entry in the provided string of blacklist 280 * addresses. Entries are separated by space, comma, or newline. 281 * If an invalid integer is encountered then the rest of the 282 * string is skipped. Updates the list pointer to the next 283 * character, or NULL if the string is exhausted or invalid. 284 */ 285 static vm_paddr_t 286 vm_page_blacklist_next(char **list, char *end) 287 { 288 vm_paddr_t bad; 289 char *cp, *pos; 290 291 if (list == NULL || *list == NULL) 292 return (0); 293 if (**list =='\0') { 294 *list = NULL; 295 return (0); 296 } 297 298 /* 299 * If there's no end pointer then the buffer is coming from 300 * the kenv and we know it's null-terminated. 301 */ 302 if (end == NULL) 303 end = *list + strlen(*list); 304 305 /* Ensure that strtoq() won't walk off the end */ 306 if (*end != '\0') { 307 if (*end == '\n' || *end == ' ' || *end == ',') 308 *end = '\0'; 309 else { 310 printf("Blacklist not terminated, skipping\n"); 311 *list = NULL; 312 return (0); 313 } 314 } 315 316 for (pos = *list; *pos != '\0'; pos = cp) { 317 bad = strtoq(pos, &cp, 0); 318 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 319 if (bad == 0) { 320 if (++cp < end) 321 continue; 322 else 323 break; 324 } 325 } else 326 break; 327 if (*cp == '\0' || ++cp >= end) 328 *list = NULL; 329 else 330 *list = cp; 331 return (trunc_page(bad)); 332 } 333 printf("Garbage in RAM blacklist, skipping\n"); 334 *list = NULL; 335 return (0); 336 } 337 338 bool 339 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 340 { 341 struct vm_domain *vmd; 342 vm_page_t m; 343 bool found; 344 345 m = vm_phys_paddr_to_vm_page(pa); 346 if (m == NULL) 347 return (true); /* page does not exist, no failure */ 348 349 vmd = VM_DOMAIN(vm_phys_domain(pa)); 350 vm_domain_free_lock(vmd); 351 found = vm_phys_unfree_page(pa); 352 vm_domain_free_unlock(vmd); 353 if (found) { 354 vm_domain_freecnt_inc(vmd, -1); 355 TAILQ_INSERT_TAIL(&blacklist_head, m, plinks.q); 356 if (verbose) 357 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 358 } 359 return (found); 360 } 361 362 /* 363 * vm_page_blacklist_check: 364 * 365 * Iterate through the provided string of blacklist addresses, pulling 366 * each entry out of the physical allocator free list and putting it 367 * onto a list for reporting via the vm.page_blacklist sysctl. 368 */ 369 static void 370 vm_page_blacklist_check(char *list, char *end) 371 { 372 vm_paddr_t pa; 373 char *next; 374 375 next = list; 376 while (next != NULL) { 377 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 378 continue; 379 vm_page_blacklist_add(pa, bootverbose); 380 } 381 } 382 383 /* 384 * vm_page_blacklist_load: 385 * 386 * Search for a special module named "ram_blacklist". It'll be a 387 * plain text file provided by the user via the loader directive 388 * of the same name. 389 */ 390 static void 391 vm_page_blacklist_load(char **list, char **end) 392 { 393 void *mod; 394 u_char *ptr; 395 u_int len; 396 397 mod = NULL; 398 ptr = NULL; 399 400 mod = preload_search_by_type("ram_blacklist"); 401 if (mod != NULL) { 402 ptr = preload_fetch_addr(mod); 403 len = preload_fetch_size(mod); 404 } 405 406 if (ptr != NULL && len > 0) { 407 *list = ptr; 408 *end = ptr + len - 1; 409 } else { 410 *list = NULL; 411 *end = NULL; 412 } 413 414 return; 415 } 416 417 static int 418 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 419 { 420 vm_page_t m; 421 struct sbuf sbuf; 422 int error, first; 423 424 first = 1; 425 error = sysctl_wire_old_buffer(req, 0); 426 if (error != 0) 427 return (error); 428 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 429 TAILQ_FOREACH(m, &blacklist_head, plinks.q) { 430 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 431 (uintmax_t)m->phys_addr); 432 first = 0; 433 } 434 error = sbuf_finish(&sbuf); 435 sbuf_delete(&sbuf); 436 return (error); 437 } 438 439 /* 440 * Initialize a dummy page for use in scans of the specified paging queue. 441 * In principle, this function only needs to set the flag PG_MARKER. 442 * Nonetheless, it write busies the page as a safety precaution. 443 */ 444 void 445 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 446 { 447 448 bzero(marker, sizeof(*marker)); 449 marker->flags = PG_MARKER; 450 marker->a.flags = aflags; 451 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 452 marker->a.queue = queue; 453 } 454 455 static void 456 vm_page_domain_init(int domain) 457 { 458 struct vm_domain *vmd; 459 struct vm_pagequeue *pq; 460 int i; 461 462 vmd = VM_DOMAIN(domain); 463 bzero(vmd, sizeof(*vmd)); 464 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 465 "vm inactive pagequeue"; 466 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 467 "vm active pagequeue"; 468 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 469 "vm laundry pagequeue"; 470 *__DECONST(const char **, 471 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 472 "vm unswappable pagequeue"; 473 vmd->vmd_domain = domain; 474 vmd->vmd_page_count = 0; 475 vmd->vmd_free_count = 0; 476 vmd->vmd_segs = 0; 477 vmd->vmd_oom = false; 478 vmd->vmd_helper_threads_enabled = true; 479 for (i = 0; i < PQ_COUNT; i++) { 480 pq = &vmd->vmd_pagequeues[i]; 481 TAILQ_INIT(&pq->pq_pl); 482 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 483 MTX_DEF | MTX_DUPOK); 484 pq->pq_pdpages = 0; 485 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 486 } 487 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 488 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 489 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 490 491 /* 492 * inacthead is used to provide FIFO ordering for LRU-bypassing 493 * insertions. 494 */ 495 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 496 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 497 &vmd->vmd_inacthead, plinks.q); 498 499 /* 500 * The clock pages are used to implement active queue scanning without 501 * requeues. Scans start at clock[0], which is advanced after the scan 502 * ends. When the two clock hands meet, they are reset and scanning 503 * resumes from the head of the queue. 504 */ 505 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 506 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 507 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 508 &vmd->vmd_clock[0], plinks.q); 509 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 510 &vmd->vmd_clock[1], plinks.q); 511 } 512 513 /* 514 * Initialize a physical page in preparation for adding it to the free 515 * lists. 516 */ 517 void 518 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool) 519 { 520 m->object = NULL; 521 m->ref_count = 0; 522 m->busy_lock = VPB_FREED; 523 m->flags = m->a.flags = 0; 524 m->phys_addr = pa; 525 m->a.queue = PQ_NONE; 526 m->psind = 0; 527 m->segind = segind; 528 m->order = VM_NFREEORDER; 529 m->pool = pool; 530 m->valid = m->dirty = 0; 531 pmap_page_init(m); 532 } 533 534 #ifndef PMAP_HAS_PAGE_ARRAY 535 static vm_paddr_t 536 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 537 { 538 vm_paddr_t new_end; 539 540 /* 541 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 542 * However, because this page is allocated from KVM, out-of-bounds 543 * accesses using the direct map will not be trapped. 544 */ 545 *vaddr += PAGE_SIZE; 546 547 /* 548 * Allocate physical memory for the page structures, and map it. 549 */ 550 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 551 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 552 VM_PROT_READ | VM_PROT_WRITE); 553 vm_page_array_size = page_range; 554 555 return (new_end); 556 } 557 #endif 558 559 /* 560 * vm_page_startup: 561 * 562 * Initializes the resident memory module. Allocates physical memory for 563 * bootstrapping UMA and some data structures that are used to manage 564 * physical pages. Initializes these structures, and populates the free 565 * page queues. 566 */ 567 vm_offset_t 568 vm_page_startup(vm_offset_t vaddr) 569 { 570 struct vm_phys_seg *seg; 571 struct vm_domain *vmd; 572 vm_page_t m; 573 char *list, *listend; 574 vm_paddr_t end, high_avail, low_avail, new_end, size; 575 vm_paddr_t page_range __unused; 576 vm_paddr_t last_pa, pa, startp, endp; 577 u_long pagecount; 578 #if MINIDUMP_PAGE_TRACKING 579 u_long vm_page_dump_size; 580 #endif 581 int biggestone, i, segind; 582 #ifdef WITNESS 583 vm_offset_t mapped; 584 int witness_size; 585 #endif 586 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 587 long ii; 588 #endif 589 int pool; 590 #ifdef VM_FREEPOOL_LAZYINIT 591 int lazyinit; 592 #endif 593 594 vaddr = round_page(vaddr); 595 596 vm_phys_early_startup(); 597 biggestone = vm_phys_avail_largest(); 598 end = phys_avail[biggestone+1]; 599 600 /* 601 * Initialize the page and queue locks. 602 */ 603 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 604 for (i = 0; i < vm_ndomains; i++) 605 vm_page_domain_init(i); 606 607 new_end = end; 608 #ifdef WITNESS 609 witness_size = round_page(witness_startup_count()); 610 new_end -= witness_size; 611 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 612 VM_PROT_READ | VM_PROT_WRITE); 613 bzero((void *)mapped, witness_size); 614 witness_startup((void *)mapped); 615 #endif 616 617 #if MINIDUMP_PAGE_TRACKING 618 /* 619 * Allocate a bitmap to indicate that a random physical page 620 * needs to be included in a minidump. 621 * 622 * The amd64 port needs this to indicate which direct map pages 623 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 624 * 625 * However, i386 still needs this workspace internally within the 626 * minidump code. In theory, they are not needed on i386, but are 627 * included should the sf_buf code decide to use them. 628 */ 629 last_pa = 0; 630 vm_page_dump_pages = 0; 631 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 632 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 633 dump_avail[i] / PAGE_SIZE; 634 if (dump_avail[i + 1] > last_pa) 635 last_pa = dump_avail[i + 1]; 636 } 637 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 638 new_end -= vm_page_dump_size; 639 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 640 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 641 bzero((void *)vm_page_dump, vm_page_dump_size); 642 #if MINIDUMP_STARTUP_PAGE_TRACKING 643 /* 644 * Include the UMA bootstrap pages, witness pages and vm_page_dump 645 * in a crash dump. When pmap_map() uses the direct map, they are 646 * not automatically included. 647 */ 648 for (pa = new_end; pa < end; pa += PAGE_SIZE) 649 dump_add_page(pa); 650 #endif 651 #else 652 (void)last_pa; 653 #endif 654 phys_avail[biggestone + 1] = new_end; 655 #ifdef __amd64__ 656 /* 657 * Request that the physical pages underlying the message buffer be 658 * included in a crash dump. Since the message buffer is accessed 659 * through the direct map, they are not automatically included. 660 */ 661 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 662 last_pa = pa + round_page(msgbufsize); 663 while (pa < last_pa) { 664 dump_add_page(pa); 665 pa += PAGE_SIZE; 666 } 667 #else 668 (void)pa; 669 #endif 670 671 /* 672 * Determine the lowest and highest physical addresses and, in the case 673 * of VM_PHYSSEG_SPARSE, the exact size of the available physical 674 * memory. vm_phys_early_startup() already checked that phys_avail[] 675 * has at least one element. 676 */ 677 #ifdef VM_PHYSSEG_SPARSE 678 size = phys_avail[1] - phys_avail[0]; 679 #endif 680 low_avail = phys_avail[0]; 681 high_avail = phys_avail[1]; 682 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 683 #ifdef VM_PHYSSEG_SPARSE 684 size += phys_avail[i + 1] - phys_avail[i]; 685 #endif 686 if (phys_avail[i] < low_avail) 687 low_avail = phys_avail[i]; 688 if (phys_avail[i + 1] > high_avail) 689 high_avail = phys_avail[i + 1]; 690 } 691 for (i = 0; i < vm_phys_nsegs; i++) { 692 #ifdef VM_PHYSSEG_SPARSE 693 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 694 #endif 695 if (vm_phys_segs[i].start < low_avail) 696 low_avail = vm_phys_segs[i].start; 697 if (vm_phys_segs[i].end > high_avail) 698 high_avail = vm_phys_segs[i].end; 699 } 700 first_page = low_avail / PAGE_SIZE; 701 #ifdef VM_PHYSSEG_DENSE 702 size = high_avail - low_avail; 703 #endif 704 705 #ifdef PMAP_HAS_PAGE_ARRAY 706 pmap_page_array_startup(size / PAGE_SIZE); 707 biggestone = vm_phys_avail_largest(); 708 end = new_end = phys_avail[biggestone + 1]; 709 #else 710 #ifdef VM_PHYSSEG_DENSE 711 /* 712 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 713 * the overhead of a page structure per page only if vm_page_array is 714 * allocated from the last physical memory chunk. Otherwise, we must 715 * allocate page structures representing the physical memory 716 * underlying vm_page_array, even though they will not be used. 717 */ 718 if (new_end != high_avail) 719 page_range = size / PAGE_SIZE; 720 else 721 #endif 722 { 723 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 724 725 /* 726 * If the partial bytes remaining are large enough for 727 * a page (PAGE_SIZE) without a corresponding 728 * 'struct vm_page', then new_end will contain an 729 * extra page after subtracting the length of the VM 730 * page array. Compensate by subtracting an extra 731 * page from new_end. 732 */ 733 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 734 if (new_end == high_avail) 735 high_avail -= PAGE_SIZE; 736 new_end -= PAGE_SIZE; 737 } 738 } 739 end = new_end; 740 new_end = vm_page_array_alloc(&vaddr, end, page_range); 741 #endif 742 743 #if VM_NRESERVLEVEL > 0 744 /* 745 * Allocate physical memory for the reservation management system's 746 * data structures, and map it. 747 */ 748 new_end = vm_reserv_startup(&vaddr, new_end); 749 #endif 750 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 751 /* 752 * Include vm_page_array and vm_reserv_array in a crash dump. 753 */ 754 for (pa = new_end; pa < end; pa += PAGE_SIZE) 755 dump_add_page(pa); 756 #endif 757 phys_avail[biggestone + 1] = new_end; 758 759 /* 760 * Add physical memory segments corresponding to the available 761 * physical pages. 762 */ 763 for (i = 0; phys_avail[i + 1] != 0; i += 2) 764 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 765 766 /* 767 * Initialize the physical memory allocator. 768 */ 769 vm_phys_init(); 770 771 pool = VM_FREEPOOL_DEFAULT; 772 #ifdef VM_FREEPOOL_LAZYINIT 773 lazyinit = 1; 774 TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit); 775 if (lazyinit) 776 pool = VM_FREEPOOL_LAZYINIT; 777 #endif 778 779 /* 780 * Initialize the page structures and add every available page to the 781 * physical memory allocator's free lists. 782 */ 783 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 784 for (ii = 0; ii < vm_page_array_size; ii++) { 785 m = &vm_page_array[ii]; 786 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0, 787 VM_FREEPOOL_DEFAULT); 788 m->flags = PG_FICTITIOUS; 789 } 790 #endif 791 vm_cnt.v_page_count = 0; 792 for (segind = 0; segind < vm_phys_nsegs; segind++) { 793 seg = &vm_phys_segs[segind]; 794 795 /* 796 * Initialize pages not covered by phys_avail[], since they 797 * might be freed to the allocator at some future point, e.g., 798 * by kmem_bootstrap_free(). 799 */ 800 startp = seg->start; 801 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 802 if (startp >= seg->end) 803 break; 804 if (phys_avail[i + 1] < startp) 805 continue; 806 if (phys_avail[i] <= startp) { 807 startp = phys_avail[i + 1]; 808 continue; 809 } 810 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 811 for (endp = MIN(phys_avail[i], seg->end); 812 startp < endp; startp += PAGE_SIZE, m++) { 813 vm_page_init_page(m, startp, segind, 814 VM_FREEPOOL_DEFAULT); 815 } 816 } 817 818 /* 819 * Add the segment's pages that are covered by one of 820 * phys_avail's ranges to the free lists. 821 */ 822 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 823 if (seg->end <= phys_avail[i] || 824 seg->start >= phys_avail[i + 1]) 825 continue; 826 827 startp = MAX(seg->start, phys_avail[i]); 828 endp = MIN(seg->end, phys_avail[i + 1]); 829 pagecount = (u_long)atop(endp - startp); 830 if (pagecount == 0) 831 continue; 832 833 /* 834 * If lazy vm_page initialization is not enabled, simply 835 * initialize all of the pages in the segment covered by 836 * phys_avail. Otherwise, initialize only the first 837 * page of each run of free pages handed to the vm_phys 838 * allocator, which in turn defers initialization of 839 * pages until they are needed. 840 * 841 * This avoids blocking the boot process for long 842 * periods, which may be relevant for VMs (which ought 843 * to boot as quickly as possible) and/or systems with 844 * large amounts of physical memory. 845 */ 846 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 847 vm_page_init_page(m, startp, segind, pool); 848 if (pool == VM_FREEPOOL_DEFAULT) { 849 for (u_long j = 1; j < pagecount; j++) { 850 vm_page_init_page(&m[j], 851 startp + ptoa((vm_paddr_t)j), 852 segind, pool); 853 } 854 } 855 vmd = VM_DOMAIN(seg->domain); 856 vm_domain_free_lock(vmd); 857 vm_phys_enqueue_contig(m, pool, pagecount); 858 vm_domain_free_unlock(vmd); 859 vm_domain_freecnt_inc(vmd, pagecount); 860 vm_cnt.v_page_count += (u_int)pagecount; 861 vmd->vmd_page_count += (u_int)pagecount; 862 vmd->vmd_segs |= 1UL << segind; 863 } 864 } 865 866 /* 867 * Remove blacklisted pages from the physical memory allocator. 868 */ 869 TAILQ_INIT(&blacklist_head); 870 vm_page_blacklist_load(&list, &listend); 871 vm_page_blacklist_check(list, listend); 872 873 list = kern_getenv("vm.blacklist"); 874 vm_page_blacklist_check(list, NULL); 875 876 freeenv(list); 877 #if VM_NRESERVLEVEL > 0 878 /* 879 * Initialize the reservation management system. 880 */ 881 vm_reserv_init(); 882 #endif 883 884 return (vaddr); 885 } 886 887 void 888 vm_page_reference(vm_page_t m) 889 { 890 891 vm_page_aflag_set(m, PGA_REFERENCED); 892 } 893 894 /* 895 * vm_page_trybusy 896 * 897 * Helper routine for grab functions to trylock busy. 898 * 899 * Returns true on success and false on failure. 900 */ 901 static bool 902 vm_page_trybusy(vm_page_t m, int allocflags) 903 { 904 905 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 906 return (vm_page_trysbusy(m)); 907 else 908 return (vm_page_tryxbusy(m)); 909 } 910 911 /* 912 * vm_page_tryacquire 913 * 914 * Helper routine for grab functions to trylock busy and wire. 915 * 916 * Returns true on success and false on failure. 917 */ 918 static inline bool 919 vm_page_tryacquire(vm_page_t m, int allocflags) 920 { 921 bool locked; 922 923 locked = vm_page_trybusy(m, allocflags); 924 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 925 vm_page_wire(m); 926 return (locked); 927 } 928 929 /* 930 * vm_page_busy_acquire: 931 * 932 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 933 * and drop the object lock if necessary. 934 */ 935 bool 936 vm_page_busy_acquire(vm_page_t m, int allocflags) 937 { 938 vm_object_t obj; 939 bool locked; 940 941 /* 942 * The page-specific object must be cached because page 943 * identity can change during the sleep, causing the 944 * re-lock of a different object. 945 * It is assumed that a reference to the object is already 946 * held by the callers. 947 */ 948 obj = atomic_load_ptr(&m->object); 949 for (;;) { 950 if (vm_page_tryacquire(m, allocflags)) 951 return (true); 952 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 953 return (false); 954 if (obj != NULL) 955 locked = VM_OBJECT_WOWNED(obj); 956 else 957 locked = false; 958 MPASS(locked || vm_page_wired(m)); 959 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 960 locked) && locked) 961 VM_OBJECT_WLOCK(obj); 962 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 963 return (false); 964 KASSERT(m->object == obj || m->object == NULL, 965 ("vm_page_busy_acquire: page %p does not belong to %p", 966 m, obj)); 967 } 968 } 969 970 /* 971 * vm_page_busy_downgrade: 972 * 973 * Downgrade an exclusive busy page into a single shared busy page. 974 */ 975 void 976 vm_page_busy_downgrade(vm_page_t m) 977 { 978 u_int x; 979 980 vm_page_assert_xbusied(m); 981 982 x = vm_page_busy_fetch(m); 983 for (;;) { 984 if (atomic_fcmpset_rel_int(&m->busy_lock, 985 &x, VPB_SHARERS_WORD(1))) 986 break; 987 } 988 if ((x & VPB_BIT_WAITERS) != 0) 989 wakeup(m); 990 } 991 992 /* 993 * 994 * vm_page_busy_tryupgrade: 995 * 996 * Attempt to upgrade a single shared busy into an exclusive busy. 997 */ 998 int 999 vm_page_busy_tryupgrade(vm_page_t m) 1000 { 1001 u_int ce, x; 1002 1003 vm_page_assert_sbusied(m); 1004 1005 x = vm_page_busy_fetch(m); 1006 ce = VPB_CURTHREAD_EXCLUSIVE; 1007 for (;;) { 1008 if (VPB_SHARERS(x) > 1) 1009 return (0); 1010 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1011 ("vm_page_busy_tryupgrade: invalid lock state")); 1012 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 1013 ce | (x & VPB_BIT_WAITERS))) 1014 continue; 1015 return (1); 1016 } 1017 } 1018 1019 /* 1020 * vm_page_sbusied: 1021 * 1022 * Return a positive value if the page is shared busied, 0 otherwise. 1023 */ 1024 int 1025 vm_page_sbusied(vm_page_t m) 1026 { 1027 u_int x; 1028 1029 x = vm_page_busy_fetch(m); 1030 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1031 } 1032 1033 /* 1034 * vm_page_sunbusy: 1035 * 1036 * Shared unbusy a page. 1037 */ 1038 void 1039 vm_page_sunbusy(vm_page_t m) 1040 { 1041 u_int x; 1042 1043 vm_page_assert_sbusied(m); 1044 1045 x = vm_page_busy_fetch(m); 1046 for (;;) { 1047 KASSERT(x != VPB_FREED, 1048 ("vm_page_sunbusy: Unlocking freed page.")); 1049 if (VPB_SHARERS(x) > 1) { 1050 if (atomic_fcmpset_int(&m->busy_lock, &x, 1051 x - VPB_ONE_SHARER)) 1052 break; 1053 continue; 1054 } 1055 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1056 ("vm_page_sunbusy: invalid lock state")); 1057 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1058 continue; 1059 if ((x & VPB_BIT_WAITERS) == 0) 1060 break; 1061 wakeup(m); 1062 break; 1063 } 1064 } 1065 1066 /* 1067 * vm_page_busy_sleep: 1068 * 1069 * Sleep if the page is busy, using the page pointer as wchan. 1070 * This is used to implement the hard-path of the busying mechanism. 1071 * 1072 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1073 * will not sleep if the page is shared-busy. 1074 * 1075 * The object lock must be held on entry. 1076 * 1077 * Returns true if it slept and dropped the object lock, or false 1078 * if there was no sleep and the lock is still held. 1079 */ 1080 bool 1081 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) 1082 { 1083 vm_object_t obj; 1084 1085 obj = m->object; 1086 VM_OBJECT_ASSERT_LOCKED(obj); 1087 1088 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, 1089 true)); 1090 } 1091 1092 /* 1093 * vm_page_busy_sleep_unlocked: 1094 * 1095 * Sleep if the page is busy, using the page pointer as wchan. 1096 * This is used to implement the hard-path of busying mechanism. 1097 * 1098 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1099 * will not sleep if the page is shared-busy. 1100 * 1101 * The object lock must not be held on entry. The operation will 1102 * return if the page changes identity. 1103 */ 1104 void 1105 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1106 const char *wmesg, int allocflags) 1107 { 1108 VM_OBJECT_ASSERT_UNLOCKED(obj); 1109 1110 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); 1111 } 1112 1113 /* 1114 * _vm_page_busy_sleep: 1115 * 1116 * Internal busy sleep function. Verifies the page identity and 1117 * lockstate against parameters. Returns true if it sleeps and 1118 * false otherwise. 1119 * 1120 * allocflags uses VM_ALLOC_* flags to specify the lock required. 1121 * 1122 * If locked is true the lock will be dropped for any true returns 1123 * and held for any false returns. 1124 */ 1125 static bool 1126 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1127 const char *wmesg, int allocflags, bool locked) 1128 { 1129 bool xsleep; 1130 u_int x; 1131 1132 /* 1133 * If the object is busy we must wait for that to drain to zero 1134 * before trying the page again. 1135 */ 1136 if (obj != NULL && vm_object_busied(obj)) { 1137 if (locked) 1138 VM_OBJECT_DROP(obj); 1139 vm_object_busy_wait(obj, wmesg); 1140 return (true); 1141 } 1142 1143 if (!vm_page_busied(m)) 1144 return (false); 1145 1146 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1147 sleepq_lock(m); 1148 x = vm_page_busy_fetch(m); 1149 do { 1150 /* 1151 * If the page changes objects or becomes unlocked we can 1152 * simply return. 1153 */ 1154 if (x == VPB_UNBUSIED || 1155 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1156 m->object != obj || m->pindex != pindex) { 1157 sleepq_release(m); 1158 return (false); 1159 } 1160 if ((x & VPB_BIT_WAITERS) != 0) 1161 break; 1162 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1163 if (locked) 1164 VM_OBJECT_DROP(obj); 1165 DROP_GIANT(); 1166 sleepq_add(m, NULL, wmesg, 0, 0); 1167 sleepq_wait(m, PVM); 1168 PICKUP_GIANT(); 1169 return (true); 1170 } 1171 1172 /* 1173 * vm_page_trysbusy: 1174 * 1175 * Try to shared busy a page. 1176 * If the operation succeeds 1 is returned otherwise 0. 1177 * The operation never sleeps. 1178 */ 1179 int 1180 vm_page_trysbusy(vm_page_t m) 1181 { 1182 vm_object_t obj; 1183 u_int x; 1184 1185 obj = m->object; 1186 x = vm_page_busy_fetch(m); 1187 for (;;) { 1188 if ((x & VPB_BIT_SHARED) == 0) 1189 return (0); 1190 /* 1191 * Reduce the window for transient busies that will trigger 1192 * false negatives in vm_page_ps_test(). 1193 */ 1194 if (obj != NULL && vm_object_busied(obj)) 1195 return (0); 1196 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1197 x + VPB_ONE_SHARER)) 1198 break; 1199 } 1200 1201 /* Refetch the object now that we're guaranteed that it is stable. */ 1202 obj = m->object; 1203 if (obj != NULL && vm_object_busied(obj)) { 1204 vm_page_sunbusy(m); 1205 return (0); 1206 } 1207 return (1); 1208 } 1209 1210 /* 1211 * vm_page_tryxbusy: 1212 * 1213 * Try to exclusive busy a page. 1214 * If the operation succeeds 1 is returned otherwise 0. 1215 * The operation never sleeps. 1216 */ 1217 int 1218 vm_page_tryxbusy(vm_page_t m) 1219 { 1220 vm_object_t obj; 1221 1222 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1223 VPB_CURTHREAD_EXCLUSIVE) == 0) 1224 return (0); 1225 1226 obj = m->object; 1227 if (obj != NULL && vm_object_busied(obj)) { 1228 vm_page_xunbusy(m); 1229 return (0); 1230 } 1231 return (1); 1232 } 1233 1234 static void 1235 vm_page_xunbusy_hard_tail(vm_page_t m) 1236 { 1237 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1238 /* Wake the waiter. */ 1239 wakeup(m); 1240 } 1241 1242 /* 1243 * vm_page_xunbusy_hard: 1244 * 1245 * Called when unbusy has failed because there is a waiter. 1246 */ 1247 void 1248 vm_page_xunbusy_hard(vm_page_t m) 1249 { 1250 vm_page_assert_xbusied(m); 1251 vm_page_xunbusy_hard_tail(m); 1252 } 1253 1254 void 1255 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1256 { 1257 vm_page_assert_xbusied_unchecked(m); 1258 vm_page_xunbusy_hard_tail(m); 1259 } 1260 1261 static void 1262 vm_page_busy_free(vm_page_t m) 1263 { 1264 u_int x; 1265 1266 atomic_thread_fence_rel(); 1267 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1268 if ((x & VPB_BIT_WAITERS) != 0) 1269 wakeup(m); 1270 } 1271 1272 /* 1273 * vm_page_unhold_pages: 1274 * 1275 * Unhold each of the pages that is referenced by the given array. 1276 */ 1277 void 1278 vm_page_unhold_pages(vm_page_t *ma, int count) 1279 { 1280 1281 for (; count != 0; count--) { 1282 vm_page_unwire(*ma, PQ_ACTIVE); 1283 ma++; 1284 } 1285 } 1286 1287 vm_page_t 1288 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1289 { 1290 vm_page_t m; 1291 1292 #ifdef VM_PHYSSEG_SPARSE 1293 m = vm_phys_paddr_to_vm_page(pa); 1294 if (m == NULL) 1295 m = vm_phys_fictitious_to_vm_page(pa); 1296 return (m); 1297 #elif defined(VM_PHYSSEG_DENSE) 1298 long pi; 1299 1300 pi = atop(pa); 1301 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1302 m = &vm_page_array[pi - first_page]; 1303 return (m); 1304 } 1305 return (vm_phys_fictitious_to_vm_page(pa)); 1306 #else 1307 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1308 #endif 1309 } 1310 1311 /* 1312 * vm_page_getfake: 1313 * 1314 * Create a fictitious page with the specified physical address and 1315 * memory attribute. The memory attribute is the only the machine- 1316 * dependent aspect of a fictitious page that must be initialized. 1317 */ 1318 vm_page_t 1319 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1320 { 1321 vm_page_t m; 1322 1323 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1324 vm_page_initfake(m, paddr, memattr); 1325 return (m); 1326 } 1327 1328 void 1329 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1330 { 1331 1332 if ((m->flags & PG_FICTITIOUS) != 0) { 1333 /* 1334 * The page's memattr might have changed since the 1335 * previous initialization. Update the pmap to the 1336 * new memattr. 1337 */ 1338 goto memattr; 1339 } 1340 m->phys_addr = paddr; 1341 m->a.queue = PQ_NONE; 1342 /* Fictitious pages don't use "segind". */ 1343 m->flags = PG_FICTITIOUS; 1344 /* Fictitious pages don't use "order" or "pool". */ 1345 m->oflags = VPO_UNMANAGED; 1346 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1347 /* Fictitious pages are unevictable. */ 1348 m->ref_count = 1; 1349 pmap_page_init(m); 1350 memattr: 1351 pmap_page_set_memattr(m, memattr); 1352 } 1353 1354 /* 1355 * vm_page_putfake: 1356 * 1357 * Release a fictitious page. 1358 */ 1359 void 1360 vm_page_putfake(vm_page_t m) 1361 { 1362 1363 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1364 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1365 ("vm_page_putfake: bad page %p", m)); 1366 vm_page_assert_xbusied(m); 1367 vm_page_busy_free(m); 1368 uma_zfree(fakepg_zone, m); 1369 } 1370 1371 /* 1372 * vm_page_updatefake: 1373 * 1374 * Update the given fictitious page to the specified physical address and 1375 * memory attribute. 1376 */ 1377 void 1378 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1379 { 1380 1381 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1382 ("vm_page_updatefake: bad page %p", m)); 1383 m->phys_addr = paddr; 1384 pmap_page_set_memattr(m, memattr); 1385 } 1386 1387 /* 1388 * vm_page_free: 1389 * 1390 * Free a page. 1391 */ 1392 void 1393 vm_page_free(vm_page_t m) 1394 { 1395 1396 m->flags &= ~PG_ZERO; 1397 vm_page_free_toq(m); 1398 } 1399 1400 /* 1401 * vm_page_free_zero: 1402 * 1403 * Free a page to the zerod-pages queue 1404 */ 1405 void 1406 vm_page_free_zero(vm_page_t m) 1407 { 1408 1409 m->flags |= PG_ZERO; 1410 vm_page_free_toq(m); 1411 } 1412 1413 /* 1414 * Unbusy and handle the page queueing for a page from a getpages request that 1415 * was optionally read ahead or behind. 1416 */ 1417 void 1418 vm_page_readahead_finish(vm_page_t m) 1419 { 1420 1421 /* We shouldn't put invalid pages on queues. */ 1422 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1423 1424 /* 1425 * Since the page is not the actually needed one, whether it should 1426 * be activated or deactivated is not obvious. Empirical results 1427 * have shown that deactivating the page is usually the best choice, 1428 * unless the page is wanted by another thread. 1429 */ 1430 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1431 vm_page_activate(m); 1432 else 1433 vm_page_deactivate(m); 1434 vm_page_xunbusy_unchecked(m); 1435 } 1436 1437 /* 1438 * Destroy the identity of an invalid page and free it if possible. 1439 * This is intended to be used when reading a page from backing store fails. 1440 */ 1441 void 1442 vm_page_free_invalid(vm_page_t m) 1443 { 1444 1445 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1446 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1447 KASSERT(m->object != NULL, ("page %p has no object", m)); 1448 VM_OBJECT_ASSERT_WLOCKED(m->object); 1449 1450 /* 1451 * We may be attempting to free the page as part of the handling for an 1452 * I/O error, in which case the page was xbusied by a different thread. 1453 */ 1454 vm_page_xbusy_claim(m); 1455 1456 /* 1457 * If someone has wired this page while the object lock 1458 * was not held, then the thread that unwires is responsible 1459 * for freeing the page. Otherwise just free the page now. 1460 * The wire count of this unmapped page cannot change while 1461 * we have the page xbusy and the page's object wlocked. 1462 */ 1463 if (vm_page_remove(m)) 1464 vm_page_free(m); 1465 } 1466 1467 /* 1468 * vm_page_dirty_KBI: [ internal use only ] 1469 * 1470 * Set all bits in the page's dirty field. 1471 * 1472 * The object containing the specified page must be locked if the 1473 * call is made from the machine-independent layer. 1474 * 1475 * See vm_page_clear_dirty_mask(). 1476 * 1477 * This function should only be called by vm_page_dirty(). 1478 */ 1479 void 1480 vm_page_dirty_KBI(vm_page_t m) 1481 { 1482 1483 /* Refer to this operation by its public name. */ 1484 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1485 m->dirty = VM_PAGE_BITS_ALL; 1486 } 1487 1488 /* 1489 * Insert the given page into the given object at the given pindex. 1490 * 1491 * The procedure is marked __always_inline to suggest to the compiler to 1492 * eliminate the iter parameter and the associated alternate branch. 1493 */ 1494 static __always_inline int 1495 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1496 bool iter, struct pctrie_iter *pages) 1497 { 1498 int error; 1499 1500 VM_OBJECT_ASSERT_WLOCKED(object); 1501 KASSERT(m->object == NULL, 1502 ("vm_page_insert: page %p already inserted", m)); 1503 1504 /* 1505 * Record the object/offset pair in this page. 1506 */ 1507 m->object = object; 1508 m->pindex = pindex; 1509 m->ref_count |= VPRC_OBJREF; 1510 1511 /* 1512 * Add this page to the object's radix tree. 1513 */ 1514 if (iter) 1515 error = vm_radix_iter_insert(pages, m); 1516 else 1517 error = vm_radix_insert(&object->rtree, m); 1518 if (__predict_false(error != 0)) { 1519 m->object = NULL; 1520 m->pindex = 0; 1521 m->ref_count &= ~VPRC_OBJREF; 1522 return (1); 1523 } 1524 1525 vm_page_insert_radixdone(m, object); 1526 vm_pager_page_inserted(object, m); 1527 return (0); 1528 } 1529 1530 /* 1531 * vm_page_insert: [ internal use only ] 1532 * 1533 * Inserts the given mem entry into the object and object list. 1534 * 1535 * The object must be locked. 1536 */ 1537 int 1538 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1539 { 1540 return (vm_page_insert_lookup(m, object, pindex, false, NULL)); 1541 } 1542 1543 /* 1544 * vm_page_iter_insert: 1545 * 1546 * Tries to insert the page "m" into the specified object at offset 1547 * "pindex" using the iterator "pages". Returns 0 if the insertion was 1548 * successful. 1549 * 1550 * The object must be locked. 1551 */ 1552 int 1553 vm_page_iter_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1554 struct pctrie_iter *pages) 1555 { 1556 return (vm_page_insert_lookup(m, object, pindex, true, pages)); 1557 } 1558 1559 /* 1560 * vm_page_insert_radixdone: 1561 * 1562 * Complete page "m" insertion into the specified object after the 1563 * radix trie hooking. 1564 * 1565 * The object must be locked. 1566 */ 1567 static void 1568 vm_page_insert_radixdone(vm_page_t m, vm_object_t object) 1569 { 1570 1571 VM_OBJECT_ASSERT_WLOCKED(object); 1572 KASSERT(object != NULL && m->object == object, 1573 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1574 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1575 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1576 1577 /* 1578 * Show that the object has one more resident page. 1579 */ 1580 object->resident_page_count++; 1581 1582 /* 1583 * Hold the vnode until the last page is released. 1584 */ 1585 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1586 vhold(object->handle); 1587 1588 /* 1589 * Since we are inserting a new and possibly dirty page, 1590 * update the object's generation count. 1591 */ 1592 if (pmap_page_is_write_mapped(m)) 1593 vm_object_set_writeable_dirty(object); 1594 } 1595 1596 /* 1597 * vm_page_remove_radixdone 1598 * 1599 * Complete page "m" removal from the specified object after the radix trie 1600 * unhooking. 1601 * 1602 * The caller is responsible for updating the page's fields to reflect this 1603 * removal. 1604 */ 1605 static void 1606 vm_page_remove_radixdone(vm_page_t m) 1607 { 1608 vm_object_t object; 1609 1610 vm_page_assert_xbusied(m); 1611 object = m->object; 1612 VM_OBJECT_ASSERT_WLOCKED(object); 1613 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1614 ("page %p is missing its object ref", m)); 1615 1616 /* Deferred free of swap space. */ 1617 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1618 vm_pager_page_unswapped(m); 1619 1620 vm_pager_page_removed(object, m); 1621 m->object = NULL; 1622 1623 /* 1624 * And show that the object has one fewer resident page. 1625 */ 1626 object->resident_page_count--; 1627 1628 /* 1629 * The vnode may now be recycled. 1630 */ 1631 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1632 vdrop(object->handle); 1633 } 1634 1635 /* 1636 * vm_page_free_object_prep: 1637 * 1638 * Disassociates the given page from its VM object. 1639 * 1640 * The object must be locked, and the page must be xbusy. 1641 */ 1642 static void 1643 vm_page_free_object_prep(vm_page_t m) 1644 { 1645 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 1646 ((m->object->flags & OBJ_UNMANAGED) != 0), 1647 ("%s: managed flag mismatch for page %p", 1648 __func__, m)); 1649 vm_page_assert_xbusied(m); 1650 1651 /* 1652 * The object reference can be released without an atomic 1653 * operation. 1654 */ 1655 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 1656 m->ref_count == VPRC_OBJREF, 1657 ("%s: page %p has unexpected ref_count %u", 1658 __func__, m, m->ref_count)); 1659 vm_page_remove_radixdone(m); 1660 m->ref_count -= VPRC_OBJREF; 1661 } 1662 1663 /* 1664 * vm_page_iter_free: 1665 * 1666 * Free the given page, and use the iterator to remove it from the radix 1667 * tree. 1668 */ 1669 void 1670 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m) 1671 { 1672 vm_radix_iter_remove(pages); 1673 vm_page_free_object_prep(m); 1674 vm_page_xunbusy(m); 1675 m->flags &= ~PG_ZERO; 1676 vm_page_free_toq(m); 1677 } 1678 1679 /* 1680 * vm_page_remove: 1681 * 1682 * Removes the specified page from its containing object, but does not 1683 * invalidate any backing storage. Returns true if the object's reference 1684 * was the last reference to the page, and false otherwise. 1685 * 1686 * The object must be locked and the page must be exclusively busied. 1687 * The exclusive busy will be released on return. If this is not the 1688 * final ref and the caller does not hold a wire reference it may not 1689 * continue to access the page. 1690 */ 1691 bool 1692 vm_page_remove(vm_page_t m) 1693 { 1694 bool dropped; 1695 1696 dropped = vm_page_remove_xbusy(m); 1697 vm_page_xunbusy(m); 1698 1699 return (dropped); 1700 } 1701 1702 /* 1703 * vm_page_iter_remove: 1704 * 1705 * Remove the current page, and use the iterator to remove it from the 1706 * radix tree. 1707 */ 1708 bool 1709 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m) 1710 { 1711 bool dropped; 1712 1713 vm_radix_iter_remove(pages); 1714 vm_page_remove_radixdone(m); 1715 dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1716 vm_page_xunbusy(m); 1717 1718 return (dropped); 1719 } 1720 1721 /* 1722 * vm_page_radix_remove 1723 * 1724 * Removes the specified page from the radix tree. 1725 */ 1726 static void 1727 vm_page_radix_remove(vm_page_t m) 1728 { 1729 vm_page_t mrem __diagused; 1730 1731 mrem = vm_radix_remove(&m->object->rtree, m->pindex); 1732 KASSERT(mrem == m, 1733 ("removed page %p, expected page %p", mrem, m)); 1734 } 1735 1736 /* 1737 * vm_page_remove_xbusy 1738 * 1739 * Removes the page but leaves the xbusy held. Returns true if this 1740 * removed the final ref and false otherwise. 1741 */ 1742 bool 1743 vm_page_remove_xbusy(vm_page_t m) 1744 { 1745 1746 vm_page_radix_remove(m); 1747 vm_page_remove_radixdone(m); 1748 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1749 } 1750 1751 /* 1752 * vm_page_lookup: 1753 * 1754 * Returns the page associated with the object/offset 1755 * pair specified; if none is found, NULL is returned. 1756 * 1757 * The object must be locked. 1758 */ 1759 vm_page_t 1760 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1761 { 1762 1763 VM_OBJECT_ASSERT_LOCKED(object); 1764 return (vm_radix_lookup(&object->rtree, pindex)); 1765 } 1766 1767 /* 1768 * vm_page_iter_init: 1769 * 1770 * Initialize iterator for vm pages. 1771 */ 1772 void 1773 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object) 1774 { 1775 1776 vm_radix_iter_init(pages, &object->rtree); 1777 } 1778 1779 /* 1780 * vm_page_iter_init: 1781 * 1782 * Initialize iterator for vm pages. 1783 */ 1784 void 1785 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object, 1786 vm_pindex_t limit) 1787 { 1788 1789 vm_radix_iter_limit_init(pages, &object->rtree, limit); 1790 } 1791 1792 /* 1793 * vm_page_lookup_unlocked: 1794 * 1795 * Returns the page associated with the object/offset pair specified; 1796 * if none is found, NULL is returned. The page may be no longer be 1797 * present in the object at the time that this function returns. Only 1798 * useful for opportunistic checks such as inmem(). 1799 */ 1800 vm_page_t 1801 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1802 { 1803 1804 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1805 } 1806 1807 /* 1808 * vm_page_relookup: 1809 * 1810 * Returns a page that must already have been busied by 1811 * the caller. Used for bogus page replacement. 1812 */ 1813 vm_page_t 1814 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1815 { 1816 vm_page_t m; 1817 1818 m = vm_page_lookup_unlocked(object, pindex); 1819 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1820 m->object == object && m->pindex == pindex, 1821 ("vm_page_relookup: Invalid page %p", m)); 1822 return (m); 1823 } 1824 1825 /* 1826 * This should only be used by lockless functions for releasing transient 1827 * incorrect acquires. The page may have been freed after we acquired a 1828 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1829 * further to do. 1830 */ 1831 static void 1832 vm_page_busy_release(vm_page_t m) 1833 { 1834 u_int x; 1835 1836 x = vm_page_busy_fetch(m); 1837 for (;;) { 1838 if (x == VPB_FREED) 1839 break; 1840 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1841 if (atomic_fcmpset_int(&m->busy_lock, &x, 1842 x - VPB_ONE_SHARER)) 1843 break; 1844 continue; 1845 } 1846 KASSERT((x & VPB_BIT_SHARED) != 0 || 1847 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1848 ("vm_page_busy_release: %p xbusy not owned.", m)); 1849 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1850 continue; 1851 if ((x & VPB_BIT_WAITERS) != 0) 1852 wakeup(m); 1853 break; 1854 } 1855 } 1856 1857 /* 1858 * Uses the page mnew as a replacement for an existing page at index 1859 * pindex which must be already present in the object. 1860 * 1861 * Both pages must be exclusively busied on enter. The old page is 1862 * unbusied on exit. 1863 * 1864 * A return value of true means mold is now free. If this is not the 1865 * final ref and the caller does not hold a wire reference it may not 1866 * continue to access the page. 1867 */ 1868 static bool 1869 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1870 vm_page_t mold) 1871 { 1872 vm_page_t mret __diagused; 1873 bool dropped; 1874 1875 VM_OBJECT_ASSERT_WLOCKED(object); 1876 vm_page_assert_xbusied(mold); 1877 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1878 ("vm_page_replace: page %p already in object", mnew)); 1879 1880 /* 1881 * This function mostly follows vm_page_insert() and 1882 * vm_page_remove() without the radix, object count and vnode 1883 * dance. Double check such functions for more comments. 1884 */ 1885 1886 mnew->object = object; 1887 mnew->pindex = pindex; 1888 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1889 mret = vm_radix_replace(&object->rtree, mnew); 1890 KASSERT(mret == mold, 1891 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1892 KASSERT((mold->oflags & VPO_UNMANAGED) == 1893 (mnew->oflags & VPO_UNMANAGED), 1894 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1895 1896 mold->object = NULL; 1897 1898 /* 1899 * The object's resident_page_count does not change because we have 1900 * swapped one page for another, but the generation count should 1901 * change if the page is dirty. 1902 */ 1903 if (pmap_page_is_write_mapped(mnew)) 1904 vm_object_set_writeable_dirty(object); 1905 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1906 vm_page_xunbusy(mold); 1907 1908 return (dropped); 1909 } 1910 1911 void 1912 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1913 vm_page_t mold) 1914 { 1915 1916 vm_page_assert_xbusied(mnew); 1917 1918 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1919 vm_page_free(mold); 1920 } 1921 1922 /* 1923 * vm_page_iter_rename: 1924 * 1925 * Tries to move the specified page from its current object to a new object 1926 * and pindex, using the given iterator to remove the page from its current 1927 * object. Returns true if the move was successful, and false if the move 1928 * was aborted due to a failed memory allocation. 1929 * 1930 * Panics if a page already resides in the new object at the new pindex. 1931 * 1932 * This routine dirties the page if it is valid, as callers are expected to 1933 * transfer backing storage only after moving the page. Dirtying the page 1934 * ensures that the destination object retains the most recent copy of the 1935 * page. 1936 * 1937 * The objects must be locked. 1938 */ 1939 bool 1940 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m, 1941 vm_object_t new_object, vm_pindex_t new_pindex) 1942 { 1943 vm_pindex_t opidx; 1944 1945 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1946 ("%s: page %p is missing object ref", __func__, m)); 1947 VM_OBJECT_ASSERT_WLOCKED(m->object); 1948 VM_OBJECT_ASSERT_WLOCKED(new_object); 1949 1950 /* 1951 * Create a custom version of vm_page_insert() which does not depend 1952 * by m_prev and can cheat on the implementation aspects of the 1953 * function. 1954 */ 1955 opidx = m->pindex; 1956 m->pindex = new_pindex; 1957 if (vm_radix_insert(&new_object->rtree, m) != 0) { 1958 m->pindex = opidx; 1959 return (false); 1960 } 1961 1962 /* 1963 * The operation cannot fail anymore. 1964 */ 1965 m->pindex = opidx; 1966 vm_radix_iter_remove(old_pages); 1967 vm_page_remove_radixdone(m); 1968 1969 /* Return back to the new pindex to complete vm_page_insert(). */ 1970 m->pindex = new_pindex; 1971 m->object = new_object; 1972 1973 vm_page_insert_radixdone(m, new_object); 1974 if (vm_page_any_valid(m)) 1975 vm_page_dirty(m); 1976 vm_pager_page_inserted(new_object, m); 1977 return (true); 1978 } 1979 1980 /* 1981 * vm_page_alloc: 1982 * 1983 * Allocate and return a page that is associated with the specified 1984 * object and offset pair. By default, this page is exclusive busied. 1985 * 1986 * The caller must always specify an allocation class. 1987 * 1988 * allocation classes: 1989 * VM_ALLOC_NORMAL normal process request 1990 * VM_ALLOC_SYSTEM system *really* needs a page 1991 * VM_ALLOC_INTERRUPT interrupt time request 1992 * 1993 * optional allocation flags: 1994 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1995 * intends to allocate 1996 * VM_ALLOC_NOBUSY do not exclusive busy the page 1997 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1998 * VM_ALLOC_NOFREE page will never be freed 1999 * VM_ALLOC_NOWAIT ignored (default behavior) 2000 * VM_ALLOC_SBUSY shared busy the allocated page 2001 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 2002 * VM_ALLOC_WIRED wire the allocated page 2003 * VM_ALLOC_ZERO prefer a zeroed page 2004 */ 2005 vm_page_t 2006 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 2007 { 2008 struct pctrie_iter pages; 2009 2010 vm_page_iter_init(&pages, object); 2011 return (vm_page_alloc_iter(object, pindex, req, &pages)); 2012 } 2013 2014 /* 2015 * Allocate a page in the specified object with the given page index. If the 2016 * object lock is dropped and regained, the pages iter is reset. 2017 */ 2018 vm_page_t 2019 vm_page_alloc_iter(vm_object_t object, vm_pindex_t pindex, int req, 2020 struct pctrie_iter *pages) 2021 { 2022 struct vm_domainset_iter di; 2023 vm_page_t m; 2024 int domain; 2025 2026 if (vm_domainset_iter_page_init(&di, object, pindex, &domain, &req) != 0) 2027 return (NULL); 2028 2029 do { 2030 m = vm_page_alloc_domain_iter(object, pindex, domain, req, 2031 pages); 2032 if (m != NULL) 2033 break; 2034 } while (vm_domainset_iter_page(&di, object, &domain, pages) == 0); 2035 2036 return (m); 2037 } 2038 2039 /* 2040 * Returns true if the number of free pages exceeds the minimum 2041 * for the request class and false otherwise. 2042 */ 2043 static int 2044 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2045 { 2046 u_int limit, old, new; 2047 2048 if (req_class == VM_ALLOC_INTERRUPT) 2049 limit = 0; 2050 else if (req_class == VM_ALLOC_SYSTEM) 2051 limit = vmd->vmd_interrupt_free_min; 2052 else 2053 limit = vmd->vmd_free_reserved; 2054 2055 /* 2056 * Attempt to reserve the pages. Fail if we're below the limit. 2057 */ 2058 limit += npages; 2059 old = atomic_load_int(&vmd->vmd_free_count); 2060 do { 2061 if (old < limit) 2062 return (0); 2063 new = old - npages; 2064 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2065 2066 /* Wake the page daemon if we've crossed the threshold. */ 2067 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2068 pagedaemon_wakeup(vmd->vmd_domain); 2069 2070 /* Only update bitsets on transitions. */ 2071 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2072 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2073 vm_domain_set(vmd); 2074 2075 return (1); 2076 } 2077 2078 int 2079 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2080 { 2081 int req_class; 2082 2083 /* 2084 * The page daemon is allowed to dig deeper into the free page list. 2085 */ 2086 req_class = req & VM_ALLOC_CLASS_MASK; 2087 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2088 req_class = VM_ALLOC_SYSTEM; 2089 return (_vm_domain_allocate(vmd, req_class, npages)); 2090 } 2091 2092 vm_page_t 2093 vm_page_alloc_domain_iter(vm_object_t object, vm_pindex_t pindex, int domain, 2094 int req, struct pctrie_iter *pages) 2095 { 2096 struct vm_domain *vmd; 2097 vm_page_t m; 2098 int flags; 2099 2100 #define VM_ALLOC_COMMON (VM_ALLOC_CLASS_MASK | VM_ALLOC_NODUMP | \ 2101 VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | \ 2102 VM_ALLOC_WIRED | VM_ALLOC_ZERO) 2103 #define VPA_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2104 VM_ALLOC_NOBUSY | VM_ALLOC_NOFREE | \ 2105 VM_ALLOC_SBUSY) 2106 KASSERT((req & ~VPA_FLAGS) == 0, 2107 ("invalid request %#x", req)); 2108 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2109 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2110 ("invalid request %#x", req)); 2111 VM_OBJECT_ASSERT_WLOCKED(object); 2112 2113 flags = 0; 2114 m = NULL; 2115 if (!vm_pager_can_alloc_page(object, pindex)) 2116 return (NULL); 2117 #if VM_NRESERVLEVEL > 0 2118 again: 2119 #endif 2120 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2121 m = vm_page_alloc_nofree_domain(domain, req); 2122 if (m != NULL) 2123 goto found; 2124 } 2125 #if VM_NRESERVLEVEL > 0 2126 /* 2127 * Can we allocate the page from a reservation? 2128 */ 2129 if (vm_object_reserv(object) && 2130 (m = vm_reserv_alloc_page(object, pindex, domain, req, pages)) != 2131 NULL) { 2132 goto found; 2133 } 2134 #endif 2135 vmd = VM_DOMAIN(domain); 2136 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2137 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2138 M_NOWAIT | M_NOVM); 2139 if (m != NULL) { 2140 flags |= PG_PCPU_CACHE; 2141 goto found; 2142 } 2143 } 2144 if (vm_domain_allocate(vmd, req, 1)) { 2145 /* 2146 * If not, allocate it from the free page queues. 2147 */ 2148 vm_domain_free_lock(vmd); 2149 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2150 vm_domain_free_unlock(vmd); 2151 if (m == NULL) { 2152 vm_domain_freecnt_inc(vmd, 1); 2153 #if VM_NRESERVLEVEL > 0 2154 if (vm_reserv_reclaim_inactive(domain)) 2155 goto again; 2156 #endif 2157 } 2158 } 2159 if (m == NULL) { 2160 /* 2161 * Not allocatable, give up. 2162 */ 2163 (void)vm_domain_alloc_fail(vmd, object, req); 2164 if ((req & VM_ALLOC_WAITFAIL) != 0) 2165 pctrie_iter_reset(pages); 2166 return (NULL); 2167 } 2168 2169 /* 2170 * At this point we had better have found a good page. 2171 */ 2172 found: 2173 vm_page_dequeue(m); 2174 vm_page_alloc_check(m); 2175 2176 /* 2177 * Initialize the page. Only the PG_ZERO flag is inherited. 2178 */ 2179 flags |= m->flags & PG_ZERO; 2180 if ((req & VM_ALLOC_NODUMP) != 0) 2181 flags |= PG_NODUMP; 2182 if ((req & VM_ALLOC_NOFREE) != 0) 2183 flags |= PG_NOFREE; 2184 m->flags = flags; 2185 m->a.flags = 0; 2186 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2187 m->pool = VM_FREEPOOL_DEFAULT; 2188 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2189 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2190 else if ((req & VM_ALLOC_SBUSY) != 0) 2191 m->busy_lock = VPB_SHARERS_WORD(1); 2192 else 2193 m->busy_lock = VPB_UNBUSIED; 2194 if (req & VM_ALLOC_WIRED) { 2195 vm_wire_add(1); 2196 m->ref_count = 1; 2197 } 2198 m->a.act_count = 0; 2199 2200 if (vm_page_iter_insert(m, object, pindex, pages)) { 2201 if (req & VM_ALLOC_WIRED) { 2202 vm_wire_sub(1); 2203 m->ref_count = 0; 2204 } 2205 KASSERT(m->object == NULL, ("page %p has object", m)); 2206 m->oflags = VPO_UNMANAGED; 2207 m->busy_lock = VPB_UNBUSIED; 2208 /* Don't change PG_ZERO. */ 2209 vm_page_free_toq(m); 2210 if (req & VM_ALLOC_WAITFAIL) { 2211 VM_OBJECT_WUNLOCK(object); 2212 vm_radix_wait(); 2213 pctrie_iter_reset(pages); 2214 VM_OBJECT_WLOCK(object); 2215 } 2216 return (NULL); 2217 } 2218 2219 /* Ignore device objects; the pager sets "memattr" for them. */ 2220 if (object->memattr != VM_MEMATTR_DEFAULT && 2221 (object->flags & OBJ_FICTITIOUS) == 0) 2222 pmap_page_set_memattr(m, object->memattr); 2223 2224 return (m); 2225 } 2226 2227 /* 2228 * vm_page_alloc_contig: 2229 * 2230 * Allocate a contiguous set of physical pages of the given size "npages" 2231 * from the free lists. All of the physical pages must be at or above 2232 * the given physical address "low" and below the given physical address 2233 * "high". The given value "alignment" determines the alignment of the 2234 * first physical page in the set. If the given value "boundary" is 2235 * non-zero, then the set of physical pages cannot cross any physical 2236 * address boundary that is a multiple of that value. Both "alignment" 2237 * and "boundary" must be a power of two. 2238 * 2239 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2240 * then the memory attribute setting for the physical pages is configured 2241 * to the object's memory attribute setting. Otherwise, the memory 2242 * attribute setting for the physical pages is configured to "memattr", 2243 * overriding the object's memory attribute setting. However, if the 2244 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2245 * memory attribute setting for the physical pages cannot be configured 2246 * to VM_MEMATTR_DEFAULT. 2247 * 2248 * The specified object may not contain fictitious pages. 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 the pages 2255 * VM_ALLOC_INTERRUPT interrupt time request 2256 * 2257 * optional allocation flags: 2258 * VM_ALLOC_NOBUSY do not exclusive busy the pages 2259 * VM_ALLOC_NODUMP do not include the pages in a kernel core dump 2260 * VM_ALLOC_NORECLAIM do not reclaim after initial failure 2261 * VM_ALLOC_NOWAIT ignored (default behavior) 2262 * VM_ALLOC_SBUSY shared busy the allocated pages 2263 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 2264 * VM_ALLOC_WIRED wire the allocated pages 2265 * VM_ALLOC_ZERO prefer zeroed pages 2266 */ 2267 vm_page_t 2268 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2269 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2270 vm_paddr_t boundary, vm_memattr_t memattr) 2271 { 2272 struct vm_domainset_iter di; 2273 vm_page_t bounds[2]; 2274 vm_page_t m; 2275 int domain; 2276 int start_segind; 2277 2278 start_segind = -1; 2279 2280 if (vm_domainset_iter_page_init(&di, object, pindex, &domain, &req) != 0) 2281 return (NULL); 2282 2283 do { 2284 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2285 npages, low, high, alignment, boundary, memattr); 2286 if (m != NULL) 2287 break; 2288 if (start_segind == -1) 2289 start_segind = vm_phys_lookup_segind(low); 2290 if (vm_phys_find_range(bounds, start_segind, domain, 2291 npages, low, high) == -1) { 2292 vm_domainset_iter_ignore(&di, domain); 2293 } 2294 } while (vm_domainset_iter_page(&di, object, &domain, NULL) == 0); 2295 2296 return (m); 2297 } 2298 2299 static vm_page_t 2300 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2301 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2302 { 2303 struct vm_domain *vmd; 2304 vm_page_t m_ret; 2305 2306 /* 2307 * Can we allocate the pages without the number of free pages falling 2308 * below the lower bound for the allocation class? 2309 */ 2310 vmd = VM_DOMAIN(domain); 2311 if (!vm_domain_allocate(vmd, req, npages)) 2312 return (NULL); 2313 /* 2314 * Try to allocate the pages from the free page queues. 2315 */ 2316 vm_domain_free_lock(vmd); 2317 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2318 alignment, boundary); 2319 vm_domain_free_unlock(vmd); 2320 if (m_ret != NULL) 2321 return (m_ret); 2322 #if VM_NRESERVLEVEL > 0 2323 /* 2324 * Try to break a reservation to allocate the pages. 2325 */ 2326 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2327 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2328 high, alignment, boundary); 2329 if (m_ret != NULL) 2330 return (m_ret); 2331 } 2332 #endif 2333 vm_domain_freecnt_inc(vmd, npages); 2334 return (NULL); 2335 } 2336 2337 vm_page_t 2338 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2339 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2340 vm_paddr_t boundary, vm_memattr_t memattr) 2341 { 2342 struct pctrie_iter pages; 2343 vm_page_t m, m_ret, mpred; 2344 u_int busy_lock, flags, oflags; 2345 2346 #define VPAC_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2347 VM_ALLOC_NOBUSY | VM_ALLOC_NORECLAIM | \ 2348 VM_ALLOC_SBUSY) 2349 KASSERT((req & ~VPAC_FLAGS) == 0, 2350 ("invalid request %#x", req)); 2351 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2352 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2353 ("invalid request %#x", req)); 2354 VM_OBJECT_ASSERT_WLOCKED(object); 2355 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2356 ("vm_page_alloc_contig: object %p has fictitious pages", 2357 object)); 2358 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2359 2360 vm_page_iter_init(&pages, object); 2361 m_ret = NULL; 2362 #if VM_NRESERVLEVEL > 0 2363 /* 2364 * Can we allocate the pages from a reservation? 2365 */ 2366 if (vm_object_reserv(object)) { 2367 m_ret = vm_reserv_alloc_contig(object, pindex, domain, 2368 req, npages, low, high, alignment, boundary, &pages); 2369 } 2370 #endif 2371 if (m_ret == NULL) { 2372 m_ret = vm_page_find_contig_domain(domain, req, npages, 2373 low, high, alignment, boundary); 2374 } 2375 if (m_ret == NULL) { 2376 (void)vm_domain_alloc_fail(VM_DOMAIN(domain), object, req); 2377 return (NULL); 2378 } 2379 2380 /* 2381 * Initialize the pages. Only the PG_ZERO flag is inherited. 2382 */ 2383 flags = PG_ZERO; 2384 if ((req & VM_ALLOC_NODUMP) != 0) 2385 flags |= PG_NODUMP; 2386 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2387 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2388 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2389 else if ((req & VM_ALLOC_SBUSY) != 0) 2390 busy_lock = VPB_SHARERS_WORD(1); 2391 else 2392 busy_lock = VPB_UNBUSIED; 2393 if ((req & VM_ALLOC_WIRED) != 0) 2394 vm_wire_add(npages); 2395 if (object->memattr != VM_MEMATTR_DEFAULT && 2396 memattr == VM_MEMATTR_DEFAULT) 2397 memattr = object->memattr; 2398 for (m = m_ret; m < &m_ret[npages]; m++) { 2399 vm_page_dequeue(m); 2400 vm_page_alloc_check(m); 2401 m->a.flags = 0; 2402 m->flags = (m->flags | PG_NODUMP) & flags; 2403 m->busy_lock = busy_lock; 2404 if ((req & VM_ALLOC_WIRED) != 0) 2405 m->ref_count = 1; 2406 m->a.act_count = 0; 2407 m->oflags = oflags; 2408 m->pool = VM_FREEPOOL_DEFAULT; 2409 if (vm_page_iter_insert(m, object, pindex, &pages)) { 2410 if ((req & VM_ALLOC_WIRED) != 0) 2411 vm_wire_sub(npages); 2412 KASSERT(m->object == NULL, 2413 ("page %p has object", m)); 2414 mpred = m; 2415 for (m = m_ret; m < &m_ret[npages]; m++) { 2416 if (m <= mpred && 2417 (req & VM_ALLOC_WIRED) != 0) 2418 m->ref_count = 0; 2419 m->oflags = VPO_UNMANAGED; 2420 m->busy_lock = VPB_UNBUSIED; 2421 /* Don't change PG_ZERO. */ 2422 vm_page_free_toq(m); 2423 } 2424 if (req & VM_ALLOC_WAITFAIL) { 2425 VM_OBJECT_WUNLOCK(object); 2426 vm_radix_wait(); 2427 VM_OBJECT_WLOCK(object); 2428 } 2429 return (NULL); 2430 } 2431 if (memattr != VM_MEMATTR_DEFAULT) 2432 pmap_page_set_memattr(m, memattr); 2433 pindex++; 2434 } 2435 return (m_ret); 2436 } 2437 2438 /* 2439 * Allocate a physical page that is not intended to be inserted into a VM 2440 * object. 2441 */ 2442 vm_page_t 2443 vm_page_alloc_noobj_domain(int domain, int req) 2444 { 2445 struct vm_domain *vmd; 2446 vm_page_t m; 2447 int flags; 2448 2449 #define VPAN_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2450 VM_ALLOC_NOFREE | VM_ALLOC_WAITOK) 2451 KASSERT((req & ~VPAN_FLAGS) == 0, 2452 ("invalid request %#x", req)); 2453 2454 flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) | 2455 ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0); 2456 vmd = VM_DOMAIN(domain); 2457 again: 2458 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2459 m = vm_page_alloc_nofree_domain(domain, req); 2460 if (m != NULL) 2461 goto found; 2462 } 2463 2464 if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2465 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2466 M_NOWAIT | M_NOVM); 2467 if (m != NULL) { 2468 flags |= PG_PCPU_CACHE; 2469 goto found; 2470 } 2471 } 2472 2473 if (vm_domain_allocate(vmd, req, 1)) { 2474 vm_domain_free_lock(vmd); 2475 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2476 vm_domain_free_unlock(vmd); 2477 if (m == NULL) { 2478 vm_domain_freecnt_inc(vmd, 1); 2479 #if VM_NRESERVLEVEL > 0 2480 if (vm_reserv_reclaim_inactive(domain)) 2481 goto again; 2482 #endif 2483 } 2484 } 2485 if (m == NULL) { 2486 if (!vm_domain_alloc_fail(vmd, NULL, req)) 2487 return (NULL); 2488 goto again; 2489 } 2490 2491 found: 2492 /* 2493 * If the page comes from the free page cache, then it might still 2494 * have a pending deferred dequeue. Specifically, when the page is 2495 * imported from a different pool by vm_phys_alloc_npages(), the 2496 * second, third, etc. pages in a non-zero order set could have 2497 * pending deferred dequeues. 2498 */ 2499 vm_page_dequeue(m); 2500 vm_page_alloc_check(m); 2501 2502 /* 2503 * Consumers should not rely on a useful default pindex value. 2504 */ 2505 m->pindex = 0xdeadc0dedeadc0de; 2506 m->flags = (m->flags & PG_ZERO) | flags; 2507 m->a.flags = 0; 2508 m->oflags = VPO_UNMANAGED; 2509 m->pool = VM_FREEPOOL_DIRECT; 2510 m->busy_lock = VPB_UNBUSIED; 2511 if ((req & VM_ALLOC_WIRED) != 0) { 2512 vm_wire_add(1); 2513 m->ref_count = 1; 2514 } 2515 2516 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2517 pmap_zero_page(m); 2518 2519 return (m); 2520 } 2521 2522 #if VM_NRESERVLEVEL > 1 2523 #define VM_NOFREE_IMPORT_ORDER (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER) 2524 #elif VM_NRESERVLEVEL > 0 2525 #define VM_NOFREE_IMPORT_ORDER VM_LEVEL_0_ORDER 2526 #else 2527 #define VM_NOFREE_IMPORT_ORDER 8 2528 #endif 2529 2530 /* 2531 * Allocate a single NOFREE page. 2532 * 2533 * This routine hands out NOFREE pages from higher-order 2534 * physical memory blocks in order to reduce memory fragmentation. 2535 * When a NOFREE for a given domain chunk is used up, 2536 * the routine will try to fetch a new one from the freelists 2537 * and discard the old one. 2538 */ 2539 static vm_page_t __noinline 2540 vm_page_alloc_nofree_domain(int domain, int req) 2541 { 2542 vm_page_t m; 2543 struct vm_domain *vmd; 2544 2545 KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req)); 2546 2547 vmd = VM_DOMAIN(domain); 2548 vm_domain_free_lock(vmd); 2549 if (TAILQ_EMPTY(&vmd->vmd_nofreeq)) { 2550 int count; 2551 2552 count = 1 << VM_NOFREE_IMPORT_ORDER; 2553 if (!vm_domain_allocate(vmd, req, count)) { 2554 vm_domain_free_unlock(vmd); 2555 return (NULL); 2556 } 2557 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 2558 VM_NOFREE_IMPORT_ORDER); 2559 if (m == NULL) { 2560 vm_domain_freecnt_inc(vmd, count); 2561 vm_domain_free_unlock(vmd); 2562 return (NULL); 2563 } 2564 m->ref_count = count - 1; 2565 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, plinks.q); 2566 atomic_add_long(&nofreeq_size, count); 2567 } 2568 m = TAILQ_FIRST(&vmd->vmd_nofreeq); 2569 TAILQ_REMOVE(&vmd->vmd_nofreeq, m, plinks.q); 2570 if (m->ref_count > 0) { 2571 vm_page_t m_next; 2572 2573 m_next = &m[1]; 2574 vm_page_dequeue(m_next); 2575 m_next->ref_count = m->ref_count - 1; 2576 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m_next, plinks.q); 2577 m->ref_count = 0; 2578 } 2579 vm_domain_free_unlock(vmd); 2580 atomic_add_long(&nofreeq_size, -1); 2581 VM_CNT_INC(v_nofree_count); 2582 2583 return (m); 2584 } 2585 2586 /* 2587 * Though a NOFREE page by definition should not be freed, we support putting 2588 * them aside for future NOFREE allocations. This enables code which allocates 2589 * NOFREE pages for some purpose but then encounters an error and releases 2590 * resources. 2591 */ 2592 static void __noinline 2593 vm_page_free_nofree(struct vm_domain *vmd, vm_page_t m) 2594 { 2595 VM_CNT_ADD(v_nofree_count, -1); 2596 atomic_add_long(&nofreeq_size, 1); 2597 vm_domain_free_lock(vmd); 2598 MPASS(m->ref_count == 0); 2599 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, plinks.q); 2600 vm_domain_free_unlock(vmd); 2601 } 2602 2603 vm_page_t 2604 vm_page_alloc_noobj(int req) 2605 { 2606 struct vm_domainset_iter di; 2607 vm_page_t m; 2608 int domain; 2609 2610 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 2611 return (NULL); 2612 2613 do { 2614 m = vm_page_alloc_noobj_domain(domain, req); 2615 if (m != NULL) 2616 break; 2617 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 2618 2619 return (m); 2620 } 2621 2622 vm_page_t 2623 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2624 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2625 vm_memattr_t memattr) 2626 { 2627 struct vm_domainset_iter di; 2628 vm_page_t m; 2629 int domain; 2630 2631 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 2632 return (NULL); 2633 2634 do { 2635 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2636 high, alignment, boundary, memattr); 2637 if (m != NULL) 2638 break; 2639 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 2640 2641 return (m); 2642 } 2643 2644 vm_page_t 2645 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2646 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2647 vm_memattr_t memattr) 2648 { 2649 vm_page_t m, m_ret; 2650 u_int flags; 2651 2652 #define VPANC_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2653 VM_ALLOC_NORECLAIM | VM_ALLOC_WAITOK) 2654 KASSERT((req & ~VPANC_FLAGS) == 0, 2655 ("invalid request %#x", req)); 2656 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2657 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2658 ("invalid request %#x", req)); 2659 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2660 2661 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2662 low, high, alignment, boundary)) == NULL) { 2663 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2664 return (NULL); 2665 } 2666 2667 /* 2668 * Initialize the pages. Only the PG_ZERO flag is inherited. 2669 */ 2670 flags = PG_ZERO; 2671 if ((req & VM_ALLOC_NODUMP) != 0) 2672 flags |= PG_NODUMP; 2673 if ((req & VM_ALLOC_WIRED) != 0) 2674 vm_wire_add(npages); 2675 for (m = m_ret; m < &m_ret[npages]; m++) { 2676 vm_page_dequeue(m); 2677 vm_page_alloc_check(m); 2678 2679 /* 2680 * Consumers should not rely on a useful default pindex value. 2681 */ 2682 m->pindex = 0xdeadc0dedeadc0de; 2683 m->a.flags = 0; 2684 m->flags = (m->flags | PG_NODUMP) & flags; 2685 m->busy_lock = VPB_UNBUSIED; 2686 if ((req & VM_ALLOC_WIRED) != 0) 2687 m->ref_count = 1; 2688 m->a.act_count = 0; 2689 m->oflags = VPO_UNMANAGED; 2690 m->pool = VM_FREEPOOL_DIRECT; 2691 2692 /* 2693 * Zero the page before updating any mappings since the page is 2694 * not yet shared with any devices which might require the 2695 * non-default memory attribute. pmap_page_set_memattr() 2696 * flushes data caches before returning. 2697 */ 2698 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2699 pmap_zero_page(m); 2700 if (memattr != VM_MEMATTR_DEFAULT) 2701 pmap_page_set_memattr(m, memattr); 2702 } 2703 return (m_ret); 2704 } 2705 2706 /* 2707 * Check a page that has been freshly dequeued from a freelist. 2708 */ 2709 static void 2710 vm_page_alloc_check(vm_page_t m) 2711 { 2712 2713 KASSERT(m->object == NULL, ("page %p has object", m)); 2714 KASSERT(m->a.queue == PQ_NONE && 2715 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2716 ("page %p has unexpected queue %d, flags %#x", 2717 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2718 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2719 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2720 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2721 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2722 ("page %p has unexpected memattr %d", 2723 m, pmap_page_get_memattr(m))); 2724 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2725 pmap_vm_page_alloc_check(m); 2726 } 2727 2728 static int 2729 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2730 { 2731 struct vm_domain *vmd; 2732 struct vm_pgcache *pgcache; 2733 int i; 2734 2735 pgcache = arg; 2736 vmd = VM_DOMAIN(pgcache->domain); 2737 2738 /* 2739 * The page daemon should avoid creating extra memory pressure since its 2740 * main purpose is to replenish the store of free pages. 2741 */ 2742 if (vmd->vmd_severeset || curproc == pageproc || 2743 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2744 return (0); 2745 domain = vmd->vmd_domain; 2746 vm_domain_free_lock(vmd); 2747 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2748 (vm_page_t *)store); 2749 vm_domain_free_unlock(vmd); 2750 if (cnt != i) 2751 vm_domain_freecnt_inc(vmd, cnt - i); 2752 2753 return (i); 2754 } 2755 2756 static void 2757 vm_page_zone_release(void *arg, void **store, int cnt) 2758 { 2759 struct vm_domain *vmd; 2760 struct vm_pgcache *pgcache; 2761 vm_page_t m; 2762 int i; 2763 2764 pgcache = arg; 2765 vmd = VM_DOMAIN(pgcache->domain); 2766 vm_domain_free_lock(vmd); 2767 for (i = 0; i < cnt; i++) { 2768 m = (vm_page_t)store[i]; 2769 vm_phys_free_pages(m, pgcache->pool, 0); 2770 } 2771 vm_domain_free_unlock(vmd); 2772 vm_domain_freecnt_inc(vmd, cnt); 2773 } 2774 2775 #define VPSC_ANY 0 /* No restrictions. */ 2776 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2777 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2778 2779 /* 2780 * vm_page_scan_contig: 2781 * 2782 * Scan vm_page_array[] between the specified entries "m_start" and 2783 * "m_end" for a run of contiguous physical pages that satisfy the 2784 * specified conditions, and return the lowest page in the run. The 2785 * specified "alignment" determines the alignment of the lowest physical 2786 * page in the run. If the specified "boundary" is non-zero, then the 2787 * run of physical pages cannot span a physical address that is a 2788 * multiple of "boundary". 2789 * 2790 * "m_end" is never dereferenced, so it need not point to a vm_page 2791 * structure within vm_page_array[]. 2792 * 2793 * "npages" must be greater than zero. "m_start" and "m_end" must not 2794 * span a hole (or discontiguity) in the physical address space. Both 2795 * "alignment" and "boundary" must be a power of two. 2796 */ 2797 static vm_page_t 2798 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2799 u_long alignment, vm_paddr_t boundary, int options) 2800 { 2801 vm_object_t object; 2802 vm_paddr_t pa; 2803 vm_page_t m, m_run; 2804 #if VM_NRESERVLEVEL > 0 2805 int level; 2806 #endif 2807 int m_inc, order, run_ext, run_len; 2808 2809 KASSERT(npages > 0, ("npages is 0")); 2810 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2811 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2812 m_run = NULL; 2813 run_len = 0; 2814 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2815 KASSERT((m->flags & PG_MARKER) == 0, 2816 ("page %p is PG_MARKER", m)); 2817 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2818 ("fictitious page %p has invalid ref count", m)); 2819 2820 /* 2821 * If the current page would be the start of a run, check its 2822 * physical address against the end, alignment, and boundary 2823 * conditions. If it doesn't satisfy these conditions, either 2824 * terminate the scan or advance to the next page that 2825 * satisfies the failed condition. 2826 */ 2827 if (run_len == 0) { 2828 KASSERT(m_run == NULL, ("m_run != NULL")); 2829 if (m + npages > m_end) 2830 break; 2831 pa = VM_PAGE_TO_PHYS(m); 2832 if (!vm_addr_align_ok(pa, alignment)) { 2833 m_inc = atop(roundup2(pa, alignment) - pa); 2834 continue; 2835 } 2836 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2837 m_inc = atop(roundup2(pa, boundary) - pa); 2838 continue; 2839 } 2840 } else 2841 KASSERT(m_run != NULL, ("m_run == NULL")); 2842 2843 retry: 2844 m_inc = 1; 2845 if (vm_page_wired(m)) 2846 run_ext = 0; 2847 #if VM_NRESERVLEVEL > 0 2848 else if ((level = vm_reserv_level(m)) >= 0 && 2849 (options & VPSC_NORESERV) != 0) { 2850 run_ext = 0; 2851 /* Advance to the end of the reservation. */ 2852 pa = VM_PAGE_TO_PHYS(m); 2853 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2854 pa); 2855 } 2856 #endif 2857 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2858 /* 2859 * The page is considered eligible for relocation if 2860 * and only if it could be laundered or reclaimed by 2861 * the page daemon. 2862 */ 2863 VM_OBJECT_RLOCK(object); 2864 if (object != m->object) { 2865 VM_OBJECT_RUNLOCK(object); 2866 goto retry; 2867 } 2868 /* Don't care: PG_NODUMP, PG_ZERO. */ 2869 if ((object->flags & OBJ_SWAP) == 0 && 2870 object->type != OBJT_VNODE) { 2871 run_ext = 0; 2872 #if VM_NRESERVLEVEL > 0 2873 } else if ((options & VPSC_NOSUPER) != 0 && 2874 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2875 run_ext = 0; 2876 /* Advance to the end of the superpage. */ 2877 pa = VM_PAGE_TO_PHYS(m); 2878 m_inc = atop(roundup2(pa + 1, 2879 vm_reserv_size(level)) - pa); 2880 #endif 2881 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2882 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2883 /* 2884 * The page is allocated but eligible for 2885 * relocation. Extend the current run by one 2886 * page. 2887 */ 2888 KASSERT(pmap_page_get_memattr(m) == 2889 VM_MEMATTR_DEFAULT, 2890 ("page %p has an unexpected memattr", m)); 2891 KASSERT((m->oflags & (VPO_SWAPINPROG | 2892 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2893 ("page %p has unexpected oflags", m)); 2894 /* Don't care: PGA_NOSYNC. */ 2895 run_ext = 1; 2896 } else 2897 run_ext = 0; 2898 VM_OBJECT_RUNLOCK(object); 2899 #if VM_NRESERVLEVEL > 0 2900 } else if (level >= 0) { 2901 /* 2902 * The page is reserved but not yet allocated. In 2903 * other words, it is still free. Extend the current 2904 * run by one page. 2905 */ 2906 run_ext = 1; 2907 #endif 2908 } else if ((order = m->order) < VM_NFREEORDER) { 2909 /* 2910 * The page is enqueued in the physical memory 2911 * allocator's free page queues. Moreover, it is the 2912 * first page in a power-of-two-sized run of 2913 * contiguous free pages. Add these pages to the end 2914 * of the current run, and jump ahead. 2915 */ 2916 run_ext = 1 << order; 2917 m_inc = 1 << order; 2918 } else { 2919 /* 2920 * Skip the page for one of the following reasons: (1) 2921 * It is enqueued in the physical memory allocator's 2922 * free page queues. However, it is not the first 2923 * page in a run of contiguous free pages. (This case 2924 * rarely occurs because the scan is performed in 2925 * ascending order.) (2) It is not reserved, and it is 2926 * transitioning from free to allocated. (Conversely, 2927 * the transition from allocated to free for managed 2928 * pages is blocked by the page busy lock.) (3) It is 2929 * allocated but not contained by an object and not 2930 * wired, e.g., allocated by Xen's balloon driver. 2931 */ 2932 run_ext = 0; 2933 } 2934 2935 /* 2936 * Extend or reset the current run of pages. 2937 */ 2938 if (run_ext > 0) { 2939 if (run_len == 0) 2940 m_run = m; 2941 run_len += run_ext; 2942 } else { 2943 if (run_len > 0) { 2944 m_run = NULL; 2945 run_len = 0; 2946 } 2947 } 2948 } 2949 if (run_len >= npages) 2950 return (m_run); 2951 return (NULL); 2952 } 2953 2954 /* 2955 * vm_page_reclaim_run: 2956 * 2957 * Try to relocate each of the allocated virtual pages within the 2958 * specified run of physical pages to a new physical address. Free the 2959 * physical pages underlying the relocated virtual pages. A virtual page 2960 * is relocatable if and only if it could be laundered or reclaimed by 2961 * the page daemon. Whenever possible, a virtual page is relocated to a 2962 * physical address above "high". 2963 * 2964 * Returns 0 if every physical page within the run was already free or 2965 * just freed by a successful relocation. Otherwise, returns a non-zero 2966 * value indicating why the last attempt to relocate a virtual page was 2967 * unsuccessful. 2968 * 2969 * "req_class" must be an allocation class. 2970 */ 2971 static int 2972 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2973 vm_paddr_t high) 2974 { 2975 struct vm_domain *vmd; 2976 struct spglist free; 2977 vm_object_t object; 2978 vm_paddr_t pa; 2979 vm_page_t m, m_end, m_new; 2980 int error, order, req; 2981 2982 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2983 ("req_class is not an allocation class")); 2984 SLIST_INIT(&free); 2985 error = 0; 2986 m = m_run; 2987 m_end = m_run + npages; 2988 for (; error == 0 && m < m_end; m++) { 2989 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2990 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2991 2992 /* 2993 * Racily check for wirings. Races are handled once the object 2994 * lock is held and the page is unmapped. 2995 */ 2996 if (vm_page_wired(m)) 2997 error = EBUSY; 2998 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2999 /* 3000 * The page is relocated if and only if it could be 3001 * laundered or reclaimed by the page daemon. 3002 */ 3003 VM_OBJECT_WLOCK(object); 3004 /* Don't care: PG_NODUMP, PG_ZERO. */ 3005 if (m->object != object || 3006 ((object->flags & OBJ_SWAP) == 0 && 3007 object->type != OBJT_VNODE)) 3008 error = EINVAL; 3009 else if (object->memattr != VM_MEMATTR_DEFAULT) 3010 error = EINVAL; 3011 else if (vm_page_queue(m) != PQ_NONE && 3012 vm_page_tryxbusy(m) != 0) { 3013 if (vm_page_wired(m)) { 3014 vm_page_xunbusy(m); 3015 error = EBUSY; 3016 goto unlock; 3017 } 3018 KASSERT(pmap_page_get_memattr(m) == 3019 VM_MEMATTR_DEFAULT, 3020 ("page %p has an unexpected memattr", m)); 3021 KASSERT(m->oflags == 0, 3022 ("page %p has unexpected oflags", m)); 3023 /* Don't care: PGA_NOSYNC. */ 3024 if (!vm_page_none_valid(m)) { 3025 /* 3026 * First, try to allocate a new page 3027 * that is above "high". Failing 3028 * that, try to allocate a new page 3029 * that is below "m_run". Allocate 3030 * the new page between the end of 3031 * "m_run" and "high" only as a last 3032 * resort. 3033 */ 3034 req = req_class; 3035 if ((m->flags & PG_NODUMP) != 0) 3036 req |= VM_ALLOC_NODUMP; 3037 if (trunc_page(high) != 3038 ~(vm_paddr_t)PAGE_MASK) { 3039 m_new = 3040 vm_page_alloc_noobj_contig( 3041 req, 1, round_page(high), 3042 ~(vm_paddr_t)0, PAGE_SIZE, 3043 0, VM_MEMATTR_DEFAULT); 3044 } else 3045 m_new = NULL; 3046 if (m_new == NULL) { 3047 pa = VM_PAGE_TO_PHYS(m_run); 3048 m_new = 3049 vm_page_alloc_noobj_contig( 3050 req, 1, 0, pa - 1, 3051 PAGE_SIZE, 0, 3052 VM_MEMATTR_DEFAULT); 3053 } 3054 if (m_new == NULL) { 3055 pa += ptoa(npages); 3056 m_new = 3057 vm_page_alloc_noobj_contig( 3058 req, 1, pa, high, PAGE_SIZE, 3059 0, VM_MEMATTR_DEFAULT); 3060 } 3061 if (m_new == NULL) { 3062 vm_page_xunbusy(m); 3063 error = ENOMEM; 3064 goto unlock; 3065 } 3066 3067 /* 3068 * Unmap the page and check for new 3069 * wirings that may have been acquired 3070 * through a pmap lookup. 3071 */ 3072 if (object->ref_count != 0 && 3073 !vm_page_try_remove_all(m)) { 3074 vm_page_xunbusy(m); 3075 vm_page_free(m_new); 3076 error = EBUSY; 3077 goto unlock; 3078 } 3079 3080 /* 3081 * Replace "m" with the new page. For 3082 * vm_page_replace(), "m" must be busy 3083 * and dequeued. Finally, change "m" 3084 * as if vm_page_free() was called. 3085 */ 3086 m_new->a.flags = m->a.flags & 3087 ~PGA_QUEUE_STATE_MASK; 3088 KASSERT(m_new->oflags == VPO_UNMANAGED, 3089 ("page %p is managed", m_new)); 3090 m_new->oflags = 0; 3091 pmap_copy_page(m, m_new); 3092 m_new->valid = m->valid; 3093 m_new->dirty = m->dirty; 3094 m->flags &= ~PG_ZERO; 3095 vm_page_dequeue(m); 3096 if (vm_page_replace_hold(m_new, object, 3097 m->pindex, m) && 3098 vm_page_free_prep(m)) 3099 SLIST_INSERT_HEAD(&free, m, 3100 plinks.s.ss); 3101 3102 /* 3103 * The new page must be deactivated 3104 * before the object is unlocked. 3105 */ 3106 vm_page_deactivate(m_new); 3107 } else { 3108 m->flags &= ~PG_ZERO; 3109 vm_page_dequeue(m); 3110 if (vm_page_free_prep(m)) 3111 SLIST_INSERT_HEAD(&free, m, 3112 plinks.s.ss); 3113 KASSERT(m->dirty == 0, 3114 ("page %p is dirty", m)); 3115 } 3116 } else 3117 error = EBUSY; 3118 unlock: 3119 VM_OBJECT_WUNLOCK(object); 3120 } else { 3121 MPASS(vm_page_domain(m) == domain); 3122 vmd = VM_DOMAIN(domain); 3123 vm_domain_free_lock(vmd); 3124 order = m->order; 3125 if (order < VM_NFREEORDER) { 3126 /* 3127 * The page is enqueued in the physical memory 3128 * allocator's free page queues. Moreover, it 3129 * is the first page in a power-of-two-sized 3130 * run of contiguous free pages. Jump ahead 3131 * to the last page within that run, and 3132 * continue from there. 3133 */ 3134 m += (1 << order) - 1; 3135 } 3136 #if VM_NRESERVLEVEL > 0 3137 else if (vm_reserv_is_page_free(m)) 3138 order = 0; 3139 #endif 3140 vm_domain_free_unlock(vmd); 3141 if (order == VM_NFREEORDER) 3142 error = EINVAL; 3143 } 3144 } 3145 if ((m = SLIST_FIRST(&free)) != NULL) { 3146 int cnt; 3147 3148 vmd = VM_DOMAIN(domain); 3149 cnt = 0; 3150 vm_domain_free_lock(vmd); 3151 do { 3152 MPASS(vm_page_domain(m) == domain); 3153 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3154 vm_phys_free_pages(m, m->pool, 0); 3155 cnt++; 3156 } while ((m = SLIST_FIRST(&free)) != NULL); 3157 vm_domain_free_unlock(vmd); 3158 vm_domain_freecnt_inc(vmd, cnt); 3159 } 3160 return (error); 3161 } 3162 3163 #define NRUNS 16 3164 3165 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 3166 3167 #define MIN_RECLAIM 8 3168 3169 /* 3170 * vm_page_reclaim_contig: 3171 * 3172 * Reclaim allocated, contiguous physical memory satisfying the specified 3173 * conditions by relocating the virtual pages using that physical memory. 3174 * Returns 0 if reclamation is successful, ERANGE if the specified domain 3175 * can't possibly satisfy the reclamation request, or ENOMEM if not 3176 * currently able to reclaim the requested number of pages. Since 3177 * relocation requires the allocation of physical pages, reclamation may 3178 * fail with ENOMEM due to a shortage of free pages. When reclamation 3179 * fails in this manner, callers are expected to perform vm_wait() before 3180 * retrying a failed allocation operation, e.g., vm_page_alloc_contig(). 3181 * 3182 * The caller must always specify an allocation class through "req". 3183 * 3184 * allocation classes: 3185 * VM_ALLOC_NORMAL normal process request 3186 * VM_ALLOC_SYSTEM system *really* needs a page 3187 * VM_ALLOC_INTERRUPT interrupt time request 3188 * 3189 * The optional allocation flags are ignored. 3190 * 3191 * "npages" must be greater than zero. Both "alignment" and "boundary" 3192 * must be a power of two. 3193 */ 3194 int 3195 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3196 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3197 int desired_runs) 3198 { 3199 struct vm_domain *vmd; 3200 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; 3201 u_long count, minalign, reclaimed; 3202 int error, i, min_reclaim, nruns, options, req_class; 3203 int segind, start_segind; 3204 int ret; 3205 3206 KASSERT(npages > 0, ("npages is 0")); 3207 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3208 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3209 3210 ret = ENOMEM; 3211 3212 /* 3213 * If the caller wants to reclaim multiple runs, try to allocate 3214 * space to store the runs. If that fails, fall back to the old 3215 * behavior of just reclaiming MIN_RECLAIM pages. 3216 */ 3217 if (desired_runs > 1) 3218 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3219 M_TEMP, M_NOWAIT); 3220 else 3221 m_runs = NULL; 3222 3223 if (m_runs == NULL) { 3224 m_runs = _m_runs; 3225 nruns = NRUNS; 3226 } else { 3227 nruns = NRUNS + desired_runs - 1; 3228 } 3229 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3230 3231 /* 3232 * The caller will attempt an allocation after some runs have been 3233 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3234 * of vm_phys_alloc_contig(), round up the requested length to the next 3235 * power of two or maximum chunk size, and ensure that each run is 3236 * suitably aligned. 3237 */ 3238 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3239 npages = roundup2(npages, minalign); 3240 if (alignment < ptoa(minalign)) 3241 alignment = ptoa(minalign); 3242 3243 /* 3244 * The page daemon is allowed to dig deeper into the free page list. 3245 */ 3246 req_class = req & VM_ALLOC_CLASS_MASK; 3247 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3248 req_class = VM_ALLOC_SYSTEM; 3249 3250 start_segind = vm_phys_lookup_segind(low); 3251 3252 /* 3253 * Return if the number of free pages cannot satisfy the requested 3254 * allocation. 3255 */ 3256 vmd = VM_DOMAIN(domain); 3257 count = vmd->vmd_free_count; 3258 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3259 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3260 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3261 goto done; 3262 3263 /* 3264 * Scan up to three times, relaxing the restrictions ("options") on 3265 * the reclamation of reservations and superpages each time. 3266 */ 3267 for (options = VPSC_NORESERV;;) { 3268 bool phys_range_exists = false; 3269 3270 /* 3271 * Find the highest runs that satisfy the given constraints 3272 * and restrictions, and record them in "m_runs". 3273 */ 3274 count = 0; 3275 segind = start_segind; 3276 while ((segind = vm_phys_find_range(bounds, segind, domain, 3277 npages, low, high)) != -1) { 3278 phys_range_exists = true; 3279 while ((m_run = vm_page_scan_contig(npages, bounds[0], 3280 bounds[1], alignment, boundary, options))) { 3281 bounds[0] = m_run + npages; 3282 m_runs[RUN_INDEX(count, nruns)] = m_run; 3283 count++; 3284 } 3285 segind++; 3286 } 3287 3288 if (!phys_range_exists) { 3289 ret = ERANGE; 3290 goto done; 3291 } 3292 3293 /* 3294 * Reclaim the highest runs in LIFO (descending) order until 3295 * the number of reclaimed pages, "reclaimed", is at least 3296 * "min_reclaim". Reset "reclaimed" each time because each 3297 * reclamation is idempotent, and runs will (likely) recur 3298 * from one scan to the next as restrictions are relaxed. 3299 */ 3300 reclaimed = 0; 3301 for (i = 0; count > 0 && i < nruns; i++) { 3302 count--; 3303 m_run = m_runs[RUN_INDEX(count, nruns)]; 3304 error = vm_page_reclaim_run(req_class, domain, npages, 3305 m_run, high); 3306 if (error == 0) { 3307 reclaimed += npages; 3308 if (reclaimed >= min_reclaim) { 3309 ret = 0; 3310 goto done; 3311 } 3312 } 3313 } 3314 3315 /* 3316 * Either relax the restrictions on the next scan or return if 3317 * the last scan had no restrictions. 3318 */ 3319 if (options == VPSC_NORESERV) 3320 options = VPSC_NOSUPER; 3321 else if (options == VPSC_NOSUPER) 3322 options = VPSC_ANY; 3323 else if (options == VPSC_ANY) { 3324 if (reclaimed != 0) 3325 ret = 0; 3326 goto done; 3327 } 3328 } 3329 done: 3330 if (m_runs != _m_runs) 3331 free(m_runs, M_TEMP); 3332 return (ret); 3333 } 3334 3335 int 3336 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3337 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3338 { 3339 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, 3340 high, alignment, boundary, 1)); 3341 } 3342 3343 int 3344 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3345 u_long alignment, vm_paddr_t boundary) 3346 { 3347 struct vm_domainset_iter di; 3348 int domain, ret, status; 3349 3350 ret = ERANGE; 3351 3352 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 3353 return (ret); 3354 3355 do { 3356 status = vm_page_reclaim_contig_domain(domain, req, npages, low, 3357 high, alignment, boundary); 3358 if (status == 0) 3359 return (0); 3360 else if (status == ERANGE) 3361 vm_domainset_iter_ignore(&di, domain); 3362 else { 3363 KASSERT(status == ENOMEM, ("Unrecognized error %d " 3364 "from vm_page_reclaim_contig_domain()", status)); 3365 ret = ENOMEM; 3366 } 3367 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 3368 3369 return (ret); 3370 } 3371 3372 /* 3373 * Set the domain in the appropriate page level domainset. 3374 */ 3375 void 3376 vm_domain_set(struct vm_domain *vmd) 3377 { 3378 3379 mtx_lock(&vm_domainset_lock); 3380 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3381 vmd->vmd_minset = 1; 3382 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3383 } 3384 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3385 vmd->vmd_severeset = 1; 3386 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3387 } 3388 mtx_unlock(&vm_domainset_lock); 3389 } 3390 3391 /* 3392 * Clear the domain from the appropriate page level domainset. 3393 */ 3394 void 3395 vm_domain_clear(struct vm_domain *vmd) 3396 { 3397 3398 mtx_lock(&vm_domainset_lock); 3399 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3400 vmd->vmd_minset = 0; 3401 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3402 if (vm_min_waiters != 0) { 3403 vm_min_waiters = 0; 3404 wakeup(&vm_min_domains); 3405 } 3406 } 3407 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3408 vmd->vmd_severeset = 0; 3409 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3410 if (vm_severe_waiters != 0) { 3411 vm_severe_waiters = 0; 3412 wakeup(&vm_severe_domains); 3413 } 3414 } 3415 3416 /* 3417 * If pageout daemon needs pages, then tell it that there are 3418 * some free. 3419 */ 3420 if (vmd->vmd_pageout_pages_needed && 3421 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3422 wakeup(&vmd->vmd_pageout_pages_needed); 3423 vmd->vmd_pageout_pages_needed = 0; 3424 } 3425 3426 /* See comments in vm_wait_doms(). */ 3427 if (vm_pageproc_waiters) { 3428 vm_pageproc_waiters = 0; 3429 wakeup(&vm_pageproc_waiters); 3430 } 3431 mtx_unlock(&vm_domainset_lock); 3432 } 3433 3434 /* 3435 * Wait for free pages to exceed the min threshold globally. 3436 */ 3437 void 3438 vm_wait_min(void) 3439 { 3440 3441 mtx_lock(&vm_domainset_lock); 3442 while (vm_page_count_min()) { 3443 vm_min_waiters++; 3444 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3445 } 3446 mtx_unlock(&vm_domainset_lock); 3447 } 3448 3449 /* 3450 * Wait for free pages to exceed the severe threshold globally. 3451 */ 3452 void 3453 vm_wait_severe(void) 3454 { 3455 3456 mtx_lock(&vm_domainset_lock); 3457 while (vm_page_count_severe()) { 3458 vm_severe_waiters++; 3459 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3460 "vmwait", 0); 3461 } 3462 mtx_unlock(&vm_domainset_lock); 3463 } 3464 3465 u_int 3466 vm_wait_count(void) 3467 { 3468 3469 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3470 } 3471 3472 int 3473 vm_wait_doms(const domainset_t *wdoms, int mflags) 3474 { 3475 int error; 3476 3477 error = 0; 3478 3479 /* 3480 * We use racey wakeup synchronization to avoid expensive global 3481 * locking for the pageproc when sleeping with a non-specific vm_wait. 3482 * To handle this, we only sleep for one tick in this instance. It 3483 * is expected that most allocations for the pageproc will come from 3484 * kmem or vm_page_grab* which will use the more specific and 3485 * race-free vm_wait_domain(). 3486 */ 3487 if (curproc == pageproc) { 3488 mtx_lock(&vm_domainset_lock); 3489 vm_pageproc_waiters++; 3490 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3491 PVM | PDROP | mflags, "pageprocwait", 1); 3492 } else { 3493 /* 3494 * XXX Ideally we would wait only until the allocation could 3495 * be satisfied. This condition can cause new allocators to 3496 * consume all freed pages while old allocators wait. 3497 */ 3498 mtx_lock(&vm_domainset_lock); 3499 if (vm_page_count_min_set(wdoms)) { 3500 if (pageproc == NULL) 3501 panic("vm_wait in early boot"); 3502 vm_min_waiters++; 3503 error = msleep(&vm_min_domains, &vm_domainset_lock, 3504 PVM | PDROP | mflags, "vmwait", 0); 3505 } else 3506 mtx_unlock(&vm_domainset_lock); 3507 } 3508 return (error); 3509 } 3510 3511 /* 3512 * vm_wait_domain: 3513 * 3514 * Sleep until free pages are available for allocation. 3515 * - Called in various places after failed memory allocations. 3516 */ 3517 void 3518 vm_wait_domain(int domain) 3519 { 3520 struct vm_domain *vmd; 3521 domainset_t wdom; 3522 3523 vmd = VM_DOMAIN(domain); 3524 vm_domain_free_assert_unlocked(vmd); 3525 3526 if (curproc == pageproc) { 3527 mtx_lock(&vm_domainset_lock); 3528 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3529 vmd->vmd_pageout_pages_needed = 1; 3530 msleep(&vmd->vmd_pageout_pages_needed, 3531 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3532 } else 3533 mtx_unlock(&vm_domainset_lock); 3534 } else { 3535 DOMAINSET_ZERO(&wdom); 3536 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3537 vm_wait_doms(&wdom, 0); 3538 } 3539 } 3540 3541 static int 3542 vm_wait_flags(vm_object_t obj, int mflags) 3543 { 3544 struct domainset *d; 3545 3546 d = NULL; 3547 3548 /* 3549 * Carefully fetch pointers only once: the struct domainset 3550 * itself is ummutable but the pointer might change. 3551 */ 3552 if (obj != NULL) 3553 d = obj->domain.dr_policy; 3554 if (d == NULL) 3555 d = curthread->td_domain.dr_policy; 3556 3557 return (vm_wait_doms(&d->ds_mask, mflags)); 3558 } 3559 3560 /* 3561 * vm_wait: 3562 * 3563 * Sleep until free pages are available for allocation in the 3564 * affinity domains of the obj. If obj is NULL, the domain set 3565 * for the calling thread is used. 3566 * Called in various places after failed memory allocations. 3567 */ 3568 void 3569 vm_wait(vm_object_t obj) 3570 { 3571 (void)vm_wait_flags(obj, 0); 3572 } 3573 3574 int 3575 vm_wait_intr(vm_object_t obj) 3576 { 3577 return (vm_wait_flags(obj, PCATCH)); 3578 } 3579 3580 /* 3581 * vm_domain_alloc_fail: 3582 * 3583 * Called when a page allocation function fails. Informs the 3584 * pagedaemon and performs the requested wait. Requires the 3585 * domain_free and object lock on entry. Returns with the 3586 * object lock held and free lock released. Returns an error when 3587 * retry is necessary. 3588 * 3589 */ 3590 static int 3591 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3592 { 3593 3594 vm_domain_free_assert_unlocked(vmd); 3595 3596 atomic_add_int(&vmd->vmd_pageout_deficit, 3597 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3598 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3599 if (object != NULL) 3600 VM_OBJECT_WUNLOCK(object); 3601 vm_wait_domain(vmd->vmd_domain); 3602 if (object != NULL) 3603 VM_OBJECT_WLOCK(object); 3604 if (req & VM_ALLOC_WAITOK) 3605 return (EAGAIN); 3606 } 3607 3608 return (0); 3609 } 3610 3611 /* 3612 * vm_waitpfault: 3613 * 3614 * Sleep until free pages are available for allocation. 3615 * - Called only in vm_fault so that processes page faulting 3616 * can be easily tracked. 3617 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3618 * processes will be able to grab memory first. Do not change 3619 * this balance without careful testing first. 3620 */ 3621 void 3622 vm_waitpfault(struct domainset *dset, int timo) 3623 { 3624 3625 /* 3626 * XXX Ideally we would wait only until the allocation could 3627 * be satisfied. This condition can cause new allocators to 3628 * consume all freed pages while old allocators wait. 3629 */ 3630 mtx_lock(&vm_domainset_lock); 3631 if (vm_page_count_min_set(&dset->ds_mask)) { 3632 vm_min_waiters++; 3633 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3634 "pfault", timo); 3635 } else 3636 mtx_unlock(&vm_domainset_lock); 3637 } 3638 3639 static struct vm_pagequeue * 3640 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3641 { 3642 3643 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3644 } 3645 3646 #ifdef INVARIANTS 3647 static struct vm_pagequeue * 3648 vm_page_pagequeue(vm_page_t m) 3649 { 3650 3651 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3652 } 3653 #endif 3654 3655 static __always_inline bool 3656 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, 3657 vm_page_astate_t new) 3658 { 3659 vm_page_astate_t tmp; 3660 3661 tmp = *old; 3662 do { 3663 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3664 return (true); 3665 counter_u64_add(pqstate_commit_retries, 1); 3666 } while (old->_bits == tmp._bits); 3667 3668 return (false); 3669 } 3670 3671 /* 3672 * Do the work of committing a queue state update that moves the page out of 3673 * its current queue. 3674 */ 3675 static bool 3676 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3677 vm_page_astate_t *old, vm_page_astate_t new) 3678 { 3679 vm_page_t next; 3680 3681 vm_pagequeue_assert_locked(pq); 3682 KASSERT(vm_page_pagequeue(m) == pq, 3683 ("%s: queue %p does not match page %p", __func__, pq, m)); 3684 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3685 ("%s: invalid queue indices %d %d", 3686 __func__, old->queue, new.queue)); 3687 3688 /* 3689 * Once the queue index of the page changes there is nothing 3690 * synchronizing with further updates to the page's physical 3691 * queue state. Therefore we must speculatively remove the page 3692 * from the queue now and be prepared to roll back if the queue 3693 * state update fails. If the page is not physically enqueued then 3694 * we just update its queue index. 3695 */ 3696 if ((old->flags & PGA_ENQUEUED) != 0) { 3697 new.flags &= ~PGA_ENQUEUED; 3698 next = TAILQ_NEXT(m, plinks.q); 3699 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3700 vm_pagequeue_cnt_dec(pq); 3701 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3702 if (next == NULL) 3703 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3704 else 3705 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3706 vm_pagequeue_cnt_inc(pq); 3707 return (false); 3708 } else { 3709 return (true); 3710 } 3711 } else { 3712 return (vm_page_pqstate_fcmpset(m, old, new)); 3713 } 3714 } 3715 3716 static bool 3717 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3718 vm_page_astate_t new) 3719 { 3720 struct vm_pagequeue *pq; 3721 vm_page_astate_t as; 3722 bool ret; 3723 3724 pq = _vm_page_pagequeue(m, old->queue); 3725 3726 /* 3727 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3728 * corresponding page queue lock is held. 3729 */ 3730 vm_pagequeue_lock(pq); 3731 as = vm_page_astate_load(m); 3732 if (__predict_false(as._bits != old->_bits)) { 3733 *old = as; 3734 ret = false; 3735 } else { 3736 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3737 } 3738 vm_pagequeue_unlock(pq); 3739 return (ret); 3740 } 3741 3742 /* 3743 * Commit a queue state update that enqueues or requeues a page. 3744 */ 3745 static bool 3746 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3747 vm_page_astate_t *old, vm_page_astate_t new) 3748 { 3749 struct vm_domain *vmd; 3750 3751 vm_pagequeue_assert_locked(pq); 3752 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3753 ("%s: invalid queue indices %d %d", 3754 __func__, old->queue, new.queue)); 3755 3756 new.flags |= PGA_ENQUEUED; 3757 if (!vm_page_pqstate_fcmpset(m, old, new)) 3758 return (false); 3759 3760 if ((old->flags & PGA_ENQUEUED) != 0) 3761 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3762 else 3763 vm_pagequeue_cnt_inc(pq); 3764 3765 /* 3766 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3767 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3768 * applied, even if it was set first. 3769 */ 3770 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3771 vmd = vm_pagequeue_domain(m); 3772 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3773 ("%s: invalid page queue for page %p", __func__, m)); 3774 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3775 } else { 3776 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3777 } 3778 return (true); 3779 } 3780 3781 /* 3782 * Commit a queue state update that encodes a request for a deferred queue 3783 * operation. 3784 */ 3785 static bool 3786 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3787 vm_page_astate_t new) 3788 { 3789 3790 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3791 ("%s: invalid state, queue %d flags %x", 3792 __func__, new.queue, new.flags)); 3793 3794 if (old->_bits != new._bits && 3795 !vm_page_pqstate_fcmpset(m, old, new)) 3796 return (false); 3797 vm_page_pqbatch_submit(m, new.queue); 3798 return (true); 3799 } 3800 3801 /* 3802 * A generic queue state update function. This handles more cases than the 3803 * specialized functions above. 3804 */ 3805 bool 3806 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3807 { 3808 3809 if (old->_bits == new._bits) 3810 return (true); 3811 3812 if (old->queue != PQ_NONE && new.queue != old->queue) { 3813 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3814 return (false); 3815 if (new.queue != PQ_NONE) 3816 vm_page_pqbatch_submit(m, new.queue); 3817 } else { 3818 if (!vm_page_pqstate_fcmpset(m, old, new)) 3819 return (false); 3820 if (new.queue != PQ_NONE && 3821 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3822 vm_page_pqbatch_submit(m, new.queue); 3823 } 3824 return (true); 3825 } 3826 3827 /* 3828 * Apply deferred queue state updates to a page. 3829 */ 3830 static inline void 3831 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3832 { 3833 vm_page_astate_t new, old; 3834 3835 CRITICAL_ASSERT(curthread); 3836 vm_pagequeue_assert_locked(pq); 3837 KASSERT(queue < PQ_COUNT, 3838 ("%s: invalid queue index %d", __func__, queue)); 3839 KASSERT(pq == _vm_page_pagequeue(m, queue), 3840 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3841 3842 for (old = vm_page_astate_load(m);;) { 3843 if (__predict_false(old.queue != queue || 3844 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3845 counter_u64_add(queue_nops, 1); 3846 break; 3847 } 3848 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3849 ("%s: page %p is unmanaged", __func__, m)); 3850 3851 new = old; 3852 if ((old.flags & PGA_DEQUEUE) != 0) { 3853 new.flags &= ~PGA_QUEUE_OP_MASK; 3854 new.queue = PQ_NONE; 3855 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3856 m, &old, new))) { 3857 counter_u64_add(queue_ops, 1); 3858 break; 3859 } 3860 } else { 3861 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3862 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3863 m, &old, new))) { 3864 counter_u64_add(queue_ops, 1); 3865 break; 3866 } 3867 } 3868 } 3869 } 3870 3871 static void 3872 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3873 uint8_t queue) 3874 { 3875 int i; 3876 3877 for (i = 0; i < bq->bq_cnt; i++) 3878 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3879 vm_batchqueue_init(bq); 3880 } 3881 3882 /* 3883 * vm_page_pqbatch_submit: [ internal use only ] 3884 * 3885 * Enqueue a page in the specified page queue's batched work queue. 3886 * The caller must have encoded the requested operation in the page 3887 * structure's a.flags field. 3888 */ 3889 void 3890 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3891 { 3892 struct vm_batchqueue *bq; 3893 struct vm_pagequeue *pq; 3894 int domain, slots_remaining; 3895 3896 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3897 3898 domain = vm_page_domain(m); 3899 critical_enter(); 3900 bq = DPCPU_PTR(pqbatch[domain][queue]); 3901 slots_remaining = vm_batchqueue_insert(bq, m); 3902 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 3903 /* keep building the bq */ 3904 critical_exit(); 3905 return; 3906 } else if (slots_remaining > 0 ) { 3907 /* Try to process the bq if we can get the lock */ 3908 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3909 if (vm_pagequeue_trylock(pq)) { 3910 vm_pqbatch_process(pq, bq, queue); 3911 vm_pagequeue_unlock(pq); 3912 } 3913 critical_exit(); 3914 return; 3915 } 3916 critical_exit(); 3917 3918 /* if we make it here, the bq is full so wait for the lock */ 3919 3920 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3921 vm_pagequeue_lock(pq); 3922 critical_enter(); 3923 bq = DPCPU_PTR(pqbatch[domain][queue]); 3924 vm_pqbatch_process(pq, bq, queue); 3925 vm_pqbatch_process_page(pq, m, queue); 3926 vm_pagequeue_unlock(pq); 3927 critical_exit(); 3928 } 3929 3930 /* 3931 * vm_page_pqbatch_drain: [ internal use only ] 3932 * 3933 * Force all per-CPU page queue batch queues to be drained. This is 3934 * intended for use in severe memory shortages, to ensure that pages 3935 * do not remain stuck in the batch queues. 3936 */ 3937 void 3938 vm_page_pqbatch_drain(void) 3939 { 3940 struct thread *td; 3941 struct vm_domain *vmd; 3942 struct vm_pagequeue *pq; 3943 int cpu, domain, queue; 3944 3945 td = curthread; 3946 CPU_FOREACH(cpu) { 3947 thread_lock(td); 3948 sched_bind(td, cpu); 3949 thread_unlock(td); 3950 3951 for (domain = 0; domain < vm_ndomains; domain++) { 3952 vmd = VM_DOMAIN(domain); 3953 for (queue = 0; queue < PQ_COUNT; queue++) { 3954 pq = &vmd->vmd_pagequeues[queue]; 3955 vm_pagequeue_lock(pq); 3956 critical_enter(); 3957 vm_pqbatch_process(pq, 3958 DPCPU_PTR(pqbatch[domain][queue]), queue); 3959 critical_exit(); 3960 vm_pagequeue_unlock(pq); 3961 } 3962 } 3963 } 3964 thread_lock(td); 3965 sched_unbind(td); 3966 thread_unlock(td); 3967 } 3968 3969 /* 3970 * vm_page_dequeue_deferred: [ internal use only ] 3971 * 3972 * Request removal of the given page from its current page 3973 * queue. Physical removal from the queue may be deferred 3974 * indefinitely. 3975 */ 3976 void 3977 vm_page_dequeue_deferred(vm_page_t m) 3978 { 3979 vm_page_astate_t new, old; 3980 3981 old = vm_page_astate_load(m); 3982 do { 3983 if (old.queue == PQ_NONE) { 3984 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3985 ("%s: page %p has unexpected queue state", 3986 __func__, m)); 3987 break; 3988 } 3989 new = old; 3990 new.flags |= PGA_DEQUEUE; 3991 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3992 } 3993 3994 /* 3995 * vm_page_dequeue: 3996 * 3997 * Remove the page from whichever page queue it's in, if any, before 3998 * returning. 3999 */ 4000 void 4001 vm_page_dequeue(vm_page_t m) 4002 { 4003 vm_page_astate_t new, old; 4004 4005 old = vm_page_astate_load(m); 4006 do { 4007 if (__predict_true(old.queue == PQ_NONE)) { 4008 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4009 ("%s: page %p has unexpected queue state", 4010 __func__, m)); 4011 break; 4012 } 4013 new = old; 4014 new.flags &= ~PGA_QUEUE_OP_MASK; 4015 new.queue = PQ_NONE; 4016 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 4017 4018 } 4019 4020 /* 4021 * Schedule the given page for insertion into the specified page queue. 4022 * Physical insertion of the page may be deferred indefinitely. 4023 */ 4024 static void 4025 vm_page_enqueue(vm_page_t m, uint8_t queue) 4026 { 4027 4028 KASSERT(m->a.queue == PQ_NONE && 4029 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 4030 ("%s: page %p is already enqueued", __func__, m)); 4031 KASSERT(m->ref_count > 0, 4032 ("%s: page %p does not carry any references", __func__, m)); 4033 4034 m->a.queue = queue; 4035 if ((m->a.flags & PGA_REQUEUE) == 0) 4036 vm_page_aflag_set(m, PGA_REQUEUE); 4037 vm_page_pqbatch_submit(m, queue); 4038 } 4039 4040 /* 4041 * vm_page_free_prep: 4042 * 4043 * Prepares the given page to be put on the free list, 4044 * disassociating it from any VM object. The caller may return 4045 * the page to the free list only if this function returns true. 4046 * 4047 * The object, if it exists, must be locked, and then the page must 4048 * be xbusy. Otherwise the page must be not busied. A managed 4049 * page must be unmapped. 4050 */ 4051 static bool 4052 vm_page_free_prep(vm_page_t m) 4053 { 4054 4055 /* 4056 * Synchronize with threads that have dropped a reference to this 4057 * page. 4058 */ 4059 atomic_thread_fence_acq(); 4060 4061 #ifdef INVARIANTS 4062 if (vm_check_pg_zero && (m->flags & PG_ZERO) != 0) { 4063 struct sf_buf *sf; 4064 unsigned long *p; 4065 int i; 4066 4067 sched_pin(); 4068 sf = sf_buf_alloc(m, SFB_CPUPRIVATE | SFB_NOWAIT); 4069 if (sf != NULL) { 4070 p = (unsigned long *)sf_buf_kva(sf); 4071 for (i = 0; i < PAGE_SIZE / sizeof(*p); i++, p++) { 4072 KASSERT(*p == 0, 4073 ("zerocheck failed page %p PG_ZERO %d %jx", 4074 m, i, (uintmax_t)*p)); 4075 } 4076 sf_buf_free(sf); 4077 } 4078 sched_unpin(); 4079 } 4080 #endif 4081 if ((m->oflags & VPO_UNMANAGED) == 0) { 4082 KASSERT(!pmap_page_is_mapped(m), 4083 ("vm_page_free_prep: freeing mapped page %p", m)); 4084 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 4085 ("vm_page_free_prep: mapping flags set in page %p", m)); 4086 } else { 4087 KASSERT(m->a.queue == PQ_NONE, 4088 ("vm_page_free_prep: unmanaged page %p is queued", m)); 4089 } 4090 VM_CNT_INC(v_tfree); 4091 4092 if (m->object != NULL) { 4093 vm_page_radix_remove(m); 4094 vm_page_free_object_prep(m); 4095 } else 4096 vm_page_assert_unbusied(m); 4097 4098 vm_page_busy_free(m); 4099 4100 /* 4101 * If fictitious remove object association and 4102 * return. 4103 */ 4104 if ((m->flags & PG_FICTITIOUS) != 0) { 4105 KASSERT(m->ref_count == 1, 4106 ("fictitious page %p is referenced", m)); 4107 KASSERT(m->a.queue == PQ_NONE, 4108 ("fictitious page %p is queued", m)); 4109 return (false); 4110 } 4111 4112 /* 4113 * Pages need not be dequeued before they are returned to the physical 4114 * memory allocator, but they must at least be marked for a deferred 4115 * dequeue. 4116 */ 4117 if ((m->oflags & VPO_UNMANAGED) == 0) 4118 vm_page_dequeue_deferred(m); 4119 4120 m->valid = 0; 4121 vm_page_undirty(m); 4122 4123 if (m->ref_count != 0) 4124 panic("vm_page_free_prep: page %p has references", m); 4125 4126 /* 4127 * Restore the default memory attribute to the page. 4128 */ 4129 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 4130 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 4131 4132 #if VM_NRESERVLEVEL > 0 4133 /* 4134 * Determine whether the page belongs to a reservation. If the page was 4135 * allocated from a per-CPU cache, it cannot belong to a reservation, so 4136 * as an optimization, we avoid the check in that case. 4137 */ 4138 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 4139 return (false); 4140 #endif 4141 4142 return (true); 4143 } 4144 4145 /* 4146 * vm_page_free_toq: 4147 * 4148 * Returns the given page to the free list, disassociating it 4149 * from any VM object. 4150 * 4151 * The object must be locked. The page must be exclusively busied if it 4152 * belongs to an object. 4153 */ 4154 static void 4155 vm_page_free_toq(vm_page_t m) 4156 { 4157 struct vm_domain *vmd; 4158 uma_zone_t zone; 4159 4160 if (!vm_page_free_prep(m)) 4161 return; 4162 4163 vmd = vm_pagequeue_domain(m); 4164 if (__predict_false((m->flags & PG_NOFREE) != 0)) { 4165 vm_page_free_nofree(vmd, m); 4166 return; 4167 } 4168 zone = vmd->vmd_pgcache[m->pool].zone; 4169 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4170 uma_zfree(zone, m); 4171 return; 4172 } 4173 vm_domain_free_lock(vmd); 4174 vm_phys_free_pages(m, m->pool, 0); 4175 vm_domain_free_unlock(vmd); 4176 vm_domain_freecnt_inc(vmd, 1); 4177 } 4178 4179 /* 4180 * vm_page_free_pages_toq: 4181 * 4182 * Returns a list of pages to the free list, disassociating it 4183 * from any VM object. In other words, this is equivalent to 4184 * calling vm_page_free_toq() for each page of a list of VM objects. 4185 */ 4186 int 4187 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4188 { 4189 vm_page_t m; 4190 int count; 4191 4192 if (SLIST_EMPTY(free)) 4193 return (0); 4194 4195 count = 0; 4196 while ((m = SLIST_FIRST(free)) != NULL) { 4197 count++; 4198 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4199 vm_page_free_toq(m); 4200 } 4201 4202 if (update_wire_count) 4203 vm_wire_sub(count); 4204 return (count); 4205 } 4206 4207 /* 4208 * Mark this page as wired down. For managed pages, this prevents reclamation 4209 * by the page daemon, or when the containing object, if any, is destroyed. 4210 */ 4211 void 4212 vm_page_wire(vm_page_t m) 4213 { 4214 u_int old; 4215 4216 #ifdef INVARIANTS 4217 if (m->object != NULL && !vm_page_busied(m) && 4218 !vm_object_busied(m->object)) 4219 VM_OBJECT_ASSERT_LOCKED(m->object); 4220 #endif 4221 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4222 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4223 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4224 4225 old = atomic_fetchadd_int(&m->ref_count, 1); 4226 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4227 ("vm_page_wire: counter overflow for page %p", m)); 4228 if (VPRC_WIRE_COUNT(old) == 0) { 4229 if ((m->oflags & VPO_UNMANAGED) == 0) 4230 vm_page_aflag_set(m, PGA_DEQUEUE); 4231 vm_wire_add(1); 4232 } 4233 } 4234 4235 /* 4236 * Attempt to wire a mapped page following a pmap lookup of that page. 4237 * This may fail if a thread is concurrently tearing down mappings of the page. 4238 * The transient failure is acceptable because it translates to the 4239 * failure of the caller pmap_extract_and_hold(), which should be then 4240 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4241 */ 4242 bool 4243 vm_page_wire_mapped(vm_page_t m) 4244 { 4245 u_int old; 4246 4247 old = atomic_load_int(&m->ref_count); 4248 do { 4249 KASSERT(old > 0, 4250 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4251 if ((old & VPRC_BLOCKED) != 0) 4252 return (false); 4253 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4254 4255 if (VPRC_WIRE_COUNT(old) == 0) { 4256 if ((m->oflags & VPO_UNMANAGED) == 0) 4257 vm_page_aflag_set(m, PGA_DEQUEUE); 4258 vm_wire_add(1); 4259 } 4260 return (true); 4261 } 4262 4263 /* 4264 * Release a wiring reference to a managed page. If the page still belongs to 4265 * an object, update its position in the page queues to reflect the reference. 4266 * If the wiring was the last reference to the page, free the page. 4267 */ 4268 static void 4269 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4270 { 4271 u_int old; 4272 4273 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4274 ("%s: page %p is unmanaged", __func__, m)); 4275 4276 /* 4277 * Update LRU state before releasing the wiring reference. 4278 * Use a release store when updating the reference count to 4279 * synchronize with vm_page_free_prep(). 4280 */ 4281 old = atomic_load_int(&m->ref_count); 4282 do { 4283 u_int count; 4284 4285 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4286 ("vm_page_unwire: wire count underflow for page %p", m)); 4287 4288 count = old & ~VPRC_BLOCKED; 4289 if (count > VPRC_OBJREF + 1) { 4290 /* 4291 * The page has at least one other wiring reference. An 4292 * earlier iteration of this loop may have called 4293 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4294 * re-set it if necessary. 4295 */ 4296 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4297 vm_page_aflag_set(m, PGA_DEQUEUE); 4298 } else if (count == VPRC_OBJREF + 1) { 4299 /* 4300 * This is the last wiring. Clear PGA_DEQUEUE and 4301 * update the page's queue state to reflect the 4302 * reference. If the page does not belong to an object 4303 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4304 * clear leftover queue state. 4305 */ 4306 vm_page_release_toq(m, nqueue, noreuse); 4307 } else if (count == 1) { 4308 vm_page_aflag_clear(m, PGA_DEQUEUE); 4309 } 4310 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4311 4312 if (VPRC_WIRE_COUNT(old) == 1) { 4313 vm_wire_sub(1); 4314 if (old == 1) 4315 vm_page_free(m); 4316 } 4317 } 4318 4319 /* 4320 * Release one wiring of the specified page, potentially allowing it to be 4321 * paged out. 4322 * 4323 * Only managed pages belonging to an object can be paged out. If the number 4324 * of wirings transitions to zero and the page is eligible for page out, then 4325 * the page is added to the specified paging queue. If the released wiring 4326 * represented the last reference to the page, the page is freed. 4327 */ 4328 void 4329 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4330 { 4331 4332 KASSERT(nqueue < PQ_COUNT, 4333 ("vm_page_unwire: invalid queue %u request for page %p", 4334 nqueue, m)); 4335 4336 if ((m->oflags & VPO_UNMANAGED) != 0) { 4337 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4338 vm_page_free(m); 4339 return; 4340 } 4341 vm_page_unwire_managed(m, nqueue, false); 4342 } 4343 4344 /* 4345 * Unwire a page without (re-)inserting it into a page queue. It is up 4346 * to the caller to enqueue, requeue, or free the page as appropriate. 4347 * In most cases involving managed pages, vm_page_unwire() should be used 4348 * instead. 4349 */ 4350 bool 4351 vm_page_unwire_noq(vm_page_t m) 4352 { 4353 u_int old; 4354 4355 old = vm_page_drop(m, 1); 4356 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4357 ("%s: counter underflow for page %p", __func__, m)); 4358 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4359 ("%s: missing ref on fictitious page %p", __func__, m)); 4360 4361 if (VPRC_WIRE_COUNT(old) > 1) 4362 return (false); 4363 if ((m->oflags & VPO_UNMANAGED) == 0) 4364 vm_page_aflag_clear(m, PGA_DEQUEUE); 4365 vm_wire_sub(1); 4366 return (true); 4367 } 4368 4369 /* 4370 * Ensure that the page ends up in the specified page queue. If the page is 4371 * active or being moved to the active queue, ensure that its act_count is 4372 * at least ACT_INIT but do not otherwise mess with it. 4373 */ 4374 static __always_inline void 4375 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4376 { 4377 vm_page_astate_t old, new; 4378 4379 KASSERT(m->ref_count > 0, 4380 ("%s: page %p does not carry any references", __func__, m)); 4381 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4382 ("%s: invalid flags %x", __func__, nflag)); 4383 4384 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4385 return; 4386 4387 old = vm_page_astate_load(m); 4388 do { 4389 if ((old.flags & PGA_DEQUEUE) != 0) 4390 break; 4391 new = old; 4392 new.flags &= ~PGA_QUEUE_OP_MASK; 4393 if (nqueue == PQ_ACTIVE) 4394 new.act_count = max(old.act_count, ACT_INIT); 4395 if (old.queue == nqueue) { 4396 /* 4397 * There is no need to requeue pages already in the 4398 * active queue. 4399 */ 4400 if (nqueue != PQ_ACTIVE || 4401 (old.flags & PGA_ENQUEUED) == 0) 4402 new.flags |= nflag; 4403 } else { 4404 new.flags |= nflag; 4405 new.queue = nqueue; 4406 } 4407 } while (!vm_page_pqstate_commit(m, &old, new)); 4408 } 4409 4410 /* 4411 * Put the specified page on the active list (if appropriate). 4412 */ 4413 void 4414 vm_page_activate(vm_page_t m) 4415 { 4416 4417 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4418 } 4419 4420 /* 4421 * Move the specified page to the tail of the inactive queue, or requeue 4422 * the page if it is already in the inactive queue. 4423 */ 4424 void 4425 vm_page_deactivate(vm_page_t m) 4426 { 4427 4428 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4429 } 4430 4431 void 4432 vm_page_deactivate_noreuse(vm_page_t m) 4433 { 4434 4435 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4436 } 4437 4438 /* 4439 * Put a page in the laundry, or requeue it if it is already there. 4440 */ 4441 void 4442 vm_page_launder(vm_page_t m) 4443 { 4444 4445 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4446 } 4447 4448 /* 4449 * Put a page in the PQ_UNSWAPPABLE holding queue. 4450 */ 4451 void 4452 vm_page_unswappable(vm_page_t m) 4453 { 4454 4455 VM_OBJECT_ASSERT_LOCKED(m->object); 4456 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4457 ("page %p already unswappable", m)); 4458 4459 vm_page_dequeue(m); 4460 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4461 } 4462 4463 /* 4464 * Release a page back to the page queues in preparation for unwiring. 4465 */ 4466 static void 4467 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4468 { 4469 vm_page_astate_t old, new; 4470 uint16_t nflag; 4471 4472 /* 4473 * Use a check of the valid bits to determine whether we should 4474 * accelerate reclamation of the page. The object lock might not be 4475 * held here, in which case the check is racy. At worst we will either 4476 * accelerate reclamation of a valid page and violate LRU, or 4477 * unnecessarily defer reclamation of an invalid page. 4478 * 4479 * If we were asked to not cache the page, place it near the head of the 4480 * inactive queue so that is reclaimed sooner. 4481 */ 4482 if (noreuse || vm_page_none_valid(m)) { 4483 nqueue = PQ_INACTIVE; 4484 nflag = PGA_REQUEUE_HEAD; 4485 } else { 4486 nflag = PGA_REQUEUE; 4487 } 4488 4489 old = vm_page_astate_load(m); 4490 do { 4491 new = old; 4492 4493 /* 4494 * If the page is already in the active queue and we are not 4495 * trying to accelerate reclamation, simply mark it as 4496 * referenced and avoid any queue operations. 4497 */ 4498 new.flags &= ~PGA_QUEUE_OP_MASK; 4499 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4500 (old.flags & PGA_ENQUEUED) != 0) 4501 new.flags |= PGA_REFERENCED; 4502 else { 4503 new.flags |= nflag; 4504 new.queue = nqueue; 4505 } 4506 } while (!vm_page_pqstate_commit(m, &old, new)); 4507 } 4508 4509 /* 4510 * Unwire a page and either attempt to free it or re-add it to the page queues. 4511 */ 4512 void 4513 vm_page_release(vm_page_t m, int flags) 4514 { 4515 vm_object_t object; 4516 4517 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4518 ("vm_page_release: page %p is unmanaged", m)); 4519 4520 if ((flags & VPR_TRYFREE) != 0) { 4521 for (;;) { 4522 object = atomic_load_ptr(&m->object); 4523 if (object == NULL) 4524 break; 4525 /* Depends on type-stability. */ 4526 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4527 break; 4528 if (object == m->object) { 4529 vm_page_release_locked(m, flags); 4530 VM_OBJECT_WUNLOCK(object); 4531 return; 4532 } 4533 VM_OBJECT_WUNLOCK(object); 4534 } 4535 } 4536 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4537 } 4538 4539 /* See vm_page_release(). */ 4540 void 4541 vm_page_release_locked(vm_page_t m, int flags) 4542 { 4543 4544 VM_OBJECT_ASSERT_WLOCKED(m->object); 4545 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4546 ("vm_page_release_locked: page %p is unmanaged", m)); 4547 4548 if (vm_page_unwire_noq(m)) { 4549 if ((flags & VPR_TRYFREE) != 0 && 4550 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4551 m->dirty == 0 && vm_page_tryxbusy(m)) { 4552 /* 4553 * An unlocked lookup may have wired the page before the 4554 * busy lock was acquired, in which case the page must 4555 * not be freed. 4556 */ 4557 if (__predict_true(!vm_page_wired(m))) { 4558 vm_page_free(m); 4559 return; 4560 } 4561 vm_page_xunbusy(m); 4562 } else { 4563 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4564 } 4565 } 4566 } 4567 4568 static bool 4569 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4570 { 4571 u_int old; 4572 4573 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4574 ("vm_page_try_blocked_op: page %p has no object", m)); 4575 KASSERT(vm_page_busied(m), 4576 ("vm_page_try_blocked_op: page %p is not busy", m)); 4577 VM_OBJECT_ASSERT_LOCKED(m->object); 4578 4579 old = atomic_load_int(&m->ref_count); 4580 do { 4581 KASSERT(old != 0, 4582 ("vm_page_try_blocked_op: page %p has no references", m)); 4583 KASSERT((old & VPRC_BLOCKED) == 0, 4584 ("vm_page_try_blocked_op: page %p blocks wirings", m)); 4585 if (VPRC_WIRE_COUNT(old) != 0) 4586 return (false); 4587 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4588 4589 (op)(m); 4590 4591 /* 4592 * If the object is read-locked, new wirings may be created via an 4593 * object lookup. 4594 */ 4595 old = vm_page_drop(m, VPRC_BLOCKED); 4596 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4597 old == (VPRC_BLOCKED | VPRC_OBJREF), 4598 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4599 old, m)); 4600 return (true); 4601 } 4602 4603 /* 4604 * Atomically check for wirings and remove all mappings of the page. 4605 */ 4606 bool 4607 vm_page_try_remove_all(vm_page_t m) 4608 { 4609 4610 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4611 } 4612 4613 /* 4614 * Atomically check for wirings and remove all writeable mappings of the page. 4615 */ 4616 bool 4617 vm_page_try_remove_write(vm_page_t m) 4618 { 4619 4620 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4621 } 4622 4623 /* 4624 * vm_page_advise 4625 * 4626 * Apply the specified advice to the given page. 4627 */ 4628 void 4629 vm_page_advise(vm_page_t m, int advice) 4630 { 4631 4632 VM_OBJECT_ASSERT_WLOCKED(m->object); 4633 vm_page_assert_xbusied(m); 4634 4635 if (advice == MADV_FREE) 4636 /* 4637 * Mark the page clean. This will allow the page to be freed 4638 * without first paging it out. MADV_FREE pages are often 4639 * quickly reused by malloc(3), so we do not do anything that 4640 * would result in a page fault on a later access. 4641 */ 4642 vm_page_undirty(m); 4643 else if (advice != MADV_DONTNEED) { 4644 if (advice == MADV_WILLNEED) 4645 vm_page_activate(m); 4646 return; 4647 } 4648 4649 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4650 vm_page_dirty(m); 4651 4652 /* 4653 * Clear any references to the page. Otherwise, the page daemon will 4654 * immediately reactivate the page. 4655 */ 4656 vm_page_aflag_clear(m, PGA_REFERENCED); 4657 4658 /* 4659 * Place clean pages near the head of the inactive queue rather than 4660 * the tail, thus defeating the queue's LRU operation and ensuring that 4661 * the page will be reused quickly. Dirty pages not already in the 4662 * laundry are moved there. 4663 */ 4664 if (m->dirty == 0) 4665 vm_page_deactivate_noreuse(m); 4666 else if (!vm_page_in_laundry(m)) 4667 vm_page_launder(m); 4668 } 4669 4670 /* 4671 * vm_page_grab_release 4672 * 4673 * Helper routine for grab functions to release busy on return. 4674 */ 4675 static inline void 4676 vm_page_grab_release(vm_page_t m, int allocflags) 4677 { 4678 4679 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4680 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4681 vm_page_sunbusy(m); 4682 else 4683 vm_page_xunbusy(m); 4684 } 4685 } 4686 4687 /* 4688 * vm_page_grab_sleep 4689 * 4690 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4691 * if the caller should retry and false otherwise. 4692 * 4693 * If the object is locked on entry the object will be unlocked with 4694 * false returns and still locked but possibly having been dropped 4695 * with true returns. 4696 */ 4697 static bool 4698 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4699 const char *wmesg, int allocflags, bool locked) 4700 { 4701 4702 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4703 return (false); 4704 4705 /* 4706 * Reference the page before unlocking and sleeping so that 4707 * the page daemon is less likely to reclaim it. 4708 */ 4709 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4710 vm_page_reference(m); 4711 4712 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4713 locked) 4714 VM_OBJECT_WLOCK(object); 4715 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4716 return (false); 4717 4718 return (true); 4719 } 4720 4721 /* 4722 * Assert that the grab flags are valid. 4723 */ 4724 static inline void 4725 vm_page_grab_check(int allocflags) 4726 { 4727 4728 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4729 (allocflags & VM_ALLOC_WIRED) != 0, 4730 ("vm_page_grab*: the pages must be busied or wired")); 4731 4732 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4733 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4734 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4735 } 4736 4737 /* 4738 * Calculate the page allocation flags for grab. 4739 */ 4740 static inline int 4741 vm_page_grab_pflags(int allocflags) 4742 { 4743 int pflags; 4744 4745 pflags = allocflags & 4746 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4747 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); 4748 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4749 pflags |= VM_ALLOC_WAITFAIL; 4750 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4751 pflags |= VM_ALLOC_SBUSY; 4752 4753 return (pflags); 4754 } 4755 4756 /* 4757 * Grab a page, waiting until we are woken up due to the page changing state. 4758 * We keep on waiting, if the page continues to be in the object, unless 4759 * allocflags forbid waiting. 4760 * 4761 * The object must be locked on entry. This routine may sleep. The lock will, 4762 * however, be released and reacquired if the routine sleeps. 4763 * 4764 * Return a grabbed page, or NULL. Set *found if a page was found, whether or 4765 * not it was grabbed. 4766 */ 4767 static inline vm_page_t 4768 vm_page_grab_lookup(vm_object_t object, vm_pindex_t pindex, int allocflags, 4769 bool *found, struct pctrie_iter *pages) 4770 { 4771 vm_page_t m; 4772 4773 while ((*found = (m = vm_radix_iter_lookup(pages, pindex)) != NULL) && 4774 !vm_page_tryacquire(m, allocflags)) { 4775 if (!vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4776 allocflags, true)) 4777 return (NULL); 4778 pctrie_iter_reset(pages); 4779 } 4780 return (m); 4781 } 4782 4783 /* 4784 * Grab a page. Use an iterator parameter. Keep on waiting, as long as the page 4785 * exists in the object. If the page doesn't exist, first allocate it and then 4786 * conditionally zero it. 4787 * 4788 * The object must be locked on entry. This routine may sleep. The lock will, 4789 * however, be released and reacquired if the routine sleeps. 4790 */ 4791 vm_page_t 4792 vm_page_grab_iter(vm_object_t object, vm_pindex_t pindex, int allocflags, 4793 struct pctrie_iter *pages) 4794 { 4795 vm_page_t m; 4796 bool found; 4797 4798 VM_OBJECT_ASSERT_WLOCKED(object); 4799 vm_page_grab_check(allocflags); 4800 4801 while ((m = vm_page_grab_lookup( 4802 object, pindex, allocflags, &found, pages)) == NULL) { 4803 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4804 return (NULL); 4805 if (found && 4806 (allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4807 return (NULL); 4808 m = vm_page_alloc_iter(object, pindex, 4809 vm_page_grab_pflags(allocflags), pages); 4810 if (m != NULL) { 4811 if ((allocflags & VM_ALLOC_ZERO) != 0 && 4812 (m->flags & PG_ZERO) == 0) 4813 pmap_zero_page(m); 4814 break; 4815 } 4816 if ((allocflags & 4817 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4818 return (NULL); 4819 } 4820 vm_page_grab_release(m, allocflags); 4821 4822 return (m); 4823 } 4824 4825 /* 4826 * Grab a page. Keep on waiting, as long as the page exists in the object. If 4827 * the page doesn't exist, first allocate it and then conditionally zero it. 4828 * 4829 * The object must be locked on entry. This routine may sleep. The lock will, 4830 * however, be released and reacquired if the routine sleeps. 4831 */ 4832 vm_page_t 4833 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4834 { 4835 struct pctrie_iter pages; 4836 4837 VM_OBJECT_ASSERT_WLOCKED(object); 4838 vm_page_iter_init(&pages, object); 4839 return (vm_page_grab_iter(object, pindex, allocflags, &pages)); 4840 } 4841 4842 /* 4843 * Attempt to validate a page, locklessly acquiring it if necessary, given a 4844 * (object, pindex) tuple and either an invalided page or NULL. The resulting 4845 * page will be validated against the identity tuple, and busied or wired as 4846 * requested. A NULL page returned guarantees that the page was not in radix at 4847 * the time of the call but callers must perform higher level synchronization or 4848 * retry the operation under a lock if they require an atomic answer. This is 4849 * the only lock free validation routine, other routines can depend on the 4850 * resulting page state. 4851 * 4852 * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to 4853 * caller flags. 4854 */ 4855 #define PAGE_NOT_ACQUIRED ((vm_page_t)1) 4856 static vm_page_t 4857 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m, 4858 int allocflags) 4859 { 4860 if (m == NULL) 4861 m = vm_page_lookup_unlocked(object, pindex); 4862 for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) { 4863 if (vm_page_trybusy(m, allocflags)) { 4864 if (m->object == object && m->pindex == pindex) { 4865 if ((allocflags & VM_ALLOC_WIRED) != 0) 4866 vm_page_wire(m); 4867 vm_page_grab_release(m, allocflags); 4868 break; 4869 } 4870 /* relookup. */ 4871 vm_page_busy_release(m); 4872 cpu_spinwait(); 4873 continue; 4874 } 4875 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4876 allocflags, false)) 4877 return (PAGE_NOT_ACQUIRED); 4878 } 4879 return (m); 4880 } 4881 4882 /* 4883 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4884 * is not set. 4885 */ 4886 vm_page_t 4887 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4888 { 4889 vm_page_t m; 4890 4891 vm_page_grab_check(allocflags); 4892 m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags); 4893 if (m == PAGE_NOT_ACQUIRED) 4894 return (NULL); 4895 if (m != NULL) 4896 return (m); 4897 4898 /* 4899 * The radix lockless lookup should never return a false negative 4900 * errors. If the user specifies NOCREAT they are guaranteed there 4901 * was no page present at the instant of the call. A NOCREAT caller 4902 * must handle create races gracefully. 4903 */ 4904 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4905 return (NULL); 4906 4907 VM_OBJECT_WLOCK(object); 4908 m = vm_page_grab(object, pindex, allocflags); 4909 VM_OBJECT_WUNLOCK(object); 4910 4911 return (m); 4912 } 4913 4914 /* 4915 * Grab a page and make it valid, paging in if necessary. Use an iterator 4916 * parameter. Pages missing from their pager are zero filled and validated. If 4917 * a VM_ALLOC_COUNT is supplied and the page is not valid as many as 4918 * VM_INITIAL_PAGEIN pages can be brought in simultaneously. Additional pages 4919 * will be left on a paging queue but will neither be wired nor busy regardless 4920 * of allocflags. 4921 */ 4922 int 4923 vm_page_grab_valid_iter(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, 4924 int allocflags, struct pctrie_iter *pages) 4925 { 4926 vm_page_t m; 4927 vm_page_t ma[VM_INITIAL_PAGEIN]; 4928 int after, ahead, i, pflags, rv; 4929 4930 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4931 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4932 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4933 KASSERT((allocflags & 4934 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4935 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4936 VM_OBJECT_ASSERT_WLOCKED(object); 4937 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4938 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4939 pflags |= VM_ALLOC_WAITFAIL; 4940 4941 retrylookup: 4942 if ((m = vm_radix_iter_lookup(pages, pindex)) != NULL) { 4943 /* 4944 * If the page is fully valid it can only become invalid 4945 * with the object lock held. If it is not valid it can 4946 * become valid with the busy lock held. Therefore, we 4947 * may unnecessarily lock the exclusive busy here if we 4948 * race with I/O completion not using the object lock. 4949 * However, we will not end up with an invalid page and a 4950 * shared lock. 4951 */ 4952 if (!vm_page_trybusy(m, 4953 vm_page_all_valid(m) ? allocflags : 0)) { 4954 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4955 allocflags, true); 4956 pctrie_iter_reset(pages); 4957 goto retrylookup; 4958 } 4959 if (vm_page_all_valid(m)) 4960 goto out; 4961 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4962 vm_page_busy_release(m); 4963 *mp = NULL; 4964 return (VM_PAGER_FAIL); 4965 } 4966 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4967 *mp = NULL; 4968 return (VM_PAGER_FAIL); 4969 } else { 4970 m = vm_page_alloc_iter(object, pindex, pflags, pages); 4971 if (m == NULL) { 4972 if (!vm_pager_can_alloc_page(object, pindex)) { 4973 *mp = NULL; 4974 return (VM_PAGER_AGAIN); 4975 } 4976 goto retrylookup; 4977 } 4978 } 4979 4980 vm_page_assert_xbusied(m); 4981 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4982 after = MIN(after, VM_INITIAL_PAGEIN); 4983 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4984 after = MAX(after, 1); 4985 ma[0] = m; 4986 pctrie_iter_reset(pages); 4987 for (i = 1; i < after; i++) { 4988 m = vm_radix_iter_lookup_ge(pages, pindex + i); 4989 ahead = after; 4990 if (m != NULL) 4991 ahead = MIN(ahead, m->pindex - pindex); 4992 for (; i < ahead; i++) { 4993 ma[i] = vm_page_alloc_iter(object, pindex + i, 4994 VM_ALLOC_NORMAL, pages); 4995 if (ma[i] == NULL) 4996 break; 4997 } 4998 if (m == NULL || m->pindex != pindex + i || 4999 vm_page_any_valid(m) || !vm_page_tryxbusy(m)) 5000 break; 5001 ma[i] = m; 5002 } 5003 after = i; 5004 vm_object_pip_add(object, after); 5005 VM_OBJECT_WUNLOCK(object); 5006 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 5007 pctrie_iter_reset(pages); 5008 VM_OBJECT_WLOCK(object); 5009 vm_object_pip_wakeupn(object, after); 5010 /* Pager may have replaced a page. */ 5011 m = ma[0]; 5012 if (rv != VM_PAGER_OK) { 5013 for (i = 0; i < after; i++) { 5014 if (!vm_page_wired(ma[i])) 5015 vm_page_free(ma[i]); 5016 else 5017 vm_page_xunbusy(ma[i]); 5018 } 5019 *mp = NULL; 5020 return (rv); 5021 } 5022 for (i = 1; i < after; i++) 5023 vm_page_readahead_finish(ma[i]); 5024 MPASS(vm_page_all_valid(m)); 5025 } else { 5026 vm_page_zero_invalid(m, TRUE); 5027 pctrie_iter_reset(pages); 5028 } 5029 out: 5030 if ((allocflags & VM_ALLOC_WIRED) != 0) 5031 vm_page_wire(m); 5032 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 5033 vm_page_busy_downgrade(m); 5034 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 5035 vm_page_busy_release(m); 5036 *mp = m; 5037 return (VM_PAGER_OK); 5038 } 5039 5040 /* 5041 * Grab a page and make it valid, paging in if necessary. Pages missing from 5042 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 5043 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 5044 * in simultaneously. Additional pages will be left on a paging queue but 5045 * will neither be wired nor busy regardless of allocflags. 5046 */ 5047 int 5048 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, 5049 int allocflags) 5050 { 5051 struct pctrie_iter pages; 5052 5053 VM_OBJECT_ASSERT_WLOCKED(object); 5054 vm_page_iter_init(&pages, object); 5055 return (vm_page_grab_valid_iter(mp, object, pindex, allocflags, 5056 &pages)); 5057 } 5058 5059 /* 5060 * Grab a page. Keep on waiting, as long as the page exists in the object. If 5061 * the page doesn't exist, and the pager has it, allocate it and zero part of 5062 * it. 5063 * 5064 * The object must be locked on entry. This routine may sleep. The lock will, 5065 * however, be released and reacquired if the routine sleeps. 5066 */ 5067 int 5068 vm_page_grab_zero_partial(vm_object_t object, vm_pindex_t pindex, int base, 5069 int end) 5070 { 5071 struct pctrie_iter pages; 5072 vm_page_t m; 5073 int allocflags, rv; 5074 bool found; 5075 5076 VM_OBJECT_ASSERT_WLOCKED(object); 5077 KASSERT(base >= 0, ("%s: base %d", __func__, base)); 5078 KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base, 5079 end)); 5080 5081 allocflags = VM_ALLOC_NOCREAT | VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL; 5082 vm_page_iter_init(&pages, object); 5083 while ((m = vm_page_grab_lookup( 5084 object, pindex, allocflags, &found, &pages)) == NULL) { 5085 if (!vm_pager_has_page(object, pindex, NULL, NULL)) 5086 return (0); 5087 m = vm_page_alloc_iter(object, pindex, 5088 vm_page_grab_pflags(allocflags), &pages); 5089 if (m != NULL) { 5090 vm_object_pip_add(object, 1); 5091 VM_OBJECT_WUNLOCK(object); 5092 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 5093 VM_OBJECT_WLOCK(object); 5094 vm_object_pip_wakeup(object); 5095 if (rv != VM_PAGER_OK) { 5096 vm_page_free(m); 5097 return (EIO); 5098 } 5099 5100 /* 5101 * Since the page was not resident, and therefore not 5102 * recently accessed, immediately enqueue it for 5103 * asynchronous laundering. The current operation is 5104 * not regarded as an access. 5105 */ 5106 vm_page_launder(m); 5107 break; 5108 } 5109 } 5110 5111 pmap_zero_page_area(m, base, end - base); 5112 KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m)); 5113 vm_page_set_dirty(m); 5114 vm_page_xunbusy(m); 5115 return (0); 5116 } 5117 5118 /* 5119 * Locklessly grab a valid page. If the page is not valid or not yet 5120 * allocated this will fall back to the object lock method. 5121 */ 5122 int 5123 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 5124 vm_pindex_t pindex, int allocflags) 5125 { 5126 vm_page_t m; 5127 int flags; 5128 int error; 5129 5130 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 5131 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 5132 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 5133 "mismatch")); 5134 KASSERT((allocflags & 5135 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 5136 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 5137 5138 /* 5139 * Attempt a lockless lookup and busy. We need at least an sbusy 5140 * before we can inspect the valid field and return a wired page. 5141 */ 5142 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 5143 vm_page_grab_check(flags); 5144 m = vm_page_acquire_unlocked(object, pindex, NULL, flags); 5145 if (m == PAGE_NOT_ACQUIRED) 5146 return (VM_PAGER_FAIL); 5147 if (m != NULL) { 5148 if (vm_page_all_valid(m)) { 5149 if ((allocflags & VM_ALLOC_WIRED) != 0) 5150 vm_page_wire(m); 5151 vm_page_grab_release(m, allocflags); 5152 *mp = m; 5153 return (VM_PAGER_OK); 5154 } 5155 vm_page_busy_release(m); 5156 } 5157 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5158 *mp = NULL; 5159 return (VM_PAGER_FAIL); 5160 } 5161 VM_OBJECT_WLOCK(object); 5162 error = vm_page_grab_valid(mp, object, pindex, allocflags); 5163 VM_OBJECT_WUNLOCK(object); 5164 5165 return (error); 5166 } 5167 5168 /* 5169 * Return the specified range of pages from the given object. For each 5170 * page offset within the range, if a page already exists within the object 5171 * at that offset and it is busy, then wait for it to change state. If, 5172 * instead, the page doesn't exist, then allocate it. 5173 * 5174 * The caller must always specify an allocation class. 5175 * 5176 * allocation classes: 5177 * VM_ALLOC_NORMAL normal process request 5178 * VM_ALLOC_SYSTEM system *really* needs the pages 5179 * VM_ALLOC_INTERRUPT interrupt time request 5180 * 5181 * The caller must always specify that the pages are to be busied and/or 5182 * wired. 5183 * 5184 * optional allocation flags: 5185 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 5186 * VM_ALLOC_NOBUSY do not exclusive busy the pages 5187 * VM_ALLOC_NODUMP do not include the pages in a kernel core dump 5188 * VM_ALLOC_NOFREE pages will never be freed 5189 * VM_ALLOC_NOWAIT do not sleep 5190 * VM_ALLOC_SBUSY set pages to sbusy state 5191 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 5192 * VM_ALLOC_WAITOK ignored (default behavior) 5193 * VM_ALLOC_WIRED wire the pages 5194 * VM_ALLOC_ZERO zero and validate any invalid pages 5195 * 5196 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 5197 * may return a partial prefix of the requested range. 5198 */ 5199 int 5200 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 5201 vm_page_t *ma, int count) 5202 { 5203 struct pctrie_iter pages; 5204 vm_page_t m; 5205 int pflags; 5206 int ahead, i; 5207 5208 VM_OBJECT_ASSERT_WLOCKED(object); 5209 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 5210 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 5211 KASSERT(count > 0, 5212 ("vm_page_grab_pages: invalid page count %d", count)); 5213 vm_page_grab_check(allocflags); 5214 5215 pflags = vm_page_grab_pflags(allocflags); 5216 i = 0; 5217 vm_page_iter_init(&pages, object); 5218 retrylookup: 5219 ahead = -1; 5220 for (; i < count; i++) { 5221 if (ahead < 0) { 5222 ahead = vm_radix_iter_lookup_range( 5223 &pages, pindex + i, &ma[i], count - i); 5224 } 5225 if (ahead-- > 0) { 5226 m = ma[i]; 5227 if (!vm_page_tryacquire(m, allocflags)) { 5228 if (vm_page_grab_sleep(object, m, pindex + i, 5229 "grbmaw", allocflags, true)) { 5230 pctrie_iter_reset(&pages); 5231 goto retrylookup; 5232 } 5233 break; 5234 } 5235 } else { 5236 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 5237 break; 5238 m = vm_page_alloc_iter(object, pindex + i, 5239 pflags | VM_ALLOC_COUNT(count - i), &pages); 5240 /* pages was reset if alloc_iter lost the lock. */ 5241 if (m == NULL) { 5242 if ((allocflags & (VM_ALLOC_NOWAIT | 5243 VM_ALLOC_WAITFAIL)) != 0) 5244 break; 5245 goto retrylookup; 5246 } 5247 ma[i] = m; 5248 } 5249 if (vm_page_none_valid(m) && 5250 (allocflags & VM_ALLOC_ZERO) != 0) { 5251 if ((m->flags & PG_ZERO) == 0) 5252 pmap_zero_page(m); 5253 vm_page_valid(m); 5254 } 5255 vm_page_grab_release(m, allocflags); 5256 } 5257 return (i); 5258 } 5259 5260 /* 5261 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 5262 * and will fall back to the locked variant to handle allocation. 5263 */ 5264 int 5265 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 5266 int allocflags, vm_page_t *ma, int count) 5267 { 5268 vm_page_t m; 5269 int flags; 5270 int i, num_fetched; 5271 5272 KASSERT(count > 0, 5273 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 5274 vm_page_grab_check(allocflags); 5275 5276 /* 5277 * Modify flags for lockless acquire to hold the page until we 5278 * set it valid if necessary. 5279 */ 5280 flags = allocflags & ~VM_ALLOC_NOBUSY; 5281 vm_page_grab_check(flags); 5282 num_fetched = vm_radix_lookup_range_unlocked(&object->rtree, pindex, 5283 ma, count); 5284 for (i = 0; i < num_fetched; i++, pindex++) { 5285 m = vm_page_acquire_unlocked(object, pindex, ma[i], flags); 5286 if (m == PAGE_NOT_ACQUIRED) 5287 return (i); 5288 if (m == NULL) 5289 break; 5290 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 5291 if ((m->flags & PG_ZERO) == 0) 5292 pmap_zero_page(m); 5293 vm_page_valid(m); 5294 } 5295 /* m will still be wired or busy according to flags. */ 5296 vm_page_grab_release(m, allocflags); 5297 /* vm_page_acquire_unlocked() may not return ma[i]. */ 5298 ma[i] = m; 5299 } 5300 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 5301 return (i); 5302 count -= i; 5303 VM_OBJECT_WLOCK(object); 5304 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 5305 VM_OBJECT_WUNLOCK(object); 5306 5307 return (i); 5308 } 5309 5310 /* 5311 * Mapping function for valid or dirty bits in a page. 5312 * 5313 * Inputs are required to range within a page. 5314 */ 5315 vm_page_bits_t 5316 vm_page_bits(int base, int size) 5317 { 5318 int first_bit; 5319 int last_bit; 5320 5321 KASSERT( 5322 base + size <= PAGE_SIZE, 5323 ("vm_page_bits: illegal base/size %d/%d", base, size) 5324 ); 5325 5326 if (size == 0) /* handle degenerate case */ 5327 return (0); 5328 5329 first_bit = base >> DEV_BSHIFT; 5330 last_bit = (base + size - 1) >> DEV_BSHIFT; 5331 5332 return (((vm_page_bits_t)2 << last_bit) - 5333 ((vm_page_bits_t)1 << first_bit)); 5334 } 5335 5336 void 5337 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5338 { 5339 5340 #if PAGE_SIZE == 32768 5341 atomic_set_64((uint64_t *)bits, set); 5342 #elif PAGE_SIZE == 16384 5343 atomic_set_32((uint32_t *)bits, set); 5344 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5345 atomic_set_16((uint16_t *)bits, set); 5346 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5347 atomic_set_8((uint8_t *)bits, set); 5348 #else /* PAGE_SIZE <= 8192 */ 5349 uintptr_t addr; 5350 int shift; 5351 5352 addr = (uintptr_t)bits; 5353 /* 5354 * Use a trick to perform a 32-bit atomic on the 5355 * containing aligned word, to not depend on the existence 5356 * of atomic_{set, clear}_{8, 16}. 5357 */ 5358 shift = addr & (sizeof(uint32_t) - 1); 5359 #if BYTE_ORDER == BIG_ENDIAN 5360 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5361 #else 5362 shift *= NBBY; 5363 #endif 5364 addr &= ~(sizeof(uint32_t) - 1); 5365 atomic_set_32((uint32_t *)addr, set << shift); 5366 #endif /* PAGE_SIZE */ 5367 } 5368 5369 static inline void 5370 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5371 { 5372 5373 #if PAGE_SIZE == 32768 5374 atomic_clear_64((uint64_t *)bits, clear); 5375 #elif PAGE_SIZE == 16384 5376 atomic_clear_32((uint32_t *)bits, clear); 5377 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5378 atomic_clear_16((uint16_t *)bits, clear); 5379 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5380 atomic_clear_8((uint8_t *)bits, clear); 5381 #else /* PAGE_SIZE <= 8192 */ 5382 uintptr_t addr; 5383 int shift; 5384 5385 addr = (uintptr_t)bits; 5386 /* 5387 * Use a trick to perform a 32-bit atomic on the 5388 * containing aligned word, to not depend on the existence 5389 * of atomic_{set, clear}_{8, 16}. 5390 */ 5391 shift = addr & (sizeof(uint32_t) - 1); 5392 #if BYTE_ORDER == BIG_ENDIAN 5393 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5394 #else 5395 shift *= NBBY; 5396 #endif 5397 addr &= ~(sizeof(uint32_t) - 1); 5398 atomic_clear_32((uint32_t *)addr, clear << shift); 5399 #endif /* PAGE_SIZE */ 5400 } 5401 5402 static inline vm_page_bits_t 5403 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5404 { 5405 #if PAGE_SIZE == 32768 5406 uint64_t old; 5407 5408 old = *bits; 5409 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5410 return (old); 5411 #elif PAGE_SIZE == 16384 5412 uint32_t old; 5413 5414 old = *bits; 5415 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5416 return (old); 5417 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5418 uint16_t old; 5419 5420 old = *bits; 5421 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5422 return (old); 5423 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5424 uint8_t old; 5425 5426 old = *bits; 5427 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5428 return (old); 5429 #else /* PAGE_SIZE <= 4096*/ 5430 uintptr_t addr; 5431 uint32_t old, new, mask; 5432 int shift; 5433 5434 addr = (uintptr_t)bits; 5435 /* 5436 * Use a trick to perform a 32-bit atomic on the 5437 * containing aligned word, to not depend on the existence 5438 * of atomic_{set, swap, clear}_{8, 16}. 5439 */ 5440 shift = addr & (sizeof(uint32_t) - 1); 5441 #if BYTE_ORDER == BIG_ENDIAN 5442 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5443 #else 5444 shift *= NBBY; 5445 #endif 5446 addr &= ~(sizeof(uint32_t) - 1); 5447 mask = VM_PAGE_BITS_ALL << shift; 5448 5449 old = *bits; 5450 do { 5451 new = old & ~mask; 5452 new |= newbits << shift; 5453 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5454 return (old >> shift); 5455 #endif /* PAGE_SIZE */ 5456 } 5457 5458 /* 5459 * vm_page_set_valid_range: 5460 * 5461 * Sets portions of a page valid. The arguments are expected 5462 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5463 * of any partial chunks touched by the range. The invalid portion of 5464 * such chunks will be zeroed. 5465 * 5466 * (base + size) must be less then or equal to PAGE_SIZE. 5467 */ 5468 void 5469 vm_page_set_valid_range(vm_page_t m, int base, int size) 5470 { 5471 int endoff, frag; 5472 vm_page_bits_t pagebits; 5473 5474 vm_page_assert_busied(m); 5475 if (size == 0) /* handle degenerate case */ 5476 return; 5477 5478 /* 5479 * If the base is not DEV_BSIZE aligned and the valid 5480 * bit is clear, we have to zero out a portion of the 5481 * first block. 5482 */ 5483 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5484 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5485 pmap_zero_page_area(m, frag, base - frag); 5486 5487 /* 5488 * If the ending offset is not DEV_BSIZE aligned and the 5489 * valid bit is clear, we have to zero out a portion of 5490 * the last block. 5491 */ 5492 endoff = base + size; 5493 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5494 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5495 pmap_zero_page_area(m, endoff, 5496 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5497 5498 /* 5499 * Assert that no previously invalid block that is now being validated 5500 * is already dirty. 5501 */ 5502 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5503 ("vm_page_set_valid_range: page %p is dirty", m)); 5504 5505 /* 5506 * Set valid bits inclusive of any overlap. 5507 */ 5508 pagebits = vm_page_bits(base, size); 5509 if (vm_page_xbusied(m)) 5510 m->valid |= pagebits; 5511 else 5512 vm_page_bits_set(m, &m->valid, pagebits); 5513 } 5514 5515 /* 5516 * Set the page dirty bits and free the invalid swap space if 5517 * present. Returns the previous dirty bits. 5518 */ 5519 vm_page_bits_t 5520 vm_page_set_dirty(vm_page_t m) 5521 { 5522 vm_page_bits_t old; 5523 5524 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5525 5526 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5527 old = m->dirty; 5528 m->dirty = VM_PAGE_BITS_ALL; 5529 } else 5530 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5531 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5532 vm_pager_page_unswapped(m); 5533 5534 return (old); 5535 } 5536 5537 /* 5538 * Clear the given bits from the specified page's dirty field. 5539 */ 5540 static __inline void 5541 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5542 { 5543 5544 vm_page_assert_busied(m); 5545 5546 /* 5547 * If the page is xbusied and not write mapped we are the 5548 * only thread that can modify dirty bits. Otherwise, The pmap 5549 * layer can call vm_page_dirty() without holding a distinguished 5550 * lock. The combination of page busy and atomic operations 5551 * suffice to guarantee consistency of the page dirty field. 5552 */ 5553 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5554 m->dirty &= ~pagebits; 5555 else 5556 vm_page_bits_clear(m, &m->dirty, pagebits); 5557 } 5558 5559 /* 5560 * vm_page_set_validclean: 5561 * 5562 * Sets portions of a page valid and clean. The arguments are expected 5563 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5564 * of any partial chunks touched by the range. The invalid portion of 5565 * such chunks will be zero'd. 5566 * 5567 * (base + size) must be less then or equal to PAGE_SIZE. 5568 */ 5569 void 5570 vm_page_set_validclean(vm_page_t m, int base, int size) 5571 { 5572 vm_page_bits_t oldvalid, pagebits; 5573 int endoff, frag; 5574 5575 vm_page_assert_busied(m); 5576 if (size == 0) /* handle degenerate case */ 5577 return; 5578 5579 /* 5580 * If the base is not DEV_BSIZE aligned and the valid 5581 * bit is clear, we have to zero out a portion of the 5582 * first block. 5583 */ 5584 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5585 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5586 pmap_zero_page_area(m, frag, base - frag); 5587 5588 /* 5589 * If the ending offset is not DEV_BSIZE aligned and the 5590 * valid bit is clear, we have to zero out a portion of 5591 * the last block. 5592 */ 5593 endoff = base + size; 5594 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5595 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5596 pmap_zero_page_area(m, endoff, 5597 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5598 5599 /* 5600 * Set valid, clear dirty bits. If validating the entire 5601 * page we can safely clear the pmap modify bit. We also 5602 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5603 * takes a write fault on a MAP_NOSYNC memory area the flag will 5604 * be set again. 5605 * 5606 * We set valid bits inclusive of any overlap, but we can only 5607 * clear dirty bits for DEV_BSIZE chunks that are fully within 5608 * the range. 5609 */ 5610 oldvalid = m->valid; 5611 pagebits = vm_page_bits(base, size); 5612 if (vm_page_xbusied(m)) 5613 m->valid |= pagebits; 5614 else 5615 vm_page_bits_set(m, &m->valid, pagebits); 5616 #if 0 /* NOT YET */ 5617 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5618 frag = DEV_BSIZE - frag; 5619 base += frag; 5620 size -= frag; 5621 if (size < 0) 5622 size = 0; 5623 } 5624 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5625 #endif 5626 if (base == 0 && size == PAGE_SIZE) { 5627 /* 5628 * The page can only be modified within the pmap if it is 5629 * mapped, and it can only be mapped if it was previously 5630 * fully valid. 5631 */ 5632 if (oldvalid == VM_PAGE_BITS_ALL) 5633 /* 5634 * Perform the pmap_clear_modify() first. Otherwise, 5635 * a concurrent pmap operation, such as 5636 * pmap_protect(), could clear a modification in the 5637 * pmap and set the dirty field on the page before 5638 * pmap_clear_modify() had begun and after the dirty 5639 * field was cleared here. 5640 */ 5641 pmap_clear_modify(m); 5642 m->dirty = 0; 5643 vm_page_aflag_clear(m, PGA_NOSYNC); 5644 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5645 m->dirty &= ~pagebits; 5646 else 5647 vm_page_clear_dirty_mask(m, pagebits); 5648 } 5649 5650 void 5651 vm_page_clear_dirty(vm_page_t m, int base, int size) 5652 { 5653 5654 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5655 } 5656 5657 /* 5658 * vm_page_set_invalid: 5659 * 5660 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5661 * valid and dirty bits for the effected areas are cleared. 5662 */ 5663 void 5664 vm_page_set_invalid(vm_page_t m, int base, int size) 5665 { 5666 vm_page_bits_t bits; 5667 vm_object_t object; 5668 5669 /* 5670 * The object lock is required so that pages can't be mapped 5671 * read-only while we're in the process of invalidating them. 5672 */ 5673 object = m->object; 5674 VM_OBJECT_ASSERT_WLOCKED(object); 5675 vm_page_assert_busied(m); 5676 5677 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5678 size >= object->un_pager.vnp.vnp_size) 5679 bits = VM_PAGE_BITS_ALL; 5680 else 5681 bits = vm_page_bits(base, size); 5682 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5683 pmap_remove_all(m); 5684 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5685 !pmap_page_is_mapped(m), 5686 ("vm_page_set_invalid: page %p is mapped", m)); 5687 if (vm_page_xbusied(m)) { 5688 m->valid &= ~bits; 5689 m->dirty &= ~bits; 5690 } else { 5691 vm_page_bits_clear(m, &m->valid, bits); 5692 vm_page_bits_clear(m, &m->dirty, bits); 5693 } 5694 } 5695 5696 /* 5697 * vm_page_invalid: 5698 * 5699 * Invalidates the entire page. The page must be busy, unmapped, and 5700 * the enclosing object must be locked. The object locks protects 5701 * against concurrent read-only pmap enter which is done without 5702 * busy. 5703 */ 5704 void 5705 vm_page_invalid(vm_page_t m) 5706 { 5707 5708 vm_page_assert_busied(m); 5709 VM_OBJECT_ASSERT_WLOCKED(m->object); 5710 MPASS(!pmap_page_is_mapped(m)); 5711 5712 if (vm_page_xbusied(m)) 5713 m->valid = 0; 5714 else 5715 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5716 } 5717 5718 /* 5719 * vm_page_zero_invalid() 5720 * 5721 * The kernel assumes that the invalid portions of a page contain 5722 * garbage, but such pages can be mapped into memory by user code. 5723 * When this occurs, we must zero out the non-valid portions of the 5724 * page so user code sees what it expects. 5725 * 5726 * Pages are most often semi-valid when the end of a file is mapped 5727 * into memory and the file's size is not page aligned. 5728 */ 5729 void 5730 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5731 { 5732 int b; 5733 int i; 5734 5735 /* 5736 * Scan the valid bits looking for invalid sections that 5737 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5738 * valid bit may be set ) have already been zeroed by 5739 * vm_page_set_validclean(). 5740 */ 5741 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5742 if (i == (PAGE_SIZE / DEV_BSIZE) || 5743 (m->valid & ((vm_page_bits_t)1 << i))) { 5744 if (i > b) { 5745 pmap_zero_page_area(m, 5746 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5747 } 5748 b = i + 1; 5749 } 5750 } 5751 5752 /* 5753 * setvalid is TRUE when we can safely set the zero'd areas 5754 * as being valid. We can do this if there are no cache consistency 5755 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5756 */ 5757 if (setvalid) 5758 vm_page_valid(m); 5759 } 5760 5761 /* 5762 * vm_page_is_valid: 5763 * 5764 * Is (partial) page valid? Note that the case where size == 0 5765 * will return FALSE in the degenerate case where the page is 5766 * entirely invalid, and TRUE otherwise. 5767 * 5768 * Some callers envoke this routine without the busy lock held and 5769 * handle races via higher level locks. Typical callers should 5770 * hold a busy lock to prevent invalidation. 5771 */ 5772 int 5773 vm_page_is_valid(vm_page_t m, int base, int size) 5774 { 5775 vm_page_bits_t bits; 5776 5777 bits = vm_page_bits(base, size); 5778 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5779 } 5780 5781 /* 5782 * Returns true if all of the specified predicates are true for the entire 5783 * (super)page and false otherwise. 5784 */ 5785 bool 5786 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m) 5787 { 5788 vm_object_t object; 5789 int i, npages; 5790 5791 object = m->object; 5792 if (skip_m != NULL && skip_m->object != object) 5793 return (false); 5794 VM_OBJECT_ASSERT_LOCKED(object); 5795 KASSERT(psind <= m->psind, 5796 ("psind %d > psind %d of m %p", psind, m->psind, m)); 5797 npages = atop(pagesizes[psind]); 5798 5799 /* 5800 * The physically contiguous pages that make up a superpage, i.e., a 5801 * page with a page size index ("psind") greater than zero, will 5802 * occupy adjacent entries in vm_page_array[]. 5803 */ 5804 for (i = 0; i < npages; i++) { 5805 /* Always test object consistency, including "skip_m". */ 5806 if (m[i].object != object) 5807 return (false); 5808 if (&m[i] == skip_m) 5809 continue; 5810 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5811 return (false); 5812 if ((flags & PS_ALL_DIRTY) != 0) { 5813 /* 5814 * Calling vm_page_test_dirty() or pmap_is_modified() 5815 * might stop this case from spuriously returning 5816 * "false". However, that would require a write lock 5817 * on the object containing "m[i]". 5818 */ 5819 if (m[i].dirty != VM_PAGE_BITS_ALL) 5820 return (false); 5821 } 5822 if ((flags & PS_ALL_VALID) != 0 && 5823 m[i].valid != VM_PAGE_BITS_ALL) 5824 return (false); 5825 } 5826 return (true); 5827 } 5828 5829 /* 5830 * Set the page's dirty bits if the page is modified. 5831 */ 5832 void 5833 vm_page_test_dirty(vm_page_t m) 5834 { 5835 5836 vm_page_assert_busied(m); 5837 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5838 vm_page_dirty(m); 5839 } 5840 5841 void 5842 vm_page_valid(vm_page_t m) 5843 { 5844 5845 vm_page_assert_busied(m); 5846 if (vm_page_xbusied(m)) 5847 m->valid = VM_PAGE_BITS_ALL; 5848 else 5849 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5850 } 5851 5852 #ifdef INVARIANTS 5853 void 5854 vm_page_object_busy_assert(vm_page_t m) 5855 { 5856 5857 /* 5858 * Certain of the page's fields may only be modified by the 5859 * holder of a page or object busy. 5860 */ 5861 if (m->object != NULL && !vm_page_busied(m)) 5862 VM_OBJECT_ASSERT_BUSY(m->object); 5863 } 5864 5865 void 5866 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5867 { 5868 5869 if ((bits & PGA_WRITEABLE) == 0) 5870 return; 5871 5872 /* 5873 * The PGA_WRITEABLE flag can only be set if the page is 5874 * managed, is exclusively busied or the object is locked. 5875 * Currently, this flag is only set by pmap_enter(). 5876 */ 5877 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5878 ("PGA_WRITEABLE on unmanaged page")); 5879 if (!vm_page_xbusied(m)) 5880 VM_OBJECT_ASSERT_BUSY(m->object); 5881 } 5882 #endif 5883 5884 #include "opt_ddb.h" 5885 #ifdef DDB 5886 #include <sys/kernel.h> 5887 5888 #include <ddb/ddb.h> 5889 5890 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5891 { 5892 5893 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5894 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5895 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5896 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5897 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5898 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5899 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5900 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5901 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5902 } 5903 5904 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5905 { 5906 int dom; 5907 5908 db_printf("pq_free %d\n", vm_free_count()); 5909 for (dom = 0; dom < vm_ndomains; dom++) { 5910 db_printf( 5911 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5912 dom, 5913 vm_dom[dom].vmd_page_count, 5914 vm_dom[dom].vmd_free_count, 5915 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5916 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5917 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5918 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5919 } 5920 } 5921 5922 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5923 { 5924 vm_page_t m; 5925 boolean_t phys, virt; 5926 5927 if (!have_addr) { 5928 db_printf("show pginfo addr\n"); 5929 return; 5930 } 5931 5932 phys = strchr(modif, 'p') != NULL; 5933 virt = strchr(modif, 'v') != NULL; 5934 if (virt) 5935 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5936 else if (phys) 5937 m = PHYS_TO_VM_PAGE(addr); 5938 else 5939 m = (vm_page_t)addr; 5940 db_printf( 5941 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5942 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5943 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5944 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5945 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5946 } 5947 #endif /* DDB */ 5948