1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * The Mach Operating System project at Carnegie-Mellon University. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 36 /*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Resident memory management module. 65 */ 66 67 #include <sys/cdefs.h> 68 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/counter.h> 73 #include <sys/domainset.h> 74 #include <sys/kernel.h> 75 #include <sys/limits.h> 76 #include <sys/linker.h> 77 #include <sys/lock.h> 78 #include <sys/malloc.h> 79 #include <sys/mman.h> 80 #include <sys/msgbuf.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> 83 #include <sys/rwlock.h> 84 #include <sys/sleepqueue.h> 85 #include <sys/sbuf.h> 86 #include <sys/sched.h> 87 #include <sys/smp.h> 88 #include <sys/sysctl.h> 89 #include <sys/vmmeter.h> 90 #include <sys/vnode.h> 91 92 #include <vm/vm.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_param.h> 95 #include <vm/vm_domainset.h> 96 #include <vm/vm_kern.h> 97 #include <vm/vm_map.h> 98 #include <vm/vm_object.h> 99 #include <vm/vm_page.h> 100 #include <vm/vm_pageout.h> 101 #include <vm/vm_phys.h> 102 #include <vm/vm_pagequeue.h> 103 #include <vm/vm_pager.h> 104 #include <vm/vm_radix.h> 105 #include <vm/vm_reserv.h> 106 #include <vm/vm_extern.h> 107 #include <vm/vm_dumpset.h> 108 #include <vm/uma.h> 109 #include <vm/uma_int.h> 110 111 #include <machine/md_var.h> 112 113 struct vm_domain vm_dom[MAXMEMDOM]; 114 115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 116 117 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 118 /* The following fields are protected by the domainset lock. */ 119 domainset_t __exclusive_cache_line vm_min_domains; 120 domainset_t __exclusive_cache_line vm_severe_domains; 121 static int vm_min_waiters; 122 static int vm_severe_waiters; 123 static int vm_pageproc_waiters; 124 125 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 126 "VM page statistics"); 127 128 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 129 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 130 CTLFLAG_RD, &pqstate_commit_retries, 131 "Number of failed per-page atomic queue state updates"); 132 133 static COUNTER_U64_DEFINE_EARLY(queue_ops); 134 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 135 CTLFLAG_RD, &queue_ops, 136 "Number of batched queue operations"); 137 138 static COUNTER_U64_DEFINE_EARLY(queue_nops); 139 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 140 CTLFLAG_RD, &queue_nops, 141 "Number of batched queue operations with no effects"); 142 143 static unsigned long nofreeq_size; 144 SYSCTL_ULONG(_vm_stats_page, OID_AUTO, nofreeq_size, CTLFLAG_RD, 145 &nofreeq_size, 0, 146 "Size of the nofree queue"); 147 148 /* 149 * bogus page -- for I/O to/from partially complete buffers, 150 * or for paging into sparsely invalid regions. 151 */ 152 vm_page_t bogus_page; 153 154 vm_page_t vm_page_array; 155 long vm_page_array_size; 156 long first_page; 157 158 struct bitset *vm_page_dump; 159 long vm_page_dump_pages; 160 161 static TAILQ_HEAD(, vm_page) blacklist_head; 162 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 163 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 164 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 165 166 static uma_zone_t fakepg_zone; 167 168 static void vm_page_alloc_check(vm_page_t m); 169 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req); 170 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 171 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 172 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 173 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 174 static bool vm_page_free_prep(vm_page_t m); 175 static void vm_page_free_toq(vm_page_t m); 176 static void vm_page_init(void *dummy); 177 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object); 178 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 179 const uint16_t nflag); 180 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 181 vm_page_t m_run, vm_paddr_t high); 182 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 183 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 184 int req); 185 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 186 int flags); 187 static void vm_page_zone_release(void *arg, void **store, int cnt); 188 189 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 190 191 static void 192 vm_page_init(void *dummy) 193 { 194 195 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 196 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 197 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE); 198 } 199 200 static int pgcache_zone_max_pcpu; 201 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu, 202 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0, 203 "Per-CPU page cache size"); 204 205 /* 206 * The cache page zone is initialized later since we need to be able to allocate 207 * pages before UMA is fully initialized. 208 */ 209 static void 210 vm_page_init_cache_zones(void *dummy __unused) 211 { 212 struct vm_domain *vmd; 213 struct vm_pgcache *pgcache; 214 int cache, domain, maxcache, pool; 215 216 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu); 217 maxcache = pgcache_zone_max_pcpu * mp_ncpus; 218 for (domain = 0; domain < vm_ndomains; domain++) { 219 vmd = VM_DOMAIN(domain); 220 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 221 #ifdef VM_FREEPOOL_LAZYINIT 222 if (pool == VM_FREEPOOL_LAZYINIT) 223 continue; 224 #endif 225 pgcache = &vmd->vmd_pgcache[pool]; 226 pgcache->domain = domain; 227 pgcache->pool = pool; 228 pgcache->zone = uma_zcache_create("vm pgcache", 229 PAGE_SIZE, NULL, NULL, NULL, NULL, 230 vm_page_zone_import, vm_page_zone_release, pgcache, 231 UMA_ZONE_VM); 232 233 /* 234 * Limit each pool's zone to 0.1% of the pages in the 235 * domain. 236 */ 237 cache = maxcache != 0 ? maxcache : 238 vmd->vmd_page_count / 1000; 239 uma_zone_set_maxcache(pgcache->zone, cache); 240 } 241 } 242 } 243 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 244 245 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 246 #if PAGE_SIZE == 32768 247 #ifdef CTASSERT 248 CTASSERT(sizeof(u_long) >= 8); 249 #endif 250 #endif 251 252 /* 253 * vm_set_page_size: 254 * 255 * Sets the page size, perhaps based upon the memory 256 * size. Must be called before any use of page-size 257 * dependent functions. 258 */ 259 void 260 vm_set_page_size(void) 261 { 262 if (vm_cnt.v_page_size == 0) 263 vm_cnt.v_page_size = PAGE_SIZE; 264 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 265 panic("vm_set_page_size: page size not a power of two"); 266 } 267 268 /* 269 * vm_page_blacklist_next: 270 * 271 * Find the next entry in the provided string of blacklist 272 * addresses. Entries are separated by space, comma, or newline. 273 * If an invalid integer is encountered then the rest of the 274 * string is skipped. Updates the list pointer to the next 275 * character, or NULL if the string is exhausted or invalid. 276 */ 277 static vm_paddr_t 278 vm_page_blacklist_next(char **list, char *end) 279 { 280 vm_paddr_t bad; 281 char *cp, *pos; 282 283 if (list == NULL || *list == NULL) 284 return (0); 285 if (**list =='\0') { 286 *list = NULL; 287 return (0); 288 } 289 290 /* 291 * If there's no end pointer then the buffer is coming from 292 * the kenv and we know it's null-terminated. 293 */ 294 if (end == NULL) 295 end = *list + strlen(*list); 296 297 /* Ensure that strtoq() won't walk off the end */ 298 if (*end != '\0') { 299 if (*end == '\n' || *end == ' ' || *end == ',') 300 *end = '\0'; 301 else { 302 printf("Blacklist not terminated, skipping\n"); 303 *list = NULL; 304 return (0); 305 } 306 } 307 308 for (pos = *list; *pos != '\0'; pos = cp) { 309 bad = strtoq(pos, &cp, 0); 310 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 311 if (bad == 0) { 312 if (++cp < end) 313 continue; 314 else 315 break; 316 } 317 } else 318 break; 319 if (*cp == '\0' || ++cp >= end) 320 *list = NULL; 321 else 322 *list = cp; 323 return (trunc_page(bad)); 324 } 325 printf("Garbage in RAM blacklist, skipping\n"); 326 *list = NULL; 327 return (0); 328 } 329 330 bool 331 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 332 { 333 struct vm_domain *vmd; 334 vm_page_t m; 335 bool found; 336 337 m = vm_phys_paddr_to_vm_page(pa); 338 if (m == NULL) 339 return (true); /* page does not exist, no failure */ 340 341 vmd = VM_DOMAIN(vm_phys_domain(pa)); 342 vm_domain_free_lock(vmd); 343 found = vm_phys_unfree_page(pa); 344 vm_domain_free_unlock(vmd); 345 if (found) { 346 vm_domain_freecnt_inc(vmd, -1); 347 TAILQ_INSERT_TAIL(&blacklist_head, m, plinks.q); 348 if (verbose) 349 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 350 } 351 return (found); 352 } 353 354 /* 355 * vm_page_blacklist_check: 356 * 357 * Iterate through the provided string of blacklist addresses, pulling 358 * each entry out of the physical allocator free list and putting it 359 * onto a list for reporting via the vm.page_blacklist sysctl. 360 */ 361 static void 362 vm_page_blacklist_check(char *list, char *end) 363 { 364 vm_paddr_t pa; 365 char *next; 366 367 next = list; 368 while (next != NULL) { 369 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 370 continue; 371 vm_page_blacklist_add(pa, bootverbose); 372 } 373 } 374 375 /* 376 * vm_page_blacklist_load: 377 * 378 * Search for a special module named "ram_blacklist". It'll be a 379 * plain text file provided by the user via the loader directive 380 * of the same name. 381 */ 382 static void 383 vm_page_blacklist_load(char **list, char **end) 384 { 385 void *mod; 386 u_char *ptr; 387 u_int len; 388 389 mod = NULL; 390 ptr = NULL; 391 392 mod = preload_search_by_type("ram_blacklist"); 393 if (mod != NULL) { 394 ptr = preload_fetch_addr(mod); 395 len = preload_fetch_size(mod); 396 } 397 398 if (ptr != NULL && len > 0) { 399 *list = ptr; 400 *end = ptr + len - 1; 401 } else { 402 *list = NULL; 403 *end = NULL; 404 } 405 406 return; 407 } 408 409 static int 410 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 411 { 412 vm_page_t m; 413 struct sbuf sbuf; 414 int error, first; 415 416 first = 1; 417 error = sysctl_wire_old_buffer(req, 0); 418 if (error != 0) 419 return (error); 420 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 421 TAILQ_FOREACH(m, &blacklist_head, plinks.q) { 422 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 423 (uintmax_t)m->phys_addr); 424 first = 0; 425 } 426 error = sbuf_finish(&sbuf); 427 sbuf_delete(&sbuf); 428 return (error); 429 } 430 431 /* 432 * Initialize a dummy page for use in scans of the specified paging queue. 433 * In principle, this function only needs to set the flag PG_MARKER. 434 * Nonetheless, it write busies the page as a safety precaution. 435 */ 436 void 437 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 438 { 439 440 bzero(marker, sizeof(*marker)); 441 marker->flags = PG_MARKER; 442 marker->a.flags = aflags; 443 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 444 marker->a.queue = queue; 445 } 446 447 static void 448 vm_page_domain_init(int domain) 449 { 450 struct vm_domain *vmd; 451 struct vm_pagequeue *pq; 452 int i; 453 454 vmd = VM_DOMAIN(domain); 455 bzero(vmd, sizeof(*vmd)); 456 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 457 "vm inactive pagequeue"; 458 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 459 "vm active pagequeue"; 460 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 461 "vm laundry pagequeue"; 462 *__DECONST(const char **, 463 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 464 "vm unswappable pagequeue"; 465 vmd->vmd_domain = domain; 466 vmd->vmd_page_count = 0; 467 vmd->vmd_free_count = 0; 468 vmd->vmd_segs = 0; 469 vmd->vmd_oom = false; 470 vmd->vmd_helper_threads_enabled = true; 471 for (i = 0; i < PQ_COUNT; i++) { 472 pq = &vmd->vmd_pagequeues[i]; 473 TAILQ_INIT(&pq->pq_pl); 474 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 475 MTX_DEF | MTX_DUPOK); 476 pq->pq_pdpages = 0; 477 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 478 } 479 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 480 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 481 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 482 483 /* 484 * inacthead is used to provide FIFO ordering for LRU-bypassing 485 * insertions. 486 */ 487 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 488 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 489 &vmd->vmd_inacthead, plinks.q); 490 491 /* 492 * The clock pages are used to implement active queue scanning without 493 * requeues. Scans start at clock[0], which is advanced after the scan 494 * ends. When the two clock hands meet, they are reset and scanning 495 * resumes from the head of the queue. 496 */ 497 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 498 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 499 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 500 &vmd->vmd_clock[0], plinks.q); 501 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 502 &vmd->vmd_clock[1], plinks.q); 503 } 504 505 /* 506 * Initialize a physical page in preparation for adding it to the free 507 * lists. 508 */ 509 void 510 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool) 511 { 512 m->object = NULL; 513 m->ref_count = 0; 514 m->busy_lock = VPB_FREED; 515 m->flags = m->a.flags = 0; 516 m->phys_addr = pa; 517 m->a.queue = PQ_NONE; 518 m->psind = 0; 519 m->segind = segind; 520 m->order = VM_NFREEORDER; 521 m->pool = pool; 522 m->valid = m->dirty = 0; 523 pmap_page_init(m); 524 } 525 526 #ifndef PMAP_HAS_PAGE_ARRAY 527 static vm_paddr_t 528 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 529 { 530 vm_paddr_t new_end; 531 532 /* 533 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 534 * However, because this page is allocated from KVM, out-of-bounds 535 * accesses using the direct map will not be trapped. 536 */ 537 *vaddr += PAGE_SIZE; 538 539 /* 540 * Allocate physical memory for the page structures, and map it. 541 */ 542 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 543 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 544 VM_PROT_READ | VM_PROT_WRITE); 545 vm_page_array_size = page_range; 546 547 return (new_end); 548 } 549 #endif 550 551 /* 552 * vm_page_startup: 553 * 554 * Initializes the resident memory module. Allocates physical memory for 555 * bootstrapping UMA and some data structures that are used to manage 556 * physical pages. Initializes these structures, and populates the free 557 * page queues. 558 */ 559 vm_offset_t 560 vm_page_startup(vm_offset_t vaddr) 561 { 562 struct vm_phys_seg *seg; 563 struct vm_domain *vmd; 564 vm_page_t m; 565 char *list, *listend; 566 vm_paddr_t end, high_avail, low_avail, new_end, size; 567 vm_paddr_t page_range __unused; 568 vm_paddr_t last_pa, pa, startp, endp; 569 u_long pagecount; 570 #if MINIDUMP_PAGE_TRACKING 571 u_long vm_page_dump_size; 572 #endif 573 int biggestone, i, segind; 574 #ifdef WITNESS 575 vm_offset_t mapped; 576 int witness_size; 577 #endif 578 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 579 long ii; 580 #endif 581 int pool; 582 #ifdef VM_FREEPOOL_LAZYINIT 583 int lazyinit; 584 #endif 585 586 vaddr = round_page(vaddr); 587 588 vm_phys_early_startup(); 589 biggestone = vm_phys_avail_largest(); 590 end = phys_avail[biggestone+1]; 591 592 /* 593 * Initialize the page and queue locks. 594 */ 595 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 596 for (i = 0; i < vm_ndomains; i++) 597 vm_page_domain_init(i); 598 599 new_end = end; 600 #ifdef WITNESS 601 witness_size = round_page(witness_startup_count()); 602 new_end -= witness_size; 603 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 604 VM_PROT_READ | VM_PROT_WRITE); 605 bzero((void *)mapped, witness_size); 606 witness_startup((void *)mapped); 607 #endif 608 609 #if MINIDUMP_PAGE_TRACKING 610 /* 611 * Allocate a bitmap to indicate that a random physical page 612 * needs to be included in a minidump. 613 * 614 * The amd64 port needs this to indicate which direct map pages 615 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 616 * 617 * However, i386 still needs this workspace internally within the 618 * minidump code. In theory, they are not needed on i386, but are 619 * included should the sf_buf code decide to use them. 620 */ 621 last_pa = 0; 622 vm_page_dump_pages = 0; 623 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 624 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 625 dump_avail[i] / PAGE_SIZE; 626 if (dump_avail[i + 1] > last_pa) 627 last_pa = dump_avail[i + 1]; 628 } 629 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 630 new_end -= vm_page_dump_size; 631 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 632 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 633 bzero((void *)vm_page_dump, vm_page_dump_size); 634 #if MINIDUMP_STARTUP_PAGE_TRACKING 635 /* 636 * Include the UMA bootstrap pages, witness pages and vm_page_dump 637 * in a crash dump. When pmap_map() uses the direct map, they are 638 * not automatically included. 639 */ 640 for (pa = new_end; pa < end; pa += PAGE_SIZE) 641 dump_add_page(pa); 642 #endif 643 #else 644 (void)last_pa; 645 #endif 646 phys_avail[biggestone + 1] = new_end; 647 #ifdef __amd64__ 648 /* 649 * Request that the physical pages underlying the message buffer be 650 * included in a crash dump. Since the message buffer is accessed 651 * through the direct map, they are not automatically included. 652 */ 653 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 654 last_pa = pa + round_page(msgbufsize); 655 while (pa < last_pa) { 656 dump_add_page(pa); 657 pa += PAGE_SIZE; 658 } 659 #else 660 (void)pa; 661 #endif 662 663 /* 664 * Determine the lowest and highest physical addresses and, in the case 665 * of VM_PHYSSEG_SPARSE, the exact size of the available physical 666 * memory. vm_phys_early_startup() already checked that phys_avail[] 667 * has at least one element. 668 */ 669 #ifdef VM_PHYSSEG_SPARSE 670 size = phys_avail[1] - phys_avail[0]; 671 #endif 672 low_avail = phys_avail[0]; 673 high_avail = phys_avail[1]; 674 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 675 #ifdef VM_PHYSSEG_SPARSE 676 size += phys_avail[i + 1] - phys_avail[i]; 677 #endif 678 if (phys_avail[i] < low_avail) 679 low_avail = phys_avail[i]; 680 if (phys_avail[i + 1] > high_avail) 681 high_avail = phys_avail[i + 1]; 682 } 683 for (i = 0; i < vm_phys_nsegs; i++) { 684 #ifdef VM_PHYSSEG_SPARSE 685 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 686 #endif 687 if (vm_phys_segs[i].start < low_avail) 688 low_avail = vm_phys_segs[i].start; 689 if (vm_phys_segs[i].end > high_avail) 690 high_avail = vm_phys_segs[i].end; 691 } 692 first_page = low_avail / PAGE_SIZE; 693 #ifdef VM_PHYSSEG_DENSE 694 size = high_avail - low_avail; 695 #endif 696 697 #ifdef PMAP_HAS_PAGE_ARRAY 698 pmap_page_array_startup(size / PAGE_SIZE); 699 biggestone = vm_phys_avail_largest(); 700 end = new_end = phys_avail[biggestone + 1]; 701 #else 702 #ifdef VM_PHYSSEG_DENSE 703 /* 704 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 705 * the overhead of a page structure per page only if vm_page_array is 706 * allocated from the last physical memory chunk. Otherwise, we must 707 * allocate page structures representing the physical memory 708 * underlying vm_page_array, even though they will not be used. 709 */ 710 if (new_end != high_avail) 711 page_range = size / PAGE_SIZE; 712 else 713 #endif 714 { 715 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 716 717 /* 718 * If the partial bytes remaining are large enough for 719 * a page (PAGE_SIZE) without a corresponding 720 * 'struct vm_page', then new_end will contain an 721 * extra page after subtracting the length of the VM 722 * page array. Compensate by subtracting an extra 723 * page from new_end. 724 */ 725 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 726 if (new_end == high_avail) 727 high_avail -= PAGE_SIZE; 728 new_end -= PAGE_SIZE; 729 } 730 } 731 end = new_end; 732 new_end = vm_page_array_alloc(&vaddr, end, page_range); 733 #endif 734 735 #if VM_NRESERVLEVEL > 0 736 /* 737 * Allocate physical memory for the reservation management system's 738 * data structures, and map it. 739 */ 740 new_end = vm_reserv_startup(&vaddr, new_end); 741 #endif 742 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 743 /* 744 * Include vm_page_array and vm_reserv_array in a crash dump. 745 */ 746 for (pa = new_end; pa < end; pa += PAGE_SIZE) 747 dump_add_page(pa); 748 #endif 749 phys_avail[biggestone + 1] = new_end; 750 751 /* 752 * Add physical memory segments corresponding to the available 753 * physical pages. 754 */ 755 for (i = 0; phys_avail[i + 1] != 0; i += 2) 756 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 757 758 /* 759 * Initialize the physical memory allocator. 760 */ 761 vm_phys_init(); 762 763 pool = VM_FREEPOOL_DEFAULT; 764 #ifdef VM_FREEPOOL_LAZYINIT 765 lazyinit = 1; 766 TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit); 767 if (lazyinit) 768 pool = VM_FREEPOOL_LAZYINIT; 769 #endif 770 771 /* 772 * Initialize the page structures and add every available page to the 773 * physical memory allocator's free lists. 774 */ 775 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 776 for (ii = 0; ii < vm_page_array_size; ii++) { 777 m = &vm_page_array[ii]; 778 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0, 779 VM_FREEPOOL_DEFAULT); 780 m->flags = PG_FICTITIOUS; 781 } 782 #endif 783 vm_cnt.v_page_count = 0; 784 for (segind = 0; segind < vm_phys_nsegs; segind++) { 785 seg = &vm_phys_segs[segind]; 786 787 /* 788 * Initialize pages not covered by phys_avail[], since they 789 * might be freed to the allocator at some future point, e.g., 790 * by kmem_bootstrap_free(). 791 */ 792 startp = seg->start; 793 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 794 if (startp >= seg->end) 795 break; 796 if (phys_avail[i + 1] < startp) 797 continue; 798 if (phys_avail[i] <= startp) { 799 startp = phys_avail[i + 1]; 800 continue; 801 } 802 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 803 for (endp = MIN(phys_avail[i], seg->end); 804 startp < endp; startp += PAGE_SIZE, m++) { 805 vm_page_init_page(m, startp, segind, 806 VM_FREEPOOL_DEFAULT); 807 } 808 } 809 810 /* 811 * Add the segment's pages that are covered by one of 812 * phys_avail's ranges to the free lists. 813 */ 814 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 815 if (seg->end <= phys_avail[i] || 816 seg->start >= phys_avail[i + 1]) 817 continue; 818 819 startp = MAX(seg->start, phys_avail[i]); 820 endp = MIN(seg->end, phys_avail[i + 1]); 821 pagecount = (u_long)atop(endp - startp); 822 if (pagecount == 0) 823 continue; 824 825 /* 826 * If lazy vm_page initialization is not enabled, simply 827 * initialize all of the pages in the segment covered by 828 * phys_avail. Otherwise, initialize only the first 829 * page of each run of free pages handed to the vm_phys 830 * allocator, which in turn defers initialization of 831 * pages until they are needed. 832 * 833 * This avoids blocking the boot process for long 834 * periods, which may be relevant for VMs (which ought 835 * to boot as quickly as possible) and/or systems with 836 * large amounts of physical memory. 837 */ 838 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 839 vm_page_init_page(m, startp, segind, pool); 840 if (pool == VM_FREEPOOL_DEFAULT) { 841 for (u_long j = 1; j < pagecount; j++) { 842 vm_page_init_page(&m[j], 843 startp + ptoa((vm_paddr_t)j), 844 segind, pool); 845 } 846 } 847 vmd = VM_DOMAIN(seg->domain); 848 vm_domain_free_lock(vmd); 849 vm_phys_enqueue_contig(m, pool, pagecount); 850 vm_domain_free_unlock(vmd); 851 vm_domain_freecnt_inc(vmd, pagecount); 852 vm_cnt.v_page_count += (u_int)pagecount; 853 vmd->vmd_page_count += (u_int)pagecount; 854 vmd->vmd_segs |= 1UL << segind; 855 } 856 } 857 858 /* 859 * Remove blacklisted pages from the physical memory allocator. 860 */ 861 TAILQ_INIT(&blacklist_head); 862 vm_page_blacklist_load(&list, &listend); 863 vm_page_blacklist_check(list, listend); 864 865 list = kern_getenv("vm.blacklist"); 866 vm_page_blacklist_check(list, NULL); 867 868 freeenv(list); 869 #if VM_NRESERVLEVEL > 0 870 /* 871 * Initialize the reservation management system. 872 */ 873 vm_reserv_init(); 874 #endif 875 876 return (vaddr); 877 } 878 879 void 880 vm_page_reference(vm_page_t m) 881 { 882 883 vm_page_aflag_set(m, PGA_REFERENCED); 884 } 885 886 /* 887 * vm_page_trybusy 888 * 889 * Helper routine for grab functions to trylock busy. 890 * 891 * Returns true on success and false on failure. 892 */ 893 static bool 894 vm_page_trybusy(vm_page_t m, int allocflags) 895 { 896 897 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 898 return (vm_page_trysbusy(m)); 899 else 900 return (vm_page_tryxbusy(m)); 901 } 902 903 /* 904 * vm_page_tryacquire 905 * 906 * Helper routine for grab functions to trylock busy and wire. 907 * 908 * Returns true on success and false on failure. 909 */ 910 static inline bool 911 vm_page_tryacquire(vm_page_t m, int allocflags) 912 { 913 bool locked; 914 915 locked = vm_page_trybusy(m, allocflags); 916 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 917 vm_page_wire(m); 918 return (locked); 919 } 920 921 /* 922 * vm_page_busy_acquire: 923 * 924 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 925 * and drop the object lock if necessary. 926 */ 927 bool 928 vm_page_busy_acquire(vm_page_t m, int allocflags) 929 { 930 vm_object_t obj; 931 bool locked; 932 933 /* 934 * The page-specific object must be cached because page 935 * identity can change during the sleep, causing the 936 * re-lock of a different object. 937 * It is assumed that a reference to the object is already 938 * held by the callers. 939 */ 940 obj = atomic_load_ptr(&m->object); 941 for (;;) { 942 if (vm_page_tryacquire(m, allocflags)) 943 return (true); 944 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 945 return (false); 946 if (obj != NULL) 947 locked = VM_OBJECT_WOWNED(obj); 948 else 949 locked = false; 950 MPASS(locked || vm_page_wired(m)); 951 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 952 locked) && locked) 953 VM_OBJECT_WLOCK(obj); 954 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 955 return (false); 956 KASSERT(m->object == obj || m->object == NULL, 957 ("vm_page_busy_acquire: page %p does not belong to %p", 958 m, obj)); 959 } 960 } 961 962 /* 963 * vm_page_busy_downgrade: 964 * 965 * Downgrade an exclusive busy page into a single shared busy page. 966 */ 967 void 968 vm_page_busy_downgrade(vm_page_t m) 969 { 970 u_int x; 971 972 vm_page_assert_xbusied(m); 973 974 x = vm_page_busy_fetch(m); 975 for (;;) { 976 if (atomic_fcmpset_rel_int(&m->busy_lock, 977 &x, VPB_SHARERS_WORD(1))) 978 break; 979 } 980 if ((x & VPB_BIT_WAITERS) != 0) 981 wakeup(m); 982 } 983 984 /* 985 * 986 * vm_page_busy_tryupgrade: 987 * 988 * Attempt to upgrade a single shared busy into an exclusive busy. 989 */ 990 int 991 vm_page_busy_tryupgrade(vm_page_t m) 992 { 993 u_int ce, x; 994 995 vm_page_assert_sbusied(m); 996 997 x = vm_page_busy_fetch(m); 998 ce = VPB_CURTHREAD_EXCLUSIVE; 999 for (;;) { 1000 if (VPB_SHARERS(x) > 1) 1001 return (0); 1002 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1003 ("vm_page_busy_tryupgrade: invalid lock state")); 1004 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 1005 ce | (x & VPB_BIT_WAITERS))) 1006 continue; 1007 return (1); 1008 } 1009 } 1010 1011 /* 1012 * vm_page_sbusied: 1013 * 1014 * Return a positive value if the page is shared busied, 0 otherwise. 1015 */ 1016 int 1017 vm_page_sbusied(vm_page_t m) 1018 { 1019 u_int x; 1020 1021 x = vm_page_busy_fetch(m); 1022 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1023 } 1024 1025 /* 1026 * vm_page_sunbusy: 1027 * 1028 * Shared unbusy a page. 1029 */ 1030 void 1031 vm_page_sunbusy(vm_page_t m) 1032 { 1033 u_int x; 1034 1035 vm_page_assert_sbusied(m); 1036 1037 x = vm_page_busy_fetch(m); 1038 for (;;) { 1039 KASSERT(x != VPB_FREED, 1040 ("vm_page_sunbusy: Unlocking freed page.")); 1041 if (VPB_SHARERS(x) > 1) { 1042 if (atomic_fcmpset_int(&m->busy_lock, &x, 1043 x - VPB_ONE_SHARER)) 1044 break; 1045 continue; 1046 } 1047 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1048 ("vm_page_sunbusy: invalid lock state")); 1049 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1050 continue; 1051 if ((x & VPB_BIT_WAITERS) == 0) 1052 break; 1053 wakeup(m); 1054 break; 1055 } 1056 } 1057 1058 /* 1059 * vm_page_busy_sleep: 1060 * 1061 * Sleep if the page is busy, using the page pointer as wchan. 1062 * This is used to implement the hard-path of the busying mechanism. 1063 * 1064 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1065 * will not sleep if the page is shared-busy. 1066 * 1067 * The object lock must be held on entry. 1068 * 1069 * Returns true if it slept and dropped the object lock, or false 1070 * if there was no sleep and the lock is still held. 1071 */ 1072 bool 1073 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) 1074 { 1075 vm_object_t obj; 1076 1077 obj = m->object; 1078 VM_OBJECT_ASSERT_LOCKED(obj); 1079 1080 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, 1081 true)); 1082 } 1083 1084 /* 1085 * vm_page_busy_sleep_unlocked: 1086 * 1087 * Sleep if the page is busy, using the page pointer as wchan. 1088 * This is used to implement the hard-path of busying mechanism. 1089 * 1090 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1091 * will not sleep if the page is shared-busy. 1092 * 1093 * The object lock must not be held on entry. The operation will 1094 * return if the page changes identity. 1095 */ 1096 void 1097 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1098 const char *wmesg, int allocflags) 1099 { 1100 VM_OBJECT_ASSERT_UNLOCKED(obj); 1101 1102 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); 1103 } 1104 1105 /* 1106 * _vm_page_busy_sleep: 1107 * 1108 * Internal busy sleep function. Verifies the page identity and 1109 * lockstate against parameters. Returns true if it sleeps and 1110 * false otherwise. 1111 * 1112 * allocflags uses VM_ALLOC_* flags to specify the lock required. 1113 * 1114 * If locked is true the lock will be dropped for any true returns 1115 * and held for any false returns. 1116 */ 1117 static bool 1118 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1119 const char *wmesg, int allocflags, bool locked) 1120 { 1121 bool xsleep; 1122 u_int x; 1123 1124 /* 1125 * If the object is busy we must wait for that to drain to zero 1126 * before trying the page again. 1127 */ 1128 if (obj != NULL && vm_object_busied(obj)) { 1129 if (locked) 1130 VM_OBJECT_DROP(obj); 1131 vm_object_busy_wait(obj, wmesg); 1132 return (true); 1133 } 1134 1135 if (!vm_page_busied(m)) 1136 return (false); 1137 1138 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1139 sleepq_lock(m); 1140 x = vm_page_busy_fetch(m); 1141 do { 1142 /* 1143 * If the page changes objects or becomes unlocked we can 1144 * simply return. 1145 */ 1146 if (x == VPB_UNBUSIED || 1147 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1148 m->object != obj || m->pindex != pindex) { 1149 sleepq_release(m); 1150 return (false); 1151 } 1152 if ((x & VPB_BIT_WAITERS) != 0) 1153 break; 1154 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1155 if (locked) 1156 VM_OBJECT_DROP(obj); 1157 DROP_GIANT(); 1158 sleepq_add(m, NULL, wmesg, 0, 0); 1159 sleepq_wait(m, PVM); 1160 PICKUP_GIANT(); 1161 return (true); 1162 } 1163 1164 /* 1165 * vm_page_trysbusy: 1166 * 1167 * Try to shared busy a page. 1168 * If the operation succeeds 1 is returned otherwise 0. 1169 * The operation never sleeps. 1170 */ 1171 int 1172 vm_page_trysbusy(vm_page_t m) 1173 { 1174 vm_object_t obj; 1175 u_int x; 1176 1177 obj = m->object; 1178 x = vm_page_busy_fetch(m); 1179 for (;;) { 1180 if ((x & VPB_BIT_SHARED) == 0) 1181 return (0); 1182 /* 1183 * Reduce the window for transient busies that will trigger 1184 * false negatives in vm_page_ps_test(). 1185 */ 1186 if (obj != NULL && vm_object_busied(obj)) 1187 return (0); 1188 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1189 x + VPB_ONE_SHARER)) 1190 break; 1191 } 1192 1193 /* Refetch the object now that we're guaranteed that it is stable. */ 1194 obj = m->object; 1195 if (obj != NULL && vm_object_busied(obj)) { 1196 vm_page_sunbusy(m); 1197 return (0); 1198 } 1199 return (1); 1200 } 1201 1202 /* 1203 * vm_page_tryxbusy: 1204 * 1205 * Try to exclusive busy a page. 1206 * If the operation succeeds 1 is returned otherwise 0. 1207 * The operation never sleeps. 1208 */ 1209 int 1210 vm_page_tryxbusy(vm_page_t m) 1211 { 1212 vm_object_t obj; 1213 1214 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1215 VPB_CURTHREAD_EXCLUSIVE) == 0) 1216 return (0); 1217 1218 obj = m->object; 1219 if (obj != NULL && vm_object_busied(obj)) { 1220 vm_page_xunbusy(m); 1221 return (0); 1222 } 1223 return (1); 1224 } 1225 1226 static void 1227 vm_page_xunbusy_hard_tail(vm_page_t m) 1228 { 1229 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1230 /* Wake the waiter. */ 1231 wakeup(m); 1232 } 1233 1234 /* 1235 * vm_page_xunbusy_hard: 1236 * 1237 * Called when unbusy has failed because there is a waiter. 1238 */ 1239 void 1240 vm_page_xunbusy_hard(vm_page_t m) 1241 { 1242 vm_page_assert_xbusied(m); 1243 vm_page_xunbusy_hard_tail(m); 1244 } 1245 1246 void 1247 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1248 { 1249 vm_page_assert_xbusied_unchecked(m); 1250 vm_page_xunbusy_hard_tail(m); 1251 } 1252 1253 static void 1254 vm_page_busy_free(vm_page_t m) 1255 { 1256 u_int x; 1257 1258 atomic_thread_fence_rel(); 1259 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1260 if ((x & VPB_BIT_WAITERS) != 0) 1261 wakeup(m); 1262 } 1263 1264 /* 1265 * vm_page_unhold_pages: 1266 * 1267 * Unhold each of the pages that is referenced by the given array. 1268 */ 1269 void 1270 vm_page_unhold_pages(vm_page_t *ma, int count) 1271 { 1272 1273 for (; count != 0; count--) { 1274 vm_page_unwire(*ma, PQ_ACTIVE); 1275 ma++; 1276 } 1277 } 1278 1279 vm_page_t 1280 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1281 { 1282 vm_page_t m; 1283 1284 #ifdef VM_PHYSSEG_SPARSE 1285 m = vm_phys_paddr_to_vm_page(pa); 1286 if (m == NULL) 1287 m = vm_phys_fictitious_to_vm_page(pa); 1288 return (m); 1289 #elif defined(VM_PHYSSEG_DENSE) 1290 long pi; 1291 1292 pi = atop(pa); 1293 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1294 m = &vm_page_array[pi - first_page]; 1295 return (m); 1296 } 1297 return (vm_phys_fictitious_to_vm_page(pa)); 1298 #else 1299 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1300 #endif 1301 } 1302 1303 /* 1304 * vm_page_getfake: 1305 * 1306 * Create a fictitious page with the specified physical address and 1307 * memory attribute. The memory attribute is the only the machine- 1308 * dependent aspect of a fictitious page that must be initialized. 1309 */ 1310 vm_page_t 1311 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1312 { 1313 vm_page_t m; 1314 1315 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1316 vm_page_initfake(m, paddr, memattr); 1317 return (m); 1318 } 1319 1320 void 1321 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1322 { 1323 1324 if ((m->flags & PG_FICTITIOUS) != 0) { 1325 /* 1326 * The page's memattr might have changed since the 1327 * previous initialization. Update the pmap to the 1328 * new memattr. 1329 */ 1330 goto memattr; 1331 } 1332 m->phys_addr = paddr; 1333 m->a.queue = PQ_NONE; 1334 /* Fictitious pages don't use "segind". */ 1335 m->flags = PG_FICTITIOUS; 1336 /* Fictitious pages don't use "order" or "pool". */ 1337 m->oflags = VPO_UNMANAGED; 1338 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1339 /* Fictitious pages are unevictable. */ 1340 m->ref_count = 1; 1341 pmap_page_init(m); 1342 memattr: 1343 pmap_page_set_memattr(m, memattr); 1344 } 1345 1346 /* 1347 * vm_page_putfake: 1348 * 1349 * Release a fictitious page. 1350 */ 1351 void 1352 vm_page_putfake(vm_page_t m) 1353 { 1354 1355 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1356 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1357 ("vm_page_putfake: bad page %p", m)); 1358 vm_page_assert_xbusied(m); 1359 vm_page_busy_free(m); 1360 uma_zfree(fakepg_zone, m); 1361 } 1362 1363 /* 1364 * vm_page_updatefake: 1365 * 1366 * Update the given fictitious page to the specified physical address and 1367 * memory attribute. 1368 */ 1369 void 1370 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1371 { 1372 1373 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1374 ("vm_page_updatefake: bad page %p", m)); 1375 m->phys_addr = paddr; 1376 pmap_page_set_memattr(m, memattr); 1377 } 1378 1379 /* 1380 * vm_page_free: 1381 * 1382 * Free a page. 1383 */ 1384 void 1385 vm_page_free(vm_page_t m) 1386 { 1387 1388 m->flags &= ~PG_ZERO; 1389 vm_page_free_toq(m); 1390 } 1391 1392 /* 1393 * vm_page_free_zero: 1394 * 1395 * Free a page to the zerod-pages queue 1396 */ 1397 void 1398 vm_page_free_zero(vm_page_t m) 1399 { 1400 1401 m->flags |= PG_ZERO; 1402 vm_page_free_toq(m); 1403 } 1404 1405 /* 1406 * Unbusy and handle the page queueing for a page from a getpages request that 1407 * was optionally read ahead or behind. 1408 */ 1409 void 1410 vm_page_readahead_finish(vm_page_t m) 1411 { 1412 1413 /* We shouldn't put invalid pages on queues. */ 1414 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1415 1416 /* 1417 * Since the page is not the actually needed one, whether it should 1418 * be activated or deactivated is not obvious. Empirical results 1419 * have shown that deactivating the page is usually the best choice, 1420 * unless the page is wanted by another thread. 1421 */ 1422 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1423 vm_page_activate(m); 1424 else 1425 vm_page_deactivate(m); 1426 vm_page_xunbusy_unchecked(m); 1427 } 1428 1429 /* 1430 * Destroy the identity of an invalid page and free it if possible. 1431 * This is intended to be used when reading a page from backing store fails. 1432 */ 1433 void 1434 vm_page_free_invalid(vm_page_t m) 1435 { 1436 1437 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1438 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1439 KASSERT(m->object != NULL, ("page %p has no object", m)); 1440 VM_OBJECT_ASSERT_WLOCKED(m->object); 1441 1442 /* 1443 * We may be attempting to free the page as part of the handling for an 1444 * I/O error, in which case the page was xbusied by a different thread. 1445 */ 1446 vm_page_xbusy_claim(m); 1447 1448 /* 1449 * If someone has wired this page while the object lock 1450 * was not held, then the thread that unwires is responsible 1451 * for freeing the page. Otherwise just free the page now. 1452 * The wire count of this unmapped page cannot change while 1453 * we have the page xbusy and the page's object wlocked. 1454 */ 1455 if (vm_page_remove(m)) 1456 vm_page_free(m); 1457 } 1458 1459 /* 1460 * vm_page_dirty_KBI: [ internal use only ] 1461 * 1462 * Set all bits in the page's dirty field. 1463 * 1464 * The object containing the specified page must be locked if the 1465 * call is made from the machine-independent layer. 1466 * 1467 * See vm_page_clear_dirty_mask(). 1468 * 1469 * This function should only be called by vm_page_dirty(). 1470 */ 1471 void 1472 vm_page_dirty_KBI(vm_page_t m) 1473 { 1474 1475 /* Refer to this operation by its public name. */ 1476 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1477 m->dirty = VM_PAGE_BITS_ALL; 1478 } 1479 1480 /* 1481 * Insert the given page into the given object at the given pindex. 1482 * 1483 * The procedure is marked __always_inline to suggest to the compiler to 1484 * eliminate the iter parameter and the associated alternate branch. 1485 */ 1486 static __always_inline int 1487 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1488 bool iter, struct pctrie_iter *pages) 1489 { 1490 int error; 1491 1492 VM_OBJECT_ASSERT_WLOCKED(object); 1493 KASSERT(m->object == NULL, 1494 ("vm_page_insert: page %p already inserted", m)); 1495 1496 /* 1497 * Record the object/offset pair in this page. 1498 */ 1499 m->object = object; 1500 m->pindex = pindex; 1501 m->ref_count |= VPRC_OBJREF; 1502 1503 /* 1504 * Add this page to the object's radix tree. 1505 */ 1506 if (iter) 1507 error = vm_radix_iter_insert(pages, m); 1508 else 1509 error = vm_radix_insert(&object->rtree, m); 1510 if (__predict_false(error != 0)) { 1511 m->object = NULL; 1512 m->pindex = 0; 1513 m->ref_count &= ~VPRC_OBJREF; 1514 return (1); 1515 } 1516 1517 vm_page_insert_radixdone(m, object); 1518 vm_pager_page_inserted(object, m); 1519 return (0); 1520 } 1521 1522 /* 1523 * vm_page_insert: [ internal use only ] 1524 * 1525 * Inserts the given mem entry into the object and object list. 1526 * 1527 * The object must be locked. 1528 */ 1529 int 1530 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1531 { 1532 return (vm_page_insert_lookup(m, object, pindex, false, NULL)); 1533 } 1534 1535 /* 1536 * vm_page_iter_insert: 1537 * 1538 * Tries to insert the page "m" into the specified object at offset 1539 * "pindex" using the iterator "pages". Returns 0 if the insertion was 1540 * successful. 1541 * 1542 * The object must be locked. 1543 */ 1544 int 1545 vm_page_iter_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1546 struct pctrie_iter *pages) 1547 { 1548 return (vm_page_insert_lookup(m, object, pindex, true, pages)); 1549 } 1550 1551 /* 1552 * vm_page_insert_radixdone: 1553 * 1554 * Complete page "m" insertion into the specified object after the 1555 * radix trie hooking. 1556 * 1557 * The object must be locked. 1558 */ 1559 static void 1560 vm_page_insert_radixdone(vm_page_t m, vm_object_t object) 1561 { 1562 1563 VM_OBJECT_ASSERT_WLOCKED(object); 1564 KASSERT(object != NULL && m->object == object, 1565 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1566 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1567 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1568 1569 /* 1570 * Show that the object has one more resident page. 1571 */ 1572 object->resident_page_count++; 1573 1574 /* 1575 * Hold the vnode until the last page is released. 1576 */ 1577 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1578 vhold(object->handle); 1579 1580 /* 1581 * Since we are inserting a new and possibly dirty page, 1582 * update the object's generation count. 1583 */ 1584 if (pmap_page_is_write_mapped(m)) 1585 vm_object_set_writeable_dirty(object); 1586 } 1587 1588 /* 1589 * vm_page_remove_radixdone 1590 * 1591 * Complete page "m" removal from the specified object after the radix trie 1592 * unhooking. 1593 * 1594 * The caller is responsible for updating the page's fields to reflect this 1595 * removal. 1596 */ 1597 static void 1598 vm_page_remove_radixdone(vm_page_t m) 1599 { 1600 vm_object_t object; 1601 1602 vm_page_assert_xbusied(m); 1603 object = m->object; 1604 VM_OBJECT_ASSERT_WLOCKED(object); 1605 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1606 ("page %p is missing its object ref", m)); 1607 1608 /* Deferred free of swap space. */ 1609 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1610 vm_pager_page_unswapped(m); 1611 1612 vm_pager_page_removed(object, m); 1613 m->object = NULL; 1614 1615 /* 1616 * And show that the object has one fewer resident page. 1617 */ 1618 object->resident_page_count--; 1619 1620 /* 1621 * The vnode may now be recycled. 1622 */ 1623 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1624 vdrop(object->handle); 1625 } 1626 1627 /* 1628 * vm_page_free_object_prep: 1629 * 1630 * Disassociates the given page from its VM object. 1631 * 1632 * The object must be locked, and the page must be xbusy. 1633 */ 1634 static void 1635 vm_page_free_object_prep(vm_page_t m) 1636 { 1637 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 1638 ((m->object->flags & OBJ_UNMANAGED) != 0), 1639 ("%s: managed flag mismatch for page %p", 1640 __func__, m)); 1641 vm_page_assert_xbusied(m); 1642 1643 /* 1644 * The object reference can be released without an atomic 1645 * operation. 1646 */ 1647 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 1648 m->ref_count == VPRC_OBJREF, 1649 ("%s: page %p has unexpected ref_count %u", 1650 __func__, m, m->ref_count)); 1651 vm_page_remove_radixdone(m); 1652 m->ref_count -= VPRC_OBJREF; 1653 } 1654 1655 /* 1656 * vm_page_iter_free: 1657 * 1658 * Free the given page, and use the iterator to remove it from the radix 1659 * tree. 1660 */ 1661 void 1662 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m) 1663 { 1664 vm_radix_iter_remove(pages); 1665 vm_page_free_object_prep(m); 1666 vm_page_xunbusy(m); 1667 m->flags &= ~PG_ZERO; 1668 vm_page_free_toq(m); 1669 } 1670 1671 /* 1672 * vm_page_remove: 1673 * 1674 * Removes the specified page from its containing object, but does not 1675 * invalidate any backing storage. Returns true if the object's reference 1676 * was the last reference to the page, and false otherwise. 1677 * 1678 * The object must be locked and the page must be exclusively busied. 1679 * The exclusive busy will be released on return. If this is not the 1680 * final ref and the caller does not hold a wire reference it may not 1681 * continue to access the page. 1682 */ 1683 bool 1684 vm_page_remove(vm_page_t m) 1685 { 1686 bool dropped; 1687 1688 dropped = vm_page_remove_xbusy(m); 1689 vm_page_xunbusy(m); 1690 1691 return (dropped); 1692 } 1693 1694 /* 1695 * vm_page_iter_remove: 1696 * 1697 * Remove the current page, and use the iterator to remove it from the 1698 * radix tree. 1699 */ 1700 bool 1701 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m) 1702 { 1703 bool dropped; 1704 1705 vm_radix_iter_remove(pages); 1706 vm_page_remove_radixdone(m); 1707 dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1708 vm_page_xunbusy(m); 1709 1710 return (dropped); 1711 } 1712 1713 /* 1714 * vm_page_radix_remove 1715 * 1716 * Removes the specified page from the radix tree. 1717 */ 1718 static void 1719 vm_page_radix_remove(vm_page_t m) 1720 { 1721 vm_page_t mrem __diagused; 1722 1723 mrem = vm_radix_remove(&m->object->rtree, m->pindex); 1724 KASSERT(mrem == m, 1725 ("removed page %p, expected page %p", mrem, m)); 1726 } 1727 1728 /* 1729 * vm_page_remove_xbusy 1730 * 1731 * Removes the page but leaves the xbusy held. Returns true if this 1732 * removed the final ref and false otherwise. 1733 */ 1734 bool 1735 vm_page_remove_xbusy(vm_page_t m) 1736 { 1737 1738 vm_page_radix_remove(m); 1739 vm_page_remove_radixdone(m); 1740 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1741 } 1742 1743 /* 1744 * vm_page_lookup: 1745 * 1746 * Returns the page associated with the object/offset 1747 * pair specified; if none is found, NULL is returned. 1748 * 1749 * The object must be locked. 1750 */ 1751 vm_page_t 1752 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1753 { 1754 1755 VM_OBJECT_ASSERT_LOCKED(object); 1756 return (vm_radix_lookup(&object->rtree, pindex)); 1757 } 1758 1759 /* 1760 * vm_page_iter_init: 1761 * 1762 * Initialize iterator for vm pages. 1763 */ 1764 void 1765 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object) 1766 { 1767 1768 vm_radix_iter_init(pages, &object->rtree); 1769 } 1770 1771 /* 1772 * vm_page_iter_init: 1773 * 1774 * Initialize iterator for vm pages. 1775 */ 1776 void 1777 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object, 1778 vm_pindex_t limit) 1779 { 1780 1781 vm_radix_iter_limit_init(pages, &object->rtree, limit); 1782 } 1783 1784 /* 1785 * vm_page_lookup_unlocked: 1786 * 1787 * Returns the page associated with the object/offset pair specified; 1788 * if none is found, NULL is returned. The page may be no longer be 1789 * present in the object at the time that this function returns. Only 1790 * useful for opportunistic checks such as inmem(). 1791 */ 1792 vm_page_t 1793 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1794 { 1795 1796 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1797 } 1798 1799 /* 1800 * vm_page_relookup: 1801 * 1802 * Returns a page that must already have been busied by 1803 * the caller. Used for bogus page replacement. 1804 */ 1805 vm_page_t 1806 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1807 { 1808 vm_page_t m; 1809 1810 m = vm_page_lookup_unlocked(object, pindex); 1811 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1812 m->object == object && m->pindex == pindex, 1813 ("vm_page_relookup: Invalid page %p", m)); 1814 return (m); 1815 } 1816 1817 /* 1818 * This should only be used by lockless functions for releasing transient 1819 * incorrect acquires. The page may have been freed after we acquired a 1820 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1821 * further to do. 1822 */ 1823 static void 1824 vm_page_busy_release(vm_page_t m) 1825 { 1826 u_int x; 1827 1828 x = vm_page_busy_fetch(m); 1829 for (;;) { 1830 if (x == VPB_FREED) 1831 break; 1832 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1833 if (atomic_fcmpset_int(&m->busy_lock, &x, 1834 x - VPB_ONE_SHARER)) 1835 break; 1836 continue; 1837 } 1838 KASSERT((x & VPB_BIT_SHARED) != 0 || 1839 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1840 ("vm_page_busy_release: %p xbusy not owned.", m)); 1841 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1842 continue; 1843 if ((x & VPB_BIT_WAITERS) != 0) 1844 wakeup(m); 1845 break; 1846 } 1847 } 1848 1849 /* 1850 * Uses the page mnew as a replacement for an existing page at index 1851 * pindex which must be already present in the object. 1852 * 1853 * Both pages must be exclusively busied on enter. The old page is 1854 * unbusied on exit. 1855 * 1856 * A return value of true means mold is now free. If this is not the 1857 * final ref and the caller does not hold a wire reference it may not 1858 * continue to access the page. 1859 */ 1860 static bool 1861 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1862 vm_page_t mold) 1863 { 1864 vm_page_t mret __diagused; 1865 bool dropped; 1866 1867 VM_OBJECT_ASSERT_WLOCKED(object); 1868 vm_page_assert_xbusied(mold); 1869 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1870 ("vm_page_replace: page %p already in object", mnew)); 1871 1872 /* 1873 * This function mostly follows vm_page_insert() and 1874 * vm_page_remove() without the radix, object count and vnode 1875 * dance. Double check such functions for more comments. 1876 */ 1877 1878 mnew->object = object; 1879 mnew->pindex = pindex; 1880 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1881 mret = vm_radix_replace(&object->rtree, mnew); 1882 KASSERT(mret == mold, 1883 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1884 KASSERT((mold->oflags & VPO_UNMANAGED) == 1885 (mnew->oflags & VPO_UNMANAGED), 1886 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1887 1888 mold->object = NULL; 1889 1890 /* 1891 * The object's resident_page_count does not change because we have 1892 * swapped one page for another, but the generation count should 1893 * change if the page is dirty. 1894 */ 1895 if (pmap_page_is_write_mapped(mnew)) 1896 vm_object_set_writeable_dirty(object); 1897 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1898 vm_page_xunbusy(mold); 1899 1900 return (dropped); 1901 } 1902 1903 void 1904 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1905 vm_page_t mold) 1906 { 1907 1908 vm_page_assert_xbusied(mnew); 1909 1910 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1911 vm_page_free(mold); 1912 } 1913 1914 /* 1915 * vm_page_iter_rename: 1916 * 1917 * Tries to move the specified page from its current object to a new object 1918 * and pindex, using the given iterator to remove the page from its current 1919 * object. Returns true if the move was successful, and false if the move 1920 * was aborted due to a failed memory allocation. 1921 * 1922 * Panics if a page already resides in the new object at the new pindex. 1923 * 1924 * This routine dirties the page if it is valid, as callers are expected to 1925 * transfer backing storage only after moving the page. Dirtying the page 1926 * ensures that the destination object retains the most recent copy of the 1927 * page. 1928 * 1929 * The objects must be locked. 1930 */ 1931 bool 1932 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m, 1933 vm_object_t new_object, vm_pindex_t new_pindex) 1934 { 1935 vm_pindex_t opidx; 1936 1937 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1938 ("%s: page %p is missing object ref", __func__, m)); 1939 VM_OBJECT_ASSERT_WLOCKED(m->object); 1940 VM_OBJECT_ASSERT_WLOCKED(new_object); 1941 1942 /* 1943 * Create a custom version of vm_page_insert() which does not depend 1944 * by m_prev and can cheat on the implementation aspects of the 1945 * function. 1946 */ 1947 opidx = m->pindex; 1948 m->pindex = new_pindex; 1949 if (vm_radix_insert(&new_object->rtree, m) != 0) { 1950 m->pindex = opidx; 1951 return (false); 1952 } 1953 1954 /* 1955 * The operation cannot fail anymore. 1956 */ 1957 m->pindex = opidx; 1958 vm_radix_iter_remove(old_pages); 1959 vm_page_remove_radixdone(m); 1960 1961 /* Return back to the new pindex to complete vm_page_insert(). */ 1962 m->pindex = new_pindex; 1963 m->object = new_object; 1964 1965 vm_page_insert_radixdone(m, new_object); 1966 if (vm_page_any_valid(m)) 1967 vm_page_dirty(m); 1968 vm_pager_page_inserted(new_object, m); 1969 return (true); 1970 } 1971 1972 /* 1973 * vm_page_alloc: 1974 * 1975 * Allocate and return a page that is associated with the specified 1976 * object and offset pair. By default, this page is exclusive busied. 1977 * 1978 * The caller must always specify an allocation class. 1979 * 1980 * allocation classes: 1981 * VM_ALLOC_NORMAL normal process request 1982 * VM_ALLOC_SYSTEM system *really* needs a page 1983 * VM_ALLOC_INTERRUPT interrupt time request 1984 * 1985 * optional allocation flags: 1986 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1987 * intends to allocate 1988 * VM_ALLOC_NOBUSY do not exclusive busy the page 1989 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1990 * VM_ALLOC_NOFREE page will never be freed 1991 * VM_ALLOC_NOWAIT ignored (default behavior) 1992 * VM_ALLOC_SBUSY shared busy the allocated page 1993 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 1994 * VM_ALLOC_WIRED wire the allocated page 1995 * VM_ALLOC_ZERO prefer a zeroed page 1996 */ 1997 vm_page_t 1998 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1999 { 2000 struct pctrie_iter pages; 2001 2002 vm_page_iter_init(&pages, object); 2003 return (vm_page_alloc_iter(object, pindex, req, &pages)); 2004 } 2005 2006 /* 2007 * Allocate a page in the specified object with the given page index. If the 2008 * object lock is dropped and regained, the pages iter is reset. 2009 */ 2010 vm_page_t 2011 vm_page_alloc_iter(vm_object_t object, vm_pindex_t pindex, int req, 2012 struct pctrie_iter *pages) 2013 { 2014 struct vm_domainset_iter di; 2015 vm_page_t m; 2016 int domain; 2017 2018 if (vm_domainset_iter_page_init(&di, object, pindex, &domain, &req) != 0) 2019 return (NULL); 2020 2021 do { 2022 m = vm_page_alloc_domain_iter(object, pindex, domain, req, 2023 pages); 2024 if (m != NULL) 2025 break; 2026 } while (vm_domainset_iter_page(&di, object, &domain, pages) == 0); 2027 2028 return (m); 2029 } 2030 2031 /* 2032 * Returns true if the number of free pages exceeds the minimum 2033 * for the request class and false otherwise. 2034 */ 2035 static int 2036 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2037 { 2038 u_int limit, old, new; 2039 2040 if (req_class == VM_ALLOC_INTERRUPT) 2041 limit = 0; 2042 else if (req_class == VM_ALLOC_SYSTEM) 2043 limit = vmd->vmd_interrupt_free_min; 2044 else 2045 limit = vmd->vmd_free_reserved; 2046 2047 /* 2048 * Attempt to reserve the pages. Fail if we're below the limit. 2049 */ 2050 limit += npages; 2051 old = atomic_load_int(&vmd->vmd_free_count); 2052 do { 2053 if (old < limit) 2054 return (0); 2055 new = old - npages; 2056 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2057 2058 /* Wake the page daemon if we've crossed the threshold. */ 2059 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2060 pagedaemon_wakeup(vmd->vmd_domain); 2061 2062 /* Only update bitsets on transitions. */ 2063 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2064 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2065 vm_domain_set(vmd); 2066 2067 return (1); 2068 } 2069 2070 int 2071 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2072 { 2073 int req_class; 2074 2075 /* 2076 * The page daemon is allowed to dig deeper into the free page list. 2077 */ 2078 req_class = req & VM_ALLOC_CLASS_MASK; 2079 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2080 req_class = VM_ALLOC_SYSTEM; 2081 return (_vm_domain_allocate(vmd, req_class, npages)); 2082 } 2083 2084 vm_page_t 2085 vm_page_alloc_domain_iter(vm_object_t object, vm_pindex_t pindex, int domain, 2086 int req, struct pctrie_iter *pages) 2087 { 2088 struct vm_domain *vmd; 2089 vm_page_t m; 2090 int flags; 2091 2092 #define VM_ALLOC_COMMON (VM_ALLOC_CLASS_MASK | VM_ALLOC_NODUMP | \ 2093 VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | \ 2094 VM_ALLOC_WIRED | VM_ALLOC_ZERO) 2095 #define VPA_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2096 VM_ALLOC_NOBUSY | VM_ALLOC_NOFREE | \ 2097 VM_ALLOC_SBUSY) 2098 KASSERT((req & ~VPA_FLAGS) == 0, 2099 ("invalid request %#x", req)); 2100 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2101 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2102 ("invalid request %#x", req)); 2103 VM_OBJECT_ASSERT_WLOCKED(object); 2104 2105 flags = 0; 2106 m = NULL; 2107 if (!vm_pager_can_alloc_page(object, pindex)) 2108 return (NULL); 2109 #if VM_NRESERVLEVEL > 0 2110 again: 2111 #endif 2112 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2113 m = vm_page_alloc_nofree_domain(domain, req); 2114 if (m != NULL) 2115 goto found; 2116 } 2117 #if VM_NRESERVLEVEL > 0 2118 /* 2119 * Can we allocate the page from a reservation? 2120 */ 2121 if (vm_object_reserv(object) && 2122 (m = vm_reserv_alloc_page(object, pindex, domain, req, pages)) != 2123 NULL) { 2124 goto found; 2125 } 2126 #endif 2127 vmd = VM_DOMAIN(domain); 2128 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2129 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2130 M_NOWAIT | M_NOVM); 2131 if (m != NULL) { 2132 flags |= PG_PCPU_CACHE; 2133 goto found; 2134 } 2135 } 2136 if (vm_domain_allocate(vmd, req, 1)) { 2137 /* 2138 * If not, allocate it from the free page queues. 2139 */ 2140 vm_domain_free_lock(vmd); 2141 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2142 vm_domain_free_unlock(vmd); 2143 if (m == NULL) { 2144 vm_domain_freecnt_inc(vmd, 1); 2145 #if VM_NRESERVLEVEL > 0 2146 if (vm_reserv_reclaim_inactive(domain)) 2147 goto again; 2148 #endif 2149 } 2150 } 2151 if (m == NULL) { 2152 /* 2153 * Not allocatable, give up. 2154 */ 2155 (void)vm_domain_alloc_fail(vmd, object, req); 2156 if ((req & VM_ALLOC_WAITFAIL) != 0) 2157 pctrie_iter_reset(pages); 2158 return (NULL); 2159 } 2160 2161 /* 2162 * At this point we had better have found a good page. 2163 */ 2164 found: 2165 vm_page_dequeue(m); 2166 vm_page_alloc_check(m); 2167 2168 /* 2169 * Initialize the page. Only the PG_ZERO flag is inherited. 2170 */ 2171 flags |= m->flags & PG_ZERO; 2172 if ((req & VM_ALLOC_NODUMP) != 0) 2173 flags |= PG_NODUMP; 2174 if ((req & VM_ALLOC_NOFREE) != 0) 2175 flags |= PG_NOFREE; 2176 m->flags = flags; 2177 m->a.flags = 0; 2178 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2179 m->pool = VM_FREEPOOL_DEFAULT; 2180 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2181 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2182 else if ((req & VM_ALLOC_SBUSY) != 0) 2183 m->busy_lock = VPB_SHARERS_WORD(1); 2184 else 2185 m->busy_lock = VPB_UNBUSIED; 2186 if (req & VM_ALLOC_WIRED) { 2187 vm_wire_add(1); 2188 m->ref_count = 1; 2189 } 2190 m->a.act_count = 0; 2191 2192 if (vm_page_iter_insert(m, object, pindex, pages)) { 2193 if (req & VM_ALLOC_WIRED) { 2194 vm_wire_sub(1); 2195 m->ref_count = 0; 2196 } 2197 KASSERT(m->object == NULL, ("page %p has object", m)); 2198 m->oflags = VPO_UNMANAGED; 2199 m->busy_lock = VPB_UNBUSIED; 2200 /* Don't change PG_ZERO. */ 2201 vm_page_free_toq(m); 2202 if (req & VM_ALLOC_WAITFAIL) { 2203 VM_OBJECT_WUNLOCK(object); 2204 vm_radix_wait(); 2205 pctrie_iter_reset(pages); 2206 VM_OBJECT_WLOCK(object); 2207 } 2208 return (NULL); 2209 } 2210 2211 /* Ignore device objects; the pager sets "memattr" for them. */ 2212 if (object->memattr != VM_MEMATTR_DEFAULT && 2213 (object->flags & OBJ_FICTITIOUS) == 0) 2214 pmap_page_set_memattr(m, object->memattr); 2215 2216 return (m); 2217 } 2218 2219 /* 2220 * vm_page_alloc_contig: 2221 * 2222 * Allocate a contiguous set of physical pages of the given size "npages" 2223 * from the free lists. All of the physical pages must be at or above 2224 * the given physical address "low" and below the given physical address 2225 * "high". The given value "alignment" determines the alignment of the 2226 * first physical page in the set. If the given value "boundary" is 2227 * non-zero, then the set of physical pages cannot cross any physical 2228 * address boundary that is a multiple of that value. Both "alignment" 2229 * and "boundary" must be a power of two. 2230 * 2231 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2232 * then the memory attribute setting for the physical pages is configured 2233 * to the object's memory attribute setting. Otherwise, the memory 2234 * attribute setting for the physical pages is configured to "memattr", 2235 * overriding the object's memory attribute setting. However, if the 2236 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2237 * memory attribute setting for the physical pages cannot be configured 2238 * to VM_MEMATTR_DEFAULT. 2239 * 2240 * The specified object may not contain fictitious pages. 2241 * 2242 * The caller must always specify an allocation class. 2243 * 2244 * allocation classes: 2245 * VM_ALLOC_NORMAL normal process request 2246 * VM_ALLOC_SYSTEM system *really* needs the pages 2247 * VM_ALLOC_INTERRUPT interrupt time request 2248 * 2249 * optional allocation flags: 2250 * VM_ALLOC_NOBUSY do not exclusive busy the pages 2251 * VM_ALLOC_NODUMP do not include the pages in a kernel core dump 2252 * VM_ALLOC_NORECLAIM do not reclaim after initial failure 2253 * VM_ALLOC_NOWAIT ignored (default behavior) 2254 * VM_ALLOC_SBUSY shared busy the allocated pages 2255 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 2256 * VM_ALLOC_WIRED wire the allocated pages 2257 * VM_ALLOC_ZERO prefer zeroed pages 2258 */ 2259 vm_page_t 2260 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2261 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2262 vm_paddr_t boundary, vm_memattr_t memattr) 2263 { 2264 struct vm_domainset_iter di; 2265 vm_page_t bounds[2]; 2266 vm_page_t m; 2267 int domain; 2268 int start_segind; 2269 2270 start_segind = -1; 2271 2272 if (vm_domainset_iter_page_init(&di, object, pindex, &domain, &req) != 0) 2273 return (NULL); 2274 2275 do { 2276 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2277 npages, low, high, alignment, boundary, memattr); 2278 if (m != NULL) 2279 break; 2280 if (start_segind == -1) 2281 start_segind = vm_phys_lookup_segind(low); 2282 if (vm_phys_find_range(bounds, start_segind, domain, 2283 npages, low, high) == -1) { 2284 vm_domainset_iter_ignore(&di, domain); 2285 } 2286 } while (vm_domainset_iter_page(&di, object, &domain, NULL) == 0); 2287 2288 return (m); 2289 } 2290 2291 static vm_page_t 2292 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2293 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2294 { 2295 struct vm_domain *vmd; 2296 vm_page_t m_ret; 2297 2298 /* 2299 * Can we allocate the pages without the number of free pages falling 2300 * below the lower bound for the allocation class? 2301 */ 2302 vmd = VM_DOMAIN(domain); 2303 if (!vm_domain_allocate(vmd, req, npages)) 2304 return (NULL); 2305 /* 2306 * Try to allocate the pages from the free page queues. 2307 */ 2308 vm_domain_free_lock(vmd); 2309 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2310 alignment, boundary); 2311 vm_domain_free_unlock(vmd); 2312 if (m_ret != NULL) 2313 return (m_ret); 2314 #if VM_NRESERVLEVEL > 0 2315 /* 2316 * Try to break a reservation to allocate the pages. 2317 */ 2318 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2319 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2320 high, alignment, boundary); 2321 if (m_ret != NULL) 2322 return (m_ret); 2323 } 2324 #endif 2325 vm_domain_freecnt_inc(vmd, npages); 2326 return (NULL); 2327 } 2328 2329 vm_page_t 2330 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2331 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2332 vm_paddr_t boundary, vm_memattr_t memattr) 2333 { 2334 struct pctrie_iter pages; 2335 vm_page_t m, m_ret, mpred; 2336 u_int busy_lock, flags, oflags; 2337 2338 #define VPAC_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2339 VM_ALLOC_NOBUSY | VM_ALLOC_NORECLAIM | \ 2340 VM_ALLOC_SBUSY) 2341 KASSERT((req & ~VPAC_FLAGS) == 0, 2342 ("invalid request %#x", req)); 2343 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2344 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2345 ("invalid request %#x", req)); 2346 VM_OBJECT_ASSERT_WLOCKED(object); 2347 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2348 ("vm_page_alloc_contig: object %p has fictitious pages", 2349 object)); 2350 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2351 2352 vm_page_iter_init(&pages, object); 2353 m_ret = NULL; 2354 #if VM_NRESERVLEVEL > 0 2355 /* 2356 * Can we allocate the pages from a reservation? 2357 */ 2358 if (vm_object_reserv(object)) { 2359 m_ret = vm_reserv_alloc_contig(object, pindex, domain, 2360 req, npages, low, high, alignment, boundary, &pages); 2361 } 2362 #endif 2363 if (m_ret == NULL) { 2364 m_ret = vm_page_find_contig_domain(domain, req, npages, 2365 low, high, alignment, boundary); 2366 } 2367 if (m_ret == NULL) { 2368 (void)vm_domain_alloc_fail(VM_DOMAIN(domain), object, req); 2369 return (NULL); 2370 } 2371 2372 /* 2373 * Initialize the pages. Only the PG_ZERO flag is inherited. 2374 */ 2375 flags = PG_ZERO; 2376 if ((req & VM_ALLOC_NODUMP) != 0) 2377 flags |= PG_NODUMP; 2378 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2379 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2380 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2381 else if ((req & VM_ALLOC_SBUSY) != 0) 2382 busy_lock = VPB_SHARERS_WORD(1); 2383 else 2384 busy_lock = VPB_UNBUSIED; 2385 if ((req & VM_ALLOC_WIRED) != 0) 2386 vm_wire_add(npages); 2387 if (object->memattr != VM_MEMATTR_DEFAULT && 2388 memattr == VM_MEMATTR_DEFAULT) 2389 memattr = object->memattr; 2390 for (m = m_ret; m < &m_ret[npages]; m++) { 2391 vm_page_dequeue(m); 2392 vm_page_alloc_check(m); 2393 m->a.flags = 0; 2394 m->flags = (m->flags | PG_NODUMP) & flags; 2395 m->busy_lock = busy_lock; 2396 if ((req & VM_ALLOC_WIRED) != 0) 2397 m->ref_count = 1; 2398 m->a.act_count = 0; 2399 m->oflags = oflags; 2400 m->pool = VM_FREEPOOL_DEFAULT; 2401 if (vm_page_iter_insert(m, object, pindex, &pages)) { 2402 if ((req & VM_ALLOC_WIRED) != 0) 2403 vm_wire_sub(npages); 2404 KASSERT(m->object == NULL, 2405 ("page %p has object", m)); 2406 mpred = m; 2407 for (m = m_ret; m < &m_ret[npages]; m++) { 2408 if (m <= mpred && 2409 (req & VM_ALLOC_WIRED) != 0) 2410 m->ref_count = 0; 2411 m->oflags = VPO_UNMANAGED; 2412 m->busy_lock = VPB_UNBUSIED; 2413 /* Don't change PG_ZERO. */ 2414 vm_page_free_toq(m); 2415 } 2416 if (req & VM_ALLOC_WAITFAIL) { 2417 VM_OBJECT_WUNLOCK(object); 2418 vm_radix_wait(); 2419 VM_OBJECT_WLOCK(object); 2420 } 2421 return (NULL); 2422 } 2423 if (memattr != VM_MEMATTR_DEFAULT) 2424 pmap_page_set_memattr(m, memattr); 2425 pindex++; 2426 } 2427 return (m_ret); 2428 } 2429 2430 /* 2431 * Allocate a physical page that is not intended to be inserted into a VM 2432 * object. 2433 */ 2434 vm_page_t 2435 vm_page_alloc_noobj_domain(int domain, int req) 2436 { 2437 struct vm_domain *vmd; 2438 vm_page_t m; 2439 int flags; 2440 2441 #define VPAN_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2442 VM_ALLOC_NOFREE | VM_ALLOC_WAITOK) 2443 KASSERT((req & ~VPAN_FLAGS) == 0, 2444 ("invalid request %#x", req)); 2445 2446 flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) | 2447 ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0); 2448 vmd = VM_DOMAIN(domain); 2449 again: 2450 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2451 m = vm_page_alloc_nofree_domain(domain, req); 2452 if (m != NULL) 2453 goto found; 2454 } 2455 2456 if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2457 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2458 M_NOWAIT | M_NOVM); 2459 if (m != NULL) { 2460 flags |= PG_PCPU_CACHE; 2461 goto found; 2462 } 2463 } 2464 2465 if (vm_domain_allocate(vmd, req, 1)) { 2466 vm_domain_free_lock(vmd); 2467 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2468 vm_domain_free_unlock(vmd); 2469 if (m == NULL) { 2470 vm_domain_freecnt_inc(vmd, 1); 2471 #if VM_NRESERVLEVEL > 0 2472 if (vm_reserv_reclaim_inactive(domain)) 2473 goto again; 2474 #endif 2475 } 2476 } 2477 if (m == NULL) { 2478 if (!vm_domain_alloc_fail(vmd, NULL, req)) 2479 return (NULL); 2480 goto again; 2481 } 2482 2483 found: 2484 /* 2485 * If the page comes from the free page cache, then it might still 2486 * have a pending deferred dequeue. Specifically, when the page is 2487 * imported from a different pool by vm_phys_alloc_npages(), the 2488 * second, third, etc. pages in a non-zero order set could have 2489 * pending deferred dequeues. 2490 */ 2491 vm_page_dequeue(m); 2492 vm_page_alloc_check(m); 2493 2494 /* 2495 * Consumers should not rely on a useful default pindex value. 2496 */ 2497 m->pindex = 0xdeadc0dedeadc0de; 2498 m->flags = (m->flags & PG_ZERO) | flags; 2499 m->a.flags = 0; 2500 m->oflags = VPO_UNMANAGED; 2501 m->pool = VM_FREEPOOL_DIRECT; 2502 m->busy_lock = VPB_UNBUSIED; 2503 if ((req & VM_ALLOC_WIRED) != 0) { 2504 vm_wire_add(1); 2505 m->ref_count = 1; 2506 } 2507 2508 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2509 pmap_zero_page(m); 2510 2511 return (m); 2512 } 2513 2514 #if VM_NRESERVLEVEL > 1 2515 #define VM_NOFREE_IMPORT_ORDER (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER) 2516 #elif VM_NRESERVLEVEL > 0 2517 #define VM_NOFREE_IMPORT_ORDER VM_LEVEL_0_ORDER 2518 #else 2519 #define VM_NOFREE_IMPORT_ORDER 8 2520 #endif 2521 2522 /* 2523 * Allocate a single NOFREE page. 2524 * 2525 * This routine hands out NOFREE pages from higher-order 2526 * physical memory blocks in order to reduce memory fragmentation. 2527 * When a NOFREE for a given domain chunk is used up, 2528 * the routine will try to fetch a new one from the freelists 2529 * and discard the old one. 2530 */ 2531 static vm_page_t __noinline 2532 vm_page_alloc_nofree_domain(int domain, int req) 2533 { 2534 vm_page_t m; 2535 struct vm_domain *vmd; 2536 2537 KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req)); 2538 2539 vmd = VM_DOMAIN(domain); 2540 vm_domain_free_lock(vmd); 2541 if (TAILQ_EMPTY(&vmd->vmd_nofreeq)) { 2542 int count; 2543 2544 count = 1 << VM_NOFREE_IMPORT_ORDER; 2545 if (!vm_domain_allocate(vmd, req, count)) { 2546 vm_domain_free_unlock(vmd); 2547 return (NULL); 2548 } 2549 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 2550 VM_NOFREE_IMPORT_ORDER); 2551 if (m == NULL) { 2552 vm_domain_freecnt_inc(vmd, count); 2553 vm_domain_free_unlock(vmd); 2554 return (NULL); 2555 } 2556 m->ref_count = count - 1; 2557 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, plinks.q); 2558 atomic_add_long(&nofreeq_size, count); 2559 } 2560 m = TAILQ_FIRST(&vmd->vmd_nofreeq); 2561 TAILQ_REMOVE(&vmd->vmd_nofreeq, m, plinks.q); 2562 if (m->ref_count > 0) { 2563 vm_page_t m_next; 2564 2565 m_next = &m[1]; 2566 vm_page_dequeue(m_next); 2567 m_next->ref_count = m->ref_count - 1; 2568 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m_next, plinks.q); 2569 m->ref_count = 0; 2570 } 2571 vm_domain_free_unlock(vmd); 2572 atomic_add_long(&nofreeq_size, -1); 2573 VM_CNT_INC(v_nofree_count); 2574 2575 return (m); 2576 } 2577 2578 /* 2579 * Though a NOFREE page by definition should not be freed, we support putting 2580 * them aside for future NOFREE allocations. This enables code which allocates 2581 * NOFREE pages for some purpose but then encounters an error and releases 2582 * resources. 2583 */ 2584 static void __noinline 2585 vm_page_free_nofree(struct vm_domain *vmd, vm_page_t m) 2586 { 2587 VM_CNT_ADD(v_nofree_count, -1); 2588 atomic_add_long(&nofreeq_size, 1); 2589 vm_domain_free_lock(vmd); 2590 MPASS(m->ref_count == 0); 2591 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, plinks.q); 2592 vm_domain_free_unlock(vmd); 2593 } 2594 2595 vm_page_t 2596 vm_page_alloc_noobj(int req) 2597 { 2598 struct vm_domainset_iter di; 2599 vm_page_t m; 2600 int domain; 2601 2602 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 2603 return (NULL); 2604 2605 do { 2606 m = vm_page_alloc_noobj_domain(domain, req); 2607 if (m != NULL) 2608 break; 2609 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 2610 2611 return (m); 2612 } 2613 2614 vm_page_t 2615 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2616 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2617 vm_memattr_t memattr) 2618 { 2619 struct vm_domainset_iter di; 2620 vm_page_t m; 2621 int domain; 2622 2623 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 2624 return (NULL); 2625 2626 do { 2627 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2628 high, alignment, boundary, memattr); 2629 if (m != NULL) 2630 break; 2631 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 2632 2633 return (m); 2634 } 2635 2636 vm_page_t 2637 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2638 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2639 vm_memattr_t memattr) 2640 { 2641 vm_page_t m, m_ret; 2642 u_int flags; 2643 2644 #define VPANC_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2645 VM_ALLOC_NORECLAIM | VM_ALLOC_WAITOK) 2646 KASSERT((req & ~VPANC_FLAGS) == 0, 2647 ("invalid request %#x", req)); 2648 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2649 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2650 ("invalid request %#x", req)); 2651 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2652 2653 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2654 low, high, alignment, boundary)) == NULL) { 2655 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2656 return (NULL); 2657 } 2658 2659 /* 2660 * Initialize the pages. Only the PG_ZERO flag is inherited. 2661 */ 2662 flags = PG_ZERO; 2663 if ((req & VM_ALLOC_NODUMP) != 0) 2664 flags |= PG_NODUMP; 2665 if ((req & VM_ALLOC_WIRED) != 0) 2666 vm_wire_add(npages); 2667 for (m = m_ret; m < &m_ret[npages]; m++) { 2668 vm_page_dequeue(m); 2669 vm_page_alloc_check(m); 2670 2671 /* 2672 * Consumers should not rely on a useful default pindex value. 2673 */ 2674 m->pindex = 0xdeadc0dedeadc0de; 2675 m->a.flags = 0; 2676 m->flags = (m->flags | PG_NODUMP) & flags; 2677 m->busy_lock = VPB_UNBUSIED; 2678 if ((req & VM_ALLOC_WIRED) != 0) 2679 m->ref_count = 1; 2680 m->a.act_count = 0; 2681 m->oflags = VPO_UNMANAGED; 2682 m->pool = VM_FREEPOOL_DIRECT; 2683 2684 /* 2685 * Zero the page before updating any mappings since the page is 2686 * not yet shared with any devices which might require the 2687 * non-default memory attribute. pmap_page_set_memattr() 2688 * flushes data caches before returning. 2689 */ 2690 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2691 pmap_zero_page(m); 2692 if (memattr != VM_MEMATTR_DEFAULT) 2693 pmap_page_set_memattr(m, memattr); 2694 } 2695 return (m_ret); 2696 } 2697 2698 /* 2699 * Check a page that has been freshly dequeued from a freelist. 2700 */ 2701 static void 2702 vm_page_alloc_check(vm_page_t m) 2703 { 2704 2705 KASSERT(m->object == NULL, ("page %p has object", m)); 2706 KASSERT(m->a.queue == PQ_NONE && 2707 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2708 ("page %p has unexpected queue %d, flags %#x", 2709 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2710 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2711 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2712 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2713 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2714 ("page %p has unexpected memattr %d", 2715 m, pmap_page_get_memattr(m))); 2716 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2717 pmap_vm_page_alloc_check(m); 2718 } 2719 2720 static int 2721 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2722 { 2723 struct vm_domain *vmd; 2724 struct vm_pgcache *pgcache; 2725 int i; 2726 2727 pgcache = arg; 2728 vmd = VM_DOMAIN(pgcache->domain); 2729 2730 /* 2731 * The page daemon should avoid creating extra memory pressure since its 2732 * main purpose is to replenish the store of free pages. 2733 */ 2734 if (vmd->vmd_severeset || curproc == pageproc || 2735 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2736 return (0); 2737 domain = vmd->vmd_domain; 2738 vm_domain_free_lock(vmd); 2739 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2740 (vm_page_t *)store); 2741 vm_domain_free_unlock(vmd); 2742 if (cnt != i) 2743 vm_domain_freecnt_inc(vmd, cnt - i); 2744 2745 return (i); 2746 } 2747 2748 static void 2749 vm_page_zone_release(void *arg, void **store, int cnt) 2750 { 2751 struct vm_domain *vmd; 2752 struct vm_pgcache *pgcache; 2753 vm_page_t m; 2754 int i; 2755 2756 pgcache = arg; 2757 vmd = VM_DOMAIN(pgcache->domain); 2758 vm_domain_free_lock(vmd); 2759 for (i = 0; i < cnt; i++) { 2760 m = (vm_page_t)store[i]; 2761 vm_phys_free_pages(m, pgcache->pool, 0); 2762 } 2763 vm_domain_free_unlock(vmd); 2764 vm_domain_freecnt_inc(vmd, cnt); 2765 } 2766 2767 #define VPSC_ANY 0 /* No restrictions. */ 2768 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2769 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2770 2771 /* 2772 * vm_page_scan_contig: 2773 * 2774 * Scan vm_page_array[] between the specified entries "m_start" and 2775 * "m_end" for a run of contiguous physical pages that satisfy the 2776 * specified conditions, and return the lowest page in the run. The 2777 * specified "alignment" determines the alignment of the lowest physical 2778 * page in the run. If the specified "boundary" is non-zero, then the 2779 * run of physical pages cannot span a physical address that is a 2780 * multiple of "boundary". 2781 * 2782 * "m_end" is never dereferenced, so it need not point to a vm_page 2783 * structure within vm_page_array[]. 2784 * 2785 * "npages" must be greater than zero. "m_start" and "m_end" must not 2786 * span a hole (or discontiguity) in the physical address space. Both 2787 * "alignment" and "boundary" must be a power of two. 2788 */ 2789 static vm_page_t 2790 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2791 u_long alignment, vm_paddr_t boundary, int options) 2792 { 2793 vm_object_t object; 2794 vm_paddr_t pa; 2795 vm_page_t m, m_run; 2796 #if VM_NRESERVLEVEL > 0 2797 int level; 2798 #endif 2799 int m_inc, order, run_ext, run_len; 2800 2801 KASSERT(npages > 0, ("npages is 0")); 2802 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2803 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2804 m_run = NULL; 2805 run_len = 0; 2806 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2807 KASSERT((m->flags & PG_MARKER) == 0, 2808 ("page %p is PG_MARKER", m)); 2809 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2810 ("fictitious page %p has invalid ref count", m)); 2811 2812 /* 2813 * If the current page would be the start of a run, check its 2814 * physical address against the end, alignment, and boundary 2815 * conditions. If it doesn't satisfy these conditions, either 2816 * terminate the scan or advance to the next page that 2817 * satisfies the failed condition. 2818 */ 2819 if (run_len == 0) { 2820 KASSERT(m_run == NULL, ("m_run != NULL")); 2821 if (m + npages > m_end) 2822 break; 2823 pa = VM_PAGE_TO_PHYS(m); 2824 if (!vm_addr_align_ok(pa, alignment)) { 2825 m_inc = atop(roundup2(pa, alignment) - pa); 2826 continue; 2827 } 2828 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2829 m_inc = atop(roundup2(pa, boundary) - pa); 2830 continue; 2831 } 2832 } else 2833 KASSERT(m_run != NULL, ("m_run == NULL")); 2834 2835 retry: 2836 m_inc = 1; 2837 if (vm_page_wired(m)) 2838 run_ext = 0; 2839 #if VM_NRESERVLEVEL > 0 2840 else if ((level = vm_reserv_level(m)) >= 0 && 2841 (options & VPSC_NORESERV) != 0) { 2842 run_ext = 0; 2843 /* Advance to the end of the reservation. */ 2844 pa = VM_PAGE_TO_PHYS(m); 2845 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2846 pa); 2847 } 2848 #endif 2849 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2850 /* 2851 * The page is considered eligible for relocation if 2852 * and only if it could be laundered or reclaimed by 2853 * the page daemon. 2854 */ 2855 VM_OBJECT_RLOCK(object); 2856 if (object != m->object) { 2857 VM_OBJECT_RUNLOCK(object); 2858 goto retry; 2859 } 2860 /* Don't care: PG_NODUMP, PG_ZERO. */ 2861 if ((object->flags & OBJ_SWAP) == 0 && 2862 object->type != OBJT_VNODE) { 2863 run_ext = 0; 2864 #if VM_NRESERVLEVEL > 0 2865 } else if ((options & VPSC_NOSUPER) != 0 && 2866 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2867 run_ext = 0; 2868 /* Advance to the end of the superpage. */ 2869 pa = VM_PAGE_TO_PHYS(m); 2870 m_inc = atop(roundup2(pa + 1, 2871 vm_reserv_size(level)) - pa); 2872 #endif 2873 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2874 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2875 /* 2876 * The page is allocated but eligible for 2877 * relocation. Extend the current run by one 2878 * page. 2879 */ 2880 KASSERT(pmap_page_get_memattr(m) == 2881 VM_MEMATTR_DEFAULT, 2882 ("page %p has an unexpected memattr", m)); 2883 KASSERT((m->oflags & (VPO_SWAPINPROG | 2884 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2885 ("page %p has unexpected oflags", m)); 2886 /* Don't care: PGA_NOSYNC. */ 2887 run_ext = 1; 2888 } else 2889 run_ext = 0; 2890 VM_OBJECT_RUNLOCK(object); 2891 #if VM_NRESERVLEVEL > 0 2892 } else if (level >= 0) { 2893 /* 2894 * The page is reserved but not yet allocated. In 2895 * other words, it is still free. Extend the current 2896 * run by one page. 2897 */ 2898 run_ext = 1; 2899 #endif 2900 } else if ((order = m->order) < VM_NFREEORDER) { 2901 /* 2902 * The page is enqueued in the physical memory 2903 * allocator's free page queues. Moreover, it is the 2904 * first page in a power-of-two-sized run of 2905 * contiguous free pages. Add these pages to the end 2906 * of the current run, and jump ahead. 2907 */ 2908 run_ext = 1 << order; 2909 m_inc = 1 << order; 2910 } else { 2911 /* 2912 * Skip the page for one of the following reasons: (1) 2913 * It is enqueued in the physical memory allocator's 2914 * free page queues. However, it is not the first 2915 * page in a run of contiguous free pages. (This case 2916 * rarely occurs because the scan is performed in 2917 * ascending order.) (2) It is not reserved, and it is 2918 * transitioning from free to allocated. (Conversely, 2919 * the transition from allocated to free for managed 2920 * pages is blocked by the page busy lock.) (3) It is 2921 * allocated but not contained by an object and not 2922 * wired, e.g., allocated by Xen's balloon driver. 2923 */ 2924 run_ext = 0; 2925 } 2926 2927 /* 2928 * Extend or reset the current run of pages. 2929 */ 2930 if (run_ext > 0) { 2931 if (run_len == 0) 2932 m_run = m; 2933 run_len += run_ext; 2934 } else { 2935 if (run_len > 0) { 2936 m_run = NULL; 2937 run_len = 0; 2938 } 2939 } 2940 } 2941 if (run_len >= npages) 2942 return (m_run); 2943 return (NULL); 2944 } 2945 2946 /* 2947 * vm_page_reclaim_run: 2948 * 2949 * Try to relocate each of the allocated virtual pages within the 2950 * specified run of physical pages to a new physical address. Free the 2951 * physical pages underlying the relocated virtual pages. A virtual page 2952 * is relocatable if and only if it could be laundered or reclaimed by 2953 * the page daemon. Whenever possible, a virtual page is relocated to a 2954 * physical address above "high". 2955 * 2956 * Returns 0 if every physical page within the run was already free or 2957 * just freed by a successful relocation. Otherwise, returns a non-zero 2958 * value indicating why the last attempt to relocate a virtual page was 2959 * unsuccessful. 2960 * 2961 * "req_class" must be an allocation class. 2962 */ 2963 static int 2964 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2965 vm_paddr_t high) 2966 { 2967 struct vm_domain *vmd; 2968 struct spglist free; 2969 vm_object_t object; 2970 vm_paddr_t pa; 2971 vm_page_t m, m_end, m_new; 2972 int error, order, req; 2973 2974 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2975 ("req_class is not an allocation class")); 2976 SLIST_INIT(&free); 2977 error = 0; 2978 m = m_run; 2979 m_end = m_run + npages; 2980 for (; error == 0 && m < m_end; m++) { 2981 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2982 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2983 2984 /* 2985 * Racily check for wirings. Races are handled once the object 2986 * lock is held and the page is unmapped. 2987 */ 2988 if (vm_page_wired(m)) 2989 error = EBUSY; 2990 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2991 /* 2992 * The page is relocated if and only if it could be 2993 * laundered or reclaimed by the page daemon. 2994 */ 2995 VM_OBJECT_WLOCK(object); 2996 /* Don't care: PG_NODUMP, PG_ZERO. */ 2997 if (m->object != object || 2998 ((object->flags & OBJ_SWAP) == 0 && 2999 object->type != OBJT_VNODE)) 3000 error = EINVAL; 3001 else if (object->memattr != VM_MEMATTR_DEFAULT) 3002 error = EINVAL; 3003 else if (vm_page_queue(m) != PQ_NONE && 3004 vm_page_tryxbusy(m) != 0) { 3005 if (vm_page_wired(m)) { 3006 vm_page_xunbusy(m); 3007 error = EBUSY; 3008 goto unlock; 3009 } 3010 KASSERT(pmap_page_get_memattr(m) == 3011 VM_MEMATTR_DEFAULT, 3012 ("page %p has an unexpected memattr", m)); 3013 KASSERT(m->oflags == 0, 3014 ("page %p has unexpected oflags", m)); 3015 /* Don't care: PGA_NOSYNC. */ 3016 if (!vm_page_none_valid(m)) { 3017 /* 3018 * First, try to allocate a new page 3019 * that is above "high". Failing 3020 * that, try to allocate a new page 3021 * that is below "m_run". Allocate 3022 * the new page between the end of 3023 * "m_run" and "high" only as a last 3024 * resort. 3025 */ 3026 req = req_class; 3027 if ((m->flags & PG_NODUMP) != 0) 3028 req |= VM_ALLOC_NODUMP; 3029 if (trunc_page(high) != 3030 ~(vm_paddr_t)PAGE_MASK) { 3031 m_new = 3032 vm_page_alloc_noobj_contig( 3033 req, 1, round_page(high), 3034 ~(vm_paddr_t)0, PAGE_SIZE, 3035 0, VM_MEMATTR_DEFAULT); 3036 } else 3037 m_new = NULL; 3038 if (m_new == NULL) { 3039 pa = VM_PAGE_TO_PHYS(m_run); 3040 m_new = 3041 vm_page_alloc_noobj_contig( 3042 req, 1, 0, pa - 1, 3043 PAGE_SIZE, 0, 3044 VM_MEMATTR_DEFAULT); 3045 } 3046 if (m_new == NULL) { 3047 pa += ptoa(npages); 3048 m_new = 3049 vm_page_alloc_noobj_contig( 3050 req, 1, pa, high, PAGE_SIZE, 3051 0, VM_MEMATTR_DEFAULT); 3052 } 3053 if (m_new == NULL) { 3054 vm_page_xunbusy(m); 3055 error = ENOMEM; 3056 goto unlock; 3057 } 3058 3059 /* 3060 * Unmap the page and check for new 3061 * wirings that may have been acquired 3062 * through a pmap lookup. 3063 */ 3064 if (object->ref_count != 0 && 3065 !vm_page_try_remove_all(m)) { 3066 vm_page_xunbusy(m); 3067 vm_page_free(m_new); 3068 error = EBUSY; 3069 goto unlock; 3070 } 3071 3072 /* 3073 * Replace "m" with the new page. For 3074 * vm_page_replace(), "m" must be busy 3075 * and dequeued. Finally, change "m" 3076 * as if vm_page_free() was called. 3077 */ 3078 m_new->a.flags = m->a.flags & 3079 ~PGA_QUEUE_STATE_MASK; 3080 KASSERT(m_new->oflags == VPO_UNMANAGED, 3081 ("page %p is managed", m_new)); 3082 m_new->oflags = 0; 3083 pmap_copy_page(m, m_new); 3084 m_new->valid = m->valid; 3085 m_new->dirty = m->dirty; 3086 m->flags &= ~PG_ZERO; 3087 vm_page_dequeue(m); 3088 if (vm_page_replace_hold(m_new, object, 3089 m->pindex, m) && 3090 vm_page_free_prep(m)) 3091 SLIST_INSERT_HEAD(&free, m, 3092 plinks.s.ss); 3093 3094 /* 3095 * The new page must be deactivated 3096 * before the object is unlocked. 3097 */ 3098 vm_page_deactivate(m_new); 3099 } else { 3100 m->flags &= ~PG_ZERO; 3101 vm_page_dequeue(m); 3102 if (vm_page_free_prep(m)) 3103 SLIST_INSERT_HEAD(&free, m, 3104 plinks.s.ss); 3105 KASSERT(m->dirty == 0, 3106 ("page %p is dirty", m)); 3107 } 3108 } else 3109 error = EBUSY; 3110 unlock: 3111 VM_OBJECT_WUNLOCK(object); 3112 } else { 3113 MPASS(vm_page_domain(m) == domain); 3114 vmd = VM_DOMAIN(domain); 3115 vm_domain_free_lock(vmd); 3116 order = m->order; 3117 if (order < VM_NFREEORDER) { 3118 /* 3119 * The page is enqueued in the physical memory 3120 * allocator's free page queues. Moreover, it 3121 * is the first page in a power-of-two-sized 3122 * run of contiguous free pages. Jump ahead 3123 * to the last page within that run, and 3124 * continue from there. 3125 */ 3126 m += (1 << order) - 1; 3127 } 3128 #if VM_NRESERVLEVEL > 0 3129 else if (vm_reserv_is_page_free(m)) 3130 order = 0; 3131 #endif 3132 vm_domain_free_unlock(vmd); 3133 if (order == VM_NFREEORDER) 3134 error = EINVAL; 3135 } 3136 } 3137 if ((m = SLIST_FIRST(&free)) != NULL) { 3138 int cnt; 3139 3140 vmd = VM_DOMAIN(domain); 3141 cnt = 0; 3142 vm_domain_free_lock(vmd); 3143 do { 3144 MPASS(vm_page_domain(m) == domain); 3145 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3146 vm_phys_free_pages(m, m->pool, 0); 3147 cnt++; 3148 } while ((m = SLIST_FIRST(&free)) != NULL); 3149 vm_domain_free_unlock(vmd); 3150 vm_domain_freecnt_inc(vmd, cnt); 3151 } 3152 return (error); 3153 } 3154 3155 #define NRUNS 16 3156 3157 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 3158 3159 #define MIN_RECLAIM 8 3160 3161 /* 3162 * vm_page_reclaim_contig: 3163 * 3164 * Reclaim allocated, contiguous physical memory satisfying the specified 3165 * conditions by relocating the virtual pages using that physical memory. 3166 * Returns 0 if reclamation is successful, ERANGE if the specified domain 3167 * can't possibly satisfy the reclamation request, or ENOMEM if not 3168 * currently able to reclaim the requested number of pages. Since 3169 * relocation requires the allocation of physical pages, reclamation may 3170 * fail with ENOMEM due to a shortage of free pages. When reclamation 3171 * fails in this manner, callers are expected to perform vm_wait() before 3172 * retrying a failed allocation operation, e.g., vm_page_alloc_contig(). 3173 * 3174 * The caller must always specify an allocation class through "req". 3175 * 3176 * allocation classes: 3177 * VM_ALLOC_NORMAL normal process request 3178 * VM_ALLOC_SYSTEM system *really* needs a page 3179 * VM_ALLOC_INTERRUPT interrupt time request 3180 * 3181 * The optional allocation flags are ignored. 3182 * 3183 * "npages" must be greater than zero. Both "alignment" and "boundary" 3184 * must be a power of two. 3185 */ 3186 int 3187 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3188 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3189 int desired_runs) 3190 { 3191 struct vm_domain *vmd; 3192 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; 3193 u_long count, minalign, reclaimed; 3194 int error, i, min_reclaim, nruns, options, req_class; 3195 int segind, start_segind; 3196 int ret; 3197 3198 KASSERT(npages > 0, ("npages is 0")); 3199 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3200 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3201 3202 ret = ENOMEM; 3203 3204 /* 3205 * If the caller wants to reclaim multiple runs, try to allocate 3206 * space to store the runs. If that fails, fall back to the old 3207 * behavior of just reclaiming MIN_RECLAIM pages. 3208 */ 3209 if (desired_runs > 1) 3210 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3211 M_TEMP, M_NOWAIT); 3212 else 3213 m_runs = NULL; 3214 3215 if (m_runs == NULL) { 3216 m_runs = _m_runs; 3217 nruns = NRUNS; 3218 } else { 3219 nruns = NRUNS + desired_runs - 1; 3220 } 3221 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3222 3223 /* 3224 * The caller will attempt an allocation after some runs have been 3225 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3226 * of vm_phys_alloc_contig(), round up the requested length to the next 3227 * power of two or maximum chunk size, and ensure that each run is 3228 * suitably aligned. 3229 */ 3230 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3231 npages = roundup2(npages, minalign); 3232 if (alignment < ptoa(minalign)) 3233 alignment = ptoa(minalign); 3234 3235 /* 3236 * The page daemon is allowed to dig deeper into the free page list. 3237 */ 3238 req_class = req & VM_ALLOC_CLASS_MASK; 3239 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3240 req_class = VM_ALLOC_SYSTEM; 3241 3242 start_segind = vm_phys_lookup_segind(low); 3243 3244 /* 3245 * Return if the number of free pages cannot satisfy the requested 3246 * allocation. 3247 */ 3248 vmd = VM_DOMAIN(domain); 3249 count = vmd->vmd_free_count; 3250 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3251 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3252 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3253 goto done; 3254 3255 /* 3256 * Scan up to three times, relaxing the restrictions ("options") on 3257 * the reclamation of reservations and superpages each time. 3258 */ 3259 for (options = VPSC_NORESERV;;) { 3260 bool phys_range_exists = false; 3261 3262 /* 3263 * Find the highest runs that satisfy the given constraints 3264 * and restrictions, and record them in "m_runs". 3265 */ 3266 count = 0; 3267 segind = start_segind; 3268 while ((segind = vm_phys_find_range(bounds, segind, domain, 3269 npages, low, high)) != -1) { 3270 phys_range_exists = true; 3271 while ((m_run = vm_page_scan_contig(npages, bounds[0], 3272 bounds[1], alignment, boundary, options))) { 3273 bounds[0] = m_run + npages; 3274 m_runs[RUN_INDEX(count, nruns)] = m_run; 3275 count++; 3276 } 3277 segind++; 3278 } 3279 3280 if (!phys_range_exists) { 3281 ret = ERANGE; 3282 goto done; 3283 } 3284 3285 /* 3286 * Reclaim the highest runs in LIFO (descending) order until 3287 * the number of reclaimed pages, "reclaimed", is at least 3288 * "min_reclaim". Reset "reclaimed" each time because each 3289 * reclamation is idempotent, and runs will (likely) recur 3290 * from one scan to the next as restrictions are relaxed. 3291 */ 3292 reclaimed = 0; 3293 for (i = 0; count > 0 && i < nruns; i++) { 3294 count--; 3295 m_run = m_runs[RUN_INDEX(count, nruns)]; 3296 error = vm_page_reclaim_run(req_class, domain, npages, 3297 m_run, high); 3298 if (error == 0) { 3299 reclaimed += npages; 3300 if (reclaimed >= min_reclaim) { 3301 ret = 0; 3302 goto done; 3303 } 3304 } 3305 } 3306 3307 /* 3308 * Either relax the restrictions on the next scan or return if 3309 * the last scan had no restrictions. 3310 */ 3311 if (options == VPSC_NORESERV) 3312 options = VPSC_NOSUPER; 3313 else if (options == VPSC_NOSUPER) 3314 options = VPSC_ANY; 3315 else if (options == VPSC_ANY) { 3316 if (reclaimed != 0) 3317 ret = 0; 3318 goto done; 3319 } 3320 } 3321 done: 3322 if (m_runs != _m_runs) 3323 free(m_runs, M_TEMP); 3324 return (ret); 3325 } 3326 3327 int 3328 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3329 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3330 { 3331 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, 3332 high, alignment, boundary, 1)); 3333 } 3334 3335 int 3336 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3337 u_long alignment, vm_paddr_t boundary) 3338 { 3339 struct vm_domainset_iter di; 3340 int domain, ret, status; 3341 3342 ret = ERANGE; 3343 3344 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 3345 return (ret); 3346 3347 do { 3348 status = vm_page_reclaim_contig_domain(domain, req, npages, low, 3349 high, alignment, boundary); 3350 if (status == 0) 3351 return (0); 3352 else if (status == ERANGE) 3353 vm_domainset_iter_ignore(&di, domain); 3354 else { 3355 KASSERT(status == ENOMEM, ("Unrecognized error %d " 3356 "from vm_page_reclaim_contig_domain()", status)); 3357 ret = ENOMEM; 3358 } 3359 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 3360 3361 return (ret); 3362 } 3363 3364 /* 3365 * Set the domain in the appropriate page level domainset. 3366 */ 3367 void 3368 vm_domain_set(struct vm_domain *vmd) 3369 { 3370 3371 mtx_lock(&vm_domainset_lock); 3372 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3373 vmd->vmd_minset = 1; 3374 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3375 } 3376 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3377 vmd->vmd_severeset = 1; 3378 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3379 } 3380 mtx_unlock(&vm_domainset_lock); 3381 } 3382 3383 /* 3384 * Clear the domain from the appropriate page level domainset. 3385 */ 3386 void 3387 vm_domain_clear(struct vm_domain *vmd) 3388 { 3389 3390 mtx_lock(&vm_domainset_lock); 3391 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3392 vmd->vmd_minset = 0; 3393 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3394 if (vm_min_waiters != 0) { 3395 vm_min_waiters = 0; 3396 wakeup(&vm_min_domains); 3397 } 3398 } 3399 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3400 vmd->vmd_severeset = 0; 3401 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3402 if (vm_severe_waiters != 0) { 3403 vm_severe_waiters = 0; 3404 wakeup(&vm_severe_domains); 3405 } 3406 } 3407 3408 /* 3409 * If pageout daemon needs pages, then tell it that there are 3410 * some free. 3411 */ 3412 if (vmd->vmd_pageout_pages_needed && 3413 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3414 wakeup(&vmd->vmd_pageout_pages_needed); 3415 vmd->vmd_pageout_pages_needed = 0; 3416 } 3417 3418 /* See comments in vm_wait_doms(). */ 3419 if (vm_pageproc_waiters) { 3420 vm_pageproc_waiters = 0; 3421 wakeup(&vm_pageproc_waiters); 3422 } 3423 mtx_unlock(&vm_domainset_lock); 3424 } 3425 3426 /* 3427 * Wait for free pages to exceed the min threshold globally. 3428 */ 3429 void 3430 vm_wait_min(void) 3431 { 3432 3433 mtx_lock(&vm_domainset_lock); 3434 while (vm_page_count_min()) { 3435 vm_min_waiters++; 3436 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3437 } 3438 mtx_unlock(&vm_domainset_lock); 3439 } 3440 3441 /* 3442 * Wait for free pages to exceed the severe threshold globally. 3443 */ 3444 void 3445 vm_wait_severe(void) 3446 { 3447 3448 mtx_lock(&vm_domainset_lock); 3449 while (vm_page_count_severe()) { 3450 vm_severe_waiters++; 3451 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3452 "vmwait", 0); 3453 } 3454 mtx_unlock(&vm_domainset_lock); 3455 } 3456 3457 u_int 3458 vm_wait_count(void) 3459 { 3460 3461 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3462 } 3463 3464 int 3465 vm_wait_doms(const domainset_t *wdoms, int mflags) 3466 { 3467 int error; 3468 3469 error = 0; 3470 3471 /* 3472 * We use racey wakeup synchronization to avoid expensive global 3473 * locking for the pageproc when sleeping with a non-specific vm_wait. 3474 * To handle this, we only sleep for one tick in this instance. It 3475 * is expected that most allocations for the pageproc will come from 3476 * kmem or vm_page_grab* which will use the more specific and 3477 * race-free vm_wait_domain(). 3478 */ 3479 if (curproc == pageproc) { 3480 mtx_lock(&vm_domainset_lock); 3481 vm_pageproc_waiters++; 3482 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3483 PVM | PDROP | mflags, "pageprocwait", 1); 3484 } else { 3485 /* 3486 * XXX Ideally we would wait only until the allocation could 3487 * be satisfied. This condition can cause new allocators to 3488 * consume all freed pages while old allocators wait. 3489 */ 3490 mtx_lock(&vm_domainset_lock); 3491 if (vm_page_count_min_set(wdoms)) { 3492 if (pageproc == NULL) 3493 panic("vm_wait in early boot"); 3494 vm_min_waiters++; 3495 error = msleep(&vm_min_domains, &vm_domainset_lock, 3496 PVM | PDROP | mflags, "vmwait", 0); 3497 } else 3498 mtx_unlock(&vm_domainset_lock); 3499 } 3500 return (error); 3501 } 3502 3503 /* 3504 * vm_wait_domain: 3505 * 3506 * Sleep until free pages are available for allocation. 3507 * - Called in various places after failed memory allocations. 3508 */ 3509 void 3510 vm_wait_domain(int domain) 3511 { 3512 struct vm_domain *vmd; 3513 domainset_t wdom; 3514 3515 vmd = VM_DOMAIN(domain); 3516 vm_domain_free_assert_unlocked(vmd); 3517 3518 if (curproc == pageproc) { 3519 mtx_lock(&vm_domainset_lock); 3520 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3521 vmd->vmd_pageout_pages_needed = 1; 3522 msleep(&vmd->vmd_pageout_pages_needed, 3523 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3524 } else 3525 mtx_unlock(&vm_domainset_lock); 3526 } else { 3527 DOMAINSET_ZERO(&wdom); 3528 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3529 vm_wait_doms(&wdom, 0); 3530 } 3531 } 3532 3533 static int 3534 vm_wait_flags(vm_object_t obj, int mflags) 3535 { 3536 struct domainset *d; 3537 3538 d = NULL; 3539 3540 /* 3541 * Carefully fetch pointers only once: the struct domainset 3542 * itself is ummutable but the pointer might change. 3543 */ 3544 if (obj != NULL) 3545 d = obj->domain.dr_policy; 3546 if (d == NULL) 3547 d = curthread->td_domain.dr_policy; 3548 3549 return (vm_wait_doms(&d->ds_mask, mflags)); 3550 } 3551 3552 /* 3553 * vm_wait: 3554 * 3555 * Sleep until free pages are available for allocation in the 3556 * affinity domains of the obj. If obj is NULL, the domain set 3557 * for the calling thread is used. 3558 * Called in various places after failed memory allocations. 3559 */ 3560 void 3561 vm_wait(vm_object_t obj) 3562 { 3563 (void)vm_wait_flags(obj, 0); 3564 } 3565 3566 int 3567 vm_wait_intr(vm_object_t obj) 3568 { 3569 return (vm_wait_flags(obj, PCATCH)); 3570 } 3571 3572 /* 3573 * vm_domain_alloc_fail: 3574 * 3575 * Called when a page allocation function fails. Informs the 3576 * pagedaemon and performs the requested wait. Requires the 3577 * domain_free and object lock on entry. Returns with the 3578 * object lock held and free lock released. Returns an error when 3579 * retry is necessary. 3580 * 3581 */ 3582 static int 3583 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3584 { 3585 3586 vm_domain_free_assert_unlocked(vmd); 3587 3588 atomic_add_int(&vmd->vmd_pageout_deficit, 3589 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3590 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3591 if (object != NULL) 3592 VM_OBJECT_WUNLOCK(object); 3593 vm_wait_domain(vmd->vmd_domain); 3594 if (object != NULL) 3595 VM_OBJECT_WLOCK(object); 3596 if (req & VM_ALLOC_WAITOK) 3597 return (EAGAIN); 3598 } 3599 3600 return (0); 3601 } 3602 3603 /* 3604 * vm_waitpfault: 3605 * 3606 * Sleep until free pages are available for allocation. 3607 * - Called only in vm_fault so that processes page faulting 3608 * can be easily tracked. 3609 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3610 * processes will be able to grab memory first. Do not change 3611 * this balance without careful testing first. 3612 */ 3613 void 3614 vm_waitpfault(struct domainset *dset, int timo) 3615 { 3616 3617 /* 3618 * XXX Ideally we would wait only until the allocation could 3619 * be satisfied. This condition can cause new allocators to 3620 * consume all freed pages while old allocators wait. 3621 */ 3622 mtx_lock(&vm_domainset_lock); 3623 if (vm_page_count_min_set(&dset->ds_mask)) { 3624 vm_min_waiters++; 3625 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3626 "pfault", timo); 3627 } else 3628 mtx_unlock(&vm_domainset_lock); 3629 } 3630 3631 static struct vm_pagequeue * 3632 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3633 { 3634 3635 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3636 } 3637 3638 #ifdef INVARIANTS 3639 static struct vm_pagequeue * 3640 vm_page_pagequeue(vm_page_t m) 3641 { 3642 3643 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3644 } 3645 #endif 3646 3647 static __always_inline bool 3648 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, 3649 vm_page_astate_t new) 3650 { 3651 vm_page_astate_t tmp; 3652 3653 tmp = *old; 3654 do { 3655 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3656 return (true); 3657 counter_u64_add(pqstate_commit_retries, 1); 3658 } while (old->_bits == tmp._bits); 3659 3660 return (false); 3661 } 3662 3663 /* 3664 * Do the work of committing a queue state update that moves the page out of 3665 * its current queue. 3666 */ 3667 static bool 3668 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3669 vm_page_astate_t *old, vm_page_astate_t new) 3670 { 3671 vm_page_t next; 3672 3673 vm_pagequeue_assert_locked(pq); 3674 KASSERT(vm_page_pagequeue(m) == pq, 3675 ("%s: queue %p does not match page %p", __func__, pq, m)); 3676 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3677 ("%s: invalid queue indices %d %d", 3678 __func__, old->queue, new.queue)); 3679 3680 /* 3681 * Once the queue index of the page changes there is nothing 3682 * synchronizing with further updates to the page's physical 3683 * queue state. Therefore we must speculatively remove the page 3684 * from the queue now and be prepared to roll back if the queue 3685 * state update fails. If the page is not physically enqueued then 3686 * we just update its queue index. 3687 */ 3688 if ((old->flags & PGA_ENQUEUED) != 0) { 3689 new.flags &= ~PGA_ENQUEUED; 3690 next = TAILQ_NEXT(m, plinks.q); 3691 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3692 vm_pagequeue_cnt_dec(pq); 3693 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3694 if (next == NULL) 3695 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3696 else 3697 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3698 vm_pagequeue_cnt_inc(pq); 3699 return (false); 3700 } else { 3701 return (true); 3702 } 3703 } else { 3704 return (vm_page_pqstate_fcmpset(m, old, new)); 3705 } 3706 } 3707 3708 static bool 3709 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3710 vm_page_astate_t new) 3711 { 3712 struct vm_pagequeue *pq; 3713 vm_page_astate_t as; 3714 bool ret; 3715 3716 pq = _vm_page_pagequeue(m, old->queue); 3717 3718 /* 3719 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3720 * corresponding page queue lock is held. 3721 */ 3722 vm_pagequeue_lock(pq); 3723 as = vm_page_astate_load(m); 3724 if (__predict_false(as._bits != old->_bits)) { 3725 *old = as; 3726 ret = false; 3727 } else { 3728 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3729 } 3730 vm_pagequeue_unlock(pq); 3731 return (ret); 3732 } 3733 3734 /* 3735 * Commit a queue state update that enqueues or requeues a page. 3736 */ 3737 static bool 3738 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3739 vm_page_astate_t *old, vm_page_astate_t new) 3740 { 3741 struct vm_domain *vmd; 3742 3743 vm_pagequeue_assert_locked(pq); 3744 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3745 ("%s: invalid queue indices %d %d", 3746 __func__, old->queue, new.queue)); 3747 3748 new.flags |= PGA_ENQUEUED; 3749 if (!vm_page_pqstate_fcmpset(m, old, new)) 3750 return (false); 3751 3752 if ((old->flags & PGA_ENQUEUED) != 0) 3753 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3754 else 3755 vm_pagequeue_cnt_inc(pq); 3756 3757 /* 3758 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3759 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3760 * applied, even if it was set first. 3761 */ 3762 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3763 vmd = vm_pagequeue_domain(m); 3764 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3765 ("%s: invalid page queue for page %p", __func__, m)); 3766 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3767 } else { 3768 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3769 } 3770 return (true); 3771 } 3772 3773 /* 3774 * Commit a queue state update that encodes a request for a deferred queue 3775 * operation. 3776 */ 3777 static bool 3778 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3779 vm_page_astate_t new) 3780 { 3781 3782 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3783 ("%s: invalid state, queue %d flags %x", 3784 __func__, new.queue, new.flags)); 3785 3786 if (old->_bits != new._bits && 3787 !vm_page_pqstate_fcmpset(m, old, new)) 3788 return (false); 3789 vm_page_pqbatch_submit(m, new.queue); 3790 return (true); 3791 } 3792 3793 /* 3794 * A generic queue state update function. This handles more cases than the 3795 * specialized functions above. 3796 */ 3797 bool 3798 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3799 { 3800 3801 if (old->_bits == new._bits) 3802 return (true); 3803 3804 if (old->queue != PQ_NONE && new.queue != old->queue) { 3805 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3806 return (false); 3807 if (new.queue != PQ_NONE) 3808 vm_page_pqbatch_submit(m, new.queue); 3809 } else { 3810 if (!vm_page_pqstate_fcmpset(m, old, new)) 3811 return (false); 3812 if (new.queue != PQ_NONE && 3813 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3814 vm_page_pqbatch_submit(m, new.queue); 3815 } 3816 return (true); 3817 } 3818 3819 /* 3820 * Apply deferred queue state updates to a page. 3821 */ 3822 static inline void 3823 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3824 { 3825 vm_page_astate_t new, old; 3826 3827 CRITICAL_ASSERT(curthread); 3828 vm_pagequeue_assert_locked(pq); 3829 KASSERT(queue < PQ_COUNT, 3830 ("%s: invalid queue index %d", __func__, queue)); 3831 KASSERT(pq == _vm_page_pagequeue(m, queue), 3832 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3833 3834 for (old = vm_page_astate_load(m);;) { 3835 if (__predict_false(old.queue != queue || 3836 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3837 counter_u64_add(queue_nops, 1); 3838 break; 3839 } 3840 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3841 ("%s: page %p is unmanaged", __func__, m)); 3842 3843 new = old; 3844 if ((old.flags & PGA_DEQUEUE) != 0) { 3845 new.flags &= ~PGA_QUEUE_OP_MASK; 3846 new.queue = PQ_NONE; 3847 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3848 m, &old, new))) { 3849 counter_u64_add(queue_ops, 1); 3850 break; 3851 } 3852 } else { 3853 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3854 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3855 m, &old, new))) { 3856 counter_u64_add(queue_ops, 1); 3857 break; 3858 } 3859 } 3860 } 3861 } 3862 3863 static void 3864 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3865 uint8_t queue) 3866 { 3867 int i; 3868 3869 for (i = 0; i < bq->bq_cnt; i++) 3870 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3871 vm_batchqueue_init(bq); 3872 } 3873 3874 /* 3875 * vm_page_pqbatch_submit: [ internal use only ] 3876 * 3877 * Enqueue a page in the specified page queue's batched work queue. 3878 * The caller must have encoded the requested operation in the page 3879 * structure's a.flags field. 3880 */ 3881 void 3882 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3883 { 3884 struct vm_batchqueue *bq; 3885 struct vm_pagequeue *pq; 3886 int domain, slots_remaining; 3887 3888 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3889 3890 domain = vm_page_domain(m); 3891 critical_enter(); 3892 bq = DPCPU_PTR(pqbatch[domain][queue]); 3893 slots_remaining = vm_batchqueue_insert(bq, m); 3894 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 3895 /* keep building the bq */ 3896 critical_exit(); 3897 return; 3898 } else if (slots_remaining > 0 ) { 3899 /* Try to process the bq if we can get the lock */ 3900 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3901 if (vm_pagequeue_trylock(pq)) { 3902 vm_pqbatch_process(pq, bq, queue); 3903 vm_pagequeue_unlock(pq); 3904 } 3905 critical_exit(); 3906 return; 3907 } 3908 critical_exit(); 3909 3910 /* if we make it here, the bq is full so wait for the lock */ 3911 3912 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3913 vm_pagequeue_lock(pq); 3914 critical_enter(); 3915 bq = DPCPU_PTR(pqbatch[domain][queue]); 3916 vm_pqbatch_process(pq, bq, queue); 3917 vm_pqbatch_process_page(pq, m, queue); 3918 vm_pagequeue_unlock(pq); 3919 critical_exit(); 3920 } 3921 3922 /* 3923 * vm_page_pqbatch_drain: [ internal use only ] 3924 * 3925 * Force all per-CPU page queue batch queues to be drained. This is 3926 * intended for use in severe memory shortages, to ensure that pages 3927 * do not remain stuck in the batch queues. 3928 */ 3929 void 3930 vm_page_pqbatch_drain(void) 3931 { 3932 struct thread *td; 3933 struct vm_domain *vmd; 3934 struct vm_pagequeue *pq; 3935 int cpu, domain, queue; 3936 3937 td = curthread; 3938 CPU_FOREACH(cpu) { 3939 thread_lock(td); 3940 sched_bind(td, cpu); 3941 thread_unlock(td); 3942 3943 for (domain = 0; domain < vm_ndomains; domain++) { 3944 vmd = VM_DOMAIN(domain); 3945 for (queue = 0; queue < PQ_COUNT; queue++) { 3946 pq = &vmd->vmd_pagequeues[queue]; 3947 vm_pagequeue_lock(pq); 3948 critical_enter(); 3949 vm_pqbatch_process(pq, 3950 DPCPU_PTR(pqbatch[domain][queue]), queue); 3951 critical_exit(); 3952 vm_pagequeue_unlock(pq); 3953 } 3954 } 3955 } 3956 thread_lock(td); 3957 sched_unbind(td); 3958 thread_unlock(td); 3959 } 3960 3961 /* 3962 * vm_page_dequeue_deferred: [ internal use only ] 3963 * 3964 * Request removal of the given page from its current page 3965 * queue. Physical removal from the queue may be deferred 3966 * indefinitely. 3967 */ 3968 void 3969 vm_page_dequeue_deferred(vm_page_t m) 3970 { 3971 vm_page_astate_t new, old; 3972 3973 old = vm_page_astate_load(m); 3974 do { 3975 if (old.queue == PQ_NONE) { 3976 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 3977 ("%s: page %p has unexpected queue state", 3978 __func__, m)); 3979 break; 3980 } 3981 new = old; 3982 new.flags |= PGA_DEQUEUE; 3983 } while (!vm_page_pqstate_commit_request(m, &old, new)); 3984 } 3985 3986 /* 3987 * vm_page_dequeue: 3988 * 3989 * Remove the page from whichever page queue it's in, if any, before 3990 * returning. 3991 */ 3992 void 3993 vm_page_dequeue(vm_page_t m) 3994 { 3995 vm_page_astate_t new, old; 3996 3997 old = vm_page_astate_load(m); 3998 do { 3999 if (__predict_true(old.queue == PQ_NONE)) { 4000 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4001 ("%s: page %p has unexpected queue state", 4002 __func__, m)); 4003 break; 4004 } 4005 new = old; 4006 new.flags &= ~PGA_QUEUE_OP_MASK; 4007 new.queue = PQ_NONE; 4008 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 4009 4010 } 4011 4012 /* 4013 * Schedule the given page for insertion into the specified page queue. 4014 * Physical insertion of the page may be deferred indefinitely. 4015 */ 4016 static void 4017 vm_page_enqueue(vm_page_t m, uint8_t queue) 4018 { 4019 4020 KASSERT(m->a.queue == PQ_NONE && 4021 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 4022 ("%s: page %p is already enqueued", __func__, m)); 4023 KASSERT(m->ref_count > 0, 4024 ("%s: page %p does not carry any references", __func__, m)); 4025 4026 m->a.queue = queue; 4027 if ((m->a.flags & PGA_REQUEUE) == 0) 4028 vm_page_aflag_set(m, PGA_REQUEUE); 4029 vm_page_pqbatch_submit(m, queue); 4030 } 4031 4032 /* 4033 * vm_page_free_prep: 4034 * 4035 * Prepares the given page to be put on the free list, 4036 * disassociating it from any VM object. The caller may return 4037 * the page to the free list only if this function returns true. 4038 * 4039 * The object, if it exists, must be locked, and then the page must 4040 * be xbusy. Otherwise the page must be not busied. A managed 4041 * page must be unmapped. 4042 */ 4043 static bool 4044 vm_page_free_prep(vm_page_t m) 4045 { 4046 4047 /* 4048 * Synchronize with threads that have dropped a reference to this 4049 * page. 4050 */ 4051 atomic_thread_fence_acq(); 4052 4053 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 4054 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 4055 uint64_t *p; 4056 int i; 4057 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 4058 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 4059 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 4060 m, i, (uintmax_t)*p)); 4061 } 4062 #endif 4063 if ((m->oflags & VPO_UNMANAGED) == 0) { 4064 KASSERT(!pmap_page_is_mapped(m), 4065 ("vm_page_free_prep: freeing mapped page %p", m)); 4066 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 4067 ("vm_page_free_prep: mapping flags set in page %p", m)); 4068 } else { 4069 KASSERT(m->a.queue == PQ_NONE, 4070 ("vm_page_free_prep: unmanaged page %p is queued", m)); 4071 } 4072 VM_CNT_INC(v_tfree); 4073 4074 if (m->object != NULL) { 4075 vm_page_radix_remove(m); 4076 vm_page_free_object_prep(m); 4077 } else 4078 vm_page_assert_unbusied(m); 4079 4080 vm_page_busy_free(m); 4081 4082 /* 4083 * If fictitious remove object association and 4084 * return. 4085 */ 4086 if ((m->flags & PG_FICTITIOUS) != 0) { 4087 KASSERT(m->ref_count == 1, 4088 ("fictitious page %p is referenced", m)); 4089 KASSERT(m->a.queue == PQ_NONE, 4090 ("fictitious page %p is queued", m)); 4091 return (false); 4092 } 4093 4094 /* 4095 * Pages need not be dequeued before they are returned to the physical 4096 * memory allocator, but they must at least be marked for a deferred 4097 * dequeue. 4098 */ 4099 if ((m->oflags & VPO_UNMANAGED) == 0) 4100 vm_page_dequeue_deferred(m); 4101 4102 m->valid = 0; 4103 vm_page_undirty(m); 4104 4105 if (m->ref_count != 0) 4106 panic("vm_page_free_prep: page %p has references", m); 4107 4108 /* 4109 * Restore the default memory attribute to the page. 4110 */ 4111 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 4112 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 4113 4114 #if VM_NRESERVLEVEL > 0 4115 /* 4116 * Determine whether the page belongs to a reservation. If the page was 4117 * allocated from a per-CPU cache, it cannot belong to a reservation, so 4118 * as an optimization, we avoid the check in that case. 4119 */ 4120 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 4121 return (false); 4122 #endif 4123 4124 return (true); 4125 } 4126 4127 /* 4128 * vm_page_free_toq: 4129 * 4130 * Returns the given page to the free list, disassociating it 4131 * from any VM object. 4132 * 4133 * The object must be locked. The page must be exclusively busied if it 4134 * belongs to an object. 4135 */ 4136 static void 4137 vm_page_free_toq(vm_page_t m) 4138 { 4139 struct vm_domain *vmd; 4140 uma_zone_t zone; 4141 4142 if (!vm_page_free_prep(m)) 4143 return; 4144 4145 vmd = vm_pagequeue_domain(m); 4146 if (__predict_false((m->flags & PG_NOFREE) != 0)) { 4147 vm_page_free_nofree(vmd, m); 4148 return; 4149 } 4150 zone = vmd->vmd_pgcache[m->pool].zone; 4151 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4152 uma_zfree(zone, m); 4153 return; 4154 } 4155 vm_domain_free_lock(vmd); 4156 vm_phys_free_pages(m, m->pool, 0); 4157 vm_domain_free_unlock(vmd); 4158 vm_domain_freecnt_inc(vmd, 1); 4159 } 4160 4161 /* 4162 * vm_page_free_pages_toq: 4163 * 4164 * Returns a list of pages to the free list, disassociating it 4165 * from any VM object. In other words, this is equivalent to 4166 * calling vm_page_free_toq() for each page of a list of VM objects. 4167 */ 4168 int 4169 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4170 { 4171 vm_page_t m; 4172 int count; 4173 4174 if (SLIST_EMPTY(free)) 4175 return (0); 4176 4177 count = 0; 4178 while ((m = SLIST_FIRST(free)) != NULL) { 4179 count++; 4180 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4181 vm_page_free_toq(m); 4182 } 4183 4184 if (update_wire_count) 4185 vm_wire_sub(count); 4186 return (count); 4187 } 4188 4189 /* 4190 * Mark this page as wired down. For managed pages, this prevents reclamation 4191 * by the page daemon, or when the containing object, if any, is destroyed. 4192 */ 4193 void 4194 vm_page_wire(vm_page_t m) 4195 { 4196 u_int old; 4197 4198 #ifdef INVARIANTS 4199 if (m->object != NULL && !vm_page_busied(m) && 4200 !vm_object_busied(m->object)) 4201 VM_OBJECT_ASSERT_LOCKED(m->object); 4202 #endif 4203 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4204 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4205 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4206 4207 old = atomic_fetchadd_int(&m->ref_count, 1); 4208 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4209 ("vm_page_wire: counter overflow for page %p", m)); 4210 if (VPRC_WIRE_COUNT(old) == 0) { 4211 if ((m->oflags & VPO_UNMANAGED) == 0) 4212 vm_page_aflag_set(m, PGA_DEQUEUE); 4213 vm_wire_add(1); 4214 } 4215 } 4216 4217 /* 4218 * Attempt to wire a mapped page following a pmap lookup of that page. 4219 * This may fail if a thread is concurrently tearing down mappings of the page. 4220 * The transient failure is acceptable because it translates to the 4221 * failure of the caller pmap_extract_and_hold(), which should be then 4222 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4223 */ 4224 bool 4225 vm_page_wire_mapped(vm_page_t m) 4226 { 4227 u_int old; 4228 4229 old = atomic_load_int(&m->ref_count); 4230 do { 4231 KASSERT(old > 0, 4232 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4233 if ((old & VPRC_BLOCKED) != 0) 4234 return (false); 4235 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4236 4237 if (VPRC_WIRE_COUNT(old) == 0) { 4238 if ((m->oflags & VPO_UNMANAGED) == 0) 4239 vm_page_aflag_set(m, PGA_DEQUEUE); 4240 vm_wire_add(1); 4241 } 4242 return (true); 4243 } 4244 4245 /* 4246 * Release a wiring reference to a managed page. If the page still belongs to 4247 * an object, update its position in the page queues to reflect the reference. 4248 * If the wiring was the last reference to the page, free the page. 4249 */ 4250 static void 4251 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4252 { 4253 u_int old; 4254 4255 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4256 ("%s: page %p is unmanaged", __func__, m)); 4257 4258 /* 4259 * Update LRU state before releasing the wiring reference. 4260 * Use a release store when updating the reference count to 4261 * synchronize with vm_page_free_prep(). 4262 */ 4263 old = atomic_load_int(&m->ref_count); 4264 do { 4265 u_int count; 4266 4267 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4268 ("vm_page_unwire: wire count underflow for page %p", m)); 4269 4270 count = old & ~VPRC_BLOCKED; 4271 if (count > VPRC_OBJREF + 1) { 4272 /* 4273 * The page has at least one other wiring reference. An 4274 * earlier iteration of this loop may have called 4275 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4276 * re-set it if necessary. 4277 */ 4278 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4279 vm_page_aflag_set(m, PGA_DEQUEUE); 4280 } else if (count == VPRC_OBJREF + 1) { 4281 /* 4282 * This is the last wiring. Clear PGA_DEQUEUE and 4283 * update the page's queue state to reflect the 4284 * reference. If the page does not belong to an object 4285 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4286 * clear leftover queue state. 4287 */ 4288 vm_page_release_toq(m, nqueue, noreuse); 4289 } else if (count == 1) { 4290 vm_page_aflag_clear(m, PGA_DEQUEUE); 4291 } 4292 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4293 4294 if (VPRC_WIRE_COUNT(old) == 1) { 4295 vm_wire_sub(1); 4296 if (old == 1) 4297 vm_page_free(m); 4298 } 4299 } 4300 4301 /* 4302 * Release one wiring of the specified page, potentially allowing it to be 4303 * paged out. 4304 * 4305 * Only managed pages belonging to an object can be paged out. If the number 4306 * of wirings transitions to zero and the page is eligible for page out, then 4307 * the page is added to the specified paging queue. If the released wiring 4308 * represented the last reference to the page, the page is freed. 4309 */ 4310 void 4311 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4312 { 4313 4314 KASSERT(nqueue < PQ_COUNT, 4315 ("vm_page_unwire: invalid queue %u request for page %p", 4316 nqueue, m)); 4317 4318 if ((m->oflags & VPO_UNMANAGED) != 0) { 4319 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4320 vm_page_free(m); 4321 return; 4322 } 4323 vm_page_unwire_managed(m, nqueue, false); 4324 } 4325 4326 /* 4327 * Unwire a page without (re-)inserting it into a page queue. It is up 4328 * to the caller to enqueue, requeue, or free the page as appropriate. 4329 * In most cases involving managed pages, vm_page_unwire() should be used 4330 * instead. 4331 */ 4332 bool 4333 vm_page_unwire_noq(vm_page_t m) 4334 { 4335 u_int old; 4336 4337 old = vm_page_drop(m, 1); 4338 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4339 ("%s: counter underflow for page %p", __func__, m)); 4340 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4341 ("%s: missing ref on fictitious page %p", __func__, m)); 4342 4343 if (VPRC_WIRE_COUNT(old) > 1) 4344 return (false); 4345 if ((m->oflags & VPO_UNMANAGED) == 0) 4346 vm_page_aflag_clear(m, PGA_DEQUEUE); 4347 vm_wire_sub(1); 4348 return (true); 4349 } 4350 4351 /* 4352 * Ensure that the page ends up in the specified page queue. If the page is 4353 * active or being moved to the active queue, ensure that its act_count is 4354 * at least ACT_INIT but do not otherwise mess with it. 4355 */ 4356 static __always_inline void 4357 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4358 { 4359 vm_page_astate_t old, new; 4360 4361 KASSERT(m->ref_count > 0, 4362 ("%s: page %p does not carry any references", __func__, m)); 4363 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4364 ("%s: invalid flags %x", __func__, nflag)); 4365 4366 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4367 return; 4368 4369 old = vm_page_astate_load(m); 4370 do { 4371 if ((old.flags & PGA_DEQUEUE) != 0) 4372 break; 4373 new = old; 4374 new.flags &= ~PGA_QUEUE_OP_MASK; 4375 if (nqueue == PQ_ACTIVE) 4376 new.act_count = max(old.act_count, ACT_INIT); 4377 if (old.queue == nqueue) { 4378 /* 4379 * There is no need to requeue pages already in the 4380 * active queue. 4381 */ 4382 if (nqueue != PQ_ACTIVE || 4383 (old.flags & PGA_ENQUEUED) == 0) 4384 new.flags |= nflag; 4385 } else { 4386 new.flags |= nflag; 4387 new.queue = nqueue; 4388 } 4389 } while (!vm_page_pqstate_commit(m, &old, new)); 4390 } 4391 4392 /* 4393 * Put the specified page on the active list (if appropriate). 4394 */ 4395 void 4396 vm_page_activate(vm_page_t m) 4397 { 4398 4399 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4400 } 4401 4402 /* 4403 * Move the specified page to the tail of the inactive queue, or requeue 4404 * the page if it is already in the inactive queue. 4405 */ 4406 void 4407 vm_page_deactivate(vm_page_t m) 4408 { 4409 4410 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4411 } 4412 4413 void 4414 vm_page_deactivate_noreuse(vm_page_t m) 4415 { 4416 4417 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4418 } 4419 4420 /* 4421 * Put a page in the laundry, or requeue it if it is already there. 4422 */ 4423 void 4424 vm_page_launder(vm_page_t m) 4425 { 4426 4427 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4428 } 4429 4430 /* 4431 * Put a page in the PQ_UNSWAPPABLE holding queue. 4432 */ 4433 void 4434 vm_page_unswappable(vm_page_t m) 4435 { 4436 4437 VM_OBJECT_ASSERT_LOCKED(m->object); 4438 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4439 ("page %p already unswappable", m)); 4440 4441 vm_page_dequeue(m); 4442 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4443 } 4444 4445 /* 4446 * Release a page back to the page queues in preparation for unwiring. 4447 */ 4448 static void 4449 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4450 { 4451 vm_page_astate_t old, new; 4452 uint16_t nflag; 4453 4454 /* 4455 * Use a check of the valid bits to determine whether we should 4456 * accelerate reclamation of the page. The object lock might not be 4457 * held here, in which case the check is racy. At worst we will either 4458 * accelerate reclamation of a valid page and violate LRU, or 4459 * unnecessarily defer reclamation of an invalid page. 4460 * 4461 * If we were asked to not cache the page, place it near the head of the 4462 * inactive queue so that is reclaimed sooner. 4463 */ 4464 if (noreuse || vm_page_none_valid(m)) { 4465 nqueue = PQ_INACTIVE; 4466 nflag = PGA_REQUEUE_HEAD; 4467 } else { 4468 nflag = PGA_REQUEUE; 4469 } 4470 4471 old = vm_page_astate_load(m); 4472 do { 4473 new = old; 4474 4475 /* 4476 * If the page is already in the active queue and we are not 4477 * trying to accelerate reclamation, simply mark it as 4478 * referenced and avoid any queue operations. 4479 */ 4480 new.flags &= ~PGA_QUEUE_OP_MASK; 4481 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4482 (old.flags & PGA_ENQUEUED) != 0) 4483 new.flags |= PGA_REFERENCED; 4484 else { 4485 new.flags |= nflag; 4486 new.queue = nqueue; 4487 } 4488 } while (!vm_page_pqstate_commit(m, &old, new)); 4489 } 4490 4491 /* 4492 * Unwire a page and either attempt to free it or re-add it to the page queues. 4493 */ 4494 void 4495 vm_page_release(vm_page_t m, int flags) 4496 { 4497 vm_object_t object; 4498 4499 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4500 ("vm_page_release: page %p is unmanaged", m)); 4501 4502 if ((flags & VPR_TRYFREE) != 0) { 4503 for (;;) { 4504 object = atomic_load_ptr(&m->object); 4505 if (object == NULL) 4506 break; 4507 /* Depends on type-stability. */ 4508 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4509 break; 4510 if (object == m->object) { 4511 vm_page_release_locked(m, flags); 4512 VM_OBJECT_WUNLOCK(object); 4513 return; 4514 } 4515 VM_OBJECT_WUNLOCK(object); 4516 } 4517 } 4518 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4519 } 4520 4521 /* See vm_page_release(). */ 4522 void 4523 vm_page_release_locked(vm_page_t m, int flags) 4524 { 4525 4526 VM_OBJECT_ASSERT_WLOCKED(m->object); 4527 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4528 ("vm_page_release_locked: page %p is unmanaged", m)); 4529 4530 if (vm_page_unwire_noq(m)) { 4531 if ((flags & VPR_TRYFREE) != 0 && 4532 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4533 m->dirty == 0 && vm_page_tryxbusy(m)) { 4534 /* 4535 * An unlocked lookup may have wired the page before the 4536 * busy lock was acquired, in which case the page must 4537 * not be freed. 4538 */ 4539 if (__predict_true(!vm_page_wired(m))) { 4540 vm_page_free(m); 4541 return; 4542 } 4543 vm_page_xunbusy(m); 4544 } else { 4545 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4546 } 4547 } 4548 } 4549 4550 static bool 4551 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4552 { 4553 u_int old; 4554 4555 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4556 ("vm_page_try_blocked_op: page %p has no object", m)); 4557 KASSERT(vm_page_busied(m), 4558 ("vm_page_try_blocked_op: page %p is not busy", m)); 4559 VM_OBJECT_ASSERT_LOCKED(m->object); 4560 4561 old = atomic_load_int(&m->ref_count); 4562 do { 4563 KASSERT(old != 0, 4564 ("vm_page_try_blocked_op: page %p has no references", m)); 4565 KASSERT((old & VPRC_BLOCKED) == 0, 4566 ("vm_page_try_blocked_op: page %p blocks wirings", m)); 4567 if (VPRC_WIRE_COUNT(old) != 0) 4568 return (false); 4569 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4570 4571 (op)(m); 4572 4573 /* 4574 * If the object is read-locked, new wirings may be created via an 4575 * object lookup. 4576 */ 4577 old = vm_page_drop(m, VPRC_BLOCKED); 4578 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4579 old == (VPRC_BLOCKED | VPRC_OBJREF), 4580 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4581 old, m)); 4582 return (true); 4583 } 4584 4585 /* 4586 * Atomically check for wirings and remove all mappings of the page. 4587 */ 4588 bool 4589 vm_page_try_remove_all(vm_page_t m) 4590 { 4591 4592 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4593 } 4594 4595 /* 4596 * Atomically check for wirings and remove all writeable mappings of the page. 4597 */ 4598 bool 4599 vm_page_try_remove_write(vm_page_t m) 4600 { 4601 4602 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4603 } 4604 4605 /* 4606 * vm_page_advise 4607 * 4608 * Apply the specified advice to the given page. 4609 */ 4610 void 4611 vm_page_advise(vm_page_t m, int advice) 4612 { 4613 4614 VM_OBJECT_ASSERT_WLOCKED(m->object); 4615 vm_page_assert_xbusied(m); 4616 4617 if (advice == MADV_FREE) 4618 /* 4619 * Mark the page clean. This will allow the page to be freed 4620 * without first paging it out. MADV_FREE pages are often 4621 * quickly reused by malloc(3), so we do not do anything that 4622 * would result in a page fault on a later access. 4623 */ 4624 vm_page_undirty(m); 4625 else if (advice != MADV_DONTNEED) { 4626 if (advice == MADV_WILLNEED) 4627 vm_page_activate(m); 4628 return; 4629 } 4630 4631 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4632 vm_page_dirty(m); 4633 4634 /* 4635 * Clear any references to the page. Otherwise, the page daemon will 4636 * immediately reactivate the page. 4637 */ 4638 vm_page_aflag_clear(m, PGA_REFERENCED); 4639 4640 /* 4641 * Place clean pages near the head of the inactive queue rather than 4642 * the tail, thus defeating the queue's LRU operation and ensuring that 4643 * the page will be reused quickly. Dirty pages not already in the 4644 * laundry are moved there. 4645 */ 4646 if (m->dirty == 0) 4647 vm_page_deactivate_noreuse(m); 4648 else if (!vm_page_in_laundry(m)) 4649 vm_page_launder(m); 4650 } 4651 4652 /* 4653 * vm_page_grab_release 4654 * 4655 * Helper routine for grab functions to release busy on return. 4656 */ 4657 static inline void 4658 vm_page_grab_release(vm_page_t m, int allocflags) 4659 { 4660 4661 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4662 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4663 vm_page_sunbusy(m); 4664 else 4665 vm_page_xunbusy(m); 4666 } 4667 } 4668 4669 /* 4670 * vm_page_grab_sleep 4671 * 4672 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4673 * if the caller should retry and false otherwise. 4674 * 4675 * If the object is locked on entry the object will be unlocked with 4676 * false returns and still locked but possibly having been dropped 4677 * with true returns. 4678 */ 4679 static bool 4680 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4681 const char *wmesg, int allocflags, bool locked) 4682 { 4683 4684 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4685 return (false); 4686 4687 /* 4688 * Reference the page before unlocking and sleeping so that 4689 * the page daemon is less likely to reclaim it. 4690 */ 4691 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4692 vm_page_reference(m); 4693 4694 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4695 locked) 4696 VM_OBJECT_WLOCK(object); 4697 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4698 return (false); 4699 4700 return (true); 4701 } 4702 4703 /* 4704 * Assert that the grab flags are valid. 4705 */ 4706 static inline void 4707 vm_page_grab_check(int allocflags) 4708 { 4709 4710 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4711 (allocflags & VM_ALLOC_WIRED) != 0, 4712 ("vm_page_grab*: the pages must be busied or wired")); 4713 4714 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4715 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4716 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4717 } 4718 4719 /* 4720 * Calculate the page allocation flags for grab. 4721 */ 4722 static inline int 4723 vm_page_grab_pflags(int allocflags) 4724 { 4725 int pflags; 4726 4727 pflags = allocflags & 4728 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4729 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); 4730 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4731 pflags |= VM_ALLOC_WAITFAIL; 4732 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4733 pflags |= VM_ALLOC_SBUSY; 4734 4735 return (pflags); 4736 } 4737 4738 /* 4739 * Grab a page, waiting until we are woken up due to the page changing state. 4740 * We keep on waiting, if the page continues to be in the object, unless 4741 * allocflags forbid waiting. 4742 * 4743 * The object must be locked on entry. This routine may sleep. The lock will, 4744 * however, be released and reacquired if the routine sleeps. 4745 * 4746 * Return a grabbed page, or NULL. Set *found if a page was found, whether or 4747 * not it was grabbed. 4748 */ 4749 static inline vm_page_t 4750 vm_page_grab_lookup(vm_object_t object, vm_pindex_t pindex, int allocflags, 4751 bool *found, struct pctrie_iter *pages) 4752 { 4753 vm_page_t m; 4754 4755 while ((*found = (m = vm_radix_iter_lookup(pages, pindex)) != NULL) && 4756 !vm_page_tryacquire(m, allocflags)) { 4757 if (!vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4758 allocflags, true)) 4759 return (NULL); 4760 pctrie_iter_reset(pages); 4761 } 4762 return (m); 4763 } 4764 4765 /* 4766 * Grab a page. Use an iterator parameter. Keep on waiting, as long as the page 4767 * exists in the object. If the page doesn't exist, first allocate it and then 4768 * conditionally zero it. 4769 * 4770 * The object must be locked on entry. This routine may sleep. The lock will, 4771 * however, be released and reacquired if the routine sleeps. 4772 */ 4773 vm_page_t 4774 vm_page_grab_iter(vm_object_t object, vm_pindex_t pindex, int allocflags, 4775 struct pctrie_iter *pages) 4776 { 4777 vm_page_t m; 4778 bool found; 4779 4780 VM_OBJECT_ASSERT_WLOCKED(object); 4781 vm_page_grab_check(allocflags); 4782 4783 while ((m = vm_page_grab_lookup( 4784 object, pindex, allocflags, &found, pages)) == NULL) { 4785 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4786 return (NULL); 4787 if (found && 4788 (allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4789 return (NULL); 4790 m = vm_page_alloc_iter(object, pindex, 4791 vm_page_grab_pflags(allocflags), pages); 4792 if (m != NULL) { 4793 if ((allocflags & VM_ALLOC_ZERO) != 0 && 4794 (m->flags & PG_ZERO) == 0) 4795 pmap_zero_page(m); 4796 break; 4797 } 4798 if ((allocflags & 4799 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4800 return (NULL); 4801 } 4802 vm_page_grab_release(m, allocflags); 4803 4804 return (m); 4805 } 4806 4807 /* 4808 * Grab a page. Keep on waiting, as long as the page exists in the object. If 4809 * the page doesn't exist, first allocate it and then conditionally zero it. 4810 * 4811 * The object must be locked on entry. This routine may sleep. The lock will, 4812 * however, be released and reacquired if the routine sleeps. 4813 */ 4814 vm_page_t 4815 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4816 { 4817 struct pctrie_iter pages; 4818 4819 VM_OBJECT_ASSERT_WLOCKED(object); 4820 vm_page_iter_init(&pages, object); 4821 return (vm_page_grab_iter(object, pindex, allocflags, &pages)); 4822 } 4823 4824 /* 4825 * Attempt to validate a page, locklessly acquiring it if necessary, given a 4826 * (object, pindex) tuple and either an invalided page or NULL. The resulting 4827 * page will be validated against the identity tuple, and busied or wired as 4828 * requested. A NULL page returned guarantees that the page was not in radix at 4829 * the time of the call but callers must perform higher level synchronization or 4830 * retry the operation under a lock if they require an atomic answer. This is 4831 * the only lock free validation routine, other routines can depend on the 4832 * resulting page state. 4833 * 4834 * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to 4835 * caller flags. 4836 */ 4837 #define PAGE_NOT_ACQUIRED ((vm_page_t)1) 4838 static vm_page_t 4839 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m, 4840 int allocflags) 4841 { 4842 if (m == NULL) 4843 m = vm_page_lookup_unlocked(object, pindex); 4844 for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) { 4845 if (vm_page_trybusy(m, allocflags)) { 4846 if (m->object == object && m->pindex == pindex) { 4847 if ((allocflags & VM_ALLOC_WIRED) != 0) 4848 vm_page_wire(m); 4849 vm_page_grab_release(m, allocflags); 4850 break; 4851 } 4852 /* relookup. */ 4853 vm_page_busy_release(m); 4854 cpu_spinwait(); 4855 continue; 4856 } 4857 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4858 allocflags, false)) 4859 return (PAGE_NOT_ACQUIRED); 4860 } 4861 return (m); 4862 } 4863 4864 /* 4865 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4866 * is not set. 4867 */ 4868 vm_page_t 4869 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4870 { 4871 vm_page_t m; 4872 4873 vm_page_grab_check(allocflags); 4874 m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags); 4875 if (m == PAGE_NOT_ACQUIRED) 4876 return (NULL); 4877 if (m != NULL) 4878 return (m); 4879 4880 /* 4881 * The radix lockless lookup should never return a false negative 4882 * errors. If the user specifies NOCREAT they are guaranteed there 4883 * was no page present at the instant of the call. A NOCREAT caller 4884 * must handle create races gracefully. 4885 */ 4886 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4887 return (NULL); 4888 4889 VM_OBJECT_WLOCK(object); 4890 m = vm_page_grab(object, pindex, allocflags); 4891 VM_OBJECT_WUNLOCK(object); 4892 4893 return (m); 4894 } 4895 4896 /* 4897 * Grab a page and make it valid, paging in if necessary. Use an iterator 4898 * parameter. Pages missing from their pager are zero filled and validated. If 4899 * a VM_ALLOC_COUNT is supplied and the page is not valid as many as 4900 * VM_INITIAL_PAGEIN pages can be brought in simultaneously. Additional pages 4901 * will be left on a paging queue but will neither be wired nor busy regardless 4902 * of allocflags. 4903 */ 4904 int 4905 vm_page_grab_valid_iter(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, 4906 int allocflags, struct pctrie_iter *pages) 4907 { 4908 vm_page_t m; 4909 vm_page_t ma[VM_INITIAL_PAGEIN]; 4910 int after, ahead, i, pflags, rv; 4911 4912 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4913 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4914 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4915 KASSERT((allocflags & 4916 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4917 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4918 VM_OBJECT_ASSERT_WLOCKED(object); 4919 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4920 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4921 pflags |= VM_ALLOC_WAITFAIL; 4922 4923 retrylookup: 4924 if ((m = vm_radix_iter_lookup(pages, pindex)) != NULL) { 4925 /* 4926 * If the page is fully valid it can only become invalid 4927 * with the object lock held. If it is not valid it can 4928 * become valid with the busy lock held. Therefore, we 4929 * may unnecessarily lock the exclusive busy here if we 4930 * race with I/O completion not using the object lock. 4931 * However, we will not end up with an invalid page and a 4932 * shared lock. 4933 */ 4934 if (!vm_page_trybusy(m, 4935 vm_page_all_valid(m) ? allocflags : 0)) { 4936 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4937 allocflags, true); 4938 pctrie_iter_reset(pages); 4939 goto retrylookup; 4940 } 4941 if (vm_page_all_valid(m)) 4942 goto out; 4943 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4944 vm_page_busy_release(m); 4945 *mp = NULL; 4946 return (VM_PAGER_FAIL); 4947 } 4948 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4949 *mp = NULL; 4950 return (VM_PAGER_FAIL); 4951 } else { 4952 m = vm_page_alloc_iter(object, pindex, pflags, pages); 4953 if (m == NULL) { 4954 if (!vm_pager_can_alloc_page(object, pindex)) { 4955 *mp = NULL; 4956 return (VM_PAGER_AGAIN); 4957 } 4958 goto retrylookup; 4959 } 4960 } 4961 4962 vm_page_assert_xbusied(m); 4963 if (vm_pager_has_page(object, pindex, NULL, &after)) { 4964 after = MIN(after, VM_INITIAL_PAGEIN); 4965 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 4966 after = MAX(after, 1); 4967 ma[0] = m; 4968 pctrie_iter_reset(pages); 4969 for (i = 1; i < after; i++) { 4970 m = vm_radix_iter_lookup_ge(pages, pindex + i); 4971 ahead = after; 4972 if (m != NULL) 4973 ahead = MIN(ahead, m->pindex - pindex); 4974 for (; i < ahead; i++) { 4975 ma[i] = vm_page_alloc_iter(object, pindex + i, 4976 VM_ALLOC_NORMAL, pages); 4977 if (ma[i] == NULL) 4978 break; 4979 } 4980 if (m == NULL || m->pindex != pindex + i || 4981 vm_page_any_valid(m) || !vm_page_tryxbusy(m)) 4982 break; 4983 ma[i] = m; 4984 } 4985 after = i; 4986 vm_object_pip_add(object, after); 4987 VM_OBJECT_WUNLOCK(object); 4988 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 4989 pctrie_iter_reset(pages); 4990 VM_OBJECT_WLOCK(object); 4991 vm_object_pip_wakeupn(object, after); 4992 /* Pager may have replaced a page. */ 4993 m = ma[0]; 4994 if (rv != VM_PAGER_OK) { 4995 for (i = 0; i < after; i++) { 4996 if (!vm_page_wired(ma[i])) 4997 vm_page_free(ma[i]); 4998 else 4999 vm_page_xunbusy(ma[i]); 5000 } 5001 *mp = NULL; 5002 return (rv); 5003 } 5004 for (i = 1; i < after; i++) 5005 vm_page_readahead_finish(ma[i]); 5006 MPASS(vm_page_all_valid(m)); 5007 } else { 5008 vm_page_zero_invalid(m, TRUE); 5009 pctrie_iter_reset(pages); 5010 } 5011 out: 5012 if ((allocflags & VM_ALLOC_WIRED) != 0) 5013 vm_page_wire(m); 5014 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 5015 vm_page_busy_downgrade(m); 5016 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 5017 vm_page_busy_release(m); 5018 *mp = m; 5019 return (VM_PAGER_OK); 5020 } 5021 5022 /* 5023 * Grab a page and make it valid, paging in if necessary. Pages missing from 5024 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 5025 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 5026 * in simultaneously. Additional pages will be left on a paging queue but 5027 * will neither be wired nor busy regardless of allocflags. 5028 */ 5029 int 5030 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, 5031 int allocflags) 5032 { 5033 struct pctrie_iter pages; 5034 5035 VM_OBJECT_ASSERT_WLOCKED(object); 5036 vm_page_iter_init(&pages, object); 5037 return (vm_page_grab_valid_iter(mp, object, pindex, allocflags, 5038 &pages)); 5039 } 5040 5041 /* 5042 * Grab a page. Keep on waiting, as long as the page exists in the object. If 5043 * the page doesn't exist, and the pager has it, allocate it and zero part of 5044 * it. 5045 * 5046 * The object must be locked on entry. This routine may sleep. The lock will, 5047 * however, be released and reacquired if the routine sleeps. 5048 */ 5049 int 5050 vm_page_grab_zero_partial(vm_object_t object, vm_pindex_t pindex, int base, 5051 int end) 5052 { 5053 struct pctrie_iter pages; 5054 vm_page_t m; 5055 int allocflags, rv; 5056 bool found; 5057 5058 VM_OBJECT_ASSERT_WLOCKED(object); 5059 KASSERT(base >= 0, ("%s: base %d", __func__, base)); 5060 KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base, 5061 end)); 5062 5063 allocflags = VM_ALLOC_NOCREAT | VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL; 5064 vm_page_iter_init(&pages, object); 5065 while ((m = vm_page_grab_lookup( 5066 object, pindex, allocflags, &found, &pages)) == NULL) { 5067 if (!vm_pager_has_page(object, pindex, NULL, NULL)) 5068 return (0); 5069 m = vm_page_alloc_iter(object, pindex, 5070 vm_page_grab_pflags(allocflags), &pages); 5071 if (m != NULL) { 5072 vm_object_pip_add(object, 1); 5073 VM_OBJECT_WUNLOCK(object); 5074 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 5075 VM_OBJECT_WLOCK(object); 5076 vm_object_pip_wakeup(object); 5077 if (rv != VM_PAGER_OK) { 5078 vm_page_free(m); 5079 return (EIO); 5080 } 5081 5082 /* 5083 * Since the page was not resident, and therefore not 5084 * recently accessed, immediately enqueue it for 5085 * asynchronous laundering. The current operation is 5086 * not regarded as an access. 5087 */ 5088 vm_page_launder(m); 5089 break; 5090 } 5091 } 5092 5093 pmap_zero_page_area(m, base, end - base); 5094 KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m)); 5095 vm_page_set_dirty(m); 5096 vm_page_xunbusy(m); 5097 return (0); 5098 } 5099 5100 /* 5101 * Locklessly grab a valid page. If the page is not valid or not yet 5102 * allocated this will fall back to the object lock method. 5103 */ 5104 int 5105 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 5106 vm_pindex_t pindex, int allocflags) 5107 { 5108 vm_page_t m; 5109 int flags; 5110 int error; 5111 5112 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 5113 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 5114 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 5115 "mismatch")); 5116 KASSERT((allocflags & 5117 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 5118 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 5119 5120 /* 5121 * Attempt a lockless lookup and busy. We need at least an sbusy 5122 * before we can inspect the valid field and return a wired page. 5123 */ 5124 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 5125 vm_page_grab_check(flags); 5126 m = vm_page_acquire_unlocked(object, pindex, NULL, flags); 5127 if (m == PAGE_NOT_ACQUIRED) 5128 return (VM_PAGER_FAIL); 5129 if (m != NULL) { 5130 if (vm_page_all_valid(m)) { 5131 if ((allocflags & VM_ALLOC_WIRED) != 0) 5132 vm_page_wire(m); 5133 vm_page_grab_release(m, allocflags); 5134 *mp = m; 5135 return (VM_PAGER_OK); 5136 } 5137 vm_page_busy_release(m); 5138 } 5139 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5140 *mp = NULL; 5141 return (VM_PAGER_FAIL); 5142 } 5143 VM_OBJECT_WLOCK(object); 5144 error = vm_page_grab_valid(mp, object, pindex, allocflags); 5145 VM_OBJECT_WUNLOCK(object); 5146 5147 return (error); 5148 } 5149 5150 /* 5151 * Return the specified range of pages from the given object. For each 5152 * page offset within the range, if a page already exists within the object 5153 * at that offset and it is busy, then wait for it to change state. If, 5154 * instead, the page doesn't exist, then allocate it. 5155 * 5156 * The caller must always specify an allocation class. 5157 * 5158 * allocation classes: 5159 * VM_ALLOC_NORMAL normal process request 5160 * VM_ALLOC_SYSTEM system *really* needs the pages 5161 * VM_ALLOC_INTERRUPT interrupt time request 5162 * 5163 * The caller must always specify that the pages are to be busied and/or 5164 * wired. 5165 * 5166 * optional allocation flags: 5167 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 5168 * VM_ALLOC_NOBUSY do not exclusive busy the pages 5169 * VM_ALLOC_NODUMP do not include the pages in a kernel core dump 5170 * VM_ALLOC_NOFREE pages will never be freed 5171 * VM_ALLOC_NOWAIT do not sleep 5172 * VM_ALLOC_SBUSY set pages to sbusy state 5173 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 5174 * VM_ALLOC_WAITOK ignored (default behavior) 5175 * VM_ALLOC_WIRED wire the pages 5176 * VM_ALLOC_ZERO zero and validate any invalid pages 5177 * 5178 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 5179 * may return a partial prefix of the requested range. 5180 */ 5181 int 5182 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 5183 vm_page_t *ma, int count) 5184 { 5185 struct pctrie_iter pages; 5186 vm_page_t m; 5187 int pflags; 5188 int ahead, i; 5189 5190 VM_OBJECT_ASSERT_WLOCKED(object); 5191 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 5192 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 5193 KASSERT(count > 0, 5194 ("vm_page_grab_pages: invalid page count %d", count)); 5195 vm_page_grab_check(allocflags); 5196 5197 pflags = vm_page_grab_pflags(allocflags); 5198 i = 0; 5199 vm_page_iter_init(&pages, object); 5200 retrylookup: 5201 ahead = -1; 5202 for (; i < count; i++) { 5203 if (ahead < 0) { 5204 ahead = vm_radix_iter_lookup_range( 5205 &pages, pindex + i, &ma[i], count - i); 5206 } 5207 if (ahead-- > 0) { 5208 m = ma[i]; 5209 if (!vm_page_tryacquire(m, allocflags)) { 5210 if (vm_page_grab_sleep(object, m, pindex + i, 5211 "grbmaw", allocflags, true)) { 5212 pctrie_iter_reset(&pages); 5213 goto retrylookup; 5214 } 5215 break; 5216 } 5217 } else { 5218 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 5219 break; 5220 m = vm_page_alloc_iter(object, pindex + i, 5221 pflags | VM_ALLOC_COUNT(count - i), &pages); 5222 /* pages was reset if alloc_iter lost the lock. */ 5223 if (m == NULL) { 5224 if ((allocflags & (VM_ALLOC_NOWAIT | 5225 VM_ALLOC_WAITFAIL)) != 0) 5226 break; 5227 goto retrylookup; 5228 } 5229 ma[i] = m; 5230 } 5231 if (vm_page_none_valid(m) && 5232 (allocflags & VM_ALLOC_ZERO) != 0) { 5233 if ((m->flags & PG_ZERO) == 0) 5234 pmap_zero_page(m); 5235 vm_page_valid(m); 5236 } 5237 vm_page_grab_release(m, allocflags); 5238 } 5239 return (i); 5240 } 5241 5242 /* 5243 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 5244 * and will fall back to the locked variant to handle allocation. 5245 */ 5246 int 5247 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 5248 int allocflags, vm_page_t *ma, int count) 5249 { 5250 vm_page_t m; 5251 int flags; 5252 int i, num_fetched; 5253 5254 KASSERT(count > 0, 5255 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 5256 vm_page_grab_check(allocflags); 5257 5258 /* 5259 * Modify flags for lockless acquire to hold the page until we 5260 * set it valid if necessary. 5261 */ 5262 flags = allocflags & ~VM_ALLOC_NOBUSY; 5263 vm_page_grab_check(flags); 5264 num_fetched = vm_radix_lookup_range_unlocked(&object->rtree, pindex, 5265 ma, count); 5266 for (i = 0; i < num_fetched; i++, pindex++) { 5267 m = vm_page_acquire_unlocked(object, pindex, ma[i], flags); 5268 if (m == PAGE_NOT_ACQUIRED) 5269 return (i); 5270 if (m == NULL) 5271 break; 5272 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 5273 if ((m->flags & PG_ZERO) == 0) 5274 pmap_zero_page(m); 5275 vm_page_valid(m); 5276 } 5277 /* m will still be wired or busy according to flags. */ 5278 vm_page_grab_release(m, allocflags); 5279 /* vm_page_acquire_unlocked() may not return ma[i]. */ 5280 ma[i] = m; 5281 } 5282 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 5283 return (i); 5284 count -= i; 5285 VM_OBJECT_WLOCK(object); 5286 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 5287 VM_OBJECT_WUNLOCK(object); 5288 5289 return (i); 5290 } 5291 5292 /* 5293 * Mapping function for valid or dirty bits in a page. 5294 * 5295 * Inputs are required to range within a page. 5296 */ 5297 vm_page_bits_t 5298 vm_page_bits(int base, int size) 5299 { 5300 int first_bit; 5301 int last_bit; 5302 5303 KASSERT( 5304 base + size <= PAGE_SIZE, 5305 ("vm_page_bits: illegal base/size %d/%d", base, size) 5306 ); 5307 5308 if (size == 0) /* handle degenerate case */ 5309 return (0); 5310 5311 first_bit = base >> DEV_BSHIFT; 5312 last_bit = (base + size - 1) >> DEV_BSHIFT; 5313 5314 return (((vm_page_bits_t)2 << last_bit) - 5315 ((vm_page_bits_t)1 << first_bit)); 5316 } 5317 5318 void 5319 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5320 { 5321 5322 #if PAGE_SIZE == 32768 5323 atomic_set_64((uint64_t *)bits, set); 5324 #elif PAGE_SIZE == 16384 5325 atomic_set_32((uint32_t *)bits, set); 5326 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5327 atomic_set_16((uint16_t *)bits, set); 5328 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5329 atomic_set_8((uint8_t *)bits, set); 5330 #else /* PAGE_SIZE <= 8192 */ 5331 uintptr_t addr; 5332 int shift; 5333 5334 addr = (uintptr_t)bits; 5335 /* 5336 * Use a trick to perform a 32-bit atomic on the 5337 * containing aligned word, to not depend on the existence 5338 * of atomic_{set, clear}_{8, 16}. 5339 */ 5340 shift = addr & (sizeof(uint32_t) - 1); 5341 #if BYTE_ORDER == BIG_ENDIAN 5342 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5343 #else 5344 shift *= NBBY; 5345 #endif 5346 addr &= ~(sizeof(uint32_t) - 1); 5347 atomic_set_32((uint32_t *)addr, set << shift); 5348 #endif /* PAGE_SIZE */ 5349 } 5350 5351 static inline void 5352 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5353 { 5354 5355 #if PAGE_SIZE == 32768 5356 atomic_clear_64((uint64_t *)bits, clear); 5357 #elif PAGE_SIZE == 16384 5358 atomic_clear_32((uint32_t *)bits, clear); 5359 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5360 atomic_clear_16((uint16_t *)bits, clear); 5361 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5362 atomic_clear_8((uint8_t *)bits, clear); 5363 #else /* PAGE_SIZE <= 8192 */ 5364 uintptr_t addr; 5365 int shift; 5366 5367 addr = (uintptr_t)bits; 5368 /* 5369 * Use a trick to perform a 32-bit atomic on the 5370 * containing aligned word, to not depend on the existence 5371 * of atomic_{set, clear}_{8, 16}. 5372 */ 5373 shift = addr & (sizeof(uint32_t) - 1); 5374 #if BYTE_ORDER == BIG_ENDIAN 5375 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5376 #else 5377 shift *= NBBY; 5378 #endif 5379 addr &= ~(sizeof(uint32_t) - 1); 5380 atomic_clear_32((uint32_t *)addr, clear << shift); 5381 #endif /* PAGE_SIZE */ 5382 } 5383 5384 static inline vm_page_bits_t 5385 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5386 { 5387 #if PAGE_SIZE == 32768 5388 uint64_t old; 5389 5390 old = *bits; 5391 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5392 return (old); 5393 #elif PAGE_SIZE == 16384 5394 uint32_t old; 5395 5396 old = *bits; 5397 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5398 return (old); 5399 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5400 uint16_t old; 5401 5402 old = *bits; 5403 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5404 return (old); 5405 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5406 uint8_t old; 5407 5408 old = *bits; 5409 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5410 return (old); 5411 #else /* PAGE_SIZE <= 4096*/ 5412 uintptr_t addr; 5413 uint32_t old, new, mask; 5414 int shift; 5415 5416 addr = (uintptr_t)bits; 5417 /* 5418 * Use a trick to perform a 32-bit atomic on the 5419 * containing aligned word, to not depend on the existence 5420 * of atomic_{set, swap, clear}_{8, 16}. 5421 */ 5422 shift = addr & (sizeof(uint32_t) - 1); 5423 #if BYTE_ORDER == BIG_ENDIAN 5424 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5425 #else 5426 shift *= NBBY; 5427 #endif 5428 addr &= ~(sizeof(uint32_t) - 1); 5429 mask = VM_PAGE_BITS_ALL << shift; 5430 5431 old = *bits; 5432 do { 5433 new = old & ~mask; 5434 new |= newbits << shift; 5435 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5436 return (old >> shift); 5437 #endif /* PAGE_SIZE */ 5438 } 5439 5440 /* 5441 * vm_page_set_valid_range: 5442 * 5443 * Sets portions of a page valid. The arguments are expected 5444 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5445 * of any partial chunks touched by the range. The invalid portion of 5446 * such chunks will be zeroed. 5447 * 5448 * (base + size) must be less then or equal to PAGE_SIZE. 5449 */ 5450 void 5451 vm_page_set_valid_range(vm_page_t m, int base, int size) 5452 { 5453 int endoff, frag; 5454 vm_page_bits_t pagebits; 5455 5456 vm_page_assert_busied(m); 5457 if (size == 0) /* handle degenerate case */ 5458 return; 5459 5460 /* 5461 * If the base is not DEV_BSIZE aligned and the valid 5462 * bit is clear, we have to zero out a portion of the 5463 * first block. 5464 */ 5465 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5466 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5467 pmap_zero_page_area(m, frag, base - frag); 5468 5469 /* 5470 * If the ending offset is not DEV_BSIZE aligned and the 5471 * valid bit is clear, we have to zero out a portion of 5472 * the last block. 5473 */ 5474 endoff = base + size; 5475 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5476 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5477 pmap_zero_page_area(m, endoff, 5478 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5479 5480 /* 5481 * Assert that no previously invalid block that is now being validated 5482 * is already dirty. 5483 */ 5484 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5485 ("vm_page_set_valid_range: page %p is dirty", m)); 5486 5487 /* 5488 * Set valid bits inclusive of any overlap. 5489 */ 5490 pagebits = vm_page_bits(base, size); 5491 if (vm_page_xbusied(m)) 5492 m->valid |= pagebits; 5493 else 5494 vm_page_bits_set(m, &m->valid, pagebits); 5495 } 5496 5497 /* 5498 * Set the page dirty bits and free the invalid swap space if 5499 * present. Returns the previous dirty bits. 5500 */ 5501 vm_page_bits_t 5502 vm_page_set_dirty(vm_page_t m) 5503 { 5504 vm_page_bits_t old; 5505 5506 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5507 5508 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5509 old = m->dirty; 5510 m->dirty = VM_PAGE_BITS_ALL; 5511 } else 5512 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5513 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5514 vm_pager_page_unswapped(m); 5515 5516 return (old); 5517 } 5518 5519 /* 5520 * Clear the given bits from the specified page's dirty field. 5521 */ 5522 static __inline void 5523 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5524 { 5525 5526 vm_page_assert_busied(m); 5527 5528 /* 5529 * If the page is xbusied and not write mapped we are the 5530 * only thread that can modify dirty bits. Otherwise, The pmap 5531 * layer can call vm_page_dirty() without holding a distinguished 5532 * lock. The combination of page busy and atomic operations 5533 * suffice to guarantee consistency of the page dirty field. 5534 */ 5535 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5536 m->dirty &= ~pagebits; 5537 else 5538 vm_page_bits_clear(m, &m->dirty, pagebits); 5539 } 5540 5541 /* 5542 * vm_page_set_validclean: 5543 * 5544 * Sets portions of a page valid and clean. The arguments are expected 5545 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5546 * of any partial chunks touched by the range. The invalid portion of 5547 * such chunks will be zero'd. 5548 * 5549 * (base + size) must be less then or equal to PAGE_SIZE. 5550 */ 5551 void 5552 vm_page_set_validclean(vm_page_t m, int base, int size) 5553 { 5554 vm_page_bits_t oldvalid, pagebits; 5555 int endoff, frag; 5556 5557 vm_page_assert_busied(m); 5558 if (size == 0) /* handle degenerate case */ 5559 return; 5560 5561 /* 5562 * If the base is not DEV_BSIZE aligned and the valid 5563 * bit is clear, we have to zero out a portion of the 5564 * first block. 5565 */ 5566 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5567 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5568 pmap_zero_page_area(m, frag, base - frag); 5569 5570 /* 5571 * If the ending offset is not DEV_BSIZE aligned and the 5572 * valid bit is clear, we have to zero out a portion of 5573 * the last block. 5574 */ 5575 endoff = base + size; 5576 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5577 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5578 pmap_zero_page_area(m, endoff, 5579 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5580 5581 /* 5582 * Set valid, clear dirty bits. If validating the entire 5583 * page we can safely clear the pmap modify bit. We also 5584 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5585 * takes a write fault on a MAP_NOSYNC memory area the flag will 5586 * be set again. 5587 * 5588 * We set valid bits inclusive of any overlap, but we can only 5589 * clear dirty bits for DEV_BSIZE chunks that are fully within 5590 * the range. 5591 */ 5592 oldvalid = m->valid; 5593 pagebits = vm_page_bits(base, size); 5594 if (vm_page_xbusied(m)) 5595 m->valid |= pagebits; 5596 else 5597 vm_page_bits_set(m, &m->valid, pagebits); 5598 #if 0 /* NOT YET */ 5599 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5600 frag = DEV_BSIZE - frag; 5601 base += frag; 5602 size -= frag; 5603 if (size < 0) 5604 size = 0; 5605 } 5606 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5607 #endif 5608 if (base == 0 && size == PAGE_SIZE) { 5609 /* 5610 * The page can only be modified within the pmap if it is 5611 * mapped, and it can only be mapped if it was previously 5612 * fully valid. 5613 */ 5614 if (oldvalid == VM_PAGE_BITS_ALL) 5615 /* 5616 * Perform the pmap_clear_modify() first. Otherwise, 5617 * a concurrent pmap operation, such as 5618 * pmap_protect(), could clear a modification in the 5619 * pmap and set the dirty field on the page before 5620 * pmap_clear_modify() had begun and after the dirty 5621 * field was cleared here. 5622 */ 5623 pmap_clear_modify(m); 5624 m->dirty = 0; 5625 vm_page_aflag_clear(m, PGA_NOSYNC); 5626 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5627 m->dirty &= ~pagebits; 5628 else 5629 vm_page_clear_dirty_mask(m, pagebits); 5630 } 5631 5632 void 5633 vm_page_clear_dirty(vm_page_t m, int base, int size) 5634 { 5635 5636 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5637 } 5638 5639 /* 5640 * vm_page_set_invalid: 5641 * 5642 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5643 * valid and dirty bits for the effected areas are cleared. 5644 */ 5645 void 5646 vm_page_set_invalid(vm_page_t m, int base, int size) 5647 { 5648 vm_page_bits_t bits; 5649 vm_object_t object; 5650 5651 /* 5652 * The object lock is required so that pages can't be mapped 5653 * read-only while we're in the process of invalidating them. 5654 */ 5655 object = m->object; 5656 VM_OBJECT_ASSERT_WLOCKED(object); 5657 vm_page_assert_busied(m); 5658 5659 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5660 size >= object->un_pager.vnp.vnp_size) 5661 bits = VM_PAGE_BITS_ALL; 5662 else 5663 bits = vm_page_bits(base, size); 5664 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5665 pmap_remove_all(m); 5666 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5667 !pmap_page_is_mapped(m), 5668 ("vm_page_set_invalid: page %p is mapped", m)); 5669 if (vm_page_xbusied(m)) { 5670 m->valid &= ~bits; 5671 m->dirty &= ~bits; 5672 } else { 5673 vm_page_bits_clear(m, &m->valid, bits); 5674 vm_page_bits_clear(m, &m->dirty, bits); 5675 } 5676 } 5677 5678 /* 5679 * vm_page_invalid: 5680 * 5681 * Invalidates the entire page. The page must be busy, unmapped, and 5682 * the enclosing object must be locked. The object locks protects 5683 * against concurrent read-only pmap enter which is done without 5684 * busy. 5685 */ 5686 void 5687 vm_page_invalid(vm_page_t m) 5688 { 5689 5690 vm_page_assert_busied(m); 5691 VM_OBJECT_ASSERT_WLOCKED(m->object); 5692 MPASS(!pmap_page_is_mapped(m)); 5693 5694 if (vm_page_xbusied(m)) 5695 m->valid = 0; 5696 else 5697 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5698 } 5699 5700 /* 5701 * vm_page_zero_invalid() 5702 * 5703 * The kernel assumes that the invalid portions of a page contain 5704 * garbage, but such pages can be mapped into memory by user code. 5705 * When this occurs, we must zero out the non-valid portions of the 5706 * page so user code sees what it expects. 5707 * 5708 * Pages are most often semi-valid when the end of a file is mapped 5709 * into memory and the file's size is not page aligned. 5710 */ 5711 void 5712 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5713 { 5714 int b; 5715 int i; 5716 5717 /* 5718 * Scan the valid bits looking for invalid sections that 5719 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5720 * valid bit may be set ) have already been zeroed by 5721 * vm_page_set_validclean(). 5722 */ 5723 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5724 if (i == (PAGE_SIZE / DEV_BSIZE) || 5725 (m->valid & ((vm_page_bits_t)1 << i))) { 5726 if (i > b) { 5727 pmap_zero_page_area(m, 5728 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5729 } 5730 b = i + 1; 5731 } 5732 } 5733 5734 /* 5735 * setvalid is TRUE when we can safely set the zero'd areas 5736 * as being valid. We can do this if there are no cache consistency 5737 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5738 */ 5739 if (setvalid) 5740 vm_page_valid(m); 5741 } 5742 5743 /* 5744 * vm_page_is_valid: 5745 * 5746 * Is (partial) page valid? Note that the case where size == 0 5747 * will return FALSE in the degenerate case where the page is 5748 * entirely invalid, and TRUE otherwise. 5749 * 5750 * Some callers envoke this routine without the busy lock held and 5751 * handle races via higher level locks. Typical callers should 5752 * hold a busy lock to prevent invalidation. 5753 */ 5754 int 5755 vm_page_is_valid(vm_page_t m, int base, int size) 5756 { 5757 vm_page_bits_t bits; 5758 5759 bits = vm_page_bits(base, size); 5760 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5761 } 5762 5763 /* 5764 * Returns true if all of the specified predicates are true for the entire 5765 * (super)page and false otherwise. 5766 */ 5767 bool 5768 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m) 5769 { 5770 vm_object_t object; 5771 int i, npages; 5772 5773 object = m->object; 5774 if (skip_m != NULL && skip_m->object != object) 5775 return (false); 5776 VM_OBJECT_ASSERT_LOCKED(object); 5777 KASSERT(psind <= m->psind, 5778 ("psind %d > psind %d of m %p", psind, m->psind, m)); 5779 npages = atop(pagesizes[psind]); 5780 5781 /* 5782 * The physically contiguous pages that make up a superpage, i.e., a 5783 * page with a page size index ("psind") greater than zero, will 5784 * occupy adjacent entries in vm_page_array[]. 5785 */ 5786 for (i = 0; i < npages; i++) { 5787 /* Always test object consistency, including "skip_m". */ 5788 if (m[i].object != object) 5789 return (false); 5790 if (&m[i] == skip_m) 5791 continue; 5792 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5793 return (false); 5794 if ((flags & PS_ALL_DIRTY) != 0) { 5795 /* 5796 * Calling vm_page_test_dirty() or pmap_is_modified() 5797 * might stop this case from spuriously returning 5798 * "false". However, that would require a write lock 5799 * on the object containing "m[i]". 5800 */ 5801 if (m[i].dirty != VM_PAGE_BITS_ALL) 5802 return (false); 5803 } 5804 if ((flags & PS_ALL_VALID) != 0 && 5805 m[i].valid != VM_PAGE_BITS_ALL) 5806 return (false); 5807 } 5808 return (true); 5809 } 5810 5811 /* 5812 * Set the page's dirty bits if the page is modified. 5813 */ 5814 void 5815 vm_page_test_dirty(vm_page_t m) 5816 { 5817 5818 vm_page_assert_busied(m); 5819 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5820 vm_page_dirty(m); 5821 } 5822 5823 void 5824 vm_page_valid(vm_page_t m) 5825 { 5826 5827 vm_page_assert_busied(m); 5828 if (vm_page_xbusied(m)) 5829 m->valid = VM_PAGE_BITS_ALL; 5830 else 5831 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5832 } 5833 5834 #ifdef INVARIANTS 5835 void 5836 vm_page_object_busy_assert(vm_page_t m) 5837 { 5838 5839 /* 5840 * Certain of the page's fields may only be modified by the 5841 * holder of a page or object busy. 5842 */ 5843 if (m->object != NULL && !vm_page_busied(m)) 5844 VM_OBJECT_ASSERT_BUSY(m->object); 5845 } 5846 5847 void 5848 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5849 { 5850 5851 if ((bits & PGA_WRITEABLE) == 0) 5852 return; 5853 5854 /* 5855 * The PGA_WRITEABLE flag can only be set if the page is 5856 * managed, is exclusively busied or the object is locked. 5857 * Currently, this flag is only set by pmap_enter(). 5858 */ 5859 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5860 ("PGA_WRITEABLE on unmanaged page")); 5861 if (!vm_page_xbusied(m)) 5862 VM_OBJECT_ASSERT_BUSY(m->object); 5863 } 5864 #endif 5865 5866 #include "opt_ddb.h" 5867 #ifdef DDB 5868 #include <sys/kernel.h> 5869 5870 #include <ddb/ddb.h> 5871 5872 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5873 { 5874 5875 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5876 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5877 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5878 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5879 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5880 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5881 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5882 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5883 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5884 } 5885 5886 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5887 { 5888 int dom; 5889 5890 db_printf("pq_free %d\n", vm_free_count()); 5891 for (dom = 0; dom < vm_ndomains; dom++) { 5892 db_printf( 5893 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5894 dom, 5895 vm_dom[dom].vmd_page_count, 5896 vm_dom[dom].vmd_free_count, 5897 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5898 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5899 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5900 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5901 } 5902 } 5903 5904 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5905 { 5906 vm_page_t m; 5907 boolean_t phys, virt; 5908 5909 if (!have_addr) { 5910 db_printf("show pginfo addr\n"); 5911 return; 5912 } 5913 5914 phys = strchr(modif, 'p') != NULL; 5915 virt = strchr(modif, 'v') != NULL; 5916 if (virt) 5917 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5918 else if (phys) 5919 m = PHYS_TO_VM_PAGE(addr); 5920 else 5921 m = (vm_page_t)addr; 5922 db_printf( 5923 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5924 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5925 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5926 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5927 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5928 } 5929 #endif /* DDB */ 5930