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 pa_lock[PA_LOCK_COUNT]; 118 119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 120 /* The following fields are protected by the domainset lock. */ 121 domainset_t __exclusive_cache_line vm_min_domains; 122 domainset_t __exclusive_cache_line vm_severe_domains; 123 static int vm_min_waiters; 124 static int vm_severe_waiters; 125 static int vm_pageproc_waiters; 126 127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 128 "VM page statistics"); 129 130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 132 CTLFLAG_RD, &pqstate_commit_retries, 133 "Number of failed per-page atomic queue state updates"); 134 135 static COUNTER_U64_DEFINE_EARLY(queue_ops); 136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 137 CTLFLAG_RD, &queue_ops, 138 "Number of batched queue operations"); 139 140 static COUNTER_U64_DEFINE_EARLY(queue_nops); 141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 142 CTLFLAG_RD, &queue_nops, 143 "Number of batched queue operations with no effects"); 144 145 /* 146 * bogus page -- for I/O to/from partially complete buffers, 147 * or for paging into sparsely invalid regions. 148 */ 149 vm_page_t bogus_page; 150 151 vm_page_t vm_page_array; 152 long vm_page_array_size; 153 long first_page; 154 155 struct bitset *vm_page_dump; 156 long vm_page_dump_pages; 157 158 static TAILQ_HEAD(, vm_page) blacklist_head; 159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 161 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 162 163 static uma_zone_t fakepg_zone; 164 165 static void vm_page_alloc_check(vm_page_t m); 166 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req); 167 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 168 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 169 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 170 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 171 static bool vm_page_free_prep(vm_page_t m); 172 static void vm_page_free_toq(vm_page_t m); 173 static void vm_page_init(void *dummy); 174 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 175 vm_pindex_t pindex, vm_page_t mpred); 176 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 177 vm_page_t mpred); 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 pgcache = &vmd->vmd_pgcache[pool]; 222 pgcache->domain = domain; 223 pgcache->pool = pool; 224 pgcache->zone = uma_zcache_create("vm pgcache", 225 PAGE_SIZE, NULL, NULL, NULL, NULL, 226 vm_page_zone_import, vm_page_zone_release, pgcache, 227 UMA_ZONE_VM); 228 229 /* 230 * Limit each pool's zone to 0.1% of the pages in the 231 * domain. 232 */ 233 cache = maxcache != 0 ? maxcache : 234 vmd->vmd_page_count / 1000; 235 uma_zone_set_maxcache(pgcache->zone, cache); 236 } 237 } 238 } 239 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 240 241 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 242 #if PAGE_SIZE == 32768 243 #ifdef CTASSERT 244 CTASSERT(sizeof(u_long) >= 8); 245 #endif 246 #endif 247 248 /* 249 * vm_set_page_size: 250 * 251 * Sets the page size, perhaps based upon the memory 252 * size. Must be called before any use of page-size 253 * dependent functions. 254 */ 255 void 256 vm_set_page_size(void) 257 { 258 if (vm_cnt.v_page_size == 0) 259 vm_cnt.v_page_size = PAGE_SIZE; 260 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 261 panic("vm_set_page_size: page size not a power of two"); 262 } 263 264 /* 265 * vm_page_blacklist_next: 266 * 267 * Find the next entry in the provided string of blacklist 268 * addresses. Entries are separated by space, comma, or newline. 269 * If an invalid integer is encountered then the rest of the 270 * string is skipped. Updates the list pointer to the next 271 * character, or NULL if the string is exhausted or invalid. 272 */ 273 static vm_paddr_t 274 vm_page_blacklist_next(char **list, char *end) 275 { 276 vm_paddr_t bad; 277 char *cp, *pos; 278 279 if (list == NULL || *list == NULL) 280 return (0); 281 if (**list =='\0') { 282 *list = NULL; 283 return (0); 284 } 285 286 /* 287 * If there's no end pointer then the buffer is coming from 288 * the kenv and we know it's null-terminated. 289 */ 290 if (end == NULL) 291 end = *list + strlen(*list); 292 293 /* Ensure that strtoq() won't walk off the end */ 294 if (*end != '\0') { 295 if (*end == '\n' || *end == ' ' || *end == ',') 296 *end = '\0'; 297 else { 298 printf("Blacklist not terminated, skipping\n"); 299 *list = NULL; 300 return (0); 301 } 302 } 303 304 for (pos = *list; *pos != '\0'; pos = cp) { 305 bad = strtoq(pos, &cp, 0); 306 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 307 if (bad == 0) { 308 if (++cp < end) 309 continue; 310 else 311 break; 312 } 313 } else 314 break; 315 if (*cp == '\0' || ++cp >= end) 316 *list = NULL; 317 else 318 *list = cp; 319 return (trunc_page(bad)); 320 } 321 printf("Garbage in RAM blacklist, skipping\n"); 322 *list = NULL; 323 return (0); 324 } 325 326 bool 327 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 328 { 329 struct vm_domain *vmd; 330 vm_page_t m; 331 bool found; 332 333 m = vm_phys_paddr_to_vm_page(pa); 334 if (m == NULL) 335 return (true); /* page does not exist, no failure */ 336 337 vmd = VM_DOMAIN(vm_phys_domain(pa)); 338 vm_domain_free_lock(vmd); 339 found = vm_phys_unfree_page(pa); 340 vm_domain_free_unlock(vmd); 341 if (found) { 342 vm_domain_freecnt_inc(vmd, -1); 343 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 344 if (verbose) 345 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 346 } 347 return (found); 348 } 349 350 /* 351 * vm_page_blacklist_check: 352 * 353 * Iterate through the provided string of blacklist addresses, pulling 354 * each entry out of the physical allocator free list and putting it 355 * onto a list for reporting via the vm.page_blacklist sysctl. 356 */ 357 static void 358 vm_page_blacklist_check(char *list, char *end) 359 { 360 vm_paddr_t pa; 361 char *next; 362 363 next = list; 364 while (next != NULL) { 365 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 366 continue; 367 vm_page_blacklist_add(pa, bootverbose); 368 } 369 } 370 371 /* 372 * vm_page_blacklist_load: 373 * 374 * Search for a special module named "ram_blacklist". It'll be a 375 * plain text file provided by the user via the loader directive 376 * of the same name. 377 */ 378 static void 379 vm_page_blacklist_load(char **list, char **end) 380 { 381 void *mod; 382 u_char *ptr; 383 u_int len; 384 385 mod = NULL; 386 ptr = NULL; 387 388 mod = preload_search_by_type("ram_blacklist"); 389 if (mod != NULL) { 390 ptr = preload_fetch_addr(mod); 391 len = preload_fetch_size(mod); 392 } 393 *list = ptr; 394 if (ptr != NULL) 395 *end = ptr + len; 396 else 397 *end = NULL; 398 return; 399 } 400 401 static int 402 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 403 { 404 vm_page_t m; 405 struct sbuf sbuf; 406 int error, first; 407 408 first = 1; 409 error = sysctl_wire_old_buffer(req, 0); 410 if (error != 0) 411 return (error); 412 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 413 TAILQ_FOREACH(m, &blacklist_head, listq) { 414 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 415 (uintmax_t)m->phys_addr); 416 first = 0; 417 } 418 error = sbuf_finish(&sbuf); 419 sbuf_delete(&sbuf); 420 return (error); 421 } 422 423 /* 424 * Initialize a dummy page for use in scans of the specified paging queue. 425 * In principle, this function only needs to set the flag PG_MARKER. 426 * Nonetheless, it write busies the page as a safety precaution. 427 */ 428 void 429 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 430 { 431 432 bzero(marker, sizeof(*marker)); 433 marker->flags = PG_MARKER; 434 marker->a.flags = aflags; 435 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 436 marker->a.queue = queue; 437 } 438 439 static void 440 vm_page_domain_init(int domain) 441 { 442 struct vm_domain *vmd; 443 struct vm_pagequeue *pq; 444 int i; 445 446 vmd = VM_DOMAIN(domain); 447 bzero(vmd, sizeof(*vmd)); 448 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 449 "vm inactive pagequeue"; 450 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 451 "vm active pagequeue"; 452 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 453 "vm laundry pagequeue"; 454 *__DECONST(const char **, 455 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 456 "vm unswappable pagequeue"; 457 vmd->vmd_domain = domain; 458 vmd->vmd_page_count = 0; 459 vmd->vmd_free_count = 0; 460 vmd->vmd_segs = 0; 461 vmd->vmd_oom = false; 462 vmd->vmd_helper_threads_enabled = true; 463 for (i = 0; i < PQ_COUNT; i++) { 464 pq = &vmd->vmd_pagequeues[i]; 465 TAILQ_INIT(&pq->pq_pl); 466 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 467 MTX_DEF | MTX_DUPOK); 468 pq->pq_pdpages = 0; 469 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 470 } 471 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 472 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 473 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 474 475 /* 476 * inacthead is used to provide FIFO ordering for LRU-bypassing 477 * insertions. 478 */ 479 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 480 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 481 &vmd->vmd_inacthead, plinks.q); 482 483 /* 484 * The clock pages are used to implement active queue scanning without 485 * requeues. Scans start at clock[0], which is advanced after the scan 486 * ends. When the two clock hands meet, they are reset and scanning 487 * resumes from the head of the queue. 488 */ 489 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 490 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 491 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 492 &vmd->vmd_clock[0], plinks.q); 493 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 494 &vmd->vmd_clock[1], plinks.q); 495 } 496 497 /* 498 * Initialize a physical page in preparation for adding it to the free 499 * lists. 500 */ 501 void 502 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool) 503 { 504 m->object = NULL; 505 m->ref_count = 0; 506 m->busy_lock = VPB_FREED; 507 m->flags = m->a.flags = 0; 508 m->phys_addr = pa; 509 m->a.queue = PQ_NONE; 510 m->psind = 0; 511 m->segind = segind; 512 m->order = VM_NFREEORDER; 513 m->pool = pool; 514 m->valid = m->dirty = 0; 515 pmap_page_init(m); 516 } 517 518 #ifndef PMAP_HAS_PAGE_ARRAY 519 static vm_paddr_t 520 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 521 { 522 vm_paddr_t new_end; 523 524 /* 525 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 526 * However, because this page is allocated from KVM, out-of-bounds 527 * accesses using the direct map will not be trapped. 528 */ 529 *vaddr += PAGE_SIZE; 530 531 /* 532 * Allocate physical memory for the page structures, and map it. 533 */ 534 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 535 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 536 VM_PROT_READ | VM_PROT_WRITE); 537 vm_page_array_size = page_range; 538 539 return (new_end); 540 } 541 #endif 542 543 /* 544 * vm_page_startup: 545 * 546 * Initializes the resident memory module. Allocates physical memory for 547 * bootstrapping UMA and some data structures that are used to manage 548 * physical pages. Initializes these structures, and populates the free 549 * page queues. 550 */ 551 vm_offset_t 552 vm_page_startup(vm_offset_t vaddr) 553 { 554 struct vm_phys_seg *seg; 555 struct vm_domain *vmd; 556 vm_page_t m; 557 char *list, *listend; 558 vm_paddr_t end, high_avail, low_avail, new_end, size; 559 vm_paddr_t page_range __unused; 560 vm_paddr_t last_pa, pa, startp, endp; 561 u_long pagecount; 562 #if MINIDUMP_PAGE_TRACKING 563 u_long vm_page_dump_size; 564 #endif 565 int biggestone, i, segind; 566 #ifdef WITNESS 567 vm_offset_t mapped; 568 int witness_size; 569 #endif 570 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 571 long ii; 572 #endif 573 int pool; 574 #ifdef VM_FREEPOOL_LAZYINIT 575 int lazyinit; 576 #endif 577 578 vaddr = round_page(vaddr); 579 580 vm_phys_early_startup(); 581 biggestone = vm_phys_avail_largest(); 582 end = phys_avail[biggestone+1]; 583 584 /* 585 * Initialize the page and queue locks. 586 */ 587 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 588 for (i = 0; i < PA_LOCK_COUNT; i++) 589 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 590 for (i = 0; i < vm_ndomains; i++) 591 vm_page_domain_init(i); 592 593 new_end = end; 594 #ifdef WITNESS 595 witness_size = round_page(witness_startup_count()); 596 new_end -= witness_size; 597 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 598 VM_PROT_READ | VM_PROT_WRITE); 599 bzero((void *)mapped, witness_size); 600 witness_startup((void *)mapped); 601 #endif 602 603 #if MINIDUMP_PAGE_TRACKING 604 /* 605 * Allocate a bitmap to indicate that a random physical page 606 * needs to be included in a minidump. 607 * 608 * The amd64 port needs this to indicate which direct map pages 609 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 610 * 611 * However, i386 still needs this workspace internally within the 612 * minidump code. In theory, they are not needed on i386, but are 613 * included should the sf_buf code decide to use them. 614 */ 615 last_pa = 0; 616 vm_page_dump_pages = 0; 617 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 618 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 619 dump_avail[i] / PAGE_SIZE; 620 if (dump_avail[i + 1] > last_pa) 621 last_pa = dump_avail[i + 1]; 622 } 623 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 624 new_end -= vm_page_dump_size; 625 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 626 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 627 bzero((void *)vm_page_dump, vm_page_dump_size); 628 #if MINIDUMP_STARTUP_PAGE_TRACKING 629 /* 630 * Include the UMA bootstrap pages, witness pages and vm_page_dump 631 * in a crash dump. When pmap_map() uses the direct map, they are 632 * not automatically included. 633 */ 634 for (pa = new_end; pa < end; pa += PAGE_SIZE) 635 dump_add_page(pa); 636 #endif 637 #else 638 (void)last_pa; 639 #endif 640 phys_avail[biggestone + 1] = new_end; 641 #ifdef __amd64__ 642 /* 643 * Request that the physical pages underlying the message buffer be 644 * included in a crash dump. Since the message buffer is accessed 645 * through the direct map, they are not automatically included. 646 */ 647 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 648 last_pa = pa + round_page(msgbufsize); 649 while (pa < last_pa) { 650 dump_add_page(pa); 651 pa += PAGE_SIZE; 652 } 653 #else 654 (void)pa; 655 #endif 656 657 /* 658 * Determine the lowest and highest physical addresses and, in the case 659 * of VM_PHYSSEG_SPARSE, the exact size of the available physical 660 * memory. vm_phys_early_startup() already checked that phys_avail[] 661 * has at least one element. 662 */ 663 #ifdef VM_PHYSSEG_SPARSE 664 size = phys_avail[1] - phys_avail[0]; 665 #endif 666 low_avail = phys_avail[0]; 667 high_avail = phys_avail[1]; 668 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 669 #ifdef VM_PHYSSEG_SPARSE 670 size += phys_avail[i + 1] - phys_avail[i]; 671 #endif 672 if (phys_avail[i] < low_avail) 673 low_avail = phys_avail[i]; 674 if (phys_avail[i + 1] > high_avail) 675 high_avail = phys_avail[i + 1]; 676 } 677 for (i = 0; i < vm_phys_nsegs; i++) { 678 #ifdef VM_PHYSSEG_SPARSE 679 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 680 #endif 681 if (vm_phys_segs[i].start < low_avail) 682 low_avail = vm_phys_segs[i].start; 683 if (vm_phys_segs[i].end > high_avail) 684 high_avail = vm_phys_segs[i].end; 685 } 686 first_page = low_avail / PAGE_SIZE; 687 #ifdef VM_PHYSSEG_DENSE 688 size = high_avail - low_avail; 689 #endif 690 691 #ifdef PMAP_HAS_PAGE_ARRAY 692 pmap_page_array_startup(size / PAGE_SIZE); 693 biggestone = vm_phys_avail_largest(); 694 end = new_end = phys_avail[biggestone + 1]; 695 #else 696 #ifdef VM_PHYSSEG_DENSE 697 /* 698 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 699 * the overhead of a page structure per page only if vm_page_array is 700 * allocated from the last physical memory chunk. Otherwise, we must 701 * allocate page structures representing the physical memory 702 * underlying vm_page_array, even though they will not be used. 703 */ 704 if (new_end != high_avail) 705 page_range = size / PAGE_SIZE; 706 else 707 #endif 708 { 709 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 710 711 /* 712 * If the partial bytes remaining are large enough for 713 * a page (PAGE_SIZE) without a corresponding 714 * 'struct vm_page', then new_end will contain an 715 * extra page after subtracting the length of the VM 716 * page array. Compensate by subtracting an extra 717 * page from new_end. 718 */ 719 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 720 if (new_end == high_avail) 721 high_avail -= PAGE_SIZE; 722 new_end -= PAGE_SIZE; 723 } 724 } 725 end = new_end; 726 new_end = vm_page_array_alloc(&vaddr, end, page_range); 727 #endif 728 729 #if VM_NRESERVLEVEL > 0 730 /* 731 * Allocate physical memory for the reservation management system's 732 * data structures, and map it. 733 */ 734 new_end = vm_reserv_startup(&vaddr, new_end); 735 #endif 736 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 737 /* 738 * Include vm_page_array and vm_reserv_array in a crash dump. 739 */ 740 for (pa = new_end; pa < end; pa += PAGE_SIZE) 741 dump_add_page(pa); 742 #endif 743 phys_avail[biggestone + 1] = new_end; 744 745 /* 746 * Add physical memory segments corresponding to the available 747 * physical pages. 748 */ 749 for (i = 0; phys_avail[i + 1] != 0; i += 2) 750 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 751 752 /* 753 * Initialize the physical memory allocator. 754 */ 755 vm_phys_init(); 756 757 pool = VM_FREEPOOL_DEFAULT; 758 #ifdef VM_FREEPOOL_LAZYINIT 759 lazyinit = 1; 760 TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit); 761 if (lazyinit) 762 pool = VM_FREEPOOL_LAZYINIT; 763 #endif 764 765 /* 766 * Initialize the page structures and add every available page to the 767 * physical memory allocator's free lists. 768 */ 769 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 770 for (ii = 0; ii < vm_page_array_size; ii++) { 771 m = &vm_page_array[ii]; 772 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0, 773 VM_FREEPOOL_DEFAULT); 774 m->flags = PG_FICTITIOUS; 775 } 776 #endif 777 vm_cnt.v_page_count = 0; 778 for (segind = 0; segind < vm_phys_nsegs; segind++) { 779 seg = &vm_phys_segs[segind]; 780 781 /* 782 * Initialize pages not covered by phys_avail[], since they 783 * might be freed to the allocator at some future point, e.g., 784 * by kmem_bootstrap_free(). 785 */ 786 startp = seg->start; 787 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 788 if (startp >= seg->end) 789 break; 790 if (phys_avail[i + 1] < startp) 791 continue; 792 if (phys_avail[i] <= startp) { 793 startp = phys_avail[i + 1]; 794 continue; 795 } 796 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 797 for (endp = MIN(phys_avail[i], seg->end); 798 startp < endp; startp += PAGE_SIZE, m++) { 799 vm_page_init_page(m, startp, segind, 800 VM_FREEPOOL_DEFAULT); 801 } 802 } 803 804 /* 805 * Add the segment's pages that are covered by one of 806 * phys_avail's ranges to the free lists. 807 */ 808 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 809 if (seg->end <= phys_avail[i] || 810 seg->start >= phys_avail[i + 1]) 811 continue; 812 813 startp = MAX(seg->start, phys_avail[i]); 814 endp = MIN(seg->end, phys_avail[i + 1]); 815 pagecount = (u_long)atop(endp - startp); 816 if (pagecount == 0) 817 continue; 818 819 /* 820 * If lazy vm_page initialization is not enabled, simply 821 * initialize all of the pages in the segment covered by 822 * phys_avail. Otherwise, initialize only the first 823 * page of each run of free pages handed to the vm_phys 824 * allocator, which in turn defers initialization of 825 * pages until they are needed. 826 * 827 * This avoids blocking the boot process for long 828 * periods, which may be relevant for VMs (which ought 829 * to boot as quickly as possible) and/or systems with 830 * large amounts of physical memory. 831 */ 832 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 833 vm_page_init_page(m, startp, segind, pool); 834 if (pool == VM_FREEPOOL_DEFAULT) { 835 for (u_long j = 1; j < pagecount; j++) { 836 vm_page_init_page(&m[j], 837 startp + ptoa((vm_paddr_t)j), 838 segind, pool); 839 } 840 } 841 vmd = VM_DOMAIN(seg->domain); 842 vm_domain_free_lock(vmd); 843 vm_phys_enqueue_contig(m, pool, pagecount); 844 vm_domain_free_unlock(vmd); 845 vm_domain_freecnt_inc(vmd, pagecount); 846 vm_cnt.v_page_count += (u_int)pagecount; 847 vmd->vmd_page_count += (u_int)pagecount; 848 vmd->vmd_segs |= 1UL << segind; 849 } 850 } 851 852 /* 853 * Remove blacklisted pages from the physical memory allocator. 854 */ 855 TAILQ_INIT(&blacklist_head); 856 vm_page_blacklist_load(&list, &listend); 857 vm_page_blacklist_check(list, listend); 858 859 list = kern_getenv("vm.blacklist"); 860 vm_page_blacklist_check(list, NULL); 861 862 freeenv(list); 863 #if VM_NRESERVLEVEL > 0 864 /* 865 * Initialize the reservation management system. 866 */ 867 vm_reserv_init(); 868 #endif 869 870 return (vaddr); 871 } 872 873 void 874 vm_page_reference(vm_page_t m) 875 { 876 877 vm_page_aflag_set(m, PGA_REFERENCED); 878 } 879 880 /* 881 * vm_page_trybusy 882 * 883 * Helper routine for grab functions to trylock busy. 884 * 885 * Returns true on success and false on failure. 886 */ 887 static bool 888 vm_page_trybusy(vm_page_t m, int allocflags) 889 { 890 891 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 892 return (vm_page_trysbusy(m)); 893 else 894 return (vm_page_tryxbusy(m)); 895 } 896 897 /* 898 * vm_page_tryacquire 899 * 900 * Helper routine for grab functions to trylock busy and wire. 901 * 902 * Returns true on success and false on failure. 903 */ 904 static inline bool 905 vm_page_tryacquire(vm_page_t m, int allocflags) 906 { 907 bool locked; 908 909 locked = vm_page_trybusy(m, allocflags); 910 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 911 vm_page_wire(m); 912 return (locked); 913 } 914 915 /* 916 * vm_page_busy_acquire: 917 * 918 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 919 * and drop the object lock if necessary. 920 */ 921 bool 922 vm_page_busy_acquire(vm_page_t m, int allocflags) 923 { 924 vm_object_t obj; 925 bool locked; 926 927 /* 928 * The page-specific object must be cached because page 929 * identity can change during the sleep, causing the 930 * re-lock of a different object. 931 * It is assumed that a reference to the object is already 932 * held by the callers. 933 */ 934 obj = atomic_load_ptr(&m->object); 935 for (;;) { 936 if (vm_page_tryacquire(m, allocflags)) 937 return (true); 938 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 939 return (false); 940 if (obj != NULL) 941 locked = VM_OBJECT_WOWNED(obj); 942 else 943 locked = false; 944 MPASS(locked || vm_page_wired(m)); 945 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 946 locked) && locked) 947 VM_OBJECT_WLOCK(obj); 948 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 949 return (false); 950 KASSERT(m->object == obj || m->object == NULL, 951 ("vm_page_busy_acquire: page %p does not belong to %p", 952 m, obj)); 953 } 954 } 955 956 /* 957 * vm_page_busy_downgrade: 958 * 959 * Downgrade an exclusive busy page into a single shared busy page. 960 */ 961 void 962 vm_page_busy_downgrade(vm_page_t m) 963 { 964 u_int x; 965 966 vm_page_assert_xbusied(m); 967 968 x = vm_page_busy_fetch(m); 969 for (;;) { 970 if (atomic_fcmpset_rel_int(&m->busy_lock, 971 &x, VPB_SHARERS_WORD(1))) 972 break; 973 } 974 if ((x & VPB_BIT_WAITERS) != 0) 975 wakeup(m); 976 } 977 978 /* 979 * 980 * vm_page_busy_tryupgrade: 981 * 982 * Attempt to upgrade a single shared busy into an exclusive busy. 983 */ 984 int 985 vm_page_busy_tryupgrade(vm_page_t m) 986 { 987 u_int ce, x; 988 989 vm_page_assert_sbusied(m); 990 991 x = vm_page_busy_fetch(m); 992 ce = VPB_CURTHREAD_EXCLUSIVE; 993 for (;;) { 994 if (VPB_SHARERS(x) > 1) 995 return (0); 996 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 997 ("vm_page_busy_tryupgrade: invalid lock state")); 998 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 999 ce | (x & VPB_BIT_WAITERS))) 1000 continue; 1001 return (1); 1002 } 1003 } 1004 1005 /* 1006 * vm_page_sbusied: 1007 * 1008 * Return a positive value if the page is shared busied, 0 otherwise. 1009 */ 1010 int 1011 vm_page_sbusied(vm_page_t m) 1012 { 1013 u_int x; 1014 1015 x = vm_page_busy_fetch(m); 1016 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1017 } 1018 1019 /* 1020 * vm_page_sunbusy: 1021 * 1022 * Shared unbusy a page. 1023 */ 1024 void 1025 vm_page_sunbusy(vm_page_t m) 1026 { 1027 u_int x; 1028 1029 vm_page_assert_sbusied(m); 1030 1031 x = vm_page_busy_fetch(m); 1032 for (;;) { 1033 KASSERT(x != VPB_FREED, 1034 ("vm_page_sunbusy: Unlocking freed page.")); 1035 if (VPB_SHARERS(x) > 1) { 1036 if (atomic_fcmpset_int(&m->busy_lock, &x, 1037 x - VPB_ONE_SHARER)) 1038 break; 1039 continue; 1040 } 1041 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1042 ("vm_page_sunbusy: invalid lock state")); 1043 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1044 continue; 1045 if ((x & VPB_BIT_WAITERS) == 0) 1046 break; 1047 wakeup(m); 1048 break; 1049 } 1050 } 1051 1052 /* 1053 * vm_page_busy_sleep: 1054 * 1055 * Sleep if the page is busy, using the page pointer as wchan. 1056 * This is used to implement the hard-path of the busying mechanism. 1057 * 1058 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1059 * will not sleep if the page is shared-busy. 1060 * 1061 * The object lock must be held on entry. 1062 * 1063 * Returns true if it slept and dropped the object lock, or false 1064 * if there was no sleep and the lock is still held. 1065 */ 1066 bool 1067 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) 1068 { 1069 vm_object_t obj; 1070 1071 obj = m->object; 1072 VM_OBJECT_ASSERT_LOCKED(obj); 1073 1074 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, 1075 true)); 1076 } 1077 1078 /* 1079 * vm_page_busy_sleep_unlocked: 1080 * 1081 * Sleep if the page is busy, using the page pointer as wchan. 1082 * This is used to implement the hard-path of busying mechanism. 1083 * 1084 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1085 * will not sleep if the page is shared-busy. 1086 * 1087 * The object lock must not be held on entry. The operation will 1088 * return if the page changes identity. 1089 */ 1090 void 1091 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1092 const char *wmesg, int allocflags) 1093 { 1094 VM_OBJECT_ASSERT_UNLOCKED(obj); 1095 1096 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); 1097 } 1098 1099 /* 1100 * _vm_page_busy_sleep: 1101 * 1102 * Internal busy sleep function. Verifies the page identity and 1103 * lockstate against parameters. Returns true if it sleeps and 1104 * false otherwise. 1105 * 1106 * allocflags uses VM_ALLOC_* flags to specify the lock required. 1107 * 1108 * If locked is true the lock will be dropped for any true returns 1109 * and held for any false returns. 1110 */ 1111 static bool 1112 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1113 const char *wmesg, int allocflags, bool locked) 1114 { 1115 bool xsleep; 1116 u_int x; 1117 1118 /* 1119 * If the object is busy we must wait for that to drain to zero 1120 * before trying the page again. 1121 */ 1122 if (obj != NULL && vm_object_busied(obj)) { 1123 if (locked) 1124 VM_OBJECT_DROP(obj); 1125 vm_object_busy_wait(obj, wmesg); 1126 return (true); 1127 } 1128 1129 if (!vm_page_busied(m)) 1130 return (false); 1131 1132 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1133 sleepq_lock(m); 1134 x = vm_page_busy_fetch(m); 1135 do { 1136 /* 1137 * If the page changes objects or becomes unlocked we can 1138 * simply return. 1139 */ 1140 if (x == VPB_UNBUSIED || 1141 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1142 m->object != obj || m->pindex != pindex) { 1143 sleepq_release(m); 1144 return (false); 1145 } 1146 if ((x & VPB_BIT_WAITERS) != 0) 1147 break; 1148 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1149 if (locked) 1150 VM_OBJECT_DROP(obj); 1151 DROP_GIANT(); 1152 sleepq_add(m, NULL, wmesg, 0, 0); 1153 sleepq_wait(m, PVM); 1154 PICKUP_GIANT(); 1155 return (true); 1156 } 1157 1158 /* 1159 * vm_page_trysbusy: 1160 * 1161 * Try to shared busy a page. 1162 * If the operation succeeds 1 is returned otherwise 0. 1163 * The operation never sleeps. 1164 */ 1165 int 1166 vm_page_trysbusy(vm_page_t m) 1167 { 1168 vm_object_t obj; 1169 u_int x; 1170 1171 obj = m->object; 1172 x = vm_page_busy_fetch(m); 1173 for (;;) { 1174 if ((x & VPB_BIT_SHARED) == 0) 1175 return (0); 1176 /* 1177 * Reduce the window for transient busies that will trigger 1178 * false negatives in vm_page_ps_test(). 1179 */ 1180 if (obj != NULL && vm_object_busied(obj)) 1181 return (0); 1182 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1183 x + VPB_ONE_SHARER)) 1184 break; 1185 } 1186 1187 /* Refetch the object now that we're guaranteed that it is stable. */ 1188 obj = m->object; 1189 if (obj != NULL && vm_object_busied(obj)) { 1190 vm_page_sunbusy(m); 1191 return (0); 1192 } 1193 return (1); 1194 } 1195 1196 /* 1197 * vm_page_tryxbusy: 1198 * 1199 * Try to exclusive busy a page. 1200 * If the operation succeeds 1 is returned otherwise 0. 1201 * The operation never sleeps. 1202 */ 1203 int 1204 vm_page_tryxbusy(vm_page_t m) 1205 { 1206 vm_object_t obj; 1207 1208 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1209 VPB_CURTHREAD_EXCLUSIVE) == 0) 1210 return (0); 1211 1212 obj = m->object; 1213 if (obj != NULL && vm_object_busied(obj)) { 1214 vm_page_xunbusy(m); 1215 return (0); 1216 } 1217 return (1); 1218 } 1219 1220 static void 1221 vm_page_xunbusy_hard_tail(vm_page_t m) 1222 { 1223 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1224 /* Wake the waiter. */ 1225 wakeup(m); 1226 } 1227 1228 /* 1229 * vm_page_xunbusy_hard: 1230 * 1231 * Called when unbusy has failed because there is a waiter. 1232 */ 1233 void 1234 vm_page_xunbusy_hard(vm_page_t m) 1235 { 1236 vm_page_assert_xbusied(m); 1237 vm_page_xunbusy_hard_tail(m); 1238 } 1239 1240 void 1241 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1242 { 1243 vm_page_assert_xbusied_unchecked(m); 1244 vm_page_xunbusy_hard_tail(m); 1245 } 1246 1247 static void 1248 vm_page_busy_free(vm_page_t m) 1249 { 1250 u_int x; 1251 1252 atomic_thread_fence_rel(); 1253 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1254 if ((x & VPB_BIT_WAITERS) != 0) 1255 wakeup(m); 1256 } 1257 1258 /* 1259 * vm_page_unhold_pages: 1260 * 1261 * Unhold each of the pages that is referenced by the given array. 1262 */ 1263 void 1264 vm_page_unhold_pages(vm_page_t *ma, int count) 1265 { 1266 1267 for (; count != 0; count--) { 1268 vm_page_unwire(*ma, PQ_ACTIVE); 1269 ma++; 1270 } 1271 } 1272 1273 vm_page_t 1274 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1275 { 1276 vm_page_t m; 1277 1278 #ifdef VM_PHYSSEG_SPARSE 1279 m = vm_phys_paddr_to_vm_page(pa); 1280 if (m == NULL) 1281 m = vm_phys_fictitious_to_vm_page(pa); 1282 return (m); 1283 #elif defined(VM_PHYSSEG_DENSE) 1284 long pi; 1285 1286 pi = atop(pa); 1287 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1288 m = &vm_page_array[pi - first_page]; 1289 return (m); 1290 } 1291 return (vm_phys_fictitious_to_vm_page(pa)); 1292 #else 1293 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1294 #endif 1295 } 1296 1297 /* 1298 * vm_page_getfake: 1299 * 1300 * Create a fictitious page with the specified physical address and 1301 * memory attribute. The memory attribute is the only the machine- 1302 * dependent aspect of a fictitious page that must be initialized. 1303 */ 1304 vm_page_t 1305 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1306 { 1307 vm_page_t m; 1308 1309 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1310 vm_page_initfake(m, paddr, memattr); 1311 return (m); 1312 } 1313 1314 void 1315 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1316 { 1317 1318 if ((m->flags & PG_FICTITIOUS) != 0) { 1319 /* 1320 * The page's memattr might have changed since the 1321 * previous initialization. Update the pmap to the 1322 * new memattr. 1323 */ 1324 goto memattr; 1325 } 1326 m->phys_addr = paddr; 1327 m->a.queue = PQ_NONE; 1328 /* Fictitious pages don't use "segind". */ 1329 m->flags = PG_FICTITIOUS; 1330 /* Fictitious pages don't use "order" or "pool". */ 1331 m->oflags = VPO_UNMANAGED; 1332 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1333 /* Fictitious pages are unevictable. */ 1334 m->ref_count = 1; 1335 pmap_page_init(m); 1336 memattr: 1337 pmap_page_set_memattr(m, memattr); 1338 } 1339 1340 /* 1341 * vm_page_putfake: 1342 * 1343 * Release a fictitious page. 1344 */ 1345 void 1346 vm_page_putfake(vm_page_t m) 1347 { 1348 1349 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1350 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1351 ("vm_page_putfake: bad page %p", m)); 1352 vm_page_assert_xbusied(m); 1353 vm_page_busy_free(m); 1354 uma_zfree(fakepg_zone, m); 1355 } 1356 1357 /* 1358 * vm_page_updatefake: 1359 * 1360 * Update the given fictitious page to the specified physical address and 1361 * memory attribute. 1362 */ 1363 void 1364 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1365 { 1366 1367 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1368 ("vm_page_updatefake: bad page %p", m)); 1369 m->phys_addr = paddr; 1370 pmap_page_set_memattr(m, memattr); 1371 } 1372 1373 /* 1374 * vm_page_free: 1375 * 1376 * Free a page. 1377 */ 1378 void 1379 vm_page_free(vm_page_t m) 1380 { 1381 1382 m->flags &= ~PG_ZERO; 1383 vm_page_free_toq(m); 1384 } 1385 1386 /* 1387 * vm_page_free_zero: 1388 * 1389 * Free a page to the zerod-pages queue 1390 */ 1391 void 1392 vm_page_free_zero(vm_page_t m) 1393 { 1394 1395 m->flags |= PG_ZERO; 1396 vm_page_free_toq(m); 1397 } 1398 1399 /* 1400 * Unbusy and handle the page queueing for a page from a getpages request that 1401 * was optionally read ahead or behind. 1402 */ 1403 void 1404 vm_page_readahead_finish(vm_page_t m) 1405 { 1406 1407 /* We shouldn't put invalid pages on queues. */ 1408 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1409 1410 /* 1411 * Since the page is not the actually needed one, whether it should 1412 * be activated or deactivated is not obvious. Empirical results 1413 * have shown that deactivating the page is usually the best choice, 1414 * unless the page is wanted by another thread. 1415 */ 1416 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1417 vm_page_activate(m); 1418 else 1419 vm_page_deactivate(m); 1420 vm_page_xunbusy_unchecked(m); 1421 } 1422 1423 /* 1424 * Destroy the identity of an invalid page and free it if possible. 1425 * This is intended to be used when reading a page from backing store fails. 1426 */ 1427 void 1428 vm_page_free_invalid(vm_page_t m) 1429 { 1430 1431 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1432 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1433 KASSERT(m->object != NULL, ("page %p has no object", m)); 1434 VM_OBJECT_ASSERT_WLOCKED(m->object); 1435 1436 /* 1437 * We may be attempting to free the page as part of the handling for an 1438 * I/O error, in which case the page was xbusied by a different thread. 1439 */ 1440 vm_page_xbusy_claim(m); 1441 1442 /* 1443 * If someone has wired this page while the object lock 1444 * was not held, then the thread that unwires is responsible 1445 * for freeing the page. Otherwise just free the page now. 1446 * The wire count of this unmapped page cannot change while 1447 * we have the page xbusy and the page's object wlocked. 1448 */ 1449 if (vm_page_remove(m)) 1450 vm_page_free(m); 1451 } 1452 1453 /* 1454 * vm_page_dirty_KBI: [ internal use only ] 1455 * 1456 * Set all bits in the page's dirty field. 1457 * 1458 * The object containing the specified page must be locked if the 1459 * call is made from the machine-independent layer. 1460 * 1461 * See vm_page_clear_dirty_mask(). 1462 * 1463 * This function should only be called by vm_page_dirty(). 1464 */ 1465 void 1466 vm_page_dirty_KBI(vm_page_t m) 1467 { 1468 1469 /* Refer to this operation by its public name. */ 1470 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1471 m->dirty = VM_PAGE_BITS_ALL; 1472 } 1473 1474 /* 1475 * Insert the given page into the given object at the given pindex. mpred is 1476 * used for memq linkage. From vm_page_insert, lookup is true, mpred is 1477 * initially NULL, and this procedure looks it up. From vm_page_insert_after 1478 * and vm_page_iter_insert, lookup is false and mpred is known to the caller 1479 * to be valid, and may be NULL if this will be the page with the lowest 1480 * pindex. 1481 * 1482 * The procedure is marked __always_inline to suggest to the compiler to 1483 * eliminate the lookup parameter and the associated alternate branch. 1484 */ 1485 static __always_inline int 1486 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1487 struct pctrie_iter *pages, bool iter, vm_page_t mpred, bool lookup) 1488 { 1489 int error; 1490 1491 VM_OBJECT_ASSERT_WLOCKED(object); 1492 KASSERT(m->object == NULL, 1493 ("vm_page_insert: page %p already inserted", m)); 1494 1495 /* 1496 * Record the object/offset pair in this page. 1497 */ 1498 m->object = object; 1499 m->pindex = pindex; 1500 m->ref_count |= VPRC_OBJREF; 1501 1502 /* 1503 * Add this page to the object's radix tree, and look up mpred if 1504 * needed. 1505 */ 1506 if (iter) { 1507 KASSERT(!lookup, ("%s: cannot lookup mpred", __func__)); 1508 error = vm_radix_iter_insert(pages, m); 1509 } else if (lookup) 1510 error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred); 1511 else 1512 error = vm_radix_insert(&object->rtree, m); 1513 if (__predict_false(error != 0)) { 1514 m->object = NULL; 1515 m->pindex = 0; 1516 m->ref_count &= ~VPRC_OBJREF; 1517 return (1); 1518 } 1519 1520 /* 1521 * Now link into the object's ordered list of backed pages. 1522 */ 1523 vm_page_insert_radixdone(m, object, mpred); 1524 vm_pager_page_inserted(object, m); 1525 return (0); 1526 } 1527 1528 /* 1529 * vm_page_insert: [ internal use only ] 1530 * 1531 * Inserts the given mem entry into the object and object list. 1532 * 1533 * The object must be locked. 1534 */ 1535 int 1536 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1537 { 1538 return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL, 1539 true)); 1540 } 1541 1542 /* 1543 * vm_page_insert_after: 1544 * 1545 * Inserts the page "m" into the specified object at offset "pindex". 1546 * 1547 * The page "mpred" must immediately precede the offset "pindex" within 1548 * the specified object. 1549 * 1550 * The object must be locked. 1551 */ 1552 static int 1553 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1554 vm_page_t mpred) 1555 { 1556 return (vm_page_insert_lookup(m, object, pindex, NULL, false, mpred, 1557 false)); 1558 } 1559 1560 /* 1561 * vm_page_iter_insert: 1562 * 1563 * Tries to insert the page "m" into the specified object at offset 1564 * "pindex" using the iterator "pages". Returns 0 if the insertion was 1565 * successful. 1566 * 1567 * The page "mpred" must immediately precede the offset "pindex" within 1568 * the specified object. 1569 * 1570 * The object must be locked. 1571 */ 1572 static int 1573 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object, 1574 vm_pindex_t pindex, vm_page_t mpred) 1575 { 1576 return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred, 1577 false)); 1578 } 1579 1580 /* 1581 * vm_page_insert_radixdone: 1582 * 1583 * Complete page "m" insertion into the specified object after the 1584 * radix trie hooking. 1585 * 1586 * The page "mpred" must precede the offset "m->pindex" within the 1587 * specified object. 1588 * 1589 * The object must be locked. 1590 */ 1591 static void 1592 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1593 { 1594 1595 VM_OBJECT_ASSERT_WLOCKED(object); 1596 KASSERT(object != NULL && m->object == object, 1597 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1598 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1599 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1600 if (mpred != NULL) { 1601 KASSERT(mpred->object == object, 1602 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1603 KASSERT(mpred->pindex < m->pindex, 1604 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1605 KASSERT(TAILQ_NEXT(mpred, listq) == NULL || 1606 m->pindex < TAILQ_NEXT(mpred, listq)->pindex, 1607 ("vm_page_insert_radixdone: pindex doesn't precede msucc")); 1608 } else { 1609 KASSERT(TAILQ_EMPTY(&object->memq) || 1610 m->pindex < TAILQ_FIRST(&object->memq)->pindex, 1611 ("vm_page_insert_radixdone: no mpred but not first page")); 1612 } 1613 1614 if (mpred != NULL) 1615 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1616 else 1617 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1618 1619 /* 1620 * Show that the object has one more resident page. 1621 */ 1622 object->resident_page_count++; 1623 1624 /* 1625 * Hold the vnode until the last page is released. 1626 */ 1627 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1628 vhold(object->handle); 1629 1630 /* 1631 * Since we are inserting a new and possibly dirty page, 1632 * update the object's generation count. 1633 */ 1634 if (pmap_page_is_write_mapped(m)) 1635 vm_object_set_writeable_dirty(object); 1636 } 1637 1638 /* 1639 * vm_page_remove_radixdone 1640 * 1641 * Complete page "m" removal from the specified object after the radix trie 1642 * unhooking. 1643 * 1644 * The caller is responsible for updating the page's fields to reflect this 1645 * removal. 1646 */ 1647 static void 1648 vm_page_remove_radixdone(vm_page_t m) 1649 { 1650 vm_object_t object; 1651 1652 vm_page_assert_xbusied(m); 1653 object = m->object; 1654 VM_OBJECT_ASSERT_WLOCKED(object); 1655 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1656 ("page %p is missing its object ref", m)); 1657 1658 /* Deferred free of swap space. */ 1659 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1660 vm_pager_page_unswapped(m); 1661 1662 vm_pager_page_removed(object, m); 1663 m->object = NULL; 1664 1665 /* 1666 * Now remove from the object's list of backed pages. 1667 */ 1668 TAILQ_REMOVE(&object->memq, m, listq); 1669 1670 /* 1671 * And show that the object has one fewer resident page. 1672 */ 1673 object->resident_page_count--; 1674 1675 /* 1676 * The vnode may now be recycled. 1677 */ 1678 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1679 vdrop(object->handle); 1680 } 1681 1682 /* 1683 * vm_page_free_object_prep: 1684 * 1685 * Disassociates the given page from its VM object. 1686 * 1687 * The object must be locked, and the page must be xbusy. 1688 */ 1689 static void 1690 vm_page_free_object_prep(vm_page_t m) 1691 { 1692 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 1693 ((m->object->flags & OBJ_UNMANAGED) != 0), 1694 ("%s: managed flag mismatch for page %p", 1695 __func__, m)); 1696 vm_page_assert_xbusied(m); 1697 1698 /* 1699 * The object reference can be released without an atomic 1700 * operation. 1701 */ 1702 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 1703 m->ref_count == VPRC_OBJREF, 1704 ("%s: page %p has unexpected ref_count %u", 1705 __func__, m, m->ref_count)); 1706 vm_page_remove_radixdone(m); 1707 m->ref_count -= VPRC_OBJREF; 1708 } 1709 1710 /* 1711 * vm_page_iter_free: 1712 * 1713 * Free the given page, and use the iterator to remove it from the radix 1714 * tree. 1715 */ 1716 void 1717 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m) 1718 { 1719 vm_radix_iter_remove(pages); 1720 vm_page_free_object_prep(m); 1721 vm_page_xunbusy(m); 1722 m->flags &= ~PG_ZERO; 1723 vm_page_free_toq(m); 1724 } 1725 1726 /* 1727 * vm_page_remove: 1728 * 1729 * Removes the specified page from its containing object, but does not 1730 * invalidate any backing storage. Returns true if the object's reference 1731 * was the last reference to the page, and false otherwise. 1732 * 1733 * The object must be locked and the page must be exclusively busied. 1734 * The exclusive busy will be released on return. If this is not the 1735 * final ref and the caller does not hold a wire reference it may not 1736 * continue to access the page. 1737 */ 1738 bool 1739 vm_page_remove(vm_page_t m) 1740 { 1741 bool dropped; 1742 1743 dropped = vm_page_remove_xbusy(m); 1744 vm_page_xunbusy(m); 1745 1746 return (dropped); 1747 } 1748 1749 /* 1750 * vm_page_iter_remove: 1751 * 1752 * Remove the current page, and use the iterator to remove it from the 1753 * radix tree. 1754 */ 1755 bool 1756 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m) 1757 { 1758 bool dropped; 1759 1760 vm_radix_iter_remove(pages); 1761 vm_page_remove_radixdone(m); 1762 dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1763 vm_page_xunbusy(m); 1764 1765 return (dropped); 1766 } 1767 1768 /* 1769 * vm_page_radix_remove 1770 * 1771 * Removes the specified page from the radix tree. 1772 */ 1773 static void 1774 vm_page_radix_remove(vm_page_t m) 1775 { 1776 vm_page_t mrem __diagused; 1777 1778 mrem = vm_radix_remove(&m->object->rtree, m->pindex); 1779 KASSERT(mrem == m, 1780 ("removed page %p, expected page %p", mrem, m)); 1781 } 1782 1783 /* 1784 * vm_page_remove_xbusy 1785 * 1786 * Removes the page but leaves the xbusy held. Returns true if this 1787 * removed the final ref and false otherwise. 1788 */ 1789 bool 1790 vm_page_remove_xbusy(vm_page_t m) 1791 { 1792 1793 vm_page_radix_remove(m); 1794 vm_page_remove_radixdone(m); 1795 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1796 } 1797 1798 /* 1799 * vm_page_lookup: 1800 * 1801 * Returns the page associated with the object/offset 1802 * pair specified; if none is found, NULL is returned. 1803 * 1804 * The object must be locked. 1805 */ 1806 vm_page_t 1807 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1808 { 1809 1810 VM_OBJECT_ASSERT_LOCKED(object); 1811 return (vm_radix_lookup(&object->rtree, pindex)); 1812 } 1813 1814 /* 1815 * vm_page_iter_init: 1816 * 1817 * Initialize iterator for vm pages. 1818 */ 1819 void 1820 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object) 1821 { 1822 1823 vm_radix_iter_init(pages, &object->rtree); 1824 } 1825 1826 /* 1827 * vm_page_iter_init: 1828 * 1829 * Initialize iterator for vm pages. 1830 */ 1831 void 1832 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object, 1833 vm_pindex_t limit) 1834 { 1835 1836 vm_radix_iter_limit_init(pages, &object->rtree, limit); 1837 } 1838 1839 /* 1840 * vm_page_lookup_unlocked: 1841 * 1842 * Returns the page associated with the object/offset pair specified; 1843 * if none is found, NULL is returned. The page may be no longer be 1844 * present in the object at the time that this function returns. Only 1845 * useful for opportunistic checks such as inmem(). 1846 */ 1847 vm_page_t 1848 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1849 { 1850 1851 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1852 } 1853 1854 /* 1855 * vm_page_relookup: 1856 * 1857 * Returns a page that must already have been busied by 1858 * the caller. Used for bogus page replacement. 1859 */ 1860 vm_page_t 1861 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1862 { 1863 vm_page_t m; 1864 1865 m = vm_page_lookup_unlocked(object, pindex); 1866 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1867 m->object == object && m->pindex == pindex, 1868 ("vm_page_relookup: Invalid page %p", m)); 1869 return (m); 1870 } 1871 1872 /* 1873 * This should only be used by lockless functions for releasing transient 1874 * incorrect acquires. The page may have been freed after we acquired a 1875 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1876 * further to do. 1877 */ 1878 static void 1879 vm_page_busy_release(vm_page_t m) 1880 { 1881 u_int x; 1882 1883 x = vm_page_busy_fetch(m); 1884 for (;;) { 1885 if (x == VPB_FREED) 1886 break; 1887 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1888 if (atomic_fcmpset_int(&m->busy_lock, &x, 1889 x - VPB_ONE_SHARER)) 1890 break; 1891 continue; 1892 } 1893 KASSERT((x & VPB_BIT_SHARED) != 0 || 1894 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1895 ("vm_page_busy_release: %p xbusy not owned.", m)); 1896 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1897 continue; 1898 if ((x & VPB_BIT_WAITERS) != 0) 1899 wakeup(m); 1900 break; 1901 } 1902 } 1903 1904 /* 1905 * vm_page_find_least: 1906 * 1907 * Returns the page associated with the object with least pindex 1908 * greater than or equal to the parameter pindex, or NULL. 1909 * 1910 * The object must be locked. 1911 */ 1912 vm_page_t 1913 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1914 { 1915 vm_page_t m; 1916 1917 VM_OBJECT_ASSERT_LOCKED(object); 1918 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1919 m = vm_radix_lookup_ge(&object->rtree, pindex); 1920 return (m); 1921 } 1922 1923 /* 1924 * Returns the given page's successor (by pindex) within the object if it is 1925 * resident; if none is found, NULL is returned. 1926 * 1927 * The object must be locked. 1928 */ 1929 vm_page_t 1930 vm_page_next(vm_page_t m) 1931 { 1932 vm_page_t next; 1933 1934 VM_OBJECT_ASSERT_LOCKED(m->object); 1935 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1936 MPASS(next->object == m->object); 1937 if (next->pindex != m->pindex + 1) 1938 next = NULL; 1939 } 1940 return (next); 1941 } 1942 1943 /* 1944 * Returns the given page's predecessor (by pindex) within the object if it is 1945 * resident; if none is found, NULL is returned. 1946 * 1947 * The object must be locked. 1948 */ 1949 vm_page_t 1950 vm_page_prev(vm_page_t m) 1951 { 1952 vm_page_t prev; 1953 1954 VM_OBJECT_ASSERT_LOCKED(m->object); 1955 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1956 MPASS(prev->object == m->object); 1957 if (prev->pindex != m->pindex - 1) 1958 prev = NULL; 1959 } 1960 return (prev); 1961 } 1962 1963 /* 1964 * Uses the page mnew as a replacement for an existing page at index 1965 * pindex which must be already present in the object. 1966 * 1967 * Both pages must be exclusively busied on enter. The old page is 1968 * unbusied on exit. 1969 * 1970 * A return value of true means mold is now free. If this is not the 1971 * final ref and the caller does not hold a wire reference it may not 1972 * continue to access the page. 1973 */ 1974 static bool 1975 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1976 vm_page_t mold) 1977 { 1978 vm_page_t mret __diagused; 1979 bool dropped; 1980 1981 VM_OBJECT_ASSERT_WLOCKED(object); 1982 vm_page_assert_xbusied(mold); 1983 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1984 ("vm_page_replace: page %p already in object", mnew)); 1985 1986 /* 1987 * This function mostly follows vm_page_insert() and 1988 * vm_page_remove() without the radix, object count and vnode 1989 * dance. Double check such functions for more comments. 1990 */ 1991 1992 mnew->object = object; 1993 mnew->pindex = pindex; 1994 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1995 mret = vm_radix_replace(&object->rtree, mnew); 1996 KASSERT(mret == mold, 1997 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1998 KASSERT((mold->oflags & VPO_UNMANAGED) == 1999 (mnew->oflags & VPO_UNMANAGED), 2000 ("vm_page_replace: mismatched VPO_UNMANAGED")); 2001 2002 /* Keep the resident page list in sorted order. */ 2003 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 2004 TAILQ_REMOVE(&object->memq, mold, listq); 2005 mold->object = NULL; 2006 2007 /* 2008 * The object's resident_page_count does not change because we have 2009 * swapped one page for another, but the generation count should 2010 * change if the page is dirty. 2011 */ 2012 if (pmap_page_is_write_mapped(mnew)) 2013 vm_object_set_writeable_dirty(object); 2014 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 2015 vm_page_xunbusy(mold); 2016 2017 return (dropped); 2018 } 2019 2020 void 2021 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 2022 vm_page_t mold) 2023 { 2024 2025 vm_page_assert_xbusied(mnew); 2026 2027 if (vm_page_replace_hold(mnew, object, pindex, mold)) 2028 vm_page_free(mold); 2029 } 2030 2031 /* 2032 * vm_page_iter_rename: 2033 * 2034 * Tries to move the specified page from its current object to a new object 2035 * and pindex, using the given iterator to remove the page from its current 2036 * object. Returns true if the move was successful, and false if the move 2037 * was aborted due to a failed memory allocation. 2038 * 2039 * Panics if a page already resides in the new object at the new pindex. 2040 * 2041 * Note: swap associated with the page must be invalidated by the move. We 2042 * have to do this for several reasons: (1) we aren't freeing the 2043 * page, (2) we are dirtying the page, (3) the VM system is probably 2044 * moving the page from object A to B, and will then later move 2045 * the backing store from A to B and we can't have a conflict. 2046 * 2047 * Note: we *always* dirty the page. It is necessary both for the 2048 * fact that we moved it, and because we may be invalidating 2049 * swap. 2050 * 2051 * The objects must be locked. 2052 */ 2053 bool 2054 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m, 2055 vm_object_t new_object, vm_pindex_t new_pindex) 2056 { 2057 vm_page_t mpred; 2058 vm_pindex_t opidx; 2059 2060 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 2061 ("%s: page %p is missing object ref", __func__, m)); 2062 VM_OBJECT_ASSERT_WLOCKED(m->object); 2063 VM_OBJECT_ASSERT_WLOCKED(new_object); 2064 2065 /* 2066 * Create a custom version of vm_page_insert() which does not depend 2067 * by m_prev and can cheat on the implementation aspects of the 2068 * function. 2069 */ 2070 opidx = m->pindex; 2071 m->pindex = new_pindex; 2072 if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) { 2073 m->pindex = opidx; 2074 return (false); 2075 } 2076 2077 /* 2078 * The operation cannot fail anymore. The removal must happen before 2079 * the listq iterator is tainted. 2080 */ 2081 m->pindex = opidx; 2082 vm_radix_iter_remove(old_pages); 2083 vm_page_remove_radixdone(m); 2084 2085 /* Return back to the new pindex to complete vm_page_insert(). */ 2086 m->pindex = new_pindex; 2087 m->object = new_object; 2088 2089 vm_page_insert_radixdone(m, new_object, mpred); 2090 vm_page_dirty(m); 2091 vm_pager_page_inserted(new_object, m); 2092 return (true); 2093 } 2094 2095 /* 2096 * vm_page_mpred: 2097 * 2098 * Return the greatest page of the object with index <= pindex, 2099 * or NULL, if there is none. Assumes object lock is held. 2100 */ 2101 vm_page_t 2102 vm_page_mpred(vm_object_t object, vm_pindex_t pindex) 2103 { 2104 return (vm_radix_lookup_le(&object->rtree, pindex)); 2105 } 2106 2107 /* 2108 * vm_page_alloc: 2109 * 2110 * Allocate and return a page that is associated with the specified 2111 * object and offset pair. By default, this page is exclusive busied. 2112 * 2113 * The caller must always specify an allocation class. 2114 * 2115 * allocation classes: 2116 * VM_ALLOC_NORMAL normal process request 2117 * VM_ALLOC_SYSTEM system *really* needs a page 2118 * VM_ALLOC_INTERRUPT interrupt time request 2119 * 2120 * optional allocation flags: 2121 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2122 * intends to allocate 2123 * VM_ALLOC_NOBUSY do not exclusive busy the page 2124 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2125 * VM_ALLOC_SBUSY shared busy the allocated page 2126 * VM_ALLOC_WIRED wire the allocated page 2127 * VM_ALLOC_ZERO prefer a zeroed page 2128 */ 2129 vm_page_t 2130 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 2131 { 2132 2133 return (vm_page_alloc_after(object, pindex, req, 2134 vm_page_mpred(object, pindex))); 2135 } 2136 2137 /* 2138 * Allocate a page in the specified object with the given page index. To 2139 * optimize insertion of the page into the object, the caller must also specify 2140 * the resident page in the object with largest index smaller than the given 2141 * page index, or NULL if no such page exists. 2142 */ 2143 vm_page_t 2144 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 2145 int req, vm_page_t mpred) 2146 { 2147 struct vm_domainset_iter di; 2148 vm_page_t m; 2149 int domain; 2150 2151 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2152 do { 2153 m = vm_page_alloc_domain_after(object, pindex, domain, req, 2154 mpred); 2155 if (m != NULL) 2156 break; 2157 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2158 2159 return (m); 2160 } 2161 2162 /* 2163 * Returns true if the number of free pages exceeds the minimum 2164 * for the request class and false otherwise. 2165 */ 2166 static int 2167 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2168 { 2169 u_int limit, old, new; 2170 2171 if (req_class == VM_ALLOC_INTERRUPT) 2172 limit = 0; 2173 else if (req_class == VM_ALLOC_SYSTEM) 2174 limit = vmd->vmd_interrupt_free_min; 2175 else 2176 limit = vmd->vmd_free_reserved; 2177 2178 /* 2179 * Attempt to reserve the pages. Fail if we're below the limit. 2180 */ 2181 limit += npages; 2182 old = atomic_load_int(&vmd->vmd_free_count); 2183 do { 2184 if (old < limit) 2185 return (0); 2186 new = old - npages; 2187 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2188 2189 /* Wake the page daemon if we've crossed the threshold. */ 2190 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2191 pagedaemon_wakeup(vmd->vmd_domain); 2192 2193 /* Only update bitsets on transitions. */ 2194 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2195 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2196 vm_domain_set(vmd); 2197 2198 return (1); 2199 } 2200 2201 int 2202 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2203 { 2204 int req_class; 2205 2206 /* 2207 * The page daemon is allowed to dig deeper into the free page list. 2208 */ 2209 req_class = req & VM_ALLOC_CLASS_MASK; 2210 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2211 req_class = VM_ALLOC_SYSTEM; 2212 return (_vm_domain_allocate(vmd, req_class, npages)); 2213 } 2214 2215 vm_page_t 2216 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 2217 int req, vm_page_t mpred) 2218 { 2219 struct vm_domain *vmd; 2220 vm_page_t m; 2221 int flags; 2222 2223 #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2224 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ 2225 VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ 2226 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | \ 2227 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK) 2228 KASSERT((req & ~VPA_FLAGS) == 0, 2229 ("invalid request %#x", req)); 2230 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2231 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2232 ("invalid request %#x", req)); 2233 KASSERT(mpred == NULL || mpred->pindex < pindex, 2234 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2235 (uintmax_t)pindex)); 2236 VM_OBJECT_ASSERT_WLOCKED(object); 2237 2238 flags = 0; 2239 m = NULL; 2240 if (!vm_pager_can_alloc_page(object, pindex)) 2241 return (NULL); 2242 again: 2243 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2244 m = vm_page_alloc_nofree_domain(domain, req); 2245 if (m != NULL) 2246 goto found; 2247 } 2248 #if VM_NRESERVLEVEL > 0 2249 /* 2250 * Can we allocate the page from a reservation? 2251 */ 2252 if (vm_object_reserv(object) && 2253 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 2254 NULL) { 2255 goto found; 2256 } 2257 #endif 2258 vmd = VM_DOMAIN(domain); 2259 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2260 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2261 M_NOWAIT | M_NOVM); 2262 if (m != NULL) { 2263 flags |= PG_PCPU_CACHE; 2264 goto found; 2265 } 2266 } 2267 if (vm_domain_allocate(vmd, req, 1)) { 2268 /* 2269 * If not, allocate it from the free page queues. 2270 */ 2271 vm_domain_free_lock(vmd); 2272 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2273 vm_domain_free_unlock(vmd); 2274 if (m == NULL) { 2275 vm_domain_freecnt_inc(vmd, 1); 2276 #if VM_NRESERVLEVEL > 0 2277 if (vm_reserv_reclaim_inactive(domain)) 2278 goto again; 2279 #endif 2280 } 2281 } 2282 if (m == NULL) { 2283 /* 2284 * Not allocatable, give up. 2285 */ 2286 if (vm_domain_alloc_fail(vmd, object, req)) 2287 goto again; 2288 return (NULL); 2289 } 2290 2291 /* 2292 * At this point we had better have found a good page. 2293 */ 2294 found: 2295 vm_page_dequeue(m); 2296 vm_page_alloc_check(m); 2297 2298 /* 2299 * Initialize the page. Only the PG_ZERO flag is inherited. 2300 */ 2301 flags |= m->flags & PG_ZERO; 2302 if ((req & VM_ALLOC_NODUMP) != 0) 2303 flags |= PG_NODUMP; 2304 if ((req & VM_ALLOC_NOFREE) != 0) 2305 flags |= PG_NOFREE; 2306 m->flags = flags; 2307 m->a.flags = 0; 2308 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2309 m->pool = VM_FREEPOOL_DEFAULT; 2310 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2311 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2312 else if ((req & VM_ALLOC_SBUSY) != 0) 2313 m->busy_lock = VPB_SHARERS_WORD(1); 2314 else 2315 m->busy_lock = VPB_UNBUSIED; 2316 if (req & VM_ALLOC_WIRED) { 2317 vm_wire_add(1); 2318 m->ref_count = 1; 2319 } 2320 m->a.act_count = 0; 2321 2322 if (vm_page_insert_after(m, object, pindex, mpred)) { 2323 if (req & VM_ALLOC_WIRED) { 2324 vm_wire_sub(1); 2325 m->ref_count = 0; 2326 } 2327 KASSERT(m->object == NULL, ("page %p has object", m)); 2328 m->oflags = VPO_UNMANAGED; 2329 m->busy_lock = VPB_UNBUSIED; 2330 /* Don't change PG_ZERO. */ 2331 vm_page_free_toq(m); 2332 if (req & VM_ALLOC_WAITFAIL) { 2333 VM_OBJECT_WUNLOCK(object); 2334 vm_radix_wait(); 2335 VM_OBJECT_WLOCK(object); 2336 } 2337 return (NULL); 2338 } 2339 2340 /* Ignore device objects; the pager sets "memattr" for them. */ 2341 if (object->memattr != VM_MEMATTR_DEFAULT && 2342 (object->flags & OBJ_FICTITIOUS) == 0) 2343 pmap_page_set_memattr(m, object->memattr); 2344 2345 return (m); 2346 } 2347 2348 /* 2349 * vm_page_alloc_contig: 2350 * 2351 * Allocate a contiguous set of physical pages of the given size "npages" 2352 * from the free lists. All of the physical pages must be at or above 2353 * the given physical address "low" and below the given physical address 2354 * "high". The given value "alignment" determines the alignment of the 2355 * first physical page in the set. If the given value "boundary" is 2356 * non-zero, then the set of physical pages cannot cross any physical 2357 * address boundary that is a multiple of that value. Both "alignment" 2358 * and "boundary" must be a power of two. 2359 * 2360 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2361 * then the memory attribute setting for the physical pages is configured 2362 * to the object's memory attribute setting. Otherwise, the memory 2363 * attribute setting for the physical pages is configured to "memattr", 2364 * overriding the object's memory attribute setting. However, if the 2365 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2366 * memory attribute setting for the physical pages cannot be configured 2367 * to VM_MEMATTR_DEFAULT. 2368 * 2369 * The specified object may not contain fictitious pages. 2370 * 2371 * The caller must always specify an allocation class. 2372 * 2373 * allocation classes: 2374 * VM_ALLOC_NORMAL normal process request 2375 * VM_ALLOC_SYSTEM system *really* needs a page 2376 * VM_ALLOC_INTERRUPT interrupt time request 2377 * 2378 * optional allocation flags: 2379 * VM_ALLOC_NOBUSY do not exclusive busy the page 2380 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2381 * VM_ALLOC_SBUSY shared busy the allocated page 2382 * VM_ALLOC_WIRED wire the allocated page 2383 * VM_ALLOC_ZERO prefer a zeroed page 2384 */ 2385 vm_page_t 2386 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2387 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2388 vm_paddr_t boundary, vm_memattr_t memattr) 2389 { 2390 struct vm_domainset_iter di; 2391 vm_page_t bounds[2]; 2392 vm_page_t m; 2393 int domain; 2394 int start_segind; 2395 2396 start_segind = -1; 2397 2398 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2399 do { 2400 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2401 npages, low, high, alignment, boundary, memattr); 2402 if (m != NULL) 2403 break; 2404 if (start_segind == -1) 2405 start_segind = vm_phys_lookup_segind(low); 2406 if (vm_phys_find_range(bounds, start_segind, domain, 2407 npages, low, high) == -1) { 2408 vm_domainset_iter_ignore(&di, domain); 2409 } 2410 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2411 2412 return (m); 2413 } 2414 2415 static vm_page_t 2416 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2417 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2418 { 2419 struct vm_domain *vmd; 2420 vm_page_t m_ret; 2421 2422 /* 2423 * Can we allocate the pages without the number of free pages falling 2424 * below the lower bound for the allocation class? 2425 */ 2426 vmd = VM_DOMAIN(domain); 2427 if (!vm_domain_allocate(vmd, req, npages)) 2428 return (NULL); 2429 /* 2430 * Try to allocate the pages from the free page queues. 2431 */ 2432 vm_domain_free_lock(vmd); 2433 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2434 alignment, boundary); 2435 vm_domain_free_unlock(vmd); 2436 if (m_ret != NULL) 2437 return (m_ret); 2438 #if VM_NRESERVLEVEL > 0 2439 /* 2440 * Try to break a reservation to allocate the pages. 2441 */ 2442 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2443 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2444 high, alignment, boundary); 2445 if (m_ret != NULL) 2446 return (m_ret); 2447 } 2448 #endif 2449 vm_domain_freecnt_inc(vmd, npages); 2450 return (NULL); 2451 } 2452 2453 vm_page_t 2454 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2455 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2456 vm_paddr_t boundary, vm_memattr_t memattr) 2457 { 2458 struct pctrie_iter pages; 2459 vm_page_t m, m_ret, mpred; 2460 u_int busy_lock, flags, oflags; 2461 2462 #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) 2463 KASSERT((req & ~VPAC_FLAGS) == 0, 2464 ("invalid request %#x", req)); 2465 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2466 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2467 ("invalid request %#x", req)); 2468 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2469 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2470 ("invalid request %#x", req)); 2471 VM_OBJECT_ASSERT_WLOCKED(object); 2472 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2473 ("vm_page_alloc_contig: object %p has fictitious pages", 2474 object)); 2475 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2476 2477 vm_page_iter_init(&pages, object); 2478 mpred = vm_radix_iter_lookup_le(&pages, pindex); 2479 KASSERT(mpred == NULL || mpred->pindex != pindex, 2480 ("vm_page_alloc_contig: pindex already allocated")); 2481 for (;;) { 2482 #if VM_NRESERVLEVEL > 0 2483 /* 2484 * Can we allocate the pages from a reservation? 2485 */ 2486 if (vm_object_reserv(object) && 2487 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2488 mpred, npages, low, high, alignment, boundary)) != NULL) { 2489 break; 2490 } 2491 #endif 2492 if ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2493 low, high, alignment, boundary)) != NULL) 2494 break; 2495 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req)) 2496 return (NULL); 2497 } 2498 2499 /* 2500 * Initialize the pages. Only the PG_ZERO flag is inherited. 2501 */ 2502 flags = PG_ZERO; 2503 if ((req & VM_ALLOC_NODUMP) != 0) 2504 flags |= PG_NODUMP; 2505 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2506 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2507 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2508 else if ((req & VM_ALLOC_SBUSY) != 0) 2509 busy_lock = VPB_SHARERS_WORD(1); 2510 else 2511 busy_lock = VPB_UNBUSIED; 2512 if ((req & VM_ALLOC_WIRED) != 0) 2513 vm_wire_add(npages); 2514 if (object->memattr != VM_MEMATTR_DEFAULT && 2515 memattr == VM_MEMATTR_DEFAULT) 2516 memattr = object->memattr; 2517 for (m = m_ret; m < &m_ret[npages]; m++) { 2518 vm_page_dequeue(m); 2519 vm_page_alloc_check(m); 2520 m->a.flags = 0; 2521 m->flags = (m->flags | PG_NODUMP) & flags; 2522 m->busy_lock = busy_lock; 2523 if ((req & VM_ALLOC_WIRED) != 0) 2524 m->ref_count = 1; 2525 m->a.act_count = 0; 2526 m->oflags = oflags; 2527 m->pool = VM_FREEPOOL_DEFAULT; 2528 if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) { 2529 if ((req & VM_ALLOC_WIRED) != 0) 2530 vm_wire_sub(npages); 2531 KASSERT(m->object == NULL, 2532 ("page %p has object", m)); 2533 mpred = m; 2534 for (m = m_ret; m < &m_ret[npages]; m++) { 2535 if (m <= mpred && 2536 (req & VM_ALLOC_WIRED) != 0) 2537 m->ref_count = 0; 2538 m->oflags = VPO_UNMANAGED; 2539 m->busy_lock = VPB_UNBUSIED; 2540 /* Don't change PG_ZERO. */ 2541 vm_page_free_toq(m); 2542 } 2543 if (req & VM_ALLOC_WAITFAIL) { 2544 VM_OBJECT_WUNLOCK(object); 2545 vm_radix_wait(); 2546 VM_OBJECT_WLOCK(object); 2547 } 2548 return (NULL); 2549 } 2550 mpred = m; 2551 if (memattr != VM_MEMATTR_DEFAULT) 2552 pmap_page_set_memattr(m, memattr); 2553 pindex++; 2554 } 2555 return (m_ret); 2556 } 2557 2558 /* 2559 * Allocate a physical page that is not intended to be inserted into a VM 2560 * object. 2561 */ 2562 vm_page_t 2563 vm_page_alloc_noobj_domain(int domain, int req) 2564 { 2565 struct vm_domain *vmd; 2566 vm_page_t m; 2567 int flags; 2568 2569 #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2570 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ 2571 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ 2572 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | \ 2573 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK) 2574 KASSERT((req & ~VPAN_FLAGS) == 0, 2575 ("invalid request %#x", req)); 2576 2577 flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) | 2578 ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0); 2579 vmd = VM_DOMAIN(domain); 2580 again: 2581 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2582 m = vm_page_alloc_nofree_domain(domain, req); 2583 if (m != NULL) 2584 goto found; 2585 } 2586 2587 if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2588 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2589 M_NOWAIT | M_NOVM); 2590 if (m != NULL) { 2591 flags |= PG_PCPU_CACHE; 2592 goto found; 2593 } 2594 } 2595 2596 if (vm_domain_allocate(vmd, req, 1)) { 2597 vm_domain_free_lock(vmd); 2598 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2599 vm_domain_free_unlock(vmd); 2600 if (m == NULL) { 2601 vm_domain_freecnt_inc(vmd, 1); 2602 #if VM_NRESERVLEVEL > 0 2603 if (vm_reserv_reclaim_inactive(domain)) 2604 goto again; 2605 #endif 2606 } 2607 } 2608 if (m == NULL) { 2609 if (vm_domain_alloc_fail(vmd, NULL, req)) 2610 goto again; 2611 return (NULL); 2612 } 2613 2614 found: 2615 vm_page_dequeue(m); 2616 vm_page_alloc_check(m); 2617 2618 /* 2619 * Consumers should not rely on a useful default pindex value. 2620 */ 2621 m->pindex = 0xdeadc0dedeadc0de; 2622 m->flags = (m->flags & PG_ZERO) | flags; 2623 m->a.flags = 0; 2624 m->oflags = VPO_UNMANAGED; 2625 m->pool = VM_FREEPOOL_DIRECT; 2626 m->busy_lock = VPB_UNBUSIED; 2627 if ((req & VM_ALLOC_WIRED) != 0) { 2628 vm_wire_add(1); 2629 m->ref_count = 1; 2630 } 2631 2632 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2633 pmap_zero_page(m); 2634 2635 return (m); 2636 } 2637 2638 #if VM_NRESERVLEVEL > 1 2639 #define VM_NOFREE_IMPORT_ORDER (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER) 2640 #elif VM_NRESERVLEVEL > 0 2641 #define VM_NOFREE_IMPORT_ORDER VM_LEVEL_0_ORDER 2642 #else 2643 #define VM_NOFREE_IMPORT_ORDER 8 2644 #endif 2645 2646 /* 2647 * Allocate a single NOFREE page. 2648 * 2649 * This routine hands out NOFREE pages from higher-order 2650 * physical memory blocks in order to reduce memory fragmentation. 2651 * When a NOFREE for a given domain chunk is used up, 2652 * the routine will try to fetch a new one from the freelists 2653 * and discard the old one. 2654 */ 2655 static vm_page_t 2656 vm_page_alloc_nofree_domain(int domain, int req) 2657 { 2658 vm_page_t m; 2659 struct vm_domain *vmd; 2660 struct vm_nofreeq *nqp; 2661 2662 KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req)); 2663 2664 vmd = VM_DOMAIN(domain); 2665 nqp = &vmd->vmd_nofreeq; 2666 vm_domain_free_lock(vmd); 2667 if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) { 2668 if (!vm_domain_allocate(vmd, req, 2669 1 << VM_NOFREE_IMPORT_ORDER)) { 2670 vm_domain_free_unlock(vmd); 2671 return (NULL); 2672 } 2673 nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 2674 VM_NOFREE_IMPORT_ORDER); 2675 if (nqp->ma == NULL) { 2676 vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER); 2677 vm_domain_free_unlock(vmd); 2678 return (NULL); 2679 } 2680 nqp->offs = 0; 2681 } 2682 m = &nqp->ma[nqp->offs++]; 2683 vm_domain_free_unlock(vmd); 2684 VM_CNT_ADD(v_nofree_count, 1); 2685 2686 return (m); 2687 } 2688 2689 vm_page_t 2690 vm_page_alloc_noobj(int req) 2691 { 2692 struct vm_domainset_iter di; 2693 vm_page_t m; 2694 int domain; 2695 2696 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2697 do { 2698 m = vm_page_alloc_noobj_domain(domain, req); 2699 if (m != NULL) 2700 break; 2701 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2702 2703 return (m); 2704 } 2705 2706 vm_page_t 2707 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2708 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2709 vm_memattr_t memattr) 2710 { 2711 struct vm_domainset_iter di; 2712 vm_page_t m; 2713 int domain; 2714 2715 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2716 do { 2717 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2718 high, alignment, boundary, memattr); 2719 if (m != NULL) 2720 break; 2721 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2722 2723 return (m); 2724 } 2725 2726 vm_page_t 2727 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2728 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2729 vm_memattr_t memattr) 2730 { 2731 vm_page_t m, m_ret; 2732 u_int flags; 2733 2734 #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) 2735 KASSERT((req & ~VPANC_FLAGS) == 0, 2736 ("invalid request %#x", req)); 2737 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2738 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2739 ("invalid request %#x", req)); 2740 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2741 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2742 ("invalid request %#x", req)); 2743 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2744 2745 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2746 low, high, alignment, boundary)) == NULL) { 2747 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2748 return (NULL); 2749 } 2750 2751 /* 2752 * Initialize the pages. Only the PG_ZERO flag is inherited. 2753 */ 2754 flags = PG_ZERO; 2755 if ((req & VM_ALLOC_NODUMP) != 0) 2756 flags |= PG_NODUMP; 2757 if ((req & VM_ALLOC_WIRED) != 0) 2758 vm_wire_add(npages); 2759 for (m = m_ret; m < &m_ret[npages]; m++) { 2760 vm_page_dequeue(m); 2761 vm_page_alloc_check(m); 2762 2763 /* 2764 * Consumers should not rely on a useful default pindex value. 2765 */ 2766 m->pindex = 0xdeadc0dedeadc0de; 2767 m->a.flags = 0; 2768 m->flags = (m->flags | PG_NODUMP) & flags; 2769 m->busy_lock = VPB_UNBUSIED; 2770 if ((req & VM_ALLOC_WIRED) != 0) 2771 m->ref_count = 1; 2772 m->a.act_count = 0; 2773 m->oflags = VPO_UNMANAGED; 2774 m->pool = VM_FREEPOOL_DIRECT; 2775 2776 /* 2777 * Zero the page before updating any mappings since the page is 2778 * not yet shared with any devices which might require the 2779 * non-default memory attribute. pmap_page_set_memattr() 2780 * flushes data caches before returning. 2781 */ 2782 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2783 pmap_zero_page(m); 2784 if (memattr != VM_MEMATTR_DEFAULT) 2785 pmap_page_set_memattr(m, memattr); 2786 } 2787 return (m_ret); 2788 } 2789 2790 /* 2791 * Check a page that has been freshly dequeued from a freelist. 2792 */ 2793 static void 2794 vm_page_alloc_check(vm_page_t m) 2795 { 2796 2797 KASSERT(m->object == NULL, ("page %p has object", m)); 2798 KASSERT(m->a.queue == PQ_NONE && 2799 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2800 ("page %p has unexpected queue %d, flags %#x", 2801 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2802 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2803 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2804 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2805 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2806 ("page %p has unexpected memattr %d", 2807 m, pmap_page_get_memattr(m))); 2808 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2809 pmap_vm_page_alloc_check(m); 2810 } 2811 2812 static int 2813 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2814 { 2815 struct vm_domain *vmd; 2816 struct vm_pgcache *pgcache; 2817 int i; 2818 2819 pgcache = arg; 2820 vmd = VM_DOMAIN(pgcache->domain); 2821 2822 /* 2823 * The page daemon should avoid creating extra memory pressure since its 2824 * main purpose is to replenish the store of free pages. 2825 */ 2826 if (vmd->vmd_severeset || curproc == pageproc || 2827 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2828 return (0); 2829 domain = vmd->vmd_domain; 2830 vm_domain_free_lock(vmd); 2831 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2832 (vm_page_t *)store); 2833 vm_domain_free_unlock(vmd); 2834 if (cnt != i) 2835 vm_domain_freecnt_inc(vmd, cnt - i); 2836 2837 return (i); 2838 } 2839 2840 static void 2841 vm_page_zone_release(void *arg, void **store, int cnt) 2842 { 2843 struct vm_domain *vmd; 2844 struct vm_pgcache *pgcache; 2845 vm_page_t m; 2846 int i; 2847 2848 pgcache = arg; 2849 vmd = VM_DOMAIN(pgcache->domain); 2850 vm_domain_free_lock(vmd); 2851 for (i = 0; i < cnt; i++) { 2852 m = (vm_page_t)store[i]; 2853 vm_phys_free_pages(m, pgcache->pool, 0); 2854 } 2855 vm_domain_free_unlock(vmd); 2856 vm_domain_freecnt_inc(vmd, cnt); 2857 } 2858 2859 #define VPSC_ANY 0 /* No restrictions. */ 2860 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2861 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2862 2863 /* 2864 * vm_page_scan_contig: 2865 * 2866 * Scan vm_page_array[] between the specified entries "m_start" and 2867 * "m_end" for a run of contiguous physical pages that satisfy the 2868 * specified conditions, and return the lowest page in the run. The 2869 * specified "alignment" determines the alignment of the lowest physical 2870 * page in the run. If the specified "boundary" is non-zero, then the 2871 * run of physical pages cannot span a physical address that is a 2872 * multiple of "boundary". 2873 * 2874 * "m_end" is never dereferenced, so it need not point to a vm_page 2875 * structure within vm_page_array[]. 2876 * 2877 * "npages" must be greater than zero. "m_start" and "m_end" must not 2878 * span a hole (or discontiguity) in the physical address space. Both 2879 * "alignment" and "boundary" must be a power of two. 2880 */ 2881 static vm_page_t 2882 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2883 u_long alignment, vm_paddr_t boundary, int options) 2884 { 2885 vm_object_t object; 2886 vm_paddr_t pa; 2887 vm_page_t m, m_run; 2888 #if VM_NRESERVLEVEL > 0 2889 int level; 2890 #endif 2891 int m_inc, order, run_ext, run_len; 2892 2893 KASSERT(npages > 0, ("npages is 0")); 2894 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2895 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2896 m_run = NULL; 2897 run_len = 0; 2898 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2899 KASSERT((m->flags & PG_MARKER) == 0, 2900 ("page %p is PG_MARKER", m)); 2901 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2902 ("fictitious page %p has invalid ref count", m)); 2903 2904 /* 2905 * If the current page would be the start of a run, check its 2906 * physical address against the end, alignment, and boundary 2907 * conditions. If it doesn't satisfy these conditions, either 2908 * terminate the scan or advance to the next page that 2909 * satisfies the failed condition. 2910 */ 2911 if (run_len == 0) { 2912 KASSERT(m_run == NULL, ("m_run != NULL")); 2913 if (m + npages > m_end) 2914 break; 2915 pa = VM_PAGE_TO_PHYS(m); 2916 if (!vm_addr_align_ok(pa, alignment)) { 2917 m_inc = atop(roundup2(pa, alignment) - pa); 2918 continue; 2919 } 2920 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2921 m_inc = atop(roundup2(pa, boundary) - pa); 2922 continue; 2923 } 2924 } else 2925 KASSERT(m_run != NULL, ("m_run == NULL")); 2926 2927 retry: 2928 m_inc = 1; 2929 if (vm_page_wired(m)) 2930 run_ext = 0; 2931 #if VM_NRESERVLEVEL > 0 2932 else if ((level = vm_reserv_level(m)) >= 0 && 2933 (options & VPSC_NORESERV) != 0) { 2934 run_ext = 0; 2935 /* Advance to the end of the reservation. */ 2936 pa = VM_PAGE_TO_PHYS(m); 2937 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2938 pa); 2939 } 2940 #endif 2941 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2942 /* 2943 * The page is considered eligible for relocation if 2944 * and only if it could be laundered or reclaimed by 2945 * the page daemon. 2946 */ 2947 VM_OBJECT_RLOCK(object); 2948 if (object != m->object) { 2949 VM_OBJECT_RUNLOCK(object); 2950 goto retry; 2951 } 2952 /* Don't care: PG_NODUMP, PG_ZERO. */ 2953 if ((object->flags & OBJ_SWAP) == 0 && 2954 object->type != OBJT_VNODE) { 2955 run_ext = 0; 2956 #if VM_NRESERVLEVEL > 0 2957 } else if ((options & VPSC_NOSUPER) != 0 && 2958 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2959 run_ext = 0; 2960 /* Advance to the end of the superpage. */ 2961 pa = VM_PAGE_TO_PHYS(m); 2962 m_inc = atop(roundup2(pa + 1, 2963 vm_reserv_size(level)) - pa); 2964 #endif 2965 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2966 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2967 /* 2968 * The page is allocated but eligible for 2969 * relocation. Extend the current run by one 2970 * page. 2971 */ 2972 KASSERT(pmap_page_get_memattr(m) == 2973 VM_MEMATTR_DEFAULT, 2974 ("page %p has an unexpected memattr", m)); 2975 KASSERT((m->oflags & (VPO_SWAPINPROG | 2976 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2977 ("page %p has unexpected oflags", m)); 2978 /* Don't care: PGA_NOSYNC. */ 2979 run_ext = 1; 2980 } else 2981 run_ext = 0; 2982 VM_OBJECT_RUNLOCK(object); 2983 #if VM_NRESERVLEVEL > 0 2984 } else if (level >= 0) { 2985 /* 2986 * The page is reserved but not yet allocated. In 2987 * other words, it is still free. Extend the current 2988 * run by one page. 2989 */ 2990 run_ext = 1; 2991 #endif 2992 } else if ((order = m->order) < VM_NFREEORDER) { 2993 /* 2994 * The page is enqueued in the physical memory 2995 * allocator's free page queues. Moreover, it is the 2996 * first page in a power-of-two-sized run of 2997 * contiguous free pages. Add these pages to the end 2998 * of the current run, and jump ahead. 2999 */ 3000 run_ext = 1 << order; 3001 m_inc = 1 << order; 3002 } else { 3003 /* 3004 * Skip the page for one of the following reasons: (1) 3005 * It is enqueued in the physical memory allocator's 3006 * free page queues. However, it is not the first 3007 * page in a run of contiguous free pages. (This case 3008 * rarely occurs because the scan is performed in 3009 * ascending order.) (2) It is not reserved, and it is 3010 * transitioning from free to allocated. (Conversely, 3011 * the transition from allocated to free for managed 3012 * pages is blocked by the page busy lock.) (3) It is 3013 * allocated but not contained by an object and not 3014 * wired, e.g., allocated by Xen's balloon driver. 3015 */ 3016 run_ext = 0; 3017 } 3018 3019 /* 3020 * Extend or reset the current run of pages. 3021 */ 3022 if (run_ext > 0) { 3023 if (run_len == 0) 3024 m_run = m; 3025 run_len += run_ext; 3026 } else { 3027 if (run_len > 0) { 3028 m_run = NULL; 3029 run_len = 0; 3030 } 3031 } 3032 } 3033 if (run_len >= npages) 3034 return (m_run); 3035 return (NULL); 3036 } 3037 3038 /* 3039 * vm_page_reclaim_run: 3040 * 3041 * Try to relocate each of the allocated virtual pages within the 3042 * specified run of physical pages to a new physical address. Free the 3043 * physical pages underlying the relocated virtual pages. A virtual page 3044 * is relocatable if and only if it could be laundered or reclaimed by 3045 * the page daemon. Whenever possible, a virtual page is relocated to a 3046 * physical address above "high". 3047 * 3048 * Returns 0 if every physical page within the run was already free or 3049 * just freed by a successful relocation. Otherwise, returns a non-zero 3050 * value indicating why the last attempt to relocate a virtual page was 3051 * unsuccessful. 3052 * 3053 * "req_class" must be an allocation class. 3054 */ 3055 static int 3056 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 3057 vm_paddr_t high) 3058 { 3059 struct vm_domain *vmd; 3060 struct spglist free; 3061 vm_object_t object; 3062 vm_paddr_t pa; 3063 vm_page_t m, m_end, m_new; 3064 int error, order, req; 3065 3066 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 3067 ("req_class is not an allocation class")); 3068 SLIST_INIT(&free); 3069 error = 0; 3070 m = m_run; 3071 m_end = m_run + npages; 3072 for (; error == 0 && m < m_end; m++) { 3073 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 3074 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 3075 3076 /* 3077 * Racily check for wirings. Races are handled once the object 3078 * lock is held and the page is unmapped. 3079 */ 3080 if (vm_page_wired(m)) 3081 error = EBUSY; 3082 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 3083 /* 3084 * The page is relocated if and only if it could be 3085 * laundered or reclaimed by the page daemon. 3086 */ 3087 VM_OBJECT_WLOCK(object); 3088 /* Don't care: PG_NODUMP, PG_ZERO. */ 3089 if (m->object != object || 3090 ((object->flags & OBJ_SWAP) == 0 && 3091 object->type != OBJT_VNODE)) 3092 error = EINVAL; 3093 else if (object->memattr != VM_MEMATTR_DEFAULT) 3094 error = EINVAL; 3095 else if (vm_page_queue(m) != PQ_NONE && 3096 vm_page_tryxbusy(m) != 0) { 3097 if (vm_page_wired(m)) { 3098 vm_page_xunbusy(m); 3099 error = EBUSY; 3100 goto unlock; 3101 } 3102 KASSERT(pmap_page_get_memattr(m) == 3103 VM_MEMATTR_DEFAULT, 3104 ("page %p has an unexpected memattr", m)); 3105 KASSERT(m->oflags == 0, 3106 ("page %p has unexpected oflags", m)); 3107 /* Don't care: PGA_NOSYNC. */ 3108 if (!vm_page_none_valid(m)) { 3109 /* 3110 * First, try to allocate a new page 3111 * that is above "high". Failing 3112 * that, try to allocate a new page 3113 * that is below "m_run". Allocate 3114 * the new page between the end of 3115 * "m_run" and "high" only as a last 3116 * resort. 3117 */ 3118 req = req_class; 3119 if ((m->flags & PG_NODUMP) != 0) 3120 req |= VM_ALLOC_NODUMP; 3121 if (trunc_page(high) != 3122 ~(vm_paddr_t)PAGE_MASK) { 3123 m_new = 3124 vm_page_alloc_noobj_contig( 3125 req, 1, round_page(high), 3126 ~(vm_paddr_t)0, PAGE_SIZE, 3127 0, VM_MEMATTR_DEFAULT); 3128 } else 3129 m_new = NULL; 3130 if (m_new == NULL) { 3131 pa = VM_PAGE_TO_PHYS(m_run); 3132 m_new = 3133 vm_page_alloc_noobj_contig( 3134 req, 1, 0, pa - 1, 3135 PAGE_SIZE, 0, 3136 VM_MEMATTR_DEFAULT); 3137 } 3138 if (m_new == NULL) { 3139 pa += ptoa(npages); 3140 m_new = 3141 vm_page_alloc_noobj_contig( 3142 req, 1, pa, high, PAGE_SIZE, 3143 0, VM_MEMATTR_DEFAULT); 3144 } 3145 if (m_new == NULL) { 3146 vm_page_xunbusy(m); 3147 error = ENOMEM; 3148 goto unlock; 3149 } 3150 3151 /* 3152 * Unmap the page and check for new 3153 * wirings that may have been acquired 3154 * through a pmap lookup. 3155 */ 3156 if (object->ref_count != 0 && 3157 !vm_page_try_remove_all(m)) { 3158 vm_page_xunbusy(m); 3159 vm_page_free(m_new); 3160 error = EBUSY; 3161 goto unlock; 3162 } 3163 3164 /* 3165 * Replace "m" with the new page. For 3166 * vm_page_replace(), "m" must be busy 3167 * and dequeued. Finally, change "m" 3168 * as if vm_page_free() was called. 3169 */ 3170 m_new->a.flags = m->a.flags & 3171 ~PGA_QUEUE_STATE_MASK; 3172 KASSERT(m_new->oflags == VPO_UNMANAGED, 3173 ("page %p is managed", m_new)); 3174 m_new->oflags = 0; 3175 pmap_copy_page(m, m_new); 3176 m_new->valid = m->valid; 3177 m_new->dirty = m->dirty; 3178 m->flags &= ~PG_ZERO; 3179 vm_page_dequeue(m); 3180 if (vm_page_replace_hold(m_new, object, 3181 m->pindex, m) && 3182 vm_page_free_prep(m)) 3183 SLIST_INSERT_HEAD(&free, m, 3184 plinks.s.ss); 3185 3186 /* 3187 * The new page must be deactivated 3188 * before the object is unlocked. 3189 */ 3190 vm_page_deactivate(m_new); 3191 } else { 3192 m->flags &= ~PG_ZERO; 3193 vm_page_dequeue(m); 3194 if (vm_page_free_prep(m)) 3195 SLIST_INSERT_HEAD(&free, m, 3196 plinks.s.ss); 3197 KASSERT(m->dirty == 0, 3198 ("page %p is dirty", m)); 3199 } 3200 } else 3201 error = EBUSY; 3202 unlock: 3203 VM_OBJECT_WUNLOCK(object); 3204 } else { 3205 MPASS(vm_page_domain(m) == domain); 3206 vmd = VM_DOMAIN(domain); 3207 vm_domain_free_lock(vmd); 3208 order = m->order; 3209 if (order < VM_NFREEORDER) { 3210 /* 3211 * The page is enqueued in the physical memory 3212 * allocator's free page queues. Moreover, it 3213 * is the first page in a power-of-two-sized 3214 * run of contiguous free pages. Jump ahead 3215 * to the last page within that run, and 3216 * continue from there. 3217 */ 3218 m += (1 << order) - 1; 3219 } 3220 #if VM_NRESERVLEVEL > 0 3221 else if (vm_reserv_is_page_free(m)) 3222 order = 0; 3223 #endif 3224 vm_domain_free_unlock(vmd); 3225 if (order == VM_NFREEORDER) 3226 error = EINVAL; 3227 } 3228 } 3229 if ((m = SLIST_FIRST(&free)) != NULL) { 3230 int cnt; 3231 3232 vmd = VM_DOMAIN(domain); 3233 cnt = 0; 3234 vm_domain_free_lock(vmd); 3235 do { 3236 MPASS(vm_page_domain(m) == domain); 3237 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3238 vm_phys_free_pages(m, m->pool, 0); 3239 cnt++; 3240 } while ((m = SLIST_FIRST(&free)) != NULL); 3241 vm_domain_free_unlock(vmd); 3242 vm_domain_freecnt_inc(vmd, cnt); 3243 } 3244 return (error); 3245 } 3246 3247 #define NRUNS 16 3248 3249 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 3250 3251 #define MIN_RECLAIM 8 3252 3253 /* 3254 * vm_page_reclaim_contig: 3255 * 3256 * Reclaim allocated, contiguous physical memory satisfying the specified 3257 * conditions by relocating the virtual pages using that physical memory. 3258 * Returns 0 if reclamation is successful, ERANGE if the specified domain 3259 * can't possibly satisfy the reclamation request, or ENOMEM if not 3260 * currently able to reclaim the requested number of pages. Since 3261 * relocation requires the allocation of physical pages, reclamation may 3262 * fail with ENOMEM due to a shortage of free pages. When reclamation 3263 * fails in this manner, callers are expected to perform vm_wait() before 3264 * retrying a failed allocation operation, e.g., vm_page_alloc_contig(). 3265 * 3266 * The caller must always specify an allocation class through "req". 3267 * 3268 * allocation classes: 3269 * VM_ALLOC_NORMAL normal process request 3270 * VM_ALLOC_SYSTEM system *really* needs a page 3271 * VM_ALLOC_INTERRUPT interrupt time request 3272 * 3273 * The optional allocation flags are ignored. 3274 * 3275 * "npages" must be greater than zero. Both "alignment" and "boundary" 3276 * must be a power of two. 3277 */ 3278 int 3279 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3280 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3281 int desired_runs) 3282 { 3283 struct vm_domain *vmd; 3284 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; 3285 u_long count, minalign, reclaimed; 3286 int error, i, min_reclaim, nruns, options, req_class; 3287 int segind, start_segind; 3288 int ret; 3289 3290 KASSERT(npages > 0, ("npages is 0")); 3291 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3292 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3293 3294 ret = ENOMEM; 3295 3296 /* 3297 * If the caller wants to reclaim multiple runs, try to allocate 3298 * space to store the runs. If that fails, fall back to the old 3299 * behavior of just reclaiming MIN_RECLAIM pages. 3300 */ 3301 if (desired_runs > 1) 3302 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3303 M_TEMP, M_NOWAIT); 3304 else 3305 m_runs = NULL; 3306 3307 if (m_runs == NULL) { 3308 m_runs = _m_runs; 3309 nruns = NRUNS; 3310 } else { 3311 nruns = NRUNS + desired_runs - 1; 3312 } 3313 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3314 3315 /* 3316 * The caller will attempt an allocation after some runs have been 3317 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3318 * of vm_phys_alloc_contig(), round up the requested length to the next 3319 * power of two or maximum chunk size, and ensure that each run is 3320 * suitably aligned. 3321 */ 3322 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3323 npages = roundup2(npages, minalign); 3324 if (alignment < ptoa(minalign)) 3325 alignment = ptoa(minalign); 3326 3327 /* 3328 * The page daemon is allowed to dig deeper into the free page list. 3329 */ 3330 req_class = req & VM_ALLOC_CLASS_MASK; 3331 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3332 req_class = VM_ALLOC_SYSTEM; 3333 3334 start_segind = vm_phys_lookup_segind(low); 3335 3336 /* 3337 * Return if the number of free pages cannot satisfy the requested 3338 * allocation. 3339 */ 3340 vmd = VM_DOMAIN(domain); 3341 count = vmd->vmd_free_count; 3342 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3343 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3344 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3345 goto done; 3346 3347 /* 3348 * Scan up to three times, relaxing the restrictions ("options") on 3349 * the reclamation of reservations and superpages each time. 3350 */ 3351 for (options = VPSC_NORESERV;;) { 3352 bool phys_range_exists = false; 3353 3354 /* 3355 * Find the highest runs that satisfy the given constraints 3356 * and restrictions, and record them in "m_runs". 3357 */ 3358 count = 0; 3359 segind = start_segind; 3360 while ((segind = vm_phys_find_range(bounds, segind, domain, 3361 npages, low, high)) != -1) { 3362 phys_range_exists = true; 3363 while ((m_run = vm_page_scan_contig(npages, bounds[0], 3364 bounds[1], alignment, boundary, options))) { 3365 bounds[0] = m_run + npages; 3366 m_runs[RUN_INDEX(count, nruns)] = m_run; 3367 count++; 3368 } 3369 segind++; 3370 } 3371 3372 if (!phys_range_exists) { 3373 ret = ERANGE; 3374 goto done; 3375 } 3376 3377 /* 3378 * Reclaim the highest runs in LIFO (descending) order until 3379 * the number of reclaimed pages, "reclaimed", is at least 3380 * "min_reclaim". Reset "reclaimed" each time because each 3381 * reclamation is idempotent, and runs will (likely) recur 3382 * from one scan to the next as restrictions are relaxed. 3383 */ 3384 reclaimed = 0; 3385 for (i = 0; count > 0 && i < nruns; i++) { 3386 count--; 3387 m_run = m_runs[RUN_INDEX(count, nruns)]; 3388 error = vm_page_reclaim_run(req_class, domain, npages, 3389 m_run, high); 3390 if (error == 0) { 3391 reclaimed += npages; 3392 if (reclaimed >= min_reclaim) { 3393 ret = 0; 3394 goto done; 3395 } 3396 } 3397 } 3398 3399 /* 3400 * Either relax the restrictions on the next scan or return if 3401 * the last scan had no restrictions. 3402 */ 3403 if (options == VPSC_NORESERV) 3404 options = VPSC_NOSUPER; 3405 else if (options == VPSC_NOSUPER) 3406 options = VPSC_ANY; 3407 else if (options == VPSC_ANY) { 3408 if (reclaimed != 0) 3409 ret = 0; 3410 goto done; 3411 } 3412 } 3413 done: 3414 if (m_runs != _m_runs) 3415 free(m_runs, M_TEMP); 3416 return (ret); 3417 } 3418 3419 int 3420 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3421 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3422 { 3423 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high, 3424 alignment, boundary, 1)); 3425 } 3426 3427 int 3428 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3429 u_long alignment, vm_paddr_t boundary) 3430 { 3431 struct vm_domainset_iter di; 3432 int domain, ret, status; 3433 3434 ret = ERANGE; 3435 3436 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3437 do { 3438 status = vm_page_reclaim_contig_domain(domain, req, npages, low, 3439 high, alignment, boundary); 3440 if (status == 0) 3441 return (0); 3442 else if (status == ERANGE) 3443 vm_domainset_iter_ignore(&di, domain); 3444 else { 3445 KASSERT(status == ENOMEM, ("Unrecognized error %d " 3446 "from vm_page_reclaim_contig_domain()", status)); 3447 ret = ENOMEM; 3448 } 3449 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3450 3451 return (ret); 3452 } 3453 3454 /* 3455 * Set the domain in the appropriate page level domainset. 3456 */ 3457 void 3458 vm_domain_set(struct vm_domain *vmd) 3459 { 3460 3461 mtx_lock(&vm_domainset_lock); 3462 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3463 vmd->vmd_minset = 1; 3464 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3465 } 3466 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3467 vmd->vmd_severeset = 1; 3468 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3469 } 3470 mtx_unlock(&vm_domainset_lock); 3471 } 3472 3473 /* 3474 * Clear the domain from the appropriate page level domainset. 3475 */ 3476 void 3477 vm_domain_clear(struct vm_domain *vmd) 3478 { 3479 3480 mtx_lock(&vm_domainset_lock); 3481 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3482 vmd->vmd_minset = 0; 3483 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3484 if (vm_min_waiters != 0) { 3485 vm_min_waiters = 0; 3486 wakeup(&vm_min_domains); 3487 } 3488 } 3489 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3490 vmd->vmd_severeset = 0; 3491 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3492 if (vm_severe_waiters != 0) { 3493 vm_severe_waiters = 0; 3494 wakeup(&vm_severe_domains); 3495 } 3496 } 3497 3498 /* 3499 * If pageout daemon needs pages, then tell it that there are 3500 * some free. 3501 */ 3502 if (vmd->vmd_pageout_pages_needed && 3503 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3504 wakeup(&vmd->vmd_pageout_pages_needed); 3505 vmd->vmd_pageout_pages_needed = 0; 3506 } 3507 3508 /* See comments in vm_wait_doms(). */ 3509 if (vm_pageproc_waiters) { 3510 vm_pageproc_waiters = 0; 3511 wakeup(&vm_pageproc_waiters); 3512 } 3513 mtx_unlock(&vm_domainset_lock); 3514 } 3515 3516 /* 3517 * Wait for free pages to exceed the min threshold globally. 3518 */ 3519 void 3520 vm_wait_min(void) 3521 { 3522 3523 mtx_lock(&vm_domainset_lock); 3524 while (vm_page_count_min()) { 3525 vm_min_waiters++; 3526 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3527 } 3528 mtx_unlock(&vm_domainset_lock); 3529 } 3530 3531 /* 3532 * Wait for free pages to exceed the severe threshold globally. 3533 */ 3534 void 3535 vm_wait_severe(void) 3536 { 3537 3538 mtx_lock(&vm_domainset_lock); 3539 while (vm_page_count_severe()) { 3540 vm_severe_waiters++; 3541 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3542 "vmwait", 0); 3543 } 3544 mtx_unlock(&vm_domainset_lock); 3545 } 3546 3547 u_int 3548 vm_wait_count(void) 3549 { 3550 3551 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3552 } 3553 3554 int 3555 vm_wait_doms(const domainset_t *wdoms, int mflags) 3556 { 3557 int error; 3558 3559 error = 0; 3560 3561 /* 3562 * We use racey wakeup synchronization to avoid expensive global 3563 * locking for the pageproc when sleeping with a non-specific vm_wait. 3564 * To handle this, we only sleep for one tick in this instance. It 3565 * is expected that most allocations for the pageproc will come from 3566 * kmem or vm_page_grab* which will use the more specific and 3567 * race-free vm_wait_domain(). 3568 */ 3569 if (curproc == pageproc) { 3570 mtx_lock(&vm_domainset_lock); 3571 vm_pageproc_waiters++; 3572 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3573 PVM | PDROP | mflags, "pageprocwait", 1); 3574 } else { 3575 /* 3576 * XXX Ideally we would wait only until the allocation could 3577 * be satisfied. This condition can cause new allocators to 3578 * consume all freed pages while old allocators wait. 3579 */ 3580 mtx_lock(&vm_domainset_lock); 3581 if (vm_page_count_min_set(wdoms)) { 3582 if (pageproc == NULL) 3583 panic("vm_wait in early boot"); 3584 vm_min_waiters++; 3585 error = msleep(&vm_min_domains, &vm_domainset_lock, 3586 PVM | PDROP | mflags, "vmwait", 0); 3587 } else 3588 mtx_unlock(&vm_domainset_lock); 3589 } 3590 return (error); 3591 } 3592 3593 /* 3594 * vm_wait_domain: 3595 * 3596 * Sleep until free pages are available for allocation. 3597 * - Called in various places after failed memory allocations. 3598 */ 3599 void 3600 vm_wait_domain(int domain) 3601 { 3602 struct vm_domain *vmd; 3603 domainset_t wdom; 3604 3605 vmd = VM_DOMAIN(domain); 3606 vm_domain_free_assert_unlocked(vmd); 3607 3608 if (curproc == pageproc) { 3609 mtx_lock(&vm_domainset_lock); 3610 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3611 vmd->vmd_pageout_pages_needed = 1; 3612 msleep(&vmd->vmd_pageout_pages_needed, 3613 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3614 } else 3615 mtx_unlock(&vm_domainset_lock); 3616 } else { 3617 DOMAINSET_ZERO(&wdom); 3618 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3619 vm_wait_doms(&wdom, 0); 3620 } 3621 } 3622 3623 static int 3624 vm_wait_flags(vm_object_t obj, int mflags) 3625 { 3626 struct domainset *d; 3627 3628 d = NULL; 3629 3630 /* 3631 * Carefully fetch pointers only once: the struct domainset 3632 * itself is ummutable but the pointer might change. 3633 */ 3634 if (obj != NULL) 3635 d = obj->domain.dr_policy; 3636 if (d == NULL) 3637 d = curthread->td_domain.dr_policy; 3638 3639 return (vm_wait_doms(&d->ds_mask, mflags)); 3640 } 3641 3642 /* 3643 * vm_wait: 3644 * 3645 * Sleep until free pages are available for allocation in the 3646 * affinity domains of the obj. If obj is NULL, the domain set 3647 * for the calling thread is used. 3648 * Called in various places after failed memory allocations. 3649 */ 3650 void 3651 vm_wait(vm_object_t obj) 3652 { 3653 (void)vm_wait_flags(obj, 0); 3654 } 3655 3656 int 3657 vm_wait_intr(vm_object_t obj) 3658 { 3659 return (vm_wait_flags(obj, PCATCH)); 3660 } 3661 3662 /* 3663 * vm_domain_alloc_fail: 3664 * 3665 * Called when a page allocation function fails. Informs the 3666 * pagedaemon and performs the requested wait. Requires the 3667 * domain_free and object lock on entry. Returns with the 3668 * object lock held and free lock released. Returns an error when 3669 * retry is necessary. 3670 * 3671 */ 3672 static int 3673 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3674 { 3675 3676 vm_domain_free_assert_unlocked(vmd); 3677 3678 atomic_add_int(&vmd->vmd_pageout_deficit, 3679 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3680 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3681 if (object != NULL) 3682 VM_OBJECT_WUNLOCK(object); 3683 vm_wait_domain(vmd->vmd_domain); 3684 if (object != NULL) 3685 VM_OBJECT_WLOCK(object); 3686 if (req & VM_ALLOC_WAITOK) 3687 return (EAGAIN); 3688 } 3689 3690 return (0); 3691 } 3692 3693 /* 3694 * vm_waitpfault: 3695 * 3696 * Sleep until free pages are available for allocation. 3697 * - Called only in vm_fault so that processes page faulting 3698 * can be easily tracked. 3699 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3700 * processes will be able to grab memory first. Do not change 3701 * this balance without careful testing first. 3702 */ 3703 void 3704 vm_waitpfault(struct domainset *dset, int timo) 3705 { 3706 3707 /* 3708 * XXX Ideally we would wait only until the allocation could 3709 * be satisfied. This condition can cause new allocators to 3710 * consume all freed pages while old allocators wait. 3711 */ 3712 mtx_lock(&vm_domainset_lock); 3713 if (vm_page_count_min_set(&dset->ds_mask)) { 3714 vm_min_waiters++; 3715 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3716 "pfault", timo); 3717 } else 3718 mtx_unlock(&vm_domainset_lock); 3719 } 3720 3721 static struct vm_pagequeue * 3722 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3723 { 3724 3725 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3726 } 3727 3728 #ifdef INVARIANTS 3729 static struct vm_pagequeue * 3730 vm_page_pagequeue(vm_page_t m) 3731 { 3732 3733 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3734 } 3735 #endif 3736 3737 static __always_inline bool 3738 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3739 { 3740 vm_page_astate_t tmp; 3741 3742 tmp = *old; 3743 do { 3744 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3745 return (true); 3746 counter_u64_add(pqstate_commit_retries, 1); 3747 } while (old->_bits == tmp._bits); 3748 3749 return (false); 3750 } 3751 3752 /* 3753 * Do the work of committing a queue state update that moves the page out of 3754 * its current queue. 3755 */ 3756 static bool 3757 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3758 vm_page_astate_t *old, vm_page_astate_t new) 3759 { 3760 vm_page_t next; 3761 3762 vm_pagequeue_assert_locked(pq); 3763 KASSERT(vm_page_pagequeue(m) == pq, 3764 ("%s: queue %p does not match page %p", __func__, pq, m)); 3765 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3766 ("%s: invalid queue indices %d %d", 3767 __func__, old->queue, new.queue)); 3768 3769 /* 3770 * Once the queue index of the page changes there is nothing 3771 * synchronizing with further updates to the page's physical 3772 * queue state. Therefore we must speculatively remove the page 3773 * from the queue now and be prepared to roll back if the queue 3774 * state update fails. If the page is not physically enqueued then 3775 * we just update its queue index. 3776 */ 3777 if ((old->flags & PGA_ENQUEUED) != 0) { 3778 new.flags &= ~PGA_ENQUEUED; 3779 next = TAILQ_NEXT(m, plinks.q); 3780 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3781 vm_pagequeue_cnt_dec(pq); 3782 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3783 if (next == NULL) 3784 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3785 else 3786 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3787 vm_pagequeue_cnt_inc(pq); 3788 return (false); 3789 } else { 3790 return (true); 3791 } 3792 } else { 3793 return (vm_page_pqstate_fcmpset(m, old, new)); 3794 } 3795 } 3796 3797 static bool 3798 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3799 vm_page_astate_t new) 3800 { 3801 struct vm_pagequeue *pq; 3802 vm_page_astate_t as; 3803 bool ret; 3804 3805 pq = _vm_page_pagequeue(m, old->queue); 3806 3807 /* 3808 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3809 * corresponding page queue lock is held. 3810 */ 3811 vm_pagequeue_lock(pq); 3812 as = vm_page_astate_load(m); 3813 if (__predict_false(as._bits != old->_bits)) { 3814 *old = as; 3815 ret = false; 3816 } else { 3817 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3818 } 3819 vm_pagequeue_unlock(pq); 3820 return (ret); 3821 } 3822 3823 /* 3824 * Commit a queue state update that enqueues or requeues a page. 3825 */ 3826 static bool 3827 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3828 vm_page_astate_t *old, vm_page_astate_t new) 3829 { 3830 struct vm_domain *vmd; 3831 3832 vm_pagequeue_assert_locked(pq); 3833 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3834 ("%s: invalid queue indices %d %d", 3835 __func__, old->queue, new.queue)); 3836 3837 new.flags |= PGA_ENQUEUED; 3838 if (!vm_page_pqstate_fcmpset(m, old, new)) 3839 return (false); 3840 3841 if ((old->flags & PGA_ENQUEUED) != 0) 3842 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3843 else 3844 vm_pagequeue_cnt_inc(pq); 3845 3846 /* 3847 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3848 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3849 * applied, even if it was set first. 3850 */ 3851 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3852 vmd = vm_pagequeue_domain(m); 3853 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3854 ("%s: invalid page queue for page %p", __func__, m)); 3855 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3856 } else { 3857 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3858 } 3859 return (true); 3860 } 3861 3862 /* 3863 * Commit a queue state update that encodes a request for a deferred queue 3864 * operation. 3865 */ 3866 static bool 3867 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3868 vm_page_astate_t new) 3869 { 3870 3871 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3872 ("%s: invalid state, queue %d flags %x", 3873 __func__, new.queue, new.flags)); 3874 3875 if (old->_bits != new._bits && 3876 !vm_page_pqstate_fcmpset(m, old, new)) 3877 return (false); 3878 vm_page_pqbatch_submit(m, new.queue); 3879 return (true); 3880 } 3881 3882 /* 3883 * A generic queue state update function. This handles more cases than the 3884 * specialized functions above. 3885 */ 3886 bool 3887 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3888 { 3889 3890 if (old->_bits == new._bits) 3891 return (true); 3892 3893 if (old->queue != PQ_NONE && new.queue != old->queue) { 3894 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3895 return (false); 3896 if (new.queue != PQ_NONE) 3897 vm_page_pqbatch_submit(m, new.queue); 3898 } else { 3899 if (!vm_page_pqstate_fcmpset(m, old, new)) 3900 return (false); 3901 if (new.queue != PQ_NONE && 3902 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3903 vm_page_pqbatch_submit(m, new.queue); 3904 } 3905 return (true); 3906 } 3907 3908 /* 3909 * Apply deferred queue state updates to a page. 3910 */ 3911 static inline void 3912 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3913 { 3914 vm_page_astate_t new, old; 3915 3916 CRITICAL_ASSERT(curthread); 3917 vm_pagequeue_assert_locked(pq); 3918 KASSERT(queue < PQ_COUNT, 3919 ("%s: invalid queue index %d", __func__, queue)); 3920 KASSERT(pq == _vm_page_pagequeue(m, queue), 3921 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3922 3923 for (old = vm_page_astate_load(m);;) { 3924 if (__predict_false(old.queue != queue || 3925 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3926 counter_u64_add(queue_nops, 1); 3927 break; 3928 } 3929 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3930 ("%s: page %p is unmanaged", __func__, m)); 3931 3932 new = old; 3933 if ((old.flags & PGA_DEQUEUE) != 0) { 3934 new.flags &= ~PGA_QUEUE_OP_MASK; 3935 new.queue = PQ_NONE; 3936 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3937 m, &old, new))) { 3938 counter_u64_add(queue_ops, 1); 3939 break; 3940 } 3941 } else { 3942 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3943 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3944 m, &old, new))) { 3945 counter_u64_add(queue_ops, 1); 3946 break; 3947 } 3948 } 3949 } 3950 } 3951 3952 static void 3953 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3954 uint8_t queue) 3955 { 3956 int i; 3957 3958 for (i = 0; i < bq->bq_cnt; i++) 3959 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3960 vm_batchqueue_init(bq); 3961 } 3962 3963 /* 3964 * vm_page_pqbatch_submit: [ internal use only ] 3965 * 3966 * Enqueue a page in the specified page queue's batched work queue. 3967 * The caller must have encoded the requested operation in the page 3968 * structure's a.flags field. 3969 */ 3970 void 3971 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3972 { 3973 struct vm_batchqueue *bq; 3974 struct vm_pagequeue *pq; 3975 int domain, slots_remaining; 3976 3977 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3978 3979 domain = vm_page_domain(m); 3980 critical_enter(); 3981 bq = DPCPU_PTR(pqbatch[domain][queue]); 3982 slots_remaining = vm_batchqueue_insert(bq, m); 3983 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 3984 /* keep building the bq */ 3985 critical_exit(); 3986 return; 3987 } else if (slots_remaining > 0 ) { 3988 /* Try to process the bq if we can get the lock */ 3989 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3990 if (vm_pagequeue_trylock(pq)) { 3991 vm_pqbatch_process(pq, bq, queue); 3992 vm_pagequeue_unlock(pq); 3993 } 3994 critical_exit(); 3995 return; 3996 } 3997 critical_exit(); 3998 3999 /* if we make it here, the bq is full so wait for the lock */ 4000 4001 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 4002 vm_pagequeue_lock(pq); 4003 critical_enter(); 4004 bq = DPCPU_PTR(pqbatch[domain][queue]); 4005 vm_pqbatch_process(pq, bq, queue); 4006 vm_pqbatch_process_page(pq, m, queue); 4007 vm_pagequeue_unlock(pq); 4008 critical_exit(); 4009 } 4010 4011 /* 4012 * vm_page_pqbatch_drain: [ internal use only ] 4013 * 4014 * Force all per-CPU page queue batch queues to be drained. This is 4015 * intended for use in severe memory shortages, to ensure that pages 4016 * do not remain stuck in the batch queues. 4017 */ 4018 void 4019 vm_page_pqbatch_drain(void) 4020 { 4021 struct thread *td; 4022 struct vm_domain *vmd; 4023 struct vm_pagequeue *pq; 4024 int cpu, domain, queue; 4025 4026 td = curthread; 4027 CPU_FOREACH(cpu) { 4028 thread_lock(td); 4029 sched_bind(td, cpu); 4030 thread_unlock(td); 4031 4032 for (domain = 0; domain < vm_ndomains; domain++) { 4033 vmd = VM_DOMAIN(domain); 4034 for (queue = 0; queue < PQ_COUNT; queue++) { 4035 pq = &vmd->vmd_pagequeues[queue]; 4036 vm_pagequeue_lock(pq); 4037 critical_enter(); 4038 vm_pqbatch_process(pq, 4039 DPCPU_PTR(pqbatch[domain][queue]), queue); 4040 critical_exit(); 4041 vm_pagequeue_unlock(pq); 4042 } 4043 } 4044 } 4045 thread_lock(td); 4046 sched_unbind(td); 4047 thread_unlock(td); 4048 } 4049 4050 /* 4051 * vm_page_dequeue_deferred: [ internal use only ] 4052 * 4053 * Request removal of the given page from its current page 4054 * queue. Physical removal from the queue may be deferred 4055 * indefinitely. 4056 */ 4057 void 4058 vm_page_dequeue_deferred(vm_page_t m) 4059 { 4060 vm_page_astate_t new, old; 4061 4062 old = vm_page_astate_load(m); 4063 do { 4064 if (old.queue == PQ_NONE) { 4065 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4066 ("%s: page %p has unexpected queue state", 4067 __func__, m)); 4068 break; 4069 } 4070 new = old; 4071 new.flags |= PGA_DEQUEUE; 4072 } while (!vm_page_pqstate_commit_request(m, &old, new)); 4073 } 4074 4075 /* 4076 * vm_page_dequeue: 4077 * 4078 * Remove the page from whichever page queue it's in, if any, before 4079 * returning. 4080 */ 4081 void 4082 vm_page_dequeue(vm_page_t m) 4083 { 4084 vm_page_astate_t new, old; 4085 4086 old = vm_page_astate_load(m); 4087 do { 4088 if (old.queue == PQ_NONE) { 4089 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4090 ("%s: page %p has unexpected queue state", 4091 __func__, m)); 4092 break; 4093 } 4094 new = old; 4095 new.flags &= ~PGA_QUEUE_OP_MASK; 4096 new.queue = PQ_NONE; 4097 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 4098 4099 } 4100 4101 /* 4102 * Schedule the given page for insertion into the specified page queue. 4103 * Physical insertion of the page may be deferred indefinitely. 4104 */ 4105 static void 4106 vm_page_enqueue(vm_page_t m, uint8_t queue) 4107 { 4108 4109 KASSERT(m->a.queue == PQ_NONE && 4110 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 4111 ("%s: page %p is already enqueued", __func__, m)); 4112 KASSERT(m->ref_count > 0, 4113 ("%s: page %p does not carry any references", __func__, m)); 4114 4115 m->a.queue = queue; 4116 if ((m->a.flags & PGA_REQUEUE) == 0) 4117 vm_page_aflag_set(m, PGA_REQUEUE); 4118 vm_page_pqbatch_submit(m, queue); 4119 } 4120 4121 /* 4122 * vm_page_free_prep: 4123 * 4124 * Prepares the given page to be put on the free list, 4125 * disassociating it from any VM object. The caller may return 4126 * the page to the free list only if this function returns true. 4127 * 4128 * The object, if it exists, must be locked, and then the page must 4129 * be xbusy. Otherwise the page must be not busied. A managed 4130 * page must be unmapped. 4131 */ 4132 static bool 4133 vm_page_free_prep(vm_page_t m) 4134 { 4135 4136 /* 4137 * Synchronize with threads that have dropped a reference to this 4138 * page. 4139 */ 4140 atomic_thread_fence_acq(); 4141 4142 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 4143 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 4144 uint64_t *p; 4145 int i; 4146 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 4147 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 4148 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 4149 m, i, (uintmax_t)*p)); 4150 } 4151 #endif 4152 KASSERT((m->flags & PG_NOFREE) == 0, 4153 ("%s: attempting to free a PG_NOFREE page", __func__)); 4154 if ((m->oflags & VPO_UNMANAGED) == 0) { 4155 KASSERT(!pmap_page_is_mapped(m), 4156 ("vm_page_free_prep: freeing mapped page %p", m)); 4157 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 4158 ("vm_page_free_prep: mapping flags set in page %p", m)); 4159 } else { 4160 KASSERT(m->a.queue == PQ_NONE, 4161 ("vm_page_free_prep: unmanaged page %p is queued", m)); 4162 } 4163 VM_CNT_INC(v_tfree); 4164 4165 if (m->object != NULL) { 4166 vm_page_radix_remove(m); 4167 vm_page_free_object_prep(m); 4168 } else 4169 vm_page_assert_unbusied(m); 4170 4171 vm_page_busy_free(m); 4172 4173 /* 4174 * If fictitious remove object association and 4175 * return. 4176 */ 4177 if ((m->flags & PG_FICTITIOUS) != 0) { 4178 KASSERT(m->ref_count == 1, 4179 ("fictitious page %p is referenced", m)); 4180 KASSERT(m->a.queue == PQ_NONE, 4181 ("fictitious page %p is queued", m)); 4182 return (false); 4183 } 4184 4185 /* 4186 * Pages need not be dequeued before they are returned to the physical 4187 * memory allocator, but they must at least be marked for a deferred 4188 * dequeue. 4189 */ 4190 if ((m->oflags & VPO_UNMANAGED) == 0) 4191 vm_page_dequeue_deferred(m); 4192 4193 m->valid = 0; 4194 vm_page_undirty(m); 4195 4196 if (m->ref_count != 0) 4197 panic("vm_page_free_prep: page %p has references", m); 4198 4199 /* 4200 * Restore the default memory attribute to the page. 4201 */ 4202 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 4203 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 4204 4205 #if VM_NRESERVLEVEL > 0 4206 /* 4207 * Determine whether the page belongs to a reservation. If the page was 4208 * allocated from a per-CPU cache, it cannot belong to a reservation, so 4209 * as an optimization, we avoid the check in that case. 4210 */ 4211 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 4212 return (false); 4213 #endif 4214 4215 return (true); 4216 } 4217 4218 /* 4219 * vm_page_free_toq: 4220 * 4221 * Returns the given page to the free list, disassociating it 4222 * from any VM object. 4223 * 4224 * The object must be locked. The page must be exclusively busied if it 4225 * belongs to an object. 4226 */ 4227 static void 4228 vm_page_free_toq(vm_page_t m) 4229 { 4230 struct vm_domain *vmd; 4231 uma_zone_t zone; 4232 4233 if (!vm_page_free_prep(m)) 4234 return; 4235 4236 vmd = vm_pagequeue_domain(m); 4237 zone = vmd->vmd_pgcache[m->pool].zone; 4238 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4239 uma_zfree(zone, m); 4240 return; 4241 } 4242 vm_domain_free_lock(vmd); 4243 vm_phys_free_pages(m, m->pool, 0); 4244 vm_domain_free_unlock(vmd); 4245 vm_domain_freecnt_inc(vmd, 1); 4246 } 4247 4248 /* 4249 * vm_page_free_pages_toq: 4250 * 4251 * Returns a list of pages to the free list, disassociating it 4252 * from any VM object. In other words, this is equivalent to 4253 * calling vm_page_free_toq() for each page of a list of VM objects. 4254 */ 4255 int 4256 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4257 { 4258 vm_page_t m; 4259 int count; 4260 4261 if (SLIST_EMPTY(free)) 4262 return (0); 4263 4264 count = 0; 4265 while ((m = SLIST_FIRST(free)) != NULL) { 4266 count++; 4267 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4268 vm_page_free_toq(m); 4269 } 4270 4271 if (update_wire_count) 4272 vm_wire_sub(count); 4273 return (count); 4274 } 4275 4276 /* 4277 * Mark this page as wired down. For managed pages, this prevents reclamation 4278 * by the page daemon, or when the containing object, if any, is destroyed. 4279 */ 4280 void 4281 vm_page_wire(vm_page_t m) 4282 { 4283 u_int old; 4284 4285 #ifdef INVARIANTS 4286 if (m->object != NULL && !vm_page_busied(m) && 4287 !vm_object_busied(m->object)) 4288 VM_OBJECT_ASSERT_LOCKED(m->object); 4289 #endif 4290 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4291 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4292 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4293 4294 old = atomic_fetchadd_int(&m->ref_count, 1); 4295 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4296 ("vm_page_wire: counter overflow for page %p", m)); 4297 if (VPRC_WIRE_COUNT(old) == 0) { 4298 if ((m->oflags & VPO_UNMANAGED) == 0) 4299 vm_page_aflag_set(m, PGA_DEQUEUE); 4300 vm_wire_add(1); 4301 } 4302 } 4303 4304 /* 4305 * Attempt to wire a mapped page following a pmap lookup of that page. 4306 * This may fail if a thread is concurrently tearing down mappings of the page. 4307 * The transient failure is acceptable because it translates to the 4308 * failure of the caller pmap_extract_and_hold(), which should be then 4309 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4310 */ 4311 bool 4312 vm_page_wire_mapped(vm_page_t m) 4313 { 4314 u_int old; 4315 4316 old = atomic_load_int(&m->ref_count); 4317 do { 4318 KASSERT(old > 0, 4319 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4320 if ((old & VPRC_BLOCKED) != 0) 4321 return (false); 4322 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4323 4324 if (VPRC_WIRE_COUNT(old) == 0) { 4325 if ((m->oflags & VPO_UNMANAGED) == 0) 4326 vm_page_aflag_set(m, PGA_DEQUEUE); 4327 vm_wire_add(1); 4328 } 4329 return (true); 4330 } 4331 4332 /* 4333 * Release a wiring reference to a managed page. If the page still belongs to 4334 * an object, update its position in the page queues to reflect the reference. 4335 * If the wiring was the last reference to the page, free the page. 4336 */ 4337 static void 4338 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4339 { 4340 u_int old; 4341 4342 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4343 ("%s: page %p is unmanaged", __func__, m)); 4344 4345 /* 4346 * Update LRU state before releasing the wiring reference. 4347 * Use a release store when updating the reference count to 4348 * synchronize with vm_page_free_prep(). 4349 */ 4350 old = atomic_load_int(&m->ref_count); 4351 do { 4352 u_int count; 4353 4354 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4355 ("vm_page_unwire: wire count underflow for page %p", m)); 4356 4357 count = old & ~VPRC_BLOCKED; 4358 if (count > VPRC_OBJREF + 1) { 4359 /* 4360 * The page has at least one other wiring reference. An 4361 * earlier iteration of this loop may have called 4362 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4363 * re-set it if necessary. 4364 */ 4365 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4366 vm_page_aflag_set(m, PGA_DEQUEUE); 4367 } else if (count == VPRC_OBJREF + 1) { 4368 /* 4369 * This is the last wiring. Clear PGA_DEQUEUE and 4370 * update the page's queue state to reflect the 4371 * reference. If the page does not belong to an object 4372 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4373 * clear leftover queue state. 4374 */ 4375 vm_page_release_toq(m, nqueue, noreuse); 4376 } else if (count == 1) { 4377 vm_page_aflag_clear(m, PGA_DEQUEUE); 4378 } 4379 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4380 4381 if (VPRC_WIRE_COUNT(old) == 1) { 4382 vm_wire_sub(1); 4383 if (old == 1) 4384 vm_page_free(m); 4385 } 4386 } 4387 4388 /* 4389 * Release one wiring of the specified page, potentially allowing it to be 4390 * paged out. 4391 * 4392 * Only managed pages belonging to an object can be paged out. If the number 4393 * of wirings transitions to zero and the page is eligible for page out, then 4394 * the page is added to the specified paging queue. If the released wiring 4395 * represented the last reference to the page, the page is freed. 4396 */ 4397 void 4398 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4399 { 4400 4401 KASSERT(nqueue < PQ_COUNT, 4402 ("vm_page_unwire: invalid queue %u request for page %p", 4403 nqueue, m)); 4404 4405 if ((m->oflags & VPO_UNMANAGED) != 0) { 4406 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4407 vm_page_free(m); 4408 return; 4409 } 4410 vm_page_unwire_managed(m, nqueue, false); 4411 } 4412 4413 /* 4414 * Unwire a page without (re-)inserting it into a page queue. It is up 4415 * to the caller to enqueue, requeue, or free the page as appropriate. 4416 * In most cases involving managed pages, vm_page_unwire() should be used 4417 * instead. 4418 */ 4419 bool 4420 vm_page_unwire_noq(vm_page_t m) 4421 { 4422 u_int old; 4423 4424 old = vm_page_drop(m, 1); 4425 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4426 ("%s: counter underflow for page %p", __func__, m)); 4427 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4428 ("%s: missing ref on fictitious page %p", __func__, m)); 4429 4430 if (VPRC_WIRE_COUNT(old) > 1) 4431 return (false); 4432 if ((m->oflags & VPO_UNMANAGED) == 0) 4433 vm_page_aflag_clear(m, PGA_DEQUEUE); 4434 vm_wire_sub(1); 4435 return (true); 4436 } 4437 4438 /* 4439 * Ensure that the page ends up in the specified page queue. If the page is 4440 * active or being moved to the active queue, ensure that its act_count is 4441 * at least ACT_INIT but do not otherwise mess with it. 4442 */ 4443 static __always_inline void 4444 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4445 { 4446 vm_page_astate_t old, new; 4447 4448 KASSERT(m->ref_count > 0, 4449 ("%s: page %p does not carry any references", __func__, m)); 4450 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4451 ("%s: invalid flags %x", __func__, nflag)); 4452 4453 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4454 return; 4455 4456 old = vm_page_astate_load(m); 4457 do { 4458 if ((old.flags & PGA_DEQUEUE) != 0) 4459 break; 4460 new = old; 4461 new.flags &= ~PGA_QUEUE_OP_MASK; 4462 if (nqueue == PQ_ACTIVE) 4463 new.act_count = max(old.act_count, ACT_INIT); 4464 if (old.queue == nqueue) { 4465 /* 4466 * There is no need to requeue pages already in the 4467 * active queue. 4468 */ 4469 if (nqueue != PQ_ACTIVE || 4470 (old.flags & PGA_ENQUEUED) == 0) 4471 new.flags |= nflag; 4472 } else { 4473 new.flags |= nflag; 4474 new.queue = nqueue; 4475 } 4476 } while (!vm_page_pqstate_commit(m, &old, new)); 4477 } 4478 4479 /* 4480 * Put the specified page on the active list (if appropriate). 4481 */ 4482 void 4483 vm_page_activate(vm_page_t m) 4484 { 4485 4486 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4487 } 4488 4489 /* 4490 * Move the specified page to the tail of the inactive queue, or requeue 4491 * the page if it is already in the inactive queue. 4492 */ 4493 void 4494 vm_page_deactivate(vm_page_t m) 4495 { 4496 4497 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4498 } 4499 4500 void 4501 vm_page_deactivate_noreuse(vm_page_t m) 4502 { 4503 4504 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4505 } 4506 4507 /* 4508 * Put a page in the laundry, or requeue it if it is already there. 4509 */ 4510 void 4511 vm_page_launder(vm_page_t m) 4512 { 4513 4514 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4515 } 4516 4517 /* 4518 * Put a page in the PQ_UNSWAPPABLE holding queue. 4519 */ 4520 void 4521 vm_page_unswappable(vm_page_t m) 4522 { 4523 4524 VM_OBJECT_ASSERT_LOCKED(m->object); 4525 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4526 ("page %p already unswappable", m)); 4527 4528 vm_page_dequeue(m); 4529 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4530 } 4531 4532 /* 4533 * Release a page back to the page queues in preparation for unwiring. 4534 */ 4535 static void 4536 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4537 { 4538 vm_page_astate_t old, new; 4539 uint16_t nflag; 4540 4541 /* 4542 * Use a check of the valid bits to determine whether we should 4543 * accelerate reclamation of the page. The object lock might not be 4544 * held here, in which case the check is racy. At worst we will either 4545 * accelerate reclamation of a valid page and violate LRU, or 4546 * unnecessarily defer reclamation of an invalid page. 4547 * 4548 * If we were asked to not cache the page, place it near the head of the 4549 * inactive queue so that is reclaimed sooner. 4550 */ 4551 if (noreuse || vm_page_none_valid(m)) { 4552 nqueue = PQ_INACTIVE; 4553 nflag = PGA_REQUEUE_HEAD; 4554 } else { 4555 nflag = PGA_REQUEUE; 4556 } 4557 4558 old = vm_page_astate_load(m); 4559 do { 4560 new = old; 4561 4562 /* 4563 * If the page is already in the active queue and we are not 4564 * trying to accelerate reclamation, simply mark it as 4565 * referenced and avoid any queue operations. 4566 */ 4567 new.flags &= ~PGA_QUEUE_OP_MASK; 4568 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4569 (old.flags & PGA_ENQUEUED) != 0) 4570 new.flags |= PGA_REFERENCED; 4571 else { 4572 new.flags |= nflag; 4573 new.queue = nqueue; 4574 } 4575 } while (!vm_page_pqstate_commit(m, &old, new)); 4576 } 4577 4578 /* 4579 * Unwire a page and either attempt to free it or re-add it to the page queues. 4580 */ 4581 void 4582 vm_page_release(vm_page_t m, int flags) 4583 { 4584 vm_object_t object; 4585 4586 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4587 ("vm_page_release: page %p is unmanaged", m)); 4588 4589 if ((flags & VPR_TRYFREE) != 0) { 4590 for (;;) { 4591 object = atomic_load_ptr(&m->object); 4592 if (object == NULL) 4593 break; 4594 /* Depends on type-stability. */ 4595 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4596 break; 4597 if (object == m->object) { 4598 vm_page_release_locked(m, flags); 4599 VM_OBJECT_WUNLOCK(object); 4600 return; 4601 } 4602 VM_OBJECT_WUNLOCK(object); 4603 } 4604 } 4605 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4606 } 4607 4608 /* See vm_page_release(). */ 4609 void 4610 vm_page_release_locked(vm_page_t m, int flags) 4611 { 4612 4613 VM_OBJECT_ASSERT_WLOCKED(m->object); 4614 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4615 ("vm_page_release_locked: page %p is unmanaged", m)); 4616 4617 if (vm_page_unwire_noq(m)) { 4618 if ((flags & VPR_TRYFREE) != 0 && 4619 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4620 m->dirty == 0 && vm_page_tryxbusy(m)) { 4621 /* 4622 * An unlocked lookup may have wired the page before the 4623 * busy lock was acquired, in which case the page must 4624 * not be freed. 4625 */ 4626 if (__predict_true(!vm_page_wired(m))) { 4627 vm_page_free(m); 4628 return; 4629 } 4630 vm_page_xunbusy(m); 4631 } else { 4632 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4633 } 4634 } 4635 } 4636 4637 static bool 4638 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4639 { 4640 u_int old; 4641 4642 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4643 ("vm_page_try_blocked_op: page %p has no object", m)); 4644 KASSERT(vm_page_busied(m), 4645 ("vm_page_try_blocked_op: page %p is not busy", m)); 4646 VM_OBJECT_ASSERT_LOCKED(m->object); 4647 4648 old = atomic_load_int(&m->ref_count); 4649 do { 4650 KASSERT(old != 0, 4651 ("vm_page_try_blocked_op: page %p has no references", m)); 4652 KASSERT((old & VPRC_BLOCKED) == 0, 4653 ("vm_page_try_blocked_op: page %p blocks wirings", m)); 4654 if (VPRC_WIRE_COUNT(old) != 0) 4655 return (false); 4656 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4657 4658 (op)(m); 4659 4660 /* 4661 * If the object is read-locked, new wirings may be created via an 4662 * object lookup. 4663 */ 4664 old = vm_page_drop(m, VPRC_BLOCKED); 4665 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4666 old == (VPRC_BLOCKED | VPRC_OBJREF), 4667 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4668 old, m)); 4669 return (true); 4670 } 4671 4672 /* 4673 * Atomically check for wirings and remove all mappings of the page. 4674 */ 4675 bool 4676 vm_page_try_remove_all(vm_page_t m) 4677 { 4678 4679 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4680 } 4681 4682 /* 4683 * Atomically check for wirings and remove all writeable mappings of the page. 4684 */ 4685 bool 4686 vm_page_try_remove_write(vm_page_t m) 4687 { 4688 4689 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4690 } 4691 4692 /* 4693 * vm_page_advise 4694 * 4695 * Apply the specified advice to the given page. 4696 */ 4697 void 4698 vm_page_advise(vm_page_t m, int advice) 4699 { 4700 4701 VM_OBJECT_ASSERT_WLOCKED(m->object); 4702 vm_page_assert_xbusied(m); 4703 4704 if (advice == MADV_FREE) 4705 /* 4706 * Mark the page clean. This will allow the page to be freed 4707 * without first paging it out. MADV_FREE pages are often 4708 * quickly reused by malloc(3), so we do not do anything that 4709 * would result in a page fault on a later access. 4710 */ 4711 vm_page_undirty(m); 4712 else if (advice != MADV_DONTNEED) { 4713 if (advice == MADV_WILLNEED) 4714 vm_page_activate(m); 4715 return; 4716 } 4717 4718 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4719 vm_page_dirty(m); 4720 4721 /* 4722 * Clear any references to the page. Otherwise, the page daemon will 4723 * immediately reactivate the page. 4724 */ 4725 vm_page_aflag_clear(m, PGA_REFERENCED); 4726 4727 /* 4728 * Place clean pages near the head of the inactive queue rather than 4729 * the tail, thus defeating the queue's LRU operation and ensuring that 4730 * the page will be reused quickly. Dirty pages not already in the 4731 * laundry are moved there. 4732 */ 4733 if (m->dirty == 0) 4734 vm_page_deactivate_noreuse(m); 4735 else if (!vm_page_in_laundry(m)) 4736 vm_page_launder(m); 4737 } 4738 4739 /* 4740 * vm_page_grab_release 4741 * 4742 * Helper routine for grab functions to release busy on return. 4743 */ 4744 static inline void 4745 vm_page_grab_release(vm_page_t m, int allocflags) 4746 { 4747 4748 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4749 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4750 vm_page_sunbusy(m); 4751 else 4752 vm_page_xunbusy(m); 4753 } 4754 } 4755 4756 /* 4757 * vm_page_grab_sleep 4758 * 4759 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4760 * if the caller should retry and false otherwise. 4761 * 4762 * If the object is locked on entry the object will be unlocked with 4763 * false returns and still locked but possibly having been dropped 4764 * with true returns. 4765 */ 4766 static bool 4767 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4768 const char *wmesg, int allocflags, bool locked) 4769 { 4770 4771 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4772 return (false); 4773 4774 /* 4775 * Reference the page before unlocking and sleeping so that 4776 * the page daemon is less likely to reclaim it. 4777 */ 4778 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4779 vm_page_reference(m); 4780 4781 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4782 locked) 4783 VM_OBJECT_WLOCK(object); 4784 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4785 return (false); 4786 4787 return (true); 4788 } 4789 4790 /* 4791 * Assert that the grab flags are valid. 4792 */ 4793 static inline void 4794 vm_page_grab_check(int allocflags) 4795 { 4796 4797 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4798 (allocflags & VM_ALLOC_WIRED) != 0, 4799 ("vm_page_grab*: the pages must be busied or wired")); 4800 4801 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4802 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4803 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4804 } 4805 4806 /* 4807 * Calculate the page allocation flags for grab. 4808 */ 4809 static inline int 4810 vm_page_grab_pflags(int allocflags) 4811 { 4812 int pflags; 4813 4814 pflags = allocflags & 4815 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4816 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY); 4817 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4818 pflags |= VM_ALLOC_WAITFAIL; 4819 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4820 pflags |= VM_ALLOC_SBUSY; 4821 4822 return (pflags); 4823 } 4824 4825 /* 4826 * Grab a page, waiting until we are woken up due to the page changing state. 4827 * We keep on waiting, if the page continues to be in the object, unless 4828 * allocflags forbid waiting. 4829 * 4830 * The object must be locked on entry. This routine may sleep. The lock will, 4831 * however, be released and reacquired if the routine sleeps. 4832 * 4833 * Return a grabbed page, or NULL. Set *found if a page was found, whether or 4834 * not it was grabbed. 4835 */ 4836 static inline vm_page_t 4837 vm_page_grab_lookup(struct pctrie_iter *pages, vm_object_t object, 4838 vm_pindex_t pindex, int allocflags, bool *found) 4839 { 4840 vm_page_t m; 4841 4842 while ((*found = (m = vm_radix_iter_lookup(pages, pindex)) != NULL) && 4843 !vm_page_tryacquire(m, allocflags)) { 4844 if (!vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4845 allocflags, true)) 4846 return (NULL); 4847 pctrie_iter_reset(pages); 4848 } 4849 return (m); 4850 } 4851 4852 /* 4853 * Grab a page. Keep on waiting, as long as the page exists in the object. If 4854 * the page doesn't exist, first allocate it and then conditionally zero it. 4855 * 4856 * The object must be locked on entry. This routine may sleep. The lock will, 4857 * however, be released and reacquired if the routine sleeps. 4858 */ 4859 vm_page_t 4860 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4861 { 4862 struct pctrie_iter pages; 4863 vm_page_t m, mpred; 4864 bool found; 4865 4866 VM_OBJECT_ASSERT_WLOCKED(object); 4867 vm_page_grab_check(allocflags); 4868 4869 vm_page_iter_init(&pages, object); 4870 while ((m = vm_page_grab_lookup( 4871 &pages, object, pindex, allocflags, &found)) == NULL) { 4872 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4873 return (NULL); 4874 if (found && 4875 (allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4876 return (NULL); 4877 mpred = vm_radix_iter_lookup_le(&pages, pindex); 4878 m = vm_page_alloc_after(object, pindex, 4879 vm_page_grab_pflags(allocflags), mpred); 4880 if (m != NULL) { 4881 if ((allocflags & VM_ALLOC_ZERO) != 0 && 4882 (m->flags & PG_ZERO) == 0) 4883 pmap_zero_page(m); 4884 break; 4885 } 4886 if ((allocflags & 4887 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4888 return (NULL); 4889 } 4890 vm_page_grab_release(m, allocflags); 4891 4892 return (m); 4893 } 4894 4895 /* 4896 * Attempt to validate a page, locklessly acquiring it if necessary, given a 4897 * (object, pindex) tuple and either an invalided page or NULL. The resulting 4898 * page will be validated against the identity tuple, and busied or wired as 4899 * requested. A NULL page returned guarantees that the page was not in radix at 4900 * the time of the call but callers must perform higher level synchronization or 4901 * retry the operation under a lock if they require an atomic answer. This is 4902 * the only lock free validation routine, other routines can depend on the 4903 * resulting page state. 4904 * 4905 * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to 4906 * caller flags. 4907 */ 4908 #define PAGE_NOT_ACQUIRED ((vm_page_t)1) 4909 static vm_page_t 4910 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m, 4911 int allocflags) 4912 { 4913 if (m == NULL) 4914 m = vm_page_lookup_unlocked(object, pindex); 4915 for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) { 4916 if (vm_page_trybusy(m, allocflags)) { 4917 if (m->object == object && m->pindex == pindex) { 4918 if ((allocflags & VM_ALLOC_WIRED) != 0) 4919 vm_page_wire(m); 4920 vm_page_grab_release(m, allocflags); 4921 break; 4922 } 4923 /* relookup. */ 4924 vm_page_busy_release(m); 4925 cpu_spinwait(); 4926 continue; 4927 } 4928 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4929 allocflags, false)) 4930 return (PAGE_NOT_ACQUIRED); 4931 } 4932 return (m); 4933 } 4934 4935 /* 4936 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4937 * is not set. 4938 */ 4939 vm_page_t 4940 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4941 { 4942 vm_page_t m; 4943 4944 vm_page_grab_check(allocflags); 4945 m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags); 4946 if (m == PAGE_NOT_ACQUIRED) 4947 return (NULL); 4948 if (m != NULL) 4949 return (m); 4950 4951 /* 4952 * The radix lockless lookup should never return a false negative 4953 * errors. If the user specifies NOCREAT they are guaranteed there 4954 * was no page present at the instant of the call. A NOCREAT caller 4955 * must handle create races gracefully. 4956 */ 4957 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4958 return (NULL); 4959 4960 VM_OBJECT_WLOCK(object); 4961 m = vm_page_grab(object, pindex, allocflags); 4962 VM_OBJECT_WUNLOCK(object); 4963 4964 return (m); 4965 } 4966 4967 /* 4968 * Grab a page and make it valid, paging in if necessary. Pages missing from 4969 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4970 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4971 * in simultaneously. Additional pages will be left on a paging queue but 4972 * will neither be wired nor busy regardless of allocflags. 4973 */ 4974 int 4975 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4976 { 4977 vm_page_t m; 4978 vm_page_t ma[VM_INITIAL_PAGEIN]; 4979 int after, i, pflags, rv; 4980 4981 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4982 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4983 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4984 KASSERT((allocflags & 4985 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4986 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4987 VM_OBJECT_ASSERT_WLOCKED(object); 4988 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4989 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4990 pflags |= VM_ALLOC_WAITFAIL; 4991 4992 retrylookup: 4993 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4994 /* 4995 * If the page is fully valid it can only become invalid 4996 * with the object lock held. If it is not valid it can 4997 * become valid with the busy lock held. Therefore, we 4998 * may unnecessarily lock the exclusive busy here if we 4999 * race with I/O completion not using the object lock. 5000 * However, we will not end up with an invalid page and a 5001 * shared lock. 5002 */ 5003 if (!vm_page_trybusy(m, 5004 vm_page_all_valid(m) ? allocflags : 0)) { 5005 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 5006 allocflags, true); 5007 goto retrylookup; 5008 } 5009 if (vm_page_all_valid(m)) 5010 goto out; 5011 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5012 vm_page_busy_release(m); 5013 *mp = NULL; 5014 return (VM_PAGER_FAIL); 5015 } 5016 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5017 *mp = NULL; 5018 return (VM_PAGER_FAIL); 5019 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { 5020 if (!vm_pager_can_alloc_page(object, pindex)) { 5021 *mp = NULL; 5022 return (VM_PAGER_AGAIN); 5023 } 5024 goto retrylookup; 5025 } 5026 5027 vm_page_assert_xbusied(m); 5028 if (vm_pager_has_page(object, pindex, NULL, &after)) { 5029 after = MIN(after, VM_INITIAL_PAGEIN); 5030 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 5031 after = MAX(after, 1); 5032 ma[0] = m; 5033 for (i = 1; i < after; i++) { 5034 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 5035 if (vm_page_any_valid(ma[i]) || 5036 !vm_page_tryxbusy(ma[i])) 5037 break; 5038 } else { 5039 ma[i] = vm_page_alloc_after(object, 5040 m->pindex + i, VM_ALLOC_NORMAL, ma[i - 1]); 5041 if (ma[i] == NULL) 5042 break; 5043 } 5044 } 5045 after = i; 5046 vm_object_pip_add(object, after); 5047 VM_OBJECT_WUNLOCK(object); 5048 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 5049 VM_OBJECT_WLOCK(object); 5050 vm_object_pip_wakeupn(object, after); 5051 /* Pager may have replaced a page. */ 5052 m = ma[0]; 5053 if (rv != VM_PAGER_OK) { 5054 for (i = 0; i < after; i++) { 5055 if (!vm_page_wired(ma[i])) 5056 vm_page_free(ma[i]); 5057 else 5058 vm_page_xunbusy(ma[i]); 5059 } 5060 *mp = NULL; 5061 return (rv); 5062 } 5063 for (i = 1; i < after; i++) 5064 vm_page_readahead_finish(ma[i]); 5065 MPASS(vm_page_all_valid(m)); 5066 } else { 5067 vm_page_zero_invalid(m, TRUE); 5068 } 5069 out: 5070 if ((allocflags & VM_ALLOC_WIRED) != 0) 5071 vm_page_wire(m); 5072 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 5073 vm_page_busy_downgrade(m); 5074 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 5075 vm_page_busy_release(m); 5076 *mp = m; 5077 return (VM_PAGER_OK); 5078 } 5079 5080 /* 5081 * Grab a page. Keep on waiting, as long as the page exists in the object. If 5082 * the page doesn't exist, and the pager has it, allocate it and zero part of 5083 * it. 5084 * 5085 * The object must be locked on entry. This routine may sleep. The lock will, 5086 * however, be released and reacquired if the routine sleeps. 5087 */ 5088 int 5089 vm_page_grab_zero_partial(vm_object_t object, vm_pindex_t pindex, int base, 5090 int end) 5091 { 5092 struct pctrie_iter pages; 5093 vm_page_t m, mpred; 5094 int allocflags, rv; 5095 bool found; 5096 5097 VM_OBJECT_ASSERT_WLOCKED(object); 5098 KASSERT(base >= 0, ("%s: base %d", __func__, base)); 5099 KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base, 5100 end)); 5101 5102 allocflags = VM_ALLOC_NOCREAT | VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL; 5103 vm_page_iter_init(&pages, object); 5104 while ((m = vm_page_grab_lookup( 5105 &pages, object, pindex, allocflags, &found)) == NULL) { 5106 if (!vm_pager_has_page(object, pindex, NULL, NULL)) 5107 return (0); 5108 mpred = vm_radix_iter_lookup_le(&pages, pindex); 5109 m = vm_page_alloc_after(object, pindex, 5110 vm_page_grab_pflags(allocflags), mpred); 5111 if (m != NULL) { 5112 vm_object_pip_add(object, 1); 5113 VM_OBJECT_WUNLOCK(object); 5114 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 5115 VM_OBJECT_WLOCK(object); 5116 vm_object_pip_wakeup(object); 5117 if (rv != VM_PAGER_OK) { 5118 vm_page_free(m); 5119 return (EIO); 5120 } 5121 5122 /* 5123 * Since the page was not resident, and therefore not 5124 * recently accessed, immediately enqueue it for 5125 * asynchronous laundering. The current operation is 5126 * not regarded as an access. 5127 */ 5128 vm_page_launder(m); 5129 break; 5130 } 5131 } 5132 5133 pmap_zero_page_area(m, base, end - base); 5134 KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m)); 5135 vm_page_set_dirty(m); 5136 vm_page_xunbusy(m); 5137 return (0); 5138 } 5139 5140 /* 5141 * Locklessly grab a valid page. If the page is not valid or not yet 5142 * allocated this will fall back to the object lock method. 5143 */ 5144 int 5145 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 5146 vm_pindex_t pindex, int allocflags) 5147 { 5148 vm_page_t m; 5149 int flags; 5150 int error; 5151 5152 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 5153 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 5154 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 5155 "mismatch")); 5156 KASSERT((allocflags & 5157 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 5158 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 5159 5160 /* 5161 * Attempt a lockless lookup and busy. We need at least an sbusy 5162 * before we can inspect the valid field and return a wired page. 5163 */ 5164 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 5165 vm_page_grab_check(flags); 5166 m = vm_page_acquire_unlocked(object, pindex, NULL, flags); 5167 if (m == PAGE_NOT_ACQUIRED) 5168 return (VM_PAGER_FAIL); 5169 if (m != NULL) { 5170 if (vm_page_all_valid(m)) { 5171 if ((allocflags & VM_ALLOC_WIRED) != 0) 5172 vm_page_wire(m); 5173 vm_page_grab_release(m, allocflags); 5174 *mp = m; 5175 return (VM_PAGER_OK); 5176 } 5177 vm_page_busy_release(m); 5178 } 5179 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5180 *mp = NULL; 5181 return (VM_PAGER_FAIL); 5182 } 5183 VM_OBJECT_WLOCK(object); 5184 error = vm_page_grab_valid(mp, object, pindex, allocflags); 5185 VM_OBJECT_WUNLOCK(object); 5186 5187 return (error); 5188 } 5189 5190 /* 5191 * Return the specified range of pages from the given object. For each 5192 * page offset within the range, if a page already exists within the object 5193 * at that offset and it is busy, then wait for it to change state. If, 5194 * instead, the page doesn't exist, then allocate it. 5195 * 5196 * The caller must always specify an allocation class. 5197 * 5198 * allocation classes: 5199 * VM_ALLOC_NORMAL normal process request 5200 * VM_ALLOC_SYSTEM system *really* needs the pages 5201 * 5202 * The caller must always specify that the pages are to be busied and/or 5203 * wired. 5204 * 5205 * optional allocation flags: 5206 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 5207 * VM_ALLOC_NOBUSY do not exclusive busy the page 5208 * VM_ALLOC_NOWAIT do not sleep 5209 * VM_ALLOC_SBUSY set page to sbusy state 5210 * VM_ALLOC_WIRED wire the pages 5211 * VM_ALLOC_ZERO zero and validate any invalid pages 5212 * 5213 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 5214 * may return a partial prefix of the requested range. 5215 */ 5216 int 5217 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 5218 vm_page_t *ma, int count) 5219 { 5220 vm_page_t m, mpred; 5221 int pflags; 5222 int i; 5223 5224 VM_OBJECT_ASSERT_WLOCKED(object); 5225 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 5226 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 5227 KASSERT(count > 0, 5228 ("vm_page_grab_pages: invalid page count %d", count)); 5229 vm_page_grab_check(allocflags); 5230 5231 pflags = vm_page_grab_pflags(allocflags); 5232 i = 0; 5233 retrylookup: 5234 m = vm_page_mpred(object, pindex + i); 5235 if (m == NULL || m->pindex != pindex + i) { 5236 mpred = m; 5237 m = NULL; 5238 } else 5239 mpred = TAILQ_PREV(m, pglist, listq); 5240 for (; i < count; i++) { 5241 if (m != NULL) { 5242 if (!vm_page_tryacquire(m, allocflags)) { 5243 if (vm_page_grab_sleep(object, m, pindex + i, 5244 "grbmaw", allocflags, true)) 5245 goto retrylookup; 5246 break; 5247 } 5248 } else { 5249 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 5250 break; 5251 m = vm_page_alloc_after(object, pindex + i, 5252 pflags | VM_ALLOC_COUNT(count - i), mpred); 5253 if (m == NULL) { 5254 if ((allocflags & (VM_ALLOC_NOWAIT | 5255 VM_ALLOC_WAITFAIL)) != 0) 5256 break; 5257 goto retrylookup; 5258 } 5259 } 5260 if (vm_page_none_valid(m) && 5261 (allocflags & VM_ALLOC_ZERO) != 0) { 5262 if ((m->flags & PG_ZERO) == 0) 5263 pmap_zero_page(m); 5264 vm_page_valid(m); 5265 } 5266 vm_page_grab_release(m, allocflags); 5267 ma[i] = mpred = m; 5268 m = vm_page_next(m); 5269 } 5270 return (i); 5271 } 5272 5273 /* 5274 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 5275 * and will fall back to the locked variant to handle allocation. 5276 */ 5277 int 5278 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 5279 int allocflags, vm_page_t *ma, int count) 5280 { 5281 vm_page_t m; 5282 int flags; 5283 int i; 5284 5285 KASSERT(count > 0, 5286 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 5287 vm_page_grab_check(allocflags); 5288 5289 /* 5290 * Modify flags for lockless acquire to hold the page until we 5291 * set it valid if necessary. 5292 */ 5293 flags = allocflags & ~VM_ALLOC_NOBUSY; 5294 vm_page_grab_check(flags); 5295 m = NULL; 5296 for (i = 0; i < count; i++, pindex++) { 5297 /* 5298 * We may see a false NULL here because the previous page has 5299 * been removed or just inserted and the list is loaded without 5300 * barriers. Switch to radix to verify. 5301 */ 5302 if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex || 5303 atomic_load_ptr(&m->object) != object) { 5304 /* 5305 * This guarantees the result is instantaneously 5306 * correct. 5307 */ 5308 m = NULL; 5309 } 5310 m = vm_page_acquire_unlocked(object, pindex, m, flags); 5311 if (m == PAGE_NOT_ACQUIRED) 5312 return (i); 5313 if (m == NULL) 5314 break; 5315 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 5316 if ((m->flags & PG_ZERO) == 0) 5317 pmap_zero_page(m); 5318 vm_page_valid(m); 5319 } 5320 /* m will still be wired or busy according to flags. */ 5321 vm_page_grab_release(m, allocflags); 5322 ma[i] = m; 5323 m = TAILQ_NEXT(m, listq); 5324 } 5325 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 5326 return (i); 5327 count -= i; 5328 VM_OBJECT_WLOCK(object); 5329 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 5330 VM_OBJECT_WUNLOCK(object); 5331 5332 return (i); 5333 } 5334 5335 /* 5336 * Mapping function for valid or dirty bits in a page. 5337 * 5338 * Inputs are required to range within a page. 5339 */ 5340 vm_page_bits_t 5341 vm_page_bits(int base, int size) 5342 { 5343 int first_bit; 5344 int last_bit; 5345 5346 KASSERT( 5347 base + size <= PAGE_SIZE, 5348 ("vm_page_bits: illegal base/size %d/%d", base, size) 5349 ); 5350 5351 if (size == 0) /* handle degenerate case */ 5352 return (0); 5353 5354 first_bit = base >> DEV_BSHIFT; 5355 last_bit = (base + size - 1) >> DEV_BSHIFT; 5356 5357 return (((vm_page_bits_t)2 << last_bit) - 5358 ((vm_page_bits_t)1 << first_bit)); 5359 } 5360 5361 void 5362 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5363 { 5364 5365 #if PAGE_SIZE == 32768 5366 atomic_set_64((uint64_t *)bits, set); 5367 #elif PAGE_SIZE == 16384 5368 atomic_set_32((uint32_t *)bits, set); 5369 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5370 atomic_set_16((uint16_t *)bits, set); 5371 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5372 atomic_set_8((uint8_t *)bits, set); 5373 #else /* PAGE_SIZE <= 8192 */ 5374 uintptr_t addr; 5375 int shift; 5376 5377 addr = (uintptr_t)bits; 5378 /* 5379 * Use a trick to perform a 32-bit atomic on the 5380 * containing aligned word, to not depend on the existence 5381 * of atomic_{set, clear}_{8, 16}. 5382 */ 5383 shift = addr & (sizeof(uint32_t) - 1); 5384 #if BYTE_ORDER == BIG_ENDIAN 5385 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5386 #else 5387 shift *= NBBY; 5388 #endif 5389 addr &= ~(sizeof(uint32_t) - 1); 5390 atomic_set_32((uint32_t *)addr, set << shift); 5391 #endif /* PAGE_SIZE */ 5392 } 5393 5394 static inline void 5395 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5396 { 5397 5398 #if PAGE_SIZE == 32768 5399 atomic_clear_64((uint64_t *)bits, clear); 5400 #elif PAGE_SIZE == 16384 5401 atomic_clear_32((uint32_t *)bits, clear); 5402 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5403 atomic_clear_16((uint16_t *)bits, clear); 5404 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5405 atomic_clear_8((uint8_t *)bits, clear); 5406 #else /* PAGE_SIZE <= 8192 */ 5407 uintptr_t addr; 5408 int shift; 5409 5410 addr = (uintptr_t)bits; 5411 /* 5412 * Use a trick to perform a 32-bit atomic on the 5413 * containing aligned word, to not depend on the existence 5414 * of atomic_{set, clear}_{8, 16}. 5415 */ 5416 shift = addr & (sizeof(uint32_t) - 1); 5417 #if BYTE_ORDER == BIG_ENDIAN 5418 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5419 #else 5420 shift *= NBBY; 5421 #endif 5422 addr &= ~(sizeof(uint32_t) - 1); 5423 atomic_clear_32((uint32_t *)addr, clear << shift); 5424 #endif /* PAGE_SIZE */ 5425 } 5426 5427 static inline vm_page_bits_t 5428 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5429 { 5430 #if PAGE_SIZE == 32768 5431 uint64_t old; 5432 5433 old = *bits; 5434 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5435 return (old); 5436 #elif PAGE_SIZE == 16384 5437 uint32_t old; 5438 5439 old = *bits; 5440 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5441 return (old); 5442 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5443 uint16_t old; 5444 5445 old = *bits; 5446 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5447 return (old); 5448 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5449 uint8_t old; 5450 5451 old = *bits; 5452 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5453 return (old); 5454 #else /* PAGE_SIZE <= 4096*/ 5455 uintptr_t addr; 5456 uint32_t old, new, mask; 5457 int shift; 5458 5459 addr = (uintptr_t)bits; 5460 /* 5461 * Use a trick to perform a 32-bit atomic on the 5462 * containing aligned word, to not depend on the existence 5463 * of atomic_{set, swap, clear}_{8, 16}. 5464 */ 5465 shift = addr & (sizeof(uint32_t) - 1); 5466 #if BYTE_ORDER == BIG_ENDIAN 5467 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5468 #else 5469 shift *= NBBY; 5470 #endif 5471 addr &= ~(sizeof(uint32_t) - 1); 5472 mask = VM_PAGE_BITS_ALL << shift; 5473 5474 old = *bits; 5475 do { 5476 new = old & ~mask; 5477 new |= newbits << shift; 5478 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5479 return (old >> shift); 5480 #endif /* PAGE_SIZE */ 5481 } 5482 5483 /* 5484 * vm_page_set_valid_range: 5485 * 5486 * Sets portions of a page valid. The arguments are expected 5487 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5488 * of any partial chunks touched by the range. The invalid portion of 5489 * such chunks will be zeroed. 5490 * 5491 * (base + size) must be less then or equal to PAGE_SIZE. 5492 */ 5493 void 5494 vm_page_set_valid_range(vm_page_t m, int base, int size) 5495 { 5496 int endoff, frag; 5497 vm_page_bits_t pagebits; 5498 5499 vm_page_assert_busied(m); 5500 if (size == 0) /* handle degenerate case */ 5501 return; 5502 5503 /* 5504 * If the base is not DEV_BSIZE aligned and the valid 5505 * bit is clear, we have to zero out a portion of the 5506 * first block. 5507 */ 5508 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5509 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5510 pmap_zero_page_area(m, frag, base - frag); 5511 5512 /* 5513 * If the ending offset is not DEV_BSIZE aligned and the 5514 * valid bit is clear, we have to zero out a portion of 5515 * the last block. 5516 */ 5517 endoff = base + size; 5518 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5519 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5520 pmap_zero_page_area(m, endoff, 5521 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5522 5523 /* 5524 * Assert that no previously invalid block that is now being validated 5525 * is already dirty. 5526 */ 5527 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5528 ("vm_page_set_valid_range: page %p is dirty", m)); 5529 5530 /* 5531 * Set valid bits inclusive of any overlap. 5532 */ 5533 pagebits = vm_page_bits(base, size); 5534 if (vm_page_xbusied(m)) 5535 m->valid |= pagebits; 5536 else 5537 vm_page_bits_set(m, &m->valid, pagebits); 5538 } 5539 5540 /* 5541 * Set the page dirty bits and free the invalid swap space if 5542 * present. Returns the previous dirty bits. 5543 */ 5544 vm_page_bits_t 5545 vm_page_set_dirty(vm_page_t m) 5546 { 5547 vm_page_bits_t old; 5548 5549 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5550 5551 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5552 old = m->dirty; 5553 m->dirty = VM_PAGE_BITS_ALL; 5554 } else 5555 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5556 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5557 vm_pager_page_unswapped(m); 5558 5559 return (old); 5560 } 5561 5562 /* 5563 * Clear the given bits from the specified page's dirty field. 5564 */ 5565 static __inline void 5566 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5567 { 5568 5569 vm_page_assert_busied(m); 5570 5571 /* 5572 * If the page is xbusied and not write mapped we are the 5573 * only thread that can modify dirty bits. Otherwise, The pmap 5574 * layer can call vm_page_dirty() without holding a distinguished 5575 * lock. The combination of page busy and atomic operations 5576 * suffice to guarantee consistency of the page dirty field. 5577 */ 5578 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5579 m->dirty &= ~pagebits; 5580 else 5581 vm_page_bits_clear(m, &m->dirty, pagebits); 5582 } 5583 5584 /* 5585 * vm_page_set_validclean: 5586 * 5587 * Sets portions of a page valid and clean. The arguments are expected 5588 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5589 * of any partial chunks touched by the range. The invalid portion of 5590 * such chunks will be zero'd. 5591 * 5592 * (base + size) must be less then or equal to PAGE_SIZE. 5593 */ 5594 void 5595 vm_page_set_validclean(vm_page_t m, int base, int size) 5596 { 5597 vm_page_bits_t oldvalid, pagebits; 5598 int endoff, frag; 5599 5600 vm_page_assert_busied(m); 5601 if (size == 0) /* handle degenerate case */ 5602 return; 5603 5604 /* 5605 * If the base is not DEV_BSIZE aligned and the valid 5606 * bit is clear, we have to zero out a portion of the 5607 * first block. 5608 */ 5609 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5610 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5611 pmap_zero_page_area(m, frag, base - frag); 5612 5613 /* 5614 * If the ending offset is not DEV_BSIZE aligned and the 5615 * valid bit is clear, we have to zero out a portion of 5616 * the last block. 5617 */ 5618 endoff = base + size; 5619 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5620 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5621 pmap_zero_page_area(m, endoff, 5622 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5623 5624 /* 5625 * Set valid, clear dirty bits. If validating the entire 5626 * page we can safely clear the pmap modify bit. We also 5627 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5628 * takes a write fault on a MAP_NOSYNC memory area the flag will 5629 * be set again. 5630 * 5631 * We set valid bits inclusive of any overlap, but we can only 5632 * clear dirty bits for DEV_BSIZE chunks that are fully within 5633 * the range. 5634 */ 5635 oldvalid = m->valid; 5636 pagebits = vm_page_bits(base, size); 5637 if (vm_page_xbusied(m)) 5638 m->valid |= pagebits; 5639 else 5640 vm_page_bits_set(m, &m->valid, pagebits); 5641 #if 0 /* NOT YET */ 5642 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5643 frag = DEV_BSIZE - frag; 5644 base += frag; 5645 size -= frag; 5646 if (size < 0) 5647 size = 0; 5648 } 5649 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5650 #endif 5651 if (base == 0 && size == PAGE_SIZE) { 5652 /* 5653 * The page can only be modified within the pmap if it is 5654 * mapped, and it can only be mapped if it was previously 5655 * fully valid. 5656 */ 5657 if (oldvalid == VM_PAGE_BITS_ALL) 5658 /* 5659 * Perform the pmap_clear_modify() first. Otherwise, 5660 * a concurrent pmap operation, such as 5661 * pmap_protect(), could clear a modification in the 5662 * pmap and set the dirty field on the page before 5663 * pmap_clear_modify() had begun and after the dirty 5664 * field was cleared here. 5665 */ 5666 pmap_clear_modify(m); 5667 m->dirty = 0; 5668 vm_page_aflag_clear(m, PGA_NOSYNC); 5669 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5670 m->dirty &= ~pagebits; 5671 else 5672 vm_page_clear_dirty_mask(m, pagebits); 5673 } 5674 5675 void 5676 vm_page_clear_dirty(vm_page_t m, int base, int size) 5677 { 5678 5679 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5680 } 5681 5682 /* 5683 * vm_page_set_invalid: 5684 * 5685 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5686 * valid and dirty bits for the effected areas are cleared. 5687 */ 5688 void 5689 vm_page_set_invalid(vm_page_t m, int base, int size) 5690 { 5691 vm_page_bits_t bits; 5692 vm_object_t object; 5693 5694 /* 5695 * The object lock is required so that pages can't be mapped 5696 * read-only while we're in the process of invalidating them. 5697 */ 5698 object = m->object; 5699 VM_OBJECT_ASSERT_WLOCKED(object); 5700 vm_page_assert_busied(m); 5701 5702 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5703 size >= object->un_pager.vnp.vnp_size) 5704 bits = VM_PAGE_BITS_ALL; 5705 else 5706 bits = vm_page_bits(base, size); 5707 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5708 pmap_remove_all(m); 5709 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5710 !pmap_page_is_mapped(m), 5711 ("vm_page_set_invalid: page %p is mapped", m)); 5712 if (vm_page_xbusied(m)) { 5713 m->valid &= ~bits; 5714 m->dirty &= ~bits; 5715 } else { 5716 vm_page_bits_clear(m, &m->valid, bits); 5717 vm_page_bits_clear(m, &m->dirty, bits); 5718 } 5719 } 5720 5721 /* 5722 * vm_page_invalid: 5723 * 5724 * Invalidates the entire page. The page must be busy, unmapped, and 5725 * the enclosing object must be locked. The object locks protects 5726 * against concurrent read-only pmap enter which is done without 5727 * busy. 5728 */ 5729 void 5730 vm_page_invalid(vm_page_t m) 5731 { 5732 5733 vm_page_assert_busied(m); 5734 VM_OBJECT_ASSERT_WLOCKED(m->object); 5735 MPASS(!pmap_page_is_mapped(m)); 5736 5737 if (vm_page_xbusied(m)) 5738 m->valid = 0; 5739 else 5740 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5741 } 5742 5743 /* 5744 * vm_page_zero_invalid() 5745 * 5746 * The kernel assumes that the invalid portions of a page contain 5747 * garbage, but such pages can be mapped into memory by user code. 5748 * When this occurs, we must zero out the non-valid portions of the 5749 * page so user code sees what it expects. 5750 * 5751 * Pages are most often semi-valid when the end of a file is mapped 5752 * into memory and the file's size is not page aligned. 5753 */ 5754 void 5755 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5756 { 5757 int b; 5758 int i; 5759 5760 /* 5761 * Scan the valid bits looking for invalid sections that 5762 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5763 * valid bit may be set ) have already been zeroed by 5764 * vm_page_set_validclean(). 5765 */ 5766 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5767 if (i == (PAGE_SIZE / DEV_BSIZE) || 5768 (m->valid & ((vm_page_bits_t)1 << i))) { 5769 if (i > b) { 5770 pmap_zero_page_area(m, 5771 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5772 } 5773 b = i + 1; 5774 } 5775 } 5776 5777 /* 5778 * setvalid is TRUE when we can safely set the zero'd areas 5779 * as being valid. We can do this if there are no cache consistency 5780 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5781 */ 5782 if (setvalid) 5783 vm_page_valid(m); 5784 } 5785 5786 /* 5787 * vm_page_is_valid: 5788 * 5789 * Is (partial) page valid? Note that the case where size == 0 5790 * will return FALSE in the degenerate case where the page is 5791 * entirely invalid, and TRUE otherwise. 5792 * 5793 * Some callers envoke this routine without the busy lock held and 5794 * handle races via higher level locks. Typical callers should 5795 * hold a busy lock to prevent invalidation. 5796 */ 5797 int 5798 vm_page_is_valid(vm_page_t m, int base, int size) 5799 { 5800 vm_page_bits_t bits; 5801 5802 bits = vm_page_bits(base, size); 5803 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5804 } 5805 5806 /* 5807 * Returns true if all of the specified predicates are true for the entire 5808 * (super)page and false otherwise. 5809 */ 5810 bool 5811 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m) 5812 { 5813 vm_object_t object; 5814 int i, npages; 5815 5816 object = m->object; 5817 if (skip_m != NULL && skip_m->object != object) 5818 return (false); 5819 VM_OBJECT_ASSERT_LOCKED(object); 5820 KASSERT(psind <= m->psind, 5821 ("psind %d > psind %d of m %p", psind, m->psind, m)); 5822 npages = atop(pagesizes[psind]); 5823 5824 /* 5825 * The physically contiguous pages that make up a superpage, i.e., a 5826 * page with a page size index ("psind") greater than zero, will 5827 * occupy adjacent entries in vm_page_array[]. 5828 */ 5829 for (i = 0; i < npages; i++) { 5830 /* Always test object consistency, including "skip_m". */ 5831 if (m[i].object != object) 5832 return (false); 5833 if (&m[i] == skip_m) 5834 continue; 5835 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5836 return (false); 5837 if ((flags & PS_ALL_DIRTY) != 0) { 5838 /* 5839 * Calling vm_page_test_dirty() or pmap_is_modified() 5840 * might stop this case from spuriously returning 5841 * "false". However, that would require a write lock 5842 * on the object containing "m[i]". 5843 */ 5844 if (m[i].dirty != VM_PAGE_BITS_ALL) 5845 return (false); 5846 } 5847 if ((flags & PS_ALL_VALID) != 0 && 5848 m[i].valid != VM_PAGE_BITS_ALL) 5849 return (false); 5850 } 5851 return (true); 5852 } 5853 5854 /* 5855 * Set the page's dirty bits if the page is modified. 5856 */ 5857 void 5858 vm_page_test_dirty(vm_page_t m) 5859 { 5860 5861 vm_page_assert_busied(m); 5862 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5863 vm_page_dirty(m); 5864 } 5865 5866 void 5867 vm_page_valid(vm_page_t m) 5868 { 5869 5870 vm_page_assert_busied(m); 5871 if (vm_page_xbusied(m)) 5872 m->valid = VM_PAGE_BITS_ALL; 5873 else 5874 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5875 } 5876 5877 void 5878 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5879 { 5880 5881 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5882 } 5883 5884 void 5885 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5886 { 5887 5888 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5889 } 5890 5891 int 5892 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5893 { 5894 5895 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5896 } 5897 5898 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5899 void 5900 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5901 { 5902 5903 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5904 } 5905 5906 void 5907 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5908 { 5909 5910 mtx_assert_(vm_page_lockptr(m), a, file, line); 5911 } 5912 #endif 5913 5914 #ifdef INVARIANTS 5915 void 5916 vm_page_object_busy_assert(vm_page_t m) 5917 { 5918 5919 /* 5920 * Certain of the page's fields may only be modified by the 5921 * holder of a page or object busy. 5922 */ 5923 if (m->object != NULL && !vm_page_busied(m)) 5924 VM_OBJECT_ASSERT_BUSY(m->object); 5925 } 5926 5927 void 5928 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5929 { 5930 5931 if ((bits & PGA_WRITEABLE) == 0) 5932 return; 5933 5934 /* 5935 * The PGA_WRITEABLE flag can only be set if the page is 5936 * managed, is exclusively busied or the object is locked. 5937 * Currently, this flag is only set by pmap_enter(). 5938 */ 5939 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5940 ("PGA_WRITEABLE on unmanaged page")); 5941 if (!vm_page_xbusied(m)) 5942 VM_OBJECT_ASSERT_BUSY(m->object); 5943 } 5944 #endif 5945 5946 #include "opt_ddb.h" 5947 #ifdef DDB 5948 #include <sys/kernel.h> 5949 5950 #include <ddb/ddb.h> 5951 5952 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5953 { 5954 5955 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5956 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5957 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5958 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5959 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5960 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5961 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5962 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5963 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5964 } 5965 5966 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5967 { 5968 int dom; 5969 5970 db_printf("pq_free %d\n", vm_free_count()); 5971 for (dom = 0; dom < vm_ndomains; dom++) { 5972 db_printf( 5973 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5974 dom, 5975 vm_dom[dom].vmd_page_count, 5976 vm_dom[dom].vmd_free_count, 5977 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5978 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5979 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5980 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5981 } 5982 } 5983 5984 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5985 { 5986 vm_page_t m; 5987 boolean_t phys, virt; 5988 5989 if (!have_addr) { 5990 db_printf("show pginfo addr\n"); 5991 return; 5992 } 5993 5994 phys = strchr(modif, 'p') != NULL; 5995 virt = strchr(modif, 'v') != NULL; 5996 if (virt) 5997 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5998 else if (phys) 5999 m = PHYS_TO_VM_PAGE(addr); 6000 else 6001 m = (vm_page_t)addr; 6002 db_printf( 6003 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 6004 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 6005 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 6006 m->a.queue, m->ref_count, m->a.flags, m->oflags, 6007 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 6008 } 6009 #endif /* DDB */ 6010