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 vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 166 int req, vm_page_t mpred); 167 static void vm_page_alloc_check(vm_page_t m); 168 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req); 169 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 170 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 171 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 172 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 173 static bool vm_page_free_prep(vm_page_t m); 174 static void vm_page_free_toq(vm_page_t m); 175 static void vm_page_init(void *dummy); 176 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 177 vm_pindex_t pindex, vm_page_t mpred); 178 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 179 vm_page_t mpred); 180 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 181 const uint16_t nflag); 182 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 183 vm_page_t m_run, vm_paddr_t high); 184 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 185 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 186 int req); 187 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 188 int flags); 189 static void vm_page_zone_release(void *arg, void **store, int cnt); 190 191 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 192 193 static void 194 vm_page_init(void *dummy) 195 { 196 197 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 198 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 199 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE); 200 } 201 202 static int pgcache_zone_max_pcpu; 203 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu, 204 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0, 205 "Per-CPU page cache size"); 206 207 /* 208 * The cache page zone is initialized later since we need to be able to allocate 209 * pages before UMA is fully initialized. 210 */ 211 static void 212 vm_page_init_cache_zones(void *dummy __unused) 213 { 214 struct vm_domain *vmd; 215 struct vm_pgcache *pgcache; 216 int cache, domain, maxcache, pool; 217 218 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu); 219 maxcache = pgcache_zone_max_pcpu * mp_ncpus; 220 for (domain = 0; domain < vm_ndomains; domain++) { 221 vmd = VM_DOMAIN(domain); 222 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 223 pgcache = &vmd->vmd_pgcache[pool]; 224 pgcache->domain = domain; 225 pgcache->pool = pool; 226 pgcache->zone = uma_zcache_create("vm pgcache", 227 PAGE_SIZE, NULL, NULL, NULL, NULL, 228 vm_page_zone_import, vm_page_zone_release, pgcache, 229 UMA_ZONE_VM); 230 231 /* 232 * Limit each pool's zone to 0.1% of the pages in the 233 * domain. 234 */ 235 cache = maxcache != 0 ? maxcache : 236 vmd->vmd_page_count / 1000; 237 uma_zone_set_maxcache(pgcache->zone, cache); 238 } 239 } 240 } 241 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 242 243 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 244 #if PAGE_SIZE == 32768 245 #ifdef CTASSERT 246 CTASSERT(sizeof(u_long) >= 8); 247 #endif 248 #endif 249 250 /* 251 * vm_set_page_size: 252 * 253 * Sets the page size, perhaps based upon the memory 254 * size. Must be called before any use of page-size 255 * dependent functions. 256 */ 257 void 258 vm_set_page_size(void) 259 { 260 if (vm_cnt.v_page_size == 0) 261 vm_cnt.v_page_size = PAGE_SIZE; 262 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 263 panic("vm_set_page_size: page size not a power of two"); 264 } 265 266 /* 267 * vm_page_blacklist_next: 268 * 269 * Find the next entry in the provided string of blacklist 270 * addresses. Entries are separated by space, comma, or newline. 271 * If an invalid integer is encountered then the rest of the 272 * string is skipped. Updates the list pointer to the next 273 * character, or NULL if the string is exhausted or invalid. 274 */ 275 static vm_paddr_t 276 vm_page_blacklist_next(char **list, char *end) 277 { 278 vm_paddr_t bad; 279 char *cp, *pos; 280 281 if (list == NULL || *list == NULL) 282 return (0); 283 if (**list =='\0') { 284 *list = NULL; 285 return (0); 286 } 287 288 /* 289 * If there's no end pointer then the buffer is coming from 290 * the kenv and we know it's null-terminated. 291 */ 292 if (end == NULL) 293 end = *list + strlen(*list); 294 295 /* Ensure that strtoq() won't walk off the end */ 296 if (*end != '\0') { 297 if (*end == '\n' || *end == ' ' || *end == ',') 298 *end = '\0'; 299 else { 300 printf("Blacklist not terminated, skipping\n"); 301 *list = NULL; 302 return (0); 303 } 304 } 305 306 for (pos = *list; *pos != '\0'; pos = cp) { 307 bad = strtoq(pos, &cp, 0); 308 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 309 if (bad == 0) { 310 if (++cp < end) 311 continue; 312 else 313 break; 314 } 315 } else 316 break; 317 if (*cp == '\0' || ++cp >= end) 318 *list = NULL; 319 else 320 *list = cp; 321 return (trunc_page(bad)); 322 } 323 printf("Garbage in RAM blacklist, skipping\n"); 324 *list = NULL; 325 return (0); 326 } 327 328 bool 329 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 330 { 331 struct vm_domain *vmd; 332 vm_page_t m; 333 bool found; 334 335 m = vm_phys_paddr_to_vm_page(pa); 336 if (m == NULL) 337 return (true); /* page does not exist, no failure */ 338 339 vmd = VM_DOMAIN(vm_phys_domain(pa)); 340 vm_domain_free_lock(vmd); 341 found = vm_phys_unfree_page(pa); 342 vm_domain_free_unlock(vmd); 343 if (found) { 344 vm_domain_freecnt_inc(vmd, -1); 345 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 346 if (verbose) 347 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 348 } 349 return (found); 350 } 351 352 /* 353 * vm_page_blacklist_check: 354 * 355 * Iterate through the provided string of blacklist addresses, pulling 356 * each entry out of the physical allocator free list and putting it 357 * onto a list for reporting via the vm.page_blacklist sysctl. 358 */ 359 static void 360 vm_page_blacklist_check(char *list, char *end) 361 { 362 vm_paddr_t pa; 363 char *next; 364 365 next = list; 366 while (next != NULL) { 367 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 368 continue; 369 vm_page_blacklist_add(pa, bootverbose); 370 } 371 } 372 373 /* 374 * vm_page_blacklist_load: 375 * 376 * Search for a special module named "ram_blacklist". It'll be a 377 * plain text file provided by the user via the loader directive 378 * of the same name. 379 */ 380 static void 381 vm_page_blacklist_load(char **list, char **end) 382 { 383 void *mod; 384 u_char *ptr; 385 u_int len; 386 387 mod = NULL; 388 ptr = NULL; 389 390 mod = preload_search_by_type("ram_blacklist"); 391 if (mod != NULL) { 392 ptr = preload_fetch_addr(mod); 393 len = preload_fetch_size(mod); 394 } 395 *list = ptr; 396 if (ptr != NULL) 397 *end = ptr + len; 398 else 399 *end = NULL; 400 return; 401 } 402 403 static int 404 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 405 { 406 vm_page_t m; 407 struct sbuf sbuf; 408 int error, first; 409 410 first = 1; 411 error = sysctl_wire_old_buffer(req, 0); 412 if (error != 0) 413 return (error); 414 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 415 TAILQ_FOREACH(m, &blacklist_head, listq) { 416 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 417 (uintmax_t)m->phys_addr); 418 first = 0; 419 } 420 error = sbuf_finish(&sbuf); 421 sbuf_delete(&sbuf); 422 return (error); 423 } 424 425 /* 426 * Initialize a dummy page for use in scans of the specified paging queue. 427 * In principle, this function only needs to set the flag PG_MARKER. 428 * Nonetheless, it write busies the page as a safety precaution. 429 */ 430 void 431 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 432 { 433 434 bzero(marker, sizeof(*marker)); 435 marker->flags = PG_MARKER; 436 marker->a.flags = aflags; 437 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 438 marker->a.queue = queue; 439 } 440 441 static void 442 vm_page_domain_init(int domain) 443 { 444 struct vm_domain *vmd; 445 struct vm_pagequeue *pq; 446 int i; 447 448 vmd = VM_DOMAIN(domain); 449 bzero(vmd, sizeof(*vmd)); 450 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 451 "vm inactive pagequeue"; 452 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 453 "vm active pagequeue"; 454 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 455 "vm laundry pagequeue"; 456 *__DECONST(const char **, 457 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 458 "vm unswappable pagequeue"; 459 vmd->vmd_domain = domain; 460 vmd->vmd_page_count = 0; 461 vmd->vmd_free_count = 0; 462 vmd->vmd_segs = 0; 463 vmd->vmd_oom = false; 464 vmd->vmd_helper_threads_enabled = true; 465 for (i = 0; i < PQ_COUNT; i++) { 466 pq = &vmd->vmd_pagequeues[i]; 467 TAILQ_INIT(&pq->pq_pl); 468 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 469 MTX_DEF | MTX_DUPOK); 470 pq->pq_pdpages = 0; 471 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 472 } 473 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 474 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 475 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 476 477 /* 478 * inacthead is used to provide FIFO ordering for LRU-bypassing 479 * insertions. 480 */ 481 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 482 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 483 &vmd->vmd_inacthead, plinks.q); 484 485 /* 486 * The clock pages are used to implement active queue scanning without 487 * requeues. Scans start at clock[0], which is advanced after the scan 488 * ends. When the two clock hands meet, they are reset and scanning 489 * resumes from the head of the queue. 490 */ 491 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 492 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 493 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 494 &vmd->vmd_clock[0], plinks.q); 495 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 496 &vmd->vmd_clock[1], plinks.q); 497 } 498 499 /* 500 * Initialize a physical page in preparation for adding it to the free 501 * lists. 502 */ 503 void 504 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool) 505 { 506 m->object = NULL; 507 m->ref_count = 0; 508 m->busy_lock = VPB_FREED; 509 m->flags = m->a.flags = 0; 510 m->phys_addr = pa; 511 m->a.queue = PQ_NONE; 512 m->psind = 0; 513 m->segind = segind; 514 m->order = VM_NFREEORDER; 515 m->pool = pool; 516 m->valid = m->dirty = 0; 517 pmap_page_init(m); 518 } 519 520 #ifndef PMAP_HAS_PAGE_ARRAY 521 static vm_paddr_t 522 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 523 { 524 vm_paddr_t new_end; 525 526 /* 527 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 528 * However, because this page is allocated from KVM, out-of-bounds 529 * accesses using the direct map will not be trapped. 530 */ 531 *vaddr += PAGE_SIZE; 532 533 /* 534 * Allocate physical memory for the page structures, and map it. 535 */ 536 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 537 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 538 VM_PROT_READ | VM_PROT_WRITE); 539 vm_page_array_size = page_range; 540 541 return (new_end); 542 } 543 #endif 544 545 /* 546 * vm_page_startup: 547 * 548 * Initializes the resident memory module. Allocates physical memory for 549 * bootstrapping UMA and some data structures that are used to manage 550 * physical pages. Initializes these structures, and populates the free 551 * page queues. 552 */ 553 vm_offset_t 554 vm_page_startup(vm_offset_t vaddr) 555 { 556 struct vm_phys_seg *seg; 557 struct vm_domain *vmd; 558 vm_page_t m; 559 char *list, *listend; 560 vm_paddr_t end, high_avail, low_avail, new_end, size; 561 vm_paddr_t page_range __unused; 562 vm_paddr_t last_pa, pa, startp, endp; 563 u_long pagecount; 564 #if MINIDUMP_PAGE_TRACKING 565 u_long vm_page_dump_size; 566 #endif 567 int biggestone, i, segind; 568 #ifdef WITNESS 569 vm_offset_t mapped; 570 int witness_size; 571 #endif 572 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 573 long ii; 574 #endif 575 #ifdef VM_FREEPOOL_LAZYINIT 576 int lazyinit; 577 #endif 578 579 vaddr = round_page(vaddr); 580 581 vm_phys_early_startup(); 582 biggestone = vm_phys_avail_largest(); 583 end = phys_avail[biggestone+1]; 584 585 /* 586 * Initialize the page and queue locks. 587 */ 588 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 589 for (i = 0; i < PA_LOCK_COUNT; i++) 590 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 591 for (i = 0; i < vm_ndomains; i++) 592 vm_page_domain_init(i); 593 594 new_end = end; 595 #ifdef WITNESS 596 witness_size = round_page(witness_startup_count()); 597 new_end -= witness_size; 598 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 599 VM_PROT_READ | VM_PROT_WRITE); 600 bzero((void *)mapped, witness_size); 601 witness_startup((void *)mapped); 602 #endif 603 604 #if MINIDUMP_PAGE_TRACKING 605 /* 606 * Allocate a bitmap to indicate that a random physical page 607 * needs to be included in a minidump. 608 * 609 * The amd64 port needs this to indicate which direct map pages 610 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 611 * 612 * However, i386 still needs this workspace internally within the 613 * minidump code. In theory, they are not needed on i386, but are 614 * included should the sf_buf code decide to use them. 615 */ 616 last_pa = 0; 617 vm_page_dump_pages = 0; 618 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 619 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 620 dump_avail[i] / PAGE_SIZE; 621 if (dump_avail[i + 1] > last_pa) 622 last_pa = dump_avail[i + 1]; 623 } 624 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 625 new_end -= vm_page_dump_size; 626 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 627 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 628 bzero((void *)vm_page_dump, vm_page_dump_size); 629 #if MINIDUMP_STARTUP_PAGE_TRACKING 630 /* 631 * Include the UMA bootstrap pages, witness pages and vm_page_dump 632 * in a crash dump. When pmap_map() uses the direct map, they are 633 * not automatically included. 634 */ 635 for (pa = new_end; pa < end; pa += PAGE_SIZE) 636 dump_add_page(pa); 637 #endif 638 #else 639 (void)last_pa; 640 #endif 641 phys_avail[biggestone + 1] = new_end; 642 #ifdef __amd64__ 643 /* 644 * Request that the physical pages underlying the message buffer be 645 * included in a crash dump. Since the message buffer is accessed 646 * through the direct map, they are not automatically included. 647 */ 648 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 649 last_pa = pa + round_page(msgbufsize); 650 while (pa < last_pa) { 651 dump_add_page(pa); 652 pa += PAGE_SIZE; 653 } 654 #endif 655 /* 656 * Compute the number of pages of memory that will be available for 657 * use, taking into account the overhead of a page structure per page. 658 * In other words, solve 659 * "available physical memory" - round_page(page_range * 660 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 661 * for page_range. 662 */ 663 low_avail = phys_avail[0]; 664 high_avail = phys_avail[1]; 665 for (i = 0; i < vm_phys_nsegs; i++) { 666 if (vm_phys_segs[i].start < low_avail) 667 low_avail = vm_phys_segs[i].start; 668 if (vm_phys_segs[i].end > high_avail) 669 high_avail = vm_phys_segs[i].end; 670 } 671 /* Skip the first chunk. It is already accounted for. */ 672 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 673 if (phys_avail[i] < low_avail) 674 low_avail = phys_avail[i]; 675 if (phys_avail[i + 1] > high_avail) 676 high_avail = phys_avail[i + 1]; 677 } 678 first_page = low_avail / PAGE_SIZE; 679 #ifdef VM_PHYSSEG_SPARSE 680 size = 0; 681 for (i = 0; i < vm_phys_nsegs; i++) 682 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 683 for (i = 0; phys_avail[i + 1] != 0; i += 2) 684 size += phys_avail[i + 1] - phys_avail[i]; 685 #elif defined(VM_PHYSSEG_DENSE) 686 size = high_avail - low_avail; 687 #else 688 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 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 if (vm_phys_avail_size(i) != 0) 751 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 752 753 /* 754 * Initialize the physical memory allocator. 755 */ 756 vm_phys_init(); 757 758 #ifdef VM_FREEPOOL_LAZYINIT 759 lazyinit = 1; 760 TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit); 761 #endif 762 763 /* 764 * Initialize the page structures and add every available page to the 765 * physical memory allocator's free lists. 766 */ 767 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 768 for (ii = 0; ii < vm_page_array_size; ii++) { 769 m = &vm_page_array[ii]; 770 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0, 771 VM_FREEPOOL_DEFAULT); 772 m->flags = PG_FICTITIOUS; 773 } 774 #endif 775 vm_cnt.v_page_count = 0; 776 for (segind = 0; segind < vm_phys_nsegs; segind++) { 777 seg = &vm_phys_segs[segind]; 778 779 /* 780 * If lazy vm_page initialization is not enabled, simply 781 * initialize all of the pages in the segment. Otherwise, we 782 * only initialize: 783 * 1. Pages not covered by phys_avail[], since they might be 784 * freed to the allocator at some future point, e.g., by 785 * kmem_bootstrap_free(). 786 * 2. The first page of each run of free pages handed to the 787 * vm_phys allocator, which in turn defers initialization 788 * of pages until they are needed. 789 * This avoids blocking the boot process for long periods, which 790 * may be relevant for VMs (which ought to boot as quickly as 791 * possible) and/or systems with large amounts of physical 792 * memory. 793 */ 794 #ifdef VM_FREEPOOL_LAZYINIT 795 if (lazyinit) { 796 startp = seg->start; 797 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 798 if (startp >= seg->end) 799 break; 800 801 if (phys_avail[i + 1] < startp) 802 continue; 803 if (phys_avail[i] <= startp) { 804 startp = phys_avail[i + 1]; 805 continue; 806 } 807 808 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 809 for (endp = MIN(phys_avail[i], seg->end); 810 startp < endp; startp += PAGE_SIZE, m++) { 811 vm_page_init_page(m, startp, segind, 812 VM_FREEPOOL_DEFAULT); 813 } 814 } 815 } else 816 #endif 817 for (m = seg->first_page, pa = seg->start; 818 pa < seg->end; m++, pa += PAGE_SIZE) { 819 vm_page_init_page(m, pa, segind, 820 VM_FREEPOOL_DEFAULT); 821 } 822 823 /* 824 * Add the segment's pages that are covered by one of 825 * phys_avail's ranges to the free lists. 826 */ 827 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 828 if (seg->end <= phys_avail[i] || 829 seg->start >= phys_avail[i + 1]) 830 continue; 831 832 startp = MAX(seg->start, phys_avail[i]); 833 endp = MIN(seg->end, phys_avail[i + 1]); 834 pagecount = (u_long)atop(endp - startp); 835 if (pagecount == 0) 836 continue; 837 838 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 839 #ifdef VM_FREEPOOL_LAZYINIT 840 if (lazyinit) { 841 vm_page_init_page(m, startp, segind, 842 VM_FREEPOOL_LAZYINIT); 843 } 844 #endif 845 vmd = VM_DOMAIN(seg->domain); 846 vm_domain_free_lock(vmd); 847 vm_phys_enqueue_contig(m, pagecount); 848 vm_domain_free_unlock(vmd); 849 vm_domain_freecnt_inc(vmd, pagecount); 850 vm_cnt.v_page_count += (u_int)pagecount; 851 vmd->vmd_page_count += (u_int)pagecount; 852 vmd->vmd_segs |= 1UL << segind; 853 } 854 } 855 856 /* 857 * Remove blacklisted pages from the physical memory allocator. 858 */ 859 TAILQ_INIT(&blacklist_head); 860 vm_page_blacklist_load(&list, &listend); 861 vm_page_blacklist_check(list, listend); 862 863 list = kern_getenv("vm.blacklist"); 864 vm_page_blacklist_check(list, NULL); 865 866 freeenv(list); 867 #if VM_NRESERVLEVEL > 0 868 /* 869 * Initialize the reservation management system. 870 */ 871 vm_reserv_init(); 872 #endif 873 874 return (vaddr); 875 } 876 877 void 878 vm_page_reference(vm_page_t m) 879 { 880 881 vm_page_aflag_set(m, PGA_REFERENCED); 882 } 883 884 /* 885 * vm_page_trybusy 886 * 887 * Helper routine for grab functions to trylock busy. 888 * 889 * Returns true on success and false on failure. 890 */ 891 static bool 892 vm_page_trybusy(vm_page_t m, int allocflags) 893 { 894 895 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 896 return (vm_page_trysbusy(m)); 897 else 898 return (vm_page_tryxbusy(m)); 899 } 900 901 /* 902 * vm_page_tryacquire 903 * 904 * Helper routine for grab functions to trylock busy and wire. 905 * 906 * Returns true on success and false on failure. 907 */ 908 static inline bool 909 vm_page_tryacquire(vm_page_t m, int allocflags) 910 { 911 bool locked; 912 913 locked = vm_page_trybusy(m, allocflags); 914 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 915 vm_page_wire(m); 916 return (locked); 917 } 918 919 /* 920 * vm_page_busy_acquire: 921 * 922 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 923 * and drop the object lock if necessary. 924 */ 925 bool 926 vm_page_busy_acquire(vm_page_t m, int allocflags) 927 { 928 vm_object_t obj; 929 bool locked; 930 931 /* 932 * The page-specific object must be cached because page 933 * identity can change during the sleep, causing the 934 * re-lock of a different object. 935 * It is assumed that a reference to the object is already 936 * held by the callers. 937 */ 938 obj = atomic_load_ptr(&m->object); 939 for (;;) { 940 if (vm_page_tryacquire(m, allocflags)) 941 return (true); 942 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 943 return (false); 944 if (obj != NULL) 945 locked = VM_OBJECT_WOWNED(obj); 946 else 947 locked = false; 948 MPASS(locked || vm_page_wired(m)); 949 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 950 locked) && locked) 951 VM_OBJECT_WLOCK(obj); 952 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 953 return (false); 954 KASSERT(m->object == obj || m->object == NULL, 955 ("vm_page_busy_acquire: page %p does not belong to %p", 956 m, obj)); 957 } 958 } 959 960 /* 961 * vm_page_busy_downgrade: 962 * 963 * Downgrade an exclusive busy page into a single shared busy page. 964 */ 965 void 966 vm_page_busy_downgrade(vm_page_t m) 967 { 968 u_int x; 969 970 vm_page_assert_xbusied(m); 971 972 x = vm_page_busy_fetch(m); 973 for (;;) { 974 if (atomic_fcmpset_rel_int(&m->busy_lock, 975 &x, VPB_SHARERS_WORD(1))) 976 break; 977 } 978 if ((x & VPB_BIT_WAITERS) != 0) 979 wakeup(m); 980 } 981 982 /* 983 * 984 * vm_page_busy_tryupgrade: 985 * 986 * Attempt to upgrade a single shared busy into an exclusive busy. 987 */ 988 int 989 vm_page_busy_tryupgrade(vm_page_t m) 990 { 991 u_int ce, x; 992 993 vm_page_assert_sbusied(m); 994 995 x = vm_page_busy_fetch(m); 996 ce = VPB_CURTHREAD_EXCLUSIVE; 997 for (;;) { 998 if (VPB_SHARERS(x) > 1) 999 return (0); 1000 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1001 ("vm_page_busy_tryupgrade: invalid lock state")); 1002 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 1003 ce | (x & VPB_BIT_WAITERS))) 1004 continue; 1005 return (1); 1006 } 1007 } 1008 1009 /* 1010 * vm_page_sbusied: 1011 * 1012 * Return a positive value if the page is shared busied, 0 otherwise. 1013 */ 1014 int 1015 vm_page_sbusied(vm_page_t m) 1016 { 1017 u_int x; 1018 1019 x = vm_page_busy_fetch(m); 1020 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1021 } 1022 1023 /* 1024 * vm_page_sunbusy: 1025 * 1026 * Shared unbusy a page. 1027 */ 1028 void 1029 vm_page_sunbusy(vm_page_t m) 1030 { 1031 u_int x; 1032 1033 vm_page_assert_sbusied(m); 1034 1035 x = vm_page_busy_fetch(m); 1036 for (;;) { 1037 KASSERT(x != VPB_FREED, 1038 ("vm_page_sunbusy: Unlocking freed page.")); 1039 if (VPB_SHARERS(x) > 1) { 1040 if (atomic_fcmpset_int(&m->busy_lock, &x, 1041 x - VPB_ONE_SHARER)) 1042 break; 1043 continue; 1044 } 1045 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1046 ("vm_page_sunbusy: invalid lock state")); 1047 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1048 continue; 1049 if ((x & VPB_BIT_WAITERS) == 0) 1050 break; 1051 wakeup(m); 1052 break; 1053 } 1054 } 1055 1056 /* 1057 * vm_page_busy_sleep: 1058 * 1059 * Sleep if the page is busy, using the page pointer as wchan. 1060 * This is used to implement the hard-path of the busying mechanism. 1061 * 1062 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1063 * will not sleep if the page is shared-busy. 1064 * 1065 * The object lock must be held on entry. 1066 * 1067 * Returns true if it slept and dropped the object lock, or false 1068 * if there was no sleep and the lock is still held. 1069 */ 1070 bool 1071 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) 1072 { 1073 vm_object_t obj; 1074 1075 obj = m->object; 1076 VM_OBJECT_ASSERT_LOCKED(obj); 1077 1078 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, 1079 true)); 1080 } 1081 1082 /* 1083 * vm_page_busy_sleep_unlocked: 1084 * 1085 * Sleep if the page is busy, using the page pointer as wchan. 1086 * This is used to implement the hard-path of busying mechanism. 1087 * 1088 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1089 * will not sleep if the page is shared-busy. 1090 * 1091 * The object lock must not be held on entry. The operation will 1092 * return if the page changes identity. 1093 */ 1094 void 1095 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1096 const char *wmesg, int allocflags) 1097 { 1098 VM_OBJECT_ASSERT_UNLOCKED(obj); 1099 1100 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); 1101 } 1102 1103 /* 1104 * _vm_page_busy_sleep: 1105 * 1106 * Internal busy sleep function. Verifies the page identity and 1107 * lockstate against parameters. Returns true if it sleeps and 1108 * false otherwise. 1109 * 1110 * allocflags uses VM_ALLOC_* flags to specify the lock required. 1111 * 1112 * If locked is true the lock will be dropped for any true returns 1113 * and held for any false returns. 1114 */ 1115 static bool 1116 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1117 const char *wmesg, int allocflags, bool locked) 1118 { 1119 bool xsleep; 1120 u_int x; 1121 1122 /* 1123 * If the object is busy we must wait for that to drain to zero 1124 * before trying the page again. 1125 */ 1126 if (obj != NULL && vm_object_busied(obj)) { 1127 if (locked) 1128 VM_OBJECT_DROP(obj); 1129 vm_object_busy_wait(obj, wmesg); 1130 return (true); 1131 } 1132 1133 if (!vm_page_busied(m)) 1134 return (false); 1135 1136 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1137 sleepq_lock(m); 1138 x = vm_page_busy_fetch(m); 1139 do { 1140 /* 1141 * If the page changes objects or becomes unlocked we can 1142 * simply return. 1143 */ 1144 if (x == VPB_UNBUSIED || 1145 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1146 m->object != obj || m->pindex != pindex) { 1147 sleepq_release(m); 1148 return (false); 1149 } 1150 if ((x & VPB_BIT_WAITERS) != 0) 1151 break; 1152 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1153 if (locked) 1154 VM_OBJECT_DROP(obj); 1155 DROP_GIANT(); 1156 sleepq_add(m, NULL, wmesg, 0, 0); 1157 sleepq_wait(m, PVM); 1158 PICKUP_GIANT(); 1159 return (true); 1160 } 1161 1162 /* 1163 * vm_page_trysbusy: 1164 * 1165 * Try to shared busy a page. 1166 * If the operation succeeds 1 is returned otherwise 0. 1167 * The operation never sleeps. 1168 */ 1169 int 1170 vm_page_trysbusy(vm_page_t m) 1171 { 1172 vm_object_t obj; 1173 u_int x; 1174 1175 obj = m->object; 1176 x = vm_page_busy_fetch(m); 1177 for (;;) { 1178 if ((x & VPB_BIT_SHARED) == 0) 1179 return (0); 1180 /* 1181 * Reduce the window for transient busies that will trigger 1182 * false negatives in vm_page_ps_test(). 1183 */ 1184 if (obj != NULL && vm_object_busied(obj)) 1185 return (0); 1186 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1187 x + VPB_ONE_SHARER)) 1188 break; 1189 } 1190 1191 /* Refetch the object now that we're guaranteed that it is stable. */ 1192 obj = m->object; 1193 if (obj != NULL && vm_object_busied(obj)) { 1194 vm_page_sunbusy(m); 1195 return (0); 1196 } 1197 return (1); 1198 } 1199 1200 /* 1201 * vm_page_tryxbusy: 1202 * 1203 * Try to exclusive busy a page. 1204 * If the operation succeeds 1 is returned otherwise 0. 1205 * The operation never sleeps. 1206 */ 1207 int 1208 vm_page_tryxbusy(vm_page_t m) 1209 { 1210 vm_object_t obj; 1211 1212 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1213 VPB_CURTHREAD_EXCLUSIVE) == 0) 1214 return (0); 1215 1216 obj = m->object; 1217 if (obj != NULL && vm_object_busied(obj)) { 1218 vm_page_xunbusy(m); 1219 return (0); 1220 } 1221 return (1); 1222 } 1223 1224 static void 1225 vm_page_xunbusy_hard_tail(vm_page_t m) 1226 { 1227 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1228 /* Wake the waiter. */ 1229 wakeup(m); 1230 } 1231 1232 /* 1233 * vm_page_xunbusy_hard: 1234 * 1235 * Called when unbusy has failed because there is a waiter. 1236 */ 1237 void 1238 vm_page_xunbusy_hard(vm_page_t m) 1239 { 1240 vm_page_assert_xbusied(m); 1241 vm_page_xunbusy_hard_tail(m); 1242 } 1243 1244 void 1245 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1246 { 1247 vm_page_assert_xbusied_unchecked(m); 1248 vm_page_xunbusy_hard_tail(m); 1249 } 1250 1251 static void 1252 vm_page_busy_free(vm_page_t m) 1253 { 1254 u_int x; 1255 1256 atomic_thread_fence_rel(); 1257 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1258 if ((x & VPB_BIT_WAITERS) != 0) 1259 wakeup(m); 1260 } 1261 1262 /* 1263 * vm_page_unhold_pages: 1264 * 1265 * Unhold each of the pages that is referenced by the given array. 1266 */ 1267 void 1268 vm_page_unhold_pages(vm_page_t *ma, int count) 1269 { 1270 1271 for (; count != 0; count--) { 1272 vm_page_unwire(*ma, PQ_ACTIVE); 1273 ma++; 1274 } 1275 } 1276 1277 vm_page_t 1278 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1279 { 1280 vm_page_t m; 1281 1282 #ifdef VM_PHYSSEG_SPARSE 1283 m = vm_phys_paddr_to_vm_page(pa); 1284 if (m == NULL) 1285 m = vm_phys_fictitious_to_vm_page(pa); 1286 return (m); 1287 #elif defined(VM_PHYSSEG_DENSE) 1288 long pi; 1289 1290 pi = atop(pa); 1291 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1292 m = &vm_page_array[pi - first_page]; 1293 return (m); 1294 } 1295 return (vm_phys_fictitious_to_vm_page(pa)); 1296 #else 1297 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1298 #endif 1299 } 1300 1301 /* 1302 * vm_page_getfake: 1303 * 1304 * Create a fictitious page with the specified physical address and 1305 * memory attribute. The memory attribute is the only the machine- 1306 * dependent aspect of a fictitious page that must be initialized. 1307 */ 1308 vm_page_t 1309 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1310 { 1311 vm_page_t m; 1312 1313 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1314 vm_page_initfake(m, paddr, memattr); 1315 return (m); 1316 } 1317 1318 void 1319 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1320 { 1321 1322 if ((m->flags & PG_FICTITIOUS) != 0) { 1323 /* 1324 * The page's memattr might have changed since the 1325 * previous initialization. Update the pmap to the 1326 * new memattr. 1327 */ 1328 goto memattr; 1329 } 1330 m->phys_addr = paddr; 1331 m->a.queue = PQ_NONE; 1332 /* Fictitious pages don't use "segind". */ 1333 m->flags = PG_FICTITIOUS; 1334 /* Fictitious pages don't use "order" or "pool". */ 1335 m->oflags = VPO_UNMANAGED; 1336 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1337 /* Fictitious pages are unevictable. */ 1338 m->ref_count = 1; 1339 pmap_page_init(m); 1340 memattr: 1341 pmap_page_set_memattr(m, memattr); 1342 } 1343 1344 /* 1345 * vm_page_putfake: 1346 * 1347 * Release a fictitious page. 1348 */ 1349 void 1350 vm_page_putfake(vm_page_t m) 1351 { 1352 1353 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1354 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1355 ("vm_page_putfake: bad page %p", m)); 1356 vm_page_assert_xbusied(m); 1357 vm_page_busy_free(m); 1358 uma_zfree(fakepg_zone, m); 1359 } 1360 1361 /* 1362 * vm_page_updatefake: 1363 * 1364 * Update the given fictitious page to the specified physical address and 1365 * memory attribute. 1366 */ 1367 void 1368 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1369 { 1370 1371 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1372 ("vm_page_updatefake: bad page %p", m)); 1373 m->phys_addr = paddr; 1374 pmap_page_set_memattr(m, memattr); 1375 } 1376 1377 /* 1378 * vm_page_free: 1379 * 1380 * Free a page. 1381 */ 1382 void 1383 vm_page_free(vm_page_t m) 1384 { 1385 1386 m->flags &= ~PG_ZERO; 1387 vm_page_free_toq(m); 1388 } 1389 1390 /* 1391 * vm_page_free_zero: 1392 * 1393 * Free a page to the zerod-pages queue 1394 */ 1395 void 1396 vm_page_free_zero(vm_page_t m) 1397 { 1398 1399 m->flags |= PG_ZERO; 1400 vm_page_free_toq(m); 1401 } 1402 1403 /* 1404 * Unbusy and handle the page queueing for a page from a getpages request that 1405 * was optionally read ahead or behind. 1406 */ 1407 void 1408 vm_page_readahead_finish(vm_page_t m) 1409 { 1410 1411 /* We shouldn't put invalid pages on queues. */ 1412 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1413 1414 /* 1415 * Since the page is not the actually needed one, whether it should 1416 * be activated or deactivated is not obvious. Empirical results 1417 * have shown that deactivating the page is usually the best choice, 1418 * unless the page is wanted by another thread. 1419 */ 1420 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1421 vm_page_activate(m); 1422 else 1423 vm_page_deactivate(m); 1424 vm_page_xunbusy_unchecked(m); 1425 } 1426 1427 /* 1428 * Destroy the identity of an invalid page and free it if possible. 1429 * This is intended to be used when reading a page from backing store fails. 1430 */ 1431 void 1432 vm_page_free_invalid(vm_page_t m) 1433 { 1434 1435 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1436 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1437 KASSERT(m->object != NULL, ("page %p has no object", m)); 1438 VM_OBJECT_ASSERT_WLOCKED(m->object); 1439 1440 /* 1441 * We may be attempting to free the page as part of the handling for an 1442 * I/O error, in which case the page was xbusied by a different thread. 1443 */ 1444 vm_page_xbusy_claim(m); 1445 1446 /* 1447 * If someone has wired this page while the object lock 1448 * was not held, then the thread that unwires is responsible 1449 * for freeing the page. Otherwise just free the page now. 1450 * The wire count of this unmapped page cannot change while 1451 * we have the page xbusy and the page's object wlocked. 1452 */ 1453 if (vm_page_remove(m)) 1454 vm_page_free(m); 1455 } 1456 1457 /* 1458 * vm_page_dirty_KBI: [ internal use only ] 1459 * 1460 * Set all bits in the page's dirty field. 1461 * 1462 * The object containing the specified page must be locked if the 1463 * call is made from the machine-independent layer. 1464 * 1465 * See vm_page_clear_dirty_mask(). 1466 * 1467 * This function should only be called by vm_page_dirty(). 1468 */ 1469 void 1470 vm_page_dirty_KBI(vm_page_t m) 1471 { 1472 1473 /* Refer to this operation by its public name. */ 1474 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1475 m->dirty = VM_PAGE_BITS_ALL; 1476 } 1477 1478 /* 1479 * Insert the given page into the given object at the given pindex. mpred is 1480 * used for memq linkage. From vm_page_insert, lookup is true, mpred is 1481 * initially NULL, and this procedure looks it up. From vm_page_insert_after 1482 * and vm_page_iter_insert, lookup is false and mpred is known to the caller 1483 * to be valid, and may be NULL if this will be the page with the lowest 1484 * pindex. 1485 * 1486 * The procedure is marked __always_inline to suggest to the compiler to 1487 * eliminate the lookup parameter and the associated alternate branch. 1488 */ 1489 static __always_inline int 1490 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1491 struct pctrie_iter *pages, bool iter, vm_page_t mpred, bool lookup) 1492 { 1493 int error; 1494 1495 VM_OBJECT_ASSERT_WLOCKED(object); 1496 KASSERT(m->object == NULL, 1497 ("vm_page_insert: page %p already inserted", m)); 1498 1499 /* 1500 * Record the object/offset pair in this page. 1501 */ 1502 m->object = object; 1503 m->pindex = pindex; 1504 m->ref_count |= VPRC_OBJREF; 1505 1506 /* 1507 * Add this page to the object's radix tree, and look up mpred if 1508 * needed. 1509 */ 1510 if (iter) { 1511 KASSERT(!lookup, ("%s: cannot lookup mpred", __func__)); 1512 error = vm_radix_iter_insert(pages, m); 1513 } else if (lookup) 1514 error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred); 1515 else 1516 error = vm_radix_insert(&object->rtree, m); 1517 if (__predict_false(error != 0)) { 1518 m->object = NULL; 1519 m->pindex = 0; 1520 m->ref_count &= ~VPRC_OBJREF; 1521 return (1); 1522 } 1523 1524 /* 1525 * Now link into the object's ordered list of backed pages. 1526 */ 1527 vm_page_insert_radixdone(m, object, mpred); 1528 vm_pager_page_inserted(object, m); 1529 return (0); 1530 } 1531 1532 /* 1533 * vm_page_insert: [ internal use only ] 1534 * 1535 * Inserts the given mem entry into the object and object list. 1536 * 1537 * The object must be locked. 1538 */ 1539 int 1540 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1541 { 1542 return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL, 1543 true)); 1544 } 1545 1546 /* 1547 * vm_page_insert_after: 1548 * 1549 * Inserts the page "m" into the specified object at offset "pindex". 1550 * 1551 * The page "mpred" must immediately precede the offset "pindex" within 1552 * the specified object. 1553 * 1554 * The object must be locked. 1555 */ 1556 static int 1557 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1558 vm_page_t mpred) 1559 { 1560 return (vm_page_insert_lookup(m, object, pindex, NULL, false, mpred, 1561 false)); 1562 } 1563 1564 /* 1565 * vm_page_iter_insert: 1566 * 1567 * Tries to insert the page "m" into the specified object at offset 1568 * "pindex" using the iterator "pages". Returns 0 if the insertion was 1569 * successful. 1570 * 1571 * The page "mpred" must immediately precede the offset "pindex" within 1572 * the specified object. 1573 * 1574 * The object must be locked. 1575 */ 1576 static int 1577 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object, 1578 vm_pindex_t pindex, vm_page_t mpred) 1579 { 1580 return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred, 1581 false)); 1582 } 1583 1584 /* 1585 * vm_page_insert_radixdone: 1586 * 1587 * Complete page "m" insertion into the specified object after the 1588 * radix trie hooking. 1589 * 1590 * The page "mpred" must precede the offset "m->pindex" within the 1591 * specified object. 1592 * 1593 * The object must be locked. 1594 */ 1595 static void 1596 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1597 { 1598 1599 VM_OBJECT_ASSERT_WLOCKED(object); 1600 KASSERT(object != NULL && m->object == object, 1601 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1602 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1603 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1604 if (mpred != NULL) { 1605 KASSERT(mpred->object == object, 1606 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1607 KASSERT(mpred->pindex < m->pindex, 1608 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1609 KASSERT(TAILQ_NEXT(mpred, listq) == NULL || 1610 m->pindex < TAILQ_NEXT(mpred, listq)->pindex, 1611 ("vm_page_insert_radixdone: pindex doesn't precede msucc")); 1612 } else { 1613 KASSERT(TAILQ_EMPTY(&object->memq) || 1614 m->pindex < TAILQ_FIRST(&object->memq)->pindex, 1615 ("vm_page_insert_radixdone: no mpred but not first page")); 1616 } 1617 1618 if (mpred != NULL) 1619 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1620 else 1621 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1622 1623 /* 1624 * Show that the object has one more resident page. 1625 */ 1626 object->resident_page_count++; 1627 1628 /* 1629 * Hold the vnode until the last page is released. 1630 */ 1631 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1632 vhold(object->handle); 1633 1634 /* 1635 * Since we are inserting a new and possibly dirty page, 1636 * update the object's generation count. 1637 */ 1638 if (pmap_page_is_write_mapped(m)) 1639 vm_object_set_writeable_dirty(object); 1640 } 1641 1642 /* 1643 * vm_page_remove_radixdone 1644 * 1645 * Complete page "m" removal from the specified object after the radix trie 1646 * unhooking. 1647 * 1648 * The caller is responsible for updating the page's fields to reflect this 1649 * removal. 1650 */ 1651 static void 1652 vm_page_remove_radixdone(vm_page_t m) 1653 { 1654 vm_object_t object; 1655 1656 vm_page_assert_xbusied(m); 1657 object = m->object; 1658 VM_OBJECT_ASSERT_WLOCKED(object); 1659 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1660 ("page %p is missing its object ref", m)); 1661 1662 /* Deferred free of swap space. */ 1663 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1664 vm_pager_page_unswapped(m); 1665 1666 vm_pager_page_removed(object, m); 1667 m->object = NULL; 1668 1669 /* 1670 * Now remove from the object's list of backed pages. 1671 */ 1672 TAILQ_REMOVE(&object->memq, m, listq); 1673 1674 /* 1675 * And show that the object has one fewer resident page. 1676 */ 1677 object->resident_page_count--; 1678 1679 /* 1680 * The vnode may now be recycled. 1681 */ 1682 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1683 vdrop(object->handle); 1684 } 1685 1686 /* 1687 * vm_page_free_object_prep: 1688 * 1689 * Disassociates the given page from its VM object. 1690 * 1691 * The object must be locked, and the page must be xbusy. 1692 */ 1693 static void 1694 vm_page_free_object_prep(vm_page_t m) 1695 { 1696 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 1697 ((m->object->flags & OBJ_UNMANAGED) != 0), 1698 ("%s: managed flag mismatch for page %p", 1699 __func__, m)); 1700 vm_page_assert_xbusied(m); 1701 1702 /* 1703 * The object reference can be released without an atomic 1704 * operation. 1705 */ 1706 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 1707 m->ref_count == VPRC_OBJREF, 1708 ("%s: page %p has unexpected ref_count %u", 1709 __func__, m, m->ref_count)); 1710 vm_page_remove_radixdone(m); 1711 m->ref_count -= VPRC_OBJREF; 1712 } 1713 1714 /* 1715 * vm_page_iter_free: 1716 * 1717 * Free the given page, and use the iterator to remove it from the radix 1718 * tree. 1719 */ 1720 void 1721 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m) 1722 { 1723 vm_radix_iter_remove(pages); 1724 vm_page_free_object_prep(m); 1725 vm_page_xunbusy(m); 1726 m->flags &= ~PG_ZERO; 1727 vm_page_free_toq(m); 1728 } 1729 1730 /* 1731 * vm_page_remove: 1732 * 1733 * Removes the specified page from its containing object, but does not 1734 * invalidate any backing storage. Returns true if the object's reference 1735 * was the last reference to the page, and false otherwise. 1736 * 1737 * The object must be locked and the page must be exclusively busied. 1738 * The exclusive busy will be released on return. If this is not the 1739 * final ref and the caller does not hold a wire reference it may not 1740 * continue to access the page. 1741 */ 1742 bool 1743 vm_page_remove(vm_page_t m) 1744 { 1745 bool dropped; 1746 1747 dropped = vm_page_remove_xbusy(m); 1748 vm_page_xunbusy(m); 1749 1750 return (dropped); 1751 } 1752 1753 /* 1754 * vm_page_iter_remove: 1755 * 1756 * Remove the current page, and use the iterator to remove it from the 1757 * radix tree. 1758 */ 1759 bool 1760 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m) 1761 { 1762 bool dropped; 1763 1764 vm_radix_iter_remove(pages); 1765 vm_page_remove_radixdone(m); 1766 dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1767 vm_page_xunbusy(m); 1768 1769 return (dropped); 1770 } 1771 1772 /* 1773 * vm_page_radix_remove 1774 * 1775 * Removes the specified page from the radix tree. 1776 */ 1777 static void 1778 vm_page_radix_remove(vm_page_t m) 1779 { 1780 vm_page_t mrem __diagused; 1781 1782 mrem = vm_radix_remove(&m->object->rtree, m->pindex); 1783 KASSERT(mrem == m, 1784 ("removed page %p, expected page %p", mrem, m)); 1785 } 1786 1787 /* 1788 * vm_page_remove_xbusy 1789 * 1790 * Removes the page but leaves the xbusy held. Returns true if this 1791 * removed the final ref and false otherwise. 1792 */ 1793 bool 1794 vm_page_remove_xbusy(vm_page_t m) 1795 { 1796 1797 vm_page_radix_remove(m); 1798 vm_page_remove_radixdone(m); 1799 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1800 } 1801 1802 /* 1803 * vm_page_lookup: 1804 * 1805 * Returns the page associated with the object/offset 1806 * pair specified; if none is found, NULL is returned. 1807 * 1808 * The object must be locked. 1809 */ 1810 vm_page_t 1811 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1812 { 1813 1814 VM_OBJECT_ASSERT_LOCKED(object); 1815 return (vm_radix_lookup(&object->rtree, pindex)); 1816 } 1817 1818 /* 1819 * vm_page_iter_init: 1820 * 1821 * Initialize iterator for vm pages. 1822 */ 1823 void 1824 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object) 1825 { 1826 1827 vm_radix_iter_init(pages, &object->rtree); 1828 } 1829 1830 /* 1831 * vm_page_iter_init: 1832 * 1833 * Initialize iterator for vm pages. 1834 */ 1835 void 1836 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object, 1837 vm_pindex_t limit) 1838 { 1839 1840 vm_radix_iter_limit_init(pages, &object->rtree, limit); 1841 } 1842 1843 /* 1844 * vm_page_iter_lookup: 1845 * 1846 * Returns the page associated with the object/offset pair specified, and 1847 * stores the path to its position; if none is found, NULL is returned. 1848 * 1849 * The iter pctrie must be locked. 1850 */ 1851 vm_page_t 1852 vm_page_iter_lookup(struct pctrie_iter *pages, vm_pindex_t pindex) 1853 { 1854 1855 return (vm_radix_iter_lookup(pages, pindex)); 1856 } 1857 1858 /* 1859 * vm_page_lookup_unlocked: 1860 * 1861 * Returns the page associated with the object/offset pair specified; 1862 * if none is found, NULL is returned. The page may be no longer be 1863 * present in the object at the time that this function returns. Only 1864 * useful for opportunistic checks such as inmem(). 1865 */ 1866 vm_page_t 1867 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1868 { 1869 1870 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1871 } 1872 1873 /* 1874 * vm_page_relookup: 1875 * 1876 * Returns a page that must already have been busied by 1877 * the caller. Used for bogus page replacement. 1878 */ 1879 vm_page_t 1880 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1881 { 1882 vm_page_t m; 1883 1884 m = vm_page_lookup_unlocked(object, pindex); 1885 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1886 m->object == object && m->pindex == pindex, 1887 ("vm_page_relookup: Invalid page %p", m)); 1888 return (m); 1889 } 1890 1891 /* 1892 * This should only be used by lockless functions for releasing transient 1893 * incorrect acquires. The page may have been freed after we acquired a 1894 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1895 * further to do. 1896 */ 1897 static void 1898 vm_page_busy_release(vm_page_t m) 1899 { 1900 u_int x; 1901 1902 x = vm_page_busy_fetch(m); 1903 for (;;) { 1904 if (x == VPB_FREED) 1905 break; 1906 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1907 if (atomic_fcmpset_int(&m->busy_lock, &x, 1908 x - VPB_ONE_SHARER)) 1909 break; 1910 continue; 1911 } 1912 KASSERT((x & VPB_BIT_SHARED) != 0 || 1913 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1914 ("vm_page_busy_release: %p xbusy not owned.", m)); 1915 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1916 continue; 1917 if ((x & VPB_BIT_WAITERS) != 0) 1918 wakeup(m); 1919 break; 1920 } 1921 } 1922 1923 /* 1924 * vm_page_find_least: 1925 * 1926 * Returns the page associated with the object with least pindex 1927 * greater than or equal to the parameter pindex, or NULL. 1928 * 1929 * The object must be locked. 1930 */ 1931 vm_page_t 1932 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1933 { 1934 vm_page_t m; 1935 1936 VM_OBJECT_ASSERT_LOCKED(object); 1937 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1938 m = vm_radix_lookup_ge(&object->rtree, pindex); 1939 return (m); 1940 } 1941 1942 /* 1943 * vm_page_iter_lookup_ge: 1944 * 1945 * Returns the page associated with the object with least pindex 1946 * greater than or equal to the parameter pindex, or NULL. Initializes the 1947 * iterator to point to that page. 1948 * 1949 * The iter pctrie must be locked. 1950 */ 1951 vm_page_t 1952 vm_page_iter_lookup_ge(struct pctrie_iter *pages, vm_pindex_t pindex) 1953 { 1954 1955 return (vm_radix_iter_lookup_ge(pages, pindex)); 1956 } 1957 1958 /* 1959 * Returns the given page's successor (by pindex) within the object if it is 1960 * resident; if none is found, NULL is returned. 1961 * 1962 * The object must be locked. 1963 */ 1964 vm_page_t 1965 vm_page_next(vm_page_t m) 1966 { 1967 vm_page_t next; 1968 1969 VM_OBJECT_ASSERT_LOCKED(m->object); 1970 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1971 MPASS(next->object == m->object); 1972 if (next->pindex != m->pindex + 1) 1973 next = NULL; 1974 } 1975 return (next); 1976 } 1977 1978 /* 1979 * Returns the given page's predecessor (by pindex) within the object if it is 1980 * resident; if none is found, NULL is returned. 1981 * 1982 * The object must be locked. 1983 */ 1984 vm_page_t 1985 vm_page_prev(vm_page_t m) 1986 { 1987 vm_page_t prev; 1988 1989 VM_OBJECT_ASSERT_LOCKED(m->object); 1990 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1991 MPASS(prev->object == m->object); 1992 if (prev->pindex != m->pindex - 1) 1993 prev = NULL; 1994 } 1995 return (prev); 1996 } 1997 1998 /* 1999 * Uses the page mnew as a replacement for an existing page at index 2000 * pindex which must be already present in the object. 2001 * 2002 * Both pages must be exclusively busied on enter. The old page is 2003 * unbusied on exit. 2004 * 2005 * A return value of true means mold is now free. If this is not the 2006 * final ref and the caller does not hold a wire reference it may not 2007 * continue to access the page. 2008 */ 2009 static bool 2010 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 2011 vm_page_t mold) 2012 { 2013 vm_page_t mret __diagused; 2014 bool dropped; 2015 2016 VM_OBJECT_ASSERT_WLOCKED(object); 2017 vm_page_assert_xbusied(mold); 2018 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 2019 ("vm_page_replace: page %p already in object", mnew)); 2020 2021 /* 2022 * This function mostly follows vm_page_insert() and 2023 * vm_page_remove() without the radix, object count and vnode 2024 * dance. Double check such functions for more comments. 2025 */ 2026 2027 mnew->object = object; 2028 mnew->pindex = pindex; 2029 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 2030 mret = vm_radix_replace(&object->rtree, mnew); 2031 KASSERT(mret == mold, 2032 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 2033 KASSERT((mold->oflags & VPO_UNMANAGED) == 2034 (mnew->oflags & VPO_UNMANAGED), 2035 ("vm_page_replace: mismatched VPO_UNMANAGED")); 2036 2037 /* Keep the resident page list in sorted order. */ 2038 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 2039 TAILQ_REMOVE(&object->memq, mold, listq); 2040 mold->object = NULL; 2041 2042 /* 2043 * The object's resident_page_count does not change because we have 2044 * swapped one page for another, but the generation count should 2045 * change if the page is dirty. 2046 */ 2047 if (pmap_page_is_write_mapped(mnew)) 2048 vm_object_set_writeable_dirty(object); 2049 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 2050 vm_page_xunbusy(mold); 2051 2052 return (dropped); 2053 } 2054 2055 void 2056 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 2057 vm_page_t mold) 2058 { 2059 2060 vm_page_assert_xbusied(mnew); 2061 2062 if (vm_page_replace_hold(mnew, object, pindex, mold)) 2063 vm_page_free(mold); 2064 } 2065 2066 /* 2067 * vm_page_iter_rename: 2068 * 2069 * Tries to move the specified page from its current object to a new object 2070 * and pindex, using the given iterator to remove the page from its current 2071 * object. Returns true if the move was successful, and false if the move 2072 * was aborted due to a failed memory allocation. 2073 * 2074 * Panics if a page already resides in the new object at the new pindex. 2075 * 2076 * Note: swap associated with the page must be invalidated by the move. We 2077 * have to do this for several reasons: (1) we aren't freeing the 2078 * page, (2) we are dirtying the page, (3) the VM system is probably 2079 * moving the page from object A to B, and will then later move 2080 * the backing store from A to B and we can't have a conflict. 2081 * 2082 * Note: we *always* dirty the page. It is necessary both for the 2083 * fact that we moved it, and because we may be invalidating 2084 * swap. 2085 * 2086 * The objects must be locked. 2087 */ 2088 bool 2089 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m, 2090 vm_object_t new_object, vm_pindex_t new_pindex) 2091 { 2092 vm_page_t mpred; 2093 vm_pindex_t opidx; 2094 2095 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 2096 ("%s: page %p is missing object ref", __func__, m)); 2097 VM_OBJECT_ASSERT_WLOCKED(m->object); 2098 VM_OBJECT_ASSERT_WLOCKED(new_object); 2099 2100 /* 2101 * Create a custom version of vm_page_insert() which does not depend 2102 * by m_prev and can cheat on the implementation aspects of the 2103 * function. 2104 */ 2105 opidx = m->pindex; 2106 m->pindex = new_pindex; 2107 if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) { 2108 m->pindex = opidx; 2109 return (false); 2110 } 2111 2112 /* 2113 * The operation cannot fail anymore. The removal must happen before 2114 * the listq iterator is tainted. 2115 */ 2116 m->pindex = opidx; 2117 vm_radix_iter_remove(old_pages); 2118 vm_page_remove_radixdone(m); 2119 2120 /* Return back to the new pindex to complete vm_page_insert(). */ 2121 m->pindex = new_pindex; 2122 m->object = new_object; 2123 2124 vm_page_insert_radixdone(m, new_object, mpred); 2125 vm_page_dirty(m); 2126 vm_pager_page_inserted(new_object, m); 2127 return (true); 2128 } 2129 2130 /* 2131 * vm_page_mpred: 2132 * 2133 * Return the greatest page of the object with index <= pindex, 2134 * or NULL, if there is none. Assumes object lock is held. 2135 */ 2136 vm_page_t 2137 vm_page_mpred(vm_object_t object, vm_pindex_t pindex) 2138 { 2139 return (vm_radix_lookup_le(&object->rtree, pindex)); 2140 } 2141 2142 /* 2143 * vm_page_alloc: 2144 * 2145 * Allocate and return a page that is associated with the specified 2146 * object and offset pair. By default, this page is exclusive busied. 2147 * 2148 * The caller must always specify an allocation class. 2149 * 2150 * allocation classes: 2151 * VM_ALLOC_NORMAL normal process request 2152 * VM_ALLOC_SYSTEM system *really* needs a page 2153 * VM_ALLOC_INTERRUPT interrupt time request 2154 * 2155 * optional allocation flags: 2156 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2157 * intends to allocate 2158 * VM_ALLOC_NOBUSY do not exclusive busy the page 2159 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2160 * VM_ALLOC_SBUSY shared busy the allocated page 2161 * VM_ALLOC_WIRED wire the allocated page 2162 * VM_ALLOC_ZERO prefer a zeroed page 2163 */ 2164 vm_page_t 2165 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 2166 { 2167 2168 return (vm_page_alloc_after(object, pindex, req, 2169 vm_page_mpred(object, pindex))); 2170 } 2171 2172 /* 2173 * Allocate a page in the specified object with the given page index. To 2174 * optimize insertion of the page into the object, the caller must also specify 2175 * the resident page in the object with largest index smaller than the given 2176 * page index, or NULL if no such page exists. 2177 */ 2178 static vm_page_t 2179 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 2180 int req, vm_page_t mpred) 2181 { 2182 struct vm_domainset_iter di; 2183 vm_page_t m; 2184 int domain; 2185 2186 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2187 do { 2188 m = vm_page_alloc_domain_after(object, pindex, domain, req, 2189 mpred); 2190 if (m != NULL) 2191 break; 2192 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2193 2194 return (m); 2195 } 2196 2197 /* 2198 * Returns true if the number of free pages exceeds the minimum 2199 * for the request class and false otherwise. 2200 */ 2201 static int 2202 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2203 { 2204 u_int limit, old, new; 2205 2206 if (req_class == VM_ALLOC_INTERRUPT) 2207 limit = 0; 2208 else if (req_class == VM_ALLOC_SYSTEM) 2209 limit = vmd->vmd_interrupt_free_min; 2210 else 2211 limit = vmd->vmd_free_reserved; 2212 2213 /* 2214 * Attempt to reserve the pages. Fail if we're below the limit. 2215 */ 2216 limit += npages; 2217 old = atomic_load_int(&vmd->vmd_free_count); 2218 do { 2219 if (old < limit) 2220 return (0); 2221 new = old - npages; 2222 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2223 2224 /* Wake the page daemon if we've crossed the threshold. */ 2225 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2226 pagedaemon_wakeup(vmd->vmd_domain); 2227 2228 /* Only update bitsets on transitions. */ 2229 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2230 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2231 vm_domain_set(vmd); 2232 2233 return (1); 2234 } 2235 2236 int 2237 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2238 { 2239 int req_class; 2240 2241 /* 2242 * The page daemon is allowed to dig deeper into the free page list. 2243 */ 2244 req_class = req & VM_ALLOC_CLASS_MASK; 2245 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2246 req_class = VM_ALLOC_SYSTEM; 2247 return (_vm_domain_allocate(vmd, req_class, npages)); 2248 } 2249 2250 vm_page_t 2251 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 2252 int req, vm_page_t mpred) 2253 { 2254 struct vm_domain *vmd; 2255 vm_page_t m; 2256 int flags; 2257 2258 #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2259 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ 2260 VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ 2261 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | \ 2262 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK) 2263 KASSERT((req & ~VPA_FLAGS) == 0, 2264 ("invalid request %#x", req)); 2265 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2266 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2267 ("invalid request %#x", req)); 2268 KASSERT(mpred == NULL || mpred->pindex < pindex, 2269 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2270 (uintmax_t)pindex)); 2271 VM_OBJECT_ASSERT_WLOCKED(object); 2272 2273 flags = 0; 2274 m = NULL; 2275 if (!vm_pager_can_alloc_page(object, pindex)) 2276 return (NULL); 2277 again: 2278 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2279 m = vm_page_alloc_nofree_domain(domain, req); 2280 if (m != NULL) 2281 goto found; 2282 } 2283 #if VM_NRESERVLEVEL > 0 2284 /* 2285 * Can we allocate the page from a reservation? 2286 */ 2287 if (vm_object_reserv(object) && 2288 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 2289 NULL) { 2290 goto found; 2291 } 2292 #endif 2293 vmd = VM_DOMAIN(domain); 2294 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2295 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2296 M_NOWAIT | M_NOVM); 2297 if (m != NULL) { 2298 flags |= PG_PCPU_CACHE; 2299 goto found; 2300 } 2301 } 2302 if (vm_domain_allocate(vmd, req, 1)) { 2303 /* 2304 * If not, allocate it from the free page queues. 2305 */ 2306 vm_domain_free_lock(vmd); 2307 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2308 vm_domain_free_unlock(vmd); 2309 if (m == NULL) { 2310 vm_domain_freecnt_inc(vmd, 1); 2311 #if VM_NRESERVLEVEL > 0 2312 if (vm_reserv_reclaim_inactive(domain)) 2313 goto again; 2314 #endif 2315 } 2316 } 2317 if (m == NULL) { 2318 /* 2319 * Not allocatable, give up. 2320 */ 2321 if (vm_domain_alloc_fail(vmd, object, req)) 2322 goto again; 2323 return (NULL); 2324 } 2325 2326 /* 2327 * At this point we had better have found a good page. 2328 */ 2329 found: 2330 vm_page_dequeue(m); 2331 vm_page_alloc_check(m); 2332 2333 /* 2334 * Initialize the page. Only the PG_ZERO flag is inherited. 2335 */ 2336 flags |= m->flags & PG_ZERO; 2337 if ((req & VM_ALLOC_NODUMP) != 0) 2338 flags |= PG_NODUMP; 2339 if ((req & VM_ALLOC_NOFREE) != 0) 2340 flags |= PG_NOFREE; 2341 m->flags = flags; 2342 m->a.flags = 0; 2343 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2344 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2345 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2346 else if ((req & VM_ALLOC_SBUSY) != 0) 2347 m->busy_lock = VPB_SHARERS_WORD(1); 2348 else 2349 m->busy_lock = VPB_UNBUSIED; 2350 if (req & VM_ALLOC_WIRED) { 2351 vm_wire_add(1); 2352 m->ref_count = 1; 2353 } 2354 m->a.act_count = 0; 2355 2356 if (vm_page_insert_after(m, object, pindex, mpred)) { 2357 if (req & VM_ALLOC_WIRED) { 2358 vm_wire_sub(1); 2359 m->ref_count = 0; 2360 } 2361 KASSERT(m->object == NULL, ("page %p has object", m)); 2362 m->oflags = VPO_UNMANAGED; 2363 m->busy_lock = VPB_UNBUSIED; 2364 /* Don't change PG_ZERO. */ 2365 vm_page_free_toq(m); 2366 if (req & VM_ALLOC_WAITFAIL) { 2367 VM_OBJECT_WUNLOCK(object); 2368 vm_radix_wait(); 2369 VM_OBJECT_WLOCK(object); 2370 } 2371 return (NULL); 2372 } 2373 2374 /* Ignore device objects; the pager sets "memattr" for them. */ 2375 if (object->memattr != VM_MEMATTR_DEFAULT && 2376 (object->flags & OBJ_FICTITIOUS) == 0) 2377 pmap_page_set_memattr(m, object->memattr); 2378 2379 return (m); 2380 } 2381 2382 /* 2383 * vm_page_alloc_contig: 2384 * 2385 * Allocate a contiguous set of physical pages of the given size "npages" 2386 * from the free lists. All of the physical pages must be at or above 2387 * the given physical address "low" and below the given physical address 2388 * "high". The given value "alignment" determines the alignment of the 2389 * first physical page in the set. If the given value "boundary" is 2390 * non-zero, then the set of physical pages cannot cross any physical 2391 * address boundary that is a multiple of that value. Both "alignment" 2392 * and "boundary" must be a power of two. 2393 * 2394 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2395 * then the memory attribute setting for the physical pages is configured 2396 * to the object's memory attribute setting. Otherwise, the memory 2397 * attribute setting for the physical pages is configured to "memattr", 2398 * overriding the object's memory attribute setting. However, if the 2399 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2400 * memory attribute setting for the physical pages cannot be configured 2401 * to VM_MEMATTR_DEFAULT. 2402 * 2403 * The specified object may not contain fictitious pages. 2404 * 2405 * The caller must always specify an allocation class. 2406 * 2407 * allocation classes: 2408 * VM_ALLOC_NORMAL normal process request 2409 * VM_ALLOC_SYSTEM system *really* needs a page 2410 * VM_ALLOC_INTERRUPT interrupt time request 2411 * 2412 * optional allocation flags: 2413 * VM_ALLOC_NOBUSY do not exclusive busy the page 2414 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2415 * VM_ALLOC_SBUSY shared busy the allocated page 2416 * VM_ALLOC_WIRED wire the allocated page 2417 * VM_ALLOC_ZERO prefer a zeroed page 2418 */ 2419 vm_page_t 2420 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2421 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2422 vm_paddr_t boundary, vm_memattr_t memattr) 2423 { 2424 struct vm_domainset_iter di; 2425 vm_page_t bounds[2]; 2426 vm_page_t m; 2427 int domain; 2428 int start_segind; 2429 2430 start_segind = -1; 2431 2432 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2433 do { 2434 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2435 npages, low, high, alignment, boundary, memattr); 2436 if (m != NULL) 2437 break; 2438 if (start_segind == -1) 2439 start_segind = vm_phys_lookup_segind(low); 2440 if (vm_phys_find_range(bounds, start_segind, domain, 2441 npages, low, high) == -1) { 2442 vm_domainset_iter_ignore(&di, domain); 2443 } 2444 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2445 2446 return (m); 2447 } 2448 2449 static vm_page_t 2450 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2451 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2452 { 2453 struct vm_domain *vmd; 2454 vm_page_t m_ret; 2455 2456 /* 2457 * Can we allocate the pages without the number of free pages falling 2458 * below the lower bound for the allocation class? 2459 */ 2460 vmd = VM_DOMAIN(domain); 2461 if (!vm_domain_allocate(vmd, req, npages)) 2462 return (NULL); 2463 /* 2464 * Try to allocate the pages from the free page queues. 2465 */ 2466 vm_domain_free_lock(vmd); 2467 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2468 alignment, boundary); 2469 vm_domain_free_unlock(vmd); 2470 if (m_ret != NULL) 2471 return (m_ret); 2472 #if VM_NRESERVLEVEL > 0 2473 /* 2474 * Try to break a reservation to allocate the pages. 2475 */ 2476 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2477 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2478 high, alignment, boundary); 2479 if (m_ret != NULL) 2480 return (m_ret); 2481 } 2482 #endif 2483 vm_domain_freecnt_inc(vmd, npages); 2484 return (NULL); 2485 } 2486 2487 vm_page_t 2488 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2489 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2490 vm_paddr_t boundary, vm_memattr_t memattr) 2491 { 2492 struct pctrie_iter pages; 2493 vm_page_t m, m_ret, mpred; 2494 u_int busy_lock, flags, oflags; 2495 2496 #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) 2497 KASSERT((req & ~VPAC_FLAGS) == 0, 2498 ("invalid request %#x", req)); 2499 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2500 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2501 ("invalid request %#x", req)); 2502 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2503 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2504 ("invalid request %#x", req)); 2505 VM_OBJECT_ASSERT_WLOCKED(object); 2506 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2507 ("vm_page_alloc_contig: object %p has fictitious pages", 2508 object)); 2509 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2510 2511 vm_page_iter_init(&pages, object); 2512 mpred = vm_radix_iter_lookup_le(&pages, pindex); 2513 KASSERT(mpred == NULL || mpred->pindex != pindex, 2514 ("vm_page_alloc_contig: pindex already allocated")); 2515 for (;;) { 2516 #if VM_NRESERVLEVEL > 0 2517 /* 2518 * Can we allocate the pages from a reservation? 2519 */ 2520 if (vm_object_reserv(object) && 2521 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2522 mpred, npages, low, high, alignment, boundary)) != NULL) { 2523 break; 2524 } 2525 #endif 2526 if ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2527 low, high, alignment, boundary)) != NULL) 2528 break; 2529 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req)) 2530 return (NULL); 2531 } 2532 2533 /* 2534 * Initialize the pages. Only the PG_ZERO flag is inherited. 2535 */ 2536 flags = PG_ZERO; 2537 if ((req & VM_ALLOC_NODUMP) != 0) 2538 flags |= PG_NODUMP; 2539 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2540 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2541 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2542 else if ((req & VM_ALLOC_SBUSY) != 0) 2543 busy_lock = VPB_SHARERS_WORD(1); 2544 else 2545 busy_lock = VPB_UNBUSIED; 2546 if ((req & VM_ALLOC_WIRED) != 0) 2547 vm_wire_add(npages); 2548 if (object->memattr != VM_MEMATTR_DEFAULT && 2549 memattr == VM_MEMATTR_DEFAULT) 2550 memattr = object->memattr; 2551 for (m = m_ret; m < &m_ret[npages]; m++) { 2552 vm_page_dequeue(m); 2553 vm_page_alloc_check(m); 2554 m->a.flags = 0; 2555 m->flags = (m->flags | PG_NODUMP) & flags; 2556 m->busy_lock = busy_lock; 2557 if ((req & VM_ALLOC_WIRED) != 0) 2558 m->ref_count = 1; 2559 m->a.act_count = 0; 2560 m->oflags = oflags; 2561 if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) { 2562 if ((req & VM_ALLOC_WIRED) != 0) 2563 vm_wire_sub(npages); 2564 KASSERT(m->object == NULL, 2565 ("page %p has object", m)); 2566 mpred = m; 2567 for (m = m_ret; m < &m_ret[npages]; m++) { 2568 if (m <= mpred && 2569 (req & VM_ALLOC_WIRED) != 0) 2570 m->ref_count = 0; 2571 m->oflags = VPO_UNMANAGED; 2572 m->busy_lock = VPB_UNBUSIED; 2573 /* Don't change PG_ZERO. */ 2574 vm_page_free_toq(m); 2575 } 2576 if (req & VM_ALLOC_WAITFAIL) { 2577 VM_OBJECT_WUNLOCK(object); 2578 vm_radix_wait(); 2579 VM_OBJECT_WLOCK(object); 2580 } 2581 return (NULL); 2582 } 2583 mpred = m; 2584 if (memattr != VM_MEMATTR_DEFAULT) 2585 pmap_page_set_memattr(m, memattr); 2586 pindex++; 2587 } 2588 return (m_ret); 2589 } 2590 2591 /* 2592 * Allocate a physical page that is not intended to be inserted into a VM 2593 * object. 2594 */ 2595 vm_page_t 2596 vm_page_alloc_noobj_domain(int domain, int req) 2597 { 2598 struct vm_domain *vmd; 2599 vm_page_t m; 2600 int flags; 2601 2602 #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2603 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ 2604 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ 2605 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | \ 2606 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK) 2607 KASSERT((req & ~VPAN_FLAGS) == 0, 2608 ("invalid request %#x", req)); 2609 2610 flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) | 2611 ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0); 2612 vmd = VM_DOMAIN(domain); 2613 again: 2614 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2615 m = vm_page_alloc_nofree_domain(domain, req); 2616 if (m != NULL) 2617 goto found; 2618 } 2619 2620 if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2621 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2622 M_NOWAIT | M_NOVM); 2623 if (m != NULL) { 2624 flags |= PG_PCPU_CACHE; 2625 goto found; 2626 } 2627 } 2628 2629 if (vm_domain_allocate(vmd, req, 1)) { 2630 vm_domain_free_lock(vmd); 2631 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2632 vm_domain_free_unlock(vmd); 2633 if (m == NULL) { 2634 vm_domain_freecnt_inc(vmd, 1); 2635 #if VM_NRESERVLEVEL > 0 2636 if (vm_reserv_reclaim_inactive(domain)) 2637 goto again; 2638 #endif 2639 } 2640 } 2641 if (m == NULL) { 2642 if (vm_domain_alloc_fail(vmd, NULL, req)) 2643 goto again; 2644 return (NULL); 2645 } 2646 2647 found: 2648 vm_page_dequeue(m); 2649 vm_page_alloc_check(m); 2650 2651 /* 2652 * Consumers should not rely on a useful default pindex value. 2653 */ 2654 m->pindex = 0xdeadc0dedeadc0de; 2655 m->flags = (m->flags & PG_ZERO) | flags; 2656 m->a.flags = 0; 2657 m->oflags = VPO_UNMANAGED; 2658 m->busy_lock = VPB_UNBUSIED; 2659 if ((req & VM_ALLOC_WIRED) != 0) { 2660 vm_wire_add(1); 2661 m->ref_count = 1; 2662 } 2663 2664 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2665 pmap_zero_page(m); 2666 2667 return (m); 2668 } 2669 2670 #if VM_NRESERVLEVEL > 1 2671 #define VM_NOFREE_IMPORT_ORDER (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER) 2672 #elif VM_NRESERVLEVEL > 0 2673 #define VM_NOFREE_IMPORT_ORDER VM_LEVEL_0_ORDER 2674 #else 2675 #define VM_NOFREE_IMPORT_ORDER 8 2676 #endif 2677 2678 /* 2679 * Allocate a single NOFREE page. 2680 * 2681 * This routine hands out NOFREE pages from higher-order 2682 * physical memory blocks in order to reduce memory fragmentation. 2683 * When a NOFREE for a given domain chunk is used up, 2684 * the routine will try to fetch a new one from the freelists 2685 * and discard the old one. 2686 */ 2687 static vm_page_t 2688 vm_page_alloc_nofree_domain(int domain, int req) 2689 { 2690 vm_page_t m; 2691 struct vm_domain *vmd; 2692 struct vm_nofreeq *nqp; 2693 2694 KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req)); 2695 2696 vmd = VM_DOMAIN(domain); 2697 nqp = &vmd->vmd_nofreeq; 2698 vm_domain_free_lock(vmd); 2699 if (nqp->offs >= (1 << VM_NOFREE_IMPORT_ORDER) || nqp->ma == NULL) { 2700 if (!vm_domain_allocate(vmd, req, 2701 1 << VM_NOFREE_IMPORT_ORDER)) { 2702 vm_domain_free_unlock(vmd); 2703 return (NULL); 2704 } 2705 nqp->ma = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 2706 VM_NOFREE_IMPORT_ORDER); 2707 if (nqp->ma == NULL) { 2708 vm_domain_freecnt_inc(vmd, 1 << VM_NOFREE_IMPORT_ORDER); 2709 vm_domain_free_unlock(vmd); 2710 return (NULL); 2711 } 2712 nqp->offs = 0; 2713 } 2714 m = &nqp->ma[nqp->offs++]; 2715 vm_domain_free_unlock(vmd); 2716 VM_CNT_ADD(v_nofree_count, 1); 2717 2718 return (m); 2719 } 2720 2721 vm_page_t 2722 vm_page_alloc_noobj(int req) 2723 { 2724 struct vm_domainset_iter di; 2725 vm_page_t m; 2726 int domain; 2727 2728 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2729 do { 2730 m = vm_page_alloc_noobj_domain(domain, req); 2731 if (m != NULL) 2732 break; 2733 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2734 2735 return (m); 2736 } 2737 2738 vm_page_t 2739 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2740 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2741 vm_memattr_t memattr) 2742 { 2743 struct vm_domainset_iter di; 2744 vm_page_t m; 2745 int domain; 2746 2747 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2748 do { 2749 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2750 high, alignment, boundary, memattr); 2751 if (m != NULL) 2752 break; 2753 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2754 2755 return (m); 2756 } 2757 2758 vm_page_t 2759 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2760 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2761 vm_memattr_t memattr) 2762 { 2763 vm_page_t m, m_ret; 2764 u_int flags; 2765 2766 #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) 2767 KASSERT((req & ~VPANC_FLAGS) == 0, 2768 ("invalid request %#x", req)); 2769 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2770 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2771 ("invalid request %#x", req)); 2772 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2773 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2774 ("invalid request %#x", req)); 2775 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2776 2777 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2778 low, high, alignment, boundary)) == NULL) { 2779 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2780 return (NULL); 2781 } 2782 2783 /* 2784 * Initialize the pages. Only the PG_ZERO flag is inherited. 2785 */ 2786 flags = PG_ZERO; 2787 if ((req & VM_ALLOC_NODUMP) != 0) 2788 flags |= PG_NODUMP; 2789 if ((req & VM_ALLOC_WIRED) != 0) 2790 vm_wire_add(npages); 2791 for (m = m_ret; m < &m_ret[npages]; m++) { 2792 vm_page_dequeue(m); 2793 vm_page_alloc_check(m); 2794 2795 /* 2796 * Consumers should not rely on a useful default pindex value. 2797 */ 2798 m->pindex = 0xdeadc0dedeadc0de; 2799 m->a.flags = 0; 2800 m->flags = (m->flags | PG_NODUMP) & flags; 2801 m->busy_lock = VPB_UNBUSIED; 2802 if ((req & VM_ALLOC_WIRED) != 0) 2803 m->ref_count = 1; 2804 m->a.act_count = 0; 2805 m->oflags = VPO_UNMANAGED; 2806 2807 /* 2808 * Zero the page before updating any mappings since the page is 2809 * not yet shared with any devices which might require the 2810 * non-default memory attribute. pmap_page_set_memattr() 2811 * flushes data caches before returning. 2812 */ 2813 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2814 pmap_zero_page(m); 2815 if (memattr != VM_MEMATTR_DEFAULT) 2816 pmap_page_set_memattr(m, memattr); 2817 } 2818 return (m_ret); 2819 } 2820 2821 /* 2822 * Check a page that has been freshly dequeued from a freelist. 2823 */ 2824 static void 2825 vm_page_alloc_check(vm_page_t m) 2826 { 2827 2828 KASSERT(m->object == NULL, ("page %p has object", m)); 2829 KASSERT(m->a.queue == PQ_NONE && 2830 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2831 ("page %p has unexpected queue %d, flags %#x", 2832 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2833 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2834 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2835 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2836 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2837 ("page %p has unexpected memattr %d", 2838 m, pmap_page_get_memattr(m))); 2839 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2840 pmap_vm_page_alloc_check(m); 2841 } 2842 2843 static int 2844 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2845 { 2846 struct vm_domain *vmd; 2847 struct vm_pgcache *pgcache; 2848 int i; 2849 2850 pgcache = arg; 2851 vmd = VM_DOMAIN(pgcache->domain); 2852 2853 /* 2854 * The page daemon should avoid creating extra memory pressure since its 2855 * main purpose is to replenish the store of free pages. 2856 */ 2857 if (vmd->vmd_severeset || curproc == pageproc || 2858 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2859 return (0); 2860 domain = vmd->vmd_domain; 2861 vm_domain_free_lock(vmd); 2862 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2863 (vm_page_t *)store); 2864 vm_domain_free_unlock(vmd); 2865 if (cnt != i) 2866 vm_domain_freecnt_inc(vmd, cnt - i); 2867 2868 return (i); 2869 } 2870 2871 static void 2872 vm_page_zone_release(void *arg, void **store, int cnt) 2873 { 2874 struct vm_domain *vmd; 2875 struct vm_pgcache *pgcache; 2876 vm_page_t m; 2877 int i; 2878 2879 pgcache = arg; 2880 vmd = VM_DOMAIN(pgcache->domain); 2881 vm_domain_free_lock(vmd); 2882 for (i = 0; i < cnt; i++) { 2883 m = (vm_page_t)store[i]; 2884 vm_phys_free_pages(m, 0); 2885 } 2886 vm_domain_free_unlock(vmd); 2887 vm_domain_freecnt_inc(vmd, cnt); 2888 } 2889 2890 #define VPSC_ANY 0 /* No restrictions. */ 2891 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2892 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2893 2894 /* 2895 * vm_page_scan_contig: 2896 * 2897 * Scan vm_page_array[] between the specified entries "m_start" and 2898 * "m_end" for a run of contiguous physical pages that satisfy the 2899 * specified conditions, and return the lowest page in the run. The 2900 * specified "alignment" determines the alignment of the lowest physical 2901 * page in the run. If the specified "boundary" is non-zero, then the 2902 * run of physical pages cannot span a physical address that is a 2903 * multiple of "boundary". 2904 * 2905 * "m_end" is never dereferenced, so it need not point to a vm_page 2906 * structure within vm_page_array[]. 2907 * 2908 * "npages" must be greater than zero. "m_start" and "m_end" must not 2909 * span a hole (or discontiguity) in the physical address space. Both 2910 * "alignment" and "boundary" must be a power of two. 2911 */ 2912 static vm_page_t 2913 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2914 u_long alignment, vm_paddr_t boundary, int options) 2915 { 2916 vm_object_t object; 2917 vm_paddr_t pa; 2918 vm_page_t m, m_run; 2919 #if VM_NRESERVLEVEL > 0 2920 int level; 2921 #endif 2922 int m_inc, order, run_ext, run_len; 2923 2924 KASSERT(npages > 0, ("npages is 0")); 2925 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2926 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2927 m_run = NULL; 2928 run_len = 0; 2929 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2930 KASSERT((m->flags & PG_MARKER) == 0, 2931 ("page %p is PG_MARKER", m)); 2932 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2933 ("fictitious page %p has invalid ref count", m)); 2934 2935 /* 2936 * If the current page would be the start of a run, check its 2937 * physical address against the end, alignment, and boundary 2938 * conditions. If it doesn't satisfy these conditions, either 2939 * terminate the scan or advance to the next page that 2940 * satisfies the failed condition. 2941 */ 2942 if (run_len == 0) { 2943 KASSERT(m_run == NULL, ("m_run != NULL")); 2944 if (m + npages > m_end) 2945 break; 2946 pa = VM_PAGE_TO_PHYS(m); 2947 if (!vm_addr_align_ok(pa, alignment)) { 2948 m_inc = atop(roundup2(pa, alignment) - pa); 2949 continue; 2950 } 2951 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2952 m_inc = atop(roundup2(pa, boundary) - pa); 2953 continue; 2954 } 2955 } else 2956 KASSERT(m_run != NULL, ("m_run == NULL")); 2957 2958 retry: 2959 m_inc = 1; 2960 if (vm_page_wired(m)) 2961 run_ext = 0; 2962 #if VM_NRESERVLEVEL > 0 2963 else if ((level = vm_reserv_level(m)) >= 0 && 2964 (options & VPSC_NORESERV) != 0) { 2965 run_ext = 0; 2966 /* Advance to the end of the reservation. */ 2967 pa = VM_PAGE_TO_PHYS(m); 2968 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2969 pa); 2970 } 2971 #endif 2972 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2973 /* 2974 * The page is considered eligible for relocation if 2975 * and only if it could be laundered or reclaimed by 2976 * the page daemon. 2977 */ 2978 VM_OBJECT_RLOCK(object); 2979 if (object != m->object) { 2980 VM_OBJECT_RUNLOCK(object); 2981 goto retry; 2982 } 2983 /* Don't care: PG_NODUMP, PG_ZERO. */ 2984 if ((object->flags & OBJ_SWAP) == 0 && 2985 object->type != OBJT_VNODE) { 2986 run_ext = 0; 2987 #if VM_NRESERVLEVEL > 0 2988 } else if ((options & VPSC_NOSUPER) != 0 && 2989 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2990 run_ext = 0; 2991 /* Advance to the end of the superpage. */ 2992 pa = VM_PAGE_TO_PHYS(m); 2993 m_inc = atop(roundup2(pa + 1, 2994 vm_reserv_size(level)) - pa); 2995 #endif 2996 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2997 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2998 /* 2999 * The page is allocated but eligible for 3000 * relocation. Extend the current run by one 3001 * page. 3002 */ 3003 KASSERT(pmap_page_get_memattr(m) == 3004 VM_MEMATTR_DEFAULT, 3005 ("page %p has an unexpected memattr", m)); 3006 KASSERT((m->oflags & (VPO_SWAPINPROG | 3007 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 3008 ("page %p has unexpected oflags", m)); 3009 /* Don't care: PGA_NOSYNC. */ 3010 run_ext = 1; 3011 } else 3012 run_ext = 0; 3013 VM_OBJECT_RUNLOCK(object); 3014 #if VM_NRESERVLEVEL > 0 3015 } else if (level >= 0) { 3016 /* 3017 * The page is reserved but not yet allocated. In 3018 * other words, it is still free. Extend the current 3019 * run by one page. 3020 */ 3021 run_ext = 1; 3022 #endif 3023 } else if ((order = m->order) < VM_NFREEORDER) { 3024 /* 3025 * The page is enqueued in the physical memory 3026 * allocator's free page queues. Moreover, it is the 3027 * first page in a power-of-two-sized run of 3028 * contiguous free pages. Add these pages to the end 3029 * of the current run, and jump ahead. 3030 */ 3031 run_ext = 1 << order; 3032 m_inc = 1 << order; 3033 } else { 3034 /* 3035 * Skip the page for one of the following reasons: (1) 3036 * It is enqueued in the physical memory allocator's 3037 * free page queues. However, it is not the first 3038 * page in a run of contiguous free pages. (This case 3039 * rarely occurs because the scan is performed in 3040 * ascending order.) (2) It is not reserved, and it is 3041 * transitioning from free to allocated. (Conversely, 3042 * the transition from allocated to free for managed 3043 * pages is blocked by the page busy lock.) (3) It is 3044 * allocated but not contained by an object and not 3045 * wired, e.g., allocated by Xen's balloon driver. 3046 */ 3047 run_ext = 0; 3048 } 3049 3050 /* 3051 * Extend or reset the current run of pages. 3052 */ 3053 if (run_ext > 0) { 3054 if (run_len == 0) 3055 m_run = m; 3056 run_len += run_ext; 3057 } else { 3058 if (run_len > 0) { 3059 m_run = NULL; 3060 run_len = 0; 3061 } 3062 } 3063 } 3064 if (run_len >= npages) 3065 return (m_run); 3066 return (NULL); 3067 } 3068 3069 /* 3070 * vm_page_reclaim_run: 3071 * 3072 * Try to relocate each of the allocated virtual pages within the 3073 * specified run of physical pages to a new physical address. Free the 3074 * physical pages underlying the relocated virtual pages. A virtual page 3075 * is relocatable if and only if it could be laundered or reclaimed by 3076 * the page daemon. Whenever possible, a virtual page is relocated to a 3077 * physical address above "high". 3078 * 3079 * Returns 0 if every physical page within the run was already free or 3080 * just freed by a successful relocation. Otherwise, returns a non-zero 3081 * value indicating why the last attempt to relocate a virtual page was 3082 * unsuccessful. 3083 * 3084 * "req_class" must be an allocation class. 3085 */ 3086 static int 3087 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 3088 vm_paddr_t high) 3089 { 3090 struct vm_domain *vmd; 3091 struct spglist free; 3092 vm_object_t object; 3093 vm_paddr_t pa; 3094 vm_page_t m, m_end, m_new; 3095 int error, order, req; 3096 3097 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 3098 ("req_class is not an allocation class")); 3099 SLIST_INIT(&free); 3100 error = 0; 3101 m = m_run; 3102 m_end = m_run + npages; 3103 for (; error == 0 && m < m_end; m++) { 3104 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 3105 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 3106 3107 /* 3108 * Racily check for wirings. Races are handled once the object 3109 * lock is held and the page is unmapped. 3110 */ 3111 if (vm_page_wired(m)) 3112 error = EBUSY; 3113 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 3114 /* 3115 * The page is relocated if and only if it could be 3116 * laundered or reclaimed by the page daemon. 3117 */ 3118 VM_OBJECT_WLOCK(object); 3119 /* Don't care: PG_NODUMP, PG_ZERO. */ 3120 if (m->object != object || 3121 ((object->flags & OBJ_SWAP) == 0 && 3122 object->type != OBJT_VNODE)) 3123 error = EINVAL; 3124 else if (object->memattr != VM_MEMATTR_DEFAULT) 3125 error = EINVAL; 3126 else if (vm_page_queue(m) != PQ_NONE && 3127 vm_page_tryxbusy(m) != 0) { 3128 if (vm_page_wired(m)) { 3129 vm_page_xunbusy(m); 3130 error = EBUSY; 3131 goto unlock; 3132 } 3133 KASSERT(pmap_page_get_memattr(m) == 3134 VM_MEMATTR_DEFAULT, 3135 ("page %p has an unexpected memattr", m)); 3136 KASSERT(m->oflags == 0, 3137 ("page %p has unexpected oflags", m)); 3138 /* Don't care: PGA_NOSYNC. */ 3139 if (!vm_page_none_valid(m)) { 3140 /* 3141 * First, try to allocate a new page 3142 * that is above "high". Failing 3143 * that, try to allocate a new page 3144 * that is below "m_run". Allocate 3145 * the new page between the end of 3146 * "m_run" and "high" only as a last 3147 * resort. 3148 */ 3149 req = req_class; 3150 if ((m->flags & PG_NODUMP) != 0) 3151 req |= VM_ALLOC_NODUMP; 3152 if (trunc_page(high) != 3153 ~(vm_paddr_t)PAGE_MASK) { 3154 m_new = 3155 vm_page_alloc_noobj_contig( 3156 req, 1, round_page(high), 3157 ~(vm_paddr_t)0, PAGE_SIZE, 3158 0, VM_MEMATTR_DEFAULT); 3159 } else 3160 m_new = NULL; 3161 if (m_new == NULL) { 3162 pa = VM_PAGE_TO_PHYS(m_run); 3163 m_new = 3164 vm_page_alloc_noobj_contig( 3165 req, 1, 0, pa - 1, 3166 PAGE_SIZE, 0, 3167 VM_MEMATTR_DEFAULT); 3168 } 3169 if (m_new == NULL) { 3170 pa += ptoa(npages); 3171 m_new = 3172 vm_page_alloc_noobj_contig( 3173 req, 1, pa, high, PAGE_SIZE, 3174 0, VM_MEMATTR_DEFAULT); 3175 } 3176 if (m_new == NULL) { 3177 vm_page_xunbusy(m); 3178 error = ENOMEM; 3179 goto unlock; 3180 } 3181 3182 /* 3183 * Unmap the page and check for new 3184 * wirings that may have been acquired 3185 * through a pmap lookup. 3186 */ 3187 if (object->ref_count != 0 && 3188 !vm_page_try_remove_all(m)) { 3189 vm_page_xunbusy(m); 3190 vm_page_free(m_new); 3191 error = EBUSY; 3192 goto unlock; 3193 } 3194 3195 /* 3196 * Replace "m" with the new page. For 3197 * vm_page_replace(), "m" must be busy 3198 * and dequeued. Finally, change "m" 3199 * as if vm_page_free() was called. 3200 */ 3201 m_new->a.flags = m->a.flags & 3202 ~PGA_QUEUE_STATE_MASK; 3203 KASSERT(m_new->oflags == VPO_UNMANAGED, 3204 ("page %p is managed", m_new)); 3205 m_new->oflags = 0; 3206 pmap_copy_page(m, m_new); 3207 m_new->valid = m->valid; 3208 m_new->dirty = m->dirty; 3209 m->flags &= ~PG_ZERO; 3210 vm_page_dequeue(m); 3211 if (vm_page_replace_hold(m_new, object, 3212 m->pindex, m) && 3213 vm_page_free_prep(m)) 3214 SLIST_INSERT_HEAD(&free, m, 3215 plinks.s.ss); 3216 3217 /* 3218 * The new page must be deactivated 3219 * before the object is unlocked. 3220 */ 3221 vm_page_deactivate(m_new); 3222 } else { 3223 m->flags &= ~PG_ZERO; 3224 vm_page_dequeue(m); 3225 if (vm_page_free_prep(m)) 3226 SLIST_INSERT_HEAD(&free, m, 3227 plinks.s.ss); 3228 KASSERT(m->dirty == 0, 3229 ("page %p is dirty", m)); 3230 } 3231 } else 3232 error = EBUSY; 3233 unlock: 3234 VM_OBJECT_WUNLOCK(object); 3235 } else { 3236 MPASS(vm_page_domain(m) == domain); 3237 vmd = VM_DOMAIN(domain); 3238 vm_domain_free_lock(vmd); 3239 order = m->order; 3240 if (order < VM_NFREEORDER) { 3241 /* 3242 * The page is enqueued in the physical memory 3243 * allocator's free page queues. Moreover, it 3244 * is the first page in a power-of-two-sized 3245 * run of contiguous free pages. Jump ahead 3246 * to the last page within that run, and 3247 * continue from there. 3248 */ 3249 m += (1 << order) - 1; 3250 } 3251 #if VM_NRESERVLEVEL > 0 3252 else if (vm_reserv_is_page_free(m)) 3253 order = 0; 3254 #endif 3255 vm_domain_free_unlock(vmd); 3256 if (order == VM_NFREEORDER) 3257 error = EINVAL; 3258 } 3259 } 3260 if ((m = SLIST_FIRST(&free)) != NULL) { 3261 int cnt; 3262 3263 vmd = VM_DOMAIN(domain); 3264 cnt = 0; 3265 vm_domain_free_lock(vmd); 3266 do { 3267 MPASS(vm_page_domain(m) == domain); 3268 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3269 vm_phys_free_pages(m, 0); 3270 cnt++; 3271 } while ((m = SLIST_FIRST(&free)) != NULL); 3272 vm_domain_free_unlock(vmd); 3273 vm_domain_freecnt_inc(vmd, cnt); 3274 } 3275 return (error); 3276 } 3277 3278 #define NRUNS 16 3279 3280 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 3281 3282 #define MIN_RECLAIM 8 3283 3284 /* 3285 * vm_page_reclaim_contig: 3286 * 3287 * Reclaim allocated, contiguous physical memory satisfying the specified 3288 * conditions by relocating the virtual pages using that physical memory. 3289 * Returns 0 if reclamation is successful, ERANGE if the specified domain 3290 * can't possibly satisfy the reclamation request, or ENOMEM if not 3291 * currently able to reclaim the requested number of pages. Since 3292 * relocation requires the allocation of physical pages, reclamation may 3293 * fail with ENOMEM due to a shortage of free pages. When reclamation 3294 * fails in this manner, callers are expected to perform vm_wait() before 3295 * retrying a failed allocation operation, e.g., vm_page_alloc_contig(). 3296 * 3297 * The caller must always specify an allocation class through "req". 3298 * 3299 * allocation classes: 3300 * VM_ALLOC_NORMAL normal process request 3301 * VM_ALLOC_SYSTEM system *really* needs a page 3302 * VM_ALLOC_INTERRUPT interrupt time request 3303 * 3304 * The optional allocation flags are ignored. 3305 * 3306 * "npages" must be greater than zero. Both "alignment" and "boundary" 3307 * must be a power of two. 3308 */ 3309 int 3310 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3311 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3312 int desired_runs) 3313 { 3314 struct vm_domain *vmd; 3315 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; 3316 u_long count, minalign, reclaimed; 3317 int error, i, min_reclaim, nruns, options, req_class; 3318 int segind, start_segind; 3319 int ret; 3320 3321 KASSERT(npages > 0, ("npages is 0")); 3322 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3323 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3324 3325 ret = ENOMEM; 3326 3327 /* 3328 * If the caller wants to reclaim multiple runs, try to allocate 3329 * space to store the runs. If that fails, fall back to the old 3330 * behavior of just reclaiming MIN_RECLAIM pages. 3331 */ 3332 if (desired_runs > 1) 3333 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3334 M_TEMP, M_NOWAIT); 3335 else 3336 m_runs = NULL; 3337 3338 if (m_runs == NULL) { 3339 m_runs = _m_runs; 3340 nruns = NRUNS; 3341 } else { 3342 nruns = NRUNS + desired_runs - 1; 3343 } 3344 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3345 3346 /* 3347 * The caller will attempt an allocation after some runs have been 3348 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3349 * of vm_phys_alloc_contig(), round up the requested length to the next 3350 * power of two or maximum chunk size, and ensure that each run is 3351 * suitably aligned. 3352 */ 3353 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3354 npages = roundup2(npages, minalign); 3355 if (alignment < ptoa(minalign)) 3356 alignment = ptoa(minalign); 3357 3358 /* 3359 * The page daemon is allowed to dig deeper into the free page list. 3360 */ 3361 req_class = req & VM_ALLOC_CLASS_MASK; 3362 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3363 req_class = VM_ALLOC_SYSTEM; 3364 3365 start_segind = vm_phys_lookup_segind(low); 3366 3367 /* 3368 * Return if the number of free pages cannot satisfy the requested 3369 * allocation. 3370 */ 3371 vmd = VM_DOMAIN(domain); 3372 count = vmd->vmd_free_count; 3373 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3374 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3375 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3376 goto done; 3377 3378 /* 3379 * Scan up to three times, relaxing the restrictions ("options") on 3380 * the reclamation of reservations and superpages each time. 3381 */ 3382 for (options = VPSC_NORESERV;;) { 3383 bool phys_range_exists = false; 3384 3385 /* 3386 * Find the highest runs that satisfy the given constraints 3387 * and restrictions, and record them in "m_runs". 3388 */ 3389 count = 0; 3390 segind = start_segind; 3391 while ((segind = vm_phys_find_range(bounds, segind, domain, 3392 npages, low, high)) != -1) { 3393 phys_range_exists = true; 3394 while ((m_run = vm_page_scan_contig(npages, bounds[0], 3395 bounds[1], alignment, boundary, options))) { 3396 bounds[0] = m_run + npages; 3397 m_runs[RUN_INDEX(count, nruns)] = m_run; 3398 count++; 3399 } 3400 segind++; 3401 } 3402 3403 if (!phys_range_exists) { 3404 ret = ERANGE; 3405 goto done; 3406 } 3407 3408 /* 3409 * Reclaim the highest runs in LIFO (descending) order until 3410 * the number of reclaimed pages, "reclaimed", is at least 3411 * "min_reclaim". Reset "reclaimed" each time because each 3412 * reclamation is idempotent, and runs will (likely) recur 3413 * from one scan to the next as restrictions are relaxed. 3414 */ 3415 reclaimed = 0; 3416 for (i = 0; count > 0 && i < nruns; i++) { 3417 count--; 3418 m_run = m_runs[RUN_INDEX(count, nruns)]; 3419 error = vm_page_reclaim_run(req_class, domain, npages, 3420 m_run, high); 3421 if (error == 0) { 3422 reclaimed += npages; 3423 if (reclaimed >= min_reclaim) { 3424 ret = 0; 3425 goto done; 3426 } 3427 } 3428 } 3429 3430 /* 3431 * Either relax the restrictions on the next scan or return if 3432 * the last scan had no restrictions. 3433 */ 3434 if (options == VPSC_NORESERV) 3435 options = VPSC_NOSUPER; 3436 else if (options == VPSC_NOSUPER) 3437 options = VPSC_ANY; 3438 else if (options == VPSC_ANY) { 3439 if (reclaimed != 0) 3440 ret = 0; 3441 goto done; 3442 } 3443 } 3444 done: 3445 if (m_runs != _m_runs) 3446 free(m_runs, M_TEMP); 3447 return (ret); 3448 } 3449 3450 int 3451 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3452 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3453 { 3454 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high, 3455 alignment, boundary, 1)); 3456 } 3457 3458 int 3459 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3460 u_long alignment, vm_paddr_t boundary) 3461 { 3462 struct vm_domainset_iter di; 3463 int domain, ret, status; 3464 3465 ret = ERANGE; 3466 3467 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3468 do { 3469 status = vm_page_reclaim_contig_domain(domain, req, npages, low, 3470 high, alignment, boundary); 3471 if (status == 0) 3472 return (0); 3473 else if (status == ERANGE) 3474 vm_domainset_iter_ignore(&di, domain); 3475 else { 3476 KASSERT(status == ENOMEM, ("Unrecognized error %d " 3477 "from vm_page_reclaim_contig_domain()", status)); 3478 ret = ENOMEM; 3479 } 3480 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3481 3482 return (ret); 3483 } 3484 3485 /* 3486 * Set the domain in the appropriate page level domainset. 3487 */ 3488 void 3489 vm_domain_set(struct vm_domain *vmd) 3490 { 3491 3492 mtx_lock(&vm_domainset_lock); 3493 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3494 vmd->vmd_minset = 1; 3495 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3496 } 3497 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3498 vmd->vmd_severeset = 1; 3499 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3500 } 3501 mtx_unlock(&vm_domainset_lock); 3502 } 3503 3504 /* 3505 * Clear the domain from the appropriate page level domainset. 3506 */ 3507 void 3508 vm_domain_clear(struct vm_domain *vmd) 3509 { 3510 3511 mtx_lock(&vm_domainset_lock); 3512 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3513 vmd->vmd_minset = 0; 3514 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3515 if (vm_min_waiters != 0) { 3516 vm_min_waiters = 0; 3517 wakeup(&vm_min_domains); 3518 } 3519 } 3520 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3521 vmd->vmd_severeset = 0; 3522 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3523 if (vm_severe_waiters != 0) { 3524 vm_severe_waiters = 0; 3525 wakeup(&vm_severe_domains); 3526 } 3527 } 3528 3529 /* 3530 * If pageout daemon needs pages, then tell it that there are 3531 * some free. 3532 */ 3533 if (vmd->vmd_pageout_pages_needed && 3534 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3535 wakeup(&vmd->vmd_pageout_pages_needed); 3536 vmd->vmd_pageout_pages_needed = 0; 3537 } 3538 3539 /* See comments in vm_wait_doms(). */ 3540 if (vm_pageproc_waiters) { 3541 vm_pageproc_waiters = 0; 3542 wakeup(&vm_pageproc_waiters); 3543 } 3544 mtx_unlock(&vm_domainset_lock); 3545 } 3546 3547 /* 3548 * Wait for free pages to exceed the min threshold globally. 3549 */ 3550 void 3551 vm_wait_min(void) 3552 { 3553 3554 mtx_lock(&vm_domainset_lock); 3555 while (vm_page_count_min()) { 3556 vm_min_waiters++; 3557 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3558 } 3559 mtx_unlock(&vm_domainset_lock); 3560 } 3561 3562 /* 3563 * Wait for free pages to exceed the severe threshold globally. 3564 */ 3565 void 3566 vm_wait_severe(void) 3567 { 3568 3569 mtx_lock(&vm_domainset_lock); 3570 while (vm_page_count_severe()) { 3571 vm_severe_waiters++; 3572 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3573 "vmwait", 0); 3574 } 3575 mtx_unlock(&vm_domainset_lock); 3576 } 3577 3578 u_int 3579 vm_wait_count(void) 3580 { 3581 3582 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3583 } 3584 3585 int 3586 vm_wait_doms(const domainset_t *wdoms, int mflags) 3587 { 3588 int error; 3589 3590 error = 0; 3591 3592 /* 3593 * We use racey wakeup synchronization to avoid expensive global 3594 * locking for the pageproc when sleeping with a non-specific vm_wait. 3595 * To handle this, we only sleep for one tick in this instance. It 3596 * is expected that most allocations for the pageproc will come from 3597 * kmem or vm_page_grab* which will use the more specific and 3598 * race-free vm_wait_domain(). 3599 */ 3600 if (curproc == pageproc) { 3601 mtx_lock(&vm_domainset_lock); 3602 vm_pageproc_waiters++; 3603 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3604 PVM | PDROP | mflags, "pageprocwait", 1); 3605 } else { 3606 /* 3607 * XXX Ideally we would wait only until the allocation could 3608 * be satisfied. This condition can cause new allocators to 3609 * consume all freed pages while old allocators wait. 3610 */ 3611 mtx_lock(&vm_domainset_lock); 3612 if (vm_page_count_min_set(wdoms)) { 3613 if (pageproc == NULL) 3614 panic("vm_wait in early boot"); 3615 vm_min_waiters++; 3616 error = msleep(&vm_min_domains, &vm_domainset_lock, 3617 PVM | PDROP | mflags, "vmwait", 0); 3618 } else 3619 mtx_unlock(&vm_domainset_lock); 3620 } 3621 return (error); 3622 } 3623 3624 /* 3625 * vm_wait_domain: 3626 * 3627 * Sleep until free pages are available for allocation. 3628 * - Called in various places after failed memory allocations. 3629 */ 3630 void 3631 vm_wait_domain(int domain) 3632 { 3633 struct vm_domain *vmd; 3634 domainset_t wdom; 3635 3636 vmd = VM_DOMAIN(domain); 3637 vm_domain_free_assert_unlocked(vmd); 3638 3639 if (curproc == pageproc) { 3640 mtx_lock(&vm_domainset_lock); 3641 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3642 vmd->vmd_pageout_pages_needed = 1; 3643 msleep(&vmd->vmd_pageout_pages_needed, 3644 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3645 } else 3646 mtx_unlock(&vm_domainset_lock); 3647 } else { 3648 DOMAINSET_ZERO(&wdom); 3649 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3650 vm_wait_doms(&wdom, 0); 3651 } 3652 } 3653 3654 static int 3655 vm_wait_flags(vm_object_t obj, int mflags) 3656 { 3657 struct domainset *d; 3658 3659 d = NULL; 3660 3661 /* 3662 * Carefully fetch pointers only once: the struct domainset 3663 * itself is ummutable but the pointer might change. 3664 */ 3665 if (obj != NULL) 3666 d = obj->domain.dr_policy; 3667 if (d == NULL) 3668 d = curthread->td_domain.dr_policy; 3669 3670 return (vm_wait_doms(&d->ds_mask, mflags)); 3671 } 3672 3673 /* 3674 * vm_wait: 3675 * 3676 * Sleep until free pages are available for allocation in the 3677 * affinity domains of the obj. If obj is NULL, the domain set 3678 * for the calling thread is used. 3679 * Called in various places after failed memory allocations. 3680 */ 3681 void 3682 vm_wait(vm_object_t obj) 3683 { 3684 (void)vm_wait_flags(obj, 0); 3685 } 3686 3687 int 3688 vm_wait_intr(vm_object_t obj) 3689 { 3690 return (vm_wait_flags(obj, PCATCH)); 3691 } 3692 3693 /* 3694 * vm_domain_alloc_fail: 3695 * 3696 * Called when a page allocation function fails. Informs the 3697 * pagedaemon and performs the requested wait. Requires the 3698 * domain_free and object lock on entry. Returns with the 3699 * object lock held and free lock released. Returns an error when 3700 * retry is necessary. 3701 * 3702 */ 3703 static int 3704 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3705 { 3706 3707 vm_domain_free_assert_unlocked(vmd); 3708 3709 atomic_add_int(&vmd->vmd_pageout_deficit, 3710 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3711 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3712 if (object != NULL) 3713 VM_OBJECT_WUNLOCK(object); 3714 vm_wait_domain(vmd->vmd_domain); 3715 if (object != NULL) 3716 VM_OBJECT_WLOCK(object); 3717 if (req & VM_ALLOC_WAITOK) 3718 return (EAGAIN); 3719 } 3720 3721 return (0); 3722 } 3723 3724 /* 3725 * vm_waitpfault: 3726 * 3727 * Sleep until free pages are available for allocation. 3728 * - Called only in vm_fault so that processes page faulting 3729 * can be easily tracked. 3730 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3731 * processes will be able to grab memory first. Do not change 3732 * this balance without careful testing first. 3733 */ 3734 void 3735 vm_waitpfault(struct domainset *dset, int timo) 3736 { 3737 3738 /* 3739 * XXX Ideally we would wait only until the allocation could 3740 * be satisfied. This condition can cause new allocators to 3741 * consume all freed pages while old allocators wait. 3742 */ 3743 mtx_lock(&vm_domainset_lock); 3744 if (vm_page_count_min_set(&dset->ds_mask)) { 3745 vm_min_waiters++; 3746 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3747 "pfault", timo); 3748 } else 3749 mtx_unlock(&vm_domainset_lock); 3750 } 3751 3752 static struct vm_pagequeue * 3753 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3754 { 3755 3756 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3757 } 3758 3759 #ifdef INVARIANTS 3760 static struct vm_pagequeue * 3761 vm_page_pagequeue(vm_page_t m) 3762 { 3763 3764 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3765 } 3766 #endif 3767 3768 static __always_inline bool 3769 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3770 { 3771 vm_page_astate_t tmp; 3772 3773 tmp = *old; 3774 do { 3775 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3776 return (true); 3777 counter_u64_add(pqstate_commit_retries, 1); 3778 } while (old->_bits == tmp._bits); 3779 3780 return (false); 3781 } 3782 3783 /* 3784 * Do the work of committing a queue state update that moves the page out of 3785 * its current queue. 3786 */ 3787 static bool 3788 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3789 vm_page_astate_t *old, vm_page_astate_t new) 3790 { 3791 vm_page_t next; 3792 3793 vm_pagequeue_assert_locked(pq); 3794 KASSERT(vm_page_pagequeue(m) == pq, 3795 ("%s: queue %p does not match page %p", __func__, pq, m)); 3796 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3797 ("%s: invalid queue indices %d %d", 3798 __func__, old->queue, new.queue)); 3799 3800 /* 3801 * Once the queue index of the page changes there is nothing 3802 * synchronizing with further updates to the page's physical 3803 * queue state. Therefore we must speculatively remove the page 3804 * from the queue now and be prepared to roll back if the queue 3805 * state update fails. If the page is not physically enqueued then 3806 * we just update its queue index. 3807 */ 3808 if ((old->flags & PGA_ENQUEUED) != 0) { 3809 new.flags &= ~PGA_ENQUEUED; 3810 next = TAILQ_NEXT(m, plinks.q); 3811 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3812 vm_pagequeue_cnt_dec(pq); 3813 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3814 if (next == NULL) 3815 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3816 else 3817 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3818 vm_pagequeue_cnt_inc(pq); 3819 return (false); 3820 } else { 3821 return (true); 3822 } 3823 } else { 3824 return (vm_page_pqstate_fcmpset(m, old, new)); 3825 } 3826 } 3827 3828 static bool 3829 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3830 vm_page_astate_t new) 3831 { 3832 struct vm_pagequeue *pq; 3833 vm_page_astate_t as; 3834 bool ret; 3835 3836 pq = _vm_page_pagequeue(m, old->queue); 3837 3838 /* 3839 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3840 * corresponding page queue lock is held. 3841 */ 3842 vm_pagequeue_lock(pq); 3843 as = vm_page_astate_load(m); 3844 if (__predict_false(as._bits != old->_bits)) { 3845 *old = as; 3846 ret = false; 3847 } else { 3848 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3849 } 3850 vm_pagequeue_unlock(pq); 3851 return (ret); 3852 } 3853 3854 /* 3855 * Commit a queue state update that enqueues or requeues a page. 3856 */ 3857 static bool 3858 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3859 vm_page_astate_t *old, vm_page_astate_t new) 3860 { 3861 struct vm_domain *vmd; 3862 3863 vm_pagequeue_assert_locked(pq); 3864 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3865 ("%s: invalid queue indices %d %d", 3866 __func__, old->queue, new.queue)); 3867 3868 new.flags |= PGA_ENQUEUED; 3869 if (!vm_page_pqstate_fcmpset(m, old, new)) 3870 return (false); 3871 3872 if ((old->flags & PGA_ENQUEUED) != 0) 3873 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3874 else 3875 vm_pagequeue_cnt_inc(pq); 3876 3877 /* 3878 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3879 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3880 * applied, even if it was set first. 3881 */ 3882 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3883 vmd = vm_pagequeue_domain(m); 3884 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3885 ("%s: invalid page queue for page %p", __func__, m)); 3886 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3887 } else { 3888 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3889 } 3890 return (true); 3891 } 3892 3893 /* 3894 * Commit a queue state update that encodes a request for a deferred queue 3895 * operation. 3896 */ 3897 static bool 3898 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3899 vm_page_astate_t new) 3900 { 3901 3902 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3903 ("%s: invalid state, queue %d flags %x", 3904 __func__, new.queue, new.flags)); 3905 3906 if (old->_bits != new._bits && 3907 !vm_page_pqstate_fcmpset(m, old, new)) 3908 return (false); 3909 vm_page_pqbatch_submit(m, new.queue); 3910 return (true); 3911 } 3912 3913 /* 3914 * A generic queue state update function. This handles more cases than the 3915 * specialized functions above. 3916 */ 3917 bool 3918 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3919 { 3920 3921 if (old->_bits == new._bits) 3922 return (true); 3923 3924 if (old->queue != PQ_NONE && new.queue != old->queue) { 3925 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3926 return (false); 3927 if (new.queue != PQ_NONE) 3928 vm_page_pqbatch_submit(m, new.queue); 3929 } else { 3930 if (!vm_page_pqstate_fcmpset(m, old, new)) 3931 return (false); 3932 if (new.queue != PQ_NONE && 3933 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3934 vm_page_pqbatch_submit(m, new.queue); 3935 } 3936 return (true); 3937 } 3938 3939 /* 3940 * Apply deferred queue state updates to a page. 3941 */ 3942 static inline void 3943 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3944 { 3945 vm_page_astate_t new, old; 3946 3947 CRITICAL_ASSERT(curthread); 3948 vm_pagequeue_assert_locked(pq); 3949 KASSERT(queue < PQ_COUNT, 3950 ("%s: invalid queue index %d", __func__, queue)); 3951 KASSERT(pq == _vm_page_pagequeue(m, queue), 3952 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3953 3954 for (old = vm_page_astate_load(m);;) { 3955 if (__predict_false(old.queue != queue || 3956 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3957 counter_u64_add(queue_nops, 1); 3958 break; 3959 } 3960 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3961 ("%s: page %p is unmanaged", __func__, m)); 3962 3963 new = old; 3964 if ((old.flags & PGA_DEQUEUE) != 0) { 3965 new.flags &= ~PGA_QUEUE_OP_MASK; 3966 new.queue = PQ_NONE; 3967 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3968 m, &old, new))) { 3969 counter_u64_add(queue_ops, 1); 3970 break; 3971 } 3972 } else { 3973 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3974 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3975 m, &old, new))) { 3976 counter_u64_add(queue_ops, 1); 3977 break; 3978 } 3979 } 3980 } 3981 } 3982 3983 static void 3984 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3985 uint8_t queue) 3986 { 3987 int i; 3988 3989 for (i = 0; i < bq->bq_cnt; i++) 3990 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3991 vm_batchqueue_init(bq); 3992 } 3993 3994 /* 3995 * vm_page_pqbatch_submit: [ internal use only ] 3996 * 3997 * Enqueue a page in the specified page queue's batched work queue. 3998 * The caller must have encoded the requested operation in the page 3999 * structure's a.flags field. 4000 */ 4001 void 4002 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 4003 { 4004 struct vm_batchqueue *bq; 4005 struct vm_pagequeue *pq; 4006 int domain, slots_remaining; 4007 4008 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 4009 4010 domain = vm_page_domain(m); 4011 critical_enter(); 4012 bq = DPCPU_PTR(pqbatch[domain][queue]); 4013 slots_remaining = vm_batchqueue_insert(bq, m); 4014 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 4015 /* keep building the bq */ 4016 critical_exit(); 4017 return; 4018 } else if (slots_remaining > 0 ) { 4019 /* Try to process the bq if we can get the lock */ 4020 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 4021 if (vm_pagequeue_trylock(pq)) { 4022 vm_pqbatch_process(pq, bq, queue); 4023 vm_pagequeue_unlock(pq); 4024 } 4025 critical_exit(); 4026 return; 4027 } 4028 critical_exit(); 4029 4030 /* if we make it here, the bq is full so wait for the lock */ 4031 4032 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 4033 vm_pagequeue_lock(pq); 4034 critical_enter(); 4035 bq = DPCPU_PTR(pqbatch[domain][queue]); 4036 vm_pqbatch_process(pq, bq, queue); 4037 vm_pqbatch_process_page(pq, m, queue); 4038 vm_pagequeue_unlock(pq); 4039 critical_exit(); 4040 } 4041 4042 /* 4043 * vm_page_pqbatch_drain: [ internal use only ] 4044 * 4045 * Force all per-CPU page queue batch queues to be drained. This is 4046 * intended for use in severe memory shortages, to ensure that pages 4047 * do not remain stuck in the batch queues. 4048 */ 4049 void 4050 vm_page_pqbatch_drain(void) 4051 { 4052 struct thread *td; 4053 struct vm_domain *vmd; 4054 struct vm_pagequeue *pq; 4055 int cpu, domain, queue; 4056 4057 td = curthread; 4058 CPU_FOREACH(cpu) { 4059 thread_lock(td); 4060 sched_bind(td, cpu); 4061 thread_unlock(td); 4062 4063 for (domain = 0; domain < vm_ndomains; domain++) { 4064 vmd = VM_DOMAIN(domain); 4065 for (queue = 0; queue < PQ_COUNT; queue++) { 4066 pq = &vmd->vmd_pagequeues[queue]; 4067 vm_pagequeue_lock(pq); 4068 critical_enter(); 4069 vm_pqbatch_process(pq, 4070 DPCPU_PTR(pqbatch[domain][queue]), queue); 4071 critical_exit(); 4072 vm_pagequeue_unlock(pq); 4073 } 4074 } 4075 } 4076 thread_lock(td); 4077 sched_unbind(td); 4078 thread_unlock(td); 4079 } 4080 4081 /* 4082 * vm_page_dequeue_deferred: [ internal use only ] 4083 * 4084 * Request removal of the given page from its current page 4085 * queue. Physical removal from the queue may be deferred 4086 * indefinitely. 4087 */ 4088 void 4089 vm_page_dequeue_deferred(vm_page_t m) 4090 { 4091 vm_page_astate_t new, old; 4092 4093 old = vm_page_astate_load(m); 4094 do { 4095 if (old.queue == PQ_NONE) { 4096 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4097 ("%s: page %p has unexpected queue state", 4098 __func__, m)); 4099 break; 4100 } 4101 new = old; 4102 new.flags |= PGA_DEQUEUE; 4103 } while (!vm_page_pqstate_commit_request(m, &old, new)); 4104 } 4105 4106 /* 4107 * vm_page_dequeue: 4108 * 4109 * Remove the page from whichever page queue it's in, if any, before 4110 * returning. 4111 */ 4112 void 4113 vm_page_dequeue(vm_page_t m) 4114 { 4115 vm_page_astate_t new, old; 4116 4117 old = vm_page_astate_load(m); 4118 do { 4119 if (old.queue == PQ_NONE) { 4120 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4121 ("%s: page %p has unexpected queue state", 4122 __func__, m)); 4123 break; 4124 } 4125 new = old; 4126 new.flags &= ~PGA_QUEUE_OP_MASK; 4127 new.queue = PQ_NONE; 4128 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 4129 4130 } 4131 4132 /* 4133 * Schedule the given page for insertion into the specified page queue. 4134 * Physical insertion of the page may be deferred indefinitely. 4135 */ 4136 static void 4137 vm_page_enqueue(vm_page_t m, uint8_t queue) 4138 { 4139 4140 KASSERT(m->a.queue == PQ_NONE && 4141 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 4142 ("%s: page %p is already enqueued", __func__, m)); 4143 KASSERT(m->ref_count > 0, 4144 ("%s: page %p does not carry any references", __func__, m)); 4145 4146 m->a.queue = queue; 4147 if ((m->a.flags & PGA_REQUEUE) == 0) 4148 vm_page_aflag_set(m, PGA_REQUEUE); 4149 vm_page_pqbatch_submit(m, queue); 4150 } 4151 4152 /* 4153 * vm_page_free_prep: 4154 * 4155 * Prepares the given page to be put on the free list, 4156 * disassociating it from any VM object. The caller may return 4157 * the page to the free list only if this function returns true. 4158 * 4159 * The object, if it exists, must be locked, and then the page must 4160 * be xbusy. Otherwise the page must be not busied. A managed 4161 * page must be unmapped. 4162 */ 4163 static bool 4164 vm_page_free_prep(vm_page_t m) 4165 { 4166 4167 /* 4168 * Synchronize with threads that have dropped a reference to this 4169 * page. 4170 */ 4171 atomic_thread_fence_acq(); 4172 4173 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 4174 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 4175 uint64_t *p; 4176 int i; 4177 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 4178 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 4179 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 4180 m, i, (uintmax_t)*p)); 4181 } 4182 #endif 4183 KASSERT((m->flags & PG_NOFREE) == 0, 4184 ("%s: attempting to free a PG_NOFREE page", __func__)); 4185 if ((m->oflags & VPO_UNMANAGED) == 0) { 4186 KASSERT(!pmap_page_is_mapped(m), 4187 ("vm_page_free_prep: freeing mapped page %p", m)); 4188 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 4189 ("vm_page_free_prep: mapping flags set in page %p", m)); 4190 } else { 4191 KASSERT(m->a.queue == PQ_NONE, 4192 ("vm_page_free_prep: unmanaged page %p is queued", m)); 4193 } 4194 VM_CNT_INC(v_tfree); 4195 4196 if (m->object != NULL) { 4197 vm_page_radix_remove(m); 4198 vm_page_free_object_prep(m); 4199 } else 4200 vm_page_assert_unbusied(m); 4201 4202 vm_page_busy_free(m); 4203 4204 /* 4205 * If fictitious remove object association and 4206 * return. 4207 */ 4208 if ((m->flags & PG_FICTITIOUS) != 0) { 4209 KASSERT(m->ref_count == 1, 4210 ("fictitious page %p is referenced", m)); 4211 KASSERT(m->a.queue == PQ_NONE, 4212 ("fictitious page %p is queued", m)); 4213 return (false); 4214 } 4215 4216 /* 4217 * Pages need not be dequeued before they are returned to the physical 4218 * memory allocator, but they must at least be marked for a deferred 4219 * dequeue. 4220 */ 4221 if ((m->oflags & VPO_UNMANAGED) == 0) 4222 vm_page_dequeue_deferred(m); 4223 4224 m->valid = 0; 4225 vm_page_undirty(m); 4226 4227 if (m->ref_count != 0) 4228 panic("vm_page_free_prep: page %p has references", m); 4229 4230 /* 4231 * Restore the default memory attribute to the page. 4232 */ 4233 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 4234 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 4235 4236 #if VM_NRESERVLEVEL > 0 4237 /* 4238 * Determine whether the page belongs to a reservation. If the page was 4239 * allocated from a per-CPU cache, it cannot belong to a reservation, so 4240 * as an optimization, we avoid the check in that case. 4241 */ 4242 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 4243 return (false); 4244 #endif 4245 4246 return (true); 4247 } 4248 4249 /* 4250 * vm_page_free_toq: 4251 * 4252 * Returns the given page to the free list, disassociating it 4253 * from any VM object. 4254 * 4255 * The object must be locked. The page must be exclusively busied if it 4256 * belongs to an object. 4257 */ 4258 static void 4259 vm_page_free_toq(vm_page_t m) 4260 { 4261 struct vm_domain *vmd; 4262 uma_zone_t zone; 4263 4264 if (!vm_page_free_prep(m)) 4265 return; 4266 4267 vmd = vm_pagequeue_domain(m); 4268 zone = vmd->vmd_pgcache[m->pool].zone; 4269 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4270 uma_zfree(zone, m); 4271 return; 4272 } 4273 vm_domain_free_lock(vmd); 4274 vm_phys_free_pages(m, 0); 4275 vm_domain_free_unlock(vmd); 4276 vm_domain_freecnt_inc(vmd, 1); 4277 } 4278 4279 /* 4280 * vm_page_free_pages_toq: 4281 * 4282 * Returns a list of pages to the free list, disassociating it 4283 * from any VM object. In other words, this is equivalent to 4284 * calling vm_page_free_toq() for each page of a list of VM objects. 4285 */ 4286 int 4287 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4288 { 4289 vm_page_t m; 4290 int count; 4291 4292 if (SLIST_EMPTY(free)) 4293 return (0); 4294 4295 count = 0; 4296 while ((m = SLIST_FIRST(free)) != NULL) { 4297 count++; 4298 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4299 vm_page_free_toq(m); 4300 } 4301 4302 if (update_wire_count) 4303 vm_wire_sub(count); 4304 return (count); 4305 } 4306 4307 /* 4308 * Mark this page as wired down. For managed pages, this prevents reclamation 4309 * by the page daemon, or when the containing object, if any, is destroyed. 4310 */ 4311 void 4312 vm_page_wire(vm_page_t m) 4313 { 4314 u_int old; 4315 4316 #ifdef INVARIANTS 4317 if (m->object != NULL && !vm_page_busied(m) && 4318 !vm_object_busied(m->object)) 4319 VM_OBJECT_ASSERT_LOCKED(m->object); 4320 #endif 4321 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4322 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4323 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4324 4325 old = atomic_fetchadd_int(&m->ref_count, 1); 4326 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4327 ("vm_page_wire: counter overflow for page %p", m)); 4328 if (VPRC_WIRE_COUNT(old) == 0) { 4329 if ((m->oflags & VPO_UNMANAGED) == 0) 4330 vm_page_aflag_set(m, PGA_DEQUEUE); 4331 vm_wire_add(1); 4332 } 4333 } 4334 4335 /* 4336 * Attempt to wire a mapped page following a pmap lookup of that page. 4337 * This may fail if a thread is concurrently tearing down mappings of the page. 4338 * The transient failure is acceptable because it translates to the 4339 * failure of the caller pmap_extract_and_hold(), which should be then 4340 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4341 */ 4342 bool 4343 vm_page_wire_mapped(vm_page_t m) 4344 { 4345 u_int old; 4346 4347 old = atomic_load_int(&m->ref_count); 4348 do { 4349 KASSERT(old > 0, 4350 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4351 if ((old & VPRC_BLOCKED) != 0) 4352 return (false); 4353 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4354 4355 if (VPRC_WIRE_COUNT(old) == 0) { 4356 if ((m->oflags & VPO_UNMANAGED) == 0) 4357 vm_page_aflag_set(m, PGA_DEQUEUE); 4358 vm_wire_add(1); 4359 } 4360 return (true); 4361 } 4362 4363 /* 4364 * Release a wiring reference to a managed page. If the page still belongs to 4365 * an object, update its position in the page queues to reflect the reference. 4366 * If the wiring was the last reference to the page, free the page. 4367 */ 4368 static void 4369 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4370 { 4371 u_int old; 4372 4373 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4374 ("%s: page %p is unmanaged", __func__, m)); 4375 4376 /* 4377 * Update LRU state before releasing the wiring reference. 4378 * Use a release store when updating the reference count to 4379 * synchronize with vm_page_free_prep(). 4380 */ 4381 old = atomic_load_int(&m->ref_count); 4382 do { 4383 u_int count; 4384 4385 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4386 ("vm_page_unwire: wire count underflow for page %p", m)); 4387 4388 count = old & ~VPRC_BLOCKED; 4389 if (count > VPRC_OBJREF + 1) { 4390 /* 4391 * The page has at least one other wiring reference. An 4392 * earlier iteration of this loop may have called 4393 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4394 * re-set it if necessary. 4395 */ 4396 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4397 vm_page_aflag_set(m, PGA_DEQUEUE); 4398 } else if (count == VPRC_OBJREF + 1) { 4399 /* 4400 * This is the last wiring. Clear PGA_DEQUEUE and 4401 * update the page's queue state to reflect the 4402 * reference. If the page does not belong to an object 4403 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4404 * clear leftover queue state. 4405 */ 4406 vm_page_release_toq(m, nqueue, noreuse); 4407 } else if (count == 1) { 4408 vm_page_aflag_clear(m, PGA_DEQUEUE); 4409 } 4410 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4411 4412 if (VPRC_WIRE_COUNT(old) == 1) { 4413 vm_wire_sub(1); 4414 if (old == 1) 4415 vm_page_free(m); 4416 } 4417 } 4418 4419 /* 4420 * Release one wiring of the specified page, potentially allowing it to be 4421 * paged out. 4422 * 4423 * Only managed pages belonging to an object can be paged out. If the number 4424 * of wirings transitions to zero and the page is eligible for page out, then 4425 * the page is added to the specified paging queue. If the released wiring 4426 * represented the last reference to the page, the page is freed. 4427 */ 4428 void 4429 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4430 { 4431 4432 KASSERT(nqueue < PQ_COUNT, 4433 ("vm_page_unwire: invalid queue %u request for page %p", 4434 nqueue, m)); 4435 4436 if ((m->oflags & VPO_UNMANAGED) != 0) { 4437 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4438 vm_page_free(m); 4439 return; 4440 } 4441 vm_page_unwire_managed(m, nqueue, false); 4442 } 4443 4444 /* 4445 * Unwire a page without (re-)inserting it into a page queue. It is up 4446 * to the caller to enqueue, requeue, or free the page as appropriate. 4447 * In most cases involving managed pages, vm_page_unwire() should be used 4448 * instead. 4449 */ 4450 bool 4451 vm_page_unwire_noq(vm_page_t m) 4452 { 4453 u_int old; 4454 4455 old = vm_page_drop(m, 1); 4456 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4457 ("%s: counter underflow for page %p", __func__, m)); 4458 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4459 ("%s: missing ref on fictitious page %p", __func__, m)); 4460 4461 if (VPRC_WIRE_COUNT(old) > 1) 4462 return (false); 4463 if ((m->oflags & VPO_UNMANAGED) == 0) 4464 vm_page_aflag_clear(m, PGA_DEQUEUE); 4465 vm_wire_sub(1); 4466 return (true); 4467 } 4468 4469 /* 4470 * Ensure that the page ends up in the specified page queue. If the page is 4471 * active or being moved to the active queue, ensure that its act_count is 4472 * at least ACT_INIT but do not otherwise mess with it. 4473 */ 4474 static __always_inline void 4475 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4476 { 4477 vm_page_astate_t old, new; 4478 4479 KASSERT(m->ref_count > 0, 4480 ("%s: page %p does not carry any references", __func__, m)); 4481 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4482 ("%s: invalid flags %x", __func__, nflag)); 4483 4484 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4485 return; 4486 4487 old = vm_page_astate_load(m); 4488 do { 4489 if ((old.flags & PGA_DEQUEUE) != 0) 4490 break; 4491 new = old; 4492 new.flags &= ~PGA_QUEUE_OP_MASK; 4493 if (nqueue == PQ_ACTIVE) 4494 new.act_count = max(old.act_count, ACT_INIT); 4495 if (old.queue == nqueue) { 4496 /* 4497 * There is no need to requeue pages already in the 4498 * active queue. 4499 */ 4500 if (nqueue != PQ_ACTIVE || 4501 (old.flags & PGA_ENQUEUED) == 0) 4502 new.flags |= nflag; 4503 } else { 4504 new.flags |= nflag; 4505 new.queue = nqueue; 4506 } 4507 } while (!vm_page_pqstate_commit(m, &old, new)); 4508 } 4509 4510 /* 4511 * Put the specified page on the active list (if appropriate). 4512 */ 4513 void 4514 vm_page_activate(vm_page_t m) 4515 { 4516 4517 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4518 } 4519 4520 /* 4521 * Move the specified page to the tail of the inactive queue, or requeue 4522 * the page if it is already in the inactive queue. 4523 */ 4524 void 4525 vm_page_deactivate(vm_page_t m) 4526 { 4527 4528 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4529 } 4530 4531 void 4532 vm_page_deactivate_noreuse(vm_page_t m) 4533 { 4534 4535 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4536 } 4537 4538 /* 4539 * Put a page in the laundry, or requeue it if it is already there. 4540 */ 4541 void 4542 vm_page_launder(vm_page_t m) 4543 { 4544 4545 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4546 } 4547 4548 /* 4549 * Put a page in the PQ_UNSWAPPABLE holding queue. 4550 */ 4551 void 4552 vm_page_unswappable(vm_page_t m) 4553 { 4554 4555 VM_OBJECT_ASSERT_LOCKED(m->object); 4556 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4557 ("page %p already unswappable", m)); 4558 4559 vm_page_dequeue(m); 4560 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4561 } 4562 4563 /* 4564 * Release a page back to the page queues in preparation for unwiring. 4565 */ 4566 static void 4567 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4568 { 4569 vm_page_astate_t old, new; 4570 uint16_t nflag; 4571 4572 /* 4573 * Use a check of the valid bits to determine whether we should 4574 * accelerate reclamation of the page. The object lock might not be 4575 * held here, in which case the check is racy. At worst we will either 4576 * accelerate reclamation of a valid page and violate LRU, or 4577 * unnecessarily defer reclamation of an invalid page. 4578 * 4579 * If we were asked to not cache the page, place it near the head of the 4580 * inactive queue so that is reclaimed sooner. 4581 */ 4582 if (noreuse || vm_page_none_valid(m)) { 4583 nqueue = PQ_INACTIVE; 4584 nflag = PGA_REQUEUE_HEAD; 4585 } else { 4586 nflag = PGA_REQUEUE; 4587 } 4588 4589 old = vm_page_astate_load(m); 4590 do { 4591 new = old; 4592 4593 /* 4594 * If the page is already in the active queue and we are not 4595 * trying to accelerate reclamation, simply mark it as 4596 * referenced and avoid any queue operations. 4597 */ 4598 new.flags &= ~PGA_QUEUE_OP_MASK; 4599 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4600 (old.flags & PGA_ENQUEUED) != 0) 4601 new.flags |= PGA_REFERENCED; 4602 else { 4603 new.flags |= nflag; 4604 new.queue = nqueue; 4605 } 4606 } while (!vm_page_pqstate_commit(m, &old, new)); 4607 } 4608 4609 /* 4610 * Unwire a page and either attempt to free it or re-add it to the page queues. 4611 */ 4612 void 4613 vm_page_release(vm_page_t m, int flags) 4614 { 4615 vm_object_t object; 4616 4617 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4618 ("vm_page_release: page %p is unmanaged", m)); 4619 4620 if ((flags & VPR_TRYFREE) != 0) { 4621 for (;;) { 4622 object = atomic_load_ptr(&m->object); 4623 if (object == NULL) 4624 break; 4625 /* Depends on type-stability. */ 4626 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4627 break; 4628 if (object == m->object) { 4629 vm_page_release_locked(m, flags); 4630 VM_OBJECT_WUNLOCK(object); 4631 return; 4632 } 4633 VM_OBJECT_WUNLOCK(object); 4634 } 4635 } 4636 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4637 } 4638 4639 /* See vm_page_release(). */ 4640 void 4641 vm_page_release_locked(vm_page_t m, int flags) 4642 { 4643 4644 VM_OBJECT_ASSERT_WLOCKED(m->object); 4645 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4646 ("vm_page_release_locked: page %p is unmanaged", m)); 4647 4648 if (vm_page_unwire_noq(m)) { 4649 if ((flags & VPR_TRYFREE) != 0 && 4650 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4651 m->dirty == 0 && vm_page_tryxbusy(m)) { 4652 /* 4653 * An unlocked lookup may have wired the page before the 4654 * busy lock was acquired, in which case the page must 4655 * not be freed. 4656 */ 4657 if (__predict_true(!vm_page_wired(m))) { 4658 vm_page_free(m); 4659 return; 4660 } 4661 vm_page_xunbusy(m); 4662 } else { 4663 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4664 } 4665 } 4666 } 4667 4668 static bool 4669 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4670 { 4671 u_int old; 4672 4673 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4674 ("vm_page_try_blocked_op: page %p has no object", m)); 4675 KASSERT(vm_page_busied(m), 4676 ("vm_page_try_blocked_op: page %p is not busy", m)); 4677 VM_OBJECT_ASSERT_LOCKED(m->object); 4678 4679 old = atomic_load_int(&m->ref_count); 4680 do { 4681 KASSERT(old != 0, 4682 ("vm_page_try_blocked_op: page %p has no references", m)); 4683 KASSERT((old & VPRC_BLOCKED) == 0, 4684 ("vm_page_try_blocked_op: page %p blocks wirings", m)); 4685 if (VPRC_WIRE_COUNT(old) != 0) 4686 return (false); 4687 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4688 4689 (op)(m); 4690 4691 /* 4692 * If the object is read-locked, new wirings may be created via an 4693 * object lookup. 4694 */ 4695 old = vm_page_drop(m, VPRC_BLOCKED); 4696 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4697 old == (VPRC_BLOCKED | VPRC_OBJREF), 4698 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4699 old, m)); 4700 return (true); 4701 } 4702 4703 /* 4704 * Atomically check for wirings and remove all mappings of the page. 4705 */ 4706 bool 4707 vm_page_try_remove_all(vm_page_t m) 4708 { 4709 4710 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4711 } 4712 4713 /* 4714 * Atomically check for wirings and remove all writeable mappings of the page. 4715 */ 4716 bool 4717 vm_page_try_remove_write(vm_page_t m) 4718 { 4719 4720 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4721 } 4722 4723 /* 4724 * vm_page_advise 4725 * 4726 * Apply the specified advice to the given page. 4727 */ 4728 void 4729 vm_page_advise(vm_page_t m, int advice) 4730 { 4731 4732 VM_OBJECT_ASSERT_WLOCKED(m->object); 4733 vm_page_assert_xbusied(m); 4734 4735 if (advice == MADV_FREE) 4736 /* 4737 * Mark the page clean. This will allow the page to be freed 4738 * without first paging it out. MADV_FREE pages are often 4739 * quickly reused by malloc(3), so we do not do anything that 4740 * would result in a page fault on a later access. 4741 */ 4742 vm_page_undirty(m); 4743 else if (advice != MADV_DONTNEED) { 4744 if (advice == MADV_WILLNEED) 4745 vm_page_activate(m); 4746 return; 4747 } 4748 4749 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4750 vm_page_dirty(m); 4751 4752 /* 4753 * Clear any references to the page. Otherwise, the page daemon will 4754 * immediately reactivate the page. 4755 */ 4756 vm_page_aflag_clear(m, PGA_REFERENCED); 4757 4758 /* 4759 * Place clean pages near the head of the inactive queue rather than 4760 * the tail, thus defeating the queue's LRU operation and ensuring that 4761 * the page will be reused quickly. Dirty pages not already in the 4762 * laundry are moved there. 4763 */ 4764 if (m->dirty == 0) 4765 vm_page_deactivate_noreuse(m); 4766 else if (!vm_page_in_laundry(m)) 4767 vm_page_launder(m); 4768 } 4769 4770 /* 4771 * vm_page_grab_release 4772 * 4773 * Helper routine for grab functions to release busy on return. 4774 */ 4775 static inline void 4776 vm_page_grab_release(vm_page_t m, int allocflags) 4777 { 4778 4779 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4780 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4781 vm_page_sunbusy(m); 4782 else 4783 vm_page_xunbusy(m); 4784 } 4785 } 4786 4787 /* 4788 * vm_page_grab_sleep 4789 * 4790 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4791 * if the caller should retry and false otherwise. 4792 * 4793 * If the object is locked on entry the object will be unlocked with 4794 * false returns and still locked but possibly having been dropped 4795 * with true returns. 4796 */ 4797 static bool 4798 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4799 const char *wmesg, int allocflags, bool locked) 4800 { 4801 4802 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4803 return (false); 4804 4805 /* 4806 * Reference the page before unlocking and sleeping so that 4807 * the page daemon is less likely to reclaim it. 4808 */ 4809 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4810 vm_page_reference(m); 4811 4812 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4813 locked) 4814 VM_OBJECT_WLOCK(object); 4815 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4816 return (false); 4817 4818 return (true); 4819 } 4820 4821 /* 4822 * Assert that the grab flags are valid. 4823 */ 4824 static inline void 4825 vm_page_grab_check(int allocflags) 4826 { 4827 4828 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4829 (allocflags & VM_ALLOC_WIRED) != 0, 4830 ("vm_page_grab*: the pages must be busied or wired")); 4831 4832 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4833 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4834 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4835 } 4836 4837 /* 4838 * Calculate the page allocation flags for grab. 4839 */ 4840 static inline int 4841 vm_page_grab_pflags(int allocflags) 4842 { 4843 int pflags; 4844 4845 pflags = allocflags & 4846 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4847 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY); 4848 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4849 pflags |= VM_ALLOC_WAITFAIL; 4850 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4851 pflags |= VM_ALLOC_SBUSY; 4852 4853 return (pflags); 4854 } 4855 4856 /* 4857 * Grab a page, waiting until we are waken up due to the page 4858 * changing state. We keep on waiting, if the page continues 4859 * to be in the object. If the page doesn't exist, first allocate it 4860 * and then conditionally zero it. 4861 * 4862 * This routine may sleep. 4863 * 4864 * The object must be locked on entry. The lock will, however, be released 4865 * and reacquired if the routine sleeps. 4866 */ 4867 vm_page_t 4868 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4869 { 4870 vm_page_t m; 4871 4872 VM_OBJECT_ASSERT_WLOCKED(object); 4873 vm_page_grab_check(allocflags); 4874 4875 retrylookup: 4876 if ((m = vm_page_lookup(object, pindex)) != NULL) { 4877 if (!vm_page_tryacquire(m, allocflags)) { 4878 if (vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4879 allocflags, true)) 4880 goto retrylookup; 4881 return (NULL); 4882 } 4883 goto out; 4884 } 4885 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4886 return (NULL); 4887 m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags)); 4888 if (m == NULL) { 4889 if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4890 return (NULL); 4891 goto retrylookup; 4892 } 4893 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 4894 pmap_zero_page(m); 4895 4896 out: 4897 vm_page_grab_release(m, allocflags); 4898 4899 return (m); 4900 } 4901 4902 /* 4903 * Attempt to validate a page, locklessly acquiring it if necessary, given a 4904 * (object, pindex) tuple and either an invalided page or NULL. The resulting 4905 * page will be validated against the identity tuple, and busied or wired as 4906 * requested. A NULL page returned guarantees that the page was not in radix at 4907 * the time of the call but callers must perform higher level synchronization or 4908 * retry the operation under a lock if they require an atomic answer. This is 4909 * the only lock free validation routine, other routines can depend on the 4910 * resulting page state. 4911 * 4912 * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to 4913 * caller flags. 4914 */ 4915 #define PAGE_NOT_ACQUIRED ((vm_page_t)1) 4916 static vm_page_t 4917 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m, 4918 int allocflags) 4919 { 4920 if (m == NULL) 4921 m = vm_page_lookup_unlocked(object, pindex); 4922 for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) { 4923 if (vm_page_trybusy(m, allocflags)) { 4924 if (m->object == object && m->pindex == pindex) { 4925 if ((allocflags & VM_ALLOC_WIRED) != 0) 4926 vm_page_wire(m); 4927 vm_page_grab_release(m, allocflags); 4928 break; 4929 } 4930 /* relookup. */ 4931 vm_page_busy_release(m); 4932 cpu_spinwait(); 4933 continue; 4934 } 4935 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4936 allocflags, false)) 4937 return (PAGE_NOT_ACQUIRED); 4938 } 4939 return (m); 4940 } 4941 4942 /* 4943 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4944 * is not set. 4945 */ 4946 vm_page_t 4947 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4948 { 4949 vm_page_t m; 4950 4951 vm_page_grab_check(allocflags); 4952 m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags); 4953 if (m == PAGE_NOT_ACQUIRED) 4954 return (NULL); 4955 if (m != NULL) 4956 return (m); 4957 4958 /* 4959 * The radix lockless lookup should never return a false negative 4960 * errors. If the user specifies NOCREAT they are guaranteed there 4961 * was no page present at the instant of the call. A NOCREAT caller 4962 * must handle create races gracefully. 4963 */ 4964 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4965 return (NULL); 4966 4967 VM_OBJECT_WLOCK(object); 4968 m = vm_page_grab(object, pindex, allocflags); 4969 VM_OBJECT_WUNLOCK(object); 4970 4971 return (m); 4972 } 4973 4974 /* 4975 * Grab a page and make it valid, paging in if necessary. Pages missing from 4976 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 4977 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 4978 * in simultaneously. Additional pages will be left on a paging queue but 4979 * will neither be wired nor busy regardless of allocflags. 4980 */ 4981 int 4982 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) 4983 { 4984 vm_page_t m; 4985 vm_page_t ma[VM_INITIAL_PAGEIN]; 4986 int after, i, pflags, rv; 4987 4988 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4989 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4990 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4991 KASSERT((allocflags & 4992 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4993 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4994 VM_OBJECT_ASSERT_WLOCKED(object); 4995 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4996 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4997 pflags |= VM_ALLOC_WAITFAIL; 4998 4999 retrylookup: 5000 if ((m = vm_page_lookup(object, pindex)) != NULL) { 5001 /* 5002 * If the page is fully valid it can only become invalid 5003 * with the object lock held. If it is not valid it can 5004 * become valid with the busy lock held. Therefore, we 5005 * may unnecessarily lock the exclusive busy here if we 5006 * race with I/O completion not using the object lock. 5007 * However, we will not end up with an invalid page and a 5008 * shared lock. 5009 */ 5010 if (!vm_page_trybusy(m, 5011 vm_page_all_valid(m) ? allocflags : 0)) { 5012 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 5013 allocflags, true); 5014 goto retrylookup; 5015 } 5016 if (vm_page_all_valid(m)) 5017 goto out; 5018 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5019 vm_page_busy_release(m); 5020 *mp = NULL; 5021 return (VM_PAGER_FAIL); 5022 } 5023 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5024 *mp = NULL; 5025 return (VM_PAGER_FAIL); 5026 } else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) { 5027 if (!vm_pager_can_alloc_page(object, pindex)) { 5028 *mp = NULL; 5029 return (VM_PAGER_AGAIN); 5030 } 5031 goto retrylookup; 5032 } 5033 5034 vm_page_assert_xbusied(m); 5035 if (vm_pager_has_page(object, pindex, NULL, &after)) { 5036 after = MIN(after, VM_INITIAL_PAGEIN); 5037 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 5038 after = MAX(after, 1); 5039 ma[0] = m; 5040 for (i = 1; i < after; i++) { 5041 if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { 5042 if (vm_page_any_valid(ma[i]) || 5043 !vm_page_tryxbusy(ma[i])) 5044 break; 5045 } else { 5046 ma[i] = vm_page_alloc(object, m->pindex + i, 5047 VM_ALLOC_NORMAL); 5048 if (ma[i] == NULL) 5049 break; 5050 } 5051 } 5052 after = i; 5053 vm_object_pip_add(object, after); 5054 VM_OBJECT_WUNLOCK(object); 5055 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 5056 VM_OBJECT_WLOCK(object); 5057 vm_object_pip_wakeupn(object, after); 5058 /* Pager may have replaced a page. */ 5059 m = ma[0]; 5060 if (rv != VM_PAGER_OK) { 5061 for (i = 0; i < after; i++) { 5062 if (!vm_page_wired(ma[i])) 5063 vm_page_free(ma[i]); 5064 else 5065 vm_page_xunbusy(ma[i]); 5066 } 5067 *mp = NULL; 5068 return (rv); 5069 } 5070 for (i = 1; i < after; i++) 5071 vm_page_readahead_finish(ma[i]); 5072 MPASS(vm_page_all_valid(m)); 5073 } else { 5074 vm_page_zero_invalid(m, TRUE); 5075 } 5076 out: 5077 if ((allocflags & VM_ALLOC_WIRED) != 0) 5078 vm_page_wire(m); 5079 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 5080 vm_page_busy_downgrade(m); 5081 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 5082 vm_page_busy_release(m); 5083 *mp = m; 5084 return (VM_PAGER_OK); 5085 } 5086 5087 /* 5088 * Locklessly grab a valid page. If the page is not valid or not yet 5089 * allocated this will fall back to the object lock method. 5090 */ 5091 int 5092 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 5093 vm_pindex_t pindex, int allocflags) 5094 { 5095 vm_page_t m; 5096 int flags; 5097 int error; 5098 5099 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 5100 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 5101 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 5102 "mismatch")); 5103 KASSERT((allocflags & 5104 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 5105 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 5106 5107 /* 5108 * Attempt a lockless lookup and busy. We need at least an sbusy 5109 * before we can inspect the valid field and return a wired page. 5110 */ 5111 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 5112 vm_page_grab_check(flags); 5113 m = vm_page_acquire_unlocked(object, pindex, NULL, flags); 5114 if (m == PAGE_NOT_ACQUIRED) 5115 return (VM_PAGER_FAIL); 5116 if (m != NULL) { 5117 if (vm_page_all_valid(m)) { 5118 if ((allocflags & VM_ALLOC_WIRED) != 0) 5119 vm_page_wire(m); 5120 vm_page_grab_release(m, allocflags); 5121 *mp = m; 5122 return (VM_PAGER_OK); 5123 } 5124 vm_page_busy_release(m); 5125 } 5126 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5127 *mp = NULL; 5128 return (VM_PAGER_FAIL); 5129 } 5130 VM_OBJECT_WLOCK(object); 5131 error = vm_page_grab_valid(mp, object, pindex, allocflags); 5132 VM_OBJECT_WUNLOCK(object); 5133 5134 return (error); 5135 } 5136 5137 /* 5138 * Return the specified range of pages from the given object. For each 5139 * page offset within the range, if a page already exists within the object 5140 * at that offset and it is busy, then wait for it to change state. If, 5141 * instead, the page doesn't exist, then allocate it. 5142 * 5143 * The caller must always specify an allocation class. 5144 * 5145 * allocation classes: 5146 * VM_ALLOC_NORMAL normal process request 5147 * VM_ALLOC_SYSTEM system *really* needs the pages 5148 * 5149 * The caller must always specify that the pages are to be busied and/or 5150 * wired. 5151 * 5152 * optional allocation flags: 5153 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 5154 * VM_ALLOC_NOBUSY do not exclusive busy the page 5155 * VM_ALLOC_NOWAIT do not sleep 5156 * VM_ALLOC_SBUSY set page to sbusy state 5157 * VM_ALLOC_WIRED wire the pages 5158 * VM_ALLOC_ZERO zero and validate any invalid pages 5159 * 5160 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 5161 * may return a partial prefix of the requested range. 5162 */ 5163 int 5164 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 5165 vm_page_t *ma, int count) 5166 { 5167 vm_page_t m, mpred; 5168 int pflags; 5169 int i; 5170 5171 VM_OBJECT_ASSERT_WLOCKED(object); 5172 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 5173 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 5174 KASSERT(count > 0, 5175 ("vm_page_grab_pages: invalid page count %d", count)); 5176 vm_page_grab_check(allocflags); 5177 5178 pflags = vm_page_grab_pflags(allocflags); 5179 i = 0; 5180 retrylookup: 5181 m = vm_page_mpred(object, pindex + i); 5182 if (m == NULL || m->pindex != pindex + i) { 5183 mpred = m; 5184 m = NULL; 5185 } else 5186 mpred = TAILQ_PREV(m, pglist, listq); 5187 for (; i < count; i++) { 5188 if (m != NULL) { 5189 if (!vm_page_tryacquire(m, allocflags)) { 5190 if (vm_page_grab_sleep(object, m, pindex + i, 5191 "grbmaw", allocflags, true)) 5192 goto retrylookup; 5193 break; 5194 } 5195 } else { 5196 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 5197 break; 5198 m = vm_page_alloc_after(object, pindex + i, 5199 pflags | VM_ALLOC_COUNT(count - i), mpred); 5200 if (m == NULL) { 5201 if ((allocflags & (VM_ALLOC_NOWAIT | 5202 VM_ALLOC_WAITFAIL)) != 0) 5203 break; 5204 goto retrylookup; 5205 } 5206 } 5207 if (vm_page_none_valid(m) && 5208 (allocflags & VM_ALLOC_ZERO) != 0) { 5209 if ((m->flags & PG_ZERO) == 0) 5210 pmap_zero_page(m); 5211 vm_page_valid(m); 5212 } 5213 vm_page_grab_release(m, allocflags); 5214 ma[i] = mpred = m; 5215 m = vm_page_next(m); 5216 } 5217 return (i); 5218 } 5219 5220 /* 5221 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 5222 * and will fall back to the locked variant to handle allocation. 5223 */ 5224 int 5225 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 5226 int allocflags, vm_page_t *ma, int count) 5227 { 5228 vm_page_t m; 5229 int flags; 5230 int i; 5231 5232 KASSERT(count > 0, 5233 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 5234 vm_page_grab_check(allocflags); 5235 5236 /* 5237 * Modify flags for lockless acquire to hold the page until we 5238 * set it valid if necessary. 5239 */ 5240 flags = allocflags & ~VM_ALLOC_NOBUSY; 5241 vm_page_grab_check(flags); 5242 m = NULL; 5243 for (i = 0; i < count; i++, pindex++) { 5244 /* 5245 * We may see a false NULL here because the previous page has 5246 * been removed or just inserted and the list is loaded without 5247 * barriers. Switch to radix to verify. 5248 */ 5249 if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex || 5250 atomic_load_ptr(&m->object) != object) { 5251 /* 5252 * This guarantees the result is instantaneously 5253 * correct. 5254 */ 5255 m = NULL; 5256 } 5257 m = vm_page_acquire_unlocked(object, pindex, m, flags); 5258 if (m == PAGE_NOT_ACQUIRED) 5259 return (i); 5260 if (m == NULL) 5261 break; 5262 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 5263 if ((m->flags & PG_ZERO) == 0) 5264 pmap_zero_page(m); 5265 vm_page_valid(m); 5266 } 5267 /* m will still be wired or busy according to flags. */ 5268 vm_page_grab_release(m, allocflags); 5269 ma[i] = m; 5270 m = TAILQ_NEXT(m, listq); 5271 } 5272 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 5273 return (i); 5274 count -= i; 5275 VM_OBJECT_WLOCK(object); 5276 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 5277 VM_OBJECT_WUNLOCK(object); 5278 5279 return (i); 5280 } 5281 5282 /* 5283 * Mapping function for valid or dirty bits in a page. 5284 * 5285 * Inputs are required to range within a page. 5286 */ 5287 vm_page_bits_t 5288 vm_page_bits(int base, int size) 5289 { 5290 int first_bit; 5291 int last_bit; 5292 5293 KASSERT( 5294 base + size <= PAGE_SIZE, 5295 ("vm_page_bits: illegal base/size %d/%d", base, size) 5296 ); 5297 5298 if (size == 0) /* handle degenerate case */ 5299 return (0); 5300 5301 first_bit = base >> DEV_BSHIFT; 5302 last_bit = (base + size - 1) >> DEV_BSHIFT; 5303 5304 return (((vm_page_bits_t)2 << last_bit) - 5305 ((vm_page_bits_t)1 << first_bit)); 5306 } 5307 5308 void 5309 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5310 { 5311 5312 #if PAGE_SIZE == 32768 5313 atomic_set_64((uint64_t *)bits, set); 5314 #elif PAGE_SIZE == 16384 5315 atomic_set_32((uint32_t *)bits, set); 5316 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5317 atomic_set_16((uint16_t *)bits, set); 5318 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5319 atomic_set_8((uint8_t *)bits, set); 5320 #else /* PAGE_SIZE <= 8192 */ 5321 uintptr_t addr; 5322 int shift; 5323 5324 addr = (uintptr_t)bits; 5325 /* 5326 * Use a trick to perform a 32-bit atomic on the 5327 * containing aligned word, to not depend on the existence 5328 * of atomic_{set, clear}_{8, 16}. 5329 */ 5330 shift = addr & (sizeof(uint32_t) - 1); 5331 #if BYTE_ORDER == BIG_ENDIAN 5332 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5333 #else 5334 shift *= NBBY; 5335 #endif 5336 addr &= ~(sizeof(uint32_t) - 1); 5337 atomic_set_32((uint32_t *)addr, set << shift); 5338 #endif /* PAGE_SIZE */ 5339 } 5340 5341 static inline void 5342 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5343 { 5344 5345 #if PAGE_SIZE == 32768 5346 atomic_clear_64((uint64_t *)bits, clear); 5347 #elif PAGE_SIZE == 16384 5348 atomic_clear_32((uint32_t *)bits, clear); 5349 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5350 atomic_clear_16((uint16_t *)bits, clear); 5351 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5352 atomic_clear_8((uint8_t *)bits, clear); 5353 #else /* PAGE_SIZE <= 8192 */ 5354 uintptr_t addr; 5355 int shift; 5356 5357 addr = (uintptr_t)bits; 5358 /* 5359 * Use a trick to perform a 32-bit atomic on the 5360 * containing aligned word, to not depend on the existence 5361 * of atomic_{set, clear}_{8, 16}. 5362 */ 5363 shift = addr & (sizeof(uint32_t) - 1); 5364 #if BYTE_ORDER == BIG_ENDIAN 5365 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5366 #else 5367 shift *= NBBY; 5368 #endif 5369 addr &= ~(sizeof(uint32_t) - 1); 5370 atomic_clear_32((uint32_t *)addr, clear << shift); 5371 #endif /* PAGE_SIZE */ 5372 } 5373 5374 static inline vm_page_bits_t 5375 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5376 { 5377 #if PAGE_SIZE == 32768 5378 uint64_t old; 5379 5380 old = *bits; 5381 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5382 return (old); 5383 #elif PAGE_SIZE == 16384 5384 uint32_t old; 5385 5386 old = *bits; 5387 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5388 return (old); 5389 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5390 uint16_t old; 5391 5392 old = *bits; 5393 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5394 return (old); 5395 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5396 uint8_t old; 5397 5398 old = *bits; 5399 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5400 return (old); 5401 #else /* PAGE_SIZE <= 4096*/ 5402 uintptr_t addr; 5403 uint32_t old, new, mask; 5404 int shift; 5405 5406 addr = (uintptr_t)bits; 5407 /* 5408 * Use a trick to perform a 32-bit atomic on the 5409 * containing aligned word, to not depend on the existence 5410 * of atomic_{set, swap, clear}_{8, 16}. 5411 */ 5412 shift = addr & (sizeof(uint32_t) - 1); 5413 #if BYTE_ORDER == BIG_ENDIAN 5414 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5415 #else 5416 shift *= NBBY; 5417 #endif 5418 addr &= ~(sizeof(uint32_t) - 1); 5419 mask = VM_PAGE_BITS_ALL << shift; 5420 5421 old = *bits; 5422 do { 5423 new = old & ~mask; 5424 new |= newbits << shift; 5425 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5426 return (old >> shift); 5427 #endif /* PAGE_SIZE */ 5428 } 5429 5430 /* 5431 * vm_page_set_valid_range: 5432 * 5433 * Sets portions of a page valid. The arguments are expected 5434 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5435 * of any partial chunks touched by the range. The invalid portion of 5436 * such chunks will be zeroed. 5437 * 5438 * (base + size) must be less then or equal to PAGE_SIZE. 5439 */ 5440 void 5441 vm_page_set_valid_range(vm_page_t m, int base, int size) 5442 { 5443 int endoff, frag; 5444 vm_page_bits_t pagebits; 5445 5446 vm_page_assert_busied(m); 5447 if (size == 0) /* handle degenerate case */ 5448 return; 5449 5450 /* 5451 * If the base is not DEV_BSIZE aligned and the valid 5452 * bit is clear, we have to zero out a portion of the 5453 * first block. 5454 */ 5455 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5456 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5457 pmap_zero_page_area(m, frag, base - frag); 5458 5459 /* 5460 * If the ending offset is not DEV_BSIZE aligned and the 5461 * valid bit is clear, we have to zero out a portion of 5462 * the last block. 5463 */ 5464 endoff = base + size; 5465 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5466 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5467 pmap_zero_page_area(m, endoff, 5468 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5469 5470 /* 5471 * Assert that no previously invalid block that is now being validated 5472 * is already dirty. 5473 */ 5474 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5475 ("vm_page_set_valid_range: page %p is dirty", m)); 5476 5477 /* 5478 * Set valid bits inclusive of any overlap. 5479 */ 5480 pagebits = vm_page_bits(base, size); 5481 if (vm_page_xbusied(m)) 5482 m->valid |= pagebits; 5483 else 5484 vm_page_bits_set(m, &m->valid, pagebits); 5485 } 5486 5487 /* 5488 * Set the page dirty bits and free the invalid swap space if 5489 * present. Returns the previous dirty bits. 5490 */ 5491 vm_page_bits_t 5492 vm_page_set_dirty(vm_page_t m) 5493 { 5494 vm_page_bits_t old; 5495 5496 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5497 5498 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5499 old = m->dirty; 5500 m->dirty = VM_PAGE_BITS_ALL; 5501 } else 5502 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5503 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5504 vm_pager_page_unswapped(m); 5505 5506 return (old); 5507 } 5508 5509 /* 5510 * Clear the given bits from the specified page's dirty field. 5511 */ 5512 static __inline void 5513 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5514 { 5515 5516 vm_page_assert_busied(m); 5517 5518 /* 5519 * If the page is xbusied and not write mapped we are the 5520 * only thread that can modify dirty bits. Otherwise, The pmap 5521 * layer can call vm_page_dirty() without holding a distinguished 5522 * lock. The combination of page busy and atomic operations 5523 * suffice to guarantee consistency of the page dirty field. 5524 */ 5525 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5526 m->dirty &= ~pagebits; 5527 else 5528 vm_page_bits_clear(m, &m->dirty, pagebits); 5529 } 5530 5531 /* 5532 * vm_page_set_validclean: 5533 * 5534 * Sets portions of a page valid and clean. The arguments are expected 5535 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5536 * of any partial chunks touched by the range. The invalid portion of 5537 * such chunks will be zero'd. 5538 * 5539 * (base + size) must be less then or equal to PAGE_SIZE. 5540 */ 5541 void 5542 vm_page_set_validclean(vm_page_t m, int base, int size) 5543 { 5544 vm_page_bits_t oldvalid, pagebits; 5545 int endoff, frag; 5546 5547 vm_page_assert_busied(m); 5548 if (size == 0) /* handle degenerate case */ 5549 return; 5550 5551 /* 5552 * If the base is not DEV_BSIZE aligned and the valid 5553 * bit is clear, we have to zero out a portion of the 5554 * first block. 5555 */ 5556 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5557 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5558 pmap_zero_page_area(m, frag, base - frag); 5559 5560 /* 5561 * If the ending offset is not DEV_BSIZE aligned and the 5562 * valid bit is clear, we have to zero out a portion of 5563 * the last block. 5564 */ 5565 endoff = base + size; 5566 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5567 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5568 pmap_zero_page_area(m, endoff, 5569 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5570 5571 /* 5572 * Set valid, clear dirty bits. If validating the entire 5573 * page we can safely clear the pmap modify bit. We also 5574 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5575 * takes a write fault on a MAP_NOSYNC memory area the flag will 5576 * be set again. 5577 * 5578 * We set valid bits inclusive of any overlap, but we can only 5579 * clear dirty bits for DEV_BSIZE chunks that are fully within 5580 * the range. 5581 */ 5582 oldvalid = m->valid; 5583 pagebits = vm_page_bits(base, size); 5584 if (vm_page_xbusied(m)) 5585 m->valid |= pagebits; 5586 else 5587 vm_page_bits_set(m, &m->valid, pagebits); 5588 #if 0 /* NOT YET */ 5589 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5590 frag = DEV_BSIZE - frag; 5591 base += frag; 5592 size -= frag; 5593 if (size < 0) 5594 size = 0; 5595 } 5596 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5597 #endif 5598 if (base == 0 && size == PAGE_SIZE) { 5599 /* 5600 * The page can only be modified within the pmap if it is 5601 * mapped, and it can only be mapped if it was previously 5602 * fully valid. 5603 */ 5604 if (oldvalid == VM_PAGE_BITS_ALL) 5605 /* 5606 * Perform the pmap_clear_modify() first. Otherwise, 5607 * a concurrent pmap operation, such as 5608 * pmap_protect(), could clear a modification in the 5609 * pmap and set the dirty field on the page before 5610 * pmap_clear_modify() had begun and after the dirty 5611 * field was cleared here. 5612 */ 5613 pmap_clear_modify(m); 5614 m->dirty = 0; 5615 vm_page_aflag_clear(m, PGA_NOSYNC); 5616 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5617 m->dirty &= ~pagebits; 5618 else 5619 vm_page_clear_dirty_mask(m, pagebits); 5620 } 5621 5622 void 5623 vm_page_clear_dirty(vm_page_t m, int base, int size) 5624 { 5625 5626 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5627 } 5628 5629 /* 5630 * vm_page_set_invalid: 5631 * 5632 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5633 * valid and dirty bits for the effected areas are cleared. 5634 */ 5635 void 5636 vm_page_set_invalid(vm_page_t m, int base, int size) 5637 { 5638 vm_page_bits_t bits; 5639 vm_object_t object; 5640 5641 /* 5642 * The object lock is required so that pages can't be mapped 5643 * read-only while we're in the process of invalidating them. 5644 */ 5645 object = m->object; 5646 VM_OBJECT_ASSERT_WLOCKED(object); 5647 vm_page_assert_busied(m); 5648 5649 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5650 size >= object->un_pager.vnp.vnp_size) 5651 bits = VM_PAGE_BITS_ALL; 5652 else 5653 bits = vm_page_bits(base, size); 5654 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5655 pmap_remove_all(m); 5656 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5657 !pmap_page_is_mapped(m), 5658 ("vm_page_set_invalid: page %p is mapped", m)); 5659 if (vm_page_xbusied(m)) { 5660 m->valid &= ~bits; 5661 m->dirty &= ~bits; 5662 } else { 5663 vm_page_bits_clear(m, &m->valid, bits); 5664 vm_page_bits_clear(m, &m->dirty, bits); 5665 } 5666 } 5667 5668 /* 5669 * vm_page_invalid: 5670 * 5671 * Invalidates the entire page. The page must be busy, unmapped, and 5672 * the enclosing object must be locked. The object locks protects 5673 * against concurrent read-only pmap enter which is done without 5674 * busy. 5675 */ 5676 void 5677 vm_page_invalid(vm_page_t m) 5678 { 5679 5680 vm_page_assert_busied(m); 5681 VM_OBJECT_ASSERT_WLOCKED(m->object); 5682 MPASS(!pmap_page_is_mapped(m)); 5683 5684 if (vm_page_xbusied(m)) 5685 m->valid = 0; 5686 else 5687 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5688 } 5689 5690 /* 5691 * vm_page_zero_invalid() 5692 * 5693 * The kernel assumes that the invalid portions of a page contain 5694 * garbage, but such pages can be mapped into memory by user code. 5695 * When this occurs, we must zero out the non-valid portions of the 5696 * page so user code sees what it expects. 5697 * 5698 * Pages are most often semi-valid when the end of a file is mapped 5699 * into memory and the file's size is not page aligned. 5700 */ 5701 void 5702 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5703 { 5704 int b; 5705 int i; 5706 5707 /* 5708 * Scan the valid bits looking for invalid sections that 5709 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5710 * valid bit may be set ) have already been zeroed by 5711 * vm_page_set_validclean(). 5712 */ 5713 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5714 if (i == (PAGE_SIZE / DEV_BSIZE) || 5715 (m->valid & ((vm_page_bits_t)1 << i))) { 5716 if (i > b) { 5717 pmap_zero_page_area(m, 5718 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5719 } 5720 b = i + 1; 5721 } 5722 } 5723 5724 /* 5725 * setvalid is TRUE when we can safely set the zero'd areas 5726 * as being valid. We can do this if there are no cache consistency 5727 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5728 */ 5729 if (setvalid) 5730 vm_page_valid(m); 5731 } 5732 5733 /* 5734 * vm_page_is_valid: 5735 * 5736 * Is (partial) page valid? Note that the case where size == 0 5737 * will return FALSE in the degenerate case where the page is 5738 * entirely invalid, and TRUE otherwise. 5739 * 5740 * Some callers envoke this routine without the busy lock held and 5741 * handle races via higher level locks. Typical callers should 5742 * hold a busy lock to prevent invalidation. 5743 */ 5744 int 5745 vm_page_is_valid(vm_page_t m, int base, int size) 5746 { 5747 vm_page_bits_t bits; 5748 5749 bits = vm_page_bits(base, size); 5750 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5751 } 5752 5753 /* 5754 * Returns true if all of the specified predicates are true for the entire 5755 * (super)page and false otherwise. 5756 */ 5757 bool 5758 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m) 5759 { 5760 vm_object_t object; 5761 int i, npages; 5762 5763 object = m->object; 5764 if (skip_m != NULL && skip_m->object != object) 5765 return (false); 5766 VM_OBJECT_ASSERT_LOCKED(object); 5767 KASSERT(psind <= m->psind, 5768 ("psind %d > psind %d of m %p", psind, m->psind, m)); 5769 npages = atop(pagesizes[psind]); 5770 5771 /* 5772 * The physically contiguous pages that make up a superpage, i.e., a 5773 * page with a page size index ("psind") greater than zero, will 5774 * occupy adjacent entries in vm_page_array[]. 5775 */ 5776 for (i = 0; i < npages; i++) { 5777 /* Always test object consistency, including "skip_m". */ 5778 if (m[i].object != object) 5779 return (false); 5780 if (&m[i] == skip_m) 5781 continue; 5782 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5783 return (false); 5784 if ((flags & PS_ALL_DIRTY) != 0) { 5785 /* 5786 * Calling vm_page_test_dirty() or pmap_is_modified() 5787 * might stop this case from spuriously returning 5788 * "false". However, that would require a write lock 5789 * on the object containing "m[i]". 5790 */ 5791 if (m[i].dirty != VM_PAGE_BITS_ALL) 5792 return (false); 5793 } 5794 if ((flags & PS_ALL_VALID) != 0 && 5795 m[i].valid != VM_PAGE_BITS_ALL) 5796 return (false); 5797 } 5798 return (true); 5799 } 5800 5801 /* 5802 * Set the page's dirty bits if the page is modified. 5803 */ 5804 void 5805 vm_page_test_dirty(vm_page_t m) 5806 { 5807 5808 vm_page_assert_busied(m); 5809 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5810 vm_page_dirty(m); 5811 } 5812 5813 void 5814 vm_page_valid(vm_page_t m) 5815 { 5816 5817 vm_page_assert_busied(m); 5818 if (vm_page_xbusied(m)) 5819 m->valid = VM_PAGE_BITS_ALL; 5820 else 5821 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5822 } 5823 5824 void 5825 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5826 { 5827 5828 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5829 } 5830 5831 void 5832 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5833 { 5834 5835 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5836 } 5837 5838 int 5839 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5840 { 5841 5842 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5843 } 5844 5845 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5846 void 5847 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5848 { 5849 5850 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5851 } 5852 5853 void 5854 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5855 { 5856 5857 mtx_assert_(vm_page_lockptr(m), a, file, line); 5858 } 5859 #endif 5860 5861 #ifdef INVARIANTS 5862 void 5863 vm_page_object_busy_assert(vm_page_t m) 5864 { 5865 5866 /* 5867 * Certain of the page's fields may only be modified by the 5868 * holder of a page or object busy. 5869 */ 5870 if (m->object != NULL && !vm_page_busied(m)) 5871 VM_OBJECT_ASSERT_BUSY(m->object); 5872 } 5873 5874 void 5875 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5876 { 5877 5878 if ((bits & PGA_WRITEABLE) == 0) 5879 return; 5880 5881 /* 5882 * The PGA_WRITEABLE flag can only be set if the page is 5883 * managed, is exclusively busied or the object is locked. 5884 * Currently, this flag is only set by pmap_enter(). 5885 */ 5886 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5887 ("PGA_WRITEABLE on unmanaged page")); 5888 if (!vm_page_xbusied(m)) 5889 VM_OBJECT_ASSERT_BUSY(m->object); 5890 } 5891 #endif 5892 5893 #include "opt_ddb.h" 5894 #ifdef DDB 5895 #include <sys/kernel.h> 5896 5897 #include <ddb/ddb.h> 5898 5899 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5900 { 5901 5902 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5903 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5904 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5905 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5906 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5907 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5908 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5909 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5910 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5911 } 5912 5913 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5914 { 5915 int dom; 5916 5917 db_printf("pq_free %d\n", vm_free_count()); 5918 for (dom = 0; dom < vm_ndomains; dom++) { 5919 db_printf( 5920 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5921 dom, 5922 vm_dom[dom].vmd_page_count, 5923 vm_dom[dom].vmd_free_count, 5924 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5925 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5926 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5927 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5928 } 5929 } 5930 5931 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5932 { 5933 vm_page_t m; 5934 boolean_t phys, virt; 5935 5936 if (!have_addr) { 5937 db_printf("show pginfo addr\n"); 5938 return; 5939 } 5940 5941 phys = strchr(modif, 'p') != NULL; 5942 virt = strchr(modif, 'v') != NULL; 5943 if (virt) 5944 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5945 else if (phys) 5946 m = PHYS_TO_VM_PAGE(addr); 5947 else 5948 m = (vm_page_t)addr; 5949 db_printf( 5950 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5951 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5952 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5953 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5954 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5955 } 5956 #endif /* DDB */ 5957