1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * The Mach Operating System project at Carnegie-Mellon University. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 36 /*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Resident memory management module. 65 */ 66 67 #include <sys/cdefs.h> 68 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/counter.h> 73 #include <sys/domainset.h> 74 #include <sys/kernel.h> 75 #include <sys/limits.h> 76 #include <sys/linker.h> 77 #include <sys/lock.h> 78 #include <sys/malloc.h> 79 #include <sys/mman.h> 80 #include <sys/msgbuf.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> 83 #include <sys/rwlock.h> 84 #include <sys/sleepqueue.h> 85 #include <sys/sbuf.h> 86 #include <sys/sched.h> 87 #include <sys/smp.h> 88 #include <sys/sysctl.h> 89 #include <sys/vmmeter.h> 90 #include <sys/vnode.h> 91 92 #include <vm/vm.h> 93 #include <vm/pmap.h> 94 #include <vm/vm_param.h> 95 #include <vm/vm_domainset.h> 96 #include <vm/vm_kern.h> 97 #include <vm/vm_map.h> 98 #include <vm/vm_object.h> 99 #include <vm/vm_page.h> 100 #include <vm/vm_pageout.h> 101 #include <vm/vm_phys.h> 102 #include <vm/vm_pagequeue.h> 103 #include <vm/vm_pager.h> 104 #include <vm/vm_radix.h> 105 #include <vm/vm_reserv.h> 106 #include <vm/vm_extern.h> 107 #include <vm/vm_dumpset.h> 108 #include <vm/uma.h> 109 #include <vm/uma_int.h> 110 111 #include <machine/md_var.h> 112 113 struct vm_domain vm_dom[MAXMEMDOM]; 114 115 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 116 117 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 118 119 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 120 /* The following fields are protected by the domainset lock. */ 121 domainset_t __exclusive_cache_line vm_min_domains; 122 domainset_t __exclusive_cache_line vm_severe_domains; 123 static int vm_min_waiters; 124 static int vm_severe_waiters; 125 static int vm_pageproc_waiters; 126 127 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 128 "VM page statistics"); 129 130 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 131 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 132 CTLFLAG_RD, &pqstate_commit_retries, 133 "Number of failed per-page atomic queue state updates"); 134 135 static COUNTER_U64_DEFINE_EARLY(queue_ops); 136 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 137 CTLFLAG_RD, &queue_ops, 138 "Number of batched queue operations"); 139 140 static COUNTER_U64_DEFINE_EARLY(queue_nops); 141 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 142 CTLFLAG_RD, &queue_nops, 143 "Number of batched queue operations with no effects"); 144 145 /* 146 * bogus page -- for I/O to/from partially complete buffers, 147 * or for paging into sparsely invalid regions. 148 */ 149 vm_page_t bogus_page; 150 151 vm_page_t vm_page_array; 152 long vm_page_array_size; 153 long first_page; 154 155 struct bitset *vm_page_dump; 156 long vm_page_dump_pages; 157 158 static TAILQ_HEAD(, vm_page) blacklist_head; 159 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 160 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 161 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 162 163 static uma_zone_t fakepg_zone; 164 165 static void vm_page_alloc_check(vm_page_t m); 166 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req); 167 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 168 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 169 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 170 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 171 static bool vm_page_free_prep(vm_page_t m); 172 static void vm_page_free_toq(vm_page_t m); 173 static void vm_page_init(void *dummy); 174 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 175 vm_page_t mpred); 176 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 177 const uint16_t nflag); 178 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 179 vm_page_t m_run, vm_paddr_t high); 180 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 181 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 182 int req); 183 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 184 int flags); 185 static void vm_page_zone_release(void *arg, void **store, int cnt); 186 187 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 188 189 static void 190 vm_page_init(void *dummy) 191 { 192 193 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 194 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 195 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE); 196 } 197 198 static int pgcache_zone_max_pcpu; 199 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu, 200 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0, 201 "Per-CPU page cache size"); 202 203 /* 204 * The cache page zone is initialized later since we need to be able to allocate 205 * pages before UMA is fully initialized. 206 */ 207 static void 208 vm_page_init_cache_zones(void *dummy __unused) 209 { 210 struct vm_domain *vmd; 211 struct vm_pgcache *pgcache; 212 int cache, domain, maxcache, pool; 213 214 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu); 215 maxcache = pgcache_zone_max_pcpu * mp_ncpus; 216 for (domain = 0; domain < vm_ndomains; domain++) { 217 vmd = VM_DOMAIN(domain); 218 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 219 pgcache = &vmd->vmd_pgcache[pool]; 220 pgcache->domain = domain; 221 pgcache->pool = pool; 222 pgcache->zone = uma_zcache_create("vm pgcache", 223 PAGE_SIZE, NULL, NULL, NULL, NULL, 224 vm_page_zone_import, vm_page_zone_release, pgcache, 225 UMA_ZONE_VM); 226 227 /* 228 * Limit each pool's zone to 0.1% of the pages in the 229 * domain. 230 */ 231 cache = maxcache != 0 ? maxcache : 232 vmd->vmd_page_count / 1000; 233 uma_zone_set_maxcache(pgcache->zone, cache); 234 } 235 } 236 } 237 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 238 239 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 240 #if PAGE_SIZE == 32768 241 #ifdef CTASSERT 242 CTASSERT(sizeof(u_long) >= 8); 243 #endif 244 #endif 245 246 /* 247 * vm_set_page_size: 248 * 249 * Sets the page size, perhaps based upon the memory 250 * size. Must be called before any use of page-size 251 * dependent functions. 252 */ 253 void 254 vm_set_page_size(void) 255 { 256 if (vm_cnt.v_page_size == 0) 257 vm_cnt.v_page_size = PAGE_SIZE; 258 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 259 panic("vm_set_page_size: page size not a power of two"); 260 } 261 262 /* 263 * vm_page_blacklist_next: 264 * 265 * Find the next entry in the provided string of blacklist 266 * addresses. Entries are separated by space, comma, or newline. 267 * If an invalid integer is encountered then the rest of the 268 * string is skipped. Updates the list pointer to the next 269 * character, or NULL if the string is exhausted or invalid. 270 */ 271 static vm_paddr_t 272 vm_page_blacklist_next(char **list, char *end) 273 { 274 vm_paddr_t bad; 275 char *cp, *pos; 276 277 if (list == NULL || *list == NULL) 278 return (0); 279 if (**list =='\0') { 280 *list = NULL; 281 return (0); 282 } 283 284 /* 285 * If there's no end pointer then the buffer is coming from 286 * the kenv and we know it's null-terminated. 287 */ 288 if (end == NULL) 289 end = *list + strlen(*list); 290 291 /* Ensure that strtoq() won't walk off the end */ 292 if (*end != '\0') { 293 if (*end == '\n' || *end == ' ' || *end == ',') 294 *end = '\0'; 295 else { 296 printf("Blacklist not terminated, skipping\n"); 297 *list = NULL; 298 return (0); 299 } 300 } 301 302 for (pos = *list; *pos != '\0'; pos = cp) { 303 bad = strtoq(pos, &cp, 0); 304 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 305 if (bad == 0) { 306 if (++cp < end) 307 continue; 308 else 309 break; 310 } 311 } else 312 break; 313 if (*cp == '\0' || ++cp >= end) 314 *list = NULL; 315 else 316 *list = cp; 317 return (trunc_page(bad)); 318 } 319 printf("Garbage in RAM blacklist, skipping\n"); 320 *list = NULL; 321 return (0); 322 } 323 324 bool 325 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 326 { 327 struct vm_domain *vmd; 328 vm_page_t m; 329 bool found; 330 331 m = vm_phys_paddr_to_vm_page(pa); 332 if (m == NULL) 333 return (true); /* page does not exist, no failure */ 334 335 vmd = VM_DOMAIN(vm_phys_domain(pa)); 336 vm_domain_free_lock(vmd); 337 found = vm_phys_unfree_page(pa); 338 vm_domain_free_unlock(vmd); 339 if (found) { 340 vm_domain_freecnt_inc(vmd, -1); 341 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 342 if (verbose) 343 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 344 } 345 return (found); 346 } 347 348 /* 349 * vm_page_blacklist_check: 350 * 351 * Iterate through the provided string of blacklist addresses, pulling 352 * each entry out of the physical allocator free list and putting it 353 * onto a list for reporting via the vm.page_blacklist sysctl. 354 */ 355 static void 356 vm_page_blacklist_check(char *list, char *end) 357 { 358 vm_paddr_t pa; 359 char *next; 360 361 next = list; 362 while (next != NULL) { 363 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 364 continue; 365 vm_page_blacklist_add(pa, bootverbose); 366 } 367 } 368 369 /* 370 * vm_page_blacklist_load: 371 * 372 * Search for a special module named "ram_blacklist". It'll be a 373 * plain text file provided by the user via the loader directive 374 * of the same name. 375 */ 376 static void 377 vm_page_blacklist_load(char **list, char **end) 378 { 379 void *mod; 380 u_char *ptr; 381 u_int len; 382 383 mod = NULL; 384 ptr = NULL; 385 386 mod = preload_search_by_type("ram_blacklist"); 387 if (mod != NULL) { 388 ptr = preload_fetch_addr(mod); 389 len = preload_fetch_size(mod); 390 } 391 *list = ptr; 392 if (ptr != NULL) 393 *end = ptr + len; 394 else 395 *end = NULL; 396 return; 397 } 398 399 static int 400 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 401 { 402 vm_page_t m; 403 struct sbuf sbuf; 404 int error, first; 405 406 first = 1; 407 error = sysctl_wire_old_buffer(req, 0); 408 if (error != 0) 409 return (error); 410 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 411 TAILQ_FOREACH(m, &blacklist_head, listq) { 412 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 413 (uintmax_t)m->phys_addr); 414 first = 0; 415 } 416 error = sbuf_finish(&sbuf); 417 sbuf_delete(&sbuf); 418 return (error); 419 } 420 421 /* 422 * Initialize a dummy page for use in scans of the specified paging queue. 423 * In principle, this function only needs to set the flag PG_MARKER. 424 * Nonetheless, it write busies the page as a safety precaution. 425 */ 426 void 427 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 428 { 429 430 bzero(marker, sizeof(*marker)); 431 marker->flags = PG_MARKER; 432 marker->a.flags = aflags; 433 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 434 marker->a.queue = queue; 435 } 436 437 static void 438 vm_page_domain_init(int domain) 439 { 440 struct vm_domain *vmd; 441 struct vm_pagequeue *pq; 442 int i; 443 444 vmd = VM_DOMAIN(domain); 445 bzero(vmd, sizeof(*vmd)); 446 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 447 "vm inactive pagequeue"; 448 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 449 "vm active pagequeue"; 450 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 451 "vm laundry pagequeue"; 452 *__DECONST(const char **, 453 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 454 "vm unswappable pagequeue"; 455 vmd->vmd_domain = domain; 456 vmd->vmd_page_count = 0; 457 vmd->vmd_free_count = 0; 458 vmd->vmd_segs = 0; 459 vmd->vmd_oom = false; 460 vmd->vmd_helper_threads_enabled = true; 461 for (i = 0; i < PQ_COUNT; i++) { 462 pq = &vmd->vmd_pagequeues[i]; 463 TAILQ_INIT(&pq->pq_pl); 464 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 465 MTX_DEF | MTX_DUPOK); 466 pq->pq_pdpages = 0; 467 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 468 } 469 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 470 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 471 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 472 473 /* 474 * inacthead is used to provide FIFO ordering for LRU-bypassing 475 * insertions. 476 */ 477 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 478 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 479 &vmd->vmd_inacthead, plinks.q); 480 481 /* 482 * The clock pages are used to implement active queue scanning without 483 * requeues. Scans start at clock[0], which is advanced after the scan 484 * ends. When the two clock hands meet, they are reset and scanning 485 * resumes from the head of the queue. 486 */ 487 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 488 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 489 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 490 &vmd->vmd_clock[0], plinks.q); 491 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 492 &vmd->vmd_clock[1], plinks.q); 493 } 494 495 /* 496 * Initialize a physical page in preparation for adding it to the free 497 * lists. 498 */ 499 void 500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool) 501 { 502 m->object = NULL; 503 m->ref_count = 0; 504 m->busy_lock = VPB_FREED; 505 m->flags = m->a.flags = 0; 506 m->phys_addr = pa; 507 m->a.queue = PQ_NONE; 508 m->psind = 0; 509 m->segind = segind; 510 m->order = VM_NFREEORDER; 511 m->pool = pool; 512 m->valid = m->dirty = 0; 513 pmap_page_init(m); 514 } 515 516 #ifndef PMAP_HAS_PAGE_ARRAY 517 static vm_paddr_t 518 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 519 { 520 vm_paddr_t new_end; 521 522 /* 523 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 524 * However, because this page is allocated from KVM, out-of-bounds 525 * accesses using the direct map will not be trapped. 526 */ 527 *vaddr += PAGE_SIZE; 528 529 /* 530 * Allocate physical memory for the page structures, and map it. 531 */ 532 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 533 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 534 VM_PROT_READ | VM_PROT_WRITE); 535 vm_page_array_size = page_range; 536 537 return (new_end); 538 } 539 #endif 540 541 /* 542 * vm_page_startup: 543 * 544 * Initializes the resident memory module. Allocates physical memory for 545 * bootstrapping UMA and some data structures that are used to manage 546 * physical pages. Initializes these structures, and populates the free 547 * page queues. 548 */ 549 vm_offset_t 550 vm_page_startup(vm_offset_t vaddr) 551 { 552 struct vm_phys_seg *seg; 553 struct vm_domain *vmd; 554 vm_page_t m; 555 char *list, *listend; 556 vm_paddr_t end, high_avail, low_avail, new_end, size; 557 vm_paddr_t page_range __unused; 558 vm_paddr_t last_pa, pa, startp, endp; 559 u_long pagecount; 560 #if MINIDUMP_PAGE_TRACKING 561 u_long vm_page_dump_size; 562 #endif 563 int biggestone, i, segind; 564 #ifdef WITNESS 565 vm_offset_t mapped; 566 int witness_size; 567 #endif 568 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 569 long ii; 570 #endif 571 int pool; 572 #ifdef VM_FREEPOOL_LAZYINIT 573 int lazyinit; 574 #endif 575 576 vaddr = round_page(vaddr); 577 578 vm_phys_early_startup(); 579 biggestone = vm_phys_avail_largest(); 580 end = phys_avail[biggestone+1]; 581 582 /* 583 * Initialize the page and queue locks. 584 */ 585 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 586 for (i = 0; i < PA_LOCK_COUNT; i++) 587 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 588 for (i = 0; i < vm_ndomains; i++) 589 vm_page_domain_init(i); 590 591 new_end = end; 592 #ifdef WITNESS 593 witness_size = round_page(witness_startup_count()); 594 new_end -= witness_size; 595 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 596 VM_PROT_READ | VM_PROT_WRITE); 597 bzero((void *)mapped, witness_size); 598 witness_startup((void *)mapped); 599 #endif 600 601 #if MINIDUMP_PAGE_TRACKING 602 /* 603 * Allocate a bitmap to indicate that a random physical page 604 * needs to be included in a minidump. 605 * 606 * The amd64 port needs this to indicate which direct map pages 607 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 608 * 609 * However, i386 still needs this workspace internally within the 610 * minidump code. In theory, they are not needed on i386, but are 611 * included should the sf_buf code decide to use them. 612 */ 613 last_pa = 0; 614 vm_page_dump_pages = 0; 615 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 616 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 617 dump_avail[i] / PAGE_SIZE; 618 if (dump_avail[i + 1] > last_pa) 619 last_pa = dump_avail[i + 1]; 620 } 621 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 622 new_end -= vm_page_dump_size; 623 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 624 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 625 bzero((void *)vm_page_dump, vm_page_dump_size); 626 #if MINIDUMP_STARTUP_PAGE_TRACKING 627 /* 628 * Include the UMA bootstrap pages, witness pages and vm_page_dump 629 * in a crash dump. When pmap_map() uses the direct map, they are 630 * not automatically included. 631 */ 632 for (pa = new_end; pa < end; pa += PAGE_SIZE) 633 dump_add_page(pa); 634 #endif 635 #else 636 (void)last_pa; 637 #endif 638 phys_avail[biggestone + 1] = new_end; 639 #ifdef __amd64__ 640 /* 641 * Request that the physical pages underlying the message buffer be 642 * included in a crash dump. Since the message buffer is accessed 643 * through the direct map, they are not automatically included. 644 */ 645 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 646 last_pa = pa + round_page(msgbufsize); 647 while (pa < last_pa) { 648 dump_add_page(pa); 649 pa += PAGE_SIZE; 650 } 651 #else 652 (void)pa; 653 #endif 654 655 /* 656 * Determine the lowest and highest physical addresses and, in the case 657 * of VM_PHYSSEG_SPARSE, the exact size of the available physical 658 * memory. vm_phys_early_startup() already checked that phys_avail[] 659 * has at least one element. 660 */ 661 #ifdef VM_PHYSSEG_SPARSE 662 size = phys_avail[1] - phys_avail[0]; 663 #endif 664 low_avail = phys_avail[0]; 665 high_avail = phys_avail[1]; 666 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 667 #ifdef VM_PHYSSEG_SPARSE 668 size += phys_avail[i + 1] - phys_avail[i]; 669 #endif 670 if (phys_avail[i] < low_avail) 671 low_avail = phys_avail[i]; 672 if (phys_avail[i + 1] > high_avail) 673 high_avail = phys_avail[i + 1]; 674 } 675 for (i = 0; i < vm_phys_nsegs; i++) { 676 #ifdef VM_PHYSSEG_SPARSE 677 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 678 #endif 679 if (vm_phys_segs[i].start < low_avail) 680 low_avail = vm_phys_segs[i].start; 681 if (vm_phys_segs[i].end > high_avail) 682 high_avail = vm_phys_segs[i].end; 683 } 684 first_page = low_avail / PAGE_SIZE; 685 #ifdef VM_PHYSSEG_DENSE 686 size = high_avail - low_avail; 687 #endif 688 689 #ifdef PMAP_HAS_PAGE_ARRAY 690 pmap_page_array_startup(size / PAGE_SIZE); 691 biggestone = vm_phys_avail_largest(); 692 end = new_end = phys_avail[biggestone + 1]; 693 #else 694 #ifdef VM_PHYSSEG_DENSE 695 /* 696 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 697 * the overhead of a page structure per page only if vm_page_array is 698 * allocated from the last physical memory chunk. Otherwise, we must 699 * allocate page structures representing the physical memory 700 * underlying vm_page_array, even though they will not be used. 701 */ 702 if (new_end != high_avail) 703 page_range = size / PAGE_SIZE; 704 else 705 #endif 706 { 707 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 708 709 /* 710 * If the partial bytes remaining are large enough for 711 * a page (PAGE_SIZE) without a corresponding 712 * 'struct vm_page', then new_end will contain an 713 * extra page after subtracting the length of the VM 714 * page array. Compensate by subtracting an extra 715 * page from new_end. 716 */ 717 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 718 if (new_end == high_avail) 719 high_avail -= PAGE_SIZE; 720 new_end -= PAGE_SIZE; 721 } 722 } 723 end = new_end; 724 new_end = vm_page_array_alloc(&vaddr, end, page_range); 725 #endif 726 727 #if VM_NRESERVLEVEL > 0 728 /* 729 * Allocate physical memory for the reservation management system's 730 * data structures, and map it. 731 */ 732 new_end = vm_reserv_startup(&vaddr, new_end); 733 #endif 734 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 735 /* 736 * Include vm_page_array and vm_reserv_array in a crash dump. 737 */ 738 for (pa = new_end; pa < end; pa += PAGE_SIZE) 739 dump_add_page(pa); 740 #endif 741 phys_avail[biggestone + 1] = new_end; 742 743 /* 744 * Add physical memory segments corresponding to the available 745 * physical pages. 746 */ 747 for (i = 0; phys_avail[i + 1] != 0; i += 2) 748 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 749 750 /* 751 * Initialize the physical memory allocator. 752 */ 753 vm_phys_init(); 754 755 pool = VM_FREEPOOL_DEFAULT; 756 #ifdef VM_FREEPOOL_LAZYINIT 757 lazyinit = 1; 758 TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit); 759 if (lazyinit) 760 pool = VM_FREEPOOL_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 * Initialize pages not covered by phys_avail[], since they 781 * might be freed to the allocator at some future point, e.g., 782 * by kmem_bootstrap_free(). 783 */ 784 startp = seg->start; 785 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 786 if (startp >= seg->end) 787 break; 788 if (phys_avail[i + 1] < startp) 789 continue; 790 if (phys_avail[i] <= startp) { 791 startp = phys_avail[i + 1]; 792 continue; 793 } 794 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 795 for (endp = MIN(phys_avail[i], seg->end); 796 startp < endp; startp += PAGE_SIZE, m++) { 797 vm_page_init_page(m, startp, segind, 798 VM_FREEPOOL_DEFAULT); 799 } 800 } 801 802 /* 803 * Add the segment's pages that are covered by one of 804 * phys_avail's ranges to the free lists. 805 */ 806 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 807 if (seg->end <= phys_avail[i] || 808 seg->start >= phys_avail[i + 1]) 809 continue; 810 811 startp = MAX(seg->start, phys_avail[i]); 812 endp = MIN(seg->end, phys_avail[i + 1]); 813 pagecount = (u_long)atop(endp - startp); 814 if (pagecount == 0) 815 continue; 816 817 /* 818 * If lazy vm_page initialization is not enabled, simply 819 * initialize all of the pages in the segment covered by 820 * phys_avail. Otherwise, initialize only the first 821 * page of each run of free pages handed to the vm_phys 822 * allocator, which in turn defers initialization of 823 * pages until they are needed. 824 * 825 * This avoids blocking the boot process for long 826 * periods, which may be relevant for VMs (which ought 827 * to boot as quickly as possible) and/or systems with 828 * large amounts of physical memory. 829 */ 830 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 831 vm_page_init_page(m, startp, segind, pool); 832 if (pool == VM_FREEPOOL_DEFAULT) { 833 for (u_long j = 1; j < pagecount; j++) { 834 vm_page_init_page(&m[j], 835 startp + ptoa((vm_paddr_t)j), 836 segind, pool); 837 } 838 } 839 vmd = VM_DOMAIN(seg->domain); 840 vm_domain_free_lock(vmd); 841 vm_phys_enqueue_contig(m, pool, pagecount); 842 vm_domain_free_unlock(vmd); 843 vm_domain_freecnt_inc(vmd, pagecount); 844 vm_cnt.v_page_count += (u_int)pagecount; 845 vmd->vmd_page_count += (u_int)pagecount; 846 vmd->vmd_segs |= 1UL << segind; 847 } 848 } 849 850 /* 851 * Remove blacklisted pages from the physical memory allocator. 852 */ 853 TAILQ_INIT(&blacklist_head); 854 vm_page_blacklist_load(&list, &listend); 855 vm_page_blacklist_check(list, listend); 856 857 list = kern_getenv("vm.blacklist"); 858 vm_page_blacklist_check(list, NULL); 859 860 freeenv(list); 861 #if VM_NRESERVLEVEL > 0 862 /* 863 * Initialize the reservation management system. 864 */ 865 vm_reserv_init(); 866 #endif 867 868 return (vaddr); 869 } 870 871 void 872 vm_page_reference(vm_page_t m) 873 { 874 875 vm_page_aflag_set(m, PGA_REFERENCED); 876 } 877 878 /* 879 * vm_page_trybusy 880 * 881 * Helper routine for grab functions to trylock busy. 882 * 883 * Returns true on success and false on failure. 884 */ 885 static bool 886 vm_page_trybusy(vm_page_t m, int allocflags) 887 { 888 889 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 890 return (vm_page_trysbusy(m)); 891 else 892 return (vm_page_tryxbusy(m)); 893 } 894 895 /* 896 * vm_page_tryacquire 897 * 898 * Helper routine for grab functions to trylock busy and wire. 899 * 900 * Returns true on success and false on failure. 901 */ 902 static inline bool 903 vm_page_tryacquire(vm_page_t m, int allocflags) 904 { 905 bool locked; 906 907 locked = vm_page_trybusy(m, allocflags); 908 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 909 vm_page_wire(m); 910 return (locked); 911 } 912 913 /* 914 * vm_page_busy_acquire: 915 * 916 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 917 * and drop the object lock if necessary. 918 */ 919 bool 920 vm_page_busy_acquire(vm_page_t m, int allocflags) 921 { 922 vm_object_t obj; 923 bool locked; 924 925 /* 926 * The page-specific object must be cached because page 927 * identity can change during the sleep, causing the 928 * re-lock of a different object. 929 * It is assumed that a reference to the object is already 930 * held by the callers. 931 */ 932 obj = atomic_load_ptr(&m->object); 933 for (;;) { 934 if (vm_page_tryacquire(m, allocflags)) 935 return (true); 936 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 937 return (false); 938 if (obj != NULL) 939 locked = VM_OBJECT_WOWNED(obj); 940 else 941 locked = false; 942 MPASS(locked || vm_page_wired(m)); 943 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 944 locked) && locked) 945 VM_OBJECT_WLOCK(obj); 946 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 947 return (false); 948 KASSERT(m->object == obj || m->object == NULL, 949 ("vm_page_busy_acquire: page %p does not belong to %p", 950 m, obj)); 951 } 952 } 953 954 /* 955 * vm_page_busy_downgrade: 956 * 957 * Downgrade an exclusive busy page into a single shared busy page. 958 */ 959 void 960 vm_page_busy_downgrade(vm_page_t m) 961 { 962 u_int x; 963 964 vm_page_assert_xbusied(m); 965 966 x = vm_page_busy_fetch(m); 967 for (;;) { 968 if (atomic_fcmpset_rel_int(&m->busy_lock, 969 &x, VPB_SHARERS_WORD(1))) 970 break; 971 } 972 if ((x & VPB_BIT_WAITERS) != 0) 973 wakeup(m); 974 } 975 976 /* 977 * 978 * vm_page_busy_tryupgrade: 979 * 980 * Attempt to upgrade a single shared busy into an exclusive busy. 981 */ 982 int 983 vm_page_busy_tryupgrade(vm_page_t m) 984 { 985 u_int ce, x; 986 987 vm_page_assert_sbusied(m); 988 989 x = vm_page_busy_fetch(m); 990 ce = VPB_CURTHREAD_EXCLUSIVE; 991 for (;;) { 992 if (VPB_SHARERS(x) > 1) 993 return (0); 994 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 995 ("vm_page_busy_tryupgrade: invalid lock state")); 996 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 997 ce | (x & VPB_BIT_WAITERS))) 998 continue; 999 return (1); 1000 } 1001 } 1002 1003 /* 1004 * vm_page_sbusied: 1005 * 1006 * Return a positive value if the page is shared busied, 0 otherwise. 1007 */ 1008 int 1009 vm_page_sbusied(vm_page_t m) 1010 { 1011 u_int x; 1012 1013 x = vm_page_busy_fetch(m); 1014 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1015 } 1016 1017 /* 1018 * vm_page_sunbusy: 1019 * 1020 * Shared unbusy a page. 1021 */ 1022 void 1023 vm_page_sunbusy(vm_page_t m) 1024 { 1025 u_int x; 1026 1027 vm_page_assert_sbusied(m); 1028 1029 x = vm_page_busy_fetch(m); 1030 for (;;) { 1031 KASSERT(x != VPB_FREED, 1032 ("vm_page_sunbusy: Unlocking freed page.")); 1033 if (VPB_SHARERS(x) > 1) { 1034 if (atomic_fcmpset_int(&m->busy_lock, &x, 1035 x - VPB_ONE_SHARER)) 1036 break; 1037 continue; 1038 } 1039 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1040 ("vm_page_sunbusy: invalid lock state")); 1041 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1042 continue; 1043 if ((x & VPB_BIT_WAITERS) == 0) 1044 break; 1045 wakeup(m); 1046 break; 1047 } 1048 } 1049 1050 /* 1051 * vm_page_busy_sleep: 1052 * 1053 * Sleep if the page is busy, using the page pointer as wchan. 1054 * This is used to implement the hard-path of the busying mechanism. 1055 * 1056 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1057 * will not sleep if the page is shared-busy. 1058 * 1059 * The object lock must be held on entry. 1060 * 1061 * Returns true if it slept and dropped the object lock, or false 1062 * if there was no sleep and the lock is still held. 1063 */ 1064 bool 1065 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) 1066 { 1067 vm_object_t obj; 1068 1069 obj = m->object; 1070 VM_OBJECT_ASSERT_LOCKED(obj); 1071 1072 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, 1073 true)); 1074 } 1075 1076 /* 1077 * vm_page_busy_sleep_unlocked: 1078 * 1079 * Sleep if the page is busy, using the page pointer as wchan. 1080 * This is used to implement the hard-path of busying mechanism. 1081 * 1082 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1083 * will not sleep if the page is shared-busy. 1084 * 1085 * The object lock must not be held on entry. The operation will 1086 * return if the page changes identity. 1087 */ 1088 void 1089 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1090 const char *wmesg, int allocflags) 1091 { 1092 VM_OBJECT_ASSERT_UNLOCKED(obj); 1093 1094 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); 1095 } 1096 1097 /* 1098 * _vm_page_busy_sleep: 1099 * 1100 * Internal busy sleep function. Verifies the page identity and 1101 * lockstate against parameters. Returns true if it sleeps and 1102 * false otherwise. 1103 * 1104 * allocflags uses VM_ALLOC_* flags to specify the lock required. 1105 * 1106 * If locked is true the lock will be dropped for any true returns 1107 * and held for any false returns. 1108 */ 1109 static bool 1110 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1111 const char *wmesg, int allocflags, bool locked) 1112 { 1113 bool xsleep; 1114 u_int x; 1115 1116 /* 1117 * If the object is busy we must wait for that to drain to zero 1118 * before trying the page again. 1119 */ 1120 if (obj != NULL && vm_object_busied(obj)) { 1121 if (locked) 1122 VM_OBJECT_DROP(obj); 1123 vm_object_busy_wait(obj, wmesg); 1124 return (true); 1125 } 1126 1127 if (!vm_page_busied(m)) 1128 return (false); 1129 1130 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1131 sleepq_lock(m); 1132 x = vm_page_busy_fetch(m); 1133 do { 1134 /* 1135 * If the page changes objects or becomes unlocked we can 1136 * simply return. 1137 */ 1138 if (x == VPB_UNBUSIED || 1139 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1140 m->object != obj || m->pindex != pindex) { 1141 sleepq_release(m); 1142 return (false); 1143 } 1144 if ((x & VPB_BIT_WAITERS) != 0) 1145 break; 1146 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1147 if (locked) 1148 VM_OBJECT_DROP(obj); 1149 DROP_GIANT(); 1150 sleepq_add(m, NULL, wmesg, 0, 0); 1151 sleepq_wait(m, PVM); 1152 PICKUP_GIANT(); 1153 return (true); 1154 } 1155 1156 /* 1157 * vm_page_trysbusy: 1158 * 1159 * Try to shared busy a page. 1160 * If the operation succeeds 1 is returned otherwise 0. 1161 * The operation never sleeps. 1162 */ 1163 int 1164 vm_page_trysbusy(vm_page_t m) 1165 { 1166 vm_object_t obj; 1167 u_int x; 1168 1169 obj = m->object; 1170 x = vm_page_busy_fetch(m); 1171 for (;;) { 1172 if ((x & VPB_BIT_SHARED) == 0) 1173 return (0); 1174 /* 1175 * Reduce the window for transient busies that will trigger 1176 * false negatives in vm_page_ps_test(). 1177 */ 1178 if (obj != NULL && vm_object_busied(obj)) 1179 return (0); 1180 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1181 x + VPB_ONE_SHARER)) 1182 break; 1183 } 1184 1185 /* Refetch the object now that we're guaranteed that it is stable. */ 1186 obj = m->object; 1187 if (obj != NULL && vm_object_busied(obj)) { 1188 vm_page_sunbusy(m); 1189 return (0); 1190 } 1191 return (1); 1192 } 1193 1194 /* 1195 * vm_page_tryxbusy: 1196 * 1197 * Try to exclusive busy a page. 1198 * If the operation succeeds 1 is returned otherwise 0. 1199 * The operation never sleeps. 1200 */ 1201 int 1202 vm_page_tryxbusy(vm_page_t m) 1203 { 1204 vm_object_t obj; 1205 1206 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1207 VPB_CURTHREAD_EXCLUSIVE) == 0) 1208 return (0); 1209 1210 obj = m->object; 1211 if (obj != NULL && vm_object_busied(obj)) { 1212 vm_page_xunbusy(m); 1213 return (0); 1214 } 1215 return (1); 1216 } 1217 1218 static void 1219 vm_page_xunbusy_hard_tail(vm_page_t m) 1220 { 1221 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1222 /* Wake the waiter. */ 1223 wakeup(m); 1224 } 1225 1226 /* 1227 * vm_page_xunbusy_hard: 1228 * 1229 * Called when unbusy has failed because there is a waiter. 1230 */ 1231 void 1232 vm_page_xunbusy_hard(vm_page_t m) 1233 { 1234 vm_page_assert_xbusied(m); 1235 vm_page_xunbusy_hard_tail(m); 1236 } 1237 1238 void 1239 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1240 { 1241 vm_page_assert_xbusied_unchecked(m); 1242 vm_page_xunbusy_hard_tail(m); 1243 } 1244 1245 static void 1246 vm_page_busy_free(vm_page_t m) 1247 { 1248 u_int x; 1249 1250 atomic_thread_fence_rel(); 1251 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1252 if ((x & VPB_BIT_WAITERS) != 0) 1253 wakeup(m); 1254 } 1255 1256 /* 1257 * vm_page_unhold_pages: 1258 * 1259 * Unhold each of the pages that is referenced by the given array. 1260 */ 1261 void 1262 vm_page_unhold_pages(vm_page_t *ma, int count) 1263 { 1264 1265 for (; count != 0; count--) { 1266 vm_page_unwire(*ma, PQ_ACTIVE); 1267 ma++; 1268 } 1269 } 1270 1271 vm_page_t 1272 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1273 { 1274 vm_page_t m; 1275 1276 #ifdef VM_PHYSSEG_SPARSE 1277 m = vm_phys_paddr_to_vm_page(pa); 1278 if (m == NULL) 1279 m = vm_phys_fictitious_to_vm_page(pa); 1280 return (m); 1281 #elif defined(VM_PHYSSEG_DENSE) 1282 long pi; 1283 1284 pi = atop(pa); 1285 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1286 m = &vm_page_array[pi - first_page]; 1287 return (m); 1288 } 1289 return (vm_phys_fictitious_to_vm_page(pa)); 1290 #else 1291 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1292 #endif 1293 } 1294 1295 /* 1296 * vm_page_getfake: 1297 * 1298 * Create a fictitious page with the specified physical address and 1299 * memory attribute. The memory attribute is the only the machine- 1300 * dependent aspect of a fictitious page that must be initialized. 1301 */ 1302 vm_page_t 1303 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1304 { 1305 vm_page_t m; 1306 1307 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1308 vm_page_initfake(m, paddr, memattr); 1309 return (m); 1310 } 1311 1312 void 1313 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1314 { 1315 1316 if ((m->flags & PG_FICTITIOUS) != 0) { 1317 /* 1318 * The page's memattr might have changed since the 1319 * previous initialization. Update the pmap to the 1320 * new memattr. 1321 */ 1322 goto memattr; 1323 } 1324 m->phys_addr = paddr; 1325 m->a.queue = PQ_NONE; 1326 /* Fictitious pages don't use "segind". */ 1327 m->flags = PG_FICTITIOUS; 1328 /* Fictitious pages don't use "order" or "pool". */ 1329 m->oflags = VPO_UNMANAGED; 1330 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1331 /* Fictitious pages are unevictable. */ 1332 m->ref_count = 1; 1333 pmap_page_init(m); 1334 memattr: 1335 pmap_page_set_memattr(m, memattr); 1336 } 1337 1338 /* 1339 * vm_page_putfake: 1340 * 1341 * Release a fictitious page. 1342 */ 1343 void 1344 vm_page_putfake(vm_page_t m) 1345 { 1346 1347 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1348 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1349 ("vm_page_putfake: bad page %p", m)); 1350 vm_page_assert_xbusied(m); 1351 vm_page_busy_free(m); 1352 uma_zfree(fakepg_zone, m); 1353 } 1354 1355 /* 1356 * vm_page_updatefake: 1357 * 1358 * Update the given fictitious page to the specified physical address and 1359 * memory attribute. 1360 */ 1361 void 1362 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1363 { 1364 1365 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1366 ("vm_page_updatefake: bad page %p", m)); 1367 m->phys_addr = paddr; 1368 pmap_page_set_memattr(m, memattr); 1369 } 1370 1371 /* 1372 * vm_page_free: 1373 * 1374 * Free a page. 1375 */ 1376 void 1377 vm_page_free(vm_page_t m) 1378 { 1379 1380 m->flags &= ~PG_ZERO; 1381 vm_page_free_toq(m); 1382 } 1383 1384 /* 1385 * vm_page_free_zero: 1386 * 1387 * Free a page to the zerod-pages queue 1388 */ 1389 void 1390 vm_page_free_zero(vm_page_t m) 1391 { 1392 1393 m->flags |= PG_ZERO; 1394 vm_page_free_toq(m); 1395 } 1396 1397 /* 1398 * Unbusy and handle the page queueing for a page from a getpages request that 1399 * was optionally read ahead or behind. 1400 */ 1401 void 1402 vm_page_readahead_finish(vm_page_t m) 1403 { 1404 1405 /* We shouldn't put invalid pages on queues. */ 1406 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1407 1408 /* 1409 * Since the page is not the actually needed one, whether it should 1410 * be activated or deactivated is not obvious. Empirical results 1411 * have shown that deactivating the page is usually the best choice, 1412 * unless the page is wanted by another thread. 1413 */ 1414 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1415 vm_page_activate(m); 1416 else 1417 vm_page_deactivate(m); 1418 vm_page_xunbusy_unchecked(m); 1419 } 1420 1421 /* 1422 * Destroy the identity of an invalid page and free it if possible. 1423 * This is intended to be used when reading a page from backing store fails. 1424 */ 1425 void 1426 vm_page_free_invalid(vm_page_t m) 1427 { 1428 1429 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1430 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1431 KASSERT(m->object != NULL, ("page %p has no object", m)); 1432 VM_OBJECT_ASSERT_WLOCKED(m->object); 1433 1434 /* 1435 * We may be attempting to free the page as part of the handling for an 1436 * I/O error, in which case the page was xbusied by a different thread. 1437 */ 1438 vm_page_xbusy_claim(m); 1439 1440 /* 1441 * If someone has wired this page while the object lock 1442 * was not held, then the thread that unwires is responsible 1443 * for freeing the page. Otherwise just free the page now. 1444 * The wire count of this unmapped page cannot change while 1445 * we have the page xbusy and the page's object wlocked. 1446 */ 1447 if (vm_page_remove(m)) 1448 vm_page_free(m); 1449 } 1450 1451 /* 1452 * vm_page_dirty_KBI: [ internal use only ] 1453 * 1454 * Set all bits in the page's dirty field. 1455 * 1456 * The object containing the specified page must be locked if the 1457 * call is made from the machine-independent layer. 1458 * 1459 * See vm_page_clear_dirty_mask(). 1460 * 1461 * This function should only be called by vm_page_dirty(). 1462 */ 1463 void 1464 vm_page_dirty_KBI(vm_page_t m) 1465 { 1466 1467 /* Refer to this operation by its public name. */ 1468 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1469 m->dirty = VM_PAGE_BITS_ALL; 1470 } 1471 1472 /* 1473 * Insert the given page into the given object at the given pindex. mpred is 1474 * used for memq linkage. From vm_page_insert, iter is false, mpred is 1475 * initially NULL, and this procedure looks it up. From vm_page_iter_insert, 1476 * iter is true and mpred is known to the caller to be valid, and may be NULL if 1477 * this will be the page with the lowest pindex. 1478 * 1479 * The procedure is marked __always_inline to suggest to the compiler to 1480 * eliminate the lookup parameter and the associated alternate branch. 1481 */ 1482 static __always_inline int 1483 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1484 struct pctrie_iter *pages, bool iter, vm_page_t mpred) 1485 { 1486 int error; 1487 1488 VM_OBJECT_ASSERT_WLOCKED(object); 1489 KASSERT(m->object == NULL, 1490 ("vm_page_insert: page %p already inserted", m)); 1491 1492 /* 1493 * Record the object/offset pair in this page. 1494 */ 1495 m->object = object; 1496 m->pindex = pindex; 1497 m->ref_count |= VPRC_OBJREF; 1498 1499 /* 1500 * Add this page to the object's radix tree, and look up mpred if 1501 * needed. 1502 */ 1503 if (iter) 1504 error = vm_radix_iter_insert(pages, m); 1505 else 1506 error = vm_radix_insert_lookup_lt(&object->rtree, m, &mpred); 1507 if (__predict_false(error != 0)) { 1508 m->object = NULL; 1509 m->pindex = 0; 1510 m->ref_count &= ~VPRC_OBJREF; 1511 return (1); 1512 } 1513 1514 /* 1515 * Now link into the object's ordered list of backed pages. 1516 */ 1517 vm_page_insert_radixdone(m, object, mpred); 1518 vm_pager_page_inserted(object, m); 1519 return (0); 1520 } 1521 1522 /* 1523 * vm_page_insert: [ internal use only ] 1524 * 1525 * Inserts the given mem entry into the object and object list. 1526 * 1527 * The object must be locked. 1528 */ 1529 int 1530 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1531 { 1532 return (vm_page_insert_lookup(m, object, pindex, NULL, false, NULL)); 1533 } 1534 1535 /* 1536 * vm_page_iter_insert: 1537 * 1538 * Tries to insert the page "m" into the specified object at offset 1539 * "pindex" using the iterator "pages". Returns 0 if the insertion was 1540 * successful. 1541 * 1542 * The page "mpred" must immediately precede the offset "pindex" within 1543 * the specified object. 1544 * 1545 * The object must be locked. 1546 */ 1547 static int 1548 vm_page_iter_insert(struct pctrie_iter *pages, vm_page_t m, vm_object_t object, 1549 vm_pindex_t pindex, vm_page_t mpred) 1550 { 1551 return (vm_page_insert_lookup(m, object, pindex, pages, true, mpred)); 1552 } 1553 1554 /* 1555 * vm_page_insert_radixdone: 1556 * 1557 * Complete page "m" insertion into the specified object after the 1558 * radix trie hooking. 1559 * 1560 * The page "mpred" must precede the offset "m->pindex" within the 1561 * specified object. 1562 * 1563 * The object must be locked. 1564 */ 1565 static void 1566 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1567 { 1568 1569 VM_OBJECT_ASSERT_WLOCKED(object); 1570 KASSERT(object != NULL && m->object == object, 1571 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1572 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1573 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1574 if (mpred != NULL) { 1575 KASSERT(mpred->object == object, 1576 ("vm_page_insert_radixdone: object doesn't contain mpred")); 1577 KASSERT(mpred->pindex < m->pindex, 1578 ("vm_page_insert_radixdone: mpred doesn't precede pindex")); 1579 KASSERT(TAILQ_NEXT(mpred, listq) == NULL || 1580 m->pindex < TAILQ_NEXT(mpred, listq)->pindex, 1581 ("vm_page_insert_radixdone: pindex doesn't precede msucc")); 1582 } else { 1583 KASSERT(TAILQ_EMPTY(&object->memq) || 1584 m->pindex < TAILQ_FIRST(&object->memq)->pindex, 1585 ("vm_page_insert_radixdone: no mpred but not first page")); 1586 } 1587 1588 if (mpred != NULL) 1589 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1590 else 1591 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1592 1593 /* 1594 * Show that the object has one more resident page. 1595 */ 1596 object->resident_page_count++; 1597 1598 /* 1599 * Hold the vnode until the last page is released. 1600 */ 1601 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1602 vhold(object->handle); 1603 1604 /* 1605 * Since we are inserting a new and possibly dirty page, 1606 * update the object's generation count. 1607 */ 1608 if (pmap_page_is_write_mapped(m)) 1609 vm_object_set_writeable_dirty(object); 1610 } 1611 1612 /* 1613 * vm_page_remove_radixdone 1614 * 1615 * Complete page "m" removal from the specified object after the radix trie 1616 * unhooking. 1617 * 1618 * The caller is responsible for updating the page's fields to reflect this 1619 * removal. 1620 */ 1621 static void 1622 vm_page_remove_radixdone(vm_page_t m) 1623 { 1624 vm_object_t object; 1625 1626 vm_page_assert_xbusied(m); 1627 object = m->object; 1628 VM_OBJECT_ASSERT_WLOCKED(object); 1629 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1630 ("page %p is missing its object ref", m)); 1631 1632 /* Deferred free of swap space. */ 1633 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1634 vm_pager_page_unswapped(m); 1635 1636 vm_pager_page_removed(object, m); 1637 m->object = NULL; 1638 1639 /* 1640 * Now remove from the object's list of backed pages. 1641 */ 1642 TAILQ_REMOVE(&object->memq, m, listq); 1643 1644 /* 1645 * And show that the object has one fewer resident page. 1646 */ 1647 object->resident_page_count--; 1648 1649 /* 1650 * The vnode may now be recycled. 1651 */ 1652 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1653 vdrop(object->handle); 1654 } 1655 1656 /* 1657 * vm_page_free_object_prep: 1658 * 1659 * Disassociates the given page from its VM object. 1660 * 1661 * The object must be locked, and the page must be xbusy. 1662 */ 1663 static void 1664 vm_page_free_object_prep(vm_page_t m) 1665 { 1666 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 1667 ((m->object->flags & OBJ_UNMANAGED) != 0), 1668 ("%s: managed flag mismatch for page %p", 1669 __func__, m)); 1670 vm_page_assert_xbusied(m); 1671 1672 /* 1673 * The object reference can be released without an atomic 1674 * operation. 1675 */ 1676 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 1677 m->ref_count == VPRC_OBJREF, 1678 ("%s: page %p has unexpected ref_count %u", 1679 __func__, m, m->ref_count)); 1680 vm_page_remove_radixdone(m); 1681 m->ref_count -= VPRC_OBJREF; 1682 } 1683 1684 /* 1685 * vm_page_iter_free: 1686 * 1687 * Free the given page, and use the iterator to remove it from the radix 1688 * tree. 1689 */ 1690 void 1691 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m) 1692 { 1693 vm_radix_iter_remove(pages); 1694 vm_page_free_object_prep(m); 1695 vm_page_xunbusy(m); 1696 m->flags &= ~PG_ZERO; 1697 vm_page_free_toq(m); 1698 } 1699 1700 /* 1701 * vm_page_remove: 1702 * 1703 * Removes the specified page from its containing object, but does not 1704 * invalidate any backing storage. Returns true if the object's reference 1705 * was the last reference to the page, and false otherwise. 1706 * 1707 * The object must be locked and the page must be exclusively busied. 1708 * The exclusive busy will be released on return. If this is not the 1709 * final ref and the caller does not hold a wire reference it may not 1710 * continue to access the page. 1711 */ 1712 bool 1713 vm_page_remove(vm_page_t m) 1714 { 1715 bool dropped; 1716 1717 dropped = vm_page_remove_xbusy(m); 1718 vm_page_xunbusy(m); 1719 1720 return (dropped); 1721 } 1722 1723 /* 1724 * vm_page_iter_remove: 1725 * 1726 * Remove the current page, and use the iterator to remove it from the 1727 * radix tree. 1728 */ 1729 bool 1730 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m) 1731 { 1732 bool dropped; 1733 1734 vm_radix_iter_remove(pages); 1735 vm_page_remove_radixdone(m); 1736 dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1737 vm_page_xunbusy(m); 1738 1739 return (dropped); 1740 } 1741 1742 /* 1743 * vm_page_radix_remove 1744 * 1745 * Removes the specified page from the radix tree. 1746 */ 1747 static void 1748 vm_page_radix_remove(vm_page_t m) 1749 { 1750 vm_page_t mrem __diagused; 1751 1752 mrem = vm_radix_remove(&m->object->rtree, m->pindex); 1753 KASSERT(mrem == m, 1754 ("removed page %p, expected page %p", mrem, m)); 1755 } 1756 1757 /* 1758 * vm_page_remove_xbusy 1759 * 1760 * Removes the page but leaves the xbusy held. Returns true if this 1761 * removed the final ref and false otherwise. 1762 */ 1763 bool 1764 vm_page_remove_xbusy(vm_page_t m) 1765 { 1766 1767 vm_page_radix_remove(m); 1768 vm_page_remove_radixdone(m); 1769 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1770 } 1771 1772 /* 1773 * vm_page_lookup: 1774 * 1775 * Returns the page associated with the object/offset 1776 * pair specified; if none is found, NULL is returned. 1777 * 1778 * The object must be locked. 1779 */ 1780 vm_page_t 1781 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1782 { 1783 1784 VM_OBJECT_ASSERT_LOCKED(object); 1785 return (vm_radix_lookup(&object->rtree, pindex)); 1786 } 1787 1788 /* 1789 * vm_page_iter_init: 1790 * 1791 * Initialize iterator for vm pages. 1792 */ 1793 void 1794 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object) 1795 { 1796 1797 vm_radix_iter_init(pages, &object->rtree); 1798 } 1799 1800 /* 1801 * vm_page_iter_init: 1802 * 1803 * Initialize iterator for vm pages. 1804 */ 1805 void 1806 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object, 1807 vm_pindex_t limit) 1808 { 1809 1810 vm_radix_iter_limit_init(pages, &object->rtree, limit); 1811 } 1812 1813 /* 1814 * vm_page_lookup_unlocked: 1815 * 1816 * Returns the page associated with the object/offset pair specified; 1817 * if none is found, NULL is returned. The page may be no longer be 1818 * present in the object at the time that this function returns. Only 1819 * useful for opportunistic checks such as inmem(). 1820 */ 1821 vm_page_t 1822 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1823 { 1824 1825 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1826 } 1827 1828 /* 1829 * vm_page_relookup: 1830 * 1831 * Returns a page that must already have been busied by 1832 * the caller. Used for bogus page replacement. 1833 */ 1834 vm_page_t 1835 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1836 { 1837 vm_page_t m; 1838 1839 m = vm_page_lookup_unlocked(object, pindex); 1840 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1841 m->object == object && m->pindex == pindex, 1842 ("vm_page_relookup: Invalid page %p", m)); 1843 return (m); 1844 } 1845 1846 /* 1847 * This should only be used by lockless functions for releasing transient 1848 * incorrect acquires. The page may have been freed after we acquired a 1849 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1850 * further to do. 1851 */ 1852 static void 1853 vm_page_busy_release(vm_page_t m) 1854 { 1855 u_int x; 1856 1857 x = vm_page_busy_fetch(m); 1858 for (;;) { 1859 if (x == VPB_FREED) 1860 break; 1861 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1862 if (atomic_fcmpset_int(&m->busy_lock, &x, 1863 x - VPB_ONE_SHARER)) 1864 break; 1865 continue; 1866 } 1867 KASSERT((x & VPB_BIT_SHARED) != 0 || 1868 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1869 ("vm_page_busy_release: %p xbusy not owned.", m)); 1870 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1871 continue; 1872 if ((x & VPB_BIT_WAITERS) != 0) 1873 wakeup(m); 1874 break; 1875 } 1876 } 1877 1878 /* 1879 * vm_page_find_least: 1880 * 1881 * Returns the page associated with the object with least pindex 1882 * greater than or equal to the parameter pindex, or NULL. 1883 * 1884 * The object must be locked. 1885 */ 1886 vm_page_t 1887 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1888 { 1889 vm_page_t m; 1890 1891 VM_OBJECT_ASSERT_LOCKED(object); 1892 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1893 m = vm_radix_lookup_ge(&object->rtree, pindex); 1894 return (m); 1895 } 1896 1897 /* 1898 * Uses the page mnew as a replacement for an existing page at index 1899 * pindex which must be already present in the object. 1900 * 1901 * Both pages must be exclusively busied on enter. The old page is 1902 * unbusied on exit. 1903 * 1904 * A return value of true means mold is now free. If this is not the 1905 * final ref and the caller does not hold a wire reference it may not 1906 * continue to access the page. 1907 */ 1908 static bool 1909 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1910 vm_page_t mold) 1911 { 1912 vm_page_t mret __diagused; 1913 bool dropped; 1914 1915 VM_OBJECT_ASSERT_WLOCKED(object); 1916 vm_page_assert_xbusied(mold); 1917 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1918 ("vm_page_replace: page %p already in object", mnew)); 1919 1920 /* 1921 * This function mostly follows vm_page_insert() and 1922 * vm_page_remove() without the radix, object count and vnode 1923 * dance. Double check such functions for more comments. 1924 */ 1925 1926 mnew->object = object; 1927 mnew->pindex = pindex; 1928 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1929 mret = vm_radix_replace(&object->rtree, mnew); 1930 KASSERT(mret == mold, 1931 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1932 KASSERT((mold->oflags & VPO_UNMANAGED) == 1933 (mnew->oflags & VPO_UNMANAGED), 1934 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1935 1936 /* Keep the resident page list in sorted order. */ 1937 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1938 TAILQ_REMOVE(&object->memq, mold, listq); 1939 mold->object = NULL; 1940 1941 /* 1942 * The object's resident_page_count does not change because we have 1943 * swapped one page for another, but the generation count should 1944 * change if the page is dirty. 1945 */ 1946 if (pmap_page_is_write_mapped(mnew)) 1947 vm_object_set_writeable_dirty(object); 1948 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1949 vm_page_xunbusy(mold); 1950 1951 return (dropped); 1952 } 1953 1954 void 1955 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1956 vm_page_t mold) 1957 { 1958 1959 vm_page_assert_xbusied(mnew); 1960 1961 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1962 vm_page_free(mold); 1963 } 1964 1965 /* 1966 * vm_page_iter_rename: 1967 * 1968 * Tries to move the specified page from its current object to a new object 1969 * and pindex, using the given iterator to remove the page from its current 1970 * object. Returns true if the move was successful, and false if the move 1971 * was aborted due to a failed memory allocation. 1972 * 1973 * Panics if a page already resides in the new object at the new pindex. 1974 * 1975 * This routine dirties the page if it is valid, as callers are expected to 1976 * transfer backing storage only after moving the page. Dirtying the page 1977 * ensures that the destination object retains the most recent copy of the 1978 * page. 1979 * 1980 * The objects must be locked. 1981 */ 1982 bool 1983 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m, 1984 vm_object_t new_object, vm_pindex_t new_pindex) 1985 { 1986 vm_page_t mpred; 1987 vm_pindex_t opidx; 1988 1989 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1990 ("%s: page %p is missing object ref", __func__, m)); 1991 VM_OBJECT_ASSERT_WLOCKED(m->object); 1992 VM_OBJECT_ASSERT_WLOCKED(new_object); 1993 1994 /* 1995 * Create a custom version of vm_page_insert() which does not depend 1996 * by m_prev and can cheat on the implementation aspects of the 1997 * function. 1998 */ 1999 opidx = m->pindex; 2000 m->pindex = new_pindex; 2001 if (vm_radix_insert_lookup_lt(&new_object->rtree, m, &mpred) != 0) { 2002 m->pindex = opidx; 2003 return (false); 2004 } 2005 2006 /* 2007 * The operation cannot fail anymore. The removal must happen before 2008 * the listq iterator is tainted. 2009 */ 2010 m->pindex = opidx; 2011 vm_radix_iter_remove(old_pages); 2012 vm_page_remove_radixdone(m); 2013 2014 /* Return back to the new pindex to complete vm_page_insert(). */ 2015 m->pindex = new_pindex; 2016 m->object = new_object; 2017 2018 vm_page_insert_radixdone(m, new_object, mpred); 2019 if (vm_page_any_valid(m)) 2020 vm_page_dirty(m); 2021 vm_pager_page_inserted(new_object, m); 2022 return (true); 2023 } 2024 2025 /* 2026 * vm_page_mpred: 2027 * 2028 * Return the greatest page of the object with index <= pindex, 2029 * or NULL, if there is none. Assumes object lock is held. 2030 */ 2031 vm_page_t 2032 vm_page_mpred(vm_object_t object, vm_pindex_t pindex) 2033 { 2034 return (vm_radix_lookup_le(&object->rtree, pindex)); 2035 } 2036 2037 /* 2038 * vm_page_alloc: 2039 * 2040 * Allocate and return a page that is associated with the specified 2041 * object and offset pair. By default, this page is exclusive busied. 2042 * 2043 * The caller must always specify an allocation class. 2044 * 2045 * allocation classes: 2046 * VM_ALLOC_NORMAL normal process request 2047 * VM_ALLOC_SYSTEM system *really* needs a page 2048 * VM_ALLOC_INTERRUPT interrupt time request 2049 * 2050 * optional allocation flags: 2051 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2052 * intends to allocate 2053 * VM_ALLOC_NOBUSY do not exclusive busy the page 2054 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2055 * VM_ALLOC_SBUSY shared busy the allocated page 2056 * VM_ALLOC_WIRED wire the allocated page 2057 * VM_ALLOC_ZERO prefer a zeroed page 2058 */ 2059 vm_page_t 2060 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 2061 { 2062 struct pctrie_iter pages; 2063 2064 vm_page_iter_init(&pages, object); 2065 return (vm_page_alloc_after(object, &pages, pindex, req, 2066 vm_page_mpred(object, pindex))); 2067 } 2068 2069 /* 2070 * Allocate a page in the specified object with the given page index. To 2071 * optimize insertion of the page into the object, the caller must also specify 2072 * the resident page in the object with largest index smaller than the given 2073 * page index, or NULL if no such page exists. 2074 */ 2075 vm_page_t 2076 vm_page_alloc_after(vm_object_t object, struct pctrie_iter *pages, 2077 vm_pindex_t pindex, int req, vm_page_t mpred) 2078 { 2079 struct vm_domainset_iter di; 2080 vm_page_t m; 2081 int domain; 2082 2083 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2084 do { 2085 m = vm_page_alloc_domain_after(object, pages, pindex, domain, 2086 req, mpred); 2087 if (m != NULL) 2088 break; 2089 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2090 2091 return (m); 2092 } 2093 2094 /* 2095 * Returns true if the number of free pages exceeds the minimum 2096 * for the request class and false otherwise. 2097 */ 2098 static int 2099 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2100 { 2101 u_int limit, old, new; 2102 2103 if (req_class == VM_ALLOC_INTERRUPT) 2104 limit = 0; 2105 else if (req_class == VM_ALLOC_SYSTEM) 2106 limit = vmd->vmd_interrupt_free_min; 2107 else 2108 limit = vmd->vmd_free_reserved; 2109 2110 /* 2111 * Attempt to reserve the pages. Fail if we're below the limit. 2112 */ 2113 limit += npages; 2114 old = atomic_load_int(&vmd->vmd_free_count); 2115 do { 2116 if (old < limit) 2117 return (0); 2118 new = old - npages; 2119 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2120 2121 /* Wake the page daemon if we've crossed the threshold. */ 2122 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2123 pagedaemon_wakeup(vmd->vmd_domain); 2124 2125 /* Only update bitsets on transitions. */ 2126 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2127 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2128 vm_domain_set(vmd); 2129 2130 return (1); 2131 } 2132 2133 int 2134 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2135 { 2136 int req_class; 2137 2138 /* 2139 * The page daemon is allowed to dig deeper into the free page list. 2140 */ 2141 req_class = req & VM_ALLOC_CLASS_MASK; 2142 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2143 req_class = VM_ALLOC_SYSTEM; 2144 return (_vm_domain_allocate(vmd, req_class, npages)); 2145 } 2146 2147 vm_page_t 2148 vm_page_alloc_domain_after(vm_object_t object, struct pctrie_iter *pages, 2149 vm_pindex_t pindex, int domain, int req, vm_page_t mpred) 2150 { 2151 struct vm_domain *vmd; 2152 vm_page_t m; 2153 int flags; 2154 2155 #define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2156 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \ 2157 VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \ 2158 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | \ 2159 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK) 2160 KASSERT((req & ~VPA_FLAGS) == 0, 2161 ("invalid request %#x", req)); 2162 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2163 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2164 ("invalid request %#x", req)); 2165 KASSERT(mpred == NULL || mpred->pindex < pindex, 2166 ("mpred %p doesn't precede pindex 0x%jx", mpred, 2167 (uintmax_t)pindex)); 2168 VM_OBJECT_ASSERT_WLOCKED(object); 2169 2170 flags = 0; 2171 m = NULL; 2172 if (!vm_pager_can_alloc_page(object, pindex)) 2173 return (NULL); 2174 again: 2175 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2176 m = vm_page_alloc_nofree_domain(domain, req); 2177 if (m != NULL) 2178 goto found; 2179 } 2180 #if VM_NRESERVLEVEL > 0 2181 /* 2182 * Can we allocate the page from a reservation? 2183 */ 2184 if (vm_object_reserv(object) && 2185 (m = vm_reserv_alloc_page(object, pages, pindex, domain, req)) != 2186 NULL) { 2187 goto found; 2188 } 2189 #endif 2190 vmd = VM_DOMAIN(domain); 2191 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2192 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2193 M_NOWAIT | M_NOVM); 2194 if (m != NULL) { 2195 flags |= PG_PCPU_CACHE; 2196 goto found; 2197 } 2198 } 2199 if (vm_domain_allocate(vmd, req, 1)) { 2200 /* 2201 * If not, allocate it from the free page queues. 2202 */ 2203 vm_domain_free_lock(vmd); 2204 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2205 vm_domain_free_unlock(vmd); 2206 if (m == NULL) { 2207 vm_domain_freecnt_inc(vmd, 1); 2208 #if VM_NRESERVLEVEL > 0 2209 if (vm_reserv_reclaim_inactive(domain)) 2210 goto again; 2211 #endif 2212 } 2213 } 2214 if (m == NULL) { 2215 /* 2216 * Not allocatable, give up. 2217 */ 2218 pctrie_iter_reset(pages); 2219 if (vm_domain_alloc_fail(vmd, object, req)) 2220 goto again; 2221 return (NULL); 2222 } 2223 2224 /* 2225 * At this point we had better have found a good page. 2226 */ 2227 found: 2228 vm_page_dequeue(m); 2229 vm_page_alloc_check(m); 2230 2231 /* 2232 * Initialize the page. Only the PG_ZERO flag is inherited. 2233 */ 2234 flags |= m->flags & PG_ZERO; 2235 if ((req & VM_ALLOC_NODUMP) != 0) 2236 flags |= PG_NODUMP; 2237 if ((req & VM_ALLOC_NOFREE) != 0) 2238 flags |= PG_NOFREE; 2239 m->flags = flags; 2240 m->a.flags = 0; 2241 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2242 m->pool = VM_FREEPOOL_DEFAULT; 2243 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2244 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2245 else if ((req & VM_ALLOC_SBUSY) != 0) 2246 m->busy_lock = VPB_SHARERS_WORD(1); 2247 else 2248 m->busy_lock = VPB_UNBUSIED; 2249 if (req & VM_ALLOC_WIRED) { 2250 vm_wire_add(1); 2251 m->ref_count = 1; 2252 } 2253 m->a.act_count = 0; 2254 2255 if (vm_page_insert_lookup(m, object, pindex, pages, true, mpred)) { 2256 if (req & VM_ALLOC_WIRED) { 2257 vm_wire_sub(1); 2258 m->ref_count = 0; 2259 } 2260 KASSERT(m->object == NULL, ("page %p has object", m)); 2261 m->oflags = VPO_UNMANAGED; 2262 m->busy_lock = VPB_UNBUSIED; 2263 /* Don't change PG_ZERO. */ 2264 vm_page_free_toq(m); 2265 if (req & VM_ALLOC_WAITFAIL) { 2266 VM_OBJECT_WUNLOCK(object); 2267 vm_radix_wait(); 2268 pctrie_iter_reset(pages); 2269 VM_OBJECT_WLOCK(object); 2270 } 2271 return (NULL); 2272 } 2273 2274 /* Ignore device objects; the pager sets "memattr" for them. */ 2275 if (object->memattr != VM_MEMATTR_DEFAULT && 2276 (object->flags & OBJ_FICTITIOUS) == 0) 2277 pmap_page_set_memattr(m, object->memattr); 2278 2279 return (m); 2280 } 2281 2282 /* 2283 * vm_page_alloc_contig: 2284 * 2285 * Allocate a contiguous set of physical pages of the given size "npages" 2286 * from the free lists. All of the physical pages must be at or above 2287 * the given physical address "low" and below the given physical address 2288 * "high". The given value "alignment" determines the alignment of the 2289 * first physical page in the set. If the given value "boundary" is 2290 * non-zero, then the set of physical pages cannot cross any physical 2291 * address boundary that is a multiple of that value. Both "alignment" 2292 * and "boundary" must be a power of two. 2293 * 2294 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2295 * then the memory attribute setting for the physical pages is configured 2296 * to the object's memory attribute setting. Otherwise, the memory 2297 * attribute setting for the physical pages is configured to "memattr", 2298 * overriding the object's memory attribute setting. However, if the 2299 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2300 * memory attribute setting for the physical pages cannot be configured 2301 * to VM_MEMATTR_DEFAULT. 2302 * 2303 * The specified object may not contain fictitious pages. 2304 * 2305 * The caller must always specify an allocation class. 2306 * 2307 * allocation classes: 2308 * VM_ALLOC_NORMAL normal process request 2309 * VM_ALLOC_SYSTEM system *really* needs a page 2310 * VM_ALLOC_INTERRUPT interrupt time request 2311 * 2312 * optional allocation flags: 2313 * VM_ALLOC_NOBUSY do not exclusive busy the page 2314 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2315 * VM_ALLOC_SBUSY shared busy the allocated page 2316 * VM_ALLOC_WIRED wire the allocated page 2317 * VM_ALLOC_ZERO prefer a zeroed page 2318 */ 2319 vm_page_t 2320 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2321 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2322 vm_paddr_t boundary, vm_memattr_t memattr) 2323 { 2324 struct vm_domainset_iter di; 2325 vm_page_t bounds[2]; 2326 vm_page_t m; 2327 int domain; 2328 int start_segind; 2329 2330 start_segind = -1; 2331 2332 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 2333 do { 2334 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2335 npages, low, high, alignment, boundary, memattr); 2336 if (m != NULL) 2337 break; 2338 if (start_segind == -1) 2339 start_segind = vm_phys_lookup_segind(low); 2340 if (vm_phys_find_range(bounds, start_segind, domain, 2341 npages, low, high) == -1) { 2342 vm_domainset_iter_ignore(&di, domain); 2343 } 2344 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 2345 2346 return (m); 2347 } 2348 2349 static vm_page_t 2350 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2351 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2352 { 2353 struct vm_domain *vmd; 2354 vm_page_t m_ret; 2355 2356 /* 2357 * Can we allocate the pages without the number of free pages falling 2358 * below the lower bound for the allocation class? 2359 */ 2360 vmd = VM_DOMAIN(domain); 2361 if (!vm_domain_allocate(vmd, req, npages)) 2362 return (NULL); 2363 /* 2364 * Try to allocate the pages from the free page queues. 2365 */ 2366 vm_domain_free_lock(vmd); 2367 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2368 alignment, boundary); 2369 vm_domain_free_unlock(vmd); 2370 if (m_ret != NULL) 2371 return (m_ret); 2372 #if VM_NRESERVLEVEL > 0 2373 /* 2374 * Try to break a reservation to allocate the pages. 2375 */ 2376 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2377 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2378 high, alignment, boundary); 2379 if (m_ret != NULL) 2380 return (m_ret); 2381 } 2382 #endif 2383 vm_domain_freecnt_inc(vmd, npages); 2384 return (NULL); 2385 } 2386 2387 vm_page_t 2388 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2389 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2390 vm_paddr_t boundary, vm_memattr_t memattr) 2391 { 2392 struct pctrie_iter pages; 2393 vm_page_t m, m_ret, mpred; 2394 u_int busy_lock, flags, oflags; 2395 2396 #define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM) 2397 KASSERT((req & ~VPAC_FLAGS) == 0, 2398 ("invalid request %#x", req)); 2399 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2400 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2401 ("invalid request %#x", req)); 2402 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2403 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2404 ("invalid request %#x", req)); 2405 VM_OBJECT_ASSERT_WLOCKED(object); 2406 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2407 ("vm_page_alloc_contig: object %p has fictitious pages", 2408 object)); 2409 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2410 2411 vm_page_iter_init(&pages, object); 2412 mpred = vm_radix_iter_lookup_lt(&pages, pindex); 2413 KASSERT(mpred == NULL || mpred->pindex != pindex, 2414 ("vm_page_alloc_contig: pindex already allocated")); 2415 for (;;) { 2416 #if VM_NRESERVLEVEL > 0 2417 /* 2418 * Can we allocate the pages from a reservation? 2419 */ 2420 if (vm_object_reserv(object) && 2421 (m_ret = vm_reserv_alloc_contig(object, &pages, pindex, 2422 domain, req, npages, low, high, alignment, boundary)) != 2423 NULL) { 2424 break; 2425 } 2426 #endif 2427 if ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2428 low, high, alignment, boundary)) != NULL) 2429 break; 2430 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req)) 2431 return (NULL); 2432 } 2433 2434 /* 2435 * Initialize the pages. Only the PG_ZERO flag is inherited. 2436 */ 2437 flags = PG_ZERO; 2438 if ((req & VM_ALLOC_NODUMP) != 0) 2439 flags |= PG_NODUMP; 2440 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2441 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2442 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2443 else if ((req & VM_ALLOC_SBUSY) != 0) 2444 busy_lock = VPB_SHARERS_WORD(1); 2445 else 2446 busy_lock = VPB_UNBUSIED; 2447 if ((req & VM_ALLOC_WIRED) != 0) 2448 vm_wire_add(npages); 2449 if (object->memattr != VM_MEMATTR_DEFAULT && 2450 memattr == VM_MEMATTR_DEFAULT) 2451 memattr = object->memattr; 2452 for (m = m_ret; m < &m_ret[npages]; m++) { 2453 vm_page_dequeue(m); 2454 vm_page_alloc_check(m); 2455 m->a.flags = 0; 2456 m->flags = (m->flags | PG_NODUMP) & flags; 2457 m->busy_lock = busy_lock; 2458 if ((req & VM_ALLOC_WIRED) != 0) 2459 m->ref_count = 1; 2460 m->a.act_count = 0; 2461 m->oflags = oflags; 2462 m->pool = VM_FREEPOOL_DEFAULT; 2463 if (vm_page_iter_insert(&pages, m, object, pindex, mpred)) { 2464 if ((req & VM_ALLOC_WIRED) != 0) 2465 vm_wire_sub(npages); 2466 KASSERT(m->object == NULL, 2467 ("page %p has object", m)); 2468 mpred = m; 2469 for (m = m_ret; m < &m_ret[npages]; m++) { 2470 if (m <= mpred && 2471 (req & VM_ALLOC_WIRED) != 0) 2472 m->ref_count = 0; 2473 m->oflags = VPO_UNMANAGED; 2474 m->busy_lock = VPB_UNBUSIED; 2475 /* Don't change PG_ZERO. */ 2476 vm_page_free_toq(m); 2477 } 2478 if (req & VM_ALLOC_WAITFAIL) { 2479 VM_OBJECT_WUNLOCK(object); 2480 vm_radix_wait(); 2481 VM_OBJECT_WLOCK(object); 2482 } 2483 return (NULL); 2484 } 2485 mpred = m; 2486 if (memattr != VM_MEMATTR_DEFAULT) 2487 pmap_page_set_memattr(m, memattr); 2488 pindex++; 2489 } 2490 return (m_ret); 2491 } 2492 2493 /* 2494 * Allocate a physical page that is not intended to be inserted into a VM 2495 * object. 2496 */ 2497 vm_page_t 2498 vm_page_alloc_noobj_domain(int domain, int req) 2499 { 2500 struct vm_domain *vmd; 2501 vm_page_t m; 2502 int flags; 2503 2504 #define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \ 2505 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \ 2506 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \ 2507 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | \ 2508 VM_ALLOC_NOFREE | VM_ALLOC_COUNT_MASK) 2509 KASSERT((req & ~VPAN_FLAGS) == 0, 2510 ("invalid request %#x", req)); 2511 2512 flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) | 2513 ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0); 2514 vmd = VM_DOMAIN(domain); 2515 again: 2516 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2517 m = vm_page_alloc_nofree_domain(domain, req); 2518 if (m != NULL) 2519 goto found; 2520 } 2521 2522 if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2523 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2524 M_NOWAIT | M_NOVM); 2525 if (m != NULL) { 2526 flags |= PG_PCPU_CACHE; 2527 goto found; 2528 } 2529 } 2530 2531 if (vm_domain_allocate(vmd, req, 1)) { 2532 vm_domain_free_lock(vmd); 2533 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2534 vm_domain_free_unlock(vmd); 2535 if (m == NULL) { 2536 vm_domain_freecnt_inc(vmd, 1); 2537 #if VM_NRESERVLEVEL > 0 2538 if (vm_reserv_reclaim_inactive(domain)) 2539 goto again; 2540 #endif 2541 } 2542 } 2543 if (m == NULL) { 2544 if (vm_domain_alloc_fail(vmd, NULL, req)) 2545 goto again; 2546 return (NULL); 2547 } 2548 2549 found: 2550 vm_page_dequeue(m); 2551 vm_page_alloc_check(m); 2552 2553 /* 2554 * Consumers should not rely on a useful default pindex value. 2555 */ 2556 m->pindex = 0xdeadc0dedeadc0de; 2557 m->flags = (m->flags & PG_ZERO) | flags; 2558 m->a.flags = 0; 2559 m->oflags = VPO_UNMANAGED; 2560 m->pool = VM_FREEPOOL_DIRECT; 2561 m->busy_lock = VPB_UNBUSIED; 2562 if ((req & VM_ALLOC_WIRED) != 0) { 2563 vm_wire_add(1); 2564 m->ref_count = 1; 2565 } 2566 2567 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2568 pmap_zero_page(m); 2569 2570 return (m); 2571 } 2572 2573 #if VM_NRESERVLEVEL > 1 2574 #define VM_NOFREE_IMPORT_ORDER (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER) 2575 #elif VM_NRESERVLEVEL > 0 2576 #define VM_NOFREE_IMPORT_ORDER VM_LEVEL_0_ORDER 2577 #else 2578 #define VM_NOFREE_IMPORT_ORDER 8 2579 #endif 2580 2581 /* 2582 * Allocate a single NOFREE page. 2583 * 2584 * This routine hands out NOFREE pages from higher-order 2585 * physical memory blocks in order to reduce memory fragmentation. 2586 * When a NOFREE for a given domain chunk is used up, 2587 * the routine will try to fetch a new one from the freelists 2588 * and discard the old one. 2589 */ 2590 static vm_page_t __noinline 2591 vm_page_alloc_nofree_domain(int domain, int req) 2592 { 2593 vm_page_t m; 2594 struct vm_domain *vmd; 2595 2596 KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req)); 2597 2598 vmd = VM_DOMAIN(domain); 2599 vm_domain_free_lock(vmd); 2600 if (TAILQ_EMPTY(&vmd->vmd_nofreeq)) { 2601 int count; 2602 2603 count = 1 << VM_NOFREE_IMPORT_ORDER; 2604 if (!vm_domain_allocate(vmd, req, count)) { 2605 vm_domain_free_unlock(vmd); 2606 return (NULL); 2607 } 2608 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 2609 VM_NOFREE_IMPORT_ORDER); 2610 if (m == NULL) { 2611 vm_domain_freecnt_inc(vmd, count); 2612 vm_domain_free_unlock(vmd); 2613 return (NULL); 2614 } 2615 m->pindex = count; 2616 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, listq); 2617 VM_CNT_ADD(v_nofree_count, count); 2618 } 2619 m = TAILQ_FIRST(&vmd->vmd_nofreeq); 2620 TAILQ_REMOVE(&vmd->vmd_nofreeq, m, listq); 2621 if (m->pindex > 1) { 2622 vm_page_t m_next; 2623 2624 m_next = &m[1]; 2625 m_next->pindex = m->pindex - 1; 2626 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m_next, listq); 2627 } 2628 vm_domain_free_unlock(vmd); 2629 VM_CNT_ADD(v_nofree_count, -1); 2630 2631 return (m); 2632 } 2633 2634 /* 2635 * Though a NOFREE page by definition should not be freed, we support putting 2636 * them aside for future NOFREE allocations. This enables code which allocates 2637 * NOFREE pages for some purpose but then encounters an error and releases 2638 * resources. 2639 */ 2640 static void __noinline 2641 vm_page_free_nofree(struct vm_domain *vmd, vm_page_t m) 2642 { 2643 vm_domain_free_lock(vmd); 2644 m->pindex = 1; 2645 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, listq); 2646 vm_domain_free_unlock(vmd); 2647 VM_CNT_ADD(v_nofree_count, 1); 2648 } 2649 2650 vm_page_t 2651 vm_page_alloc_noobj(int req) 2652 { 2653 struct vm_domainset_iter di; 2654 vm_page_t m; 2655 int domain; 2656 2657 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2658 do { 2659 m = vm_page_alloc_noobj_domain(domain, req); 2660 if (m != NULL) 2661 break; 2662 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2663 2664 return (m); 2665 } 2666 2667 vm_page_t 2668 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2669 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2670 vm_memattr_t memattr) 2671 { 2672 struct vm_domainset_iter di; 2673 vm_page_t m; 2674 int domain; 2675 2676 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2677 do { 2678 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2679 high, alignment, boundary, memattr); 2680 if (m != NULL) 2681 break; 2682 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2683 2684 return (m); 2685 } 2686 2687 vm_page_t 2688 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2689 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2690 vm_memattr_t memattr) 2691 { 2692 vm_page_t m, m_ret; 2693 u_int flags; 2694 2695 #define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM) 2696 KASSERT((req & ~VPANC_FLAGS) == 0, 2697 ("invalid request %#x", req)); 2698 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2699 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2700 ("invalid request %#x", req)); 2701 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2702 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2703 ("invalid request %#x", req)); 2704 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2705 2706 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2707 low, high, alignment, boundary)) == NULL) { 2708 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2709 return (NULL); 2710 } 2711 2712 /* 2713 * Initialize the pages. Only the PG_ZERO flag is inherited. 2714 */ 2715 flags = PG_ZERO; 2716 if ((req & VM_ALLOC_NODUMP) != 0) 2717 flags |= PG_NODUMP; 2718 if ((req & VM_ALLOC_WIRED) != 0) 2719 vm_wire_add(npages); 2720 for (m = m_ret; m < &m_ret[npages]; m++) { 2721 vm_page_dequeue(m); 2722 vm_page_alloc_check(m); 2723 2724 /* 2725 * Consumers should not rely on a useful default pindex value. 2726 */ 2727 m->pindex = 0xdeadc0dedeadc0de; 2728 m->a.flags = 0; 2729 m->flags = (m->flags | PG_NODUMP) & flags; 2730 m->busy_lock = VPB_UNBUSIED; 2731 if ((req & VM_ALLOC_WIRED) != 0) 2732 m->ref_count = 1; 2733 m->a.act_count = 0; 2734 m->oflags = VPO_UNMANAGED; 2735 m->pool = VM_FREEPOOL_DIRECT; 2736 2737 /* 2738 * Zero the page before updating any mappings since the page is 2739 * not yet shared with any devices which might require the 2740 * non-default memory attribute. pmap_page_set_memattr() 2741 * flushes data caches before returning. 2742 */ 2743 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2744 pmap_zero_page(m); 2745 if (memattr != VM_MEMATTR_DEFAULT) 2746 pmap_page_set_memattr(m, memattr); 2747 } 2748 return (m_ret); 2749 } 2750 2751 /* 2752 * Check a page that has been freshly dequeued from a freelist. 2753 */ 2754 static void 2755 vm_page_alloc_check(vm_page_t m) 2756 { 2757 2758 KASSERT(m->object == NULL, ("page %p has object", m)); 2759 KASSERT(m->a.queue == PQ_NONE && 2760 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2761 ("page %p has unexpected queue %d, flags %#x", 2762 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2763 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2764 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2765 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2766 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2767 ("page %p has unexpected memattr %d", 2768 m, pmap_page_get_memattr(m))); 2769 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2770 pmap_vm_page_alloc_check(m); 2771 } 2772 2773 static int 2774 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2775 { 2776 struct vm_domain *vmd; 2777 struct vm_pgcache *pgcache; 2778 int i; 2779 2780 pgcache = arg; 2781 vmd = VM_DOMAIN(pgcache->domain); 2782 2783 /* 2784 * The page daemon should avoid creating extra memory pressure since its 2785 * main purpose is to replenish the store of free pages. 2786 */ 2787 if (vmd->vmd_severeset || curproc == pageproc || 2788 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2789 return (0); 2790 domain = vmd->vmd_domain; 2791 vm_domain_free_lock(vmd); 2792 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2793 (vm_page_t *)store); 2794 vm_domain_free_unlock(vmd); 2795 if (cnt != i) 2796 vm_domain_freecnt_inc(vmd, cnt - i); 2797 2798 return (i); 2799 } 2800 2801 static void 2802 vm_page_zone_release(void *arg, void **store, int cnt) 2803 { 2804 struct vm_domain *vmd; 2805 struct vm_pgcache *pgcache; 2806 vm_page_t m; 2807 int i; 2808 2809 pgcache = arg; 2810 vmd = VM_DOMAIN(pgcache->domain); 2811 vm_domain_free_lock(vmd); 2812 for (i = 0; i < cnt; i++) { 2813 m = (vm_page_t)store[i]; 2814 vm_phys_free_pages(m, pgcache->pool, 0); 2815 } 2816 vm_domain_free_unlock(vmd); 2817 vm_domain_freecnt_inc(vmd, cnt); 2818 } 2819 2820 #define VPSC_ANY 0 /* No restrictions. */ 2821 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2822 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2823 2824 /* 2825 * vm_page_scan_contig: 2826 * 2827 * Scan vm_page_array[] between the specified entries "m_start" and 2828 * "m_end" for a run of contiguous physical pages that satisfy the 2829 * specified conditions, and return the lowest page in the run. The 2830 * specified "alignment" determines the alignment of the lowest physical 2831 * page in the run. If the specified "boundary" is non-zero, then the 2832 * run of physical pages cannot span a physical address that is a 2833 * multiple of "boundary". 2834 * 2835 * "m_end" is never dereferenced, so it need not point to a vm_page 2836 * structure within vm_page_array[]. 2837 * 2838 * "npages" must be greater than zero. "m_start" and "m_end" must not 2839 * span a hole (or discontiguity) in the physical address space. Both 2840 * "alignment" and "boundary" must be a power of two. 2841 */ 2842 static vm_page_t 2843 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2844 u_long alignment, vm_paddr_t boundary, int options) 2845 { 2846 vm_object_t object; 2847 vm_paddr_t pa; 2848 vm_page_t m, m_run; 2849 #if VM_NRESERVLEVEL > 0 2850 int level; 2851 #endif 2852 int m_inc, order, run_ext, run_len; 2853 2854 KASSERT(npages > 0, ("npages is 0")); 2855 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2856 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2857 m_run = NULL; 2858 run_len = 0; 2859 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2860 KASSERT((m->flags & PG_MARKER) == 0, 2861 ("page %p is PG_MARKER", m)); 2862 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2863 ("fictitious page %p has invalid ref count", m)); 2864 2865 /* 2866 * If the current page would be the start of a run, check its 2867 * physical address against the end, alignment, and boundary 2868 * conditions. If it doesn't satisfy these conditions, either 2869 * terminate the scan or advance to the next page that 2870 * satisfies the failed condition. 2871 */ 2872 if (run_len == 0) { 2873 KASSERT(m_run == NULL, ("m_run != NULL")); 2874 if (m + npages > m_end) 2875 break; 2876 pa = VM_PAGE_TO_PHYS(m); 2877 if (!vm_addr_align_ok(pa, alignment)) { 2878 m_inc = atop(roundup2(pa, alignment) - pa); 2879 continue; 2880 } 2881 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2882 m_inc = atop(roundup2(pa, boundary) - pa); 2883 continue; 2884 } 2885 } else 2886 KASSERT(m_run != NULL, ("m_run == NULL")); 2887 2888 retry: 2889 m_inc = 1; 2890 if (vm_page_wired(m)) 2891 run_ext = 0; 2892 #if VM_NRESERVLEVEL > 0 2893 else if ((level = vm_reserv_level(m)) >= 0 && 2894 (options & VPSC_NORESERV) != 0) { 2895 run_ext = 0; 2896 /* Advance to the end of the reservation. */ 2897 pa = VM_PAGE_TO_PHYS(m); 2898 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2899 pa); 2900 } 2901 #endif 2902 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2903 /* 2904 * The page is considered eligible for relocation if 2905 * and only if it could be laundered or reclaimed by 2906 * the page daemon. 2907 */ 2908 VM_OBJECT_RLOCK(object); 2909 if (object != m->object) { 2910 VM_OBJECT_RUNLOCK(object); 2911 goto retry; 2912 } 2913 /* Don't care: PG_NODUMP, PG_ZERO. */ 2914 if ((object->flags & OBJ_SWAP) == 0 && 2915 object->type != OBJT_VNODE) { 2916 run_ext = 0; 2917 #if VM_NRESERVLEVEL > 0 2918 } else if ((options & VPSC_NOSUPER) != 0 && 2919 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2920 run_ext = 0; 2921 /* Advance to the end of the superpage. */ 2922 pa = VM_PAGE_TO_PHYS(m); 2923 m_inc = atop(roundup2(pa + 1, 2924 vm_reserv_size(level)) - pa); 2925 #endif 2926 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2927 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2928 /* 2929 * The page is allocated but eligible for 2930 * relocation. Extend the current run by one 2931 * page. 2932 */ 2933 KASSERT(pmap_page_get_memattr(m) == 2934 VM_MEMATTR_DEFAULT, 2935 ("page %p has an unexpected memattr", m)); 2936 KASSERT((m->oflags & (VPO_SWAPINPROG | 2937 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2938 ("page %p has unexpected oflags", m)); 2939 /* Don't care: PGA_NOSYNC. */ 2940 run_ext = 1; 2941 } else 2942 run_ext = 0; 2943 VM_OBJECT_RUNLOCK(object); 2944 #if VM_NRESERVLEVEL > 0 2945 } else if (level >= 0) { 2946 /* 2947 * The page is reserved but not yet allocated. In 2948 * other words, it is still free. Extend the current 2949 * run by one page. 2950 */ 2951 run_ext = 1; 2952 #endif 2953 } else if ((order = m->order) < VM_NFREEORDER) { 2954 /* 2955 * The page is enqueued in the physical memory 2956 * allocator's free page queues. Moreover, it is the 2957 * first page in a power-of-two-sized run of 2958 * contiguous free pages. Add these pages to the end 2959 * of the current run, and jump ahead. 2960 */ 2961 run_ext = 1 << order; 2962 m_inc = 1 << order; 2963 } else { 2964 /* 2965 * Skip the page for one of the following reasons: (1) 2966 * It is enqueued in the physical memory allocator's 2967 * free page queues. However, it is not the first 2968 * page in a run of contiguous free pages. (This case 2969 * rarely occurs because the scan is performed in 2970 * ascending order.) (2) It is not reserved, and it is 2971 * transitioning from free to allocated. (Conversely, 2972 * the transition from allocated to free for managed 2973 * pages is blocked by the page busy lock.) (3) It is 2974 * allocated but not contained by an object and not 2975 * wired, e.g., allocated by Xen's balloon driver. 2976 */ 2977 run_ext = 0; 2978 } 2979 2980 /* 2981 * Extend or reset the current run of pages. 2982 */ 2983 if (run_ext > 0) { 2984 if (run_len == 0) 2985 m_run = m; 2986 run_len += run_ext; 2987 } else { 2988 if (run_len > 0) { 2989 m_run = NULL; 2990 run_len = 0; 2991 } 2992 } 2993 } 2994 if (run_len >= npages) 2995 return (m_run); 2996 return (NULL); 2997 } 2998 2999 /* 3000 * vm_page_reclaim_run: 3001 * 3002 * Try to relocate each of the allocated virtual pages within the 3003 * specified run of physical pages to a new physical address. Free the 3004 * physical pages underlying the relocated virtual pages. A virtual page 3005 * is relocatable if and only if it could be laundered or reclaimed by 3006 * the page daemon. Whenever possible, a virtual page is relocated to a 3007 * physical address above "high". 3008 * 3009 * Returns 0 if every physical page within the run was already free or 3010 * just freed by a successful relocation. Otherwise, returns a non-zero 3011 * value indicating why the last attempt to relocate a virtual page was 3012 * unsuccessful. 3013 * 3014 * "req_class" must be an allocation class. 3015 */ 3016 static int 3017 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 3018 vm_paddr_t high) 3019 { 3020 struct vm_domain *vmd; 3021 struct spglist free; 3022 vm_object_t object; 3023 vm_paddr_t pa; 3024 vm_page_t m, m_end, m_new; 3025 int error, order, req; 3026 3027 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 3028 ("req_class is not an allocation class")); 3029 SLIST_INIT(&free); 3030 error = 0; 3031 m = m_run; 3032 m_end = m_run + npages; 3033 for (; error == 0 && m < m_end; m++) { 3034 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 3035 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 3036 3037 /* 3038 * Racily check for wirings. Races are handled once the object 3039 * lock is held and the page is unmapped. 3040 */ 3041 if (vm_page_wired(m)) 3042 error = EBUSY; 3043 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 3044 /* 3045 * The page is relocated if and only if it could be 3046 * laundered or reclaimed by the page daemon. 3047 */ 3048 VM_OBJECT_WLOCK(object); 3049 /* Don't care: PG_NODUMP, PG_ZERO. */ 3050 if (m->object != object || 3051 ((object->flags & OBJ_SWAP) == 0 && 3052 object->type != OBJT_VNODE)) 3053 error = EINVAL; 3054 else if (object->memattr != VM_MEMATTR_DEFAULT) 3055 error = EINVAL; 3056 else if (vm_page_queue(m) != PQ_NONE && 3057 vm_page_tryxbusy(m) != 0) { 3058 if (vm_page_wired(m)) { 3059 vm_page_xunbusy(m); 3060 error = EBUSY; 3061 goto unlock; 3062 } 3063 KASSERT(pmap_page_get_memattr(m) == 3064 VM_MEMATTR_DEFAULT, 3065 ("page %p has an unexpected memattr", m)); 3066 KASSERT(m->oflags == 0, 3067 ("page %p has unexpected oflags", m)); 3068 /* Don't care: PGA_NOSYNC. */ 3069 if (!vm_page_none_valid(m)) { 3070 /* 3071 * First, try to allocate a new page 3072 * that is above "high". Failing 3073 * that, try to allocate a new page 3074 * that is below "m_run". Allocate 3075 * the new page between the end of 3076 * "m_run" and "high" only as a last 3077 * resort. 3078 */ 3079 req = req_class; 3080 if ((m->flags & PG_NODUMP) != 0) 3081 req |= VM_ALLOC_NODUMP; 3082 if (trunc_page(high) != 3083 ~(vm_paddr_t)PAGE_MASK) { 3084 m_new = 3085 vm_page_alloc_noobj_contig( 3086 req, 1, round_page(high), 3087 ~(vm_paddr_t)0, PAGE_SIZE, 3088 0, VM_MEMATTR_DEFAULT); 3089 } else 3090 m_new = NULL; 3091 if (m_new == NULL) { 3092 pa = VM_PAGE_TO_PHYS(m_run); 3093 m_new = 3094 vm_page_alloc_noobj_contig( 3095 req, 1, 0, pa - 1, 3096 PAGE_SIZE, 0, 3097 VM_MEMATTR_DEFAULT); 3098 } 3099 if (m_new == NULL) { 3100 pa += ptoa(npages); 3101 m_new = 3102 vm_page_alloc_noobj_contig( 3103 req, 1, pa, high, PAGE_SIZE, 3104 0, VM_MEMATTR_DEFAULT); 3105 } 3106 if (m_new == NULL) { 3107 vm_page_xunbusy(m); 3108 error = ENOMEM; 3109 goto unlock; 3110 } 3111 3112 /* 3113 * Unmap the page and check for new 3114 * wirings that may have been acquired 3115 * through a pmap lookup. 3116 */ 3117 if (object->ref_count != 0 && 3118 !vm_page_try_remove_all(m)) { 3119 vm_page_xunbusy(m); 3120 vm_page_free(m_new); 3121 error = EBUSY; 3122 goto unlock; 3123 } 3124 3125 /* 3126 * Replace "m" with the new page. For 3127 * vm_page_replace(), "m" must be busy 3128 * and dequeued. Finally, change "m" 3129 * as if vm_page_free() was called. 3130 */ 3131 m_new->a.flags = m->a.flags & 3132 ~PGA_QUEUE_STATE_MASK; 3133 KASSERT(m_new->oflags == VPO_UNMANAGED, 3134 ("page %p is managed", m_new)); 3135 m_new->oflags = 0; 3136 pmap_copy_page(m, m_new); 3137 m_new->valid = m->valid; 3138 m_new->dirty = m->dirty; 3139 m->flags &= ~PG_ZERO; 3140 vm_page_dequeue(m); 3141 if (vm_page_replace_hold(m_new, object, 3142 m->pindex, m) && 3143 vm_page_free_prep(m)) 3144 SLIST_INSERT_HEAD(&free, m, 3145 plinks.s.ss); 3146 3147 /* 3148 * The new page must be deactivated 3149 * before the object is unlocked. 3150 */ 3151 vm_page_deactivate(m_new); 3152 } else { 3153 m->flags &= ~PG_ZERO; 3154 vm_page_dequeue(m); 3155 if (vm_page_free_prep(m)) 3156 SLIST_INSERT_HEAD(&free, m, 3157 plinks.s.ss); 3158 KASSERT(m->dirty == 0, 3159 ("page %p is dirty", m)); 3160 } 3161 } else 3162 error = EBUSY; 3163 unlock: 3164 VM_OBJECT_WUNLOCK(object); 3165 } else { 3166 MPASS(vm_page_domain(m) == domain); 3167 vmd = VM_DOMAIN(domain); 3168 vm_domain_free_lock(vmd); 3169 order = m->order; 3170 if (order < VM_NFREEORDER) { 3171 /* 3172 * The page is enqueued in the physical memory 3173 * allocator's free page queues. Moreover, it 3174 * is the first page in a power-of-two-sized 3175 * run of contiguous free pages. Jump ahead 3176 * to the last page within that run, and 3177 * continue from there. 3178 */ 3179 m += (1 << order) - 1; 3180 } 3181 #if VM_NRESERVLEVEL > 0 3182 else if (vm_reserv_is_page_free(m)) 3183 order = 0; 3184 #endif 3185 vm_domain_free_unlock(vmd); 3186 if (order == VM_NFREEORDER) 3187 error = EINVAL; 3188 } 3189 } 3190 if ((m = SLIST_FIRST(&free)) != NULL) { 3191 int cnt; 3192 3193 vmd = VM_DOMAIN(domain); 3194 cnt = 0; 3195 vm_domain_free_lock(vmd); 3196 do { 3197 MPASS(vm_page_domain(m) == domain); 3198 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3199 vm_phys_free_pages(m, m->pool, 0); 3200 cnt++; 3201 } while ((m = SLIST_FIRST(&free)) != NULL); 3202 vm_domain_free_unlock(vmd); 3203 vm_domain_freecnt_inc(vmd, cnt); 3204 } 3205 return (error); 3206 } 3207 3208 #define NRUNS 16 3209 3210 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 3211 3212 #define MIN_RECLAIM 8 3213 3214 /* 3215 * vm_page_reclaim_contig: 3216 * 3217 * Reclaim allocated, contiguous physical memory satisfying the specified 3218 * conditions by relocating the virtual pages using that physical memory. 3219 * Returns 0 if reclamation is successful, ERANGE if the specified domain 3220 * can't possibly satisfy the reclamation request, or ENOMEM if not 3221 * currently able to reclaim the requested number of pages. Since 3222 * relocation requires the allocation of physical pages, reclamation may 3223 * fail with ENOMEM due to a shortage of free pages. When reclamation 3224 * fails in this manner, callers are expected to perform vm_wait() before 3225 * retrying a failed allocation operation, e.g., vm_page_alloc_contig(). 3226 * 3227 * The caller must always specify an allocation class through "req". 3228 * 3229 * allocation classes: 3230 * VM_ALLOC_NORMAL normal process request 3231 * VM_ALLOC_SYSTEM system *really* needs a page 3232 * VM_ALLOC_INTERRUPT interrupt time request 3233 * 3234 * The optional allocation flags are ignored. 3235 * 3236 * "npages" must be greater than zero. Both "alignment" and "boundary" 3237 * must be a power of two. 3238 */ 3239 int 3240 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3241 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3242 int desired_runs) 3243 { 3244 struct vm_domain *vmd; 3245 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; 3246 u_long count, minalign, reclaimed; 3247 int error, i, min_reclaim, nruns, options, req_class; 3248 int segind, start_segind; 3249 int ret; 3250 3251 KASSERT(npages > 0, ("npages is 0")); 3252 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3253 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3254 3255 ret = ENOMEM; 3256 3257 /* 3258 * If the caller wants to reclaim multiple runs, try to allocate 3259 * space to store the runs. If that fails, fall back to the old 3260 * behavior of just reclaiming MIN_RECLAIM pages. 3261 */ 3262 if (desired_runs > 1) 3263 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3264 M_TEMP, M_NOWAIT); 3265 else 3266 m_runs = NULL; 3267 3268 if (m_runs == NULL) { 3269 m_runs = _m_runs; 3270 nruns = NRUNS; 3271 } else { 3272 nruns = NRUNS + desired_runs - 1; 3273 } 3274 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3275 3276 /* 3277 * The caller will attempt an allocation after some runs have been 3278 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3279 * of vm_phys_alloc_contig(), round up the requested length to the next 3280 * power of two or maximum chunk size, and ensure that each run is 3281 * suitably aligned. 3282 */ 3283 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3284 npages = roundup2(npages, minalign); 3285 if (alignment < ptoa(minalign)) 3286 alignment = ptoa(minalign); 3287 3288 /* 3289 * The page daemon is allowed to dig deeper into the free page list. 3290 */ 3291 req_class = req & VM_ALLOC_CLASS_MASK; 3292 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3293 req_class = VM_ALLOC_SYSTEM; 3294 3295 start_segind = vm_phys_lookup_segind(low); 3296 3297 /* 3298 * Return if the number of free pages cannot satisfy the requested 3299 * allocation. 3300 */ 3301 vmd = VM_DOMAIN(domain); 3302 count = vmd->vmd_free_count; 3303 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3304 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3305 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3306 goto done; 3307 3308 /* 3309 * Scan up to three times, relaxing the restrictions ("options") on 3310 * the reclamation of reservations and superpages each time. 3311 */ 3312 for (options = VPSC_NORESERV;;) { 3313 bool phys_range_exists = false; 3314 3315 /* 3316 * Find the highest runs that satisfy the given constraints 3317 * and restrictions, and record them in "m_runs". 3318 */ 3319 count = 0; 3320 segind = start_segind; 3321 while ((segind = vm_phys_find_range(bounds, segind, domain, 3322 npages, low, high)) != -1) { 3323 phys_range_exists = true; 3324 while ((m_run = vm_page_scan_contig(npages, bounds[0], 3325 bounds[1], alignment, boundary, options))) { 3326 bounds[0] = m_run + npages; 3327 m_runs[RUN_INDEX(count, nruns)] = m_run; 3328 count++; 3329 } 3330 segind++; 3331 } 3332 3333 if (!phys_range_exists) { 3334 ret = ERANGE; 3335 goto done; 3336 } 3337 3338 /* 3339 * Reclaim the highest runs in LIFO (descending) order until 3340 * the number of reclaimed pages, "reclaimed", is at least 3341 * "min_reclaim". Reset "reclaimed" each time because each 3342 * reclamation is idempotent, and runs will (likely) recur 3343 * from one scan to the next as restrictions are relaxed. 3344 */ 3345 reclaimed = 0; 3346 for (i = 0; count > 0 && i < nruns; i++) { 3347 count--; 3348 m_run = m_runs[RUN_INDEX(count, nruns)]; 3349 error = vm_page_reclaim_run(req_class, domain, npages, 3350 m_run, high); 3351 if (error == 0) { 3352 reclaimed += npages; 3353 if (reclaimed >= min_reclaim) { 3354 ret = 0; 3355 goto done; 3356 } 3357 } 3358 } 3359 3360 /* 3361 * Either relax the restrictions on the next scan or return if 3362 * the last scan had no restrictions. 3363 */ 3364 if (options == VPSC_NORESERV) 3365 options = VPSC_NOSUPER; 3366 else if (options == VPSC_NOSUPER) 3367 options = VPSC_ANY; 3368 else if (options == VPSC_ANY) { 3369 if (reclaimed != 0) 3370 ret = 0; 3371 goto done; 3372 } 3373 } 3374 done: 3375 if (m_runs != _m_runs) 3376 free(m_runs, M_TEMP); 3377 return (ret); 3378 } 3379 3380 int 3381 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3382 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3383 { 3384 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, 3385 high, alignment, boundary, 1)); 3386 } 3387 3388 int 3389 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3390 u_long alignment, vm_paddr_t boundary) 3391 { 3392 struct vm_domainset_iter di; 3393 int domain, ret, status; 3394 3395 ret = ERANGE; 3396 3397 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 3398 do { 3399 status = vm_page_reclaim_contig_domain(domain, req, npages, low, 3400 high, alignment, boundary); 3401 if (status == 0) 3402 return (0); 3403 else if (status == ERANGE) 3404 vm_domainset_iter_ignore(&di, domain); 3405 else { 3406 KASSERT(status == ENOMEM, ("Unrecognized error %d " 3407 "from vm_page_reclaim_contig_domain()", status)); 3408 ret = ENOMEM; 3409 } 3410 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 3411 3412 return (ret); 3413 } 3414 3415 /* 3416 * Set the domain in the appropriate page level domainset. 3417 */ 3418 void 3419 vm_domain_set(struct vm_domain *vmd) 3420 { 3421 3422 mtx_lock(&vm_domainset_lock); 3423 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3424 vmd->vmd_minset = 1; 3425 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3426 } 3427 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3428 vmd->vmd_severeset = 1; 3429 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3430 } 3431 mtx_unlock(&vm_domainset_lock); 3432 } 3433 3434 /* 3435 * Clear the domain from the appropriate page level domainset. 3436 */ 3437 void 3438 vm_domain_clear(struct vm_domain *vmd) 3439 { 3440 3441 mtx_lock(&vm_domainset_lock); 3442 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3443 vmd->vmd_minset = 0; 3444 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3445 if (vm_min_waiters != 0) { 3446 vm_min_waiters = 0; 3447 wakeup(&vm_min_domains); 3448 } 3449 } 3450 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3451 vmd->vmd_severeset = 0; 3452 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3453 if (vm_severe_waiters != 0) { 3454 vm_severe_waiters = 0; 3455 wakeup(&vm_severe_domains); 3456 } 3457 } 3458 3459 /* 3460 * If pageout daemon needs pages, then tell it that there are 3461 * some free. 3462 */ 3463 if (vmd->vmd_pageout_pages_needed && 3464 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3465 wakeup(&vmd->vmd_pageout_pages_needed); 3466 vmd->vmd_pageout_pages_needed = 0; 3467 } 3468 3469 /* See comments in vm_wait_doms(). */ 3470 if (vm_pageproc_waiters) { 3471 vm_pageproc_waiters = 0; 3472 wakeup(&vm_pageproc_waiters); 3473 } 3474 mtx_unlock(&vm_domainset_lock); 3475 } 3476 3477 /* 3478 * Wait for free pages to exceed the min threshold globally. 3479 */ 3480 void 3481 vm_wait_min(void) 3482 { 3483 3484 mtx_lock(&vm_domainset_lock); 3485 while (vm_page_count_min()) { 3486 vm_min_waiters++; 3487 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3488 } 3489 mtx_unlock(&vm_domainset_lock); 3490 } 3491 3492 /* 3493 * Wait for free pages to exceed the severe threshold globally. 3494 */ 3495 void 3496 vm_wait_severe(void) 3497 { 3498 3499 mtx_lock(&vm_domainset_lock); 3500 while (vm_page_count_severe()) { 3501 vm_severe_waiters++; 3502 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3503 "vmwait", 0); 3504 } 3505 mtx_unlock(&vm_domainset_lock); 3506 } 3507 3508 u_int 3509 vm_wait_count(void) 3510 { 3511 3512 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3513 } 3514 3515 int 3516 vm_wait_doms(const domainset_t *wdoms, int mflags) 3517 { 3518 int error; 3519 3520 error = 0; 3521 3522 /* 3523 * We use racey wakeup synchronization to avoid expensive global 3524 * locking for the pageproc when sleeping with a non-specific vm_wait. 3525 * To handle this, we only sleep for one tick in this instance. It 3526 * is expected that most allocations for the pageproc will come from 3527 * kmem or vm_page_grab* which will use the more specific and 3528 * race-free vm_wait_domain(). 3529 */ 3530 if (curproc == pageproc) { 3531 mtx_lock(&vm_domainset_lock); 3532 vm_pageproc_waiters++; 3533 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3534 PVM | PDROP | mflags, "pageprocwait", 1); 3535 } else { 3536 /* 3537 * XXX Ideally we would wait only until the allocation could 3538 * be satisfied. This condition can cause new allocators to 3539 * consume all freed pages while old allocators wait. 3540 */ 3541 mtx_lock(&vm_domainset_lock); 3542 if (vm_page_count_min_set(wdoms)) { 3543 if (pageproc == NULL) 3544 panic("vm_wait in early boot"); 3545 vm_min_waiters++; 3546 error = msleep(&vm_min_domains, &vm_domainset_lock, 3547 PVM | PDROP | mflags, "vmwait", 0); 3548 } else 3549 mtx_unlock(&vm_domainset_lock); 3550 } 3551 return (error); 3552 } 3553 3554 /* 3555 * vm_wait_domain: 3556 * 3557 * Sleep until free pages are available for allocation. 3558 * - Called in various places after failed memory allocations. 3559 */ 3560 void 3561 vm_wait_domain(int domain) 3562 { 3563 struct vm_domain *vmd; 3564 domainset_t wdom; 3565 3566 vmd = VM_DOMAIN(domain); 3567 vm_domain_free_assert_unlocked(vmd); 3568 3569 if (curproc == pageproc) { 3570 mtx_lock(&vm_domainset_lock); 3571 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3572 vmd->vmd_pageout_pages_needed = 1; 3573 msleep(&vmd->vmd_pageout_pages_needed, 3574 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3575 } else 3576 mtx_unlock(&vm_domainset_lock); 3577 } else { 3578 DOMAINSET_ZERO(&wdom); 3579 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3580 vm_wait_doms(&wdom, 0); 3581 } 3582 } 3583 3584 static int 3585 vm_wait_flags(vm_object_t obj, int mflags) 3586 { 3587 struct domainset *d; 3588 3589 d = NULL; 3590 3591 /* 3592 * Carefully fetch pointers only once: the struct domainset 3593 * itself is ummutable but the pointer might change. 3594 */ 3595 if (obj != NULL) 3596 d = obj->domain.dr_policy; 3597 if (d == NULL) 3598 d = curthread->td_domain.dr_policy; 3599 3600 return (vm_wait_doms(&d->ds_mask, mflags)); 3601 } 3602 3603 /* 3604 * vm_wait: 3605 * 3606 * Sleep until free pages are available for allocation in the 3607 * affinity domains of the obj. If obj is NULL, the domain set 3608 * for the calling thread is used. 3609 * Called in various places after failed memory allocations. 3610 */ 3611 void 3612 vm_wait(vm_object_t obj) 3613 { 3614 (void)vm_wait_flags(obj, 0); 3615 } 3616 3617 int 3618 vm_wait_intr(vm_object_t obj) 3619 { 3620 return (vm_wait_flags(obj, PCATCH)); 3621 } 3622 3623 /* 3624 * vm_domain_alloc_fail: 3625 * 3626 * Called when a page allocation function fails. Informs the 3627 * pagedaemon and performs the requested wait. Requires the 3628 * domain_free and object lock on entry. Returns with the 3629 * object lock held and free lock released. Returns an error when 3630 * retry is necessary. 3631 * 3632 */ 3633 static int 3634 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3635 { 3636 3637 vm_domain_free_assert_unlocked(vmd); 3638 3639 atomic_add_int(&vmd->vmd_pageout_deficit, 3640 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3641 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3642 if (object != NULL) 3643 VM_OBJECT_WUNLOCK(object); 3644 vm_wait_domain(vmd->vmd_domain); 3645 if (object != NULL) 3646 VM_OBJECT_WLOCK(object); 3647 if (req & VM_ALLOC_WAITOK) 3648 return (EAGAIN); 3649 } 3650 3651 return (0); 3652 } 3653 3654 /* 3655 * vm_waitpfault: 3656 * 3657 * Sleep until free pages are available for allocation. 3658 * - Called only in vm_fault so that processes page faulting 3659 * can be easily tracked. 3660 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3661 * processes will be able to grab memory first. Do not change 3662 * this balance without careful testing first. 3663 */ 3664 void 3665 vm_waitpfault(struct domainset *dset, int timo) 3666 { 3667 3668 /* 3669 * XXX Ideally we would wait only until the allocation could 3670 * be satisfied. This condition can cause new allocators to 3671 * consume all freed pages while old allocators wait. 3672 */ 3673 mtx_lock(&vm_domainset_lock); 3674 if (vm_page_count_min_set(&dset->ds_mask)) { 3675 vm_min_waiters++; 3676 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3677 "pfault", timo); 3678 } else 3679 mtx_unlock(&vm_domainset_lock); 3680 } 3681 3682 static struct vm_pagequeue * 3683 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3684 { 3685 3686 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3687 } 3688 3689 #ifdef INVARIANTS 3690 static struct vm_pagequeue * 3691 vm_page_pagequeue(vm_page_t m) 3692 { 3693 3694 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3695 } 3696 #endif 3697 3698 static __always_inline bool 3699 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, 3700 vm_page_astate_t new) 3701 { 3702 vm_page_astate_t tmp; 3703 3704 tmp = *old; 3705 do { 3706 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3707 return (true); 3708 counter_u64_add(pqstate_commit_retries, 1); 3709 } while (old->_bits == tmp._bits); 3710 3711 return (false); 3712 } 3713 3714 /* 3715 * Do the work of committing a queue state update that moves the page out of 3716 * its current queue. 3717 */ 3718 static bool 3719 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3720 vm_page_astate_t *old, vm_page_astate_t new) 3721 { 3722 vm_page_t next; 3723 3724 vm_pagequeue_assert_locked(pq); 3725 KASSERT(vm_page_pagequeue(m) == pq, 3726 ("%s: queue %p does not match page %p", __func__, pq, m)); 3727 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3728 ("%s: invalid queue indices %d %d", 3729 __func__, old->queue, new.queue)); 3730 3731 /* 3732 * Once the queue index of the page changes there is nothing 3733 * synchronizing with further updates to the page's physical 3734 * queue state. Therefore we must speculatively remove the page 3735 * from the queue now and be prepared to roll back if the queue 3736 * state update fails. If the page is not physically enqueued then 3737 * we just update its queue index. 3738 */ 3739 if ((old->flags & PGA_ENQUEUED) != 0) { 3740 new.flags &= ~PGA_ENQUEUED; 3741 next = TAILQ_NEXT(m, plinks.q); 3742 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3743 vm_pagequeue_cnt_dec(pq); 3744 if (!vm_page_pqstate_fcmpset(m, old, new)) { 3745 if (next == NULL) 3746 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3747 else 3748 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3749 vm_pagequeue_cnt_inc(pq); 3750 return (false); 3751 } else { 3752 return (true); 3753 } 3754 } else { 3755 return (vm_page_pqstate_fcmpset(m, old, new)); 3756 } 3757 } 3758 3759 static bool 3760 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3761 vm_page_astate_t new) 3762 { 3763 struct vm_pagequeue *pq; 3764 vm_page_astate_t as; 3765 bool ret; 3766 3767 pq = _vm_page_pagequeue(m, old->queue); 3768 3769 /* 3770 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3771 * corresponding page queue lock is held. 3772 */ 3773 vm_pagequeue_lock(pq); 3774 as = vm_page_astate_load(m); 3775 if (__predict_false(as._bits != old->_bits)) { 3776 *old = as; 3777 ret = false; 3778 } else { 3779 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3780 } 3781 vm_pagequeue_unlock(pq); 3782 return (ret); 3783 } 3784 3785 /* 3786 * Commit a queue state update that enqueues or requeues a page. 3787 */ 3788 static bool 3789 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3790 vm_page_astate_t *old, vm_page_astate_t new) 3791 { 3792 struct vm_domain *vmd; 3793 3794 vm_pagequeue_assert_locked(pq); 3795 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3796 ("%s: invalid queue indices %d %d", 3797 __func__, old->queue, new.queue)); 3798 3799 new.flags |= PGA_ENQUEUED; 3800 if (!vm_page_pqstate_fcmpset(m, old, new)) 3801 return (false); 3802 3803 if ((old->flags & PGA_ENQUEUED) != 0) 3804 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3805 else 3806 vm_pagequeue_cnt_inc(pq); 3807 3808 /* 3809 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3810 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3811 * applied, even if it was set first. 3812 */ 3813 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3814 vmd = vm_pagequeue_domain(m); 3815 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3816 ("%s: invalid page queue for page %p", __func__, m)); 3817 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3818 } else { 3819 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3820 } 3821 return (true); 3822 } 3823 3824 /* 3825 * Commit a queue state update that encodes a request for a deferred queue 3826 * operation. 3827 */ 3828 static bool 3829 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3830 vm_page_astate_t new) 3831 { 3832 3833 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3834 ("%s: invalid state, queue %d flags %x", 3835 __func__, new.queue, new.flags)); 3836 3837 if (old->_bits != new._bits && 3838 !vm_page_pqstate_fcmpset(m, old, new)) 3839 return (false); 3840 vm_page_pqbatch_submit(m, new.queue); 3841 return (true); 3842 } 3843 3844 /* 3845 * A generic queue state update function. This handles more cases than the 3846 * specialized functions above. 3847 */ 3848 bool 3849 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3850 { 3851 3852 if (old->_bits == new._bits) 3853 return (true); 3854 3855 if (old->queue != PQ_NONE && new.queue != old->queue) { 3856 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3857 return (false); 3858 if (new.queue != PQ_NONE) 3859 vm_page_pqbatch_submit(m, new.queue); 3860 } else { 3861 if (!vm_page_pqstate_fcmpset(m, old, new)) 3862 return (false); 3863 if (new.queue != PQ_NONE && 3864 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3865 vm_page_pqbatch_submit(m, new.queue); 3866 } 3867 return (true); 3868 } 3869 3870 /* 3871 * Apply deferred queue state updates to a page. 3872 */ 3873 static inline void 3874 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3875 { 3876 vm_page_astate_t new, old; 3877 3878 CRITICAL_ASSERT(curthread); 3879 vm_pagequeue_assert_locked(pq); 3880 KASSERT(queue < PQ_COUNT, 3881 ("%s: invalid queue index %d", __func__, queue)); 3882 KASSERT(pq == _vm_page_pagequeue(m, queue), 3883 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3884 3885 for (old = vm_page_astate_load(m);;) { 3886 if (__predict_false(old.queue != queue || 3887 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3888 counter_u64_add(queue_nops, 1); 3889 break; 3890 } 3891 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3892 ("%s: page %p is unmanaged", __func__, m)); 3893 3894 new = old; 3895 if ((old.flags & PGA_DEQUEUE) != 0) { 3896 new.flags &= ~PGA_QUEUE_OP_MASK; 3897 new.queue = PQ_NONE; 3898 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3899 m, &old, new))) { 3900 counter_u64_add(queue_ops, 1); 3901 break; 3902 } 3903 } else { 3904 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3905 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3906 m, &old, new))) { 3907 counter_u64_add(queue_ops, 1); 3908 break; 3909 } 3910 } 3911 } 3912 } 3913 3914 static void 3915 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3916 uint8_t queue) 3917 { 3918 int i; 3919 3920 for (i = 0; i < bq->bq_cnt; i++) 3921 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3922 vm_batchqueue_init(bq); 3923 } 3924 3925 /* 3926 * vm_page_pqbatch_submit: [ internal use only ] 3927 * 3928 * Enqueue a page in the specified page queue's batched work queue. 3929 * The caller must have encoded the requested operation in the page 3930 * structure's a.flags field. 3931 */ 3932 void 3933 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3934 { 3935 struct vm_batchqueue *bq; 3936 struct vm_pagequeue *pq; 3937 int domain, slots_remaining; 3938 3939 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3940 3941 domain = vm_page_domain(m); 3942 critical_enter(); 3943 bq = DPCPU_PTR(pqbatch[domain][queue]); 3944 slots_remaining = vm_batchqueue_insert(bq, m); 3945 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 3946 /* keep building the bq */ 3947 critical_exit(); 3948 return; 3949 } else if (slots_remaining > 0 ) { 3950 /* Try to process the bq if we can get the lock */ 3951 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3952 if (vm_pagequeue_trylock(pq)) { 3953 vm_pqbatch_process(pq, bq, queue); 3954 vm_pagequeue_unlock(pq); 3955 } 3956 critical_exit(); 3957 return; 3958 } 3959 critical_exit(); 3960 3961 /* if we make it here, the bq is full so wait for the lock */ 3962 3963 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3964 vm_pagequeue_lock(pq); 3965 critical_enter(); 3966 bq = DPCPU_PTR(pqbatch[domain][queue]); 3967 vm_pqbatch_process(pq, bq, queue); 3968 vm_pqbatch_process_page(pq, m, queue); 3969 vm_pagequeue_unlock(pq); 3970 critical_exit(); 3971 } 3972 3973 /* 3974 * vm_page_pqbatch_drain: [ internal use only ] 3975 * 3976 * Force all per-CPU page queue batch queues to be drained. This is 3977 * intended for use in severe memory shortages, to ensure that pages 3978 * do not remain stuck in the batch queues. 3979 */ 3980 void 3981 vm_page_pqbatch_drain(void) 3982 { 3983 struct thread *td; 3984 struct vm_domain *vmd; 3985 struct vm_pagequeue *pq; 3986 int cpu, domain, queue; 3987 3988 td = curthread; 3989 CPU_FOREACH(cpu) { 3990 thread_lock(td); 3991 sched_bind(td, cpu); 3992 thread_unlock(td); 3993 3994 for (domain = 0; domain < vm_ndomains; domain++) { 3995 vmd = VM_DOMAIN(domain); 3996 for (queue = 0; queue < PQ_COUNT; queue++) { 3997 pq = &vmd->vmd_pagequeues[queue]; 3998 vm_pagequeue_lock(pq); 3999 critical_enter(); 4000 vm_pqbatch_process(pq, 4001 DPCPU_PTR(pqbatch[domain][queue]), queue); 4002 critical_exit(); 4003 vm_pagequeue_unlock(pq); 4004 } 4005 } 4006 } 4007 thread_lock(td); 4008 sched_unbind(td); 4009 thread_unlock(td); 4010 } 4011 4012 /* 4013 * vm_page_dequeue_deferred: [ internal use only ] 4014 * 4015 * Request removal of the given page from its current page 4016 * queue. Physical removal from the queue may be deferred 4017 * indefinitely. 4018 */ 4019 void 4020 vm_page_dequeue_deferred(vm_page_t m) 4021 { 4022 vm_page_astate_t new, old; 4023 4024 old = vm_page_astate_load(m); 4025 do { 4026 if (old.queue == PQ_NONE) { 4027 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4028 ("%s: page %p has unexpected queue state", 4029 __func__, m)); 4030 break; 4031 } 4032 new = old; 4033 new.flags |= PGA_DEQUEUE; 4034 } while (!vm_page_pqstate_commit_request(m, &old, new)); 4035 } 4036 4037 /* 4038 * vm_page_dequeue: 4039 * 4040 * Remove the page from whichever page queue it's in, if any, before 4041 * returning. 4042 */ 4043 void 4044 vm_page_dequeue(vm_page_t m) 4045 { 4046 vm_page_astate_t new, old; 4047 4048 old = vm_page_astate_load(m); 4049 do { 4050 if (old.queue == PQ_NONE) { 4051 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4052 ("%s: page %p has unexpected queue state", 4053 __func__, m)); 4054 break; 4055 } 4056 new = old; 4057 new.flags &= ~PGA_QUEUE_OP_MASK; 4058 new.queue = PQ_NONE; 4059 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 4060 4061 } 4062 4063 /* 4064 * Schedule the given page for insertion into the specified page queue. 4065 * Physical insertion of the page may be deferred indefinitely. 4066 */ 4067 static void 4068 vm_page_enqueue(vm_page_t m, uint8_t queue) 4069 { 4070 4071 KASSERT(m->a.queue == PQ_NONE && 4072 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 4073 ("%s: page %p is already enqueued", __func__, m)); 4074 KASSERT(m->ref_count > 0, 4075 ("%s: page %p does not carry any references", __func__, m)); 4076 4077 m->a.queue = queue; 4078 if ((m->a.flags & PGA_REQUEUE) == 0) 4079 vm_page_aflag_set(m, PGA_REQUEUE); 4080 vm_page_pqbatch_submit(m, queue); 4081 } 4082 4083 /* 4084 * vm_page_free_prep: 4085 * 4086 * Prepares the given page to be put on the free list, 4087 * disassociating it from any VM object. The caller may return 4088 * the page to the free list only if this function returns true. 4089 * 4090 * The object, if it exists, must be locked, and then the page must 4091 * be xbusy. Otherwise the page must be not busied. A managed 4092 * page must be unmapped. 4093 */ 4094 static bool 4095 vm_page_free_prep(vm_page_t m) 4096 { 4097 4098 /* 4099 * Synchronize with threads that have dropped a reference to this 4100 * page. 4101 */ 4102 atomic_thread_fence_acq(); 4103 4104 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 4105 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 4106 uint64_t *p; 4107 int i; 4108 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 4109 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 4110 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 4111 m, i, (uintmax_t)*p)); 4112 } 4113 #endif 4114 if ((m->oflags & VPO_UNMANAGED) == 0) { 4115 KASSERT(!pmap_page_is_mapped(m), 4116 ("vm_page_free_prep: freeing mapped page %p", m)); 4117 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 4118 ("vm_page_free_prep: mapping flags set in page %p", m)); 4119 } else { 4120 KASSERT(m->a.queue == PQ_NONE, 4121 ("vm_page_free_prep: unmanaged page %p is queued", m)); 4122 } 4123 VM_CNT_INC(v_tfree); 4124 4125 if (m->object != NULL) { 4126 vm_page_radix_remove(m); 4127 vm_page_free_object_prep(m); 4128 } else 4129 vm_page_assert_unbusied(m); 4130 4131 vm_page_busy_free(m); 4132 4133 /* 4134 * If fictitious remove object association and 4135 * return. 4136 */ 4137 if ((m->flags & PG_FICTITIOUS) != 0) { 4138 KASSERT(m->ref_count == 1, 4139 ("fictitious page %p is referenced", m)); 4140 KASSERT(m->a.queue == PQ_NONE, 4141 ("fictitious page %p is queued", m)); 4142 return (false); 4143 } 4144 4145 /* 4146 * Pages need not be dequeued before they are returned to the physical 4147 * memory allocator, but they must at least be marked for a deferred 4148 * dequeue. 4149 */ 4150 if ((m->oflags & VPO_UNMANAGED) == 0) 4151 vm_page_dequeue_deferred(m); 4152 4153 m->valid = 0; 4154 vm_page_undirty(m); 4155 4156 if (m->ref_count != 0) 4157 panic("vm_page_free_prep: page %p has references", m); 4158 4159 /* 4160 * Restore the default memory attribute to the page. 4161 */ 4162 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 4163 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 4164 4165 #if VM_NRESERVLEVEL > 0 4166 /* 4167 * Determine whether the page belongs to a reservation. If the page was 4168 * allocated from a per-CPU cache, it cannot belong to a reservation, so 4169 * as an optimization, we avoid the check in that case. 4170 */ 4171 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 4172 return (false); 4173 #endif 4174 4175 return (true); 4176 } 4177 4178 /* 4179 * vm_page_free_toq: 4180 * 4181 * Returns the given page to the free list, disassociating it 4182 * from any VM object. 4183 * 4184 * The object must be locked. The page must be exclusively busied if it 4185 * belongs to an object. 4186 */ 4187 static void 4188 vm_page_free_toq(vm_page_t m) 4189 { 4190 struct vm_domain *vmd; 4191 uma_zone_t zone; 4192 4193 if (!vm_page_free_prep(m)) 4194 return; 4195 4196 vmd = vm_pagequeue_domain(m); 4197 if (__predict_false((m->flags & PG_NOFREE) != 0)) { 4198 vm_page_free_nofree(vmd, m); 4199 return; 4200 } 4201 zone = vmd->vmd_pgcache[m->pool].zone; 4202 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4203 uma_zfree(zone, m); 4204 return; 4205 } 4206 vm_domain_free_lock(vmd); 4207 vm_phys_free_pages(m, m->pool, 0); 4208 vm_domain_free_unlock(vmd); 4209 vm_domain_freecnt_inc(vmd, 1); 4210 } 4211 4212 /* 4213 * vm_page_free_pages_toq: 4214 * 4215 * Returns a list of pages to the free list, disassociating it 4216 * from any VM object. In other words, this is equivalent to 4217 * calling vm_page_free_toq() for each page of a list of VM objects. 4218 */ 4219 int 4220 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4221 { 4222 vm_page_t m; 4223 int count; 4224 4225 if (SLIST_EMPTY(free)) 4226 return (0); 4227 4228 count = 0; 4229 while ((m = SLIST_FIRST(free)) != NULL) { 4230 count++; 4231 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4232 vm_page_free_toq(m); 4233 } 4234 4235 if (update_wire_count) 4236 vm_wire_sub(count); 4237 return (count); 4238 } 4239 4240 /* 4241 * Mark this page as wired down. For managed pages, this prevents reclamation 4242 * by the page daemon, or when the containing object, if any, is destroyed. 4243 */ 4244 void 4245 vm_page_wire(vm_page_t m) 4246 { 4247 u_int old; 4248 4249 #ifdef INVARIANTS 4250 if (m->object != NULL && !vm_page_busied(m) && 4251 !vm_object_busied(m->object)) 4252 VM_OBJECT_ASSERT_LOCKED(m->object); 4253 #endif 4254 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4255 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4256 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4257 4258 old = atomic_fetchadd_int(&m->ref_count, 1); 4259 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4260 ("vm_page_wire: counter overflow for page %p", m)); 4261 if (VPRC_WIRE_COUNT(old) == 0) { 4262 if ((m->oflags & VPO_UNMANAGED) == 0) 4263 vm_page_aflag_set(m, PGA_DEQUEUE); 4264 vm_wire_add(1); 4265 } 4266 } 4267 4268 /* 4269 * Attempt to wire a mapped page following a pmap lookup of that page. 4270 * This may fail if a thread is concurrently tearing down mappings of the page. 4271 * The transient failure is acceptable because it translates to the 4272 * failure of the caller pmap_extract_and_hold(), which should be then 4273 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4274 */ 4275 bool 4276 vm_page_wire_mapped(vm_page_t m) 4277 { 4278 u_int old; 4279 4280 old = atomic_load_int(&m->ref_count); 4281 do { 4282 KASSERT(old > 0, 4283 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4284 if ((old & VPRC_BLOCKED) != 0) 4285 return (false); 4286 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4287 4288 if (VPRC_WIRE_COUNT(old) == 0) { 4289 if ((m->oflags & VPO_UNMANAGED) == 0) 4290 vm_page_aflag_set(m, PGA_DEQUEUE); 4291 vm_wire_add(1); 4292 } 4293 return (true); 4294 } 4295 4296 /* 4297 * Release a wiring reference to a managed page. If the page still belongs to 4298 * an object, update its position in the page queues to reflect the reference. 4299 * If the wiring was the last reference to the page, free the page. 4300 */ 4301 static void 4302 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4303 { 4304 u_int old; 4305 4306 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4307 ("%s: page %p is unmanaged", __func__, m)); 4308 4309 /* 4310 * Update LRU state before releasing the wiring reference. 4311 * Use a release store when updating the reference count to 4312 * synchronize with vm_page_free_prep(). 4313 */ 4314 old = atomic_load_int(&m->ref_count); 4315 do { 4316 u_int count; 4317 4318 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4319 ("vm_page_unwire: wire count underflow for page %p", m)); 4320 4321 count = old & ~VPRC_BLOCKED; 4322 if (count > VPRC_OBJREF + 1) { 4323 /* 4324 * The page has at least one other wiring reference. An 4325 * earlier iteration of this loop may have called 4326 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4327 * re-set it if necessary. 4328 */ 4329 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4330 vm_page_aflag_set(m, PGA_DEQUEUE); 4331 } else if (count == VPRC_OBJREF + 1) { 4332 /* 4333 * This is the last wiring. Clear PGA_DEQUEUE and 4334 * update the page's queue state to reflect the 4335 * reference. If the page does not belong to an object 4336 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4337 * clear leftover queue state. 4338 */ 4339 vm_page_release_toq(m, nqueue, noreuse); 4340 } else if (count == 1) { 4341 vm_page_aflag_clear(m, PGA_DEQUEUE); 4342 } 4343 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4344 4345 if (VPRC_WIRE_COUNT(old) == 1) { 4346 vm_wire_sub(1); 4347 if (old == 1) 4348 vm_page_free(m); 4349 } 4350 } 4351 4352 /* 4353 * Release one wiring of the specified page, potentially allowing it to be 4354 * paged out. 4355 * 4356 * Only managed pages belonging to an object can be paged out. If the number 4357 * of wirings transitions to zero and the page is eligible for page out, then 4358 * the page is added to the specified paging queue. If the released wiring 4359 * represented the last reference to the page, the page is freed. 4360 */ 4361 void 4362 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4363 { 4364 4365 KASSERT(nqueue < PQ_COUNT, 4366 ("vm_page_unwire: invalid queue %u request for page %p", 4367 nqueue, m)); 4368 4369 if ((m->oflags & VPO_UNMANAGED) != 0) { 4370 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4371 vm_page_free(m); 4372 return; 4373 } 4374 vm_page_unwire_managed(m, nqueue, false); 4375 } 4376 4377 /* 4378 * Unwire a page without (re-)inserting it into a page queue. It is up 4379 * to the caller to enqueue, requeue, or free the page as appropriate. 4380 * In most cases involving managed pages, vm_page_unwire() should be used 4381 * instead. 4382 */ 4383 bool 4384 vm_page_unwire_noq(vm_page_t m) 4385 { 4386 u_int old; 4387 4388 old = vm_page_drop(m, 1); 4389 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4390 ("%s: counter underflow for page %p", __func__, m)); 4391 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4392 ("%s: missing ref on fictitious page %p", __func__, m)); 4393 4394 if (VPRC_WIRE_COUNT(old) > 1) 4395 return (false); 4396 if ((m->oflags & VPO_UNMANAGED) == 0) 4397 vm_page_aflag_clear(m, PGA_DEQUEUE); 4398 vm_wire_sub(1); 4399 return (true); 4400 } 4401 4402 /* 4403 * Ensure that the page ends up in the specified page queue. If the page is 4404 * active or being moved to the active queue, ensure that its act_count is 4405 * at least ACT_INIT but do not otherwise mess with it. 4406 */ 4407 static __always_inline void 4408 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4409 { 4410 vm_page_astate_t old, new; 4411 4412 KASSERT(m->ref_count > 0, 4413 ("%s: page %p does not carry any references", __func__, m)); 4414 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4415 ("%s: invalid flags %x", __func__, nflag)); 4416 4417 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4418 return; 4419 4420 old = vm_page_astate_load(m); 4421 do { 4422 if ((old.flags & PGA_DEQUEUE) != 0) 4423 break; 4424 new = old; 4425 new.flags &= ~PGA_QUEUE_OP_MASK; 4426 if (nqueue == PQ_ACTIVE) 4427 new.act_count = max(old.act_count, ACT_INIT); 4428 if (old.queue == nqueue) { 4429 /* 4430 * There is no need to requeue pages already in the 4431 * active queue. 4432 */ 4433 if (nqueue != PQ_ACTIVE || 4434 (old.flags & PGA_ENQUEUED) == 0) 4435 new.flags |= nflag; 4436 } else { 4437 new.flags |= nflag; 4438 new.queue = nqueue; 4439 } 4440 } while (!vm_page_pqstate_commit(m, &old, new)); 4441 } 4442 4443 /* 4444 * Put the specified page on the active list (if appropriate). 4445 */ 4446 void 4447 vm_page_activate(vm_page_t m) 4448 { 4449 4450 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4451 } 4452 4453 /* 4454 * Move the specified page to the tail of the inactive queue, or requeue 4455 * the page if it is already in the inactive queue. 4456 */ 4457 void 4458 vm_page_deactivate(vm_page_t m) 4459 { 4460 4461 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4462 } 4463 4464 void 4465 vm_page_deactivate_noreuse(vm_page_t m) 4466 { 4467 4468 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4469 } 4470 4471 /* 4472 * Put a page in the laundry, or requeue it if it is already there. 4473 */ 4474 void 4475 vm_page_launder(vm_page_t m) 4476 { 4477 4478 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4479 } 4480 4481 /* 4482 * Put a page in the PQ_UNSWAPPABLE holding queue. 4483 */ 4484 void 4485 vm_page_unswappable(vm_page_t m) 4486 { 4487 4488 VM_OBJECT_ASSERT_LOCKED(m->object); 4489 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4490 ("page %p already unswappable", m)); 4491 4492 vm_page_dequeue(m); 4493 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4494 } 4495 4496 /* 4497 * Release a page back to the page queues in preparation for unwiring. 4498 */ 4499 static void 4500 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4501 { 4502 vm_page_astate_t old, new; 4503 uint16_t nflag; 4504 4505 /* 4506 * Use a check of the valid bits to determine whether we should 4507 * accelerate reclamation of the page. The object lock might not be 4508 * held here, in which case the check is racy. At worst we will either 4509 * accelerate reclamation of a valid page and violate LRU, or 4510 * unnecessarily defer reclamation of an invalid page. 4511 * 4512 * If we were asked to not cache the page, place it near the head of the 4513 * inactive queue so that is reclaimed sooner. 4514 */ 4515 if (noreuse || vm_page_none_valid(m)) { 4516 nqueue = PQ_INACTIVE; 4517 nflag = PGA_REQUEUE_HEAD; 4518 } else { 4519 nflag = PGA_REQUEUE; 4520 } 4521 4522 old = vm_page_astate_load(m); 4523 do { 4524 new = old; 4525 4526 /* 4527 * If the page is already in the active queue and we are not 4528 * trying to accelerate reclamation, simply mark it as 4529 * referenced and avoid any queue operations. 4530 */ 4531 new.flags &= ~PGA_QUEUE_OP_MASK; 4532 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4533 (old.flags & PGA_ENQUEUED) != 0) 4534 new.flags |= PGA_REFERENCED; 4535 else { 4536 new.flags |= nflag; 4537 new.queue = nqueue; 4538 } 4539 } while (!vm_page_pqstate_commit(m, &old, new)); 4540 } 4541 4542 /* 4543 * Unwire a page and either attempt to free it or re-add it to the page queues. 4544 */ 4545 void 4546 vm_page_release(vm_page_t m, int flags) 4547 { 4548 vm_object_t object; 4549 4550 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4551 ("vm_page_release: page %p is unmanaged", m)); 4552 4553 if ((flags & VPR_TRYFREE) != 0) { 4554 for (;;) { 4555 object = atomic_load_ptr(&m->object); 4556 if (object == NULL) 4557 break; 4558 /* Depends on type-stability. */ 4559 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4560 break; 4561 if (object == m->object) { 4562 vm_page_release_locked(m, flags); 4563 VM_OBJECT_WUNLOCK(object); 4564 return; 4565 } 4566 VM_OBJECT_WUNLOCK(object); 4567 } 4568 } 4569 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4570 } 4571 4572 /* See vm_page_release(). */ 4573 void 4574 vm_page_release_locked(vm_page_t m, int flags) 4575 { 4576 4577 VM_OBJECT_ASSERT_WLOCKED(m->object); 4578 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4579 ("vm_page_release_locked: page %p is unmanaged", m)); 4580 4581 if (vm_page_unwire_noq(m)) { 4582 if ((flags & VPR_TRYFREE) != 0 && 4583 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4584 m->dirty == 0 && vm_page_tryxbusy(m)) { 4585 /* 4586 * An unlocked lookup may have wired the page before the 4587 * busy lock was acquired, in which case the page must 4588 * not be freed. 4589 */ 4590 if (__predict_true(!vm_page_wired(m))) { 4591 vm_page_free(m); 4592 return; 4593 } 4594 vm_page_xunbusy(m); 4595 } else { 4596 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4597 } 4598 } 4599 } 4600 4601 static bool 4602 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4603 { 4604 u_int old; 4605 4606 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4607 ("vm_page_try_blocked_op: page %p has no object", m)); 4608 KASSERT(vm_page_busied(m), 4609 ("vm_page_try_blocked_op: page %p is not busy", m)); 4610 VM_OBJECT_ASSERT_LOCKED(m->object); 4611 4612 old = atomic_load_int(&m->ref_count); 4613 do { 4614 KASSERT(old != 0, 4615 ("vm_page_try_blocked_op: page %p has no references", m)); 4616 KASSERT((old & VPRC_BLOCKED) == 0, 4617 ("vm_page_try_blocked_op: page %p blocks wirings", m)); 4618 if (VPRC_WIRE_COUNT(old) != 0) 4619 return (false); 4620 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4621 4622 (op)(m); 4623 4624 /* 4625 * If the object is read-locked, new wirings may be created via an 4626 * object lookup. 4627 */ 4628 old = vm_page_drop(m, VPRC_BLOCKED); 4629 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4630 old == (VPRC_BLOCKED | VPRC_OBJREF), 4631 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4632 old, m)); 4633 return (true); 4634 } 4635 4636 /* 4637 * Atomically check for wirings and remove all mappings of the page. 4638 */ 4639 bool 4640 vm_page_try_remove_all(vm_page_t m) 4641 { 4642 4643 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4644 } 4645 4646 /* 4647 * Atomically check for wirings and remove all writeable mappings of the page. 4648 */ 4649 bool 4650 vm_page_try_remove_write(vm_page_t m) 4651 { 4652 4653 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4654 } 4655 4656 /* 4657 * vm_page_advise 4658 * 4659 * Apply the specified advice to the given page. 4660 */ 4661 void 4662 vm_page_advise(vm_page_t m, int advice) 4663 { 4664 4665 VM_OBJECT_ASSERT_WLOCKED(m->object); 4666 vm_page_assert_xbusied(m); 4667 4668 if (advice == MADV_FREE) 4669 /* 4670 * Mark the page clean. This will allow the page to be freed 4671 * without first paging it out. MADV_FREE pages are often 4672 * quickly reused by malloc(3), so we do not do anything that 4673 * would result in a page fault on a later access. 4674 */ 4675 vm_page_undirty(m); 4676 else if (advice != MADV_DONTNEED) { 4677 if (advice == MADV_WILLNEED) 4678 vm_page_activate(m); 4679 return; 4680 } 4681 4682 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4683 vm_page_dirty(m); 4684 4685 /* 4686 * Clear any references to the page. Otherwise, the page daemon will 4687 * immediately reactivate the page. 4688 */ 4689 vm_page_aflag_clear(m, PGA_REFERENCED); 4690 4691 /* 4692 * Place clean pages near the head of the inactive queue rather than 4693 * the tail, thus defeating the queue's LRU operation and ensuring that 4694 * the page will be reused quickly. Dirty pages not already in the 4695 * laundry are moved there. 4696 */ 4697 if (m->dirty == 0) 4698 vm_page_deactivate_noreuse(m); 4699 else if (!vm_page_in_laundry(m)) 4700 vm_page_launder(m); 4701 } 4702 4703 /* 4704 * vm_page_grab_release 4705 * 4706 * Helper routine for grab functions to release busy on return. 4707 */ 4708 static inline void 4709 vm_page_grab_release(vm_page_t m, int allocflags) 4710 { 4711 4712 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4713 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4714 vm_page_sunbusy(m); 4715 else 4716 vm_page_xunbusy(m); 4717 } 4718 } 4719 4720 /* 4721 * vm_page_grab_sleep 4722 * 4723 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4724 * if the caller should retry and false otherwise. 4725 * 4726 * If the object is locked on entry the object will be unlocked with 4727 * false returns and still locked but possibly having been dropped 4728 * with true returns. 4729 */ 4730 static bool 4731 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4732 const char *wmesg, int allocflags, bool locked) 4733 { 4734 4735 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4736 return (false); 4737 4738 /* 4739 * Reference the page before unlocking and sleeping so that 4740 * the page daemon is less likely to reclaim it. 4741 */ 4742 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4743 vm_page_reference(m); 4744 4745 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4746 locked) 4747 VM_OBJECT_WLOCK(object); 4748 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4749 return (false); 4750 4751 return (true); 4752 } 4753 4754 /* 4755 * Assert that the grab flags are valid. 4756 */ 4757 static inline void 4758 vm_page_grab_check(int allocflags) 4759 { 4760 4761 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4762 (allocflags & VM_ALLOC_WIRED) != 0, 4763 ("vm_page_grab*: the pages must be busied or wired")); 4764 4765 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4766 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4767 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4768 } 4769 4770 /* 4771 * Calculate the page allocation flags for grab. 4772 */ 4773 static inline int 4774 vm_page_grab_pflags(int allocflags) 4775 { 4776 int pflags; 4777 4778 pflags = allocflags & 4779 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4780 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY); 4781 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4782 pflags |= VM_ALLOC_WAITFAIL; 4783 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4784 pflags |= VM_ALLOC_SBUSY; 4785 4786 return (pflags); 4787 } 4788 4789 /* 4790 * Grab a page, waiting until we are woken up due to the page changing state. 4791 * We keep on waiting, if the page continues to be in the object, unless 4792 * allocflags forbid waiting. 4793 * 4794 * The object must be locked on entry. This routine may sleep. The lock will, 4795 * however, be released and reacquired if the routine sleeps. 4796 * 4797 * Return a grabbed page, or NULL. Set *found if a page was found, whether or 4798 * not it was grabbed. 4799 */ 4800 static inline vm_page_t 4801 vm_page_grab_lookup(struct pctrie_iter *pages, vm_object_t object, 4802 vm_pindex_t pindex, int allocflags, bool *found) 4803 { 4804 vm_page_t m; 4805 4806 while ((*found = (m = vm_radix_iter_lookup(pages, pindex)) != NULL) && 4807 !vm_page_tryacquire(m, allocflags)) { 4808 if (!vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4809 allocflags, true)) 4810 return (NULL); 4811 pctrie_iter_reset(pages); 4812 } 4813 return (m); 4814 } 4815 4816 /* 4817 * Grab a page. Use an iterator parameter. Keep on waiting, as long as the page 4818 * exists in the object. If the page doesn't exist, first allocate it and then 4819 * conditionally zero it. 4820 * 4821 * The object must be locked on entry. This routine may sleep. The lock will, 4822 * however, be released and reacquired if the routine sleeps. 4823 */ 4824 vm_page_t 4825 vm_page_grab_iter(vm_object_t object, struct pctrie_iter *pages, 4826 vm_pindex_t pindex, int allocflags) 4827 { 4828 vm_page_t m, mpred; 4829 bool found; 4830 4831 VM_OBJECT_ASSERT_WLOCKED(object); 4832 vm_page_grab_check(allocflags); 4833 4834 while ((m = vm_page_grab_lookup( 4835 pages, object, pindex, allocflags, &found)) == NULL) { 4836 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4837 return (NULL); 4838 if (found && 4839 (allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4840 return (NULL); 4841 mpred = vm_radix_iter_lookup_lt(pages, pindex); 4842 m = vm_page_alloc_after(object, pages, pindex, 4843 vm_page_grab_pflags(allocflags), mpred); 4844 if (m != NULL) { 4845 if ((allocflags & VM_ALLOC_ZERO) != 0 && 4846 (m->flags & PG_ZERO) == 0) 4847 pmap_zero_page(m); 4848 break; 4849 } 4850 if ((allocflags & 4851 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4852 return (NULL); 4853 } 4854 vm_page_grab_release(m, allocflags); 4855 4856 return (m); 4857 } 4858 4859 /* 4860 * Grab a page. Keep on waiting, as long as the page exists in the object. If 4861 * the page doesn't exist, first allocate it and then conditionally zero it. 4862 * 4863 * The object must be locked on entry. This routine may sleep. The lock will, 4864 * however, be released and reacquired if the routine sleeps. 4865 */ 4866 vm_page_t 4867 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4868 { 4869 struct pctrie_iter pages; 4870 4871 VM_OBJECT_ASSERT_WLOCKED(object); 4872 vm_page_iter_init(&pages, object); 4873 return (vm_page_grab_iter(object, &pages, pindex, allocflags)); 4874 } 4875 4876 /* 4877 * Attempt to validate a page, locklessly acquiring it if necessary, given a 4878 * (object, pindex) tuple and either an invalided page or NULL. The resulting 4879 * page will be validated against the identity tuple, and busied or wired as 4880 * requested. A NULL page returned guarantees that the page was not in radix at 4881 * the time of the call but callers must perform higher level synchronization or 4882 * retry the operation under a lock if they require an atomic answer. This is 4883 * the only lock free validation routine, other routines can depend on the 4884 * resulting page state. 4885 * 4886 * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to 4887 * caller flags. 4888 */ 4889 #define PAGE_NOT_ACQUIRED ((vm_page_t)1) 4890 static vm_page_t 4891 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m, 4892 int allocflags) 4893 { 4894 if (m == NULL) 4895 m = vm_page_lookup_unlocked(object, pindex); 4896 for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) { 4897 if (vm_page_trybusy(m, allocflags)) { 4898 if (m->object == object && m->pindex == pindex) { 4899 if ((allocflags & VM_ALLOC_WIRED) != 0) 4900 vm_page_wire(m); 4901 vm_page_grab_release(m, allocflags); 4902 break; 4903 } 4904 /* relookup. */ 4905 vm_page_busy_release(m); 4906 cpu_spinwait(); 4907 continue; 4908 } 4909 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4910 allocflags, false)) 4911 return (PAGE_NOT_ACQUIRED); 4912 } 4913 return (m); 4914 } 4915 4916 /* 4917 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4918 * is not set. 4919 */ 4920 vm_page_t 4921 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4922 { 4923 vm_page_t m; 4924 4925 vm_page_grab_check(allocflags); 4926 m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags); 4927 if (m == PAGE_NOT_ACQUIRED) 4928 return (NULL); 4929 if (m != NULL) 4930 return (m); 4931 4932 /* 4933 * The radix lockless lookup should never return a false negative 4934 * errors. If the user specifies NOCREAT they are guaranteed there 4935 * was no page present at the instant of the call. A NOCREAT caller 4936 * must handle create races gracefully. 4937 */ 4938 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4939 return (NULL); 4940 4941 VM_OBJECT_WLOCK(object); 4942 m = vm_page_grab(object, pindex, allocflags); 4943 VM_OBJECT_WUNLOCK(object); 4944 4945 return (m); 4946 } 4947 4948 /* 4949 * Grab a page and make it valid, paging in if necessary. Use an iterator 4950 * parameter. Pages missing from their pager are zero filled and validated. If 4951 * a VM_ALLOC_COUNT is supplied and the page is not valid as many as 4952 * VM_INITIAL_PAGEIN pages can be brought in simultaneously. Additional pages 4953 * will be left on a paging queue but will neither be wired nor busy regardless 4954 * of allocflags. 4955 */ 4956 int 4957 vm_page_grab_valid_iter(vm_page_t *mp, vm_object_t object, 4958 struct pctrie_iter *pages, vm_pindex_t pindex, int allocflags) 4959 { 4960 vm_page_t m, mpred; 4961 vm_page_t ma[VM_INITIAL_PAGEIN]; 4962 int after, i, pflags, rv; 4963 4964 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4965 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4966 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4967 KASSERT((allocflags & 4968 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4969 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4970 VM_OBJECT_ASSERT_WLOCKED(object); 4971 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4972 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4973 pflags |= VM_ALLOC_WAITFAIL; 4974 4975 retrylookup: 4976 if ((m = vm_radix_iter_lookup(pages, pindex)) != NULL) { 4977 /* 4978 * If the page is fully valid it can only become invalid 4979 * with the object lock held. If it is not valid it can 4980 * become valid with the busy lock held. Therefore, we 4981 * may unnecessarily lock the exclusive busy here if we 4982 * race with I/O completion not using the object lock. 4983 * However, we will not end up with an invalid page and a 4984 * shared lock. 4985 */ 4986 if (!vm_page_trybusy(m, 4987 vm_page_all_valid(m) ? allocflags : 0)) { 4988 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4989 allocflags, true); 4990 pctrie_iter_reset(pages); 4991 goto retrylookup; 4992 } 4993 if (vm_page_all_valid(m)) 4994 goto out; 4995 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 4996 vm_page_busy_release(m); 4997 *mp = NULL; 4998 return (VM_PAGER_FAIL); 4999 } 5000 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5001 *mp = NULL; 5002 return (VM_PAGER_FAIL); 5003 } else { 5004 mpred = vm_radix_iter_lookup_lt(pages, pindex); 5005 m = vm_page_alloc_after(object, pages, pindex, pflags, mpred); 5006 if (m == NULL) { 5007 if (!vm_pager_can_alloc_page(object, pindex)) { 5008 *mp = NULL; 5009 return (VM_PAGER_AGAIN); 5010 } 5011 pctrie_iter_reset(pages); 5012 goto retrylookup; 5013 } 5014 } 5015 5016 vm_page_assert_xbusied(m); 5017 if (vm_pager_has_page(object, pindex, NULL, &after)) { 5018 after = MIN(after, VM_INITIAL_PAGEIN); 5019 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 5020 after = MAX(after, 1); 5021 ma[0] = mpred = m; 5022 pctrie_iter_reset(pages); 5023 for (i = 1; i < after; i++) { 5024 m = vm_radix_iter_lookup(pages, pindex + i); 5025 if (m == NULL) { 5026 m = vm_page_alloc_after(object, pages, 5027 pindex + i, VM_ALLOC_NORMAL, mpred); 5028 if (m == NULL) 5029 break; 5030 } else if (vm_page_any_valid(m) || !vm_page_tryxbusy(m)) 5031 break; 5032 mpred = ma[i] = m; 5033 } 5034 after = i; 5035 vm_object_pip_add(object, after); 5036 VM_OBJECT_WUNLOCK(object); 5037 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 5038 VM_OBJECT_WLOCK(object); 5039 vm_object_pip_wakeupn(object, after); 5040 /* Pager may have replaced a page. */ 5041 m = ma[0]; 5042 if (rv != VM_PAGER_OK) { 5043 for (i = 0; i < after; i++) { 5044 if (!vm_page_wired(ma[i])) 5045 vm_page_free(ma[i]); 5046 else 5047 vm_page_xunbusy(ma[i]); 5048 } 5049 *mp = NULL; 5050 return (rv); 5051 } 5052 for (i = 1; i < after; i++) 5053 vm_page_readahead_finish(ma[i]); 5054 MPASS(vm_page_all_valid(m)); 5055 } else { 5056 vm_page_zero_invalid(m, TRUE); 5057 } 5058 pctrie_iter_reset(pages); 5059 out: 5060 if ((allocflags & VM_ALLOC_WIRED) != 0) 5061 vm_page_wire(m); 5062 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 5063 vm_page_busy_downgrade(m); 5064 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 5065 vm_page_busy_release(m); 5066 *mp = m; 5067 return (VM_PAGER_OK); 5068 } 5069 5070 /* 5071 * Grab a page and make it valid, paging in if necessary. Pages missing from 5072 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 5073 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 5074 * in simultaneously. Additional pages will be left on a paging queue but 5075 * will neither be wired nor busy regardless of allocflags. 5076 */ 5077 int 5078 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, 5079 int allocflags) 5080 { 5081 struct pctrie_iter pages; 5082 5083 VM_OBJECT_ASSERT_WLOCKED(object); 5084 vm_page_iter_init(&pages, object); 5085 return (vm_page_grab_valid_iter(mp, object, &pages, pindex, 5086 allocflags)); 5087 } 5088 5089 /* 5090 * Grab a page. Keep on waiting, as long as the page exists in the object. If 5091 * the page doesn't exist, and the pager has it, allocate it and zero part of 5092 * it. 5093 * 5094 * The object must be locked on entry. This routine may sleep. The lock will, 5095 * however, be released and reacquired if the routine sleeps. 5096 */ 5097 int 5098 vm_page_grab_zero_partial(vm_object_t object, vm_pindex_t pindex, int base, 5099 int end) 5100 { 5101 struct pctrie_iter pages; 5102 vm_page_t m, mpred; 5103 int allocflags, rv; 5104 bool found; 5105 5106 VM_OBJECT_ASSERT_WLOCKED(object); 5107 KASSERT(base >= 0, ("%s: base %d", __func__, base)); 5108 KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base, 5109 end)); 5110 5111 allocflags = VM_ALLOC_NOCREAT | VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL; 5112 vm_page_iter_init(&pages, object); 5113 while ((m = vm_page_grab_lookup( 5114 &pages, object, pindex, allocflags, &found)) == NULL) { 5115 if (!vm_pager_has_page(object, pindex, NULL, NULL)) 5116 return (0); 5117 mpred = vm_radix_iter_lookup_lt(&pages, pindex); 5118 m = vm_page_alloc_after(object, &pages, pindex, 5119 vm_page_grab_pflags(allocflags), mpred); 5120 if (m != NULL) { 5121 vm_object_pip_add(object, 1); 5122 VM_OBJECT_WUNLOCK(object); 5123 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 5124 VM_OBJECT_WLOCK(object); 5125 vm_object_pip_wakeup(object); 5126 if (rv != VM_PAGER_OK) { 5127 vm_page_free(m); 5128 return (EIO); 5129 } 5130 5131 /* 5132 * Since the page was not resident, and therefore not 5133 * recently accessed, immediately enqueue it for 5134 * asynchronous laundering. The current operation is 5135 * not regarded as an access. 5136 */ 5137 vm_page_launder(m); 5138 break; 5139 } 5140 } 5141 5142 pmap_zero_page_area(m, base, end - base); 5143 KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m)); 5144 vm_page_set_dirty(m); 5145 vm_page_xunbusy(m); 5146 return (0); 5147 } 5148 5149 /* 5150 * Locklessly grab a valid page. If the page is not valid or not yet 5151 * allocated this will fall back to the object lock method. 5152 */ 5153 int 5154 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 5155 vm_pindex_t pindex, int allocflags) 5156 { 5157 vm_page_t m; 5158 int flags; 5159 int error; 5160 5161 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 5162 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 5163 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 5164 "mismatch")); 5165 KASSERT((allocflags & 5166 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 5167 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 5168 5169 /* 5170 * Attempt a lockless lookup and busy. We need at least an sbusy 5171 * before we can inspect the valid field and return a wired page. 5172 */ 5173 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 5174 vm_page_grab_check(flags); 5175 m = vm_page_acquire_unlocked(object, pindex, NULL, flags); 5176 if (m == PAGE_NOT_ACQUIRED) 5177 return (VM_PAGER_FAIL); 5178 if (m != NULL) { 5179 if (vm_page_all_valid(m)) { 5180 if ((allocflags & VM_ALLOC_WIRED) != 0) 5181 vm_page_wire(m); 5182 vm_page_grab_release(m, allocflags); 5183 *mp = m; 5184 return (VM_PAGER_OK); 5185 } 5186 vm_page_busy_release(m); 5187 } 5188 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5189 *mp = NULL; 5190 return (VM_PAGER_FAIL); 5191 } 5192 VM_OBJECT_WLOCK(object); 5193 error = vm_page_grab_valid(mp, object, pindex, allocflags); 5194 VM_OBJECT_WUNLOCK(object); 5195 5196 return (error); 5197 } 5198 5199 /* 5200 * Return the specified range of pages from the given object. For each 5201 * page offset within the range, if a page already exists within the object 5202 * at that offset and it is busy, then wait for it to change state. If, 5203 * instead, the page doesn't exist, then allocate it. 5204 * 5205 * The caller must always specify an allocation class. 5206 * 5207 * allocation classes: 5208 * VM_ALLOC_NORMAL normal process request 5209 * VM_ALLOC_SYSTEM system *really* needs the pages 5210 * 5211 * The caller must always specify that the pages are to be busied and/or 5212 * wired. 5213 * 5214 * optional allocation flags: 5215 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 5216 * VM_ALLOC_NOBUSY do not exclusive busy the page 5217 * VM_ALLOC_NOWAIT do not sleep 5218 * VM_ALLOC_SBUSY set page to sbusy state 5219 * VM_ALLOC_WIRED wire the pages 5220 * VM_ALLOC_ZERO zero and validate any invalid pages 5221 * 5222 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 5223 * may return a partial prefix of the requested range. 5224 */ 5225 int 5226 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 5227 vm_page_t *ma, int count) 5228 { 5229 struct pctrie_iter pages; 5230 vm_page_t m, mpred; 5231 int pflags; 5232 int i; 5233 5234 VM_OBJECT_ASSERT_WLOCKED(object); 5235 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 5236 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 5237 KASSERT(count > 0, 5238 ("vm_page_grab_pages: invalid page count %d", count)); 5239 vm_page_grab_check(allocflags); 5240 5241 pflags = vm_page_grab_pflags(allocflags); 5242 i = 0; 5243 vm_page_iter_init(&pages, object); 5244 retrylookup: 5245 mpred = vm_radix_iter_lookup_lt(&pages, pindex + i); 5246 for (; i < count; i++) { 5247 m = vm_radix_iter_lookup(&pages, pindex + i); 5248 if (m != NULL) { 5249 if (!vm_page_tryacquire(m, allocflags)) { 5250 if (vm_page_grab_sleep(object, m, pindex + i, 5251 "grbmaw", allocflags, true)) { 5252 pctrie_iter_reset(&pages); 5253 goto retrylookup; 5254 } 5255 break; 5256 } 5257 } else { 5258 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 5259 break; 5260 m = vm_page_alloc_after(object, &pages, pindex + i, 5261 pflags | VM_ALLOC_COUNT(count - i), mpred); 5262 if (m == NULL) { 5263 if ((allocflags & (VM_ALLOC_NOWAIT | 5264 VM_ALLOC_WAITFAIL)) != 0) 5265 break; 5266 goto retrylookup; 5267 } 5268 } 5269 if (vm_page_none_valid(m) && 5270 (allocflags & VM_ALLOC_ZERO) != 0) { 5271 if ((m->flags & PG_ZERO) == 0) 5272 pmap_zero_page(m); 5273 vm_page_valid(m); 5274 } 5275 vm_page_grab_release(m, allocflags); 5276 ma[i] = mpred = m; 5277 } 5278 return (i); 5279 } 5280 5281 /* 5282 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 5283 * and will fall back to the locked variant to handle allocation. 5284 */ 5285 int 5286 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 5287 int allocflags, vm_page_t *ma, int count) 5288 { 5289 vm_page_t m; 5290 int flags; 5291 int i; 5292 5293 KASSERT(count > 0, 5294 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 5295 vm_page_grab_check(allocflags); 5296 5297 /* 5298 * Modify flags for lockless acquire to hold the page until we 5299 * set it valid if necessary. 5300 */ 5301 flags = allocflags & ~VM_ALLOC_NOBUSY; 5302 vm_page_grab_check(flags); 5303 m = NULL; 5304 for (i = 0; i < count; i++, pindex++) { 5305 /* 5306 * We may see a false NULL here because the previous page has 5307 * been removed or just inserted and the list is loaded without 5308 * barriers. Switch to radix to verify. 5309 */ 5310 if (m == NULL || QMD_IS_TRASHED(m) || m->pindex != pindex || 5311 atomic_load_ptr(&m->object) != object) { 5312 /* 5313 * This guarantees the result is instantaneously 5314 * correct. 5315 */ 5316 m = NULL; 5317 } 5318 m = vm_page_acquire_unlocked(object, pindex, m, flags); 5319 if (m == PAGE_NOT_ACQUIRED) 5320 return (i); 5321 if (m == NULL) 5322 break; 5323 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 5324 if ((m->flags & PG_ZERO) == 0) 5325 pmap_zero_page(m); 5326 vm_page_valid(m); 5327 } 5328 /* m will still be wired or busy according to flags. */ 5329 vm_page_grab_release(m, allocflags); 5330 ma[i] = m; 5331 m = TAILQ_NEXT(m, listq); 5332 } 5333 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 5334 return (i); 5335 count -= i; 5336 VM_OBJECT_WLOCK(object); 5337 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 5338 VM_OBJECT_WUNLOCK(object); 5339 5340 return (i); 5341 } 5342 5343 /* 5344 * Mapping function for valid or dirty bits in a page. 5345 * 5346 * Inputs are required to range within a page. 5347 */ 5348 vm_page_bits_t 5349 vm_page_bits(int base, int size) 5350 { 5351 int first_bit; 5352 int last_bit; 5353 5354 KASSERT( 5355 base + size <= PAGE_SIZE, 5356 ("vm_page_bits: illegal base/size %d/%d", base, size) 5357 ); 5358 5359 if (size == 0) /* handle degenerate case */ 5360 return (0); 5361 5362 first_bit = base >> DEV_BSHIFT; 5363 last_bit = (base + size - 1) >> DEV_BSHIFT; 5364 5365 return (((vm_page_bits_t)2 << last_bit) - 5366 ((vm_page_bits_t)1 << first_bit)); 5367 } 5368 5369 void 5370 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5371 { 5372 5373 #if PAGE_SIZE == 32768 5374 atomic_set_64((uint64_t *)bits, set); 5375 #elif PAGE_SIZE == 16384 5376 atomic_set_32((uint32_t *)bits, set); 5377 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) 5378 atomic_set_16((uint16_t *)bits, set); 5379 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) 5380 atomic_set_8((uint8_t *)bits, set); 5381 #else /* PAGE_SIZE <= 8192 */ 5382 uintptr_t addr; 5383 int shift; 5384 5385 addr = (uintptr_t)bits; 5386 /* 5387 * Use a trick to perform a 32-bit atomic on the 5388 * containing aligned word, to not depend on the existence 5389 * of atomic_{set, clear}_{8, 16}. 5390 */ 5391 shift = addr & (sizeof(uint32_t) - 1); 5392 #if BYTE_ORDER == BIG_ENDIAN 5393 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5394 #else 5395 shift *= NBBY; 5396 #endif 5397 addr &= ~(sizeof(uint32_t) - 1); 5398 atomic_set_32((uint32_t *)addr, set << shift); 5399 #endif /* PAGE_SIZE */ 5400 } 5401 5402 static inline void 5403 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5404 { 5405 5406 #if PAGE_SIZE == 32768 5407 atomic_clear_64((uint64_t *)bits, clear); 5408 #elif PAGE_SIZE == 16384 5409 atomic_clear_32((uint32_t *)bits, clear); 5410 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) 5411 atomic_clear_16((uint16_t *)bits, clear); 5412 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) 5413 atomic_clear_8((uint8_t *)bits, clear); 5414 #else /* PAGE_SIZE <= 8192 */ 5415 uintptr_t addr; 5416 int shift; 5417 5418 addr = (uintptr_t)bits; 5419 /* 5420 * Use a trick to perform a 32-bit atomic on the 5421 * containing aligned word, to not depend on the existence 5422 * of atomic_{set, clear}_{8, 16}. 5423 */ 5424 shift = addr & (sizeof(uint32_t) - 1); 5425 #if BYTE_ORDER == BIG_ENDIAN 5426 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5427 #else 5428 shift *= NBBY; 5429 #endif 5430 addr &= ~(sizeof(uint32_t) - 1); 5431 atomic_clear_32((uint32_t *)addr, clear << shift); 5432 #endif /* PAGE_SIZE */ 5433 } 5434 5435 static inline vm_page_bits_t 5436 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5437 { 5438 #if PAGE_SIZE == 32768 5439 uint64_t old; 5440 5441 old = *bits; 5442 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5443 return (old); 5444 #elif PAGE_SIZE == 16384 5445 uint32_t old; 5446 5447 old = *bits; 5448 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5449 return (old); 5450 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) 5451 uint16_t old; 5452 5453 old = *bits; 5454 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5455 return (old); 5456 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) 5457 uint8_t old; 5458 5459 old = *bits; 5460 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5461 return (old); 5462 #else /* PAGE_SIZE <= 4096*/ 5463 uintptr_t addr; 5464 uint32_t old, new, mask; 5465 int shift; 5466 5467 addr = (uintptr_t)bits; 5468 /* 5469 * Use a trick to perform a 32-bit atomic on the 5470 * containing aligned word, to not depend on the existence 5471 * of atomic_{set, swap, clear}_{8, 16}. 5472 */ 5473 shift = addr & (sizeof(uint32_t) - 1); 5474 #if BYTE_ORDER == BIG_ENDIAN 5475 shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; 5476 #else 5477 shift *= NBBY; 5478 #endif 5479 addr &= ~(sizeof(uint32_t) - 1); 5480 mask = VM_PAGE_BITS_ALL << shift; 5481 5482 old = *bits; 5483 do { 5484 new = old & ~mask; 5485 new |= newbits << shift; 5486 } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); 5487 return (old >> shift); 5488 #endif /* PAGE_SIZE */ 5489 } 5490 5491 /* 5492 * vm_page_set_valid_range: 5493 * 5494 * Sets portions of a page valid. The arguments are expected 5495 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5496 * of any partial chunks touched by the range. The invalid portion of 5497 * such chunks will be zeroed. 5498 * 5499 * (base + size) must be less then or equal to PAGE_SIZE. 5500 */ 5501 void 5502 vm_page_set_valid_range(vm_page_t m, int base, int size) 5503 { 5504 int endoff, frag; 5505 vm_page_bits_t pagebits; 5506 5507 vm_page_assert_busied(m); 5508 if (size == 0) /* handle degenerate case */ 5509 return; 5510 5511 /* 5512 * If the base is not DEV_BSIZE aligned and the valid 5513 * bit is clear, we have to zero out a portion of the 5514 * first block. 5515 */ 5516 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5517 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5518 pmap_zero_page_area(m, frag, base - frag); 5519 5520 /* 5521 * If the ending offset is not DEV_BSIZE aligned and the 5522 * valid bit is clear, we have to zero out a portion of 5523 * the last block. 5524 */ 5525 endoff = base + size; 5526 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5527 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5528 pmap_zero_page_area(m, endoff, 5529 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5530 5531 /* 5532 * Assert that no previously invalid block that is now being validated 5533 * is already dirty. 5534 */ 5535 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5536 ("vm_page_set_valid_range: page %p is dirty", m)); 5537 5538 /* 5539 * Set valid bits inclusive of any overlap. 5540 */ 5541 pagebits = vm_page_bits(base, size); 5542 if (vm_page_xbusied(m)) 5543 m->valid |= pagebits; 5544 else 5545 vm_page_bits_set(m, &m->valid, pagebits); 5546 } 5547 5548 /* 5549 * Set the page dirty bits and free the invalid swap space if 5550 * present. Returns the previous dirty bits. 5551 */ 5552 vm_page_bits_t 5553 vm_page_set_dirty(vm_page_t m) 5554 { 5555 vm_page_bits_t old; 5556 5557 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5558 5559 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5560 old = m->dirty; 5561 m->dirty = VM_PAGE_BITS_ALL; 5562 } else 5563 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5564 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5565 vm_pager_page_unswapped(m); 5566 5567 return (old); 5568 } 5569 5570 /* 5571 * Clear the given bits from the specified page's dirty field. 5572 */ 5573 static __inline void 5574 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5575 { 5576 5577 vm_page_assert_busied(m); 5578 5579 /* 5580 * If the page is xbusied and not write mapped we are the 5581 * only thread that can modify dirty bits. Otherwise, The pmap 5582 * layer can call vm_page_dirty() without holding a distinguished 5583 * lock. The combination of page busy and atomic operations 5584 * suffice to guarantee consistency of the page dirty field. 5585 */ 5586 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5587 m->dirty &= ~pagebits; 5588 else 5589 vm_page_bits_clear(m, &m->dirty, pagebits); 5590 } 5591 5592 /* 5593 * vm_page_set_validclean: 5594 * 5595 * Sets portions of a page valid and clean. The arguments are expected 5596 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5597 * of any partial chunks touched by the range. The invalid portion of 5598 * such chunks will be zero'd. 5599 * 5600 * (base + size) must be less then or equal to PAGE_SIZE. 5601 */ 5602 void 5603 vm_page_set_validclean(vm_page_t m, int base, int size) 5604 { 5605 vm_page_bits_t oldvalid, pagebits; 5606 int endoff, frag; 5607 5608 vm_page_assert_busied(m); 5609 if (size == 0) /* handle degenerate case */ 5610 return; 5611 5612 /* 5613 * If the base is not DEV_BSIZE aligned and the valid 5614 * bit is clear, we have to zero out a portion of the 5615 * first block. 5616 */ 5617 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5618 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5619 pmap_zero_page_area(m, frag, base - frag); 5620 5621 /* 5622 * If the ending offset is not DEV_BSIZE aligned and the 5623 * valid bit is clear, we have to zero out a portion of 5624 * the last block. 5625 */ 5626 endoff = base + size; 5627 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5628 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5629 pmap_zero_page_area(m, endoff, 5630 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5631 5632 /* 5633 * Set valid, clear dirty bits. If validating the entire 5634 * page we can safely clear the pmap modify bit. We also 5635 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5636 * takes a write fault on a MAP_NOSYNC memory area the flag will 5637 * be set again. 5638 * 5639 * We set valid bits inclusive of any overlap, but we can only 5640 * clear dirty bits for DEV_BSIZE chunks that are fully within 5641 * the range. 5642 */ 5643 oldvalid = m->valid; 5644 pagebits = vm_page_bits(base, size); 5645 if (vm_page_xbusied(m)) 5646 m->valid |= pagebits; 5647 else 5648 vm_page_bits_set(m, &m->valid, pagebits); 5649 #if 0 /* NOT YET */ 5650 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5651 frag = DEV_BSIZE - frag; 5652 base += frag; 5653 size -= frag; 5654 if (size < 0) 5655 size = 0; 5656 } 5657 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5658 #endif 5659 if (base == 0 && size == PAGE_SIZE) { 5660 /* 5661 * The page can only be modified within the pmap if it is 5662 * mapped, and it can only be mapped if it was previously 5663 * fully valid. 5664 */ 5665 if (oldvalid == VM_PAGE_BITS_ALL) 5666 /* 5667 * Perform the pmap_clear_modify() first. Otherwise, 5668 * a concurrent pmap operation, such as 5669 * pmap_protect(), could clear a modification in the 5670 * pmap and set the dirty field on the page before 5671 * pmap_clear_modify() had begun and after the dirty 5672 * field was cleared here. 5673 */ 5674 pmap_clear_modify(m); 5675 m->dirty = 0; 5676 vm_page_aflag_clear(m, PGA_NOSYNC); 5677 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5678 m->dirty &= ~pagebits; 5679 else 5680 vm_page_clear_dirty_mask(m, pagebits); 5681 } 5682 5683 void 5684 vm_page_clear_dirty(vm_page_t m, int base, int size) 5685 { 5686 5687 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5688 } 5689 5690 /* 5691 * vm_page_set_invalid: 5692 * 5693 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5694 * valid and dirty bits for the effected areas are cleared. 5695 */ 5696 void 5697 vm_page_set_invalid(vm_page_t m, int base, int size) 5698 { 5699 vm_page_bits_t bits; 5700 vm_object_t object; 5701 5702 /* 5703 * The object lock is required so that pages can't be mapped 5704 * read-only while we're in the process of invalidating them. 5705 */ 5706 object = m->object; 5707 VM_OBJECT_ASSERT_WLOCKED(object); 5708 vm_page_assert_busied(m); 5709 5710 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5711 size >= object->un_pager.vnp.vnp_size) 5712 bits = VM_PAGE_BITS_ALL; 5713 else 5714 bits = vm_page_bits(base, size); 5715 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5716 pmap_remove_all(m); 5717 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5718 !pmap_page_is_mapped(m), 5719 ("vm_page_set_invalid: page %p is mapped", m)); 5720 if (vm_page_xbusied(m)) { 5721 m->valid &= ~bits; 5722 m->dirty &= ~bits; 5723 } else { 5724 vm_page_bits_clear(m, &m->valid, bits); 5725 vm_page_bits_clear(m, &m->dirty, bits); 5726 } 5727 } 5728 5729 /* 5730 * vm_page_invalid: 5731 * 5732 * Invalidates the entire page. The page must be busy, unmapped, and 5733 * the enclosing object must be locked. The object locks protects 5734 * against concurrent read-only pmap enter which is done without 5735 * busy. 5736 */ 5737 void 5738 vm_page_invalid(vm_page_t m) 5739 { 5740 5741 vm_page_assert_busied(m); 5742 VM_OBJECT_ASSERT_WLOCKED(m->object); 5743 MPASS(!pmap_page_is_mapped(m)); 5744 5745 if (vm_page_xbusied(m)) 5746 m->valid = 0; 5747 else 5748 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5749 } 5750 5751 /* 5752 * vm_page_zero_invalid() 5753 * 5754 * The kernel assumes that the invalid portions of a page contain 5755 * garbage, but such pages can be mapped into memory by user code. 5756 * When this occurs, we must zero out the non-valid portions of the 5757 * page so user code sees what it expects. 5758 * 5759 * Pages are most often semi-valid when the end of a file is mapped 5760 * into memory and the file's size is not page aligned. 5761 */ 5762 void 5763 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5764 { 5765 int b; 5766 int i; 5767 5768 /* 5769 * Scan the valid bits looking for invalid sections that 5770 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5771 * valid bit may be set ) have already been zeroed by 5772 * vm_page_set_validclean(). 5773 */ 5774 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5775 if (i == (PAGE_SIZE / DEV_BSIZE) || 5776 (m->valid & ((vm_page_bits_t)1 << i))) { 5777 if (i > b) { 5778 pmap_zero_page_area(m, 5779 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5780 } 5781 b = i + 1; 5782 } 5783 } 5784 5785 /* 5786 * setvalid is TRUE when we can safely set the zero'd areas 5787 * as being valid. We can do this if there are no cache consistency 5788 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5789 */ 5790 if (setvalid) 5791 vm_page_valid(m); 5792 } 5793 5794 /* 5795 * vm_page_is_valid: 5796 * 5797 * Is (partial) page valid? Note that the case where size == 0 5798 * will return FALSE in the degenerate case where the page is 5799 * entirely invalid, and TRUE otherwise. 5800 * 5801 * Some callers envoke this routine without the busy lock held and 5802 * handle races via higher level locks. Typical callers should 5803 * hold a busy lock to prevent invalidation. 5804 */ 5805 int 5806 vm_page_is_valid(vm_page_t m, int base, int size) 5807 { 5808 vm_page_bits_t bits; 5809 5810 bits = vm_page_bits(base, size); 5811 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5812 } 5813 5814 /* 5815 * Returns true if all of the specified predicates are true for the entire 5816 * (super)page and false otherwise. 5817 */ 5818 bool 5819 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m) 5820 { 5821 vm_object_t object; 5822 int i, npages; 5823 5824 object = m->object; 5825 if (skip_m != NULL && skip_m->object != object) 5826 return (false); 5827 VM_OBJECT_ASSERT_LOCKED(object); 5828 KASSERT(psind <= m->psind, 5829 ("psind %d > psind %d of m %p", psind, m->psind, m)); 5830 npages = atop(pagesizes[psind]); 5831 5832 /* 5833 * The physically contiguous pages that make up a superpage, i.e., a 5834 * page with a page size index ("psind") greater than zero, will 5835 * occupy adjacent entries in vm_page_array[]. 5836 */ 5837 for (i = 0; i < npages; i++) { 5838 /* Always test object consistency, including "skip_m". */ 5839 if (m[i].object != object) 5840 return (false); 5841 if (&m[i] == skip_m) 5842 continue; 5843 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5844 return (false); 5845 if ((flags & PS_ALL_DIRTY) != 0) { 5846 /* 5847 * Calling vm_page_test_dirty() or pmap_is_modified() 5848 * might stop this case from spuriously returning 5849 * "false". However, that would require a write lock 5850 * on the object containing "m[i]". 5851 */ 5852 if (m[i].dirty != VM_PAGE_BITS_ALL) 5853 return (false); 5854 } 5855 if ((flags & PS_ALL_VALID) != 0 && 5856 m[i].valid != VM_PAGE_BITS_ALL) 5857 return (false); 5858 } 5859 return (true); 5860 } 5861 5862 /* 5863 * Set the page's dirty bits if the page is modified. 5864 */ 5865 void 5866 vm_page_test_dirty(vm_page_t m) 5867 { 5868 5869 vm_page_assert_busied(m); 5870 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5871 vm_page_dirty(m); 5872 } 5873 5874 void 5875 vm_page_valid(vm_page_t m) 5876 { 5877 5878 vm_page_assert_busied(m); 5879 if (vm_page_xbusied(m)) 5880 m->valid = VM_PAGE_BITS_ALL; 5881 else 5882 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5883 } 5884 5885 void 5886 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 5887 { 5888 5889 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 5890 } 5891 5892 void 5893 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 5894 { 5895 5896 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 5897 } 5898 5899 int 5900 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 5901 { 5902 5903 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 5904 } 5905 5906 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 5907 void 5908 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 5909 { 5910 5911 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 5912 } 5913 5914 void 5915 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 5916 { 5917 5918 mtx_assert_(vm_page_lockptr(m), a, file, line); 5919 } 5920 #endif 5921 5922 #ifdef INVARIANTS 5923 void 5924 vm_page_object_busy_assert(vm_page_t m) 5925 { 5926 5927 /* 5928 * Certain of the page's fields may only be modified by the 5929 * holder of a page or object busy. 5930 */ 5931 if (m->object != NULL && !vm_page_busied(m)) 5932 VM_OBJECT_ASSERT_BUSY(m->object); 5933 } 5934 5935 void 5936 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5937 { 5938 5939 if ((bits & PGA_WRITEABLE) == 0) 5940 return; 5941 5942 /* 5943 * The PGA_WRITEABLE flag can only be set if the page is 5944 * managed, is exclusively busied or the object is locked. 5945 * Currently, this flag is only set by pmap_enter(). 5946 */ 5947 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5948 ("PGA_WRITEABLE on unmanaged page")); 5949 if (!vm_page_xbusied(m)) 5950 VM_OBJECT_ASSERT_BUSY(m->object); 5951 } 5952 #endif 5953 5954 #include "opt_ddb.h" 5955 #ifdef DDB 5956 #include <sys/kernel.h> 5957 5958 #include <ddb/ddb.h> 5959 5960 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5961 { 5962 5963 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5964 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5965 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5966 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5967 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5968 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5969 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5970 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5971 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5972 } 5973 5974 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5975 { 5976 int dom; 5977 5978 db_printf("pq_free %d\n", vm_free_count()); 5979 for (dom = 0; dom < vm_ndomains; dom++) { 5980 db_printf( 5981 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5982 dom, 5983 vm_dom[dom].vmd_page_count, 5984 vm_dom[dom].vmd_free_count, 5985 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5986 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5987 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5988 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5989 } 5990 } 5991 5992 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5993 { 5994 vm_page_t m; 5995 boolean_t phys, virt; 5996 5997 if (!have_addr) { 5998 db_printf("show pginfo addr\n"); 5999 return; 6000 } 6001 6002 phys = strchr(modif, 'p') != NULL; 6003 virt = strchr(modif, 'v') != NULL; 6004 if (virt) 6005 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 6006 else if (phys) 6007 m = PHYS_TO_VM_PAGE(addr); 6008 else 6009 m = (vm_page_t)addr; 6010 db_printf( 6011 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 6012 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 6013 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 6014 m->a.queue, m->ref_count, m->a.flags, m->oflags, 6015 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 6016 } 6017 #endif /* DDB */ 6018