1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * The Mach Operating System project at Carnegie-Mellon University. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 3. Neither the name of the University nor the names of its contributors 20 * may be used to endorse or promote products derived from this software 21 * without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 36 /*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Resident memory management module. 65 */ 66 67 #include <sys/cdefs.h> 68 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/counter.h> 73 #include <sys/domainset.h> 74 #include <sys/kernel.h> 75 #include <sys/limits.h> 76 #include <sys/linker.h> 77 #include <sys/lock.h> 78 #include <sys/malloc.h> 79 #include <sys/mman.h> 80 #include <sys/msgbuf.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> 83 #include <sys/rwlock.h> 84 #include <sys/sleepqueue.h> 85 #include <sys/sbuf.h> 86 #include <sys/sched.h> 87 #include <sys/sf_buf.h> 88 #include <sys/smp.h> 89 #include <sys/sysctl.h> 90 #include <sys/vmmeter.h> 91 #include <sys/vnode.h> 92 93 #include <vm/vm.h> 94 #include <vm/pmap.h> 95 #include <vm/vm_param.h> 96 #include <vm/vm_domainset.h> 97 #include <vm/vm_kern.h> 98 #include <vm/vm_map.h> 99 #include <vm/vm_object.h> 100 #include <vm/vm_page.h> 101 #include <vm/vm_pageout.h> 102 #include <vm/vm_phys.h> 103 #include <vm/vm_pagequeue.h> 104 #include <vm/vm_pager.h> 105 #include <vm/vm_radix.h> 106 #include <vm/vm_reserv.h> 107 #include <vm/vm_extern.h> 108 #include <vm/vm_dumpset.h> 109 #include <vm/uma.h> 110 #include <vm/uma_int.h> 111 112 #include <machine/md_var.h> 113 #if defined(__aarch64__) 114 #include <machine/pmap.h> 115 #include <machine/rsi.h> 116 #endif 117 118 struct vm_domain vm_dom[MAXMEMDOM]; 119 120 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 121 122 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 123 /* The following fields are protected by the domainset lock. */ 124 domainset_t __exclusive_cache_line vm_min_domains; 125 domainset_t __exclusive_cache_line vm_severe_domains; 126 static int vm_min_waiters; 127 static int vm_severe_waiters; 128 static int vm_pageproc_waiters; 129 130 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 131 "VM page statistics"); 132 133 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries); 134 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, 135 CTLFLAG_RD, &pqstate_commit_retries, 136 "Number of failed per-page atomic queue state updates"); 137 138 static COUNTER_U64_DEFINE_EARLY(queue_ops); 139 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, 140 CTLFLAG_RD, &queue_ops, 141 "Number of batched queue operations"); 142 143 static COUNTER_U64_DEFINE_EARLY(queue_nops); 144 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, 145 CTLFLAG_RD, &queue_nops, 146 "Number of batched queue operations with no effects"); 147 148 static unsigned long nofreeq_size; 149 SYSCTL_ULONG(_vm_stats_page, OID_AUTO, nofreeq_size, CTLFLAG_RD, 150 &nofreeq_size, 0, 151 "Size of the nofree queue"); 152 153 #ifdef INVARIANTS 154 bool vm_check_pg_zero = false; 155 SYSCTL_BOOL(_debug, OID_AUTO, vm_check_pg_zero, CTLFLAG_RWTUN, 156 &vm_check_pg_zero, 0, 157 "verify content of freed zero-filled pages"); 158 #endif 159 160 /* 161 * bogus page -- for I/O to/from partially complete buffers, 162 * or for paging into sparsely invalid regions. 163 */ 164 vm_page_t bogus_page; 165 166 vm_page_t vm_page_array; 167 long vm_page_array_size; 168 long first_page; 169 170 struct bitset *vm_page_dump; 171 long vm_page_dump_pages; 172 173 static TAILQ_HEAD(, vm_page) blacklist_head; 174 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 175 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 176 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 177 178 static uma_zone_t fakepg_zone; 179 180 static void vm_page_alloc_check(vm_page_t m); 181 static vm_page_t vm_page_alloc_nofree_domain(int domain, int req); 182 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, 183 vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked); 184 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 185 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 186 static bool vm_page_free_prep(vm_page_t m); 187 static void vm_page_free_toq(vm_page_t m); 188 static void vm_page_init(void *dummy); 189 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object); 190 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, 191 const uint16_t nflag); 192 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 193 vm_page_t m_run, vm_paddr_t high); 194 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); 195 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 196 int req); 197 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, 198 int flags); 199 static void vm_page_zone_release(void *arg, void **store, int cnt); 200 201 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 202 203 static void 204 vm_page_init(void *dummy) 205 { 206 207 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 208 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 209 bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_NOFREE); 210 } 211 212 static int pgcache_zone_max_pcpu; 213 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu, 214 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0, 215 "Per-CPU page cache size"); 216 217 /* 218 * The cache page zone is initialized later since we need to be able to allocate 219 * pages before UMA is fully initialized. 220 */ 221 static void 222 vm_page_init_cache_zones(void *dummy __unused) 223 { 224 struct vm_domain *vmd; 225 struct vm_pgcache *pgcache; 226 int cache, domain, maxcache, pool; 227 228 TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu); 229 maxcache = pgcache_zone_max_pcpu * mp_ncpus; 230 for (domain = 0; domain < vm_ndomains; domain++) { 231 vmd = VM_DOMAIN(domain); 232 for (pool = 0; pool < VM_NFREEPOOL; pool++) { 233 #ifdef VM_FREEPOOL_LAZYINIT 234 if (pool == VM_FREEPOOL_LAZYINIT) 235 continue; 236 #endif 237 pgcache = &vmd->vmd_pgcache[pool]; 238 pgcache->domain = domain; 239 pgcache->pool = pool; 240 pgcache->zone = uma_zcache_create("vm pgcache", 241 PAGE_SIZE, NULL, NULL, NULL, NULL, 242 vm_page_zone_import, vm_page_zone_release, pgcache, 243 UMA_ZONE_VM); 244 245 /* 246 * Limit each pool's zone to 0.1% of the pages in the 247 * domain. 248 */ 249 cache = maxcache != 0 ? maxcache : 250 vmd->vmd_page_count / 1000; 251 uma_zone_set_maxcache(pgcache->zone, cache); 252 } 253 } 254 } 255 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 256 257 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 258 #if PAGE_SIZE == 32768 259 #ifdef CTASSERT 260 CTASSERT(sizeof(u_long) >= 8); 261 #endif 262 #endif 263 264 /* 265 * vm_set_page_size: 266 * 267 * Sets the page size, perhaps based upon the memory 268 * size. Must be called before any use of page-size 269 * dependent functions. 270 */ 271 void 272 vm_set_page_size(void) 273 { 274 if (vm_cnt.v_page_size == 0) 275 vm_cnt.v_page_size = PAGE_SIZE; 276 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 277 panic("vm_set_page_size: page size not a power of two"); 278 } 279 280 /* 281 * vm_page_blacklist_next: 282 * 283 * Find the next entry in the provided string of blacklist 284 * addresses. Entries are separated by space, comma, or newline. 285 * If an invalid integer is encountered then the rest of the 286 * string is skipped. Updates the list pointer to the next 287 * character, or NULL if the string is exhausted or invalid. 288 */ 289 static vm_paddr_t 290 vm_page_blacklist_next(char **list, char *end) 291 { 292 vm_paddr_t bad; 293 char *cp, *pos; 294 295 if (list == NULL || *list == NULL) 296 return (0); 297 if (**list =='\0') { 298 *list = NULL; 299 return (0); 300 } 301 302 /* 303 * If there's no end pointer then the buffer is coming from 304 * the kenv and we know it's null-terminated. 305 */ 306 if (end == NULL) 307 end = *list + strlen(*list); 308 309 /* Ensure that strtoq() won't walk off the end */ 310 if (*end != '\0') { 311 if (*end == '\n' || *end == ' ' || *end == ',') 312 *end = '\0'; 313 else { 314 printf("Blacklist not terminated, skipping\n"); 315 *list = NULL; 316 return (0); 317 } 318 } 319 320 for (pos = *list; *pos != '\0'; pos = cp) { 321 bad = strtoq(pos, &cp, 0); 322 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 323 if (bad == 0) { 324 if (++cp < end) 325 continue; 326 else 327 break; 328 } 329 } else 330 break; 331 if (*cp == '\0' || ++cp >= end) 332 *list = NULL; 333 else 334 *list = cp; 335 return (trunc_page(bad)); 336 } 337 printf("Garbage in RAM blacklist, skipping\n"); 338 *list = NULL; 339 return (0); 340 } 341 342 bool 343 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 344 { 345 struct vm_domain *vmd; 346 vm_page_t m; 347 bool found; 348 349 m = vm_phys_paddr_to_vm_page(pa); 350 if (m == NULL) 351 return (true); /* page does not exist, no failure */ 352 353 vmd = VM_DOMAIN(vm_phys_domain(pa)); 354 vm_domain_free_lock(vmd); 355 found = vm_phys_unfree_page(pa); 356 vm_domain_free_unlock(vmd); 357 if (found) { 358 vm_domain_freecnt_inc(vmd, -1); 359 TAILQ_INSERT_TAIL(&blacklist_head, m, plinks.q); 360 if (verbose) 361 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 362 } 363 return (found); 364 } 365 366 /* 367 * vm_page_blacklist_check: 368 * 369 * Iterate through the provided string of blacklist addresses, pulling 370 * each entry out of the physical allocator free list and putting it 371 * onto a list for reporting via the vm.page_blacklist sysctl. 372 */ 373 static void 374 vm_page_blacklist_check(char *list, char *end) 375 { 376 vm_paddr_t pa; 377 char *next; 378 379 next = list; 380 while (next != NULL) { 381 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 382 continue; 383 vm_page_blacklist_add(pa, bootverbose); 384 } 385 } 386 387 /* 388 * vm_page_blacklist_load: 389 * 390 * Search for a special module named "ram_blacklist". It'll be a 391 * plain text file provided by the user via the loader directive 392 * of the same name. 393 */ 394 static void 395 vm_page_blacklist_load(char **list, char **end) 396 { 397 void *mod; 398 u_char *ptr; 399 u_int len; 400 401 mod = NULL; 402 ptr = NULL; 403 404 mod = preload_search_by_type("ram_blacklist"); 405 if (mod != NULL) { 406 ptr = preload_fetch_addr(mod); 407 len = preload_fetch_size(mod); 408 } 409 410 if (ptr != NULL && len > 0) { 411 *list = ptr; 412 *end = ptr + len - 1; 413 } else { 414 *list = NULL; 415 *end = NULL; 416 } 417 418 return; 419 } 420 421 static int 422 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 423 { 424 vm_page_t m; 425 struct sbuf sbuf; 426 int error, first; 427 428 first = 1; 429 error = sysctl_wire_old_buffer(req, 0); 430 if (error != 0) 431 return (error); 432 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 433 TAILQ_FOREACH(m, &blacklist_head, plinks.q) { 434 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 435 (uintmax_t)m->phys_addr); 436 first = 0; 437 } 438 error = sbuf_finish(&sbuf); 439 sbuf_delete(&sbuf); 440 return (error); 441 } 442 443 /* 444 * Initialize a dummy page for use in scans of the specified paging queue. 445 * In principle, this function only needs to set the flag PG_MARKER. 446 * Nonetheless, it write busies the page as a safety precaution. 447 */ 448 void 449 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) 450 { 451 452 bzero(marker, sizeof(*marker)); 453 marker->flags = PG_MARKER; 454 marker->a.flags = aflags; 455 marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 456 marker->a.queue = queue; 457 } 458 459 static void 460 vm_page_domain_init(int domain) 461 { 462 struct vm_domain *vmd; 463 struct vm_pagequeue *pq; 464 int i; 465 466 vmd = VM_DOMAIN(domain); 467 bzero(vmd, sizeof(*vmd)); 468 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 469 "vm inactive pagequeue"; 470 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 471 "vm active pagequeue"; 472 *__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 473 "vm laundry pagequeue"; 474 *__DECONST(const char **, 475 &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 476 "vm unswappable pagequeue"; 477 vmd->vmd_domain = domain; 478 vmd->vmd_page_count = 0; 479 vmd->vmd_free_count = 0; 480 vmd->vmd_segs = 0; 481 vmd->vmd_oom = false; 482 vmd->vmd_helper_threads_enabled = true; 483 for (i = 0; i < PQ_COUNT; i++) { 484 pq = &vmd->vmd_pagequeues[i]; 485 TAILQ_INIT(&pq->pq_pl); 486 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 487 MTX_DEF | MTX_DUPOK); 488 pq->pq_pdpages = 0; 489 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 490 } 491 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 492 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 493 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 494 495 /* 496 * inacthead is used to provide FIFO ordering for LRU-bypassing 497 * insertions. 498 */ 499 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 500 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 501 &vmd->vmd_inacthead, plinks.q); 502 503 /* 504 * The clock pages are used to implement active queue scanning without 505 * requeues. Scans start at clock[0], which is advanced after the scan 506 * ends. When the two clock hands meet, they are reset and scanning 507 * resumes from the head of the queue. 508 */ 509 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 510 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 511 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 512 &vmd->vmd_clock[0], plinks.q); 513 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 514 &vmd->vmd_clock[1], plinks.q); 515 } 516 517 /* 518 * Initialize a physical page in preparation for adding it to the free 519 * lists. 520 */ 521 void 522 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind, int pool) 523 { 524 m->object = NULL; 525 m->ref_count = 0; 526 m->busy_lock = VPB_FREED; 527 m->flags = m->a.flags = 0; 528 m->phys_addr = pa; 529 m->a.queue = PQ_NONE; 530 m->psind = 0; 531 m->segind = segind; 532 m->order = VM_NFREEORDER; 533 m->pool = pool; 534 m->valid = m->dirty = 0; 535 pmap_page_init(m); 536 } 537 538 #ifndef PMAP_HAS_PAGE_ARRAY 539 static vm_paddr_t 540 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) 541 { 542 vm_paddr_t new_end; 543 544 /* 545 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 546 * However, because this page is allocated from KVM, out-of-bounds 547 * accesses using the direct map will not be trapped. 548 */ 549 *vaddr += PAGE_SIZE; 550 551 /* 552 * Allocate physical memory for the page structures, and map it. 553 */ 554 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 555 vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, 556 VM_PROT_READ | VM_PROT_WRITE); 557 vm_page_array_size = page_range; 558 559 return (new_end); 560 } 561 #endif 562 563 /* 564 * vm_page_startup: 565 * 566 * Initializes the resident memory module. Allocates physical memory for 567 * bootstrapping UMA and some data structures that are used to manage 568 * physical pages. Initializes these structures, and populates the free 569 * page queues. 570 */ 571 vm_offset_t 572 vm_page_startup(vm_offset_t vaddr) 573 { 574 struct vm_phys_seg *seg; 575 struct vm_domain *vmd; 576 vm_page_t m; 577 char *list, *listend; 578 vm_paddr_t end, high_avail, low_avail, new_end, size; 579 vm_paddr_t page_range __unused; 580 vm_paddr_t last_pa, pa, startp, endp; 581 u_long pagecount; 582 #if MINIDUMP_PAGE_TRACKING 583 u_long vm_page_dump_size; 584 #endif 585 int biggestone, i, segind; 586 #ifdef WITNESS 587 void *mapped; 588 u_long witness_size; 589 #endif 590 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 591 long ii; 592 #endif 593 int pool; 594 #ifdef VM_FREEPOOL_LAZYINIT 595 int lazyinit; 596 #endif 597 598 vaddr = round_page(vaddr); 599 600 vm_phys_early_startup(); 601 biggestone = vm_phys_avail_largest(); 602 end = phys_avail[biggestone+1]; 603 604 /* 605 * Initialize the page and queue locks. 606 */ 607 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 608 for (i = 0; i < vm_ndomains; i++) 609 vm_page_domain_init(i); 610 611 new_end = end; 612 #ifdef WITNESS 613 /* 614 * witness(4) support. Allocate and map memory and initialize. 615 * Advertised available memory is limited in order to avoid witness 616 * misconfiguration consuming memory needed for subsequent essential 617 * allocations. 618 */ 619 witness_size = round_page(witness_startup_count( 620 trunc_page((new_end - phys_avail[biggestone]) / 2))); 621 new_end -= witness_size; 622 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 623 VM_PROT_READ | VM_PROT_WRITE); 624 bzero(mapped, witness_size); 625 witness_startup(mapped); 626 #endif 627 628 #if MINIDUMP_PAGE_TRACKING 629 /* 630 * Allocate a bitmap to indicate that a random physical page 631 * needs to be included in a minidump. 632 * 633 * The amd64 port needs this to indicate which direct map pages 634 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 635 * 636 * However, i386 still needs this workspace internally within the 637 * minidump code. In theory, they are not needed on i386, but are 638 * included should the sf_buf code decide to use them. 639 */ 640 last_pa = 0; 641 vm_page_dump_pages = 0; 642 for (i = 0; dump_avail[i + 1] != 0; i += 2) { 643 vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) - 644 dump_avail[i] / PAGE_SIZE; 645 if (dump_avail[i + 1] > last_pa) 646 last_pa = dump_avail[i + 1]; 647 } 648 vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages)); 649 new_end -= vm_page_dump_size; 650 vm_page_dump = pmap_map(&vaddr, new_end, 651 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 652 bzero((void *)vm_page_dump, vm_page_dump_size); 653 #if MINIDUMP_STARTUP_PAGE_TRACKING 654 /* 655 * Include the UMA bootstrap pages, witness pages and vm_page_dump 656 * in a crash dump. When pmap_map() uses the direct map, they are 657 * not automatically included. 658 */ 659 for (pa = new_end; pa < end; pa += PAGE_SIZE) 660 dump_add_page(pa); 661 #endif 662 #else 663 (void)last_pa; 664 #endif 665 phys_avail[biggestone + 1] = new_end; 666 #ifdef __amd64__ 667 /* 668 * Request that the physical pages underlying the message buffer be 669 * included in a crash dump. Since the message buffer is accessed 670 * through the direct map, they are not automatically included. 671 */ 672 pa = DMAP_TO_PHYS(msgbufp->msg_ptr); 673 last_pa = pa + round_page(msgbufsize); 674 while (pa < last_pa) { 675 dump_add_page(pa); 676 pa += PAGE_SIZE; 677 } 678 #else 679 (void)pa; 680 #endif 681 682 /* 683 * Determine the lowest and highest physical addresses and, in the case 684 * of VM_PHYSSEG_SPARSE, the exact size of the available physical 685 * memory. vm_phys_early_startup() already checked that phys_avail[] 686 * has at least one element. 687 */ 688 #ifdef VM_PHYSSEG_SPARSE 689 size = phys_avail[1] - phys_avail[0]; 690 #endif 691 low_avail = phys_avail[0]; 692 high_avail = phys_avail[1]; 693 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 694 #ifdef VM_PHYSSEG_SPARSE 695 size += phys_avail[i + 1] - phys_avail[i]; 696 #endif 697 if (phys_avail[i] < low_avail) 698 low_avail = phys_avail[i]; 699 if (phys_avail[i + 1] > high_avail) 700 high_avail = phys_avail[i + 1]; 701 } 702 for (i = 0; i < vm_phys_nsegs; i++) { 703 #ifdef VM_PHYSSEG_SPARSE 704 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 705 #endif 706 if (vm_phys_segs[i].start < low_avail) 707 low_avail = vm_phys_segs[i].start; 708 if (vm_phys_segs[i].end > high_avail) 709 high_avail = vm_phys_segs[i].end; 710 } 711 first_page = low_avail / PAGE_SIZE; 712 #ifdef VM_PHYSSEG_DENSE 713 size = high_avail - low_avail; 714 #endif 715 716 #ifdef PMAP_HAS_PAGE_ARRAY 717 pmap_page_array_startup(size / PAGE_SIZE); 718 biggestone = vm_phys_avail_largest(); 719 end = new_end = phys_avail[biggestone + 1]; 720 #else 721 #ifdef VM_PHYSSEG_DENSE 722 /* 723 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 724 * the overhead of a page structure per page only if vm_page_array is 725 * allocated from the last physical memory chunk. Otherwise, we must 726 * allocate page structures representing the physical memory 727 * underlying vm_page_array, even though they will not be used. 728 */ 729 if (new_end != high_avail) 730 page_range = size / PAGE_SIZE; 731 else 732 #endif 733 { 734 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 735 736 /* 737 * If the partial bytes remaining are large enough for 738 * a page (PAGE_SIZE) without a corresponding 739 * 'struct vm_page', then new_end will contain an 740 * extra page after subtracting the length of the VM 741 * page array. Compensate by subtracting an extra 742 * page from new_end. 743 */ 744 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 745 if (new_end == high_avail) 746 high_avail -= PAGE_SIZE; 747 new_end -= PAGE_SIZE; 748 } 749 } 750 end = new_end; 751 new_end = vm_page_array_alloc(&vaddr, end, page_range); 752 #endif 753 754 #if VM_NRESERVLEVEL > 0 755 /* 756 * Allocate physical memory for the reservation management system's 757 * data structures, and map it. 758 */ 759 new_end = vm_reserv_startup(&vaddr, new_end); 760 #endif 761 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 762 /* 763 * Include vm_page_array and vm_reserv_array in a crash dump. 764 */ 765 for (pa = new_end; pa < end; pa += PAGE_SIZE) 766 dump_add_page(pa); 767 #endif 768 phys_avail[biggestone + 1] = new_end; 769 770 /* 771 * Add physical memory segments corresponding to the available 772 * physical pages. 773 */ 774 for (i = 0; phys_avail[i + 1] != 0; i += 2) 775 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 776 777 /* 778 * Initialize the physical memory allocator. 779 */ 780 vm_phys_init(); 781 782 pool = VM_FREEPOOL_DEFAULT; 783 #ifdef VM_FREEPOOL_LAZYINIT 784 lazyinit = 1; 785 TUNABLE_INT_FETCH("debug.vm.lazy_page_init", &lazyinit); 786 if (lazyinit) 787 pool = VM_FREEPOOL_LAZYINIT; 788 #endif 789 790 /* 791 * Initialize the page structures and add every available page to the 792 * physical memory allocator's free lists. 793 */ 794 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 795 for (ii = 0; ii < vm_page_array_size; ii++) { 796 m = &vm_page_array[ii]; 797 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0, 798 VM_FREEPOOL_DEFAULT); 799 m->flags = PG_FICTITIOUS; 800 } 801 #endif 802 vm_cnt.v_page_count = 0; 803 for (segind = 0; segind < vm_phys_nsegs; segind++) { 804 seg = &vm_phys_segs[segind]; 805 806 /* 807 * Initialize pages not covered by phys_avail[], since they 808 * might be freed to the allocator at some future point, e.g., 809 * by kmem_bootstrap_free(). 810 */ 811 startp = seg->start; 812 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 813 if (startp >= seg->end) 814 break; 815 if (phys_avail[i + 1] < startp) 816 continue; 817 if (phys_avail[i] <= startp) { 818 startp = phys_avail[i + 1]; 819 continue; 820 } 821 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 822 for (endp = MIN(phys_avail[i], seg->end); 823 startp < endp; startp += PAGE_SIZE, m++) { 824 vm_page_init_page(m, startp, segind, 825 VM_FREEPOOL_DEFAULT); 826 } 827 } 828 829 /* 830 * Add the segment's pages that are covered by one of 831 * phys_avail's ranges to the free lists. 832 */ 833 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 834 if (seg->end <= phys_avail[i] || 835 seg->start >= phys_avail[i + 1]) 836 continue; 837 838 startp = MAX(seg->start, phys_avail[i]); 839 endp = MIN(seg->end, phys_avail[i + 1]); 840 pagecount = (u_long)atop(endp - startp); 841 if (pagecount == 0) 842 continue; 843 844 /* 845 * If lazy vm_page initialization is not enabled, simply 846 * initialize all of the pages in the segment covered by 847 * phys_avail. Otherwise, initialize only the first 848 * page of each run of free pages handed to the vm_phys 849 * allocator, which in turn defers initialization of 850 * pages until they are needed. 851 * 852 * This avoids blocking the boot process for long 853 * periods, which may be relevant for VMs (which ought 854 * to boot as quickly as possible) and/or systems with 855 * large amounts of physical memory. 856 */ 857 m = vm_phys_seg_paddr_to_vm_page(seg, startp); 858 vm_page_init_page(m, startp, segind, pool); 859 if (pool == VM_FREEPOOL_DEFAULT) { 860 for (u_long j = 1; j < pagecount; j++) { 861 vm_page_init_page(&m[j], 862 startp + ptoa((vm_paddr_t)j), 863 segind, pool); 864 } 865 } 866 vmd = VM_DOMAIN(seg->domain); 867 vm_domain_free_lock(vmd); 868 vm_phys_enqueue_contig(m, pool, pagecount); 869 vm_domain_free_unlock(vmd); 870 vm_domain_freecnt_inc(vmd, pagecount); 871 vm_cnt.v_page_count += (u_int)pagecount; 872 vmd->vmd_page_count += (u_int)pagecount; 873 vmd->vmd_segs |= 1UL << segind; 874 } 875 } 876 877 /* 878 * Remove blacklisted pages from the physical memory allocator. 879 */ 880 TAILQ_INIT(&blacklist_head); 881 vm_page_blacklist_load(&list, &listend); 882 vm_page_blacklist_check(list, listend); 883 884 list = kern_getenv("vm.blacklist"); 885 vm_page_blacklist_check(list, NULL); 886 887 freeenv(list); 888 #if VM_NRESERVLEVEL > 0 889 /* 890 * Initialize the reservation management system. 891 */ 892 vm_reserv_init(); 893 #endif 894 895 return (vaddr); 896 } 897 898 void 899 vm_page_reference(vm_page_t m) 900 { 901 902 vm_page_aflag_set(m, PGA_REFERENCED); 903 } 904 905 /* 906 * vm_page_trybusy 907 * 908 * Helper routine for grab functions to trylock busy. 909 * 910 * Returns true on success and false on failure. 911 */ 912 static bool 913 vm_page_trybusy(vm_page_t m, int allocflags) 914 { 915 916 if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) 917 return (vm_page_trysbusy(m)); 918 else 919 return (vm_page_tryxbusy(m)); 920 } 921 922 /* 923 * vm_page_tryacquire 924 * 925 * Helper routine for grab functions to trylock busy and wire. 926 * 927 * Returns true on success and false on failure. 928 */ 929 static inline bool 930 vm_page_tryacquire(vm_page_t m, int allocflags) 931 { 932 bool locked; 933 934 locked = vm_page_trybusy(m, allocflags); 935 if (locked && (allocflags & VM_ALLOC_WIRED) != 0) 936 vm_page_wire(m); 937 return (locked); 938 } 939 940 /* 941 * vm_page_busy_acquire: 942 * 943 * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop 944 * and drop the object lock if necessary. 945 */ 946 bool 947 vm_page_busy_acquire(vm_page_t m, int allocflags) 948 { 949 vm_object_t obj; 950 bool locked; 951 952 /* 953 * The page-specific object must be cached because page 954 * identity can change during the sleep, causing the 955 * re-lock of a different object. 956 * It is assumed that a reference to the object is already 957 * held by the callers. 958 */ 959 obj = atomic_load_ptr(&m->object); 960 for (;;) { 961 if (vm_page_tryacquire(m, allocflags)) 962 return (true); 963 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 964 return (false); 965 if (obj != NULL) 966 locked = VM_OBJECT_WOWNED(obj); 967 else 968 locked = false; 969 MPASS(locked || vm_page_wired(m)); 970 if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags, 971 locked) && locked) 972 VM_OBJECT_WLOCK(obj); 973 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 974 return (false); 975 KASSERT(m->object == obj || m->object == NULL, 976 ("vm_page_busy_acquire: page %p does not belong to %p", 977 m, obj)); 978 } 979 } 980 981 /* 982 * vm_page_busy_downgrade: 983 * 984 * Downgrade an exclusive busy page into a single shared busy page. 985 */ 986 void 987 vm_page_busy_downgrade(vm_page_t m) 988 { 989 u_int x; 990 991 vm_page_assert_xbusied(m); 992 993 x = vm_page_busy_fetch(m); 994 for (;;) { 995 if (atomic_fcmpset_rel_int(&m->busy_lock, 996 &x, VPB_SHARERS_WORD(1))) 997 break; 998 } 999 if ((x & VPB_BIT_WAITERS) != 0) 1000 wakeup(m); 1001 } 1002 1003 /* 1004 * 1005 * vm_page_busy_tryupgrade: 1006 * 1007 * Attempt to upgrade a single shared busy into an exclusive busy. 1008 */ 1009 int 1010 vm_page_busy_tryupgrade(vm_page_t m) 1011 { 1012 u_int ce, x; 1013 1014 vm_page_assert_sbusied(m); 1015 1016 x = vm_page_busy_fetch(m); 1017 ce = VPB_CURTHREAD_EXCLUSIVE; 1018 for (;;) { 1019 if (VPB_SHARERS(x) > 1) 1020 return (0); 1021 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1022 ("vm_page_busy_tryupgrade: invalid lock state")); 1023 if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, 1024 ce | (x & VPB_BIT_WAITERS))) 1025 continue; 1026 return (1); 1027 } 1028 } 1029 1030 /* 1031 * vm_page_sbusied: 1032 * 1033 * Return a positive value if the page is shared busied, 0 otherwise. 1034 */ 1035 int 1036 vm_page_sbusied(vm_page_t m) 1037 { 1038 u_int x; 1039 1040 x = vm_page_busy_fetch(m); 1041 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 1042 } 1043 1044 /* 1045 * vm_page_sunbusy: 1046 * 1047 * Shared unbusy a page. 1048 */ 1049 void 1050 vm_page_sunbusy(vm_page_t m) 1051 { 1052 u_int x; 1053 1054 vm_page_assert_sbusied(m); 1055 1056 x = vm_page_busy_fetch(m); 1057 for (;;) { 1058 KASSERT(x != VPB_FREED, 1059 ("vm_page_sunbusy: Unlocking freed page.")); 1060 if (VPB_SHARERS(x) > 1) { 1061 if (atomic_fcmpset_int(&m->busy_lock, &x, 1062 x - VPB_ONE_SHARER)) 1063 break; 1064 continue; 1065 } 1066 KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), 1067 ("vm_page_sunbusy: invalid lock state")); 1068 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1069 continue; 1070 if ((x & VPB_BIT_WAITERS) == 0) 1071 break; 1072 wakeup(m); 1073 break; 1074 } 1075 } 1076 1077 /* 1078 * vm_page_busy_sleep: 1079 * 1080 * Sleep if the page is busy, using the page pointer as wchan. 1081 * This is used to implement the hard-path of the busying mechanism. 1082 * 1083 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1084 * will not sleep if the page is shared-busy. 1085 * 1086 * The object lock must be held on entry. 1087 * 1088 * Returns true if it slept and dropped the object lock, or false 1089 * if there was no sleep and the lock is still held. 1090 */ 1091 bool 1092 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags) 1093 { 1094 vm_object_t obj; 1095 1096 obj = m->object; 1097 VM_OBJECT_ASSERT_LOCKED(obj); 1098 1099 return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags, 1100 true)); 1101 } 1102 1103 /* 1104 * vm_page_busy_sleep_unlocked: 1105 * 1106 * Sleep if the page is busy, using the page pointer as wchan. 1107 * This is used to implement the hard-path of busying mechanism. 1108 * 1109 * If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function 1110 * will not sleep if the page is shared-busy. 1111 * 1112 * The object lock must not be held on entry. The operation will 1113 * return if the page changes identity. 1114 */ 1115 void 1116 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1117 const char *wmesg, int allocflags) 1118 { 1119 VM_OBJECT_ASSERT_UNLOCKED(obj); 1120 1121 (void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false); 1122 } 1123 1124 /* 1125 * _vm_page_busy_sleep: 1126 * 1127 * Internal busy sleep function. Verifies the page identity and 1128 * lockstate against parameters. Returns true if it sleeps and 1129 * false otherwise. 1130 * 1131 * allocflags uses VM_ALLOC_* flags to specify the lock required. 1132 * 1133 * If locked is true the lock will be dropped for any true returns 1134 * and held for any false returns. 1135 */ 1136 static bool 1137 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex, 1138 const char *wmesg, int allocflags, bool locked) 1139 { 1140 bool xsleep; 1141 u_int x; 1142 1143 /* 1144 * If the object is busy we must wait for that to drain to zero 1145 * before trying the page again. 1146 */ 1147 if (obj != NULL && vm_object_busied(obj)) { 1148 if (locked) 1149 VM_OBJECT_DROP(obj); 1150 vm_object_busy_wait(obj, wmesg); 1151 return (true); 1152 } 1153 1154 if (!vm_page_busied(m)) 1155 return (false); 1156 1157 xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0; 1158 sleepq_lock(m); 1159 x = vm_page_busy_fetch(m); 1160 do { 1161 /* 1162 * If the page changes objects or becomes unlocked we can 1163 * simply return. 1164 */ 1165 if (x == VPB_UNBUSIED || 1166 (xsleep && (x & VPB_BIT_SHARED) != 0) || 1167 m->object != obj || m->pindex != pindex) { 1168 sleepq_release(m); 1169 return (false); 1170 } 1171 if ((x & VPB_BIT_WAITERS) != 0) 1172 break; 1173 } while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS)); 1174 if (locked) 1175 VM_OBJECT_DROP(obj); 1176 DROP_GIANT(); 1177 sleepq_add(m, NULL, wmesg, 0, 0); 1178 sleepq_wait(m, PVM); 1179 PICKUP_GIANT(); 1180 return (true); 1181 } 1182 1183 /* 1184 * vm_page_trysbusy: 1185 * 1186 * Try to shared busy a page. 1187 * If the operation succeeds 1 is returned otherwise 0. 1188 * The operation never sleeps. 1189 */ 1190 int 1191 vm_page_trysbusy(vm_page_t m) 1192 { 1193 vm_object_t obj; 1194 u_int x; 1195 1196 obj = m->object; 1197 x = vm_page_busy_fetch(m); 1198 for (;;) { 1199 if ((x & VPB_BIT_SHARED) == 0) 1200 return (0); 1201 /* 1202 * Reduce the window for transient busies that will trigger 1203 * false negatives in vm_page_ps_test(). 1204 */ 1205 if (obj != NULL && vm_object_busied(obj)) 1206 return (0); 1207 if (atomic_fcmpset_acq_int(&m->busy_lock, &x, 1208 x + VPB_ONE_SHARER)) 1209 break; 1210 } 1211 1212 /* Refetch the object now that we're guaranteed that it is stable. */ 1213 obj = m->object; 1214 if (obj != NULL && vm_object_busied(obj)) { 1215 vm_page_sunbusy(m); 1216 return (0); 1217 } 1218 return (1); 1219 } 1220 1221 /* 1222 * vm_page_tryxbusy: 1223 * 1224 * Try to exclusive busy a page. 1225 * If the operation succeeds 1 is returned otherwise 0. 1226 * The operation never sleeps. 1227 */ 1228 int 1229 vm_page_tryxbusy(vm_page_t m) 1230 { 1231 vm_object_t obj; 1232 1233 if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED, 1234 VPB_CURTHREAD_EXCLUSIVE) == 0) 1235 return (0); 1236 1237 obj = m->object; 1238 if (obj != NULL && vm_object_busied(obj)) { 1239 vm_page_xunbusy(m); 1240 return (0); 1241 } 1242 return (1); 1243 } 1244 1245 static void 1246 vm_page_xunbusy_hard_tail(vm_page_t m) 1247 { 1248 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1249 /* Wake the waiter. */ 1250 wakeup(m); 1251 } 1252 1253 /* 1254 * vm_page_xunbusy_hard: 1255 * 1256 * Called when unbusy has failed because there is a waiter. 1257 */ 1258 void 1259 vm_page_xunbusy_hard(vm_page_t m) 1260 { 1261 vm_page_assert_xbusied(m); 1262 vm_page_xunbusy_hard_tail(m); 1263 } 1264 1265 void 1266 vm_page_xunbusy_hard_unchecked(vm_page_t m) 1267 { 1268 vm_page_assert_xbusied_unchecked(m); 1269 vm_page_xunbusy_hard_tail(m); 1270 } 1271 1272 static void 1273 vm_page_busy_free(vm_page_t m) 1274 { 1275 u_int x; 1276 1277 atomic_thread_fence_rel(); 1278 x = atomic_swap_int(&m->busy_lock, VPB_FREED); 1279 if ((x & VPB_BIT_WAITERS) != 0) 1280 wakeup(m); 1281 } 1282 1283 /* 1284 * vm_page_unhold_pages: 1285 * 1286 * Unhold each of the pages that is referenced by the given array. 1287 */ 1288 void 1289 vm_page_unhold_pages(vm_page_t *ma, int count) 1290 { 1291 1292 for (; count != 0; count--) { 1293 vm_page_unwire(*ma, PQ_ACTIVE); 1294 ma++; 1295 } 1296 } 1297 1298 vm_page_t 1299 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1300 { 1301 vm_page_t m; 1302 1303 #ifdef VM_PHYSSEG_SPARSE 1304 #if defined(__aarch64__) 1305 if (in_realm()) 1306 pa &= ~prot_ns_shared_pa; /* Mask off secure bit */ 1307 #endif 1308 m = vm_phys_paddr_to_vm_page(pa); 1309 if (m == NULL) 1310 m = vm_phys_fictitious_to_vm_page(pa); 1311 return (m); 1312 #elif defined(VM_PHYSSEG_DENSE) 1313 long pi; 1314 1315 pi = atop(pa); 1316 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1317 m = &vm_page_array[pi - first_page]; 1318 return (m); 1319 } 1320 return (vm_phys_fictitious_to_vm_page(pa)); 1321 #else 1322 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1323 #endif 1324 } 1325 1326 /* 1327 * vm_page_getfake: 1328 * 1329 * Create a fictitious page with the specified physical address and 1330 * memory attribute. The memory attribute is the only the machine- 1331 * dependent aspect of a fictitious page that must be initialized. 1332 */ 1333 vm_page_t 1334 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1335 { 1336 vm_page_t m; 1337 1338 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1339 vm_page_initfake(m, paddr, memattr); 1340 return (m); 1341 } 1342 1343 void 1344 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1345 { 1346 1347 if ((m->flags & PG_FICTITIOUS) != 0) { 1348 /* 1349 * The page's memattr might have changed since the 1350 * previous initialization. Update the pmap to the 1351 * new memattr. 1352 */ 1353 goto memattr; 1354 } 1355 m->phys_addr = paddr; 1356 m->a.queue = PQ_NONE; 1357 /* Fictitious pages don't use "segind". */ 1358 m->flags = PG_FICTITIOUS; 1359 /* Fictitious pages don't use "order" or "pool". */ 1360 m->oflags = VPO_UNMANAGED; 1361 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 1362 /* Fictitious pages are unevictable. */ 1363 m->ref_count = 1; 1364 pmap_page_init(m); 1365 memattr: 1366 pmap_page_set_memattr(m, memattr); 1367 } 1368 1369 /* 1370 * vm_page_putfake: 1371 * 1372 * Release a fictitious page. 1373 */ 1374 void 1375 vm_page_putfake(vm_page_t m) 1376 { 1377 1378 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1379 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1380 ("vm_page_putfake: bad page %p", m)); 1381 if (m->object != NULL) { 1382 vm_page_assert_xbusied(m); 1383 vm_page_busy_free(m); 1384 } 1385 uma_zfree(fakepg_zone, m); 1386 } 1387 1388 /* 1389 * vm_page_updatefake: 1390 * 1391 * Update the given fictitious page to the specified physical address and 1392 * memory attribute. 1393 */ 1394 void 1395 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1396 { 1397 1398 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1399 ("vm_page_updatefake: bad page %p", m)); 1400 m->phys_addr = paddr; 1401 pmap_page_set_memattr(m, memattr); 1402 } 1403 1404 /* 1405 * vm_page_free: 1406 * 1407 * Free a page. 1408 */ 1409 void 1410 vm_page_free(vm_page_t m) 1411 { 1412 1413 m->flags &= ~PG_ZERO; 1414 vm_page_free_toq(m); 1415 } 1416 1417 /* 1418 * vm_page_free_zero: 1419 * 1420 * Free a page to the zerod-pages queue 1421 */ 1422 void 1423 vm_page_free_zero(vm_page_t m) 1424 { 1425 1426 m->flags |= PG_ZERO; 1427 vm_page_free_toq(m); 1428 } 1429 1430 /* 1431 * Unbusy and handle the page queueing for a page from a getpages request that 1432 * was optionally read ahead or behind. 1433 */ 1434 void 1435 vm_page_readahead_finish(vm_page_t m) 1436 { 1437 1438 /* We shouldn't put invalid pages on queues. */ 1439 KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); 1440 1441 /* 1442 * Since the page is not the actually needed one, whether it should 1443 * be activated or deactivated is not obvious. Empirical results 1444 * have shown that deactivating the page is usually the best choice, 1445 * unless the page is wanted by another thread. 1446 */ 1447 if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0) 1448 vm_page_activate(m); 1449 else 1450 vm_page_deactivate(m); 1451 vm_page_xunbusy_unchecked(m); 1452 } 1453 1454 /* 1455 * Destroy the identity of an invalid page and free it if possible. 1456 * This is intended to be used when reading a page from backing store fails. 1457 */ 1458 void 1459 vm_page_free_invalid(vm_page_t m) 1460 { 1461 1462 KASSERT(vm_page_none_valid(m), ("page %p is valid", m)); 1463 KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); 1464 KASSERT(m->object != NULL, ("page %p has no object", m)); 1465 VM_OBJECT_ASSERT_WLOCKED(m->object); 1466 1467 /* 1468 * We may be attempting to free the page as part of the handling for an 1469 * I/O error, in which case the page was xbusied by a different thread. 1470 */ 1471 vm_page_xbusy_claim(m); 1472 1473 /* 1474 * If someone has wired this page while the object lock 1475 * was not held, then the thread that unwires is responsible 1476 * for freeing the page. Otherwise just free the page now. 1477 * The wire count of this unmapped page cannot change while 1478 * we have the page xbusy and the page's object wlocked. 1479 */ 1480 if (vm_page_remove(m)) 1481 vm_page_free(m); 1482 } 1483 1484 /* 1485 * vm_page_dirty_KBI: [ internal use only ] 1486 * 1487 * Set all bits in the page's dirty field. 1488 * 1489 * The object containing the specified page must be locked if the 1490 * call is made from the machine-independent layer. 1491 * 1492 * See vm_page_clear_dirty_mask(). 1493 * 1494 * This function should only be called by vm_page_dirty(). 1495 */ 1496 void 1497 vm_page_dirty_KBI(vm_page_t m) 1498 { 1499 1500 /* Refer to this operation by its public name. */ 1501 KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); 1502 m->dirty = VM_PAGE_BITS_ALL; 1503 } 1504 1505 /* 1506 * Insert the given page into the given object at the given pindex. 1507 * 1508 * The procedure is marked __always_inline to suggest to the compiler to 1509 * eliminate the iter parameter and the associated alternate branch. 1510 */ 1511 static __always_inline int 1512 vm_page_insert_lookup(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1513 bool iter, struct pctrie_iter *pages) 1514 { 1515 int error; 1516 1517 VM_OBJECT_ASSERT_WLOCKED(object); 1518 KASSERT(m->object == NULL, 1519 ("vm_page_insert: page %p already inserted", m)); 1520 1521 /* 1522 * Record the object/offset pair in this page. 1523 */ 1524 m->object = object; 1525 m->pindex = pindex; 1526 m->ref_count |= VPRC_OBJREF; 1527 1528 /* 1529 * Add this page to the object's radix tree. 1530 */ 1531 if (iter) 1532 error = vm_radix_iter_insert(pages, m); 1533 else 1534 error = vm_radix_insert(&object->rtree, m); 1535 if (__predict_false(error != 0)) { 1536 m->object = NULL; 1537 m->pindex = 0; 1538 m->ref_count &= ~VPRC_OBJREF; 1539 return (1); 1540 } 1541 1542 vm_page_insert_radixdone(m, object); 1543 vm_pager_page_inserted(object, m); 1544 return (0); 1545 } 1546 1547 /* 1548 * vm_page_insert: [ internal use only ] 1549 * 1550 * Inserts the given mem entry into the object and object list. 1551 * 1552 * The object must be locked. 1553 */ 1554 int 1555 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1556 { 1557 return (vm_page_insert_lookup(m, object, pindex, false, NULL)); 1558 } 1559 1560 /* 1561 * vm_page_iter_insert: 1562 * 1563 * Tries to insert the page "m" into the specified object at offset 1564 * "pindex" using the iterator "pages". Returns 0 if the insertion was 1565 * successful. 1566 * 1567 * The object must be locked. 1568 */ 1569 int 1570 vm_page_iter_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1571 struct pctrie_iter *pages) 1572 { 1573 return (vm_page_insert_lookup(m, object, pindex, true, pages)); 1574 } 1575 1576 /* 1577 * vm_page_insert_radixdone: 1578 * 1579 * Complete page "m" insertion into the specified object after the 1580 * radix trie hooking. 1581 * 1582 * The object must be locked. 1583 */ 1584 static void 1585 vm_page_insert_radixdone(vm_page_t m, vm_object_t object) 1586 { 1587 1588 VM_OBJECT_ASSERT_WLOCKED(object); 1589 KASSERT(object != NULL && m->object == object, 1590 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1591 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1592 ("vm_page_insert_radixdone: page %p is missing object ref", m)); 1593 1594 /* 1595 * Show that the object has one more resident page. 1596 */ 1597 object->resident_page_count++; 1598 1599 /* 1600 * Hold the vnode until the last page is released. 1601 */ 1602 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1603 vhold(object->handle); 1604 1605 /* 1606 * Since we are inserting a new and possibly dirty page, 1607 * update the object's generation count. 1608 */ 1609 if (pmap_page_is_write_mapped(m)) 1610 vm_object_set_writeable_dirty(object); 1611 } 1612 1613 /* 1614 * vm_page_remove_radixdone 1615 * 1616 * Complete page "m" removal from the specified object after the radix trie 1617 * unhooking. 1618 * 1619 * The caller is responsible for updating the page's fields to reflect this 1620 * removal. 1621 */ 1622 static void 1623 vm_page_remove_radixdone(vm_page_t m) 1624 { 1625 vm_object_t object; 1626 1627 vm_page_assert_xbusied(m); 1628 object = m->object; 1629 VM_OBJECT_ASSERT_WLOCKED(object); 1630 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1631 ("page %p is missing its object ref", m)); 1632 1633 /* Deferred free of swap space. */ 1634 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1635 vm_pager_page_unswapped(m); 1636 1637 vm_pager_page_removed(object, m); 1638 m->object = NULL; 1639 1640 /* 1641 * And show that the object has one fewer resident page. 1642 */ 1643 object->resident_page_count--; 1644 1645 /* 1646 * The vnode may now be recycled. 1647 */ 1648 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1649 vdrop(object->handle); 1650 } 1651 1652 /* 1653 * vm_page_free_object_prep: 1654 * 1655 * Disassociates the given page from its VM object. 1656 * 1657 * The object must be locked, and the page must be xbusy. 1658 */ 1659 static void 1660 vm_page_free_object_prep(vm_page_t m) 1661 { 1662 KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == 1663 ((m->object->flags & OBJ_UNMANAGED) != 0), 1664 ("%s: managed flag mismatch for page %p", 1665 __func__, m)); 1666 vm_page_assert_xbusied(m); 1667 1668 /* 1669 * The object reference can be released without an atomic 1670 * operation. 1671 */ 1672 KASSERT((m->flags & PG_FICTITIOUS) != 0 || 1673 m->ref_count == VPRC_OBJREF, 1674 ("%s: page %p has unexpected ref_count %u", 1675 __func__, m, m->ref_count)); 1676 vm_page_remove_radixdone(m); 1677 m->ref_count -= VPRC_OBJREF; 1678 } 1679 1680 /* 1681 * vm_page_iter_free: 1682 * 1683 * Free the given page, and use the iterator to remove it from the radix 1684 * tree. 1685 */ 1686 void 1687 vm_page_iter_free(struct pctrie_iter *pages, vm_page_t m) 1688 { 1689 vm_radix_iter_remove(pages); 1690 vm_page_free_object_prep(m); 1691 vm_page_xunbusy(m); 1692 m->flags &= ~PG_ZERO; 1693 vm_page_free_toq(m); 1694 } 1695 1696 /* 1697 * vm_page_remove: 1698 * 1699 * Removes the specified page from its containing object, but does not 1700 * invalidate any backing storage. Returns true if the object's reference 1701 * was the last reference to the page, and false otherwise. 1702 * 1703 * The object must be locked and the page must be exclusively busied. 1704 * The exclusive busy will be released on return. If this is not the 1705 * final ref and the caller does not hold a wire reference it may not 1706 * continue to access the page. 1707 */ 1708 bool 1709 vm_page_remove(vm_page_t m) 1710 { 1711 bool dropped; 1712 1713 dropped = vm_page_remove_xbusy(m); 1714 vm_page_xunbusy(m); 1715 1716 return (dropped); 1717 } 1718 1719 /* 1720 * vm_page_iter_remove: 1721 * 1722 * Remove the current page, and use the iterator to remove it from the 1723 * radix tree. 1724 */ 1725 bool 1726 vm_page_iter_remove(struct pctrie_iter *pages, vm_page_t m) 1727 { 1728 bool dropped; 1729 1730 vm_radix_iter_remove(pages); 1731 vm_page_remove_radixdone(m); 1732 dropped = (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1733 vm_page_xunbusy(m); 1734 1735 return (dropped); 1736 } 1737 1738 /* 1739 * vm_page_radix_remove 1740 * 1741 * Removes the specified page from the radix tree. 1742 */ 1743 static void 1744 vm_page_radix_remove(vm_page_t m) 1745 { 1746 vm_page_t mrem __diagused; 1747 1748 mrem = vm_radix_remove(&m->object->rtree, m->pindex); 1749 KASSERT(mrem == m, 1750 ("removed page %p, expected page %p", mrem, m)); 1751 } 1752 1753 /* 1754 * vm_page_remove_xbusy 1755 * 1756 * Removes the page but leaves the xbusy held. Returns true if this 1757 * removed the final ref and false otherwise. 1758 */ 1759 bool 1760 vm_page_remove_xbusy(vm_page_t m) 1761 { 1762 1763 vm_page_radix_remove(m); 1764 vm_page_remove_radixdone(m); 1765 return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); 1766 } 1767 1768 /* 1769 * vm_page_lookup: 1770 * 1771 * Returns the page associated with the object/offset 1772 * pair specified; if none is found, NULL is returned. 1773 * 1774 * The object must be locked. 1775 */ 1776 vm_page_t 1777 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1778 { 1779 1780 VM_OBJECT_ASSERT_LOCKED(object); 1781 return (vm_radix_lookup(&object->rtree, pindex)); 1782 } 1783 1784 /* 1785 * vm_page_iter_init: 1786 * 1787 * Initialize iterator for vm pages. 1788 */ 1789 void 1790 vm_page_iter_init(struct pctrie_iter *pages, vm_object_t object) 1791 { 1792 1793 vm_radix_iter_init(pages, &object->rtree); 1794 } 1795 1796 /* 1797 * vm_page_iter_init: 1798 * 1799 * Initialize iterator for vm pages. 1800 */ 1801 void 1802 vm_page_iter_limit_init(struct pctrie_iter *pages, vm_object_t object, 1803 vm_pindex_t limit) 1804 { 1805 1806 vm_radix_iter_limit_init(pages, &object->rtree, limit); 1807 } 1808 1809 /* 1810 * vm_page_lookup_unlocked: 1811 * 1812 * Returns the page associated with the object/offset pair specified; 1813 * if none is found, NULL is returned. The page may be no longer be 1814 * present in the object at the time that this function returns. Only 1815 * useful for opportunistic checks such as inmem(). 1816 */ 1817 vm_page_t 1818 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex) 1819 { 1820 1821 return (vm_radix_lookup_unlocked(&object->rtree, pindex)); 1822 } 1823 1824 /* 1825 * vm_page_relookup: 1826 * 1827 * Returns a page that must already have been busied by 1828 * the caller. Used for bogus page replacement. 1829 */ 1830 vm_page_t 1831 vm_page_relookup(vm_object_t object, vm_pindex_t pindex) 1832 { 1833 vm_page_t m; 1834 1835 m = vm_page_lookup_unlocked(object, pindex); 1836 KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) && 1837 m->object == object && m->pindex == pindex, 1838 ("vm_page_relookup: Invalid page %p", m)); 1839 return (m); 1840 } 1841 1842 /* 1843 * This should only be used by lockless functions for releasing transient 1844 * incorrect acquires. The page may have been freed after we acquired a 1845 * busy lock. In this case busy_lock == VPB_FREED and we have nothing 1846 * further to do. 1847 */ 1848 static void 1849 vm_page_busy_release(vm_page_t m) 1850 { 1851 u_int x; 1852 1853 x = vm_page_busy_fetch(m); 1854 for (;;) { 1855 if (x == VPB_FREED) 1856 break; 1857 if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) { 1858 if (atomic_fcmpset_int(&m->busy_lock, &x, 1859 x - VPB_ONE_SHARER)) 1860 break; 1861 continue; 1862 } 1863 KASSERT((x & VPB_BIT_SHARED) != 0 || 1864 (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE, 1865 ("vm_page_busy_release: %p xbusy not owned.", m)); 1866 if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) 1867 continue; 1868 if ((x & VPB_BIT_WAITERS) != 0) 1869 wakeup(m); 1870 break; 1871 } 1872 } 1873 1874 /* 1875 * Uses the page mnew as a replacement for an existing page at index 1876 * pindex which must be already present in the object. 1877 * 1878 * Both pages must be exclusively busied on enter. The old page is 1879 * unbusied on exit. 1880 * 1881 * A return value of true means mold is now free. If this is not the 1882 * final ref and the caller does not hold a wire reference it may not 1883 * continue to access the page. 1884 */ 1885 static bool 1886 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1887 vm_page_t mold) 1888 { 1889 vm_page_t mret __diagused; 1890 bool dropped; 1891 1892 VM_OBJECT_ASSERT_WLOCKED(object); 1893 vm_page_assert_xbusied(mold); 1894 KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, 1895 ("vm_page_replace: page %p already in object", mnew)); 1896 1897 /* 1898 * This function mostly follows vm_page_insert() and 1899 * vm_page_remove() without the radix, object count and vnode 1900 * dance. Double check such functions for more comments. 1901 */ 1902 1903 mnew->object = object; 1904 mnew->pindex = pindex; 1905 atomic_set_int(&mnew->ref_count, VPRC_OBJREF); 1906 mret = vm_radix_replace(&object->rtree, mnew); 1907 KASSERT(mret == mold, 1908 ("invalid page replacement, mold=%p, mret=%p", mold, mret)); 1909 KASSERT((mold->oflags & VPO_UNMANAGED) == 1910 (mnew->oflags & VPO_UNMANAGED), 1911 ("vm_page_replace: mismatched VPO_UNMANAGED")); 1912 1913 mold->object = NULL; 1914 1915 /* 1916 * The object's resident_page_count does not change because we have 1917 * swapped one page for another, but the generation count should 1918 * change if the page is dirty. 1919 */ 1920 if (pmap_page_is_write_mapped(mnew)) 1921 vm_object_set_writeable_dirty(object); 1922 dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; 1923 vm_page_xunbusy(mold); 1924 1925 return (dropped); 1926 } 1927 1928 void 1929 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, 1930 vm_page_t mold) 1931 { 1932 1933 vm_page_assert_xbusied(mnew); 1934 1935 if (vm_page_replace_hold(mnew, object, pindex, mold)) 1936 vm_page_free(mold); 1937 } 1938 1939 /* 1940 * vm_page_iter_rename: 1941 * 1942 * Tries to move the specified page from its current object to a new object 1943 * and pindex, using the given iterator to remove the page from its current 1944 * object. Returns true if the move was successful, and false if the move 1945 * was aborted due to a failed memory allocation. 1946 * 1947 * Panics if a page already resides in the new object at the new pindex. 1948 * 1949 * This routine dirties the page if it is valid, as callers are expected to 1950 * transfer backing storage only after moving the page. Dirtying the page 1951 * ensures that the destination object retains the most recent copy of the 1952 * page. 1953 * 1954 * The objects must be locked. 1955 */ 1956 bool 1957 vm_page_iter_rename(struct pctrie_iter *old_pages, vm_page_t m, 1958 vm_object_t new_object, vm_pindex_t new_pindex) 1959 { 1960 vm_pindex_t opidx; 1961 1962 KASSERT((m->ref_count & VPRC_OBJREF) != 0, 1963 ("%s: page %p is missing object ref", __func__, m)); 1964 VM_OBJECT_ASSERT_WLOCKED(m->object); 1965 VM_OBJECT_ASSERT_WLOCKED(new_object); 1966 1967 /* 1968 * Create a custom version of vm_page_insert() which does not depend 1969 * by m_prev and can cheat on the implementation aspects of the 1970 * function. 1971 */ 1972 opidx = m->pindex; 1973 m->pindex = new_pindex; 1974 if (vm_radix_insert(&new_object->rtree, m) != 0) { 1975 m->pindex = opidx; 1976 return (false); 1977 } 1978 1979 /* 1980 * The operation cannot fail anymore. 1981 */ 1982 m->pindex = opidx; 1983 vm_radix_iter_remove(old_pages); 1984 vm_page_remove_radixdone(m); 1985 1986 /* Return back to the new pindex to complete vm_page_insert(). */ 1987 m->pindex = new_pindex; 1988 m->object = new_object; 1989 1990 vm_page_insert_radixdone(m, new_object); 1991 if (vm_page_any_valid(m)) 1992 vm_page_dirty(m); 1993 vm_pager_page_inserted(new_object, m); 1994 return (true); 1995 } 1996 1997 /* 1998 * vm_page_alloc: 1999 * 2000 * Allocate and return a page that is associated with the specified 2001 * object and offset pair. By default, this page is exclusive busied. 2002 * 2003 * The caller must always specify an allocation class. 2004 * 2005 * allocation classes: 2006 * VM_ALLOC_NORMAL normal process request 2007 * VM_ALLOC_SYSTEM system *really* needs a page 2008 * VM_ALLOC_INTERRUPT interrupt time request 2009 * 2010 * optional allocation flags: 2011 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2012 * intends to allocate 2013 * VM_ALLOC_NOBUSY do not exclusive busy the page 2014 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 2015 * VM_ALLOC_NOFREE page will never be freed 2016 * VM_ALLOC_NOWAIT ignored (default behavior) 2017 * VM_ALLOC_SBUSY shared busy the allocated page 2018 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 2019 * VM_ALLOC_WIRED wire the allocated page 2020 * VM_ALLOC_ZERO prefer a zeroed page 2021 */ 2022 vm_page_t 2023 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 2024 { 2025 struct pctrie_iter pages; 2026 2027 vm_page_iter_init(&pages, object); 2028 return (vm_page_alloc_iter(object, pindex, req, &pages)); 2029 } 2030 2031 /* 2032 * Allocate a page in the specified object with the given page index. If the 2033 * object lock is dropped and regained, the pages iter is reset. 2034 */ 2035 vm_page_t 2036 vm_page_alloc_iter(vm_object_t object, vm_pindex_t pindex, int req, 2037 struct pctrie_iter *pages) 2038 { 2039 struct vm_domainset_iter di; 2040 vm_page_t m; 2041 int domain; 2042 2043 if (vm_domainset_iter_page_init(&di, object, pindex, &domain, &req) != 0) 2044 return (NULL); 2045 2046 do { 2047 m = vm_page_alloc_domain_iter(object, pindex, domain, req, 2048 pages); 2049 if (m != NULL) 2050 break; 2051 } while (vm_domainset_iter_page(&di, object, &domain, pages) == 0); 2052 2053 return (m); 2054 } 2055 2056 /* 2057 * Returns true if the number of free pages exceeds the minimum 2058 * for the request class and false otherwise. 2059 */ 2060 static int 2061 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) 2062 { 2063 u_int limit, old, new; 2064 2065 if (req_class == VM_ALLOC_INTERRUPT) 2066 limit = 0; 2067 else if (req_class == VM_ALLOC_SYSTEM) 2068 limit = vmd->vmd_interrupt_free_min; 2069 else 2070 limit = vmd->vmd_free_reserved; 2071 2072 /* 2073 * Attempt to reserve the pages. Fail if we're below the limit. 2074 */ 2075 limit += npages; 2076 old = atomic_load_int(&vmd->vmd_free_count); 2077 do { 2078 if (old < limit) 2079 return (0); 2080 new = old - npages; 2081 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 2082 2083 /* Wake the page daemon if we've crossed the threshold. */ 2084 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 2085 pagedaemon_wakeup(vmd->vmd_domain); 2086 2087 /* Only update bitsets on transitions. */ 2088 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 2089 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 2090 vm_domain_set(vmd); 2091 2092 return (1); 2093 } 2094 2095 int 2096 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 2097 { 2098 int req_class; 2099 2100 /* 2101 * The page daemon is allowed to dig deeper into the free page list. 2102 */ 2103 req_class = req & VM_ALLOC_CLASS_MASK; 2104 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2105 req_class = VM_ALLOC_SYSTEM; 2106 return (_vm_domain_allocate(vmd, req_class, npages)); 2107 } 2108 2109 vm_page_t 2110 vm_page_alloc_domain_iter(vm_object_t object, vm_pindex_t pindex, int domain, 2111 int req, struct pctrie_iter *pages) 2112 { 2113 struct vm_domain *vmd; 2114 vm_page_t m; 2115 int flags; 2116 2117 #define VM_ALLOC_COMMON (VM_ALLOC_CLASS_MASK | VM_ALLOC_NODUMP | \ 2118 VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | \ 2119 VM_ALLOC_WIRED | VM_ALLOC_ZERO) 2120 #define VPA_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2121 VM_ALLOC_NOBUSY | VM_ALLOC_NOFREE | \ 2122 VM_ALLOC_SBUSY) 2123 KASSERT((req & ~VPA_FLAGS) == 0, 2124 ("invalid request %#x", req)); 2125 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2126 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2127 ("invalid request %#x", req)); 2128 VM_OBJECT_ASSERT_WLOCKED(object); 2129 2130 flags = 0; 2131 m = NULL; 2132 if (!vm_pager_can_alloc_page(object, pindex)) 2133 return (NULL); 2134 #if VM_NRESERVLEVEL > 0 2135 again: 2136 #endif 2137 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2138 m = vm_page_alloc_nofree_domain(domain, req); 2139 if (m != NULL) 2140 goto found; 2141 } 2142 #if VM_NRESERVLEVEL > 0 2143 /* 2144 * Can we allocate the page from a reservation? 2145 */ 2146 if (vm_object_reserv(object) && 2147 (m = vm_reserv_alloc_page(object, pindex, domain, req, pages)) != 2148 NULL) { 2149 goto found; 2150 } 2151 #endif 2152 vmd = VM_DOMAIN(domain); 2153 if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) { 2154 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone, 2155 M_NOWAIT | M_NOVM); 2156 if (m != NULL) { 2157 flags |= PG_PCPU_CACHE; 2158 goto found; 2159 } 2160 } 2161 if (vm_domain_allocate(vmd, req, 1)) { 2162 /* 2163 * If not, allocate it from the free page queues. 2164 */ 2165 vm_domain_free_lock(vmd); 2166 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0); 2167 vm_domain_free_unlock(vmd); 2168 if (m == NULL) { 2169 vm_domain_freecnt_inc(vmd, 1); 2170 #if VM_NRESERVLEVEL > 0 2171 if (vm_reserv_reclaim_inactive(domain)) 2172 goto again; 2173 #endif 2174 } 2175 } 2176 if (m == NULL) { 2177 /* 2178 * Not allocatable, give up. 2179 */ 2180 (void)vm_domain_alloc_fail(vmd, object, req); 2181 if ((req & VM_ALLOC_WAITFAIL) != 0) 2182 pctrie_iter_reset(pages); 2183 return (NULL); 2184 } 2185 2186 /* 2187 * At this point we had better have found a good page. 2188 */ 2189 found: 2190 vm_page_dequeue(m); 2191 vm_page_alloc_check(m); 2192 2193 /* 2194 * Initialize the page. Only the PG_ZERO flag is inherited. 2195 */ 2196 flags |= m->flags & PG_ZERO; 2197 if ((req & VM_ALLOC_NODUMP) != 0) 2198 flags |= PG_NODUMP; 2199 if ((req & VM_ALLOC_NOFREE) != 0) 2200 flags |= PG_NOFREE; 2201 m->flags = flags; 2202 m->a.flags = 0; 2203 m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2204 m->pool = VM_FREEPOOL_DEFAULT; 2205 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2206 m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2207 else if ((req & VM_ALLOC_SBUSY) != 0) 2208 m->busy_lock = VPB_SHARERS_WORD(1); 2209 else 2210 m->busy_lock = VPB_UNBUSIED; 2211 if (req & VM_ALLOC_WIRED) { 2212 vm_wire_add(1); 2213 m->ref_count = 1; 2214 } 2215 m->a.act_count = 0; 2216 2217 if (vm_page_iter_insert(m, object, pindex, pages)) { 2218 if (req & VM_ALLOC_WIRED) { 2219 vm_wire_sub(1); 2220 m->ref_count = 0; 2221 } 2222 KASSERT(m->object == NULL, ("page %p has object", m)); 2223 m->oflags = VPO_UNMANAGED; 2224 m->busy_lock = VPB_UNBUSIED; 2225 /* Don't change PG_ZERO. */ 2226 vm_page_free_toq(m); 2227 if (req & VM_ALLOC_WAITFAIL) { 2228 VM_OBJECT_WUNLOCK(object); 2229 vm_radix_wait(); 2230 pctrie_iter_reset(pages); 2231 VM_OBJECT_WLOCK(object); 2232 } 2233 return (NULL); 2234 } 2235 2236 /* Ignore device objects; the pager sets "memattr" for them. */ 2237 if (object->memattr != VM_MEMATTR_DEFAULT && 2238 (object->flags & OBJ_FICTITIOUS) == 0) 2239 pmap_page_set_memattr(m, object->memattr); 2240 2241 return (m); 2242 } 2243 2244 /* 2245 * vm_page_alloc_contig: 2246 * 2247 * Allocate a contiguous set of physical pages of the given size "npages" 2248 * from the free lists. All of the physical pages must be at or above 2249 * the given physical address "low" and below the given physical address 2250 * "high". The given value "alignment" determines the alignment of the 2251 * first physical page in the set. If the given value "boundary" is 2252 * non-zero, then the set of physical pages cannot cross any physical 2253 * address boundary that is a multiple of that value. Both "alignment" 2254 * and "boundary" must be a power of two. 2255 * 2256 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 2257 * then the memory attribute setting for the physical pages is configured 2258 * to the object's memory attribute setting. Otherwise, the memory 2259 * attribute setting for the physical pages is configured to "memattr", 2260 * overriding the object's memory attribute setting. However, if the 2261 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 2262 * memory attribute setting for the physical pages cannot be configured 2263 * to VM_MEMATTR_DEFAULT. 2264 * 2265 * The specified object may not contain fictitious pages. 2266 * 2267 * The caller must always specify an allocation class. 2268 * 2269 * allocation classes: 2270 * VM_ALLOC_NORMAL normal process request 2271 * VM_ALLOC_SYSTEM system *really* needs the pages 2272 * VM_ALLOC_INTERRUPT interrupt time request 2273 * 2274 * optional allocation flags: 2275 * VM_ALLOC_NOBUSY do not exclusive busy the pages 2276 * VM_ALLOC_NODUMP do not include the pages in a kernel core dump 2277 * VM_ALLOC_NORECLAIM do not reclaim after initial failure 2278 * VM_ALLOC_NOWAIT ignored (default behavior) 2279 * VM_ALLOC_SBUSY shared busy the allocated pages 2280 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 2281 * VM_ALLOC_WIRED wire the allocated pages 2282 * VM_ALLOC_ZERO prefer zeroed pages 2283 */ 2284 vm_page_t 2285 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 2286 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2287 vm_paddr_t boundary, vm_memattr_t memattr) 2288 { 2289 struct vm_domainset_iter di; 2290 vm_page_t bounds[2]; 2291 vm_page_t m; 2292 int domain; 2293 int start_segind; 2294 2295 start_segind = -1; 2296 2297 if (vm_domainset_iter_page_init(&di, object, pindex, &domain, &req) != 0) 2298 return (NULL); 2299 2300 do { 2301 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 2302 npages, low, high, alignment, boundary, memattr); 2303 if (m != NULL) 2304 break; 2305 if (start_segind == -1) 2306 start_segind = vm_phys_lookup_segind(low); 2307 if (vm_phys_find_range(bounds, start_segind, domain, 2308 npages, low, high) == -1) { 2309 vm_domainset_iter_ignore(&di, domain); 2310 } 2311 } while (vm_domainset_iter_page(&di, object, &domain, NULL) == 0); 2312 2313 return (m); 2314 } 2315 2316 static vm_page_t 2317 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, 2318 vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2319 { 2320 struct vm_domain *vmd; 2321 vm_page_t m_ret; 2322 2323 /* 2324 * Can we allocate the pages without the number of free pages falling 2325 * below the lower bound for the allocation class? 2326 */ 2327 vmd = VM_DOMAIN(domain); 2328 if (!vm_domain_allocate(vmd, req, npages)) 2329 return (NULL); 2330 /* 2331 * Try to allocate the pages from the free page queues. 2332 */ 2333 vm_domain_free_lock(vmd); 2334 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2335 alignment, boundary); 2336 vm_domain_free_unlock(vmd); 2337 if (m_ret != NULL) 2338 return (m_ret); 2339 #if VM_NRESERVLEVEL > 0 2340 /* 2341 * Try to break a reservation to allocate the pages. 2342 */ 2343 if ((req & VM_ALLOC_NORECLAIM) == 0) { 2344 m_ret = vm_reserv_reclaim_contig(domain, npages, low, 2345 high, alignment, boundary); 2346 if (m_ret != NULL) 2347 return (m_ret); 2348 } 2349 #endif 2350 vm_domain_freecnt_inc(vmd, npages); 2351 return (NULL); 2352 } 2353 2354 vm_page_t 2355 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 2356 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 2357 vm_paddr_t boundary, vm_memattr_t memattr) 2358 { 2359 struct pctrie_iter pages; 2360 vm_page_t m, m_ret, mpred; 2361 u_int busy_lock, flags, oflags; 2362 2363 #define VPAC_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2364 VM_ALLOC_NOBUSY | VM_ALLOC_NORECLAIM | \ 2365 VM_ALLOC_SBUSY) 2366 KASSERT((req & ~VPAC_FLAGS) == 0, 2367 ("invalid request %#x", req)); 2368 KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 2369 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 2370 ("invalid request %#x", req)); 2371 VM_OBJECT_ASSERT_WLOCKED(object); 2372 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2373 ("vm_page_alloc_contig: object %p has fictitious pages", 2374 object)); 2375 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2376 2377 vm_page_iter_init(&pages, object); 2378 m_ret = NULL; 2379 #if VM_NRESERVLEVEL > 0 2380 /* 2381 * Can we allocate the pages from a reservation? 2382 */ 2383 if (vm_object_reserv(object)) { 2384 m_ret = vm_reserv_alloc_contig(object, pindex, domain, 2385 req, npages, low, high, alignment, boundary, &pages); 2386 } 2387 #endif 2388 if (m_ret == NULL) { 2389 m_ret = vm_page_find_contig_domain(domain, req, npages, 2390 low, high, alignment, boundary); 2391 } 2392 if (m_ret == NULL) { 2393 (void)vm_domain_alloc_fail(VM_DOMAIN(domain), object, req); 2394 return (NULL); 2395 } 2396 2397 /* 2398 * Initialize the pages. Only the PG_ZERO flag is inherited. 2399 */ 2400 flags = PG_ZERO; 2401 if ((req & VM_ALLOC_NODUMP) != 0) 2402 flags |= PG_NODUMP; 2403 oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; 2404 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2405 busy_lock = VPB_CURTHREAD_EXCLUSIVE; 2406 else if ((req & VM_ALLOC_SBUSY) != 0) 2407 busy_lock = VPB_SHARERS_WORD(1); 2408 else 2409 busy_lock = VPB_UNBUSIED; 2410 if ((req & VM_ALLOC_WIRED) != 0) 2411 vm_wire_add(npages); 2412 if (object->memattr != VM_MEMATTR_DEFAULT && 2413 memattr == VM_MEMATTR_DEFAULT) 2414 memattr = object->memattr; 2415 for (m = m_ret; m < &m_ret[npages]; m++) { 2416 vm_page_dequeue(m); 2417 vm_page_alloc_check(m); 2418 m->a.flags = 0; 2419 m->flags = (m->flags | PG_NODUMP) & flags; 2420 m->busy_lock = busy_lock; 2421 if ((req & VM_ALLOC_WIRED) != 0) 2422 m->ref_count = 1; 2423 m->a.act_count = 0; 2424 m->oflags = oflags; 2425 m->pool = VM_FREEPOOL_DEFAULT; 2426 if (vm_page_iter_insert(m, object, pindex, &pages)) { 2427 if ((req & VM_ALLOC_WIRED) != 0) 2428 vm_wire_sub(npages); 2429 KASSERT(m->object == NULL, 2430 ("page %p has object", m)); 2431 mpred = m; 2432 for (m = m_ret; m < &m_ret[npages]; m++) { 2433 if (m <= mpred && 2434 (req & VM_ALLOC_WIRED) != 0) 2435 m->ref_count = VPRC_OBJREF; 2436 m->oflags = VPO_UNMANAGED; 2437 m->busy_lock = VPB_UNBUSIED; 2438 /* Don't change PG_ZERO. */ 2439 vm_page_free_toq(m); 2440 } 2441 if (req & VM_ALLOC_WAITFAIL) { 2442 VM_OBJECT_WUNLOCK(object); 2443 vm_radix_wait(); 2444 VM_OBJECT_WLOCK(object); 2445 } 2446 return (NULL); 2447 } 2448 if (memattr != VM_MEMATTR_DEFAULT) 2449 pmap_page_set_memattr(m, memattr); 2450 pindex++; 2451 } 2452 return (m_ret); 2453 } 2454 2455 /* 2456 * Allocate a physical page that is not intended to be inserted into a VM 2457 * object. 2458 */ 2459 vm_page_t 2460 vm_page_alloc_noobj_domain(int domain, int req) 2461 { 2462 struct vm_domain *vmd; 2463 vm_page_t m; 2464 int flags; 2465 2466 #define VPAN_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2467 VM_ALLOC_NOFREE | VM_ALLOC_WAITOK) 2468 KASSERT((req & ~VPAN_FLAGS) == 0, 2469 ("invalid request %#x", req)); 2470 2471 flags = ((req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0) | 2472 ((req & VM_ALLOC_NOFREE) != 0 ? PG_NOFREE : 0); 2473 vmd = VM_DOMAIN(domain); 2474 again: 2475 if (__predict_false((req & VM_ALLOC_NOFREE) != 0)) { 2476 m = vm_page_alloc_nofree_domain(domain, req); 2477 if (m != NULL) 2478 goto found; 2479 } 2480 2481 if (vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) { 2482 m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone, 2483 M_NOWAIT | M_NOVM); 2484 if (m != NULL) { 2485 flags |= PG_PCPU_CACHE; 2486 goto found; 2487 } 2488 } 2489 2490 if (vm_domain_allocate(vmd, req, 1)) { 2491 vm_domain_free_lock(vmd); 2492 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0); 2493 vm_domain_free_unlock(vmd); 2494 if (m == NULL) { 2495 vm_domain_freecnt_inc(vmd, 1); 2496 #if VM_NRESERVLEVEL > 0 2497 if (vm_reserv_reclaim_inactive(domain)) 2498 goto again; 2499 #endif 2500 } 2501 } 2502 if (m == NULL) { 2503 if (!vm_domain_alloc_fail(vmd, NULL, req)) 2504 return (NULL); 2505 goto again; 2506 } 2507 2508 found: 2509 /* 2510 * If the page comes from the free page cache, then it might still 2511 * have a pending deferred dequeue. Specifically, when the page is 2512 * imported from a different pool by vm_phys_alloc_npages(), the 2513 * second, third, etc. pages in a non-zero order set could have 2514 * pending deferred dequeues. 2515 */ 2516 vm_page_dequeue(m); 2517 vm_page_alloc_check(m); 2518 2519 /* 2520 * Consumers should not rely on a useful default pindex value. 2521 */ 2522 m->pindex = 0xdeadc0dedeadc0de; 2523 m->flags = (m->flags & PG_ZERO) | flags; 2524 m->a.flags = 0; 2525 m->oflags = VPO_UNMANAGED; 2526 m->pool = VM_FREEPOOL_DIRECT; 2527 m->busy_lock = VPB_UNBUSIED; 2528 if ((req & VM_ALLOC_WIRED) != 0) { 2529 vm_wire_add(1); 2530 m->ref_count = 1; 2531 } 2532 2533 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2534 pmap_zero_page(m); 2535 2536 return (m); 2537 } 2538 2539 #if VM_NRESERVLEVEL > 1 2540 #define VM_NOFREE_IMPORT_ORDER (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER) 2541 #elif VM_NRESERVLEVEL > 0 2542 #define VM_NOFREE_IMPORT_ORDER VM_LEVEL_0_ORDER 2543 #else 2544 #define VM_NOFREE_IMPORT_ORDER 8 2545 #endif 2546 2547 /* 2548 * Allocate a single NOFREE page. 2549 * 2550 * This routine hands out NOFREE pages from higher-order 2551 * physical memory blocks in order to reduce memory fragmentation. 2552 * When a NOFREE for a given domain chunk is used up, 2553 * the routine will try to fetch a new one from the freelists 2554 * and discard the old one. 2555 */ 2556 static vm_page_t __noinline 2557 vm_page_alloc_nofree_domain(int domain, int req) 2558 { 2559 vm_page_t m; 2560 struct vm_domain *vmd; 2561 2562 KASSERT((req & VM_ALLOC_NOFREE) != 0, ("invalid request %#x", req)); 2563 2564 vmd = VM_DOMAIN(domain); 2565 vm_domain_free_lock(vmd); 2566 if (TAILQ_EMPTY(&vmd->vmd_nofreeq)) { 2567 int count; 2568 2569 count = 1 << VM_NOFREE_IMPORT_ORDER; 2570 if (!vm_domain_allocate(vmd, req, count)) { 2571 vm_domain_free_unlock(vmd); 2572 return (NULL); 2573 } 2574 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 2575 VM_NOFREE_IMPORT_ORDER); 2576 if (m == NULL) { 2577 vm_domain_freecnt_inc(vmd, count); 2578 vm_domain_free_unlock(vmd); 2579 return (NULL); 2580 } 2581 m->ref_count = count - 1; 2582 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, plinks.q); 2583 atomic_add_long(&nofreeq_size, count); 2584 } 2585 m = TAILQ_FIRST(&vmd->vmd_nofreeq); 2586 TAILQ_REMOVE(&vmd->vmd_nofreeq, m, plinks.q); 2587 if (m->ref_count > 0) { 2588 vm_page_t m_next; 2589 2590 m_next = &m[1]; 2591 vm_page_dequeue(m_next); 2592 m_next->ref_count = m->ref_count - 1; 2593 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m_next, plinks.q); 2594 m->ref_count = 0; 2595 } 2596 vm_domain_free_unlock(vmd); 2597 atomic_add_long(&nofreeq_size, -1); 2598 VM_CNT_INC(v_nofree_count); 2599 2600 return (m); 2601 } 2602 2603 /* 2604 * Though a NOFREE page by definition should not be freed, we support putting 2605 * them aside for future NOFREE allocations. This enables code which allocates 2606 * NOFREE pages for some purpose but then encounters an error and releases 2607 * resources. 2608 */ 2609 static void __noinline 2610 vm_page_free_nofree(struct vm_domain *vmd, vm_page_t m) 2611 { 2612 VM_CNT_ADD(v_nofree_count, -1); 2613 atomic_add_long(&nofreeq_size, 1); 2614 vm_domain_free_lock(vmd); 2615 MPASS(m->ref_count == 0); 2616 TAILQ_INSERT_HEAD(&vmd->vmd_nofreeq, m, plinks.q); 2617 vm_domain_free_unlock(vmd); 2618 } 2619 2620 vm_page_t 2621 vm_page_alloc_noobj(int req) 2622 { 2623 struct vm_domainset_iter di; 2624 vm_page_t m; 2625 int domain; 2626 2627 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 2628 return (NULL); 2629 2630 do { 2631 m = vm_page_alloc_noobj_domain(domain, req); 2632 if (m != NULL) 2633 break; 2634 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 2635 2636 return (m); 2637 } 2638 2639 vm_page_t 2640 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low, 2641 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2642 vm_memattr_t memattr) 2643 { 2644 struct vm_domainset_iter di; 2645 vm_page_t m; 2646 int domain; 2647 2648 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 2649 return (NULL); 2650 2651 do { 2652 m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low, 2653 high, alignment, boundary, memattr); 2654 if (m != NULL) 2655 break; 2656 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 2657 2658 return (m); 2659 } 2660 2661 vm_page_t 2662 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages, 2663 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 2664 vm_memattr_t memattr) 2665 { 2666 vm_page_t m, m_ret; 2667 u_int flags; 2668 2669 #define VPANC_FLAGS (VM_ALLOC_COMMON | VM_ALLOC_COUNT_MASK | \ 2670 VM_ALLOC_NORECLAIM | VM_ALLOC_WAITOK) 2671 KASSERT((req & ~VPANC_FLAGS) == 0, 2672 ("invalid request %#x", req)); 2673 KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) != 2674 (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM), 2675 ("invalid request %#x", req)); 2676 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2677 2678 while ((m_ret = vm_page_find_contig_domain(domain, req, npages, 2679 low, high, alignment, boundary)) == NULL) { 2680 if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req)) 2681 return (NULL); 2682 } 2683 2684 /* 2685 * Initialize the pages. Only the PG_ZERO flag is inherited. 2686 */ 2687 flags = PG_ZERO; 2688 if ((req & VM_ALLOC_NODUMP) != 0) 2689 flags |= PG_NODUMP; 2690 if ((req & VM_ALLOC_WIRED) != 0) 2691 vm_wire_add(npages); 2692 for (m = m_ret; m < &m_ret[npages]; m++) { 2693 vm_page_dequeue(m); 2694 vm_page_alloc_check(m); 2695 2696 /* 2697 * Consumers should not rely on a useful default pindex value. 2698 */ 2699 m->pindex = 0xdeadc0dedeadc0de; 2700 m->a.flags = 0; 2701 m->flags = (m->flags | PG_NODUMP) & flags; 2702 m->busy_lock = VPB_UNBUSIED; 2703 if ((req & VM_ALLOC_WIRED) != 0) 2704 m->ref_count = 1; 2705 m->a.act_count = 0; 2706 m->oflags = VPO_UNMANAGED; 2707 m->pool = VM_FREEPOOL_DIRECT; 2708 2709 /* 2710 * Zero the page before updating any mappings since the page is 2711 * not yet shared with any devices which might require the 2712 * non-default memory attribute. pmap_page_set_memattr() 2713 * flushes data caches before returning. 2714 */ 2715 if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0) 2716 pmap_zero_page(m); 2717 if (memattr != VM_MEMATTR_DEFAULT) 2718 pmap_page_set_memattr(m, memattr); 2719 } 2720 return (m_ret); 2721 } 2722 2723 /* 2724 * Check a page that has been freshly dequeued from a freelist. 2725 */ 2726 static void 2727 vm_page_alloc_check(vm_page_t m) 2728 { 2729 2730 KASSERT(m->object == NULL, ("page %p has object", m)); 2731 KASSERT(m->a.queue == PQ_NONE && 2732 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 2733 ("page %p has unexpected queue %d, flags %#x", 2734 m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); 2735 KASSERT(m->ref_count == 0, ("page %p has references", m)); 2736 KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m)); 2737 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2738 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2739 ("page %p has unexpected memattr %d", 2740 m, pmap_page_get_memattr(m))); 2741 KASSERT(vm_page_none_valid(m), ("free page %p is valid", m)); 2742 pmap_vm_page_alloc_check(m); 2743 } 2744 2745 static int 2746 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) 2747 { 2748 struct vm_domain *vmd; 2749 struct vm_pgcache *pgcache; 2750 int i; 2751 2752 pgcache = arg; 2753 vmd = VM_DOMAIN(pgcache->domain); 2754 2755 /* 2756 * The page daemon should avoid creating extra memory pressure since its 2757 * main purpose is to replenish the store of free pages. 2758 */ 2759 if (vmd->vmd_severeset || curproc == pageproc || 2760 !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2761 return (0); 2762 domain = vmd->vmd_domain; 2763 vm_domain_free_lock(vmd); 2764 i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, 2765 (vm_page_t *)store); 2766 vm_domain_free_unlock(vmd); 2767 if (cnt != i) 2768 vm_domain_freecnt_inc(vmd, cnt - i); 2769 2770 return (i); 2771 } 2772 2773 static void 2774 vm_page_zone_release(void *arg, void **store, int cnt) 2775 { 2776 struct vm_domain *vmd; 2777 struct vm_pgcache *pgcache; 2778 vm_page_t m; 2779 int i; 2780 2781 pgcache = arg; 2782 vmd = VM_DOMAIN(pgcache->domain); 2783 vm_domain_free_lock(vmd); 2784 for (i = 0; i < cnt; i++) { 2785 m = (vm_page_t)store[i]; 2786 vm_phys_free_pages(m, pgcache->pool, 0); 2787 } 2788 vm_domain_free_unlock(vmd); 2789 vm_domain_freecnt_inc(vmd, cnt); 2790 } 2791 2792 #define VPSC_ANY 0 /* No restrictions. */ 2793 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2794 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2795 2796 /* 2797 * vm_page_scan_contig: 2798 * 2799 * Scan vm_page_array[] between the specified entries "m_start" and 2800 * "m_end" for a run of contiguous physical pages that satisfy the 2801 * specified conditions, and return the lowest page in the run. The 2802 * specified "alignment" determines the alignment of the lowest physical 2803 * page in the run. If the specified "boundary" is non-zero, then the 2804 * run of physical pages cannot span a physical address that is a 2805 * multiple of "boundary". 2806 * 2807 * "m_end" is never dereferenced, so it need not point to a vm_page 2808 * structure within vm_page_array[]. 2809 * 2810 * "npages" must be greater than zero. "m_start" and "m_end" must not 2811 * span a hole (or discontiguity) in the physical address space. Both 2812 * "alignment" and "boundary" must be a power of two. 2813 */ 2814 static vm_page_t 2815 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2816 u_long alignment, vm_paddr_t boundary, int options) 2817 { 2818 vm_object_t object; 2819 vm_paddr_t pa; 2820 vm_page_t m, m_run; 2821 #if VM_NRESERVLEVEL > 0 2822 int level; 2823 #endif 2824 int m_inc, order, run_ext, run_len; 2825 2826 KASSERT(npages > 0, ("npages is 0")); 2827 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2828 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2829 m_run = NULL; 2830 run_len = 0; 2831 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2832 KASSERT((m->flags & PG_MARKER) == 0, 2833 ("page %p is PG_MARKER", m)); 2834 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, 2835 ("fictitious page %p has invalid ref count", m)); 2836 2837 /* 2838 * If the current page would be the start of a run, check its 2839 * physical address against the end, alignment, and boundary 2840 * conditions. If it doesn't satisfy these conditions, either 2841 * terminate the scan or advance to the next page that 2842 * satisfies the failed condition. 2843 */ 2844 if (run_len == 0) { 2845 KASSERT(m_run == NULL, ("m_run != NULL")); 2846 if (m + npages > m_end) 2847 break; 2848 pa = VM_PAGE_TO_PHYS(m); 2849 if (!vm_addr_align_ok(pa, alignment)) { 2850 m_inc = atop(roundup2(pa, alignment) - pa); 2851 continue; 2852 } 2853 if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) { 2854 m_inc = atop(roundup2(pa, boundary) - pa); 2855 continue; 2856 } 2857 } else 2858 KASSERT(m_run != NULL, ("m_run == NULL")); 2859 2860 retry: 2861 m_inc = 1; 2862 if (vm_page_wired(m)) 2863 run_ext = 0; 2864 #if VM_NRESERVLEVEL > 0 2865 else if ((level = vm_reserv_level(m)) >= 0 && 2866 (options & VPSC_NORESERV) != 0) { 2867 run_ext = 0; 2868 /* Advance to the end of the reservation. */ 2869 pa = VM_PAGE_TO_PHYS(m); 2870 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2871 pa); 2872 } 2873 #endif 2874 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 2875 /* 2876 * The page is considered eligible for relocation if 2877 * and only if it could be laundered or reclaimed by 2878 * the page daemon. 2879 */ 2880 VM_OBJECT_RLOCK(object); 2881 if (object != m->object) { 2882 VM_OBJECT_RUNLOCK(object); 2883 goto retry; 2884 } 2885 /* Don't care: PG_NODUMP, PG_ZERO. */ 2886 if ((object->flags & OBJ_SWAP) == 0 && 2887 object->type != OBJT_VNODE) { 2888 run_ext = 0; 2889 #if VM_NRESERVLEVEL > 0 2890 } else if ((options & VPSC_NOSUPER) != 0 && 2891 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2892 run_ext = 0; 2893 /* Advance to the end of the superpage. */ 2894 pa = VM_PAGE_TO_PHYS(m); 2895 m_inc = atop(roundup2(pa + 1, 2896 vm_reserv_size(level)) - pa); 2897 #endif 2898 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2899 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2900 /* 2901 * The page is allocated but eligible for 2902 * relocation. Extend the current run by one 2903 * page. 2904 */ 2905 KASSERT(pmap_page_get_memattr(m) == 2906 VM_MEMATTR_DEFAULT, 2907 ("page %p has an unexpected memattr", m)); 2908 KASSERT((m->oflags & (VPO_SWAPINPROG | 2909 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2910 ("page %p has unexpected oflags", m)); 2911 /* Don't care: PGA_NOSYNC. */ 2912 run_ext = 1; 2913 } else 2914 run_ext = 0; 2915 VM_OBJECT_RUNLOCK(object); 2916 #if VM_NRESERVLEVEL > 0 2917 } else if (level >= 0) { 2918 /* 2919 * The page is reserved but not yet allocated. In 2920 * other words, it is still free. Extend the current 2921 * run by one page. 2922 */ 2923 run_ext = 1; 2924 #endif 2925 } else if ((order = m->order) < VM_NFREEORDER) { 2926 /* 2927 * The page is enqueued in the physical memory 2928 * allocator's free page queues. Moreover, it is the 2929 * first page in a power-of-two-sized run of 2930 * contiguous free pages. Add these pages to the end 2931 * of the current run, and jump ahead. 2932 */ 2933 run_ext = 1 << order; 2934 m_inc = 1 << order; 2935 } else { 2936 /* 2937 * Skip the page for one of the following reasons: (1) 2938 * It is enqueued in the physical memory allocator's 2939 * free page queues. However, it is not the first 2940 * page in a run of contiguous free pages. (This case 2941 * rarely occurs because the scan is performed in 2942 * ascending order.) (2) It is not reserved, and it is 2943 * transitioning from free to allocated. (Conversely, 2944 * the transition from allocated to free for managed 2945 * pages is blocked by the page busy lock.) (3) It is 2946 * allocated but not contained by an object and not 2947 * wired, e.g., allocated by Xen's balloon driver. 2948 */ 2949 run_ext = 0; 2950 } 2951 2952 /* 2953 * Extend or reset the current run of pages. 2954 */ 2955 if (run_ext > 0) { 2956 if (run_len == 0) 2957 m_run = m; 2958 run_len += run_ext; 2959 } else { 2960 if (run_len > 0) { 2961 m_run = NULL; 2962 run_len = 0; 2963 } 2964 } 2965 } 2966 if (run_len >= npages) 2967 return (m_run); 2968 return (NULL); 2969 } 2970 2971 /* 2972 * vm_page_reclaim_run: 2973 * 2974 * Try to relocate each of the allocated virtual pages within the 2975 * specified run of physical pages to a new physical address. Free the 2976 * physical pages underlying the relocated virtual pages. A virtual page 2977 * is relocatable if and only if it could be laundered or reclaimed by 2978 * the page daemon. Whenever possible, a virtual page is relocated to a 2979 * physical address above "high". 2980 * 2981 * Returns 0 if every physical page within the run was already free or 2982 * just freed by a successful relocation. Otherwise, returns a non-zero 2983 * value indicating why the last attempt to relocate a virtual page was 2984 * unsuccessful. 2985 * 2986 * "req_class" must be an allocation class. 2987 */ 2988 static int 2989 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2990 vm_paddr_t high) 2991 { 2992 struct vm_domain *vmd; 2993 struct spglist free; 2994 vm_object_t object; 2995 vm_paddr_t pa; 2996 vm_page_t m, m_end, m_new; 2997 int error, order, req; 2998 2999 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 3000 ("req_class is not an allocation class")); 3001 SLIST_INIT(&free); 3002 error = 0; 3003 m = m_run; 3004 m_end = m_run + npages; 3005 for (; error == 0 && m < m_end; m++) { 3006 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 3007 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 3008 3009 /* 3010 * Racily check for wirings. Races are handled once the object 3011 * lock is held and the page is unmapped. 3012 */ 3013 if (vm_page_wired(m)) 3014 error = EBUSY; 3015 else if ((object = atomic_load_ptr(&m->object)) != NULL) { 3016 /* 3017 * The page is relocated if and only if it could be 3018 * laundered or reclaimed by the page daemon. 3019 */ 3020 VM_OBJECT_WLOCK(object); 3021 /* Don't care: PG_NODUMP, PG_ZERO. */ 3022 if (m->object != object || 3023 ((object->flags & OBJ_SWAP) == 0 && 3024 object->type != OBJT_VNODE)) 3025 error = EINVAL; 3026 else if (object->memattr != VM_MEMATTR_DEFAULT) 3027 error = EINVAL; 3028 else if (vm_page_queue(m) != PQ_NONE && 3029 vm_page_tryxbusy(m) != 0) { 3030 if (vm_page_wired(m)) { 3031 vm_page_xunbusy(m); 3032 error = EBUSY; 3033 goto unlock; 3034 } 3035 KASSERT(pmap_page_get_memattr(m) == 3036 VM_MEMATTR_DEFAULT, 3037 ("page %p has an unexpected memattr", m)); 3038 KASSERT(m->oflags == 0, 3039 ("page %p has unexpected oflags", m)); 3040 /* Don't care: PGA_NOSYNC. */ 3041 if (!vm_page_none_valid(m)) { 3042 /* 3043 * First, try to allocate a new page 3044 * that is above "high". Failing 3045 * that, try to allocate a new page 3046 * that is below "m_run". Allocate 3047 * the new page between the end of 3048 * "m_run" and "high" only as a last 3049 * resort. 3050 */ 3051 req = req_class; 3052 if ((m->flags & PG_NODUMP) != 0) 3053 req |= VM_ALLOC_NODUMP; 3054 if (trunc_page(high) != 3055 ~(vm_paddr_t)PAGE_MASK) { 3056 m_new = 3057 vm_page_alloc_noobj_contig( 3058 req, 1, round_page(high), 3059 ~(vm_paddr_t)0, PAGE_SIZE, 3060 0, VM_MEMATTR_DEFAULT); 3061 } else 3062 m_new = NULL; 3063 if (m_new == NULL) { 3064 pa = VM_PAGE_TO_PHYS(m_run); 3065 m_new = 3066 vm_page_alloc_noobj_contig( 3067 req, 1, 0, pa - 1, 3068 PAGE_SIZE, 0, 3069 VM_MEMATTR_DEFAULT); 3070 } 3071 if (m_new == NULL) { 3072 pa += ptoa(npages); 3073 m_new = 3074 vm_page_alloc_noobj_contig( 3075 req, 1, pa, high, PAGE_SIZE, 3076 0, VM_MEMATTR_DEFAULT); 3077 } 3078 if (m_new == NULL) { 3079 vm_page_xunbusy(m); 3080 error = ENOMEM; 3081 goto unlock; 3082 } 3083 3084 /* 3085 * Unmap the page and check for new 3086 * wirings that may have been acquired 3087 * through a pmap lookup. 3088 */ 3089 if (object->ref_count != 0 && 3090 !vm_page_try_remove_all(m)) { 3091 vm_page_xunbusy(m); 3092 vm_page_free(m_new); 3093 error = EBUSY; 3094 goto unlock; 3095 } 3096 3097 /* 3098 * Replace "m" with the new page. For 3099 * vm_page_replace(), "m" must be busy 3100 * and dequeued. Finally, change "m" 3101 * as if vm_page_free() was called. 3102 */ 3103 m_new->a.flags = m->a.flags & 3104 ~PGA_QUEUE_STATE_MASK; 3105 KASSERT(m_new->oflags == VPO_UNMANAGED, 3106 ("page %p is managed", m_new)); 3107 m_new->oflags = 0; 3108 pmap_copy_page(m, m_new); 3109 m_new->valid = m->valid; 3110 m_new->dirty = m->dirty; 3111 m->flags &= ~PG_ZERO; 3112 vm_page_dequeue(m); 3113 if (vm_page_replace_hold(m_new, object, 3114 m->pindex, m) && 3115 vm_page_free_prep(m)) 3116 SLIST_INSERT_HEAD(&free, m, 3117 plinks.s.ss); 3118 3119 /* 3120 * The new page must be deactivated 3121 * before the object is unlocked. 3122 */ 3123 vm_page_deactivate(m_new); 3124 } else { 3125 m->flags &= ~PG_ZERO; 3126 vm_page_dequeue(m); 3127 if (vm_page_free_prep(m)) 3128 SLIST_INSERT_HEAD(&free, m, 3129 plinks.s.ss); 3130 KASSERT(m->dirty == 0, 3131 ("page %p is dirty", m)); 3132 } 3133 } else 3134 error = EBUSY; 3135 unlock: 3136 VM_OBJECT_WUNLOCK(object); 3137 } else { 3138 MPASS(vm_page_domain(m) == domain); 3139 vmd = VM_DOMAIN(domain); 3140 vm_domain_free_lock(vmd); 3141 order = m->order; 3142 if (order < VM_NFREEORDER) { 3143 /* 3144 * The page is enqueued in the physical memory 3145 * allocator's free page queues. Moreover, it 3146 * is the first page in a power-of-two-sized 3147 * run of contiguous free pages. Jump ahead 3148 * to the last page within that run, and 3149 * continue from there. 3150 */ 3151 m += (1 << order) - 1; 3152 } 3153 #if VM_NRESERVLEVEL > 0 3154 else if (vm_reserv_is_page_free(m)) 3155 order = 0; 3156 #endif 3157 vm_domain_free_unlock(vmd); 3158 if (order == VM_NFREEORDER) 3159 error = EINVAL; 3160 } 3161 } 3162 if ((m = SLIST_FIRST(&free)) != NULL) { 3163 int cnt; 3164 3165 vmd = VM_DOMAIN(domain); 3166 cnt = 0; 3167 vm_domain_free_lock(vmd); 3168 do { 3169 MPASS(vm_page_domain(m) == domain); 3170 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 3171 vm_phys_free_pages(m, m->pool, 0); 3172 cnt++; 3173 } while ((m = SLIST_FIRST(&free)) != NULL); 3174 vm_domain_free_unlock(vmd); 3175 vm_domain_freecnt_inc(vmd, cnt); 3176 } 3177 return (error); 3178 } 3179 3180 #define NRUNS 16 3181 3182 #define RUN_INDEX(count, nruns) ((count) % (nruns)) 3183 3184 #define MIN_RECLAIM 8 3185 3186 /* 3187 * vm_page_reclaim_contig: 3188 * 3189 * Reclaim allocated, contiguous physical memory satisfying the specified 3190 * conditions by relocating the virtual pages using that physical memory. 3191 * Returns 0 if reclamation is successful, ERANGE if the specified domain 3192 * can't possibly satisfy the reclamation request, or ENOMEM if not 3193 * currently able to reclaim the requested number of pages. Since 3194 * relocation requires the allocation of physical pages, reclamation may 3195 * fail with ENOMEM due to a shortage of free pages. When reclamation 3196 * fails in this manner, callers are expected to perform vm_wait() before 3197 * retrying a failed allocation operation, e.g., vm_page_alloc_contig(). 3198 * 3199 * The caller must always specify an allocation class through "req". 3200 * 3201 * allocation classes: 3202 * VM_ALLOC_NORMAL normal process request 3203 * VM_ALLOC_SYSTEM system *really* needs a page 3204 * VM_ALLOC_INTERRUPT interrupt time request 3205 * 3206 * The optional allocation flags are ignored. 3207 * 3208 * "npages" must be greater than zero. Both "alignment" and "boundary" 3209 * must be a power of two. 3210 */ 3211 int 3212 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages, 3213 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 3214 int desired_runs) 3215 { 3216 struct vm_domain *vmd; 3217 vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs; 3218 u_long count, minalign, reclaimed; 3219 int error, i, min_reclaim, nruns, options, req_class; 3220 int segind, start_segind; 3221 int ret; 3222 3223 KASSERT(npages > 0, ("npages is 0")); 3224 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 3225 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 3226 3227 ret = ENOMEM; 3228 3229 /* 3230 * If the caller wants to reclaim multiple runs, try to allocate 3231 * space to store the runs. If that fails, fall back to the old 3232 * behavior of just reclaiming MIN_RECLAIM pages. 3233 */ 3234 if (desired_runs > 1) 3235 m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs), 3236 M_TEMP, M_NOWAIT); 3237 else 3238 m_runs = NULL; 3239 3240 if (m_runs == NULL) { 3241 m_runs = _m_runs; 3242 nruns = NRUNS; 3243 } else { 3244 nruns = NRUNS + desired_runs - 1; 3245 } 3246 min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM); 3247 3248 /* 3249 * The caller will attempt an allocation after some runs have been 3250 * reclaimed and added to the vm_phys buddy lists. Due to limitations 3251 * of vm_phys_alloc_contig(), round up the requested length to the next 3252 * power of two or maximum chunk size, and ensure that each run is 3253 * suitably aligned. 3254 */ 3255 minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1); 3256 npages = roundup2(npages, minalign); 3257 if (alignment < ptoa(minalign)) 3258 alignment = ptoa(minalign); 3259 3260 /* 3261 * The page daemon is allowed to dig deeper into the free page list. 3262 */ 3263 req_class = req & VM_ALLOC_CLASS_MASK; 3264 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 3265 req_class = VM_ALLOC_SYSTEM; 3266 3267 start_segind = vm_phys_lookup_segind(low); 3268 3269 /* 3270 * Return if the number of free pages cannot satisfy the requested 3271 * allocation. 3272 */ 3273 vmd = VM_DOMAIN(domain); 3274 count = vmd->vmd_free_count; 3275 if (count < npages + vmd->vmd_free_reserved || (count < npages + 3276 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 3277 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 3278 goto done; 3279 3280 /* 3281 * Scan up to three times, relaxing the restrictions ("options") on 3282 * the reclamation of reservations and superpages each time. 3283 */ 3284 for (options = VPSC_NORESERV;;) { 3285 bool phys_range_exists = false; 3286 3287 /* 3288 * Find the highest runs that satisfy the given constraints 3289 * and restrictions, and record them in "m_runs". 3290 */ 3291 count = 0; 3292 segind = start_segind; 3293 while ((segind = vm_phys_find_range(bounds, segind, domain, 3294 npages, low, high)) != -1) { 3295 phys_range_exists = true; 3296 while ((m_run = vm_page_scan_contig(npages, bounds[0], 3297 bounds[1], alignment, boundary, options))) { 3298 bounds[0] = m_run + npages; 3299 m_runs[RUN_INDEX(count, nruns)] = m_run; 3300 count++; 3301 } 3302 segind++; 3303 } 3304 3305 if (!phys_range_exists) { 3306 ret = ERANGE; 3307 goto done; 3308 } 3309 3310 /* 3311 * Reclaim the highest runs in LIFO (descending) order until 3312 * the number of reclaimed pages, "reclaimed", is at least 3313 * "min_reclaim". Reset "reclaimed" each time because each 3314 * reclamation is idempotent, and runs will (likely) recur 3315 * from one scan to the next as restrictions are relaxed. 3316 */ 3317 reclaimed = 0; 3318 for (i = 0; count > 0 && i < nruns; i++) { 3319 count--; 3320 m_run = m_runs[RUN_INDEX(count, nruns)]; 3321 error = vm_page_reclaim_run(req_class, domain, npages, 3322 m_run, high); 3323 if (error == 0) { 3324 reclaimed += npages; 3325 if (reclaimed >= min_reclaim) { 3326 ret = 0; 3327 goto done; 3328 } 3329 } 3330 } 3331 3332 /* 3333 * Either relax the restrictions on the next scan or return if 3334 * the last scan had no restrictions. 3335 */ 3336 if (options == VPSC_NORESERV) 3337 options = VPSC_NOSUPER; 3338 else if (options == VPSC_NOSUPER) 3339 options = VPSC_ANY; 3340 else if (options == VPSC_ANY) { 3341 if (reclaimed != 0) 3342 ret = 0; 3343 goto done; 3344 } 3345 } 3346 done: 3347 if (m_runs != _m_runs) 3348 free(m_runs, M_TEMP); 3349 return (ret); 3350 } 3351 3352 int 3353 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 3354 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 3355 { 3356 return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, 3357 high, alignment, boundary, 1)); 3358 } 3359 3360 int 3361 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 3362 u_long alignment, vm_paddr_t boundary) 3363 { 3364 struct vm_domainset_iter di; 3365 int domain, ret, status; 3366 3367 ret = ERANGE; 3368 3369 if (vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req) != 0) 3370 return (ret); 3371 3372 do { 3373 status = vm_page_reclaim_contig_domain(domain, req, npages, low, 3374 high, alignment, boundary); 3375 if (status == 0) 3376 return (0); 3377 else if (status == ERANGE) 3378 vm_domainset_iter_ignore(&di, domain); 3379 else { 3380 KASSERT(status == ENOMEM, ("Unrecognized error %d " 3381 "from vm_page_reclaim_contig_domain()", status)); 3382 ret = ENOMEM; 3383 } 3384 } while (vm_domainset_iter_page(&di, NULL, &domain, NULL) == 0); 3385 3386 return (ret); 3387 } 3388 3389 /* 3390 * Set the domain in the appropriate page level domainset. 3391 */ 3392 void 3393 vm_domain_set(struct vm_domain *vmd) 3394 { 3395 3396 mtx_lock(&vm_domainset_lock); 3397 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 3398 vmd->vmd_minset = 1; 3399 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 3400 } 3401 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 3402 vmd->vmd_severeset = 1; 3403 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 3404 } 3405 mtx_unlock(&vm_domainset_lock); 3406 } 3407 3408 /* 3409 * Clear the domain from the appropriate page level domainset. 3410 */ 3411 void 3412 vm_domain_clear(struct vm_domain *vmd) 3413 { 3414 3415 mtx_lock(&vm_domainset_lock); 3416 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 3417 vmd->vmd_minset = 0; 3418 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 3419 if (vm_min_waiters != 0) { 3420 vm_min_waiters = 0; 3421 wakeup(&vm_min_domains); 3422 } 3423 } 3424 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 3425 vmd->vmd_severeset = 0; 3426 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 3427 if (vm_severe_waiters != 0) { 3428 vm_severe_waiters = 0; 3429 wakeup(&vm_severe_domains); 3430 } 3431 } 3432 3433 /* 3434 * If pageout daemon needs pages, then tell it that there are 3435 * some free. 3436 */ 3437 if (vmd->vmd_pageout_pages_needed && 3438 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3439 wakeup(&vmd->vmd_pageout_pages_needed); 3440 vmd->vmd_pageout_pages_needed = 0; 3441 } 3442 3443 /* See comments in vm_wait_doms(). */ 3444 if (vm_pageproc_waiters) { 3445 vm_pageproc_waiters = 0; 3446 wakeup(&vm_pageproc_waiters); 3447 } 3448 mtx_unlock(&vm_domainset_lock); 3449 } 3450 3451 /* 3452 * Wait for free pages to exceed the min threshold globally. 3453 */ 3454 void 3455 vm_wait_min(void) 3456 { 3457 3458 mtx_lock(&vm_domainset_lock); 3459 while (vm_page_count_min()) { 3460 vm_min_waiters++; 3461 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 3462 } 3463 mtx_unlock(&vm_domainset_lock); 3464 } 3465 3466 /* 3467 * Wait for free pages to exceed the severe threshold globally. 3468 */ 3469 void 3470 vm_wait_severe(void) 3471 { 3472 3473 mtx_lock(&vm_domainset_lock); 3474 while (vm_page_count_severe()) { 3475 vm_severe_waiters++; 3476 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 3477 "vmwait", 0); 3478 } 3479 mtx_unlock(&vm_domainset_lock); 3480 } 3481 3482 u_int 3483 vm_wait_count(void) 3484 { 3485 3486 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 3487 } 3488 3489 int 3490 vm_wait_doms(const domainset_t *wdoms, int mflags) 3491 { 3492 int error; 3493 3494 error = 0; 3495 3496 /* 3497 * We use racey wakeup synchronization to avoid expensive global 3498 * locking for the pageproc when sleeping with a non-specific vm_wait. 3499 * To handle this, we only sleep for one tick in this instance. It 3500 * is expected that most allocations for the pageproc will come from 3501 * kmem or vm_page_grab* which will use the more specific and 3502 * race-free vm_wait_domain(). 3503 */ 3504 if (curproc == pageproc) { 3505 mtx_lock(&vm_domainset_lock); 3506 vm_pageproc_waiters++; 3507 error = msleep(&vm_pageproc_waiters, &vm_domainset_lock, 3508 PVM | PDROP | mflags, "pageprocwait", 1); 3509 } else { 3510 /* 3511 * XXX Ideally we would wait only until the allocation could 3512 * be satisfied. This condition can cause new allocators to 3513 * consume all freed pages while old allocators wait. 3514 */ 3515 mtx_lock(&vm_domainset_lock); 3516 if (vm_page_count_min_set(wdoms)) { 3517 if (pageproc == NULL) 3518 panic("vm_wait in early boot"); 3519 vm_min_waiters++; 3520 error = msleep(&vm_min_domains, &vm_domainset_lock, 3521 PVM | PDROP | mflags, "vmwait", 0); 3522 } else 3523 mtx_unlock(&vm_domainset_lock); 3524 } 3525 return (error); 3526 } 3527 3528 /* 3529 * vm_wait_domain: 3530 * 3531 * Sleep until free pages are available for allocation. 3532 * - Called in various places after failed memory allocations. 3533 */ 3534 void 3535 vm_wait_domain(int domain) 3536 { 3537 struct vm_domain *vmd; 3538 domainset_t wdom; 3539 3540 vmd = VM_DOMAIN(domain); 3541 vm_domain_free_assert_unlocked(vmd); 3542 3543 if (curproc == pageproc) { 3544 mtx_lock(&vm_domainset_lock); 3545 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 3546 vmd->vmd_pageout_pages_needed = 1; 3547 msleep(&vmd->vmd_pageout_pages_needed, 3548 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 3549 } else 3550 mtx_unlock(&vm_domainset_lock); 3551 } else { 3552 DOMAINSET_ZERO(&wdom); 3553 DOMAINSET_SET(vmd->vmd_domain, &wdom); 3554 vm_wait_doms(&wdom, 0); 3555 } 3556 } 3557 3558 static int 3559 vm_wait_flags(vm_object_t obj, int mflags) 3560 { 3561 struct domainset *d; 3562 3563 d = NULL; 3564 3565 /* 3566 * Carefully fetch pointers only once: the struct domainset 3567 * itself is ummutable but the pointer might change. 3568 */ 3569 if (obj != NULL) 3570 d = obj->domain.dr_policy; 3571 if (d == NULL) 3572 d = curthread->td_domain.dr_policy; 3573 3574 return (vm_wait_doms(&d->ds_mask, mflags)); 3575 } 3576 3577 /* 3578 * vm_wait: 3579 * 3580 * Sleep until free pages are available for allocation in the 3581 * affinity domains of the obj. If obj is NULL, the domain set 3582 * for the calling thread is used. 3583 * Called in various places after failed memory allocations. 3584 */ 3585 void 3586 vm_wait(vm_object_t obj) 3587 { 3588 (void)vm_wait_flags(obj, 0); 3589 } 3590 3591 int 3592 vm_wait_intr(vm_object_t obj) 3593 { 3594 return (vm_wait_flags(obj, PCATCH)); 3595 } 3596 3597 /* 3598 * vm_domain_alloc_fail: 3599 * 3600 * Called when a page allocation function fails. Informs the 3601 * pagedaemon and performs the requested wait. Requires the 3602 * domain_free and object lock on entry. Returns with the 3603 * object lock held and free lock released. Returns an error when 3604 * retry is necessary. 3605 * 3606 */ 3607 static int 3608 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3609 { 3610 3611 vm_domain_free_assert_unlocked(vmd); 3612 3613 atomic_add_int(&vmd->vmd_pageout_deficit, 3614 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3615 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3616 if (object != NULL) 3617 VM_OBJECT_WUNLOCK(object); 3618 vm_wait_domain(vmd->vmd_domain); 3619 if (object != NULL) 3620 VM_OBJECT_WLOCK(object); 3621 if (req & VM_ALLOC_WAITOK) 3622 return (EAGAIN); 3623 } 3624 3625 return (0); 3626 } 3627 3628 /* 3629 * vm_waitpfault: 3630 * 3631 * Sleep until free pages are available for allocation. 3632 * - Called only in vm_fault so that processes page faulting 3633 * can be easily tracked. 3634 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3635 * processes will be able to grab memory first. Do not change 3636 * this balance without careful testing first. 3637 */ 3638 void 3639 vm_waitpfault(struct domainset *dset, int timo) 3640 { 3641 3642 /* 3643 * XXX Ideally we would wait only until the allocation could 3644 * be satisfied. This condition can cause new allocators to 3645 * consume all freed pages while old allocators wait. 3646 */ 3647 mtx_lock(&vm_domainset_lock); 3648 if (vm_page_count_min_set(&dset->ds_mask)) { 3649 vm_min_waiters++; 3650 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3651 "pfault", timo); 3652 } else 3653 mtx_unlock(&vm_domainset_lock); 3654 } 3655 3656 static struct vm_pagequeue * 3657 _vm_page_pagequeue(vm_page_t m, uint8_t queue) 3658 { 3659 3660 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); 3661 } 3662 3663 #ifdef INVARIANTS 3664 static struct vm_pagequeue * 3665 vm_page_pagequeue(vm_page_t m) 3666 { 3667 3668 return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); 3669 } 3670 #endif 3671 3672 static __always_inline bool 3673 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, 3674 vm_page_astate_t new) 3675 { 3676 vm_page_astate_t tmp; 3677 3678 tmp = *old; 3679 do { 3680 if (__predict_true(vm_page_astate_fcmpset(m, old, new))) 3681 return (true); 3682 counter_u64_add(pqstate_commit_retries, 1); 3683 } while (old->_bits == tmp._bits); 3684 3685 return (false); 3686 } 3687 3688 static __always_inline bool 3689 vm_page_pqstate_fcmpset_rel(vm_page_t m, vm_page_astate_t *old, 3690 vm_page_astate_t new) 3691 { 3692 vm_page_astate_t tmp; 3693 3694 tmp = *old; 3695 do { 3696 if (__predict_true(vm_page_astate_fcmpset_rel(m, old, new))) 3697 return (true); 3698 counter_u64_add(pqstate_commit_retries, 1); 3699 } while (old->_bits == tmp._bits); 3700 3701 return (false); 3702 } 3703 3704 /* 3705 * Do the work of committing a queue state update that moves the page out of 3706 * its current queue. 3707 */ 3708 static bool 3709 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, 3710 vm_page_astate_t *old, vm_page_astate_t new) 3711 { 3712 vm_page_t next; 3713 3714 vm_pagequeue_assert_locked(pq); 3715 KASSERT(vm_page_pagequeue(m) == pq, 3716 ("%s: queue %p does not match page %p", __func__, pq, m)); 3717 KASSERT(old->queue != PQ_NONE && new.queue != old->queue, 3718 ("%s: invalid queue indices %d %d", 3719 __func__, old->queue, new.queue)); 3720 3721 /* 3722 * Once the queue index of the page changes there is nothing 3723 * synchronizing with further updates to the page's physical 3724 * queue state. Therefore we must speculatively remove the page 3725 * from the queue now and be prepared to roll back if the queue 3726 * state update fails. If the page is not physically enqueued then 3727 * we just update its queue index. 3728 */ 3729 if ((old->flags & PGA_ENQUEUED) != 0) { 3730 new.flags &= ~PGA_ENQUEUED; 3731 next = TAILQ_NEXT(m, plinks.q); 3732 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3733 vm_pagequeue_cnt_dec(pq); 3734 /* See vm_page_dequeue(). */ 3735 if (!vm_page_pqstate_fcmpset_rel(m, old, new)) { 3736 if (next == NULL) 3737 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3738 else 3739 TAILQ_INSERT_BEFORE(next, m, plinks.q); 3740 vm_pagequeue_cnt_inc(pq); 3741 return (false); 3742 } else { 3743 return (true); 3744 } 3745 } else { 3746 return (vm_page_pqstate_fcmpset(m, old, new)); 3747 } 3748 } 3749 3750 static bool 3751 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, 3752 vm_page_astate_t new) 3753 { 3754 struct vm_pagequeue *pq; 3755 vm_page_astate_t as; 3756 bool ret; 3757 3758 pq = _vm_page_pagequeue(m, old->queue); 3759 3760 /* 3761 * The queue field and PGA_ENQUEUED flag are stable only so long as the 3762 * corresponding page queue lock is held. 3763 */ 3764 vm_pagequeue_lock(pq); 3765 as = vm_page_astate_load(m); 3766 if (__predict_false(as._bits != old->_bits)) { 3767 *old = as; 3768 ret = false; 3769 } else { 3770 ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); 3771 } 3772 vm_pagequeue_unlock(pq); 3773 return (ret); 3774 } 3775 3776 /* 3777 * Commit a queue state update that enqueues or requeues a page. 3778 */ 3779 static bool 3780 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, 3781 vm_page_astate_t *old, vm_page_astate_t new) 3782 { 3783 struct vm_domain *vmd; 3784 3785 vm_pagequeue_assert_locked(pq); 3786 KASSERT(old->queue != PQ_NONE && new.queue == old->queue, 3787 ("%s: invalid queue indices %d %d", 3788 __func__, old->queue, new.queue)); 3789 3790 new.flags |= PGA_ENQUEUED; 3791 if (!vm_page_pqstate_fcmpset(m, old, new)) 3792 return (false); 3793 3794 if ((old->flags & PGA_ENQUEUED) != 0) 3795 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3796 else 3797 vm_pagequeue_cnt_inc(pq); 3798 3799 /* 3800 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if 3801 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be 3802 * applied, even if it was set first. 3803 */ 3804 if ((old->flags & PGA_REQUEUE_HEAD) != 0) { 3805 vmd = vm_pagequeue_domain(m); 3806 KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], 3807 ("%s: invalid page queue for page %p", __func__, m)); 3808 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3809 } else { 3810 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3811 } 3812 return (true); 3813 } 3814 3815 /* 3816 * Commit a queue state update that encodes a request for a deferred queue 3817 * operation. 3818 */ 3819 static bool 3820 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, 3821 vm_page_astate_t new) 3822 { 3823 3824 KASSERT(old->queue == new.queue || new.queue != PQ_NONE, 3825 ("%s: invalid state, queue %d flags %x", 3826 __func__, new.queue, new.flags)); 3827 3828 if (old->_bits != new._bits && 3829 !vm_page_pqstate_fcmpset(m, old, new)) 3830 return (false); 3831 vm_page_pqbatch_submit(m, new.queue); 3832 return (true); 3833 } 3834 3835 /* 3836 * A generic queue state update function. This handles more cases than the 3837 * specialized functions above. 3838 */ 3839 bool 3840 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) 3841 { 3842 3843 if (old->_bits == new._bits) 3844 return (true); 3845 3846 if (old->queue != PQ_NONE && new.queue != old->queue) { 3847 if (!vm_page_pqstate_commit_dequeue(m, old, new)) 3848 return (false); 3849 if (new.queue != PQ_NONE) 3850 vm_page_pqbatch_submit(m, new.queue); 3851 } else { 3852 if (!vm_page_pqstate_fcmpset(m, old, new)) 3853 return (false); 3854 if (new.queue != PQ_NONE && 3855 ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) 3856 vm_page_pqbatch_submit(m, new.queue); 3857 } 3858 return (true); 3859 } 3860 3861 /* 3862 * Apply deferred queue state updates to a page. 3863 */ 3864 static inline void 3865 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) 3866 { 3867 vm_page_astate_t new, old; 3868 3869 CRITICAL_ASSERT(curthread); 3870 vm_pagequeue_assert_locked(pq); 3871 KASSERT(queue < PQ_COUNT, 3872 ("%s: invalid queue index %d", __func__, queue)); 3873 KASSERT(pq == _vm_page_pagequeue(m, queue), 3874 ("%s: page %p does not belong to queue %p", __func__, m, pq)); 3875 3876 for (old = vm_page_astate_load(m);;) { 3877 if (__predict_false(old.queue != queue || 3878 (old.flags & PGA_QUEUE_OP_MASK) == 0)) { 3879 counter_u64_add(queue_nops, 1); 3880 break; 3881 } 3882 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3883 ("%s: page %p is unmanaged", __func__, m)); 3884 3885 new = old; 3886 if ((old.flags & PGA_DEQUEUE) != 0) { 3887 new.flags &= ~PGA_QUEUE_OP_MASK; 3888 new.queue = PQ_NONE; 3889 if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, 3890 m, &old, new))) { 3891 counter_u64_add(queue_ops, 1); 3892 break; 3893 } 3894 } else { 3895 new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); 3896 if (__predict_true(_vm_page_pqstate_commit_requeue(pq, 3897 m, &old, new))) { 3898 counter_u64_add(queue_ops, 1); 3899 break; 3900 } 3901 } 3902 } 3903 } 3904 3905 static void 3906 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3907 uint8_t queue) 3908 { 3909 int i; 3910 3911 for (i = 0; i < bq->bq_cnt; i++) 3912 vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); 3913 vm_batchqueue_init(bq); 3914 } 3915 3916 /* 3917 * vm_page_pqbatch_submit: [ internal use only ] 3918 * 3919 * Enqueue a page in the specified page queue's batched work queue. 3920 * The caller must have encoded the requested operation in the page 3921 * structure's a.flags field. 3922 */ 3923 void 3924 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) 3925 { 3926 struct vm_batchqueue *bq; 3927 struct vm_pagequeue *pq; 3928 int domain, slots_remaining; 3929 3930 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3931 3932 domain = vm_page_domain(m); 3933 critical_enter(); 3934 bq = DPCPU_PTR(pqbatch[domain][queue]); 3935 slots_remaining = vm_batchqueue_insert(bq, m); 3936 if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) { 3937 /* keep building the bq */ 3938 critical_exit(); 3939 return; 3940 } else if (slots_remaining > 0 ) { 3941 /* Try to process the bq if we can get the lock */ 3942 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3943 if (vm_pagequeue_trylock(pq)) { 3944 vm_pqbatch_process(pq, bq, queue); 3945 vm_pagequeue_unlock(pq); 3946 } 3947 critical_exit(); 3948 return; 3949 } 3950 critical_exit(); 3951 3952 /* if we make it here, the bq is full so wait for the lock */ 3953 3954 pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue]; 3955 vm_pagequeue_lock(pq); 3956 critical_enter(); 3957 bq = DPCPU_PTR(pqbatch[domain][queue]); 3958 vm_pqbatch_process(pq, bq, queue); 3959 vm_pqbatch_process_page(pq, m, queue); 3960 vm_pagequeue_unlock(pq); 3961 critical_exit(); 3962 } 3963 3964 /* 3965 * vm_page_pqbatch_drain: [ internal use only ] 3966 * 3967 * Force all per-CPU page queue batch queues to be drained. This is 3968 * intended for use in severe memory shortages, to ensure that pages 3969 * do not remain stuck in the batch queues. 3970 */ 3971 void 3972 vm_page_pqbatch_drain(void) 3973 { 3974 struct thread *td; 3975 struct vm_domain *vmd; 3976 struct vm_pagequeue *pq; 3977 int cpu, domain, queue; 3978 3979 td = curthread; 3980 CPU_FOREACH(cpu) { 3981 thread_lock(td); 3982 sched_bind(td, cpu); 3983 thread_unlock(td); 3984 3985 for (domain = 0; domain < vm_ndomains; domain++) { 3986 vmd = VM_DOMAIN(domain); 3987 for (queue = 0; queue < PQ_COUNT; queue++) { 3988 pq = &vmd->vmd_pagequeues[queue]; 3989 vm_pagequeue_lock(pq); 3990 critical_enter(); 3991 vm_pqbatch_process(pq, 3992 DPCPU_PTR(pqbatch[domain][queue]), queue); 3993 critical_exit(); 3994 vm_pagequeue_unlock(pq); 3995 } 3996 } 3997 } 3998 thread_lock(td); 3999 sched_unbind(td); 4000 thread_unlock(td); 4001 } 4002 4003 /* 4004 * vm_page_dequeue_deferred: [ internal use only ] 4005 * 4006 * Request removal of the given page from its current page 4007 * queue. Physical removal from the queue may be deferred 4008 * indefinitely. 4009 */ 4010 void 4011 vm_page_dequeue_deferred(vm_page_t m) 4012 { 4013 vm_page_astate_t new, old; 4014 4015 old = vm_page_astate_load(m); 4016 do { 4017 if (old.queue == PQ_NONE) { 4018 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4019 ("%s: page %p has unexpected queue state", 4020 __func__, m)); 4021 break; 4022 } 4023 new = old; 4024 new.flags |= PGA_DEQUEUE; 4025 } while (!vm_page_pqstate_commit_request(m, &old, new)); 4026 } 4027 4028 /* 4029 * vm_page_dequeue: 4030 * 4031 * Remove the page from whichever page queue it's in, if any, before 4032 * returning. 4033 */ 4034 void 4035 vm_page_dequeue(vm_page_t m) 4036 { 4037 vm_page_astate_t new, old; 4038 4039 /* 4040 * Synchronize with _vm_page_pqstate_commit_dequeue(): make sure 4041 * that the page's queue linkage field updates are visible before 4042 * returning. 4043 */ 4044 old = vm_page_astate_load_acq(m); 4045 do { 4046 if (__predict_true(old.queue == PQ_NONE)) { 4047 KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, 4048 ("%s: page %p has unexpected queue state", 4049 __func__, m)); 4050 break; 4051 } 4052 new = old; 4053 new.flags &= ~PGA_QUEUE_OP_MASK; 4054 new.queue = PQ_NONE; 4055 } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); 4056 4057 } 4058 4059 /* 4060 * Schedule the given page for insertion into the specified page queue. 4061 * Physical insertion of the page may be deferred indefinitely. 4062 */ 4063 static void 4064 vm_page_enqueue(vm_page_t m, uint8_t queue) 4065 { 4066 4067 KASSERT(m->a.queue == PQ_NONE && 4068 (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, 4069 ("%s: page %p is already enqueued", __func__, m)); 4070 KASSERT(m->ref_count > 0, 4071 ("%s: page %p does not carry any references", __func__, m)); 4072 4073 m->a.queue = queue; 4074 if ((m->a.flags & PGA_REQUEUE) == 0) 4075 vm_page_aflag_set(m, PGA_REQUEUE); 4076 vm_page_pqbatch_submit(m, queue); 4077 } 4078 4079 /* 4080 * vm_page_free_prep: 4081 * 4082 * Prepares the given page to be put on the free list, 4083 * disassociating it from any VM object. The caller may return 4084 * the page to the free list only if this function returns true. 4085 * 4086 * The object, if it exists, must be locked, and then the page must 4087 * be xbusy. Otherwise the page must be not busied. A managed 4088 * page must be unmapped. 4089 */ 4090 static bool 4091 vm_page_free_prep(vm_page_t m) 4092 { 4093 4094 /* 4095 * Synchronize with threads that have dropped a reference to this 4096 * page. 4097 */ 4098 atomic_thread_fence_acq(); 4099 4100 #ifdef INVARIANTS 4101 if (vm_check_pg_zero && (m->flags & PG_ZERO) != 0) { 4102 struct sf_buf *sf; 4103 unsigned long *p; 4104 int i; 4105 4106 sched_pin(); 4107 sf = sf_buf_alloc(m, SFB_CPUPRIVATE | SFB_NOWAIT); 4108 if (sf != NULL) { 4109 p = (unsigned long *)sf_buf_kva(sf); 4110 for (i = 0; i < PAGE_SIZE / sizeof(*p); i++, p++) { 4111 KASSERT(*p == 0, 4112 ("zerocheck failed page %p PG_ZERO %d %jx", 4113 m, i, (uintmax_t)*p)); 4114 } 4115 sf_buf_free(sf); 4116 } 4117 sched_unpin(); 4118 } 4119 #endif 4120 if ((m->oflags & VPO_UNMANAGED) == 0) { 4121 KASSERT(!pmap_page_is_mapped(m), 4122 ("vm_page_free_prep: freeing mapped page %p", m)); 4123 KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, 4124 ("vm_page_free_prep: mapping flags set in page %p", m)); 4125 } else { 4126 KASSERT(m->a.queue == PQ_NONE, 4127 ("vm_page_free_prep: unmanaged page %p is queued", m)); 4128 } 4129 VM_CNT_INC(v_tfree); 4130 4131 if (m->object != NULL) { 4132 vm_page_radix_remove(m); 4133 vm_page_free_object_prep(m); 4134 } else 4135 vm_page_assert_unbusied(m); 4136 4137 vm_page_busy_free(m); 4138 4139 /* 4140 * If fictitious remove object association and 4141 * return. 4142 */ 4143 if ((m->flags & PG_FICTITIOUS) != 0) { 4144 KASSERT(m->ref_count == 1, 4145 ("fictitious page %p is referenced", m)); 4146 KASSERT(m->a.queue == PQ_NONE, 4147 ("fictitious page %p is queued", m)); 4148 return (false); 4149 } 4150 4151 /* 4152 * Pages need not be dequeued before they are returned to the physical 4153 * memory allocator, but they must at least be marked for a deferred 4154 * dequeue. 4155 */ 4156 if ((m->oflags & VPO_UNMANAGED) == 0) 4157 vm_page_dequeue_deferred(m); 4158 4159 m->valid = 0; 4160 vm_page_undirty(m); 4161 4162 if (m->ref_count != 0) 4163 panic("vm_page_free_prep: page %p has references", m); 4164 4165 /* 4166 * Restore the default memory attribute to the page. 4167 */ 4168 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 4169 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 4170 4171 #if VM_NRESERVLEVEL > 0 4172 /* 4173 * Determine whether the page belongs to a reservation. If the page was 4174 * allocated from a per-CPU cache, it cannot belong to a reservation, so 4175 * as an optimization, we avoid the check in that case. 4176 */ 4177 if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) 4178 return (false); 4179 #endif 4180 4181 return (true); 4182 } 4183 4184 /* 4185 * vm_page_free_toq: 4186 * 4187 * Returns the given page to the free list, disassociating it 4188 * from any VM object. 4189 * 4190 * The object must be locked. The page must be exclusively busied if it 4191 * belongs to an object. 4192 */ 4193 static void 4194 vm_page_free_toq(vm_page_t m) 4195 { 4196 struct vm_domain *vmd; 4197 uma_zone_t zone; 4198 4199 if (!vm_page_free_prep(m)) 4200 return; 4201 4202 vmd = vm_pagequeue_domain(m); 4203 if (__predict_false((m->flags & PG_NOFREE) != 0)) { 4204 vm_page_free_nofree(vmd, m); 4205 return; 4206 } 4207 zone = vmd->vmd_pgcache[m->pool].zone; 4208 if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { 4209 uma_zfree(zone, m); 4210 return; 4211 } 4212 vm_domain_free_lock(vmd); 4213 vm_phys_free_pages(m, m->pool, 0); 4214 vm_domain_free_unlock(vmd); 4215 vm_domain_freecnt_inc(vmd, 1); 4216 } 4217 4218 /* 4219 * vm_page_free_pages_toq: 4220 * 4221 * Returns a list of pages to the free list, disassociating it 4222 * from any VM object. In other words, this is equivalent to 4223 * calling vm_page_free_toq() for each page of a list of VM objects. 4224 */ 4225 int 4226 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 4227 { 4228 vm_page_t m; 4229 int count; 4230 4231 if (SLIST_EMPTY(free)) 4232 return (0); 4233 4234 count = 0; 4235 while ((m = SLIST_FIRST(free)) != NULL) { 4236 count++; 4237 SLIST_REMOVE_HEAD(free, plinks.s.ss); 4238 vm_page_free_toq(m); 4239 } 4240 4241 if (update_wire_count) 4242 vm_wire_sub(count); 4243 return (count); 4244 } 4245 4246 /* 4247 * Mark this page as wired down. For managed pages, this prevents reclamation 4248 * by the page daemon, or when the containing object, if any, is destroyed. 4249 */ 4250 void 4251 vm_page_wire(vm_page_t m) 4252 { 4253 u_int old; 4254 4255 #ifdef INVARIANTS 4256 if (m->object != NULL && !vm_page_busied(m) && 4257 !vm_object_busied(m->object)) 4258 VM_OBJECT_ASSERT_LOCKED(m->object); 4259 #endif 4260 KASSERT((m->flags & PG_FICTITIOUS) == 0 || 4261 VPRC_WIRE_COUNT(m->ref_count) >= 1, 4262 ("vm_page_wire: fictitious page %p has zero wirings", m)); 4263 4264 old = atomic_fetchadd_int(&m->ref_count, 1); 4265 KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, 4266 ("vm_page_wire: counter overflow for page %p", m)); 4267 if (VPRC_WIRE_COUNT(old) == 0) { 4268 if ((m->oflags & VPO_UNMANAGED) == 0) 4269 vm_page_aflag_set(m, PGA_DEQUEUE); 4270 vm_wire_add(1); 4271 } 4272 } 4273 4274 /* 4275 * Attempt to wire a mapped page following a pmap lookup of that page. 4276 * This may fail if a thread is concurrently tearing down mappings of the page. 4277 * The transient failure is acceptable because it translates to the 4278 * failure of the caller pmap_extract_and_hold(), which should be then 4279 * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). 4280 */ 4281 bool 4282 vm_page_wire_mapped(vm_page_t m) 4283 { 4284 u_int old; 4285 4286 old = atomic_load_int(&m->ref_count); 4287 do { 4288 KASSERT(old > 0, 4289 ("vm_page_wire_mapped: wiring unreferenced page %p", m)); 4290 if ((old & VPRC_BLOCKED) != 0) 4291 return (false); 4292 } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); 4293 4294 if (VPRC_WIRE_COUNT(old) == 0) { 4295 if ((m->oflags & VPO_UNMANAGED) == 0) 4296 vm_page_aflag_set(m, PGA_DEQUEUE); 4297 vm_wire_add(1); 4298 } 4299 return (true); 4300 } 4301 4302 /* 4303 * Release a wiring reference to a managed page. If the page still belongs to 4304 * an object, update its position in the page queues to reflect the reference. 4305 * If the wiring was the last reference to the page, free the page. 4306 */ 4307 static void 4308 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) 4309 { 4310 u_int old; 4311 4312 KASSERT(nqueue < PQ_COUNT, 4313 ("vm_page_unwire: invalid queue %u request for page %p", 4314 nqueue, m)); 4315 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4316 ("%s: page %p is unmanaged", __func__, m)); 4317 4318 /* 4319 * Update LRU state before releasing the wiring reference. 4320 * Use a release store when updating the reference count to 4321 * synchronize with vm_page_free_prep(). 4322 */ 4323 old = atomic_load_int(&m->ref_count); 4324 do { 4325 u_int count; 4326 4327 KASSERT(VPRC_WIRE_COUNT(old) > 0, 4328 ("vm_page_unwire: wire count underflow for page %p", m)); 4329 4330 count = old & ~VPRC_BLOCKED; 4331 if (count > VPRC_OBJREF + 1) { 4332 /* 4333 * The page has at least one other wiring reference. An 4334 * earlier iteration of this loop may have called 4335 * vm_page_release_toq() and cleared PGA_DEQUEUE, so 4336 * re-set it if necessary. 4337 */ 4338 if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) 4339 vm_page_aflag_set(m, PGA_DEQUEUE); 4340 } else if (count == VPRC_OBJREF + 1) { 4341 /* 4342 * This is the last wiring. Clear PGA_DEQUEUE and 4343 * update the page's queue state to reflect the 4344 * reference. If the page does not belong to an object 4345 * (i.e., the VPRC_OBJREF bit is clear), we only need to 4346 * clear leftover queue state. 4347 */ 4348 vm_page_release_toq(m, nqueue, noreuse); 4349 } else if (count == 1) { 4350 vm_page_aflag_clear(m, PGA_DEQUEUE); 4351 } 4352 } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); 4353 4354 if (VPRC_WIRE_COUNT(old) == 1) { 4355 vm_wire_sub(1); 4356 if (old == 1) 4357 vm_page_free(m); 4358 } 4359 } 4360 4361 /* 4362 * Release one wiring of the specified page, potentially allowing it to be 4363 * paged out. 4364 * 4365 * Only managed pages belonging to an object can be paged out. If the number 4366 * of wirings transitions to zero and the page is eligible for page out, then 4367 * the page is added to the specified paging queue. If the released wiring 4368 * represented the last reference to the page, the page is freed. 4369 */ 4370 void 4371 vm_page_unwire(vm_page_t m, uint8_t nqueue) 4372 { 4373 KASSERT(nqueue < PQ_COUNT || nqueue == PQ_NONE, 4374 ("%s: invalid queue %u request for page %p", __func__, nqueue, m)); 4375 4376 if ((m->oflags & VPO_UNMANAGED) != 0) { 4377 if (vm_page_unwire_noq(m) && m->ref_count == 0) 4378 vm_page_free(m); 4379 } else { 4380 vm_page_unwire_managed(m, nqueue, false); 4381 } 4382 } 4383 4384 /* 4385 * Unwire a page without (re-)inserting it into a page queue. It is up 4386 * to the caller to enqueue, requeue, or free the page as appropriate. 4387 * In most cases involving managed pages, vm_page_unwire() should be used 4388 * instead. 4389 */ 4390 bool 4391 vm_page_unwire_noq(vm_page_t m) 4392 { 4393 u_int old; 4394 4395 old = vm_page_drop(m, 1); 4396 KASSERT(VPRC_WIRE_COUNT(old) != 0, 4397 ("%s: counter underflow for page %p", __func__, m)); 4398 KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, 4399 ("%s: missing ref on fictitious page %p", __func__, m)); 4400 4401 if (VPRC_WIRE_COUNT(old) > 1) 4402 return (false); 4403 if ((m->oflags & VPO_UNMANAGED) == 0) 4404 vm_page_aflag_clear(m, PGA_DEQUEUE); 4405 vm_wire_sub(1); 4406 return (true); 4407 } 4408 4409 /* 4410 * Ensure that the page ends up in the specified page queue. If the page is 4411 * active or being moved to the active queue, ensure that its act_count is 4412 * at least ACT_INIT but do not otherwise mess with it. 4413 */ 4414 static __always_inline void 4415 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) 4416 { 4417 vm_page_astate_t old, new; 4418 4419 KASSERT(m->ref_count > 0, 4420 ("%s: page %p does not carry any references", __func__, m)); 4421 KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, 4422 ("%s: invalid flags %x", __func__, nflag)); 4423 4424 if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m)) 4425 return; 4426 4427 old = vm_page_astate_load(m); 4428 do { 4429 if ((old.flags & PGA_DEQUEUE) != 0) 4430 break; 4431 new = old; 4432 new.flags &= ~PGA_QUEUE_OP_MASK; 4433 if (nqueue == PQ_ACTIVE) 4434 new.act_count = max(old.act_count, ACT_INIT); 4435 if (old.queue == nqueue) { 4436 /* 4437 * There is no need to requeue pages already in the 4438 * active queue. 4439 */ 4440 if (nqueue != PQ_ACTIVE || 4441 (old.flags & PGA_ENQUEUED) == 0) 4442 new.flags |= nflag; 4443 } else { 4444 new.flags |= nflag; 4445 new.queue = nqueue; 4446 } 4447 } while (!vm_page_pqstate_commit(m, &old, new)); 4448 } 4449 4450 /* 4451 * Put the specified page on the active list (if appropriate). 4452 */ 4453 void 4454 vm_page_activate(vm_page_t m) 4455 { 4456 4457 vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); 4458 } 4459 4460 /* 4461 * Move the specified page to the tail of the inactive queue, or requeue 4462 * the page if it is already in the inactive queue. 4463 */ 4464 void 4465 vm_page_deactivate(vm_page_t m) 4466 { 4467 4468 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); 4469 } 4470 4471 void 4472 vm_page_deactivate_noreuse(vm_page_t m) 4473 { 4474 4475 vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); 4476 } 4477 4478 /* 4479 * Put a page in the laundry, or requeue it if it is already there. 4480 */ 4481 void 4482 vm_page_launder(vm_page_t m) 4483 { 4484 4485 vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); 4486 } 4487 4488 /* 4489 * Put a page in the PQ_UNSWAPPABLE holding queue. 4490 */ 4491 void 4492 vm_page_unswappable(vm_page_t m) 4493 { 4494 4495 VM_OBJECT_ASSERT_LOCKED(m->object); 4496 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4497 ("page %p already unswappable", m)); 4498 4499 vm_page_dequeue(m); 4500 vm_page_enqueue(m, PQ_UNSWAPPABLE); 4501 } 4502 4503 /* 4504 * Release a page back to the page queues in preparation for unwiring. 4505 */ 4506 static void 4507 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) 4508 { 4509 vm_page_astate_t old, new; 4510 uint16_t nflag; 4511 4512 /* 4513 * Use a check of the valid bits to determine whether we should 4514 * accelerate reclamation of the page. The object lock might not be 4515 * held here, in which case the check is racy. At worst we will either 4516 * accelerate reclamation of a valid page and violate LRU, or 4517 * unnecessarily defer reclamation of an invalid page. 4518 * 4519 * If we were asked to not cache the page, place it near the head of the 4520 * inactive queue so that is reclaimed sooner. 4521 */ 4522 if (noreuse || vm_page_none_valid(m)) { 4523 nqueue = PQ_INACTIVE; 4524 nflag = PGA_REQUEUE_HEAD; 4525 } else { 4526 nflag = PGA_REQUEUE; 4527 } 4528 4529 old = vm_page_astate_load(m); 4530 do { 4531 new = old; 4532 4533 /* 4534 * If the page is already in the active queue and we are not 4535 * trying to accelerate reclamation, simply mark it as 4536 * referenced and avoid any queue operations. 4537 */ 4538 new.flags &= ~PGA_QUEUE_OP_MASK; 4539 if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE && 4540 (old.flags & PGA_ENQUEUED) != 0) 4541 new.flags |= PGA_REFERENCED; 4542 else { 4543 new.flags |= nflag; 4544 new.queue = nqueue; 4545 } 4546 } while (!vm_page_pqstate_commit(m, &old, new)); 4547 } 4548 4549 /* 4550 * Unwire a page and either attempt to free it or re-add it to the page queues. 4551 */ 4552 void 4553 vm_page_release(vm_page_t m, int flags) 4554 { 4555 if ((m->oflags & VPO_UNMANAGED) != 0) { 4556 vm_page_unwire(m, PQ_NONE); 4557 return; 4558 } 4559 4560 if ((flags & VPR_TRYFREE) != 0) { 4561 for (;;) { 4562 vm_object_t object; 4563 4564 object = atomic_load_ptr(&m->object); 4565 if (object == NULL) 4566 break; 4567 /* Depends on type-stability. */ 4568 if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) 4569 break; 4570 if (object == m->object) { 4571 vm_page_release_locked(m, flags); 4572 VM_OBJECT_WUNLOCK(object); 4573 return; 4574 } 4575 VM_OBJECT_WUNLOCK(object); 4576 } 4577 } 4578 vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); 4579 } 4580 4581 /* See vm_page_release(). */ 4582 void 4583 vm_page_release_locked(vm_page_t m, int flags) 4584 { 4585 4586 VM_OBJECT_ASSERT_WLOCKED(m->object); 4587 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4588 ("vm_page_release_locked: page %p is unmanaged", m)); 4589 4590 if (vm_page_unwire_noq(m)) { 4591 if ((flags & VPR_TRYFREE) != 0 && 4592 (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && 4593 m->dirty == 0 && vm_page_tryxbusy(m)) { 4594 /* 4595 * An unlocked lookup may have wired the page before the 4596 * busy lock was acquired, in which case the page must 4597 * not be freed. 4598 */ 4599 if (__predict_true(!vm_page_wired(m))) { 4600 vm_page_free(m); 4601 return; 4602 } 4603 vm_page_xunbusy(m); 4604 } else { 4605 vm_page_release_toq(m, PQ_INACTIVE, flags != 0); 4606 } 4607 } 4608 } 4609 4610 static bool 4611 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) 4612 { 4613 u_int old; 4614 4615 KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, 4616 ("vm_page_try_blocked_op: page %p has no object", m)); 4617 KASSERT(vm_page_busied(m), 4618 ("vm_page_try_blocked_op: page %p is not busy", m)); 4619 VM_OBJECT_ASSERT_LOCKED(m->object); 4620 4621 old = atomic_load_int(&m->ref_count); 4622 do { 4623 KASSERT(old != 0, 4624 ("vm_page_try_blocked_op: page %p has no references", m)); 4625 KASSERT((old & VPRC_BLOCKED) == 0, 4626 ("vm_page_try_blocked_op: page %p blocks wirings", m)); 4627 if (VPRC_WIRE_COUNT(old) != 0) 4628 return (false); 4629 } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); 4630 4631 (op)(m); 4632 4633 /* 4634 * If the object is read-locked, new wirings may be created via an 4635 * object lookup. 4636 */ 4637 old = vm_page_drop(m, VPRC_BLOCKED); 4638 KASSERT(!VM_OBJECT_WOWNED(m->object) || 4639 old == (VPRC_BLOCKED | VPRC_OBJREF), 4640 ("vm_page_try_blocked_op: unexpected refcount value %u for %p", 4641 old, m)); 4642 return (true); 4643 } 4644 4645 /* 4646 * Atomically check for wirings and remove all mappings of the page. 4647 */ 4648 bool 4649 vm_page_try_remove_all(vm_page_t m) 4650 { 4651 4652 return (vm_page_try_blocked_op(m, pmap_remove_all)); 4653 } 4654 4655 /* 4656 * Atomically check for wirings and remove all writeable mappings of the page. 4657 */ 4658 bool 4659 vm_page_try_remove_write(vm_page_t m) 4660 { 4661 4662 return (vm_page_try_blocked_op(m, pmap_remove_write)); 4663 } 4664 4665 /* 4666 * vm_page_advise 4667 * 4668 * Apply the specified advice to the given page. 4669 */ 4670 void 4671 vm_page_advise(vm_page_t m, int advice) 4672 { 4673 4674 VM_OBJECT_ASSERT_WLOCKED(m->object); 4675 vm_page_assert_xbusied(m); 4676 4677 if (advice == MADV_FREE) 4678 /* 4679 * Mark the page clean. This will allow the page to be freed 4680 * without first paging it out. MADV_FREE pages are often 4681 * quickly reused by malloc(3), so we do not do anything that 4682 * would result in a page fault on a later access. 4683 */ 4684 vm_page_undirty(m); 4685 else if (advice != MADV_DONTNEED) { 4686 if (advice == MADV_WILLNEED) 4687 vm_page_activate(m); 4688 return; 4689 } 4690 4691 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 4692 vm_page_dirty(m); 4693 4694 /* 4695 * Clear any references to the page. Otherwise, the page daemon will 4696 * immediately reactivate the page. 4697 */ 4698 vm_page_aflag_clear(m, PGA_REFERENCED); 4699 4700 /* 4701 * Place clean pages near the head of the inactive queue rather than 4702 * the tail, thus defeating the queue's LRU operation and ensuring that 4703 * the page will be reused quickly. Dirty pages not already in the 4704 * laundry are moved there. 4705 */ 4706 if (m->dirty == 0) 4707 vm_page_deactivate_noreuse(m); 4708 else if (!vm_page_in_laundry(m)) 4709 vm_page_launder(m); 4710 } 4711 4712 /* 4713 * vm_page_grab_release 4714 * 4715 * Helper routine for grab functions to release busy on return. 4716 */ 4717 static inline void 4718 vm_page_grab_release(vm_page_t m, int allocflags) 4719 { 4720 4721 if ((allocflags & VM_ALLOC_NOBUSY) != 0) { 4722 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4723 vm_page_sunbusy(m); 4724 else 4725 vm_page_xunbusy(m); 4726 } 4727 } 4728 4729 /* 4730 * vm_page_grab_sleep 4731 * 4732 * Sleep for busy according to VM_ALLOC_ parameters. Returns true 4733 * if the caller should retry and false otherwise. 4734 * 4735 * If the object is locked on entry the object will be unlocked with 4736 * false returns and still locked but possibly having been dropped 4737 * with true returns. 4738 */ 4739 static bool 4740 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex, 4741 const char *wmesg, int allocflags, bool locked) 4742 { 4743 4744 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4745 return (false); 4746 4747 /* 4748 * Reference the page before unlocking and sleeping so that 4749 * the page daemon is less likely to reclaim it. 4750 */ 4751 if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0) 4752 vm_page_reference(m); 4753 4754 if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) && 4755 locked) 4756 VM_OBJECT_WLOCK(object); 4757 if ((allocflags & VM_ALLOC_WAITFAIL) != 0) 4758 return (false); 4759 4760 return (true); 4761 } 4762 4763 /* 4764 * Assert that the grab flags are valid. 4765 */ 4766 static inline void 4767 vm_page_grab_check(int allocflags) 4768 { 4769 4770 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 4771 (allocflags & VM_ALLOC_WIRED) != 0, 4772 ("vm_page_grab*: the pages must be busied or wired")); 4773 4774 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4775 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4776 ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4777 } 4778 4779 /* 4780 * Calculate the page allocation flags for grab. 4781 */ 4782 static inline int 4783 vm_page_grab_pflags(int allocflags) 4784 { 4785 int pflags; 4786 4787 pflags = allocflags & 4788 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | 4789 VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); 4790 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 4791 pflags |= VM_ALLOC_WAITFAIL; 4792 if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) 4793 pflags |= VM_ALLOC_SBUSY; 4794 4795 return (pflags); 4796 } 4797 4798 /* 4799 * Grab a page, waiting until we are woken up due to the page changing state. 4800 * We keep on waiting, if the page continues to be in the object, unless 4801 * allocflags forbid waiting. 4802 * 4803 * The object must be locked on entry. This routine may sleep. The lock will, 4804 * however, be released and reacquired if the routine sleeps. 4805 * 4806 * Return a grabbed page, or NULL. Set *found if a page was found, whether or 4807 * not it was grabbed. 4808 */ 4809 static inline vm_page_t 4810 vm_page_grab_lookup(vm_object_t object, vm_pindex_t pindex, int allocflags, 4811 bool *found, struct pctrie_iter *pages) 4812 { 4813 vm_page_t m; 4814 4815 while ((*found = (m = vm_radix_iter_lookup(pages, pindex)) != NULL) && 4816 !vm_page_tryacquire(m, allocflags)) { 4817 if (!vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4818 allocflags, true)) 4819 return (NULL); 4820 pctrie_iter_reset(pages); 4821 } 4822 return (m); 4823 } 4824 4825 /* 4826 * Grab a page. Use an iterator parameter. Keep on waiting, as long as the page 4827 * exists in the object. If the page doesn't exist, first allocate it and then 4828 * conditionally zero it. 4829 * 4830 * The object must be locked on entry. This routine may sleep. The lock will, 4831 * however, be released and reacquired if the routine sleeps. 4832 */ 4833 vm_page_t 4834 vm_page_grab_iter(vm_object_t object, vm_pindex_t pindex, int allocflags, 4835 struct pctrie_iter *pages) 4836 { 4837 vm_page_t m; 4838 bool found; 4839 4840 VM_OBJECT_ASSERT_WLOCKED(object); 4841 vm_page_grab_check(allocflags); 4842 4843 while ((m = vm_page_grab_lookup( 4844 object, pindex, allocflags, &found, pages)) == NULL) { 4845 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4846 return (NULL); 4847 if (found && 4848 (allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4849 return (NULL); 4850 m = vm_page_alloc_iter(object, pindex, 4851 vm_page_grab_pflags(allocflags), pages); 4852 if (m != NULL) { 4853 if ((allocflags & VM_ALLOC_ZERO) != 0 && 4854 (m->flags & PG_ZERO) == 0) 4855 pmap_zero_page(m); 4856 break; 4857 } 4858 if ((allocflags & 4859 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0) 4860 return (NULL); 4861 } 4862 vm_page_grab_release(m, allocflags); 4863 4864 return (m); 4865 } 4866 4867 /* 4868 * Grab a page. Keep on waiting, as long as the page exists in the object. If 4869 * the page doesn't exist, first allocate it and then conditionally zero it. 4870 * 4871 * The object must be locked on entry. This routine may sleep. The lock will, 4872 * however, be released and reacquired if the routine sleeps. 4873 */ 4874 vm_page_t 4875 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 4876 { 4877 struct pctrie_iter pages; 4878 4879 VM_OBJECT_ASSERT_WLOCKED(object); 4880 vm_page_iter_init(&pages, object); 4881 return (vm_page_grab_iter(object, pindex, allocflags, &pages)); 4882 } 4883 4884 /* 4885 * Attempt to validate a page, locklessly acquiring it if necessary, given a 4886 * (object, pindex) tuple and either an invalided page or NULL. The resulting 4887 * page will be validated against the identity tuple, and busied or wired as 4888 * requested. A NULL page returned guarantees that the page was not in radix at 4889 * the time of the call but callers must perform higher level synchronization or 4890 * retry the operation under a lock if they require an atomic answer. This is 4891 * the only lock free validation routine, other routines can depend on the 4892 * resulting page state. 4893 * 4894 * The return value PAGE_NOT_ACQUIRED indicates that the operation failed due to 4895 * caller flags. 4896 */ 4897 #define PAGE_NOT_ACQUIRED ((vm_page_t)1) 4898 static vm_page_t 4899 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex, vm_page_t m, 4900 int allocflags) 4901 { 4902 if (m == NULL) 4903 m = vm_page_lookup_unlocked(object, pindex); 4904 for (; m != NULL; m = vm_page_lookup_unlocked(object, pindex)) { 4905 if (vm_page_trybusy(m, allocflags)) { 4906 if (m->object == object && m->pindex == pindex) { 4907 if ((allocflags & VM_ALLOC_WIRED) != 0) 4908 vm_page_wire(m); 4909 vm_page_grab_release(m, allocflags); 4910 break; 4911 } 4912 /* relookup. */ 4913 vm_page_busy_release(m); 4914 cpu_spinwait(); 4915 continue; 4916 } 4917 if (!vm_page_grab_sleep(object, m, pindex, "pgnslp", 4918 allocflags, false)) 4919 return (PAGE_NOT_ACQUIRED); 4920 } 4921 return (m); 4922 } 4923 4924 /* 4925 * Try to locklessly grab a page and fall back to the object lock if NOCREAT 4926 * is not set. 4927 */ 4928 vm_page_t 4929 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags) 4930 { 4931 vm_page_t m; 4932 4933 vm_page_grab_check(allocflags); 4934 m = vm_page_acquire_unlocked(object, pindex, NULL, allocflags); 4935 if (m == PAGE_NOT_ACQUIRED) 4936 return (NULL); 4937 if (m != NULL) 4938 return (m); 4939 4940 /* 4941 * The radix lockless lookup should never return a false negative 4942 * errors. If the user specifies NOCREAT they are guaranteed there 4943 * was no page present at the instant of the call. A NOCREAT caller 4944 * must handle create races gracefully. 4945 */ 4946 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 4947 return (NULL); 4948 4949 VM_OBJECT_WLOCK(object); 4950 m = vm_page_grab(object, pindex, allocflags); 4951 VM_OBJECT_WUNLOCK(object); 4952 4953 return (m); 4954 } 4955 4956 /* 4957 * Grab a page and make it valid, paging in if necessary. Use an iterator 4958 * parameter. Pages missing from their pager are zero filled and validated. If 4959 * a VM_ALLOC_COUNT is supplied and the page is not valid as many as 4960 * VM_INITIAL_PAGEIN pages can be brought in simultaneously. Additional pages 4961 * will be left on a paging queue but will neither be wired nor busy regardless 4962 * of allocflags. 4963 */ 4964 int 4965 vm_page_grab_valid_iter(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, 4966 int allocflags, struct pctrie_iter *pages) 4967 { 4968 vm_page_t m; 4969 vm_page_t ma[VM_INITIAL_PAGEIN]; 4970 int after, ahead, i, pflags, rv; 4971 4972 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 4973 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 4974 ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 4975 KASSERT((allocflags & 4976 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 4977 ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); 4978 VM_OBJECT_ASSERT_WLOCKED(object); 4979 pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY | 4980 VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY); 4981 pflags |= VM_ALLOC_WAITFAIL; 4982 4983 retrylookup: 4984 if ((m = vm_radix_iter_lookup(pages, pindex)) != NULL) { 4985 /* 4986 * If the page is fully valid it can only become invalid 4987 * with the object lock held. If it is not valid it can 4988 * become valid with the busy lock held. Therefore, we 4989 * may unnecessarily lock the exclusive busy here if we 4990 * race with I/O completion not using the object lock. 4991 * However, we will not end up with an invalid page and a 4992 * shared lock. 4993 */ 4994 if (!vm_page_trybusy(m, 4995 vm_page_all_valid(m) ? allocflags : 0)) { 4996 (void)vm_page_grab_sleep(object, m, pindex, "pgrbwt", 4997 allocflags, true); 4998 pctrie_iter_reset(pages); 4999 goto retrylookup; 5000 } 5001 if (vm_page_all_valid(m)) 5002 goto out; 5003 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5004 vm_page_busy_release(m); 5005 *mp = NULL; 5006 return (VM_PAGER_FAIL); 5007 } 5008 } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5009 *mp = NULL; 5010 return (VM_PAGER_FAIL); 5011 } else { 5012 m = vm_page_alloc_iter(object, pindex, pflags, pages); 5013 if (m == NULL) { 5014 if (!vm_pager_can_alloc_page(object, pindex)) { 5015 *mp = NULL; 5016 return (VM_PAGER_AGAIN); 5017 } 5018 goto retrylookup; 5019 } 5020 } 5021 5022 vm_page_assert_xbusied(m); 5023 if (vm_pager_has_page(object, pindex, NULL, &after)) { 5024 after = MIN(after, VM_INITIAL_PAGEIN); 5025 after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); 5026 after = MAX(after, 1); 5027 ma[0] = m; 5028 pctrie_iter_reset(pages); 5029 for (i = 1; i < after; i++) { 5030 m = vm_radix_iter_lookup_ge(pages, pindex + i); 5031 ahead = after; 5032 if (m != NULL) 5033 ahead = MIN(ahead, m->pindex - pindex); 5034 for (; i < ahead; i++) { 5035 ma[i] = vm_page_alloc_iter(object, pindex + i, 5036 VM_ALLOC_NORMAL, pages); 5037 if (ma[i] == NULL) 5038 break; 5039 } 5040 if (m == NULL || m->pindex != pindex + i || 5041 vm_page_any_valid(m) || !vm_page_tryxbusy(m)) 5042 break; 5043 ma[i] = m; 5044 } 5045 after = i; 5046 vm_object_pip_add(object, after); 5047 VM_OBJECT_WUNLOCK(object); 5048 rv = vm_pager_get_pages(object, ma, after, NULL, NULL); 5049 pctrie_iter_reset(pages); 5050 VM_OBJECT_WLOCK(object); 5051 vm_object_pip_wakeupn(object, after); 5052 /* Pager may have replaced a page. */ 5053 m = ma[0]; 5054 if (rv != VM_PAGER_OK) { 5055 for (i = 0; i < after; i++) { 5056 if (!vm_page_wired(ma[i])) 5057 vm_page_free(ma[i]); 5058 else 5059 vm_page_xunbusy(ma[i]); 5060 } 5061 *mp = NULL; 5062 return (rv); 5063 } 5064 for (i = 1; i < after; i++) 5065 vm_page_readahead_finish(ma[i]); 5066 MPASS(vm_page_all_valid(m)); 5067 } else { 5068 vm_page_zero_invalid(m, TRUE); 5069 pctrie_iter_reset(pages); 5070 } 5071 out: 5072 if ((allocflags & VM_ALLOC_WIRED) != 0) 5073 vm_page_wire(m); 5074 if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m)) 5075 vm_page_busy_downgrade(m); 5076 else if ((allocflags & VM_ALLOC_NOBUSY) != 0) 5077 vm_page_busy_release(m); 5078 *mp = m; 5079 return (VM_PAGER_OK); 5080 } 5081 5082 /* 5083 * Grab a page and make it valid, paging in if necessary. Pages missing from 5084 * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied 5085 * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought 5086 * in simultaneously. Additional pages will be left on a paging queue but 5087 * will neither be wired nor busy regardless of allocflags. 5088 */ 5089 int 5090 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, 5091 int allocflags) 5092 { 5093 struct pctrie_iter pages; 5094 5095 VM_OBJECT_ASSERT_WLOCKED(object); 5096 vm_page_iter_init(&pages, object); 5097 return (vm_page_grab_valid_iter(mp, object, pindex, allocflags, 5098 &pages)); 5099 } 5100 5101 /* 5102 * Grab a page. Keep on waiting, as long as the page exists in the object. If 5103 * the page doesn't exist, and the pager has it, allocate it and zero part of 5104 * it. 5105 * 5106 * The object must be locked on entry. This routine may sleep. The lock will, 5107 * however, be released and reacquired if the routine sleeps. 5108 */ 5109 int 5110 vm_page_grab_zero_partial(vm_object_t object, vm_pindex_t pindex, int base, 5111 int end) 5112 { 5113 struct pctrie_iter pages; 5114 vm_page_t m; 5115 int allocflags, rv; 5116 bool found; 5117 5118 VM_OBJECT_ASSERT_WLOCKED(object); 5119 KASSERT(base >= 0, ("%s: base %d", __func__, base)); 5120 KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base, 5121 end)); 5122 5123 allocflags = VM_ALLOC_NOCREAT | VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL; 5124 vm_page_iter_init(&pages, object); 5125 while ((m = vm_page_grab_lookup( 5126 object, pindex, allocflags, &found, &pages)) == NULL) { 5127 if (!vm_pager_has_page(object, pindex, NULL, NULL)) 5128 return (0); 5129 m = vm_page_alloc_iter(object, pindex, 5130 vm_page_grab_pflags(allocflags), &pages); 5131 if (m != NULL) { 5132 vm_object_pip_add(object, 1); 5133 VM_OBJECT_WUNLOCK(object); 5134 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); 5135 VM_OBJECT_WLOCK(object); 5136 vm_object_pip_wakeup(object); 5137 if (rv != VM_PAGER_OK) { 5138 vm_page_free(m); 5139 return (EIO); 5140 } 5141 5142 /* 5143 * Since the page was not resident, and therefore not 5144 * recently accessed, immediately enqueue it for 5145 * asynchronous laundering. The current operation is 5146 * not regarded as an access. 5147 */ 5148 vm_page_launder(m); 5149 break; 5150 } 5151 } 5152 5153 pmap_zero_page_area(m, base, end - base); 5154 KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m)); 5155 vm_page_set_dirty(m); 5156 vm_page_xunbusy(m); 5157 return (0); 5158 } 5159 5160 /* 5161 * Locklessly grab a valid page. If the page is not valid or not yet 5162 * allocated this will fall back to the object lock method. 5163 */ 5164 int 5165 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object, 5166 vm_pindex_t pindex, int allocflags) 5167 { 5168 vm_page_t m; 5169 int flags; 5170 int error; 5171 5172 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 5173 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 5174 ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " 5175 "mismatch")); 5176 KASSERT((allocflags & 5177 (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, 5178 ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags)); 5179 5180 /* 5181 * Attempt a lockless lookup and busy. We need at least an sbusy 5182 * before we can inspect the valid field and return a wired page. 5183 */ 5184 flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED); 5185 vm_page_grab_check(flags); 5186 m = vm_page_acquire_unlocked(object, pindex, NULL, flags); 5187 if (m == PAGE_NOT_ACQUIRED) 5188 return (VM_PAGER_FAIL); 5189 if (m != NULL) { 5190 if (vm_page_all_valid(m)) { 5191 if ((allocflags & VM_ALLOC_WIRED) != 0) 5192 vm_page_wire(m); 5193 vm_page_grab_release(m, allocflags); 5194 *mp = m; 5195 return (VM_PAGER_OK); 5196 } 5197 vm_page_busy_release(m); 5198 } 5199 if ((allocflags & VM_ALLOC_NOCREAT) != 0) { 5200 *mp = NULL; 5201 return (VM_PAGER_FAIL); 5202 } 5203 VM_OBJECT_WLOCK(object); 5204 error = vm_page_grab_valid(mp, object, pindex, allocflags); 5205 VM_OBJECT_WUNLOCK(object); 5206 5207 return (error); 5208 } 5209 5210 /* 5211 * Return the specified range of pages from the given object. For each 5212 * page offset within the range, if a page already exists within the object 5213 * at that offset and it is busy, then wait for it to change state. If, 5214 * instead, the page doesn't exist, then allocate it. 5215 * 5216 * The caller must always specify an allocation class. 5217 * 5218 * allocation classes: 5219 * VM_ALLOC_NORMAL normal process request 5220 * VM_ALLOC_SYSTEM system *really* needs the pages 5221 * VM_ALLOC_INTERRUPT interrupt time request 5222 * 5223 * The caller must always specify that the pages are to be busied and/or 5224 * wired. 5225 * 5226 * optional allocation flags: 5227 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 5228 * VM_ALLOC_NOBUSY do not exclusive busy the pages 5229 * VM_ALLOC_NODUMP do not include the pages in a kernel core dump 5230 * VM_ALLOC_NOFREE pages will never be freed 5231 * VM_ALLOC_NOWAIT do not sleep 5232 * VM_ALLOC_SBUSY set pages to sbusy state 5233 * VM_ALLOC_WAITFAIL in case of failure, sleep before returning 5234 * VM_ALLOC_WAITOK ignored (default behavior) 5235 * VM_ALLOC_WIRED wire the pages 5236 * VM_ALLOC_ZERO zero and validate any invalid pages 5237 * 5238 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 5239 * may return a partial prefix of the requested range. 5240 */ 5241 int 5242 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 5243 vm_page_t *ma, int count) 5244 { 5245 struct pctrie_iter pages; 5246 vm_page_t m; 5247 int pflags; 5248 int ahead, i; 5249 5250 VM_OBJECT_ASSERT_WLOCKED(object); 5251 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 5252 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 5253 KASSERT(count > 0, 5254 ("vm_page_grab_pages: invalid page count %d", count)); 5255 vm_page_grab_check(allocflags); 5256 5257 pflags = vm_page_grab_pflags(allocflags); 5258 i = 0; 5259 vm_page_iter_init(&pages, object); 5260 retrylookup: 5261 ahead = -1; 5262 for (; i < count; i++) { 5263 if (ahead < 0) { 5264 ahead = vm_radix_iter_lookup_range( 5265 &pages, pindex + i, &ma[i], count - i); 5266 } 5267 if (ahead-- > 0) { 5268 m = ma[i]; 5269 if (!vm_page_tryacquire(m, allocflags)) { 5270 if (vm_page_grab_sleep(object, m, pindex + i, 5271 "grbmaw", allocflags, true)) { 5272 pctrie_iter_reset(&pages); 5273 goto retrylookup; 5274 } 5275 break; 5276 } 5277 } else { 5278 if ((allocflags & VM_ALLOC_NOCREAT) != 0) 5279 break; 5280 m = vm_page_alloc_iter(object, pindex + i, 5281 pflags | VM_ALLOC_COUNT(count - i), &pages); 5282 /* pages was reset if alloc_iter lost the lock. */ 5283 if (m == NULL) { 5284 if ((allocflags & (VM_ALLOC_NOWAIT | 5285 VM_ALLOC_WAITFAIL)) != 0) 5286 break; 5287 goto retrylookup; 5288 } 5289 ma[i] = m; 5290 } 5291 if (vm_page_none_valid(m) && 5292 (allocflags & VM_ALLOC_ZERO) != 0) { 5293 if ((m->flags & PG_ZERO) == 0) 5294 pmap_zero_page(m); 5295 vm_page_valid(m); 5296 } 5297 vm_page_grab_release(m, allocflags); 5298 } 5299 return (i); 5300 } 5301 5302 /* 5303 * Unlocked variant of vm_page_grab_pages(). This accepts the same flags 5304 * and will fall back to the locked variant to handle allocation. 5305 */ 5306 int 5307 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex, 5308 int allocflags, vm_page_t *ma, int count) 5309 { 5310 vm_page_t m; 5311 int flags; 5312 int i, num_fetched; 5313 5314 KASSERT(count > 0, 5315 ("vm_page_grab_pages_unlocked: invalid page count %d", count)); 5316 vm_page_grab_check(allocflags); 5317 5318 /* 5319 * Modify flags for lockless acquire to hold the page until we 5320 * set it valid if necessary. 5321 */ 5322 flags = allocflags & ~VM_ALLOC_NOBUSY; 5323 vm_page_grab_check(flags); 5324 num_fetched = vm_radix_lookup_range_unlocked(&object->rtree, pindex, 5325 ma, count); 5326 for (i = 0; i < num_fetched; i++, pindex++) { 5327 m = vm_page_acquire_unlocked(object, pindex, ma[i], flags); 5328 if (m == PAGE_NOT_ACQUIRED) 5329 return (i); 5330 if (m == NULL) 5331 break; 5332 if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) { 5333 if ((m->flags & PG_ZERO) == 0) 5334 pmap_zero_page(m); 5335 vm_page_valid(m); 5336 } 5337 /* m will still be wired or busy according to flags. */ 5338 vm_page_grab_release(m, allocflags); 5339 /* vm_page_acquire_unlocked() may not return ma[i]. */ 5340 ma[i] = m; 5341 } 5342 if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0) 5343 return (i); 5344 count -= i; 5345 VM_OBJECT_WLOCK(object); 5346 i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count); 5347 VM_OBJECT_WUNLOCK(object); 5348 5349 return (i); 5350 } 5351 5352 /* 5353 * Mapping function for valid or dirty bits in a page. 5354 * 5355 * Inputs are required to range within a page. 5356 */ 5357 vm_page_bits_t 5358 vm_page_bits(int base, int size) 5359 { 5360 int first_bit; 5361 int last_bit; 5362 5363 KASSERT( 5364 base + size <= PAGE_SIZE, 5365 ("vm_page_bits: illegal base/size %d/%d", base, size) 5366 ); 5367 5368 if (size == 0) /* handle degenerate case */ 5369 return (0); 5370 5371 first_bit = base >> DEV_BSHIFT; 5372 last_bit = (base + size - 1) >> DEV_BSHIFT; 5373 5374 return (((vm_page_bits_t)2 << last_bit) - 5375 ((vm_page_bits_t)1 << first_bit)); 5376 } 5377 5378 void 5379 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) 5380 { 5381 #if PAGE_SIZE == 32768 5382 atomic_set_64((uint64_t *)bits, set); 5383 #elif PAGE_SIZE == 16384 5384 atomic_set_32((uint32_t *)bits, set); 5385 #elif PAGE_SIZE == 8192 5386 atomic_set_16((uint16_t *)bits, set); 5387 #elif PAGE_SIZE == 4096 5388 atomic_set_8((uint8_t *)bits, set); 5389 #else 5390 #error unhandled page size 5391 #endif 5392 } 5393 5394 static inline void 5395 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) 5396 { 5397 #if PAGE_SIZE == 32768 5398 atomic_clear_64((uint64_t *)bits, clear); 5399 #elif PAGE_SIZE == 16384 5400 atomic_clear_32((uint32_t *)bits, clear); 5401 #elif PAGE_SIZE == 8192 5402 atomic_clear_16((uint16_t *)bits, clear); 5403 #elif PAGE_SIZE == 4096 5404 atomic_clear_8((uint8_t *)bits, clear); 5405 #else 5406 #error unhandled page size 5407 #endif 5408 } 5409 5410 static inline vm_page_bits_t 5411 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) 5412 { 5413 #if PAGE_SIZE == 32768 5414 uint64_t old; 5415 5416 old = *bits; 5417 while (atomic_fcmpset_64(bits, &old, newbits) == 0); 5418 return (old); 5419 #elif PAGE_SIZE == 16384 5420 uint32_t old; 5421 5422 old = *bits; 5423 while (atomic_fcmpset_32(bits, &old, newbits) == 0); 5424 return (old); 5425 #elif PAGE_SIZE == 8192 5426 uint16_t old; 5427 5428 old = *bits; 5429 while (atomic_fcmpset_16(bits, &old, newbits) == 0); 5430 return (old); 5431 #elif PAGE_SIZE == 4096 5432 uint8_t old; 5433 5434 old = *bits; 5435 while (atomic_fcmpset_8(bits, &old, newbits) == 0); 5436 return (old); 5437 #else 5438 #error unhandled page size 5439 #endif 5440 } 5441 5442 /* 5443 * vm_page_set_valid_range: 5444 * 5445 * Sets portions of a page valid. The arguments are expected 5446 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5447 * of any partial chunks touched by the range. The invalid portion of 5448 * such chunks will be zeroed. 5449 * 5450 * (base + size) must be less then or equal to PAGE_SIZE. 5451 */ 5452 void 5453 vm_page_set_valid_range(vm_page_t m, int base, int size) 5454 { 5455 int endoff, frag; 5456 vm_page_bits_t pagebits; 5457 5458 vm_page_assert_busied(m); 5459 if (size == 0) /* handle degenerate case */ 5460 return; 5461 5462 /* 5463 * If the base is not DEV_BSIZE aligned and the valid 5464 * bit is clear, we have to zero out a portion of the 5465 * first block. 5466 */ 5467 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5468 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 5469 pmap_zero_page_area(m, frag, base - frag); 5470 5471 /* 5472 * If the ending offset is not DEV_BSIZE aligned and the 5473 * valid bit is clear, we have to zero out a portion of 5474 * the last block. 5475 */ 5476 endoff = base + size; 5477 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5478 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 5479 pmap_zero_page_area(m, endoff, 5480 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5481 5482 /* 5483 * Assert that no previously invalid block that is now being validated 5484 * is already dirty. 5485 */ 5486 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 5487 ("vm_page_set_valid_range: page %p is dirty", m)); 5488 5489 /* 5490 * Set valid bits inclusive of any overlap. 5491 */ 5492 pagebits = vm_page_bits(base, size); 5493 if (vm_page_xbusied(m)) 5494 m->valid |= pagebits; 5495 else 5496 vm_page_bits_set(m, &m->valid, pagebits); 5497 } 5498 5499 /* 5500 * Set the page dirty bits and free the invalid swap space if 5501 * present. Returns the previous dirty bits. 5502 */ 5503 vm_page_bits_t 5504 vm_page_set_dirty(vm_page_t m) 5505 { 5506 vm_page_bits_t old; 5507 5508 VM_PAGE_OBJECT_BUSY_ASSERT(m); 5509 5510 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { 5511 old = m->dirty; 5512 m->dirty = VM_PAGE_BITS_ALL; 5513 } else 5514 old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); 5515 if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) 5516 vm_pager_page_unswapped(m); 5517 5518 return (old); 5519 } 5520 5521 /* 5522 * Clear the given bits from the specified page's dirty field. 5523 */ 5524 static __inline void 5525 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 5526 { 5527 5528 vm_page_assert_busied(m); 5529 5530 /* 5531 * If the page is xbusied and not write mapped we are the 5532 * only thread that can modify dirty bits. Otherwise, The pmap 5533 * layer can call vm_page_dirty() without holding a distinguished 5534 * lock. The combination of page busy and atomic operations 5535 * suffice to guarantee consistency of the page dirty field. 5536 */ 5537 if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 5538 m->dirty &= ~pagebits; 5539 else 5540 vm_page_bits_clear(m, &m->dirty, pagebits); 5541 } 5542 5543 /* 5544 * vm_page_set_validclean: 5545 * 5546 * Sets portions of a page valid and clean. The arguments are expected 5547 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 5548 * of any partial chunks touched by the range. The invalid portion of 5549 * such chunks will be zero'd. 5550 * 5551 * (base + size) must be less then or equal to PAGE_SIZE. 5552 */ 5553 void 5554 vm_page_set_validclean(vm_page_t m, int base, int size) 5555 { 5556 vm_page_bits_t oldvalid, pagebits; 5557 int endoff, frag; 5558 5559 vm_page_assert_busied(m); 5560 if (size == 0) /* handle degenerate case */ 5561 return; 5562 5563 /* 5564 * If the base is not DEV_BSIZE aligned and the valid 5565 * bit is clear, we have to zero out a portion of the 5566 * first block. 5567 */ 5568 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 5569 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 5570 pmap_zero_page_area(m, frag, base - frag); 5571 5572 /* 5573 * If the ending offset is not DEV_BSIZE aligned and the 5574 * valid bit is clear, we have to zero out a portion of 5575 * the last block. 5576 */ 5577 endoff = base + size; 5578 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 5579 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 5580 pmap_zero_page_area(m, endoff, 5581 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 5582 5583 /* 5584 * Set valid, clear dirty bits. If validating the entire 5585 * page we can safely clear the pmap modify bit. We also 5586 * use this opportunity to clear the PGA_NOSYNC flag. If a process 5587 * takes a write fault on a MAP_NOSYNC memory area the flag will 5588 * be set again. 5589 * 5590 * We set valid bits inclusive of any overlap, but we can only 5591 * clear dirty bits for DEV_BSIZE chunks that are fully within 5592 * the range. 5593 */ 5594 oldvalid = m->valid; 5595 pagebits = vm_page_bits(base, size); 5596 if (vm_page_xbusied(m)) 5597 m->valid |= pagebits; 5598 else 5599 vm_page_bits_set(m, &m->valid, pagebits); 5600 #if 0 /* NOT YET */ 5601 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 5602 frag = DEV_BSIZE - frag; 5603 base += frag; 5604 size -= frag; 5605 if (size < 0) 5606 size = 0; 5607 } 5608 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 5609 #endif 5610 if (base == 0 && size == PAGE_SIZE) { 5611 /* 5612 * The page can only be modified within the pmap if it is 5613 * mapped, and it can only be mapped if it was previously 5614 * fully valid. 5615 */ 5616 if (oldvalid == VM_PAGE_BITS_ALL) 5617 /* 5618 * Perform the pmap_clear_modify() first. Otherwise, 5619 * a concurrent pmap operation, such as 5620 * pmap_protect(), could clear a modification in the 5621 * pmap and set the dirty field on the page before 5622 * pmap_clear_modify() had begun and after the dirty 5623 * field was cleared here. 5624 */ 5625 pmap_clear_modify(m); 5626 m->dirty = 0; 5627 vm_page_aflag_clear(m, PGA_NOSYNC); 5628 } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) 5629 m->dirty &= ~pagebits; 5630 else 5631 vm_page_clear_dirty_mask(m, pagebits); 5632 } 5633 5634 void 5635 vm_page_clear_dirty(vm_page_t m, int base, int size) 5636 { 5637 5638 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 5639 } 5640 5641 /* 5642 * vm_page_set_invalid: 5643 * 5644 * Invalidates DEV_BSIZE'd chunks within a page. Both the 5645 * valid and dirty bits for the effected areas are cleared. 5646 */ 5647 void 5648 vm_page_set_invalid(vm_page_t m, int base, int size) 5649 { 5650 vm_page_bits_t bits; 5651 vm_object_t object; 5652 5653 /* 5654 * The object lock is required so that pages can't be mapped 5655 * read-only while we're in the process of invalidating them. 5656 */ 5657 object = m->object; 5658 VM_OBJECT_ASSERT_WLOCKED(object); 5659 vm_page_assert_busied(m); 5660 5661 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 5662 size >= object->un_pager.vnp.vnp_size) 5663 bits = VM_PAGE_BITS_ALL; 5664 else 5665 bits = vm_page_bits(base, size); 5666 if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) 5667 pmap_remove_all(m); 5668 KASSERT((bits == 0 && vm_page_all_valid(m)) || 5669 !pmap_page_is_mapped(m), 5670 ("vm_page_set_invalid: page %p is mapped", m)); 5671 if (vm_page_xbusied(m)) { 5672 m->valid &= ~bits; 5673 m->dirty &= ~bits; 5674 } else { 5675 vm_page_bits_clear(m, &m->valid, bits); 5676 vm_page_bits_clear(m, &m->dirty, bits); 5677 } 5678 } 5679 5680 /* 5681 * vm_page_invalid: 5682 * 5683 * Invalidates the entire page. The page must be busy, unmapped, and 5684 * the enclosing object must be locked. The object locks protects 5685 * against concurrent read-only pmap enter which is done without 5686 * busy. 5687 */ 5688 void 5689 vm_page_invalid(vm_page_t m) 5690 { 5691 5692 vm_page_assert_busied(m); 5693 VM_OBJECT_ASSERT_WLOCKED(m->object); 5694 MPASS(!pmap_page_is_mapped(m)); 5695 5696 if (vm_page_xbusied(m)) 5697 m->valid = 0; 5698 else 5699 vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); 5700 } 5701 5702 /* 5703 * vm_page_zero_invalid() 5704 * 5705 * The kernel assumes that the invalid portions of a page contain 5706 * garbage, but such pages can be mapped into memory by user code. 5707 * When this occurs, we must zero out the non-valid portions of the 5708 * page so user code sees what it expects. 5709 * 5710 * Pages are most often semi-valid when the end of a file is mapped 5711 * into memory and the file's size is not page aligned. 5712 */ 5713 void 5714 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 5715 { 5716 int b; 5717 int i; 5718 5719 /* 5720 * Scan the valid bits looking for invalid sections that 5721 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 5722 * valid bit may be set ) have already been zeroed by 5723 * vm_page_set_validclean(). 5724 */ 5725 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 5726 if (i == (PAGE_SIZE / DEV_BSIZE) || 5727 (m->valid & ((vm_page_bits_t)1 << i))) { 5728 if (i > b) { 5729 pmap_zero_page_area(m, 5730 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 5731 } 5732 b = i + 1; 5733 } 5734 } 5735 5736 /* 5737 * setvalid is TRUE when we can safely set the zero'd areas 5738 * as being valid. We can do this if there are no cache consistency 5739 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 5740 */ 5741 if (setvalid) 5742 vm_page_valid(m); 5743 } 5744 5745 /* 5746 * vm_page_is_valid: 5747 * 5748 * Is (partial) page valid? Note that the case where size == 0 5749 * will return FALSE in the degenerate case where the page is 5750 * entirely invalid, and TRUE otherwise. 5751 * 5752 * Some callers envoke this routine without the busy lock held and 5753 * handle races via higher level locks. Typical callers should 5754 * hold a busy lock to prevent invalidation. 5755 */ 5756 int 5757 vm_page_is_valid(vm_page_t m, int base, int size) 5758 { 5759 vm_page_bits_t bits; 5760 5761 bits = vm_page_bits(base, size); 5762 return (vm_page_any_valid(m) && (m->valid & bits) == bits); 5763 } 5764 5765 /* 5766 * Returns true if all of the specified predicates are true for the entire 5767 * (super)page and false otherwise. 5768 */ 5769 bool 5770 vm_page_ps_test(vm_page_t m, int psind, int flags, vm_page_t skip_m) 5771 { 5772 vm_object_t object; 5773 int i, npages; 5774 5775 object = m->object; 5776 if (skip_m != NULL && skip_m->object != object) 5777 return (false); 5778 VM_OBJECT_ASSERT_LOCKED(object); 5779 KASSERT(psind <= m->psind, 5780 ("psind %d > psind %d of m %p", psind, m->psind, m)); 5781 npages = atop(pagesizes[psind]); 5782 5783 /* 5784 * The physically contiguous pages that make up a superpage, i.e., a 5785 * page with a page size index ("psind") greater than zero, will 5786 * occupy adjacent entries in vm_page_array[]. 5787 */ 5788 for (i = 0; i < npages; i++) { 5789 /* Always test object consistency, including "skip_m". */ 5790 if (m[i].object != object) 5791 return (false); 5792 if (&m[i] == skip_m) 5793 continue; 5794 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 5795 return (false); 5796 if ((flags & PS_ALL_DIRTY) != 0) { 5797 /* 5798 * Calling vm_page_test_dirty() or pmap_is_modified() 5799 * might stop this case from spuriously returning 5800 * "false". However, that would require a write lock 5801 * on the object containing "m[i]". 5802 */ 5803 if (m[i].dirty != VM_PAGE_BITS_ALL) 5804 return (false); 5805 } 5806 if ((flags & PS_ALL_VALID) != 0 && 5807 m[i].valid != VM_PAGE_BITS_ALL) 5808 return (false); 5809 } 5810 return (true); 5811 } 5812 5813 /* 5814 * Set the page's dirty bits if the page is modified. 5815 */ 5816 void 5817 vm_page_test_dirty(vm_page_t m) 5818 { 5819 5820 vm_page_assert_busied(m); 5821 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 5822 vm_page_dirty(m); 5823 } 5824 5825 void 5826 vm_page_valid(vm_page_t m) 5827 { 5828 5829 vm_page_assert_busied(m); 5830 if (vm_page_xbusied(m)) 5831 m->valid = VM_PAGE_BITS_ALL; 5832 else 5833 vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); 5834 } 5835 5836 #ifdef INVARIANTS 5837 void 5838 vm_page_object_busy_assert(vm_page_t m) 5839 { 5840 5841 /* 5842 * Certain of the page's fields may only be modified by the 5843 * holder of a page or object busy. 5844 */ 5845 if (m->object != NULL && !vm_page_busied(m)) 5846 VM_OBJECT_ASSERT_BUSY(m->object); 5847 } 5848 5849 void 5850 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) 5851 { 5852 5853 if ((bits & PGA_WRITEABLE) == 0) 5854 return; 5855 5856 /* 5857 * The PGA_WRITEABLE flag can only be set if the page is 5858 * managed, is exclusively busied or the object is locked. 5859 * Currently, this flag is only set by pmap_enter(). 5860 */ 5861 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 5862 ("PGA_WRITEABLE on unmanaged page")); 5863 if (!vm_page_xbusied(m)) 5864 VM_OBJECT_ASSERT_BUSY(m->object); 5865 } 5866 #endif 5867 5868 #include "opt_ddb.h" 5869 #ifdef DDB 5870 #include <sys/kernel.h> 5871 5872 #include <ddb/ddb.h> 5873 5874 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE) 5875 { 5876 5877 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 5878 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 5879 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 5880 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 5881 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 5882 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 5883 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 5884 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 5885 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 5886 } 5887 5888 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE) 5889 { 5890 int dom; 5891 5892 db_printf("pq_free %d\n", vm_free_count()); 5893 for (dom = 0; dom < vm_ndomains; dom++) { 5894 db_printf( 5895 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 5896 dom, 5897 vm_dom[dom].vmd_page_count, 5898 vm_dom[dom].vmd_free_count, 5899 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 5900 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 5901 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 5902 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 5903 } 5904 } 5905 5906 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 5907 { 5908 vm_page_t m; 5909 boolean_t phys, virt; 5910 5911 if (!have_addr) { 5912 db_printf("show pginfo addr\n"); 5913 return; 5914 } 5915 5916 phys = strchr(modif, 'p') != NULL; 5917 virt = strchr(modif, 'v') != NULL; 5918 if (virt) 5919 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 5920 else if (phys) 5921 m = PHYS_TO_VM_PAGE(addr); 5922 else 5923 m = (vm_page_t)addr; 5924 db_printf( 5925 "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n" 5926 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 5927 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 5928 m->a.queue, m->ref_count, m->a.flags, m->oflags, 5929 m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); 5930 } 5931 #endif /* DDB */ 5932