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