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 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 36 */ 37 38 /*- 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 */ 64 65 /* 66 * Resident memory management module. 67 */ 68 69 #include <sys/cdefs.h> 70 __FBSDID("$FreeBSD$"); 71 72 #include "opt_vm.h" 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/lock.h> 77 #include <sys/domainset.h> 78 #include <sys/kernel.h> 79 #include <sys/limits.h> 80 #include <sys/linker.h> 81 #include <sys/malloc.h> 82 #include <sys/mman.h> 83 #include <sys/msgbuf.h> 84 #include <sys/mutex.h> 85 #include <sys/proc.h> 86 #include <sys/rwlock.h> 87 #include <sys/sbuf.h> 88 #include <sys/sched.h> 89 #include <sys/smp.h> 90 #include <sys/sysctl.h> 91 #include <sys/vmmeter.h> 92 #include <sys/vnode.h> 93 94 #include <vm/vm.h> 95 #include <vm/pmap.h> 96 #include <vm/vm_param.h> 97 #include <vm/vm_domainset.h> 98 #include <vm/vm_kern.h> 99 #include <vm/vm_map.h> 100 #include <vm/vm_object.h> 101 #include <vm/vm_page.h> 102 #include <vm/vm_pageout.h> 103 #include <vm/vm_phys.h> 104 #include <vm/vm_pagequeue.h> 105 #include <vm/vm_pager.h> 106 #include <vm/vm_radix.h> 107 #include <vm/vm_reserv.h> 108 #include <vm/vm_extern.h> 109 #include <vm/uma.h> 110 #include <vm/uma_int.h> 111 112 #include <machine/md_var.h> 113 114 extern int uma_startup_count(int); 115 extern void uma_startup(void *, int); 116 extern int vmem_startup_count(void); 117 118 struct vm_domain vm_dom[MAXMEMDOM]; 119 120 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); 121 122 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 123 124 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 125 /* The following fields are protected by the domainset lock. */ 126 domainset_t __exclusive_cache_line vm_min_domains; 127 domainset_t __exclusive_cache_line vm_severe_domains; 128 static int vm_min_waiters; 129 static int vm_severe_waiters; 130 static int vm_pageproc_waiters; 131 132 /* 133 * bogus page -- for I/O to/from partially complete buffers, 134 * or for paging into sparsely invalid regions. 135 */ 136 vm_page_t bogus_page; 137 138 vm_page_t vm_page_array; 139 long vm_page_array_size; 140 long first_page; 141 142 static int boot_pages; 143 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 144 &boot_pages, 0, 145 "number of pages allocated for bootstrapping the VM system"); 146 147 static int pa_tryrelock_restart; 148 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 149 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 150 151 static TAILQ_HEAD(, vm_page) blacklist_head; 152 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 153 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 154 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 155 156 static uma_zone_t fakepg_zone; 157 158 static void vm_page_alloc_check(vm_page_t m); 159 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 160 static void vm_page_dequeue_complete(vm_page_t m); 161 static void vm_page_enqueue(vm_page_t m, uint8_t queue); 162 static void vm_page_init(void *dummy); 163 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 164 vm_pindex_t pindex, vm_page_t mpred); 165 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 166 vm_page_t mpred); 167 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 168 vm_page_t m_run, vm_paddr_t high); 169 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 170 int req); 171 static int vm_page_import(void *arg, void **store, int cnt, int domain, 172 int flags); 173 static void vm_page_release(void *arg, void **store, int cnt); 174 175 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 176 177 static void 178 vm_page_init(void *dummy) 179 { 180 181 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 182 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 183 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 184 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 185 } 186 187 /* 188 * The cache page zone is initialized later since we need to be able to allocate 189 * pages before UMA is fully initialized. 190 */ 191 static void 192 vm_page_init_cache_zones(void *dummy __unused) 193 { 194 struct vm_domain *vmd; 195 int i; 196 197 for (i = 0; i < vm_ndomains; i++) { 198 vmd = VM_DOMAIN(i); 199 /* 200 * Don't allow the page cache to take up more than .25% of 201 * memory. 202 */ 203 if (vmd->vmd_page_count / 400 < 256 * mp_ncpus) 204 continue; 205 vmd->vmd_pgcache = uma_zcache_create("vm pgcache", 206 sizeof(struct vm_page), NULL, NULL, NULL, NULL, 207 vm_page_import, vm_page_release, vmd, 208 UMA_ZONE_MAXBUCKET | UMA_ZONE_VM); 209 (void )uma_zone_set_maxcache(vmd->vmd_pgcache, 0); 210 } 211 } 212 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); 213 214 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 215 #if PAGE_SIZE == 32768 216 #ifdef CTASSERT 217 CTASSERT(sizeof(u_long) >= 8); 218 #endif 219 #endif 220 221 /* 222 * Try to acquire a physical address lock while a pmap is locked. If we 223 * fail to trylock we unlock and lock the pmap directly and cache the 224 * locked pa in *locked. The caller should then restart their loop in case 225 * the virtual to physical mapping has changed. 226 */ 227 int 228 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 229 { 230 vm_paddr_t lockpa; 231 232 lockpa = *locked; 233 *locked = pa; 234 if (lockpa) { 235 PA_LOCK_ASSERT(lockpa, MA_OWNED); 236 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 237 return (0); 238 PA_UNLOCK(lockpa); 239 } 240 if (PA_TRYLOCK(pa)) 241 return (0); 242 PMAP_UNLOCK(pmap); 243 atomic_add_int(&pa_tryrelock_restart, 1); 244 PA_LOCK(pa); 245 PMAP_LOCK(pmap); 246 return (EAGAIN); 247 } 248 249 /* 250 * vm_set_page_size: 251 * 252 * Sets the page size, perhaps based upon the memory 253 * size. Must be called before any use of page-size 254 * dependent functions. 255 */ 256 void 257 vm_set_page_size(void) 258 { 259 if (vm_cnt.v_page_size == 0) 260 vm_cnt.v_page_size = PAGE_SIZE; 261 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 262 panic("vm_set_page_size: page size not a power of two"); 263 } 264 265 /* 266 * vm_page_blacklist_next: 267 * 268 * Find the next entry in the provided string of blacklist 269 * addresses. Entries are separated by space, comma, or newline. 270 * If an invalid integer is encountered then the rest of the 271 * string is skipped. Updates the list pointer to the next 272 * character, or NULL if the string is exhausted or invalid. 273 */ 274 static vm_paddr_t 275 vm_page_blacklist_next(char **list, char *end) 276 { 277 vm_paddr_t bad; 278 char *cp, *pos; 279 280 if (list == NULL || *list == NULL) 281 return (0); 282 if (**list =='\0') { 283 *list = NULL; 284 return (0); 285 } 286 287 /* 288 * If there's no end pointer then the buffer is coming from 289 * the kenv and we know it's null-terminated. 290 */ 291 if (end == NULL) 292 end = *list + strlen(*list); 293 294 /* Ensure that strtoq() won't walk off the end */ 295 if (*end != '\0') { 296 if (*end == '\n' || *end == ' ' || *end == ',') 297 *end = '\0'; 298 else { 299 printf("Blacklist not terminated, skipping\n"); 300 *list = NULL; 301 return (0); 302 } 303 } 304 305 for (pos = *list; *pos != '\0'; pos = cp) { 306 bad = strtoq(pos, &cp, 0); 307 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 308 if (bad == 0) { 309 if (++cp < end) 310 continue; 311 else 312 break; 313 } 314 } else 315 break; 316 if (*cp == '\0' || ++cp >= end) 317 *list = NULL; 318 else 319 *list = cp; 320 return (trunc_page(bad)); 321 } 322 printf("Garbage in RAM blacklist, skipping\n"); 323 *list = NULL; 324 return (0); 325 } 326 327 bool 328 vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 329 { 330 struct vm_domain *vmd; 331 vm_page_t m; 332 int ret; 333 334 m = vm_phys_paddr_to_vm_page(pa); 335 if (m == NULL) 336 return (true); /* page does not exist, no failure */ 337 338 vmd = vm_pagequeue_domain(m); 339 vm_domain_free_lock(vmd); 340 ret = vm_phys_unfree_page(m); 341 vm_domain_free_unlock(vmd); 342 if (ret != 0) { 343 vm_domain_freecnt_inc(vmd, -1); 344 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 345 if (verbose) 346 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 347 } 348 return (ret); 349 } 350 351 /* 352 * vm_page_blacklist_check: 353 * 354 * Iterate through the provided string of blacklist addresses, pulling 355 * each entry out of the physical allocator free list and putting it 356 * onto a list for reporting via the vm.page_blacklist sysctl. 357 */ 358 static void 359 vm_page_blacklist_check(char *list, char *end) 360 { 361 vm_paddr_t pa; 362 char *next; 363 364 next = list; 365 while (next != NULL) { 366 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 367 continue; 368 vm_page_blacklist_add(pa, bootverbose); 369 } 370 } 371 372 /* 373 * vm_page_blacklist_load: 374 * 375 * Search for a special module named "ram_blacklist". It'll be a 376 * plain text file provided by the user via the loader directive 377 * of the same name. 378 */ 379 static void 380 vm_page_blacklist_load(char **list, char **end) 381 { 382 void *mod; 383 u_char *ptr; 384 u_int len; 385 386 mod = NULL; 387 ptr = NULL; 388 389 mod = preload_search_by_type("ram_blacklist"); 390 if (mod != NULL) { 391 ptr = preload_fetch_addr(mod); 392 len = preload_fetch_size(mod); 393 } 394 *list = ptr; 395 if (ptr != NULL) 396 *end = ptr + len; 397 else 398 *end = NULL; 399 return; 400 } 401 402 static int 403 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 404 { 405 vm_page_t m; 406 struct sbuf sbuf; 407 int error, first; 408 409 first = 1; 410 error = sysctl_wire_old_buffer(req, 0); 411 if (error != 0) 412 return (error); 413 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 414 TAILQ_FOREACH(m, &blacklist_head, listq) { 415 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 416 (uintmax_t)m->phys_addr); 417 first = 0; 418 } 419 error = sbuf_finish(&sbuf); 420 sbuf_delete(&sbuf); 421 return (error); 422 } 423 424 /* 425 * Initialize a dummy page for use in scans of the specified paging queue. 426 * In principle, this function only needs to set the flag PG_MARKER. 427 * Nonetheless, it write busies and initializes the hold count to one as 428 * safety precautions. 429 */ 430 static void 431 vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags) 432 { 433 434 bzero(marker, sizeof(*marker)); 435 marker->flags = PG_MARKER; 436 marker->aflags = aflags; 437 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 438 marker->queue = queue; 439 marker->hold_count = 1; 440 } 441 442 static void 443 vm_page_domain_init(int domain) 444 { 445 struct vm_domain *vmd; 446 struct vm_pagequeue *pq; 447 int i; 448 449 vmd = VM_DOMAIN(domain); 450 bzero(vmd, sizeof(*vmd)); 451 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 452 "vm inactive pagequeue"; 453 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 454 "vm active pagequeue"; 455 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 456 "vm laundry pagequeue"; 457 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 458 "vm unswappable pagequeue"; 459 vmd->vmd_domain = domain; 460 vmd->vmd_page_count = 0; 461 vmd->vmd_free_count = 0; 462 vmd->vmd_segs = 0; 463 vmd->vmd_oom = FALSE; 464 for (i = 0; i < PQ_COUNT; i++) { 465 pq = &vmd->vmd_pagequeues[i]; 466 TAILQ_INIT(&pq->pq_pl); 467 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 468 MTX_DEF | MTX_DUPOK); 469 pq->pq_pdpages = 0; 470 vm_page_init_marker(&vmd->vmd_markers[i], i, 0); 471 } 472 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 473 mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); 474 snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); 475 476 /* 477 * inacthead is used to provide FIFO ordering for LRU-bypassing 478 * insertions. 479 */ 480 vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); 481 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, 482 &vmd->vmd_inacthead, plinks.q); 483 484 /* 485 * The clock pages are used to implement active queue scanning without 486 * requeues. Scans start at clock[0], which is advanced after the scan 487 * ends. When the two clock hands meet, they are reset and scanning 488 * resumes from the head of the queue. 489 */ 490 vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); 491 vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); 492 TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 493 &vmd->vmd_clock[0], plinks.q); 494 TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, 495 &vmd->vmd_clock[1], plinks.q); 496 } 497 498 /* 499 * Initialize a physical page in preparation for adding it to the free 500 * lists. 501 */ 502 static void 503 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 504 { 505 506 m->object = NULL; 507 m->wire_count = 0; 508 m->busy_lock = VPB_UNBUSIED; 509 m->hold_count = 0; 510 m->flags = m->aflags = 0; 511 m->phys_addr = pa; 512 m->queue = PQ_NONE; 513 m->psind = 0; 514 m->segind = segind; 515 m->order = VM_NFREEORDER; 516 m->pool = VM_FREEPOOL_DEFAULT; 517 m->valid = m->dirty = 0; 518 pmap_page_init(m); 519 } 520 521 /* 522 * vm_page_startup: 523 * 524 * Initializes the resident memory module. Allocates physical memory for 525 * bootstrapping UMA and some data structures that are used to manage 526 * physical pages. Initializes these structures, and populates the free 527 * page queues. 528 */ 529 vm_offset_t 530 vm_page_startup(vm_offset_t vaddr) 531 { 532 struct vm_phys_seg *seg; 533 vm_page_t m; 534 char *list, *listend; 535 vm_offset_t mapped; 536 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; 537 vm_paddr_t biggestsize, last_pa, pa; 538 u_long pagecount; 539 int biggestone, i, segind; 540 #ifdef WITNESS 541 int witness_size; 542 #endif 543 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 544 long ii; 545 #endif 546 547 biggestsize = 0; 548 biggestone = 0; 549 vaddr = round_page(vaddr); 550 551 for (i = 0; phys_avail[i + 1]; i += 2) { 552 phys_avail[i] = round_page(phys_avail[i]); 553 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 554 } 555 for (i = 0; phys_avail[i + 1]; i += 2) { 556 size = phys_avail[i + 1] - phys_avail[i]; 557 if (size > biggestsize) { 558 biggestone = i; 559 biggestsize = size; 560 } 561 } 562 563 end = phys_avail[biggestone+1]; 564 565 /* 566 * Initialize the page and queue locks. 567 */ 568 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 569 for (i = 0; i < PA_LOCK_COUNT; i++) 570 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 571 for (i = 0; i < vm_ndomains; i++) 572 vm_page_domain_init(i); 573 574 /* 575 * Allocate memory for use when boot strapping the kernel memory 576 * allocator. Tell UMA how many zones we are going to create 577 * before going fully functional. UMA will add its zones. 578 * 579 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, 580 * KMAP ENTRY, MAP ENTRY, VMSPACE. 581 */ 582 boot_pages = uma_startup_count(8); 583 584 #ifndef UMA_MD_SMALL_ALLOC 585 /* vmem_startup() calls uma_prealloc(). */ 586 boot_pages += vmem_startup_count(); 587 /* vm_map_startup() calls uma_prealloc(). */ 588 boot_pages += howmany(MAX_KMAP, 589 UMA_SLAB_SPACE / sizeof(struct vm_map)); 590 591 /* 592 * Before going fully functional kmem_init() does allocation 593 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". 594 */ 595 boot_pages += 2; 596 #endif 597 /* 598 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 599 * manually fetch the value. 600 */ 601 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 602 new_end = end - (boot_pages * UMA_SLAB_SIZE); 603 new_end = trunc_page(new_end); 604 mapped = pmap_map(&vaddr, new_end, end, 605 VM_PROT_READ | VM_PROT_WRITE); 606 bzero((void *)mapped, end - new_end); 607 uma_startup((void *)mapped, boot_pages); 608 609 #ifdef WITNESS 610 witness_size = round_page(witness_startup_count()); 611 new_end -= witness_size; 612 mapped = pmap_map(&vaddr, new_end, new_end + witness_size, 613 VM_PROT_READ | VM_PROT_WRITE); 614 bzero((void *)mapped, witness_size); 615 witness_startup((void *)mapped); 616 #endif 617 618 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 619 defined(__i386__) || defined(__mips__) || defined(__riscv) 620 /* 621 * Allocate a bitmap to indicate that a random physical page 622 * needs to be included in a minidump. 623 * 624 * The amd64 port needs this to indicate which direct map pages 625 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 626 * 627 * However, i386 still needs this workspace internally within the 628 * minidump code. In theory, they are not needed on i386, but are 629 * included should the sf_buf code decide to use them. 630 */ 631 last_pa = 0; 632 for (i = 0; dump_avail[i + 1] != 0; i += 2) 633 if (dump_avail[i + 1] > last_pa) 634 last_pa = dump_avail[i + 1]; 635 page_range = last_pa / PAGE_SIZE; 636 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 637 new_end -= vm_page_dump_size; 638 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 639 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 640 bzero((void *)vm_page_dump, vm_page_dump_size); 641 #else 642 (void)last_pa; 643 #endif 644 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 645 defined(__riscv) 646 /* 647 * Include the UMA bootstrap pages, witness pages and vm_page_dump 648 * in a crash dump. When pmap_map() uses the direct map, they are 649 * not automatically included. 650 */ 651 for (pa = new_end; pa < end; pa += PAGE_SIZE) 652 dump_add_page(pa); 653 #endif 654 phys_avail[biggestone + 1] = new_end; 655 #ifdef __amd64__ 656 /* 657 * Request that the physical pages underlying the message buffer be 658 * included in a crash dump. Since the message buffer is accessed 659 * through the direct map, they are not automatically included. 660 */ 661 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 662 last_pa = pa + round_page(msgbufsize); 663 while (pa < last_pa) { 664 dump_add_page(pa); 665 pa += PAGE_SIZE; 666 } 667 #endif 668 /* 669 * Compute the number of pages of memory that will be available for 670 * use, taking into account the overhead of a page structure per page. 671 * In other words, solve 672 * "available physical memory" - round_page(page_range * 673 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 674 * for page_range. 675 */ 676 low_avail = phys_avail[0]; 677 high_avail = phys_avail[1]; 678 for (i = 0; i < vm_phys_nsegs; i++) { 679 if (vm_phys_segs[i].start < low_avail) 680 low_avail = vm_phys_segs[i].start; 681 if (vm_phys_segs[i].end > high_avail) 682 high_avail = vm_phys_segs[i].end; 683 } 684 /* Skip the first chunk. It is already accounted for. */ 685 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 686 if (phys_avail[i] < low_avail) 687 low_avail = phys_avail[i]; 688 if (phys_avail[i + 1] > high_avail) 689 high_avail = phys_avail[i + 1]; 690 } 691 first_page = low_avail / PAGE_SIZE; 692 #ifdef VM_PHYSSEG_SPARSE 693 size = 0; 694 for (i = 0; i < vm_phys_nsegs; i++) 695 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 696 for (i = 0; phys_avail[i + 1] != 0; i += 2) 697 size += phys_avail[i + 1] - phys_avail[i]; 698 #elif defined(VM_PHYSSEG_DENSE) 699 size = high_avail - low_avail; 700 #else 701 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 702 #endif 703 704 #ifdef VM_PHYSSEG_DENSE 705 /* 706 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 707 * the overhead of a page structure per page only if vm_page_array is 708 * allocated from the last physical memory chunk. Otherwise, we must 709 * allocate page structures representing the physical memory 710 * underlying vm_page_array, even though they will not be used. 711 */ 712 if (new_end != high_avail) 713 page_range = size / PAGE_SIZE; 714 else 715 #endif 716 { 717 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 718 719 /* 720 * If the partial bytes remaining are large enough for 721 * a page (PAGE_SIZE) without a corresponding 722 * 'struct vm_page', then new_end will contain an 723 * extra page after subtracting the length of the VM 724 * page array. Compensate by subtracting an extra 725 * page from new_end. 726 */ 727 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 728 if (new_end == high_avail) 729 high_avail -= PAGE_SIZE; 730 new_end -= PAGE_SIZE; 731 } 732 } 733 end = new_end; 734 735 /* 736 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 737 * However, because this page is allocated from KVM, out-of-bounds 738 * accesses using the direct map will not be trapped. 739 */ 740 vaddr += PAGE_SIZE; 741 742 /* 743 * Allocate physical memory for the page structures, and map it. 744 */ 745 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 746 mapped = pmap_map(&vaddr, new_end, end, 747 VM_PROT_READ | VM_PROT_WRITE); 748 vm_page_array = (vm_page_t)mapped; 749 vm_page_array_size = page_range; 750 751 #if VM_NRESERVLEVEL > 0 752 /* 753 * Allocate physical memory for the reservation management system's 754 * data structures, and map it. 755 */ 756 if (high_avail == end) 757 high_avail = new_end; 758 new_end = vm_reserv_startup(&vaddr, new_end, high_avail); 759 #endif 760 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 761 defined(__riscv) 762 /* 763 * Include vm_page_array and vm_reserv_array in a crash dump. 764 */ 765 for (pa = new_end; pa < end; pa += PAGE_SIZE) 766 dump_add_page(pa); 767 #endif 768 phys_avail[biggestone + 1] = new_end; 769 770 /* 771 * Add physical memory segments corresponding to the available 772 * physical pages. 773 */ 774 for (i = 0; phys_avail[i + 1] != 0; i += 2) 775 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 776 777 /* 778 * Initialize the physical memory allocator. 779 */ 780 vm_phys_init(); 781 782 /* 783 * Initialize the page structures and add every available page to the 784 * physical memory allocator's free lists. 785 */ 786 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) 787 for (ii = 0; ii < vm_page_array_size; ii++) { 788 m = &vm_page_array[ii]; 789 vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); 790 m->flags = PG_FICTITIOUS; 791 } 792 #endif 793 vm_cnt.v_page_count = 0; 794 for (segind = 0; segind < vm_phys_nsegs; segind++) { 795 seg = &vm_phys_segs[segind]; 796 for (m = seg->first_page, pa = seg->start; pa < seg->end; 797 m++, pa += PAGE_SIZE) 798 vm_page_init_page(m, pa, segind); 799 800 /* 801 * Add the segment to the free lists only if it is covered by 802 * one of the ranges in phys_avail. Because we've added the 803 * ranges to the vm_phys_segs array, we can assume that each 804 * segment is either entirely contained in one of the ranges, 805 * or doesn't overlap any of them. 806 */ 807 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 808 struct vm_domain *vmd; 809 810 if (seg->start < phys_avail[i] || 811 seg->end > phys_avail[i + 1]) 812 continue; 813 814 m = seg->first_page; 815 pagecount = (u_long)atop(seg->end - seg->start); 816 817 vmd = VM_DOMAIN(seg->domain); 818 vm_domain_free_lock(vmd); 819 vm_phys_enqueue_contig(m, pagecount); 820 vm_domain_free_unlock(vmd); 821 vm_domain_freecnt_inc(vmd, pagecount); 822 vm_cnt.v_page_count += (u_int)pagecount; 823 824 vmd = VM_DOMAIN(seg->domain); 825 vmd->vmd_page_count += (u_int)pagecount; 826 vmd->vmd_segs |= 1UL << m->segind; 827 break; 828 } 829 } 830 831 /* 832 * Remove blacklisted pages from the physical memory allocator. 833 */ 834 TAILQ_INIT(&blacklist_head); 835 vm_page_blacklist_load(&list, &listend); 836 vm_page_blacklist_check(list, listend); 837 838 list = kern_getenv("vm.blacklist"); 839 vm_page_blacklist_check(list, NULL); 840 841 freeenv(list); 842 #if VM_NRESERVLEVEL > 0 843 /* 844 * Initialize the reservation management system. 845 */ 846 vm_reserv_init(); 847 #endif 848 849 return (vaddr); 850 } 851 852 void 853 vm_page_reference(vm_page_t m) 854 { 855 856 vm_page_aflag_set(m, PGA_REFERENCED); 857 } 858 859 /* 860 * vm_page_busy_downgrade: 861 * 862 * Downgrade an exclusive busy page into a single shared busy page. 863 */ 864 void 865 vm_page_busy_downgrade(vm_page_t m) 866 { 867 u_int x; 868 bool locked; 869 870 vm_page_assert_xbusied(m); 871 locked = mtx_owned(vm_page_lockptr(m)); 872 873 for (;;) { 874 x = m->busy_lock; 875 x &= VPB_BIT_WAITERS; 876 if (x != 0 && !locked) 877 vm_page_lock(m); 878 if (atomic_cmpset_rel_int(&m->busy_lock, 879 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 880 break; 881 if (x != 0 && !locked) 882 vm_page_unlock(m); 883 } 884 if (x != 0) { 885 wakeup(m); 886 if (!locked) 887 vm_page_unlock(m); 888 } 889 } 890 891 /* 892 * vm_page_sbusied: 893 * 894 * Return a positive value if the page is shared busied, 0 otherwise. 895 */ 896 int 897 vm_page_sbusied(vm_page_t m) 898 { 899 u_int x; 900 901 x = m->busy_lock; 902 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 903 } 904 905 /* 906 * vm_page_sunbusy: 907 * 908 * Shared unbusy a page. 909 */ 910 void 911 vm_page_sunbusy(vm_page_t m) 912 { 913 u_int x; 914 915 vm_page_lock_assert(m, MA_NOTOWNED); 916 vm_page_assert_sbusied(m); 917 918 for (;;) { 919 x = m->busy_lock; 920 if (VPB_SHARERS(x) > 1) { 921 if (atomic_cmpset_int(&m->busy_lock, x, 922 x - VPB_ONE_SHARER)) 923 break; 924 continue; 925 } 926 if ((x & VPB_BIT_WAITERS) == 0) { 927 KASSERT(x == VPB_SHARERS_WORD(1), 928 ("vm_page_sunbusy: invalid lock state")); 929 if (atomic_cmpset_int(&m->busy_lock, 930 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 931 break; 932 continue; 933 } 934 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 935 ("vm_page_sunbusy: invalid lock state for waiters")); 936 937 vm_page_lock(m); 938 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 939 vm_page_unlock(m); 940 continue; 941 } 942 wakeup(m); 943 vm_page_unlock(m); 944 break; 945 } 946 } 947 948 /* 949 * vm_page_busy_sleep: 950 * 951 * Sleep and release the page lock, using the page pointer as wchan. 952 * This is used to implement the hard-path of busying mechanism. 953 * 954 * The given page must be locked. 955 * 956 * If nonshared is true, sleep only if the page is xbusy. 957 */ 958 void 959 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 960 { 961 u_int x; 962 963 vm_page_assert_locked(m); 964 965 x = m->busy_lock; 966 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 967 ((x & VPB_BIT_WAITERS) == 0 && 968 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 969 vm_page_unlock(m); 970 return; 971 } 972 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 973 } 974 975 /* 976 * vm_page_trysbusy: 977 * 978 * Try to shared busy a page. 979 * If the operation succeeds 1 is returned otherwise 0. 980 * The operation never sleeps. 981 */ 982 int 983 vm_page_trysbusy(vm_page_t m) 984 { 985 u_int x; 986 987 for (;;) { 988 x = m->busy_lock; 989 if ((x & VPB_BIT_SHARED) == 0) 990 return (0); 991 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 992 return (1); 993 } 994 } 995 996 static void 997 vm_page_xunbusy_locked(vm_page_t m) 998 { 999 1000 vm_page_assert_xbusied(m); 1001 vm_page_assert_locked(m); 1002 1003 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1004 /* There is a waiter, do wakeup() instead of vm_page_flash(). */ 1005 wakeup(m); 1006 } 1007 1008 void 1009 vm_page_xunbusy_maybelocked(vm_page_t m) 1010 { 1011 bool lockacq; 1012 1013 vm_page_assert_xbusied(m); 1014 1015 /* 1016 * Fast path for unbusy. If it succeeds, we know that there 1017 * are no waiters, so we do not need a wakeup. 1018 */ 1019 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER, 1020 VPB_UNBUSIED)) 1021 return; 1022 1023 lockacq = !mtx_owned(vm_page_lockptr(m)); 1024 if (lockacq) 1025 vm_page_lock(m); 1026 vm_page_xunbusy_locked(m); 1027 if (lockacq) 1028 vm_page_unlock(m); 1029 } 1030 1031 /* 1032 * vm_page_xunbusy_hard: 1033 * 1034 * Called after the first try the exclusive unbusy of a page failed. 1035 * It is assumed that the waiters bit is on. 1036 */ 1037 void 1038 vm_page_xunbusy_hard(vm_page_t m) 1039 { 1040 1041 vm_page_assert_xbusied(m); 1042 1043 vm_page_lock(m); 1044 vm_page_xunbusy_locked(m); 1045 vm_page_unlock(m); 1046 } 1047 1048 /* 1049 * vm_page_flash: 1050 * 1051 * Wakeup anyone waiting for the page. 1052 * The ownership bits do not change. 1053 * 1054 * The given page must be locked. 1055 */ 1056 void 1057 vm_page_flash(vm_page_t m) 1058 { 1059 u_int x; 1060 1061 vm_page_lock_assert(m, MA_OWNED); 1062 1063 for (;;) { 1064 x = m->busy_lock; 1065 if ((x & VPB_BIT_WAITERS) == 0) 1066 return; 1067 if (atomic_cmpset_int(&m->busy_lock, x, 1068 x & (~VPB_BIT_WAITERS))) 1069 break; 1070 } 1071 wakeup(m); 1072 } 1073 1074 /* 1075 * Avoid releasing and reacquiring the same page lock. 1076 */ 1077 void 1078 vm_page_change_lock(vm_page_t m, struct mtx **mtx) 1079 { 1080 struct mtx *mtx1; 1081 1082 mtx1 = vm_page_lockptr(m); 1083 if (*mtx == mtx1) 1084 return; 1085 if (*mtx != NULL) 1086 mtx_unlock(*mtx); 1087 *mtx = mtx1; 1088 mtx_lock(mtx1); 1089 } 1090 1091 /* 1092 * Keep page from being freed by the page daemon 1093 * much of the same effect as wiring, except much lower 1094 * overhead and should be used only for *very* temporary 1095 * holding ("wiring"). 1096 */ 1097 void 1098 vm_page_hold(vm_page_t mem) 1099 { 1100 1101 vm_page_lock_assert(mem, MA_OWNED); 1102 mem->hold_count++; 1103 } 1104 1105 void 1106 vm_page_unhold(vm_page_t mem) 1107 { 1108 1109 vm_page_lock_assert(mem, MA_OWNED); 1110 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 1111 --mem->hold_count; 1112 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 1113 vm_page_free_toq(mem); 1114 } 1115 1116 /* 1117 * vm_page_unhold_pages: 1118 * 1119 * Unhold each of the pages that is referenced by the given array. 1120 */ 1121 void 1122 vm_page_unhold_pages(vm_page_t *ma, int count) 1123 { 1124 struct mtx *mtx; 1125 1126 mtx = NULL; 1127 for (; count != 0; count--) { 1128 vm_page_change_lock(*ma, &mtx); 1129 vm_page_unhold(*ma); 1130 ma++; 1131 } 1132 if (mtx != NULL) 1133 mtx_unlock(mtx); 1134 } 1135 1136 vm_page_t 1137 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1138 { 1139 vm_page_t m; 1140 1141 #ifdef VM_PHYSSEG_SPARSE 1142 m = vm_phys_paddr_to_vm_page(pa); 1143 if (m == NULL) 1144 m = vm_phys_fictitious_to_vm_page(pa); 1145 return (m); 1146 #elif defined(VM_PHYSSEG_DENSE) 1147 long pi; 1148 1149 pi = atop(pa); 1150 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1151 m = &vm_page_array[pi - first_page]; 1152 return (m); 1153 } 1154 return (vm_phys_fictitious_to_vm_page(pa)); 1155 #else 1156 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1157 #endif 1158 } 1159 1160 /* 1161 * vm_page_getfake: 1162 * 1163 * Create a fictitious page with the specified physical address and 1164 * memory attribute. The memory attribute is the only the machine- 1165 * dependent aspect of a fictitious page that must be initialized. 1166 */ 1167 vm_page_t 1168 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1169 { 1170 vm_page_t m; 1171 1172 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1173 vm_page_initfake(m, paddr, memattr); 1174 return (m); 1175 } 1176 1177 void 1178 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1179 { 1180 1181 if ((m->flags & PG_FICTITIOUS) != 0) { 1182 /* 1183 * The page's memattr might have changed since the 1184 * previous initialization. Update the pmap to the 1185 * new memattr. 1186 */ 1187 goto memattr; 1188 } 1189 m->phys_addr = paddr; 1190 m->queue = PQ_NONE; 1191 /* Fictitious pages don't use "segind". */ 1192 m->flags = PG_FICTITIOUS; 1193 /* Fictitious pages don't use "order" or "pool". */ 1194 m->oflags = VPO_UNMANAGED; 1195 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1196 m->wire_count = 1; 1197 pmap_page_init(m); 1198 memattr: 1199 pmap_page_set_memattr(m, memattr); 1200 } 1201 1202 /* 1203 * vm_page_putfake: 1204 * 1205 * Release a fictitious page. 1206 */ 1207 void 1208 vm_page_putfake(vm_page_t m) 1209 { 1210 1211 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1212 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1213 ("vm_page_putfake: bad page %p", m)); 1214 uma_zfree(fakepg_zone, m); 1215 } 1216 1217 /* 1218 * vm_page_updatefake: 1219 * 1220 * Update the given fictitious page to the specified physical address and 1221 * memory attribute. 1222 */ 1223 void 1224 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1225 { 1226 1227 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1228 ("vm_page_updatefake: bad page %p", m)); 1229 m->phys_addr = paddr; 1230 pmap_page_set_memattr(m, memattr); 1231 } 1232 1233 /* 1234 * vm_page_free: 1235 * 1236 * Free a page. 1237 */ 1238 void 1239 vm_page_free(vm_page_t m) 1240 { 1241 1242 m->flags &= ~PG_ZERO; 1243 vm_page_free_toq(m); 1244 } 1245 1246 /* 1247 * vm_page_free_zero: 1248 * 1249 * Free a page to the zerod-pages queue 1250 */ 1251 void 1252 vm_page_free_zero(vm_page_t m) 1253 { 1254 1255 m->flags |= PG_ZERO; 1256 vm_page_free_toq(m); 1257 } 1258 1259 /* 1260 * Unbusy and handle the page queueing for a page from a getpages request that 1261 * was optionally read ahead or behind. 1262 */ 1263 void 1264 vm_page_readahead_finish(vm_page_t m) 1265 { 1266 1267 /* We shouldn't put invalid pages on queues. */ 1268 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); 1269 1270 /* 1271 * Since the page is not the actually needed one, whether it should 1272 * be activated or deactivated is not obvious. Empirical results 1273 * have shown that deactivating the page is usually the best choice, 1274 * unless the page is wanted by another thread. 1275 */ 1276 vm_page_lock(m); 1277 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1278 vm_page_activate(m); 1279 else 1280 vm_page_deactivate(m); 1281 vm_page_unlock(m); 1282 vm_page_xunbusy(m); 1283 } 1284 1285 /* 1286 * vm_page_sleep_if_busy: 1287 * 1288 * Sleep and release the page queues lock if the page is busied. 1289 * Returns TRUE if the thread slept. 1290 * 1291 * The given page must be unlocked and object containing it must 1292 * be locked. 1293 */ 1294 int 1295 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1296 { 1297 vm_object_t obj; 1298 1299 vm_page_lock_assert(m, MA_NOTOWNED); 1300 VM_OBJECT_ASSERT_WLOCKED(m->object); 1301 1302 if (vm_page_busied(m)) { 1303 /* 1304 * The page-specific object must be cached because page 1305 * identity can change during the sleep, causing the 1306 * re-lock of a different object. 1307 * It is assumed that a reference to the object is already 1308 * held by the callers. 1309 */ 1310 obj = m->object; 1311 vm_page_lock(m); 1312 VM_OBJECT_WUNLOCK(obj); 1313 vm_page_busy_sleep(m, msg, false); 1314 VM_OBJECT_WLOCK(obj); 1315 return (TRUE); 1316 } 1317 return (FALSE); 1318 } 1319 1320 /* 1321 * vm_page_dirty_KBI: [ internal use only ] 1322 * 1323 * Set all bits in the page's dirty field. 1324 * 1325 * The object containing the specified page must be locked if the 1326 * call is made from the machine-independent layer. 1327 * 1328 * See vm_page_clear_dirty_mask(). 1329 * 1330 * This function should only be called by vm_page_dirty(). 1331 */ 1332 void 1333 vm_page_dirty_KBI(vm_page_t m) 1334 { 1335 1336 /* Refer to this operation by its public name. */ 1337 KASSERT(m->valid == VM_PAGE_BITS_ALL, 1338 ("vm_page_dirty: page is invalid!")); 1339 m->dirty = VM_PAGE_BITS_ALL; 1340 } 1341 1342 /* 1343 * vm_page_insert: [ internal use only ] 1344 * 1345 * Inserts the given mem entry into the object and object list. 1346 * 1347 * The object must be locked. 1348 */ 1349 int 1350 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1351 { 1352 vm_page_t mpred; 1353 1354 VM_OBJECT_ASSERT_WLOCKED(object); 1355 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1356 return (vm_page_insert_after(m, object, pindex, mpred)); 1357 } 1358 1359 /* 1360 * vm_page_insert_after: 1361 * 1362 * Inserts the page "m" into the specified object at offset "pindex". 1363 * 1364 * The page "mpred" must immediately precede the offset "pindex" within 1365 * the specified object. 1366 * 1367 * The object must be locked. 1368 */ 1369 static int 1370 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1371 vm_page_t mpred) 1372 { 1373 vm_page_t msucc; 1374 1375 VM_OBJECT_ASSERT_WLOCKED(object); 1376 KASSERT(m->object == NULL, 1377 ("vm_page_insert_after: page already inserted")); 1378 if (mpred != NULL) { 1379 KASSERT(mpred->object == object, 1380 ("vm_page_insert_after: object doesn't contain mpred")); 1381 KASSERT(mpred->pindex < pindex, 1382 ("vm_page_insert_after: mpred doesn't precede pindex")); 1383 msucc = TAILQ_NEXT(mpred, listq); 1384 } else 1385 msucc = TAILQ_FIRST(&object->memq); 1386 if (msucc != NULL) 1387 KASSERT(msucc->pindex > pindex, 1388 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1389 1390 /* 1391 * Record the object/offset pair in this page 1392 */ 1393 m->object = object; 1394 m->pindex = pindex; 1395 1396 /* 1397 * Now link into the object's ordered list of backed pages. 1398 */ 1399 if (vm_radix_insert(&object->rtree, m)) { 1400 m->object = NULL; 1401 m->pindex = 0; 1402 return (1); 1403 } 1404 vm_page_insert_radixdone(m, object, mpred); 1405 return (0); 1406 } 1407 1408 /* 1409 * vm_page_insert_radixdone: 1410 * 1411 * Complete page "m" insertion into the specified object after the 1412 * radix trie hooking. 1413 * 1414 * The page "mpred" must precede the offset "m->pindex" within the 1415 * specified object. 1416 * 1417 * The object must be locked. 1418 */ 1419 static void 1420 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1421 { 1422 1423 VM_OBJECT_ASSERT_WLOCKED(object); 1424 KASSERT(object != NULL && m->object == object, 1425 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1426 if (mpred != NULL) { 1427 KASSERT(mpred->object == object, 1428 ("vm_page_insert_after: object doesn't contain mpred")); 1429 KASSERT(mpred->pindex < m->pindex, 1430 ("vm_page_insert_after: mpred doesn't precede pindex")); 1431 } 1432 1433 if (mpred != NULL) 1434 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1435 else 1436 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1437 1438 /* 1439 * Show that the object has one more resident page. 1440 */ 1441 object->resident_page_count++; 1442 1443 /* 1444 * Hold the vnode until the last page is released. 1445 */ 1446 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1447 vhold(object->handle); 1448 1449 /* 1450 * Since we are inserting a new and possibly dirty page, 1451 * update the object's OBJ_MIGHTBEDIRTY flag. 1452 */ 1453 if (pmap_page_is_write_mapped(m)) 1454 vm_object_set_writeable_dirty(object); 1455 } 1456 1457 /* 1458 * vm_page_remove: 1459 * 1460 * Removes the specified page from its containing object, but does not 1461 * invalidate any backing storage. 1462 * 1463 * The object must be locked. The page must be locked if it is managed. 1464 */ 1465 void 1466 vm_page_remove(vm_page_t m) 1467 { 1468 vm_object_t object; 1469 vm_page_t mrem; 1470 1471 if ((m->oflags & VPO_UNMANAGED) == 0) 1472 vm_page_assert_locked(m); 1473 if ((object = m->object) == NULL) 1474 return; 1475 VM_OBJECT_ASSERT_WLOCKED(object); 1476 if (vm_page_xbusied(m)) 1477 vm_page_xunbusy_maybelocked(m); 1478 mrem = vm_radix_remove(&object->rtree, m->pindex); 1479 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1480 1481 /* 1482 * Now remove from the object's list of backed pages. 1483 */ 1484 TAILQ_REMOVE(&object->memq, m, listq); 1485 1486 /* 1487 * And show that the object has one fewer resident page. 1488 */ 1489 object->resident_page_count--; 1490 1491 /* 1492 * The vnode may now be recycled. 1493 */ 1494 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1495 vdrop(object->handle); 1496 1497 m->object = NULL; 1498 } 1499 1500 /* 1501 * vm_page_lookup: 1502 * 1503 * Returns the page associated with the object/offset 1504 * pair specified; if none is found, NULL is returned. 1505 * 1506 * The object must be locked. 1507 */ 1508 vm_page_t 1509 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1510 { 1511 1512 VM_OBJECT_ASSERT_LOCKED(object); 1513 return (vm_radix_lookup(&object->rtree, pindex)); 1514 } 1515 1516 /* 1517 * vm_page_find_least: 1518 * 1519 * Returns the page associated with the object with least pindex 1520 * greater than or equal to the parameter pindex, or NULL. 1521 * 1522 * The object must be locked. 1523 */ 1524 vm_page_t 1525 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1526 { 1527 vm_page_t m; 1528 1529 VM_OBJECT_ASSERT_LOCKED(object); 1530 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1531 m = vm_radix_lookup_ge(&object->rtree, pindex); 1532 return (m); 1533 } 1534 1535 /* 1536 * Returns the given page's successor (by pindex) within the object if it is 1537 * resident; if none is found, NULL is returned. 1538 * 1539 * The object must be locked. 1540 */ 1541 vm_page_t 1542 vm_page_next(vm_page_t m) 1543 { 1544 vm_page_t next; 1545 1546 VM_OBJECT_ASSERT_LOCKED(m->object); 1547 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1548 MPASS(next->object == m->object); 1549 if (next->pindex != m->pindex + 1) 1550 next = NULL; 1551 } 1552 return (next); 1553 } 1554 1555 /* 1556 * Returns the given page's predecessor (by pindex) within the object if it is 1557 * resident; if none is found, NULL is returned. 1558 * 1559 * The object must be locked. 1560 */ 1561 vm_page_t 1562 vm_page_prev(vm_page_t m) 1563 { 1564 vm_page_t prev; 1565 1566 VM_OBJECT_ASSERT_LOCKED(m->object); 1567 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1568 MPASS(prev->object == m->object); 1569 if (prev->pindex != m->pindex - 1) 1570 prev = NULL; 1571 } 1572 return (prev); 1573 } 1574 1575 /* 1576 * Uses the page mnew as a replacement for an existing page at index 1577 * pindex which must be already present in the object. 1578 * 1579 * The existing page must not be on a paging queue. 1580 */ 1581 vm_page_t 1582 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1583 { 1584 vm_page_t mold; 1585 1586 VM_OBJECT_ASSERT_WLOCKED(object); 1587 KASSERT(mnew->object == NULL, 1588 ("vm_page_replace: page %p already in object", mnew)); 1589 KASSERT(mnew->queue == PQ_NONE, 1590 ("vm_page_replace: new page %p is on a paging queue", mnew)); 1591 1592 /* 1593 * This function mostly follows vm_page_insert() and 1594 * vm_page_remove() without the radix, object count and vnode 1595 * dance. Double check such functions for more comments. 1596 */ 1597 1598 mnew->object = object; 1599 mnew->pindex = pindex; 1600 mold = vm_radix_replace(&object->rtree, mnew); 1601 KASSERT(mold->queue == PQ_NONE, 1602 ("vm_page_replace: old page %p is on a paging queue", mold)); 1603 1604 /* Keep the resident page list in sorted order. */ 1605 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1606 TAILQ_REMOVE(&object->memq, mold, listq); 1607 1608 mold->object = NULL; 1609 vm_page_xunbusy_maybelocked(mold); 1610 1611 /* 1612 * The object's resident_page_count does not change because we have 1613 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1614 */ 1615 if (pmap_page_is_write_mapped(mnew)) 1616 vm_object_set_writeable_dirty(object); 1617 return (mold); 1618 } 1619 1620 /* 1621 * vm_page_rename: 1622 * 1623 * Move the given memory entry from its 1624 * current object to the specified target object/offset. 1625 * 1626 * Note: swap associated with the page must be invalidated by the move. We 1627 * have to do this for several reasons: (1) we aren't freeing the 1628 * page, (2) we are dirtying the page, (3) the VM system is probably 1629 * moving the page from object A to B, and will then later move 1630 * the backing store from A to B and we can't have a conflict. 1631 * 1632 * Note: we *always* dirty the page. It is necessary both for the 1633 * fact that we moved it, and because we may be invalidating 1634 * swap. 1635 * 1636 * The objects must be locked. 1637 */ 1638 int 1639 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1640 { 1641 vm_page_t mpred; 1642 vm_pindex_t opidx; 1643 1644 VM_OBJECT_ASSERT_WLOCKED(new_object); 1645 1646 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1647 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1648 ("vm_page_rename: pindex already renamed")); 1649 1650 /* 1651 * Create a custom version of vm_page_insert() which does not depend 1652 * by m_prev and can cheat on the implementation aspects of the 1653 * function. 1654 */ 1655 opidx = m->pindex; 1656 m->pindex = new_pindex; 1657 if (vm_radix_insert(&new_object->rtree, m)) { 1658 m->pindex = opidx; 1659 return (1); 1660 } 1661 1662 /* 1663 * The operation cannot fail anymore. The removal must happen before 1664 * the listq iterator is tainted. 1665 */ 1666 m->pindex = opidx; 1667 vm_page_lock(m); 1668 vm_page_remove(m); 1669 1670 /* Return back to the new pindex to complete vm_page_insert(). */ 1671 m->pindex = new_pindex; 1672 m->object = new_object; 1673 vm_page_unlock(m); 1674 vm_page_insert_radixdone(m, new_object, mpred); 1675 vm_page_dirty(m); 1676 return (0); 1677 } 1678 1679 /* 1680 * vm_page_alloc: 1681 * 1682 * Allocate and return a page that is associated with the specified 1683 * object and offset pair. By default, this page is exclusive busied. 1684 * 1685 * The caller must always specify an allocation class. 1686 * 1687 * allocation classes: 1688 * VM_ALLOC_NORMAL normal process request 1689 * VM_ALLOC_SYSTEM system *really* needs a page 1690 * VM_ALLOC_INTERRUPT interrupt time request 1691 * 1692 * optional allocation flags: 1693 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1694 * intends to allocate 1695 * VM_ALLOC_NOBUSY do not exclusive busy the page 1696 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1697 * VM_ALLOC_NOOBJ page is not associated with an object and 1698 * should not be exclusive busy 1699 * VM_ALLOC_SBUSY shared busy the allocated page 1700 * VM_ALLOC_WIRED wire the allocated page 1701 * VM_ALLOC_ZERO prefer a zeroed page 1702 */ 1703 vm_page_t 1704 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1705 { 1706 1707 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1708 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1709 } 1710 1711 vm_page_t 1712 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1713 int req) 1714 { 1715 1716 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1717 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1718 NULL)); 1719 } 1720 1721 /* 1722 * Allocate a page in the specified object with the given page index. To 1723 * optimize insertion of the page into the object, the caller must also specifiy 1724 * the resident page in the object with largest index smaller than the given 1725 * page index, or NULL if no such page exists. 1726 */ 1727 vm_page_t 1728 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1729 int req, vm_page_t mpred) 1730 { 1731 struct vm_domainset_iter di; 1732 vm_page_t m; 1733 int domain; 1734 1735 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1736 do { 1737 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1738 mpred); 1739 if (m != NULL) 1740 break; 1741 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1742 1743 return (m); 1744 } 1745 1746 /* 1747 * Returns true if the number of free pages exceeds the minimum 1748 * for the request class and false otherwise. 1749 */ 1750 int 1751 vm_domain_allocate(struct vm_domain *vmd, int req, int npages) 1752 { 1753 u_int limit, old, new; 1754 1755 req = req & VM_ALLOC_CLASS_MASK; 1756 1757 /* 1758 * The page daemon is allowed to dig deeper into the free page list. 1759 */ 1760 if (curproc == pageproc && req != VM_ALLOC_INTERRUPT) 1761 req = VM_ALLOC_SYSTEM; 1762 if (req == VM_ALLOC_INTERRUPT) 1763 limit = 0; 1764 else if (req == VM_ALLOC_SYSTEM) 1765 limit = vmd->vmd_interrupt_free_min; 1766 else 1767 limit = vmd->vmd_free_reserved; 1768 1769 /* 1770 * Attempt to reserve the pages. Fail if we're below the limit. 1771 */ 1772 limit += npages; 1773 old = vmd->vmd_free_count; 1774 do { 1775 if (old < limit) 1776 return (0); 1777 new = old - npages; 1778 } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); 1779 1780 /* Wake the page daemon if we've crossed the threshold. */ 1781 if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) 1782 pagedaemon_wakeup(vmd->vmd_domain); 1783 1784 /* Only update bitsets on transitions. */ 1785 if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || 1786 (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) 1787 vm_domain_set(vmd); 1788 1789 return (1); 1790 } 1791 1792 vm_page_t 1793 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 1794 int req, vm_page_t mpred) 1795 { 1796 struct vm_domain *vmd; 1797 vm_page_t m; 1798 int flags; 1799 1800 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1801 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1802 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1803 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1804 ("inconsistent object(%p)/req(%x)", object, req)); 1805 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1806 ("Can't sleep and retry object insertion.")); 1807 KASSERT(mpred == NULL || mpred->pindex < pindex, 1808 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1809 (uintmax_t)pindex)); 1810 if (object != NULL) 1811 VM_OBJECT_ASSERT_WLOCKED(object); 1812 1813 again: 1814 m = NULL; 1815 #if VM_NRESERVLEVEL > 0 1816 /* 1817 * Can we allocate the page from a reservation? 1818 */ 1819 if (vm_object_reserv(object) && 1820 (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != 1821 NULL) { 1822 domain = vm_phys_domain(m); 1823 vmd = VM_DOMAIN(domain); 1824 goto found; 1825 } 1826 #endif 1827 vmd = VM_DOMAIN(domain); 1828 if (object != NULL && vmd->vmd_pgcache != NULL) { 1829 m = uma_zalloc(vmd->vmd_pgcache, M_NOWAIT); 1830 if (m != NULL) 1831 goto found; 1832 } 1833 if (vm_domain_allocate(vmd, req, 1)) { 1834 /* 1835 * If not, allocate it from the free page queues. 1836 */ 1837 vm_domain_free_lock(vmd); 1838 m = vm_phys_alloc_pages(domain, object != NULL ? 1839 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1840 vm_domain_free_unlock(vmd); 1841 if (m == NULL) { 1842 vm_domain_freecnt_inc(vmd, 1); 1843 #if VM_NRESERVLEVEL > 0 1844 if (vm_reserv_reclaim_inactive(domain)) 1845 goto again; 1846 #endif 1847 } 1848 } 1849 if (m == NULL) { 1850 /* 1851 * Not allocatable, give up. 1852 */ 1853 if (vm_domain_alloc_fail(vmd, object, req)) 1854 goto again; 1855 return (NULL); 1856 } 1857 1858 /* 1859 * At this point we had better have found a good page. 1860 */ 1861 KASSERT(m != NULL, ("missing page")); 1862 1863 found: 1864 vm_page_dequeue(m); 1865 vm_page_alloc_check(m); 1866 1867 /* 1868 * Initialize the page. Only the PG_ZERO flag is inherited. 1869 */ 1870 flags = 0; 1871 if ((req & VM_ALLOC_ZERO) != 0) 1872 flags = PG_ZERO; 1873 flags &= m->flags; 1874 if ((req & VM_ALLOC_NODUMP) != 0) 1875 flags |= PG_NODUMP; 1876 m->flags = flags; 1877 m->aflags = 0; 1878 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1879 VPO_UNMANAGED : 0; 1880 m->busy_lock = VPB_UNBUSIED; 1881 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1882 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1883 if ((req & VM_ALLOC_SBUSY) != 0) 1884 m->busy_lock = VPB_SHARERS_WORD(1); 1885 if (req & VM_ALLOC_WIRED) { 1886 /* 1887 * The page lock is not required for wiring a page until that 1888 * page is inserted into the object. 1889 */ 1890 vm_wire_add(1); 1891 m->wire_count = 1; 1892 } 1893 m->act_count = 0; 1894 1895 if (object != NULL) { 1896 if (vm_page_insert_after(m, object, pindex, mpred)) { 1897 if (req & VM_ALLOC_WIRED) { 1898 vm_wire_sub(1); 1899 m->wire_count = 0; 1900 } 1901 KASSERT(m->object == NULL, ("page %p has object", m)); 1902 m->oflags = VPO_UNMANAGED; 1903 m->busy_lock = VPB_UNBUSIED; 1904 /* Don't change PG_ZERO. */ 1905 vm_page_free_toq(m); 1906 if (req & VM_ALLOC_WAITFAIL) { 1907 VM_OBJECT_WUNLOCK(object); 1908 vm_radix_wait(); 1909 VM_OBJECT_WLOCK(object); 1910 } 1911 return (NULL); 1912 } 1913 1914 /* Ignore device objects; the pager sets "memattr" for them. */ 1915 if (object->memattr != VM_MEMATTR_DEFAULT && 1916 (object->flags & OBJ_FICTITIOUS) == 0) 1917 pmap_page_set_memattr(m, object->memattr); 1918 } else 1919 m->pindex = pindex; 1920 1921 return (m); 1922 } 1923 1924 /* 1925 * vm_page_alloc_contig: 1926 * 1927 * Allocate a contiguous set of physical pages of the given size "npages" 1928 * from the free lists. All of the physical pages must be at or above 1929 * the given physical address "low" and below the given physical address 1930 * "high". The given value "alignment" determines the alignment of the 1931 * first physical page in the set. If the given value "boundary" is 1932 * non-zero, then the set of physical pages cannot cross any physical 1933 * address boundary that is a multiple of that value. Both "alignment" 1934 * and "boundary" must be a power of two. 1935 * 1936 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1937 * then the memory attribute setting for the physical pages is configured 1938 * to the object's memory attribute setting. Otherwise, the memory 1939 * attribute setting for the physical pages is configured to "memattr", 1940 * overriding the object's memory attribute setting. However, if the 1941 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1942 * memory attribute setting for the physical pages cannot be configured 1943 * to VM_MEMATTR_DEFAULT. 1944 * 1945 * The specified object may not contain fictitious pages. 1946 * 1947 * The caller must always specify an allocation class. 1948 * 1949 * allocation classes: 1950 * VM_ALLOC_NORMAL normal process request 1951 * VM_ALLOC_SYSTEM system *really* needs a page 1952 * VM_ALLOC_INTERRUPT interrupt time request 1953 * 1954 * optional allocation flags: 1955 * VM_ALLOC_NOBUSY do not exclusive busy the page 1956 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1957 * VM_ALLOC_NOOBJ page is not associated with an object and 1958 * should not be exclusive busy 1959 * VM_ALLOC_SBUSY shared busy the allocated page 1960 * VM_ALLOC_WIRED wire the allocated page 1961 * VM_ALLOC_ZERO prefer a zeroed page 1962 */ 1963 vm_page_t 1964 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1965 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1966 vm_paddr_t boundary, vm_memattr_t memattr) 1967 { 1968 struct vm_domainset_iter di; 1969 vm_page_t m; 1970 int domain; 1971 1972 vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); 1973 do { 1974 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 1975 npages, low, high, alignment, boundary, memattr); 1976 if (m != NULL) 1977 break; 1978 } while (vm_domainset_iter_page(&di, object, &domain) == 0); 1979 1980 return (m); 1981 } 1982 1983 vm_page_t 1984 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1985 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1986 vm_paddr_t boundary, vm_memattr_t memattr) 1987 { 1988 struct vm_domain *vmd; 1989 vm_page_t m, m_ret, mpred; 1990 u_int busy_lock, flags, oflags; 1991 1992 mpred = NULL; /* XXX: pacify gcc */ 1993 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1994 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1995 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1996 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1997 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 1998 req)); 1999 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 2000 ("Can't sleep and retry object insertion.")); 2001 if (object != NULL) { 2002 VM_OBJECT_ASSERT_WLOCKED(object); 2003 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 2004 ("vm_page_alloc_contig: object %p has fictitious pages", 2005 object)); 2006 } 2007 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 2008 2009 if (object != NULL) { 2010 mpred = vm_radix_lookup_le(&object->rtree, pindex); 2011 KASSERT(mpred == NULL || mpred->pindex != pindex, 2012 ("vm_page_alloc_contig: pindex already allocated")); 2013 } 2014 2015 /* 2016 * Can we allocate the pages without the number of free pages falling 2017 * below the lower bound for the allocation class? 2018 */ 2019 again: 2020 #if VM_NRESERVLEVEL > 0 2021 /* 2022 * Can we allocate the pages from a reservation? 2023 */ 2024 if (vm_object_reserv(object) && 2025 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, 2026 mpred, npages, low, high, alignment, boundary)) != NULL) { 2027 domain = vm_phys_domain(m_ret); 2028 vmd = VM_DOMAIN(domain); 2029 goto found; 2030 } 2031 #endif 2032 m_ret = NULL; 2033 vmd = VM_DOMAIN(domain); 2034 if (vm_domain_allocate(vmd, req, npages)) { 2035 /* 2036 * allocate them from the free page queues. 2037 */ 2038 vm_domain_free_lock(vmd); 2039 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 2040 alignment, boundary); 2041 vm_domain_free_unlock(vmd); 2042 if (m_ret == NULL) { 2043 vm_domain_freecnt_inc(vmd, npages); 2044 #if VM_NRESERVLEVEL > 0 2045 if (vm_reserv_reclaim_contig(domain, npages, low, 2046 high, alignment, boundary)) 2047 goto again; 2048 #endif 2049 } 2050 } 2051 if (m_ret == NULL) { 2052 if (vm_domain_alloc_fail(vmd, object, req)) 2053 goto again; 2054 return (NULL); 2055 } 2056 #if VM_NRESERVLEVEL > 0 2057 found: 2058 #endif 2059 for (m = m_ret; m < &m_ret[npages]; m++) { 2060 vm_page_dequeue(m); 2061 vm_page_alloc_check(m); 2062 } 2063 2064 /* 2065 * Initialize the pages. Only the PG_ZERO flag is inherited. 2066 */ 2067 flags = 0; 2068 if ((req & VM_ALLOC_ZERO) != 0) 2069 flags = PG_ZERO; 2070 if ((req & VM_ALLOC_NODUMP) != 0) 2071 flags |= PG_NODUMP; 2072 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 2073 VPO_UNMANAGED : 0; 2074 busy_lock = VPB_UNBUSIED; 2075 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 2076 busy_lock = VPB_SINGLE_EXCLUSIVER; 2077 if ((req & VM_ALLOC_SBUSY) != 0) 2078 busy_lock = VPB_SHARERS_WORD(1); 2079 if ((req & VM_ALLOC_WIRED) != 0) 2080 vm_wire_add(npages); 2081 if (object != NULL) { 2082 if (object->memattr != VM_MEMATTR_DEFAULT && 2083 memattr == VM_MEMATTR_DEFAULT) 2084 memattr = object->memattr; 2085 } 2086 for (m = m_ret; m < &m_ret[npages]; m++) { 2087 m->aflags = 0; 2088 m->flags = (m->flags | PG_NODUMP) & flags; 2089 m->busy_lock = busy_lock; 2090 if ((req & VM_ALLOC_WIRED) != 0) 2091 m->wire_count = 1; 2092 m->act_count = 0; 2093 m->oflags = oflags; 2094 if (object != NULL) { 2095 if (vm_page_insert_after(m, object, pindex, mpred)) { 2096 if ((req & VM_ALLOC_WIRED) != 0) 2097 vm_wire_sub(npages); 2098 KASSERT(m->object == NULL, 2099 ("page %p has object", m)); 2100 mpred = m; 2101 for (m = m_ret; m < &m_ret[npages]; m++) { 2102 if (m <= mpred && 2103 (req & VM_ALLOC_WIRED) != 0) 2104 m->wire_count = 0; 2105 m->oflags = VPO_UNMANAGED; 2106 m->busy_lock = VPB_UNBUSIED; 2107 /* Don't change PG_ZERO. */ 2108 vm_page_free_toq(m); 2109 } 2110 if (req & VM_ALLOC_WAITFAIL) { 2111 VM_OBJECT_WUNLOCK(object); 2112 vm_radix_wait(); 2113 VM_OBJECT_WLOCK(object); 2114 } 2115 return (NULL); 2116 } 2117 mpred = m; 2118 } else 2119 m->pindex = pindex; 2120 if (memattr != VM_MEMATTR_DEFAULT) 2121 pmap_page_set_memattr(m, memattr); 2122 pindex++; 2123 } 2124 return (m_ret); 2125 } 2126 2127 /* 2128 * Check a page that has been freshly dequeued from a freelist. 2129 */ 2130 static void 2131 vm_page_alloc_check(vm_page_t m) 2132 { 2133 2134 KASSERT(m->object == NULL, ("page %p has object", m)); 2135 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 2136 ("page %p has unexpected queue %d, flags %#x", 2137 m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK))); 2138 KASSERT(!vm_page_held(m), ("page %p is held", m)); 2139 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 2140 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2141 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2142 ("page %p has unexpected memattr %d", 2143 m, pmap_page_get_memattr(m))); 2144 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2145 } 2146 2147 /* 2148 * vm_page_alloc_freelist: 2149 * 2150 * Allocate a physical page from the specified free page list. 2151 * 2152 * The caller must always specify an allocation class. 2153 * 2154 * allocation classes: 2155 * VM_ALLOC_NORMAL normal process request 2156 * VM_ALLOC_SYSTEM system *really* needs a page 2157 * VM_ALLOC_INTERRUPT interrupt time request 2158 * 2159 * optional allocation flags: 2160 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2161 * intends to allocate 2162 * VM_ALLOC_WIRED wire the allocated page 2163 * VM_ALLOC_ZERO prefer a zeroed page 2164 */ 2165 vm_page_t 2166 vm_page_alloc_freelist(int freelist, int req) 2167 { 2168 struct vm_domainset_iter di; 2169 vm_page_t m; 2170 int domain; 2171 2172 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2173 do { 2174 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2175 if (m != NULL) 2176 break; 2177 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2178 2179 return (m); 2180 } 2181 2182 vm_page_t 2183 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2184 { 2185 struct vm_domain *vmd; 2186 vm_page_t m; 2187 u_int flags; 2188 2189 m = NULL; 2190 vmd = VM_DOMAIN(domain); 2191 again: 2192 if (vm_domain_allocate(vmd, req, 1)) { 2193 vm_domain_free_lock(vmd); 2194 m = vm_phys_alloc_freelist_pages(domain, freelist, 2195 VM_FREEPOOL_DIRECT, 0); 2196 vm_domain_free_unlock(vmd); 2197 if (m == NULL) 2198 vm_domain_freecnt_inc(vmd, 1); 2199 } 2200 if (m == NULL) { 2201 if (vm_domain_alloc_fail(vmd, NULL, req)) 2202 goto again; 2203 return (NULL); 2204 } 2205 vm_page_dequeue(m); 2206 vm_page_alloc_check(m); 2207 2208 /* 2209 * Initialize the page. Only the PG_ZERO flag is inherited. 2210 */ 2211 m->aflags = 0; 2212 flags = 0; 2213 if ((req & VM_ALLOC_ZERO) != 0) 2214 flags = PG_ZERO; 2215 m->flags &= flags; 2216 if ((req & VM_ALLOC_WIRED) != 0) { 2217 /* 2218 * The page lock is not required for wiring a page that does 2219 * not belong to an object. 2220 */ 2221 vm_wire_add(1); 2222 m->wire_count = 1; 2223 } 2224 /* Unmanaged pages don't use "act_count". */ 2225 m->oflags = VPO_UNMANAGED; 2226 return (m); 2227 } 2228 2229 static int 2230 vm_page_import(void *arg, void **store, int cnt, int domain, int flags) 2231 { 2232 struct vm_domain *vmd; 2233 int i; 2234 2235 vmd = arg; 2236 /* Only import if we can bring in a full bucket. */ 2237 if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) 2238 return (0); 2239 domain = vmd->vmd_domain; 2240 vm_domain_free_lock(vmd); 2241 i = vm_phys_alloc_npages(domain, VM_FREEPOOL_DEFAULT, cnt, 2242 (vm_page_t *)store); 2243 vm_domain_free_unlock(vmd); 2244 if (cnt != i) 2245 vm_domain_freecnt_inc(vmd, cnt - i); 2246 2247 return (i); 2248 } 2249 2250 static void 2251 vm_page_release(void *arg, void **store, int cnt) 2252 { 2253 struct vm_domain *vmd; 2254 vm_page_t m; 2255 int i; 2256 2257 vmd = arg; 2258 vm_domain_free_lock(vmd); 2259 for (i = 0; i < cnt; i++) { 2260 m = (vm_page_t)store[i]; 2261 vm_phys_free_pages(m, 0); 2262 } 2263 vm_domain_free_unlock(vmd); 2264 vm_domain_freecnt_inc(vmd, cnt); 2265 } 2266 2267 #define VPSC_ANY 0 /* No restrictions. */ 2268 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2269 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2270 2271 /* 2272 * vm_page_scan_contig: 2273 * 2274 * Scan vm_page_array[] between the specified entries "m_start" and 2275 * "m_end" for a run of contiguous physical pages that satisfy the 2276 * specified conditions, and return the lowest page in the run. The 2277 * specified "alignment" determines the alignment of the lowest physical 2278 * page in the run. If the specified "boundary" is non-zero, then the 2279 * run of physical pages cannot span a physical address that is a 2280 * multiple of "boundary". 2281 * 2282 * "m_end" is never dereferenced, so it need not point to a vm_page 2283 * structure within vm_page_array[]. 2284 * 2285 * "npages" must be greater than zero. "m_start" and "m_end" must not 2286 * span a hole (or discontiguity) in the physical address space. Both 2287 * "alignment" and "boundary" must be a power of two. 2288 */ 2289 vm_page_t 2290 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2291 u_long alignment, vm_paddr_t boundary, int options) 2292 { 2293 struct mtx *m_mtx; 2294 vm_object_t object; 2295 vm_paddr_t pa; 2296 vm_page_t m, m_run; 2297 #if VM_NRESERVLEVEL > 0 2298 int level; 2299 #endif 2300 int m_inc, order, run_ext, run_len; 2301 2302 KASSERT(npages > 0, ("npages is 0")); 2303 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2304 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2305 m_run = NULL; 2306 run_len = 0; 2307 m_mtx = NULL; 2308 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2309 KASSERT((m->flags & PG_MARKER) == 0, 2310 ("page %p is PG_MARKER", m)); 2311 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1, 2312 ("fictitious page %p has invalid wire count", m)); 2313 2314 /* 2315 * If the current page would be the start of a run, check its 2316 * physical address against the end, alignment, and boundary 2317 * conditions. If it doesn't satisfy these conditions, either 2318 * terminate the scan or advance to the next page that 2319 * satisfies the failed condition. 2320 */ 2321 if (run_len == 0) { 2322 KASSERT(m_run == NULL, ("m_run != NULL")); 2323 if (m + npages > m_end) 2324 break; 2325 pa = VM_PAGE_TO_PHYS(m); 2326 if ((pa & (alignment - 1)) != 0) { 2327 m_inc = atop(roundup2(pa, alignment) - pa); 2328 continue; 2329 } 2330 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2331 boundary) != 0) { 2332 m_inc = atop(roundup2(pa, boundary) - pa); 2333 continue; 2334 } 2335 } else 2336 KASSERT(m_run != NULL, ("m_run == NULL")); 2337 2338 vm_page_change_lock(m, &m_mtx); 2339 m_inc = 1; 2340 retry: 2341 if (vm_page_held(m)) 2342 run_ext = 0; 2343 #if VM_NRESERVLEVEL > 0 2344 else if ((level = vm_reserv_level(m)) >= 0 && 2345 (options & VPSC_NORESERV) != 0) { 2346 run_ext = 0; 2347 /* Advance to the end of the reservation. */ 2348 pa = VM_PAGE_TO_PHYS(m); 2349 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2350 pa); 2351 } 2352 #endif 2353 else if ((object = m->object) != NULL) { 2354 /* 2355 * The page is considered eligible for relocation if 2356 * and only if it could be laundered or reclaimed by 2357 * the page daemon. 2358 */ 2359 if (!VM_OBJECT_TRYRLOCK(object)) { 2360 mtx_unlock(m_mtx); 2361 VM_OBJECT_RLOCK(object); 2362 mtx_lock(m_mtx); 2363 if (m->object != object) { 2364 /* 2365 * The page may have been freed. 2366 */ 2367 VM_OBJECT_RUNLOCK(object); 2368 goto retry; 2369 } else if (vm_page_held(m)) { 2370 run_ext = 0; 2371 goto unlock; 2372 } 2373 } 2374 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2375 ("page %p is PG_UNHOLDFREE", m)); 2376 /* Don't care: PG_NODUMP, PG_ZERO. */ 2377 if (object->type != OBJT_DEFAULT && 2378 object->type != OBJT_SWAP && 2379 object->type != OBJT_VNODE) { 2380 run_ext = 0; 2381 #if VM_NRESERVLEVEL > 0 2382 } else if ((options & VPSC_NOSUPER) != 0 && 2383 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2384 run_ext = 0; 2385 /* Advance to the end of the superpage. */ 2386 pa = VM_PAGE_TO_PHYS(m); 2387 m_inc = atop(roundup2(pa + 1, 2388 vm_reserv_size(level)) - pa); 2389 #endif 2390 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2391 vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { 2392 /* 2393 * The page is allocated but eligible for 2394 * relocation. Extend the current run by one 2395 * page. 2396 */ 2397 KASSERT(pmap_page_get_memattr(m) == 2398 VM_MEMATTR_DEFAULT, 2399 ("page %p has an unexpected memattr", m)); 2400 KASSERT((m->oflags & (VPO_SWAPINPROG | 2401 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2402 ("page %p has unexpected oflags", m)); 2403 /* Don't care: VPO_NOSYNC. */ 2404 run_ext = 1; 2405 } else 2406 run_ext = 0; 2407 unlock: 2408 VM_OBJECT_RUNLOCK(object); 2409 #if VM_NRESERVLEVEL > 0 2410 } else if (level >= 0) { 2411 /* 2412 * The page is reserved but not yet allocated. In 2413 * other words, it is still free. Extend the current 2414 * run by one page. 2415 */ 2416 run_ext = 1; 2417 #endif 2418 } else if ((order = m->order) < VM_NFREEORDER) { 2419 /* 2420 * The page is enqueued in the physical memory 2421 * allocator's free page queues. Moreover, it is the 2422 * first page in a power-of-two-sized run of 2423 * contiguous free pages. Add these pages to the end 2424 * of the current run, and jump ahead. 2425 */ 2426 run_ext = 1 << order; 2427 m_inc = 1 << order; 2428 } else { 2429 /* 2430 * Skip the page for one of the following reasons: (1) 2431 * It is enqueued in the physical memory allocator's 2432 * free page queues. However, it is not the first 2433 * page in a run of contiguous free pages. (This case 2434 * rarely occurs because the scan is performed in 2435 * ascending order.) (2) It is not reserved, and it is 2436 * transitioning from free to allocated. (Conversely, 2437 * the transition from allocated to free for managed 2438 * pages is blocked by the page lock.) (3) It is 2439 * allocated but not contained by an object and not 2440 * wired, e.g., allocated by Xen's balloon driver. 2441 */ 2442 run_ext = 0; 2443 } 2444 2445 /* 2446 * Extend or reset the current run of pages. 2447 */ 2448 if (run_ext > 0) { 2449 if (run_len == 0) 2450 m_run = m; 2451 run_len += run_ext; 2452 } else { 2453 if (run_len > 0) { 2454 m_run = NULL; 2455 run_len = 0; 2456 } 2457 } 2458 } 2459 if (m_mtx != NULL) 2460 mtx_unlock(m_mtx); 2461 if (run_len >= npages) 2462 return (m_run); 2463 return (NULL); 2464 } 2465 2466 /* 2467 * vm_page_reclaim_run: 2468 * 2469 * Try to relocate each of the allocated virtual pages within the 2470 * specified run of physical pages to a new physical address. Free the 2471 * physical pages underlying the relocated virtual pages. A virtual page 2472 * is relocatable if and only if it could be laundered or reclaimed by 2473 * the page daemon. Whenever possible, a virtual page is relocated to a 2474 * physical address above "high". 2475 * 2476 * Returns 0 if every physical page within the run was already free or 2477 * just freed by a successful relocation. Otherwise, returns a non-zero 2478 * value indicating why the last attempt to relocate a virtual page was 2479 * unsuccessful. 2480 * 2481 * "req_class" must be an allocation class. 2482 */ 2483 static int 2484 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2485 vm_paddr_t high) 2486 { 2487 struct vm_domain *vmd; 2488 struct mtx *m_mtx; 2489 struct spglist free; 2490 vm_object_t object; 2491 vm_paddr_t pa; 2492 vm_page_t m, m_end, m_new; 2493 int error, order, req; 2494 2495 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2496 ("req_class is not an allocation class")); 2497 SLIST_INIT(&free); 2498 error = 0; 2499 m = m_run; 2500 m_end = m_run + npages; 2501 m_mtx = NULL; 2502 for (; error == 0 && m < m_end; m++) { 2503 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2504 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2505 2506 /* 2507 * Avoid releasing and reacquiring the same page lock. 2508 */ 2509 vm_page_change_lock(m, &m_mtx); 2510 retry: 2511 if (vm_page_held(m)) 2512 error = EBUSY; 2513 else if ((object = m->object) != NULL) { 2514 /* 2515 * The page is relocated if and only if it could be 2516 * laundered or reclaimed by the page daemon. 2517 */ 2518 if (!VM_OBJECT_TRYWLOCK(object)) { 2519 mtx_unlock(m_mtx); 2520 VM_OBJECT_WLOCK(object); 2521 mtx_lock(m_mtx); 2522 if (m->object != object) { 2523 /* 2524 * The page may have been freed. 2525 */ 2526 VM_OBJECT_WUNLOCK(object); 2527 goto retry; 2528 } else if (vm_page_held(m)) { 2529 error = EBUSY; 2530 goto unlock; 2531 } 2532 } 2533 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2534 ("page %p is PG_UNHOLDFREE", m)); 2535 /* Don't care: PG_NODUMP, PG_ZERO. */ 2536 if (object->type != OBJT_DEFAULT && 2537 object->type != OBJT_SWAP && 2538 object->type != OBJT_VNODE) 2539 error = EINVAL; 2540 else if (object->memattr != VM_MEMATTR_DEFAULT) 2541 error = EINVAL; 2542 else if (vm_page_queue(m) != PQ_NONE && 2543 !vm_page_busied(m)) { 2544 KASSERT(pmap_page_get_memattr(m) == 2545 VM_MEMATTR_DEFAULT, 2546 ("page %p has an unexpected memattr", m)); 2547 KASSERT((m->oflags & (VPO_SWAPINPROG | 2548 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2549 ("page %p has unexpected oflags", m)); 2550 /* Don't care: VPO_NOSYNC. */ 2551 if (m->valid != 0) { 2552 /* 2553 * First, try to allocate a new page 2554 * that is above "high". Failing 2555 * that, try to allocate a new page 2556 * that is below "m_run". Allocate 2557 * the new page between the end of 2558 * "m_run" and "high" only as a last 2559 * resort. 2560 */ 2561 req = req_class | VM_ALLOC_NOOBJ; 2562 if ((m->flags & PG_NODUMP) != 0) 2563 req |= VM_ALLOC_NODUMP; 2564 if (trunc_page(high) != 2565 ~(vm_paddr_t)PAGE_MASK) { 2566 m_new = vm_page_alloc_contig( 2567 NULL, 0, req, 1, 2568 round_page(high), 2569 ~(vm_paddr_t)0, 2570 PAGE_SIZE, 0, 2571 VM_MEMATTR_DEFAULT); 2572 } else 2573 m_new = NULL; 2574 if (m_new == NULL) { 2575 pa = VM_PAGE_TO_PHYS(m_run); 2576 m_new = vm_page_alloc_contig( 2577 NULL, 0, req, 1, 2578 0, pa - 1, PAGE_SIZE, 0, 2579 VM_MEMATTR_DEFAULT); 2580 } 2581 if (m_new == NULL) { 2582 pa += ptoa(npages); 2583 m_new = vm_page_alloc_contig( 2584 NULL, 0, req, 1, 2585 pa, high, PAGE_SIZE, 0, 2586 VM_MEMATTR_DEFAULT); 2587 } 2588 if (m_new == NULL) { 2589 error = ENOMEM; 2590 goto unlock; 2591 } 2592 KASSERT(!vm_page_wired(m_new), 2593 ("page %p is wired", m_new)); 2594 2595 /* 2596 * Replace "m" with the new page. For 2597 * vm_page_replace(), "m" must be busy 2598 * and dequeued. Finally, change "m" 2599 * as if vm_page_free() was called. 2600 */ 2601 if (object->ref_count != 0) 2602 pmap_remove_all(m); 2603 m_new->aflags = m->aflags & 2604 ~PGA_QUEUE_STATE_MASK; 2605 KASSERT(m_new->oflags == VPO_UNMANAGED, 2606 ("page %p is managed", m_new)); 2607 m_new->oflags = m->oflags & VPO_NOSYNC; 2608 pmap_copy_page(m, m_new); 2609 m_new->valid = m->valid; 2610 m_new->dirty = m->dirty; 2611 m->flags &= ~PG_ZERO; 2612 vm_page_xbusy(m); 2613 vm_page_dequeue(m); 2614 vm_page_replace_checked(m_new, object, 2615 m->pindex, m); 2616 if (vm_page_free_prep(m)) 2617 SLIST_INSERT_HEAD(&free, m, 2618 plinks.s.ss); 2619 2620 /* 2621 * The new page must be deactivated 2622 * before the object is unlocked. 2623 */ 2624 vm_page_change_lock(m_new, &m_mtx); 2625 vm_page_deactivate(m_new); 2626 } else { 2627 m->flags &= ~PG_ZERO; 2628 vm_page_dequeue(m); 2629 if (vm_page_free_prep(m)) 2630 SLIST_INSERT_HEAD(&free, m, 2631 plinks.s.ss); 2632 KASSERT(m->dirty == 0, 2633 ("page %p is dirty", m)); 2634 } 2635 } else 2636 error = EBUSY; 2637 unlock: 2638 VM_OBJECT_WUNLOCK(object); 2639 } else { 2640 MPASS(vm_phys_domain(m) == domain); 2641 vmd = VM_DOMAIN(domain); 2642 vm_domain_free_lock(vmd); 2643 order = m->order; 2644 if (order < VM_NFREEORDER) { 2645 /* 2646 * The page is enqueued in the physical memory 2647 * allocator's free page queues. Moreover, it 2648 * is the first page in a power-of-two-sized 2649 * run of contiguous free pages. Jump ahead 2650 * to the last page within that run, and 2651 * continue from there. 2652 */ 2653 m += (1 << order) - 1; 2654 } 2655 #if VM_NRESERVLEVEL > 0 2656 else if (vm_reserv_is_page_free(m)) 2657 order = 0; 2658 #endif 2659 vm_domain_free_unlock(vmd); 2660 if (order == VM_NFREEORDER) 2661 error = EINVAL; 2662 } 2663 } 2664 if (m_mtx != NULL) 2665 mtx_unlock(m_mtx); 2666 if ((m = SLIST_FIRST(&free)) != NULL) { 2667 int cnt; 2668 2669 vmd = VM_DOMAIN(domain); 2670 cnt = 0; 2671 vm_domain_free_lock(vmd); 2672 do { 2673 MPASS(vm_phys_domain(m) == domain); 2674 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2675 vm_phys_free_pages(m, 0); 2676 cnt++; 2677 } while ((m = SLIST_FIRST(&free)) != NULL); 2678 vm_domain_free_unlock(vmd); 2679 vm_domain_freecnt_inc(vmd, cnt); 2680 } 2681 return (error); 2682 } 2683 2684 #define NRUNS 16 2685 2686 CTASSERT(powerof2(NRUNS)); 2687 2688 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2689 2690 #define MIN_RECLAIM 8 2691 2692 /* 2693 * vm_page_reclaim_contig: 2694 * 2695 * Reclaim allocated, contiguous physical memory satisfying the specified 2696 * conditions by relocating the virtual pages using that physical memory. 2697 * Returns true if reclamation is successful and false otherwise. Since 2698 * relocation requires the allocation of physical pages, reclamation may 2699 * fail due to a shortage of free pages. When reclamation fails, callers 2700 * are expected to perform vm_wait() before retrying a failed allocation 2701 * operation, e.g., vm_page_alloc_contig(). 2702 * 2703 * The caller must always specify an allocation class through "req". 2704 * 2705 * allocation classes: 2706 * VM_ALLOC_NORMAL normal process request 2707 * VM_ALLOC_SYSTEM system *really* needs a page 2708 * VM_ALLOC_INTERRUPT interrupt time request 2709 * 2710 * The optional allocation flags are ignored. 2711 * 2712 * "npages" must be greater than zero. Both "alignment" and "boundary" 2713 * must be a power of two. 2714 */ 2715 bool 2716 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2717 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2718 { 2719 struct vm_domain *vmd; 2720 vm_paddr_t curr_low; 2721 vm_page_t m_run, m_runs[NRUNS]; 2722 u_long count, reclaimed; 2723 int error, i, options, req_class; 2724 2725 KASSERT(npages > 0, ("npages is 0")); 2726 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2727 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2728 req_class = req & VM_ALLOC_CLASS_MASK; 2729 2730 /* 2731 * The page daemon is allowed to dig deeper into the free page list. 2732 */ 2733 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2734 req_class = VM_ALLOC_SYSTEM; 2735 2736 /* 2737 * Return if the number of free pages cannot satisfy the requested 2738 * allocation. 2739 */ 2740 vmd = VM_DOMAIN(domain); 2741 count = vmd->vmd_free_count; 2742 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2743 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2744 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2745 return (false); 2746 2747 /* 2748 * Scan up to three times, relaxing the restrictions ("options") on 2749 * the reclamation of reservations and superpages each time. 2750 */ 2751 for (options = VPSC_NORESERV;;) { 2752 /* 2753 * Find the highest runs that satisfy the given constraints 2754 * and restrictions, and record them in "m_runs". 2755 */ 2756 curr_low = low; 2757 count = 0; 2758 for (;;) { 2759 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2760 high, alignment, boundary, options); 2761 if (m_run == NULL) 2762 break; 2763 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2764 m_runs[RUN_INDEX(count)] = m_run; 2765 count++; 2766 } 2767 2768 /* 2769 * Reclaim the highest runs in LIFO (descending) order until 2770 * the number of reclaimed pages, "reclaimed", is at least 2771 * MIN_RECLAIM. Reset "reclaimed" each time because each 2772 * reclamation is idempotent, and runs will (likely) recur 2773 * from one scan to the next as restrictions are relaxed. 2774 */ 2775 reclaimed = 0; 2776 for (i = 0; count > 0 && i < NRUNS; i++) { 2777 count--; 2778 m_run = m_runs[RUN_INDEX(count)]; 2779 error = vm_page_reclaim_run(req_class, domain, npages, 2780 m_run, high); 2781 if (error == 0) { 2782 reclaimed += npages; 2783 if (reclaimed >= MIN_RECLAIM) 2784 return (true); 2785 } 2786 } 2787 2788 /* 2789 * Either relax the restrictions on the next scan or return if 2790 * the last scan had no restrictions. 2791 */ 2792 if (options == VPSC_NORESERV) 2793 options = VPSC_NOSUPER; 2794 else if (options == VPSC_NOSUPER) 2795 options = VPSC_ANY; 2796 else if (options == VPSC_ANY) 2797 return (reclaimed != 0); 2798 } 2799 } 2800 2801 bool 2802 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2803 u_long alignment, vm_paddr_t boundary) 2804 { 2805 struct vm_domainset_iter di; 2806 int domain; 2807 bool ret; 2808 2809 vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); 2810 do { 2811 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 2812 high, alignment, boundary); 2813 if (ret) 2814 break; 2815 } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); 2816 2817 return (ret); 2818 } 2819 2820 /* 2821 * Set the domain in the appropriate page level domainset. 2822 */ 2823 void 2824 vm_domain_set(struct vm_domain *vmd) 2825 { 2826 2827 mtx_lock(&vm_domainset_lock); 2828 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 2829 vmd->vmd_minset = 1; 2830 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 2831 } 2832 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 2833 vmd->vmd_severeset = 1; 2834 DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); 2835 } 2836 mtx_unlock(&vm_domainset_lock); 2837 } 2838 2839 /* 2840 * Clear the domain from the appropriate page level domainset. 2841 */ 2842 void 2843 vm_domain_clear(struct vm_domain *vmd) 2844 { 2845 2846 mtx_lock(&vm_domainset_lock); 2847 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 2848 vmd->vmd_minset = 0; 2849 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 2850 if (vm_min_waiters != 0) { 2851 vm_min_waiters = 0; 2852 wakeup(&vm_min_domains); 2853 } 2854 } 2855 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 2856 vmd->vmd_severeset = 0; 2857 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 2858 if (vm_severe_waiters != 0) { 2859 vm_severe_waiters = 0; 2860 wakeup(&vm_severe_domains); 2861 } 2862 } 2863 2864 /* 2865 * If pageout daemon needs pages, then tell it that there are 2866 * some free. 2867 */ 2868 if (vmd->vmd_pageout_pages_needed && 2869 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 2870 wakeup(&vmd->vmd_pageout_pages_needed); 2871 vmd->vmd_pageout_pages_needed = 0; 2872 } 2873 2874 /* See comments in vm_wait_doms(). */ 2875 if (vm_pageproc_waiters) { 2876 vm_pageproc_waiters = 0; 2877 wakeup(&vm_pageproc_waiters); 2878 } 2879 mtx_unlock(&vm_domainset_lock); 2880 } 2881 2882 /* 2883 * Wait for free pages to exceed the min threshold globally. 2884 */ 2885 void 2886 vm_wait_min(void) 2887 { 2888 2889 mtx_lock(&vm_domainset_lock); 2890 while (vm_page_count_min()) { 2891 vm_min_waiters++; 2892 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 2893 } 2894 mtx_unlock(&vm_domainset_lock); 2895 } 2896 2897 /* 2898 * Wait for free pages to exceed the severe threshold globally. 2899 */ 2900 void 2901 vm_wait_severe(void) 2902 { 2903 2904 mtx_lock(&vm_domainset_lock); 2905 while (vm_page_count_severe()) { 2906 vm_severe_waiters++; 2907 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 2908 "vmwait", 0); 2909 } 2910 mtx_unlock(&vm_domainset_lock); 2911 } 2912 2913 u_int 2914 vm_wait_count(void) 2915 { 2916 2917 return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); 2918 } 2919 2920 void 2921 vm_wait_doms(const domainset_t *wdoms) 2922 { 2923 2924 /* 2925 * We use racey wakeup synchronization to avoid expensive global 2926 * locking for the pageproc when sleeping with a non-specific vm_wait. 2927 * To handle this, we only sleep for one tick in this instance. It 2928 * is expected that most allocations for the pageproc will come from 2929 * kmem or vm_page_grab* which will use the more specific and 2930 * race-free vm_wait_domain(). 2931 */ 2932 if (curproc == pageproc) { 2933 mtx_lock(&vm_domainset_lock); 2934 vm_pageproc_waiters++; 2935 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, 2936 "pageprocwait", 1); 2937 } else { 2938 /* 2939 * XXX Ideally we would wait only until the allocation could 2940 * be satisfied. This condition can cause new allocators to 2941 * consume all freed pages while old allocators wait. 2942 */ 2943 mtx_lock(&vm_domainset_lock); 2944 if (vm_page_count_min_set(wdoms)) { 2945 vm_min_waiters++; 2946 msleep(&vm_min_domains, &vm_domainset_lock, 2947 PVM | PDROP, "vmwait", 0); 2948 } else 2949 mtx_unlock(&vm_domainset_lock); 2950 } 2951 } 2952 2953 /* 2954 * vm_wait_domain: 2955 * 2956 * Sleep until free pages are available for allocation. 2957 * - Called in various places after failed memory allocations. 2958 */ 2959 void 2960 vm_wait_domain(int domain) 2961 { 2962 struct vm_domain *vmd; 2963 domainset_t wdom; 2964 2965 vmd = VM_DOMAIN(domain); 2966 vm_domain_free_assert_unlocked(vmd); 2967 2968 if (curproc == pageproc) { 2969 mtx_lock(&vm_domainset_lock); 2970 if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { 2971 vmd->vmd_pageout_pages_needed = 1; 2972 msleep(&vmd->vmd_pageout_pages_needed, 2973 &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); 2974 } else 2975 mtx_unlock(&vm_domainset_lock); 2976 } else { 2977 if (pageproc == NULL) 2978 panic("vm_wait in early boot"); 2979 DOMAINSET_ZERO(&wdom); 2980 DOMAINSET_SET(vmd->vmd_domain, &wdom); 2981 vm_wait_doms(&wdom); 2982 } 2983 } 2984 2985 /* 2986 * vm_wait: 2987 * 2988 * Sleep until free pages are available for allocation in the 2989 * affinity domains of the obj. If obj is NULL, the domain set 2990 * for the calling thread is used. 2991 * Called in various places after failed memory allocations. 2992 */ 2993 void 2994 vm_wait(vm_object_t obj) 2995 { 2996 struct domainset *d; 2997 2998 d = NULL; 2999 3000 /* 3001 * Carefully fetch pointers only once: the struct domainset 3002 * itself is ummutable but the pointer might change. 3003 */ 3004 if (obj != NULL) 3005 d = obj->domain.dr_policy; 3006 if (d == NULL) 3007 d = curthread->td_domain.dr_policy; 3008 3009 vm_wait_doms(&d->ds_mask); 3010 } 3011 3012 /* 3013 * vm_domain_alloc_fail: 3014 * 3015 * Called when a page allocation function fails. Informs the 3016 * pagedaemon and performs the requested wait. Requires the 3017 * domain_free and object lock on entry. Returns with the 3018 * object lock held and free lock released. Returns an error when 3019 * retry is necessary. 3020 * 3021 */ 3022 static int 3023 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 3024 { 3025 3026 vm_domain_free_assert_unlocked(vmd); 3027 3028 atomic_add_int(&vmd->vmd_pageout_deficit, 3029 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 3030 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 3031 if (object != NULL) 3032 VM_OBJECT_WUNLOCK(object); 3033 vm_wait_domain(vmd->vmd_domain); 3034 if (object != NULL) 3035 VM_OBJECT_WLOCK(object); 3036 if (req & VM_ALLOC_WAITOK) 3037 return (EAGAIN); 3038 } 3039 3040 return (0); 3041 } 3042 3043 /* 3044 * vm_waitpfault: 3045 * 3046 * Sleep until free pages are available for allocation. 3047 * - Called only in vm_fault so that processes page faulting 3048 * can be easily tracked. 3049 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 3050 * processes will be able to grab memory first. Do not change 3051 * this balance without careful testing first. 3052 */ 3053 void 3054 vm_waitpfault(struct domainset *dset) 3055 { 3056 3057 /* 3058 * XXX Ideally we would wait only until the allocation could 3059 * be satisfied. This condition can cause new allocators to 3060 * consume all freed pages while old allocators wait. 3061 */ 3062 mtx_lock(&vm_domainset_lock); 3063 if (vm_page_count_min_set(&dset->ds_mask)) { 3064 vm_min_waiters++; 3065 msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, 3066 "pfault", 0); 3067 } else 3068 mtx_unlock(&vm_domainset_lock); 3069 } 3070 3071 struct vm_pagequeue * 3072 vm_page_pagequeue(vm_page_t m) 3073 { 3074 3075 return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]); 3076 } 3077 3078 static struct mtx * 3079 vm_page_pagequeue_lockptr(vm_page_t m) 3080 { 3081 uint8_t queue; 3082 3083 if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) 3084 return (NULL); 3085 return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex); 3086 } 3087 3088 static inline void 3089 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m) 3090 { 3091 struct vm_domain *vmd; 3092 uint8_t qflags; 3093 3094 CRITICAL_ASSERT(curthread); 3095 vm_pagequeue_assert_locked(pq); 3096 3097 /* 3098 * The page daemon is allowed to set m->queue = PQ_NONE without 3099 * the page queue lock held. In this case it is about to free the page, 3100 * which must not have any queue state. 3101 */ 3102 qflags = atomic_load_8(&m->aflags) & PGA_QUEUE_STATE_MASK; 3103 KASSERT(pq == vm_page_pagequeue(m) || qflags == 0, 3104 ("page %p doesn't belong to queue %p but has queue state %#x", 3105 m, pq, qflags)); 3106 3107 if ((qflags & PGA_DEQUEUE) != 0) { 3108 if (__predict_true((qflags & PGA_ENQUEUED) != 0)) { 3109 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3110 vm_pagequeue_cnt_dec(pq); 3111 } 3112 vm_page_dequeue_complete(m); 3113 } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) { 3114 if ((qflags & PGA_ENQUEUED) != 0) 3115 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3116 else { 3117 vm_pagequeue_cnt_inc(pq); 3118 vm_page_aflag_set(m, PGA_ENQUEUED); 3119 } 3120 if ((qflags & PGA_REQUEUE_HEAD) != 0) { 3121 KASSERT(m->queue == PQ_INACTIVE, 3122 ("head enqueue not supported for page %p", m)); 3123 vmd = vm_pagequeue_domain(m); 3124 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); 3125 } else 3126 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3127 3128 /* 3129 * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after 3130 * setting PGA_ENQUEUED in order to synchronize with the 3131 * page daemon. 3132 */ 3133 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); 3134 } 3135 } 3136 3137 static void 3138 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, 3139 uint8_t queue) 3140 { 3141 vm_page_t m; 3142 int i; 3143 3144 for (i = 0; i < bq->bq_cnt; i++) { 3145 m = bq->bq_pa[i]; 3146 if (__predict_false(m->queue != queue)) 3147 continue; 3148 vm_pqbatch_process_page(pq, m); 3149 } 3150 vm_batchqueue_init(bq); 3151 } 3152 3153 static void 3154 vm_pqbatch_submit_page(vm_page_t m, uint8_t queue) 3155 { 3156 struct vm_batchqueue *bq; 3157 struct vm_pagequeue *pq; 3158 int domain; 3159 3160 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3161 ("page %p is unmanaged", m)); 3162 KASSERT(mtx_owned(vm_page_lockptr(m)) || 3163 (m->object == NULL && (m->aflags & PGA_DEQUEUE) != 0), 3164 ("missing synchronization for page %p", m)); 3165 KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); 3166 3167 domain = vm_phys_domain(m); 3168 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 3169 3170 critical_enter(); 3171 bq = DPCPU_PTR(pqbatch[domain][queue]); 3172 if (vm_batchqueue_insert(bq, m)) { 3173 critical_exit(); 3174 return; 3175 } 3176 if (!vm_pagequeue_trylock(pq)) { 3177 critical_exit(); 3178 vm_pagequeue_lock(pq); 3179 critical_enter(); 3180 bq = DPCPU_PTR(pqbatch[domain][queue]); 3181 } 3182 vm_pqbatch_process(pq, bq, queue); 3183 3184 /* 3185 * The page may have been logically dequeued before we acquired the 3186 * page queue lock. In this case, since we either hold the page lock 3187 * or the page is being freed, a different thread cannot be concurrently 3188 * enqueuing the page. 3189 */ 3190 if (__predict_true(m->queue == queue)) 3191 vm_pqbatch_process_page(pq, m); 3192 else { 3193 KASSERT(m->queue == PQ_NONE, 3194 ("invalid queue transition for page %p", m)); 3195 KASSERT((m->aflags & PGA_ENQUEUED) == 0, 3196 ("page %p is enqueued with invalid queue index", m)); 3197 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); 3198 } 3199 vm_pagequeue_unlock(pq); 3200 critical_exit(); 3201 } 3202 3203 /* 3204 * vm_page_drain_pqbatch: [ internal use only ] 3205 * 3206 * Force all per-CPU page queue batch queues to be drained. This is 3207 * intended for use in severe memory shortages, to ensure that pages 3208 * do not remain stuck in the batch queues. 3209 */ 3210 void 3211 vm_page_drain_pqbatch(void) 3212 { 3213 struct thread *td; 3214 struct vm_domain *vmd; 3215 struct vm_pagequeue *pq; 3216 int cpu, domain, queue; 3217 3218 td = curthread; 3219 CPU_FOREACH(cpu) { 3220 thread_lock(td); 3221 sched_bind(td, cpu); 3222 thread_unlock(td); 3223 3224 for (domain = 0; domain < vm_ndomains; domain++) { 3225 vmd = VM_DOMAIN(domain); 3226 for (queue = 0; queue < PQ_COUNT; queue++) { 3227 pq = &vmd->vmd_pagequeues[queue]; 3228 vm_pagequeue_lock(pq); 3229 critical_enter(); 3230 vm_pqbatch_process(pq, 3231 DPCPU_PTR(pqbatch[domain][queue]), queue); 3232 critical_exit(); 3233 vm_pagequeue_unlock(pq); 3234 } 3235 } 3236 } 3237 thread_lock(td); 3238 sched_unbind(td); 3239 thread_unlock(td); 3240 } 3241 3242 /* 3243 * Complete the logical removal of a page from a page queue. We must be 3244 * careful to synchronize with the page daemon, which may be concurrently 3245 * examining the page with only the page lock held. The page must not be 3246 * in a state where it appears to be logically enqueued. 3247 */ 3248 static void 3249 vm_page_dequeue_complete(vm_page_t m) 3250 { 3251 3252 m->queue = PQ_NONE; 3253 atomic_thread_fence_rel(); 3254 vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); 3255 } 3256 3257 /* 3258 * vm_page_dequeue_deferred: [ internal use only ] 3259 * 3260 * Request removal of the given page from its current page 3261 * queue. Physical removal from the queue may be deferred 3262 * indefinitely. 3263 * 3264 * The page must be locked. 3265 */ 3266 void 3267 vm_page_dequeue_deferred(vm_page_t m) 3268 { 3269 uint8_t queue; 3270 3271 vm_page_assert_locked(m); 3272 3273 if ((queue = vm_page_queue(m)) == PQ_NONE) 3274 return; 3275 vm_page_aflag_set(m, PGA_DEQUEUE); 3276 vm_pqbatch_submit_page(m, queue); 3277 } 3278 3279 /* 3280 * A variant of vm_page_dequeue_deferred() that does not assert the page 3281 * lock and is only to be called from vm_page_free_prep(). It is just an 3282 * open-coded implementation of vm_page_dequeue_deferred(). Because the 3283 * page is being freed, we can assume that nothing else is scheduling queue 3284 * operations on this page, so we get for free the mutual exclusion that 3285 * is otherwise provided by the page lock. 3286 */ 3287 static void 3288 vm_page_dequeue_deferred_free(vm_page_t m) 3289 { 3290 uint8_t queue; 3291 3292 KASSERT(m->object == NULL, ("page %p has an object reference", m)); 3293 3294 if ((m->aflags & PGA_DEQUEUE) != 0) 3295 return; 3296 atomic_thread_fence_acq(); 3297 if ((queue = m->queue) == PQ_NONE) 3298 return; 3299 vm_page_aflag_set(m, PGA_DEQUEUE); 3300 vm_pqbatch_submit_page(m, queue); 3301 } 3302 3303 /* 3304 * vm_page_dequeue: 3305 * 3306 * Remove the page from whichever page queue it's in, if any. 3307 * The page must either be locked or unallocated. This constraint 3308 * ensures that the queue state of the page will remain consistent 3309 * after this function returns. 3310 */ 3311 void 3312 vm_page_dequeue(vm_page_t m) 3313 { 3314 struct mtx *lock, *lock1; 3315 struct vm_pagequeue *pq; 3316 uint8_t aflags; 3317 3318 KASSERT(mtx_owned(vm_page_lockptr(m)) || m->order == VM_NFREEORDER, 3319 ("page %p is allocated and unlocked", m)); 3320 3321 for (;;) { 3322 lock = vm_page_pagequeue_lockptr(m); 3323 if (lock == NULL) { 3324 /* 3325 * A thread may be concurrently executing 3326 * vm_page_dequeue_complete(). Ensure that all queue 3327 * state is cleared before we return. 3328 */ 3329 aflags = atomic_load_8(&m->aflags); 3330 if ((aflags & PGA_QUEUE_STATE_MASK) == 0) 3331 return; 3332 KASSERT((aflags & PGA_DEQUEUE) != 0, 3333 ("page %p has unexpected queue state flags %#x", 3334 m, aflags)); 3335 3336 /* 3337 * Busy wait until the thread updating queue state is 3338 * finished. Such a thread must be executing in a 3339 * critical section. 3340 */ 3341 cpu_spinwait(); 3342 continue; 3343 } 3344 mtx_lock(lock); 3345 if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock) 3346 break; 3347 mtx_unlock(lock); 3348 lock = lock1; 3349 } 3350 KASSERT(lock == vm_page_pagequeue_lockptr(m), 3351 ("%s: page %p migrated directly between queues", __func__, m)); 3352 KASSERT((m->aflags & PGA_DEQUEUE) != 0 || 3353 mtx_owned(vm_page_lockptr(m)), 3354 ("%s: queued unlocked page %p", __func__, m)); 3355 3356 if ((m->aflags & PGA_ENQUEUED) != 0) { 3357 pq = vm_page_pagequeue(m); 3358 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3359 vm_pagequeue_cnt_dec(pq); 3360 } 3361 vm_page_dequeue_complete(m); 3362 mtx_unlock(lock); 3363 } 3364 3365 /* 3366 * Schedule the given page for insertion into the specified page queue. 3367 * Physical insertion of the page may be deferred indefinitely. 3368 */ 3369 static void 3370 vm_page_enqueue(vm_page_t m, uint8_t queue) 3371 { 3372 3373 vm_page_assert_locked(m); 3374 KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, 3375 ("%s: page %p is already enqueued", __func__, m)); 3376 3377 m->queue = queue; 3378 if ((m->aflags & PGA_REQUEUE) == 0) 3379 vm_page_aflag_set(m, PGA_REQUEUE); 3380 vm_pqbatch_submit_page(m, queue); 3381 } 3382 3383 /* 3384 * vm_page_requeue: [ internal use only ] 3385 * 3386 * Schedule a requeue of the given page. 3387 * 3388 * The page must be locked. 3389 */ 3390 void 3391 vm_page_requeue(vm_page_t m) 3392 { 3393 3394 vm_page_assert_locked(m); 3395 KASSERT(vm_page_queue(m) != PQ_NONE, 3396 ("%s: page %p is not logically enqueued", __func__, m)); 3397 3398 if ((m->aflags & PGA_REQUEUE) == 0) 3399 vm_page_aflag_set(m, PGA_REQUEUE); 3400 vm_pqbatch_submit_page(m, atomic_load_8(&m->queue)); 3401 } 3402 3403 /* 3404 * vm_page_activate: 3405 * 3406 * Put the specified page on the active list (if appropriate). 3407 * Ensure that act_count is at least ACT_INIT but do not otherwise 3408 * mess with it. 3409 * 3410 * The page must be locked. 3411 */ 3412 void 3413 vm_page_activate(vm_page_t m) 3414 { 3415 3416 vm_page_assert_locked(m); 3417 3418 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0) 3419 return; 3420 if (vm_page_queue(m) == PQ_ACTIVE) { 3421 if (m->act_count < ACT_INIT) 3422 m->act_count = ACT_INIT; 3423 return; 3424 } 3425 3426 vm_page_dequeue(m); 3427 if (m->act_count < ACT_INIT) 3428 m->act_count = ACT_INIT; 3429 vm_page_enqueue(m, PQ_ACTIVE); 3430 } 3431 3432 /* 3433 * vm_page_free_prep: 3434 * 3435 * Prepares the given page to be put on the free list, 3436 * disassociating it from any VM object. The caller may return 3437 * the page to the free list only if this function returns true. 3438 * 3439 * The object must be locked. The page must be locked if it is 3440 * managed. 3441 */ 3442 bool 3443 vm_page_free_prep(vm_page_t m) 3444 { 3445 3446 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3447 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3448 uint64_t *p; 3449 int i; 3450 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3451 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3452 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3453 m, i, (uintmax_t)*p)); 3454 } 3455 #endif 3456 if ((m->oflags & VPO_UNMANAGED) == 0) { 3457 vm_page_lock_assert(m, MA_OWNED); 3458 KASSERT(!pmap_page_is_mapped(m), 3459 ("vm_page_free_prep: freeing mapped page %p", m)); 3460 } else 3461 KASSERT(m->queue == PQ_NONE, 3462 ("vm_page_free_prep: unmanaged page %p is queued", m)); 3463 VM_CNT_INC(v_tfree); 3464 3465 if (vm_page_sbusied(m)) 3466 panic("vm_page_free_prep: freeing busy page %p", m); 3467 3468 vm_page_remove(m); 3469 3470 /* 3471 * If fictitious remove object association and 3472 * return. 3473 */ 3474 if ((m->flags & PG_FICTITIOUS) != 0) { 3475 KASSERT(m->wire_count == 1, 3476 ("fictitious page %p is not wired", m)); 3477 KASSERT(m->queue == PQ_NONE, 3478 ("fictitious page %p is queued", m)); 3479 return (false); 3480 } 3481 3482 /* 3483 * Pages need not be dequeued before they are returned to the physical 3484 * memory allocator, but they must at least be marked for a deferred 3485 * dequeue. 3486 */ 3487 if ((m->oflags & VPO_UNMANAGED) == 0) 3488 vm_page_dequeue_deferred_free(m); 3489 3490 m->valid = 0; 3491 vm_page_undirty(m); 3492 3493 if (vm_page_wired(m) != 0) 3494 panic("vm_page_free_prep: freeing wired page %p", m); 3495 if (m->hold_count != 0) { 3496 m->flags &= ~PG_ZERO; 3497 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 3498 ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m)); 3499 m->flags |= PG_UNHOLDFREE; 3500 return (false); 3501 } 3502 3503 /* 3504 * Restore the default memory attribute to the page. 3505 */ 3506 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3507 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3508 3509 #if VM_NRESERVLEVEL > 0 3510 if (vm_reserv_free_page(m)) 3511 return (false); 3512 #endif 3513 3514 return (true); 3515 } 3516 3517 /* 3518 * vm_page_free_toq: 3519 * 3520 * Returns the given page to the free list, disassociating it 3521 * from any VM object. 3522 * 3523 * The object must be locked. The page must be locked if it is 3524 * managed. 3525 */ 3526 void 3527 vm_page_free_toq(vm_page_t m) 3528 { 3529 struct vm_domain *vmd; 3530 3531 if (!vm_page_free_prep(m)) 3532 return; 3533 3534 vmd = vm_pagequeue_domain(m); 3535 if (m->pool == VM_FREEPOOL_DEFAULT && vmd->vmd_pgcache != NULL) { 3536 uma_zfree(vmd->vmd_pgcache, m); 3537 return; 3538 } 3539 vm_domain_free_lock(vmd); 3540 vm_phys_free_pages(m, 0); 3541 vm_domain_free_unlock(vmd); 3542 vm_domain_freecnt_inc(vmd, 1); 3543 } 3544 3545 /* 3546 * vm_page_free_pages_toq: 3547 * 3548 * Returns a list of pages to the free list, disassociating it 3549 * from any VM object. In other words, this is equivalent to 3550 * calling vm_page_free_toq() for each page of a list of VM objects. 3551 * 3552 * The objects must be locked. The pages must be locked if it is 3553 * managed. 3554 */ 3555 void 3556 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) 3557 { 3558 vm_page_t m; 3559 int count; 3560 3561 if (SLIST_EMPTY(free)) 3562 return; 3563 3564 count = 0; 3565 while ((m = SLIST_FIRST(free)) != NULL) { 3566 count++; 3567 SLIST_REMOVE_HEAD(free, plinks.s.ss); 3568 vm_page_free_toq(m); 3569 } 3570 3571 if (update_wire_count) 3572 vm_wire_sub(count); 3573 } 3574 3575 /* 3576 * vm_page_wire: 3577 * 3578 * Mark this page as wired down. If the page is fictitious, then 3579 * its wire count must remain one. 3580 * 3581 * The page must be locked. 3582 */ 3583 void 3584 vm_page_wire(vm_page_t m) 3585 { 3586 3587 vm_page_assert_locked(m); 3588 if ((m->flags & PG_FICTITIOUS) != 0) { 3589 KASSERT(m->wire_count == 1, 3590 ("vm_page_wire: fictitious page %p's wire count isn't one", 3591 m)); 3592 return; 3593 } 3594 if (!vm_page_wired(m)) { 3595 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 3596 m->queue == PQ_NONE, 3597 ("vm_page_wire: unmanaged page %p is queued", m)); 3598 vm_wire_add(1); 3599 } 3600 m->wire_count++; 3601 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 3602 } 3603 3604 /* 3605 * vm_page_unwire: 3606 * 3607 * Release one wiring of the specified page, potentially allowing it to be 3608 * paged out. Returns TRUE if the number of wirings transitions to zero and 3609 * FALSE otherwise. 3610 * 3611 * Only managed pages belonging to an object can be paged out. If the number 3612 * of wirings transitions to zero and the page is eligible for page out, then 3613 * the page is added to the specified paging queue (unless PQ_NONE is 3614 * specified, in which case the page is dequeued if it belongs to a paging 3615 * queue). 3616 * 3617 * If a page is fictitious, then its wire count must always be one. 3618 * 3619 * A managed page must be locked. 3620 */ 3621 bool 3622 vm_page_unwire(vm_page_t m, uint8_t queue) 3623 { 3624 bool unwired; 3625 3626 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 3627 ("vm_page_unwire: invalid queue %u request for page %p", 3628 queue, m)); 3629 if ((m->oflags & VPO_UNMANAGED) == 0) 3630 vm_page_assert_locked(m); 3631 3632 unwired = vm_page_unwire_noq(m); 3633 if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL) 3634 return (unwired); 3635 3636 if (vm_page_queue(m) == queue) { 3637 if (queue == PQ_ACTIVE) 3638 vm_page_reference(m); 3639 else if (queue != PQ_NONE) 3640 vm_page_requeue(m); 3641 } else { 3642 vm_page_dequeue(m); 3643 if (queue != PQ_NONE) { 3644 vm_page_enqueue(m, queue); 3645 if (queue == PQ_ACTIVE) 3646 /* Initialize act_count. */ 3647 vm_page_activate(m); 3648 } 3649 } 3650 return (unwired); 3651 } 3652 3653 /* 3654 * 3655 * vm_page_unwire_noq: 3656 * 3657 * Unwire a page without (re-)inserting it into a page queue. It is up 3658 * to the caller to enqueue, requeue, or free the page as appropriate. 3659 * In most cases, vm_page_unwire() should be used instead. 3660 */ 3661 bool 3662 vm_page_unwire_noq(vm_page_t m) 3663 { 3664 3665 if ((m->oflags & VPO_UNMANAGED) == 0) 3666 vm_page_assert_locked(m); 3667 if ((m->flags & PG_FICTITIOUS) != 0) { 3668 KASSERT(m->wire_count == 1, 3669 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 3670 return (false); 3671 } 3672 if (!vm_page_wired(m)) 3673 panic("vm_page_unwire: page %p's wire count is zero", m); 3674 m->wire_count--; 3675 if (m->wire_count == 0) { 3676 vm_wire_sub(1); 3677 return (true); 3678 } else 3679 return (false); 3680 } 3681 3682 /* 3683 * Move the specified page to the tail of the inactive queue, or requeue 3684 * the page if it is already in the inactive queue. 3685 * 3686 * The page must be locked. 3687 */ 3688 void 3689 vm_page_deactivate(vm_page_t m) 3690 { 3691 3692 vm_page_assert_locked(m); 3693 3694 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0) 3695 return; 3696 3697 if (!vm_page_inactive(m)) { 3698 vm_page_dequeue(m); 3699 vm_page_enqueue(m, PQ_INACTIVE); 3700 } else 3701 vm_page_requeue(m); 3702 } 3703 3704 /* 3705 * Move the specified page close to the head of the inactive queue, 3706 * bypassing LRU. A marker page is used to maintain FIFO ordering. 3707 * As with regular enqueues, we use a per-CPU batch queue to reduce 3708 * contention on the page queue lock. 3709 * 3710 * The page must be locked. 3711 */ 3712 void 3713 vm_page_deactivate_noreuse(vm_page_t m) 3714 { 3715 3716 vm_page_assert_locked(m); 3717 3718 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0) 3719 return; 3720 3721 if (!vm_page_inactive(m)) { 3722 vm_page_dequeue(m); 3723 m->queue = PQ_INACTIVE; 3724 } 3725 if ((m->aflags & PGA_REQUEUE_HEAD) == 0) 3726 vm_page_aflag_set(m, PGA_REQUEUE_HEAD); 3727 vm_pqbatch_submit_page(m, PQ_INACTIVE); 3728 } 3729 3730 /* 3731 * vm_page_launder 3732 * 3733 * Put a page in the laundry, or requeue it if it is already there. 3734 */ 3735 void 3736 vm_page_launder(vm_page_t m) 3737 { 3738 3739 vm_page_assert_locked(m); 3740 if (vm_page_wired(m) || (m->oflags & VPO_UNMANAGED) != 0) 3741 return; 3742 3743 if (vm_page_in_laundry(m)) 3744 vm_page_requeue(m); 3745 else { 3746 vm_page_dequeue(m); 3747 vm_page_enqueue(m, PQ_LAUNDRY); 3748 } 3749 } 3750 3751 /* 3752 * vm_page_unswappable 3753 * 3754 * Put a page in the PQ_UNSWAPPABLE holding queue. 3755 */ 3756 void 3757 vm_page_unswappable(vm_page_t m) 3758 { 3759 3760 vm_page_assert_locked(m); 3761 KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, 3762 ("page %p already unswappable", m)); 3763 3764 vm_page_dequeue(m); 3765 vm_page_enqueue(m, PQ_UNSWAPPABLE); 3766 } 3767 3768 /* 3769 * Attempt to free the page. If it cannot be freed, do nothing. Returns true 3770 * if the page is freed and false otherwise. 3771 * 3772 * The page must be managed. The page and its containing object must be 3773 * locked. 3774 */ 3775 bool 3776 vm_page_try_to_free(vm_page_t m) 3777 { 3778 3779 vm_page_assert_locked(m); 3780 VM_OBJECT_ASSERT_WLOCKED(m->object); 3781 KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m)); 3782 if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m)) 3783 return (false); 3784 if (m->object->ref_count != 0) { 3785 pmap_remove_all(m); 3786 if (m->dirty != 0) 3787 return (false); 3788 } 3789 vm_page_free(m); 3790 return (true); 3791 } 3792 3793 /* 3794 * vm_page_advise 3795 * 3796 * Apply the specified advice to the given page. 3797 * 3798 * The object and page must be locked. 3799 */ 3800 void 3801 vm_page_advise(vm_page_t m, int advice) 3802 { 3803 3804 vm_page_assert_locked(m); 3805 VM_OBJECT_ASSERT_WLOCKED(m->object); 3806 if (advice == MADV_FREE) 3807 /* 3808 * Mark the page clean. This will allow the page to be freed 3809 * without first paging it out. MADV_FREE pages are often 3810 * quickly reused by malloc(3), so we do not do anything that 3811 * would result in a page fault on a later access. 3812 */ 3813 vm_page_undirty(m); 3814 else if (advice != MADV_DONTNEED) { 3815 if (advice == MADV_WILLNEED) 3816 vm_page_activate(m); 3817 return; 3818 } 3819 3820 /* 3821 * Clear any references to the page. Otherwise, the page daemon will 3822 * immediately reactivate the page. 3823 */ 3824 vm_page_aflag_clear(m, PGA_REFERENCED); 3825 3826 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 3827 vm_page_dirty(m); 3828 3829 /* 3830 * Place clean pages near the head of the inactive queue rather than 3831 * the tail, thus defeating the queue's LRU operation and ensuring that 3832 * the page will be reused quickly. Dirty pages not already in the 3833 * laundry are moved there. 3834 */ 3835 if (m->dirty == 0) 3836 vm_page_deactivate_noreuse(m); 3837 else if (!vm_page_in_laundry(m)) 3838 vm_page_launder(m); 3839 } 3840 3841 /* 3842 * Grab a page, waiting until we are waken up due to the page 3843 * changing state. We keep on waiting, if the page continues 3844 * to be in the object. If the page doesn't exist, first allocate it 3845 * and then conditionally zero it. 3846 * 3847 * This routine may sleep. 3848 * 3849 * The object must be locked on entry. The lock will, however, be released 3850 * and reacquired if the routine sleeps. 3851 */ 3852 vm_page_t 3853 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3854 { 3855 vm_page_t m; 3856 int sleep; 3857 int pflags; 3858 3859 VM_OBJECT_ASSERT_WLOCKED(object); 3860 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3861 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3862 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3863 pflags = allocflags & 3864 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 3865 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 3866 pflags |= VM_ALLOC_WAITFAIL; 3867 retrylookup: 3868 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3869 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3870 vm_page_xbusied(m) : vm_page_busied(m); 3871 if (sleep) { 3872 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3873 return (NULL); 3874 /* 3875 * Reference the page before unlocking and 3876 * sleeping so that the page daemon is less 3877 * likely to reclaim it. 3878 */ 3879 vm_page_aflag_set(m, PGA_REFERENCED); 3880 vm_page_lock(m); 3881 VM_OBJECT_WUNLOCK(object); 3882 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 3883 VM_ALLOC_IGN_SBUSY) != 0); 3884 VM_OBJECT_WLOCK(object); 3885 goto retrylookup; 3886 } else { 3887 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3888 vm_page_lock(m); 3889 vm_page_wire(m); 3890 vm_page_unlock(m); 3891 } 3892 if ((allocflags & 3893 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3894 vm_page_xbusy(m); 3895 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3896 vm_page_sbusy(m); 3897 return (m); 3898 } 3899 } 3900 m = vm_page_alloc(object, pindex, pflags); 3901 if (m == NULL) { 3902 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3903 return (NULL); 3904 goto retrylookup; 3905 } 3906 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3907 pmap_zero_page(m); 3908 return (m); 3909 } 3910 3911 /* 3912 * Return the specified range of pages from the given object. For each 3913 * page offset within the range, if a page already exists within the object 3914 * at that offset and it is busy, then wait for it to change state. If, 3915 * instead, the page doesn't exist, then allocate it. 3916 * 3917 * The caller must always specify an allocation class. 3918 * 3919 * allocation classes: 3920 * VM_ALLOC_NORMAL normal process request 3921 * VM_ALLOC_SYSTEM system *really* needs the pages 3922 * 3923 * The caller must always specify that the pages are to be busied and/or 3924 * wired. 3925 * 3926 * optional allocation flags: 3927 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 3928 * VM_ALLOC_NOBUSY do not exclusive busy the page 3929 * VM_ALLOC_NOWAIT do not sleep 3930 * VM_ALLOC_SBUSY set page to sbusy state 3931 * VM_ALLOC_WIRED wire the pages 3932 * VM_ALLOC_ZERO zero and validate any invalid pages 3933 * 3934 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 3935 * may return a partial prefix of the requested range. 3936 */ 3937 int 3938 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 3939 vm_page_t *ma, int count) 3940 { 3941 vm_page_t m, mpred; 3942 int pflags; 3943 int i; 3944 bool sleep; 3945 3946 VM_OBJECT_ASSERT_WLOCKED(object); 3947 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 3948 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 3949 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 3950 (allocflags & VM_ALLOC_WIRED) != 0, 3951 ("vm_page_grab_pages: the pages must be busied or wired")); 3952 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3953 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3954 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); 3955 if (count == 0) 3956 return (0); 3957 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | 3958 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY); 3959 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 3960 pflags |= VM_ALLOC_WAITFAIL; 3961 i = 0; 3962 retrylookup: 3963 m = vm_radix_lookup_le(&object->rtree, pindex + i); 3964 if (m == NULL || m->pindex != pindex + i) { 3965 mpred = m; 3966 m = NULL; 3967 } else 3968 mpred = TAILQ_PREV(m, pglist, listq); 3969 for (; i < count; i++) { 3970 if (m != NULL) { 3971 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3972 vm_page_xbusied(m) : vm_page_busied(m); 3973 if (sleep) { 3974 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3975 break; 3976 /* 3977 * Reference the page before unlocking and 3978 * sleeping so that the page daemon is less 3979 * likely to reclaim it. 3980 */ 3981 vm_page_aflag_set(m, PGA_REFERENCED); 3982 vm_page_lock(m); 3983 VM_OBJECT_WUNLOCK(object); 3984 vm_page_busy_sleep(m, "grbmaw", (allocflags & 3985 VM_ALLOC_IGN_SBUSY) != 0); 3986 VM_OBJECT_WLOCK(object); 3987 goto retrylookup; 3988 } 3989 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3990 vm_page_lock(m); 3991 vm_page_wire(m); 3992 vm_page_unlock(m); 3993 } 3994 if ((allocflags & (VM_ALLOC_NOBUSY | 3995 VM_ALLOC_SBUSY)) == 0) 3996 vm_page_xbusy(m); 3997 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3998 vm_page_sbusy(m); 3999 } else { 4000 m = vm_page_alloc_after(object, pindex + i, 4001 pflags | VM_ALLOC_COUNT(count - i), mpred); 4002 if (m == NULL) { 4003 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 4004 break; 4005 goto retrylookup; 4006 } 4007 } 4008 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) { 4009 if ((m->flags & PG_ZERO) == 0) 4010 pmap_zero_page(m); 4011 m->valid = VM_PAGE_BITS_ALL; 4012 } 4013 ma[i] = mpred = m; 4014 m = vm_page_next(m); 4015 } 4016 return (i); 4017 } 4018 4019 /* 4020 * Mapping function for valid or dirty bits in a page. 4021 * 4022 * Inputs are required to range within a page. 4023 */ 4024 vm_page_bits_t 4025 vm_page_bits(int base, int size) 4026 { 4027 int first_bit; 4028 int last_bit; 4029 4030 KASSERT( 4031 base + size <= PAGE_SIZE, 4032 ("vm_page_bits: illegal base/size %d/%d", base, size) 4033 ); 4034 4035 if (size == 0) /* handle degenerate case */ 4036 return (0); 4037 4038 first_bit = base >> DEV_BSHIFT; 4039 last_bit = (base + size - 1) >> DEV_BSHIFT; 4040 4041 return (((vm_page_bits_t)2 << last_bit) - 4042 ((vm_page_bits_t)1 << first_bit)); 4043 } 4044 4045 /* 4046 * vm_page_set_valid_range: 4047 * 4048 * Sets portions of a page valid. The arguments are expected 4049 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4050 * of any partial chunks touched by the range. The invalid portion of 4051 * such chunks will be zeroed. 4052 * 4053 * (base + size) must be less then or equal to PAGE_SIZE. 4054 */ 4055 void 4056 vm_page_set_valid_range(vm_page_t m, int base, int size) 4057 { 4058 int endoff, frag; 4059 4060 VM_OBJECT_ASSERT_WLOCKED(m->object); 4061 if (size == 0) /* handle degenerate case */ 4062 return; 4063 4064 /* 4065 * If the base is not DEV_BSIZE aligned and the valid 4066 * bit is clear, we have to zero out a portion of the 4067 * first block. 4068 */ 4069 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4070 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 4071 pmap_zero_page_area(m, frag, base - frag); 4072 4073 /* 4074 * If the ending offset is not DEV_BSIZE aligned and the 4075 * valid bit is clear, we have to zero out a portion of 4076 * the last block. 4077 */ 4078 endoff = base + size; 4079 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4080 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 4081 pmap_zero_page_area(m, endoff, 4082 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4083 4084 /* 4085 * Assert that no previously invalid block that is now being validated 4086 * is already dirty. 4087 */ 4088 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 4089 ("vm_page_set_valid_range: page %p is dirty", m)); 4090 4091 /* 4092 * Set valid bits inclusive of any overlap. 4093 */ 4094 m->valid |= vm_page_bits(base, size); 4095 } 4096 4097 /* 4098 * Clear the given bits from the specified page's dirty field. 4099 */ 4100 static __inline void 4101 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 4102 { 4103 uintptr_t addr; 4104 #if PAGE_SIZE < 16384 4105 int shift; 4106 #endif 4107 4108 /* 4109 * If the object is locked and the page is neither exclusive busy nor 4110 * write mapped, then the page's dirty field cannot possibly be 4111 * set by a concurrent pmap operation. 4112 */ 4113 VM_OBJECT_ASSERT_WLOCKED(m->object); 4114 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 4115 m->dirty &= ~pagebits; 4116 else { 4117 /* 4118 * The pmap layer can call vm_page_dirty() without 4119 * holding a distinguished lock. The combination of 4120 * the object's lock and an atomic operation suffice 4121 * to guarantee consistency of the page dirty field. 4122 * 4123 * For PAGE_SIZE == 32768 case, compiler already 4124 * properly aligns the dirty field, so no forcible 4125 * alignment is needed. Only require existence of 4126 * atomic_clear_64 when page size is 32768. 4127 */ 4128 addr = (uintptr_t)&m->dirty; 4129 #if PAGE_SIZE == 32768 4130 atomic_clear_64((uint64_t *)addr, pagebits); 4131 #elif PAGE_SIZE == 16384 4132 atomic_clear_32((uint32_t *)addr, pagebits); 4133 #else /* PAGE_SIZE <= 8192 */ 4134 /* 4135 * Use a trick to perform a 32-bit atomic on the 4136 * containing aligned word, to not depend on the existence 4137 * of atomic_clear_{8, 16}. 4138 */ 4139 shift = addr & (sizeof(uint32_t) - 1); 4140 #if BYTE_ORDER == BIG_ENDIAN 4141 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 4142 #else 4143 shift *= NBBY; 4144 #endif 4145 addr &= ~(sizeof(uint32_t) - 1); 4146 atomic_clear_32((uint32_t *)addr, pagebits << shift); 4147 #endif /* PAGE_SIZE */ 4148 } 4149 } 4150 4151 /* 4152 * vm_page_set_validclean: 4153 * 4154 * Sets portions of a page valid and clean. The arguments are expected 4155 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 4156 * of any partial chunks touched by the range. The invalid portion of 4157 * such chunks will be zero'd. 4158 * 4159 * (base + size) must be less then or equal to PAGE_SIZE. 4160 */ 4161 void 4162 vm_page_set_validclean(vm_page_t m, int base, int size) 4163 { 4164 vm_page_bits_t oldvalid, pagebits; 4165 int endoff, frag; 4166 4167 VM_OBJECT_ASSERT_WLOCKED(m->object); 4168 if (size == 0) /* handle degenerate case */ 4169 return; 4170 4171 /* 4172 * If the base is not DEV_BSIZE aligned and the valid 4173 * bit is clear, we have to zero out a portion of the 4174 * first block. 4175 */ 4176 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 4177 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 4178 pmap_zero_page_area(m, frag, base - frag); 4179 4180 /* 4181 * If the ending offset is not DEV_BSIZE aligned and the 4182 * valid bit is clear, we have to zero out a portion of 4183 * the last block. 4184 */ 4185 endoff = base + size; 4186 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 4187 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 4188 pmap_zero_page_area(m, endoff, 4189 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 4190 4191 /* 4192 * Set valid, clear dirty bits. If validating the entire 4193 * page we can safely clear the pmap modify bit. We also 4194 * use this opportunity to clear the VPO_NOSYNC flag. If a process 4195 * takes a write fault on a MAP_NOSYNC memory area the flag will 4196 * be set again. 4197 * 4198 * We set valid bits inclusive of any overlap, but we can only 4199 * clear dirty bits for DEV_BSIZE chunks that are fully within 4200 * the range. 4201 */ 4202 oldvalid = m->valid; 4203 pagebits = vm_page_bits(base, size); 4204 m->valid |= pagebits; 4205 #if 0 /* NOT YET */ 4206 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 4207 frag = DEV_BSIZE - frag; 4208 base += frag; 4209 size -= frag; 4210 if (size < 0) 4211 size = 0; 4212 } 4213 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 4214 #endif 4215 if (base == 0 && size == PAGE_SIZE) { 4216 /* 4217 * The page can only be modified within the pmap if it is 4218 * mapped, and it can only be mapped if it was previously 4219 * fully valid. 4220 */ 4221 if (oldvalid == VM_PAGE_BITS_ALL) 4222 /* 4223 * Perform the pmap_clear_modify() first. Otherwise, 4224 * a concurrent pmap operation, such as 4225 * pmap_protect(), could clear a modification in the 4226 * pmap and set the dirty field on the page before 4227 * pmap_clear_modify() had begun and after the dirty 4228 * field was cleared here. 4229 */ 4230 pmap_clear_modify(m); 4231 m->dirty = 0; 4232 m->oflags &= ~VPO_NOSYNC; 4233 } else if (oldvalid != VM_PAGE_BITS_ALL) 4234 m->dirty &= ~pagebits; 4235 else 4236 vm_page_clear_dirty_mask(m, pagebits); 4237 } 4238 4239 void 4240 vm_page_clear_dirty(vm_page_t m, int base, int size) 4241 { 4242 4243 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 4244 } 4245 4246 /* 4247 * vm_page_set_invalid: 4248 * 4249 * Invalidates DEV_BSIZE'd chunks within a page. Both the 4250 * valid and dirty bits for the effected areas are cleared. 4251 */ 4252 void 4253 vm_page_set_invalid(vm_page_t m, int base, int size) 4254 { 4255 vm_page_bits_t bits; 4256 vm_object_t object; 4257 4258 object = m->object; 4259 VM_OBJECT_ASSERT_WLOCKED(object); 4260 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 4261 size >= object->un_pager.vnp.vnp_size) 4262 bits = VM_PAGE_BITS_ALL; 4263 else 4264 bits = vm_page_bits(base, size); 4265 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 4266 bits != 0) 4267 pmap_remove_all(m); 4268 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 4269 !pmap_page_is_mapped(m), 4270 ("vm_page_set_invalid: page %p is mapped", m)); 4271 m->valid &= ~bits; 4272 m->dirty &= ~bits; 4273 } 4274 4275 /* 4276 * vm_page_zero_invalid() 4277 * 4278 * The kernel assumes that the invalid portions of a page contain 4279 * garbage, but such pages can be mapped into memory by user code. 4280 * When this occurs, we must zero out the non-valid portions of the 4281 * page so user code sees what it expects. 4282 * 4283 * Pages are most often semi-valid when the end of a file is mapped 4284 * into memory and the file's size is not page aligned. 4285 */ 4286 void 4287 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 4288 { 4289 int b; 4290 int i; 4291 4292 VM_OBJECT_ASSERT_WLOCKED(m->object); 4293 /* 4294 * Scan the valid bits looking for invalid sections that 4295 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 4296 * valid bit may be set ) have already been zeroed by 4297 * vm_page_set_validclean(). 4298 */ 4299 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 4300 if (i == (PAGE_SIZE / DEV_BSIZE) || 4301 (m->valid & ((vm_page_bits_t)1 << i))) { 4302 if (i > b) { 4303 pmap_zero_page_area(m, 4304 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 4305 } 4306 b = i + 1; 4307 } 4308 } 4309 4310 /* 4311 * setvalid is TRUE when we can safely set the zero'd areas 4312 * as being valid. We can do this if there are no cache consistancy 4313 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 4314 */ 4315 if (setvalid) 4316 m->valid = VM_PAGE_BITS_ALL; 4317 } 4318 4319 /* 4320 * vm_page_is_valid: 4321 * 4322 * Is (partial) page valid? Note that the case where size == 0 4323 * will return FALSE in the degenerate case where the page is 4324 * entirely invalid, and TRUE otherwise. 4325 */ 4326 int 4327 vm_page_is_valid(vm_page_t m, int base, int size) 4328 { 4329 vm_page_bits_t bits; 4330 4331 VM_OBJECT_ASSERT_LOCKED(m->object); 4332 bits = vm_page_bits(base, size); 4333 return (m->valid != 0 && (m->valid & bits) == bits); 4334 } 4335 4336 /* 4337 * Returns true if all of the specified predicates are true for the entire 4338 * (super)page and false otherwise. 4339 */ 4340 bool 4341 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 4342 { 4343 vm_object_t object; 4344 int i, npages; 4345 4346 object = m->object; 4347 if (skip_m != NULL && skip_m->object != object) 4348 return (false); 4349 VM_OBJECT_ASSERT_LOCKED(object); 4350 npages = atop(pagesizes[m->psind]); 4351 4352 /* 4353 * The physically contiguous pages that make up a superpage, i.e., a 4354 * page with a page size index ("psind") greater than zero, will 4355 * occupy adjacent entries in vm_page_array[]. 4356 */ 4357 for (i = 0; i < npages; i++) { 4358 /* Always test object consistency, including "skip_m". */ 4359 if (m[i].object != object) 4360 return (false); 4361 if (&m[i] == skip_m) 4362 continue; 4363 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 4364 return (false); 4365 if ((flags & PS_ALL_DIRTY) != 0) { 4366 /* 4367 * Calling vm_page_test_dirty() or pmap_is_modified() 4368 * might stop this case from spuriously returning 4369 * "false". However, that would require a write lock 4370 * on the object containing "m[i]". 4371 */ 4372 if (m[i].dirty != VM_PAGE_BITS_ALL) 4373 return (false); 4374 } 4375 if ((flags & PS_ALL_VALID) != 0 && 4376 m[i].valid != VM_PAGE_BITS_ALL) 4377 return (false); 4378 } 4379 return (true); 4380 } 4381 4382 /* 4383 * Set the page's dirty bits if the page is modified. 4384 */ 4385 void 4386 vm_page_test_dirty(vm_page_t m) 4387 { 4388 4389 VM_OBJECT_ASSERT_WLOCKED(m->object); 4390 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 4391 vm_page_dirty(m); 4392 } 4393 4394 void 4395 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 4396 { 4397 4398 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 4399 } 4400 4401 void 4402 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 4403 { 4404 4405 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 4406 } 4407 4408 int 4409 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 4410 { 4411 4412 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 4413 } 4414 4415 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 4416 void 4417 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 4418 { 4419 4420 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 4421 } 4422 4423 void 4424 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 4425 { 4426 4427 mtx_assert_(vm_page_lockptr(m), a, file, line); 4428 } 4429 #endif 4430 4431 #ifdef INVARIANTS 4432 void 4433 vm_page_object_lock_assert(vm_page_t m) 4434 { 4435 4436 /* 4437 * Certain of the page's fields may only be modified by the 4438 * holder of the containing object's lock or the exclusive busy. 4439 * holder. Unfortunately, the holder of the write busy is 4440 * not recorded, and thus cannot be checked here. 4441 */ 4442 if (m->object != NULL && !vm_page_xbusied(m)) 4443 VM_OBJECT_ASSERT_WLOCKED(m->object); 4444 } 4445 4446 void 4447 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 4448 { 4449 4450 if ((bits & PGA_WRITEABLE) == 0) 4451 return; 4452 4453 /* 4454 * The PGA_WRITEABLE flag can only be set if the page is 4455 * managed, is exclusively busied or the object is locked. 4456 * Currently, this flag is only set by pmap_enter(). 4457 */ 4458 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4459 ("PGA_WRITEABLE on unmanaged page")); 4460 if (!vm_page_xbusied(m)) 4461 VM_OBJECT_ASSERT_LOCKED(m->object); 4462 } 4463 #endif 4464 4465 #include "opt_ddb.h" 4466 #ifdef DDB 4467 #include <sys/kernel.h> 4468 4469 #include <ddb/ddb.h> 4470 4471 DB_SHOW_COMMAND(page, vm_page_print_page_info) 4472 { 4473 4474 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 4475 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 4476 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 4477 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 4478 db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); 4479 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 4480 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 4481 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 4482 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 4483 } 4484 4485 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 4486 { 4487 int dom; 4488 4489 db_printf("pq_free %d\n", vm_free_count()); 4490 for (dom = 0; dom < vm_ndomains; dom++) { 4491 db_printf( 4492 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 4493 dom, 4494 vm_dom[dom].vmd_page_count, 4495 vm_dom[dom].vmd_free_count, 4496 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 4497 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 4498 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 4499 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 4500 } 4501 } 4502 4503 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 4504 { 4505 vm_page_t m; 4506 boolean_t phys, virt; 4507 4508 if (!have_addr) { 4509 db_printf("show pginfo addr\n"); 4510 return; 4511 } 4512 4513 phys = strchr(modif, 'p') != NULL; 4514 virt = strchr(modif, 'v') != NULL; 4515 if (virt) 4516 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 4517 else if (phys) 4518 m = PHYS_TO_VM_PAGE(addr); 4519 else 4520 m = (vm_page_t)addr; 4521 db_printf( 4522 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 4523 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 4524 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 4525 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 4526 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 4527 } 4528 #endif /* DDB */ 4529