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 * GENERAL RULES ON VM_PAGE MANIPULATION 67 * 68 * - A page queue lock is required when adding or removing a page from a 69 * page queue regardless of other locks or the busy state of a page. 70 * 71 * * In general, no thread besides the page daemon can acquire or 72 * hold more than one page queue lock at a time. 73 * 74 * * The page daemon can acquire and hold any pair of page queue 75 * locks in any order. 76 * 77 * - The object lock is required when inserting or removing 78 * pages from an object (vm_page_insert() or vm_page_remove()). 79 * 80 */ 81 82 /* 83 * Resident memory management module. 84 */ 85 86 #include <sys/cdefs.h> 87 __FBSDID("$FreeBSD$"); 88 89 #include "opt_vm.h" 90 91 #include <sys/param.h> 92 #include <sys/systm.h> 93 #include <sys/lock.h> 94 #include <sys/domainset.h> 95 #include <sys/kernel.h> 96 #include <sys/limits.h> 97 #include <sys/linker.h> 98 #include <sys/malloc.h> 99 #include <sys/mman.h> 100 #include <sys/msgbuf.h> 101 #include <sys/mutex.h> 102 #include <sys/proc.h> 103 #include <sys/rwlock.h> 104 #include <sys/sbuf.h> 105 #include <sys/smp.h> 106 #include <sys/sysctl.h> 107 #include <sys/vmmeter.h> 108 #include <sys/vnode.h> 109 110 #include <vm/vm.h> 111 #include <vm/pmap.h> 112 #include <vm/vm_param.h> 113 #include <vm/vm_domainset.h> 114 #include <vm/vm_kern.h> 115 #include <vm/vm_map.h> 116 #include <vm/vm_object.h> 117 #include <vm/vm_page.h> 118 #include <vm/vm_pageout.h> 119 #include <vm/vm_phys.h> 120 #include <vm/vm_pagequeue.h> 121 #include <vm/vm_pager.h> 122 #include <vm/vm_radix.h> 123 #include <vm/vm_reserv.h> 124 #include <vm/vm_extern.h> 125 #include <vm/uma.h> 126 #include <vm/uma_int.h> 127 128 #include <machine/md_var.h> 129 130 extern int uma_startup_count(int); 131 extern void uma_startup(void *, int); 132 extern int vmem_startup_count(void); 133 134 /* 135 * Associated with page of user-allocatable memory is a 136 * page structure. 137 */ 138 139 struct vm_domain vm_dom[MAXMEMDOM]; 140 141 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 142 struct mtx_padalign __exclusive_cache_line vm_domainset_lock; 143 domainset_t __exclusive_cache_line vm_min_domains; 144 domainset_t __exclusive_cache_line vm_severe_domains; 145 static int vm_min_waiters; 146 static int vm_severe_waiters; 147 static int vm_pageproc_waiters; 148 149 150 /* 151 * bogus page -- for I/O to/from partially complete buffers, 152 * or for paging into sparsely invalid regions. 153 */ 154 vm_page_t bogus_page; 155 156 vm_page_t vm_page_array; 157 long vm_page_array_size; 158 long first_page; 159 160 static int boot_pages; 161 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 162 &boot_pages, 0, 163 "number of pages allocated for bootstrapping the VM system"); 164 165 static int pa_tryrelock_restart; 166 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 167 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 168 169 static TAILQ_HEAD(, vm_page) blacklist_head; 170 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 171 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 172 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 173 174 static uma_zone_t fakepg_zone; 175 176 static void vm_page_alloc_check(vm_page_t m); 177 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 178 static void vm_page_enqueue(uint8_t queue, vm_page_t m); 179 static void vm_page_free_phys(struct vm_domain *vmd, vm_page_t m); 180 static void vm_page_init(void *dummy); 181 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 182 vm_pindex_t pindex, vm_page_t mpred); 183 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 184 vm_page_t mpred); 185 static int vm_page_reclaim_run(int req_class, int domain, u_long npages, 186 vm_page_t m_run, vm_paddr_t high); 187 static void vm_domain_free_wakeup(struct vm_domain *); 188 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, 189 int req); 190 191 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 192 193 static void 194 vm_page_init(void *dummy) 195 { 196 197 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 198 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 199 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 200 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 201 } 202 203 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 204 #if PAGE_SIZE == 32768 205 #ifdef CTASSERT 206 CTASSERT(sizeof(u_long) >= 8); 207 #endif 208 #endif 209 210 /* 211 * Try to acquire a physical address lock while a pmap is locked. If we 212 * fail to trylock we unlock and lock the pmap directly and cache the 213 * locked pa in *locked. The caller should then restart their loop in case 214 * the virtual to physical mapping has changed. 215 */ 216 int 217 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 218 { 219 vm_paddr_t lockpa; 220 221 lockpa = *locked; 222 *locked = pa; 223 if (lockpa) { 224 PA_LOCK_ASSERT(lockpa, MA_OWNED); 225 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 226 return (0); 227 PA_UNLOCK(lockpa); 228 } 229 if (PA_TRYLOCK(pa)) 230 return (0); 231 PMAP_UNLOCK(pmap); 232 atomic_add_int(&pa_tryrelock_restart, 1); 233 PA_LOCK(pa); 234 PMAP_LOCK(pmap); 235 return (EAGAIN); 236 } 237 238 /* 239 * vm_set_page_size: 240 * 241 * Sets the page size, perhaps based upon the memory 242 * size. Must be called before any use of page-size 243 * dependent functions. 244 */ 245 void 246 vm_set_page_size(void) 247 { 248 if (vm_cnt.v_page_size == 0) 249 vm_cnt.v_page_size = PAGE_SIZE; 250 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 251 panic("vm_set_page_size: page size not a power of two"); 252 } 253 254 /* 255 * vm_page_blacklist_next: 256 * 257 * Find the next entry in the provided string of blacklist 258 * addresses. Entries are separated by space, comma, or newline. 259 * If an invalid integer is encountered then the rest of the 260 * string is skipped. Updates the list pointer to the next 261 * character, or NULL if the string is exhausted or invalid. 262 */ 263 static vm_paddr_t 264 vm_page_blacklist_next(char **list, char *end) 265 { 266 vm_paddr_t bad; 267 char *cp, *pos; 268 269 if (list == NULL || *list == NULL) 270 return (0); 271 if (**list =='\0') { 272 *list = NULL; 273 return (0); 274 } 275 276 /* 277 * If there's no end pointer then the buffer is coming from 278 * the kenv and we know it's null-terminated. 279 */ 280 if (end == NULL) 281 end = *list + strlen(*list); 282 283 /* Ensure that strtoq() won't walk off the end */ 284 if (*end != '\0') { 285 if (*end == '\n' || *end == ' ' || *end == ',') 286 *end = '\0'; 287 else { 288 printf("Blacklist not terminated, skipping\n"); 289 *list = NULL; 290 return (0); 291 } 292 } 293 294 for (pos = *list; *pos != '\0'; pos = cp) { 295 bad = strtoq(pos, &cp, 0); 296 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 297 if (bad == 0) { 298 if (++cp < end) 299 continue; 300 else 301 break; 302 } 303 } else 304 break; 305 if (*cp == '\0' || ++cp >= end) 306 *list = NULL; 307 else 308 *list = cp; 309 return (trunc_page(bad)); 310 } 311 printf("Garbage in RAM blacklist, skipping\n"); 312 *list = NULL; 313 return (0); 314 } 315 316 /* 317 * vm_page_blacklist_check: 318 * 319 * Iterate through the provided string of blacklist addresses, pulling 320 * each entry out of the physical allocator free list and putting it 321 * onto a list for reporting via the vm.page_blacklist sysctl. 322 */ 323 static void 324 vm_page_blacklist_check(char *list, char *end) 325 { 326 struct vm_domain *vmd; 327 vm_paddr_t pa; 328 vm_page_t m; 329 char *next; 330 int ret; 331 332 next = list; 333 while (next != NULL) { 334 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 335 continue; 336 m = vm_phys_paddr_to_vm_page(pa); 337 if (m == NULL) 338 continue; 339 vmd = vm_pagequeue_domain(m); 340 vm_domain_free_lock(vmd); 341 ret = vm_phys_unfree_page(m); 342 vm_domain_free_unlock(vmd); 343 if (ret == TRUE) { 344 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 345 if (bootverbose) 346 printf("Skipping page with pa 0x%jx\n", 347 (uintmax_t)pa); 348 } 349 } 350 } 351 352 /* 353 * vm_page_blacklist_load: 354 * 355 * Search for a special module named "ram_blacklist". It'll be a 356 * plain text file provided by the user via the loader directive 357 * of the same name. 358 */ 359 static void 360 vm_page_blacklist_load(char **list, char **end) 361 { 362 void *mod; 363 u_char *ptr; 364 u_int len; 365 366 mod = NULL; 367 ptr = NULL; 368 369 mod = preload_search_by_type("ram_blacklist"); 370 if (mod != NULL) { 371 ptr = preload_fetch_addr(mod); 372 len = preload_fetch_size(mod); 373 } 374 *list = ptr; 375 if (ptr != NULL) 376 *end = ptr + len; 377 else 378 *end = NULL; 379 return; 380 } 381 382 static int 383 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 384 { 385 vm_page_t m; 386 struct sbuf sbuf; 387 int error, first; 388 389 first = 1; 390 error = sysctl_wire_old_buffer(req, 0); 391 if (error != 0) 392 return (error); 393 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 394 TAILQ_FOREACH(m, &blacklist_head, listq) { 395 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 396 (uintmax_t)m->phys_addr); 397 first = 0; 398 } 399 error = sbuf_finish(&sbuf); 400 sbuf_delete(&sbuf); 401 return (error); 402 } 403 404 static void 405 vm_page_domain_init(int domain) 406 { 407 struct vm_domain *vmd; 408 struct vm_pagequeue *pq; 409 int i; 410 411 vmd = VM_DOMAIN(domain); 412 bzero(vmd, sizeof(*vmd)); 413 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 414 "vm inactive pagequeue"; 415 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 416 "vm active pagequeue"; 417 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 418 "vm laundry pagequeue"; 419 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 420 "vm unswappable pagequeue"; 421 vmd->vmd_domain = domain; 422 vmd->vmd_page_count = 0; 423 vmd->vmd_free_count = 0; 424 vmd->vmd_segs = 0; 425 vmd->vmd_oom = FALSE; 426 for (i = 0; i < PQ_COUNT; i++) { 427 pq = &vmd->vmd_pagequeues[i]; 428 TAILQ_INIT(&pq->pq_pl); 429 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 430 MTX_DEF | MTX_DUPOK); 431 } 432 mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); 433 } 434 435 /* 436 * Initialize a physical page in preparation for adding it to the free 437 * lists. 438 */ 439 static void 440 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 441 { 442 443 m->object = NULL; 444 m->wire_count = 0; 445 m->busy_lock = VPB_UNBUSIED; 446 m->hold_count = 0; 447 m->flags = 0; 448 m->phys_addr = pa; 449 m->queue = PQ_NONE; 450 m->psind = 0; 451 m->segind = segind; 452 m->order = VM_NFREEORDER; 453 m->pool = VM_FREEPOOL_DEFAULT; 454 m->valid = m->dirty = 0; 455 pmap_page_init(m); 456 } 457 458 /* 459 * vm_page_startup: 460 * 461 * Initializes the resident memory module. Allocates physical memory for 462 * bootstrapping UMA and some data structures that are used to manage 463 * physical pages. Initializes these structures, and populates the free 464 * page queues. 465 */ 466 vm_offset_t 467 vm_page_startup(vm_offset_t vaddr) 468 { 469 struct vm_phys_seg *seg; 470 vm_page_t m; 471 char *list, *listend; 472 vm_offset_t mapped; 473 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; 474 vm_paddr_t biggestsize, last_pa, pa; 475 u_long pagecount; 476 int biggestone, i, segind; 477 478 biggestsize = 0; 479 biggestone = 0; 480 vaddr = round_page(vaddr); 481 482 for (i = 0; phys_avail[i + 1]; i += 2) { 483 phys_avail[i] = round_page(phys_avail[i]); 484 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 485 } 486 for (i = 0; phys_avail[i + 1]; i += 2) { 487 size = phys_avail[i + 1] - phys_avail[i]; 488 if (size > biggestsize) { 489 biggestone = i; 490 biggestsize = size; 491 } 492 } 493 494 end = phys_avail[biggestone+1]; 495 496 /* 497 * Initialize the page and queue locks. 498 */ 499 mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); 500 for (i = 0; i < PA_LOCK_COUNT; i++) 501 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 502 for (i = 0; i < vm_ndomains; i++) 503 vm_page_domain_init(i); 504 505 /* 506 * Allocate memory for use when boot strapping the kernel memory 507 * allocator. Tell UMA how many zones we are going to create 508 * before going fully functional. UMA will add its zones. 509 * 510 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, 511 * KMAP ENTRY, MAP ENTRY, VMSPACE. 512 */ 513 boot_pages = uma_startup_count(8); 514 515 #ifndef UMA_MD_SMALL_ALLOC 516 /* vmem_startup() calls uma_prealloc(). */ 517 boot_pages += vmem_startup_count(); 518 /* vm_map_startup() calls uma_prealloc(). */ 519 boot_pages += howmany(MAX_KMAP, 520 UMA_SLAB_SPACE / sizeof(struct vm_map)); 521 522 /* 523 * Before going fully functional kmem_init() does allocation 524 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". 525 */ 526 boot_pages += 2; 527 #endif 528 /* 529 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 530 * manually fetch the value. 531 */ 532 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 533 new_end = end - (boot_pages * UMA_SLAB_SIZE); 534 new_end = trunc_page(new_end); 535 mapped = pmap_map(&vaddr, new_end, end, 536 VM_PROT_READ | VM_PROT_WRITE); 537 bzero((void *)mapped, end - new_end); 538 uma_startup((void *)mapped, boot_pages); 539 540 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 541 defined(__i386__) || defined(__mips__) 542 /* 543 * Allocate a bitmap to indicate that a random physical page 544 * needs to be included in a minidump. 545 * 546 * The amd64 port needs this to indicate which direct map pages 547 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 548 * 549 * However, i386 still needs this workspace internally within the 550 * minidump code. In theory, they are not needed on i386, but are 551 * included should the sf_buf code decide to use them. 552 */ 553 last_pa = 0; 554 for (i = 0; dump_avail[i + 1] != 0; i += 2) 555 if (dump_avail[i + 1] > last_pa) 556 last_pa = dump_avail[i + 1]; 557 page_range = last_pa / PAGE_SIZE; 558 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 559 new_end -= vm_page_dump_size; 560 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 561 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 562 bzero((void *)vm_page_dump, vm_page_dump_size); 563 #else 564 (void)last_pa; 565 #endif 566 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 567 /* 568 * Include the UMA bootstrap pages and vm_page_dump in a crash dump. 569 * When pmap_map() uses the direct map, they are not automatically 570 * included. 571 */ 572 for (pa = new_end; pa < end; pa += PAGE_SIZE) 573 dump_add_page(pa); 574 #endif 575 phys_avail[biggestone + 1] = new_end; 576 #ifdef __amd64__ 577 /* 578 * Request that the physical pages underlying the message buffer be 579 * included in a crash dump. Since the message buffer is accessed 580 * through the direct map, they are not automatically included. 581 */ 582 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 583 last_pa = pa + round_page(msgbufsize); 584 while (pa < last_pa) { 585 dump_add_page(pa); 586 pa += PAGE_SIZE; 587 } 588 #endif 589 /* 590 * Compute the number of pages of memory that will be available for 591 * use, taking into account the overhead of a page structure per page. 592 * In other words, solve 593 * "available physical memory" - round_page(page_range * 594 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 595 * for page_range. 596 */ 597 low_avail = phys_avail[0]; 598 high_avail = phys_avail[1]; 599 for (i = 0; i < vm_phys_nsegs; i++) { 600 if (vm_phys_segs[i].start < low_avail) 601 low_avail = vm_phys_segs[i].start; 602 if (vm_phys_segs[i].end > high_avail) 603 high_avail = vm_phys_segs[i].end; 604 } 605 /* Skip the first chunk. It is already accounted for. */ 606 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 607 if (phys_avail[i] < low_avail) 608 low_avail = phys_avail[i]; 609 if (phys_avail[i + 1] > high_avail) 610 high_avail = phys_avail[i + 1]; 611 } 612 first_page = low_avail / PAGE_SIZE; 613 #ifdef VM_PHYSSEG_SPARSE 614 size = 0; 615 for (i = 0; i < vm_phys_nsegs; i++) 616 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 617 for (i = 0; phys_avail[i + 1] != 0; i += 2) 618 size += phys_avail[i + 1] - phys_avail[i]; 619 #elif defined(VM_PHYSSEG_DENSE) 620 size = high_avail - low_avail; 621 #else 622 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 623 #endif 624 625 #ifdef VM_PHYSSEG_DENSE 626 /* 627 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 628 * the overhead of a page structure per page only if vm_page_array is 629 * allocated from the last physical memory chunk. Otherwise, we must 630 * allocate page structures representing the physical memory 631 * underlying vm_page_array, even though they will not be used. 632 */ 633 if (new_end != high_avail) 634 page_range = size / PAGE_SIZE; 635 else 636 #endif 637 { 638 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 639 640 /* 641 * If the partial bytes remaining are large enough for 642 * a page (PAGE_SIZE) without a corresponding 643 * 'struct vm_page', then new_end will contain an 644 * extra page after subtracting the length of the VM 645 * page array. Compensate by subtracting an extra 646 * page from new_end. 647 */ 648 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 649 if (new_end == high_avail) 650 high_avail -= PAGE_SIZE; 651 new_end -= PAGE_SIZE; 652 } 653 } 654 end = new_end; 655 656 /* 657 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 658 * However, because this page is allocated from KVM, out-of-bounds 659 * accesses using the direct map will not be trapped. 660 */ 661 vaddr += PAGE_SIZE; 662 663 /* 664 * Allocate physical memory for the page structures, and map it. 665 */ 666 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 667 mapped = pmap_map(&vaddr, new_end, end, 668 VM_PROT_READ | VM_PROT_WRITE); 669 vm_page_array = (vm_page_t)mapped; 670 vm_page_array_size = page_range; 671 672 #if VM_NRESERVLEVEL > 0 673 /* 674 * Allocate physical memory for the reservation management system's 675 * data structures, and map it. 676 */ 677 if (high_avail == end) 678 high_avail = new_end; 679 new_end = vm_reserv_startup(&vaddr, new_end, high_avail); 680 #endif 681 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 682 /* 683 * Include vm_page_array and vm_reserv_array in a crash dump. 684 */ 685 for (pa = new_end; pa < end; pa += PAGE_SIZE) 686 dump_add_page(pa); 687 #endif 688 phys_avail[biggestone + 1] = new_end; 689 690 /* 691 * Add physical memory segments corresponding to the available 692 * physical pages. 693 */ 694 for (i = 0; phys_avail[i + 1] != 0; i += 2) 695 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 696 697 /* 698 * Initialize the physical memory allocator. 699 */ 700 vm_phys_init(); 701 702 /* 703 * Initialize the page structures and add every available page to the 704 * physical memory allocator's free lists. 705 */ 706 vm_cnt.v_page_count = 0; 707 for (segind = 0; segind < vm_phys_nsegs; segind++) { 708 seg = &vm_phys_segs[segind]; 709 for (m = seg->first_page, pa = seg->start; pa < seg->end; 710 m++, pa += PAGE_SIZE) 711 vm_page_init_page(m, pa, segind); 712 713 /* 714 * Add the segment to the free lists only if it is covered by 715 * one of the ranges in phys_avail. Because we've added the 716 * ranges to the vm_phys_segs array, we can assume that each 717 * segment is either entirely contained in one of the ranges, 718 * or doesn't overlap any of them. 719 */ 720 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 721 struct vm_domain *vmd; 722 723 if (seg->start < phys_avail[i] || 724 seg->end > phys_avail[i + 1]) 725 continue; 726 727 m = seg->first_page; 728 pagecount = (u_long)atop(seg->end - seg->start); 729 730 vmd = VM_DOMAIN(seg->domain); 731 vm_domain_free_lock(vmd); 732 vm_phys_free_contig(m, pagecount); 733 vm_domain_freecnt_adj(vmd, (int)pagecount); 734 vm_domain_free_unlock(vmd); 735 vm_cnt.v_page_count += (u_int)pagecount; 736 737 vmd = VM_DOMAIN(seg->domain); 738 vmd->vmd_page_count += (u_int)pagecount; 739 vmd->vmd_segs |= 1UL << m->segind; 740 break; 741 } 742 } 743 744 /* 745 * Remove blacklisted pages from the physical memory allocator. 746 */ 747 TAILQ_INIT(&blacklist_head); 748 vm_page_blacklist_load(&list, &listend); 749 vm_page_blacklist_check(list, listend); 750 751 list = kern_getenv("vm.blacklist"); 752 vm_page_blacklist_check(list, NULL); 753 754 freeenv(list); 755 #if VM_NRESERVLEVEL > 0 756 /* 757 * Initialize the reservation management system. 758 */ 759 vm_reserv_init(); 760 #endif 761 /* 762 * Set an initial domain policy for thread0 so that allocations 763 * can work. 764 */ 765 domainset_zero(); 766 767 return (vaddr); 768 } 769 770 void 771 vm_page_reference(vm_page_t m) 772 { 773 774 vm_page_aflag_set(m, PGA_REFERENCED); 775 } 776 777 /* 778 * vm_page_busy_downgrade: 779 * 780 * Downgrade an exclusive busy page into a single shared busy page. 781 */ 782 void 783 vm_page_busy_downgrade(vm_page_t m) 784 { 785 u_int x; 786 bool locked; 787 788 vm_page_assert_xbusied(m); 789 locked = mtx_owned(vm_page_lockptr(m)); 790 791 for (;;) { 792 x = m->busy_lock; 793 x &= VPB_BIT_WAITERS; 794 if (x != 0 && !locked) 795 vm_page_lock(m); 796 if (atomic_cmpset_rel_int(&m->busy_lock, 797 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 798 break; 799 if (x != 0 && !locked) 800 vm_page_unlock(m); 801 } 802 if (x != 0) { 803 wakeup(m); 804 if (!locked) 805 vm_page_unlock(m); 806 } 807 } 808 809 /* 810 * vm_page_sbusied: 811 * 812 * Return a positive value if the page is shared busied, 0 otherwise. 813 */ 814 int 815 vm_page_sbusied(vm_page_t m) 816 { 817 u_int x; 818 819 x = m->busy_lock; 820 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 821 } 822 823 /* 824 * vm_page_sunbusy: 825 * 826 * Shared unbusy a page. 827 */ 828 void 829 vm_page_sunbusy(vm_page_t m) 830 { 831 u_int x; 832 833 vm_page_lock_assert(m, MA_NOTOWNED); 834 vm_page_assert_sbusied(m); 835 836 for (;;) { 837 x = m->busy_lock; 838 if (VPB_SHARERS(x) > 1) { 839 if (atomic_cmpset_int(&m->busy_lock, x, 840 x - VPB_ONE_SHARER)) 841 break; 842 continue; 843 } 844 if ((x & VPB_BIT_WAITERS) == 0) { 845 KASSERT(x == VPB_SHARERS_WORD(1), 846 ("vm_page_sunbusy: invalid lock state")); 847 if (atomic_cmpset_int(&m->busy_lock, 848 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 849 break; 850 continue; 851 } 852 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 853 ("vm_page_sunbusy: invalid lock state for waiters")); 854 855 vm_page_lock(m); 856 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 857 vm_page_unlock(m); 858 continue; 859 } 860 wakeup(m); 861 vm_page_unlock(m); 862 break; 863 } 864 } 865 866 /* 867 * vm_page_busy_sleep: 868 * 869 * Sleep and release the page lock, using the page pointer as wchan. 870 * This is used to implement the hard-path of busying mechanism. 871 * 872 * The given page must be locked. 873 * 874 * If nonshared is true, sleep only if the page is xbusy. 875 */ 876 void 877 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 878 { 879 u_int x; 880 881 vm_page_assert_locked(m); 882 883 x = m->busy_lock; 884 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 885 ((x & VPB_BIT_WAITERS) == 0 && 886 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 887 vm_page_unlock(m); 888 return; 889 } 890 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 891 } 892 893 /* 894 * vm_page_trysbusy: 895 * 896 * Try to shared busy a page. 897 * If the operation succeeds 1 is returned otherwise 0. 898 * The operation never sleeps. 899 */ 900 int 901 vm_page_trysbusy(vm_page_t m) 902 { 903 u_int x; 904 905 for (;;) { 906 x = m->busy_lock; 907 if ((x & VPB_BIT_SHARED) == 0) 908 return (0); 909 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 910 return (1); 911 } 912 } 913 914 static void 915 vm_page_xunbusy_locked(vm_page_t m) 916 { 917 918 vm_page_assert_xbusied(m); 919 vm_page_assert_locked(m); 920 921 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 922 /* There is a waiter, do wakeup() instead of vm_page_flash(). */ 923 wakeup(m); 924 } 925 926 void 927 vm_page_xunbusy_maybelocked(vm_page_t m) 928 { 929 bool lockacq; 930 931 vm_page_assert_xbusied(m); 932 933 /* 934 * Fast path for unbusy. If it succeeds, we know that there 935 * are no waiters, so we do not need a wakeup. 936 */ 937 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER, 938 VPB_UNBUSIED)) 939 return; 940 941 lockacq = !mtx_owned(vm_page_lockptr(m)); 942 if (lockacq) 943 vm_page_lock(m); 944 vm_page_xunbusy_locked(m); 945 if (lockacq) 946 vm_page_unlock(m); 947 } 948 949 /* 950 * vm_page_xunbusy_hard: 951 * 952 * Called after the first try the exclusive unbusy of a page failed. 953 * It is assumed that the waiters bit is on. 954 */ 955 void 956 vm_page_xunbusy_hard(vm_page_t m) 957 { 958 959 vm_page_assert_xbusied(m); 960 961 vm_page_lock(m); 962 vm_page_xunbusy_locked(m); 963 vm_page_unlock(m); 964 } 965 966 /* 967 * vm_page_flash: 968 * 969 * Wakeup anyone waiting for the page. 970 * The ownership bits do not change. 971 * 972 * The given page must be locked. 973 */ 974 void 975 vm_page_flash(vm_page_t m) 976 { 977 u_int x; 978 979 vm_page_lock_assert(m, MA_OWNED); 980 981 for (;;) { 982 x = m->busy_lock; 983 if ((x & VPB_BIT_WAITERS) == 0) 984 return; 985 if (atomic_cmpset_int(&m->busy_lock, x, 986 x & (~VPB_BIT_WAITERS))) 987 break; 988 } 989 wakeup(m); 990 } 991 992 /* 993 * Avoid releasing and reacquiring the same page lock. 994 */ 995 void 996 vm_page_change_lock(vm_page_t m, struct mtx **mtx) 997 { 998 struct mtx *mtx1; 999 1000 mtx1 = vm_page_lockptr(m); 1001 if (*mtx == mtx1) 1002 return; 1003 if (*mtx != NULL) 1004 mtx_unlock(*mtx); 1005 *mtx = mtx1; 1006 mtx_lock(mtx1); 1007 } 1008 1009 /* 1010 * Keep page from being freed by the page daemon 1011 * much of the same effect as wiring, except much lower 1012 * overhead and should be used only for *very* temporary 1013 * holding ("wiring"). 1014 */ 1015 void 1016 vm_page_hold(vm_page_t mem) 1017 { 1018 1019 vm_page_lock_assert(mem, MA_OWNED); 1020 mem->hold_count++; 1021 } 1022 1023 void 1024 vm_page_unhold(vm_page_t mem) 1025 { 1026 1027 vm_page_lock_assert(mem, MA_OWNED); 1028 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 1029 --mem->hold_count; 1030 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 1031 vm_page_free_toq(mem); 1032 } 1033 1034 /* 1035 * vm_page_unhold_pages: 1036 * 1037 * Unhold each of the pages that is referenced by the given array. 1038 */ 1039 void 1040 vm_page_unhold_pages(vm_page_t *ma, int count) 1041 { 1042 struct mtx *mtx; 1043 1044 mtx = NULL; 1045 for (; count != 0; count--) { 1046 vm_page_change_lock(*ma, &mtx); 1047 vm_page_unhold(*ma); 1048 ma++; 1049 } 1050 if (mtx != NULL) 1051 mtx_unlock(mtx); 1052 } 1053 1054 vm_page_t 1055 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1056 { 1057 vm_page_t m; 1058 1059 #ifdef VM_PHYSSEG_SPARSE 1060 m = vm_phys_paddr_to_vm_page(pa); 1061 if (m == NULL) 1062 m = vm_phys_fictitious_to_vm_page(pa); 1063 return (m); 1064 #elif defined(VM_PHYSSEG_DENSE) 1065 long pi; 1066 1067 pi = atop(pa); 1068 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1069 m = &vm_page_array[pi - first_page]; 1070 return (m); 1071 } 1072 return (vm_phys_fictitious_to_vm_page(pa)); 1073 #else 1074 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1075 #endif 1076 } 1077 1078 /* 1079 * vm_page_getfake: 1080 * 1081 * Create a fictitious page with the specified physical address and 1082 * memory attribute. The memory attribute is the only the machine- 1083 * dependent aspect of a fictitious page that must be initialized. 1084 */ 1085 vm_page_t 1086 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1087 { 1088 vm_page_t m; 1089 1090 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1091 vm_page_initfake(m, paddr, memattr); 1092 return (m); 1093 } 1094 1095 void 1096 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1097 { 1098 1099 if ((m->flags & PG_FICTITIOUS) != 0) { 1100 /* 1101 * The page's memattr might have changed since the 1102 * previous initialization. Update the pmap to the 1103 * new memattr. 1104 */ 1105 goto memattr; 1106 } 1107 m->phys_addr = paddr; 1108 m->queue = PQ_NONE; 1109 /* Fictitious pages don't use "segind". */ 1110 m->flags = PG_FICTITIOUS; 1111 /* Fictitious pages don't use "order" or "pool". */ 1112 m->oflags = VPO_UNMANAGED; 1113 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1114 m->wire_count = 1; 1115 pmap_page_init(m); 1116 memattr: 1117 pmap_page_set_memattr(m, memattr); 1118 } 1119 1120 /* 1121 * vm_page_putfake: 1122 * 1123 * Release a fictitious page. 1124 */ 1125 void 1126 vm_page_putfake(vm_page_t m) 1127 { 1128 1129 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1130 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1131 ("vm_page_putfake: bad page %p", m)); 1132 uma_zfree(fakepg_zone, m); 1133 } 1134 1135 /* 1136 * vm_page_updatefake: 1137 * 1138 * Update the given fictitious page to the specified physical address and 1139 * memory attribute. 1140 */ 1141 void 1142 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1143 { 1144 1145 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1146 ("vm_page_updatefake: bad page %p", m)); 1147 m->phys_addr = paddr; 1148 pmap_page_set_memattr(m, memattr); 1149 } 1150 1151 /* 1152 * vm_page_free: 1153 * 1154 * Free a page. 1155 */ 1156 void 1157 vm_page_free(vm_page_t m) 1158 { 1159 1160 m->flags &= ~PG_ZERO; 1161 vm_page_free_toq(m); 1162 } 1163 1164 /* 1165 * vm_page_free_zero: 1166 * 1167 * Free a page to the zerod-pages queue 1168 */ 1169 void 1170 vm_page_free_zero(vm_page_t m) 1171 { 1172 1173 m->flags |= PG_ZERO; 1174 vm_page_free_toq(m); 1175 } 1176 1177 /* 1178 * Unbusy and handle the page queueing for a page from a getpages request that 1179 * was optionally read ahead or behind. 1180 */ 1181 void 1182 vm_page_readahead_finish(vm_page_t m) 1183 { 1184 1185 /* We shouldn't put invalid pages on queues. */ 1186 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); 1187 1188 /* 1189 * Since the page is not the actually needed one, whether it should 1190 * be activated or deactivated is not obvious. Empirical results 1191 * have shown that deactivating the page is usually the best choice, 1192 * unless the page is wanted by another thread. 1193 */ 1194 vm_page_lock(m); 1195 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1196 vm_page_activate(m); 1197 else 1198 vm_page_deactivate(m); 1199 vm_page_unlock(m); 1200 vm_page_xunbusy(m); 1201 } 1202 1203 /* 1204 * vm_page_sleep_if_busy: 1205 * 1206 * Sleep and release the page queues lock if the page is busied. 1207 * Returns TRUE if the thread slept. 1208 * 1209 * The given page must be unlocked and object containing it must 1210 * be locked. 1211 */ 1212 int 1213 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1214 { 1215 vm_object_t obj; 1216 1217 vm_page_lock_assert(m, MA_NOTOWNED); 1218 VM_OBJECT_ASSERT_WLOCKED(m->object); 1219 1220 if (vm_page_busied(m)) { 1221 /* 1222 * The page-specific object must be cached because page 1223 * identity can change during the sleep, causing the 1224 * re-lock of a different object. 1225 * It is assumed that a reference to the object is already 1226 * held by the callers. 1227 */ 1228 obj = m->object; 1229 vm_page_lock(m); 1230 VM_OBJECT_WUNLOCK(obj); 1231 vm_page_busy_sleep(m, msg, false); 1232 VM_OBJECT_WLOCK(obj); 1233 return (TRUE); 1234 } 1235 return (FALSE); 1236 } 1237 1238 /* 1239 * vm_page_dirty_KBI: [ internal use only ] 1240 * 1241 * Set all bits in the page's dirty field. 1242 * 1243 * The object containing the specified page must be locked if the 1244 * call is made from the machine-independent layer. 1245 * 1246 * See vm_page_clear_dirty_mask(). 1247 * 1248 * This function should only be called by vm_page_dirty(). 1249 */ 1250 void 1251 vm_page_dirty_KBI(vm_page_t m) 1252 { 1253 1254 /* Refer to this operation by its public name. */ 1255 KASSERT(m->valid == VM_PAGE_BITS_ALL, 1256 ("vm_page_dirty: page is invalid!")); 1257 m->dirty = VM_PAGE_BITS_ALL; 1258 } 1259 1260 /* 1261 * vm_page_insert: [ internal use only ] 1262 * 1263 * Inserts the given mem entry into the object and object list. 1264 * 1265 * The object must be locked. 1266 */ 1267 int 1268 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1269 { 1270 vm_page_t mpred; 1271 1272 VM_OBJECT_ASSERT_WLOCKED(object); 1273 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1274 return (vm_page_insert_after(m, object, pindex, mpred)); 1275 } 1276 1277 /* 1278 * vm_page_insert_after: 1279 * 1280 * Inserts the page "m" into the specified object at offset "pindex". 1281 * 1282 * The page "mpred" must immediately precede the offset "pindex" within 1283 * the specified object. 1284 * 1285 * The object must be locked. 1286 */ 1287 static int 1288 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1289 vm_page_t mpred) 1290 { 1291 vm_page_t msucc; 1292 1293 VM_OBJECT_ASSERT_WLOCKED(object); 1294 KASSERT(m->object == NULL, 1295 ("vm_page_insert_after: page already inserted")); 1296 if (mpred != NULL) { 1297 KASSERT(mpred->object == object, 1298 ("vm_page_insert_after: object doesn't contain mpred")); 1299 KASSERT(mpred->pindex < pindex, 1300 ("vm_page_insert_after: mpred doesn't precede pindex")); 1301 msucc = TAILQ_NEXT(mpred, listq); 1302 } else 1303 msucc = TAILQ_FIRST(&object->memq); 1304 if (msucc != NULL) 1305 KASSERT(msucc->pindex > pindex, 1306 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1307 1308 /* 1309 * Record the object/offset pair in this page 1310 */ 1311 m->object = object; 1312 m->pindex = pindex; 1313 1314 /* 1315 * Now link into the object's ordered list of backed pages. 1316 */ 1317 if (vm_radix_insert(&object->rtree, m)) { 1318 m->object = NULL; 1319 m->pindex = 0; 1320 return (1); 1321 } 1322 vm_page_insert_radixdone(m, object, mpred); 1323 return (0); 1324 } 1325 1326 /* 1327 * vm_page_insert_radixdone: 1328 * 1329 * Complete page "m" insertion into the specified object after the 1330 * radix trie hooking. 1331 * 1332 * The page "mpred" must precede the offset "m->pindex" within the 1333 * specified object. 1334 * 1335 * The object must be locked. 1336 */ 1337 static void 1338 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1339 { 1340 1341 VM_OBJECT_ASSERT_WLOCKED(object); 1342 KASSERT(object != NULL && m->object == object, 1343 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1344 if (mpred != NULL) { 1345 KASSERT(mpred->object == object, 1346 ("vm_page_insert_after: object doesn't contain mpred")); 1347 KASSERT(mpred->pindex < m->pindex, 1348 ("vm_page_insert_after: mpred doesn't precede pindex")); 1349 } 1350 1351 if (mpred != NULL) 1352 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1353 else 1354 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1355 1356 /* 1357 * Show that the object has one more resident page. 1358 */ 1359 object->resident_page_count++; 1360 1361 /* 1362 * Hold the vnode until the last page is released. 1363 */ 1364 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1365 vhold(object->handle); 1366 1367 /* 1368 * Since we are inserting a new and possibly dirty page, 1369 * update the object's OBJ_MIGHTBEDIRTY flag. 1370 */ 1371 if (pmap_page_is_write_mapped(m)) 1372 vm_object_set_writeable_dirty(object); 1373 } 1374 1375 /* 1376 * vm_page_remove: 1377 * 1378 * Removes the specified page from its containing object, but does not 1379 * invalidate any backing storage. 1380 * 1381 * The object must be locked. The page must be locked if it is managed. 1382 */ 1383 void 1384 vm_page_remove(vm_page_t m) 1385 { 1386 vm_object_t object; 1387 vm_page_t mrem; 1388 1389 if ((m->oflags & VPO_UNMANAGED) == 0) 1390 vm_page_assert_locked(m); 1391 if ((object = m->object) == NULL) 1392 return; 1393 VM_OBJECT_ASSERT_WLOCKED(object); 1394 if (vm_page_xbusied(m)) 1395 vm_page_xunbusy_maybelocked(m); 1396 mrem = vm_radix_remove(&object->rtree, m->pindex); 1397 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1398 1399 /* 1400 * Now remove from the object's list of backed pages. 1401 */ 1402 TAILQ_REMOVE(&object->memq, m, listq); 1403 1404 /* 1405 * And show that the object has one fewer resident page. 1406 */ 1407 object->resident_page_count--; 1408 1409 /* 1410 * The vnode may now be recycled. 1411 */ 1412 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1413 vdrop(object->handle); 1414 1415 m->object = NULL; 1416 } 1417 1418 /* 1419 * vm_page_lookup: 1420 * 1421 * Returns the page associated with the object/offset 1422 * pair specified; if none is found, NULL is returned. 1423 * 1424 * The object must be locked. 1425 */ 1426 vm_page_t 1427 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1428 { 1429 1430 VM_OBJECT_ASSERT_LOCKED(object); 1431 return (vm_radix_lookup(&object->rtree, pindex)); 1432 } 1433 1434 /* 1435 * vm_page_find_least: 1436 * 1437 * Returns the page associated with the object with least pindex 1438 * greater than or equal to the parameter pindex, or NULL. 1439 * 1440 * The object must be locked. 1441 */ 1442 vm_page_t 1443 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1444 { 1445 vm_page_t m; 1446 1447 VM_OBJECT_ASSERT_LOCKED(object); 1448 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1449 m = vm_radix_lookup_ge(&object->rtree, pindex); 1450 return (m); 1451 } 1452 1453 /* 1454 * Returns the given page's successor (by pindex) within the object if it is 1455 * resident; if none is found, NULL is returned. 1456 * 1457 * The object must be locked. 1458 */ 1459 vm_page_t 1460 vm_page_next(vm_page_t m) 1461 { 1462 vm_page_t next; 1463 1464 VM_OBJECT_ASSERT_LOCKED(m->object); 1465 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1466 MPASS(next->object == m->object); 1467 if (next->pindex != m->pindex + 1) 1468 next = NULL; 1469 } 1470 return (next); 1471 } 1472 1473 /* 1474 * Returns the given page's predecessor (by pindex) within the object if it is 1475 * resident; if none is found, NULL is returned. 1476 * 1477 * The object must be locked. 1478 */ 1479 vm_page_t 1480 vm_page_prev(vm_page_t m) 1481 { 1482 vm_page_t prev; 1483 1484 VM_OBJECT_ASSERT_LOCKED(m->object); 1485 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1486 MPASS(prev->object == m->object); 1487 if (prev->pindex != m->pindex - 1) 1488 prev = NULL; 1489 } 1490 return (prev); 1491 } 1492 1493 /* 1494 * Uses the page mnew as a replacement for an existing page at index 1495 * pindex which must be already present in the object. 1496 * 1497 * The existing page must not be on a paging queue. 1498 */ 1499 vm_page_t 1500 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1501 { 1502 vm_page_t mold; 1503 1504 VM_OBJECT_ASSERT_WLOCKED(object); 1505 KASSERT(mnew->object == NULL, 1506 ("vm_page_replace: page already in object")); 1507 1508 /* 1509 * This function mostly follows vm_page_insert() and 1510 * vm_page_remove() without the radix, object count and vnode 1511 * dance. Double check such functions for more comments. 1512 */ 1513 1514 mnew->object = object; 1515 mnew->pindex = pindex; 1516 mold = vm_radix_replace(&object->rtree, mnew); 1517 KASSERT(mold->queue == PQ_NONE, 1518 ("vm_page_replace: mold is on a paging queue")); 1519 1520 /* Keep the resident page list in sorted order. */ 1521 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1522 TAILQ_REMOVE(&object->memq, mold, listq); 1523 1524 mold->object = NULL; 1525 vm_page_xunbusy_maybelocked(mold); 1526 1527 /* 1528 * The object's resident_page_count does not change because we have 1529 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1530 */ 1531 if (pmap_page_is_write_mapped(mnew)) 1532 vm_object_set_writeable_dirty(object); 1533 return (mold); 1534 } 1535 1536 /* 1537 * vm_page_rename: 1538 * 1539 * Move the given memory entry from its 1540 * current object to the specified target object/offset. 1541 * 1542 * Note: swap associated with the page must be invalidated by the move. We 1543 * have to do this for several reasons: (1) we aren't freeing the 1544 * page, (2) we are dirtying the page, (3) the VM system is probably 1545 * moving the page from object A to B, and will then later move 1546 * the backing store from A to B and we can't have a conflict. 1547 * 1548 * Note: we *always* dirty the page. It is necessary both for the 1549 * fact that we moved it, and because we may be invalidating 1550 * swap. 1551 * 1552 * The objects must be locked. 1553 */ 1554 int 1555 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1556 { 1557 vm_page_t mpred; 1558 vm_pindex_t opidx; 1559 1560 VM_OBJECT_ASSERT_WLOCKED(new_object); 1561 1562 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1563 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1564 ("vm_page_rename: pindex already renamed")); 1565 1566 /* 1567 * Create a custom version of vm_page_insert() which does not depend 1568 * by m_prev and can cheat on the implementation aspects of the 1569 * function. 1570 */ 1571 opidx = m->pindex; 1572 m->pindex = new_pindex; 1573 if (vm_radix_insert(&new_object->rtree, m)) { 1574 m->pindex = opidx; 1575 return (1); 1576 } 1577 1578 /* 1579 * The operation cannot fail anymore. The removal must happen before 1580 * the listq iterator is tainted. 1581 */ 1582 m->pindex = opidx; 1583 vm_page_lock(m); 1584 vm_page_remove(m); 1585 1586 /* Return back to the new pindex to complete vm_page_insert(). */ 1587 m->pindex = new_pindex; 1588 m->object = new_object; 1589 vm_page_unlock(m); 1590 vm_page_insert_radixdone(m, new_object, mpred); 1591 vm_page_dirty(m); 1592 return (0); 1593 } 1594 1595 /* 1596 * vm_page_alloc: 1597 * 1598 * Allocate and return a page that is associated with the specified 1599 * object and offset pair. By default, this page is exclusive busied. 1600 * 1601 * The caller must always specify an allocation class. 1602 * 1603 * allocation classes: 1604 * VM_ALLOC_NORMAL normal process request 1605 * VM_ALLOC_SYSTEM system *really* needs a page 1606 * VM_ALLOC_INTERRUPT interrupt time request 1607 * 1608 * optional allocation flags: 1609 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1610 * intends to allocate 1611 * VM_ALLOC_NOBUSY do not exclusive busy the page 1612 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1613 * VM_ALLOC_NOOBJ page is not associated with an object and 1614 * should not be exclusive busy 1615 * VM_ALLOC_SBUSY shared busy the allocated page 1616 * VM_ALLOC_WIRED wire the allocated page 1617 * VM_ALLOC_ZERO prefer a zeroed page 1618 */ 1619 vm_page_t 1620 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1621 { 1622 1623 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1624 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1625 } 1626 1627 vm_page_t 1628 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1629 int req) 1630 { 1631 1632 return (vm_page_alloc_domain_after(object, pindex, domain, req, 1633 object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : 1634 NULL)); 1635 } 1636 1637 /* 1638 * Allocate a page in the specified object with the given page index. To 1639 * optimize insertion of the page into the object, the caller must also specifiy 1640 * the resident page in the object with largest index smaller than the given 1641 * page index, or NULL if no such page exists. 1642 */ 1643 vm_page_t 1644 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, 1645 int req, vm_page_t mpred) 1646 { 1647 struct vm_domainset_iter di; 1648 vm_page_t m; 1649 int domain; 1650 1651 vm_domainset_iter_page_init(&di, object, &domain, &req); 1652 do { 1653 m = vm_page_alloc_domain_after(object, pindex, domain, req, 1654 mpred); 1655 if (m != NULL) 1656 break; 1657 } while (vm_domainset_iter_page(&di, &domain, &req) == 0); 1658 1659 return (m); 1660 } 1661 1662 /* 1663 * Returns true if the number of free pages exceeds the minimum 1664 * for the request class and false otherwise. 1665 */ 1666 int 1667 vm_domain_available(struct vm_domain *vmd, int req, int npages) 1668 { 1669 1670 vm_domain_free_assert_locked(vmd); 1671 req = req & VM_ALLOC_CLASS_MASK; 1672 1673 /* 1674 * The page daemon is allowed to dig deeper into the free page list. 1675 */ 1676 if (curproc == pageproc && req != VM_ALLOC_INTERRUPT) 1677 req = VM_ALLOC_SYSTEM; 1678 1679 if (vmd->vmd_free_count >= npages + vmd->vmd_free_reserved || 1680 (req == VM_ALLOC_SYSTEM && 1681 vmd->vmd_free_count >= npages + vmd->vmd_interrupt_free_min) || 1682 (req == VM_ALLOC_INTERRUPT && 1683 vmd->vmd_free_count >= npages)) 1684 return (1); 1685 1686 return (0); 1687 } 1688 1689 vm_page_t 1690 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, 1691 int req, vm_page_t mpred) 1692 { 1693 struct vm_domain *vmd; 1694 vm_page_t m; 1695 int flags; 1696 u_int free_count; 1697 1698 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1699 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1700 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1701 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1702 ("inconsistent object(%p)/req(%x)", object, req)); 1703 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1704 ("Can't sleep and retry object insertion.")); 1705 KASSERT(mpred == NULL || mpred->pindex < pindex, 1706 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1707 (uintmax_t)pindex)); 1708 if (object != NULL) 1709 VM_OBJECT_ASSERT_WLOCKED(object); 1710 1711 again: 1712 m = NULL; 1713 #if VM_NRESERVLEVEL > 0 1714 if (vm_object_reserv(object) && 1715 (m = vm_reserv_extend(req, object, pindex, domain, mpred)) 1716 != NULL) { 1717 domain = vm_phys_domain(m); 1718 vmd = VM_DOMAIN(domain); 1719 goto found; 1720 } 1721 #endif 1722 vmd = VM_DOMAIN(domain); 1723 vm_domain_free_lock(vmd); 1724 if (vm_domain_available(vmd, req, 1)) { 1725 /* 1726 * Can we allocate the page from a reservation? 1727 */ 1728 #if VM_NRESERVLEVEL > 0 1729 if (!vm_object_reserv(object) || 1730 (m = vm_reserv_alloc_page(object, pindex, 1731 domain, mpred)) == NULL) 1732 #endif 1733 { 1734 /* 1735 * If not, allocate it from the free page queues. 1736 */ 1737 m = vm_phys_alloc_pages(domain, object != NULL ? 1738 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1739 #if VM_NRESERVLEVEL > 0 1740 if (m == NULL && vm_reserv_reclaim_inactive(domain)) { 1741 m = vm_phys_alloc_pages(domain, 1742 object != NULL ? 1743 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1744 0); 1745 } 1746 #endif 1747 } 1748 } 1749 if (m == NULL) { 1750 /* 1751 * Not allocatable, give up. 1752 */ 1753 if (vm_domain_alloc_fail(vmd, object, req)) 1754 goto again; 1755 return (NULL); 1756 } 1757 1758 /* 1759 * At this point we had better have found a good page. 1760 */ 1761 KASSERT(m != NULL, ("missing page")); 1762 free_count = vm_domain_freecnt_adj(vmd, -1); 1763 vm_domain_free_unlock(vmd); 1764 1765 /* 1766 * Don't wakeup too often - wakeup the pageout daemon when 1767 * we would be nearly out of memory. 1768 */ 1769 if (vm_paging_needed(vmd, free_count)) 1770 pagedaemon_wakeup(vmd->vmd_domain); 1771 #if VM_NRESERVLEVEL > 0 1772 found: 1773 #endif 1774 vm_page_alloc_check(m); 1775 1776 /* 1777 * Initialize the page. Only the PG_ZERO flag is inherited. 1778 */ 1779 flags = 0; 1780 if ((req & VM_ALLOC_ZERO) != 0) 1781 flags = PG_ZERO; 1782 flags &= m->flags; 1783 if ((req & VM_ALLOC_NODUMP) != 0) 1784 flags |= PG_NODUMP; 1785 m->flags = flags; 1786 m->aflags = 0; 1787 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1788 VPO_UNMANAGED : 0; 1789 m->busy_lock = VPB_UNBUSIED; 1790 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1791 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1792 if ((req & VM_ALLOC_SBUSY) != 0) 1793 m->busy_lock = VPB_SHARERS_WORD(1); 1794 if (req & VM_ALLOC_WIRED) { 1795 /* 1796 * The page lock is not required for wiring a page until that 1797 * page is inserted into the object. 1798 */ 1799 atomic_add_int(&vm_cnt.v_wire_count, 1); 1800 m->wire_count = 1; 1801 } 1802 m->act_count = 0; 1803 1804 if (object != NULL) { 1805 if (vm_page_insert_after(m, object, pindex, mpred)) { 1806 pagedaemon_wakeup(domain); 1807 if (req & VM_ALLOC_WIRED) { 1808 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1809 m->wire_count = 0; 1810 } 1811 KASSERT(m->object == NULL, ("page %p has object", m)); 1812 m->oflags = VPO_UNMANAGED; 1813 m->busy_lock = VPB_UNBUSIED; 1814 /* Don't change PG_ZERO. */ 1815 vm_page_free_toq(m); 1816 if (req & VM_ALLOC_WAITFAIL) { 1817 VM_OBJECT_WUNLOCK(object); 1818 vm_radix_wait(); 1819 VM_OBJECT_WLOCK(object); 1820 } 1821 return (NULL); 1822 } 1823 1824 /* Ignore device objects; the pager sets "memattr" for them. */ 1825 if (object->memattr != VM_MEMATTR_DEFAULT && 1826 (object->flags & OBJ_FICTITIOUS) == 0) 1827 pmap_page_set_memattr(m, object->memattr); 1828 } else 1829 m->pindex = pindex; 1830 1831 return (m); 1832 } 1833 1834 /* 1835 * vm_page_alloc_contig: 1836 * 1837 * Allocate a contiguous set of physical pages of the given size "npages" 1838 * from the free lists. All of the physical pages must be at or above 1839 * the given physical address "low" and below the given physical address 1840 * "high". The given value "alignment" determines the alignment of the 1841 * first physical page in the set. If the given value "boundary" is 1842 * non-zero, then the set of physical pages cannot cross any physical 1843 * address boundary that is a multiple of that value. Both "alignment" 1844 * and "boundary" must be a power of two. 1845 * 1846 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1847 * then the memory attribute setting for the physical pages is configured 1848 * to the object's memory attribute setting. Otherwise, the memory 1849 * attribute setting for the physical pages is configured to "memattr", 1850 * overriding the object's memory attribute setting. However, if the 1851 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1852 * memory attribute setting for the physical pages cannot be configured 1853 * to VM_MEMATTR_DEFAULT. 1854 * 1855 * The specified object may not contain fictitious pages. 1856 * 1857 * The caller must always specify an allocation class. 1858 * 1859 * allocation classes: 1860 * VM_ALLOC_NORMAL normal process request 1861 * VM_ALLOC_SYSTEM system *really* needs a page 1862 * VM_ALLOC_INTERRUPT interrupt time request 1863 * 1864 * optional allocation flags: 1865 * VM_ALLOC_NOBUSY do not exclusive busy the page 1866 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1867 * VM_ALLOC_NOOBJ page is not associated with an object and 1868 * should not be exclusive busy 1869 * VM_ALLOC_SBUSY shared busy the allocated page 1870 * VM_ALLOC_WIRED wire the allocated page 1871 * VM_ALLOC_ZERO prefer a zeroed page 1872 */ 1873 vm_page_t 1874 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1875 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1876 vm_paddr_t boundary, vm_memattr_t memattr) 1877 { 1878 struct vm_domainset_iter di; 1879 vm_page_t m; 1880 int domain; 1881 1882 vm_domainset_iter_page_init(&di, object, &domain, &req); 1883 do { 1884 m = vm_page_alloc_contig_domain(object, pindex, domain, req, 1885 npages, low, high, alignment, boundary, memattr); 1886 if (m != NULL) 1887 break; 1888 } while (vm_domainset_iter_page(&di, &domain, &req) == 0); 1889 1890 return (m); 1891 } 1892 1893 vm_page_t 1894 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, 1895 int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1896 vm_paddr_t boundary, vm_memattr_t memattr) 1897 { 1898 struct vm_domain *vmd; 1899 vm_page_t m, m_ret, mpred; 1900 u_int busy_lock, flags, oflags; 1901 1902 mpred = NULL; /* XXX: pacify gcc */ 1903 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1904 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1905 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1906 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1907 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 1908 req)); 1909 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1910 ("Can't sleep and retry object insertion.")); 1911 if (object != NULL) { 1912 VM_OBJECT_ASSERT_WLOCKED(object); 1913 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 1914 ("vm_page_alloc_contig: object %p has fictitious pages", 1915 object)); 1916 } 1917 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1918 1919 if (object != NULL) { 1920 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1921 KASSERT(mpred == NULL || mpred->pindex != pindex, 1922 ("vm_page_alloc_contig: pindex already allocated")); 1923 } 1924 1925 /* 1926 * Can we allocate the pages without the number of free pages falling 1927 * below the lower bound for the allocation class? 1928 */ 1929 again: 1930 #if VM_NRESERVLEVEL > 0 1931 if (vm_object_reserv(object) && 1932 (m_ret = vm_reserv_extend_contig(req, object, pindex, domain, 1933 npages, low, high, alignment, boundary, mpred)) != NULL) { 1934 domain = vm_phys_domain(m_ret); 1935 vmd = VM_DOMAIN(domain); 1936 goto found; 1937 } 1938 #endif 1939 m_ret = NULL; 1940 vmd = VM_DOMAIN(domain); 1941 vm_domain_free_lock(vmd); 1942 if (vm_domain_available(vmd, req, npages)) { 1943 /* 1944 * Can we allocate the pages from a reservation? 1945 */ 1946 #if VM_NRESERVLEVEL > 0 1947 retry: 1948 if (!vm_object_reserv(object) || 1949 (m_ret = vm_reserv_alloc_contig(object, pindex, domain, 1950 npages, low, high, alignment, boundary, mpred)) == NULL) 1951 #endif 1952 /* 1953 * If not, allocate them from the free page queues. 1954 */ 1955 m_ret = vm_phys_alloc_contig(domain, npages, low, high, 1956 alignment, boundary); 1957 #if VM_NRESERVLEVEL > 0 1958 if (m_ret == NULL && vm_reserv_reclaim_contig( 1959 domain, npages, low, high, alignment, boundary)) 1960 goto retry; 1961 #endif 1962 } 1963 if (m_ret == NULL) { 1964 if (vm_domain_alloc_fail(vmd, object, req)) 1965 goto again; 1966 return (NULL); 1967 } 1968 vm_domain_freecnt_adj(vmd, -npages); 1969 vm_domain_free_unlock(vmd); 1970 #if VM_NRESERVLEVEL > 0 1971 found: 1972 #endif 1973 for (m = m_ret; m < &m_ret[npages]; m++) 1974 vm_page_alloc_check(m); 1975 1976 /* 1977 * Initialize the pages. Only the PG_ZERO flag is inherited. 1978 */ 1979 flags = 0; 1980 if ((req & VM_ALLOC_ZERO) != 0) 1981 flags = PG_ZERO; 1982 if ((req & VM_ALLOC_NODUMP) != 0) 1983 flags |= PG_NODUMP; 1984 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1985 VPO_UNMANAGED : 0; 1986 busy_lock = VPB_UNBUSIED; 1987 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1988 busy_lock = VPB_SINGLE_EXCLUSIVER; 1989 if ((req & VM_ALLOC_SBUSY) != 0) 1990 busy_lock = VPB_SHARERS_WORD(1); 1991 if ((req & VM_ALLOC_WIRED) != 0) 1992 atomic_add_int(&vm_cnt.v_wire_count, npages); 1993 if (object != NULL) { 1994 if (object->memattr != VM_MEMATTR_DEFAULT && 1995 memattr == VM_MEMATTR_DEFAULT) 1996 memattr = object->memattr; 1997 } 1998 for (m = m_ret; m < &m_ret[npages]; m++) { 1999 m->aflags = 0; 2000 m->flags = (m->flags | PG_NODUMP) & flags; 2001 m->busy_lock = busy_lock; 2002 if ((req & VM_ALLOC_WIRED) != 0) 2003 m->wire_count = 1; 2004 m->act_count = 0; 2005 m->oflags = oflags; 2006 if (object != NULL) { 2007 if (vm_page_insert_after(m, object, pindex, mpred)) { 2008 pagedaemon_wakeup(domain); 2009 if ((req & VM_ALLOC_WIRED) != 0) 2010 atomic_subtract_int( 2011 &vm_cnt.v_wire_count, npages); 2012 KASSERT(m->object == NULL, 2013 ("page %p has object", m)); 2014 mpred = m; 2015 for (m = m_ret; m < &m_ret[npages]; m++) { 2016 if (m <= mpred && 2017 (req & VM_ALLOC_WIRED) != 0) 2018 m->wire_count = 0; 2019 m->oflags = VPO_UNMANAGED; 2020 m->busy_lock = VPB_UNBUSIED; 2021 /* Don't change PG_ZERO. */ 2022 vm_page_free_toq(m); 2023 } 2024 if (req & VM_ALLOC_WAITFAIL) { 2025 VM_OBJECT_WUNLOCK(object); 2026 vm_radix_wait(); 2027 VM_OBJECT_WLOCK(object); 2028 } 2029 return (NULL); 2030 } 2031 mpred = m; 2032 } else 2033 m->pindex = pindex; 2034 if (memattr != VM_MEMATTR_DEFAULT) 2035 pmap_page_set_memattr(m, memattr); 2036 pindex++; 2037 } 2038 vmd = VM_DOMAIN(domain); 2039 if (vm_paging_needed(vmd, vmd->vmd_free_count)) 2040 pagedaemon_wakeup(domain); 2041 return (m_ret); 2042 } 2043 2044 /* 2045 * Check a page that has been freshly dequeued from a freelist. 2046 */ 2047 static void 2048 vm_page_alloc_check(vm_page_t m) 2049 { 2050 2051 KASSERT(m->object == NULL, ("page %p has object", m)); 2052 KASSERT(m->queue == PQ_NONE, 2053 ("page %p has unexpected queue %d", m, m->queue)); 2054 KASSERT(!vm_page_held(m), ("page %p is held", m)); 2055 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 2056 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 2057 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 2058 ("page %p has unexpected memattr %d", 2059 m, pmap_page_get_memattr(m))); 2060 KASSERT(m->valid == 0, ("free page %p is valid", m)); 2061 } 2062 2063 /* 2064 * vm_page_alloc_freelist: 2065 * 2066 * Allocate a physical page from the specified free page list. 2067 * 2068 * The caller must always specify an allocation class. 2069 * 2070 * allocation classes: 2071 * VM_ALLOC_NORMAL normal process request 2072 * VM_ALLOC_SYSTEM system *really* needs a page 2073 * VM_ALLOC_INTERRUPT interrupt time request 2074 * 2075 * optional allocation flags: 2076 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2077 * intends to allocate 2078 * VM_ALLOC_WIRED wire the allocated page 2079 * VM_ALLOC_ZERO prefer a zeroed page 2080 */ 2081 vm_page_t 2082 vm_page_alloc_freelist(int freelist, int req) 2083 { 2084 struct vm_domainset_iter di; 2085 vm_page_t m; 2086 int domain; 2087 2088 vm_domainset_iter_page_init(&di, kernel_object, &domain, &req); 2089 do { 2090 m = vm_page_alloc_freelist_domain(domain, freelist, req); 2091 if (m != NULL) 2092 break; 2093 } while (vm_domainset_iter_page(&di, &domain, &req) == 0); 2094 2095 return (m); 2096 } 2097 2098 vm_page_t 2099 vm_page_alloc_freelist_domain(int domain, int freelist, int req) 2100 { 2101 struct vm_domain *vmd; 2102 vm_page_t m; 2103 u_int flags, free_count; 2104 2105 /* 2106 * Do not allocate reserved pages unless the req has asked for it. 2107 */ 2108 vmd = VM_DOMAIN(domain); 2109 again: 2110 vm_domain_free_lock(vmd); 2111 if (vm_domain_available(vmd, req, 1)) 2112 m = vm_phys_alloc_freelist_pages(domain, freelist, 2113 VM_FREEPOOL_DIRECT, 0); 2114 if (m == NULL) { 2115 if (vm_domain_alloc_fail(vmd, NULL, req)) 2116 goto again; 2117 return (NULL); 2118 } 2119 free_count = vm_domain_freecnt_adj(vmd, -1); 2120 vm_domain_free_unlock(vmd); 2121 vm_page_alloc_check(m); 2122 2123 /* 2124 * Initialize the page. Only the PG_ZERO flag is inherited. 2125 */ 2126 m->aflags = 0; 2127 flags = 0; 2128 if ((req & VM_ALLOC_ZERO) != 0) 2129 flags = PG_ZERO; 2130 m->flags &= flags; 2131 if ((req & VM_ALLOC_WIRED) != 0) { 2132 /* 2133 * The page lock is not required for wiring a page that does 2134 * not belong to an object. 2135 */ 2136 atomic_add_int(&vm_cnt.v_wire_count, 1); 2137 m->wire_count = 1; 2138 } 2139 /* Unmanaged pages don't use "act_count". */ 2140 m->oflags = VPO_UNMANAGED; 2141 if (vm_paging_needed(vmd, free_count)) 2142 pagedaemon_wakeup(domain); 2143 return (m); 2144 } 2145 2146 #define VPSC_ANY 0 /* No restrictions. */ 2147 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2148 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2149 2150 /* 2151 * vm_page_scan_contig: 2152 * 2153 * Scan vm_page_array[] between the specified entries "m_start" and 2154 * "m_end" for a run of contiguous physical pages that satisfy the 2155 * specified conditions, and return the lowest page in the run. The 2156 * specified "alignment" determines the alignment of the lowest physical 2157 * page in the run. If the specified "boundary" is non-zero, then the 2158 * run of physical pages cannot span a physical address that is a 2159 * multiple of "boundary". 2160 * 2161 * "m_end" is never dereferenced, so it need not point to a vm_page 2162 * structure within vm_page_array[]. 2163 * 2164 * "npages" must be greater than zero. "m_start" and "m_end" must not 2165 * span a hole (or discontiguity) in the physical address space. Both 2166 * "alignment" and "boundary" must be a power of two. 2167 */ 2168 vm_page_t 2169 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2170 u_long alignment, vm_paddr_t boundary, int options) 2171 { 2172 struct mtx *m_mtx; 2173 vm_object_t object; 2174 vm_paddr_t pa; 2175 vm_page_t m, m_run; 2176 #if VM_NRESERVLEVEL > 0 2177 int level; 2178 #endif 2179 int m_inc, order, run_ext, run_len; 2180 2181 KASSERT(npages > 0, ("npages is 0")); 2182 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2183 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2184 m_run = NULL; 2185 run_len = 0; 2186 m_mtx = NULL; 2187 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2188 KASSERT((m->flags & PG_MARKER) == 0, 2189 ("page %p is PG_MARKER", m)); 2190 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1, 2191 ("fictitious page %p has invalid wire count", m)); 2192 2193 /* 2194 * If the current page would be the start of a run, check its 2195 * physical address against the end, alignment, and boundary 2196 * conditions. If it doesn't satisfy these conditions, either 2197 * terminate the scan or advance to the next page that 2198 * satisfies the failed condition. 2199 */ 2200 if (run_len == 0) { 2201 KASSERT(m_run == NULL, ("m_run != NULL")); 2202 if (m + npages > m_end) 2203 break; 2204 pa = VM_PAGE_TO_PHYS(m); 2205 if ((pa & (alignment - 1)) != 0) { 2206 m_inc = atop(roundup2(pa, alignment) - pa); 2207 continue; 2208 } 2209 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2210 boundary) != 0) { 2211 m_inc = atop(roundup2(pa, boundary) - pa); 2212 continue; 2213 } 2214 } else 2215 KASSERT(m_run != NULL, ("m_run == NULL")); 2216 2217 vm_page_change_lock(m, &m_mtx); 2218 m_inc = 1; 2219 retry: 2220 if (vm_page_held(m)) 2221 run_ext = 0; 2222 #if VM_NRESERVLEVEL > 0 2223 else if ((level = vm_reserv_level(m)) >= 0 && 2224 (options & VPSC_NORESERV) != 0) { 2225 run_ext = 0; 2226 /* Advance to the end of the reservation. */ 2227 pa = VM_PAGE_TO_PHYS(m); 2228 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2229 pa); 2230 } 2231 #endif 2232 else if ((object = m->object) != NULL) { 2233 /* 2234 * The page is considered eligible for relocation if 2235 * and only if it could be laundered or reclaimed by 2236 * the page daemon. 2237 */ 2238 if (!VM_OBJECT_TRYRLOCK(object)) { 2239 mtx_unlock(m_mtx); 2240 VM_OBJECT_RLOCK(object); 2241 mtx_lock(m_mtx); 2242 if (m->object != object) { 2243 /* 2244 * The page may have been freed. 2245 */ 2246 VM_OBJECT_RUNLOCK(object); 2247 goto retry; 2248 } else if (vm_page_held(m)) { 2249 run_ext = 0; 2250 goto unlock; 2251 } 2252 } 2253 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2254 ("page %p is PG_UNHOLDFREE", m)); 2255 /* Don't care: PG_NODUMP, PG_ZERO. */ 2256 if (object->type != OBJT_DEFAULT && 2257 object->type != OBJT_SWAP && 2258 object->type != OBJT_VNODE) { 2259 run_ext = 0; 2260 #if VM_NRESERVLEVEL > 0 2261 } else if ((options & VPSC_NOSUPER) != 0 && 2262 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2263 run_ext = 0; 2264 /* Advance to the end of the superpage. */ 2265 pa = VM_PAGE_TO_PHYS(m); 2266 m_inc = atop(roundup2(pa + 1, 2267 vm_reserv_size(level)) - pa); 2268 #endif 2269 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2270 m->queue != PQ_NONE && !vm_page_busied(m)) { 2271 /* 2272 * The page is allocated but eligible for 2273 * relocation. Extend the current run by one 2274 * page. 2275 */ 2276 KASSERT(pmap_page_get_memattr(m) == 2277 VM_MEMATTR_DEFAULT, 2278 ("page %p has an unexpected memattr", m)); 2279 KASSERT((m->oflags & (VPO_SWAPINPROG | 2280 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2281 ("page %p has unexpected oflags", m)); 2282 /* Don't care: VPO_NOSYNC. */ 2283 run_ext = 1; 2284 } else 2285 run_ext = 0; 2286 unlock: 2287 VM_OBJECT_RUNLOCK(object); 2288 #if VM_NRESERVLEVEL > 0 2289 } else if (level >= 0) { 2290 /* 2291 * The page is reserved but not yet allocated. In 2292 * other words, it is still free. Extend the current 2293 * run by one page. 2294 */ 2295 run_ext = 1; 2296 #endif 2297 } else if ((order = m->order) < VM_NFREEORDER) { 2298 /* 2299 * The page is enqueued in the physical memory 2300 * allocator's free page queues. Moreover, it is the 2301 * first page in a power-of-two-sized run of 2302 * contiguous free pages. Add these pages to the end 2303 * of the current run, and jump ahead. 2304 */ 2305 run_ext = 1 << order; 2306 m_inc = 1 << order; 2307 } else { 2308 /* 2309 * Skip the page for one of the following reasons: (1) 2310 * It is enqueued in the physical memory allocator's 2311 * free page queues. However, it is not the first 2312 * page in a run of contiguous free pages. (This case 2313 * rarely occurs because the scan is performed in 2314 * ascending order.) (2) It is not reserved, and it is 2315 * transitioning from free to allocated. (Conversely, 2316 * the transition from allocated to free for managed 2317 * pages is blocked by the page lock.) (3) It is 2318 * allocated but not contained by an object and not 2319 * wired, e.g., allocated by Xen's balloon driver. 2320 */ 2321 run_ext = 0; 2322 } 2323 2324 /* 2325 * Extend or reset the current run of pages. 2326 */ 2327 if (run_ext > 0) { 2328 if (run_len == 0) 2329 m_run = m; 2330 run_len += run_ext; 2331 } else { 2332 if (run_len > 0) { 2333 m_run = NULL; 2334 run_len = 0; 2335 } 2336 } 2337 } 2338 if (m_mtx != NULL) 2339 mtx_unlock(m_mtx); 2340 if (run_len >= npages) 2341 return (m_run); 2342 return (NULL); 2343 } 2344 2345 /* 2346 * vm_page_reclaim_run: 2347 * 2348 * Try to relocate each of the allocated virtual pages within the 2349 * specified run of physical pages to a new physical address. Free the 2350 * physical pages underlying the relocated virtual pages. A virtual page 2351 * is relocatable if and only if it could be laundered or reclaimed by 2352 * the page daemon. Whenever possible, a virtual page is relocated to a 2353 * physical address above "high". 2354 * 2355 * Returns 0 if every physical page within the run was already free or 2356 * just freed by a successful relocation. Otherwise, returns a non-zero 2357 * value indicating why the last attempt to relocate a virtual page was 2358 * unsuccessful. 2359 * 2360 * "req_class" must be an allocation class. 2361 */ 2362 static int 2363 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, 2364 vm_paddr_t high) 2365 { 2366 struct vm_domain *vmd; 2367 struct mtx *m_mtx; 2368 struct spglist free; 2369 vm_object_t object; 2370 vm_paddr_t pa; 2371 vm_page_t m, m_end, m_new; 2372 int error, order, req; 2373 2374 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2375 ("req_class is not an allocation class")); 2376 SLIST_INIT(&free); 2377 error = 0; 2378 m = m_run; 2379 m_end = m_run + npages; 2380 m_mtx = NULL; 2381 for (; error == 0 && m < m_end; m++) { 2382 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2383 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2384 2385 /* 2386 * Avoid releasing and reacquiring the same page lock. 2387 */ 2388 vm_page_change_lock(m, &m_mtx); 2389 retry: 2390 if (vm_page_held(m)) 2391 error = EBUSY; 2392 else if ((object = m->object) != NULL) { 2393 /* 2394 * The page is relocated if and only if it could be 2395 * laundered or reclaimed by the page daemon. 2396 */ 2397 if (!VM_OBJECT_TRYWLOCK(object)) { 2398 mtx_unlock(m_mtx); 2399 VM_OBJECT_WLOCK(object); 2400 mtx_lock(m_mtx); 2401 if (m->object != object) { 2402 /* 2403 * The page may have been freed. 2404 */ 2405 VM_OBJECT_WUNLOCK(object); 2406 goto retry; 2407 } else if (vm_page_held(m)) { 2408 error = EBUSY; 2409 goto unlock; 2410 } 2411 } 2412 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2413 ("page %p is PG_UNHOLDFREE", m)); 2414 /* Don't care: PG_NODUMP, PG_ZERO. */ 2415 if (object->type != OBJT_DEFAULT && 2416 object->type != OBJT_SWAP && 2417 object->type != OBJT_VNODE) 2418 error = EINVAL; 2419 else if (object->memattr != VM_MEMATTR_DEFAULT) 2420 error = EINVAL; 2421 else if (m->queue != PQ_NONE && !vm_page_busied(m)) { 2422 KASSERT(pmap_page_get_memattr(m) == 2423 VM_MEMATTR_DEFAULT, 2424 ("page %p has an unexpected memattr", m)); 2425 KASSERT((m->oflags & (VPO_SWAPINPROG | 2426 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2427 ("page %p has unexpected oflags", m)); 2428 /* Don't care: VPO_NOSYNC. */ 2429 if (m->valid != 0) { 2430 /* 2431 * First, try to allocate a new page 2432 * that is above "high". Failing 2433 * that, try to allocate a new page 2434 * that is below "m_run". Allocate 2435 * the new page between the end of 2436 * "m_run" and "high" only as a last 2437 * resort. 2438 */ 2439 req = req_class | VM_ALLOC_NOOBJ; 2440 if ((m->flags & PG_NODUMP) != 0) 2441 req |= VM_ALLOC_NODUMP; 2442 if (trunc_page(high) != 2443 ~(vm_paddr_t)PAGE_MASK) { 2444 m_new = vm_page_alloc_contig( 2445 NULL, 0, req, 1, 2446 round_page(high), 2447 ~(vm_paddr_t)0, 2448 PAGE_SIZE, 0, 2449 VM_MEMATTR_DEFAULT); 2450 } else 2451 m_new = NULL; 2452 if (m_new == NULL) { 2453 pa = VM_PAGE_TO_PHYS(m_run); 2454 m_new = vm_page_alloc_contig( 2455 NULL, 0, req, 1, 2456 0, pa - 1, PAGE_SIZE, 0, 2457 VM_MEMATTR_DEFAULT); 2458 } 2459 if (m_new == NULL) { 2460 pa += ptoa(npages); 2461 m_new = vm_page_alloc_contig( 2462 NULL, 0, req, 1, 2463 pa, high, PAGE_SIZE, 0, 2464 VM_MEMATTR_DEFAULT); 2465 } 2466 if (m_new == NULL) { 2467 error = ENOMEM; 2468 goto unlock; 2469 } 2470 KASSERT(m_new->wire_count == 0, 2471 ("page %p is wired", m)); 2472 2473 /* 2474 * Replace "m" with the new page. For 2475 * vm_page_replace(), "m" must be busy 2476 * and dequeued. Finally, change "m" 2477 * as if vm_page_free() was called. 2478 */ 2479 if (object->ref_count != 0) 2480 pmap_remove_all(m); 2481 m_new->aflags = m->aflags; 2482 KASSERT(m_new->oflags == VPO_UNMANAGED, 2483 ("page %p is managed", m)); 2484 m_new->oflags = m->oflags & VPO_NOSYNC; 2485 pmap_copy_page(m, m_new); 2486 m_new->valid = m->valid; 2487 m_new->dirty = m->dirty; 2488 m->flags &= ~PG_ZERO; 2489 vm_page_xbusy(m); 2490 vm_page_remque(m); 2491 vm_page_replace_checked(m_new, object, 2492 m->pindex, m); 2493 m->valid = 0; 2494 vm_page_undirty(m); 2495 2496 /* 2497 * The new page must be deactivated 2498 * before the object is unlocked. 2499 */ 2500 vm_page_change_lock(m_new, &m_mtx); 2501 vm_page_deactivate(m_new); 2502 } else { 2503 m->flags &= ~PG_ZERO; 2504 vm_page_remque(m); 2505 vm_page_remove(m); 2506 KASSERT(m->dirty == 0, 2507 ("page %p is dirty", m)); 2508 } 2509 SLIST_INSERT_HEAD(&free, m, plinks.s.ss); 2510 } else 2511 error = EBUSY; 2512 unlock: 2513 VM_OBJECT_WUNLOCK(object); 2514 } else { 2515 MPASS(vm_phys_domain(m) == domain); 2516 vmd = VM_DOMAIN(domain); 2517 vm_domain_free_lock(vmd); 2518 order = m->order; 2519 if (order < VM_NFREEORDER) { 2520 /* 2521 * The page is enqueued in the physical memory 2522 * allocator's free page queues. Moreover, it 2523 * is the first page in a power-of-two-sized 2524 * run of contiguous free pages. Jump ahead 2525 * to the last page within that run, and 2526 * continue from there. 2527 */ 2528 m += (1 << order) - 1; 2529 } 2530 #if VM_NRESERVLEVEL > 0 2531 else if (vm_reserv_is_page_free(m)) 2532 order = 0; 2533 #endif 2534 vm_domain_free_unlock(vmd); 2535 if (order == VM_NFREEORDER) 2536 error = EINVAL; 2537 } 2538 } 2539 if (m_mtx != NULL) 2540 mtx_unlock(m_mtx); 2541 if ((m = SLIST_FIRST(&free)) != NULL) { 2542 vmd = VM_DOMAIN(domain); 2543 vm_domain_free_lock(vmd); 2544 do { 2545 MPASS(vm_phys_domain(m) == domain); 2546 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2547 vm_page_free_phys(vmd, m); 2548 } while ((m = SLIST_FIRST(&free)) != NULL); 2549 vm_domain_free_wakeup(vmd); 2550 vm_domain_free_unlock(vmd); 2551 } 2552 return (error); 2553 } 2554 2555 #define NRUNS 16 2556 2557 CTASSERT(powerof2(NRUNS)); 2558 2559 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2560 2561 #define MIN_RECLAIM 8 2562 2563 /* 2564 * vm_page_reclaim_contig: 2565 * 2566 * Reclaim allocated, contiguous physical memory satisfying the specified 2567 * conditions by relocating the virtual pages using that physical memory. 2568 * Returns true if reclamation is successful and false otherwise. Since 2569 * relocation requires the allocation of physical pages, reclamation may 2570 * fail due to a shortage of free pages. When reclamation fails, callers 2571 * are expected to perform VM_WAIT before retrying a failed allocation 2572 * operation, e.g., vm_page_alloc_contig(). 2573 * 2574 * The caller must always specify an allocation class through "req". 2575 * 2576 * allocation classes: 2577 * VM_ALLOC_NORMAL normal process request 2578 * VM_ALLOC_SYSTEM system *really* needs a page 2579 * VM_ALLOC_INTERRUPT interrupt time request 2580 * 2581 * The optional allocation flags are ignored. 2582 * 2583 * "npages" must be greater than zero. Both "alignment" and "boundary" 2584 * must be a power of two. 2585 */ 2586 bool 2587 vm_page_reclaim_contig_domain(int domain, int req, u_long npages, 2588 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 2589 { 2590 struct vm_domain *vmd; 2591 vm_paddr_t curr_low; 2592 vm_page_t m_run, m_runs[NRUNS]; 2593 u_long count, reclaimed; 2594 int error, i, options, req_class; 2595 2596 KASSERT(npages > 0, ("npages is 0")); 2597 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2598 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2599 req_class = req & VM_ALLOC_CLASS_MASK; 2600 2601 /* 2602 * The page daemon is allowed to dig deeper into the free page list. 2603 */ 2604 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2605 req_class = VM_ALLOC_SYSTEM; 2606 2607 /* 2608 * Return if the number of free pages cannot satisfy the requested 2609 * allocation. 2610 */ 2611 vmd = VM_DOMAIN(domain); 2612 count = vmd->vmd_free_count; 2613 if (count < npages + vmd->vmd_free_reserved || (count < npages + 2614 vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2615 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2616 return (false); 2617 2618 /* 2619 * Scan up to three times, relaxing the restrictions ("options") on 2620 * the reclamation of reservations and superpages each time. 2621 */ 2622 for (options = VPSC_NORESERV;;) { 2623 /* 2624 * Find the highest runs that satisfy the given constraints 2625 * and restrictions, and record them in "m_runs". 2626 */ 2627 curr_low = low; 2628 count = 0; 2629 for (;;) { 2630 m_run = vm_phys_scan_contig(domain, npages, curr_low, 2631 high, alignment, boundary, options); 2632 if (m_run == NULL) 2633 break; 2634 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2635 m_runs[RUN_INDEX(count)] = m_run; 2636 count++; 2637 } 2638 2639 /* 2640 * Reclaim the highest runs in LIFO (descending) order until 2641 * the number of reclaimed pages, "reclaimed", is at least 2642 * MIN_RECLAIM. Reset "reclaimed" each time because each 2643 * reclamation is idempotent, and runs will (likely) recur 2644 * from one scan to the next as restrictions are relaxed. 2645 */ 2646 reclaimed = 0; 2647 for (i = 0; count > 0 && i < NRUNS; i++) { 2648 count--; 2649 m_run = m_runs[RUN_INDEX(count)]; 2650 error = vm_page_reclaim_run(req_class, domain, npages, 2651 m_run, high); 2652 if (error == 0) { 2653 reclaimed += npages; 2654 if (reclaimed >= MIN_RECLAIM) 2655 return (true); 2656 } 2657 } 2658 2659 /* 2660 * Either relax the restrictions on the next scan or return if 2661 * the last scan had no restrictions. 2662 */ 2663 if (options == VPSC_NORESERV) 2664 options = VPSC_NOSUPER; 2665 else if (options == VPSC_NOSUPER) 2666 options = VPSC_ANY; 2667 else if (options == VPSC_ANY) 2668 return (reclaimed != 0); 2669 } 2670 } 2671 2672 bool 2673 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2674 u_long alignment, vm_paddr_t boundary) 2675 { 2676 struct vm_domainset_iter di; 2677 int domain; 2678 bool ret; 2679 2680 vm_domainset_iter_page_init(&di, kernel_object, &domain, &req); 2681 do { 2682 ret = vm_page_reclaim_contig_domain(domain, req, npages, low, 2683 high, alignment, boundary); 2684 if (ret) 2685 break; 2686 } while (vm_domainset_iter_page(&di, &domain, &req) == 0); 2687 2688 return (ret); 2689 } 2690 2691 /* 2692 * Set the domain in the appropriate page level domainset. 2693 */ 2694 void 2695 vm_domain_set(struct vm_domain *vmd) 2696 { 2697 2698 mtx_lock(&vm_domainset_lock); 2699 if (!vmd->vmd_minset && vm_paging_min(vmd)) { 2700 vmd->vmd_minset = 1; 2701 DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); 2702 } 2703 if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { 2704 vmd->vmd_severeset = 1; 2705 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 2706 } 2707 mtx_unlock(&vm_domainset_lock); 2708 } 2709 2710 /* 2711 * Clear the domain from the appropriate page level domainset. 2712 */ 2713 static void 2714 vm_domain_clear(struct vm_domain *vmd) 2715 { 2716 2717 mtx_lock(&vm_domainset_lock); 2718 if (vmd->vmd_minset && !vm_paging_min(vmd)) { 2719 vmd->vmd_minset = 0; 2720 DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); 2721 if (vm_min_waiters != 0) { 2722 vm_min_waiters = 0; 2723 wakeup(&vm_min_domains); 2724 } 2725 } 2726 if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { 2727 vmd->vmd_severeset = 0; 2728 DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); 2729 if (vm_severe_waiters != 0) { 2730 vm_severe_waiters = 0; 2731 wakeup(&vm_severe_domains); 2732 } 2733 } 2734 mtx_unlock(&vm_domainset_lock); 2735 } 2736 2737 /* 2738 * Wait for free pages to exceed the min threshold globally. 2739 */ 2740 void 2741 vm_wait_min(void) 2742 { 2743 2744 mtx_lock(&vm_domainset_lock); 2745 while (vm_page_count_min()) { 2746 vm_min_waiters++; 2747 msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); 2748 } 2749 mtx_unlock(&vm_domainset_lock); 2750 } 2751 2752 /* 2753 * Wait for free pages to exceed the severe threshold globally. 2754 */ 2755 void 2756 vm_wait_severe(void) 2757 { 2758 2759 mtx_lock(&vm_domainset_lock); 2760 while (vm_page_count_severe()) { 2761 vm_severe_waiters++; 2762 msleep(&vm_severe_domains, &vm_domainset_lock, PVM, 2763 "vmwait", 0); 2764 } 2765 mtx_unlock(&vm_domainset_lock); 2766 } 2767 2768 u_int 2769 vm_wait_count(void) 2770 { 2771 u_int cnt; 2772 int i; 2773 2774 cnt = 0; 2775 for (i = 0; i < vm_ndomains; i++) 2776 cnt += VM_DOMAIN(i)->vmd_waiters; 2777 cnt += vm_severe_waiters + vm_min_waiters; 2778 2779 return (cnt); 2780 } 2781 2782 /* 2783 * vm_wait_domain: 2784 * 2785 * Sleep until free pages are available for allocation. 2786 * - Called in various places after failed memory allocations. 2787 */ 2788 void 2789 vm_wait_domain(int domain) 2790 { 2791 struct vm_domain *vmd; 2792 2793 vmd = VM_DOMAIN(domain); 2794 vm_domain_free_assert_locked(vmd); 2795 2796 if (curproc == pageproc) { 2797 vmd->vmd_pageout_pages_needed = 1; 2798 msleep(&vmd->vmd_pageout_pages_needed, 2799 vm_domain_free_lockptr(vmd), PDROP | PSWP, "VMWait", 0); 2800 } else { 2801 if (pageproc == NULL) 2802 panic("vm_wait in early boot"); 2803 pagedaemon_wait(domain, PVM, "vmwait"); 2804 } 2805 } 2806 2807 /* 2808 * vm_wait: (also see VM_WAIT macro) 2809 * 2810 * Sleep until free pages are available for allocation. 2811 * - Called in various places after failed memory allocations. 2812 */ 2813 void 2814 vm_wait(void) 2815 { 2816 2817 /* 2818 * We use racey wakeup synchronization to avoid expensive global 2819 * locking for the pageproc when sleeping with a non-specific vm_wait. 2820 * To handle this, we only sleep for one tick in this instance. It 2821 * is expected that most allocations for the pageproc will come from 2822 * kmem or vm_page_grab* which will use the more specific and 2823 * race-free vm_wait_domain(). 2824 */ 2825 if (curproc == pageproc) { 2826 mtx_lock(&vm_domainset_lock); 2827 vm_pageproc_waiters++; 2828 msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM, 2829 "pageprocwait", 1); 2830 mtx_unlock(&vm_domainset_lock); 2831 } else { 2832 /* 2833 * XXX Ideally we would wait only until the allocation could 2834 * be satisfied. This condition can cause new allocators to 2835 * consume all freed pages while old allocators wait. 2836 */ 2837 mtx_lock(&vm_domainset_lock); 2838 if (vm_page_count_min()) { 2839 vm_min_waiters++; 2840 msleep(&vm_min_domains, &vm_domainset_lock, PVM, 2841 "vmwait", 0); 2842 } 2843 mtx_unlock(&vm_domainset_lock); 2844 } 2845 } 2846 2847 /* 2848 * vm_domain_alloc_fail: 2849 * 2850 * Called when a page allocation function fails. Informs the 2851 * pagedaemon and performs the requested wait. Requires the 2852 * domain_free and object lock on entry. Returns with the 2853 * object lock held and free lock released. Returns an error when 2854 * retry is necessary. 2855 * 2856 */ 2857 static int 2858 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) 2859 { 2860 2861 vm_domain_free_assert_locked(vmd); 2862 2863 atomic_add_int(&vmd->vmd_pageout_deficit, 2864 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 2865 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 2866 if (object != NULL) 2867 VM_OBJECT_WUNLOCK(object); 2868 vm_wait_domain(vmd->vmd_domain); 2869 if (object != NULL) 2870 VM_OBJECT_WLOCK(object); 2871 if (req & VM_ALLOC_WAITOK) 2872 return (EAGAIN); 2873 } else { 2874 vm_domain_free_unlock(vmd); 2875 pagedaemon_wakeup(vmd->vmd_domain); 2876 } 2877 return (0); 2878 } 2879 2880 /* 2881 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2882 * 2883 * Sleep until free pages are available for allocation. 2884 * - Called only in vm_fault so that processes page faulting 2885 * can be easily tracked. 2886 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2887 * processes will be able to grab memory first. Do not change 2888 * this balance without careful testing first. 2889 */ 2890 void 2891 vm_waitpfault(void) 2892 { 2893 2894 mtx_lock(&vm_domainset_lock); 2895 if (vm_page_count_min()) { 2896 vm_min_waiters++; 2897 msleep(&vm_min_domains, &vm_domainset_lock, PUSER, "pfault", 0); 2898 } 2899 mtx_unlock(&vm_domainset_lock); 2900 } 2901 2902 struct vm_pagequeue * 2903 vm_page_pagequeue(vm_page_t m) 2904 { 2905 2906 return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]); 2907 } 2908 2909 /* 2910 * vm_page_dequeue: 2911 * 2912 * Remove the given page from its current page queue. 2913 * 2914 * The page must be locked. 2915 */ 2916 void 2917 vm_page_dequeue(vm_page_t m) 2918 { 2919 struct vm_pagequeue *pq; 2920 2921 vm_page_assert_locked(m); 2922 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2923 m)); 2924 pq = vm_page_pagequeue(m); 2925 vm_pagequeue_lock(pq); 2926 m->queue = PQ_NONE; 2927 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2928 vm_pagequeue_cnt_dec(pq); 2929 vm_pagequeue_unlock(pq); 2930 } 2931 2932 /* 2933 * vm_page_dequeue_locked: 2934 * 2935 * Remove the given page from its current page queue. 2936 * 2937 * The page and page queue must be locked. 2938 */ 2939 void 2940 vm_page_dequeue_locked(vm_page_t m) 2941 { 2942 struct vm_pagequeue *pq; 2943 2944 vm_page_lock_assert(m, MA_OWNED); 2945 pq = vm_page_pagequeue(m); 2946 vm_pagequeue_assert_locked(pq); 2947 m->queue = PQ_NONE; 2948 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2949 vm_pagequeue_cnt_dec(pq); 2950 } 2951 2952 /* 2953 * vm_page_enqueue: 2954 * 2955 * Add the given page to the specified page queue. 2956 * 2957 * The page must be locked. 2958 */ 2959 static void 2960 vm_page_enqueue(uint8_t queue, vm_page_t m) 2961 { 2962 struct vm_pagequeue *pq; 2963 2964 vm_page_lock_assert(m, MA_OWNED); 2965 KASSERT(queue < PQ_COUNT, 2966 ("vm_page_enqueue: invalid queue %u request for page %p", 2967 queue, m)); 2968 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; 2969 vm_pagequeue_lock(pq); 2970 m->queue = queue; 2971 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2972 vm_pagequeue_cnt_inc(pq); 2973 vm_pagequeue_unlock(pq); 2974 } 2975 2976 /* 2977 * vm_page_requeue: 2978 * 2979 * Move the given page to the tail of its current page queue. 2980 * 2981 * The page must be locked. 2982 */ 2983 void 2984 vm_page_requeue(vm_page_t m) 2985 { 2986 struct vm_pagequeue *pq; 2987 2988 vm_page_lock_assert(m, MA_OWNED); 2989 KASSERT(m->queue != PQ_NONE, 2990 ("vm_page_requeue: page %p is not queued", m)); 2991 pq = vm_page_pagequeue(m); 2992 vm_pagequeue_lock(pq); 2993 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2994 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2995 vm_pagequeue_unlock(pq); 2996 } 2997 2998 /* 2999 * vm_page_requeue_locked: 3000 * 3001 * Move the given page to the tail of its current page queue. 3002 * 3003 * The page queue must be locked. 3004 */ 3005 void 3006 vm_page_requeue_locked(vm_page_t m) 3007 { 3008 struct vm_pagequeue *pq; 3009 3010 KASSERT(m->queue != PQ_NONE, 3011 ("vm_page_requeue_locked: page %p is not queued", m)); 3012 pq = vm_page_pagequeue(m); 3013 vm_pagequeue_assert_locked(pq); 3014 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 3015 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3016 } 3017 3018 /* 3019 * vm_page_activate: 3020 * 3021 * Put the specified page on the active list (if appropriate). 3022 * Ensure that act_count is at least ACT_INIT but do not otherwise 3023 * mess with it. 3024 * 3025 * The page must be locked. 3026 */ 3027 void 3028 vm_page_activate(vm_page_t m) 3029 { 3030 int queue; 3031 3032 vm_page_lock_assert(m, MA_OWNED); 3033 if ((queue = m->queue) != PQ_ACTIVE) { 3034 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3035 if (m->act_count < ACT_INIT) 3036 m->act_count = ACT_INIT; 3037 if (queue != PQ_NONE) 3038 vm_page_dequeue(m); 3039 vm_page_enqueue(PQ_ACTIVE, m); 3040 } 3041 } else { 3042 if (m->act_count < ACT_INIT) 3043 m->act_count = ACT_INIT; 3044 } 3045 } 3046 3047 /* 3048 * vm_domain_free_wakeup: 3049 * 3050 * Helper routine for vm_page_free_toq(). This routine is called 3051 * when a page is added to the free queues. 3052 * 3053 * The page queues must be locked. 3054 */ 3055 static void 3056 vm_domain_free_wakeup(struct vm_domain *vmd) 3057 { 3058 3059 vm_domain_free_assert_locked(vmd); 3060 3061 /* 3062 * if pageout daemon needs pages, then tell it that there are 3063 * some free. 3064 */ 3065 if (vmd->vmd_pageout_pages_needed && 3066 vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { 3067 wakeup(&vmd->vmd_pageout_pages_needed); 3068 vmd->vmd_pageout_pages_needed = 0; 3069 } 3070 /* 3071 * wakeup processes that are waiting on memory if we hit a 3072 * high water mark. And wakeup scheduler process if we have 3073 * lots of memory. this process will swapin processes. 3074 */ 3075 if (vmd->vmd_pages_needed && !vm_paging_min(vmd)) { 3076 vmd->vmd_pages_needed = false; 3077 wakeup(&vmd->vmd_free_count); 3078 } 3079 if ((vmd->vmd_minset && !vm_paging_min(vmd)) || 3080 (vmd->vmd_severeset && !vm_paging_severe(vmd))) 3081 vm_domain_clear(vmd); 3082 3083 /* See comments in vm_wait(); */ 3084 if (vm_pageproc_waiters) { 3085 vm_pageproc_waiters = 0; 3086 wakeup(&vm_pageproc_waiters); 3087 } 3088 3089 } 3090 3091 /* 3092 * vm_page_free_prep: 3093 * 3094 * Prepares the given page to be put on the free list, 3095 * disassociating it from any VM object. The caller may return 3096 * the page to the free list only if this function returns true. 3097 * 3098 * The object must be locked. The page must be locked if it is 3099 * managed. For a queued managed page, the pagequeue_locked 3100 * argument specifies whether the page queue is already locked. 3101 */ 3102 bool 3103 vm_page_free_prep(vm_page_t m, bool pagequeue_locked) 3104 { 3105 3106 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 3107 if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { 3108 uint64_t *p; 3109 int i; 3110 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 3111 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 3112 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 3113 m, i, (uintmax_t)*p)); 3114 } 3115 #endif 3116 if ((m->oflags & VPO_UNMANAGED) == 0) { 3117 vm_page_lock_assert(m, MA_OWNED); 3118 KASSERT(!pmap_page_is_mapped(m), 3119 ("vm_page_free_toq: freeing mapped page %p", m)); 3120 } else 3121 KASSERT(m->queue == PQ_NONE, 3122 ("vm_page_free_toq: unmanaged page %p is queued", m)); 3123 VM_CNT_INC(v_tfree); 3124 3125 if (vm_page_sbusied(m)) 3126 panic("vm_page_free: freeing busy page %p", m); 3127 3128 vm_page_remove(m); 3129 3130 /* 3131 * If fictitious remove object association and 3132 * return. 3133 */ 3134 if ((m->flags & PG_FICTITIOUS) != 0) { 3135 KASSERT(m->wire_count == 1, 3136 ("fictitious page %p is not wired", m)); 3137 KASSERT(m->queue == PQ_NONE, 3138 ("fictitious page %p is queued", m)); 3139 return (false); 3140 } 3141 3142 if (m->queue != PQ_NONE) { 3143 if (pagequeue_locked) 3144 vm_page_dequeue_locked(m); 3145 else 3146 vm_page_dequeue(m); 3147 } 3148 m->valid = 0; 3149 vm_page_undirty(m); 3150 3151 if (m->wire_count != 0) 3152 panic("vm_page_free: freeing wired page %p", m); 3153 if (m->hold_count != 0) { 3154 m->flags &= ~PG_ZERO; 3155 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 3156 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 3157 m->flags |= PG_UNHOLDFREE; 3158 return (false); 3159 } 3160 3161 /* 3162 * Restore the default memory attribute to the page. 3163 */ 3164 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3165 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3166 3167 return (true); 3168 } 3169 3170 /* 3171 * Insert the page into the physical memory allocator's free page 3172 * queues. This is the last step to free a page. 3173 */ 3174 static void 3175 vm_page_free_phys(struct vm_domain *vmd, vm_page_t m) 3176 { 3177 3178 vm_domain_free_assert_locked(vmd); 3179 3180 vm_domain_freecnt_adj(vmd, 1); 3181 #if VM_NRESERVLEVEL > 0 3182 if (!vm_reserv_free_page(m)) 3183 #endif 3184 vm_phys_free_pages(m, 0); 3185 } 3186 3187 void 3188 vm_page_free_phys_pglist(struct pglist *tq) 3189 { 3190 struct vm_domain *vmd; 3191 vm_page_t m; 3192 3193 if (TAILQ_EMPTY(tq)) 3194 return; 3195 vmd = NULL; 3196 TAILQ_FOREACH(m, tq, listq) { 3197 if (vmd != vm_pagequeue_domain(m)) { 3198 if (vmd != NULL) { 3199 vm_domain_free_wakeup(vmd); 3200 vm_domain_free_unlock(vmd); 3201 } 3202 vmd = vm_pagequeue_domain(m); 3203 vm_domain_free_lock(vmd); 3204 } 3205 vm_page_free_phys(vmd, m); 3206 } 3207 if (vmd != NULL) { 3208 vm_domain_free_wakeup(vmd); 3209 vm_domain_free_unlock(vmd); 3210 } 3211 } 3212 3213 /* 3214 * vm_page_free_toq: 3215 * 3216 * Returns the given page to the free list, disassociating it 3217 * from any VM object. 3218 * 3219 * The object must be locked. The page must be locked if it is 3220 * managed. 3221 */ 3222 void 3223 vm_page_free_toq(vm_page_t m) 3224 { 3225 struct vm_domain *vmd; 3226 3227 if (!vm_page_free_prep(m, false)) 3228 return; 3229 vmd = vm_pagequeue_domain(m); 3230 vm_domain_free_lock(vmd); 3231 vm_page_free_phys(vmd, m); 3232 vm_domain_free_wakeup(vmd); 3233 vm_domain_free_unlock(vmd); 3234 } 3235 3236 /* 3237 * vm_page_wire: 3238 * 3239 * Mark this page as wired down. If the page is fictitious, then 3240 * its wire count must remain one. 3241 * 3242 * The page must be locked. 3243 */ 3244 void 3245 vm_page_wire(vm_page_t m) 3246 { 3247 3248 vm_page_assert_locked(m); 3249 if ((m->flags & PG_FICTITIOUS) != 0) { 3250 KASSERT(m->wire_count == 1, 3251 ("vm_page_wire: fictitious page %p's wire count isn't one", 3252 m)); 3253 return; 3254 } 3255 if (m->wire_count == 0) { 3256 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 3257 m->queue == PQ_NONE, 3258 ("vm_page_wire: unmanaged page %p is queued", m)); 3259 atomic_add_int(&vm_cnt.v_wire_count, 1); 3260 } 3261 m->wire_count++; 3262 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 3263 } 3264 3265 /* 3266 * vm_page_unwire: 3267 * 3268 * Release one wiring of the specified page, potentially allowing it to be 3269 * paged out. Returns TRUE if the number of wirings transitions to zero and 3270 * FALSE otherwise. 3271 * 3272 * Only managed pages belonging to an object can be paged out. If the number 3273 * of wirings transitions to zero and the page is eligible for page out, then 3274 * the page is added to the specified paging queue (unless PQ_NONE is 3275 * specified, in which case the page is dequeued if it belongs to a paging 3276 * queue). 3277 * 3278 * If a page is fictitious, then its wire count must always be one. 3279 * 3280 * A managed page must be locked. 3281 */ 3282 bool 3283 vm_page_unwire(vm_page_t m, uint8_t queue) 3284 { 3285 bool unwired; 3286 3287 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 3288 ("vm_page_unwire: invalid queue %u request for page %p", 3289 queue, m)); 3290 3291 unwired = vm_page_unwire_noq(m); 3292 if (unwired && (m->oflags & VPO_UNMANAGED) == 0 && m->object != NULL) { 3293 if (m->queue == queue) { 3294 if (queue == PQ_ACTIVE) 3295 vm_page_reference(m); 3296 else if (queue != PQ_NONE) 3297 vm_page_requeue(m); 3298 } else { 3299 vm_page_remque(m); 3300 if (queue != PQ_NONE) { 3301 vm_page_enqueue(queue, m); 3302 if (queue == PQ_ACTIVE) 3303 /* Initialize act_count. */ 3304 vm_page_activate(m); 3305 } 3306 } 3307 } 3308 return (unwired); 3309 } 3310 3311 /* 3312 * 3313 * vm_page_unwire_noq: 3314 * 3315 * Unwire a page without (re-)inserting it into a page queue. It is up 3316 * to the caller to enqueue, requeue, or free the page as appropriate. 3317 * In most cases, vm_page_unwire() should be used instead. 3318 */ 3319 bool 3320 vm_page_unwire_noq(vm_page_t m) 3321 { 3322 3323 if ((m->oflags & VPO_UNMANAGED) == 0) 3324 vm_page_assert_locked(m); 3325 if ((m->flags & PG_FICTITIOUS) != 0) { 3326 KASSERT(m->wire_count == 1, 3327 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 3328 return (false); 3329 } 3330 if (m->wire_count == 0) 3331 panic("vm_page_unwire: page %p's wire count is zero", m); 3332 m->wire_count--; 3333 if (m->wire_count == 0) { 3334 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 3335 return (true); 3336 } else 3337 return (false); 3338 } 3339 3340 /* 3341 * Move the specified page to the inactive queue. 3342 * 3343 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive 3344 * queue. However, setting "noreuse" to TRUE will accelerate the specified 3345 * page's reclamation, but it will not unmap the page from any address space. 3346 * This is implemented by inserting the page near the head of the inactive 3347 * queue, using a marker page to guide FIFO insertion ordering. 3348 * 3349 * The page must be locked. 3350 */ 3351 static inline void 3352 _vm_page_deactivate(vm_page_t m, boolean_t noreuse) 3353 { 3354 struct vm_pagequeue *pq; 3355 int queue; 3356 3357 vm_page_assert_locked(m); 3358 3359 /* 3360 * Ignore if the page is already inactive, unless it is unlikely to be 3361 * reactivated. 3362 */ 3363 if ((queue = m->queue) == PQ_INACTIVE && !noreuse) 3364 return; 3365 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3366 pq = &vm_pagequeue_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 3367 /* Avoid multiple acquisitions of the inactive queue lock. */ 3368 if (queue == PQ_INACTIVE) { 3369 vm_pagequeue_lock(pq); 3370 vm_page_dequeue_locked(m); 3371 } else { 3372 if (queue != PQ_NONE) 3373 vm_page_dequeue(m); 3374 vm_pagequeue_lock(pq); 3375 } 3376 m->queue = PQ_INACTIVE; 3377 if (noreuse) 3378 TAILQ_INSERT_BEFORE( 3379 &vm_pagequeue_domain(m)->vmd_inacthead, m, 3380 plinks.q); 3381 else 3382 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3383 vm_pagequeue_cnt_inc(pq); 3384 vm_pagequeue_unlock(pq); 3385 } 3386 } 3387 3388 /* 3389 * Move the specified page to the inactive queue. 3390 * 3391 * The page must be locked. 3392 */ 3393 void 3394 vm_page_deactivate(vm_page_t m) 3395 { 3396 3397 _vm_page_deactivate(m, FALSE); 3398 } 3399 3400 /* 3401 * Move the specified page to the inactive queue with the expectation 3402 * that it is unlikely to be reused. 3403 * 3404 * The page must be locked. 3405 */ 3406 void 3407 vm_page_deactivate_noreuse(vm_page_t m) 3408 { 3409 3410 _vm_page_deactivate(m, TRUE); 3411 } 3412 3413 /* 3414 * vm_page_launder 3415 * 3416 * Put a page in the laundry. 3417 */ 3418 void 3419 vm_page_launder(vm_page_t m) 3420 { 3421 int queue; 3422 3423 vm_page_assert_locked(m); 3424 if ((queue = m->queue) != PQ_LAUNDRY) { 3425 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3426 if (queue != PQ_NONE) 3427 vm_page_dequeue(m); 3428 vm_page_enqueue(PQ_LAUNDRY, m); 3429 } else 3430 KASSERT(queue == PQ_NONE, 3431 ("wired page %p is queued", m)); 3432 } 3433 } 3434 3435 /* 3436 * vm_page_unswappable 3437 * 3438 * Put a page in the PQ_UNSWAPPABLE holding queue. 3439 */ 3440 void 3441 vm_page_unswappable(vm_page_t m) 3442 { 3443 3444 vm_page_assert_locked(m); 3445 KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0, 3446 ("page %p already unswappable", m)); 3447 if (m->queue != PQ_NONE) 3448 vm_page_dequeue(m); 3449 vm_page_enqueue(PQ_UNSWAPPABLE, m); 3450 } 3451 3452 /* 3453 * Attempt to free the page. If it cannot be freed, do nothing. Returns true 3454 * if the page is freed and false otherwise. 3455 * 3456 * The page must be managed. The page and its containing object must be 3457 * locked. 3458 */ 3459 bool 3460 vm_page_try_to_free(vm_page_t m) 3461 { 3462 3463 vm_page_assert_locked(m); 3464 VM_OBJECT_ASSERT_WLOCKED(m->object); 3465 KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m)); 3466 if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m)) 3467 return (false); 3468 if (m->object->ref_count != 0) { 3469 pmap_remove_all(m); 3470 if (m->dirty != 0) 3471 return (false); 3472 } 3473 vm_page_free(m); 3474 return (true); 3475 } 3476 3477 /* 3478 * vm_page_advise 3479 * 3480 * Apply the specified advice to the given page. 3481 * 3482 * The object and page must be locked. 3483 */ 3484 void 3485 vm_page_advise(vm_page_t m, int advice) 3486 { 3487 3488 vm_page_assert_locked(m); 3489 VM_OBJECT_ASSERT_WLOCKED(m->object); 3490 if (advice == MADV_FREE) 3491 /* 3492 * Mark the page clean. This will allow the page to be freed 3493 * without first paging it out. MADV_FREE pages are often 3494 * quickly reused by malloc(3), so we do not do anything that 3495 * would result in a page fault on a later access. 3496 */ 3497 vm_page_undirty(m); 3498 else if (advice != MADV_DONTNEED) { 3499 if (advice == MADV_WILLNEED) 3500 vm_page_activate(m); 3501 return; 3502 } 3503 3504 /* 3505 * Clear any references to the page. Otherwise, the page daemon will 3506 * immediately reactivate the page. 3507 */ 3508 vm_page_aflag_clear(m, PGA_REFERENCED); 3509 3510 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 3511 vm_page_dirty(m); 3512 3513 /* 3514 * Place clean pages near the head of the inactive queue rather than 3515 * the tail, thus defeating the queue's LRU operation and ensuring that 3516 * the page will be reused quickly. Dirty pages not already in the 3517 * laundry are moved there. 3518 */ 3519 if (m->dirty == 0) 3520 vm_page_deactivate_noreuse(m); 3521 else 3522 vm_page_launder(m); 3523 } 3524 3525 /* 3526 * Grab a page, waiting until we are waken up due to the page 3527 * changing state. We keep on waiting, if the page continues 3528 * to be in the object. If the page doesn't exist, first allocate it 3529 * and then conditionally zero it. 3530 * 3531 * This routine may sleep. 3532 * 3533 * The object must be locked on entry. The lock will, however, be released 3534 * and reacquired if the routine sleeps. 3535 */ 3536 vm_page_t 3537 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3538 { 3539 vm_page_t m; 3540 int sleep; 3541 int pflags; 3542 3543 VM_OBJECT_ASSERT_WLOCKED(object); 3544 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3545 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3546 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3547 pflags = allocflags & 3548 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 3549 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 3550 pflags |= VM_ALLOC_WAITFAIL; 3551 retrylookup: 3552 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3553 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3554 vm_page_xbusied(m) : vm_page_busied(m); 3555 if (sleep) { 3556 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3557 return (NULL); 3558 /* 3559 * Reference the page before unlocking and 3560 * sleeping so that the page daemon is less 3561 * likely to reclaim it. 3562 */ 3563 vm_page_aflag_set(m, PGA_REFERENCED); 3564 vm_page_lock(m); 3565 VM_OBJECT_WUNLOCK(object); 3566 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 3567 VM_ALLOC_IGN_SBUSY) != 0); 3568 VM_OBJECT_WLOCK(object); 3569 goto retrylookup; 3570 } else { 3571 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3572 vm_page_lock(m); 3573 vm_page_wire(m); 3574 vm_page_unlock(m); 3575 } 3576 if ((allocflags & 3577 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3578 vm_page_xbusy(m); 3579 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3580 vm_page_sbusy(m); 3581 return (m); 3582 } 3583 } 3584 m = vm_page_alloc(object, pindex, pflags); 3585 if (m == NULL) { 3586 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3587 return (NULL); 3588 goto retrylookup; 3589 } 3590 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3591 pmap_zero_page(m); 3592 return (m); 3593 } 3594 3595 /* 3596 * Return the specified range of pages from the given object. For each 3597 * page offset within the range, if a page already exists within the object 3598 * at that offset and it is busy, then wait for it to change state. If, 3599 * instead, the page doesn't exist, then allocate it. 3600 * 3601 * The caller must always specify an allocation class. 3602 * 3603 * allocation classes: 3604 * VM_ALLOC_NORMAL normal process request 3605 * VM_ALLOC_SYSTEM system *really* needs the pages 3606 * 3607 * The caller must always specify that the pages are to be busied and/or 3608 * wired. 3609 * 3610 * optional allocation flags: 3611 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 3612 * VM_ALLOC_NOBUSY do not exclusive busy the page 3613 * VM_ALLOC_NOWAIT do not sleep 3614 * VM_ALLOC_SBUSY set page to sbusy state 3615 * VM_ALLOC_WIRED wire the pages 3616 * VM_ALLOC_ZERO zero and validate any invalid pages 3617 * 3618 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 3619 * may return a partial prefix of the requested range. 3620 */ 3621 int 3622 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 3623 vm_page_t *ma, int count) 3624 { 3625 vm_page_t m, mpred; 3626 int pflags; 3627 int i; 3628 bool sleep; 3629 3630 VM_OBJECT_ASSERT_WLOCKED(object); 3631 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 3632 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 3633 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 3634 (allocflags & VM_ALLOC_WIRED) != 0, 3635 ("vm_page_grab_pages: the pages must be busied or wired")); 3636 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3637 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3638 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); 3639 if (count == 0) 3640 return (0); 3641 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | 3642 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY); 3643 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 3644 pflags |= VM_ALLOC_WAITFAIL; 3645 i = 0; 3646 retrylookup: 3647 m = vm_radix_lookup_le(&object->rtree, pindex + i); 3648 if (m == NULL || m->pindex != pindex + i) { 3649 mpred = m; 3650 m = NULL; 3651 } else 3652 mpred = TAILQ_PREV(m, pglist, listq); 3653 for (; i < count; i++) { 3654 if (m != NULL) { 3655 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3656 vm_page_xbusied(m) : vm_page_busied(m); 3657 if (sleep) { 3658 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3659 break; 3660 /* 3661 * Reference the page before unlocking and 3662 * sleeping so that the page daemon is less 3663 * likely to reclaim it. 3664 */ 3665 vm_page_aflag_set(m, PGA_REFERENCED); 3666 vm_page_lock(m); 3667 VM_OBJECT_WUNLOCK(object); 3668 vm_page_busy_sleep(m, "grbmaw", (allocflags & 3669 VM_ALLOC_IGN_SBUSY) != 0); 3670 VM_OBJECT_WLOCK(object); 3671 goto retrylookup; 3672 } 3673 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3674 vm_page_lock(m); 3675 vm_page_wire(m); 3676 vm_page_unlock(m); 3677 } 3678 if ((allocflags & (VM_ALLOC_NOBUSY | 3679 VM_ALLOC_SBUSY)) == 0) 3680 vm_page_xbusy(m); 3681 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3682 vm_page_sbusy(m); 3683 } else { 3684 m = vm_page_alloc_after(object, pindex + i, 3685 pflags | VM_ALLOC_COUNT(count - i), mpred); 3686 if (m == NULL) { 3687 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3688 break; 3689 goto retrylookup; 3690 } 3691 } 3692 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) { 3693 if ((m->flags & PG_ZERO) == 0) 3694 pmap_zero_page(m); 3695 m->valid = VM_PAGE_BITS_ALL; 3696 } 3697 ma[i] = mpred = m; 3698 m = vm_page_next(m); 3699 } 3700 return (i); 3701 } 3702 3703 /* 3704 * Mapping function for valid or dirty bits in a page. 3705 * 3706 * Inputs are required to range within a page. 3707 */ 3708 vm_page_bits_t 3709 vm_page_bits(int base, int size) 3710 { 3711 int first_bit; 3712 int last_bit; 3713 3714 KASSERT( 3715 base + size <= PAGE_SIZE, 3716 ("vm_page_bits: illegal base/size %d/%d", base, size) 3717 ); 3718 3719 if (size == 0) /* handle degenerate case */ 3720 return (0); 3721 3722 first_bit = base >> DEV_BSHIFT; 3723 last_bit = (base + size - 1) >> DEV_BSHIFT; 3724 3725 return (((vm_page_bits_t)2 << last_bit) - 3726 ((vm_page_bits_t)1 << first_bit)); 3727 } 3728 3729 /* 3730 * vm_page_set_valid_range: 3731 * 3732 * Sets portions of a page valid. The arguments are expected 3733 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3734 * of any partial chunks touched by the range. The invalid portion of 3735 * such chunks will be zeroed. 3736 * 3737 * (base + size) must be less then or equal to PAGE_SIZE. 3738 */ 3739 void 3740 vm_page_set_valid_range(vm_page_t m, int base, int size) 3741 { 3742 int endoff, frag; 3743 3744 VM_OBJECT_ASSERT_WLOCKED(m->object); 3745 if (size == 0) /* handle degenerate case */ 3746 return; 3747 3748 /* 3749 * If the base is not DEV_BSIZE aligned and the valid 3750 * bit is clear, we have to zero out a portion of the 3751 * first block. 3752 */ 3753 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3754 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 3755 pmap_zero_page_area(m, frag, base - frag); 3756 3757 /* 3758 * If the ending offset is not DEV_BSIZE aligned and the 3759 * valid bit is clear, we have to zero out a portion of 3760 * the last block. 3761 */ 3762 endoff = base + size; 3763 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3764 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 3765 pmap_zero_page_area(m, endoff, 3766 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3767 3768 /* 3769 * Assert that no previously invalid block that is now being validated 3770 * is already dirty. 3771 */ 3772 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 3773 ("vm_page_set_valid_range: page %p is dirty", m)); 3774 3775 /* 3776 * Set valid bits inclusive of any overlap. 3777 */ 3778 m->valid |= vm_page_bits(base, size); 3779 } 3780 3781 /* 3782 * Clear the given bits from the specified page's dirty field. 3783 */ 3784 static __inline void 3785 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 3786 { 3787 uintptr_t addr; 3788 #if PAGE_SIZE < 16384 3789 int shift; 3790 #endif 3791 3792 /* 3793 * If the object is locked and the page is neither exclusive busy nor 3794 * write mapped, then the page's dirty field cannot possibly be 3795 * set by a concurrent pmap operation. 3796 */ 3797 VM_OBJECT_ASSERT_WLOCKED(m->object); 3798 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 3799 m->dirty &= ~pagebits; 3800 else { 3801 /* 3802 * The pmap layer can call vm_page_dirty() without 3803 * holding a distinguished lock. The combination of 3804 * the object's lock and an atomic operation suffice 3805 * to guarantee consistency of the page dirty field. 3806 * 3807 * For PAGE_SIZE == 32768 case, compiler already 3808 * properly aligns the dirty field, so no forcible 3809 * alignment is needed. Only require existence of 3810 * atomic_clear_64 when page size is 32768. 3811 */ 3812 addr = (uintptr_t)&m->dirty; 3813 #if PAGE_SIZE == 32768 3814 atomic_clear_64((uint64_t *)addr, pagebits); 3815 #elif PAGE_SIZE == 16384 3816 atomic_clear_32((uint32_t *)addr, pagebits); 3817 #else /* PAGE_SIZE <= 8192 */ 3818 /* 3819 * Use a trick to perform a 32-bit atomic on the 3820 * containing aligned word, to not depend on the existence 3821 * of atomic_clear_{8, 16}. 3822 */ 3823 shift = addr & (sizeof(uint32_t) - 1); 3824 #if BYTE_ORDER == BIG_ENDIAN 3825 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 3826 #else 3827 shift *= NBBY; 3828 #endif 3829 addr &= ~(sizeof(uint32_t) - 1); 3830 atomic_clear_32((uint32_t *)addr, pagebits << shift); 3831 #endif /* PAGE_SIZE */ 3832 } 3833 } 3834 3835 /* 3836 * vm_page_set_validclean: 3837 * 3838 * Sets portions of a page valid and clean. The arguments are expected 3839 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3840 * of any partial chunks touched by the range. The invalid portion of 3841 * such chunks will be zero'd. 3842 * 3843 * (base + size) must be less then or equal to PAGE_SIZE. 3844 */ 3845 void 3846 vm_page_set_validclean(vm_page_t m, int base, int size) 3847 { 3848 vm_page_bits_t oldvalid, pagebits; 3849 int endoff, frag; 3850 3851 VM_OBJECT_ASSERT_WLOCKED(m->object); 3852 if (size == 0) /* handle degenerate case */ 3853 return; 3854 3855 /* 3856 * If the base is not DEV_BSIZE aligned and the valid 3857 * bit is clear, we have to zero out a portion of the 3858 * first block. 3859 */ 3860 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3861 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 3862 pmap_zero_page_area(m, frag, base - frag); 3863 3864 /* 3865 * If the ending offset is not DEV_BSIZE aligned and the 3866 * valid bit is clear, we have to zero out a portion of 3867 * the last block. 3868 */ 3869 endoff = base + size; 3870 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3871 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 3872 pmap_zero_page_area(m, endoff, 3873 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3874 3875 /* 3876 * Set valid, clear dirty bits. If validating the entire 3877 * page we can safely clear the pmap modify bit. We also 3878 * use this opportunity to clear the VPO_NOSYNC flag. If a process 3879 * takes a write fault on a MAP_NOSYNC memory area the flag will 3880 * be set again. 3881 * 3882 * We set valid bits inclusive of any overlap, but we can only 3883 * clear dirty bits for DEV_BSIZE chunks that are fully within 3884 * the range. 3885 */ 3886 oldvalid = m->valid; 3887 pagebits = vm_page_bits(base, size); 3888 m->valid |= pagebits; 3889 #if 0 /* NOT YET */ 3890 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 3891 frag = DEV_BSIZE - frag; 3892 base += frag; 3893 size -= frag; 3894 if (size < 0) 3895 size = 0; 3896 } 3897 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 3898 #endif 3899 if (base == 0 && size == PAGE_SIZE) { 3900 /* 3901 * The page can only be modified within the pmap if it is 3902 * mapped, and it can only be mapped if it was previously 3903 * fully valid. 3904 */ 3905 if (oldvalid == VM_PAGE_BITS_ALL) 3906 /* 3907 * Perform the pmap_clear_modify() first. Otherwise, 3908 * a concurrent pmap operation, such as 3909 * pmap_protect(), could clear a modification in the 3910 * pmap and set the dirty field on the page before 3911 * pmap_clear_modify() had begun and after the dirty 3912 * field was cleared here. 3913 */ 3914 pmap_clear_modify(m); 3915 m->dirty = 0; 3916 m->oflags &= ~VPO_NOSYNC; 3917 } else if (oldvalid != VM_PAGE_BITS_ALL) 3918 m->dirty &= ~pagebits; 3919 else 3920 vm_page_clear_dirty_mask(m, pagebits); 3921 } 3922 3923 void 3924 vm_page_clear_dirty(vm_page_t m, int base, int size) 3925 { 3926 3927 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3928 } 3929 3930 /* 3931 * vm_page_set_invalid: 3932 * 3933 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3934 * valid and dirty bits for the effected areas are cleared. 3935 */ 3936 void 3937 vm_page_set_invalid(vm_page_t m, int base, int size) 3938 { 3939 vm_page_bits_t bits; 3940 vm_object_t object; 3941 3942 object = m->object; 3943 VM_OBJECT_ASSERT_WLOCKED(object); 3944 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3945 size >= object->un_pager.vnp.vnp_size) 3946 bits = VM_PAGE_BITS_ALL; 3947 else 3948 bits = vm_page_bits(base, size); 3949 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3950 bits != 0) 3951 pmap_remove_all(m); 3952 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3953 !pmap_page_is_mapped(m), 3954 ("vm_page_set_invalid: page %p is mapped", m)); 3955 m->valid &= ~bits; 3956 m->dirty &= ~bits; 3957 } 3958 3959 /* 3960 * vm_page_zero_invalid() 3961 * 3962 * The kernel assumes that the invalid portions of a page contain 3963 * garbage, but such pages can be mapped into memory by user code. 3964 * When this occurs, we must zero out the non-valid portions of the 3965 * page so user code sees what it expects. 3966 * 3967 * Pages are most often semi-valid when the end of a file is mapped 3968 * into memory and the file's size is not page aligned. 3969 */ 3970 void 3971 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3972 { 3973 int b; 3974 int i; 3975 3976 VM_OBJECT_ASSERT_WLOCKED(m->object); 3977 /* 3978 * Scan the valid bits looking for invalid sections that 3979 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3980 * valid bit may be set ) have already been zeroed by 3981 * vm_page_set_validclean(). 3982 */ 3983 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3984 if (i == (PAGE_SIZE / DEV_BSIZE) || 3985 (m->valid & ((vm_page_bits_t)1 << i))) { 3986 if (i > b) { 3987 pmap_zero_page_area(m, 3988 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3989 } 3990 b = i + 1; 3991 } 3992 } 3993 3994 /* 3995 * setvalid is TRUE when we can safely set the zero'd areas 3996 * as being valid. We can do this if there are no cache consistancy 3997 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3998 */ 3999 if (setvalid) 4000 m->valid = VM_PAGE_BITS_ALL; 4001 } 4002 4003 /* 4004 * vm_page_is_valid: 4005 * 4006 * Is (partial) page valid? Note that the case where size == 0 4007 * will return FALSE in the degenerate case where the page is 4008 * entirely invalid, and TRUE otherwise. 4009 */ 4010 int 4011 vm_page_is_valid(vm_page_t m, int base, int size) 4012 { 4013 vm_page_bits_t bits; 4014 4015 VM_OBJECT_ASSERT_LOCKED(m->object); 4016 bits = vm_page_bits(base, size); 4017 return (m->valid != 0 && (m->valid & bits) == bits); 4018 } 4019 4020 /* 4021 * Returns true if all of the specified predicates are true for the entire 4022 * (super)page and false otherwise. 4023 */ 4024 bool 4025 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 4026 { 4027 vm_object_t object; 4028 int i, npages; 4029 4030 object = m->object; 4031 VM_OBJECT_ASSERT_LOCKED(object); 4032 npages = atop(pagesizes[m->psind]); 4033 4034 /* 4035 * The physically contiguous pages that make up a superpage, i.e., a 4036 * page with a page size index ("psind") greater than zero, will 4037 * occupy adjacent entries in vm_page_array[]. 4038 */ 4039 for (i = 0; i < npages; i++) { 4040 /* Always test object consistency, including "skip_m". */ 4041 if (m[i].object != object) 4042 return (false); 4043 if (&m[i] == skip_m) 4044 continue; 4045 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 4046 return (false); 4047 if ((flags & PS_ALL_DIRTY) != 0) { 4048 /* 4049 * Calling vm_page_test_dirty() or pmap_is_modified() 4050 * might stop this case from spuriously returning 4051 * "false". However, that would require a write lock 4052 * on the object containing "m[i]". 4053 */ 4054 if (m[i].dirty != VM_PAGE_BITS_ALL) 4055 return (false); 4056 } 4057 if ((flags & PS_ALL_VALID) != 0 && 4058 m[i].valid != VM_PAGE_BITS_ALL) 4059 return (false); 4060 } 4061 return (true); 4062 } 4063 4064 /* 4065 * Set the page's dirty bits if the page is modified. 4066 */ 4067 void 4068 vm_page_test_dirty(vm_page_t m) 4069 { 4070 4071 VM_OBJECT_ASSERT_WLOCKED(m->object); 4072 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 4073 vm_page_dirty(m); 4074 } 4075 4076 void 4077 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 4078 { 4079 4080 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 4081 } 4082 4083 void 4084 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 4085 { 4086 4087 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 4088 } 4089 4090 int 4091 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 4092 { 4093 4094 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 4095 } 4096 4097 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 4098 void 4099 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 4100 { 4101 4102 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 4103 } 4104 4105 void 4106 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 4107 { 4108 4109 mtx_assert_(vm_page_lockptr(m), a, file, line); 4110 } 4111 #endif 4112 4113 #ifdef INVARIANTS 4114 void 4115 vm_page_object_lock_assert(vm_page_t m) 4116 { 4117 4118 /* 4119 * Certain of the page's fields may only be modified by the 4120 * holder of the containing object's lock or the exclusive busy. 4121 * holder. Unfortunately, the holder of the write busy is 4122 * not recorded, and thus cannot be checked here. 4123 */ 4124 if (m->object != NULL && !vm_page_xbusied(m)) 4125 VM_OBJECT_ASSERT_WLOCKED(m->object); 4126 } 4127 4128 void 4129 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 4130 { 4131 4132 if ((bits & PGA_WRITEABLE) == 0) 4133 return; 4134 4135 /* 4136 * The PGA_WRITEABLE flag can only be set if the page is 4137 * managed, is exclusively busied or the object is locked. 4138 * Currently, this flag is only set by pmap_enter(). 4139 */ 4140 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 4141 ("PGA_WRITEABLE on unmanaged page")); 4142 if (!vm_page_xbusied(m)) 4143 VM_OBJECT_ASSERT_LOCKED(m->object); 4144 } 4145 #endif 4146 4147 #include "opt_ddb.h" 4148 #ifdef DDB 4149 #include <sys/kernel.h> 4150 4151 #include <ddb/ddb.h> 4152 4153 DB_SHOW_COMMAND(page, vm_page_print_page_info) 4154 { 4155 4156 db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); 4157 db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); 4158 db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); 4159 db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); 4160 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 4161 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 4162 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 4163 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 4164 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 4165 } 4166 4167 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 4168 { 4169 int dom; 4170 4171 db_printf("pq_free %d\n", vm_free_count()); 4172 for (dom = 0; dom < vm_ndomains; dom++) { 4173 db_printf( 4174 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 4175 dom, 4176 vm_dom[dom].vmd_page_count, 4177 vm_dom[dom].vmd_free_count, 4178 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 4179 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 4180 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 4181 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 4182 } 4183 } 4184 4185 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 4186 { 4187 vm_page_t m; 4188 boolean_t phys; 4189 4190 if (!have_addr) { 4191 db_printf("show pginfo addr\n"); 4192 return; 4193 } 4194 4195 phys = strchr(modif, 'p') != NULL; 4196 if (phys) 4197 m = PHYS_TO_VM_PAGE(addr); 4198 else 4199 m = (vm_page_t)addr; 4200 db_printf( 4201 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 4202 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 4203 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 4204 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 4205 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 4206 } 4207 #endif /* DDB */ 4208