1 /*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 3. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 34 */ 35 36 /*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * GENERAL RULES ON VM_PAGE MANIPULATION 65 * 66 * - A page queue lock is required when adding or removing a page from a 67 * page queue regardless of other locks or the busy state of a page. 68 * 69 * * In general, no thread besides the page daemon can acquire or 70 * hold more than one page queue lock at a time. 71 * 72 * * The page daemon can acquire and hold any pair of page queue 73 * locks in any order. 74 * 75 * - The object lock is required when inserting or removing 76 * pages from an object (vm_page_insert() or vm_page_remove()). 77 * 78 */ 79 80 /* 81 * Resident memory management module. 82 */ 83 84 #include <sys/cdefs.h> 85 __FBSDID("$FreeBSD$"); 86 87 #include "opt_vm.h" 88 89 #include <sys/param.h> 90 #include <sys/systm.h> 91 #include <sys/lock.h> 92 #include <sys/kernel.h> 93 #include <sys/limits.h> 94 #include <sys/linker.h> 95 #include <sys/malloc.h> 96 #include <sys/mman.h> 97 #include <sys/msgbuf.h> 98 #include <sys/mutex.h> 99 #include <sys/proc.h> 100 #include <sys/rwlock.h> 101 #include <sys/sbuf.h> 102 #include <sys/smp.h> 103 #include <sys/sysctl.h> 104 #include <sys/vmmeter.h> 105 #include <sys/vnode.h> 106 107 #include <vm/vm.h> 108 #include <vm/pmap.h> 109 #include <vm/vm_param.h> 110 #include <vm/vm_kern.h> 111 #include <vm/vm_object.h> 112 #include <vm/vm_page.h> 113 #include <vm/vm_pageout.h> 114 #include <vm/vm_pager.h> 115 #include <vm/vm_phys.h> 116 #include <vm/vm_radix.h> 117 #include <vm/vm_reserv.h> 118 #include <vm/vm_extern.h> 119 #include <vm/uma.h> 120 #include <vm/uma_int.h> 121 122 #include <machine/md_var.h> 123 124 /* 125 * Associated with page of user-allocatable memory is a 126 * page structure. 127 */ 128 129 struct vm_domain vm_dom[MAXMEMDOM]; 130 struct mtx_padalign __exclusive_cache_line vm_page_queue_free_mtx; 131 132 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 133 134 /* 135 * bogus page -- for I/O to/from partially complete buffers, 136 * or for paging into sparsely invalid regions. 137 */ 138 vm_page_t bogus_page; 139 140 vm_page_t vm_page_array; 141 long vm_page_array_size; 142 long first_page; 143 144 static int boot_pages = UMA_BOOT_PAGES; 145 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 146 &boot_pages, 0, 147 "number of pages allocated for bootstrapping the VM system"); 148 149 static int pa_tryrelock_restart; 150 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 151 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 152 153 static TAILQ_HEAD(, vm_page) blacklist_head; 154 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 155 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 156 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 157 158 /* Is the page daemon waiting for free pages? */ 159 static int vm_pageout_pages_needed; 160 161 static uma_zone_t fakepg_zone; 162 163 static void vm_page_alloc_check(vm_page_t m); 164 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 165 static void vm_page_enqueue(uint8_t queue, vm_page_t m); 166 static void vm_page_free_phys(vm_page_t m); 167 static void vm_page_free_wakeup(void); 168 static void vm_page_init(void *dummy); 169 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 170 vm_pindex_t pindex, vm_page_t mpred); 171 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 172 vm_page_t mpred); 173 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 174 vm_paddr_t high); 175 176 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); 177 178 static void 179 vm_page_init(void *dummy) 180 { 181 182 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 183 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 184 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 185 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 186 } 187 188 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 189 #if PAGE_SIZE == 32768 190 #ifdef CTASSERT 191 CTASSERT(sizeof(u_long) >= 8); 192 #endif 193 #endif 194 195 /* 196 * Try to acquire a physical address lock while a pmap is locked. If we 197 * fail to trylock we unlock and lock the pmap directly and cache the 198 * locked pa in *locked. The caller should then restart their loop in case 199 * the virtual to physical mapping has changed. 200 */ 201 int 202 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 203 { 204 vm_paddr_t lockpa; 205 206 lockpa = *locked; 207 *locked = pa; 208 if (lockpa) { 209 PA_LOCK_ASSERT(lockpa, MA_OWNED); 210 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 211 return (0); 212 PA_UNLOCK(lockpa); 213 } 214 if (PA_TRYLOCK(pa)) 215 return (0); 216 PMAP_UNLOCK(pmap); 217 atomic_add_int(&pa_tryrelock_restart, 1); 218 PA_LOCK(pa); 219 PMAP_LOCK(pmap); 220 return (EAGAIN); 221 } 222 223 /* 224 * vm_set_page_size: 225 * 226 * Sets the page size, perhaps based upon the memory 227 * size. Must be called before any use of page-size 228 * dependent functions. 229 */ 230 void 231 vm_set_page_size(void) 232 { 233 if (vm_cnt.v_page_size == 0) 234 vm_cnt.v_page_size = PAGE_SIZE; 235 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 236 panic("vm_set_page_size: page size not a power of two"); 237 } 238 239 /* 240 * vm_page_blacklist_next: 241 * 242 * Find the next entry in the provided string of blacklist 243 * addresses. Entries are separated by space, comma, or newline. 244 * If an invalid integer is encountered then the rest of the 245 * string is skipped. Updates the list pointer to the next 246 * character, or NULL if the string is exhausted or invalid. 247 */ 248 static vm_paddr_t 249 vm_page_blacklist_next(char **list, char *end) 250 { 251 vm_paddr_t bad; 252 char *cp, *pos; 253 254 if (list == NULL || *list == NULL) 255 return (0); 256 if (**list =='\0') { 257 *list = NULL; 258 return (0); 259 } 260 261 /* 262 * If there's no end pointer then the buffer is coming from 263 * the kenv and we know it's null-terminated. 264 */ 265 if (end == NULL) 266 end = *list + strlen(*list); 267 268 /* Ensure that strtoq() won't walk off the end */ 269 if (*end != '\0') { 270 if (*end == '\n' || *end == ' ' || *end == ',') 271 *end = '\0'; 272 else { 273 printf("Blacklist not terminated, skipping\n"); 274 *list = NULL; 275 return (0); 276 } 277 } 278 279 for (pos = *list; *pos != '\0'; pos = cp) { 280 bad = strtoq(pos, &cp, 0); 281 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 282 if (bad == 0) { 283 if (++cp < end) 284 continue; 285 else 286 break; 287 } 288 } else 289 break; 290 if (*cp == '\0' || ++cp >= end) 291 *list = NULL; 292 else 293 *list = cp; 294 return (trunc_page(bad)); 295 } 296 printf("Garbage in RAM blacklist, skipping\n"); 297 *list = NULL; 298 return (0); 299 } 300 301 /* 302 * vm_page_blacklist_check: 303 * 304 * Iterate through the provided string of blacklist addresses, pulling 305 * each entry out of the physical allocator free list and putting it 306 * onto a list for reporting via the vm.page_blacklist sysctl. 307 */ 308 static void 309 vm_page_blacklist_check(char *list, char *end) 310 { 311 vm_paddr_t pa; 312 vm_page_t m; 313 char *next; 314 int ret; 315 316 next = list; 317 while (next != NULL) { 318 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 319 continue; 320 m = vm_phys_paddr_to_vm_page(pa); 321 if (m == NULL) 322 continue; 323 mtx_lock(&vm_page_queue_free_mtx); 324 ret = vm_phys_unfree_page(m); 325 mtx_unlock(&vm_page_queue_free_mtx); 326 if (ret == TRUE) { 327 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 328 if (bootverbose) 329 printf("Skipping page with pa 0x%jx\n", 330 (uintmax_t)pa); 331 } 332 } 333 } 334 335 /* 336 * vm_page_blacklist_load: 337 * 338 * Search for a special module named "ram_blacklist". It'll be a 339 * plain text file provided by the user via the loader directive 340 * of the same name. 341 */ 342 static void 343 vm_page_blacklist_load(char **list, char **end) 344 { 345 void *mod; 346 u_char *ptr; 347 u_int len; 348 349 mod = NULL; 350 ptr = NULL; 351 352 mod = preload_search_by_type("ram_blacklist"); 353 if (mod != NULL) { 354 ptr = preload_fetch_addr(mod); 355 len = preload_fetch_size(mod); 356 } 357 *list = ptr; 358 if (ptr != NULL) 359 *end = ptr + len; 360 else 361 *end = NULL; 362 return; 363 } 364 365 static int 366 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 367 { 368 vm_page_t m; 369 struct sbuf sbuf; 370 int error, first; 371 372 first = 1; 373 error = sysctl_wire_old_buffer(req, 0); 374 if (error != 0) 375 return (error); 376 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 377 TAILQ_FOREACH(m, &blacklist_head, listq) { 378 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 379 (uintmax_t)m->phys_addr); 380 first = 0; 381 } 382 error = sbuf_finish(&sbuf); 383 sbuf_delete(&sbuf); 384 return (error); 385 } 386 387 static void 388 vm_page_domain_init(struct vm_domain *vmd) 389 { 390 struct vm_pagequeue *pq; 391 int i; 392 393 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 394 "vm inactive pagequeue"; 395 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 396 &vm_cnt.v_inactive_count; 397 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 398 "vm active pagequeue"; 399 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 400 &vm_cnt.v_active_count; 401 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 402 "vm laundry pagequeue"; 403 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) = 404 &vm_cnt.v_laundry_count; 405 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = 406 "vm unswappable pagequeue"; 407 /* Unswappable dirty pages are counted as being in the laundry. */ 408 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_vcnt) = 409 &vm_cnt.v_laundry_count; 410 vmd->vmd_page_count = 0; 411 vmd->vmd_free_count = 0; 412 vmd->vmd_segs = 0; 413 vmd->vmd_oom = FALSE; 414 for (i = 0; i < PQ_COUNT; i++) { 415 pq = &vmd->vmd_pagequeues[i]; 416 TAILQ_INIT(&pq->pq_pl); 417 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 418 MTX_DEF | MTX_DUPOK); 419 } 420 } 421 422 /* 423 * vm_page_startup: 424 * 425 * Initializes the resident memory module. Allocates physical memory for 426 * bootstrapping UMA and some data structures that are used to manage 427 * physical pages. Initializes these structures, and populates the free 428 * page queues. 429 */ 430 vm_offset_t 431 vm_page_startup(vm_offset_t vaddr) 432 { 433 struct vm_domain *vmd; 434 struct vm_phys_seg *seg; 435 vm_page_t m; 436 char *list, *listend; 437 vm_offset_t mapped; 438 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; 439 vm_paddr_t biggestsize, last_pa, pa; 440 u_long pagecount; 441 int biggestone, i, pages_per_zone, segind; 442 443 biggestsize = 0; 444 biggestone = 0; 445 vaddr = round_page(vaddr); 446 447 for (i = 0; phys_avail[i + 1]; i += 2) { 448 phys_avail[i] = round_page(phys_avail[i]); 449 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 450 } 451 for (i = 0; phys_avail[i + 1]; i += 2) { 452 size = phys_avail[i + 1] - phys_avail[i]; 453 if (size > biggestsize) { 454 biggestone = i; 455 biggestsize = size; 456 } 457 } 458 459 end = phys_avail[biggestone+1]; 460 461 /* 462 * Initialize the page and queue locks. 463 */ 464 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 465 for (i = 0; i < PA_LOCK_COUNT; i++) 466 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 467 for (i = 0; i < vm_ndomains; i++) 468 vm_page_domain_init(&vm_dom[i]); 469 470 /* 471 * Almost all of the pages needed for bootstrapping UMA are used 472 * for zone structures, so if the number of CPUs results in those 473 * structures taking more than one page each, we set aside more pages 474 * in proportion to the zone structure size. 475 */ 476 pages_per_zone = howmany(sizeof(struct uma_zone) + 477 sizeof(struct uma_cache) * (mp_maxid + 1) + 478 roundup2(sizeof(struct uma_slab), sizeof(void *)), UMA_SLAB_SIZE); 479 if (pages_per_zone > 1) { 480 /* Reserve more pages so that we don't run out. */ 481 boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone; 482 } 483 484 /* 485 * Allocate memory for use when boot strapping the kernel memory 486 * allocator. 487 * 488 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 489 * manually fetch the value. 490 */ 491 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 492 new_end = end - (boot_pages * UMA_SLAB_SIZE); 493 new_end = trunc_page(new_end); 494 mapped = pmap_map(&vaddr, new_end, end, 495 VM_PROT_READ | VM_PROT_WRITE); 496 bzero((void *)mapped, end - new_end); 497 uma_startup((void *)mapped, boot_pages); 498 499 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 500 defined(__i386__) || defined(__mips__) 501 /* 502 * Allocate a bitmap to indicate that a random physical page 503 * needs to be included in a minidump. 504 * 505 * The amd64 port needs this to indicate which direct map pages 506 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 507 * 508 * However, i386 still needs this workspace internally within the 509 * minidump code. In theory, they are not needed on i386, but are 510 * included should the sf_buf code decide to use them. 511 */ 512 last_pa = 0; 513 for (i = 0; dump_avail[i + 1] != 0; i += 2) 514 if (dump_avail[i + 1] > last_pa) 515 last_pa = dump_avail[i + 1]; 516 page_range = last_pa / PAGE_SIZE; 517 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 518 new_end -= vm_page_dump_size; 519 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 520 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 521 bzero((void *)vm_page_dump, vm_page_dump_size); 522 #else 523 (void)last_pa; 524 #endif 525 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 526 /* 527 * Include the UMA bootstrap pages and vm_page_dump in a crash dump. 528 * When pmap_map() uses the direct map, they are not automatically 529 * included. 530 */ 531 for (pa = new_end; pa < end; pa += PAGE_SIZE) 532 dump_add_page(pa); 533 #endif 534 phys_avail[biggestone + 1] = new_end; 535 #ifdef __amd64__ 536 /* 537 * Request that the physical pages underlying the message buffer be 538 * included in a crash dump. Since the message buffer is accessed 539 * through the direct map, they are not automatically included. 540 */ 541 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 542 last_pa = pa + round_page(msgbufsize); 543 while (pa < last_pa) { 544 dump_add_page(pa); 545 pa += PAGE_SIZE; 546 } 547 #endif 548 /* 549 * Compute the number of pages of memory that will be available for 550 * use, taking into account the overhead of a page structure per page. 551 * In other words, solve 552 * "available physical memory" - round_page(page_range * 553 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 554 * for page_range. 555 */ 556 low_avail = phys_avail[0]; 557 high_avail = phys_avail[1]; 558 for (i = 0; i < vm_phys_nsegs; i++) { 559 if (vm_phys_segs[i].start < low_avail) 560 low_avail = vm_phys_segs[i].start; 561 if (vm_phys_segs[i].end > high_avail) 562 high_avail = vm_phys_segs[i].end; 563 } 564 /* Skip the first chunk. It is already accounted for. */ 565 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 566 if (phys_avail[i] < low_avail) 567 low_avail = phys_avail[i]; 568 if (phys_avail[i + 1] > high_avail) 569 high_avail = phys_avail[i + 1]; 570 } 571 first_page = low_avail / PAGE_SIZE; 572 #ifdef VM_PHYSSEG_SPARSE 573 size = 0; 574 for (i = 0; i < vm_phys_nsegs; i++) 575 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 576 for (i = 0; phys_avail[i + 1] != 0; i += 2) 577 size += phys_avail[i + 1] - phys_avail[i]; 578 #elif defined(VM_PHYSSEG_DENSE) 579 size = high_avail - low_avail; 580 #else 581 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 582 #endif 583 584 #ifdef VM_PHYSSEG_DENSE 585 /* 586 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 587 * the overhead of a page structure per page only if vm_page_array is 588 * allocated from the last physical memory chunk. Otherwise, we must 589 * allocate page structures representing the physical memory 590 * underlying vm_page_array, even though they will not be used. 591 */ 592 if (new_end != high_avail) 593 page_range = size / PAGE_SIZE; 594 else 595 #endif 596 { 597 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 598 599 /* 600 * If the partial bytes remaining are large enough for 601 * a page (PAGE_SIZE) without a corresponding 602 * 'struct vm_page', then new_end will contain an 603 * extra page after subtracting the length of the VM 604 * page array. Compensate by subtracting an extra 605 * page from new_end. 606 */ 607 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 608 if (new_end == high_avail) 609 high_avail -= PAGE_SIZE; 610 new_end -= PAGE_SIZE; 611 } 612 } 613 end = new_end; 614 615 /* 616 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 617 * However, because this page is allocated from KVM, out-of-bounds 618 * accesses using the direct map will not be trapped. 619 */ 620 vaddr += PAGE_SIZE; 621 622 /* 623 * Allocate physical memory for the page structures, and map it. 624 */ 625 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 626 mapped = pmap_map(&vaddr, new_end, end, 627 VM_PROT_READ | VM_PROT_WRITE); 628 vm_page_array = (vm_page_t)mapped; 629 vm_page_array_size = page_range; 630 631 #if VM_NRESERVLEVEL > 0 632 /* 633 * Allocate physical memory for the reservation management system's 634 * data structures, and map it. 635 */ 636 if (high_avail == end) 637 high_avail = new_end; 638 new_end = vm_reserv_startup(&vaddr, new_end, high_avail); 639 #endif 640 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 641 /* 642 * Include vm_page_array and vm_reserv_array in a crash dump. 643 */ 644 for (pa = new_end; pa < end; pa += PAGE_SIZE) 645 dump_add_page(pa); 646 #endif 647 phys_avail[biggestone + 1] = new_end; 648 649 /* 650 * Add physical memory segments corresponding to the available 651 * physical pages. 652 */ 653 for (i = 0; phys_avail[i + 1] != 0; i += 2) 654 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 655 656 /* 657 * Initialize the physical memory allocator. 658 */ 659 vm_phys_init(); 660 661 /* 662 * Initialize the page structures and add every available page to the 663 * physical memory allocator's free lists. 664 */ 665 vm_cnt.v_page_count = 0; 666 vm_cnt.v_free_count = 0; 667 for (segind = 0; segind < vm_phys_nsegs; segind++) { 668 seg = &vm_phys_segs[segind]; 669 for (pa = seg->start; pa < seg->end; pa += PAGE_SIZE) 670 vm_phys_init_page(pa); 671 672 /* 673 * Add the segment to the free lists only if it is covered by 674 * one of the ranges in phys_avail. Because we've added the 675 * ranges to the vm_phys_segs array, we can assume that each 676 * segment is either entirely contained in one of the ranges, 677 * or doesn't overlap any of them. 678 */ 679 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 680 if (seg->start < phys_avail[i] || 681 seg->end > phys_avail[i + 1]) 682 continue; 683 684 m = seg->first_page; 685 pagecount = (u_long)atop(seg->end - seg->start); 686 687 mtx_lock(&vm_page_queue_free_mtx); 688 vm_phys_free_contig(m, pagecount); 689 vm_phys_freecnt_adj(m, (int)pagecount); 690 mtx_unlock(&vm_page_queue_free_mtx); 691 vm_cnt.v_page_count += (u_int)pagecount; 692 693 vmd = &vm_dom[seg->domain]; 694 vmd->vmd_page_count += (u_int)pagecount; 695 vmd->vmd_segs |= 1UL << m->segind; 696 break; 697 } 698 } 699 700 /* 701 * Remove blacklisted pages from the physical memory allocator. 702 */ 703 TAILQ_INIT(&blacklist_head); 704 vm_page_blacklist_load(&list, &listend); 705 vm_page_blacklist_check(list, listend); 706 707 list = kern_getenv("vm.blacklist"); 708 vm_page_blacklist_check(list, NULL); 709 710 freeenv(list); 711 #if VM_NRESERVLEVEL > 0 712 /* 713 * Initialize the reservation management system. 714 */ 715 vm_reserv_init(); 716 #endif 717 return (vaddr); 718 } 719 720 void 721 vm_page_reference(vm_page_t m) 722 { 723 724 vm_page_aflag_set(m, PGA_REFERENCED); 725 } 726 727 /* 728 * vm_page_busy_downgrade: 729 * 730 * Downgrade an exclusive busy page into a single shared busy page. 731 */ 732 void 733 vm_page_busy_downgrade(vm_page_t m) 734 { 735 u_int x; 736 bool locked; 737 738 vm_page_assert_xbusied(m); 739 locked = mtx_owned(vm_page_lockptr(m)); 740 741 for (;;) { 742 x = m->busy_lock; 743 x &= VPB_BIT_WAITERS; 744 if (x != 0 && !locked) 745 vm_page_lock(m); 746 if (atomic_cmpset_rel_int(&m->busy_lock, 747 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 748 break; 749 if (x != 0 && !locked) 750 vm_page_unlock(m); 751 } 752 if (x != 0) { 753 wakeup(m); 754 if (!locked) 755 vm_page_unlock(m); 756 } 757 } 758 759 /* 760 * vm_page_sbusied: 761 * 762 * Return a positive value if the page is shared busied, 0 otherwise. 763 */ 764 int 765 vm_page_sbusied(vm_page_t m) 766 { 767 u_int x; 768 769 x = m->busy_lock; 770 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 771 } 772 773 /* 774 * vm_page_sunbusy: 775 * 776 * Shared unbusy a page. 777 */ 778 void 779 vm_page_sunbusy(vm_page_t m) 780 { 781 u_int x; 782 783 vm_page_lock_assert(m, MA_NOTOWNED); 784 vm_page_assert_sbusied(m); 785 786 for (;;) { 787 x = m->busy_lock; 788 if (VPB_SHARERS(x) > 1) { 789 if (atomic_cmpset_int(&m->busy_lock, x, 790 x - VPB_ONE_SHARER)) 791 break; 792 continue; 793 } 794 if ((x & VPB_BIT_WAITERS) == 0) { 795 KASSERT(x == VPB_SHARERS_WORD(1), 796 ("vm_page_sunbusy: invalid lock state")); 797 if (atomic_cmpset_int(&m->busy_lock, 798 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 799 break; 800 continue; 801 } 802 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 803 ("vm_page_sunbusy: invalid lock state for waiters")); 804 805 vm_page_lock(m); 806 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 807 vm_page_unlock(m); 808 continue; 809 } 810 wakeup(m); 811 vm_page_unlock(m); 812 break; 813 } 814 } 815 816 /* 817 * vm_page_busy_sleep: 818 * 819 * Sleep and release the page lock, using the page pointer as wchan. 820 * This is used to implement the hard-path of busying mechanism. 821 * 822 * The given page must be locked. 823 * 824 * If nonshared is true, sleep only if the page is xbusy. 825 */ 826 void 827 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 828 { 829 u_int x; 830 831 vm_page_assert_locked(m); 832 833 x = m->busy_lock; 834 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 835 ((x & VPB_BIT_WAITERS) == 0 && 836 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 837 vm_page_unlock(m); 838 return; 839 } 840 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 841 } 842 843 /* 844 * vm_page_trysbusy: 845 * 846 * Try to shared busy a page. 847 * If the operation succeeds 1 is returned otherwise 0. 848 * The operation never sleeps. 849 */ 850 int 851 vm_page_trysbusy(vm_page_t m) 852 { 853 u_int x; 854 855 for (;;) { 856 x = m->busy_lock; 857 if ((x & VPB_BIT_SHARED) == 0) 858 return (0); 859 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 860 return (1); 861 } 862 } 863 864 static void 865 vm_page_xunbusy_locked(vm_page_t m) 866 { 867 868 vm_page_assert_xbusied(m); 869 vm_page_assert_locked(m); 870 871 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 872 /* There is a waiter, do wakeup() instead of vm_page_flash(). */ 873 wakeup(m); 874 } 875 876 void 877 vm_page_xunbusy_maybelocked(vm_page_t m) 878 { 879 bool lockacq; 880 881 vm_page_assert_xbusied(m); 882 883 /* 884 * Fast path for unbusy. If it succeeds, we know that there 885 * are no waiters, so we do not need a wakeup. 886 */ 887 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER, 888 VPB_UNBUSIED)) 889 return; 890 891 lockacq = !mtx_owned(vm_page_lockptr(m)); 892 if (lockacq) 893 vm_page_lock(m); 894 vm_page_xunbusy_locked(m); 895 if (lockacq) 896 vm_page_unlock(m); 897 } 898 899 /* 900 * vm_page_xunbusy_hard: 901 * 902 * Called after the first try the exclusive unbusy of a page failed. 903 * It is assumed that the waiters bit is on. 904 */ 905 void 906 vm_page_xunbusy_hard(vm_page_t m) 907 { 908 909 vm_page_assert_xbusied(m); 910 911 vm_page_lock(m); 912 vm_page_xunbusy_locked(m); 913 vm_page_unlock(m); 914 } 915 916 /* 917 * vm_page_flash: 918 * 919 * Wakeup anyone waiting for the page. 920 * The ownership bits do not change. 921 * 922 * The given page must be locked. 923 */ 924 void 925 vm_page_flash(vm_page_t m) 926 { 927 u_int x; 928 929 vm_page_lock_assert(m, MA_OWNED); 930 931 for (;;) { 932 x = m->busy_lock; 933 if ((x & VPB_BIT_WAITERS) == 0) 934 return; 935 if (atomic_cmpset_int(&m->busy_lock, x, 936 x & (~VPB_BIT_WAITERS))) 937 break; 938 } 939 wakeup(m); 940 } 941 942 /* 943 * Avoid releasing and reacquiring the same page lock. 944 */ 945 void 946 vm_page_change_lock(vm_page_t m, struct mtx **mtx) 947 { 948 struct mtx *mtx1; 949 950 mtx1 = vm_page_lockptr(m); 951 if (*mtx == mtx1) 952 return; 953 if (*mtx != NULL) 954 mtx_unlock(*mtx); 955 *mtx = mtx1; 956 mtx_lock(mtx1); 957 } 958 959 /* 960 * Keep page from being freed by the page daemon 961 * much of the same effect as wiring, except much lower 962 * overhead and should be used only for *very* temporary 963 * holding ("wiring"). 964 */ 965 void 966 vm_page_hold(vm_page_t mem) 967 { 968 969 vm_page_lock_assert(mem, MA_OWNED); 970 mem->hold_count++; 971 } 972 973 void 974 vm_page_unhold(vm_page_t mem) 975 { 976 977 vm_page_lock_assert(mem, MA_OWNED); 978 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 979 --mem->hold_count; 980 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 981 vm_page_free_toq(mem); 982 } 983 984 /* 985 * vm_page_unhold_pages: 986 * 987 * Unhold each of the pages that is referenced by the given array. 988 */ 989 void 990 vm_page_unhold_pages(vm_page_t *ma, int count) 991 { 992 struct mtx *mtx; 993 994 mtx = NULL; 995 for (; count != 0; count--) { 996 vm_page_change_lock(*ma, &mtx); 997 vm_page_unhold(*ma); 998 ma++; 999 } 1000 if (mtx != NULL) 1001 mtx_unlock(mtx); 1002 } 1003 1004 vm_page_t 1005 PHYS_TO_VM_PAGE(vm_paddr_t pa) 1006 { 1007 vm_page_t m; 1008 1009 #ifdef VM_PHYSSEG_SPARSE 1010 m = vm_phys_paddr_to_vm_page(pa); 1011 if (m == NULL) 1012 m = vm_phys_fictitious_to_vm_page(pa); 1013 return (m); 1014 #elif defined(VM_PHYSSEG_DENSE) 1015 long pi; 1016 1017 pi = atop(pa); 1018 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1019 m = &vm_page_array[pi - first_page]; 1020 return (m); 1021 } 1022 return (vm_phys_fictitious_to_vm_page(pa)); 1023 #else 1024 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1025 #endif 1026 } 1027 1028 /* 1029 * vm_page_getfake: 1030 * 1031 * Create a fictitious page with the specified physical address and 1032 * memory attribute. The memory attribute is the only the machine- 1033 * dependent aspect of a fictitious page that must be initialized. 1034 */ 1035 vm_page_t 1036 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1037 { 1038 vm_page_t m; 1039 1040 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1041 vm_page_initfake(m, paddr, memattr); 1042 return (m); 1043 } 1044 1045 void 1046 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1047 { 1048 1049 if ((m->flags & PG_FICTITIOUS) != 0) { 1050 /* 1051 * The page's memattr might have changed since the 1052 * previous initialization. Update the pmap to the 1053 * new memattr. 1054 */ 1055 goto memattr; 1056 } 1057 m->phys_addr = paddr; 1058 m->queue = PQ_NONE; 1059 /* Fictitious pages don't use "segind". */ 1060 m->flags = PG_FICTITIOUS; 1061 /* Fictitious pages don't use "order" or "pool". */ 1062 m->oflags = VPO_UNMANAGED; 1063 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1064 m->wire_count = 1; 1065 pmap_page_init(m); 1066 memattr: 1067 pmap_page_set_memattr(m, memattr); 1068 } 1069 1070 /* 1071 * vm_page_putfake: 1072 * 1073 * Release a fictitious page. 1074 */ 1075 void 1076 vm_page_putfake(vm_page_t m) 1077 { 1078 1079 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1080 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1081 ("vm_page_putfake: bad page %p", m)); 1082 uma_zfree(fakepg_zone, m); 1083 } 1084 1085 /* 1086 * vm_page_updatefake: 1087 * 1088 * Update the given fictitious page to the specified physical address and 1089 * memory attribute. 1090 */ 1091 void 1092 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1093 { 1094 1095 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1096 ("vm_page_updatefake: bad page %p", m)); 1097 m->phys_addr = paddr; 1098 pmap_page_set_memattr(m, memattr); 1099 } 1100 1101 /* 1102 * vm_page_free: 1103 * 1104 * Free a page. 1105 */ 1106 void 1107 vm_page_free(vm_page_t m) 1108 { 1109 1110 m->flags &= ~PG_ZERO; 1111 vm_page_free_toq(m); 1112 } 1113 1114 /* 1115 * vm_page_free_zero: 1116 * 1117 * Free a page to the zerod-pages queue 1118 */ 1119 void 1120 vm_page_free_zero(vm_page_t m) 1121 { 1122 1123 m->flags |= PG_ZERO; 1124 vm_page_free_toq(m); 1125 } 1126 1127 /* 1128 * Unbusy and handle the page queueing for a page from a getpages request that 1129 * was optionally read ahead or behind. 1130 */ 1131 void 1132 vm_page_readahead_finish(vm_page_t m) 1133 { 1134 1135 /* We shouldn't put invalid pages on queues. */ 1136 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); 1137 1138 /* 1139 * Since the page is not the actually needed one, whether it should 1140 * be activated or deactivated is not obvious. Empirical results 1141 * have shown that deactivating the page is usually the best choice, 1142 * unless the page is wanted by another thread. 1143 */ 1144 vm_page_lock(m); 1145 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1146 vm_page_activate(m); 1147 else 1148 vm_page_deactivate(m); 1149 vm_page_unlock(m); 1150 vm_page_xunbusy(m); 1151 } 1152 1153 /* 1154 * vm_page_sleep_if_busy: 1155 * 1156 * Sleep and release the page queues lock if the page is busied. 1157 * Returns TRUE if the thread slept. 1158 * 1159 * The given page must be unlocked and object containing it must 1160 * be locked. 1161 */ 1162 int 1163 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1164 { 1165 vm_object_t obj; 1166 1167 vm_page_lock_assert(m, MA_NOTOWNED); 1168 VM_OBJECT_ASSERT_WLOCKED(m->object); 1169 1170 if (vm_page_busied(m)) { 1171 /* 1172 * The page-specific object must be cached because page 1173 * identity can change during the sleep, causing the 1174 * re-lock of a different object. 1175 * It is assumed that a reference to the object is already 1176 * held by the callers. 1177 */ 1178 obj = m->object; 1179 vm_page_lock(m); 1180 VM_OBJECT_WUNLOCK(obj); 1181 vm_page_busy_sleep(m, msg, false); 1182 VM_OBJECT_WLOCK(obj); 1183 return (TRUE); 1184 } 1185 return (FALSE); 1186 } 1187 1188 /* 1189 * vm_page_dirty_KBI: [ internal use only ] 1190 * 1191 * Set all bits in the page's dirty field. 1192 * 1193 * The object containing the specified page must be locked if the 1194 * call is made from the machine-independent layer. 1195 * 1196 * See vm_page_clear_dirty_mask(). 1197 * 1198 * This function should only be called by vm_page_dirty(). 1199 */ 1200 void 1201 vm_page_dirty_KBI(vm_page_t m) 1202 { 1203 1204 /* Refer to this operation by its public name. */ 1205 KASSERT(m->valid == VM_PAGE_BITS_ALL, 1206 ("vm_page_dirty: page is invalid!")); 1207 m->dirty = VM_PAGE_BITS_ALL; 1208 } 1209 1210 /* 1211 * vm_page_insert: [ internal use only ] 1212 * 1213 * Inserts the given mem entry into the object and object list. 1214 * 1215 * The object must be locked. 1216 */ 1217 int 1218 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1219 { 1220 vm_page_t mpred; 1221 1222 VM_OBJECT_ASSERT_WLOCKED(object); 1223 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1224 return (vm_page_insert_after(m, object, pindex, mpred)); 1225 } 1226 1227 /* 1228 * vm_page_insert_after: 1229 * 1230 * Inserts the page "m" into the specified object at offset "pindex". 1231 * 1232 * The page "mpred" must immediately precede the offset "pindex" within 1233 * the specified object. 1234 * 1235 * The object must be locked. 1236 */ 1237 static int 1238 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1239 vm_page_t mpred) 1240 { 1241 vm_page_t msucc; 1242 1243 VM_OBJECT_ASSERT_WLOCKED(object); 1244 KASSERT(m->object == NULL, 1245 ("vm_page_insert_after: page already inserted")); 1246 if (mpred != NULL) { 1247 KASSERT(mpred->object == object, 1248 ("vm_page_insert_after: object doesn't contain mpred")); 1249 KASSERT(mpred->pindex < pindex, 1250 ("vm_page_insert_after: mpred doesn't precede pindex")); 1251 msucc = TAILQ_NEXT(mpred, listq); 1252 } else 1253 msucc = TAILQ_FIRST(&object->memq); 1254 if (msucc != NULL) 1255 KASSERT(msucc->pindex > pindex, 1256 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1257 1258 /* 1259 * Record the object/offset pair in this page 1260 */ 1261 m->object = object; 1262 m->pindex = pindex; 1263 1264 /* 1265 * Now link into the object's ordered list of backed pages. 1266 */ 1267 if (vm_radix_insert(&object->rtree, m)) { 1268 m->object = NULL; 1269 m->pindex = 0; 1270 return (1); 1271 } 1272 vm_page_insert_radixdone(m, object, mpred); 1273 return (0); 1274 } 1275 1276 /* 1277 * vm_page_insert_radixdone: 1278 * 1279 * Complete page "m" insertion into the specified object after the 1280 * radix trie hooking. 1281 * 1282 * The page "mpred" must precede the offset "m->pindex" within the 1283 * specified object. 1284 * 1285 * The object must be locked. 1286 */ 1287 static void 1288 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1289 { 1290 1291 VM_OBJECT_ASSERT_WLOCKED(object); 1292 KASSERT(object != NULL && m->object == object, 1293 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1294 if (mpred != NULL) { 1295 KASSERT(mpred->object == object, 1296 ("vm_page_insert_after: object doesn't contain mpred")); 1297 KASSERT(mpred->pindex < m->pindex, 1298 ("vm_page_insert_after: mpred doesn't precede pindex")); 1299 } 1300 1301 if (mpred != NULL) 1302 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1303 else 1304 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1305 1306 /* 1307 * Show that the object has one more resident page. 1308 */ 1309 object->resident_page_count++; 1310 1311 /* 1312 * Hold the vnode until the last page is released. 1313 */ 1314 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1315 vhold(object->handle); 1316 1317 /* 1318 * Since we are inserting a new and possibly dirty page, 1319 * update the object's OBJ_MIGHTBEDIRTY flag. 1320 */ 1321 if (pmap_page_is_write_mapped(m)) 1322 vm_object_set_writeable_dirty(object); 1323 } 1324 1325 /* 1326 * vm_page_remove: 1327 * 1328 * Removes the specified page from its containing object, but does not 1329 * invalidate any backing storage. 1330 * 1331 * The object must be locked. The page must be locked if it is managed. 1332 */ 1333 void 1334 vm_page_remove(vm_page_t m) 1335 { 1336 vm_object_t object; 1337 vm_page_t mrem; 1338 1339 if ((m->oflags & VPO_UNMANAGED) == 0) 1340 vm_page_assert_locked(m); 1341 if ((object = m->object) == NULL) 1342 return; 1343 VM_OBJECT_ASSERT_WLOCKED(object); 1344 if (vm_page_xbusied(m)) 1345 vm_page_xunbusy_maybelocked(m); 1346 mrem = vm_radix_remove(&object->rtree, m->pindex); 1347 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1348 1349 /* 1350 * Now remove from the object's list of backed pages. 1351 */ 1352 TAILQ_REMOVE(&object->memq, m, listq); 1353 1354 /* 1355 * And show that the object has one fewer resident page. 1356 */ 1357 object->resident_page_count--; 1358 1359 /* 1360 * The vnode may now be recycled. 1361 */ 1362 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1363 vdrop(object->handle); 1364 1365 m->object = NULL; 1366 } 1367 1368 /* 1369 * vm_page_lookup: 1370 * 1371 * Returns the page associated with the object/offset 1372 * pair specified; if none is found, NULL is returned. 1373 * 1374 * The object must be locked. 1375 */ 1376 vm_page_t 1377 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1378 { 1379 1380 VM_OBJECT_ASSERT_LOCKED(object); 1381 return (vm_radix_lookup(&object->rtree, pindex)); 1382 } 1383 1384 /* 1385 * vm_page_find_least: 1386 * 1387 * Returns the page associated with the object with least pindex 1388 * greater than or equal to the parameter pindex, or NULL. 1389 * 1390 * The object must be locked. 1391 */ 1392 vm_page_t 1393 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1394 { 1395 vm_page_t m; 1396 1397 VM_OBJECT_ASSERT_LOCKED(object); 1398 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1399 m = vm_radix_lookup_ge(&object->rtree, pindex); 1400 return (m); 1401 } 1402 1403 /* 1404 * Returns the given page's successor (by pindex) within the object if it is 1405 * resident; if none is found, NULL is returned. 1406 * 1407 * The object must be locked. 1408 */ 1409 vm_page_t 1410 vm_page_next(vm_page_t m) 1411 { 1412 vm_page_t next; 1413 1414 VM_OBJECT_ASSERT_LOCKED(m->object); 1415 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1416 MPASS(next->object == m->object); 1417 if (next->pindex != m->pindex + 1) 1418 next = NULL; 1419 } 1420 return (next); 1421 } 1422 1423 /* 1424 * Returns the given page's predecessor (by pindex) within the object if it is 1425 * resident; if none is found, NULL is returned. 1426 * 1427 * The object must be locked. 1428 */ 1429 vm_page_t 1430 vm_page_prev(vm_page_t m) 1431 { 1432 vm_page_t prev; 1433 1434 VM_OBJECT_ASSERT_LOCKED(m->object); 1435 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1436 MPASS(prev->object == m->object); 1437 if (prev->pindex != m->pindex - 1) 1438 prev = NULL; 1439 } 1440 return (prev); 1441 } 1442 1443 /* 1444 * Uses the page mnew as a replacement for an existing page at index 1445 * pindex which must be already present in the object. 1446 * 1447 * The existing page must not be on a paging queue. 1448 */ 1449 vm_page_t 1450 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1451 { 1452 vm_page_t mold; 1453 1454 VM_OBJECT_ASSERT_WLOCKED(object); 1455 KASSERT(mnew->object == NULL, 1456 ("vm_page_replace: page already in object")); 1457 1458 /* 1459 * This function mostly follows vm_page_insert() and 1460 * vm_page_remove() without the radix, object count and vnode 1461 * dance. Double check such functions for more comments. 1462 */ 1463 1464 mnew->object = object; 1465 mnew->pindex = pindex; 1466 mold = vm_radix_replace(&object->rtree, mnew); 1467 KASSERT(mold->queue == PQ_NONE, 1468 ("vm_page_replace: mold is on a paging queue")); 1469 1470 /* Keep the resident page list in sorted order. */ 1471 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1472 TAILQ_REMOVE(&object->memq, mold, listq); 1473 1474 mold->object = NULL; 1475 vm_page_xunbusy_maybelocked(mold); 1476 1477 /* 1478 * The object's resident_page_count does not change because we have 1479 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1480 */ 1481 if (pmap_page_is_write_mapped(mnew)) 1482 vm_object_set_writeable_dirty(object); 1483 return (mold); 1484 } 1485 1486 /* 1487 * vm_page_rename: 1488 * 1489 * Move the given memory entry from its 1490 * current object to the specified target object/offset. 1491 * 1492 * Note: swap associated with the page must be invalidated by the move. We 1493 * have to do this for several reasons: (1) we aren't freeing the 1494 * page, (2) we are dirtying the page, (3) the VM system is probably 1495 * moving the page from object A to B, and will then later move 1496 * the backing store from A to B and we can't have a conflict. 1497 * 1498 * Note: we *always* dirty the page. It is necessary both for the 1499 * fact that we moved it, and because we may be invalidating 1500 * swap. 1501 * 1502 * The objects must be locked. 1503 */ 1504 int 1505 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1506 { 1507 vm_page_t mpred; 1508 vm_pindex_t opidx; 1509 1510 VM_OBJECT_ASSERT_WLOCKED(new_object); 1511 1512 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1513 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1514 ("vm_page_rename: pindex already renamed")); 1515 1516 /* 1517 * Create a custom version of vm_page_insert() which does not depend 1518 * by m_prev and can cheat on the implementation aspects of the 1519 * function. 1520 */ 1521 opidx = m->pindex; 1522 m->pindex = new_pindex; 1523 if (vm_radix_insert(&new_object->rtree, m)) { 1524 m->pindex = opidx; 1525 return (1); 1526 } 1527 1528 /* 1529 * The operation cannot fail anymore. The removal must happen before 1530 * the listq iterator is tainted. 1531 */ 1532 m->pindex = opidx; 1533 vm_page_lock(m); 1534 vm_page_remove(m); 1535 1536 /* Return back to the new pindex to complete vm_page_insert(). */ 1537 m->pindex = new_pindex; 1538 m->object = new_object; 1539 vm_page_unlock(m); 1540 vm_page_insert_radixdone(m, new_object, mpred); 1541 vm_page_dirty(m); 1542 return (0); 1543 } 1544 1545 /* 1546 * vm_page_alloc: 1547 * 1548 * Allocate and return a page that is associated with the specified 1549 * object and offset pair. By default, this page is exclusive busied. 1550 * 1551 * The caller must always specify an allocation class. 1552 * 1553 * allocation classes: 1554 * VM_ALLOC_NORMAL normal process request 1555 * VM_ALLOC_SYSTEM system *really* needs a page 1556 * VM_ALLOC_INTERRUPT interrupt time request 1557 * 1558 * optional allocation flags: 1559 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1560 * intends to allocate 1561 * VM_ALLOC_NOBUSY do not exclusive busy the page 1562 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1563 * VM_ALLOC_NOOBJ page is not associated with an object and 1564 * should not be exclusive busy 1565 * VM_ALLOC_SBUSY shared busy the allocated page 1566 * VM_ALLOC_WIRED wire the allocated page 1567 * VM_ALLOC_ZERO prefer a zeroed page 1568 * 1569 * This routine may not sleep. 1570 */ 1571 vm_page_t 1572 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1573 { 1574 1575 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1576 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1577 } 1578 1579 /* 1580 * Allocate a page in the specified object with the given page index. To 1581 * optimize insertion of the page into the object, the caller must also specifiy 1582 * the resident page in the object with largest index smaller than the given 1583 * page index, or NULL if no such page exists. 1584 */ 1585 vm_page_t 1586 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, 1587 vm_page_t mpred) 1588 { 1589 vm_page_t m; 1590 int flags, req_class; 1591 u_int free_count; 1592 1593 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1594 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1595 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1596 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1597 ("inconsistent object(%p)/req(%x)", object, req)); 1598 KASSERT(mpred == NULL || mpred->pindex < pindex, 1599 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1600 (uintmax_t)pindex)); 1601 if (object != NULL) 1602 VM_OBJECT_ASSERT_WLOCKED(object); 1603 1604 req_class = req & VM_ALLOC_CLASS_MASK; 1605 1606 /* 1607 * The page daemon is allowed to dig deeper into the free page list. 1608 */ 1609 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1610 req_class = VM_ALLOC_SYSTEM; 1611 1612 /* 1613 * Allocate a page if the number of free pages exceeds the minimum 1614 * for the request class. 1615 */ 1616 mtx_lock(&vm_page_queue_free_mtx); 1617 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1618 (req_class == VM_ALLOC_SYSTEM && 1619 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1620 (req_class == VM_ALLOC_INTERRUPT && 1621 vm_cnt.v_free_count > 0)) { 1622 /* 1623 * Can we allocate the page from a reservation? 1624 */ 1625 #if VM_NRESERVLEVEL > 0 1626 if (object == NULL || (object->flags & (OBJ_COLORED | 1627 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1628 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) 1629 #endif 1630 { 1631 /* 1632 * If not, allocate it from the free page queues. 1633 */ 1634 m = vm_phys_alloc_pages(object != NULL ? 1635 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1636 #if VM_NRESERVLEVEL > 0 1637 if (m == NULL && vm_reserv_reclaim_inactive()) { 1638 m = vm_phys_alloc_pages(object != NULL ? 1639 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1640 0); 1641 } 1642 #endif 1643 } 1644 } else { 1645 /* 1646 * Not allocatable, give up. 1647 */ 1648 mtx_unlock(&vm_page_queue_free_mtx); 1649 atomic_add_int(&vm_pageout_deficit, 1650 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1651 pagedaemon_wakeup(); 1652 return (NULL); 1653 } 1654 1655 /* 1656 * At this point we had better have found a good page. 1657 */ 1658 KASSERT(m != NULL, ("missing page")); 1659 free_count = vm_phys_freecnt_adj(m, -1); 1660 mtx_unlock(&vm_page_queue_free_mtx); 1661 vm_page_alloc_check(m); 1662 1663 /* 1664 * Initialize the page. Only the PG_ZERO flag is inherited. 1665 */ 1666 flags = 0; 1667 if ((req & VM_ALLOC_ZERO) != 0) 1668 flags = PG_ZERO; 1669 flags &= m->flags; 1670 if ((req & VM_ALLOC_NODUMP) != 0) 1671 flags |= PG_NODUMP; 1672 m->flags = flags; 1673 m->aflags = 0; 1674 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1675 VPO_UNMANAGED : 0; 1676 m->busy_lock = VPB_UNBUSIED; 1677 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1678 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1679 if ((req & VM_ALLOC_SBUSY) != 0) 1680 m->busy_lock = VPB_SHARERS_WORD(1); 1681 if (req & VM_ALLOC_WIRED) { 1682 /* 1683 * The page lock is not required for wiring a page until that 1684 * page is inserted into the object. 1685 */ 1686 atomic_add_int(&vm_cnt.v_wire_count, 1); 1687 m->wire_count = 1; 1688 } 1689 m->act_count = 0; 1690 1691 if (object != NULL) { 1692 if (vm_page_insert_after(m, object, pindex, mpred)) { 1693 pagedaemon_wakeup(); 1694 if (req & VM_ALLOC_WIRED) { 1695 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1696 m->wire_count = 0; 1697 } 1698 KASSERT(m->object == NULL, ("page %p has object", m)); 1699 m->oflags = VPO_UNMANAGED; 1700 m->busy_lock = VPB_UNBUSIED; 1701 /* Don't change PG_ZERO. */ 1702 vm_page_free_toq(m); 1703 return (NULL); 1704 } 1705 1706 /* Ignore device objects; the pager sets "memattr" for them. */ 1707 if (object->memattr != VM_MEMATTR_DEFAULT && 1708 (object->flags & OBJ_FICTITIOUS) == 0) 1709 pmap_page_set_memattr(m, object->memattr); 1710 } else 1711 m->pindex = pindex; 1712 1713 /* 1714 * Don't wakeup too often - wakeup the pageout daemon when 1715 * we would be nearly out of memory. 1716 */ 1717 if (vm_paging_needed(free_count)) 1718 pagedaemon_wakeup(); 1719 1720 return (m); 1721 } 1722 1723 /* 1724 * vm_page_alloc_contig: 1725 * 1726 * Allocate a contiguous set of physical pages of the given size "npages" 1727 * from the free lists. All of the physical pages must be at or above 1728 * the given physical address "low" and below the given physical address 1729 * "high". The given value "alignment" determines the alignment of the 1730 * first physical page in the set. If the given value "boundary" is 1731 * non-zero, then the set of physical pages cannot cross any physical 1732 * address boundary that is a multiple of that value. Both "alignment" 1733 * and "boundary" must be a power of two. 1734 * 1735 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1736 * then the memory attribute setting for the physical pages is configured 1737 * to the object's memory attribute setting. Otherwise, the memory 1738 * attribute setting for the physical pages is configured to "memattr", 1739 * overriding the object's memory attribute setting. However, if the 1740 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1741 * memory attribute setting for the physical pages cannot be configured 1742 * to VM_MEMATTR_DEFAULT. 1743 * 1744 * The specified object may not contain fictitious pages. 1745 * 1746 * The caller must always specify an allocation class. 1747 * 1748 * allocation classes: 1749 * VM_ALLOC_NORMAL normal process request 1750 * VM_ALLOC_SYSTEM system *really* needs a page 1751 * VM_ALLOC_INTERRUPT interrupt time request 1752 * 1753 * optional allocation flags: 1754 * VM_ALLOC_NOBUSY do not exclusive busy the page 1755 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1756 * VM_ALLOC_NOOBJ page is not associated with an object and 1757 * should not be exclusive busy 1758 * VM_ALLOC_SBUSY shared busy the allocated page 1759 * VM_ALLOC_WIRED wire the allocated page 1760 * VM_ALLOC_ZERO prefer a zeroed page 1761 * 1762 * This routine may not sleep. 1763 */ 1764 vm_page_t 1765 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1766 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1767 vm_paddr_t boundary, vm_memattr_t memattr) 1768 { 1769 vm_page_t m, m_ret, mpred; 1770 u_int busy_lock, flags, oflags; 1771 int req_class; 1772 1773 mpred = NULL; /* XXX: pacify gcc */ 1774 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1775 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1776 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1777 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1778 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 1779 req)); 1780 if (object != NULL) { 1781 VM_OBJECT_ASSERT_WLOCKED(object); 1782 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 1783 ("vm_page_alloc_contig: object %p has fictitious pages", 1784 object)); 1785 } 1786 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1787 req_class = req & VM_ALLOC_CLASS_MASK; 1788 1789 /* 1790 * The page daemon is allowed to dig deeper into the free page list. 1791 */ 1792 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1793 req_class = VM_ALLOC_SYSTEM; 1794 1795 if (object != NULL) { 1796 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1797 KASSERT(mpred == NULL || mpred->pindex != pindex, 1798 ("vm_page_alloc_contig: pindex already allocated")); 1799 } 1800 1801 /* 1802 * Can we allocate the pages without the number of free pages falling 1803 * below the lower bound for the allocation class? 1804 */ 1805 mtx_lock(&vm_page_queue_free_mtx); 1806 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved || 1807 (req_class == VM_ALLOC_SYSTEM && 1808 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) || 1809 (req_class == VM_ALLOC_INTERRUPT && 1810 vm_cnt.v_free_count >= npages)) { 1811 /* 1812 * Can we allocate the pages from a reservation? 1813 */ 1814 #if VM_NRESERVLEVEL > 0 1815 retry: 1816 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1817 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1818 low, high, alignment, boundary, mpred)) == NULL) 1819 #endif 1820 /* 1821 * If not, allocate them from the free page queues. 1822 */ 1823 m_ret = vm_phys_alloc_contig(npages, low, high, 1824 alignment, boundary); 1825 } else { 1826 mtx_unlock(&vm_page_queue_free_mtx); 1827 atomic_add_int(&vm_pageout_deficit, npages); 1828 pagedaemon_wakeup(); 1829 return (NULL); 1830 } 1831 if (m_ret != NULL) 1832 vm_phys_freecnt_adj(m_ret, -npages); 1833 else { 1834 #if VM_NRESERVLEVEL > 0 1835 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1836 boundary)) 1837 goto retry; 1838 #endif 1839 } 1840 mtx_unlock(&vm_page_queue_free_mtx); 1841 if (m_ret == NULL) 1842 return (NULL); 1843 for (m = m_ret; m < &m_ret[npages]; m++) 1844 vm_page_alloc_check(m); 1845 1846 /* 1847 * Initialize the pages. Only the PG_ZERO flag is inherited. 1848 */ 1849 flags = 0; 1850 if ((req & VM_ALLOC_ZERO) != 0) 1851 flags = PG_ZERO; 1852 if ((req & VM_ALLOC_NODUMP) != 0) 1853 flags |= PG_NODUMP; 1854 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1855 VPO_UNMANAGED : 0; 1856 busy_lock = VPB_UNBUSIED; 1857 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1858 busy_lock = VPB_SINGLE_EXCLUSIVER; 1859 if ((req & VM_ALLOC_SBUSY) != 0) 1860 busy_lock = VPB_SHARERS_WORD(1); 1861 if ((req & VM_ALLOC_WIRED) != 0) 1862 atomic_add_int(&vm_cnt.v_wire_count, npages); 1863 if (object != NULL) { 1864 if (object->memattr != VM_MEMATTR_DEFAULT && 1865 memattr == VM_MEMATTR_DEFAULT) 1866 memattr = object->memattr; 1867 } 1868 for (m = m_ret; m < &m_ret[npages]; m++) { 1869 m->aflags = 0; 1870 m->flags = (m->flags | PG_NODUMP) & flags; 1871 m->busy_lock = busy_lock; 1872 if ((req & VM_ALLOC_WIRED) != 0) 1873 m->wire_count = 1; 1874 m->act_count = 0; 1875 m->oflags = oflags; 1876 if (object != NULL) { 1877 if (vm_page_insert_after(m, object, pindex, mpred)) { 1878 pagedaemon_wakeup(); 1879 if ((req & VM_ALLOC_WIRED) != 0) 1880 atomic_subtract_int( 1881 &vm_cnt.v_wire_count, npages); 1882 KASSERT(m->object == NULL, 1883 ("page %p has object", m)); 1884 mpred = m; 1885 for (m = m_ret; m < &m_ret[npages]; m++) { 1886 if (m <= mpred && 1887 (req & VM_ALLOC_WIRED) != 0) 1888 m->wire_count = 0; 1889 m->oflags = VPO_UNMANAGED; 1890 m->busy_lock = VPB_UNBUSIED; 1891 /* Don't change PG_ZERO. */ 1892 vm_page_free_toq(m); 1893 } 1894 return (NULL); 1895 } 1896 mpred = m; 1897 } else 1898 m->pindex = pindex; 1899 if (memattr != VM_MEMATTR_DEFAULT) 1900 pmap_page_set_memattr(m, memattr); 1901 pindex++; 1902 } 1903 if (vm_paging_needed(vm_cnt.v_free_count)) 1904 pagedaemon_wakeup(); 1905 return (m_ret); 1906 } 1907 1908 /* 1909 * Check a page that has been freshly dequeued from a freelist. 1910 */ 1911 static void 1912 vm_page_alloc_check(vm_page_t m) 1913 { 1914 1915 KASSERT(m->object == NULL, ("page %p has object", m)); 1916 KASSERT(m->queue == PQ_NONE, 1917 ("page %p has unexpected queue %d", m, m->queue)); 1918 KASSERT(m->wire_count == 0, ("page %p is wired", m)); 1919 KASSERT(m->hold_count == 0, ("page %p is held", m)); 1920 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 1921 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 1922 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1923 ("page %p has unexpected memattr %d", 1924 m, pmap_page_get_memattr(m))); 1925 KASSERT(m->valid == 0, ("free page %p is valid", m)); 1926 } 1927 1928 /* 1929 * vm_page_alloc_freelist: 1930 * 1931 * Allocate a physical page from the specified free page list. 1932 * 1933 * The caller must always specify an allocation class. 1934 * 1935 * allocation classes: 1936 * VM_ALLOC_NORMAL normal process request 1937 * VM_ALLOC_SYSTEM system *really* needs a page 1938 * VM_ALLOC_INTERRUPT interrupt time request 1939 * 1940 * optional allocation flags: 1941 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1942 * intends to allocate 1943 * VM_ALLOC_WIRED wire the allocated page 1944 * VM_ALLOC_ZERO prefer a zeroed page 1945 * 1946 * This routine may not sleep. 1947 */ 1948 vm_page_t 1949 vm_page_alloc_freelist(int flind, int req) 1950 { 1951 vm_page_t m; 1952 u_int flags, free_count; 1953 int req_class; 1954 1955 req_class = req & VM_ALLOC_CLASS_MASK; 1956 1957 /* 1958 * The page daemon is allowed to dig deeper into the free page list. 1959 */ 1960 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1961 req_class = VM_ALLOC_SYSTEM; 1962 1963 /* 1964 * Do not allocate reserved pages unless the req has asked for it. 1965 */ 1966 mtx_lock(&vm_page_queue_free_mtx); 1967 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1968 (req_class == VM_ALLOC_SYSTEM && 1969 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1970 (req_class == VM_ALLOC_INTERRUPT && 1971 vm_cnt.v_free_count > 0)) 1972 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1973 else { 1974 mtx_unlock(&vm_page_queue_free_mtx); 1975 atomic_add_int(&vm_pageout_deficit, 1976 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1977 pagedaemon_wakeup(); 1978 return (NULL); 1979 } 1980 if (m == NULL) { 1981 mtx_unlock(&vm_page_queue_free_mtx); 1982 return (NULL); 1983 } 1984 free_count = vm_phys_freecnt_adj(m, -1); 1985 mtx_unlock(&vm_page_queue_free_mtx); 1986 vm_page_alloc_check(m); 1987 1988 /* 1989 * Initialize the page. Only the PG_ZERO flag is inherited. 1990 */ 1991 m->aflags = 0; 1992 flags = 0; 1993 if ((req & VM_ALLOC_ZERO) != 0) 1994 flags = PG_ZERO; 1995 m->flags &= flags; 1996 if ((req & VM_ALLOC_WIRED) != 0) { 1997 /* 1998 * The page lock is not required for wiring a page that does 1999 * not belong to an object. 2000 */ 2001 atomic_add_int(&vm_cnt.v_wire_count, 1); 2002 m->wire_count = 1; 2003 } 2004 /* Unmanaged pages don't use "act_count". */ 2005 m->oflags = VPO_UNMANAGED; 2006 if (vm_paging_needed(free_count)) 2007 pagedaemon_wakeup(); 2008 return (m); 2009 } 2010 2011 #define VPSC_ANY 0 /* No restrictions. */ 2012 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2013 #define VPSC_NOSUPER 2 /* Skip superpages. */ 2014 2015 /* 2016 * vm_page_scan_contig: 2017 * 2018 * Scan vm_page_array[] between the specified entries "m_start" and 2019 * "m_end" for a run of contiguous physical pages that satisfy the 2020 * specified conditions, and return the lowest page in the run. The 2021 * specified "alignment" determines the alignment of the lowest physical 2022 * page in the run. If the specified "boundary" is non-zero, then the 2023 * run of physical pages cannot span a physical address that is a 2024 * multiple of "boundary". 2025 * 2026 * "m_end" is never dereferenced, so it need not point to a vm_page 2027 * structure within vm_page_array[]. 2028 * 2029 * "npages" must be greater than zero. "m_start" and "m_end" must not 2030 * span a hole (or discontiguity) in the physical address space. Both 2031 * "alignment" and "boundary" must be a power of two. 2032 */ 2033 vm_page_t 2034 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2035 u_long alignment, vm_paddr_t boundary, int options) 2036 { 2037 struct mtx *m_mtx; 2038 vm_object_t object; 2039 vm_paddr_t pa; 2040 vm_page_t m, m_run; 2041 #if VM_NRESERVLEVEL > 0 2042 int level; 2043 #endif 2044 int m_inc, order, run_ext, run_len; 2045 2046 KASSERT(npages > 0, ("npages is 0")); 2047 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2048 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2049 m_run = NULL; 2050 run_len = 0; 2051 m_mtx = NULL; 2052 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2053 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2054 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2055 2056 /* 2057 * If the current page would be the start of a run, check its 2058 * physical address against the end, alignment, and boundary 2059 * conditions. If it doesn't satisfy these conditions, either 2060 * terminate the scan or advance to the next page that 2061 * satisfies the failed condition. 2062 */ 2063 if (run_len == 0) { 2064 KASSERT(m_run == NULL, ("m_run != NULL")); 2065 if (m + npages > m_end) 2066 break; 2067 pa = VM_PAGE_TO_PHYS(m); 2068 if ((pa & (alignment - 1)) != 0) { 2069 m_inc = atop(roundup2(pa, alignment) - pa); 2070 continue; 2071 } 2072 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2073 boundary) != 0) { 2074 m_inc = atop(roundup2(pa, boundary) - pa); 2075 continue; 2076 } 2077 } else 2078 KASSERT(m_run != NULL, ("m_run == NULL")); 2079 2080 vm_page_change_lock(m, &m_mtx); 2081 m_inc = 1; 2082 retry: 2083 if (m->wire_count != 0 || m->hold_count != 0) 2084 run_ext = 0; 2085 #if VM_NRESERVLEVEL > 0 2086 else if ((level = vm_reserv_level(m)) >= 0 && 2087 (options & VPSC_NORESERV) != 0) { 2088 run_ext = 0; 2089 /* Advance to the end of the reservation. */ 2090 pa = VM_PAGE_TO_PHYS(m); 2091 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2092 pa); 2093 } 2094 #endif 2095 else if ((object = m->object) != NULL) { 2096 /* 2097 * The page is considered eligible for relocation if 2098 * and only if it could be laundered or reclaimed by 2099 * the page daemon. 2100 */ 2101 if (!VM_OBJECT_TRYRLOCK(object)) { 2102 mtx_unlock(m_mtx); 2103 VM_OBJECT_RLOCK(object); 2104 mtx_lock(m_mtx); 2105 if (m->object != object) { 2106 /* 2107 * The page may have been freed. 2108 */ 2109 VM_OBJECT_RUNLOCK(object); 2110 goto retry; 2111 } else if (m->wire_count != 0 || 2112 m->hold_count != 0) { 2113 run_ext = 0; 2114 goto unlock; 2115 } 2116 } 2117 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2118 ("page %p is PG_UNHOLDFREE", m)); 2119 /* Don't care: PG_NODUMP, PG_ZERO. */ 2120 if (object->type != OBJT_DEFAULT && 2121 object->type != OBJT_SWAP && 2122 object->type != OBJT_VNODE) { 2123 run_ext = 0; 2124 #if VM_NRESERVLEVEL > 0 2125 } else if ((options & VPSC_NOSUPER) != 0 && 2126 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2127 run_ext = 0; 2128 /* Advance to the end of the superpage. */ 2129 pa = VM_PAGE_TO_PHYS(m); 2130 m_inc = atop(roundup2(pa + 1, 2131 vm_reserv_size(level)) - pa); 2132 #endif 2133 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2134 m->queue != PQ_NONE && !vm_page_busied(m)) { 2135 /* 2136 * The page is allocated but eligible for 2137 * relocation. Extend the current run by one 2138 * page. 2139 */ 2140 KASSERT(pmap_page_get_memattr(m) == 2141 VM_MEMATTR_DEFAULT, 2142 ("page %p has an unexpected memattr", m)); 2143 KASSERT((m->oflags & (VPO_SWAPINPROG | 2144 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2145 ("page %p has unexpected oflags", m)); 2146 /* Don't care: VPO_NOSYNC. */ 2147 run_ext = 1; 2148 } else 2149 run_ext = 0; 2150 unlock: 2151 VM_OBJECT_RUNLOCK(object); 2152 #if VM_NRESERVLEVEL > 0 2153 } else if (level >= 0) { 2154 /* 2155 * The page is reserved but not yet allocated. In 2156 * other words, it is still free. Extend the current 2157 * run by one page. 2158 */ 2159 run_ext = 1; 2160 #endif 2161 } else if ((order = m->order) < VM_NFREEORDER) { 2162 /* 2163 * The page is enqueued in the physical memory 2164 * allocator's free page queues. Moreover, it is the 2165 * first page in a power-of-two-sized run of 2166 * contiguous free pages. Add these pages to the end 2167 * of the current run, and jump ahead. 2168 */ 2169 run_ext = 1 << order; 2170 m_inc = 1 << order; 2171 } else { 2172 /* 2173 * Skip the page for one of the following reasons: (1) 2174 * It is enqueued in the physical memory allocator's 2175 * free page queues. However, it is not the first 2176 * page in a run of contiguous free pages. (This case 2177 * rarely occurs because the scan is performed in 2178 * ascending order.) (2) It is not reserved, and it is 2179 * transitioning from free to allocated. (Conversely, 2180 * the transition from allocated to free for managed 2181 * pages is blocked by the page lock.) (3) It is 2182 * allocated but not contained by an object and not 2183 * wired, e.g., allocated by Xen's balloon driver. 2184 */ 2185 run_ext = 0; 2186 } 2187 2188 /* 2189 * Extend or reset the current run of pages. 2190 */ 2191 if (run_ext > 0) { 2192 if (run_len == 0) 2193 m_run = m; 2194 run_len += run_ext; 2195 } else { 2196 if (run_len > 0) { 2197 m_run = NULL; 2198 run_len = 0; 2199 } 2200 } 2201 } 2202 if (m_mtx != NULL) 2203 mtx_unlock(m_mtx); 2204 if (run_len >= npages) 2205 return (m_run); 2206 return (NULL); 2207 } 2208 2209 /* 2210 * vm_page_reclaim_run: 2211 * 2212 * Try to relocate each of the allocated virtual pages within the 2213 * specified run of physical pages to a new physical address. Free the 2214 * physical pages underlying the relocated virtual pages. A virtual page 2215 * is relocatable if and only if it could be laundered or reclaimed by 2216 * the page daemon. Whenever possible, a virtual page is relocated to a 2217 * physical address above "high". 2218 * 2219 * Returns 0 if every physical page within the run was already free or 2220 * just freed by a successful relocation. Otherwise, returns a non-zero 2221 * value indicating why the last attempt to relocate a virtual page was 2222 * unsuccessful. 2223 * 2224 * "req_class" must be an allocation class. 2225 */ 2226 static int 2227 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 2228 vm_paddr_t high) 2229 { 2230 struct mtx *m_mtx; 2231 struct spglist free; 2232 vm_object_t object; 2233 vm_paddr_t pa; 2234 vm_page_t m, m_end, m_new; 2235 int error, order, req; 2236 2237 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2238 ("req_class is not an allocation class")); 2239 SLIST_INIT(&free); 2240 error = 0; 2241 m = m_run; 2242 m_end = m_run + npages; 2243 m_mtx = NULL; 2244 for (; error == 0 && m < m_end; m++) { 2245 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2246 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2247 2248 /* 2249 * Avoid releasing and reacquiring the same page lock. 2250 */ 2251 vm_page_change_lock(m, &m_mtx); 2252 retry: 2253 if (m->wire_count != 0 || m->hold_count != 0) 2254 error = EBUSY; 2255 else if ((object = m->object) != NULL) { 2256 /* 2257 * The page is relocated if and only if it could be 2258 * laundered or reclaimed by the page daemon. 2259 */ 2260 if (!VM_OBJECT_TRYWLOCK(object)) { 2261 mtx_unlock(m_mtx); 2262 VM_OBJECT_WLOCK(object); 2263 mtx_lock(m_mtx); 2264 if (m->object != object) { 2265 /* 2266 * The page may have been freed. 2267 */ 2268 VM_OBJECT_WUNLOCK(object); 2269 goto retry; 2270 } else if (m->wire_count != 0 || 2271 m->hold_count != 0) { 2272 error = EBUSY; 2273 goto unlock; 2274 } 2275 } 2276 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2277 ("page %p is PG_UNHOLDFREE", m)); 2278 /* Don't care: PG_NODUMP, PG_ZERO. */ 2279 if (object->type != OBJT_DEFAULT && 2280 object->type != OBJT_SWAP && 2281 object->type != OBJT_VNODE) 2282 error = EINVAL; 2283 else if (object->memattr != VM_MEMATTR_DEFAULT) 2284 error = EINVAL; 2285 else if (m->queue != PQ_NONE && !vm_page_busied(m)) { 2286 KASSERT(pmap_page_get_memattr(m) == 2287 VM_MEMATTR_DEFAULT, 2288 ("page %p has an unexpected memattr", m)); 2289 KASSERT((m->oflags & (VPO_SWAPINPROG | 2290 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2291 ("page %p has unexpected oflags", m)); 2292 /* Don't care: VPO_NOSYNC. */ 2293 if (m->valid != 0) { 2294 /* 2295 * First, try to allocate a new page 2296 * that is above "high". Failing 2297 * that, try to allocate a new page 2298 * that is below "m_run". Allocate 2299 * the new page between the end of 2300 * "m_run" and "high" only as a last 2301 * resort. 2302 */ 2303 req = req_class | VM_ALLOC_NOOBJ; 2304 if ((m->flags & PG_NODUMP) != 0) 2305 req |= VM_ALLOC_NODUMP; 2306 if (trunc_page(high) != 2307 ~(vm_paddr_t)PAGE_MASK) { 2308 m_new = vm_page_alloc_contig( 2309 NULL, 0, req, 1, 2310 round_page(high), 2311 ~(vm_paddr_t)0, 2312 PAGE_SIZE, 0, 2313 VM_MEMATTR_DEFAULT); 2314 } else 2315 m_new = NULL; 2316 if (m_new == NULL) { 2317 pa = VM_PAGE_TO_PHYS(m_run); 2318 m_new = vm_page_alloc_contig( 2319 NULL, 0, req, 1, 2320 0, pa - 1, PAGE_SIZE, 0, 2321 VM_MEMATTR_DEFAULT); 2322 } 2323 if (m_new == NULL) { 2324 pa += ptoa(npages); 2325 m_new = vm_page_alloc_contig( 2326 NULL, 0, req, 1, 2327 pa, high, PAGE_SIZE, 0, 2328 VM_MEMATTR_DEFAULT); 2329 } 2330 if (m_new == NULL) { 2331 error = ENOMEM; 2332 goto unlock; 2333 } 2334 KASSERT(m_new->wire_count == 0, 2335 ("page %p is wired", m)); 2336 2337 /* 2338 * Replace "m" with the new page. For 2339 * vm_page_replace(), "m" must be busy 2340 * and dequeued. Finally, change "m" 2341 * as if vm_page_free() was called. 2342 */ 2343 if (object->ref_count != 0) 2344 pmap_remove_all(m); 2345 m_new->aflags = m->aflags; 2346 KASSERT(m_new->oflags == VPO_UNMANAGED, 2347 ("page %p is managed", m)); 2348 m_new->oflags = m->oflags & VPO_NOSYNC; 2349 pmap_copy_page(m, m_new); 2350 m_new->valid = m->valid; 2351 m_new->dirty = m->dirty; 2352 m->flags &= ~PG_ZERO; 2353 vm_page_xbusy(m); 2354 vm_page_remque(m); 2355 vm_page_replace_checked(m_new, object, 2356 m->pindex, m); 2357 m->valid = 0; 2358 vm_page_undirty(m); 2359 2360 /* 2361 * The new page must be deactivated 2362 * before the object is unlocked. 2363 */ 2364 vm_page_change_lock(m_new, &m_mtx); 2365 vm_page_deactivate(m_new); 2366 } else { 2367 m->flags &= ~PG_ZERO; 2368 vm_page_remque(m); 2369 vm_page_remove(m); 2370 KASSERT(m->dirty == 0, 2371 ("page %p is dirty", m)); 2372 } 2373 SLIST_INSERT_HEAD(&free, m, plinks.s.ss); 2374 } else 2375 error = EBUSY; 2376 unlock: 2377 VM_OBJECT_WUNLOCK(object); 2378 } else { 2379 mtx_lock(&vm_page_queue_free_mtx); 2380 order = m->order; 2381 if (order < VM_NFREEORDER) { 2382 /* 2383 * The page is enqueued in the physical memory 2384 * allocator's free page queues. Moreover, it 2385 * is the first page in a power-of-two-sized 2386 * run of contiguous free pages. Jump ahead 2387 * to the last page within that run, and 2388 * continue from there. 2389 */ 2390 m += (1 << order) - 1; 2391 } 2392 #if VM_NRESERVLEVEL > 0 2393 else if (vm_reserv_is_page_free(m)) 2394 order = 0; 2395 #endif 2396 mtx_unlock(&vm_page_queue_free_mtx); 2397 if (order == VM_NFREEORDER) 2398 error = EINVAL; 2399 } 2400 } 2401 if (m_mtx != NULL) 2402 mtx_unlock(m_mtx); 2403 if ((m = SLIST_FIRST(&free)) != NULL) { 2404 mtx_lock(&vm_page_queue_free_mtx); 2405 do { 2406 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2407 vm_page_free_phys(m); 2408 } while ((m = SLIST_FIRST(&free)) != NULL); 2409 vm_page_free_wakeup(); 2410 mtx_unlock(&vm_page_queue_free_mtx); 2411 } 2412 return (error); 2413 } 2414 2415 #define NRUNS 16 2416 2417 CTASSERT(powerof2(NRUNS)); 2418 2419 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2420 2421 #define MIN_RECLAIM 8 2422 2423 /* 2424 * vm_page_reclaim_contig: 2425 * 2426 * Reclaim allocated, contiguous physical memory satisfying the specified 2427 * conditions by relocating the virtual pages using that physical memory. 2428 * Returns true if reclamation is successful and false otherwise. Since 2429 * relocation requires the allocation of physical pages, reclamation may 2430 * fail due to a shortage of free pages. When reclamation fails, callers 2431 * are expected to perform VM_WAIT before retrying a failed allocation 2432 * operation, e.g., vm_page_alloc_contig(). 2433 * 2434 * The caller must always specify an allocation class through "req". 2435 * 2436 * allocation classes: 2437 * VM_ALLOC_NORMAL normal process request 2438 * VM_ALLOC_SYSTEM system *really* needs a page 2439 * VM_ALLOC_INTERRUPT interrupt time request 2440 * 2441 * The optional allocation flags are ignored. 2442 * 2443 * "npages" must be greater than zero. Both "alignment" and "boundary" 2444 * must be a power of two. 2445 */ 2446 bool 2447 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2448 u_long alignment, vm_paddr_t boundary) 2449 { 2450 vm_paddr_t curr_low; 2451 vm_page_t m_run, m_runs[NRUNS]; 2452 u_long count, reclaimed; 2453 int error, i, options, req_class; 2454 2455 KASSERT(npages > 0, ("npages is 0")); 2456 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2457 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2458 req_class = req & VM_ALLOC_CLASS_MASK; 2459 2460 /* 2461 * The page daemon is allowed to dig deeper into the free page list. 2462 */ 2463 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2464 req_class = VM_ALLOC_SYSTEM; 2465 2466 /* 2467 * Return if the number of free pages cannot satisfy the requested 2468 * allocation. 2469 */ 2470 count = vm_cnt.v_free_count; 2471 if (count < npages + vm_cnt.v_free_reserved || (count < npages + 2472 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2473 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2474 return (false); 2475 2476 /* 2477 * Scan up to three times, relaxing the restrictions ("options") on 2478 * the reclamation of reservations and superpages each time. 2479 */ 2480 for (options = VPSC_NORESERV;;) { 2481 /* 2482 * Find the highest runs that satisfy the given constraints 2483 * and restrictions, and record them in "m_runs". 2484 */ 2485 curr_low = low; 2486 count = 0; 2487 for (;;) { 2488 m_run = vm_phys_scan_contig(npages, curr_low, high, 2489 alignment, boundary, options); 2490 if (m_run == NULL) 2491 break; 2492 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2493 m_runs[RUN_INDEX(count)] = m_run; 2494 count++; 2495 } 2496 2497 /* 2498 * Reclaim the highest runs in LIFO (descending) order until 2499 * the number of reclaimed pages, "reclaimed", is at least 2500 * MIN_RECLAIM. Reset "reclaimed" each time because each 2501 * reclamation is idempotent, and runs will (likely) recur 2502 * from one scan to the next as restrictions are relaxed. 2503 */ 2504 reclaimed = 0; 2505 for (i = 0; count > 0 && i < NRUNS; i++) { 2506 count--; 2507 m_run = m_runs[RUN_INDEX(count)]; 2508 error = vm_page_reclaim_run(req_class, npages, m_run, 2509 high); 2510 if (error == 0) { 2511 reclaimed += npages; 2512 if (reclaimed >= MIN_RECLAIM) 2513 return (true); 2514 } 2515 } 2516 2517 /* 2518 * Either relax the restrictions on the next scan or return if 2519 * the last scan had no restrictions. 2520 */ 2521 if (options == VPSC_NORESERV) 2522 options = VPSC_NOSUPER; 2523 else if (options == VPSC_NOSUPER) 2524 options = VPSC_ANY; 2525 else if (options == VPSC_ANY) 2526 return (reclaimed != 0); 2527 } 2528 } 2529 2530 /* 2531 * vm_wait: (also see VM_WAIT macro) 2532 * 2533 * Sleep until free pages are available for allocation. 2534 * - Called in various places before memory allocations. 2535 */ 2536 void 2537 vm_wait(void) 2538 { 2539 2540 mtx_lock(&vm_page_queue_free_mtx); 2541 if (curproc == pageproc) { 2542 vm_pageout_pages_needed = 1; 2543 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2544 PDROP | PSWP, "VMWait", 0); 2545 } else { 2546 if (__predict_false(pageproc == NULL)) 2547 panic("vm_wait in early boot"); 2548 if (!vm_pageout_wanted) { 2549 vm_pageout_wanted = true; 2550 wakeup(&vm_pageout_wanted); 2551 } 2552 vm_pages_needed = true; 2553 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2554 "vmwait", 0); 2555 } 2556 } 2557 2558 /* 2559 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2560 * 2561 * Sleep until free pages are available for allocation. 2562 * - Called only in vm_fault so that processes page faulting 2563 * can be easily tracked. 2564 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2565 * processes will be able to grab memory first. Do not change 2566 * this balance without careful testing first. 2567 */ 2568 void 2569 vm_waitpfault(void) 2570 { 2571 2572 mtx_lock(&vm_page_queue_free_mtx); 2573 if (!vm_pageout_wanted) { 2574 vm_pageout_wanted = true; 2575 wakeup(&vm_pageout_wanted); 2576 } 2577 vm_pages_needed = true; 2578 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2579 "pfault", 0); 2580 } 2581 2582 struct vm_pagequeue * 2583 vm_page_pagequeue(vm_page_t m) 2584 { 2585 2586 if (vm_page_in_laundry(m)) 2587 return (&vm_dom[0].vmd_pagequeues[m->queue]); 2588 else 2589 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2590 } 2591 2592 /* 2593 * vm_page_dequeue: 2594 * 2595 * Remove the given page from its current page queue. 2596 * 2597 * The page must be locked. 2598 */ 2599 void 2600 vm_page_dequeue(vm_page_t m) 2601 { 2602 struct vm_pagequeue *pq; 2603 2604 vm_page_assert_locked(m); 2605 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2606 m)); 2607 pq = vm_page_pagequeue(m); 2608 vm_pagequeue_lock(pq); 2609 m->queue = PQ_NONE; 2610 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2611 vm_pagequeue_cnt_dec(pq); 2612 vm_pagequeue_unlock(pq); 2613 } 2614 2615 /* 2616 * vm_page_dequeue_locked: 2617 * 2618 * Remove the given page from its current page queue. 2619 * 2620 * The page and page queue must be locked. 2621 */ 2622 void 2623 vm_page_dequeue_locked(vm_page_t m) 2624 { 2625 struct vm_pagequeue *pq; 2626 2627 vm_page_lock_assert(m, MA_OWNED); 2628 pq = vm_page_pagequeue(m); 2629 vm_pagequeue_assert_locked(pq); 2630 m->queue = PQ_NONE; 2631 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2632 vm_pagequeue_cnt_dec(pq); 2633 } 2634 2635 /* 2636 * vm_page_enqueue: 2637 * 2638 * Add the given page to the specified page queue. 2639 * 2640 * The page must be locked. 2641 */ 2642 static void 2643 vm_page_enqueue(uint8_t queue, vm_page_t m) 2644 { 2645 struct vm_pagequeue *pq; 2646 2647 vm_page_lock_assert(m, MA_OWNED); 2648 KASSERT(queue < PQ_COUNT, 2649 ("vm_page_enqueue: invalid queue %u request for page %p", 2650 queue, m)); 2651 if (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE) 2652 pq = &vm_dom[0].vmd_pagequeues[queue]; 2653 else 2654 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2655 vm_pagequeue_lock(pq); 2656 m->queue = queue; 2657 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2658 vm_pagequeue_cnt_inc(pq); 2659 vm_pagequeue_unlock(pq); 2660 } 2661 2662 /* 2663 * vm_page_requeue: 2664 * 2665 * Move the given page to the tail of its current page queue. 2666 * 2667 * The page must be locked. 2668 */ 2669 void 2670 vm_page_requeue(vm_page_t m) 2671 { 2672 struct vm_pagequeue *pq; 2673 2674 vm_page_lock_assert(m, MA_OWNED); 2675 KASSERT(m->queue != PQ_NONE, 2676 ("vm_page_requeue: page %p is not queued", m)); 2677 pq = vm_page_pagequeue(m); 2678 vm_pagequeue_lock(pq); 2679 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2680 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2681 vm_pagequeue_unlock(pq); 2682 } 2683 2684 /* 2685 * vm_page_requeue_locked: 2686 * 2687 * Move the given page to the tail of its current page queue. 2688 * 2689 * The page queue must be locked. 2690 */ 2691 void 2692 vm_page_requeue_locked(vm_page_t m) 2693 { 2694 struct vm_pagequeue *pq; 2695 2696 KASSERT(m->queue != PQ_NONE, 2697 ("vm_page_requeue_locked: page %p is not queued", m)); 2698 pq = vm_page_pagequeue(m); 2699 vm_pagequeue_assert_locked(pq); 2700 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2701 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2702 } 2703 2704 /* 2705 * vm_page_activate: 2706 * 2707 * Put the specified page on the active list (if appropriate). 2708 * Ensure that act_count is at least ACT_INIT but do not otherwise 2709 * mess with it. 2710 * 2711 * The page must be locked. 2712 */ 2713 void 2714 vm_page_activate(vm_page_t m) 2715 { 2716 int queue; 2717 2718 vm_page_lock_assert(m, MA_OWNED); 2719 if ((queue = m->queue) != PQ_ACTIVE) { 2720 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2721 if (m->act_count < ACT_INIT) 2722 m->act_count = ACT_INIT; 2723 if (queue != PQ_NONE) 2724 vm_page_dequeue(m); 2725 vm_page_enqueue(PQ_ACTIVE, m); 2726 } else 2727 KASSERT(queue == PQ_NONE, 2728 ("vm_page_activate: wired page %p is queued", m)); 2729 } else { 2730 if (m->act_count < ACT_INIT) 2731 m->act_count = ACT_INIT; 2732 } 2733 } 2734 2735 /* 2736 * vm_page_free_wakeup: 2737 * 2738 * Helper routine for vm_page_free_toq(). This routine is called 2739 * when a page is added to the free queues. 2740 * 2741 * The page queues must be locked. 2742 */ 2743 static void 2744 vm_page_free_wakeup(void) 2745 { 2746 2747 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2748 /* 2749 * if pageout daemon needs pages, then tell it that there are 2750 * some free. 2751 */ 2752 if (vm_pageout_pages_needed && 2753 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) { 2754 wakeup(&vm_pageout_pages_needed); 2755 vm_pageout_pages_needed = 0; 2756 } 2757 /* 2758 * wakeup processes that are waiting on memory if we hit a 2759 * high water mark. And wakeup scheduler process if we have 2760 * lots of memory. this process will swapin processes. 2761 */ 2762 if (vm_pages_needed && !vm_page_count_min()) { 2763 vm_pages_needed = false; 2764 wakeup(&vm_cnt.v_free_count); 2765 } 2766 } 2767 2768 /* 2769 * vm_page_free_prep: 2770 * 2771 * Prepares the given page to be put on the free list, 2772 * disassociating it from any VM object. The caller may return 2773 * the page to the free list only if this function returns true. 2774 * 2775 * The object must be locked. The page must be locked if it is 2776 * managed. For a queued managed page, the pagequeue_locked 2777 * argument specifies whether the page queue is already locked. 2778 */ 2779 bool 2780 vm_page_free_prep(vm_page_t m, bool pagequeue_locked) 2781 { 2782 2783 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 2784 if ((m->flags & PG_ZERO) != 0) { 2785 uint64_t *p; 2786 int i; 2787 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 2788 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 2789 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 2790 m, i, (uintmax_t)*p)); 2791 } 2792 #endif 2793 if ((m->oflags & VPO_UNMANAGED) == 0) { 2794 vm_page_lock_assert(m, MA_OWNED); 2795 KASSERT(!pmap_page_is_mapped(m), 2796 ("vm_page_free_toq: freeing mapped page %p", m)); 2797 } else 2798 KASSERT(m->queue == PQ_NONE, 2799 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2800 VM_CNT_INC(v_tfree); 2801 2802 if (vm_page_sbusied(m)) 2803 panic("vm_page_free: freeing busy page %p", m); 2804 2805 vm_page_remove(m); 2806 2807 /* 2808 * If fictitious remove object association and 2809 * return. 2810 */ 2811 if ((m->flags & PG_FICTITIOUS) != 0) { 2812 KASSERT(m->wire_count == 1, 2813 ("fictitious page %p is not wired", m)); 2814 KASSERT(m->queue == PQ_NONE, 2815 ("fictitious page %p is queued", m)); 2816 return (false); 2817 } 2818 2819 if (m->queue != PQ_NONE) { 2820 if (pagequeue_locked) 2821 vm_page_dequeue_locked(m); 2822 else 2823 vm_page_dequeue(m); 2824 } 2825 m->valid = 0; 2826 vm_page_undirty(m); 2827 2828 if (m->wire_count != 0) 2829 panic("vm_page_free: freeing wired page %p", m); 2830 if (m->hold_count != 0) { 2831 m->flags &= ~PG_ZERO; 2832 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2833 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2834 m->flags |= PG_UNHOLDFREE; 2835 return (false); 2836 } 2837 2838 /* 2839 * Restore the default memory attribute to the page. 2840 */ 2841 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2842 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2843 2844 return (true); 2845 } 2846 2847 /* 2848 * Insert the page into the physical memory allocator's free page 2849 * queues. This is the last step to free a page. 2850 */ 2851 static void 2852 vm_page_free_phys(vm_page_t m) 2853 { 2854 2855 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2856 2857 vm_phys_freecnt_adj(m, 1); 2858 #if VM_NRESERVLEVEL > 0 2859 if (!vm_reserv_free_page(m)) 2860 #endif 2861 vm_phys_free_pages(m, 0); 2862 } 2863 2864 void 2865 vm_page_free_phys_pglist(struct pglist *tq) 2866 { 2867 vm_page_t m; 2868 2869 if (TAILQ_EMPTY(tq)) 2870 return; 2871 mtx_lock(&vm_page_queue_free_mtx); 2872 TAILQ_FOREACH(m, tq, listq) 2873 vm_page_free_phys(m); 2874 vm_page_free_wakeup(); 2875 mtx_unlock(&vm_page_queue_free_mtx); 2876 } 2877 2878 /* 2879 * vm_page_free_toq: 2880 * 2881 * Returns the given page to the free list, disassociating it 2882 * from any VM object. 2883 * 2884 * The object must be locked. The page must be locked if it is 2885 * managed. 2886 */ 2887 void 2888 vm_page_free_toq(vm_page_t m) 2889 { 2890 2891 if (!vm_page_free_prep(m, false)) 2892 return; 2893 mtx_lock(&vm_page_queue_free_mtx); 2894 vm_page_free_phys(m); 2895 vm_page_free_wakeup(); 2896 mtx_unlock(&vm_page_queue_free_mtx); 2897 } 2898 2899 /* 2900 * vm_page_wire: 2901 * 2902 * Mark this page as wired down by yet 2903 * another map, removing it from paging queues 2904 * as necessary. 2905 * 2906 * If the page is fictitious, then its wire count must remain one. 2907 * 2908 * The page must be locked. 2909 */ 2910 void 2911 vm_page_wire(vm_page_t m) 2912 { 2913 2914 /* 2915 * Only bump the wire statistics if the page is not already wired, 2916 * and only unqueue the page if it is on some queue (if it is unmanaged 2917 * it is already off the queues). 2918 */ 2919 vm_page_lock_assert(m, MA_OWNED); 2920 if ((m->flags & PG_FICTITIOUS) != 0) { 2921 KASSERT(m->wire_count == 1, 2922 ("vm_page_wire: fictitious page %p's wire count isn't one", 2923 m)); 2924 return; 2925 } 2926 if (m->wire_count == 0) { 2927 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2928 m->queue == PQ_NONE, 2929 ("vm_page_wire: unmanaged page %p is queued", m)); 2930 vm_page_remque(m); 2931 atomic_add_int(&vm_cnt.v_wire_count, 1); 2932 } 2933 m->wire_count++; 2934 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2935 } 2936 2937 /* 2938 * vm_page_unwire: 2939 * 2940 * Release one wiring of the specified page, potentially allowing it to be 2941 * paged out. Returns TRUE if the number of wirings transitions to zero and 2942 * FALSE otherwise. 2943 * 2944 * Only managed pages belonging to an object can be paged out. If the number 2945 * of wirings transitions to zero and the page is eligible for page out, then 2946 * the page is added to the specified paging queue (unless PQ_NONE is 2947 * specified). 2948 * 2949 * If a page is fictitious, then its wire count must always be one. 2950 * 2951 * A managed page must be locked. 2952 */ 2953 boolean_t 2954 vm_page_unwire(vm_page_t m, uint8_t queue) 2955 { 2956 2957 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 2958 ("vm_page_unwire: invalid queue %u request for page %p", 2959 queue, m)); 2960 if ((m->oflags & VPO_UNMANAGED) == 0) 2961 vm_page_assert_locked(m); 2962 if ((m->flags & PG_FICTITIOUS) != 0) { 2963 KASSERT(m->wire_count == 1, 2964 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2965 return (FALSE); 2966 } 2967 if (m->wire_count > 0) { 2968 m->wire_count--; 2969 if (m->wire_count == 0) { 2970 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 2971 if ((m->oflags & VPO_UNMANAGED) == 0 && 2972 m->object != NULL && queue != PQ_NONE) 2973 vm_page_enqueue(queue, m); 2974 return (TRUE); 2975 } else 2976 return (FALSE); 2977 } else 2978 panic("vm_page_unwire: page %p's wire count is zero", m); 2979 } 2980 2981 /* 2982 * Move the specified page to the inactive queue. 2983 * 2984 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive 2985 * queue. However, setting "noreuse" to TRUE will accelerate the specified 2986 * page's reclamation, but it will not unmap the page from any address space. 2987 * This is implemented by inserting the page near the head of the inactive 2988 * queue, using a marker page to guide FIFO insertion ordering. 2989 * 2990 * The page must be locked. 2991 */ 2992 static inline void 2993 _vm_page_deactivate(vm_page_t m, boolean_t noreuse) 2994 { 2995 struct vm_pagequeue *pq; 2996 int queue; 2997 2998 vm_page_assert_locked(m); 2999 3000 /* 3001 * Ignore if the page is already inactive, unless it is unlikely to be 3002 * reactivated. 3003 */ 3004 if ((queue = m->queue) == PQ_INACTIVE && !noreuse) 3005 return; 3006 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3007 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 3008 /* Avoid multiple acquisitions of the inactive queue lock. */ 3009 if (queue == PQ_INACTIVE) { 3010 vm_pagequeue_lock(pq); 3011 vm_page_dequeue_locked(m); 3012 } else { 3013 if (queue != PQ_NONE) 3014 vm_page_dequeue(m); 3015 vm_pagequeue_lock(pq); 3016 } 3017 m->queue = PQ_INACTIVE; 3018 if (noreuse) 3019 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead, 3020 m, plinks.q); 3021 else 3022 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3023 vm_pagequeue_cnt_inc(pq); 3024 vm_pagequeue_unlock(pq); 3025 } 3026 } 3027 3028 /* 3029 * Move the specified page to the inactive queue. 3030 * 3031 * The page must be locked. 3032 */ 3033 void 3034 vm_page_deactivate(vm_page_t m) 3035 { 3036 3037 _vm_page_deactivate(m, FALSE); 3038 } 3039 3040 /* 3041 * Move the specified page to the inactive queue with the expectation 3042 * that it is unlikely to be reused. 3043 * 3044 * The page must be locked. 3045 */ 3046 void 3047 vm_page_deactivate_noreuse(vm_page_t m) 3048 { 3049 3050 _vm_page_deactivate(m, TRUE); 3051 } 3052 3053 /* 3054 * vm_page_launder 3055 * 3056 * Put a page in the laundry. 3057 */ 3058 void 3059 vm_page_launder(vm_page_t m) 3060 { 3061 int queue; 3062 3063 vm_page_assert_locked(m); 3064 if ((queue = m->queue) != PQ_LAUNDRY) { 3065 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3066 if (queue != PQ_NONE) 3067 vm_page_dequeue(m); 3068 vm_page_enqueue(PQ_LAUNDRY, m); 3069 } else 3070 KASSERT(queue == PQ_NONE, 3071 ("wired page %p is queued", m)); 3072 } 3073 } 3074 3075 /* 3076 * vm_page_unswappable 3077 * 3078 * Put a page in the PQ_UNSWAPPABLE holding queue. 3079 */ 3080 void 3081 vm_page_unswappable(vm_page_t m) 3082 { 3083 3084 vm_page_assert_locked(m); 3085 KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0, 3086 ("page %p already unswappable", m)); 3087 if (m->queue != PQ_NONE) 3088 vm_page_dequeue(m); 3089 vm_page_enqueue(PQ_UNSWAPPABLE, m); 3090 } 3091 3092 /* 3093 * Attempt to free the page. If it cannot be freed, do nothing. Returns true 3094 * if the page is freed and false otherwise. 3095 * 3096 * The page must be managed. The page and its containing object must be 3097 * locked. 3098 */ 3099 bool 3100 vm_page_try_to_free(vm_page_t m) 3101 { 3102 3103 vm_page_assert_locked(m); 3104 VM_OBJECT_ASSERT_WLOCKED(m->object); 3105 KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m)); 3106 if (m->dirty != 0 || m->hold_count != 0 || m->wire_count != 0 || 3107 vm_page_busied(m)) 3108 return (false); 3109 if (m->object->ref_count != 0) { 3110 pmap_remove_all(m); 3111 if (m->dirty != 0) 3112 return (false); 3113 } 3114 vm_page_free(m); 3115 return (true); 3116 } 3117 3118 /* 3119 * vm_page_advise 3120 * 3121 * Apply the specified advice to the given page. 3122 * 3123 * The object and page must be locked. 3124 */ 3125 void 3126 vm_page_advise(vm_page_t m, int advice) 3127 { 3128 3129 vm_page_assert_locked(m); 3130 VM_OBJECT_ASSERT_WLOCKED(m->object); 3131 if (advice == MADV_FREE) 3132 /* 3133 * Mark the page clean. This will allow the page to be freed 3134 * without first paging it out. MADV_FREE pages are often 3135 * quickly reused by malloc(3), so we do not do anything that 3136 * would result in a page fault on a later access. 3137 */ 3138 vm_page_undirty(m); 3139 else if (advice != MADV_DONTNEED) { 3140 if (advice == MADV_WILLNEED) 3141 vm_page_activate(m); 3142 return; 3143 } 3144 3145 /* 3146 * Clear any references to the page. Otherwise, the page daemon will 3147 * immediately reactivate the page. 3148 */ 3149 vm_page_aflag_clear(m, PGA_REFERENCED); 3150 3151 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 3152 vm_page_dirty(m); 3153 3154 /* 3155 * Place clean pages near the head of the inactive queue rather than 3156 * the tail, thus defeating the queue's LRU operation and ensuring that 3157 * the page will be reused quickly. Dirty pages not already in the 3158 * laundry are moved there. 3159 */ 3160 if (m->dirty == 0) 3161 vm_page_deactivate_noreuse(m); 3162 else 3163 vm_page_launder(m); 3164 } 3165 3166 /* 3167 * Grab a page, waiting until we are waken up due to the page 3168 * changing state. We keep on waiting, if the page continues 3169 * to be in the object. If the page doesn't exist, first allocate it 3170 * and then conditionally zero it. 3171 * 3172 * This routine may sleep. 3173 * 3174 * The object must be locked on entry. The lock will, however, be released 3175 * and reacquired if the routine sleeps. 3176 */ 3177 vm_page_t 3178 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3179 { 3180 vm_page_t m; 3181 int sleep; 3182 3183 VM_OBJECT_ASSERT_WLOCKED(object); 3184 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3185 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3186 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3187 retrylookup: 3188 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3189 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3190 vm_page_xbusied(m) : vm_page_busied(m); 3191 if (sleep) { 3192 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3193 return (NULL); 3194 /* 3195 * Reference the page before unlocking and 3196 * sleeping so that the page daemon is less 3197 * likely to reclaim it. 3198 */ 3199 vm_page_aflag_set(m, PGA_REFERENCED); 3200 vm_page_lock(m); 3201 VM_OBJECT_WUNLOCK(object); 3202 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 3203 VM_ALLOC_IGN_SBUSY) != 0); 3204 VM_OBJECT_WLOCK(object); 3205 goto retrylookup; 3206 } else { 3207 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3208 vm_page_lock(m); 3209 vm_page_wire(m); 3210 vm_page_unlock(m); 3211 } 3212 if ((allocflags & 3213 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3214 vm_page_xbusy(m); 3215 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3216 vm_page_sbusy(m); 3217 return (m); 3218 } 3219 } 3220 m = vm_page_alloc(object, pindex, allocflags); 3221 if (m == NULL) { 3222 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3223 return (NULL); 3224 VM_OBJECT_WUNLOCK(object); 3225 VM_WAIT; 3226 VM_OBJECT_WLOCK(object); 3227 goto retrylookup; 3228 } 3229 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3230 pmap_zero_page(m); 3231 return (m); 3232 } 3233 3234 /* 3235 * Return the specified range of pages from the given object. For each 3236 * page offset within the range, if a page already exists within the object 3237 * at that offset and it is busy, then wait for it to change state. If, 3238 * instead, the page doesn't exist, then allocate it. 3239 * 3240 * The caller must always specify an allocation class. 3241 * 3242 * allocation classes: 3243 * VM_ALLOC_NORMAL normal process request 3244 * VM_ALLOC_SYSTEM system *really* needs the pages 3245 * 3246 * The caller must always specify that the pages are to be busied and/or 3247 * wired. 3248 * 3249 * optional allocation flags: 3250 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 3251 * VM_ALLOC_NOBUSY do not exclusive busy the page 3252 * VM_ALLOC_NOWAIT do not sleep 3253 * VM_ALLOC_SBUSY set page to sbusy state 3254 * VM_ALLOC_WIRED wire the pages 3255 * VM_ALLOC_ZERO zero and validate any invalid pages 3256 * 3257 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 3258 * may return a partial prefix of the requested range. 3259 */ 3260 int 3261 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 3262 vm_page_t *ma, int count) 3263 { 3264 vm_page_t m, mpred; 3265 int i; 3266 bool sleep; 3267 3268 VM_OBJECT_ASSERT_WLOCKED(object); 3269 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 3270 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 3271 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 3272 (allocflags & VM_ALLOC_WIRED) != 0, 3273 ("vm_page_grab_pages: the pages must be busied or wired")); 3274 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3275 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3276 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); 3277 if (count == 0) 3278 return (0); 3279 i = 0; 3280 retrylookup: 3281 m = vm_radix_lookup_le(&object->rtree, pindex + i); 3282 if (m == NULL || m->pindex != pindex + i) { 3283 mpred = m; 3284 m = NULL; 3285 } else 3286 mpred = TAILQ_PREV(m, pglist, listq); 3287 for (; i < count; i++) { 3288 if (m != NULL) { 3289 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3290 vm_page_xbusied(m) : vm_page_busied(m); 3291 if (sleep) { 3292 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3293 break; 3294 /* 3295 * Reference the page before unlocking and 3296 * sleeping so that the page daemon is less 3297 * likely to reclaim it. 3298 */ 3299 vm_page_aflag_set(m, PGA_REFERENCED); 3300 vm_page_lock(m); 3301 VM_OBJECT_WUNLOCK(object); 3302 vm_page_busy_sleep(m, "grbmaw", (allocflags & 3303 VM_ALLOC_IGN_SBUSY) != 0); 3304 VM_OBJECT_WLOCK(object); 3305 goto retrylookup; 3306 } 3307 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3308 vm_page_lock(m); 3309 vm_page_wire(m); 3310 vm_page_unlock(m); 3311 } 3312 if ((allocflags & (VM_ALLOC_NOBUSY | 3313 VM_ALLOC_SBUSY)) == 0) 3314 vm_page_xbusy(m); 3315 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3316 vm_page_sbusy(m); 3317 } else { 3318 m = vm_page_alloc_after(object, pindex + i, 3319 (allocflags & ~VM_ALLOC_IGN_SBUSY) | 3320 VM_ALLOC_COUNT(count - i), mpred); 3321 if (m == NULL) { 3322 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3323 break; 3324 VM_OBJECT_WUNLOCK(object); 3325 VM_WAIT; 3326 VM_OBJECT_WLOCK(object); 3327 goto retrylookup; 3328 } 3329 } 3330 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) { 3331 if ((m->flags & PG_ZERO) == 0) 3332 pmap_zero_page(m); 3333 m->valid = VM_PAGE_BITS_ALL; 3334 } 3335 ma[i] = mpred = m; 3336 m = vm_page_next(m); 3337 } 3338 return (i); 3339 } 3340 3341 /* 3342 * Mapping function for valid or dirty bits in a page. 3343 * 3344 * Inputs are required to range within a page. 3345 */ 3346 vm_page_bits_t 3347 vm_page_bits(int base, int size) 3348 { 3349 int first_bit; 3350 int last_bit; 3351 3352 KASSERT( 3353 base + size <= PAGE_SIZE, 3354 ("vm_page_bits: illegal base/size %d/%d", base, size) 3355 ); 3356 3357 if (size == 0) /* handle degenerate case */ 3358 return (0); 3359 3360 first_bit = base >> DEV_BSHIFT; 3361 last_bit = (base + size - 1) >> DEV_BSHIFT; 3362 3363 return (((vm_page_bits_t)2 << last_bit) - 3364 ((vm_page_bits_t)1 << first_bit)); 3365 } 3366 3367 /* 3368 * vm_page_set_valid_range: 3369 * 3370 * Sets portions of a page valid. The arguments are expected 3371 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3372 * of any partial chunks touched by the range. The invalid portion of 3373 * such chunks will be zeroed. 3374 * 3375 * (base + size) must be less then or equal to PAGE_SIZE. 3376 */ 3377 void 3378 vm_page_set_valid_range(vm_page_t m, int base, int size) 3379 { 3380 int endoff, frag; 3381 3382 VM_OBJECT_ASSERT_WLOCKED(m->object); 3383 if (size == 0) /* handle degenerate case */ 3384 return; 3385 3386 /* 3387 * If the base is not DEV_BSIZE aligned and the valid 3388 * bit is clear, we have to zero out a portion of the 3389 * first block. 3390 */ 3391 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3392 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 3393 pmap_zero_page_area(m, frag, base - frag); 3394 3395 /* 3396 * If the ending offset is not DEV_BSIZE aligned and the 3397 * valid bit is clear, we have to zero out a portion of 3398 * the last block. 3399 */ 3400 endoff = base + size; 3401 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3402 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 3403 pmap_zero_page_area(m, endoff, 3404 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3405 3406 /* 3407 * Assert that no previously invalid block that is now being validated 3408 * is already dirty. 3409 */ 3410 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 3411 ("vm_page_set_valid_range: page %p is dirty", m)); 3412 3413 /* 3414 * Set valid bits inclusive of any overlap. 3415 */ 3416 m->valid |= vm_page_bits(base, size); 3417 } 3418 3419 /* 3420 * Clear the given bits from the specified page's dirty field. 3421 */ 3422 static __inline void 3423 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 3424 { 3425 uintptr_t addr; 3426 #if PAGE_SIZE < 16384 3427 int shift; 3428 #endif 3429 3430 /* 3431 * If the object is locked and the page is neither exclusive busy nor 3432 * write mapped, then the page's dirty field cannot possibly be 3433 * set by a concurrent pmap operation. 3434 */ 3435 VM_OBJECT_ASSERT_WLOCKED(m->object); 3436 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 3437 m->dirty &= ~pagebits; 3438 else { 3439 /* 3440 * The pmap layer can call vm_page_dirty() without 3441 * holding a distinguished lock. The combination of 3442 * the object's lock and an atomic operation suffice 3443 * to guarantee consistency of the page dirty field. 3444 * 3445 * For PAGE_SIZE == 32768 case, compiler already 3446 * properly aligns the dirty field, so no forcible 3447 * alignment is needed. Only require existence of 3448 * atomic_clear_64 when page size is 32768. 3449 */ 3450 addr = (uintptr_t)&m->dirty; 3451 #if PAGE_SIZE == 32768 3452 atomic_clear_64((uint64_t *)addr, pagebits); 3453 #elif PAGE_SIZE == 16384 3454 atomic_clear_32((uint32_t *)addr, pagebits); 3455 #else /* PAGE_SIZE <= 8192 */ 3456 /* 3457 * Use a trick to perform a 32-bit atomic on the 3458 * containing aligned word, to not depend on the existence 3459 * of atomic_clear_{8, 16}. 3460 */ 3461 shift = addr & (sizeof(uint32_t) - 1); 3462 #if BYTE_ORDER == BIG_ENDIAN 3463 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 3464 #else 3465 shift *= NBBY; 3466 #endif 3467 addr &= ~(sizeof(uint32_t) - 1); 3468 atomic_clear_32((uint32_t *)addr, pagebits << shift); 3469 #endif /* PAGE_SIZE */ 3470 } 3471 } 3472 3473 /* 3474 * vm_page_set_validclean: 3475 * 3476 * Sets portions of a page valid and clean. The arguments are expected 3477 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3478 * of any partial chunks touched by the range. The invalid portion of 3479 * such chunks will be zero'd. 3480 * 3481 * (base + size) must be less then or equal to PAGE_SIZE. 3482 */ 3483 void 3484 vm_page_set_validclean(vm_page_t m, int base, int size) 3485 { 3486 vm_page_bits_t oldvalid, pagebits; 3487 int endoff, frag; 3488 3489 VM_OBJECT_ASSERT_WLOCKED(m->object); 3490 if (size == 0) /* handle degenerate case */ 3491 return; 3492 3493 /* 3494 * If the base is not DEV_BSIZE aligned and the valid 3495 * bit is clear, we have to zero out a portion of the 3496 * first block. 3497 */ 3498 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3499 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 3500 pmap_zero_page_area(m, frag, base - frag); 3501 3502 /* 3503 * If the ending offset is not DEV_BSIZE aligned and the 3504 * valid bit is clear, we have to zero out a portion of 3505 * the last block. 3506 */ 3507 endoff = base + size; 3508 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3509 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 3510 pmap_zero_page_area(m, endoff, 3511 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3512 3513 /* 3514 * Set valid, clear dirty bits. If validating the entire 3515 * page we can safely clear the pmap modify bit. We also 3516 * use this opportunity to clear the VPO_NOSYNC flag. If a process 3517 * takes a write fault on a MAP_NOSYNC memory area the flag will 3518 * be set again. 3519 * 3520 * We set valid bits inclusive of any overlap, but we can only 3521 * clear dirty bits for DEV_BSIZE chunks that are fully within 3522 * the range. 3523 */ 3524 oldvalid = m->valid; 3525 pagebits = vm_page_bits(base, size); 3526 m->valid |= pagebits; 3527 #if 0 /* NOT YET */ 3528 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 3529 frag = DEV_BSIZE - frag; 3530 base += frag; 3531 size -= frag; 3532 if (size < 0) 3533 size = 0; 3534 } 3535 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 3536 #endif 3537 if (base == 0 && size == PAGE_SIZE) { 3538 /* 3539 * The page can only be modified within the pmap if it is 3540 * mapped, and it can only be mapped if it was previously 3541 * fully valid. 3542 */ 3543 if (oldvalid == VM_PAGE_BITS_ALL) 3544 /* 3545 * Perform the pmap_clear_modify() first. Otherwise, 3546 * a concurrent pmap operation, such as 3547 * pmap_protect(), could clear a modification in the 3548 * pmap and set the dirty field on the page before 3549 * pmap_clear_modify() had begun and after the dirty 3550 * field was cleared here. 3551 */ 3552 pmap_clear_modify(m); 3553 m->dirty = 0; 3554 m->oflags &= ~VPO_NOSYNC; 3555 } else if (oldvalid != VM_PAGE_BITS_ALL) 3556 m->dirty &= ~pagebits; 3557 else 3558 vm_page_clear_dirty_mask(m, pagebits); 3559 } 3560 3561 void 3562 vm_page_clear_dirty(vm_page_t m, int base, int size) 3563 { 3564 3565 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3566 } 3567 3568 /* 3569 * vm_page_set_invalid: 3570 * 3571 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3572 * valid and dirty bits for the effected areas are cleared. 3573 */ 3574 void 3575 vm_page_set_invalid(vm_page_t m, int base, int size) 3576 { 3577 vm_page_bits_t bits; 3578 vm_object_t object; 3579 3580 object = m->object; 3581 VM_OBJECT_ASSERT_WLOCKED(object); 3582 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3583 size >= object->un_pager.vnp.vnp_size) 3584 bits = VM_PAGE_BITS_ALL; 3585 else 3586 bits = vm_page_bits(base, size); 3587 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3588 bits != 0) 3589 pmap_remove_all(m); 3590 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3591 !pmap_page_is_mapped(m), 3592 ("vm_page_set_invalid: page %p is mapped", m)); 3593 m->valid &= ~bits; 3594 m->dirty &= ~bits; 3595 } 3596 3597 /* 3598 * vm_page_zero_invalid() 3599 * 3600 * The kernel assumes that the invalid portions of a page contain 3601 * garbage, but such pages can be mapped into memory by user code. 3602 * When this occurs, we must zero out the non-valid portions of the 3603 * page so user code sees what it expects. 3604 * 3605 * Pages are most often semi-valid when the end of a file is mapped 3606 * into memory and the file's size is not page aligned. 3607 */ 3608 void 3609 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3610 { 3611 int b; 3612 int i; 3613 3614 VM_OBJECT_ASSERT_WLOCKED(m->object); 3615 /* 3616 * Scan the valid bits looking for invalid sections that 3617 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3618 * valid bit may be set ) have already been zeroed by 3619 * vm_page_set_validclean(). 3620 */ 3621 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3622 if (i == (PAGE_SIZE / DEV_BSIZE) || 3623 (m->valid & ((vm_page_bits_t)1 << i))) { 3624 if (i > b) { 3625 pmap_zero_page_area(m, 3626 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3627 } 3628 b = i + 1; 3629 } 3630 } 3631 3632 /* 3633 * setvalid is TRUE when we can safely set the zero'd areas 3634 * as being valid. We can do this if there are no cache consistancy 3635 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3636 */ 3637 if (setvalid) 3638 m->valid = VM_PAGE_BITS_ALL; 3639 } 3640 3641 /* 3642 * vm_page_is_valid: 3643 * 3644 * Is (partial) page valid? Note that the case where size == 0 3645 * will return FALSE in the degenerate case where the page is 3646 * entirely invalid, and TRUE otherwise. 3647 */ 3648 int 3649 vm_page_is_valid(vm_page_t m, int base, int size) 3650 { 3651 vm_page_bits_t bits; 3652 3653 VM_OBJECT_ASSERT_LOCKED(m->object); 3654 bits = vm_page_bits(base, size); 3655 return (m->valid != 0 && (m->valid & bits) == bits); 3656 } 3657 3658 /* 3659 * Returns true if all of the specified predicates are true for the entire 3660 * (super)page and false otherwise. 3661 */ 3662 bool 3663 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 3664 { 3665 vm_object_t object; 3666 int i, npages; 3667 3668 object = m->object; 3669 VM_OBJECT_ASSERT_LOCKED(object); 3670 npages = atop(pagesizes[m->psind]); 3671 3672 /* 3673 * The physically contiguous pages that make up a superpage, i.e., a 3674 * page with a page size index ("psind") greater than zero, will 3675 * occupy adjacent entries in vm_page_array[]. 3676 */ 3677 for (i = 0; i < npages; i++) { 3678 /* Always test object consistency, including "skip_m". */ 3679 if (m[i].object != object) 3680 return (false); 3681 if (&m[i] == skip_m) 3682 continue; 3683 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 3684 return (false); 3685 if ((flags & PS_ALL_DIRTY) != 0) { 3686 /* 3687 * Calling vm_page_test_dirty() or pmap_is_modified() 3688 * might stop this case from spuriously returning 3689 * "false". However, that would require a write lock 3690 * on the object containing "m[i]". 3691 */ 3692 if (m[i].dirty != VM_PAGE_BITS_ALL) 3693 return (false); 3694 } 3695 if ((flags & PS_ALL_VALID) != 0 && 3696 m[i].valid != VM_PAGE_BITS_ALL) 3697 return (false); 3698 } 3699 return (true); 3700 } 3701 3702 /* 3703 * Set the page's dirty bits if the page is modified. 3704 */ 3705 void 3706 vm_page_test_dirty(vm_page_t m) 3707 { 3708 3709 VM_OBJECT_ASSERT_WLOCKED(m->object); 3710 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3711 vm_page_dirty(m); 3712 } 3713 3714 void 3715 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3716 { 3717 3718 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3719 } 3720 3721 void 3722 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3723 { 3724 3725 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3726 } 3727 3728 int 3729 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3730 { 3731 3732 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3733 } 3734 3735 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3736 void 3737 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3738 { 3739 3740 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3741 } 3742 3743 void 3744 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3745 { 3746 3747 mtx_assert_(vm_page_lockptr(m), a, file, line); 3748 } 3749 #endif 3750 3751 #ifdef INVARIANTS 3752 void 3753 vm_page_object_lock_assert(vm_page_t m) 3754 { 3755 3756 /* 3757 * Certain of the page's fields may only be modified by the 3758 * holder of the containing object's lock or the exclusive busy. 3759 * holder. Unfortunately, the holder of the write busy is 3760 * not recorded, and thus cannot be checked here. 3761 */ 3762 if (m->object != NULL && !vm_page_xbusied(m)) 3763 VM_OBJECT_ASSERT_WLOCKED(m->object); 3764 } 3765 3766 void 3767 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3768 { 3769 3770 if ((bits & PGA_WRITEABLE) == 0) 3771 return; 3772 3773 /* 3774 * The PGA_WRITEABLE flag can only be set if the page is 3775 * managed, is exclusively busied or the object is locked. 3776 * Currently, this flag is only set by pmap_enter(). 3777 */ 3778 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3779 ("PGA_WRITEABLE on unmanaged page")); 3780 if (!vm_page_xbusied(m)) 3781 VM_OBJECT_ASSERT_LOCKED(m->object); 3782 } 3783 #endif 3784 3785 #include "opt_ddb.h" 3786 #ifdef DDB 3787 #include <sys/kernel.h> 3788 3789 #include <ddb/ddb.h> 3790 3791 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3792 { 3793 3794 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count); 3795 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count); 3796 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count); 3797 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count); 3798 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 3799 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 3800 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 3801 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 3802 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 3803 } 3804 3805 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3806 { 3807 int dom; 3808 3809 db_printf("pq_free %d\n", vm_cnt.v_free_count); 3810 for (dom = 0; dom < vm_ndomains; dom++) { 3811 db_printf( 3812 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", 3813 dom, 3814 vm_dom[dom].vmd_page_count, 3815 vm_dom[dom].vmd_free_count, 3816 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3817 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3818 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, 3819 vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); 3820 } 3821 } 3822 3823 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3824 { 3825 vm_page_t m; 3826 boolean_t phys; 3827 3828 if (!have_addr) { 3829 db_printf("show pginfo addr\n"); 3830 return; 3831 } 3832 3833 phys = strchr(modif, 'p') != NULL; 3834 if (phys) 3835 m = PHYS_TO_VM_PAGE(addr); 3836 else 3837 m = (vm_page_t)addr; 3838 db_printf( 3839 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3840 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3841 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3842 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3843 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3844 } 3845 #endif /* DDB */ 3846