vm_page.c revision 320605
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 * 4. 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: stable/11/sys/vm/vm_page.c 320605 2017-07-03 16:40:05Z markj $"); 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 129struct vm_domain vm_dom[MAXMEMDOM]; 130struct mtx_padalign vm_page_queue_free_mtx; 131 132struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 133 134vm_page_t vm_page_array; 135long vm_page_array_size; 136long first_page; 137int vm_page_zero_count; 138 139static int boot_pages = UMA_BOOT_PAGES; 140SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 141 &boot_pages, 0, 142 "number of pages allocated for bootstrapping the VM system"); 143 144static int pa_tryrelock_restart; 145SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 146 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 147 148static TAILQ_HEAD(, vm_page) blacklist_head; 149static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 150SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 151 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 152 153/* Is the page daemon waiting for free pages? */ 154static int vm_pageout_pages_needed; 155 156static uma_zone_t fakepg_zone; 157 158static void vm_page_alloc_check(vm_page_t m); 159static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 160static void vm_page_enqueue(uint8_t queue, vm_page_t m); 161static void vm_page_free_wakeup(void); 162static void vm_page_init_fakepg(void *dummy); 163static int vm_page_insert_after(vm_page_t m, vm_object_t object, 164 vm_pindex_t pindex, vm_page_t mpred); 165static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 166 vm_page_t mpred); 167static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 168 vm_paddr_t high); 169 170SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 171 172static void 173vm_page_init_fakepg(void *dummy) 174{ 175 176 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 177 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 178} 179 180/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 181#if PAGE_SIZE == 32768 182#ifdef CTASSERT 183CTASSERT(sizeof(u_long) >= 8); 184#endif 185#endif 186 187/* 188 * Try to acquire a physical address lock while a pmap is locked. If we 189 * fail to trylock we unlock and lock the pmap directly and cache the 190 * locked pa in *locked. The caller should then restart their loop in case 191 * the virtual to physical mapping has changed. 192 */ 193int 194vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 195{ 196 vm_paddr_t lockpa; 197 198 lockpa = *locked; 199 *locked = pa; 200 if (lockpa) { 201 PA_LOCK_ASSERT(lockpa, MA_OWNED); 202 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 203 return (0); 204 PA_UNLOCK(lockpa); 205 } 206 if (PA_TRYLOCK(pa)) 207 return (0); 208 PMAP_UNLOCK(pmap); 209 atomic_add_int(&pa_tryrelock_restart, 1); 210 PA_LOCK(pa); 211 PMAP_LOCK(pmap); 212 return (EAGAIN); 213} 214 215/* 216 * vm_set_page_size: 217 * 218 * Sets the page size, perhaps based upon the memory 219 * size. Must be called before any use of page-size 220 * dependent functions. 221 */ 222void 223vm_set_page_size(void) 224{ 225 if (vm_cnt.v_page_size == 0) 226 vm_cnt.v_page_size = PAGE_SIZE; 227 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 228 panic("vm_set_page_size: page size not a power of two"); 229} 230 231/* 232 * vm_page_blacklist_next: 233 * 234 * Find the next entry in the provided string of blacklist 235 * addresses. Entries are separated by space, comma, or newline. 236 * If an invalid integer is encountered then the rest of the 237 * string is skipped. Updates the list pointer to the next 238 * character, or NULL if the string is exhausted or invalid. 239 */ 240static vm_paddr_t 241vm_page_blacklist_next(char **list, char *end) 242{ 243 vm_paddr_t bad; 244 char *cp, *pos; 245 246 if (list == NULL || *list == NULL) 247 return (0); 248 if (**list =='\0') { 249 *list = NULL; 250 return (0); 251 } 252 253 /* 254 * If there's no end pointer then the buffer is coming from 255 * the kenv and we know it's null-terminated. 256 */ 257 if (end == NULL) 258 end = *list + strlen(*list); 259 260 /* Ensure that strtoq() won't walk off the end */ 261 if (*end != '\0') { 262 if (*end == '\n' || *end == ' ' || *end == ',') 263 *end = '\0'; 264 else { 265 printf("Blacklist not terminated, skipping\n"); 266 *list = NULL; 267 return (0); 268 } 269 } 270 271 for (pos = *list; *pos != '\0'; pos = cp) { 272 bad = strtoq(pos, &cp, 0); 273 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 274 if (bad == 0) { 275 if (++cp < end) 276 continue; 277 else 278 break; 279 } 280 } else 281 break; 282 if (*cp == '\0' || ++cp >= end) 283 *list = NULL; 284 else 285 *list = cp; 286 return (trunc_page(bad)); 287 } 288 printf("Garbage in RAM blacklist, skipping\n"); 289 *list = NULL; 290 return (0); 291} 292 293/* 294 * vm_page_blacklist_check: 295 * 296 * Iterate through the provided string of blacklist addresses, pulling 297 * each entry out of the physical allocator free list and putting it 298 * onto a list for reporting via the vm.page_blacklist sysctl. 299 */ 300static void 301vm_page_blacklist_check(char *list, char *end) 302{ 303 vm_paddr_t pa; 304 vm_page_t m; 305 char *next; 306 int ret; 307 308 next = list; 309 while (next != NULL) { 310 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 311 continue; 312 m = vm_phys_paddr_to_vm_page(pa); 313 if (m == NULL) 314 continue; 315 mtx_lock(&vm_page_queue_free_mtx); 316 ret = vm_phys_unfree_page(m); 317 mtx_unlock(&vm_page_queue_free_mtx); 318 if (ret == TRUE) { 319 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 320 if (bootverbose) 321 printf("Skipping page with pa 0x%jx\n", 322 (uintmax_t)pa); 323 } 324 } 325} 326 327/* 328 * vm_page_blacklist_load: 329 * 330 * Search for a special module named "ram_blacklist". It'll be a 331 * plain text file provided by the user via the loader directive 332 * of the same name. 333 */ 334static void 335vm_page_blacklist_load(char **list, char **end) 336{ 337 void *mod; 338 u_char *ptr; 339 u_int len; 340 341 mod = NULL; 342 ptr = NULL; 343 344 mod = preload_search_by_type("ram_blacklist"); 345 if (mod != NULL) { 346 ptr = preload_fetch_addr(mod); 347 len = preload_fetch_size(mod); 348 } 349 *list = ptr; 350 if (ptr != NULL) 351 *end = ptr + len; 352 else 353 *end = NULL; 354 return; 355} 356 357static int 358sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 359{ 360 vm_page_t m; 361 struct sbuf sbuf; 362 int error, first; 363 364 first = 1; 365 error = sysctl_wire_old_buffer(req, 0); 366 if (error != 0) 367 return (error); 368 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 369 TAILQ_FOREACH(m, &blacklist_head, listq) { 370 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 371 (uintmax_t)m->phys_addr); 372 first = 0; 373 } 374 error = sbuf_finish(&sbuf); 375 sbuf_delete(&sbuf); 376 return (error); 377} 378 379static void 380vm_page_domain_init(struct vm_domain *vmd) 381{ 382 struct vm_pagequeue *pq; 383 int i; 384 385 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 386 "vm inactive pagequeue"; 387 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 388 &vm_cnt.v_inactive_count; 389 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 390 "vm active pagequeue"; 391 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 392 &vm_cnt.v_active_count; 393 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 394 "vm laundry pagequeue"; 395 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) = 396 &vm_cnt.v_laundry_count; 397 vmd->vmd_page_count = 0; 398 vmd->vmd_free_count = 0; 399 vmd->vmd_segs = 0; 400 vmd->vmd_oom = FALSE; 401 for (i = 0; i < PQ_COUNT; i++) { 402 pq = &vmd->vmd_pagequeues[i]; 403 TAILQ_INIT(&pq->pq_pl); 404 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 405 MTX_DEF | MTX_DUPOK); 406 } 407} 408 409/* 410 * vm_page_startup: 411 * 412 * Initializes the resident memory module. Allocates physical memory for 413 * bootstrapping UMA and some data structures that are used to manage 414 * physical pages. Initializes these structures, and populates the free 415 * page queues. 416 */ 417vm_offset_t 418vm_page_startup(vm_offset_t vaddr) 419{ 420 vm_offset_t mapped; 421 vm_paddr_t high_avail, low_avail, page_range, size; 422 vm_paddr_t new_end; 423 int i; 424 vm_paddr_t pa; 425 vm_paddr_t last_pa; 426 char *list, *listend; 427 vm_paddr_t end; 428 vm_paddr_t biggestsize; 429 int biggestone; 430 int pages_per_zone; 431 432 biggestsize = 0; 433 biggestone = 0; 434 vaddr = round_page(vaddr); 435 436 for (i = 0; phys_avail[i + 1]; i += 2) { 437 phys_avail[i] = round_page(phys_avail[i]); 438 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 439 } 440 for (i = 0; phys_avail[i + 1]; i += 2) { 441 size = phys_avail[i + 1] - phys_avail[i]; 442 if (size > biggestsize) { 443 biggestone = i; 444 biggestsize = size; 445 } 446 } 447 448 end = phys_avail[biggestone+1]; 449 450 /* 451 * Initialize the page and queue locks. 452 */ 453 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 454 for (i = 0; i < PA_LOCK_COUNT; i++) 455 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 456 for (i = 0; i < vm_ndomains; i++) 457 vm_page_domain_init(&vm_dom[i]); 458 459 /* 460 * Almost all of the pages needed for bootstrapping UMA are used 461 * for zone structures, so if the number of CPUs results in those 462 * structures taking more than one page each, we set aside more pages 463 * in proportion to the zone structure size. 464 */ 465 pages_per_zone = howmany(sizeof(struct uma_zone) + 466 sizeof(struct uma_cache) * (mp_maxid + 1), UMA_SLAB_SIZE); 467 if (pages_per_zone > 1) { 468 /* Reserve more pages so that we don't run out. */ 469 boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone; 470 } 471 472 /* 473 * Allocate memory for use when boot strapping the kernel memory 474 * allocator. 475 * 476 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 477 * manually fetch the value. 478 */ 479 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 480 new_end = end - (boot_pages * UMA_SLAB_SIZE); 481 new_end = trunc_page(new_end); 482 mapped = pmap_map(&vaddr, new_end, end, 483 VM_PROT_READ | VM_PROT_WRITE); 484 bzero((void *)mapped, end - new_end); 485 uma_startup((void *)mapped, boot_pages); 486 487#if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 488 defined(__i386__) || defined(__mips__) 489 /* 490 * Allocate a bitmap to indicate that a random physical page 491 * needs to be included in a minidump. 492 * 493 * The amd64 port needs this to indicate which direct map pages 494 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 495 * 496 * However, i386 still needs this workspace internally within the 497 * minidump code. In theory, they are not needed on i386, but are 498 * included should the sf_buf code decide to use them. 499 */ 500 last_pa = 0; 501 for (i = 0; dump_avail[i + 1] != 0; i += 2) 502 if (dump_avail[i + 1] > last_pa) 503 last_pa = dump_avail[i + 1]; 504 page_range = last_pa / PAGE_SIZE; 505 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 506 new_end -= vm_page_dump_size; 507 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 508 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 509 bzero((void *)vm_page_dump, vm_page_dump_size); 510#endif 511#if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 512 /* 513 * Include the UMA bootstrap pages and vm_page_dump in a crash dump. 514 * When pmap_map() uses the direct map, they are not automatically 515 * included. 516 */ 517 for (pa = new_end; pa < end; pa += PAGE_SIZE) 518 dump_add_page(pa); 519#endif 520 phys_avail[biggestone + 1] = new_end; 521#ifdef __amd64__ 522 /* 523 * Request that the physical pages underlying the message buffer be 524 * included in a crash dump. Since the message buffer is accessed 525 * through the direct map, they are not automatically included. 526 */ 527 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 528 last_pa = pa + round_page(msgbufsize); 529 while (pa < last_pa) { 530 dump_add_page(pa); 531 pa += PAGE_SIZE; 532 } 533#endif 534 /* 535 * Compute the number of pages of memory that will be available for 536 * use, taking into account the overhead of a page structure per page. 537 * In other words, solve 538 * "available physical memory" - round_page(page_range * 539 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 540 * for page_range. 541 */ 542 low_avail = phys_avail[0]; 543 high_avail = phys_avail[1]; 544 for (i = 0; i < vm_phys_nsegs; i++) { 545 if (vm_phys_segs[i].start < low_avail) 546 low_avail = vm_phys_segs[i].start; 547 if (vm_phys_segs[i].end > high_avail) 548 high_avail = vm_phys_segs[i].end; 549 } 550 /* Skip the first chunk. It is already accounted for. */ 551 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 552 if (phys_avail[i] < low_avail) 553 low_avail = phys_avail[i]; 554 if (phys_avail[i + 1] > high_avail) 555 high_avail = phys_avail[i + 1]; 556 } 557 first_page = low_avail / PAGE_SIZE; 558#ifdef VM_PHYSSEG_SPARSE 559 size = 0; 560 for (i = 0; i < vm_phys_nsegs; i++) 561 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 562 for (i = 0; phys_avail[i + 1] != 0; i += 2) 563 size += phys_avail[i + 1] - phys_avail[i]; 564#elif defined(VM_PHYSSEG_DENSE) 565 size = high_avail - low_avail; 566#else 567#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 568#endif 569 570#ifdef VM_PHYSSEG_DENSE 571 /* 572 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 573 * the overhead of a page structure per page only if vm_page_array is 574 * allocated from the last physical memory chunk. Otherwise, we must 575 * allocate page structures representing the physical memory 576 * underlying vm_page_array, even though they will not be used. 577 */ 578 if (new_end != high_avail) 579 page_range = size / PAGE_SIZE; 580 else 581#endif 582 { 583 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 584 585 /* 586 * If the partial bytes remaining are large enough for 587 * a page (PAGE_SIZE) without a corresponding 588 * 'struct vm_page', then new_end will contain an 589 * extra page after subtracting the length of the VM 590 * page array. Compensate by subtracting an extra 591 * page from new_end. 592 */ 593 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 594 if (new_end == high_avail) 595 high_avail -= PAGE_SIZE; 596 new_end -= PAGE_SIZE; 597 } 598 } 599 end = new_end; 600 601 /* 602 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 603 * However, because this page is allocated from KVM, out-of-bounds 604 * accesses using the direct map will not be trapped. 605 */ 606 vaddr += PAGE_SIZE; 607 608 /* 609 * Allocate physical memory for the page structures, and map it. 610 */ 611 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 612 mapped = pmap_map(&vaddr, new_end, end, 613 VM_PROT_READ | VM_PROT_WRITE); 614 vm_page_array = (vm_page_t) mapped; 615#if VM_NRESERVLEVEL > 0 616 /* 617 * Allocate physical memory for the reservation management system's 618 * data structures, and map it. 619 */ 620 if (high_avail == end) 621 high_avail = new_end; 622 new_end = vm_reserv_startup(&vaddr, new_end, high_avail); 623#endif 624#if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 625 /* 626 * Include vm_page_array and vm_reserv_array in a crash dump. 627 */ 628 for (pa = new_end; pa < end; pa += PAGE_SIZE) 629 dump_add_page(pa); 630#endif 631 phys_avail[biggestone + 1] = new_end; 632 633 /* 634 * Add physical memory segments corresponding to the available 635 * physical pages. 636 */ 637 for (i = 0; phys_avail[i + 1] != 0; i += 2) 638 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 639 640 /* 641 * Clear all of the page structures 642 */ 643 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 644 for (i = 0; i < page_range; i++) 645 vm_page_array[i].order = VM_NFREEORDER; 646 vm_page_array_size = page_range; 647 648 /* 649 * Initialize the physical memory allocator. 650 */ 651 vm_phys_init(); 652 653 /* 654 * Add every available physical page that is not blacklisted to 655 * the free lists. 656 */ 657 vm_cnt.v_page_count = 0; 658 vm_cnt.v_free_count = 0; 659 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 660 pa = phys_avail[i]; 661 last_pa = phys_avail[i + 1]; 662 while (pa < last_pa) { 663 vm_phys_add_page(pa); 664 pa += PAGE_SIZE; 665 } 666 } 667 668 TAILQ_INIT(&blacklist_head); 669 vm_page_blacklist_load(&list, &listend); 670 vm_page_blacklist_check(list, listend); 671 672 list = kern_getenv("vm.blacklist"); 673 vm_page_blacklist_check(list, NULL); 674 675 freeenv(list); 676#if VM_NRESERVLEVEL > 0 677 /* 678 * Initialize the reservation management system. 679 */ 680 vm_reserv_init(); 681#endif 682 return (vaddr); 683} 684 685void 686vm_page_reference(vm_page_t m) 687{ 688 689 vm_page_aflag_set(m, PGA_REFERENCED); 690} 691 692/* 693 * vm_page_busy_downgrade: 694 * 695 * Downgrade an exclusive busy page into a single shared busy page. 696 */ 697void 698vm_page_busy_downgrade(vm_page_t m) 699{ 700 u_int x; 701 bool locked; 702 703 vm_page_assert_xbusied(m); 704 locked = mtx_owned(vm_page_lockptr(m)); 705 706 for (;;) { 707 x = m->busy_lock; 708 x &= VPB_BIT_WAITERS; 709 if (x != 0 && !locked) 710 vm_page_lock(m); 711 if (atomic_cmpset_rel_int(&m->busy_lock, 712 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 713 break; 714 if (x != 0 && !locked) 715 vm_page_unlock(m); 716 } 717 if (x != 0) { 718 wakeup(m); 719 if (!locked) 720 vm_page_unlock(m); 721 } 722} 723 724/* 725 * vm_page_sbusied: 726 * 727 * Return a positive value if the page is shared busied, 0 otherwise. 728 */ 729int 730vm_page_sbusied(vm_page_t m) 731{ 732 u_int x; 733 734 x = m->busy_lock; 735 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 736} 737 738/* 739 * vm_page_sunbusy: 740 * 741 * Shared unbusy a page. 742 */ 743void 744vm_page_sunbusy(vm_page_t m) 745{ 746 u_int x; 747 748 vm_page_assert_sbusied(m); 749 750 for (;;) { 751 x = m->busy_lock; 752 if (VPB_SHARERS(x) > 1) { 753 if (atomic_cmpset_int(&m->busy_lock, x, 754 x - VPB_ONE_SHARER)) 755 break; 756 continue; 757 } 758 if ((x & VPB_BIT_WAITERS) == 0) { 759 KASSERT(x == VPB_SHARERS_WORD(1), 760 ("vm_page_sunbusy: invalid lock state")); 761 if (atomic_cmpset_int(&m->busy_lock, 762 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 763 break; 764 continue; 765 } 766 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 767 ("vm_page_sunbusy: invalid lock state for waiters")); 768 769 vm_page_lock(m); 770 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 771 vm_page_unlock(m); 772 continue; 773 } 774 wakeup(m); 775 vm_page_unlock(m); 776 break; 777 } 778} 779 780/* 781 * vm_page_busy_sleep: 782 * 783 * Sleep and release the page lock, using the page pointer as wchan. 784 * This is used to implement the hard-path of busying mechanism. 785 * 786 * The given page must be locked. 787 * 788 * If nonshared is true, sleep only if the page is xbusy. 789 */ 790void 791vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 792{ 793 u_int x; 794 795 vm_page_assert_locked(m); 796 797 x = m->busy_lock; 798 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 799 ((x & VPB_BIT_WAITERS) == 0 && 800 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 801 vm_page_unlock(m); 802 return; 803 } 804 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 805} 806 807/* 808 * vm_page_trysbusy: 809 * 810 * Try to shared busy a page. 811 * If the operation succeeds 1 is returned otherwise 0. 812 * The operation never sleeps. 813 */ 814int 815vm_page_trysbusy(vm_page_t m) 816{ 817 u_int x; 818 819 for (;;) { 820 x = m->busy_lock; 821 if ((x & VPB_BIT_SHARED) == 0) 822 return (0); 823 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 824 return (1); 825 } 826} 827 828static void 829vm_page_xunbusy_locked(vm_page_t m) 830{ 831 832 vm_page_assert_xbusied(m); 833 vm_page_assert_locked(m); 834 835 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 836 /* There is a waiter, do wakeup() instead of vm_page_flash(). */ 837 wakeup(m); 838} 839 840void 841vm_page_xunbusy_maybelocked(vm_page_t m) 842{ 843 bool lockacq; 844 845 vm_page_assert_xbusied(m); 846 847 /* 848 * Fast path for unbusy. If it succeeds, we know that there 849 * are no waiters, so we do not need a wakeup. 850 */ 851 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER, 852 VPB_UNBUSIED)) 853 return; 854 855 lockacq = !mtx_owned(vm_page_lockptr(m)); 856 if (lockacq) 857 vm_page_lock(m); 858 vm_page_xunbusy_locked(m); 859 if (lockacq) 860 vm_page_unlock(m); 861} 862 863/* 864 * vm_page_xunbusy_hard: 865 * 866 * Called after the first try the exclusive unbusy of a page failed. 867 * It is assumed that the waiters bit is on. 868 */ 869void 870vm_page_xunbusy_hard(vm_page_t m) 871{ 872 873 vm_page_assert_xbusied(m); 874 875 vm_page_lock(m); 876 vm_page_xunbusy_locked(m); 877 vm_page_unlock(m); 878} 879 880/* 881 * vm_page_flash: 882 * 883 * Wakeup anyone waiting for the page. 884 * The ownership bits do not change. 885 * 886 * The given page must be locked. 887 */ 888void 889vm_page_flash(vm_page_t m) 890{ 891 u_int x; 892 893 vm_page_lock_assert(m, MA_OWNED); 894 895 for (;;) { 896 x = m->busy_lock; 897 if ((x & VPB_BIT_WAITERS) == 0) 898 return; 899 if (atomic_cmpset_int(&m->busy_lock, x, 900 x & (~VPB_BIT_WAITERS))) 901 break; 902 } 903 wakeup(m); 904} 905 906/* 907 * Keep page from being freed by the page daemon 908 * much of the same effect as wiring, except much lower 909 * overhead and should be used only for *very* temporary 910 * holding ("wiring"). 911 */ 912void 913vm_page_hold(vm_page_t mem) 914{ 915 916 vm_page_lock_assert(mem, MA_OWNED); 917 mem->hold_count++; 918} 919 920void 921vm_page_unhold(vm_page_t mem) 922{ 923 924 vm_page_lock_assert(mem, MA_OWNED); 925 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 926 --mem->hold_count; 927 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 928 vm_page_free_toq(mem); 929} 930 931/* 932 * vm_page_unhold_pages: 933 * 934 * Unhold each of the pages that is referenced by the given array. 935 */ 936void 937vm_page_unhold_pages(vm_page_t *ma, int count) 938{ 939 struct mtx *mtx, *new_mtx; 940 941 mtx = NULL; 942 for (; count != 0; count--) { 943 /* 944 * Avoid releasing and reacquiring the same page lock. 945 */ 946 new_mtx = vm_page_lockptr(*ma); 947 if (mtx != new_mtx) { 948 if (mtx != NULL) 949 mtx_unlock(mtx); 950 mtx = new_mtx; 951 mtx_lock(mtx); 952 } 953 vm_page_unhold(*ma); 954 ma++; 955 } 956 if (mtx != NULL) 957 mtx_unlock(mtx); 958} 959 960vm_page_t 961PHYS_TO_VM_PAGE(vm_paddr_t pa) 962{ 963 vm_page_t m; 964 965#ifdef VM_PHYSSEG_SPARSE 966 m = vm_phys_paddr_to_vm_page(pa); 967 if (m == NULL) 968 m = vm_phys_fictitious_to_vm_page(pa); 969 return (m); 970#elif defined(VM_PHYSSEG_DENSE) 971 long pi; 972 973 pi = atop(pa); 974 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 975 m = &vm_page_array[pi - first_page]; 976 return (m); 977 } 978 return (vm_phys_fictitious_to_vm_page(pa)); 979#else 980#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 981#endif 982} 983 984/* 985 * vm_page_getfake: 986 * 987 * Create a fictitious page with the specified physical address and 988 * memory attribute. The memory attribute is the only the machine- 989 * dependent aspect of a fictitious page that must be initialized. 990 */ 991vm_page_t 992vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 993{ 994 vm_page_t m; 995 996 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 997 vm_page_initfake(m, paddr, memattr); 998 return (m); 999} 1000 1001void 1002vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1003{ 1004 1005 if ((m->flags & PG_FICTITIOUS) != 0) { 1006 /* 1007 * The page's memattr might have changed since the 1008 * previous initialization. Update the pmap to the 1009 * new memattr. 1010 */ 1011 goto memattr; 1012 } 1013 m->phys_addr = paddr; 1014 m->queue = PQ_NONE; 1015 /* Fictitious pages don't use "segind". */ 1016 m->flags = PG_FICTITIOUS; 1017 /* Fictitious pages don't use "order" or "pool". */ 1018 m->oflags = VPO_UNMANAGED; 1019 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1020 m->wire_count = 1; 1021 pmap_page_init(m); 1022memattr: 1023 pmap_page_set_memattr(m, memattr); 1024} 1025 1026/* 1027 * vm_page_putfake: 1028 * 1029 * Release a fictitious page. 1030 */ 1031void 1032vm_page_putfake(vm_page_t m) 1033{ 1034 1035 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1036 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1037 ("vm_page_putfake: bad page %p", m)); 1038 uma_zfree(fakepg_zone, m); 1039} 1040 1041/* 1042 * vm_page_updatefake: 1043 * 1044 * Update the given fictitious page to the specified physical address and 1045 * memory attribute. 1046 */ 1047void 1048vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1049{ 1050 1051 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1052 ("vm_page_updatefake: bad page %p", m)); 1053 m->phys_addr = paddr; 1054 pmap_page_set_memattr(m, memattr); 1055} 1056 1057/* 1058 * vm_page_free: 1059 * 1060 * Free a page. 1061 */ 1062void 1063vm_page_free(vm_page_t m) 1064{ 1065 1066 m->flags &= ~PG_ZERO; 1067 vm_page_free_toq(m); 1068} 1069 1070/* 1071 * vm_page_free_zero: 1072 * 1073 * Free a page to the zerod-pages queue 1074 */ 1075void 1076vm_page_free_zero(vm_page_t m) 1077{ 1078 1079 m->flags |= PG_ZERO; 1080 vm_page_free_toq(m); 1081} 1082 1083/* 1084 * Unbusy and handle the page queueing for a page from a getpages request that 1085 * was optionally read ahead or behind. 1086 */ 1087void 1088vm_page_readahead_finish(vm_page_t m) 1089{ 1090 1091 /* We shouldn't put invalid pages on queues. */ 1092 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); 1093 1094 /* 1095 * Since the page is not the actually needed one, whether it should 1096 * be activated or deactivated is not obvious. Empirical results 1097 * have shown that deactivating the page is usually the best choice, 1098 * unless the page is wanted by another thread. 1099 */ 1100 vm_page_lock(m); 1101 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1102 vm_page_activate(m); 1103 else 1104 vm_page_deactivate(m); 1105 vm_page_unlock(m); 1106 vm_page_xunbusy(m); 1107} 1108 1109/* 1110 * vm_page_sleep_if_busy: 1111 * 1112 * Sleep and release the page queues lock if the page is busied. 1113 * Returns TRUE if the thread slept. 1114 * 1115 * The given page must be unlocked and object containing it must 1116 * be locked. 1117 */ 1118int 1119vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1120{ 1121 vm_object_t obj; 1122 1123 vm_page_lock_assert(m, MA_NOTOWNED); 1124 VM_OBJECT_ASSERT_WLOCKED(m->object); 1125 1126 if (vm_page_busied(m)) { 1127 /* 1128 * The page-specific object must be cached because page 1129 * identity can change during the sleep, causing the 1130 * re-lock of a different object. 1131 * It is assumed that a reference to the object is already 1132 * held by the callers. 1133 */ 1134 obj = m->object; 1135 vm_page_lock(m); 1136 VM_OBJECT_WUNLOCK(obj); 1137 vm_page_busy_sleep(m, msg, false); 1138 VM_OBJECT_WLOCK(obj); 1139 return (TRUE); 1140 } 1141 return (FALSE); 1142} 1143 1144/* 1145 * vm_page_dirty_KBI: [ internal use only ] 1146 * 1147 * Set all bits in the page's dirty field. 1148 * 1149 * The object containing the specified page must be locked if the 1150 * call is made from the machine-independent layer. 1151 * 1152 * See vm_page_clear_dirty_mask(). 1153 * 1154 * This function should only be called by vm_page_dirty(). 1155 */ 1156void 1157vm_page_dirty_KBI(vm_page_t m) 1158{ 1159 1160 /* Refer to this operation by its public name. */ 1161 KASSERT(m->valid == VM_PAGE_BITS_ALL, 1162 ("vm_page_dirty: page is invalid!")); 1163 m->dirty = VM_PAGE_BITS_ALL; 1164} 1165 1166/* 1167 * vm_page_insert: [ internal use only ] 1168 * 1169 * Inserts the given mem entry into the object and object list. 1170 * 1171 * The object must be locked. 1172 */ 1173int 1174vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1175{ 1176 vm_page_t mpred; 1177 1178 VM_OBJECT_ASSERT_WLOCKED(object); 1179 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1180 return (vm_page_insert_after(m, object, pindex, mpred)); 1181} 1182 1183/* 1184 * vm_page_insert_after: 1185 * 1186 * Inserts the page "m" into the specified object at offset "pindex". 1187 * 1188 * The page "mpred" must immediately precede the offset "pindex" within 1189 * the specified object. 1190 * 1191 * The object must be locked. 1192 */ 1193static int 1194vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1195 vm_page_t mpred) 1196{ 1197 vm_page_t msucc; 1198 1199 VM_OBJECT_ASSERT_WLOCKED(object); 1200 KASSERT(m->object == NULL, 1201 ("vm_page_insert_after: page already inserted")); 1202 if (mpred != NULL) { 1203 KASSERT(mpred->object == object, 1204 ("vm_page_insert_after: object doesn't contain mpred")); 1205 KASSERT(mpred->pindex < pindex, 1206 ("vm_page_insert_after: mpred doesn't precede pindex")); 1207 msucc = TAILQ_NEXT(mpred, listq); 1208 } else 1209 msucc = TAILQ_FIRST(&object->memq); 1210 if (msucc != NULL) 1211 KASSERT(msucc->pindex > pindex, 1212 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1213 1214 /* 1215 * Record the object/offset pair in this page 1216 */ 1217 m->object = object; 1218 m->pindex = pindex; 1219 1220 /* 1221 * Now link into the object's ordered list of backed pages. 1222 */ 1223 if (vm_radix_insert(&object->rtree, m)) { 1224 m->object = NULL; 1225 m->pindex = 0; 1226 return (1); 1227 } 1228 vm_page_insert_radixdone(m, object, mpred); 1229 return (0); 1230} 1231 1232/* 1233 * vm_page_insert_radixdone: 1234 * 1235 * Complete page "m" insertion into the specified object after the 1236 * radix trie hooking. 1237 * 1238 * The page "mpred" must precede the offset "m->pindex" within the 1239 * specified object. 1240 * 1241 * The object must be locked. 1242 */ 1243static void 1244vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1245{ 1246 1247 VM_OBJECT_ASSERT_WLOCKED(object); 1248 KASSERT(object != NULL && m->object == object, 1249 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1250 if (mpred != NULL) { 1251 KASSERT(mpred->object == object, 1252 ("vm_page_insert_after: object doesn't contain mpred")); 1253 KASSERT(mpred->pindex < m->pindex, 1254 ("vm_page_insert_after: mpred doesn't precede pindex")); 1255 } 1256 1257 if (mpred != NULL) 1258 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1259 else 1260 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1261 1262 /* 1263 * Show that the object has one more resident page. 1264 */ 1265 object->resident_page_count++; 1266 1267 /* 1268 * Hold the vnode until the last page is released. 1269 */ 1270 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1271 vhold(object->handle); 1272 1273 /* 1274 * Since we are inserting a new and possibly dirty page, 1275 * update the object's OBJ_MIGHTBEDIRTY flag. 1276 */ 1277 if (pmap_page_is_write_mapped(m)) 1278 vm_object_set_writeable_dirty(object); 1279} 1280 1281/* 1282 * vm_page_remove: 1283 * 1284 * Removes the specified page from its containing object, but does not 1285 * invalidate any backing storage. 1286 * 1287 * The object must be locked. The page must be locked if it is managed. 1288 */ 1289void 1290vm_page_remove(vm_page_t m) 1291{ 1292 vm_object_t object; 1293 vm_page_t mrem; 1294 1295 if ((m->oflags & VPO_UNMANAGED) == 0) 1296 vm_page_assert_locked(m); 1297 if ((object = m->object) == NULL) 1298 return; 1299 VM_OBJECT_ASSERT_WLOCKED(object); 1300 if (vm_page_xbusied(m)) 1301 vm_page_xunbusy_maybelocked(m); 1302 mrem = vm_radix_remove(&object->rtree, m->pindex); 1303 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1304 1305 /* 1306 * Now remove from the object's list of backed pages. 1307 */ 1308 TAILQ_REMOVE(&object->memq, m, listq); 1309 1310 /* 1311 * And show that the object has one fewer resident page. 1312 */ 1313 object->resident_page_count--; 1314 1315 /* 1316 * The vnode may now be recycled. 1317 */ 1318 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1319 vdrop(object->handle); 1320 1321 m->object = NULL; 1322} 1323 1324/* 1325 * vm_page_lookup: 1326 * 1327 * Returns the page associated with the object/offset 1328 * pair specified; if none is found, NULL is returned. 1329 * 1330 * The object must be locked. 1331 */ 1332vm_page_t 1333vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1334{ 1335 1336 VM_OBJECT_ASSERT_LOCKED(object); 1337 return (vm_radix_lookup(&object->rtree, pindex)); 1338} 1339 1340/* 1341 * vm_page_find_least: 1342 * 1343 * Returns the page associated with the object with least pindex 1344 * greater than or equal to the parameter pindex, or NULL. 1345 * 1346 * The object must be locked. 1347 */ 1348vm_page_t 1349vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1350{ 1351 vm_page_t m; 1352 1353 VM_OBJECT_ASSERT_LOCKED(object); 1354 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1355 m = vm_radix_lookup_ge(&object->rtree, pindex); 1356 return (m); 1357} 1358 1359/* 1360 * Returns the given page's successor (by pindex) within the object if it is 1361 * resident; if none is found, NULL is returned. 1362 * 1363 * The object must be locked. 1364 */ 1365vm_page_t 1366vm_page_next(vm_page_t m) 1367{ 1368 vm_page_t next; 1369 1370 VM_OBJECT_ASSERT_LOCKED(m->object); 1371 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1372 MPASS(next->object == m->object); 1373 if (next->pindex != m->pindex + 1) 1374 next = NULL; 1375 } 1376 return (next); 1377} 1378 1379/* 1380 * Returns the given page's predecessor (by pindex) within the object if it is 1381 * resident; if none is found, NULL is returned. 1382 * 1383 * The object must be locked. 1384 */ 1385vm_page_t 1386vm_page_prev(vm_page_t m) 1387{ 1388 vm_page_t prev; 1389 1390 VM_OBJECT_ASSERT_LOCKED(m->object); 1391 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1392 MPASS(prev->object == m->object); 1393 if (prev->pindex != m->pindex - 1) 1394 prev = NULL; 1395 } 1396 return (prev); 1397} 1398 1399/* 1400 * Uses the page mnew as a replacement for an existing page at index 1401 * pindex which must be already present in the object. 1402 * 1403 * The existing page must not be on a paging queue. 1404 */ 1405vm_page_t 1406vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1407{ 1408 vm_page_t mold; 1409 1410 VM_OBJECT_ASSERT_WLOCKED(object); 1411 KASSERT(mnew->object == NULL, 1412 ("vm_page_replace: page already in object")); 1413 1414 /* 1415 * This function mostly follows vm_page_insert() and 1416 * vm_page_remove() without the radix, object count and vnode 1417 * dance. Double check such functions for more comments. 1418 */ 1419 1420 mnew->object = object; 1421 mnew->pindex = pindex; 1422 mold = vm_radix_replace(&object->rtree, mnew); 1423 KASSERT(mold->queue == PQ_NONE, 1424 ("vm_page_replace: mold is on a paging queue")); 1425 1426 /* Keep the resident page list in sorted order. */ 1427 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1428 TAILQ_REMOVE(&object->memq, mold, listq); 1429 1430 mold->object = NULL; 1431 vm_page_xunbusy_maybelocked(mold); 1432 1433 /* 1434 * The object's resident_page_count does not change because we have 1435 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1436 */ 1437 if (pmap_page_is_write_mapped(mnew)) 1438 vm_object_set_writeable_dirty(object); 1439 return (mold); 1440} 1441 1442/* 1443 * vm_page_rename: 1444 * 1445 * Move the given memory entry from its 1446 * current object to the specified target object/offset. 1447 * 1448 * Note: swap associated with the page must be invalidated by the move. We 1449 * have to do this for several reasons: (1) we aren't freeing the 1450 * page, (2) we are dirtying the page, (3) the VM system is probably 1451 * moving the page from object A to B, and will then later move 1452 * the backing store from A to B and we can't have a conflict. 1453 * 1454 * Note: we *always* dirty the page. It is necessary both for the 1455 * fact that we moved it, and because we may be invalidating 1456 * swap. 1457 * 1458 * The objects must be locked. 1459 */ 1460int 1461vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1462{ 1463 vm_page_t mpred; 1464 vm_pindex_t opidx; 1465 1466 VM_OBJECT_ASSERT_WLOCKED(new_object); 1467 1468 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1469 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1470 ("vm_page_rename: pindex already renamed")); 1471 1472 /* 1473 * Create a custom version of vm_page_insert() which does not depend 1474 * by m_prev and can cheat on the implementation aspects of the 1475 * function. 1476 */ 1477 opidx = m->pindex; 1478 m->pindex = new_pindex; 1479 if (vm_radix_insert(&new_object->rtree, m)) { 1480 m->pindex = opidx; 1481 return (1); 1482 } 1483 1484 /* 1485 * The operation cannot fail anymore. The removal must happen before 1486 * the listq iterator is tainted. 1487 */ 1488 m->pindex = opidx; 1489 vm_page_lock(m); 1490 vm_page_remove(m); 1491 1492 /* Return back to the new pindex to complete vm_page_insert(). */ 1493 m->pindex = new_pindex; 1494 m->object = new_object; 1495 vm_page_unlock(m); 1496 vm_page_insert_radixdone(m, new_object, mpred); 1497 vm_page_dirty(m); 1498 return (0); 1499} 1500 1501/* 1502 * vm_page_alloc: 1503 * 1504 * Allocate and return a page that is associated with the specified 1505 * object and offset pair. By default, this page is exclusive busied. 1506 * 1507 * The caller must always specify an allocation class. 1508 * 1509 * allocation classes: 1510 * VM_ALLOC_NORMAL normal process request 1511 * VM_ALLOC_SYSTEM system *really* needs a page 1512 * VM_ALLOC_INTERRUPT interrupt time request 1513 * 1514 * optional allocation flags: 1515 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1516 * intends to allocate 1517 * VM_ALLOC_NOBUSY do not exclusive busy the page 1518 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1519 * VM_ALLOC_NOOBJ page is not associated with an object and 1520 * should not be exclusive busy 1521 * VM_ALLOC_SBUSY shared busy the allocated page 1522 * VM_ALLOC_WIRED wire the allocated page 1523 * VM_ALLOC_ZERO prefer a zeroed page 1524 * 1525 * This routine may not sleep. 1526 */ 1527vm_page_t 1528vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1529{ 1530 vm_page_t m, mpred; 1531 int flags, req_class; 1532 1533 mpred = NULL; /* XXX: pacify gcc */ 1534 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1535 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1536 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1537 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1538 ("vm_page_alloc: inconsistent object(%p)/req(%x)", object, req)); 1539 if (object != NULL) 1540 VM_OBJECT_ASSERT_WLOCKED(object); 1541 1542 if (__predict_false((req & VM_ALLOC_IFCACHED) != 0)) 1543 return (NULL); 1544 1545 req_class = req & VM_ALLOC_CLASS_MASK; 1546 1547 /* 1548 * The page daemon is allowed to dig deeper into the free page list. 1549 */ 1550 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1551 req_class = VM_ALLOC_SYSTEM; 1552 1553 if (object != NULL) { 1554 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1555 KASSERT(mpred == NULL || mpred->pindex != pindex, 1556 ("vm_page_alloc: pindex already allocated")); 1557 } 1558 1559 /* 1560 * Allocate a page if the number of free pages exceeds the minimum 1561 * for the request class. 1562 */ 1563 mtx_lock(&vm_page_queue_free_mtx); 1564 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1565 (req_class == VM_ALLOC_SYSTEM && 1566 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1567 (req_class == VM_ALLOC_INTERRUPT && 1568 vm_cnt.v_free_count > 0)) { 1569 /* 1570 * Can we allocate the page from a reservation? 1571 */ 1572#if VM_NRESERVLEVEL > 0 1573 if (object == NULL || (object->flags & (OBJ_COLORED | 1574 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1575 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) 1576#endif 1577 { 1578 /* 1579 * If not, allocate it from the free page queues. 1580 */ 1581 m = vm_phys_alloc_pages(object != NULL ? 1582 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1583#if VM_NRESERVLEVEL > 0 1584 if (m == NULL && vm_reserv_reclaim_inactive()) { 1585 m = vm_phys_alloc_pages(object != NULL ? 1586 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1587 0); 1588 } 1589#endif 1590 } 1591 } else { 1592 /* 1593 * Not allocatable, give up. 1594 */ 1595 mtx_unlock(&vm_page_queue_free_mtx); 1596 atomic_add_int(&vm_pageout_deficit, 1597 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1598 pagedaemon_wakeup(); 1599 return (NULL); 1600 } 1601 1602 /* 1603 * At this point we had better have found a good page. 1604 */ 1605 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1606 vm_phys_freecnt_adj(m, -1); 1607 if ((m->flags & PG_ZERO) != 0) 1608 vm_page_zero_count--; 1609 mtx_unlock(&vm_page_queue_free_mtx); 1610 vm_page_alloc_check(m); 1611 1612 /* 1613 * Initialize the page. Only the PG_ZERO flag is inherited. 1614 */ 1615 flags = 0; 1616 if ((req & VM_ALLOC_ZERO) != 0) 1617 flags = PG_ZERO; 1618 flags &= m->flags; 1619 if ((req & VM_ALLOC_NODUMP) != 0) 1620 flags |= PG_NODUMP; 1621 m->flags = flags; 1622 m->aflags = 0; 1623 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1624 VPO_UNMANAGED : 0; 1625 m->busy_lock = VPB_UNBUSIED; 1626 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1627 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1628 if ((req & VM_ALLOC_SBUSY) != 0) 1629 m->busy_lock = VPB_SHARERS_WORD(1); 1630 if (req & VM_ALLOC_WIRED) { 1631 /* 1632 * The page lock is not required for wiring a page until that 1633 * page is inserted into the object. 1634 */ 1635 atomic_add_int(&vm_cnt.v_wire_count, 1); 1636 m->wire_count = 1; 1637 } 1638 m->act_count = 0; 1639 1640 if (object != NULL) { 1641 if (vm_page_insert_after(m, object, pindex, mpred)) { 1642 pagedaemon_wakeup(); 1643 if (req & VM_ALLOC_WIRED) { 1644 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1645 m->wire_count = 0; 1646 } 1647 KASSERT(m->object == NULL, ("page %p has object", m)); 1648 m->oflags = VPO_UNMANAGED; 1649 m->busy_lock = VPB_UNBUSIED; 1650 /* Don't change PG_ZERO. */ 1651 vm_page_free_toq(m); 1652 return (NULL); 1653 } 1654 1655 /* Ignore device objects; the pager sets "memattr" for them. */ 1656 if (object->memattr != VM_MEMATTR_DEFAULT && 1657 (object->flags & OBJ_FICTITIOUS) == 0) 1658 pmap_page_set_memattr(m, object->memattr); 1659 } else 1660 m->pindex = pindex; 1661 1662 /* 1663 * Don't wakeup too often - wakeup the pageout daemon when 1664 * we would be nearly out of memory. 1665 */ 1666 if (vm_paging_needed()) 1667 pagedaemon_wakeup(); 1668 1669 return (m); 1670} 1671 1672/* 1673 * vm_page_alloc_contig: 1674 * 1675 * Allocate a contiguous set of physical pages of the given size "npages" 1676 * from the free lists. All of the physical pages must be at or above 1677 * the given physical address "low" and below the given physical address 1678 * "high". The given value "alignment" determines the alignment of the 1679 * first physical page in the set. If the given value "boundary" is 1680 * non-zero, then the set of physical pages cannot cross any physical 1681 * address boundary that is a multiple of that value. Both "alignment" 1682 * and "boundary" must be a power of two. 1683 * 1684 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1685 * then the memory attribute setting for the physical pages is configured 1686 * to the object's memory attribute setting. Otherwise, the memory 1687 * attribute setting for the physical pages is configured to "memattr", 1688 * overriding the object's memory attribute setting. However, if the 1689 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1690 * memory attribute setting for the physical pages cannot be configured 1691 * to VM_MEMATTR_DEFAULT. 1692 * 1693 * The specified object may not contain fictitious pages. 1694 * 1695 * The caller must always specify an allocation class. 1696 * 1697 * allocation classes: 1698 * VM_ALLOC_NORMAL normal process request 1699 * VM_ALLOC_SYSTEM system *really* needs a page 1700 * VM_ALLOC_INTERRUPT interrupt time request 1701 * 1702 * optional allocation flags: 1703 * VM_ALLOC_NOBUSY do not exclusive busy the page 1704 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1705 * VM_ALLOC_NOOBJ page is not associated with an object and 1706 * should not be exclusive busy 1707 * VM_ALLOC_SBUSY shared busy the allocated page 1708 * VM_ALLOC_WIRED wire the allocated page 1709 * VM_ALLOC_ZERO prefer a zeroed page 1710 * 1711 * This routine may not sleep. 1712 */ 1713vm_page_t 1714vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1715 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1716 vm_paddr_t boundary, vm_memattr_t memattr) 1717{ 1718 vm_page_t m, m_ret, mpred; 1719 u_int busy_lock, flags, oflags; 1720 int req_class; 1721 1722 mpred = NULL; /* XXX: pacify gcc */ 1723 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1724 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1725 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1726 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1727 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 1728 req)); 1729 if (object != NULL) { 1730 VM_OBJECT_ASSERT_WLOCKED(object); 1731 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 1732 ("vm_page_alloc_contig: object %p has fictitious pages", 1733 object)); 1734 } 1735 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1736 req_class = req & VM_ALLOC_CLASS_MASK; 1737 1738 /* 1739 * The page daemon is allowed to dig deeper into the free page list. 1740 */ 1741 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1742 req_class = VM_ALLOC_SYSTEM; 1743 1744 if (object != NULL) { 1745 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1746 KASSERT(mpred == NULL || mpred->pindex != pindex, 1747 ("vm_page_alloc_contig: pindex already allocated")); 1748 } 1749 1750 /* 1751 * Can we allocate the pages without the number of free pages falling 1752 * below the lower bound for the allocation class? 1753 */ 1754 mtx_lock(&vm_page_queue_free_mtx); 1755 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved || 1756 (req_class == VM_ALLOC_SYSTEM && 1757 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) || 1758 (req_class == VM_ALLOC_INTERRUPT && 1759 vm_cnt.v_free_count >= npages)) { 1760 /* 1761 * Can we allocate the pages from a reservation? 1762 */ 1763#if VM_NRESERVLEVEL > 0 1764retry: 1765 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1766 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1767 low, high, alignment, boundary, mpred)) == NULL) 1768#endif 1769 /* 1770 * If not, allocate them from the free page queues. 1771 */ 1772 m_ret = vm_phys_alloc_contig(npages, low, high, 1773 alignment, boundary); 1774 } else { 1775 mtx_unlock(&vm_page_queue_free_mtx); 1776 atomic_add_int(&vm_pageout_deficit, npages); 1777 pagedaemon_wakeup(); 1778 return (NULL); 1779 } 1780 if (m_ret != NULL) { 1781 vm_phys_freecnt_adj(m_ret, -npages); 1782 for (m = m_ret; m < &m_ret[npages]; m++) 1783 if ((m->flags & PG_ZERO) != 0) 1784 vm_page_zero_count--; 1785 } else { 1786#if VM_NRESERVLEVEL > 0 1787 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1788 boundary)) 1789 goto retry; 1790#endif 1791 } 1792 mtx_unlock(&vm_page_queue_free_mtx); 1793 if (m_ret == NULL) 1794 return (NULL); 1795 for (m = m_ret; m < &m_ret[npages]; m++) 1796 vm_page_alloc_check(m); 1797 1798 /* 1799 * Initialize the pages. Only the PG_ZERO flag is inherited. 1800 */ 1801 flags = 0; 1802 if ((req & VM_ALLOC_ZERO) != 0) 1803 flags = PG_ZERO; 1804 if ((req & VM_ALLOC_NODUMP) != 0) 1805 flags |= PG_NODUMP; 1806 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1807 VPO_UNMANAGED : 0; 1808 busy_lock = VPB_UNBUSIED; 1809 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1810 busy_lock = VPB_SINGLE_EXCLUSIVER; 1811 if ((req & VM_ALLOC_SBUSY) != 0) 1812 busy_lock = VPB_SHARERS_WORD(1); 1813 if ((req & VM_ALLOC_WIRED) != 0) 1814 atomic_add_int(&vm_cnt.v_wire_count, npages); 1815 if (object != NULL) { 1816 if (object->memattr != VM_MEMATTR_DEFAULT && 1817 memattr == VM_MEMATTR_DEFAULT) 1818 memattr = object->memattr; 1819 } 1820 for (m = m_ret; m < &m_ret[npages]; m++) { 1821 m->aflags = 0; 1822 m->flags = (m->flags | PG_NODUMP) & flags; 1823 m->busy_lock = busy_lock; 1824 if ((req & VM_ALLOC_WIRED) != 0) 1825 m->wire_count = 1; 1826 m->act_count = 0; 1827 m->oflags = oflags; 1828 if (object != NULL) { 1829 if (vm_page_insert_after(m, object, pindex, mpred)) { 1830 pagedaemon_wakeup(); 1831 if ((req & VM_ALLOC_WIRED) != 0) 1832 atomic_subtract_int( 1833 &vm_cnt.v_wire_count, npages); 1834 KASSERT(m->object == NULL, 1835 ("page %p has object", m)); 1836 mpred = m; 1837 for (m = m_ret; m < &m_ret[npages]; m++) { 1838 if (m <= mpred && 1839 (req & VM_ALLOC_WIRED) != 0) 1840 m->wire_count = 0; 1841 m->oflags = VPO_UNMANAGED; 1842 m->busy_lock = VPB_UNBUSIED; 1843 /* Don't change PG_ZERO. */ 1844 vm_page_free_toq(m); 1845 } 1846 return (NULL); 1847 } 1848 mpred = m; 1849 } else 1850 m->pindex = pindex; 1851 if (memattr != VM_MEMATTR_DEFAULT) 1852 pmap_page_set_memattr(m, memattr); 1853 pindex++; 1854 } 1855 if (vm_paging_needed()) 1856 pagedaemon_wakeup(); 1857 return (m_ret); 1858} 1859 1860/* 1861 * Check a page that has been freshly dequeued from a freelist. 1862 */ 1863static void 1864vm_page_alloc_check(vm_page_t m) 1865{ 1866 1867 KASSERT(m->object == NULL, ("page %p has object", m)); 1868 KASSERT(m->queue == PQ_NONE, 1869 ("page %p has unexpected queue %d", m, m->queue)); 1870 KASSERT(m->wire_count == 0, ("page %p is wired", m)); 1871 KASSERT(m->hold_count == 0, ("page %p is held", m)); 1872 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 1873 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 1874 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1875 ("page %p has unexpected memattr %d", 1876 m, pmap_page_get_memattr(m))); 1877 KASSERT(m->valid == 0, ("free page %p is valid", m)); 1878} 1879 1880/* 1881 * vm_page_alloc_freelist: 1882 * 1883 * Allocate a physical page from the specified free page list. 1884 * 1885 * The caller must always specify an allocation class. 1886 * 1887 * allocation classes: 1888 * VM_ALLOC_NORMAL normal process request 1889 * VM_ALLOC_SYSTEM system *really* needs a page 1890 * VM_ALLOC_INTERRUPT interrupt time request 1891 * 1892 * optional allocation flags: 1893 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1894 * intends to allocate 1895 * VM_ALLOC_WIRED wire the allocated page 1896 * VM_ALLOC_ZERO prefer a zeroed page 1897 * 1898 * This routine may not sleep. 1899 */ 1900vm_page_t 1901vm_page_alloc_freelist(int flind, int req) 1902{ 1903 vm_page_t m; 1904 u_int flags; 1905 int req_class; 1906 1907 req_class = req & VM_ALLOC_CLASS_MASK; 1908 1909 /* 1910 * The page daemon is allowed to dig deeper into the free page list. 1911 */ 1912 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1913 req_class = VM_ALLOC_SYSTEM; 1914 1915 /* 1916 * Do not allocate reserved pages unless the req has asked for it. 1917 */ 1918 mtx_lock(&vm_page_queue_free_mtx); 1919 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1920 (req_class == VM_ALLOC_SYSTEM && 1921 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1922 (req_class == VM_ALLOC_INTERRUPT && 1923 vm_cnt.v_free_count > 0)) 1924 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1925 else { 1926 mtx_unlock(&vm_page_queue_free_mtx); 1927 atomic_add_int(&vm_pageout_deficit, 1928 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1929 pagedaemon_wakeup(); 1930 return (NULL); 1931 } 1932 if (m == NULL) { 1933 mtx_unlock(&vm_page_queue_free_mtx); 1934 return (NULL); 1935 } 1936 vm_phys_freecnt_adj(m, -1); 1937 if ((m->flags & PG_ZERO) != 0) 1938 vm_page_zero_count--; 1939 mtx_unlock(&vm_page_queue_free_mtx); 1940 vm_page_alloc_check(m); 1941 1942 /* 1943 * Initialize the page. Only the PG_ZERO flag is inherited. 1944 */ 1945 m->aflags = 0; 1946 flags = 0; 1947 if ((req & VM_ALLOC_ZERO) != 0) 1948 flags = PG_ZERO; 1949 m->flags &= flags; 1950 if ((req & VM_ALLOC_WIRED) != 0) { 1951 /* 1952 * The page lock is not required for wiring a page that does 1953 * not belong to an object. 1954 */ 1955 atomic_add_int(&vm_cnt.v_wire_count, 1); 1956 m->wire_count = 1; 1957 } 1958 /* Unmanaged pages don't use "act_count". */ 1959 m->oflags = VPO_UNMANAGED; 1960 if (vm_paging_needed()) 1961 pagedaemon_wakeup(); 1962 return (m); 1963} 1964 1965#define VPSC_ANY 0 /* No restrictions. */ 1966#define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 1967#define VPSC_NOSUPER 2 /* Skip superpages. */ 1968 1969/* 1970 * vm_page_scan_contig: 1971 * 1972 * Scan vm_page_array[] between the specified entries "m_start" and 1973 * "m_end" for a run of contiguous physical pages that satisfy the 1974 * specified conditions, and return the lowest page in the run. The 1975 * specified "alignment" determines the alignment of the lowest physical 1976 * page in the run. If the specified "boundary" is non-zero, then the 1977 * run of physical pages cannot span a physical address that is a 1978 * multiple of "boundary". 1979 * 1980 * "m_end" is never dereferenced, so it need not point to a vm_page 1981 * structure within vm_page_array[]. 1982 * 1983 * "npages" must be greater than zero. "m_start" and "m_end" must not 1984 * span a hole (or discontiguity) in the physical address space. Both 1985 * "alignment" and "boundary" must be a power of two. 1986 */ 1987vm_page_t 1988vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 1989 u_long alignment, vm_paddr_t boundary, int options) 1990{ 1991 struct mtx *m_mtx, *new_mtx; 1992 vm_object_t object; 1993 vm_paddr_t pa; 1994 vm_page_t m, m_run; 1995#if VM_NRESERVLEVEL > 0 1996 int level; 1997#endif 1998 int m_inc, order, run_ext, run_len; 1999 2000 KASSERT(npages > 0, ("npages is 0")); 2001 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2002 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2003 m_run = NULL; 2004 run_len = 0; 2005 m_mtx = NULL; 2006 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2007 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2008 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2009 2010 /* 2011 * If the current page would be the start of a run, check its 2012 * physical address against the end, alignment, and boundary 2013 * conditions. If it doesn't satisfy these conditions, either 2014 * terminate the scan or advance to the next page that 2015 * satisfies the failed condition. 2016 */ 2017 if (run_len == 0) { 2018 KASSERT(m_run == NULL, ("m_run != NULL")); 2019 if (m + npages > m_end) 2020 break; 2021 pa = VM_PAGE_TO_PHYS(m); 2022 if ((pa & (alignment - 1)) != 0) { 2023 m_inc = atop(roundup2(pa, alignment) - pa); 2024 continue; 2025 } 2026 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2027 boundary) != 0) { 2028 m_inc = atop(roundup2(pa, boundary) - pa); 2029 continue; 2030 } 2031 } else 2032 KASSERT(m_run != NULL, ("m_run == NULL")); 2033 2034 /* 2035 * Avoid releasing and reacquiring the same page lock. 2036 */ 2037 new_mtx = vm_page_lockptr(m); 2038 if (m_mtx != new_mtx) { 2039 if (m_mtx != NULL) 2040 mtx_unlock(m_mtx); 2041 m_mtx = new_mtx; 2042 mtx_lock(m_mtx); 2043 } 2044 m_inc = 1; 2045retry: 2046 if (m->wire_count != 0 || m->hold_count != 0) 2047 run_ext = 0; 2048#if VM_NRESERVLEVEL > 0 2049 else if ((level = vm_reserv_level(m)) >= 0 && 2050 (options & VPSC_NORESERV) != 0) { 2051 run_ext = 0; 2052 /* Advance to the end of the reservation. */ 2053 pa = VM_PAGE_TO_PHYS(m); 2054 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2055 pa); 2056 } 2057#endif 2058 else if ((object = m->object) != NULL) { 2059 /* 2060 * The page is considered eligible for relocation if 2061 * and only if it could be laundered or reclaimed by 2062 * the page daemon. 2063 */ 2064 if (!VM_OBJECT_TRYRLOCK(object)) { 2065 mtx_unlock(m_mtx); 2066 VM_OBJECT_RLOCK(object); 2067 mtx_lock(m_mtx); 2068 if (m->object != object) { 2069 /* 2070 * The page may have been freed. 2071 */ 2072 VM_OBJECT_RUNLOCK(object); 2073 goto retry; 2074 } else if (m->wire_count != 0 || 2075 m->hold_count != 0) { 2076 run_ext = 0; 2077 goto unlock; 2078 } 2079 } 2080 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2081 ("page %p is PG_UNHOLDFREE", m)); 2082 /* Don't care: PG_NODUMP, PG_ZERO. */ 2083 if (object->type != OBJT_DEFAULT && 2084 object->type != OBJT_SWAP && 2085 object->type != OBJT_VNODE) { 2086 run_ext = 0; 2087#if VM_NRESERVLEVEL > 0 2088 } else if ((options & VPSC_NOSUPER) != 0 && 2089 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2090 run_ext = 0; 2091 /* Advance to the end of the superpage. */ 2092 pa = VM_PAGE_TO_PHYS(m); 2093 m_inc = atop(roundup2(pa + 1, 2094 vm_reserv_size(level)) - pa); 2095#endif 2096 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2097 m->queue != PQ_NONE && !vm_page_busied(m)) { 2098 /* 2099 * The page is allocated but eligible for 2100 * relocation. Extend the current run by one 2101 * page. 2102 */ 2103 KASSERT(pmap_page_get_memattr(m) == 2104 VM_MEMATTR_DEFAULT, 2105 ("page %p has an unexpected memattr", m)); 2106 KASSERT((m->oflags & (VPO_SWAPINPROG | 2107 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2108 ("page %p has unexpected oflags", m)); 2109 /* Don't care: VPO_NOSYNC. */ 2110 run_ext = 1; 2111 } else 2112 run_ext = 0; 2113unlock: 2114 VM_OBJECT_RUNLOCK(object); 2115#if VM_NRESERVLEVEL > 0 2116 } else if (level >= 0) { 2117 /* 2118 * The page is reserved but not yet allocated. In 2119 * other words, it is still free. Extend the current 2120 * run by one page. 2121 */ 2122 run_ext = 1; 2123#endif 2124 } else if ((order = m->order) < VM_NFREEORDER) { 2125 /* 2126 * The page is enqueued in the physical memory 2127 * allocator's free page queues. Moreover, it is the 2128 * first page in a power-of-two-sized run of 2129 * contiguous free pages. Add these pages to the end 2130 * of the current run, and jump ahead. 2131 */ 2132 run_ext = 1 << order; 2133 m_inc = 1 << order; 2134 } else { 2135 /* 2136 * Skip the page for one of the following reasons: (1) 2137 * It is enqueued in the physical memory allocator's 2138 * free page queues. However, it is not the first 2139 * page in a run of contiguous free pages. (This case 2140 * rarely occurs because the scan is performed in 2141 * ascending order.) (2) It is not reserved, and it is 2142 * transitioning from free to allocated. (Conversely, 2143 * the transition from allocated to free for managed 2144 * pages is blocked by the page lock.) (3) It is 2145 * allocated but not contained by an object and not 2146 * wired, e.g., allocated by Xen's balloon driver. 2147 */ 2148 run_ext = 0; 2149 } 2150 2151 /* 2152 * Extend or reset the current run of pages. 2153 */ 2154 if (run_ext > 0) { 2155 if (run_len == 0) 2156 m_run = m; 2157 run_len += run_ext; 2158 } else { 2159 if (run_len > 0) { 2160 m_run = NULL; 2161 run_len = 0; 2162 } 2163 } 2164 } 2165 if (m_mtx != NULL) 2166 mtx_unlock(m_mtx); 2167 if (run_len >= npages) 2168 return (m_run); 2169 return (NULL); 2170} 2171 2172/* 2173 * vm_page_reclaim_run: 2174 * 2175 * Try to relocate each of the allocated virtual pages within the 2176 * specified run of physical pages to a new physical address. Free the 2177 * physical pages underlying the relocated virtual pages. A virtual page 2178 * is relocatable if and only if it could be laundered or reclaimed by 2179 * the page daemon. Whenever possible, a virtual page is relocated to a 2180 * physical address above "high". 2181 * 2182 * Returns 0 if every physical page within the run was already free or 2183 * just freed by a successful relocation. Otherwise, returns a non-zero 2184 * value indicating why the last attempt to relocate a virtual page was 2185 * unsuccessful. 2186 * 2187 * "req_class" must be an allocation class. 2188 */ 2189static int 2190vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 2191 vm_paddr_t high) 2192{ 2193 struct mtx *m_mtx, *new_mtx; 2194 struct spglist free; 2195 vm_object_t object; 2196 vm_paddr_t pa; 2197 vm_page_t m, m_end, m_new; 2198 int error, order, req; 2199 2200 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2201 ("req_class is not an allocation class")); 2202 SLIST_INIT(&free); 2203 error = 0; 2204 m = m_run; 2205 m_end = m_run + npages; 2206 m_mtx = NULL; 2207 for (; error == 0 && m < m_end; m++) { 2208 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2209 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2210 2211 /* 2212 * Avoid releasing and reacquiring the same page lock. 2213 */ 2214 new_mtx = vm_page_lockptr(m); 2215 if (m_mtx != new_mtx) { 2216 if (m_mtx != NULL) 2217 mtx_unlock(m_mtx); 2218 m_mtx = new_mtx; 2219 mtx_lock(m_mtx); 2220 } 2221retry: 2222 if (m->wire_count != 0 || m->hold_count != 0) 2223 error = EBUSY; 2224 else if ((object = m->object) != NULL) { 2225 /* 2226 * The page is relocated if and only if it could be 2227 * laundered or reclaimed by the page daemon. 2228 */ 2229 if (!VM_OBJECT_TRYWLOCK(object)) { 2230 mtx_unlock(m_mtx); 2231 VM_OBJECT_WLOCK(object); 2232 mtx_lock(m_mtx); 2233 if (m->object != object) { 2234 /* 2235 * The page may have been freed. 2236 */ 2237 VM_OBJECT_WUNLOCK(object); 2238 goto retry; 2239 } else if (m->wire_count != 0 || 2240 m->hold_count != 0) { 2241 error = EBUSY; 2242 goto unlock; 2243 } 2244 } 2245 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2246 ("page %p is PG_UNHOLDFREE", m)); 2247 /* Don't care: PG_NODUMP, PG_ZERO. */ 2248 if (object->type != OBJT_DEFAULT && 2249 object->type != OBJT_SWAP && 2250 object->type != OBJT_VNODE) 2251 error = EINVAL; 2252 else if (object->memattr != VM_MEMATTR_DEFAULT) 2253 error = EINVAL; 2254 else if (m->queue != PQ_NONE && !vm_page_busied(m)) { 2255 KASSERT(pmap_page_get_memattr(m) == 2256 VM_MEMATTR_DEFAULT, 2257 ("page %p has an unexpected memattr", m)); 2258 KASSERT((m->oflags & (VPO_SWAPINPROG | 2259 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2260 ("page %p has unexpected oflags", m)); 2261 /* Don't care: VPO_NOSYNC. */ 2262 if (m->valid != 0) { 2263 /* 2264 * First, try to allocate a new page 2265 * that is above "high". Failing 2266 * that, try to allocate a new page 2267 * that is below "m_run". Allocate 2268 * the new page between the end of 2269 * "m_run" and "high" only as a last 2270 * resort. 2271 */ 2272 req = req_class | VM_ALLOC_NOOBJ; 2273 if ((m->flags & PG_NODUMP) != 0) 2274 req |= VM_ALLOC_NODUMP; 2275 if (trunc_page(high) != 2276 ~(vm_paddr_t)PAGE_MASK) { 2277 m_new = vm_page_alloc_contig( 2278 NULL, 0, req, 1, 2279 round_page(high), 2280 ~(vm_paddr_t)0, 2281 PAGE_SIZE, 0, 2282 VM_MEMATTR_DEFAULT); 2283 } else 2284 m_new = NULL; 2285 if (m_new == NULL) { 2286 pa = VM_PAGE_TO_PHYS(m_run); 2287 m_new = vm_page_alloc_contig( 2288 NULL, 0, req, 1, 2289 0, pa - 1, PAGE_SIZE, 0, 2290 VM_MEMATTR_DEFAULT); 2291 } 2292 if (m_new == NULL) { 2293 pa += ptoa(npages); 2294 m_new = vm_page_alloc_contig( 2295 NULL, 0, req, 1, 2296 pa, high, PAGE_SIZE, 0, 2297 VM_MEMATTR_DEFAULT); 2298 } 2299 if (m_new == NULL) { 2300 error = ENOMEM; 2301 goto unlock; 2302 } 2303 KASSERT(m_new->wire_count == 0, 2304 ("page %p is wired", m)); 2305 2306 /* 2307 * Replace "m" with the new page. For 2308 * vm_page_replace(), "m" must be busy 2309 * and dequeued. Finally, change "m" 2310 * as if vm_page_free() was called. 2311 */ 2312 if (object->ref_count != 0) 2313 pmap_remove_all(m); 2314 m_new->aflags = m->aflags; 2315 KASSERT(m_new->oflags == VPO_UNMANAGED, 2316 ("page %p is managed", m)); 2317 m_new->oflags = m->oflags & VPO_NOSYNC; 2318 pmap_copy_page(m, m_new); 2319 m_new->valid = m->valid; 2320 m_new->dirty = m->dirty; 2321 m->flags &= ~PG_ZERO; 2322 vm_page_xbusy(m); 2323 vm_page_remque(m); 2324 vm_page_replace_checked(m_new, object, 2325 m->pindex, m); 2326 m->valid = 0; 2327 vm_page_undirty(m); 2328 2329 /* 2330 * The new page must be deactivated 2331 * before the object is unlocked. 2332 */ 2333 new_mtx = vm_page_lockptr(m_new); 2334 if (m_mtx != new_mtx) { 2335 mtx_unlock(m_mtx); 2336 m_mtx = new_mtx; 2337 mtx_lock(m_mtx); 2338 } 2339 vm_page_deactivate(m_new); 2340 } else { 2341 m->flags &= ~PG_ZERO; 2342 vm_page_remque(m); 2343 vm_page_remove(m); 2344 KASSERT(m->dirty == 0, 2345 ("page %p is dirty", m)); 2346 } 2347 SLIST_INSERT_HEAD(&free, m, plinks.s.ss); 2348 } else 2349 error = EBUSY; 2350unlock: 2351 VM_OBJECT_WUNLOCK(object); 2352 } else { 2353 mtx_lock(&vm_page_queue_free_mtx); 2354 order = m->order; 2355 if (order < VM_NFREEORDER) { 2356 /* 2357 * The page is enqueued in the physical memory 2358 * allocator's free page queues. Moreover, it 2359 * is the first page in a power-of-two-sized 2360 * run of contiguous free pages. Jump ahead 2361 * to the last page within that run, and 2362 * continue from there. 2363 */ 2364 m += (1 << order) - 1; 2365 } 2366#if VM_NRESERVLEVEL > 0 2367 else if (vm_reserv_is_page_free(m)) 2368 order = 0; 2369#endif 2370 mtx_unlock(&vm_page_queue_free_mtx); 2371 if (order == VM_NFREEORDER) 2372 error = EINVAL; 2373 } 2374 } 2375 if (m_mtx != NULL) 2376 mtx_unlock(m_mtx); 2377 if ((m = SLIST_FIRST(&free)) != NULL) { 2378 mtx_lock(&vm_page_queue_free_mtx); 2379 do { 2380 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2381 vm_phys_freecnt_adj(m, 1); 2382#if VM_NRESERVLEVEL > 0 2383 if (!vm_reserv_free_page(m)) 2384#else 2385 if (true) 2386#endif 2387 vm_phys_free_pages(m, 0); 2388 } while ((m = SLIST_FIRST(&free)) != NULL); 2389 vm_page_zero_idle_wakeup(); 2390 vm_page_free_wakeup(); 2391 mtx_unlock(&vm_page_queue_free_mtx); 2392 } 2393 return (error); 2394} 2395 2396#define NRUNS 16 2397 2398CTASSERT(powerof2(NRUNS)); 2399 2400#define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2401 2402#define MIN_RECLAIM 8 2403 2404/* 2405 * vm_page_reclaim_contig: 2406 * 2407 * Reclaim allocated, contiguous physical memory satisfying the specified 2408 * conditions by relocating the virtual pages using that physical memory. 2409 * Returns true if reclamation is successful and false otherwise. Since 2410 * relocation requires the allocation of physical pages, reclamation may 2411 * fail due to a shortage of free pages. When reclamation fails, callers 2412 * are expected to perform VM_WAIT before retrying a failed allocation 2413 * operation, e.g., vm_page_alloc_contig(). 2414 * 2415 * The caller must always specify an allocation class through "req". 2416 * 2417 * allocation classes: 2418 * VM_ALLOC_NORMAL normal process request 2419 * VM_ALLOC_SYSTEM system *really* needs a page 2420 * VM_ALLOC_INTERRUPT interrupt time request 2421 * 2422 * The optional allocation flags are ignored. 2423 * 2424 * "npages" must be greater than zero. Both "alignment" and "boundary" 2425 * must be a power of two. 2426 */ 2427bool 2428vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2429 u_long alignment, vm_paddr_t boundary) 2430{ 2431 vm_paddr_t curr_low; 2432 vm_page_t m_run, m_runs[NRUNS]; 2433 u_long count, reclaimed; 2434 int error, i, options, req_class; 2435 2436 KASSERT(npages > 0, ("npages is 0")); 2437 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2438 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2439 req_class = req & VM_ALLOC_CLASS_MASK; 2440 2441 /* 2442 * The page daemon is allowed to dig deeper into the free page list. 2443 */ 2444 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2445 req_class = VM_ALLOC_SYSTEM; 2446 2447 /* 2448 * Return if the number of free pages cannot satisfy the requested 2449 * allocation. 2450 */ 2451 count = vm_cnt.v_free_count; 2452 if (count < npages + vm_cnt.v_free_reserved || (count < npages + 2453 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2454 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2455 return (false); 2456 2457 /* 2458 * Scan up to three times, relaxing the restrictions ("options") on 2459 * the reclamation of reservations and superpages each time. 2460 */ 2461 for (options = VPSC_NORESERV;;) { 2462 /* 2463 * Find the highest runs that satisfy the given constraints 2464 * and restrictions, and record them in "m_runs". 2465 */ 2466 curr_low = low; 2467 count = 0; 2468 for (;;) { 2469 m_run = vm_phys_scan_contig(npages, curr_low, high, 2470 alignment, boundary, options); 2471 if (m_run == NULL) 2472 break; 2473 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2474 m_runs[RUN_INDEX(count)] = m_run; 2475 count++; 2476 } 2477 2478 /* 2479 * Reclaim the highest runs in LIFO (descending) order until 2480 * the number of reclaimed pages, "reclaimed", is at least 2481 * MIN_RECLAIM. Reset "reclaimed" each time because each 2482 * reclamation is idempotent, and runs will (likely) recur 2483 * from one scan to the next as restrictions are relaxed. 2484 */ 2485 reclaimed = 0; 2486 for (i = 0; count > 0 && i < NRUNS; i++) { 2487 count--; 2488 m_run = m_runs[RUN_INDEX(count)]; 2489 error = vm_page_reclaim_run(req_class, npages, m_run, 2490 high); 2491 if (error == 0) { 2492 reclaimed += npages; 2493 if (reclaimed >= MIN_RECLAIM) 2494 return (true); 2495 } 2496 } 2497 2498 /* 2499 * Either relax the restrictions on the next scan or return if 2500 * the last scan had no restrictions. 2501 */ 2502 if (options == VPSC_NORESERV) 2503 options = VPSC_NOSUPER; 2504 else if (options == VPSC_NOSUPER) 2505 options = VPSC_ANY; 2506 else if (options == VPSC_ANY) 2507 return (reclaimed != 0); 2508 } 2509} 2510 2511/* 2512 * vm_wait: (also see VM_WAIT macro) 2513 * 2514 * Sleep until free pages are available for allocation. 2515 * - Called in various places before memory allocations. 2516 */ 2517void 2518vm_wait(void) 2519{ 2520 2521 mtx_lock(&vm_page_queue_free_mtx); 2522 if (curproc == pageproc) { 2523 vm_pageout_pages_needed = 1; 2524 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2525 PDROP | PSWP, "VMWait", 0); 2526 } else { 2527 if (__predict_false(pageproc == NULL)) 2528 panic("vm_wait in early boot"); 2529 if (!vm_pageout_wanted) { 2530 vm_pageout_wanted = true; 2531 wakeup(&vm_pageout_wanted); 2532 } 2533 vm_pages_needed = true; 2534 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2535 "vmwait", 0); 2536 } 2537} 2538 2539/* 2540 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2541 * 2542 * Sleep until free pages are available for allocation. 2543 * - Called only in vm_fault so that processes page faulting 2544 * can be easily tracked. 2545 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2546 * processes will be able to grab memory first. Do not change 2547 * this balance without careful testing first. 2548 */ 2549void 2550vm_waitpfault(void) 2551{ 2552 2553 mtx_lock(&vm_page_queue_free_mtx); 2554 if (!vm_pageout_wanted) { 2555 vm_pageout_wanted = true; 2556 wakeup(&vm_pageout_wanted); 2557 } 2558 vm_pages_needed = true; 2559 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2560 "pfault", 0); 2561} 2562 2563struct vm_pagequeue * 2564vm_page_pagequeue(vm_page_t m) 2565{ 2566 2567 if (vm_page_in_laundry(m)) 2568 return (&vm_dom[0].vmd_pagequeues[m->queue]); 2569 else 2570 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2571} 2572 2573/* 2574 * vm_page_dequeue: 2575 * 2576 * Remove the given page from its current page queue. 2577 * 2578 * The page must be locked. 2579 */ 2580void 2581vm_page_dequeue(vm_page_t m) 2582{ 2583 struct vm_pagequeue *pq; 2584 2585 vm_page_assert_locked(m); 2586 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2587 m)); 2588 pq = vm_page_pagequeue(m); 2589 vm_pagequeue_lock(pq); 2590 m->queue = PQ_NONE; 2591 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2592 vm_pagequeue_cnt_dec(pq); 2593 vm_pagequeue_unlock(pq); 2594} 2595 2596/* 2597 * vm_page_dequeue_locked: 2598 * 2599 * Remove the given page from its current page queue. 2600 * 2601 * The page and page queue must be locked. 2602 */ 2603void 2604vm_page_dequeue_locked(vm_page_t m) 2605{ 2606 struct vm_pagequeue *pq; 2607 2608 vm_page_lock_assert(m, MA_OWNED); 2609 pq = vm_page_pagequeue(m); 2610 vm_pagequeue_assert_locked(pq); 2611 m->queue = PQ_NONE; 2612 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2613 vm_pagequeue_cnt_dec(pq); 2614} 2615 2616/* 2617 * vm_page_enqueue: 2618 * 2619 * Add the given page to the specified page queue. 2620 * 2621 * The page must be locked. 2622 */ 2623static void 2624vm_page_enqueue(uint8_t queue, vm_page_t m) 2625{ 2626 struct vm_pagequeue *pq; 2627 2628 vm_page_lock_assert(m, MA_OWNED); 2629 KASSERT(queue < PQ_COUNT, 2630 ("vm_page_enqueue: invalid queue %u request for page %p", 2631 queue, m)); 2632 if (queue == PQ_LAUNDRY) 2633 pq = &vm_dom[0].vmd_pagequeues[queue]; 2634 else 2635 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2636 vm_pagequeue_lock(pq); 2637 m->queue = queue; 2638 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2639 vm_pagequeue_cnt_inc(pq); 2640 vm_pagequeue_unlock(pq); 2641} 2642 2643/* 2644 * vm_page_requeue: 2645 * 2646 * Move the given page to the tail of its current page queue. 2647 * 2648 * The page must be locked. 2649 */ 2650void 2651vm_page_requeue(vm_page_t m) 2652{ 2653 struct vm_pagequeue *pq; 2654 2655 vm_page_lock_assert(m, MA_OWNED); 2656 KASSERT(m->queue != PQ_NONE, 2657 ("vm_page_requeue: page %p is not queued", m)); 2658 pq = vm_page_pagequeue(m); 2659 vm_pagequeue_lock(pq); 2660 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2661 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2662 vm_pagequeue_unlock(pq); 2663} 2664 2665/* 2666 * vm_page_requeue_locked: 2667 * 2668 * Move the given page to the tail of its current page queue. 2669 * 2670 * The page queue must be locked. 2671 */ 2672void 2673vm_page_requeue_locked(vm_page_t m) 2674{ 2675 struct vm_pagequeue *pq; 2676 2677 KASSERT(m->queue != PQ_NONE, 2678 ("vm_page_requeue_locked: page %p is not queued", m)); 2679 pq = vm_page_pagequeue(m); 2680 vm_pagequeue_assert_locked(pq); 2681 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2682 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2683} 2684 2685/* 2686 * vm_page_activate: 2687 * 2688 * Put the specified page on the active list (if appropriate). 2689 * Ensure that act_count is at least ACT_INIT but do not otherwise 2690 * mess with it. 2691 * 2692 * The page must be locked. 2693 */ 2694void 2695vm_page_activate(vm_page_t m) 2696{ 2697 int queue; 2698 2699 vm_page_lock_assert(m, MA_OWNED); 2700 if ((queue = m->queue) != PQ_ACTIVE) { 2701 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2702 if (m->act_count < ACT_INIT) 2703 m->act_count = ACT_INIT; 2704 if (queue != PQ_NONE) 2705 vm_page_dequeue(m); 2706 vm_page_enqueue(PQ_ACTIVE, m); 2707 } else 2708 KASSERT(queue == PQ_NONE, 2709 ("vm_page_activate: wired page %p is queued", m)); 2710 } else { 2711 if (m->act_count < ACT_INIT) 2712 m->act_count = ACT_INIT; 2713 } 2714} 2715 2716/* 2717 * vm_page_free_wakeup: 2718 * 2719 * Helper routine for vm_page_free_toq(). This routine is called 2720 * when a page is added to the free queues. 2721 * 2722 * The page queues must be locked. 2723 */ 2724static inline void 2725vm_page_free_wakeup(void) 2726{ 2727 2728 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2729 /* 2730 * if pageout daemon needs pages, then tell it that there are 2731 * some free. 2732 */ 2733 if (vm_pageout_pages_needed && 2734 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) { 2735 wakeup(&vm_pageout_pages_needed); 2736 vm_pageout_pages_needed = 0; 2737 } 2738 /* 2739 * wakeup processes that are waiting on memory if we hit a 2740 * high water mark. And wakeup scheduler process if we have 2741 * lots of memory. this process will swapin processes. 2742 */ 2743 if (vm_pages_needed && !vm_page_count_min()) { 2744 vm_pages_needed = false; 2745 wakeup(&vm_cnt.v_free_count); 2746 } 2747} 2748 2749/* 2750 * vm_page_free_toq: 2751 * 2752 * Returns the given page to the free list, 2753 * disassociating it with any VM object. 2754 * 2755 * The object must be locked. The page must be locked if it is managed. 2756 */ 2757void 2758vm_page_free_toq(vm_page_t m) 2759{ 2760 2761 if ((m->oflags & VPO_UNMANAGED) == 0) { 2762 vm_page_lock_assert(m, MA_OWNED); 2763 KASSERT(!pmap_page_is_mapped(m), 2764 ("vm_page_free_toq: freeing mapped page %p", m)); 2765 } else 2766 KASSERT(m->queue == PQ_NONE, 2767 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2768 PCPU_INC(cnt.v_tfree); 2769 2770 if (vm_page_sbusied(m)) 2771 panic("vm_page_free: freeing busy page %p", m); 2772 2773 /* 2774 * Unqueue, then remove page. Note that we cannot destroy 2775 * the page here because we do not want to call the pager's 2776 * callback routine until after we've put the page on the 2777 * appropriate free queue. 2778 */ 2779 vm_page_remque(m); 2780 vm_page_remove(m); 2781 2782 /* 2783 * If fictitious remove object association and 2784 * return, otherwise delay object association removal. 2785 */ 2786 if ((m->flags & PG_FICTITIOUS) != 0) { 2787 return; 2788 } 2789 2790 m->valid = 0; 2791 vm_page_undirty(m); 2792 2793 if (m->wire_count != 0) 2794 panic("vm_page_free: freeing wired page %p", m); 2795 if (m->hold_count != 0) { 2796 m->flags &= ~PG_ZERO; 2797 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2798 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2799 m->flags |= PG_UNHOLDFREE; 2800 } else { 2801 /* 2802 * Restore the default memory attribute to the page. 2803 */ 2804 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2805 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2806 2807 /* 2808 * Insert the page into the physical memory allocator's free 2809 * page queues. 2810 */ 2811 mtx_lock(&vm_page_queue_free_mtx); 2812 vm_phys_freecnt_adj(m, 1); 2813#if VM_NRESERVLEVEL > 0 2814 if (!vm_reserv_free_page(m)) 2815#else 2816 if (TRUE) 2817#endif 2818 vm_phys_free_pages(m, 0); 2819 if ((m->flags & PG_ZERO) != 0) 2820 ++vm_page_zero_count; 2821 else 2822 vm_page_zero_idle_wakeup(); 2823 vm_page_free_wakeup(); 2824 mtx_unlock(&vm_page_queue_free_mtx); 2825 } 2826} 2827 2828/* 2829 * vm_page_wire: 2830 * 2831 * Mark this page as wired down by yet 2832 * another map, removing it from paging queues 2833 * as necessary. 2834 * 2835 * If the page is fictitious, then its wire count must remain one. 2836 * 2837 * The page must be locked. 2838 */ 2839void 2840vm_page_wire(vm_page_t m) 2841{ 2842 2843 /* 2844 * Only bump the wire statistics if the page is not already wired, 2845 * and only unqueue the page if it is on some queue (if it is unmanaged 2846 * it is already off the queues). 2847 */ 2848 vm_page_lock_assert(m, MA_OWNED); 2849 if ((m->flags & PG_FICTITIOUS) != 0) { 2850 KASSERT(m->wire_count == 1, 2851 ("vm_page_wire: fictitious page %p's wire count isn't one", 2852 m)); 2853 return; 2854 } 2855 if (m->wire_count == 0) { 2856 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2857 m->queue == PQ_NONE, 2858 ("vm_page_wire: unmanaged page %p is queued", m)); 2859 vm_page_remque(m); 2860 atomic_add_int(&vm_cnt.v_wire_count, 1); 2861 } 2862 m->wire_count++; 2863 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2864} 2865 2866/* 2867 * vm_page_unwire: 2868 * 2869 * Release one wiring of the specified page, potentially allowing it to be 2870 * paged out. Returns TRUE if the number of wirings transitions to zero and 2871 * FALSE otherwise. 2872 * 2873 * Only managed pages belonging to an object can be paged out. If the number 2874 * of wirings transitions to zero and the page is eligible for page out, then 2875 * the page is added to the specified paging queue (unless PQ_NONE is 2876 * specified). 2877 * 2878 * If a page is fictitious, then its wire count must always be one. 2879 * 2880 * A managed page must be locked. 2881 */ 2882boolean_t 2883vm_page_unwire(vm_page_t m, uint8_t queue) 2884{ 2885 2886 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 2887 ("vm_page_unwire: invalid queue %u request for page %p", 2888 queue, m)); 2889 if ((m->oflags & VPO_UNMANAGED) == 0) 2890 vm_page_assert_locked(m); 2891 if ((m->flags & PG_FICTITIOUS) != 0) { 2892 KASSERT(m->wire_count == 1, 2893 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2894 return (FALSE); 2895 } 2896 if (m->wire_count > 0) { 2897 m->wire_count--; 2898 if (m->wire_count == 0) { 2899 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 2900 if ((m->oflags & VPO_UNMANAGED) == 0 && 2901 m->object != NULL && queue != PQ_NONE) 2902 vm_page_enqueue(queue, m); 2903 return (TRUE); 2904 } else 2905 return (FALSE); 2906 } else 2907 panic("vm_page_unwire: page %p's wire count is zero", m); 2908} 2909 2910/* 2911 * Move the specified page to the inactive queue. 2912 * 2913 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive 2914 * queue. However, setting "noreuse" to TRUE will accelerate the specified 2915 * page's reclamation, but it will not unmap the page from any address space. 2916 * This is implemented by inserting the page near the head of the inactive 2917 * queue, using a marker page to guide FIFO insertion ordering. 2918 * 2919 * The page must be locked. 2920 */ 2921static inline void 2922_vm_page_deactivate(vm_page_t m, boolean_t noreuse) 2923{ 2924 struct vm_pagequeue *pq; 2925 int queue; 2926 2927 vm_page_assert_locked(m); 2928 2929 /* 2930 * Ignore if the page is already inactive, unless it is unlikely to be 2931 * reactivated. 2932 */ 2933 if ((queue = m->queue) == PQ_INACTIVE && !noreuse) 2934 return; 2935 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2936 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2937 /* Avoid multiple acquisitions of the inactive queue lock. */ 2938 if (queue == PQ_INACTIVE) { 2939 vm_pagequeue_lock(pq); 2940 vm_page_dequeue_locked(m); 2941 } else { 2942 if (queue != PQ_NONE) 2943 vm_page_dequeue(m); 2944 vm_pagequeue_lock(pq); 2945 } 2946 m->queue = PQ_INACTIVE; 2947 if (noreuse) 2948 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead, 2949 m, plinks.q); 2950 else 2951 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2952 vm_pagequeue_cnt_inc(pq); 2953 vm_pagequeue_unlock(pq); 2954 } 2955} 2956 2957/* 2958 * Move the specified page to the inactive queue. 2959 * 2960 * The page must be locked. 2961 */ 2962void 2963vm_page_deactivate(vm_page_t m) 2964{ 2965 2966 _vm_page_deactivate(m, FALSE); 2967} 2968 2969/* 2970 * Move the specified page to the inactive queue with the expectation 2971 * that it is unlikely to be reused. 2972 * 2973 * The page must be locked. 2974 */ 2975void 2976vm_page_deactivate_noreuse(vm_page_t m) 2977{ 2978 2979 _vm_page_deactivate(m, TRUE); 2980} 2981 2982/* 2983 * vm_page_launder 2984 * 2985 * Put a page in the laundry. 2986 */ 2987void 2988vm_page_launder(vm_page_t m) 2989{ 2990 int queue; 2991 2992 vm_page_assert_locked(m); 2993 if ((queue = m->queue) != PQ_LAUNDRY) { 2994 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2995 if (queue != PQ_NONE) 2996 vm_page_dequeue(m); 2997 vm_page_enqueue(PQ_LAUNDRY, m); 2998 } else 2999 KASSERT(queue == PQ_NONE, 3000 ("wired page %p is queued", m)); 3001 } 3002} 3003 3004/* 3005 * vm_page_try_to_free() 3006 * 3007 * Attempt to free the page. If we cannot free it, we do nothing. 3008 * 1 is returned on success, 0 on failure. 3009 */ 3010int 3011vm_page_try_to_free(vm_page_t m) 3012{ 3013 3014 vm_page_lock_assert(m, MA_OWNED); 3015 if (m->object != NULL) 3016 VM_OBJECT_ASSERT_WLOCKED(m->object); 3017 if (m->dirty || m->hold_count || m->wire_count || 3018 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 3019 return (0); 3020 pmap_remove_all(m); 3021 if (m->dirty) 3022 return (0); 3023 vm_page_free(m); 3024 return (1); 3025} 3026 3027/* 3028 * vm_page_advise 3029 * 3030 * Deactivate or do nothing, as appropriate. 3031 * 3032 * The object and page must be locked. 3033 */ 3034void 3035vm_page_advise(vm_page_t m, int advice) 3036{ 3037 3038 vm_page_assert_locked(m); 3039 VM_OBJECT_ASSERT_WLOCKED(m->object); 3040 if (advice == MADV_FREE) 3041 /* 3042 * Mark the page clean. This will allow the page to be freed 3043 * without first paging it out. MADV_FREE pages are often 3044 * quickly reused by malloc(3), so we do not do anything that 3045 * would result in a page fault on a later access. 3046 */ 3047 vm_page_undirty(m); 3048 else if (advice != MADV_DONTNEED) 3049 return; 3050 3051 /* 3052 * Clear any references to the page. Otherwise, the page daemon will 3053 * immediately reactivate the page. 3054 */ 3055 vm_page_aflag_clear(m, PGA_REFERENCED); 3056 3057 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 3058 vm_page_dirty(m); 3059 3060 /* 3061 * Place clean pages near the head of the inactive queue rather than 3062 * the tail, thus defeating the queue's LRU operation and ensuring that 3063 * the page will be reused quickly. Dirty pages not already in the 3064 * laundry are moved there. 3065 */ 3066 if (m->dirty == 0) 3067 vm_page_deactivate_noreuse(m); 3068 else 3069 vm_page_launder(m); 3070} 3071 3072/* 3073 * Grab a page, waiting until we are waken up due to the page 3074 * changing state. We keep on waiting, if the page continues 3075 * to be in the object. If the page doesn't exist, first allocate it 3076 * and then conditionally zero it. 3077 * 3078 * This routine may sleep. 3079 * 3080 * The object must be locked on entry. The lock will, however, be released 3081 * and reacquired if the routine sleeps. 3082 */ 3083vm_page_t 3084vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3085{ 3086 vm_page_t m; 3087 int sleep; 3088 3089 VM_OBJECT_ASSERT_WLOCKED(object); 3090 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3091 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3092 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3093retrylookup: 3094 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3095 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3096 vm_page_xbusied(m) : vm_page_busied(m); 3097 if (sleep) { 3098 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3099 return (NULL); 3100 /* 3101 * Reference the page before unlocking and 3102 * sleeping so that the page daemon is less 3103 * likely to reclaim it. 3104 */ 3105 vm_page_aflag_set(m, PGA_REFERENCED); 3106 vm_page_lock(m); 3107 VM_OBJECT_WUNLOCK(object); 3108 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 3109 VM_ALLOC_IGN_SBUSY) != 0); 3110 VM_OBJECT_WLOCK(object); 3111 goto retrylookup; 3112 } else { 3113 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3114 vm_page_lock(m); 3115 vm_page_wire(m); 3116 vm_page_unlock(m); 3117 } 3118 if ((allocflags & 3119 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3120 vm_page_xbusy(m); 3121 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3122 vm_page_sbusy(m); 3123 return (m); 3124 } 3125 } 3126 m = vm_page_alloc(object, pindex, allocflags); 3127 if (m == NULL) { 3128 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3129 return (NULL); 3130 VM_OBJECT_WUNLOCK(object); 3131 VM_WAIT; 3132 VM_OBJECT_WLOCK(object); 3133 goto retrylookup; 3134 } 3135 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3136 pmap_zero_page(m); 3137 return (m); 3138} 3139 3140/* 3141 * Mapping function for valid or dirty bits in a page. 3142 * 3143 * Inputs are required to range within a page. 3144 */ 3145vm_page_bits_t 3146vm_page_bits(int base, int size) 3147{ 3148 int first_bit; 3149 int last_bit; 3150 3151 KASSERT( 3152 base + size <= PAGE_SIZE, 3153 ("vm_page_bits: illegal base/size %d/%d", base, size) 3154 ); 3155 3156 if (size == 0) /* handle degenerate case */ 3157 return (0); 3158 3159 first_bit = base >> DEV_BSHIFT; 3160 last_bit = (base + size - 1) >> DEV_BSHIFT; 3161 3162 return (((vm_page_bits_t)2 << last_bit) - 3163 ((vm_page_bits_t)1 << first_bit)); 3164} 3165 3166/* 3167 * vm_page_set_valid_range: 3168 * 3169 * Sets portions of a page valid. The arguments are expected 3170 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3171 * of any partial chunks touched by the range. The invalid portion of 3172 * such chunks will be zeroed. 3173 * 3174 * (base + size) must be less then or equal to PAGE_SIZE. 3175 */ 3176void 3177vm_page_set_valid_range(vm_page_t m, int base, int size) 3178{ 3179 int endoff, frag; 3180 3181 VM_OBJECT_ASSERT_WLOCKED(m->object); 3182 if (size == 0) /* handle degenerate case */ 3183 return; 3184 3185 /* 3186 * If the base is not DEV_BSIZE aligned and the valid 3187 * bit is clear, we have to zero out a portion of the 3188 * first block. 3189 */ 3190 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3191 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 3192 pmap_zero_page_area(m, frag, base - frag); 3193 3194 /* 3195 * If the ending offset is not DEV_BSIZE aligned and the 3196 * valid bit is clear, we have to zero out a portion of 3197 * the last block. 3198 */ 3199 endoff = base + size; 3200 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3201 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 3202 pmap_zero_page_area(m, endoff, 3203 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3204 3205 /* 3206 * Assert that no previously invalid block that is now being validated 3207 * is already dirty. 3208 */ 3209 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 3210 ("vm_page_set_valid_range: page %p is dirty", m)); 3211 3212 /* 3213 * Set valid bits inclusive of any overlap. 3214 */ 3215 m->valid |= vm_page_bits(base, size); 3216} 3217 3218/* 3219 * Clear the given bits from the specified page's dirty field. 3220 */ 3221static __inline void 3222vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 3223{ 3224 uintptr_t addr; 3225#if PAGE_SIZE < 16384 3226 int shift; 3227#endif 3228 3229 /* 3230 * If the object is locked and the page is neither exclusive busy nor 3231 * write mapped, then the page's dirty field cannot possibly be 3232 * set by a concurrent pmap operation. 3233 */ 3234 VM_OBJECT_ASSERT_WLOCKED(m->object); 3235 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 3236 m->dirty &= ~pagebits; 3237 else { 3238 /* 3239 * The pmap layer can call vm_page_dirty() without 3240 * holding a distinguished lock. The combination of 3241 * the object's lock and an atomic operation suffice 3242 * to guarantee consistency of the page dirty field. 3243 * 3244 * For PAGE_SIZE == 32768 case, compiler already 3245 * properly aligns the dirty field, so no forcible 3246 * alignment is needed. Only require existence of 3247 * atomic_clear_64 when page size is 32768. 3248 */ 3249 addr = (uintptr_t)&m->dirty; 3250#if PAGE_SIZE == 32768 3251 atomic_clear_64((uint64_t *)addr, pagebits); 3252#elif PAGE_SIZE == 16384 3253 atomic_clear_32((uint32_t *)addr, pagebits); 3254#else /* PAGE_SIZE <= 8192 */ 3255 /* 3256 * Use a trick to perform a 32-bit atomic on the 3257 * containing aligned word, to not depend on the existence 3258 * of atomic_clear_{8, 16}. 3259 */ 3260 shift = addr & (sizeof(uint32_t) - 1); 3261#if BYTE_ORDER == BIG_ENDIAN 3262 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 3263#else 3264 shift *= NBBY; 3265#endif 3266 addr &= ~(sizeof(uint32_t) - 1); 3267 atomic_clear_32((uint32_t *)addr, pagebits << shift); 3268#endif /* PAGE_SIZE */ 3269 } 3270} 3271 3272/* 3273 * vm_page_set_validclean: 3274 * 3275 * Sets portions of a page valid and clean. The arguments are expected 3276 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3277 * of any partial chunks touched by the range. The invalid portion of 3278 * such chunks will be zero'd. 3279 * 3280 * (base + size) must be less then or equal to PAGE_SIZE. 3281 */ 3282void 3283vm_page_set_validclean(vm_page_t m, int base, int size) 3284{ 3285 vm_page_bits_t oldvalid, pagebits; 3286 int endoff, frag; 3287 3288 VM_OBJECT_ASSERT_WLOCKED(m->object); 3289 if (size == 0) /* handle degenerate case */ 3290 return; 3291 3292 /* 3293 * If the base is not DEV_BSIZE aligned and the valid 3294 * bit is clear, we have to zero out a portion of the 3295 * first block. 3296 */ 3297 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3298 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 3299 pmap_zero_page_area(m, frag, base - frag); 3300 3301 /* 3302 * If the ending offset is not DEV_BSIZE aligned and the 3303 * valid bit is clear, we have to zero out a portion of 3304 * the last block. 3305 */ 3306 endoff = base + size; 3307 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3308 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 3309 pmap_zero_page_area(m, endoff, 3310 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3311 3312 /* 3313 * Set valid, clear dirty bits. If validating the entire 3314 * page we can safely clear the pmap modify bit. We also 3315 * use this opportunity to clear the VPO_NOSYNC flag. If a process 3316 * takes a write fault on a MAP_NOSYNC memory area the flag will 3317 * be set again. 3318 * 3319 * We set valid bits inclusive of any overlap, but we can only 3320 * clear dirty bits for DEV_BSIZE chunks that are fully within 3321 * the range. 3322 */ 3323 oldvalid = m->valid; 3324 pagebits = vm_page_bits(base, size); 3325 m->valid |= pagebits; 3326#if 0 /* NOT YET */ 3327 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 3328 frag = DEV_BSIZE - frag; 3329 base += frag; 3330 size -= frag; 3331 if (size < 0) 3332 size = 0; 3333 } 3334 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 3335#endif 3336 if (base == 0 && size == PAGE_SIZE) { 3337 /* 3338 * The page can only be modified within the pmap if it is 3339 * mapped, and it can only be mapped if it was previously 3340 * fully valid. 3341 */ 3342 if (oldvalid == VM_PAGE_BITS_ALL) 3343 /* 3344 * Perform the pmap_clear_modify() first. Otherwise, 3345 * a concurrent pmap operation, such as 3346 * pmap_protect(), could clear a modification in the 3347 * pmap and set the dirty field on the page before 3348 * pmap_clear_modify() had begun and after the dirty 3349 * field was cleared here. 3350 */ 3351 pmap_clear_modify(m); 3352 m->dirty = 0; 3353 m->oflags &= ~VPO_NOSYNC; 3354 } else if (oldvalid != VM_PAGE_BITS_ALL) 3355 m->dirty &= ~pagebits; 3356 else 3357 vm_page_clear_dirty_mask(m, pagebits); 3358} 3359 3360void 3361vm_page_clear_dirty(vm_page_t m, int base, int size) 3362{ 3363 3364 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3365} 3366 3367/* 3368 * vm_page_set_invalid: 3369 * 3370 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3371 * valid and dirty bits for the effected areas are cleared. 3372 */ 3373void 3374vm_page_set_invalid(vm_page_t m, int base, int size) 3375{ 3376 vm_page_bits_t bits; 3377 vm_object_t object; 3378 3379 object = m->object; 3380 VM_OBJECT_ASSERT_WLOCKED(object); 3381 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3382 size >= object->un_pager.vnp.vnp_size) 3383 bits = VM_PAGE_BITS_ALL; 3384 else 3385 bits = vm_page_bits(base, size); 3386 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3387 bits != 0) 3388 pmap_remove_all(m); 3389 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3390 !pmap_page_is_mapped(m), 3391 ("vm_page_set_invalid: page %p is mapped", m)); 3392 m->valid &= ~bits; 3393 m->dirty &= ~bits; 3394} 3395 3396/* 3397 * vm_page_zero_invalid() 3398 * 3399 * The kernel assumes that the invalid portions of a page contain 3400 * garbage, but such pages can be mapped into memory by user code. 3401 * When this occurs, we must zero out the non-valid portions of the 3402 * page so user code sees what it expects. 3403 * 3404 * Pages are most often semi-valid when the end of a file is mapped 3405 * into memory and the file's size is not page aligned. 3406 */ 3407void 3408vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3409{ 3410 int b; 3411 int i; 3412 3413 VM_OBJECT_ASSERT_WLOCKED(m->object); 3414 /* 3415 * Scan the valid bits looking for invalid sections that 3416 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3417 * valid bit may be set ) have already been zeroed by 3418 * vm_page_set_validclean(). 3419 */ 3420 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3421 if (i == (PAGE_SIZE / DEV_BSIZE) || 3422 (m->valid & ((vm_page_bits_t)1 << i))) { 3423 if (i > b) { 3424 pmap_zero_page_area(m, 3425 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3426 } 3427 b = i + 1; 3428 } 3429 } 3430 3431 /* 3432 * setvalid is TRUE when we can safely set the zero'd areas 3433 * as being valid. We can do this if there are no cache consistancy 3434 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3435 */ 3436 if (setvalid) 3437 m->valid = VM_PAGE_BITS_ALL; 3438} 3439 3440/* 3441 * vm_page_is_valid: 3442 * 3443 * Is (partial) page valid? Note that the case where size == 0 3444 * will return FALSE in the degenerate case where the page is 3445 * entirely invalid, and TRUE otherwise. 3446 */ 3447int 3448vm_page_is_valid(vm_page_t m, int base, int size) 3449{ 3450 vm_page_bits_t bits; 3451 3452 VM_OBJECT_ASSERT_LOCKED(m->object); 3453 bits = vm_page_bits(base, size); 3454 return (m->valid != 0 && (m->valid & bits) == bits); 3455} 3456 3457/* 3458 * vm_page_ps_is_valid: 3459 * 3460 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3461 */ 3462boolean_t 3463vm_page_ps_is_valid(vm_page_t m) 3464{ 3465 int i, npages; 3466 3467 VM_OBJECT_ASSERT_LOCKED(m->object); 3468 npages = atop(pagesizes[m->psind]); 3469 3470 /* 3471 * The physically contiguous pages that make up a superpage, i.e., a 3472 * page with a page size index ("psind") greater than zero, will 3473 * occupy adjacent entries in vm_page_array[]. 3474 */ 3475 for (i = 0; i < npages; i++) { 3476 if (m[i].valid != VM_PAGE_BITS_ALL) 3477 return (FALSE); 3478 } 3479 return (TRUE); 3480} 3481 3482/* 3483 * Set the page's dirty bits if the page is modified. 3484 */ 3485void 3486vm_page_test_dirty(vm_page_t m) 3487{ 3488 3489 VM_OBJECT_ASSERT_WLOCKED(m->object); 3490 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3491 vm_page_dirty(m); 3492} 3493 3494void 3495vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3496{ 3497 3498 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3499} 3500 3501void 3502vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3503{ 3504 3505 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3506} 3507 3508int 3509vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3510{ 3511 3512 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3513} 3514 3515#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3516void 3517vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3518{ 3519 3520 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3521} 3522 3523void 3524vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3525{ 3526 3527 mtx_assert_(vm_page_lockptr(m), a, file, line); 3528} 3529#endif 3530 3531#ifdef INVARIANTS 3532void 3533vm_page_object_lock_assert(vm_page_t m) 3534{ 3535 3536 /* 3537 * Certain of the page's fields may only be modified by the 3538 * holder of the containing object's lock or the exclusive busy. 3539 * holder. Unfortunately, the holder of the write busy is 3540 * not recorded, and thus cannot be checked here. 3541 */ 3542 if (m->object != NULL && !vm_page_xbusied(m)) 3543 VM_OBJECT_ASSERT_WLOCKED(m->object); 3544} 3545 3546void 3547vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3548{ 3549 3550 if ((bits & PGA_WRITEABLE) == 0) 3551 return; 3552 3553 /* 3554 * The PGA_WRITEABLE flag can only be set if the page is 3555 * managed, is exclusively busied or the object is locked. 3556 * Currently, this flag is only set by pmap_enter(). 3557 */ 3558 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3559 ("PGA_WRITEABLE on unmanaged page")); 3560 if (!vm_page_xbusied(m)) 3561 VM_OBJECT_ASSERT_LOCKED(m->object); 3562} 3563#endif 3564 3565#include "opt_ddb.h" 3566#ifdef DDB 3567#include <sys/kernel.h> 3568 3569#include <ddb/ddb.h> 3570 3571DB_SHOW_COMMAND(page, vm_page_print_page_info) 3572{ 3573 3574 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count); 3575 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count); 3576 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count); 3577 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count); 3578 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 3579 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 3580 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 3581 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 3582 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 3583} 3584 3585DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3586{ 3587 int dom; 3588 3589 db_printf("pq_free %d\n", vm_cnt.v_free_count); 3590 for (dom = 0; dom < vm_ndomains; dom++) { 3591 db_printf( 3592 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d\n", 3593 dom, 3594 vm_dom[dom].vmd_page_count, 3595 vm_dom[dom].vmd_free_count, 3596 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3597 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3598 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt); 3599 } 3600} 3601 3602DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3603{ 3604 vm_page_t m; 3605 boolean_t phys; 3606 3607 if (!have_addr) { 3608 db_printf("show pginfo addr\n"); 3609 return; 3610 } 3611 3612 phys = strchr(modif, 'p') != NULL; 3613 if (phys) 3614 m = PHYS_TO_VM_PAGE(addr); 3615 else 3616 m = (vm_page_t)addr; 3617 db_printf( 3618 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3619 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3620 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3621 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3622 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3623} 3624#endif /* DDB */ 3625