vm_page.c revision 311515
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/10/sys/vm/vm_page.c 311515 2017-01-06 12:11:16Z kib $"); 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/malloc.h> 95#include <sys/mman.h> 96#include <sys/msgbuf.h> 97#include <sys/mutex.h> 98#include <sys/proc.h> 99#include <sys/rwlock.h> 100#include <sys/sysctl.h> 101#include <sys/vmmeter.h> 102#include <sys/vnode.h> 103 104#include <vm/vm.h> 105#include <vm/pmap.h> 106#include <vm/vm_param.h> 107#include <vm/vm_kern.h> 108#include <vm/vm_object.h> 109#include <vm/vm_page.h> 110#include <vm/vm_pageout.h> 111#include <vm/vm_pager.h> 112#include <vm/vm_phys.h> 113#include <vm/vm_radix.h> 114#include <vm/vm_reserv.h> 115#include <vm/vm_extern.h> 116#include <vm/uma.h> 117#include <vm/uma_int.h> 118 119#include <machine/md_var.h> 120 121/* 122 * Associated with page of user-allocatable memory is a 123 * page structure. 124 */ 125 126struct vm_domain vm_dom[MAXMEMDOM]; 127struct mtx_padalign vm_page_queue_free_mtx; 128 129struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 130 131vm_page_t vm_page_array; 132long vm_page_array_size; 133long first_page; 134int vm_page_zero_count; 135 136static int boot_pages = UMA_BOOT_PAGES; 137TUNABLE_INT("vm.boot_pages", &boot_pages); 138SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 139 "number of pages allocated for bootstrapping the VM system"); 140 141static int pa_tryrelock_restart; 142SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 143 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 144 145static uma_zone_t fakepg_zone; 146 147static struct vnode *vm_page_alloc_init(vm_page_t m); 148static void vm_page_cache_turn_free(vm_page_t m); 149static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 150static void vm_page_enqueue(int queue, vm_page_t m); 151static void vm_page_init_fakepg(void *dummy); 152static int vm_page_insert_after(vm_page_t m, vm_object_t object, 153 vm_pindex_t pindex, vm_page_t mpred); 154static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 155 vm_page_t mpred); 156 157SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 158 159static void 160vm_page_init_fakepg(void *dummy) 161{ 162 163 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 164 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 165} 166 167/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 168#if PAGE_SIZE == 32768 169#ifdef CTASSERT 170CTASSERT(sizeof(u_long) >= 8); 171#endif 172#endif 173 174/* 175 * Try to acquire a physical address lock while a pmap is locked. If we 176 * fail to trylock we unlock and lock the pmap directly and cache the 177 * locked pa in *locked. The caller should then restart their loop in case 178 * the virtual to physical mapping has changed. 179 */ 180int 181vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 182{ 183 vm_paddr_t lockpa; 184 185 lockpa = *locked; 186 *locked = pa; 187 if (lockpa) { 188 PA_LOCK_ASSERT(lockpa, MA_OWNED); 189 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 190 return (0); 191 PA_UNLOCK(lockpa); 192 } 193 if (PA_TRYLOCK(pa)) 194 return (0); 195 PMAP_UNLOCK(pmap); 196 atomic_add_int(&pa_tryrelock_restart, 1); 197 PA_LOCK(pa); 198 PMAP_LOCK(pmap); 199 return (EAGAIN); 200} 201 202/* 203 * vm_set_page_size: 204 * 205 * Sets the page size, perhaps based upon the memory 206 * size. Must be called before any use of page-size 207 * dependent functions. 208 */ 209void 210vm_set_page_size(void) 211{ 212 if (cnt.v_page_size == 0) 213 cnt.v_page_size = PAGE_SIZE; 214 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 215 panic("vm_set_page_size: page size not a power of two"); 216} 217 218/* 219 * vm_page_blacklist_lookup: 220 * 221 * See if a physical address in this page has been listed 222 * in the blacklist tunable. Entries in the tunable are 223 * separated by spaces or commas. If an invalid integer is 224 * encountered then the rest of the string is skipped. 225 */ 226static int 227vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 228{ 229 vm_paddr_t bad; 230 char *cp, *pos; 231 232 for (pos = list; *pos != '\0'; pos = cp) { 233 bad = strtoq(pos, &cp, 0); 234 if (*cp != '\0') { 235 if (*cp == ' ' || *cp == ',') { 236 cp++; 237 if (cp == pos) 238 continue; 239 } else 240 break; 241 } 242 if (pa == trunc_page(bad)) 243 return (1); 244 } 245 return (0); 246} 247 248static void 249vm_page_domain_init(struct vm_domain *vmd) 250{ 251 struct vm_pagequeue *pq; 252 int i; 253 254 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 255 "vm inactive pagequeue"; 256 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 257 &cnt.v_inactive_count; 258 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 259 "vm active pagequeue"; 260 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 261 &cnt.v_active_count; 262 vmd->vmd_page_count = 0; 263 vmd->vmd_free_count = 0; 264 vmd->vmd_segs = 0; 265 vmd->vmd_oom = FALSE; 266 vmd->vmd_pass = 0; 267 for (i = 0; i < PQ_COUNT; i++) { 268 pq = &vmd->vmd_pagequeues[i]; 269 TAILQ_INIT(&pq->pq_pl); 270 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 271 MTX_DEF | MTX_DUPOK); 272 } 273} 274 275/* 276 * vm_page_startup: 277 * 278 * Initializes the resident memory module. 279 * 280 * Allocates memory for the page cells, and 281 * for the object/offset-to-page hash table headers. 282 * Each page cell is initialized and placed on the free list. 283 */ 284vm_offset_t 285vm_page_startup(vm_offset_t vaddr) 286{ 287 vm_offset_t mapped; 288 vm_paddr_t page_range; 289 vm_paddr_t new_end; 290 int i; 291 vm_paddr_t pa; 292 vm_paddr_t last_pa; 293 char *list; 294 295 /* the biggest memory array is the second group of pages */ 296 vm_paddr_t end; 297 vm_paddr_t biggestsize; 298 vm_paddr_t low_water, high_water; 299 int biggestone; 300 301 biggestsize = 0; 302 biggestone = 0; 303 vaddr = round_page(vaddr); 304 305 for (i = 0; phys_avail[i + 1]; i += 2) { 306 phys_avail[i] = round_page(phys_avail[i]); 307 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 308 } 309 310#ifdef XEN 311 /* 312 * There is no obvious reason why i386 PV Xen needs vm_page structs 313 * created for these pseudo-physical addresses. XXX 314 */ 315 vm_phys_add_seg(0, phys_avail[0]); 316#endif 317 318 low_water = phys_avail[0]; 319 high_water = phys_avail[1]; 320 321 for (i = 0; i < vm_phys_nsegs; i++) { 322 if (vm_phys_segs[i].start < low_water) 323 low_water = vm_phys_segs[i].start; 324 if (vm_phys_segs[i].end > high_water) 325 high_water = vm_phys_segs[i].end; 326 } 327 for (i = 0; phys_avail[i + 1]; i += 2) { 328 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 329 330 if (size > biggestsize) { 331 biggestone = i; 332 biggestsize = size; 333 } 334 if (phys_avail[i] < low_water) 335 low_water = phys_avail[i]; 336 if (phys_avail[i + 1] > high_water) 337 high_water = phys_avail[i + 1]; 338 } 339 340 end = phys_avail[biggestone+1]; 341 342 /* 343 * Initialize the page and queue locks. 344 */ 345 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 346 for (i = 0; i < PA_LOCK_COUNT; i++) 347 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 348 for (i = 0; i < vm_ndomains; i++) 349 vm_page_domain_init(&vm_dom[i]); 350 351 /* 352 * Allocate memory for use when boot strapping the kernel memory 353 * allocator. 354 */ 355 new_end = end - (boot_pages * UMA_SLAB_SIZE); 356 new_end = trunc_page(new_end); 357 mapped = pmap_map(&vaddr, new_end, end, 358 VM_PROT_READ | VM_PROT_WRITE); 359 bzero((void *)mapped, end - new_end); 360 uma_startup((void *)mapped, boot_pages); 361 362#if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 363 defined(__mips__) 364 /* 365 * Allocate a bitmap to indicate that a random physical page 366 * needs to be included in a minidump. 367 * 368 * The amd64 port needs this to indicate which direct map pages 369 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 370 * 371 * However, i386 still needs this workspace internally within the 372 * minidump code. In theory, they are not needed on i386, but are 373 * included should the sf_buf code decide to use them. 374 */ 375 last_pa = 0; 376 for (i = 0; dump_avail[i + 1] != 0; i += 2) 377 if (dump_avail[i + 1] > last_pa) 378 last_pa = dump_avail[i + 1]; 379 page_range = last_pa / PAGE_SIZE; 380 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 381 new_end -= vm_page_dump_size; 382 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 383 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 384 bzero((void *)vm_page_dump, vm_page_dump_size); 385#endif 386#ifdef __amd64__ 387 /* 388 * Request that the physical pages underlying the message buffer be 389 * included in a crash dump. Since the message buffer is accessed 390 * through the direct map, they are not automatically included. 391 */ 392 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 393 last_pa = pa + round_page(msgbufsize); 394 while (pa < last_pa) { 395 dump_add_page(pa); 396 pa += PAGE_SIZE; 397 } 398#endif 399 /* 400 * Compute the number of pages of memory that will be available for 401 * use (taking into account the overhead of a page structure per 402 * page). 403 */ 404 first_page = low_water / PAGE_SIZE; 405#ifdef VM_PHYSSEG_SPARSE 406 page_range = 0; 407 for (i = 0; i < vm_phys_nsegs; i++) { 408 page_range += atop(vm_phys_segs[i].end - 409 vm_phys_segs[i].start); 410 } 411 for (i = 0; phys_avail[i + 1] != 0; i += 2) 412 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 413#elif defined(VM_PHYSSEG_DENSE) 414 page_range = high_water / PAGE_SIZE - first_page; 415#else 416#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 417#endif 418 end = new_end; 419 420 /* 421 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 422 */ 423 vaddr += PAGE_SIZE; 424 425 /* 426 * Initialize the mem entry structures now, and put them in the free 427 * queue. 428 */ 429 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 430 mapped = pmap_map(&vaddr, new_end, end, 431 VM_PROT_READ | VM_PROT_WRITE); 432 vm_page_array = (vm_page_t) mapped; 433#if VM_NRESERVLEVEL > 0 434 /* 435 * Allocate memory for the reservation management system's data 436 * structures. 437 */ 438 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 439#endif 440#if defined(__amd64__) || defined(__mips__) 441 /* 442 * pmap_map on amd64 and mips can come out of the direct-map, not kvm 443 * like i386, so the pages must be tracked for a crashdump to include 444 * this data. This includes the vm_page_array and the early UMA 445 * bootstrap pages. 446 */ 447 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 448 dump_add_page(pa); 449#endif 450 phys_avail[biggestone + 1] = new_end; 451 452 /* 453 * Add physical memory segments corresponding to the available 454 * physical pages. 455 */ 456 for (i = 0; phys_avail[i + 1] != 0; i += 2) 457 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 458 459 /* 460 * Clear all of the page structures 461 */ 462 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 463 for (i = 0; i < page_range; i++) 464 vm_page_array[i].order = VM_NFREEORDER; 465 vm_page_array_size = page_range; 466 467 /* 468 * Initialize the physical memory allocator. 469 */ 470 vm_phys_init(); 471 472 /* 473 * Add every available physical page that is not blacklisted to 474 * the free lists. 475 */ 476 cnt.v_page_count = 0; 477 cnt.v_free_count = 0; 478 list = getenv("vm.blacklist"); 479 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 480 pa = phys_avail[i]; 481 last_pa = phys_avail[i + 1]; 482 while (pa < last_pa) { 483 if (list != NULL && 484 vm_page_blacklist_lookup(list, pa)) 485 printf("Skipping page with pa 0x%jx\n", 486 (uintmax_t)pa); 487 else 488 vm_phys_add_page(pa); 489 pa += PAGE_SIZE; 490 } 491 } 492 freeenv(list); 493#if VM_NRESERVLEVEL > 0 494 /* 495 * Initialize the reservation management system. 496 */ 497 vm_reserv_init(); 498#endif 499 return (vaddr); 500} 501 502void 503vm_page_reference(vm_page_t m) 504{ 505 506 vm_page_aflag_set(m, PGA_REFERENCED); 507} 508 509/* 510 * vm_page_busy_downgrade: 511 * 512 * Downgrade an exclusive busy page into a single shared busy page. 513 */ 514void 515vm_page_busy_downgrade(vm_page_t m) 516{ 517 u_int x; 518 bool locked; 519 520 vm_page_assert_xbusied(m); 521 locked = mtx_owned(vm_page_lockptr(m)); 522 523 for (;;) { 524 x = m->busy_lock; 525 x &= VPB_BIT_WAITERS; 526 if (x != 0 && !locked) 527 vm_page_lock(m); 528 if (atomic_cmpset_rel_int(&m->busy_lock, 529 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 530 break; 531 if (x != 0 && !locked) 532 vm_page_unlock(m); 533 } 534 if (x != 0) { 535 wakeup(m); 536 if (!locked) 537 vm_page_unlock(m); 538 } 539} 540 541/* 542 * vm_page_sbusied: 543 * 544 * Return a positive value if the page is shared busied, 0 otherwise. 545 */ 546int 547vm_page_sbusied(vm_page_t m) 548{ 549 u_int x; 550 551 x = m->busy_lock; 552 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 553} 554 555/* 556 * vm_page_sunbusy: 557 * 558 * Shared unbusy a page. 559 */ 560void 561vm_page_sunbusy(vm_page_t m) 562{ 563 u_int x; 564 565 vm_page_assert_sbusied(m); 566 567 for (;;) { 568 x = m->busy_lock; 569 if (VPB_SHARERS(x) > 1) { 570 if (atomic_cmpset_int(&m->busy_lock, x, 571 x - VPB_ONE_SHARER)) 572 break; 573 continue; 574 } 575 if ((x & VPB_BIT_WAITERS) == 0) { 576 KASSERT(x == VPB_SHARERS_WORD(1), 577 ("vm_page_sunbusy: invalid lock state")); 578 if (atomic_cmpset_int(&m->busy_lock, 579 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 580 break; 581 continue; 582 } 583 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 584 ("vm_page_sunbusy: invalid lock state for waiters")); 585 586 vm_page_lock(m); 587 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 588 vm_page_unlock(m); 589 continue; 590 } 591 wakeup(m); 592 vm_page_unlock(m); 593 break; 594 } 595} 596 597/* 598 * vm_page_busy_sleep: 599 * 600 * Sleep and release the page lock, using the page pointer as wchan. 601 * This is used to implement the hard-path of busying mechanism. 602 * 603 * The given page must be locked. 604 * 605 * If nonshared is true, sleep only if the page is xbusy. 606 */ 607void 608vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 609{ 610 u_int x; 611 612 vm_page_assert_locked(m); 613 614 x = m->busy_lock; 615 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 616 ((x & VPB_BIT_WAITERS) == 0 && 617 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 618 vm_page_unlock(m); 619 return; 620 } 621 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 622} 623 624/* 625 * vm_page_trysbusy: 626 * 627 * Try to shared busy a page. 628 * If the operation succeeds 1 is returned otherwise 0. 629 * The operation never sleeps. 630 */ 631int 632vm_page_trysbusy(vm_page_t m) 633{ 634 u_int x; 635 636 for (;;) { 637 x = m->busy_lock; 638 if ((x & VPB_BIT_SHARED) == 0) 639 return (0); 640 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 641 return (1); 642 } 643} 644 645/* 646 * vm_page_xunbusy_hard: 647 * 648 * Called after the first try the exclusive unbusy of a page failed. 649 * It is assumed that the waiters bit is on. 650 */ 651void 652vm_page_xunbusy_hard(vm_page_t m) 653{ 654 655 vm_page_assert_xbusied(m); 656 657 vm_page_lock(m); 658 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 659 wakeup(m); 660 vm_page_unlock(m); 661} 662 663/* 664 * vm_page_flash: 665 * 666 * Wakeup anyone waiting for the page. 667 * The ownership bits do not change. 668 * 669 * The given page must be locked. 670 */ 671void 672vm_page_flash(vm_page_t m) 673{ 674 u_int x; 675 676 vm_page_lock_assert(m, MA_OWNED); 677 678 for (;;) { 679 x = m->busy_lock; 680 if ((x & VPB_BIT_WAITERS) == 0) 681 return; 682 if (atomic_cmpset_int(&m->busy_lock, x, 683 x & (~VPB_BIT_WAITERS))) 684 break; 685 } 686 wakeup(m); 687} 688 689/* 690 * Keep page from being freed by the page daemon 691 * much of the same effect as wiring, except much lower 692 * overhead and should be used only for *very* temporary 693 * holding ("wiring"). 694 */ 695void 696vm_page_hold(vm_page_t mem) 697{ 698 699 vm_page_lock_assert(mem, MA_OWNED); 700 mem->hold_count++; 701} 702 703void 704vm_page_unhold(vm_page_t mem) 705{ 706 707 vm_page_lock_assert(mem, MA_OWNED); 708 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 709 --mem->hold_count; 710 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 711 vm_page_free_toq(mem); 712} 713 714/* 715 * vm_page_unhold_pages: 716 * 717 * Unhold each of the pages that is referenced by the given array. 718 */ 719void 720vm_page_unhold_pages(vm_page_t *ma, int count) 721{ 722 struct mtx *mtx, *new_mtx; 723 724 mtx = NULL; 725 for (; count != 0; count--) { 726 /* 727 * Avoid releasing and reacquiring the same page lock. 728 */ 729 new_mtx = vm_page_lockptr(*ma); 730 if (mtx != new_mtx) { 731 if (mtx != NULL) 732 mtx_unlock(mtx); 733 mtx = new_mtx; 734 mtx_lock(mtx); 735 } 736 vm_page_unhold(*ma); 737 ma++; 738 } 739 if (mtx != NULL) 740 mtx_unlock(mtx); 741} 742 743vm_page_t 744PHYS_TO_VM_PAGE(vm_paddr_t pa) 745{ 746 vm_page_t m; 747 748#ifdef VM_PHYSSEG_SPARSE 749 m = vm_phys_paddr_to_vm_page(pa); 750 if (m == NULL) 751 m = vm_phys_fictitious_to_vm_page(pa); 752 return (m); 753#elif defined(VM_PHYSSEG_DENSE) 754 long pi; 755 756 pi = atop(pa); 757 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 758 m = &vm_page_array[pi - first_page]; 759 return (m); 760 } 761 return (vm_phys_fictitious_to_vm_page(pa)); 762#else 763#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 764#endif 765} 766 767/* 768 * vm_page_getfake: 769 * 770 * Create a fictitious page with the specified physical address and 771 * memory attribute. The memory attribute is the only the machine- 772 * dependent aspect of a fictitious page that must be initialized. 773 */ 774vm_page_t 775vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 776{ 777 vm_page_t m; 778 779 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 780 vm_page_initfake(m, paddr, memattr); 781 return (m); 782} 783 784void 785vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 786{ 787 788 if ((m->flags & PG_FICTITIOUS) != 0) { 789 /* 790 * The page's memattr might have changed since the 791 * previous initialization. Update the pmap to the 792 * new memattr. 793 */ 794 goto memattr; 795 } 796 m->phys_addr = paddr; 797 m->queue = PQ_NONE; 798 /* Fictitious pages don't use "segind". */ 799 m->flags = PG_FICTITIOUS; 800 /* Fictitious pages don't use "order" or "pool". */ 801 m->oflags = VPO_UNMANAGED; 802 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 803 m->wire_count = 1; 804 pmap_page_init(m); 805memattr: 806 pmap_page_set_memattr(m, memattr); 807} 808 809/* 810 * vm_page_putfake: 811 * 812 * Release a fictitious page. 813 */ 814void 815vm_page_putfake(vm_page_t m) 816{ 817 818 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 819 KASSERT((m->flags & PG_FICTITIOUS) != 0, 820 ("vm_page_putfake: bad page %p", m)); 821 uma_zfree(fakepg_zone, m); 822} 823 824/* 825 * vm_page_updatefake: 826 * 827 * Update the given fictitious page to the specified physical address and 828 * memory attribute. 829 */ 830void 831vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 832{ 833 834 KASSERT((m->flags & PG_FICTITIOUS) != 0, 835 ("vm_page_updatefake: bad page %p", m)); 836 m->phys_addr = paddr; 837 pmap_page_set_memattr(m, memattr); 838} 839 840/* 841 * vm_page_free: 842 * 843 * Free a page. 844 */ 845void 846vm_page_free(vm_page_t m) 847{ 848 849 m->flags &= ~PG_ZERO; 850 vm_page_free_toq(m); 851} 852 853/* 854 * vm_page_free_zero: 855 * 856 * Free a page to the zerod-pages queue 857 */ 858void 859vm_page_free_zero(vm_page_t m) 860{ 861 862 m->flags |= PG_ZERO; 863 vm_page_free_toq(m); 864} 865 866/* 867 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 868 * array which is not the request page. 869 */ 870void 871vm_page_readahead_finish(vm_page_t m) 872{ 873 874 if (m->valid != 0) { 875 /* 876 * Since the page is not the requested page, whether 877 * it should be activated or deactivated is not 878 * obvious. Empirical results have shown that 879 * deactivating the page is usually the best choice, 880 * unless the page is wanted by another thread. 881 */ 882 vm_page_lock(m); 883 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 884 vm_page_activate(m); 885 else 886 vm_page_deactivate(m); 887 vm_page_unlock(m); 888 vm_page_xunbusy(m); 889 } else { 890 /* 891 * Free the completely invalid page. Such page state 892 * occurs due to the short read operation which did 893 * not covered our page at all, or in case when a read 894 * error happens. 895 */ 896 vm_page_lock(m); 897 vm_page_free(m); 898 vm_page_unlock(m); 899 } 900} 901 902/* 903 * vm_page_sleep_if_busy: 904 * 905 * Sleep and release the page queues lock if the page is busied. 906 * Returns TRUE if the thread slept. 907 * 908 * The given page must be unlocked and object containing it must 909 * be locked. 910 */ 911int 912vm_page_sleep_if_busy(vm_page_t m, const char *msg) 913{ 914 vm_object_t obj; 915 916 vm_page_lock_assert(m, MA_NOTOWNED); 917 VM_OBJECT_ASSERT_WLOCKED(m->object); 918 919 if (vm_page_busied(m)) { 920 /* 921 * The page-specific object must be cached because page 922 * identity can change during the sleep, causing the 923 * re-lock of a different object. 924 * It is assumed that a reference to the object is already 925 * held by the callers. 926 */ 927 obj = m->object; 928 vm_page_lock(m); 929 VM_OBJECT_WUNLOCK(obj); 930 vm_page_busy_sleep(m, msg, false); 931 VM_OBJECT_WLOCK(obj); 932 return (TRUE); 933 } 934 return (FALSE); 935} 936 937/* 938 * vm_page_dirty_KBI: [ internal use only ] 939 * 940 * Set all bits in the page's dirty field. 941 * 942 * The object containing the specified page must be locked if the 943 * call is made from the machine-independent layer. 944 * 945 * See vm_page_clear_dirty_mask(). 946 * 947 * This function should only be called by vm_page_dirty(). 948 */ 949void 950vm_page_dirty_KBI(vm_page_t m) 951{ 952 953 /* These assertions refer to this operation by its public name. */ 954 KASSERT((m->flags & PG_CACHED) == 0, 955 ("vm_page_dirty: page in cache!")); 956 KASSERT(!VM_PAGE_IS_FREE(m), 957 ("vm_page_dirty: page is free!")); 958 KASSERT(m->valid == VM_PAGE_BITS_ALL, 959 ("vm_page_dirty: page is invalid!")); 960 m->dirty = VM_PAGE_BITS_ALL; 961} 962 963/* 964 * vm_page_insert: [ internal use only ] 965 * 966 * Inserts the given mem entry into the object and object list. 967 * 968 * The object must be locked. 969 */ 970int 971vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 972{ 973 vm_page_t mpred; 974 975 VM_OBJECT_ASSERT_WLOCKED(object); 976 mpred = vm_radix_lookup_le(&object->rtree, pindex); 977 return (vm_page_insert_after(m, object, pindex, mpred)); 978} 979 980/* 981 * vm_page_insert_after: 982 * 983 * Inserts the page "m" into the specified object at offset "pindex". 984 * 985 * The page "mpred" must immediately precede the offset "pindex" within 986 * the specified object. 987 * 988 * The object must be locked. 989 */ 990static int 991vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 992 vm_page_t mpred) 993{ 994 vm_page_t msucc; 995 996 VM_OBJECT_ASSERT_WLOCKED(object); 997 KASSERT(m->object == NULL, 998 ("vm_page_insert_after: page already inserted")); 999 if (mpred != NULL) { 1000 KASSERT(mpred->object == object, 1001 ("vm_page_insert_after: object doesn't contain mpred")); 1002 KASSERT(mpred->pindex < pindex, 1003 ("vm_page_insert_after: mpred doesn't precede pindex")); 1004 msucc = TAILQ_NEXT(mpred, listq); 1005 } else 1006 msucc = TAILQ_FIRST(&object->memq); 1007 if (msucc != NULL) 1008 KASSERT(msucc->pindex > pindex, 1009 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1010 1011 /* 1012 * Record the object/offset pair in this page 1013 */ 1014 m->object = object; 1015 m->pindex = pindex; 1016 1017 /* 1018 * Now link into the object's ordered list of backed pages. 1019 */ 1020 if (vm_radix_insert(&object->rtree, m)) { 1021 m->object = NULL; 1022 m->pindex = 0; 1023 return (1); 1024 } 1025 vm_page_insert_radixdone(m, object, mpred); 1026 return (0); 1027} 1028 1029/* 1030 * vm_page_insert_radixdone: 1031 * 1032 * Complete page "m" insertion into the specified object after the 1033 * radix trie hooking. 1034 * 1035 * The page "mpred" must precede the offset "m->pindex" within the 1036 * specified object. 1037 * 1038 * The object must be locked. 1039 */ 1040static void 1041vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1042{ 1043 1044 VM_OBJECT_ASSERT_WLOCKED(object); 1045 KASSERT(object != NULL && m->object == object, 1046 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1047 if (mpred != NULL) { 1048 KASSERT(mpred->object == object, 1049 ("vm_page_insert_after: object doesn't contain mpred")); 1050 KASSERT(mpred->pindex < m->pindex, 1051 ("vm_page_insert_after: mpred doesn't precede pindex")); 1052 } 1053 1054 if (mpred != NULL) 1055 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1056 else 1057 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1058 1059 /* 1060 * Show that the object has one more resident page. 1061 */ 1062 object->resident_page_count++; 1063 1064 /* 1065 * Hold the vnode until the last page is released. 1066 */ 1067 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1068 vhold(object->handle); 1069 1070 /* 1071 * Since we are inserting a new and possibly dirty page, 1072 * update the object's OBJ_MIGHTBEDIRTY flag. 1073 */ 1074 if (pmap_page_is_write_mapped(m)) 1075 vm_object_set_writeable_dirty(object); 1076} 1077 1078/* 1079 * vm_page_remove: 1080 * 1081 * Removes the given mem entry from the object/offset-page 1082 * table and the object page list, but do not invalidate/terminate 1083 * the backing store. 1084 * 1085 * The object must be locked. The page must be locked if it is managed. 1086 */ 1087void 1088vm_page_remove(vm_page_t m) 1089{ 1090 vm_object_t object; 1091 boolean_t lockacq; 1092 1093 if ((m->oflags & VPO_UNMANAGED) == 0) 1094 vm_page_lock_assert(m, MA_OWNED); 1095 if ((object = m->object) == NULL) 1096 return; 1097 VM_OBJECT_ASSERT_WLOCKED(object); 1098 if (vm_page_xbusied(m)) { 1099 lockacq = FALSE; 1100 if ((m->oflags & VPO_UNMANAGED) != 0 && 1101 !mtx_owned(vm_page_lockptr(m))) { 1102 lockacq = TRUE; 1103 vm_page_lock(m); 1104 } 1105 vm_page_flash(m); 1106 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1107 if (lockacq) 1108 vm_page_unlock(m); 1109 } 1110 1111 /* 1112 * Now remove from the object's list of backed pages. 1113 */ 1114 vm_radix_remove(&object->rtree, m->pindex); 1115 TAILQ_REMOVE(&object->memq, m, listq); 1116 1117 /* 1118 * And show that the object has one fewer resident page. 1119 */ 1120 object->resident_page_count--; 1121 1122 /* 1123 * The vnode may now be recycled. 1124 */ 1125 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1126 vdrop(object->handle); 1127 1128 m->object = NULL; 1129} 1130 1131/* 1132 * vm_page_lookup: 1133 * 1134 * Returns the page associated with the object/offset 1135 * pair specified; if none is found, NULL is returned. 1136 * 1137 * The object must be locked. 1138 */ 1139vm_page_t 1140vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1141{ 1142 1143 VM_OBJECT_ASSERT_LOCKED(object); 1144 return (vm_radix_lookup(&object->rtree, pindex)); 1145} 1146 1147/* 1148 * vm_page_find_least: 1149 * 1150 * Returns the page associated with the object with least pindex 1151 * greater than or equal to the parameter pindex, or NULL. 1152 * 1153 * The object must be locked. 1154 */ 1155vm_page_t 1156vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1157{ 1158 vm_page_t m; 1159 1160 VM_OBJECT_ASSERT_LOCKED(object); 1161 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1162 m = vm_radix_lookup_ge(&object->rtree, pindex); 1163 return (m); 1164} 1165 1166/* 1167 * Returns the given page's successor (by pindex) within the object if it is 1168 * resident; if none is found, NULL is returned. 1169 * 1170 * The object must be locked. 1171 */ 1172vm_page_t 1173vm_page_next(vm_page_t m) 1174{ 1175 vm_page_t next; 1176 1177 VM_OBJECT_ASSERT_WLOCKED(m->object); 1178 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1179 MPASS(next->object == m->object); 1180 if (next->pindex != m->pindex + 1) 1181 next = NULL; 1182 } 1183 return (next); 1184} 1185 1186/* 1187 * Returns the given page's predecessor (by pindex) within the object if it is 1188 * resident; if none is found, NULL is returned. 1189 * 1190 * The object must be locked. 1191 */ 1192vm_page_t 1193vm_page_prev(vm_page_t m) 1194{ 1195 vm_page_t prev; 1196 1197 VM_OBJECT_ASSERT_WLOCKED(m->object); 1198 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1199 MPASS(prev->object == m->object); 1200 if (prev->pindex != m->pindex - 1) 1201 prev = NULL; 1202 } 1203 return (prev); 1204} 1205 1206/* 1207 * Uses the page mnew as a replacement for an existing page at index 1208 * pindex which must be already present in the object. 1209 * 1210 * The existing page must not be on a paging queue. 1211 */ 1212vm_page_t 1213vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1214{ 1215 vm_page_t mold, mpred; 1216 1217 VM_OBJECT_ASSERT_WLOCKED(object); 1218 1219 /* 1220 * This function mostly follows vm_page_insert() and 1221 * vm_page_remove() without the radix, object count and vnode 1222 * dance. Double check such functions for more comments. 1223 */ 1224 mpred = vm_radix_lookup(&object->rtree, pindex); 1225 KASSERT(mpred != NULL, 1226 ("vm_page_replace: replacing page not present with pindex")); 1227 mpred = TAILQ_PREV(mpred, respgs, listq); 1228 if (mpred != NULL) 1229 KASSERT(mpred->pindex < pindex, 1230 ("vm_page_insert_after: mpred doesn't precede pindex")); 1231 1232 mnew->object = object; 1233 mnew->pindex = pindex; 1234 mold = vm_radix_replace(&object->rtree, mnew); 1235 KASSERT(mold->queue == PQ_NONE, 1236 ("vm_page_replace: mold is on a paging queue")); 1237 1238 /* Detach the old page from the resident tailq. */ 1239 TAILQ_REMOVE(&object->memq, mold, listq); 1240 1241 mold->object = NULL; 1242 vm_page_xunbusy(mold); 1243 1244 /* Insert the new page in the resident tailq. */ 1245 if (mpred != NULL) 1246 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq); 1247 else 1248 TAILQ_INSERT_HEAD(&object->memq, mnew, listq); 1249 if (pmap_page_is_write_mapped(mnew)) 1250 vm_object_set_writeable_dirty(object); 1251 return (mold); 1252} 1253 1254/* 1255 * vm_page_rename: 1256 * 1257 * Move the given memory entry from its 1258 * current object to the specified target object/offset. 1259 * 1260 * Note: swap associated with the page must be invalidated by the move. We 1261 * have to do this for several reasons: (1) we aren't freeing the 1262 * page, (2) we are dirtying the page, (3) the VM system is probably 1263 * moving the page from object A to B, and will then later move 1264 * the backing store from A to B and we can't have a conflict. 1265 * 1266 * Note: we *always* dirty the page. It is necessary both for the 1267 * fact that we moved it, and because we may be invalidating 1268 * swap. If the page is on the cache, we have to deactivate it 1269 * or vm_page_dirty() will panic. Dirty pages are not allowed 1270 * on the cache. 1271 * 1272 * The objects must be locked. 1273 */ 1274int 1275vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1276{ 1277 vm_page_t mpred; 1278 vm_pindex_t opidx; 1279 1280 VM_OBJECT_ASSERT_WLOCKED(new_object); 1281 1282 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1283 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1284 ("vm_page_rename: pindex already renamed")); 1285 1286 /* 1287 * Create a custom version of vm_page_insert() which does not depend 1288 * by m_prev and can cheat on the implementation aspects of the 1289 * function. 1290 */ 1291 opidx = m->pindex; 1292 m->pindex = new_pindex; 1293 if (vm_radix_insert(&new_object->rtree, m)) { 1294 m->pindex = opidx; 1295 return (1); 1296 } 1297 1298 /* 1299 * The operation cannot fail anymore. The removal must happen before 1300 * the listq iterator is tainted. 1301 */ 1302 m->pindex = opidx; 1303 vm_page_lock(m); 1304 vm_page_remove(m); 1305 1306 /* Return back to the new pindex to complete vm_page_insert(). */ 1307 m->pindex = new_pindex; 1308 m->object = new_object; 1309 vm_page_unlock(m); 1310 vm_page_insert_radixdone(m, new_object, mpred); 1311 vm_page_dirty(m); 1312 return (0); 1313} 1314 1315/* 1316 * Convert all of the given object's cached pages that have a 1317 * pindex within the given range into free pages. If the value 1318 * zero is given for "end", then the range's upper bound is 1319 * infinity. If the given object is backed by a vnode and it 1320 * transitions from having one or more cached pages to none, the 1321 * vnode's hold count is reduced. 1322 */ 1323void 1324vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1325{ 1326 vm_page_t m; 1327 boolean_t empty; 1328 1329 mtx_lock(&vm_page_queue_free_mtx); 1330 if (__predict_false(vm_radix_is_empty(&object->cache))) { 1331 mtx_unlock(&vm_page_queue_free_mtx); 1332 return; 1333 } 1334 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) { 1335 if (end != 0 && m->pindex >= end) 1336 break; 1337 vm_radix_remove(&object->cache, m->pindex); 1338 vm_page_cache_turn_free(m); 1339 } 1340 empty = vm_radix_is_empty(&object->cache); 1341 mtx_unlock(&vm_page_queue_free_mtx); 1342 if (object->type == OBJT_VNODE && empty) 1343 vdrop(object->handle); 1344} 1345 1346/* 1347 * Returns the cached page that is associated with the given 1348 * object and offset. If, however, none exists, returns NULL. 1349 * 1350 * The free page queue must be locked. 1351 */ 1352static inline vm_page_t 1353vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1354{ 1355 1356 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1357 return (vm_radix_lookup(&object->cache, pindex)); 1358} 1359 1360/* 1361 * Remove the given cached page from its containing object's 1362 * collection of cached pages. 1363 * 1364 * The free page queue must be locked. 1365 */ 1366static void 1367vm_page_cache_remove(vm_page_t m) 1368{ 1369 1370 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1371 KASSERT((m->flags & PG_CACHED) != 0, 1372 ("vm_page_cache_remove: page %p is not cached", m)); 1373 vm_radix_remove(&m->object->cache, m->pindex); 1374 m->object = NULL; 1375 cnt.v_cache_count--; 1376} 1377 1378/* 1379 * Transfer all of the cached pages with offset greater than or 1380 * equal to 'offidxstart' from the original object's cache to the 1381 * new object's cache. However, any cached pages with offset 1382 * greater than or equal to the new object's size are kept in the 1383 * original object. Initially, the new object's cache must be 1384 * empty. Offset 'offidxstart' in the original object must 1385 * correspond to offset zero in the new object. 1386 * 1387 * The new object must be locked. 1388 */ 1389void 1390vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1391 vm_object_t new_object) 1392{ 1393 vm_page_t m; 1394 1395 /* 1396 * Insertion into an object's collection of cached pages 1397 * requires the object to be locked. In contrast, removal does 1398 * not. 1399 */ 1400 VM_OBJECT_ASSERT_WLOCKED(new_object); 1401 KASSERT(vm_radix_is_empty(&new_object->cache), 1402 ("vm_page_cache_transfer: object %p has cached pages", 1403 new_object)); 1404 mtx_lock(&vm_page_queue_free_mtx); 1405 while ((m = vm_radix_lookup_ge(&orig_object->cache, 1406 offidxstart)) != NULL) { 1407 /* 1408 * Transfer all of the pages with offset greater than or 1409 * equal to 'offidxstart' from the original object's 1410 * cache to the new object's cache. 1411 */ 1412 if ((m->pindex - offidxstart) >= new_object->size) 1413 break; 1414 vm_radix_remove(&orig_object->cache, m->pindex); 1415 /* Update the page's object and offset. */ 1416 m->object = new_object; 1417 m->pindex -= offidxstart; 1418 if (vm_radix_insert(&new_object->cache, m)) 1419 vm_page_cache_turn_free(m); 1420 } 1421 mtx_unlock(&vm_page_queue_free_mtx); 1422} 1423 1424/* 1425 * Returns TRUE if a cached page is associated with the given object and 1426 * offset, and FALSE otherwise. 1427 * 1428 * The object must be locked. 1429 */ 1430boolean_t 1431vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1432{ 1433 vm_page_t m; 1434 1435 /* 1436 * Insertion into an object's collection of cached pages requires the 1437 * object to be locked. Therefore, if the object is locked and the 1438 * object's collection is empty, there is no need to acquire the free 1439 * page queues lock in order to prove that the specified page doesn't 1440 * exist. 1441 */ 1442 VM_OBJECT_ASSERT_WLOCKED(object); 1443 if (__predict_true(vm_object_cache_is_empty(object))) 1444 return (FALSE); 1445 mtx_lock(&vm_page_queue_free_mtx); 1446 m = vm_page_cache_lookup(object, pindex); 1447 mtx_unlock(&vm_page_queue_free_mtx); 1448 return (m != NULL); 1449} 1450 1451/* 1452 * vm_page_alloc: 1453 * 1454 * Allocate and return a page that is associated with the specified 1455 * object and offset pair. By default, this page is exclusive busied. 1456 * 1457 * The caller must always specify an allocation class. 1458 * 1459 * allocation classes: 1460 * VM_ALLOC_NORMAL normal process request 1461 * VM_ALLOC_SYSTEM system *really* needs a page 1462 * VM_ALLOC_INTERRUPT interrupt time request 1463 * 1464 * optional allocation flags: 1465 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1466 * intends to allocate 1467 * VM_ALLOC_IFCACHED return page only if it is cached 1468 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1469 * is cached 1470 * VM_ALLOC_NOBUSY do not exclusive busy the page 1471 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1472 * VM_ALLOC_NOOBJ page is not associated with an object and 1473 * should not be exclusive busy 1474 * VM_ALLOC_SBUSY shared busy the allocated page 1475 * VM_ALLOC_WIRED wire the allocated page 1476 * VM_ALLOC_ZERO prefer a zeroed page 1477 * 1478 * This routine may not sleep. 1479 */ 1480vm_page_t 1481vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1482{ 1483 struct vnode *vp = NULL; 1484 vm_object_t m_object; 1485 vm_page_t m, mpred; 1486 int flags, req_class; 1487 1488 mpred = 0; /* XXX: pacify gcc */ 1489 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1490 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1491 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1492 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1493 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1494 req)); 1495 if (object != NULL) 1496 VM_OBJECT_ASSERT_WLOCKED(object); 1497 1498 req_class = req & VM_ALLOC_CLASS_MASK; 1499 1500 /* 1501 * The page daemon is allowed to dig deeper into the free page list. 1502 */ 1503 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1504 req_class = VM_ALLOC_SYSTEM; 1505 1506 if (object != NULL) { 1507 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1508 KASSERT(mpred == NULL || mpred->pindex != pindex, 1509 ("vm_page_alloc: pindex already allocated")); 1510 } 1511 1512 /* 1513 * The page allocation request can came from consumers which already 1514 * hold the free page queue mutex, like vm_page_insert() in 1515 * vm_page_cache(). 1516 */ 1517 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1518 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1519 (req_class == VM_ALLOC_SYSTEM && 1520 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1521 (req_class == VM_ALLOC_INTERRUPT && 1522 cnt.v_free_count + cnt.v_cache_count > 0)) { 1523 /* 1524 * Allocate from the free queue if the number of free pages 1525 * exceeds the minimum for the request class. 1526 */ 1527 if (object != NULL && 1528 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1529 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1530 mtx_unlock(&vm_page_queue_free_mtx); 1531 return (NULL); 1532 } 1533 if (vm_phys_unfree_page(m)) 1534 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1535#if VM_NRESERVLEVEL > 0 1536 else if (!vm_reserv_reactivate_page(m)) 1537#else 1538 else 1539#endif 1540 panic("vm_page_alloc: cache page %p is missing" 1541 " from the free queue", m); 1542 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1543 mtx_unlock(&vm_page_queue_free_mtx); 1544 return (NULL); 1545#if VM_NRESERVLEVEL > 0 1546 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1547 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1548 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) { 1549#else 1550 } else { 1551#endif 1552 m = vm_phys_alloc_pages(object != NULL ? 1553 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1554#if VM_NRESERVLEVEL > 0 1555 if (m == NULL && vm_reserv_reclaim_inactive()) { 1556 m = vm_phys_alloc_pages(object != NULL ? 1557 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1558 0); 1559 } 1560#endif 1561 } 1562 } else { 1563 /* 1564 * Not allocatable, give up. 1565 */ 1566 mtx_unlock(&vm_page_queue_free_mtx); 1567 atomic_add_int(&vm_pageout_deficit, 1568 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1569 pagedaemon_wakeup(); 1570 return (NULL); 1571 } 1572 1573 /* 1574 * At this point we had better have found a good page. 1575 */ 1576 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1577 KASSERT(m->queue == PQ_NONE, 1578 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1579 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1580 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1581 KASSERT(!vm_page_busied(m), ("vm_page_alloc: page %p is busy", m)); 1582 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1583 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1584 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1585 pmap_page_get_memattr(m))); 1586 if ((m->flags & PG_CACHED) != 0) { 1587 KASSERT((m->flags & PG_ZERO) == 0, 1588 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1589 KASSERT(m->valid != 0, 1590 ("vm_page_alloc: cached page %p is invalid", m)); 1591 if (m->object == object && m->pindex == pindex) 1592 cnt.v_reactivated++; 1593 else 1594 m->valid = 0; 1595 m_object = m->object; 1596 vm_page_cache_remove(m); 1597 if (m_object->type == OBJT_VNODE && 1598 vm_object_cache_is_empty(m_object)) 1599 vp = m_object->handle; 1600 } else { 1601 KASSERT(VM_PAGE_IS_FREE(m), 1602 ("vm_page_alloc: page %p is not free", m)); 1603 KASSERT(m->valid == 0, 1604 ("vm_page_alloc: free page %p is valid", m)); 1605 vm_phys_freecnt_adj(m, -1); 1606 } 1607 1608 /* 1609 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag 1610 * must be cleared before the free page queues lock is released. 1611 */ 1612 flags = 0; 1613 if (m->flags & PG_ZERO) { 1614 vm_page_zero_count--; 1615 if (req & VM_ALLOC_ZERO) 1616 flags = PG_ZERO; 1617 } 1618 if (req & VM_ALLOC_NODUMP) 1619 flags |= PG_NODUMP; 1620 m->flags = flags; 1621 mtx_unlock(&vm_page_queue_free_mtx); 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(&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 /* See the comment below about hold count. */ 1643 if (vp != NULL) 1644 vdrop(vp); 1645 pagedaemon_wakeup(); 1646 if (req & VM_ALLOC_WIRED) { 1647 atomic_subtract_int(&cnt.v_wire_count, 1); 1648 m->wire_count = 0; 1649 } 1650 m->object = NULL; 1651 m->oflags = VPO_UNMANAGED; 1652 m->busy_lock = VPB_UNBUSIED; 1653 vm_page_free(m); 1654 return (NULL); 1655 } 1656 1657 /* Ignore device objects; the pager sets "memattr" for them. */ 1658 if (object->memattr != VM_MEMATTR_DEFAULT && 1659 (object->flags & OBJ_FICTITIOUS) == 0) 1660 pmap_page_set_memattr(m, object->memattr); 1661 } else 1662 m->pindex = pindex; 1663 1664 /* 1665 * The following call to vdrop() must come after the above call 1666 * to vm_page_insert() in case both affect the same object and 1667 * vnode. Otherwise, the affected vnode's hold count could 1668 * temporarily become zero. 1669 */ 1670 if (vp != NULL) 1671 vdrop(vp); 1672 1673 /* 1674 * Don't wakeup too often - wakeup the pageout daemon when 1675 * we would be nearly out of memory. 1676 */ 1677 if (vm_paging_needed()) 1678 pagedaemon_wakeup(); 1679 1680 return (m); 1681} 1682 1683static void 1684vm_page_alloc_contig_vdrop(struct spglist *lst) 1685{ 1686 1687 while (!SLIST_EMPTY(lst)) { 1688 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv); 1689 SLIST_REMOVE_HEAD(lst, plinks.s.ss); 1690 } 1691} 1692 1693/* 1694 * vm_page_alloc_contig: 1695 * 1696 * Allocate a contiguous set of physical pages of the given size "npages" 1697 * from the free lists. All of the physical pages must be at or above 1698 * the given physical address "low" and below the given physical address 1699 * "high". The given value "alignment" determines the alignment of the 1700 * first physical page in the set. If the given value "boundary" is 1701 * non-zero, then the set of physical pages cannot cross any physical 1702 * address boundary that is a multiple of that value. Both "alignment" 1703 * and "boundary" must be a power of two. 1704 * 1705 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1706 * then the memory attribute setting for the physical pages is configured 1707 * to the object's memory attribute setting. Otherwise, the memory 1708 * attribute setting for the physical pages is configured to "memattr", 1709 * overriding the object's memory attribute setting. However, if the 1710 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1711 * memory attribute setting for the physical pages cannot be configured 1712 * to VM_MEMATTR_DEFAULT. 1713 * 1714 * The caller must always specify an allocation class. 1715 * 1716 * allocation classes: 1717 * VM_ALLOC_NORMAL normal process request 1718 * VM_ALLOC_SYSTEM system *really* needs a page 1719 * VM_ALLOC_INTERRUPT interrupt time request 1720 * 1721 * optional allocation flags: 1722 * VM_ALLOC_NOBUSY do not exclusive busy the page 1723 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1724 * VM_ALLOC_NOOBJ page is not associated with an object and 1725 * should not be exclusive busy 1726 * VM_ALLOC_SBUSY shared busy the allocated page 1727 * VM_ALLOC_WIRED wire the allocated page 1728 * VM_ALLOC_ZERO prefer a zeroed page 1729 * 1730 * This routine may not sleep. 1731 */ 1732vm_page_t 1733vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1734 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1735 vm_paddr_t boundary, vm_memattr_t memattr) 1736{ 1737 struct vnode *drop; 1738 struct spglist deferred_vdrop_list; 1739 vm_page_t m, m_tmp, m_ret; 1740 u_int flags, oflags; 1741 int req_class; 1742 1743 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1744 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1745 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1746 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1747 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1748 req)); 1749 if (object != NULL) { 1750 VM_OBJECT_ASSERT_WLOCKED(object); 1751 KASSERT(object->type == OBJT_PHYS, 1752 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1753 object)); 1754 } 1755 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1756 req_class = req & VM_ALLOC_CLASS_MASK; 1757 1758 /* 1759 * The page daemon is allowed to dig deeper into the free page list. 1760 */ 1761 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1762 req_class = VM_ALLOC_SYSTEM; 1763 1764 SLIST_INIT(&deferred_vdrop_list); 1765 mtx_lock(&vm_page_queue_free_mtx); 1766 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1767 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1768 cnt.v_free_count + cnt.v_cache_count >= npages + 1769 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1770 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1771#if VM_NRESERVLEVEL > 0 1772retry: 1773 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1774 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1775 low, high, alignment, boundary)) == NULL) 1776#endif 1777 m_ret = vm_phys_alloc_contig(npages, low, high, 1778 alignment, boundary); 1779 } else { 1780 mtx_unlock(&vm_page_queue_free_mtx); 1781 atomic_add_int(&vm_pageout_deficit, npages); 1782 pagedaemon_wakeup(); 1783 return (NULL); 1784 } 1785 if (m_ret != NULL) 1786 for (m = m_ret; m < &m_ret[npages]; m++) { 1787 drop = vm_page_alloc_init(m); 1788 if (drop != NULL) { 1789 /* 1790 * Enqueue the vnode for deferred vdrop(). 1791 */ 1792 m->plinks.s.pv = drop; 1793 SLIST_INSERT_HEAD(&deferred_vdrop_list, m, 1794 plinks.s.ss); 1795 } 1796 } 1797 else { 1798#if VM_NRESERVLEVEL > 0 1799 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1800 boundary)) 1801 goto retry; 1802#endif 1803 } 1804 mtx_unlock(&vm_page_queue_free_mtx); 1805 if (m_ret == NULL) 1806 return (NULL); 1807 1808 /* 1809 * Initialize the pages. Only the PG_ZERO flag is inherited. 1810 */ 1811 flags = 0; 1812 if ((req & VM_ALLOC_ZERO) != 0) 1813 flags = PG_ZERO; 1814 if ((req & VM_ALLOC_NODUMP) != 0) 1815 flags |= PG_NODUMP; 1816 if ((req & VM_ALLOC_WIRED) != 0) 1817 atomic_add_int(&cnt.v_wire_count, npages); 1818 oflags = VPO_UNMANAGED; 1819 if (object != NULL) { 1820 if (object->memattr != VM_MEMATTR_DEFAULT && 1821 memattr == VM_MEMATTR_DEFAULT) 1822 memattr = object->memattr; 1823 } 1824 for (m = m_ret; m < &m_ret[npages]; m++) { 1825 m->aflags = 0; 1826 m->flags = (m->flags | PG_NODUMP) & flags; 1827 m->busy_lock = VPB_UNBUSIED; 1828 if (object != NULL) { 1829 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1830 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1831 if ((req & VM_ALLOC_SBUSY) != 0) 1832 m->busy_lock = VPB_SHARERS_WORD(1); 1833 } 1834 if ((req & VM_ALLOC_WIRED) != 0) 1835 m->wire_count = 1; 1836 /* Unmanaged pages don't use "act_count". */ 1837 m->oflags = oflags; 1838 if (object != NULL) { 1839 if (vm_page_insert(m, object, pindex)) { 1840 vm_page_alloc_contig_vdrop( 1841 &deferred_vdrop_list); 1842 if (vm_paging_needed()) 1843 pagedaemon_wakeup(); 1844 if ((req & VM_ALLOC_WIRED) != 0) 1845 atomic_subtract_int(&cnt.v_wire_count, 1846 npages); 1847 for (m_tmp = m, m = m_ret; 1848 m < &m_ret[npages]; m++) { 1849 if ((req & VM_ALLOC_WIRED) != 0) 1850 m->wire_count = 0; 1851 if (m >= m_tmp) { 1852 m->object = NULL; 1853 m->oflags |= VPO_UNMANAGED; 1854 } 1855 m->busy_lock = VPB_UNBUSIED; 1856 vm_page_free(m); 1857 } 1858 return (NULL); 1859 } 1860 } else 1861 m->pindex = pindex; 1862 if (memattr != VM_MEMATTR_DEFAULT) 1863 pmap_page_set_memattr(m, memattr); 1864 pindex++; 1865 } 1866 vm_page_alloc_contig_vdrop(&deferred_vdrop_list); 1867 if (vm_paging_needed()) 1868 pagedaemon_wakeup(); 1869 return (m_ret); 1870} 1871 1872/* 1873 * Initialize a page that has been freshly dequeued from a freelist. 1874 * The caller has to drop the vnode returned, if it is not NULL. 1875 * 1876 * This function may only be used to initialize unmanaged pages. 1877 * 1878 * To be called with vm_page_queue_free_mtx held. 1879 */ 1880static struct vnode * 1881vm_page_alloc_init(vm_page_t m) 1882{ 1883 struct vnode *drop; 1884 vm_object_t m_object; 1885 1886 KASSERT(m->queue == PQ_NONE, 1887 ("vm_page_alloc_init: page %p has unexpected queue %d", 1888 m, m->queue)); 1889 KASSERT(m->wire_count == 0, 1890 ("vm_page_alloc_init: page %p is wired", m)); 1891 KASSERT(m->hold_count == 0, 1892 ("vm_page_alloc_init: page %p is held", m)); 1893 KASSERT(!vm_page_busied(m), 1894 ("vm_page_alloc_init: page %p is busy", m)); 1895 KASSERT(m->dirty == 0, 1896 ("vm_page_alloc_init: page %p is dirty", m)); 1897 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1898 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1899 m, pmap_page_get_memattr(m))); 1900 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1901 drop = NULL; 1902 if ((m->flags & PG_CACHED) != 0) { 1903 KASSERT((m->flags & PG_ZERO) == 0, 1904 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1905 m->valid = 0; 1906 m_object = m->object; 1907 vm_page_cache_remove(m); 1908 if (m_object->type == OBJT_VNODE && 1909 vm_object_cache_is_empty(m_object)) 1910 drop = m_object->handle; 1911 } else { 1912 KASSERT(VM_PAGE_IS_FREE(m), 1913 ("vm_page_alloc_init: page %p is not free", m)); 1914 KASSERT(m->valid == 0, 1915 ("vm_page_alloc_init: free page %p is valid", m)); 1916 vm_phys_freecnt_adj(m, -1); 1917 if ((m->flags & PG_ZERO) != 0) 1918 vm_page_zero_count--; 1919 } 1920 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1921 m->flags &= PG_ZERO; 1922 return (drop); 1923} 1924 1925/* 1926 * vm_page_alloc_freelist: 1927 * 1928 * Allocate a physical page from the specified free page list. 1929 * 1930 * The caller must always specify an allocation class. 1931 * 1932 * allocation classes: 1933 * VM_ALLOC_NORMAL normal process request 1934 * VM_ALLOC_SYSTEM system *really* needs a page 1935 * VM_ALLOC_INTERRUPT interrupt time request 1936 * 1937 * optional allocation flags: 1938 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1939 * intends to allocate 1940 * VM_ALLOC_WIRED wire the allocated page 1941 * VM_ALLOC_ZERO prefer a zeroed page 1942 * 1943 * This routine may not sleep. 1944 */ 1945vm_page_t 1946vm_page_alloc_freelist(int flind, int req) 1947{ 1948 struct vnode *drop; 1949 vm_page_t m; 1950 u_int flags; 1951 int req_class; 1952 1953 req_class = req & VM_ALLOC_CLASS_MASK; 1954 1955 /* 1956 * The page daemon is allowed to dig deeper into the free page list. 1957 */ 1958 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1959 req_class = VM_ALLOC_SYSTEM; 1960 1961 /* 1962 * Do not allocate reserved pages unless the req has asked for it. 1963 */ 1964 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1965 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1966 (req_class == VM_ALLOC_SYSTEM && 1967 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1968 (req_class == VM_ALLOC_INTERRUPT && 1969 cnt.v_free_count + cnt.v_cache_count > 0)) 1970 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1971 else { 1972 mtx_unlock(&vm_page_queue_free_mtx); 1973 atomic_add_int(&vm_pageout_deficit, 1974 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1975 pagedaemon_wakeup(); 1976 return (NULL); 1977 } 1978 if (m == NULL) { 1979 mtx_unlock(&vm_page_queue_free_mtx); 1980 return (NULL); 1981 } 1982 drop = vm_page_alloc_init(m); 1983 mtx_unlock(&vm_page_queue_free_mtx); 1984 1985 /* 1986 * Initialize the page. Only the PG_ZERO flag is inherited. 1987 */ 1988 m->aflags = 0; 1989 flags = 0; 1990 if ((req & VM_ALLOC_ZERO) != 0) 1991 flags = PG_ZERO; 1992 m->flags &= flags; 1993 if ((req & VM_ALLOC_WIRED) != 0) { 1994 /* 1995 * The page lock is not required for wiring a page that does 1996 * not belong to an object. 1997 */ 1998 atomic_add_int(&cnt.v_wire_count, 1); 1999 m->wire_count = 1; 2000 } 2001 /* Unmanaged pages don't use "act_count". */ 2002 m->oflags = VPO_UNMANAGED; 2003 if (drop != NULL) 2004 vdrop(drop); 2005 if (vm_paging_needed()) 2006 pagedaemon_wakeup(); 2007 return (m); 2008} 2009 2010/* 2011 * vm_wait: (also see VM_WAIT macro) 2012 * 2013 * Sleep until free pages are available for allocation. 2014 * - Called in various places before memory allocations. 2015 */ 2016void 2017vm_wait(void) 2018{ 2019 2020 mtx_lock(&vm_page_queue_free_mtx); 2021 if (curproc == pageproc) { 2022 vm_pageout_pages_needed = 1; 2023 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2024 PDROP | PSWP, "VMWait", 0); 2025 } else { 2026 if (!vm_pages_needed) { 2027 vm_pages_needed = 1; 2028 wakeup(&vm_pages_needed); 2029 } 2030 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2031 "vmwait", 0); 2032 } 2033} 2034 2035/* 2036 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2037 * 2038 * Sleep until free pages are available for allocation. 2039 * - Called only in vm_fault so that processes page faulting 2040 * can be easily tracked. 2041 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2042 * processes will be able to grab memory first. Do not change 2043 * this balance without careful testing first. 2044 */ 2045void 2046vm_waitpfault(void) 2047{ 2048 2049 mtx_lock(&vm_page_queue_free_mtx); 2050 if (!vm_pages_needed) { 2051 vm_pages_needed = 1; 2052 wakeup(&vm_pages_needed); 2053 } 2054 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2055 "pfault", 0); 2056} 2057 2058struct vm_pagequeue * 2059vm_page_pagequeue(vm_page_t m) 2060{ 2061 2062 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2063} 2064 2065/* 2066 * vm_page_dequeue: 2067 * 2068 * Remove the given page from its current page queue. 2069 * 2070 * The page must be locked. 2071 */ 2072void 2073vm_page_dequeue(vm_page_t m) 2074{ 2075 struct vm_pagequeue *pq; 2076 2077 vm_page_lock_assert(m, MA_OWNED); 2078 KASSERT(m->queue != PQ_NONE, 2079 ("vm_page_dequeue: page %p is not queued", m)); 2080 pq = vm_page_pagequeue(m); 2081 vm_pagequeue_lock(pq); 2082 m->queue = PQ_NONE; 2083 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2084 vm_pagequeue_cnt_dec(pq); 2085 vm_pagequeue_unlock(pq); 2086} 2087 2088/* 2089 * vm_page_dequeue_locked: 2090 * 2091 * Remove the given page from its current page queue. 2092 * 2093 * The page and page queue must be locked. 2094 */ 2095void 2096vm_page_dequeue_locked(vm_page_t m) 2097{ 2098 struct vm_pagequeue *pq; 2099 2100 vm_page_lock_assert(m, MA_OWNED); 2101 pq = vm_page_pagequeue(m); 2102 vm_pagequeue_assert_locked(pq); 2103 m->queue = PQ_NONE; 2104 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2105 vm_pagequeue_cnt_dec(pq); 2106} 2107 2108/* 2109 * vm_page_enqueue: 2110 * 2111 * Add the given page to the specified page queue. 2112 * 2113 * The page must be locked. 2114 */ 2115static void 2116vm_page_enqueue(int queue, vm_page_t m) 2117{ 2118 struct vm_pagequeue *pq; 2119 2120 vm_page_lock_assert(m, MA_OWNED); 2121 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2122 vm_pagequeue_lock(pq); 2123 m->queue = queue; 2124 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2125 vm_pagequeue_cnt_inc(pq); 2126 vm_pagequeue_unlock(pq); 2127} 2128 2129/* 2130 * vm_page_requeue: 2131 * 2132 * Move the given page to the tail of its current page queue. 2133 * 2134 * The page must be locked. 2135 */ 2136void 2137vm_page_requeue(vm_page_t m) 2138{ 2139 struct vm_pagequeue *pq; 2140 2141 vm_page_lock_assert(m, MA_OWNED); 2142 KASSERT(m->queue != PQ_NONE, 2143 ("vm_page_requeue: page %p is not queued", m)); 2144 pq = vm_page_pagequeue(m); 2145 vm_pagequeue_lock(pq); 2146 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2147 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2148 vm_pagequeue_unlock(pq); 2149} 2150 2151/* 2152 * vm_page_requeue_locked: 2153 * 2154 * Move the given page to the tail of its current page queue. 2155 * 2156 * The page queue must be locked. 2157 */ 2158void 2159vm_page_requeue_locked(vm_page_t m) 2160{ 2161 struct vm_pagequeue *pq; 2162 2163 KASSERT(m->queue != PQ_NONE, 2164 ("vm_page_requeue_locked: page %p is not queued", m)); 2165 pq = vm_page_pagequeue(m); 2166 vm_pagequeue_assert_locked(pq); 2167 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2168 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2169} 2170 2171/* 2172 * vm_page_activate: 2173 * 2174 * Put the specified page on the active list (if appropriate). 2175 * Ensure that act_count is at least ACT_INIT but do not otherwise 2176 * mess with it. 2177 * 2178 * The page must be locked. 2179 */ 2180void 2181vm_page_activate(vm_page_t m) 2182{ 2183 int queue; 2184 2185 vm_page_lock_assert(m, MA_OWNED); 2186 if ((queue = m->queue) != PQ_ACTIVE) { 2187 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2188 if (m->act_count < ACT_INIT) 2189 m->act_count = ACT_INIT; 2190 if (queue != PQ_NONE) 2191 vm_page_dequeue(m); 2192 vm_page_enqueue(PQ_ACTIVE, m); 2193 } else 2194 KASSERT(queue == PQ_NONE, 2195 ("vm_page_activate: wired page %p is queued", m)); 2196 } else { 2197 if (m->act_count < ACT_INIT) 2198 m->act_count = ACT_INIT; 2199 } 2200} 2201 2202/* 2203 * vm_page_free_wakeup: 2204 * 2205 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2206 * routine is called when a page has been added to the cache or free 2207 * queues. 2208 * 2209 * The page queues must be locked. 2210 */ 2211static inline void 2212vm_page_free_wakeup(void) 2213{ 2214 2215 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2216 /* 2217 * if pageout daemon needs pages, then tell it that there are 2218 * some free. 2219 */ 2220 if (vm_pageout_pages_needed && 2221 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2222 wakeup(&vm_pageout_pages_needed); 2223 vm_pageout_pages_needed = 0; 2224 } 2225 /* 2226 * wakeup processes that are waiting on memory if we hit a 2227 * high water mark. And wakeup scheduler process if we have 2228 * lots of memory. this process will swapin processes. 2229 */ 2230 if (vm_pages_needed && !vm_page_count_min()) { 2231 vm_pages_needed = 0; 2232 wakeup(&cnt.v_free_count); 2233 } 2234} 2235 2236/* 2237 * Turn a cached page into a free page, by changing its attributes. 2238 * Keep the statistics up-to-date. 2239 * 2240 * The free page queue must be locked. 2241 */ 2242static void 2243vm_page_cache_turn_free(vm_page_t m) 2244{ 2245 2246 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2247 2248 m->object = NULL; 2249 m->valid = 0; 2250 /* Clear PG_CACHED and set PG_FREE. */ 2251 m->flags ^= PG_CACHED | PG_FREE; 2252 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 2253 ("vm_page_cache_free: page %p has inconsistent flags", m)); 2254 cnt.v_cache_count--; 2255 vm_phys_freecnt_adj(m, 1); 2256} 2257 2258/* 2259 * vm_page_free_toq: 2260 * 2261 * Returns the given page to the free list, 2262 * disassociating it with any VM object. 2263 * 2264 * The object must be locked. The page must be locked if it is managed. 2265 */ 2266void 2267vm_page_free_toq(vm_page_t m) 2268{ 2269 2270 if ((m->oflags & VPO_UNMANAGED) == 0) { 2271 vm_page_lock_assert(m, MA_OWNED); 2272 KASSERT(!pmap_page_is_mapped(m), 2273 ("vm_page_free_toq: freeing mapped page %p", m)); 2274 } else 2275 KASSERT(m->queue == PQ_NONE, 2276 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2277 PCPU_INC(cnt.v_tfree); 2278 2279 if (VM_PAGE_IS_FREE(m)) 2280 panic("vm_page_free: freeing free page %p", m); 2281 else if (vm_page_sbusied(m)) 2282 panic("vm_page_free: freeing busy page %p", m); 2283 2284 /* 2285 * Unqueue, then remove page. Note that we cannot destroy 2286 * the page here because we do not want to call the pager's 2287 * callback routine until after we've put the page on the 2288 * appropriate free queue. 2289 */ 2290 vm_page_remque(m); 2291 vm_page_remove(m); 2292 2293 /* 2294 * If fictitious remove object association and 2295 * return, otherwise delay object association removal. 2296 */ 2297 if ((m->flags & PG_FICTITIOUS) != 0) { 2298 return; 2299 } 2300 2301 m->valid = 0; 2302 vm_page_undirty(m); 2303 2304 if (m->wire_count != 0) 2305 panic("vm_page_free: freeing wired page %p", m); 2306 if (m->hold_count != 0) { 2307 m->flags &= ~PG_ZERO; 2308 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2309 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2310 m->flags |= PG_UNHOLDFREE; 2311 } else { 2312 /* 2313 * Restore the default memory attribute to the page. 2314 */ 2315 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2316 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2317 2318 /* 2319 * Insert the page into the physical memory allocator's 2320 * cache/free page queues. 2321 */ 2322 mtx_lock(&vm_page_queue_free_mtx); 2323 m->flags |= PG_FREE; 2324 vm_phys_freecnt_adj(m, 1); 2325#if VM_NRESERVLEVEL > 0 2326 if (!vm_reserv_free_page(m)) 2327#else 2328 if (TRUE) 2329#endif 2330 vm_phys_free_pages(m, 0); 2331 if ((m->flags & PG_ZERO) != 0) 2332 ++vm_page_zero_count; 2333 else 2334 vm_page_zero_idle_wakeup(); 2335 vm_page_free_wakeup(); 2336 mtx_unlock(&vm_page_queue_free_mtx); 2337 } 2338} 2339 2340/* 2341 * vm_page_wire: 2342 * 2343 * Mark this page as wired down by yet 2344 * another map, removing it from paging queues 2345 * as necessary. 2346 * 2347 * If the page is fictitious, then its wire count must remain one. 2348 * 2349 * The page must be locked. 2350 */ 2351void 2352vm_page_wire(vm_page_t m) 2353{ 2354 2355 /* 2356 * Only bump the wire statistics if the page is not already wired, 2357 * and only unqueue the page if it is on some queue (if it is unmanaged 2358 * it is already off the queues). 2359 */ 2360 vm_page_lock_assert(m, MA_OWNED); 2361 if ((m->flags & PG_FICTITIOUS) != 0) { 2362 KASSERT(m->wire_count == 1, 2363 ("vm_page_wire: fictitious page %p's wire count isn't one", 2364 m)); 2365 return; 2366 } 2367 if (m->wire_count == 0) { 2368 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2369 m->queue == PQ_NONE, 2370 ("vm_page_wire: unmanaged page %p is queued", m)); 2371 vm_page_remque(m); 2372 atomic_add_int(&cnt.v_wire_count, 1); 2373 } 2374 m->wire_count++; 2375 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2376} 2377 2378/* 2379 * vm_page_unwire: 2380 * 2381 * Release one wiring of the specified page, potentially enabling it to be 2382 * paged again. If paging is enabled, then the value of the parameter 2383 * "activate" determines to which queue the page is added. If "activate" is 2384 * non-zero, then the page is added to the active queue. Otherwise, it is 2385 * added to the inactive queue. 2386 * 2387 * However, unless the page belongs to an object, it is not enqueued because 2388 * it cannot be paged out. 2389 * 2390 * If a page is fictitious, then its wire count must always be one. 2391 * 2392 * A managed page must be locked. 2393 */ 2394void 2395vm_page_unwire(vm_page_t m, int activate) 2396{ 2397 2398 if ((m->oflags & VPO_UNMANAGED) == 0) 2399 vm_page_lock_assert(m, MA_OWNED); 2400 if ((m->flags & PG_FICTITIOUS) != 0) { 2401 KASSERT(m->wire_count == 1, 2402 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2403 return; 2404 } 2405 if (m->wire_count > 0) { 2406 m->wire_count--; 2407 if (m->wire_count == 0) { 2408 atomic_subtract_int(&cnt.v_wire_count, 1); 2409 if ((m->oflags & VPO_UNMANAGED) != 0 || 2410 m->object == NULL) 2411 return; 2412 if (!activate) 2413 m->flags &= ~PG_WINATCFLS; 2414 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2415 } 2416 } else 2417 panic("vm_page_unwire: page %p's wire count is zero", m); 2418} 2419 2420/* 2421 * Move the specified page to the inactive queue. 2422 * 2423 * Many pages placed on the inactive queue should actually go 2424 * into the cache, but it is difficult to figure out which. What 2425 * we do instead, if the inactive target is well met, is to put 2426 * clean pages at the head of the inactive queue instead of the tail. 2427 * This will cause them to be moved to the cache more quickly and 2428 * if not actively re-referenced, reclaimed more quickly. If we just 2429 * stick these pages at the end of the inactive queue, heavy filesystem 2430 * meta-data accesses can cause an unnecessary paging load on memory bound 2431 * processes. This optimization causes one-time-use metadata to be 2432 * reused more quickly. 2433 * 2434 * Normally athead is 0 resulting in LRU operation. athead is set 2435 * to 1 if we want this page to be 'as if it were placed in the cache', 2436 * except without unmapping it from the process address space. 2437 * 2438 * The page must be locked. 2439 */ 2440static inline void 2441_vm_page_deactivate(vm_page_t m, int athead) 2442{ 2443 struct vm_pagequeue *pq; 2444 int queue; 2445 2446 vm_page_assert_locked(m); 2447 2448 /* 2449 * Ignore if the page is already inactive, unless it is unlikely to be 2450 * reactivated. 2451 */ 2452 if ((queue = m->queue) == PQ_INACTIVE && !athead) 2453 return; 2454 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2455 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2456 /* Avoid multiple acquisitions of the inactive queue lock. */ 2457 if (queue == PQ_INACTIVE) { 2458 vm_pagequeue_lock(pq); 2459 vm_page_dequeue_locked(m); 2460 } else { 2461 if (queue != PQ_NONE) 2462 vm_page_dequeue(m); 2463 m->flags &= ~PG_WINATCFLS; 2464 vm_pagequeue_lock(pq); 2465 } 2466 m->queue = PQ_INACTIVE; 2467 if (athead) 2468 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q); 2469 else 2470 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2471 vm_pagequeue_cnt_inc(pq); 2472 vm_pagequeue_unlock(pq); 2473 } 2474} 2475 2476/* 2477 * Move the specified page to the inactive queue. 2478 * 2479 * The page must be locked. 2480 */ 2481void 2482vm_page_deactivate(vm_page_t m) 2483{ 2484 2485 _vm_page_deactivate(m, 0); 2486} 2487 2488/* 2489 * vm_page_try_to_cache: 2490 * 2491 * Returns 0 on failure, 1 on success 2492 */ 2493int 2494vm_page_try_to_cache(vm_page_t m) 2495{ 2496 2497 vm_page_lock_assert(m, MA_OWNED); 2498 VM_OBJECT_ASSERT_WLOCKED(m->object); 2499 if (m->dirty || m->hold_count || m->wire_count || 2500 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2501 return (0); 2502 pmap_remove_all(m); 2503 if (m->dirty) 2504 return (0); 2505 vm_page_cache(m); 2506 return (1); 2507} 2508 2509/* 2510 * vm_page_try_to_free() 2511 * 2512 * Attempt to free the page. If we cannot free it, we do nothing. 2513 * 1 is returned on success, 0 on failure. 2514 */ 2515int 2516vm_page_try_to_free(vm_page_t m) 2517{ 2518 2519 vm_page_lock_assert(m, MA_OWNED); 2520 if (m->object != NULL) 2521 VM_OBJECT_ASSERT_WLOCKED(m->object); 2522 if (m->dirty || m->hold_count || m->wire_count || 2523 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2524 return (0); 2525 pmap_remove_all(m); 2526 if (m->dirty) 2527 return (0); 2528 vm_page_free(m); 2529 return (1); 2530} 2531 2532/* 2533 * vm_page_cache 2534 * 2535 * Put the specified page onto the page cache queue (if appropriate). 2536 * 2537 * The object and page must be locked. 2538 */ 2539void 2540vm_page_cache(vm_page_t m) 2541{ 2542 vm_object_t object; 2543 boolean_t cache_was_empty; 2544 2545 vm_page_lock_assert(m, MA_OWNED); 2546 object = m->object; 2547 VM_OBJECT_ASSERT_WLOCKED(object); 2548 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) || 2549 m->hold_count || m->wire_count) 2550 panic("vm_page_cache: attempting to cache busy page"); 2551 KASSERT(!pmap_page_is_mapped(m), 2552 ("vm_page_cache: page %p is mapped", m)); 2553 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2554 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2555 (object->type == OBJT_SWAP && 2556 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2557 /* 2558 * Hypothesis: A cache-elgible page belonging to a 2559 * default object or swap object but without a backing 2560 * store must be zero filled. 2561 */ 2562 vm_page_free(m); 2563 return; 2564 } 2565 KASSERT((m->flags & PG_CACHED) == 0, 2566 ("vm_page_cache: page %p is already cached", m)); 2567 2568 /* 2569 * Remove the page from the paging queues. 2570 */ 2571 vm_page_remque(m); 2572 2573 /* 2574 * Remove the page from the object's collection of resident 2575 * pages. 2576 */ 2577 vm_radix_remove(&object->rtree, m->pindex); 2578 TAILQ_REMOVE(&object->memq, m, listq); 2579 object->resident_page_count--; 2580 2581 /* 2582 * Restore the default memory attribute to the page. 2583 */ 2584 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2585 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2586 2587 /* 2588 * Insert the page into the object's collection of cached pages 2589 * and the physical memory allocator's cache/free page queues. 2590 */ 2591 m->flags &= ~PG_ZERO; 2592 mtx_lock(&vm_page_queue_free_mtx); 2593 cache_was_empty = vm_radix_is_empty(&object->cache); 2594 if (vm_radix_insert(&object->cache, m)) { 2595 mtx_unlock(&vm_page_queue_free_mtx); 2596 if (object->type == OBJT_VNODE && 2597 object->resident_page_count == 0) 2598 vdrop(object->handle); 2599 m->object = NULL; 2600 vm_page_free(m); 2601 return; 2602 } 2603 2604 /* 2605 * The above call to vm_radix_insert() could reclaim the one pre- 2606 * existing cached page from this object, resulting in a call to 2607 * vdrop(). 2608 */ 2609 if (!cache_was_empty) 2610 cache_was_empty = vm_radix_is_singleton(&object->cache); 2611 2612 m->flags |= PG_CACHED; 2613 cnt.v_cache_count++; 2614 PCPU_INC(cnt.v_tcached); 2615#if VM_NRESERVLEVEL > 0 2616 if (!vm_reserv_free_page(m)) { 2617#else 2618 if (TRUE) { 2619#endif 2620 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2621 vm_phys_free_pages(m, 0); 2622 } 2623 vm_page_free_wakeup(); 2624 mtx_unlock(&vm_page_queue_free_mtx); 2625 2626 /* 2627 * Increment the vnode's hold count if this is the object's only 2628 * cached page. Decrement the vnode's hold count if this was 2629 * the object's only resident page. 2630 */ 2631 if (object->type == OBJT_VNODE) { 2632 if (cache_was_empty && object->resident_page_count != 0) 2633 vhold(object->handle); 2634 else if (!cache_was_empty && object->resident_page_count == 0) 2635 vdrop(object->handle); 2636 } 2637} 2638 2639/* 2640 * vm_page_advise 2641 * 2642 * Deactivate or do nothing, as appropriate. This routine is used 2643 * by madvise() and vop_stdadvise(). 2644 * 2645 * The object and page must be locked. 2646 */ 2647void 2648vm_page_advise(vm_page_t m, int advice) 2649{ 2650 2651 vm_page_assert_locked(m); 2652 VM_OBJECT_ASSERT_WLOCKED(m->object); 2653 if (advice == MADV_FREE) 2654 /* 2655 * Mark the page clean. This will allow the page to be freed 2656 * up by the system. However, such pages are often reused 2657 * quickly by malloc() so we do not do anything that would 2658 * cause a page fault if we can help it. 2659 * 2660 * Specifically, we do not try to actually free the page now 2661 * nor do we try to put it in the cache (which would cause a 2662 * page fault on reuse). 2663 * 2664 * But we do make the page as freeable as we can without 2665 * actually taking the step of unmapping it. 2666 */ 2667 vm_page_undirty(m); 2668 else if (advice != MADV_DONTNEED) 2669 return; 2670 2671 /* 2672 * Clear any references to the page. Otherwise, the page daemon will 2673 * immediately reactivate the page. 2674 */ 2675 vm_page_aflag_clear(m, PGA_REFERENCED); 2676 2677 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 2678 vm_page_dirty(m); 2679 2680 /* 2681 * Place clean pages at the head of the inactive queue rather than the 2682 * tail, thus defeating the queue's LRU operation and ensuring that the 2683 * page will be reused quickly. 2684 */ 2685 _vm_page_deactivate(m, m->dirty == 0); 2686} 2687 2688/* 2689 * Grab a page, waiting until we are waken up due to the page 2690 * changing state. We keep on waiting, if the page continues 2691 * to be in the object. If the page doesn't exist, first allocate it 2692 * and then conditionally zero it. 2693 * 2694 * This routine may sleep. 2695 * 2696 * The object must be locked on entry. The lock will, however, be released 2697 * and reacquired if the routine sleeps. 2698 */ 2699vm_page_t 2700vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2701{ 2702 vm_page_t m; 2703 int sleep; 2704 2705 VM_OBJECT_ASSERT_WLOCKED(object); 2706 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 2707 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 2708 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 2709retrylookup: 2710 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2711 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 2712 vm_page_xbusied(m) : vm_page_busied(m); 2713 if (sleep) { 2714 /* 2715 * Reference the page before unlocking and 2716 * sleeping so that the page daemon is less 2717 * likely to reclaim it. 2718 */ 2719 vm_page_aflag_set(m, PGA_REFERENCED); 2720 vm_page_lock(m); 2721 VM_OBJECT_WUNLOCK(object); 2722 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 2723 VM_ALLOC_IGN_SBUSY) != 0); 2724 VM_OBJECT_WLOCK(object); 2725 goto retrylookup; 2726 } else { 2727 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2728 vm_page_lock(m); 2729 vm_page_wire(m); 2730 vm_page_unlock(m); 2731 } 2732 if ((allocflags & 2733 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2734 vm_page_xbusy(m); 2735 if ((allocflags & VM_ALLOC_SBUSY) != 0) 2736 vm_page_sbusy(m); 2737 return (m); 2738 } 2739 } 2740 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY); 2741 if (m == NULL) { 2742 VM_OBJECT_WUNLOCK(object); 2743 VM_WAIT; 2744 VM_OBJECT_WLOCK(object); 2745 goto retrylookup; 2746 } else if (m->valid != 0) 2747 return (m); 2748 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2749 pmap_zero_page(m); 2750 return (m); 2751} 2752 2753/* 2754 * Mapping function for valid or dirty bits in a page. 2755 * 2756 * Inputs are required to range within a page. 2757 */ 2758vm_page_bits_t 2759vm_page_bits(int base, int size) 2760{ 2761 int first_bit; 2762 int last_bit; 2763 2764 KASSERT( 2765 base + size <= PAGE_SIZE, 2766 ("vm_page_bits: illegal base/size %d/%d", base, size) 2767 ); 2768 2769 if (size == 0) /* handle degenerate case */ 2770 return (0); 2771 2772 first_bit = base >> DEV_BSHIFT; 2773 last_bit = (base + size - 1) >> DEV_BSHIFT; 2774 2775 return (((vm_page_bits_t)2 << last_bit) - 2776 ((vm_page_bits_t)1 << first_bit)); 2777} 2778 2779/* 2780 * vm_page_set_valid_range: 2781 * 2782 * Sets portions of a page valid. The arguments are expected 2783 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2784 * of any partial chunks touched by the range. The invalid portion of 2785 * such chunks will be zeroed. 2786 * 2787 * (base + size) must be less then or equal to PAGE_SIZE. 2788 */ 2789void 2790vm_page_set_valid_range(vm_page_t m, int base, int size) 2791{ 2792 int endoff, frag; 2793 2794 VM_OBJECT_ASSERT_WLOCKED(m->object); 2795 if (size == 0) /* handle degenerate case */ 2796 return; 2797 2798 /* 2799 * If the base is not DEV_BSIZE aligned and the valid 2800 * bit is clear, we have to zero out a portion of the 2801 * first block. 2802 */ 2803 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2804 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2805 pmap_zero_page_area(m, frag, base - frag); 2806 2807 /* 2808 * If the ending offset is not DEV_BSIZE aligned and the 2809 * valid bit is clear, we have to zero out a portion of 2810 * the last block. 2811 */ 2812 endoff = base + size; 2813 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2814 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2815 pmap_zero_page_area(m, endoff, 2816 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2817 2818 /* 2819 * Assert that no previously invalid block that is now being validated 2820 * is already dirty. 2821 */ 2822 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2823 ("vm_page_set_valid_range: page %p is dirty", m)); 2824 2825 /* 2826 * Set valid bits inclusive of any overlap. 2827 */ 2828 m->valid |= vm_page_bits(base, size); 2829} 2830 2831/* 2832 * Clear the given bits from the specified page's dirty field. 2833 */ 2834static __inline void 2835vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2836{ 2837 uintptr_t addr; 2838#if PAGE_SIZE < 16384 2839 int shift; 2840#endif 2841 2842 /* 2843 * If the object is locked and the page is neither exclusive busy nor 2844 * write mapped, then the page's dirty field cannot possibly be 2845 * set by a concurrent pmap operation. 2846 */ 2847 VM_OBJECT_ASSERT_WLOCKED(m->object); 2848 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 2849 m->dirty &= ~pagebits; 2850 else { 2851 /* 2852 * The pmap layer can call vm_page_dirty() without 2853 * holding a distinguished lock. The combination of 2854 * the object's lock and an atomic operation suffice 2855 * to guarantee consistency of the page dirty field. 2856 * 2857 * For PAGE_SIZE == 32768 case, compiler already 2858 * properly aligns the dirty field, so no forcible 2859 * alignment is needed. Only require existence of 2860 * atomic_clear_64 when page size is 32768. 2861 */ 2862 addr = (uintptr_t)&m->dirty; 2863#if PAGE_SIZE == 32768 2864 atomic_clear_64((uint64_t *)addr, pagebits); 2865#elif PAGE_SIZE == 16384 2866 atomic_clear_32((uint32_t *)addr, pagebits); 2867#else /* PAGE_SIZE <= 8192 */ 2868 /* 2869 * Use a trick to perform a 32-bit atomic on the 2870 * containing aligned word, to not depend on the existence 2871 * of atomic_clear_{8, 16}. 2872 */ 2873 shift = addr & (sizeof(uint32_t) - 1); 2874#if BYTE_ORDER == BIG_ENDIAN 2875 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2876#else 2877 shift *= NBBY; 2878#endif 2879 addr &= ~(sizeof(uint32_t) - 1); 2880 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2881#endif /* PAGE_SIZE */ 2882 } 2883} 2884 2885/* 2886 * vm_page_set_validclean: 2887 * 2888 * Sets portions of a page valid and clean. The arguments are expected 2889 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2890 * of any partial chunks touched by the range. The invalid portion of 2891 * such chunks will be zero'd. 2892 * 2893 * (base + size) must be less then or equal to PAGE_SIZE. 2894 */ 2895void 2896vm_page_set_validclean(vm_page_t m, int base, int size) 2897{ 2898 vm_page_bits_t oldvalid, pagebits; 2899 int endoff, frag; 2900 2901 VM_OBJECT_ASSERT_WLOCKED(m->object); 2902 if (size == 0) /* handle degenerate case */ 2903 return; 2904 2905 /* 2906 * If the base is not DEV_BSIZE aligned and the valid 2907 * bit is clear, we have to zero out a portion of the 2908 * first block. 2909 */ 2910 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2911 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2912 pmap_zero_page_area(m, frag, base - frag); 2913 2914 /* 2915 * If the ending offset is not DEV_BSIZE aligned and the 2916 * valid bit is clear, we have to zero out a portion of 2917 * the last block. 2918 */ 2919 endoff = base + size; 2920 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2921 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2922 pmap_zero_page_area(m, endoff, 2923 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2924 2925 /* 2926 * Set valid, clear dirty bits. If validating the entire 2927 * page we can safely clear the pmap modify bit. We also 2928 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2929 * takes a write fault on a MAP_NOSYNC memory area the flag will 2930 * be set again. 2931 * 2932 * We set valid bits inclusive of any overlap, but we can only 2933 * clear dirty bits for DEV_BSIZE chunks that are fully within 2934 * the range. 2935 */ 2936 oldvalid = m->valid; 2937 pagebits = vm_page_bits(base, size); 2938 m->valid |= pagebits; 2939#if 0 /* NOT YET */ 2940 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2941 frag = DEV_BSIZE - frag; 2942 base += frag; 2943 size -= frag; 2944 if (size < 0) 2945 size = 0; 2946 } 2947 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2948#endif 2949 if (base == 0 && size == PAGE_SIZE) { 2950 /* 2951 * The page can only be modified within the pmap if it is 2952 * mapped, and it can only be mapped if it was previously 2953 * fully valid. 2954 */ 2955 if (oldvalid == VM_PAGE_BITS_ALL) 2956 /* 2957 * Perform the pmap_clear_modify() first. Otherwise, 2958 * a concurrent pmap operation, such as 2959 * pmap_protect(), could clear a modification in the 2960 * pmap and set the dirty field on the page before 2961 * pmap_clear_modify() had begun and after the dirty 2962 * field was cleared here. 2963 */ 2964 pmap_clear_modify(m); 2965 m->dirty = 0; 2966 m->oflags &= ~VPO_NOSYNC; 2967 } else if (oldvalid != VM_PAGE_BITS_ALL) 2968 m->dirty &= ~pagebits; 2969 else 2970 vm_page_clear_dirty_mask(m, pagebits); 2971} 2972 2973void 2974vm_page_clear_dirty(vm_page_t m, int base, int size) 2975{ 2976 2977 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2978} 2979 2980/* 2981 * vm_page_set_invalid: 2982 * 2983 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2984 * valid and dirty bits for the effected areas are cleared. 2985 */ 2986void 2987vm_page_set_invalid(vm_page_t m, int base, int size) 2988{ 2989 vm_page_bits_t bits; 2990 vm_object_t object; 2991 2992 object = m->object; 2993 VM_OBJECT_ASSERT_WLOCKED(object); 2994 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 2995 size >= object->un_pager.vnp.vnp_size) 2996 bits = VM_PAGE_BITS_ALL; 2997 else 2998 bits = vm_page_bits(base, size); 2999 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3000 bits != 0) 3001 pmap_remove_all(m); 3002 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3003 !pmap_page_is_mapped(m), 3004 ("vm_page_set_invalid: page %p is mapped", m)); 3005 m->valid &= ~bits; 3006 m->dirty &= ~bits; 3007} 3008 3009/* 3010 * vm_page_zero_invalid() 3011 * 3012 * The kernel assumes that the invalid portions of a page contain 3013 * garbage, but such pages can be mapped into memory by user code. 3014 * When this occurs, we must zero out the non-valid portions of the 3015 * page so user code sees what it expects. 3016 * 3017 * Pages are most often semi-valid when the end of a file is mapped 3018 * into memory and the file's size is not page aligned. 3019 */ 3020void 3021vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3022{ 3023 int b; 3024 int i; 3025 3026 VM_OBJECT_ASSERT_WLOCKED(m->object); 3027 /* 3028 * Scan the valid bits looking for invalid sections that 3029 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3030 * valid bit may be set ) have already been zeroed by 3031 * vm_page_set_validclean(). 3032 */ 3033 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3034 if (i == (PAGE_SIZE / DEV_BSIZE) || 3035 (m->valid & ((vm_page_bits_t)1 << i))) { 3036 if (i > b) { 3037 pmap_zero_page_area(m, 3038 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3039 } 3040 b = i + 1; 3041 } 3042 } 3043 3044 /* 3045 * setvalid is TRUE when we can safely set the zero'd areas 3046 * as being valid. We can do this if there are no cache consistancy 3047 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3048 */ 3049 if (setvalid) 3050 m->valid = VM_PAGE_BITS_ALL; 3051} 3052 3053/* 3054 * vm_page_is_valid: 3055 * 3056 * Is (partial) page valid? Note that the case where size == 0 3057 * will return FALSE in the degenerate case where the page is 3058 * entirely invalid, and TRUE otherwise. 3059 */ 3060int 3061vm_page_is_valid(vm_page_t m, int base, int size) 3062{ 3063 vm_page_bits_t bits; 3064 3065 VM_OBJECT_ASSERT_LOCKED(m->object); 3066 bits = vm_page_bits(base, size); 3067 return (m->valid != 0 && (m->valid & bits) == bits); 3068} 3069 3070/* 3071 * vm_page_ps_is_valid: 3072 * 3073 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3074 */ 3075boolean_t 3076vm_page_ps_is_valid(vm_page_t m) 3077{ 3078 int i, npages; 3079 3080 VM_OBJECT_ASSERT_LOCKED(m->object); 3081 npages = atop(pagesizes[m->psind]); 3082 3083 /* 3084 * The physically contiguous pages that make up a superpage, i.e., a 3085 * page with a page size index ("psind") greater than zero, will 3086 * occupy adjacent entries in vm_page_array[]. 3087 */ 3088 for (i = 0; i < npages; i++) { 3089 if (m[i].valid != VM_PAGE_BITS_ALL) 3090 return (FALSE); 3091 } 3092 return (TRUE); 3093} 3094 3095/* 3096 * Set the page's dirty bits if the page is modified. 3097 */ 3098void 3099vm_page_test_dirty(vm_page_t m) 3100{ 3101 3102 VM_OBJECT_ASSERT_WLOCKED(m->object); 3103 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3104 vm_page_dirty(m); 3105} 3106 3107void 3108vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3109{ 3110 3111 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3112} 3113 3114void 3115vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3116{ 3117 3118 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3119} 3120 3121int 3122vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3123{ 3124 3125 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3126} 3127 3128#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3129void 3130vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3131{ 3132 3133 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3134} 3135 3136void 3137vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3138{ 3139 3140 mtx_assert_(vm_page_lockptr(m), a, file, line); 3141} 3142#endif 3143 3144#ifdef INVARIANTS 3145void 3146vm_page_object_lock_assert(vm_page_t m) 3147{ 3148 3149 /* 3150 * Certain of the page's fields may only be modified by the 3151 * holder of the containing object's lock or the exclusive busy. 3152 * holder. Unfortunately, the holder of the write busy is 3153 * not recorded, and thus cannot be checked here. 3154 */ 3155 if (m->object != NULL && !vm_page_xbusied(m)) 3156 VM_OBJECT_ASSERT_WLOCKED(m->object); 3157} 3158 3159void 3160vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3161{ 3162 3163 if ((bits & PGA_WRITEABLE) == 0) 3164 return; 3165 3166 /* 3167 * The PGA_WRITEABLE flag can only be set if the page is 3168 * managed, is exclusively busied or the object is locked. 3169 * Currently, this flag is only set by pmap_enter(). 3170 */ 3171 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3172 ("PGA_WRITEABLE on unmanaged page")); 3173 if (!vm_page_xbusied(m)) 3174 VM_OBJECT_ASSERT_LOCKED(m->object); 3175} 3176#endif 3177 3178#include "opt_ddb.h" 3179#ifdef DDB 3180#include <sys/kernel.h> 3181 3182#include <ddb/ddb.h> 3183 3184DB_SHOW_COMMAND(page, vm_page_print_page_info) 3185{ 3186 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3187 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3188 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3189 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3190 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3191 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3192 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3193 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3194 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3195 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3196} 3197 3198DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3199{ 3200 int dom; 3201 3202 db_printf("pq_free %d pq_cache %d\n", 3203 cnt.v_free_count, cnt.v_cache_count); 3204 for (dom = 0; dom < vm_ndomains; dom++) { 3205 db_printf( 3206 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n", 3207 dom, 3208 vm_dom[dom].vmd_page_count, 3209 vm_dom[dom].vmd_free_count, 3210 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3211 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3212 vm_dom[dom].vmd_pass); 3213 } 3214} 3215 3216DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3217{ 3218 vm_page_t m; 3219 boolean_t phys; 3220 3221 if (!have_addr) { 3222 db_printf("show pginfo addr\n"); 3223 return; 3224 } 3225 3226 phys = strchr(modif, 'p') != NULL; 3227 if (phys) 3228 m = PHYS_TO_VM_PAGE(addr); 3229 else 3230 m = (vm_page_t)addr; 3231 db_printf( 3232 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3233 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3234 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3235 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3236 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3237} 3238#endif /* DDB */ 3239