vm_page.c revision 304092
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 304092 2016-08-14 19:01:15Z markj $"); 86 87#include "opt_vm.h" 88 89#include <sys/param.h> 90#include <sys/systm.h> 91#include <sys/lock.h> 92#include <sys/kernel.h> 93#include <sys/limits.h> 94#include <sys/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; 518 519 vm_page_assert_xbusied(m); 520 521 for (;;) { 522 x = m->busy_lock; 523 x &= VPB_BIT_WAITERS; 524 if (atomic_cmpset_rel_int(&m->busy_lock, 525 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x)) 526 break; 527 } 528} 529 530/* 531 * vm_page_sbusied: 532 * 533 * Return a positive value if the page is shared busied, 0 otherwise. 534 */ 535int 536vm_page_sbusied(vm_page_t m) 537{ 538 u_int x; 539 540 x = m->busy_lock; 541 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 542} 543 544/* 545 * vm_page_sunbusy: 546 * 547 * Shared unbusy a page. 548 */ 549void 550vm_page_sunbusy(vm_page_t m) 551{ 552 u_int x; 553 554 vm_page_assert_sbusied(m); 555 556 for (;;) { 557 x = m->busy_lock; 558 if (VPB_SHARERS(x) > 1) { 559 if (atomic_cmpset_int(&m->busy_lock, x, 560 x - VPB_ONE_SHARER)) 561 break; 562 continue; 563 } 564 if ((x & VPB_BIT_WAITERS) == 0) { 565 KASSERT(x == VPB_SHARERS_WORD(1), 566 ("vm_page_sunbusy: invalid lock state")); 567 if (atomic_cmpset_int(&m->busy_lock, 568 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 569 break; 570 continue; 571 } 572 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 573 ("vm_page_sunbusy: invalid lock state for waiters")); 574 575 vm_page_lock(m); 576 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 577 vm_page_unlock(m); 578 continue; 579 } 580 wakeup(m); 581 vm_page_unlock(m); 582 break; 583 } 584} 585 586/* 587 * vm_page_busy_sleep: 588 * 589 * Sleep and release the page lock, using the page pointer as wchan. 590 * This is used to implement the hard-path of busying mechanism. 591 * 592 * The given page must be locked. 593 */ 594void 595vm_page_busy_sleep(vm_page_t m, const char *wmesg) 596{ 597 u_int x; 598 599 vm_page_lock_assert(m, MA_OWNED); 600 601 x = m->busy_lock; 602 if (x == VPB_UNBUSIED) { 603 vm_page_unlock(m); 604 return; 605 } 606 if ((x & VPB_BIT_WAITERS) == 0 && 607 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) { 608 vm_page_unlock(m); 609 return; 610 } 611 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 612} 613 614/* 615 * vm_page_trysbusy: 616 * 617 * Try to shared busy a page. 618 * If the operation succeeds 1 is returned otherwise 0. 619 * The operation never sleeps. 620 */ 621int 622vm_page_trysbusy(vm_page_t m) 623{ 624 u_int x; 625 626 for (;;) { 627 x = m->busy_lock; 628 if ((x & VPB_BIT_SHARED) == 0) 629 return (0); 630 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 631 return (1); 632 } 633} 634 635/* 636 * vm_page_xunbusy_hard: 637 * 638 * Called after the first try the exclusive unbusy of a page failed. 639 * It is assumed that the waiters bit is on. 640 */ 641void 642vm_page_xunbusy_hard(vm_page_t m) 643{ 644 645 vm_page_assert_xbusied(m); 646 647 vm_page_lock(m); 648 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 649 wakeup(m); 650 vm_page_unlock(m); 651} 652 653/* 654 * vm_page_flash: 655 * 656 * Wakeup anyone waiting for the page. 657 * The ownership bits do not change. 658 * 659 * The given page must be locked. 660 */ 661void 662vm_page_flash(vm_page_t m) 663{ 664 u_int x; 665 666 vm_page_lock_assert(m, MA_OWNED); 667 668 for (;;) { 669 x = m->busy_lock; 670 if ((x & VPB_BIT_WAITERS) == 0) 671 return; 672 if (atomic_cmpset_int(&m->busy_lock, x, 673 x & (~VPB_BIT_WAITERS))) 674 break; 675 } 676 wakeup(m); 677} 678 679/* 680 * Keep page from being freed by the page daemon 681 * much of the same effect as wiring, except much lower 682 * overhead and should be used only for *very* temporary 683 * holding ("wiring"). 684 */ 685void 686vm_page_hold(vm_page_t mem) 687{ 688 689 vm_page_lock_assert(mem, MA_OWNED); 690 mem->hold_count++; 691} 692 693void 694vm_page_unhold(vm_page_t mem) 695{ 696 697 vm_page_lock_assert(mem, MA_OWNED); 698 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 699 --mem->hold_count; 700 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 701 vm_page_free_toq(mem); 702} 703 704/* 705 * vm_page_unhold_pages: 706 * 707 * Unhold each of the pages that is referenced by the given array. 708 */ 709void 710vm_page_unhold_pages(vm_page_t *ma, int count) 711{ 712 struct mtx *mtx, *new_mtx; 713 714 mtx = NULL; 715 for (; count != 0; count--) { 716 /* 717 * Avoid releasing and reacquiring the same page lock. 718 */ 719 new_mtx = vm_page_lockptr(*ma); 720 if (mtx != new_mtx) { 721 if (mtx != NULL) 722 mtx_unlock(mtx); 723 mtx = new_mtx; 724 mtx_lock(mtx); 725 } 726 vm_page_unhold(*ma); 727 ma++; 728 } 729 if (mtx != NULL) 730 mtx_unlock(mtx); 731} 732 733vm_page_t 734PHYS_TO_VM_PAGE(vm_paddr_t pa) 735{ 736 vm_page_t m; 737 738#ifdef VM_PHYSSEG_SPARSE 739 m = vm_phys_paddr_to_vm_page(pa); 740 if (m == NULL) 741 m = vm_phys_fictitious_to_vm_page(pa); 742 return (m); 743#elif defined(VM_PHYSSEG_DENSE) 744 long pi; 745 746 pi = atop(pa); 747 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 748 m = &vm_page_array[pi - first_page]; 749 return (m); 750 } 751 return (vm_phys_fictitious_to_vm_page(pa)); 752#else 753#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 754#endif 755} 756 757/* 758 * vm_page_getfake: 759 * 760 * Create a fictitious page with the specified physical address and 761 * memory attribute. The memory attribute is the only the machine- 762 * dependent aspect of a fictitious page that must be initialized. 763 */ 764vm_page_t 765vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 766{ 767 vm_page_t m; 768 769 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 770 vm_page_initfake(m, paddr, memattr); 771 return (m); 772} 773 774void 775vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 776{ 777 778 if ((m->flags & PG_FICTITIOUS) != 0) { 779 /* 780 * The page's memattr might have changed since the 781 * previous initialization. Update the pmap to the 782 * new memattr. 783 */ 784 goto memattr; 785 } 786 m->phys_addr = paddr; 787 m->queue = PQ_NONE; 788 /* Fictitious pages don't use "segind". */ 789 m->flags = PG_FICTITIOUS; 790 /* Fictitious pages don't use "order" or "pool". */ 791 m->oflags = VPO_UNMANAGED; 792 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 793 m->wire_count = 1; 794 pmap_page_init(m); 795memattr: 796 pmap_page_set_memattr(m, memattr); 797} 798 799/* 800 * vm_page_putfake: 801 * 802 * Release a fictitious page. 803 */ 804void 805vm_page_putfake(vm_page_t m) 806{ 807 808 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 809 KASSERT((m->flags & PG_FICTITIOUS) != 0, 810 ("vm_page_putfake: bad page %p", m)); 811 uma_zfree(fakepg_zone, m); 812} 813 814/* 815 * vm_page_updatefake: 816 * 817 * Update the given fictitious page to the specified physical address and 818 * memory attribute. 819 */ 820void 821vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 822{ 823 824 KASSERT((m->flags & PG_FICTITIOUS) != 0, 825 ("vm_page_updatefake: bad page %p", m)); 826 m->phys_addr = paddr; 827 pmap_page_set_memattr(m, memattr); 828} 829 830/* 831 * vm_page_free: 832 * 833 * Free a page. 834 */ 835void 836vm_page_free(vm_page_t m) 837{ 838 839 m->flags &= ~PG_ZERO; 840 vm_page_free_toq(m); 841} 842 843/* 844 * vm_page_free_zero: 845 * 846 * Free a page to the zerod-pages queue 847 */ 848void 849vm_page_free_zero(vm_page_t m) 850{ 851 852 m->flags |= PG_ZERO; 853 vm_page_free_toq(m); 854} 855 856/* 857 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 858 * array which is not the request page. 859 */ 860void 861vm_page_readahead_finish(vm_page_t m) 862{ 863 864 if (m->valid != 0) { 865 /* 866 * Since the page is not the requested page, whether 867 * it should be activated or deactivated is not 868 * obvious. Empirical results have shown that 869 * deactivating the page is usually the best choice, 870 * unless the page is wanted by another thread. 871 */ 872 vm_page_lock(m); 873 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 874 vm_page_activate(m); 875 else 876 vm_page_deactivate(m); 877 vm_page_unlock(m); 878 vm_page_xunbusy(m); 879 } else { 880 /* 881 * Free the completely invalid page. Such page state 882 * occurs due to the short read operation which did 883 * not covered our page at all, or in case when a read 884 * error happens. 885 */ 886 vm_page_lock(m); 887 vm_page_free(m); 888 vm_page_unlock(m); 889 } 890} 891 892/* 893 * vm_page_sleep_if_busy: 894 * 895 * Sleep and release the page queues lock if the page is busied. 896 * Returns TRUE if the thread slept. 897 * 898 * The given page must be unlocked and object containing it must 899 * be locked. 900 */ 901int 902vm_page_sleep_if_busy(vm_page_t m, const char *msg) 903{ 904 vm_object_t obj; 905 906 vm_page_lock_assert(m, MA_NOTOWNED); 907 VM_OBJECT_ASSERT_WLOCKED(m->object); 908 909 if (vm_page_busied(m)) { 910 /* 911 * The page-specific object must be cached because page 912 * identity can change during the sleep, causing the 913 * re-lock of a different object. 914 * It is assumed that a reference to the object is already 915 * held by the callers. 916 */ 917 obj = m->object; 918 vm_page_lock(m); 919 VM_OBJECT_WUNLOCK(obj); 920 vm_page_busy_sleep(m, msg); 921 VM_OBJECT_WLOCK(obj); 922 return (TRUE); 923 } 924 return (FALSE); 925} 926 927/* 928 * vm_page_dirty_KBI: [ internal use only ] 929 * 930 * Set all bits in the page's dirty field. 931 * 932 * The object containing the specified page must be locked if the 933 * call is made from the machine-independent layer. 934 * 935 * See vm_page_clear_dirty_mask(). 936 * 937 * This function should only be called by vm_page_dirty(). 938 */ 939void 940vm_page_dirty_KBI(vm_page_t m) 941{ 942 943 /* These assertions refer to this operation by its public name. */ 944 KASSERT((m->flags & PG_CACHED) == 0, 945 ("vm_page_dirty: page in cache!")); 946 KASSERT(!VM_PAGE_IS_FREE(m), 947 ("vm_page_dirty: page is free!")); 948 KASSERT(m->valid == VM_PAGE_BITS_ALL, 949 ("vm_page_dirty: page is invalid!")); 950 m->dirty = VM_PAGE_BITS_ALL; 951} 952 953/* 954 * vm_page_insert: [ internal use only ] 955 * 956 * Inserts the given mem entry into the object and object list. 957 * 958 * The object must be locked. 959 */ 960int 961vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 962{ 963 vm_page_t mpred; 964 965 VM_OBJECT_ASSERT_WLOCKED(object); 966 mpred = vm_radix_lookup_le(&object->rtree, pindex); 967 return (vm_page_insert_after(m, object, pindex, mpred)); 968} 969 970/* 971 * vm_page_insert_after: 972 * 973 * Inserts the page "m" into the specified object at offset "pindex". 974 * 975 * The page "mpred" must immediately precede the offset "pindex" within 976 * the specified object. 977 * 978 * The object must be locked. 979 */ 980static int 981vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 982 vm_page_t mpred) 983{ 984 vm_page_t msucc; 985 986 VM_OBJECT_ASSERT_WLOCKED(object); 987 KASSERT(m->object == NULL, 988 ("vm_page_insert_after: page already inserted")); 989 if (mpred != NULL) { 990 KASSERT(mpred->object == object, 991 ("vm_page_insert_after: object doesn't contain mpred")); 992 KASSERT(mpred->pindex < pindex, 993 ("vm_page_insert_after: mpred doesn't precede pindex")); 994 msucc = TAILQ_NEXT(mpred, listq); 995 } else 996 msucc = TAILQ_FIRST(&object->memq); 997 if (msucc != NULL) 998 KASSERT(msucc->pindex > pindex, 999 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1000 1001 /* 1002 * Record the object/offset pair in this page 1003 */ 1004 m->object = object; 1005 m->pindex = pindex; 1006 1007 /* 1008 * Now link into the object's ordered list of backed pages. 1009 */ 1010 if (vm_radix_insert(&object->rtree, m)) { 1011 m->object = NULL; 1012 m->pindex = 0; 1013 return (1); 1014 } 1015 vm_page_insert_radixdone(m, object, mpred); 1016 return (0); 1017} 1018 1019/* 1020 * vm_page_insert_radixdone: 1021 * 1022 * Complete page "m" insertion into the specified object after the 1023 * radix trie hooking. 1024 * 1025 * The page "mpred" must precede the offset "m->pindex" within the 1026 * specified object. 1027 * 1028 * The object must be locked. 1029 */ 1030static void 1031vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1032{ 1033 1034 VM_OBJECT_ASSERT_WLOCKED(object); 1035 KASSERT(object != NULL && m->object == object, 1036 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1037 if (mpred != NULL) { 1038 KASSERT(mpred->object == object, 1039 ("vm_page_insert_after: object doesn't contain mpred")); 1040 KASSERT(mpred->pindex < m->pindex, 1041 ("vm_page_insert_after: mpred doesn't precede pindex")); 1042 } 1043 1044 if (mpred != NULL) 1045 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1046 else 1047 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1048 1049 /* 1050 * Show that the object has one more resident page. 1051 */ 1052 object->resident_page_count++; 1053 1054 /* 1055 * Hold the vnode until the last page is released. 1056 */ 1057 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1058 vhold(object->handle); 1059 1060 /* 1061 * Since we are inserting a new and possibly dirty page, 1062 * update the object's OBJ_MIGHTBEDIRTY flag. 1063 */ 1064 if (pmap_page_is_write_mapped(m)) 1065 vm_object_set_writeable_dirty(object); 1066} 1067 1068/* 1069 * vm_page_remove: 1070 * 1071 * Removes the given mem entry from the object/offset-page 1072 * table and the object page list, but do not invalidate/terminate 1073 * the backing store. 1074 * 1075 * The object must be locked. The page must be locked if it is managed. 1076 */ 1077void 1078vm_page_remove(vm_page_t m) 1079{ 1080 vm_object_t object; 1081 boolean_t lockacq; 1082 1083 if ((m->oflags & VPO_UNMANAGED) == 0) 1084 vm_page_lock_assert(m, MA_OWNED); 1085 if ((object = m->object) == NULL) 1086 return; 1087 VM_OBJECT_ASSERT_WLOCKED(object); 1088 if (vm_page_xbusied(m)) { 1089 lockacq = FALSE; 1090 if ((m->oflags & VPO_UNMANAGED) != 0 && 1091 !mtx_owned(vm_page_lockptr(m))) { 1092 lockacq = TRUE; 1093 vm_page_lock(m); 1094 } 1095 vm_page_flash(m); 1096 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1097 if (lockacq) 1098 vm_page_unlock(m); 1099 } 1100 1101 /* 1102 * Now remove from the object's list of backed pages. 1103 */ 1104 vm_radix_remove(&object->rtree, m->pindex); 1105 TAILQ_REMOVE(&object->memq, m, listq); 1106 1107 /* 1108 * And show that the object has one fewer resident page. 1109 */ 1110 object->resident_page_count--; 1111 1112 /* 1113 * The vnode may now be recycled. 1114 */ 1115 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1116 vdrop(object->handle); 1117 1118 m->object = NULL; 1119} 1120 1121/* 1122 * vm_page_lookup: 1123 * 1124 * Returns the page associated with the object/offset 1125 * pair specified; if none is found, NULL is returned. 1126 * 1127 * The object must be locked. 1128 */ 1129vm_page_t 1130vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1131{ 1132 1133 VM_OBJECT_ASSERT_LOCKED(object); 1134 return (vm_radix_lookup(&object->rtree, pindex)); 1135} 1136 1137/* 1138 * vm_page_find_least: 1139 * 1140 * Returns the page associated with the object with least pindex 1141 * greater than or equal to the parameter pindex, or NULL. 1142 * 1143 * The object must be locked. 1144 */ 1145vm_page_t 1146vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1147{ 1148 vm_page_t m; 1149 1150 VM_OBJECT_ASSERT_LOCKED(object); 1151 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1152 m = vm_radix_lookup_ge(&object->rtree, pindex); 1153 return (m); 1154} 1155 1156/* 1157 * Returns the given page's successor (by pindex) within the object if it is 1158 * resident; if none is found, NULL is returned. 1159 * 1160 * The object must be locked. 1161 */ 1162vm_page_t 1163vm_page_next(vm_page_t m) 1164{ 1165 vm_page_t next; 1166 1167 VM_OBJECT_ASSERT_WLOCKED(m->object); 1168 if ((next = TAILQ_NEXT(m, listq)) != NULL && 1169 next->pindex != m->pindex + 1) 1170 next = NULL; 1171 return (next); 1172} 1173 1174/* 1175 * Returns the given page's predecessor (by pindex) within the object if it is 1176 * resident; if none is found, NULL is returned. 1177 * 1178 * The object must be locked. 1179 */ 1180vm_page_t 1181vm_page_prev(vm_page_t m) 1182{ 1183 vm_page_t prev; 1184 1185 VM_OBJECT_ASSERT_WLOCKED(m->object); 1186 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1187 prev->pindex != m->pindex - 1) 1188 prev = NULL; 1189 return (prev); 1190} 1191 1192/* 1193 * Uses the page mnew as a replacement for an existing page at index 1194 * pindex which must be already present in the object. 1195 * 1196 * The existing page must not be on a paging queue. 1197 */ 1198vm_page_t 1199vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1200{ 1201 vm_page_t mold, mpred; 1202 1203 VM_OBJECT_ASSERT_WLOCKED(object); 1204 1205 /* 1206 * This function mostly follows vm_page_insert() and 1207 * vm_page_remove() without the radix, object count and vnode 1208 * dance. Double check such functions for more comments. 1209 */ 1210 mpred = vm_radix_lookup(&object->rtree, pindex); 1211 KASSERT(mpred != NULL, 1212 ("vm_page_replace: replacing page not present with pindex")); 1213 mpred = TAILQ_PREV(mpred, respgs, listq); 1214 if (mpred != NULL) 1215 KASSERT(mpred->pindex < pindex, 1216 ("vm_page_insert_after: mpred doesn't precede pindex")); 1217 1218 mnew->object = object; 1219 mnew->pindex = pindex; 1220 mold = vm_radix_replace(&object->rtree, mnew); 1221 KASSERT(mold->queue == PQ_NONE, 1222 ("vm_page_replace: mold is on a paging queue")); 1223 1224 /* Detach the old page from the resident tailq. */ 1225 TAILQ_REMOVE(&object->memq, mold, listq); 1226 1227 mold->object = NULL; 1228 vm_page_xunbusy(mold); 1229 1230 /* Insert the new page in the resident tailq. */ 1231 if (mpred != NULL) 1232 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq); 1233 else 1234 TAILQ_INSERT_HEAD(&object->memq, mnew, listq); 1235 if (pmap_page_is_write_mapped(mnew)) 1236 vm_object_set_writeable_dirty(object); 1237 return (mold); 1238} 1239 1240/* 1241 * vm_page_rename: 1242 * 1243 * Move the given memory entry from its 1244 * current object to the specified target object/offset. 1245 * 1246 * Note: swap associated with the page must be invalidated by the move. We 1247 * have to do this for several reasons: (1) we aren't freeing the 1248 * page, (2) we are dirtying the page, (3) the VM system is probably 1249 * moving the page from object A to B, and will then later move 1250 * the backing store from A to B and we can't have a conflict. 1251 * 1252 * Note: we *always* dirty the page. It is necessary both for the 1253 * fact that we moved it, and because we may be invalidating 1254 * swap. If the page is on the cache, we have to deactivate it 1255 * or vm_page_dirty() will panic. Dirty pages are not allowed 1256 * on the cache. 1257 * 1258 * The objects must be locked. 1259 */ 1260int 1261vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1262{ 1263 vm_page_t mpred; 1264 vm_pindex_t opidx; 1265 1266 VM_OBJECT_ASSERT_WLOCKED(new_object); 1267 1268 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1269 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1270 ("vm_page_rename: pindex already renamed")); 1271 1272 /* 1273 * Create a custom version of vm_page_insert() which does not depend 1274 * by m_prev and can cheat on the implementation aspects of the 1275 * function. 1276 */ 1277 opidx = m->pindex; 1278 m->pindex = new_pindex; 1279 if (vm_radix_insert(&new_object->rtree, m)) { 1280 m->pindex = opidx; 1281 return (1); 1282 } 1283 1284 /* 1285 * The operation cannot fail anymore. The removal must happen before 1286 * the listq iterator is tainted. 1287 */ 1288 m->pindex = opidx; 1289 vm_page_lock(m); 1290 vm_page_remove(m); 1291 1292 /* Return back to the new pindex to complete vm_page_insert(). */ 1293 m->pindex = new_pindex; 1294 m->object = new_object; 1295 vm_page_unlock(m); 1296 vm_page_insert_radixdone(m, new_object, mpred); 1297 vm_page_dirty(m); 1298 return (0); 1299} 1300 1301/* 1302 * Convert all of the given object's cached pages that have a 1303 * pindex within the given range into free pages. If the value 1304 * zero is given for "end", then the range's upper bound is 1305 * infinity. If the given object is backed by a vnode and it 1306 * transitions from having one or more cached pages to none, the 1307 * vnode's hold count is reduced. 1308 */ 1309void 1310vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1311{ 1312 vm_page_t m; 1313 boolean_t empty; 1314 1315 mtx_lock(&vm_page_queue_free_mtx); 1316 if (__predict_false(vm_radix_is_empty(&object->cache))) { 1317 mtx_unlock(&vm_page_queue_free_mtx); 1318 return; 1319 } 1320 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) { 1321 if (end != 0 && m->pindex >= end) 1322 break; 1323 vm_radix_remove(&object->cache, m->pindex); 1324 vm_page_cache_turn_free(m); 1325 } 1326 empty = vm_radix_is_empty(&object->cache); 1327 mtx_unlock(&vm_page_queue_free_mtx); 1328 if (object->type == OBJT_VNODE && empty) 1329 vdrop(object->handle); 1330} 1331 1332/* 1333 * Returns the cached page that is associated with the given 1334 * object and offset. If, however, none exists, returns NULL. 1335 * 1336 * The free page queue must be locked. 1337 */ 1338static inline vm_page_t 1339vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1340{ 1341 1342 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1343 return (vm_radix_lookup(&object->cache, pindex)); 1344} 1345 1346/* 1347 * Remove the given cached page from its containing object's 1348 * collection of cached pages. 1349 * 1350 * The free page queue must be locked. 1351 */ 1352static void 1353vm_page_cache_remove(vm_page_t m) 1354{ 1355 1356 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1357 KASSERT((m->flags & PG_CACHED) != 0, 1358 ("vm_page_cache_remove: page %p is not cached", m)); 1359 vm_radix_remove(&m->object->cache, m->pindex); 1360 m->object = NULL; 1361 cnt.v_cache_count--; 1362} 1363 1364/* 1365 * Transfer all of the cached pages with offset greater than or 1366 * equal to 'offidxstart' from the original object's cache to the 1367 * new object's cache. However, any cached pages with offset 1368 * greater than or equal to the new object's size are kept in the 1369 * original object. Initially, the new object's cache must be 1370 * empty. Offset 'offidxstart' in the original object must 1371 * correspond to offset zero in the new object. 1372 * 1373 * The new object must be locked. 1374 */ 1375void 1376vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1377 vm_object_t new_object) 1378{ 1379 vm_page_t m; 1380 1381 /* 1382 * Insertion into an object's collection of cached pages 1383 * requires the object to be locked. In contrast, removal does 1384 * not. 1385 */ 1386 VM_OBJECT_ASSERT_WLOCKED(new_object); 1387 KASSERT(vm_radix_is_empty(&new_object->cache), 1388 ("vm_page_cache_transfer: object %p has cached pages", 1389 new_object)); 1390 mtx_lock(&vm_page_queue_free_mtx); 1391 while ((m = vm_radix_lookup_ge(&orig_object->cache, 1392 offidxstart)) != NULL) { 1393 /* 1394 * Transfer all of the pages with offset greater than or 1395 * equal to 'offidxstart' from the original object's 1396 * cache to the new object's cache. 1397 */ 1398 if ((m->pindex - offidxstart) >= new_object->size) 1399 break; 1400 vm_radix_remove(&orig_object->cache, m->pindex); 1401 /* Update the page's object and offset. */ 1402 m->object = new_object; 1403 m->pindex -= offidxstart; 1404 if (vm_radix_insert(&new_object->cache, m)) 1405 vm_page_cache_turn_free(m); 1406 } 1407 mtx_unlock(&vm_page_queue_free_mtx); 1408} 1409 1410/* 1411 * Returns TRUE if a cached page is associated with the given object and 1412 * offset, and FALSE otherwise. 1413 * 1414 * The object must be locked. 1415 */ 1416boolean_t 1417vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1418{ 1419 vm_page_t m; 1420 1421 /* 1422 * Insertion into an object's collection of cached pages requires the 1423 * object to be locked. Therefore, if the object is locked and the 1424 * object's collection is empty, there is no need to acquire the free 1425 * page queues lock in order to prove that the specified page doesn't 1426 * exist. 1427 */ 1428 VM_OBJECT_ASSERT_WLOCKED(object); 1429 if (__predict_true(vm_object_cache_is_empty(object))) 1430 return (FALSE); 1431 mtx_lock(&vm_page_queue_free_mtx); 1432 m = vm_page_cache_lookup(object, pindex); 1433 mtx_unlock(&vm_page_queue_free_mtx); 1434 return (m != NULL); 1435} 1436 1437/* 1438 * vm_page_alloc: 1439 * 1440 * Allocate and return a page that is associated with the specified 1441 * object and offset pair. By default, this page is exclusive busied. 1442 * 1443 * The caller must always specify an allocation class. 1444 * 1445 * allocation classes: 1446 * VM_ALLOC_NORMAL normal process request 1447 * VM_ALLOC_SYSTEM system *really* needs a page 1448 * VM_ALLOC_INTERRUPT interrupt time request 1449 * 1450 * optional allocation flags: 1451 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1452 * intends to allocate 1453 * VM_ALLOC_IFCACHED return page only if it is cached 1454 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1455 * is cached 1456 * VM_ALLOC_NOBUSY do not exclusive busy the page 1457 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1458 * VM_ALLOC_NOOBJ page is not associated with an object and 1459 * should not be exclusive busy 1460 * VM_ALLOC_SBUSY shared busy the allocated page 1461 * VM_ALLOC_WIRED wire the allocated page 1462 * VM_ALLOC_ZERO prefer a zeroed page 1463 * 1464 * This routine may not sleep. 1465 */ 1466vm_page_t 1467vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1468{ 1469 struct vnode *vp = NULL; 1470 vm_object_t m_object; 1471 vm_page_t m, mpred; 1472 int flags, req_class; 1473 1474 mpred = 0; /* XXX: pacify gcc */ 1475 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1476 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1477 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1478 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1479 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1480 req)); 1481 if (object != NULL) 1482 VM_OBJECT_ASSERT_WLOCKED(object); 1483 1484 req_class = req & VM_ALLOC_CLASS_MASK; 1485 1486 /* 1487 * The page daemon is allowed to dig deeper into the free page list. 1488 */ 1489 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1490 req_class = VM_ALLOC_SYSTEM; 1491 1492 if (object != NULL) { 1493 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1494 KASSERT(mpred == NULL || mpred->pindex != pindex, 1495 ("vm_page_alloc: pindex already allocated")); 1496 } 1497 1498 /* 1499 * The page allocation request can came from consumers which already 1500 * hold the free page queue mutex, like vm_page_insert() in 1501 * vm_page_cache(). 1502 */ 1503 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1504 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1505 (req_class == VM_ALLOC_SYSTEM && 1506 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1507 (req_class == VM_ALLOC_INTERRUPT && 1508 cnt.v_free_count + cnt.v_cache_count > 0)) { 1509 /* 1510 * Allocate from the free queue if the number of free pages 1511 * exceeds the minimum for the request class. 1512 */ 1513 if (object != NULL && 1514 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1515 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1516 mtx_unlock(&vm_page_queue_free_mtx); 1517 return (NULL); 1518 } 1519 if (vm_phys_unfree_page(m)) 1520 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1521#if VM_NRESERVLEVEL > 0 1522 else if (!vm_reserv_reactivate_page(m)) 1523#else 1524 else 1525#endif 1526 panic("vm_page_alloc: cache page %p is missing" 1527 " from the free queue", m); 1528 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1529 mtx_unlock(&vm_page_queue_free_mtx); 1530 return (NULL); 1531#if VM_NRESERVLEVEL > 0 1532 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1533 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1534 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) { 1535#else 1536 } else { 1537#endif 1538 m = vm_phys_alloc_pages(object != NULL ? 1539 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1540#if VM_NRESERVLEVEL > 0 1541 if (m == NULL && vm_reserv_reclaim_inactive()) { 1542 m = vm_phys_alloc_pages(object != NULL ? 1543 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1544 0); 1545 } 1546#endif 1547 } 1548 } else { 1549 /* 1550 * Not allocatable, give up. 1551 */ 1552 mtx_unlock(&vm_page_queue_free_mtx); 1553 atomic_add_int(&vm_pageout_deficit, 1554 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1555 pagedaemon_wakeup(); 1556 return (NULL); 1557 } 1558 1559 /* 1560 * At this point we had better have found a good page. 1561 */ 1562 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1563 KASSERT(m->queue == PQ_NONE, 1564 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1565 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1566 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1567 KASSERT(!vm_page_sbusied(m), 1568 ("vm_page_alloc: page %p is busy", m)); 1569 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1570 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1571 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1572 pmap_page_get_memattr(m))); 1573 if ((m->flags & PG_CACHED) != 0) { 1574 KASSERT((m->flags & PG_ZERO) == 0, 1575 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1576 KASSERT(m->valid != 0, 1577 ("vm_page_alloc: cached page %p is invalid", m)); 1578 if (m->object == object && m->pindex == pindex) 1579 cnt.v_reactivated++; 1580 else 1581 m->valid = 0; 1582 m_object = m->object; 1583 vm_page_cache_remove(m); 1584 if (m_object->type == OBJT_VNODE && 1585 vm_object_cache_is_empty(m_object)) 1586 vp = m_object->handle; 1587 } else { 1588 KASSERT(VM_PAGE_IS_FREE(m), 1589 ("vm_page_alloc: page %p is not free", m)); 1590 KASSERT(m->valid == 0, 1591 ("vm_page_alloc: free page %p is valid", m)); 1592 vm_phys_freecnt_adj(m, -1); 1593 } 1594 1595 /* 1596 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag 1597 * must be cleared before the free page queues lock is released. 1598 */ 1599 flags = 0; 1600 if (m->flags & PG_ZERO) { 1601 vm_page_zero_count--; 1602 if (req & VM_ALLOC_ZERO) 1603 flags = PG_ZERO; 1604 } 1605 if (req & VM_ALLOC_NODUMP) 1606 flags |= PG_NODUMP; 1607 m->flags = flags; 1608 mtx_unlock(&vm_page_queue_free_mtx); 1609 m->aflags = 0; 1610 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1611 VPO_UNMANAGED : 0; 1612 m->busy_lock = VPB_UNBUSIED; 1613 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1614 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1615 if ((req & VM_ALLOC_SBUSY) != 0) 1616 m->busy_lock = VPB_SHARERS_WORD(1); 1617 if (req & VM_ALLOC_WIRED) { 1618 /* 1619 * The page lock is not required for wiring a page until that 1620 * page is inserted into the object. 1621 */ 1622 atomic_add_int(&cnt.v_wire_count, 1); 1623 m->wire_count = 1; 1624 } 1625 m->act_count = 0; 1626 1627 if (object != NULL) { 1628 if (vm_page_insert_after(m, object, pindex, mpred)) { 1629 /* See the comment below about hold count. */ 1630 if (vp != NULL) 1631 vdrop(vp); 1632 pagedaemon_wakeup(); 1633 if (req & VM_ALLOC_WIRED) { 1634 atomic_subtract_int(&cnt.v_wire_count, 1); 1635 m->wire_count = 0; 1636 } 1637 m->object = NULL; 1638 m->oflags = VPO_UNMANAGED; 1639 m->busy_lock = VPB_UNBUSIED; 1640 vm_page_free(m); 1641 return (NULL); 1642 } 1643 1644 /* Ignore device objects; the pager sets "memattr" for them. */ 1645 if (object->memattr != VM_MEMATTR_DEFAULT && 1646 (object->flags & OBJ_FICTITIOUS) == 0) 1647 pmap_page_set_memattr(m, object->memattr); 1648 } else 1649 m->pindex = pindex; 1650 1651 /* 1652 * The following call to vdrop() must come after the above call 1653 * to vm_page_insert() in case both affect the same object and 1654 * vnode. Otherwise, the affected vnode's hold count could 1655 * temporarily become zero. 1656 */ 1657 if (vp != NULL) 1658 vdrop(vp); 1659 1660 /* 1661 * Don't wakeup too often - wakeup the pageout daemon when 1662 * we would be nearly out of memory. 1663 */ 1664 if (vm_paging_needed()) 1665 pagedaemon_wakeup(); 1666 1667 return (m); 1668} 1669 1670static void 1671vm_page_alloc_contig_vdrop(struct spglist *lst) 1672{ 1673 1674 while (!SLIST_EMPTY(lst)) { 1675 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv); 1676 SLIST_REMOVE_HEAD(lst, plinks.s.ss); 1677 } 1678} 1679 1680/* 1681 * vm_page_alloc_contig: 1682 * 1683 * Allocate a contiguous set of physical pages of the given size "npages" 1684 * from the free lists. All of the physical pages must be at or above 1685 * the given physical address "low" and below the given physical address 1686 * "high". The given value "alignment" determines the alignment of the 1687 * first physical page in the set. If the given value "boundary" is 1688 * non-zero, then the set of physical pages cannot cross any physical 1689 * address boundary that is a multiple of that value. Both "alignment" 1690 * and "boundary" must be a power of two. 1691 * 1692 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1693 * then the memory attribute setting for the physical pages is configured 1694 * to the object's memory attribute setting. Otherwise, the memory 1695 * attribute setting for the physical pages is configured to "memattr", 1696 * overriding the object's memory attribute setting. However, if the 1697 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1698 * memory attribute setting for the physical pages cannot be configured 1699 * to VM_MEMATTR_DEFAULT. 1700 * 1701 * The caller must always specify an allocation class. 1702 * 1703 * allocation classes: 1704 * VM_ALLOC_NORMAL normal process request 1705 * VM_ALLOC_SYSTEM system *really* needs a page 1706 * VM_ALLOC_INTERRUPT interrupt time request 1707 * 1708 * optional allocation flags: 1709 * VM_ALLOC_NOBUSY do not exclusive busy the page 1710 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1711 * VM_ALLOC_NOOBJ page is not associated with an object and 1712 * should not be exclusive busy 1713 * VM_ALLOC_SBUSY shared busy the allocated page 1714 * VM_ALLOC_WIRED wire the allocated page 1715 * VM_ALLOC_ZERO prefer a zeroed page 1716 * 1717 * This routine may not sleep. 1718 */ 1719vm_page_t 1720vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1721 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1722 vm_paddr_t boundary, vm_memattr_t memattr) 1723{ 1724 struct vnode *drop; 1725 struct spglist deferred_vdrop_list; 1726 vm_page_t m, m_tmp, m_ret; 1727 u_int flags, oflags; 1728 int req_class; 1729 1730 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1731 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1732 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1733 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1734 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1735 req)); 1736 if (object != NULL) { 1737 VM_OBJECT_ASSERT_WLOCKED(object); 1738 KASSERT(object->type == OBJT_PHYS, 1739 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1740 object)); 1741 } 1742 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1743 req_class = req & VM_ALLOC_CLASS_MASK; 1744 1745 /* 1746 * The page daemon is allowed to dig deeper into the free page list. 1747 */ 1748 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1749 req_class = VM_ALLOC_SYSTEM; 1750 1751 SLIST_INIT(&deferred_vdrop_list); 1752 mtx_lock(&vm_page_queue_free_mtx); 1753 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1754 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1755 cnt.v_free_count + cnt.v_cache_count >= npages + 1756 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1757 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1758#if VM_NRESERVLEVEL > 0 1759retry: 1760 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1761 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1762 low, high, alignment, boundary)) == NULL) 1763#endif 1764 m_ret = vm_phys_alloc_contig(npages, low, high, 1765 alignment, boundary); 1766 } else { 1767 mtx_unlock(&vm_page_queue_free_mtx); 1768 atomic_add_int(&vm_pageout_deficit, npages); 1769 pagedaemon_wakeup(); 1770 return (NULL); 1771 } 1772 if (m_ret != NULL) 1773 for (m = m_ret; m < &m_ret[npages]; m++) { 1774 drop = vm_page_alloc_init(m); 1775 if (drop != NULL) { 1776 /* 1777 * Enqueue the vnode for deferred vdrop(). 1778 */ 1779 m->plinks.s.pv = drop; 1780 SLIST_INSERT_HEAD(&deferred_vdrop_list, m, 1781 plinks.s.ss); 1782 } 1783 } 1784 else { 1785#if VM_NRESERVLEVEL > 0 1786 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1787 boundary)) 1788 goto retry; 1789#endif 1790 } 1791 mtx_unlock(&vm_page_queue_free_mtx); 1792 if (m_ret == NULL) 1793 return (NULL); 1794 1795 /* 1796 * Initialize the pages. Only the PG_ZERO flag is inherited. 1797 */ 1798 flags = 0; 1799 if ((req & VM_ALLOC_ZERO) != 0) 1800 flags = PG_ZERO; 1801 if ((req & VM_ALLOC_NODUMP) != 0) 1802 flags |= PG_NODUMP; 1803 if ((req & VM_ALLOC_WIRED) != 0) 1804 atomic_add_int(&cnt.v_wire_count, npages); 1805 oflags = VPO_UNMANAGED; 1806 if (object != NULL) { 1807 if (object->memattr != VM_MEMATTR_DEFAULT && 1808 memattr == VM_MEMATTR_DEFAULT) 1809 memattr = object->memattr; 1810 } 1811 for (m = m_ret; m < &m_ret[npages]; m++) { 1812 m->aflags = 0; 1813 m->flags = (m->flags | PG_NODUMP) & flags; 1814 m->busy_lock = VPB_UNBUSIED; 1815 if (object != NULL) { 1816 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1817 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1818 if ((req & VM_ALLOC_SBUSY) != 0) 1819 m->busy_lock = VPB_SHARERS_WORD(1); 1820 } 1821 if ((req & VM_ALLOC_WIRED) != 0) 1822 m->wire_count = 1; 1823 /* Unmanaged pages don't use "act_count". */ 1824 m->oflags = oflags; 1825 if (object != NULL) { 1826 if (vm_page_insert(m, object, pindex)) { 1827 vm_page_alloc_contig_vdrop( 1828 &deferred_vdrop_list); 1829 if (vm_paging_needed()) 1830 pagedaemon_wakeup(); 1831 if ((req & VM_ALLOC_WIRED) != 0) 1832 atomic_subtract_int(&cnt.v_wire_count, 1833 npages); 1834 for (m_tmp = m, m = m_ret; 1835 m < &m_ret[npages]; m++) { 1836 if ((req & VM_ALLOC_WIRED) != 0) 1837 m->wire_count = 0; 1838 if (m >= m_tmp) { 1839 m->object = NULL; 1840 m->oflags |= VPO_UNMANAGED; 1841 } 1842 m->busy_lock = VPB_UNBUSIED; 1843 vm_page_free(m); 1844 } 1845 return (NULL); 1846 } 1847 } else 1848 m->pindex = pindex; 1849 if (memattr != VM_MEMATTR_DEFAULT) 1850 pmap_page_set_memattr(m, memattr); 1851 pindex++; 1852 } 1853 vm_page_alloc_contig_vdrop(&deferred_vdrop_list); 1854 if (vm_paging_needed()) 1855 pagedaemon_wakeup(); 1856 return (m_ret); 1857} 1858 1859/* 1860 * Initialize a page that has been freshly dequeued from a freelist. 1861 * The caller has to drop the vnode returned, if it is not NULL. 1862 * 1863 * This function may only be used to initialize unmanaged pages. 1864 * 1865 * To be called with vm_page_queue_free_mtx held. 1866 */ 1867static struct vnode * 1868vm_page_alloc_init(vm_page_t m) 1869{ 1870 struct vnode *drop; 1871 vm_object_t m_object; 1872 1873 KASSERT(m->queue == PQ_NONE, 1874 ("vm_page_alloc_init: page %p has unexpected queue %d", 1875 m, m->queue)); 1876 KASSERT(m->wire_count == 0, 1877 ("vm_page_alloc_init: page %p is wired", m)); 1878 KASSERT(m->hold_count == 0, 1879 ("vm_page_alloc_init: page %p is held", m)); 1880 KASSERT(!vm_page_sbusied(m), 1881 ("vm_page_alloc_init: page %p is busy", m)); 1882 KASSERT(m->dirty == 0, 1883 ("vm_page_alloc_init: page %p is dirty", m)); 1884 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1885 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1886 m, pmap_page_get_memattr(m))); 1887 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1888 drop = NULL; 1889 if ((m->flags & PG_CACHED) != 0) { 1890 KASSERT((m->flags & PG_ZERO) == 0, 1891 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1892 m->valid = 0; 1893 m_object = m->object; 1894 vm_page_cache_remove(m); 1895 if (m_object->type == OBJT_VNODE && 1896 vm_object_cache_is_empty(m_object)) 1897 drop = m_object->handle; 1898 } else { 1899 KASSERT(VM_PAGE_IS_FREE(m), 1900 ("vm_page_alloc_init: page %p is not free", m)); 1901 KASSERT(m->valid == 0, 1902 ("vm_page_alloc_init: free page %p is valid", m)); 1903 vm_phys_freecnt_adj(m, -1); 1904 if ((m->flags & PG_ZERO) != 0) 1905 vm_page_zero_count--; 1906 } 1907 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1908 m->flags &= PG_ZERO; 1909 return (drop); 1910} 1911 1912/* 1913 * vm_page_alloc_freelist: 1914 * 1915 * Allocate a physical page from the specified free page list. 1916 * 1917 * The caller must always specify an allocation class. 1918 * 1919 * allocation classes: 1920 * VM_ALLOC_NORMAL normal process request 1921 * VM_ALLOC_SYSTEM system *really* needs a page 1922 * VM_ALLOC_INTERRUPT interrupt time request 1923 * 1924 * optional allocation flags: 1925 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1926 * intends to allocate 1927 * VM_ALLOC_WIRED wire the allocated page 1928 * VM_ALLOC_ZERO prefer a zeroed page 1929 * 1930 * This routine may not sleep. 1931 */ 1932vm_page_t 1933vm_page_alloc_freelist(int flind, int req) 1934{ 1935 struct vnode *drop; 1936 vm_page_t m; 1937 u_int flags; 1938 int req_class; 1939 1940 req_class = req & VM_ALLOC_CLASS_MASK; 1941 1942 /* 1943 * The page daemon is allowed to dig deeper into the free page list. 1944 */ 1945 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1946 req_class = VM_ALLOC_SYSTEM; 1947 1948 /* 1949 * Do not allocate reserved pages unless the req has asked for it. 1950 */ 1951 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1952 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1953 (req_class == VM_ALLOC_SYSTEM && 1954 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1955 (req_class == VM_ALLOC_INTERRUPT && 1956 cnt.v_free_count + cnt.v_cache_count > 0)) 1957 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1958 else { 1959 mtx_unlock(&vm_page_queue_free_mtx); 1960 atomic_add_int(&vm_pageout_deficit, 1961 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1962 pagedaemon_wakeup(); 1963 return (NULL); 1964 } 1965 if (m == NULL) { 1966 mtx_unlock(&vm_page_queue_free_mtx); 1967 return (NULL); 1968 } 1969 drop = vm_page_alloc_init(m); 1970 mtx_unlock(&vm_page_queue_free_mtx); 1971 1972 /* 1973 * Initialize the page. Only the PG_ZERO flag is inherited. 1974 */ 1975 m->aflags = 0; 1976 flags = 0; 1977 if ((req & VM_ALLOC_ZERO) != 0) 1978 flags = PG_ZERO; 1979 m->flags &= flags; 1980 if ((req & VM_ALLOC_WIRED) != 0) { 1981 /* 1982 * The page lock is not required for wiring a page that does 1983 * not belong to an object. 1984 */ 1985 atomic_add_int(&cnt.v_wire_count, 1); 1986 m->wire_count = 1; 1987 } 1988 /* Unmanaged pages don't use "act_count". */ 1989 m->oflags = VPO_UNMANAGED; 1990 if (drop != NULL) 1991 vdrop(drop); 1992 if (vm_paging_needed()) 1993 pagedaemon_wakeup(); 1994 return (m); 1995} 1996 1997/* 1998 * vm_wait: (also see VM_WAIT macro) 1999 * 2000 * Sleep until free pages are available for allocation. 2001 * - Called in various places before memory allocations. 2002 */ 2003void 2004vm_wait(void) 2005{ 2006 2007 mtx_lock(&vm_page_queue_free_mtx); 2008 if (curproc == pageproc) { 2009 vm_pageout_pages_needed = 1; 2010 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2011 PDROP | PSWP, "VMWait", 0); 2012 } else { 2013 if (!vm_pages_needed) { 2014 vm_pages_needed = 1; 2015 wakeup(&vm_pages_needed); 2016 } 2017 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2018 "vmwait", 0); 2019 } 2020} 2021 2022/* 2023 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2024 * 2025 * Sleep until free pages are available for allocation. 2026 * - Called only in vm_fault so that processes page faulting 2027 * can be easily tracked. 2028 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2029 * processes will be able to grab memory first. Do not change 2030 * this balance without careful testing first. 2031 */ 2032void 2033vm_waitpfault(void) 2034{ 2035 2036 mtx_lock(&vm_page_queue_free_mtx); 2037 if (!vm_pages_needed) { 2038 vm_pages_needed = 1; 2039 wakeup(&vm_pages_needed); 2040 } 2041 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2042 "pfault", 0); 2043} 2044 2045struct vm_pagequeue * 2046vm_page_pagequeue(vm_page_t m) 2047{ 2048 2049 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2050} 2051 2052/* 2053 * vm_page_dequeue: 2054 * 2055 * Remove the given page from its current page queue. 2056 * 2057 * The page must be locked. 2058 */ 2059void 2060vm_page_dequeue(vm_page_t m) 2061{ 2062 struct vm_pagequeue *pq; 2063 2064 vm_page_lock_assert(m, MA_OWNED); 2065 KASSERT(m->queue != PQ_NONE, 2066 ("vm_page_dequeue: page %p is not queued", m)); 2067 pq = vm_page_pagequeue(m); 2068 vm_pagequeue_lock(pq); 2069 m->queue = PQ_NONE; 2070 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2071 vm_pagequeue_cnt_dec(pq); 2072 vm_pagequeue_unlock(pq); 2073} 2074 2075/* 2076 * vm_page_dequeue_locked: 2077 * 2078 * Remove the given page from its current page queue. 2079 * 2080 * The page and page queue must be locked. 2081 */ 2082void 2083vm_page_dequeue_locked(vm_page_t m) 2084{ 2085 struct vm_pagequeue *pq; 2086 2087 vm_page_lock_assert(m, MA_OWNED); 2088 pq = vm_page_pagequeue(m); 2089 vm_pagequeue_assert_locked(pq); 2090 m->queue = PQ_NONE; 2091 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2092 vm_pagequeue_cnt_dec(pq); 2093} 2094 2095/* 2096 * vm_page_enqueue: 2097 * 2098 * Add the given page to the specified page queue. 2099 * 2100 * The page must be locked. 2101 */ 2102static void 2103vm_page_enqueue(int queue, vm_page_t m) 2104{ 2105 struct vm_pagequeue *pq; 2106 2107 vm_page_lock_assert(m, MA_OWNED); 2108 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2109 vm_pagequeue_lock(pq); 2110 m->queue = queue; 2111 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2112 vm_pagequeue_cnt_inc(pq); 2113 vm_pagequeue_unlock(pq); 2114} 2115 2116/* 2117 * vm_page_requeue: 2118 * 2119 * Move the given page to the tail of its current page queue. 2120 * 2121 * The page must be locked. 2122 */ 2123void 2124vm_page_requeue(vm_page_t m) 2125{ 2126 struct vm_pagequeue *pq; 2127 2128 vm_page_lock_assert(m, MA_OWNED); 2129 KASSERT(m->queue != PQ_NONE, 2130 ("vm_page_requeue: page %p is not queued", m)); 2131 pq = vm_page_pagequeue(m); 2132 vm_pagequeue_lock(pq); 2133 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2134 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2135 vm_pagequeue_unlock(pq); 2136} 2137 2138/* 2139 * vm_page_requeue_locked: 2140 * 2141 * Move the given page to the tail of its current page queue. 2142 * 2143 * The page queue must be locked. 2144 */ 2145void 2146vm_page_requeue_locked(vm_page_t m) 2147{ 2148 struct vm_pagequeue *pq; 2149 2150 KASSERT(m->queue != PQ_NONE, 2151 ("vm_page_requeue_locked: page %p is not queued", m)); 2152 pq = vm_page_pagequeue(m); 2153 vm_pagequeue_assert_locked(pq); 2154 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2155 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2156} 2157 2158/* 2159 * vm_page_activate: 2160 * 2161 * Put the specified page on the active list (if appropriate). 2162 * Ensure that act_count is at least ACT_INIT but do not otherwise 2163 * mess with it. 2164 * 2165 * The page must be locked. 2166 */ 2167void 2168vm_page_activate(vm_page_t m) 2169{ 2170 int queue; 2171 2172 vm_page_lock_assert(m, MA_OWNED); 2173 if ((queue = m->queue) != PQ_ACTIVE) { 2174 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2175 if (m->act_count < ACT_INIT) 2176 m->act_count = ACT_INIT; 2177 if (queue != PQ_NONE) 2178 vm_page_dequeue(m); 2179 vm_page_enqueue(PQ_ACTIVE, m); 2180 } else 2181 KASSERT(queue == PQ_NONE, 2182 ("vm_page_activate: wired page %p is queued", m)); 2183 } else { 2184 if (m->act_count < ACT_INIT) 2185 m->act_count = ACT_INIT; 2186 } 2187} 2188 2189/* 2190 * vm_page_free_wakeup: 2191 * 2192 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2193 * routine is called when a page has been added to the cache or free 2194 * queues. 2195 * 2196 * The page queues must be locked. 2197 */ 2198static inline void 2199vm_page_free_wakeup(void) 2200{ 2201 2202 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2203 /* 2204 * if pageout daemon needs pages, then tell it that there are 2205 * some free. 2206 */ 2207 if (vm_pageout_pages_needed && 2208 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2209 wakeup(&vm_pageout_pages_needed); 2210 vm_pageout_pages_needed = 0; 2211 } 2212 /* 2213 * wakeup processes that are waiting on memory if we hit a 2214 * high water mark. And wakeup scheduler process if we have 2215 * lots of memory. this process will swapin processes. 2216 */ 2217 if (vm_pages_needed && !vm_page_count_min()) { 2218 vm_pages_needed = 0; 2219 wakeup(&cnt.v_free_count); 2220 } 2221} 2222 2223/* 2224 * Turn a cached page into a free page, by changing its attributes. 2225 * Keep the statistics up-to-date. 2226 * 2227 * The free page queue must be locked. 2228 */ 2229static void 2230vm_page_cache_turn_free(vm_page_t m) 2231{ 2232 2233 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2234 2235 m->object = NULL; 2236 m->valid = 0; 2237 /* Clear PG_CACHED and set PG_FREE. */ 2238 m->flags ^= PG_CACHED | PG_FREE; 2239 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 2240 ("vm_page_cache_free: page %p has inconsistent flags", m)); 2241 cnt.v_cache_count--; 2242 vm_phys_freecnt_adj(m, 1); 2243} 2244 2245/* 2246 * vm_page_free_toq: 2247 * 2248 * Returns the given page to the free list, 2249 * disassociating it with any VM object. 2250 * 2251 * The object must be locked. The page must be locked if it is managed. 2252 */ 2253void 2254vm_page_free_toq(vm_page_t m) 2255{ 2256 2257 if ((m->oflags & VPO_UNMANAGED) == 0) { 2258 vm_page_lock_assert(m, MA_OWNED); 2259 KASSERT(!pmap_page_is_mapped(m), 2260 ("vm_page_free_toq: freeing mapped page %p", m)); 2261 } else 2262 KASSERT(m->queue == PQ_NONE, 2263 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2264 PCPU_INC(cnt.v_tfree); 2265 2266 if (VM_PAGE_IS_FREE(m)) 2267 panic("vm_page_free: freeing free page %p", m); 2268 else if (vm_page_sbusied(m)) 2269 panic("vm_page_free: freeing busy page %p", m); 2270 2271 /* 2272 * Unqueue, then remove page. Note that we cannot destroy 2273 * the page here because we do not want to call the pager's 2274 * callback routine until after we've put the page on the 2275 * appropriate free queue. 2276 */ 2277 vm_page_remque(m); 2278 vm_page_remove(m); 2279 2280 /* 2281 * If fictitious remove object association and 2282 * return, otherwise delay object association removal. 2283 */ 2284 if ((m->flags & PG_FICTITIOUS) != 0) { 2285 return; 2286 } 2287 2288 m->valid = 0; 2289 vm_page_undirty(m); 2290 2291 if (m->wire_count != 0) 2292 panic("vm_page_free: freeing wired page %p", m); 2293 if (m->hold_count != 0) { 2294 m->flags &= ~PG_ZERO; 2295 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2296 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2297 m->flags |= PG_UNHOLDFREE; 2298 } else { 2299 /* 2300 * Restore the default memory attribute to the page. 2301 */ 2302 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2303 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2304 2305 /* 2306 * Insert the page into the physical memory allocator's 2307 * cache/free page queues. 2308 */ 2309 mtx_lock(&vm_page_queue_free_mtx); 2310 m->flags |= PG_FREE; 2311 vm_phys_freecnt_adj(m, 1); 2312#if VM_NRESERVLEVEL > 0 2313 if (!vm_reserv_free_page(m)) 2314#else 2315 if (TRUE) 2316#endif 2317 vm_phys_free_pages(m, 0); 2318 if ((m->flags & PG_ZERO) != 0) 2319 ++vm_page_zero_count; 2320 else 2321 vm_page_zero_idle_wakeup(); 2322 vm_page_free_wakeup(); 2323 mtx_unlock(&vm_page_queue_free_mtx); 2324 } 2325} 2326 2327/* 2328 * vm_page_wire: 2329 * 2330 * Mark this page as wired down by yet 2331 * another map, removing it from paging queues 2332 * as necessary. 2333 * 2334 * If the page is fictitious, then its wire count must remain one. 2335 * 2336 * The page must be locked. 2337 */ 2338void 2339vm_page_wire(vm_page_t m) 2340{ 2341 2342 /* 2343 * Only bump the wire statistics if the page is not already wired, 2344 * and only unqueue the page if it is on some queue (if it is unmanaged 2345 * it is already off the queues). 2346 */ 2347 vm_page_lock_assert(m, MA_OWNED); 2348 if ((m->flags & PG_FICTITIOUS) != 0) { 2349 KASSERT(m->wire_count == 1, 2350 ("vm_page_wire: fictitious page %p's wire count isn't one", 2351 m)); 2352 return; 2353 } 2354 if (m->wire_count == 0) { 2355 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2356 m->queue == PQ_NONE, 2357 ("vm_page_wire: unmanaged page %p is queued", m)); 2358 vm_page_remque(m); 2359 atomic_add_int(&cnt.v_wire_count, 1); 2360 } 2361 m->wire_count++; 2362 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2363} 2364 2365/* 2366 * vm_page_unwire: 2367 * 2368 * Release one wiring of the specified page, potentially enabling it to be 2369 * paged again. If paging is enabled, then the value of the parameter 2370 * "activate" determines to which queue the page is added. If "activate" is 2371 * non-zero, then the page is added to the active queue. Otherwise, it is 2372 * added to the inactive queue. 2373 * 2374 * However, unless the page belongs to an object, it is not enqueued because 2375 * it cannot be paged out. 2376 * 2377 * If a page is fictitious, then its wire count must always be one. 2378 * 2379 * A managed page must be locked. 2380 */ 2381void 2382vm_page_unwire(vm_page_t m, int activate) 2383{ 2384 2385 if ((m->oflags & VPO_UNMANAGED) == 0) 2386 vm_page_lock_assert(m, MA_OWNED); 2387 if ((m->flags & PG_FICTITIOUS) != 0) { 2388 KASSERT(m->wire_count == 1, 2389 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2390 return; 2391 } 2392 if (m->wire_count > 0) { 2393 m->wire_count--; 2394 if (m->wire_count == 0) { 2395 atomic_subtract_int(&cnt.v_wire_count, 1); 2396 if ((m->oflags & VPO_UNMANAGED) != 0 || 2397 m->object == NULL) 2398 return; 2399 if (!activate) 2400 m->flags &= ~PG_WINATCFLS; 2401 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2402 } 2403 } else 2404 panic("vm_page_unwire: page %p's wire count is zero", m); 2405} 2406 2407/* 2408 * Move the specified page to the inactive queue. 2409 * 2410 * Many pages placed on the inactive queue should actually go 2411 * into the cache, but it is difficult to figure out which. What 2412 * we do instead, if the inactive target is well met, is to put 2413 * clean pages at the head of the inactive queue instead of the tail. 2414 * This will cause them to be moved to the cache more quickly and 2415 * if not actively re-referenced, reclaimed more quickly. If we just 2416 * stick these pages at the end of the inactive queue, heavy filesystem 2417 * meta-data accesses can cause an unnecessary paging load on memory bound 2418 * processes. This optimization causes one-time-use metadata to be 2419 * reused more quickly. 2420 * 2421 * Normally athead is 0 resulting in LRU operation. athead is set 2422 * to 1 if we want this page to be 'as if it were placed in the cache', 2423 * except without unmapping it from the process address space. 2424 * 2425 * The page must be locked. 2426 */ 2427static inline void 2428_vm_page_deactivate(vm_page_t m, int athead) 2429{ 2430 struct vm_pagequeue *pq; 2431 int queue; 2432 2433 vm_page_assert_locked(m); 2434 2435 /* 2436 * Ignore if the page is already inactive, unless it is unlikely to be 2437 * reactivated. 2438 */ 2439 if ((queue = m->queue) == PQ_INACTIVE && !athead) 2440 return; 2441 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2442 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2443 /* Avoid multiple acquisitions of the inactive queue lock. */ 2444 if (queue == PQ_INACTIVE) { 2445 vm_pagequeue_lock(pq); 2446 vm_page_dequeue_locked(m); 2447 } else { 2448 if (queue != PQ_NONE) 2449 vm_page_dequeue(m); 2450 m->flags &= ~PG_WINATCFLS; 2451 vm_pagequeue_lock(pq); 2452 } 2453 m->queue = PQ_INACTIVE; 2454 if (athead) 2455 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q); 2456 else 2457 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2458 vm_pagequeue_cnt_inc(pq); 2459 vm_pagequeue_unlock(pq); 2460 } 2461} 2462 2463/* 2464 * Move the specified page to the inactive queue. 2465 * 2466 * The page must be locked. 2467 */ 2468void 2469vm_page_deactivate(vm_page_t m) 2470{ 2471 2472 _vm_page_deactivate(m, 0); 2473} 2474 2475/* 2476 * vm_page_try_to_cache: 2477 * 2478 * Returns 0 on failure, 1 on success 2479 */ 2480int 2481vm_page_try_to_cache(vm_page_t m) 2482{ 2483 2484 vm_page_lock_assert(m, MA_OWNED); 2485 VM_OBJECT_ASSERT_WLOCKED(m->object); 2486 if (m->dirty || m->hold_count || m->wire_count || 2487 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2488 return (0); 2489 pmap_remove_all(m); 2490 if (m->dirty) 2491 return (0); 2492 vm_page_cache(m); 2493 return (1); 2494} 2495 2496/* 2497 * vm_page_try_to_free() 2498 * 2499 * Attempt to free the page. If we cannot free it, we do nothing. 2500 * 1 is returned on success, 0 on failure. 2501 */ 2502int 2503vm_page_try_to_free(vm_page_t m) 2504{ 2505 2506 vm_page_lock_assert(m, MA_OWNED); 2507 if (m->object != NULL) 2508 VM_OBJECT_ASSERT_WLOCKED(m->object); 2509 if (m->dirty || m->hold_count || m->wire_count || 2510 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2511 return (0); 2512 pmap_remove_all(m); 2513 if (m->dirty) 2514 return (0); 2515 vm_page_free(m); 2516 return (1); 2517} 2518 2519/* 2520 * vm_page_cache 2521 * 2522 * Put the specified page onto the page cache queue (if appropriate). 2523 * 2524 * The object and page must be locked. 2525 */ 2526void 2527vm_page_cache(vm_page_t m) 2528{ 2529 vm_object_t object; 2530 boolean_t cache_was_empty; 2531 2532 vm_page_lock_assert(m, MA_OWNED); 2533 object = m->object; 2534 VM_OBJECT_ASSERT_WLOCKED(object); 2535 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) || 2536 m->hold_count || m->wire_count) 2537 panic("vm_page_cache: attempting to cache busy page"); 2538 KASSERT(!pmap_page_is_mapped(m), 2539 ("vm_page_cache: page %p is mapped", m)); 2540 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2541 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2542 (object->type == OBJT_SWAP && 2543 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2544 /* 2545 * Hypothesis: A cache-elgible page belonging to a 2546 * default object or swap object but without a backing 2547 * store must be zero filled. 2548 */ 2549 vm_page_free(m); 2550 return; 2551 } 2552 KASSERT((m->flags & PG_CACHED) == 0, 2553 ("vm_page_cache: page %p is already cached", m)); 2554 2555 /* 2556 * Remove the page from the paging queues. 2557 */ 2558 vm_page_remque(m); 2559 2560 /* 2561 * Remove the page from the object's collection of resident 2562 * pages. 2563 */ 2564 vm_radix_remove(&object->rtree, m->pindex); 2565 TAILQ_REMOVE(&object->memq, m, listq); 2566 object->resident_page_count--; 2567 2568 /* 2569 * Restore the default memory attribute to the page. 2570 */ 2571 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2572 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2573 2574 /* 2575 * Insert the page into the object's collection of cached pages 2576 * and the physical memory allocator's cache/free page queues. 2577 */ 2578 m->flags &= ~PG_ZERO; 2579 mtx_lock(&vm_page_queue_free_mtx); 2580 cache_was_empty = vm_radix_is_empty(&object->cache); 2581 if (vm_radix_insert(&object->cache, m)) { 2582 mtx_unlock(&vm_page_queue_free_mtx); 2583 if (object->type == OBJT_VNODE && 2584 object->resident_page_count == 0) 2585 vdrop(object->handle); 2586 m->object = NULL; 2587 vm_page_free(m); 2588 return; 2589 } 2590 2591 /* 2592 * The above call to vm_radix_insert() could reclaim the one pre- 2593 * existing cached page from this object, resulting in a call to 2594 * vdrop(). 2595 */ 2596 if (!cache_was_empty) 2597 cache_was_empty = vm_radix_is_singleton(&object->cache); 2598 2599 m->flags |= PG_CACHED; 2600 cnt.v_cache_count++; 2601 PCPU_INC(cnt.v_tcached); 2602#if VM_NRESERVLEVEL > 0 2603 if (!vm_reserv_free_page(m)) { 2604#else 2605 if (TRUE) { 2606#endif 2607 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2608 vm_phys_free_pages(m, 0); 2609 } 2610 vm_page_free_wakeup(); 2611 mtx_unlock(&vm_page_queue_free_mtx); 2612 2613 /* 2614 * Increment the vnode's hold count if this is the object's only 2615 * cached page. Decrement the vnode's hold count if this was 2616 * the object's only resident page. 2617 */ 2618 if (object->type == OBJT_VNODE) { 2619 if (cache_was_empty && object->resident_page_count != 0) 2620 vhold(object->handle); 2621 else if (!cache_was_empty && object->resident_page_count == 0) 2622 vdrop(object->handle); 2623 } 2624} 2625 2626/* 2627 * vm_page_advise 2628 * 2629 * Deactivate or do nothing, as appropriate. This routine is used 2630 * by madvise() and vop_stdadvise(). 2631 * 2632 * The object and page must be locked. 2633 */ 2634void 2635vm_page_advise(vm_page_t m, int advice) 2636{ 2637 2638 vm_page_assert_locked(m); 2639 VM_OBJECT_ASSERT_WLOCKED(m->object); 2640 if (advice == MADV_FREE) 2641 /* 2642 * Mark the page clean. This will allow the page to be freed 2643 * up by the system. However, such pages are often reused 2644 * quickly by malloc() so we do not do anything that would 2645 * cause a page fault if we can help it. 2646 * 2647 * Specifically, we do not try to actually free the page now 2648 * nor do we try to put it in the cache (which would cause a 2649 * page fault on reuse). 2650 * 2651 * But we do make the page as freeable as we can without 2652 * actually taking the step of unmapping it. 2653 */ 2654 vm_page_undirty(m); 2655 else if (advice != MADV_DONTNEED) 2656 return; 2657 2658 /* 2659 * Clear any references to the page. Otherwise, the page daemon will 2660 * immediately reactivate the page. 2661 */ 2662 vm_page_aflag_clear(m, PGA_REFERENCED); 2663 2664 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 2665 vm_page_dirty(m); 2666 2667 /* 2668 * Place clean pages at the head of the inactive queue rather than the 2669 * tail, thus defeating the queue's LRU operation and ensuring that the 2670 * page will be reused quickly. 2671 */ 2672 _vm_page_deactivate(m, m->dirty == 0); 2673} 2674 2675/* 2676 * Grab a page, waiting until we are waken up due to the page 2677 * changing state. We keep on waiting, if the page continues 2678 * to be in the object. If the page doesn't exist, first allocate it 2679 * and then conditionally zero it. 2680 * 2681 * This routine may sleep. 2682 * 2683 * The object must be locked on entry. The lock will, however, be released 2684 * and reacquired if the routine sleeps. 2685 */ 2686vm_page_t 2687vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2688{ 2689 vm_page_t m; 2690 int sleep; 2691 2692 VM_OBJECT_ASSERT_WLOCKED(object); 2693 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 2694 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 2695 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 2696retrylookup: 2697 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2698 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 2699 vm_page_xbusied(m) : vm_page_busied(m); 2700 if (sleep) { 2701 /* 2702 * Reference the page before unlocking and 2703 * sleeping so that the page daemon is less 2704 * likely to reclaim it. 2705 */ 2706 vm_page_aflag_set(m, PGA_REFERENCED); 2707 vm_page_lock(m); 2708 VM_OBJECT_WUNLOCK(object); 2709 vm_page_busy_sleep(m, "pgrbwt"); 2710 VM_OBJECT_WLOCK(object); 2711 goto retrylookup; 2712 } else { 2713 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2714 vm_page_lock(m); 2715 vm_page_wire(m); 2716 vm_page_unlock(m); 2717 } 2718 if ((allocflags & 2719 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2720 vm_page_xbusy(m); 2721 if ((allocflags & VM_ALLOC_SBUSY) != 0) 2722 vm_page_sbusy(m); 2723 return (m); 2724 } 2725 } 2726 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY); 2727 if (m == NULL) { 2728 VM_OBJECT_WUNLOCK(object); 2729 VM_WAIT; 2730 VM_OBJECT_WLOCK(object); 2731 goto retrylookup; 2732 } else if (m->valid != 0) 2733 return (m); 2734 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2735 pmap_zero_page(m); 2736 return (m); 2737} 2738 2739/* 2740 * Mapping function for valid or dirty bits in a page. 2741 * 2742 * Inputs are required to range within a page. 2743 */ 2744vm_page_bits_t 2745vm_page_bits(int base, int size) 2746{ 2747 int first_bit; 2748 int last_bit; 2749 2750 KASSERT( 2751 base + size <= PAGE_SIZE, 2752 ("vm_page_bits: illegal base/size %d/%d", base, size) 2753 ); 2754 2755 if (size == 0) /* handle degenerate case */ 2756 return (0); 2757 2758 first_bit = base >> DEV_BSHIFT; 2759 last_bit = (base + size - 1) >> DEV_BSHIFT; 2760 2761 return (((vm_page_bits_t)2 << last_bit) - 2762 ((vm_page_bits_t)1 << first_bit)); 2763} 2764 2765/* 2766 * vm_page_set_valid_range: 2767 * 2768 * Sets portions of a page valid. The arguments are expected 2769 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2770 * of any partial chunks touched by the range. The invalid portion of 2771 * such chunks will be zeroed. 2772 * 2773 * (base + size) must be less then or equal to PAGE_SIZE. 2774 */ 2775void 2776vm_page_set_valid_range(vm_page_t m, int base, int size) 2777{ 2778 int endoff, frag; 2779 2780 VM_OBJECT_ASSERT_WLOCKED(m->object); 2781 if (size == 0) /* handle degenerate case */ 2782 return; 2783 2784 /* 2785 * If the base is not DEV_BSIZE aligned and the valid 2786 * bit is clear, we have to zero out a portion of the 2787 * first block. 2788 */ 2789 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2790 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2791 pmap_zero_page_area(m, frag, base - frag); 2792 2793 /* 2794 * If the ending offset is not DEV_BSIZE aligned and the 2795 * valid bit is clear, we have to zero out a portion of 2796 * the last block. 2797 */ 2798 endoff = base + size; 2799 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2800 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2801 pmap_zero_page_area(m, endoff, 2802 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2803 2804 /* 2805 * Assert that no previously invalid block that is now being validated 2806 * is already dirty. 2807 */ 2808 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2809 ("vm_page_set_valid_range: page %p is dirty", m)); 2810 2811 /* 2812 * Set valid bits inclusive of any overlap. 2813 */ 2814 m->valid |= vm_page_bits(base, size); 2815} 2816 2817/* 2818 * Clear the given bits from the specified page's dirty field. 2819 */ 2820static __inline void 2821vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2822{ 2823 uintptr_t addr; 2824#if PAGE_SIZE < 16384 2825 int shift; 2826#endif 2827 2828 /* 2829 * If the object is locked and the page is neither exclusive busy nor 2830 * write mapped, then the page's dirty field cannot possibly be 2831 * set by a concurrent pmap operation. 2832 */ 2833 VM_OBJECT_ASSERT_WLOCKED(m->object); 2834 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 2835 m->dirty &= ~pagebits; 2836 else { 2837 /* 2838 * The pmap layer can call vm_page_dirty() without 2839 * holding a distinguished lock. The combination of 2840 * the object's lock and an atomic operation suffice 2841 * to guarantee consistency of the page dirty field. 2842 * 2843 * For PAGE_SIZE == 32768 case, compiler already 2844 * properly aligns the dirty field, so no forcible 2845 * alignment is needed. Only require existence of 2846 * atomic_clear_64 when page size is 32768. 2847 */ 2848 addr = (uintptr_t)&m->dirty; 2849#if PAGE_SIZE == 32768 2850 atomic_clear_64((uint64_t *)addr, pagebits); 2851#elif PAGE_SIZE == 16384 2852 atomic_clear_32((uint32_t *)addr, pagebits); 2853#else /* PAGE_SIZE <= 8192 */ 2854 /* 2855 * Use a trick to perform a 32-bit atomic on the 2856 * containing aligned word, to not depend on the existence 2857 * of atomic_clear_{8, 16}. 2858 */ 2859 shift = addr & (sizeof(uint32_t) - 1); 2860#if BYTE_ORDER == BIG_ENDIAN 2861 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2862#else 2863 shift *= NBBY; 2864#endif 2865 addr &= ~(sizeof(uint32_t) - 1); 2866 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2867#endif /* PAGE_SIZE */ 2868 } 2869} 2870 2871/* 2872 * vm_page_set_validclean: 2873 * 2874 * Sets portions of a page valid and clean. The arguments are expected 2875 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2876 * of any partial chunks touched by the range. The invalid portion of 2877 * such chunks will be zero'd. 2878 * 2879 * (base + size) must be less then or equal to PAGE_SIZE. 2880 */ 2881void 2882vm_page_set_validclean(vm_page_t m, int base, int size) 2883{ 2884 vm_page_bits_t oldvalid, pagebits; 2885 int endoff, frag; 2886 2887 VM_OBJECT_ASSERT_WLOCKED(m->object); 2888 if (size == 0) /* handle degenerate case */ 2889 return; 2890 2891 /* 2892 * If the base is not DEV_BSIZE aligned and the valid 2893 * bit is clear, we have to zero out a portion of the 2894 * first block. 2895 */ 2896 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2897 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2898 pmap_zero_page_area(m, frag, base - frag); 2899 2900 /* 2901 * If the ending offset is not DEV_BSIZE aligned and the 2902 * valid bit is clear, we have to zero out a portion of 2903 * the last block. 2904 */ 2905 endoff = base + size; 2906 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2907 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2908 pmap_zero_page_area(m, endoff, 2909 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2910 2911 /* 2912 * Set valid, clear dirty bits. If validating the entire 2913 * page we can safely clear the pmap modify bit. We also 2914 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2915 * takes a write fault on a MAP_NOSYNC memory area the flag will 2916 * be set again. 2917 * 2918 * We set valid bits inclusive of any overlap, but we can only 2919 * clear dirty bits for DEV_BSIZE chunks that are fully within 2920 * the range. 2921 */ 2922 oldvalid = m->valid; 2923 pagebits = vm_page_bits(base, size); 2924 m->valid |= pagebits; 2925#if 0 /* NOT YET */ 2926 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2927 frag = DEV_BSIZE - frag; 2928 base += frag; 2929 size -= frag; 2930 if (size < 0) 2931 size = 0; 2932 } 2933 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2934#endif 2935 if (base == 0 && size == PAGE_SIZE) { 2936 /* 2937 * The page can only be modified within the pmap if it is 2938 * mapped, and it can only be mapped if it was previously 2939 * fully valid. 2940 */ 2941 if (oldvalid == VM_PAGE_BITS_ALL) 2942 /* 2943 * Perform the pmap_clear_modify() first. Otherwise, 2944 * a concurrent pmap operation, such as 2945 * pmap_protect(), could clear a modification in the 2946 * pmap and set the dirty field on the page before 2947 * pmap_clear_modify() had begun and after the dirty 2948 * field was cleared here. 2949 */ 2950 pmap_clear_modify(m); 2951 m->dirty = 0; 2952 m->oflags &= ~VPO_NOSYNC; 2953 } else if (oldvalid != VM_PAGE_BITS_ALL) 2954 m->dirty &= ~pagebits; 2955 else 2956 vm_page_clear_dirty_mask(m, pagebits); 2957} 2958 2959void 2960vm_page_clear_dirty(vm_page_t m, int base, int size) 2961{ 2962 2963 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2964} 2965 2966/* 2967 * vm_page_set_invalid: 2968 * 2969 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2970 * valid and dirty bits for the effected areas are cleared. 2971 */ 2972void 2973vm_page_set_invalid(vm_page_t m, int base, int size) 2974{ 2975 vm_page_bits_t bits; 2976 vm_object_t object; 2977 2978 object = m->object; 2979 VM_OBJECT_ASSERT_WLOCKED(object); 2980 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 2981 size >= object->un_pager.vnp.vnp_size) 2982 bits = VM_PAGE_BITS_ALL; 2983 else 2984 bits = vm_page_bits(base, size); 2985 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 2986 bits != 0) 2987 pmap_remove_all(m); 2988 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 2989 !pmap_page_is_mapped(m), 2990 ("vm_page_set_invalid: page %p is mapped", m)); 2991 m->valid &= ~bits; 2992 m->dirty &= ~bits; 2993} 2994 2995/* 2996 * vm_page_zero_invalid() 2997 * 2998 * The kernel assumes that the invalid portions of a page contain 2999 * garbage, but such pages can be mapped into memory by user code. 3000 * When this occurs, we must zero out the non-valid portions of the 3001 * page so user code sees what it expects. 3002 * 3003 * Pages are most often semi-valid when the end of a file is mapped 3004 * into memory and the file's size is not page aligned. 3005 */ 3006void 3007vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3008{ 3009 int b; 3010 int i; 3011 3012 VM_OBJECT_ASSERT_WLOCKED(m->object); 3013 /* 3014 * Scan the valid bits looking for invalid sections that 3015 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3016 * valid bit may be set ) have already been zeroed by 3017 * vm_page_set_validclean(). 3018 */ 3019 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3020 if (i == (PAGE_SIZE / DEV_BSIZE) || 3021 (m->valid & ((vm_page_bits_t)1 << i))) { 3022 if (i > b) { 3023 pmap_zero_page_area(m, 3024 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3025 } 3026 b = i + 1; 3027 } 3028 } 3029 3030 /* 3031 * setvalid is TRUE when we can safely set the zero'd areas 3032 * as being valid. We can do this if there are no cache consistancy 3033 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3034 */ 3035 if (setvalid) 3036 m->valid = VM_PAGE_BITS_ALL; 3037} 3038 3039/* 3040 * vm_page_is_valid: 3041 * 3042 * Is (partial) page valid? Note that the case where size == 0 3043 * will return FALSE in the degenerate case where the page is 3044 * entirely invalid, and TRUE otherwise. 3045 */ 3046int 3047vm_page_is_valid(vm_page_t m, int base, int size) 3048{ 3049 vm_page_bits_t bits; 3050 3051 VM_OBJECT_ASSERT_LOCKED(m->object); 3052 bits = vm_page_bits(base, size); 3053 return (m->valid != 0 && (m->valid & bits) == bits); 3054} 3055 3056/* 3057 * vm_page_ps_is_valid: 3058 * 3059 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3060 */ 3061boolean_t 3062vm_page_ps_is_valid(vm_page_t m) 3063{ 3064 int i, npages; 3065 3066 VM_OBJECT_ASSERT_LOCKED(m->object); 3067 npages = atop(pagesizes[m->psind]); 3068 3069 /* 3070 * The physically contiguous pages that make up a superpage, i.e., a 3071 * page with a page size index ("psind") greater than zero, will 3072 * occupy adjacent entries in vm_page_array[]. 3073 */ 3074 for (i = 0; i < npages; i++) { 3075 if (m[i].valid != VM_PAGE_BITS_ALL) 3076 return (FALSE); 3077 } 3078 return (TRUE); 3079} 3080 3081/* 3082 * Set the page's dirty bits if the page is modified. 3083 */ 3084void 3085vm_page_test_dirty(vm_page_t m) 3086{ 3087 3088 VM_OBJECT_ASSERT_WLOCKED(m->object); 3089 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3090 vm_page_dirty(m); 3091} 3092 3093void 3094vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3095{ 3096 3097 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3098} 3099 3100void 3101vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3102{ 3103 3104 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3105} 3106 3107int 3108vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3109{ 3110 3111 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3112} 3113 3114#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3115void 3116vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3117{ 3118 3119 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3120} 3121 3122void 3123vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3124{ 3125 3126 mtx_assert_(vm_page_lockptr(m), a, file, line); 3127} 3128#endif 3129 3130#ifdef INVARIANTS 3131void 3132vm_page_object_lock_assert(vm_page_t m) 3133{ 3134 3135 /* 3136 * Certain of the page's fields may only be modified by the 3137 * holder of the containing object's lock or the exclusive busy. 3138 * holder. Unfortunately, the holder of the write busy is 3139 * not recorded, and thus cannot be checked here. 3140 */ 3141 if (m->object != NULL && !vm_page_xbusied(m)) 3142 VM_OBJECT_ASSERT_WLOCKED(m->object); 3143} 3144 3145void 3146vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3147{ 3148 3149 if ((bits & PGA_WRITEABLE) == 0) 3150 return; 3151 3152 /* 3153 * The PGA_WRITEABLE flag can only be set if the page is 3154 * managed, is exclusively busied or the object is locked. 3155 * Currently, this flag is only set by pmap_enter(). 3156 */ 3157 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3158 ("PGA_WRITEABLE on unmanaged page")); 3159 if (!vm_page_xbusied(m)) 3160 VM_OBJECT_ASSERT_LOCKED(m->object); 3161} 3162#endif 3163 3164#include "opt_ddb.h" 3165#ifdef DDB 3166#include <sys/kernel.h> 3167 3168#include <ddb/ddb.h> 3169 3170DB_SHOW_COMMAND(page, vm_page_print_page_info) 3171{ 3172 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3173 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3174 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3175 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3176 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3177 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3178 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3179 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3180 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3181 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3182} 3183 3184DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3185{ 3186 int dom; 3187 3188 db_printf("pq_free %d pq_cache %d\n", 3189 cnt.v_free_count, cnt.v_cache_count); 3190 for (dom = 0; dom < vm_ndomains; dom++) { 3191 db_printf( 3192 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n", 3193 dom, 3194 vm_dom[dom].vmd_page_count, 3195 vm_dom[dom].vmd_free_count, 3196 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3197 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3198 vm_dom[dom].vmd_pass); 3199 } 3200} 3201 3202DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3203{ 3204 vm_page_t m; 3205 boolean_t phys; 3206 3207 if (!have_addr) { 3208 db_printf("show pginfo addr\n"); 3209 return; 3210 } 3211 3212 phys = strchr(modif, 'p') != NULL; 3213 if (phys) 3214 m = PHYS_TO_VM_PAGE(addr); 3215 else 3216 m = (vm_page_t)addr; 3217 db_printf( 3218 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3219 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3220 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3221 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3222 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3223} 3224#endif /* DDB */ 3225