vm_page.c revision 283598
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 283598 2015-05-27 08:30:16Z kib $"); 86 87#include "opt_vm.h" 88 89#include <sys/param.h> 90#include <sys/systm.h> 91#include <sys/lock.h> 92#include <sys/kernel.h> 93#include <sys/limits.h> 94#include <sys/malloc.h> 95#include <sys/mman.h> 96#include <sys/msgbuf.h> 97#include <sys/mutex.h> 98#include <sys/proc.h> 99#include <sys/rwlock.h> 100#include <sys/sysctl.h> 101#include <sys/vmmeter.h> 102#include <sys/vnode.h> 103 104#include <vm/vm.h> 105#include <vm/pmap.h> 106#include <vm/vm_param.h> 107#include <vm/vm_kern.h> 108#include <vm/vm_object.h> 109#include <vm/vm_page.h> 110#include <vm/vm_pageout.h> 111#include <vm/vm_pager.h> 112#include <vm/vm_phys.h> 113#include <vm/vm_radix.h> 114#include <vm/vm_reserv.h> 115#include <vm/vm_extern.h> 116#include <vm/uma.h> 117#include <vm/uma_int.h> 118 119#include <machine/md_var.h> 120 121/* 122 * Associated with page of user-allocatable memory is a 123 * page structure. 124 */ 125 126struct vm_domain vm_dom[MAXMEMDOM]; 127struct mtx_padalign vm_page_queue_free_mtx; 128 129struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 130 131vm_page_t vm_page_array; 132long vm_page_array_size; 133long first_page; 134int vm_page_zero_count; 135 136static int boot_pages = UMA_BOOT_PAGES; 137TUNABLE_INT("vm.boot_pages", &boot_pages); 138SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 139 "number of pages allocated for bootstrapping the VM system"); 140 141static int pa_tryrelock_restart; 142SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 143 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 144 145static uma_zone_t fakepg_zone; 146 147static struct vnode *vm_page_alloc_init(vm_page_t m); 148static void vm_page_cache_turn_free(vm_page_t m); 149static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 150static void vm_page_enqueue(int queue, vm_page_t m); 151static void vm_page_init_fakepg(void *dummy); 152static int vm_page_insert_after(vm_page_t m, vm_object_t object, 153 vm_pindex_t pindex, vm_page_t mpred); 154static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 155 vm_page_t mpred); 156 157SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 158 159static void 160vm_page_init_fakepg(void *dummy) 161{ 162 163 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 164 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 165} 166 167/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 168#if PAGE_SIZE == 32768 169#ifdef CTASSERT 170CTASSERT(sizeof(u_long) >= 8); 171#endif 172#endif 173 174/* 175 * Try to acquire a physical address lock while a pmap is locked. If we 176 * fail to trylock we unlock and lock the pmap directly and cache the 177 * locked pa in *locked. The caller should then restart their loop in case 178 * the virtual to physical mapping has changed. 179 */ 180int 181vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 182{ 183 vm_paddr_t lockpa; 184 185 lockpa = *locked; 186 *locked = pa; 187 if (lockpa) { 188 PA_LOCK_ASSERT(lockpa, MA_OWNED); 189 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 190 return (0); 191 PA_UNLOCK(lockpa); 192 } 193 if (PA_TRYLOCK(pa)) 194 return (0); 195 PMAP_UNLOCK(pmap); 196 atomic_add_int(&pa_tryrelock_restart, 1); 197 PA_LOCK(pa); 198 PMAP_LOCK(pmap); 199 return (EAGAIN); 200} 201 202/* 203 * vm_set_page_size: 204 * 205 * Sets the page size, perhaps based upon the memory 206 * size. Must be called before any use of page-size 207 * dependent functions. 208 */ 209void 210vm_set_page_size(void) 211{ 212 if (cnt.v_page_size == 0) 213 cnt.v_page_size = PAGE_SIZE; 214 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 215 panic("vm_set_page_size: page size not a power of two"); 216} 217 218/* 219 * vm_page_blacklist_lookup: 220 * 221 * See if a physical address in this page has been listed 222 * in the blacklist tunable. Entries in the tunable are 223 * separated by spaces or commas. If an invalid integer is 224 * encountered then the rest of the string is skipped. 225 */ 226static int 227vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 228{ 229 vm_paddr_t bad; 230 char *cp, *pos; 231 232 for (pos = list; *pos != '\0'; pos = cp) { 233 bad = strtoq(pos, &cp, 0); 234 if (*cp != '\0') { 235 if (*cp == ' ' || *cp == ',') { 236 cp++; 237 if (cp == pos) 238 continue; 239 } else 240 break; 241 } 242 if (pa == trunc_page(bad)) 243 return (1); 244 } 245 return (0); 246} 247 248static void 249vm_page_domain_init(struct vm_domain *vmd) 250{ 251 struct vm_pagequeue *pq; 252 int i; 253 254 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 255 "vm inactive pagequeue"; 256 *__DECONST(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(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_pindex_t sidx; 985 vm_object_t sobj; 986 vm_page_t msucc; 987 988 VM_OBJECT_ASSERT_WLOCKED(object); 989 KASSERT(m->object == NULL, 990 ("vm_page_insert_after: page already inserted")); 991 if (mpred != NULL) { 992 KASSERT(mpred->object == object, 993 ("vm_page_insert_after: object doesn't contain mpred")); 994 KASSERT(mpred->pindex < pindex, 995 ("vm_page_insert_after: mpred doesn't precede pindex")); 996 msucc = TAILQ_NEXT(mpred, listq); 997 } else 998 msucc = TAILQ_FIRST(&object->memq); 999 if (msucc != NULL) 1000 KASSERT(msucc->pindex > pindex, 1001 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1002 1003 /* 1004 * Record the object/offset pair in this page 1005 */ 1006 sobj = m->object; 1007 sidx = m->pindex; 1008 m->object = object; 1009 m->pindex = pindex; 1010 1011 /* 1012 * Now link into the object's ordered list of backed pages. 1013 */ 1014 if (vm_radix_insert(&object->rtree, m)) { 1015 m->object = sobj; 1016 m->pindex = sidx; 1017 return (1); 1018 } 1019 vm_page_insert_radixdone(m, object, mpred); 1020 return (0); 1021} 1022 1023/* 1024 * vm_page_insert_radixdone: 1025 * 1026 * Complete page "m" insertion into the specified object after the 1027 * radix trie hooking. 1028 * 1029 * The page "mpred" must precede the offset "m->pindex" within the 1030 * specified object. 1031 * 1032 * The object must be locked. 1033 */ 1034static void 1035vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1036{ 1037 1038 VM_OBJECT_ASSERT_WLOCKED(object); 1039 KASSERT(object != NULL && m->object == object, 1040 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1041 if (mpred != NULL) { 1042 KASSERT(mpred->object == object, 1043 ("vm_page_insert_after: object doesn't contain mpred")); 1044 KASSERT(mpred->pindex < m->pindex, 1045 ("vm_page_insert_after: mpred doesn't precede pindex")); 1046 } 1047 1048 if (mpred != NULL) 1049 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1050 else 1051 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1052 1053 /* 1054 * Show that the object has one more resident page. 1055 */ 1056 object->resident_page_count++; 1057 1058 /* 1059 * Hold the vnode until the last page is released. 1060 */ 1061 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1062 vhold(object->handle); 1063 1064 /* 1065 * Since we are inserting a new and possibly dirty page, 1066 * update the object's OBJ_MIGHTBEDIRTY flag. 1067 */ 1068 if (pmap_page_is_write_mapped(m)) 1069 vm_object_set_writeable_dirty(object); 1070} 1071 1072/* 1073 * vm_page_remove: 1074 * 1075 * Removes the given mem entry from the object/offset-page 1076 * table and the object page list, but do not invalidate/terminate 1077 * the backing store. 1078 * 1079 * The object must be locked. The page must be locked if it is managed. 1080 */ 1081void 1082vm_page_remove(vm_page_t m) 1083{ 1084 vm_object_t object; 1085 boolean_t lockacq; 1086 1087 if ((m->oflags & VPO_UNMANAGED) == 0) 1088 vm_page_lock_assert(m, MA_OWNED); 1089 if ((object = m->object) == NULL) 1090 return; 1091 VM_OBJECT_ASSERT_WLOCKED(object); 1092 if (vm_page_xbusied(m)) { 1093 lockacq = FALSE; 1094 if ((m->oflags & VPO_UNMANAGED) != 0 && 1095 !mtx_owned(vm_page_lockptr(m))) { 1096 lockacq = TRUE; 1097 vm_page_lock(m); 1098 } 1099 vm_page_flash(m); 1100 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1101 if (lockacq) 1102 vm_page_unlock(m); 1103 } 1104 1105 /* 1106 * Now remove from the object's list of backed pages. 1107 */ 1108 vm_radix_remove(&object->rtree, m->pindex); 1109 TAILQ_REMOVE(&object->memq, m, listq); 1110 1111 /* 1112 * And show that the object has one fewer resident page. 1113 */ 1114 object->resident_page_count--; 1115 1116 /* 1117 * The vnode may now be recycled. 1118 */ 1119 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1120 vdrop(object->handle); 1121 1122 m->object = NULL; 1123} 1124 1125/* 1126 * vm_page_lookup: 1127 * 1128 * Returns the page associated with the object/offset 1129 * pair specified; if none is found, NULL is returned. 1130 * 1131 * The object must be locked. 1132 */ 1133vm_page_t 1134vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1135{ 1136 1137 VM_OBJECT_ASSERT_LOCKED(object); 1138 return (vm_radix_lookup(&object->rtree, pindex)); 1139} 1140 1141/* 1142 * vm_page_find_least: 1143 * 1144 * Returns the page associated with the object with least pindex 1145 * greater than or equal to the parameter pindex, or NULL. 1146 * 1147 * The object must be locked. 1148 */ 1149vm_page_t 1150vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1151{ 1152 vm_page_t m; 1153 1154 VM_OBJECT_ASSERT_LOCKED(object); 1155 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1156 m = vm_radix_lookup_ge(&object->rtree, pindex); 1157 return (m); 1158} 1159 1160/* 1161 * Returns the given page's successor (by pindex) within the object if it is 1162 * resident; if none is found, NULL is returned. 1163 * 1164 * The object must be locked. 1165 */ 1166vm_page_t 1167vm_page_next(vm_page_t m) 1168{ 1169 vm_page_t next; 1170 1171 VM_OBJECT_ASSERT_WLOCKED(m->object); 1172 if ((next = TAILQ_NEXT(m, listq)) != NULL && 1173 next->pindex != m->pindex + 1) 1174 next = NULL; 1175 return (next); 1176} 1177 1178/* 1179 * Returns the given page's predecessor (by pindex) within the object if it is 1180 * resident; if none is found, NULL is returned. 1181 * 1182 * The object must be locked. 1183 */ 1184vm_page_t 1185vm_page_prev(vm_page_t m) 1186{ 1187 vm_page_t prev; 1188 1189 VM_OBJECT_ASSERT_WLOCKED(m->object); 1190 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1191 prev->pindex != m->pindex - 1) 1192 prev = NULL; 1193 return (prev); 1194} 1195 1196/* 1197 * Uses the page mnew as a replacement for an existing page at index 1198 * pindex which must be already present in the object. 1199 * 1200 * The existing page must not be on a paging queue. 1201 */ 1202vm_page_t 1203vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1204{ 1205 vm_page_t mold, mpred; 1206 1207 VM_OBJECT_ASSERT_WLOCKED(object); 1208 1209 /* 1210 * This function mostly follows vm_page_insert() and 1211 * vm_page_remove() without the radix, object count and vnode 1212 * dance. Double check such functions for more comments. 1213 */ 1214 mpred = vm_radix_lookup(&object->rtree, pindex); 1215 KASSERT(mpred != NULL, 1216 ("vm_page_replace: replacing page not present with pindex")); 1217 mpred = TAILQ_PREV(mpred, respgs, listq); 1218 if (mpred != NULL) 1219 KASSERT(mpred->pindex < pindex, 1220 ("vm_page_insert_after: mpred doesn't precede pindex")); 1221 1222 mnew->object = object; 1223 mnew->pindex = pindex; 1224 mold = vm_radix_replace(&object->rtree, mnew); 1225 KASSERT(mold->queue == PQ_NONE, 1226 ("vm_page_replace: mold is on a paging queue")); 1227 1228 /* Detach the old page from the resident tailq. */ 1229 TAILQ_REMOVE(&object->memq, mold, listq); 1230 1231 mold->object = NULL; 1232 vm_page_xunbusy(mold); 1233 1234 /* Insert the new page in the resident tailq. */ 1235 if (mpred != NULL) 1236 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq); 1237 else 1238 TAILQ_INSERT_HEAD(&object->memq, mnew, listq); 1239 if (pmap_page_is_write_mapped(mnew)) 1240 vm_object_set_writeable_dirty(object); 1241 return (mold); 1242} 1243 1244/* 1245 * vm_page_rename: 1246 * 1247 * Move the given memory entry from its 1248 * current object to the specified target object/offset. 1249 * 1250 * Note: swap associated with the page must be invalidated by the move. We 1251 * have to do this for several reasons: (1) we aren't freeing the 1252 * page, (2) we are dirtying the page, (3) the VM system is probably 1253 * moving the page from object A to B, and will then later move 1254 * the backing store from A to B and we can't have a conflict. 1255 * 1256 * Note: we *always* dirty the page. It is necessary both for the 1257 * fact that we moved it, and because we may be invalidating 1258 * swap. If the page is on the cache, we have to deactivate it 1259 * or vm_page_dirty() will panic. Dirty pages are not allowed 1260 * on the cache. 1261 * 1262 * The objects must be locked. 1263 */ 1264int 1265vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1266{ 1267 vm_page_t mpred; 1268 vm_pindex_t opidx; 1269 1270 VM_OBJECT_ASSERT_WLOCKED(new_object); 1271 1272 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1273 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1274 ("vm_page_rename: pindex already renamed")); 1275 1276 /* 1277 * Create a custom version of vm_page_insert() which does not depend 1278 * by m_prev and can cheat on the implementation aspects of the 1279 * function. 1280 */ 1281 opidx = m->pindex; 1282 m->pindex = new_pindex; 1283 if (vm_radix_insert(&new_object->rtree, m)) { 1284 m->pindex = opidx; 1285 return (1); 1286 } 1287 1288 /* 1289 * The operation cannot fail anymore. The removal must happen before 1290 * the listq iterator is tainted. 1291 */ 1292 m->pindex = opidx; 1293 vm_page_lock(m); 1294 vm_page_remove(m); 1295 1296 /* Return back to the new pindex to complete vm_page_insert(). */ 1297 m->pindex = new_pindex; 1298 m->object = new_object; 1299 vm_page_unlock(m); 1300 vm_page_insert_radixdone(m, new_object, mpred); 1301 vm_page_dirty(m); 1302 return (0); 1303} 1304 1305/* 1306 * Convert all of the given object's cached pages that have a 1307 * pindex within the given range into free pages. If the value 1308 * zero is given for "end", then the range's upper bound is 1309 * infinity. If the given object is backed by a vnode and it 1310 * transitions from having one or more cached pages to none, the 1311 * vnode's hold count is reduced. 1312 */ 1313void 1314vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1315{ 1316 vm_page_t m; 1317 boolean_t empty; 1318 1319 mtx_lock(&vm_page_queue_free_mtx); 1320 if (__predict_false(vm_radix_is_empty(&object->cache))) { 1321 mtx_unlock(&vm_page_queue_free_mtx); 1322 return; 1323 } 1324 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) { 1325 if (end != 0 && m->pindex >= end) 1326 break; 1327 vm_radix_remove(&object->cache, m->pindex); 1328 vm_page_cache_turn_free(m); 1329 } 1330 empty = vm_radix_is_empty(&object->cache); 1331 mtx_unlock(&vm_page_queue_free_mtx); 1332 if (object->type == OBJT_VNODE && empty) 1333 vdrop(object->handle); 1334} 1335 1336/* 1337 * Returns the cached page that is associated with the given 1338 * object and offset. If, however, none exists, returns NULL. 1339 * 1340 * The free page queue must be locked. 1341 */ 1342static inline vm_page_t 1343vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1344{ 1345 1346 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1347 return (vm_radix_lookup(&object->cache, pindex)); 1348} 1349 1350/* 1351 * Remove the given cached page from its containing object's 1352 * collection of cached pages. 1353 * 1354 * The free page queue must be locked. 1355 */ 1356static void 1357vm_page_cache_remove(vm_page_t m) 1358{ 1359 1360 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1361 KASSERT((m->flags & PG_CACHED) != 0, 1362 ("vm_page_cache_remove: page %p is not cached", m)); 1363 vm_radix_remove(&m->object->cache, m->pindex); 1364 m->object = NULL; 1365 cnt.v_cache_count--; 1366} 1367 1368/* 1369 * Transfer all of the cached pages with offset greater than or 1370 * equal to 'offidxstart' from the original object's cache to the 1371 * new object's cache. However, any cached pages with offset 1372 * greater than or equal to the new object's size are kept in the 1373 * original object. Initially, the new object's cache must be 1374 * empty. Offset 'offidxstart' in the original object must 1375 * correspond to offset zero in the new object. 1376 * 1377 * The new object must be locked. 1378 */ 1379void 1380vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1381 vm_object_t new_object) 1382{ 1383 vm_page_t m; 1384 1385 /* 1386 * Insertion into an object's collection of cached pages 1387 * requires the object to be locked. In contrast, removal does 1388 * not. 1389 */ 1390 VM_OBJECT_ASSERT_WLOCKED(new_object); 1391 KASSERT(vm_radix_is_empty(&new_object->cache), 1392 ("vm_page_cache_transfer: object %p has cached pages", 1393 new_object)); 1394 mtx_lock(&vm_page_queue_free_mtx); 1395 while ((m = vm_radix_lookup_ge(&orig_object->cache, 1396 offidxstart)) != NULL) { 1397 /* 1398 * Transfer all of the pages with offset greater than or 1399 * equal to 'offidxstart' from the original object's 1400 * cache to the new object's cache. 1401 */ 1402 if ((m->pindex - offidxstart) >= new_object->size) 1403 break; 1404 vm_radix_remove(&orig_object->cache, m->pindex); 1405 /* Update the page's object and offset. */ 1406 m->object = new_object; 1407 m->pindex -= offidxstart; 1408 if (vm_radix_insert(&new_object->cache, m)) 1409 vm_page_cache_turn_free(m); 1410 } 1411 mtx_unlock(&vm_page_queue_free_mtx); 1412} 1413 1414/* 1415 * Returns TRUE if a cached page is associated with the given object and 1416 * offset, and FALSE otherwise. 1417 * 1418 * The object must be locked. 1419 */ 1420boolean_t 1421vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1422{ 1423 vm_page_t m; 1424 1425 /* 1426 * Insertion into an object's collection of cached pages requires the 1427 * object to be locked. Therefore, if the object is locked and the 1428 * object's collection is empty, there is no need to acquire the free 1429 * page queues lock in order to prove that the specified page doesn't 1430 * exist. 1431 */ 1432 VM_OBJECT_ASSERT_WLOCKED(object); 1433 if (__predict_true(vm_object_cache_is_empty(object))) 1434 return (FALSE); 1435 mtx_lock(&vm_page_queue_free_mtx); 1436 m = vm_page_cache_lookup(object, pindex); 1437 mtx_unlock(&vm_page_queue_free_mtx); 1438 return (m != NULL); 1439} 1440 1441/* 1442 * vm_page_alloc: 1443 * 1444 * Allocate and return a page that is associated with the specified 1445 * object and offset pair. By default, this page is exclusive busied. 1446 * 1447 * The caller must always specify an allocation class. 1448 * 1449 * allocation classes: 1450 * VM_ALLOC_NORMAL normal process request 1451 * VM_ALLOC_SYSTEM system *really* needs a page 1452 * VM_ALLOC_INTERRUPT interrupt time request 1453 * 1454 * optional allocation flags: 1455 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1456 * intends to allocate 1457 * VM_ALLOC_IFCACHED return page only if it is cached 1458 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1459 * is cached 1460 * VM_ALLOC_NOBUSY do not exclusive busy the page 1461 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1462 * VM_ALLOC_NOOBJ page is not associated with an object and 1463 * should not be exclusive busy 1464 * VM_ALLOC_SBUSY shared busy the allocated page 1465 * VM_ALLOC_WIRED wire the allocated page 1466 * VM_ALLOC_ZERO prefer a zeroed page 1467 * 1468 * This routine may not sleep. 1469 */ 1470vm_page_t 1471vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1472{ 1473 struct vnode *vp = NULL; 1474 vm_object_t m_object; 1475 vm_page_t m, mpred; 1476 int flags, req_class; 1477 1478 mpred = 0; /* XXX: pacify gcc */ 1479 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1480 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1481 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1482 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1483 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1484 req)); 1485 if (object != NULL) 1486 VM_OBJECT_ASSERT_WLOCKED(object); 1487 1488 req_class = req & VM_ALLOC_CLASS_MASK; 1489 1490 /* 1491 * The page daemon is allowed to dig deeper into the free page list. 1492 */ 1493 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1494 req_class = VM_ALLOC_SYSTEM; 1495 1496 if (object != NULL) { 1497 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1498 KASSERT(mpred == NULL || mpred->pindex != pindex, 1499 ("vm_page_alloc: pindex already allocated")); 1500 } 1501 1502 /* 1503 * The page allocation request can came from consumers which already 1504 * hold the free page queue mutex, like vm_page_insert() in 1505 * vm_page_cache(). 1506 */ 1507 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1508 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1509 (req_class == VM_ALLOC_SYSTEM && 1510 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1511 (req_class == VM_ALLOC_INTERRUPT && 1512 cnt.v_free_count + cnt.v_cache_count > 0)) { 1513 /* 1514 * Allocate from the free queue if the number of free pages 1515 * exceeds the minimum for the request class. 1516 */ 1517 if (object != NULL && 1518 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1519 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1520 mtx_unlock(&vm_page_queue_free_mtx); 1521 return (NULL); 1522 } 1523 if (vm_phys_unfree_page(m)) 1524 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1525#if VM_NRESERVLEVEL > 0 1526 else if (!vm_reserv_reactivate_page(m)) 1527#else 1528 else 1529#endif 1530 panic("vm_page_alloc: cache page %p is missing" 1531 " from the free queue", m); 1532 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1533 mtx_unlock(&vm_page_queue_free_mtx); 1534 return (NULL); 1535#if VM_NRESERVLEVEL > 0 1536 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1537 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1538 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) { 1539#else 1540 } else { 1541#endif 1542 m = vm_phys_alloc_pages(object != NULL ? 1543 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1544#if VM_NRESERVLEVEL > 0 1545 if (m == NULL && vm_reserv_reclaim_inactive()) { 1546 m = vm_phys_alloc_pages(object != NULL ? 1547 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1548 0); 1549 } 1550#endif 1551 } 1552 } else { 1553 /* 1554 * Not allocatable, give up. 1555 */ 1556 mtx_unlock(&vm_page_queue_free_mtx); 1557 atomic_add_int(&vm_pageout_deficit, 1558 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1559 pagedaemon_wakeup(); 1560 return (NULL); 1561 } 1562 1563 /* 1564 * At this point we had better have found a good page. 1565 */ 1566 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1567 KASSERT(m->queue == PQ_NONE, 1568 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1569 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1570 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1571 KASSERT(!vm_page_sbusied(m), 1572 ("vm_page_alloc: page %p is busy", m)); 1573 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1574 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1575 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1576 pmap_page_get_memattr(m))); 1577 if ((m->flags & PG_CACHED) != 0) { 1578 KASSERT((m->flags & PG_ZERO) == 0, 1579 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1580 KASSERT(m->valid != 0, 1581 ("vm_page_alloc: cached page %p is invalid", m)); 1582 if (m->object == object && m->pindex == pindex) 1583 cnt.v_reactivated++; 1584 else 1585 m->valid = 0; 1586 m_object = m->object; 1587 vm_page_cache_remove(m); 1588 if (m_object->type == OBJT_VNODE && 1589 vm_object_cache_is_empty(m_object)) 1590 vp = m_object->handle; 1591 } else { 1592 KASSERT(VM_PAGE_IS_FREE(m), 1593 ("vm_page_alloc: page %p is not free", m)); 1594 KASSERT(m->valid == 0, 1595 ("vm_page_alloc: free page %p is valid", m)); 1596 vm_phys_freecnt_adj(m, -1); 1597 } 1598 1599 /* 1600 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag 1601 * must be cleared before the free page queues lock is released. 1602 */ 1603 flags = 0; 1604 if (m->flags & PG_ZERO) { 1605 vm_page_zero_count--; 1606 if (req & VM_ALLOC_ZERO) 1607 flags = PG_ZERO; 1608 } 1609 if (req & VM_ALLOC_NODUMP) 1610 flags |= PG_NODUMP; 1611 m->flags = flags; 1612 mtx_unlock(&vm_page_queue_free_mtx); 1613 m->aflags = 0; 1614 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1615 VPO_UNMANAGED : 0; 1616 m->busy_lock = VPB_UNBUSIED; 1617 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1618 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1619 if ((req & VM_ALLOC_SBUSY) != 0) 1620 m->busy_lock = VPB_SHARERS_WORD(1); 1621 if (req & VM_ALLOC_WIRED) { 1622 /* 1623 * The page lock is not required for wiring a page until that 1624 * page is inserted into the object. 1625 */ 1626 atomic_add_int(&cnt.v_wire_count, 1); 1627 m->wire_count = 1; 1628 } 1629 m->act_count = 0; 1630 1631 if (object != NULL) { 1632 if (vm_page_insert_after(m, object, pindex, mpred)) { 1633 /* See the comment below about hold count. */ 1634 if (vp != NULL) 1635 vdrop(vp); 1636 pagedaemon_wakeup(); 1637 if (req & VM_ALLOC_WIRED) { 1638 atomic_subtract_int(&cnt.v_wire_count, 1); 1639 m->wire_count = 0; 1640 } 1641 m->object = NULL; 1642 m->oflags = VPO_UNMANAGED; 1643 vm_page_free(m); 1644 return (NULL); 1645 } 1646 1647 /* Ignore device objects; the pager sets "memattr" for them. */ 1648 if (object->memattr != VM_MEMATTR_DEFAULT && 1649 (object->flags & OBJ_FICTITIOUS) == 0) 1650 pmap_page_set_memattr(m, object->memattr); 1651 } else 1652 m->pindex = pindex; 1653 1654 /* 1655 * The following call to vdrop() must come after the above call 1656 * to vm_page_insert() in case both affect the same object and 1657 * vnode. Otherwise, the affected vnode's hold count could 1658 * temporarily become zero. 1659 */ 1660 if (vp != NULL) 1661 vdrop(vp); 1662 1663 /* 1664 * Don't wakeup too often - wakeup the pageout daemon when 1665 * we would be nearly out of memory. 1666 */ 1667 if (vm_paging_needed()) 1668 pagedaemon_wakeup(); 1669 1670 return (m); 1671} 1672 1673static void 1674vm_page_alloc_contig_vdrop(struct spglist *lst) 1675{ 1676 1677 while (!SLIST_EMPTY(lst)) { 1678 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv); 1679 SLIST_REMOVE_HEAD(lst, plinks.s.ss); 1680 } 1681} 1682 1683/* 1684 * vm_page_alloc_contig: 1685 * 1686 * Allocate a contiguous set of physical pages of the given size "npages" 1687 * from the free lists. All of the physical pages must be at or above 1688 * the given physical address "low" and below the given physical address 1689 * "high". The given value "alignment" determines the alignment of the 1690 * first physical page in the set. If the given value "boundary" is 1691 * non-zero, then the set of physical pages cannot cross any physical 1692 * address boundary that is a multiple of that value. Both "alignment" 1693 * and "boundary" must be a power of two. 1694 * 1695 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1696 * then the memory attribute setting for the physical pages is configured 1697 * to the object's memory attribute setting. Otherwise, the memory 1698 * attribute setting for the physical pages is configured to "memattr", 1699 * overriding the object's memory attribute setting. However, if the 1700 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1701 * memory attribute setting for the physical pages cannot be configured 1702 * to VM_MEMATTR_DEFAULT. 1703 * 1704 * The caller must always specify an allocation class. 1705 * 1706 * allocation classes: 1707 * VM_ALLOC_NORMAL normal process request 1708 * VM_ALLOC_SYSTEM system *really* needs a page 1709 * VM_ALLOC_INTERRUPT interrupt time request 1710 * 1711 * optional allocation flags: 1712 * VM_ALLOC_NOBUSY do not exclusive busy the page 1713 * VM_ALLOC_NOOBJ page is not associated with an object and 1714 * should not be exclusive busy 1715 * VM_ALLOC_SBUSY shared busy the allocated page 1716 * VM_ALLOC_WIRED wire the allocated page 1717 * VM_ALLOC_ZERO prefer a zeroed page 1718 * 1719 * This routine may not sleep. 1720 */ 1721vm_page_t 1722vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1723 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1724 vm_paddr_t boundary, vm_memattr_t memattr) 1725{ 1726 struct vnode *drop; 1727 struct spglist deferred_vdrop_list; 1728 vm_page_t m, m_tmp, m_ret; 1729 u_int flags, oflags; 1730 int req_class; 1731 1732 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1733 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1734 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1735 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1736 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1737 req)); 1738 if (object != NULL) { 1739 VM_OBJECT_ASSERT_WLOCKED(object); 1740 KASSERT(object->type == OBJT_PHYS, 1741 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1742 object)); 1743 } 1744 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1745 req_class = req & VM_ALLOC_CLASS_MASK; 1746 1747 /* 1748 * The page daemon is allowed to dig deeper into the free page list. 1749 */ 1750 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1751 req_class = VM_ALLOC_SYSTEM; 1752 1753 SLIST_INIT(&deferred_vdrop_list); 1754 mtx_lock(&vm_page_queue_free_mtx); 1755 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1756 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1757 cnt.v_free_count + cnt.v_cache_count >= npages + 1758 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1759 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1760#if VM_NRESERVLEVEL > 0 1761retry: 1762 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1763 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1764 low, high, alignment, boundary)) == NULL) 1765#endif 1766 m_ret = vm_phys_alloc_contig(npages, low, high, 1767 alignment, boundary); 1768 } else { 1769 mtx_unlock(&vm_page_queue_free_mtx); 1770 atomic_add_int(&vm_pageout_deficit, npages); 1771 pagedaemon_wakeup(); 1772 return (NULL); 1773 } 1774 if (m_ret != NULL) 1775 for (m = m_ret; m < &m_ret[npages]; m++) { 1776 drop = vm_page_alloc_init(m); 1777 if (drop != NULL) { 1778 /* 1779 * Enqueue the vnode for deferred vdrop(). 1780 */ 1781 m->plinks.s.pv = drop; 1782 SLIST_INSERT_HEAD(&deferred_vdrop_list, m, 1783 plinks.s.ss); 1784 } 1785 } 1786 else { 1787#if VM_NRESERVLEVEL > 0 1788 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1789 boundary)) 1790 goto retry; 1791#endif 1792 } 1793 mtx_unlock(&vm_page_queue_free_mtx); 1794 if (m_ret == NULL) 1795 return (NULL); 1796 1797 /* 1798 * Initialize the pages. Only the PG_ZERO flag is inherited. 1799 */ 1800 flags = 0; 1801 if ((req & VM_ALLOC_ZERO) != 0) 1802 flags = PG_ZERO; 1803 if ((req & VM_ALLOC_NODUMP) != 0) 1804 flags |= PG_NODUMP; 1805 if ((req & VM_ALLOC_WIRED) != 0) 1806 atomic_add_int(&cnt.v_wire_count, npages); 1807 oflags = VPO_UNMANAGED; 1808 if (object != NULL) { 1809 if (object->memattr != VM_MEMATTR_DEFAULT && 1810 memattr == VM_MEMATTR_DEFAULT) 1811 memattr = object->memattr; 1812 } 1813 for (m = m_ret; m < &m_ret[npages]; m++) { 1814 m->aflags = 0; 1815 m->flags = (m->flags | PG_NODUMP) & flags; 1816 m->busy_lock = VPB_UNBUSIED; 1817 if (object != NULL) { 1818 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1819 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1820 if ((req & VM_ALLOC_SBUSY) != 0) 1821 m->busy_lock = VPB_SHARERS_WORD(1); 1822 } 1823 if ((req & VM_ALLOC_WIRED) != 0) 1824 m->wire_count = 1; 1825 /* Unmanaged pages don't use "act_count". */ 1826 m->oflags = oflags; 1827 if (object != NULL) { 1828 if (vm_page_insert(m, object, pindex)) { 1829 vm_page_alloc_contig_vdrop( 1830 &deferred_vdrop_list); 1831 if (vm_paging_needed()) 1832 pagedaemon_wakeup(); 1833 if ((req & VM_ALLOC_WIRED) != 0) 1834 atomic_subtract_int(&cnt.v_wire_count, 1835 npages); 1836 for (m_tmp = m, m = m_ret; 1837 m < &m_ret[npages]; m++) { 1838 if ((req & VM_ALLOC_WIRED) != 0) 1839 m->wire_count = 0; 1840 if (m >= m_tmp) 1841 m->object = NULL; 1842 vm_page_free(m); 1843 } 1844 return (NULL); 1845 } 1846 } else 1847 m->pindex = pindex; 1848 if (memattr != VM_MEMATTR_DEFAULT) 1849 pmap_page_set_memattr(m, memattr); 1850 pindex++; 1851 } 1852 vm_page_alloc_contig_vdrop(&deferred_vdrop_list); 1853 if (vm_paging_needed()) 1854 pagedaemon_wakeup(); 1855 return (m_ret); 1856} 1857 1858/* 1859 * Initialize a page that has been freshly dequeued from a freelist. 1860 * The caller has to drop the vnode returned, if it is not NULL. 1861 * 1862 * This function may only be used to initialize unmanaged pages. 1863 * 1864 * To be called with vm_page_queue_free_mtx held. 1865 */ 1866static struct vnode * 1867vm_page_alloc_init(vm_page_t m) 1868{ 1869 struct vnode *drop; 1870 vm_object_t m_object; 1871 1872 KASSERT(m->queue == PQ_NONE, 1873 ("vm_page_alloc_init: page %p has unexpected queue %d", 1874 m, m->queue)); 1875 KASSERT(m->wire_count == 0, 1876 ("vm_page_alloc_init: page %p is wired", m)); 1877 KASSERT(m->hold_count == 0, 1878 ("vm_page_alloc_init: page %p is held", m)); 1879 KASSERT(!vm_page_sbusied(m), 1880 ("vm_page_alloc_init: page %p is busy", m)); 1881 KASSERT(m->dirty == 0, 1882 ("vm_page_alloc_init: page %p is dirty", m)); 1883 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1884 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1885 m, pmap_page_get_memattr(m))); 1886 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1887 drop = NULL; 1888 if ((m->flags & PG_CACHED) != 0) { 1889 KASSERT((m->flags & PG_ZERO) == 0, 1890 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1891 m->valid = 0; 1892 m_object = m->object; 1893 vm_page_cache_remove(m); 1894 if (m_object->type == OBJT_VNODE && 1895 vm_object_cache_is_empty(m_object)) 1896 drop = m_object->handle; 1897 } else { 1898 KASSERT(VM_PAGE_IS_FREE(m), 1899 ("vm_page_alloc_init: page %p is not free", m)); 1900 KASSERT(m->valid == 0, 1901 ("vm_page_alloc_init: free page %p is valid", m)); 1902 vm_phys_freecnt_adj(m, -1); 1903 if ((m->flags & PG_ZERO) != 0) 1904 vm_page_zero_count--; 1905 } 1906 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1907 m->flags &= PG_ZERO; 1908 return (drop); 1909} 1910 1911/* 1912 * vm_page_alloc_freelist: 1913 * 1914 * Allocate a physical page from the specified free page list. 1915 * 1916 * The caller must always specify an allocation class. 1917 * 1918 * allocation classes: 1919 * VM_ALLOC_NORMAL normal process request 1920 * VM_ALLOC_SYSTEM system *really* needs a page 1921 * VM_ALLOC_INTERRUPT interrupt time request 1922 * 1923 * optional allocation flags: 1924 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1925 * intends to allocate 1926 * VM_ALLOC_WIRED wire the allocated page 1927 * VM_ALLOC_ZERO prefer a zeroed page 1928 * 1929 * This routine may not sleep. 1930 */ 1931vm_page_t 1932vm_page_alloc_freelist(int flind, int req) 1933{ 1934 struct vnode *drop; 1935 vm_page_t m; 1936 u_int flags; 1937 int req_class; 1938 1939 req_class = req & VM_ALLOC_CLASS_MASK; 1940 1941 /* 1942 * The page daemon is allowed to dig deeper into the free page list. 1943 */ 1944 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1945 req_class = VM_ALLOC_SYSTEM; 1946 1947 /* 1948 * Do not allocate reserved pages unless the req has asked for it. 1949 */ 1950 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1951 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1952 (req_class == VM_ALLOC_SYSTEM && 1953 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1954 (req_class == VM_ALLOC_INTERRUPT && 1955 cnt.v_free_count + cnt.v_cache_count > 0)) 1956 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1957 else { 1958 mtx_unlock(&vm_page_queue_free_mtx); 1959 atomic_add_int(&vm_pageout_deficit, 1960 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1961 pagedaemon_wakeup(); 1962 return (NULL); 1963 } 1964 if (m == NULL) { 1965 mtx_unlock(&vm_page_queue_free_mtx); 1966 return (NULL); 1967 } 1968 drop = vm_page_alloc_init(m); 1969 mtx_unlock(&vm_page_queue_free_mtx); 1970 1971 /* 1972 * Initialize the page. Only the PG_ZERO flag is inherited. 1973 */ 1974 m->aflags = 0; 1975 flags = 0; 1976 if ((req & VM_ALLOC_ZERO) != 0) 1977 flags = PG_ZERO; 1978 m->flags &= flags; 1979 if ((req & VM_ALLOC_WIRED) != 0) { 1980 /* 1981 * The page lock is not required for wiring a page that does 1982 * not belong to an object. 1983 */ 1984 atomic_add_int(&cnt.v_wire_count, 1); 1985 m->wire_count = 1; 1986 } 1987 /* Unmanaged pages don't use "act_count". */ 1988 m->oflags = VPO_UNMANAGED; 1989 if (drop != NULL) 1990 vdrop(drop); 1991 if (vm_paging_needed()) 1992 pagedaemon_wakeup(); 1993 return (m); 1994} 1995 1996/* 1997 * vm_wait: (also see VM_WAIT macro) 1998 * 1999 * Sleep until free pages are available for allocation. 2000 * - Called in various places before memory allocations. 2001 */ 2002void 2003vm_wait(void) 2004{ 2005 2006 mtx_lock(&vm_page_queue_free_mtx); 2007 if (curproc == pageproc) { 2008 vm_pageout_pages_needed = 1; 2009 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2010 PDROP | PSWP, "VMWait", 0); 2011 } else { 2012 if (!vm_pages_needed) { 2013 vm_pages_needed = 1; 2014 wakeup(&vm_pages_needed); 2015 } 2016 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2017 "vmwait", 0); 2018 } 2019} 2020 2021/* 2022 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2023 * 2024 * Sleep until free pages are available for allocation. 2025 * - Called only in vm_fault so that processes page faulting 2026 * can be easily tracked. 2027 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2028 * processes will be able to grab memory first. Do not change 2029 * this balance without careful testing first. 2030 */ 2031void 2032vm_waitpfault(void) 2033{ 2034 2035 mtx_lock(&vm_page_queue_free_mtx); 2036 if (!vm_pages_needed) { 2037 vm_pages_needed = 1; 2038 wakeup(&vm_pages_needed); 2039 } 2040 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2041 "pfault", 0); 2042} 2043 2044struct vm_pagequeue * 2045vm_page_pagequeue(vm_page_t m) 2046{ 2047 2048 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2049} 2050 2051/* 2052 * vm_page_dequeue: 2053 * 2054 * Remove the given page from its current page queue. 2055 * 2056 * The page must be locked. 2057 */ 2058void 2059vm_page_dequeue(vm_page_t m) 2060{ 2061 struct vm_pagequeue *pq; 2062 2063 vm_page_lock_assert(m, MA_OWNED); 2064 KASSERT(m->queue != PQ_NONE, 2065 ("vm_page_dequeue: page %p is not queued", m)); 2066 pq = vm_page_pagequeue(m); 2067 vm_pagequeue_lock(pq); 2068 m->queue = PQ_NONE; 2069 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2070 vm_pagequeue_cnt_dec(pq); 2071 vm_pagequeue_unlock(pq); 2072} 2073 2074/* 2075 * vm_page_dequeue_locked: 2076 * 2077 * Remove the given page from its current page queue. 2078 * 2079 * The page and page queue must be locked. 2080 */ 2081void 2082vm_page_dequeue_locked(vm_page_t m) 2083{ 2084 struct vm_pagequeue *pq; 2085 2086 vm_page_lock_assert(m, MA_OWNED); 2087 pq = vm_page_pagequeue(m); 2088 vm_pagequeue_assert_locked(pq); 2089 m->queue = PQ_NONE; 2090 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2091 vm_pagequeue_cnt_dec(pq); 2092} 2093 2094/* 2095 * vm_page_enqueue: 2096 * 2097 * Add the given page to the specified page queue. 2098 * 2099 * The page must be locked. 2100 */ 2101static void 2102vm_page_enqueue(int queue, vm_page_t m) 2103{ 2104 struct vm_pagequeue *pq; 2105 2106 vm_page_lock_assert(m, MA_OWNED); 2107 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2108 vm_pagequeue_lock(pq); 2109 m->queue = queue; 2110 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2111 vm_pagequeue_cnt_inc(pq); 2112 vm_pagequeue_unlock(pq); 2113} 2114 2115/* 2116 * vm_page_requeue: 2117 * 2118 * Move the given page to the tail of its current page queue. 2119 * 2120 * The page must be locked. 2121 */ 2122void 2123vm_page_requeue(vm_page_t m) 2124{ 2125 struct vm_pagequeue *pq; 2126 2127 vm_page_lock_assert(m, MA_OWNED); 2128 KASSERT(m->queue != PQ_NONE, 2129 ("vm_page_requeue: page %p is not queued", m)); 2130 pq = vm_page_pagequeue(m); 2131 vm_pagequeue_lock(pq); 2132 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2133 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2134 vm_pagequeue_unlock(pq); 2135} 2136 2137/* 2138 * vm_page_requeue_locked: 2139 * 2140 * Move the given page to the tail of its current page queue. 2141 * 2142 * The page queue must be locked. 2143 */ 2144void 2145vm_page_requeue_locked(vm_page_t m) 2146{ 2147 struct vm_pagequeue *pq; 2148 2149 KASSERT(m->queue != PQ_NONE, 2150 ("vm_page_requeue_locked: page %p is not queued", m)); 2151 pq = vm_page_pagequeue(m); 2152 vm_pagequeue_assert_locked(pq); 2153 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2154 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2155} 2156 2157/* 2158 * vm_page_activate: 2159 * 2160 * Put the specified page on the active list (if appropriate). 2161 * Ensure that act_count is at least ACT_INIT but do not otherwise 2162 * mess with it. 2163 * 2164 * The page must be locked. 2165 */ 2166void 2167vm_page_activate(vm_page_t m) 2168{ 2169 int queue; 2170 2171 vm_page_lock_assert(m, MA_OWNED); 2172 if ((queue = m->queue) != PQ_ACTIVE) { 2173 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2174 if (m->act_count < ACT_INIT) 2175 m->act_count = ACT_INIT; 2176 if (queue != PQ_NONE) 2177 vm_page_dequeue(m); 2178 vm_page_enqueue(PQ_ACTIVE, m); 2179 } else 2180 KASSERT(queue == PQ_NONE, 2181 ("vm_page_activate: wired page %p is queued", m)); 2182 } else { 2183 if (m->act_count < ACT_INIT) 2184 m->act_count = ACT_INIT; 2185 } 2186} 2187 2188/* 2189 * vm_page_free_wakeup: 2190 * 2191 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2192 * routine is called when a page has been added to the cache or free 2193 * queues. 2194 * 2195 * The page queues must be locked. 2196 */ 2197static inline void 2198vm_page_free_wakeup(void) 2199{ 2200 2201 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2202 /* 2203 * if pageout daemon needs pages, then tell it that there are 2204 * some free. 2205 */ 2206 if (vm_pageout_pages_needed && 2207 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2208 wakeup(&vm_pageout_pages_needed); 2209 vm_pageout_pages_needed = 0; 2210 } 2211 /* 2212 * wakeup processes that are waiting on memory if we hit a 2213 * high water mark. And wakeup scheduler process if we have 2214 * lots of memory. this process will swapin processes. 2215 */ 2216 if (vm_pages_needed && !vm_page_count_min()) { 2217 vm_pages_needed = 0; 2218 wakeup(&cnt.v_free_count); 2219 } 2220} 2221 2222/* 2223 * Turn a cached page into a free page, by changing its attributes. 2224 * Keep the statistics up-to-date. 2225 * 2226 * The free page queue must be locked. 2227 */ 2228static void 2229vm_page_cache_turn_free(vm_page_t m) 2230{ 2231 2232 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2233 2234 m->object = NULL; 2235 m->valid = 0; 2236 /* Clear PG_CACHED and set PG_FREE. */ 2237 m->flags ^= PG_CACHED | PG_FREE; 2238 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 2239 ("vm_page_cache_free: page %p has inconsistent flags", m)); 2240 cnt.v_cache_count--; 2241 vm_phys_freecnt_adj(m, 1); 2242} 2243 2244/* 2245 * vm_page_free_toq: 2246 * 2247 * Returns the given page to the free list, 2248 * disassociating it with any VM object. 2249 * 2250 * The object must be locked. The page must be locked if it is managed. 2251 */ 2252void 2253vm_page_free_toq(vm_page_t m) 2254{ 2255 2256 if ((m->oflags & VPO_UNMANAGED) == 0) { 2257 vm_page_lock_assert(m, MA_OWNED); 2258 KASSERT(!pmap_page_is_mapped(m), 2259 ("vm_page_free_toq: freeing mapped page %p", m)); 2260 } else 2261 KASSERT(m->queue == PQ_NONE, 2262 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2263 PCPU_INC(cnt.v_tfree); 2264 2265 if (VM_PAGE_IS_FREE(m)) 2266 panic("vm_page_free: freeing free page %p", m); 2267 else if (vm_page_sbusied(m)) 2268 panic("vm_page_free: freeing busy page %p", m); 2269 2270 /* 2271 * Unqueue, then remove page. Note that we cannot destroy 2272 * the page here because we do not want to call the pager's 2273 * callback routine until after we've put the page on the 2274 * appropriate free queue. 2275 */ 2276 vm_page_remque(m); 2277 vm_page_remove(m); 2278 2279 /* 2280 * If fictitious remove object association and 2281 * return, otherwise delay object association removal. 2282 */ 2283 if ((m->flags & PG_FICTITIOUS) != 0) { 2284 return; 2285 } 2286 2287 m->valid = 0; 2288 vm_page_undirty(m); 2289 2290 if (m->wire_count != 0) 2291 panic("vm_page_free: freeing wired page %p", m); 2292 if (m->hold_count != 0) { 2293 m->flags &= ~PG_ZERO; 2294 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2295 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2296 m->flags |= PG_UNHOLDFREE; 2297 } else { 2298 /* 2299 * Restore the default memory attribute to the page. 2300 */ 2301 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2302 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2303 2304 /* 2305 * Insert the page into the physical memory allocator's 2306 * cache/free page queues. 2307 */ 2308 mtx_lock(&vm_page_queue_free_mtx); 2309 m->flags |= PG_FREE; 2310 vm_phys_freecnt_adj(m, 1); 2311#if VM_NRESERVLEVEL > 0 2312 if (!vm_reserv_free_page(m)) 2313#else 2314 if (TRUE) 2315#endif 2316 vm_phys_free_pages(m, 0); 2317 if ((m->flags & PG_ZERO) != 0) 2318 ++vm_page_zero_count; 2319 else 2320 vm_page_zero_idle_wakeup(); 2321 vm_page_free_wakeup(); 2322 mtx_unlock(&vm_page_queue_free_mtx); 2323 } 2324} 2325 2326/* 2327 * vm_page_wire: 2328 * 2329 * Mark this page as wired down by yet 2330 * another map, removing it from paging queues 2331 * as necessary. 2332 * 2333 * If the page is fictitious, then its wire count must remain one. 2334 * 2335 * The page must be locked. 2336 */ 2337void 2338vm_page_wire(vm_page_t m) 2339{ 2340 2341 /* 2342 * Only bump the wire statistics if the page is not already wired, 2343 * and only unqueue the page if it is on some queue (if it is unmanaged 2344 * it is already off the queues). 2345 */ 2346 vm_page_lock_assert(m, MA_OWNED); 2347 if ((m->flags & PG_FICTITIOUS) != 0) { 2348 KASSERT(m->wire_count == 1, 2349 ("vm_page_wire: fictitious page %p's wire count isn't one", 2350 m)); 2351 return; 2352 } 2353 if (m->wire_count == 0) { 2354 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2355 m->queue == PQ_NONE, 2356 ("vm_page_wire: unmanaged page %p is queued", m)); 2357 vm_page_remque(m); 2358 atomic_add_int(&cnt.v_wire_count, 1); 2359 } 2360 m->wire_count++; 2361 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2362} 2363 2364/* 2365 * vm_page_unwire: 2366 * 2367 * Release one wiring of the specified page, potentially enabling it to be 2368 * paged again. If paging is enabled, then the value of the parameter 2369 * "activate" determines to which queue the page is added. If "activate" is 2370 * non-zero, then the page is added to the active queue. Otherwise, it is 2371 * added to the inactive queue. 2372 * 2373 * However, unless the page belongs to an object, it is not enqueued because 2374 * it cannot be paged out. 2375 * 2376 * If a page is fictitious, then its wire count must always be one. 2377 * 2378 * A managed page must be locked. 2379 */ 2380void 2381vm_page_unwire(vm_page_t m, int activate) 2382{ 2383 2384 if ((m->oflags & VPO_UNMANAGED) == 0) 2385 vm_page_lock_assert(m, MA_OWNED); 2386 if ((m->flags & PG_FICTITIOUS) != 0) { 2387 KASSERT(m->wire_count == 1, 2388 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2389 return; 2390 } 2391 if (m->wire_count > 0) { 2392 m->wire_count--; 2393 if (m->wire_count == 0) { 2394 atomic_subtract_int(&cnt.v_wire_count, 1); 2395 if ((m->oflags & VPO_UNMANAGED) != 0 || 2396 m->object == NULL) 2397 return; 2398 if (!activate) 2399 m->flags &= ~PG_WINATCFLS; 2400 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2401 } 2402 } else 2403 panic("vm_page_unwire: page %p's wire count is zero", m); 2404} 2405 2406/* 2407 * Move the specified page to the inactive queue. 2408 * 2409 * Many pages placed on the inactive queue should actually go 2410 * into the cache, but it is difficult to figure out which. What 2411 * we do instead, if the inactive target is well met, is to put 2412 * clean pages at the head of the inactive queue instead of the tail. 2413 * This will cause them to be moved to the cache more quickly and 2414 * if not actively re-referenced, reclaimed more quickly. If we just 2415 * stick these pages at the end of the inactive queue, heavy filesystem 2416 * meta-data accesses can cause an unnecessary paging load on memory bound 2417 * processes. This optimization causes one-time-use metadata to be 2418 * reused more quickly. 2419 * 2420 * Normally athead is 0 resulting in LRU operation. athead is set 2421 * to 1 if we want this page to be 'as if it were placed in the cache', 2422 * except without unmapping it from the process address space. 2423 * 2424 * The page must be locked. 2425 */ 2426static inline void 2427_vm_page_deactivate(vm_page_t m, int athead) 2428{ 2429 struct vm_pagequeue *pq; 2430 int queue; 2431 2432 vm_page_lock_assert(m, MA_OWNED); 2433 2434 /* 2435 * Ignore if already inactive. 2436 */ 2437 if ((queue = m->queue) == PQ_INACTIVE) 2438 return; 2439 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2440 if (queue != PQ_NONE) 2441 vm_page_dequeue(m); 2442 m->flags &= ~PG_WINATCFLS; 2443 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2444 vm_pagequeue_lock(pq); 2445 m->queue = PQ_INACTIVE; 2446 if (athead) 2447 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q); 2448 else 2449 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2450 vm_pagequeue_cnt_inc(pq); 2451 vm_pagequeue_unlock(pq); 2452 } 2453} 2454 2455/* 2456 * Move the specified page to the inactive queue. 2457 * 2458 * The page must be locked. 2459 */ 2460void 2461vm_page_deactivate(vm_page_t m) 2462{ 2463 2464 _vm_page_deactivate(m, 0); 2465} 2466 2467/* 2468 * vm_page_try_to_cache: 2469 * 2470 * Returns 0 on failure, 1 on success 2471 */ 2472int 2473vm_page_try_to_cache(vm_page_t m) 2474{ 2475 2476 vm_page_lock_assert(m, MA_OWNED); 2477 VM_OBJECT_ASSERT_WLOCKED(m->object); 2478 if (m->dirty || m->hold_count || m->wire_count || 2479 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2480 return (0); 2481 pmap_remove_all(m); 2482 if (m->dirty) 2483 return (0); 2484 vm_page_cache(m); 2485 return (1); 2486} 2487 2488/* 2489 * vm_page_try_to_free() 2490 * 2491 * Attempt to free the page. If we cannot free it, we do nothing. 2492 * 1 is returned on success, 0 on failure. 2493 */ 2494int 2495vm_page_try_to_free(vm_page_t m) 2496{ 2497 2498 vm_page_lock_assert(m, MA_OWNED); 2499 if (m->object != NULL) 2500 VM_OBJECT_ASSERT_WLOCKED(m->object); 2501 if (m->dirty || m->hold_count || m->wire_count || 2502 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2503 return (0); 2504 pmap_remove_all(m); 2505 if (m->dirty) 2506 return (0); 2507 vm_page_free(m); 2508 return (1); 2509} 2510 2511/* 2512 * vm_page_cache 2513 * 2514 * Put the specified page onto the page cache queue (if appropriate). 2515 * 2516 * The object and page must be locked. 2517 */ 2518void 2519vm_page_cache(vm_page_t m) 2520{ 2521 vm_object_t object; 2522 boolean_t cache_was_empty; 2523 2524 vm_page_lock_assert(m, MA_OWNED); 2525 object = m->object; 2526 VM_OBJECT_ASSERT_WLOCKED(object); 2527 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) || 2528 m->hold_count || m->wire_count) 2529 panic("vm_page_cache: attempting to cache busy page"); 2530 KASSERT(!pmap_page_is_mapped(m), 2531 ("vm_page_cache: page %p is mapped", m)); 2532 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2533 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2534 (object->type == OBJT_SWAP && 2535 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2536 /* 2537 * Hypothesis: A cache-elgible page belonging to a 2538 * default object or swap object but without a backing 2539 * store must be zero filled. 2540 */ 2541 vm_page_free(m); 2542 return; 2543 } 2544 KASSERT((m->flags & PG_CACHED) == 0, 2545 ("vm_page_cache: page %p is already cached", m)); 2546 2547 /* 2548 * Remove the page from the paging queues. 2549 */ 2550 vm_page_remque(m); 2551 2552 /* 2553 * Remove the page from the object's collection of resident 2554 * pages. 2555 */ 2556 vm_radix_remove(&object->rtree, m->pindex); 2557 TAILQ_REMOVE(&object->memq, m, listq); 2558 object->resident_page_count--; 2559 2560 /* 2561 * Restore the default memory attribute to the page. 2562 */ 2563 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2564 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2565 2566 /* 2567 * Insert the page into the object's collection of cached pages 2568 * and the physical memory allocator's cache/free page queues. 2569 */ 2570 m->flags &= ~PG_ZERO; 2571 mtx_lock(&vm_page_queue_free_mtx); 2572 cache_was_empty = vm_radix_is_empty(&object->cache); 2573 if (vm_radix_insert(&object->cache, m)) { 2574 mtx_unlock(&vm_page_queue_free_mtx); 2575 if (object->resident_page_count == 0) 2576 vdrop(object->handle); 2577 m->object = NULL; 2578 vm_page_free(m); 2579 return; 2580 } 2581 2582 /* 2583 * The above call to vm_radix_insert() could reclaim the one pre- 2584 * existing cached page from this object, resulting in a call to 2585 * vdrop(). 2586 */ 2587 if (!cache_was_empty) 2588 cache_was_empty = vm_radix_is_singleton(&object->cache); 2589 2590 m->flags |= PG_CACHED; 2591 cnt.v_cache_count++; 2592 PCPU_INC(cnt.v_tcached); 2593#if VM_NRESERVLEVEL > 0 2594 if (!vm_reserv_free_page(m)) { 2595#else 2596 if (TRUE) { 2597#endif 2598 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2599 vm_phys_free_pages(m, 0); 2600 } 2601 vm_page_free_wakeup(); 2602 mtx_unlock(&vm_page_queue_free_mtx); 2603 2604 /* 2605 * Increment the vnode's hold count if this is the object's only 2606 * cached page. Decrement the vnode's hold count if this was 2607 * the object's only resident page. 2608 */ 2609 if (object->type == OBJT_VNODE) { 2610 if (cache_was_empty && object->resident_page_count != 0) 2611 vhold(object->handle); 2612 else if (!cache_was_empty && object->resident_page_count == 0) 2613 vdrop(object->handle); 2614 } 2615} 2616 2617/* 2618 * vm_page_advise 2619 * 2620 * Cache, deactivate, or do nothing as appropriate. This routine 2621 * is used by madvise(). 2622 * 2623 * Generally speaking we want to move the page into the cache so 2624 * it gets reused quickly. However, this can result in a silly syndrome 2625 * due to the page recycling too quickly. Small objects will not be 2626 * fully cached. On the other hand, if we move the page to the inactive 2627 * queue we wind up with a problem whereby very large objects 2628 * unnecessarily blow away our inactive and cache queues. 2629 * 2630 * The solution is to move the pages based on a fixed weighting. We 2631 * either leave them alone, deactivate them, or move them to the cache, 2632 * where moving them to the cache has the highest weighting. 2633 * By forcing some pages into other queues we eventually force the 2634 * system to balance the queues, potentially recovering other unrelated 2635 * space from active. The idea is to not force this to happen too 2636 * often. 2637 * 2638 * The object and page must be locked. 2639 */ 2640void 2641vm_page_advise(vm_page_t m, int advice) 2642{ 2643 int dnw, head; 2644 2645 vm_page_assert_locked(m); 2646 VM_OBJECT_ASSERT_WLOCKED(m->object); 2647 if (advice == MADV_FREE) { 2648 /* 2649 * Mark the page clean. This will allow the page to be freed 2650 * up by the system. However, such pages are often reused 2651 * quickly by malloc() so we do not do anything that would 2652 * cause a page fault if we can help it. 2653 * 2654 * Specifically, we do not try to actually free the page now 2655 * nor do we try to put it in the cache (which would cause a 2656 * page fault on reuse). 2657 * 2658 * But we do make the page is freeable as we can without 2659 * actually taking the step of unmapping it. 2660 */ 2661 m->dirty = 0; 2662 m->act_count = 0; 2663 } else if (advice != MADV_DONTNEED) 2664 return; 2665 dnw = PCPU_GET(dnweight); 2666 PCPU_INC(dnweight); 2667 2668 /* 2669 * Occasionally leave the page alone. 2670 */ 2671 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2672 if (m->act_count >= ACT_INIT) 2673 --m->act_count; 2674 return; 2675 } 2676 2677 /* 2678 * Clear any references to the page. Otherwise, the page daemon will 2679 * immediately reactivate the page. 2680 */ 2681 vm_page_aflag_clear(m, PGA_REFERENCED); 2682 2683 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 2684 vm_page_dirty(m); 2685 2686 if (m->dirty || (dnw & 0x0070) == 0) { 2687 /* 2688 * Deactivate the page 3 times out of 32. 2689 */ 2690 head = 0; 2691 } else { 2692 /* 2693 * Cache the page 28 times out of every 32. Note that 2694 * the page is deactivated instead of cached, but placed 2695 * at the head of the queue instead of the tail. 2696 */ 2697 head = 1; 2698 } 2699 _vm_page_deactivate(m, head); 2700} 2701 2702/* 2703 * Grab a page, waiting until we are waken up due to the page 2704 * changing state. We keep on waiting, if the page continues 2705 * to be in the object. If the page doesn't exist, first allocate it 2706 * and then conditionally zero it. 2707 * 2708 * This routine may sleep. 2709 * 2710 * The object must be locked on entry. The lock will, however, be released 2711 * and reacquired if the routine sleeps. 2712 */ 2713vm_page_t 2714vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2715{ 2716 vm_page_t m; 2717 int sleep; 2718 2719 VM_OBJECT_ASSERT_WLOCKED(object); 2720 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 2721 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 2722 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 2723retrylookup: 2724 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2725 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 2726 vm_page_xbusied(m) : vm_page_busied(m); 2727 if (sleep) { 2728 /* 2729 * Reference the page before unlocking and 2730 * sleeping so that the page daemon is less 2731 * likely to reclaim it. 2732 */ 2733 vm_page_aflag_set(m, PGA_REFERENCED); 2734 vm_page_lock(m); 2735 VM_OBJECT_WUNLOCK(object); 2736 vm_page_busy_sleep(m, "pgrbwt"); 2737 VM_OBJECT_WLOCK(object); 2738 goto retrylookup; 2739 } else { 2740 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2741 vm_page_lock(m); 2742 vm_page_wire(m); 2743 vm_page_unlock(m); 2744 } 2745 if ((allocflags & 2746 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2747 vm_page_xbusy(m); 2748 if ((allocflags & VM_ALLOC_SBUSY) != 0) 2749 vm_page_sbusy(m); 2750 return (m); 2751 } 2752 } 2753 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY); 2754 if (m == NULL) { 2755 VM_OBJECT_WUNLOCK(object); 2756 VM_WAIT; 2757 VM_OBJECT_WLOCK(object); 2758 goto retrylookup; 2759 } else if (m->valid != 0) 2760 return (m); 2761 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2762 pmap_zero_page(m); 2763 return (m); 2764} 2765 2766/* 2767 * Mapping function for valid or dirty bits in a page. 2768 * 2769 * Inputs are required to range within a page. 2770 */ 2771vm_page_bits_t 2772vm_page_bits(int base, int size) 2773{ 2774 int first_bit; 2775 int last_bit; 2776 2777 KASSERT( 2778 base + size <= PAGE_SIZE, 2779 ("vm_page_bits: illegal base/size %d/%d", base, size) 2780 ); 2781 2782 if (size == 0) /* handle degenerate case */ 2783 return (0); 2784 2785 first_bit = base >> DEV_BSHIFT; 2786 last_bit = (base + size - 1) >> DEV_BSHIFT; 2787 2788 return (((vm_page_bits_t)2 << last_bit) - 2789 ((vm_page_bits_t)1 << first_bit)); 2790} 2791 2792/* 2793 * vm_page_set_valid_range: 2794 * 2795 * Sets portions of a page valid. The arguments are expected 2796 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2797 * of any partial chunks touched by the range. The invalid portion of 2798 * such chunks will be zeroed. 2799 * 2800 * (base + size) must be less then or equal to PAGE_SIZE. 2801 */ 2802void 2803vm_page_set_valid_range(vm_page_t m, int base, int size) 2804{ 2805 int endoff, frag; 2806 2807 VM_OBJECT_ASSERT_WLOCKED(m->object); 2808 if (size == 0) /* handle degenerate case */ 2809 return; 2810 2811 /* 2812 * If the base is not DEV_BSIZE aligned and the valid 2813 * bit is clear, we have to zero out a portion of the 2814 * first block. 2815 */ 2816 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2817 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2818 pmap_zero_page_area(m, frag, base - frag); 2819 2820 /* 2821 * If the ending offset is not DEV_BSIZE aligned and the 2822 * valid bit is clear, we have to zero out a portion of 2823 * the last block. 2824 */ 2825 endoff = base + size; 2826 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2827 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2828 pmap_zero_page_area(m, endoff, 2829 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2830 2831 /* 2832 * Assert that no previously invalid block that is now being validated 2833 * is already dirty. 2834 */ 2835 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2836 ("vm_page_set_valid_range: page %p is dirty", m)); 2837 2838 /* 2839 * Set valid bits inclusive of any overlap. 2840 */ 2841 m->valid |= vm_page_bits(base, size); 2842} 2843 2844/* 2845 * Clear the given bits from the specified page's dirty field. 2846 */ 2847static __inline void 2848vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2849{ 2850 uintptr_t addr; 2851#if PAGE_SIZE < 16384 2852 int shift; 2853#endif 2854 2855 /* 2856 * If the object is locked and the page is neither exclusive busy nor 2857 * write mapped, then the page's dirty field cannot possibly be 2858 * set by a concurrent pmap operation. 2859 */ 2860 VM_OBJECT_ASSERT_WLOCKED(m->object); 2861 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 2862 m->dirty &= ~pagebits; 2863 else { 2864 /* 2865 * The pmap layer can call vm_page_dirty() without 2866 * holding a distinguished lock. The combination of 2867 * the object's lock and an atomic operation suffice 2868 * to guarantee consistency of the page dirty field. 2869 * 2870 * For PAGE_SIZE == 32768 case, compiler already 2871 * properly aligns the dirty field, so no forcible 2872 * alignment is needed. Only require existence of 2873 * atomic_clear_64 when page size is 32768. 2874 */ 2875 addr = (uintptr_t)&m->dirty; 2876#if PAGE_SIZE == 32768 2877 atomic_clear_64((uint64_t *)addr, pagebits); 2878#elif PAGE_SIZE == 16384 2879 atomic_clear_32((uint32_t *)addr, pagebits); 2880#else /* PAGE_SIZE <= 8192 */ 2881 /* 2882 * Use a trick to perform a 32-bit atomic on the 2883 * containing aligned word, to not depend on the existence 2884 * of atomic_clear_{8, 16}. 2885 */ 2886 shift = addr & (sizeof(uint32_t) - 1); 2887#if BYTE_ORDER == BIG_ENDIAN 2888 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2889#else 2890 shift *= NBBY; 2891#endif 2892 addr &= ~(sizeof(uint32_t) - 1); 2893 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2894#endif /* PAGE_SIZE */ 2895 } 2896} 2897 2898/* 2899 * vm_page_set_validclean: 2900 * 2901 * Sets portions of a page valid and clean. The arguments are expected 2902 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2903 * of any partial chunks touched by the range. The invalid portion of 2904 * such chunks will be zero'd. 2905 * 2906 * (base + size) must be less then or equal to PAGE_SIZE. 2907 */ 2908void 2909vm_page_set_validclean(vm_page_t m, int base, int size) 2910{ 2911 vm_page_bits_t oldvalid, pagebits; 2912 int endoff, frag; 2913 2914 VM_OBJECT_ASSERT_WLOCKED(m->object); 2915 if (size == 0) /* handle degenerate case */ 2916 return; 2917 2918 /* 2919 * If the base is not DEV_BSIZE aligned and the valid 2920 * bit is clear, we have to zero out a portion of the 2921 * first block. 2922 */ 2923 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2924 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2925 pmap_zero_page_area(m, frag, base - frag); 2926 2927 /* 2928 * If the ending offset is not DEV_BSIZE aligned and the 2929 * valid bit is clear, we have to zero out a portion of 2930 * the last block. 2931 */ 2932 endoff = base + size; 2933 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2934 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2935 pmap_zero_page_area(m, endoff, 2936 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2937 2938 /* 2939 * Set valid, clear dirty bits. If validating the entire 2940 * page we can safely clear the pmap modify bit. We also 2941 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2942 * takes a write fault on a MAP_NOSYNC memory area the flag will 2943 * be set again. 2944 * 2945 * We set valid bits inclusive of any overlap, but we can only 2946 * clear dirty bits for DEV_BSIZE chunks that are fully within 2947 * the range. 2948 */ 2949 oldvalid = m->valid; 2950 pagebits = vm_page_bits(base, size); 2951 m->valid |= pagebits; 2952#if 0 /* NOT YET */ 2953 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2954 frag = DEV_BSIZE - frag; 2955 base += frag; 2956 size -= frag; 2957 if (size < 0) 2958 size = 0; 2959 } 2960 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2961#endif 2962 if (base == 0 && size == PAGE_SIZE) { 2963 /* 2964 * The page can only be modified within the pmap if it is 2965 * mapped, and it can only be mapped if it was previously 2966 * fully valid. 2967 */ 2968 if (oldvalid == VM_PAGE_BITS_ALL) 2969 /* 2970 * Perform the pmap_clear_modify() first. Otherwise, 2971 * a concurrent pmap operation, such as 2972 * pmap_protect(), could clear a modification in the 2973 * pmap and set the dirty field on the page before 2974 * pmap_clear_modify() had begun and after the dirty 2975 * field was cleared here. 2976 */ 2977 pmap_clear_modify(m); 2978 m->dirty = 0; 2979 m->oflags &= ~VPO_NOSYNC; 2980 } else if (oldvalid != VM_PAGE_BITS_ALL) 2981 m->dirty &= ~pagebits; 2982 else 2983 vm_page_clear_dirty_mask(m, pagebits); 2984} 2985 2986void 2987vm_page_clear_dirty(vm_page_t m, int base, int size) 2988{ 2989 2990 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2991} 2992 2993/* 2994 * vm_page_set_invalid: 2995 * 2996 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2997 * valid and dirty bits for the effected areas are cleared. 2998 */ 2999void 3000vm_page_set_invalid(vm_page_t m, int base, int size) 3001{ 3002 vm_page_bits_t bits; 3003 vm_object_t object; 3004 3005 object = m->object; 3006 VM_OBJECT_ASSERT_WLOCKED(object); 3007 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3008 size >= object->un_pager.vnp.vnp_size) 3009 bits = VM_PAGE_BITS_ALL; 3010 else 3011 bits = vm_page_bits(base, size); 3012 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 3013 pmap_remove_all(m); 3014 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3015 !pmap_page_is_mapped(m), 3016 ("vm_page_set_invalid: page %p is mapped", m)); 3017 m->valid &= ~bits; 3018 m->dirty &= ~bits; 3019} 3020 3021/* 3022 * vm_page_zero_invalid() 3023 * 3024 * The kernel assumes that the invalid portions of a page contain 3025 * garbage, but such pages can be mapped into memory by user code. 3026 * When this occurs, we must zero out the non-valid portions of the 3027 * page so user code sees what it expects. 3028 * 3029 * Pages are most often semi-valid when the end of a file is mapped 3030 * into memory and the file's size is not page aligned. 3031 */ 3032void 3033vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3034{ 3035 int b; 3036 int i; 3037 3038 VM_OBJECT_ASSERT_WLOCKED(m->object); 3039 /* 3040 * Scan the valid bits looking for invalid sections that 3041 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3042 * valid bit may be set ) have already been zeroed by 3043 * vm_page_set_validclean(). 3044 */ 3045 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3046 if (i == (PAGE_SIZE / DEV_BSIZE) || 3047 (m->valid & ((vm_page_bits_t)1 << i))) { 3048 if (i > b) { 3049 pmap_zero_page_area(m, 3050 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3051 } 3052 b = i + 1; 3053 } 3054 } 3055 3056 /* 3057 * setvalid is TRUE when we can safely set the zero'd areas 3058 * as being valid. We can do this if there are no cache consistancy 3059 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3060 */ 3061 if (setvalid) 3062 m->valid = VM_PAGE_BITS_ALL; 3063} 3064 3065/* 3066 * vm_page_is_valid: 3067 * 3068 * Is (partial) page valid? Note that the case where size == 0 3069 * will return FALSE in the degenerate case where the page is 3070 * entirely invalid, and TRUE otherwise. 3071 */ 3072int 3073vm_page_is_valid(vm_page_t m, int base, int size) 3074{ 3075 vm_page_bits_t bits; 3076 3077 VM_OBJECT_ASSERT_LOCKED(m->object); 3078 bits = vm_page_bits(base, size); 3079 return (m->valid != 0 && (m->valid & bits) == bits); 3080} 3081 3082/* 3083 * vm_page_ps_is_valid: 3084 * 3085 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3086 */ 3087boolean_t 3088vm_page_ps_is_valid(vm_page_t m) 3089{ 3090 int i, npages; 3091 3092 VM_OBJECT_ASSERT_LOCKED(m->object); 3093 npages = atop(pagesizes[m->psind]); 3094 3095 /* 3096 * The physically contiguous pages that make up a superpage, i.e., a 3097 * page with a page size index ("psind") greater than zero, will 3098 * occupy adjacent entries in vm_page_array[]. 3099 */ 3100 for (i = 0; i < npages; i++) { 3101 if (m[i].valid != VM_PAGE_BITS_ALL) 3102 return (FALSE); 3103 } 3104 return (TRUE); 3105} 3106 3107/* 3108 * Set the page's dirty bits if the page is modified. 3109 */ 3110void 3111vm_page_test_dirty(vm_page_t m) 3112{ 3113 3114 VM_OBJECT_ASSERT_WLOCKED(m->object); 3115 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3116 vm_page_dirty(m); 3117} 3118 3119void 3120vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3121{ 3122 3123 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3124} 3125 3126void 3127vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3128{ 3129 3130 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3131} 3132 3133int 3134vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3135{ 3136 3137 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3138} 3139 3140#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3141void 3142vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3143{ 3144 3145 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3146} 3147 3148void 3149vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3150{ 3151 3152 mtx_assert_(vm_page_lockptr(m), a, file, line); 3153} 3154#endif 3155 3156#ifdef INVARIANTS 3157void 3158vm_page_object_lock_assert(vm_page_t m) 3159{ 3160 3161 /* 3162 * Certain of the page's fields may only be modified by the 3163 * holder of the containing object's lock or the exclusive busy. 3164 * holder. Unfortunately, the holder of the write busy is 3165 * not recorded, and thus cannot be checked here. 3166 */ 3167 if (m->object != NULL && !vm_page_xbusied(m)) 3168 VM_OBJECT_ASSERT_WLOCKED(m->object); 3169} 3170 3171void 3172vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3173{ 3174 3175 if ((bits & PGA_WRITEABLE) == 0) 3176 return; 3177 3178 /* 3179 * The PGA_WRITEABLE flag can only be set if the page is 3180 * managed, is exclusively busied or the object is locked. 3181 * Currently, this flag is only set by pmap_enter(). 3182 */ 3183 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3184 ("PGA_WRITEABLE on unmanaged page")); 3185 if (!vm_page_xbusied(m)) 3186 VM_OBJECT_ASSERT_LOCKED(m->object); 3187} 3188#endif 3189 3190#include "opt_ddb.h" 3191#ifdef DDB 3192#include <sys/kernel.h> 3193 3194#include <ddb/ddb.h> 3195 3196DB_SHOW_COMMAND(page, vm_page_print_page_info) 3197{ 3198 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3199 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3200 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3201 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3202 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3203 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3204 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3205 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3206 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3207 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3208} 3209 3210DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3211{ 3212 int dom; 3213 3214 db_printf("pq_free %d pq_cache %d\n", 3215 cnt.v_free_count, cnt.v_cache_count); 3216 for (dom = 0; dom < vm_ndomains; dom++) { 3217 db_printf( 3218 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n", 3219 dom, 3220 vm_dom[dom].vmd_page_count, 3221 vm_dom[dom].vmd_free_count, 3222 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3223 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3224 vm_dom[dom].vmd_pass); 3225 } 3226} 3227 3228DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3229{ 3230 vm_page_t m; 3231 boolean_t phys; 3232 3233 if (!have_addr) { 3234 db_printf("show pginfo addr\n"); 3235 return; 3236 } 3237 3238 phys = strchr(modif, 'p') != NULL; 3239 if (phys) 3240 m = PHYS_TO_VM_PAGE(addr); 3241 else 3242 m = (vm_page_t)addr; 3243 db_printf( 3244 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3245 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3246 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3247 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3248 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3249} 3250#endif /* DDB */ 3251