vm_page.c revision 209407
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 pageq mutex is required when adding or removing a page from a 67 * page queue (vm_page_queue[]), regardless of other mutexes or the 68 * busy state of a page. 69 * 70 * - a hash chain mutex is required when associating or disassociating 71 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 72 * regardless of other mutexes or the busy state of a page. 73 * 74 * - either a hash chain mutex OR a busied page is required in order 75 * to modify the page flags. A hash chain mutex must be obtained in 76 * order to busy a page. A page's flags cannot be modified by a 77 * hash chain mutex if the page is marked busy. 78 * 79 * - The object memq mutex is held when inserting or removing 80 * pages from an object (vm_page_insert() or vm_page_remove()). This 81 * is different from the object's main mutex. 82 * 83 * Generally speaking, you have to be aware of side effects when running 84 * vm_page ops. A vm_page_lookup() will return with the hash chain 85 * locked, whether it was able to lookup the page or not. vm_page_free(), 86 * vm_page_cache(), vm_page_activate(), and a number of other routines 87 * will release the hash chain mutex for you. Intermediate manipulation 88 * routines such as vm_page_flag_set() expect the hash chain to be held 89 * on entry and the hash chain will remain held on return. 90 * 91 * pageq scanning can only occur with the pageq in question locked. 92 * We have a known bottleneck with the active queue, but the cache 93 * and free queues are actually arrays already. 94 */ 95 96/* 97 * Resident memory management module. 98 */ 99 100#include <sys/cdefs.h> 101__FBSDID("$FreeBSD: head/sys/vm/vm_page.c 209407 2010-06-21 23:27:24Z alc $"); 102 103#include "opt_vm.h" 104 105#include <sys/param.h> 106#include <sys/systm.h> 107#include <sys/lock.h> 108#include <sys/kernel.h> 109#include <sys/limits.h> 110#include <sys/malloc.h> 111#include <sys/msgbuf.h> 112#include <sys/mutex.h> 113#include <sys/proc.h> 114#include <sys/sysctl.h> 115#include <sys/vmmeter.h> 116#include <sys/vnode.h> 117 118#include <vm/vm.h> 119#include <vm/pmap.h> 120#include <vm/vm_param.h> 121#include <vm/vm_kern.h> 122#include <vm/vm_object.h> 123#include <vm/vm_page.h> 124#include <vm/vm_pageout.h> 125#include <vm/vm_pager.h> 126#include <vm/vm_phys.h> 127#include <vm/vm_reserv.h> 128#include <vm/vm_extern.h> 129#include <vm/uma.h> 130#include <vm/uma_int.h> 131 132#include <machine/md_var.h> 133 134#if defined(__amd64__) || defined (__i386__) 135extern struct sysctl_oid_list sysctl__vm_pmap_children; 136#else 137SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); 138#endif 139 140static uint64_t pmap_tryrelock_calls; 141SYSCTL_QUAD(_vm_pmap, OID_AUTO, tryrelock_calls, CTLFLAG_RD, 142 &pmap_tryrelock_calls, 0, "Number of tryrelock calls"); 143 144static int pmap_tryrelock_restart; 145SYSCTL_INT(_vm_pmap, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 146 &pmap_tryrelock_restart, 0, "Number of tryrelock restarts"); 147 148static int pmap_tryrelock_race; 149SYSCTL_INT(_vm_pmap, OID_AUTO, tryrelock_race, CTLFLAG_RD, 150 &pmap_tryrelock_race, 0, "Number of tryrelock pmap race cases"); 151 152/* 153 * Associated with page of user-allocatable memory is a 154 * page structure. 155 */ 156 157struct vpgqueues vm_page_queues[PQ_COUNT]; 158struct vpglocks vm_page_queue_lock; 159struct vpglocks vm_page_queue_free_lock; 160 161struct vpglocks pa_lock[PA_LOCK_COUNT] __aligned(CACHE_LINE_SIZE); 162 163vm_page_t vm_page_array = 0; 164int vm_page_array_size = 0; 165long first_page = 0; 166int vm_page_zero_count = 0; 167 168static int boot_pages = UMA_BOOT_PAGES; 169TUNABLE_INT("vm.boot_pages", &boot_pages); 170SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 171 "number of pages allocated for bootstrapping the VM system"); 172 173static void vm_page_clear_dirty_mask(vm_page_t m, int pagebits); 174static void vm_page_queue_remove(int queue, vm_page_t m); 175static void vm_page_enqueue(int queue, vm_page_t m); 176 177/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 178#if PAGE_SIZE == 32768 179#ifdef CTASSERT 180CTASSERT(sizeof(u_long) >= 8); 181#endif 182#endif 183 184/* 185 * Try to acquire a physical address lock while a pmap is locked. If we 186 * fail to trylock we unlock and lock the pmap directly and cache the 187 * locked pa in *locked. The caller should then restart their loop in case 188 * the virtual to physical mapping has changed. 189 */ 190int 191vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 192{ 193 vm_paddr_t lockpa; 194 uint32_t gen_count; 195 196 gen_count = pmap->pm_gen_count; 197 atomic_add_long((volatile long *)&pmap_tryrelock_calls, 1); 198 lockpa = *locked; 199 *locked = pa; 200 if (lockpa) { 201 PA_LOCK_ASSERT(lockpa, MA_OWNED); 202 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 203 return (0); 204 PA_UNLOCK(lockpa); 205 } 206 if (PA_TRYLOCK(pa)) 207 return (0); 208 PMAP_UNLOCK(pmap); 209 atomic_add_int((volatile int *)&pmap_tryrelock_restart, 1); 210 PA_LOCK(pa); 211 PMAP_LOCK(pmap); 212 213 if (pmap->pm_gen_count != gen_count + 1) { 214 pmap->pm_retries++; 215 atomic_add_int((volatile int *)&pmap_tryrelock_race, 1); 216 return (EAGAIN); 217 } 218 return (0); 219} 220 221/* 222 * vm_set_page_size: 223 * 224 * Sets the page size, perhaps based upon the memory 225 * size. Must be called before any use of page-size 226 * dependent functions. 227 */ 228void 229vm_set_page_size(void) 230{ 231 if (cnt.v_page_size == 0) 232 cnt.v_page_size = PAGE_SIZE; 233 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 234 panic("vm_set_page_size: page size not a power of two"); 235} 236 237/* 238 * vm_page_blacklist_lookup: 239 * 240 * See if a physical address in this page has been listed 241 * in the blacklist tunable. Entries in the tunable are 242 * separated by spaces or commas. If an invalid integer is 243 * encountered then the rest of the string is skipped. 244 */ 245static int 246vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 247{ 248 vm_paddr_t bad; 249 char *cp, *pos; 250 251 for (pos = list; *pos != '\0'; pos = cp) { 252 bad = strtoq(pos, &cp, 0); 253 if (*cp != '\0') { 254 if (*cp == ' ' || *cp == ',') { 255 cp++; 256 if (cp == pos) 257 continue; 258 } else 259 break; 260 } 261 if (pa == trunc_page(bad)) 262 return (1); 263 } 264 return (0); 265} 266 267/* 268 * vm_page_startup: 269 * 270 * Initializes the resident memory module. 271 * 272 * Allocates memory for the page cells, and 273 * for the object/offset-to-page hash table headers. 274 * Each page cell is initialized and placed on the free list. 275 */ 276vm_offset_t 277vm_page_startup(vm_offset_t vaddr) 278{ 279 vm_offset_t mapped; 280 vm_paddr_t page_range; 281 vm_paddr_t new_end; 282 int i; 283 vm_paddr_t pa; 284 int nblocks; 285 vm_paddr_t last_pa; 286 char *list; 287 288 /* the biggest memory array is the second group of pages */ 289 vm_paddr_t end; 290 vm_paddr_t biggestsize; 291 vm_paddr_t low_water, high_water; 292 int biggestone; 293 294 biggestsize = 0; 295 biggestone = 0; 296 nblocks = 0; 297 vaddr = round_page(vaddr); 298 299 for (i = 0; phys_avail[i + 1]; i += 2) { 300 phys_avail[i] = round_page(phys_avail[i]); 301 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 302 } 303 304 low_water = phys_avail[0]; 305 high_water = phys_avail[1]; 306 307 for (i = 0; phys_avail[i + 1]; i += 2) { 308 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 309 310 if (size > biggestsize) { 311 biggestone = i; 312 biggestsize = size; 313 } 314 if (phys_avail[i] < low_water) 315 low_water = phys_avail[i]; 316 if (phys_avail[i + 1] > high_water) 317 high_water = phys_avail[i + 1]; 318 ++nblocks; 319 } 320 321#ifdef XEN 322 low_water = 0; 323#endif 324 325 end = phys_avail[biggestone+1]; 326 327 /* 328 * Initialize the locks. 329 */ 330 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 331 MTX_RECURSE); 332 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 333 MTX_DEF); 334 335 /* Setup page locks. */ 336 for (i = 0; i < PA_LOCK_COUNT; i++) 337 mtx_init(&pa_lock[i].data, "page lock", NULL, 338 MTX_DEF | MTX_RECURSE | MTX_DUPOK); 339 340 /* 341 * Initialize the queue headers for the hold queue, the active queue, 342 * and the inactive queue. 343 */ 344 for (i = 0; i < PQ_COUNT; i++) 345 TAILQ_INIT(&vm_page_queues[i].pl); 346 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 347 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 348 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count; 349 350 /* 351 * Allocate memory for use when boot strapping the kernel memory 352 * allocator. 353 */ 354 new_end = end - (boot_pages * UMA_SLAB_SIZE); 355 new_end = trunc_page(new_end); 356 mapped = pmap_map(&vaddr, new_end, end, 357 VM_PROT_READ | VM_PROT_WRITE); 358 bzero((void *)mapped, end - new_end); 359 uma_startup((void *)mapped, boot_pages); 360 361#if defined(__amd64__) || defined(__i386__) || defined(__arm__) 362 /* 363 * Allocate a bitmap to indicate that a random physical page 364 * needs to be included in a minidump. 365 * 366 * The amd64 port needs this to indicate which direct map pages 367 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 368 * 369 * However, i386 still needs this workspace internally within the 370 * minidump code. In theory, they are not needed on i386, but are 371 * included should the sf_buf code decide to use them. 372 */ 373 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 374 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 375 new_end -= vm_page_dump_size; 376 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 377 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 378 bzero((void *)vm_page_dump, vm_page_dump_size); 379#endif 380#ifdef __amd64__ 381 /* 382 * Request that the physical pages underlying the message buffer be 383 * included in a crash dump. Since the message buffer is accessed 384 * through the direct map, they are not automatically included. 385 */ 386 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 387 last_pa = pa + round_page(MSGBUF_SIZE); 388 while (pa < last_pa) { 389 dump_add_page(pa); 390 pa += PAGE_SIZE; 391 } 392#endif 393 /* 394 * Compute the number of pages of memory that will be available for 395 * use (taking into account the overhead of a page structure per 396 * page). 397 */ 398 first_page = low_water / PAGE_SIZE; 399#ifdef VM_PHYSSEG_SPARSE 400 page_range = 0; 401 for (i = 0; phys_avail[i + 1] != 0; i += 2) 402 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 403#elif defined(VM_PHYSSEG_DENSE) 404 page_range = high_water / PAGE_SIZE - first_page; 405#else 406#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 407#endif 408 end = new_end; 409 410 /* 411 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 412 */ 413 vaddr += PAGE_SIZE; 414 415 /* 416 * Initialize the mem entry structures now, and put them in the free 417 * queue. 418 */ 419 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 420 mapped = pmap_map(&vaddr, new_end, end, 421 VM_PROT_READ | VM_PROT_WRITE); 422 vm_page_array = (vm_page_t) mapped; 423#if VM_NRESERVLEVEL > 0 424 /* 425 * Allocate memory for the reservation management system's data 426 * structures. 427 */ 428 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 429#endif 430#ifdef __amd64__ 431 /* 432 * pmap_map on amd64 comes out of the direct-map, not kvm like i386, 433 * so the pages must be tracked for a crashdump to include this data. 434 * This includes the vm_page_array and the early UMA bootstrap pages. 435 */ 436 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 437 dump_add_page(pa); 438#endif 439 phys_avail[biggestone + 1] = new_end; 440 441 /* 442 * Clear all of the page structures 443 */ 444 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 445 for (i = 0; i < page_range; i++) 446 vm_page_array[i].order = VM_NFREEORDER; 447 vm_page_array_size = page_range; 448 449 /* 450 * Initialize the physical memory allocator. 451 */ 452 vm_phys_init(); 453 454 /* 455 * Add every available physical page that is not blacklisted to 456 * the free lists. 457 */ 458 cnt.v_page_count = 0; 459 cnt.v_free_count = 0; 460 list = getenv("vm.blacklist"); 461 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 462 pa = phys_avail[i]; 463 last_pa = phys_avail[i + 1]; 464 while (pa < last_pa) { 465 if (list != NULL && 466 vm_page_blacklist_lookup(list, pa)) 467 printf("Skipping page with pa 0x%jx\n", 468 (uintmax_t)pa); 469 else 470 vm_phys_add_page(pa); 471 pa += PAGE_SIZE; 472 } 473 } 474 freeenv(list); 475#if VM_NRESERVLEVEL > 0 476 /* 477 * Initialize the reservation management system. 478 */ 479 vm_reserv_init(); 480#endif 481 return (vaddr); 482} 483 484void 485vm_page_flag_set(vm_page_t m, unsigned short bits) 486{ 487 488 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 489 /* 490 * The PG_WRITEABLE flag can only be set if the page is managed and 491 * VPO_BUSY. Currently, this flag is only set by pmap_enter(). 492 */ 493 KASSERT((bits & PG_WRITEABLE) == 0 || 494 ((m->flags & (PG_UNMANAGED | PG_FICTITIOUS)) == 0 && 495 (m->oflags & VPO_BUSY) != 0), ("PG_WRITEABLE and !VPO_BUSY")); 496 m->flags |= bits; 497} 498 499void 500vm_page_flag_clear(vm_page_t m, unsigned short bits) 501{ 502 503 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 504 /* 505 * The PG_REFERENCED flag can only be cleared if the object 506 * containing the page is locked. 507 */ 508 KASSERT((bits & PG_REFERENCED) == 0 || VM_OBJECT_LOCKED(m->object), 509 ("PG_REFERENCED and !VM_OBJECT_LOCKED")); 510 m->flags &= ~bits; 511} 512 513void 514vm_page_busy(vm_page_t m) 515{ 516 517 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 518 KASSERT((m->oflags & VPO_BUSY) == 0, 519 ("vm_page_busy: page already busy!!!")); 520 m->oflags |= VPO_BUSY; 521} 522 523/* 524 * vm_page_flash: 525 * 526 * wakeup anyone waiting for the page. 527 */ 528void 529vm_page_flash(vm_page_t m) 530{ 531 532 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 533 if (m->oflags & VPO_WANTED) { 534 m->oflags &= ~VPO_WANTED; 535 wakeup(m); 536 } 537} 538 539/* 540 * vm_page_wakeup: 541 * 542 * clear the VPO_BUSY flag and wakeup anyone waiting for the 543 * page. 544 * 545 */ 546void 547vm_page_wakeup(vm_page_t m) 548{ 549 550 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 551 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 552 m->oflags &= ~VPO_BUSY; 553 vm_page_flash(m); 554} 555 556void 557vm_page_io_start(vm_page_t m) 558{ 559 560 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 561 m->busy++; 562} 563 564void 565vm_page_io_finish(vm_page_t m) 566{ 567 568 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 569 m->busy--; 570 if (m->busy == 0) 571 vm_page_flash(m); 572} 573 574/* 575 * Keep page from being freed by the page daemon 576 * much of the same effect as wiring, except much lower 577 * overhead and should be used only for *very* temporary 578 * holding ("wiring"). 579 */ 580void 581vm_page_hold(vm_page_t mem) 582{ 583 584 vm_page_lock_assert(mem, MA_OWNED); 585 mem->hold_count++; 586} 587 588void 589vm_page_unhold(vm_page_t mem) 590{ 591 592 vm_page_lock_assert(mem, MA_OWNED); 593 --mem->hold_count; 594 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 595 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD)) 596 vm_page_free_toq(mem); 597} 598 599/* 600 * vm_page_free: 601 * 602 * Free a page. 603 */ 604void 605vm_page_free(vm_page_t m) 606{ 607 608 m->flags &= ~PG_ZERO; 609 vm_page_free_toq(m); 610} 611 612/* 613 * vm_page_free_zero: 614 * 615 * Free a page to the zerod-pages queue 616 */ 617void 618vm_page_free_zero(vm_page_t m) 619{ 620 621 m->flags |= PG_ZERO; 622 vm_page_free_toq(m); 623} 624 625/* 626 * vm_page_sleep: 627 * 628 * Sleep and release the page and page queues locks. 629 * 630 * The object containing the given page must be locked. 631 */ 632void 633vm_page_sleep(vm_page_t m, const char *msg) 634{ 635 636 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 637 if (mtx_owned(&vm_page_queue_mtx)) 638 vm_page_unlock_queues(); 639 if (mtx_owned(vm_page_lockptr(m))) 640 vm_page_unlock(m); 641 642 /* 643 * It's possible that while we sleep, the page will get 644 * unbusied and freed. If we are holding the object 645 * lock, we will assume we hold a reference to the object 646 * such that even if m->object changes, we can re-lock 647 * it. 648 */ 649 m->oflags |= VPO_WANTED; 650 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 651} 652 653/* 654 * vm_page_dirty: 655 * 656 * make page all dirty 657 */ 658void 659vm_page_dirty(vm_page_t m) 660{ 661 662 KASSERT((m->flags & PG_CACHED) == 0, 663 ("vm_page_dirty: page in cache!")); 664 KASSERT(!VM_PAGE_IS_FREE(m), 665 ("vm_page_dirty: page is free!")); 666 KASSERT(m->valid == VM_PAGE_BITS_ALL, 667 ("vm_page_dirty: page is invalid!")); 668 m->dirty = VM_PAGE_BITS_ALL; 669} 670 671/* 672 * vm_page_splay: 673 * 674 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 675 * the vm_page containing the given pindex. If, however, that 676 * pindex is not found in the vm_object, returns a vm_page that is 677 * adjacent to the pindex, coming before or after it. 678 */ 679vm_page_t 680vm_page_splay(vm_pindex_t pindex, vm_page_t root) 681{ 682 struct vm_page dummy; 683 vm_page_t lefttreemax, righttreemin, y; 684 685 if (root == NULL) 686 return (root); 687 lefttreemax = righttreemin = &dummy; 688 for (;; root = y) { 689 if (pindex < root->pindex) { 690 if ((y = root->left) == NULL) 691 break; 692 if (pindex < y->pindex) { 693 /* Rotate right. */ 694 root->left = y->right; 695 y->right = root; 696 root = y; 697 if ((y = root->left) == NULL) 698 break; 699 } 700 /* Link into the new root's right tree. */ 701 righttreemin->left = root; 702 righttreemin = root; 703 } else if (pindex > root->pindex) { 704 if ((y = root->right) == NULL) 705 break; 706 if (pindex > y->pindex) { 707 /* Rotate left. */ 708 root->right = y->left; 709 y->left = root; 710 root = y; 711 if ((y = root->right) == NULL) 712 break; 713 } 714 /* Link into the new root's left tree. */ 715 lefttreemax->right = root; 716 lefttreemax = root; 717 } else 718 break; 719 } 720 /* Assemble the new root. */ 721 lefttreemax->right = root->left; 722 righttreemin->left = root->right; 723 root->left = dummy.right; 724 root->right = dummy.left; 725 return (root); 726} 727 728/* 729 * vm_page_insert: [ internal use only ] 730 * 731 * Inserts the given mem entry into the object and object list. 732 * 733 * The pagetables are not updated but will presumably fault the page 734 * in if necessary, or if a kernel page the caller will at some point 735 * enter the page into the kernel's pmap. We are not allowed to block 736 * here so we *can't* do this anyway. 737 * 738 * The object and page must be locked. 739 * This routine may not block. 740 */ 741void 742vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 743{ 744 vm_page_t root; 745 746 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 747 if (m->object != NULL) 748 panic("vm_page_insert: page already inserted"); 749 750 /* 751 * Record the object/offset pair in this page 752 */ 753 m->object = object; 754 m->pindex = pindex; 755 756 /* 757 * Now link into the object's ordered list of backed pages. 758 */ 759 root = object->root; 760 if (root == NULL) { 761 m->left = NULL; 762 m->right = NULL; 763 TAILQ_INSERT_TAIL(&object->memq, m, listq); 764 } else { 765 root = vm_page_splay(pindex, root); 766 if (pindex < root->pindex) { 767 m->left = root->left; 768 m->right = root; 769 root->left = NULL; 770 TAILQ_INSERT_BEFORE(root, m, listq); 771 } else if (pindex == root->pindex) 772 panic("vm_page_insert: offset already allocated"); 773 else { 774 m->right = root->right; 775 m->left = root; 776 root->right = NULL; 777 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 778 } 779 } 780 object->root = m; 781 object->generation++; 782 783 /* 784 * show that the object has one more resident page. 785 */ 786 object->resident_page_count++; 787 /* 788 * Hold the vnode until the last page is released. 789 */ 790 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 791 vhold((struct vnode *)object->handle); 792 793 /* 794 * Since we are inserting a new and possibly dirty page, 795 * update the object's OBJ_MIGHTBEDIRTY flag. 796 */ 797 if (m->flags & PG_WRITEABLE) 798 vm_object_set_writeable_dirty(object); 799} 800 801/* 802 * vm_page_remove: 803 * NOTE: used by device pager as well -wfj 804 * 805 * Removes the given mem entry from the object/offset-page 806 * table and the object page list, but do not invalidate/terminate 807 * the backing store. 808 * 809 * The object and page must be locked. 810 * The underlying pmap entry (if any) is NOT removed here. 811 * This routine may not block. 812 */ 813void 814vm_page_remove(vm_page_t m) 815{ 816 vm_object_t object; 817 vm_page_t root; 818 819 if ((m->flags & PG_UNMANAGED) == 0) 820 vm_page_lock_assert(m, MA_OWNED); 821 if ((object = m->object) == NULL) 822 return; 823 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 824 if (m->oflags & VPO_BUSY) { 825 m->oflags &= ~VPO_BUSY; 826 vm_page_flash(m); 827 } 828 829 /* 830 * Now remove from the object's list of backed pages. 831 */ 832 if (m != object->root) 833 vm_page_splay(m->pindex, object->root); 834 if (m->left == NULL) 835 root = m->right; 836 else { 837 root = vm_page_splay(m->pindex, m->left); 838 root->right = m->right; 839 } 840 object->root = root; 841 TAILQ_REMOVE(&object->memq, m, listq); 842 843 /* 844 * And show that the object has one fewer resident page. 845 */ 846 object->resident_page_count--; 847 object->generation++; 848 /* 849 * The vnode may now be recycled. 850 */ 851 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 852 vdrop((struct vnode *)object->handle); 853 854 m->object = NULL; 855} 856 857/* 858 * vm_page_lookup: 859 * 860 * Returns the page associated with the object/offset 861 * pair specified; if none is found, NULL is returned. 862 * 863 * The object must be locked. 864 * This routine may not block. 865 * This is a critical path routine 866 */ 867vm_page_t 868vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 869{ 870 vm_page_t m; 871 872 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 873 if ((m = object->root) != NULL && m->pindex != pindex) { 874 m = vm_page_splay(pindex, m); 875 if ((object->root = m)->pindex != pindex) 876 m = NULL; 877 } 878 return (m); 879} 880 881/* 882 * Returns the given page's successor (by pindex) within the object if it is 883 * resident; if none is found, NULL is returned. 884 * 885 * The object must be locked. 886 */ 887vm_page_t 888vm_page_next(vm_page_t m) 889{ 890 vm_page_t next; 891 892 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 893 if ((next = TAILQ_NEXT(m, listq)) != NULL && 894 next->pindex != m->pindex + 1) 895 next = NULL; 896 return (next); 897} 898 899/* 900 * Returns the given page's predecessor (by pindex) within the object if it is 901 * resident; if none is found, NULL is returned. 902 * 903 * The object must be locked. 904 */ 905vm_page_t 906vm_page_prev(vm_page_t m) 907{ 908 vm_page_t prev; 909 910 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 911 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 912 prev->pindex != m->pindex - 1) 913 prev = NULL; 914 return (prev); 915} 916 917/* 918 * vm_page_rename: 919 * 920 * Move the given memory entry from its 921 * current object to the specified target object/offset. 922 * 923 * The object must be locked. 924 * This routine may not block. 925 * 926 * Note: swap associated with the page must be invalidated by the move. We 927 * have to do this for several reasons: (1) we aren't freeing the 928 * page, (2) we are dirtying the page, (3) the VM system is probably 929 * moving the page from object A to B, and will then later move 930 * the backing store from A to B and we can't have a conflict. 931 * 932 * Note: we *always* dirty the page. It is necessary both for the 933 * fact that we moved it, and because we may be invalidating 934 * swap. If the page is on the cache, we have to deactivate it 935 * or vm_page_dirty() will panic. Dirty pages are not allowed 936 * on the cache. 937 */ 938void 939vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 940{ 941 942 vm_page_remove(m); 943 vm_page_insert(m, new_object, new_pindex); 944 vm_page_dirty(m); 945} 946 947/* 948 * Convert all of the given object's cached pages that have a 949 * pindex within the given range into free pages. If the value 950 * zero is given for "end", then the range's upper bound is 951 * infinity. If the given object is backed by a vnode and it 952 * transitions from having one or more cached pages to none, the 953 * vnode's hold count is reduced. 954 */ 955void 956vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 957{ 958 vm_page_t m, m_next; 959 boolean_t empty; 960 961 mtx_lock(&vm_page_queue_free_mtx); 962 if (__predict_false(object->cache == NULL)) { 963 mtx_unlock(&vm_page_queue_free_mtx); 964 return; 965 } 966 m = object->cache = vm_page_splay(start, object->cache); 967 if (m->pindex < start) { 968 if (m->right == NULL) 969 m = NULL; 970 else { 971 m_next = vm_page_splay(start, m->right); 972 m_next->left = m; 973 m->right = NULL; 974 m = object->cache = m_next; 975 } 976 } 977 978 /* 979 * At this point, "m" is either (1) a reference to the page 980 * with the least pindex that is greater than or equal to 981 * "start" or (2) NULL. 982 */ 983 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 984 /* 985 * Find "m"'s successor and remove "m" from the 986 * object's cache. 987 */ 988 if (m->right == NULL) { 989 object->cache = m->left; 990 m_next = NULL; 991 } else { 992 m_next = vm_page_splay(start, m->right); 993 m_next->left = m->left; 994 object->cache = m_next; 995 } 996 /* Convert "m" to a free page. */ 997 m->object = NULL; 998 m->valid = 0; 999 /* Clear PG_CACHED and set PG_FREE. */ 1000 m->flags ^= PG_CACHED | PG_FREE; 1001 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 1002 ("vm_page_cache_free: page %p has inconsistent flags", m)); 1003 cnt.v_cache_count--; 1004 cnt.v_free_count++; 1005 } 1006 empty = object->cache == NULL; 1007 mtx_unlock(&vm_page_queue_free_mtx); 1008 if (object->type == OBJT_VNODE && empty) 1009 vdrop(object->handle); 1010} 1011 1012/* 1013 * Returns the cached page that is associated with the given 1014 * object and offset. If, however, none exists, returns NULL. 1015 * 1016 * The free page queue must be locked. 1017 */ 1018static inline vm_page_t 1019vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1020{ 1021 vm_page_t m; 1022 1023 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1024 if ((m = object->cache) != NULL && m->pindex != pindex) { 1025 m = vm_page_splay(pindex, m); 1026 if ((object->cache = m)->pindex != pindex) 1027 m = NULL; 1028 } 1029 return (m); 1030} 1031 1032/* 1033 * Remove the given cached page from its containing object's 1034 * collection of cached pages. 1035 * 1036 * The free page queue must be locked. 1037 */ 1038void 1039vm_page_cache_remove(vm_page_t m) 1040{ 1041 vm_object_t object; 1042 vm_page_t root; 1043 1044 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1045 KASSERT((m->flags & PG_CACHED) != 0, 1046 ("vm_page_cache_remove: page %p is not cached", m)); 1047 object = m->object; 1048 if (m != object->cache) { 1049 root = vm_page_splay(m->pindex, object->cache); 1050 KASSERT(root == m, 1051 ("vm_page_cache_remove: page %p is not cached in object %p", 1052 m, object)); 1053 } 1054 if (m->left == NULL) 1055 root = m->right; 1056 else if (m->right == NULL) 1057 root = m->left; 1058 else { 1059 root = vm_page_splay(m->pindex, m->left); 1060 root->right = m->right; 1061 } 1062 object->cache = root; 1063 m->object = NULL; 1064 cnt.v_cache_count--; 1065} 1066 1067/* 1068 * Transfer all of the cached pages with offset greater than or 1069 * equal to 'offidxstart' from the original object's cache to the 1070 * new object's cache. However, any cached pages with offset 1071 * greater than or equal to the new object's size are kept in the 1072 * original object. Initially, the new object's cache must be 1073 * empty. Offset 'offidxstart' in the original object must 1074 * correspond to offset zero in the new object. 1075 * 1076 * The new object must be locked. 1077 */ 1078void 1079vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1080 vm_object_t new_object) 1081{ 1082 vm_page_t m, m_next; 1083 1084 /* 1085 * Insertion into an object's collection of cached pages 1086 * requires the object to be locked. In contrast, removal does 1087 * not. 1088 */ 1089 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 1090 KASSERT(new_object->cache == NULL, 1091 ("vm_page_cache_transfer: object %p has cached pages", 1092 new_object)); 1093 mtx_lock(&vm_page_queue_free_mtx); 1094 if ((m = orig_object->cache) != NULL) { 1095 /* 1096 * Transfer all of the pages with offset greater than or 1097 * equal to 'offidxstart' from the original object's 1098 * cache to the new object's cache. 1099 */ 1100 m = vm_page_splay(offidxstart, m); 1101 if (m->pindex < offidxstart) { 1102 orig_object->cache = m; 1103 new_object->cache = m->right; 1104 m->right = NULL; 1105 } else { 1106 orig_object->cache = m->left; 1107 new_object->cache = m; 1108 m->left = NULL; 1109 } 1110 while ((m = new_object->cache) != NULL) { 1111 if ((m->pindex - offidxstart) >= new_object->size) { 1112 /* 1113 * Return all of the cached pages with 1114 * offset greater than or equal to the 1115 * new object's size to the original 1116 * object's cache. 1117 */ 1118 new_object->cache = m->left; 1119 m->left = orig_object->cache; 1120 orig_object->cache = m; 1121 break; 1122 } 1123 m_next = vm_page_splay(m->pindex, m->right); 1124 /* Update the page's object and offset. */ 1125 m->object = new_object; 1126 m->pindex -= offidxstart; 1127 if (m_next == NULL) 1128 break; 1129 m->right = NULL; 1130 m_next->left = m; 1131 new_object->cache = m_next; 1132 } 1133 KASSERT(new_object->cache == NULL || 1134 new_object->type == OBJT_SWAP, 1135 ("vm_page_cache_transfer: object %p's type is incompatible" 1136 " with cached pages", new_object)); 1137 } 1138 mtx_unlock(&vm_page_queue_free_mtx); 1139} 1140 1141/* 1142 * vm_page_alloc: 1143 * 1144 * Allocate and return a memory cell associated 1145 * with this VM object/offset pair. 1146 * 1147 * page_req classes: 1148 * VM_ALLOC_NORMAL normal process request 1149 * VM_ALLOC_SYSTEM system *really* needs a page 1150 * VM_ALLOC_INTERRUPT interrupt time request 1151 * VM_ALLOC_ZERO zero page 1152 * VM_ALLOC_WIRED wire the allocated page 1153 * VM_ALLOC_NOOBJ page is not associated with a vm object 1154 * VM_ALLOC_NOBUSY do not set the page busy 1155 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1156 * is cached 1157 * 1158 * This routine may not sleep. 1159 */ 1160vm_page_t 1161vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1162{ 1163 struct vnode *vp = NULL; 1164 vm_object_t m_object; 1165 vm_page_t m; 1166 int flags, page_req; 1167 1168 page_req = req & VM_ALLOC_CLASS_MASK; 1169 KASSERT(curthread->td_intr_nesting_level == 0 || 1170 page_req == VM_ALLOC_INTERRUPT, 1171 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 1172 1173 if ((req & VM_ALLOC_NOOBJ) == 0) { 1174 KASSERT(object != NULL, 1175 ("vm_page_alloc: NULL object.")); 1176 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1177 } 1178 1179 /* 1180 * The pager is allowed to eat deeper into the free page list. 1181 */ 1182 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 1183 page_req = VM_ALLOC_SYSTEM; 1184 }; 1185 1186 mtx_lock(&vm_page_queue_free_mtx); 1187 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1188 (page_req == VM_ALLOC_SYSTEM && 1189 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1190 (page_req == VM_ALLOC_INTERRUPT && 1191 cnt.v_free_count + cnt.v_cache_count > 0)) { 1192 /* 1193 * Allocate from the free queue if the number of free pages 1194 * exceeds the minimum for the request class. 1195 */ 1196 if (object != NULL && 1197 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1198 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1199 mtx_unlock(&vm_page_queue_free_mtx); 1200 return (NULL); 1201 } 1202 if (vm_phys_unfree_page(m)) 1203 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1204#if VM_NRESERVLEVEL > 0 1205 else if (!vm_reserv_reactivate_page(m)) 1206#else 1207 else 1208#endif 1209 panic("vm_page_alloc: cache page %p is missing" 1210 " from the free queue", m); 1211 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1212 mtx_unlock(&vm_page_queue_free_mtx); 1213 return (NULL); 1214#if VM_NRESERVLEVEL > 0 1215 } else if (object == NULL || object->type == OBJT_DEVICE || 1216 object->type == OBJT_SG || 1217 (object->flags & OBJ_COLORED) == 0 || 1218 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1219#else 1220 } else { 1221#endif 1222 m = vm_phys_alloc_pages(object != NULL ? 1223 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1224#if VM_NRESERVLEVEL > 0 1225 if (m == NULL && vm_reserv_reclaim_inactive()) { 1226 m = vm_phys_alloc_pages(object != NULL ? 1227 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1228 0); 1229 } 1230#endif 1231 } 1232 } else { 1233 /* 1234 * Not allocatable, give up. 1235 */ 1236 mtx_unlock(&vm_page_queue_free_mtx); 1237 atomic_add_int(&vm_pageout_deficit, 1); 1238 pagedaemon_wakeup(); 1239 return (NULL); 1240 } 1241 1242 /* 1243 * At this point we had better have found a good page. 1244 */ 1245 1246 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1247 KASSERT(m->queue == PQ_NONE, 1248 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1249 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1250 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1251 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1252 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1253 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1254 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1255 pmap_page_get_memattr(m))); 1256 if ((m->flags & PG_CACHED) != 0) { 1257 KASSERT(m->valid != 0, 1258 ("vm_page_alloc: cached page %p is invalid", m)); 1259 if (m->object == object && m->pindex == pindex) 1260 cnt.v_reactivated++; 1261 else 1262 m->valid = 0; 1263 m_object = m->object; 1264 vm_page_cache_remove(m); 1265 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1266 vp = m_object->handle; 1267 } else { 1268 KASSERT(VM_PAGE_IS_FREE(m), 1269 ("vm_page_alloc: page %p is not free", m)); 1270 KASSERT(m->valid == 0, 1271 ("vm_page_alloc: free page %p is valid", m)); 1272 cnt.v_free_count--; 1273 } 1274 1275 /* 1276 * Initialize structure. Only the PG_ZERO flag is inherited. 1277 */ 1278 flags = 0; 1279 if (m->flags & PG_ZERO) { 1280 vm_page_zero_count--; 1281 if (req & VM_ALLOC_ZERO) 1282 flags = PG_ZERO; 1283 } 1284 if (object == NULL || object->type == OBJT_PHYS) 1285 flags |= PG_UNMANAGED; 1286 m->flags = flags; 1287 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 1288 m->oflags = 0; 1289 else 1290 m->oflags = VPO_BUSY; 1291 if (req & VM_ALLOC_WIRED) { 1292 atomic_add_int(&cnt.v_wire_count, 1); 1293 m->wire_count = 1; 1294 } 1295 m->act_count = 0; 1296 mtx_unlock(&vm_page_queue_free_mtx); 1297 1298 if (object != NULL) { 1299 /* Ignore device objects; the pager sets "memattr" for them. */ 1300 if (object->memattr != VM_MEMATTR_DEFAULT && 1301 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1302 pmap_page_set_memattr(m, object->memattr); 1303 vm_page_insert(m, object, pindex); 1304 } else 1305 m->pindex = pindex; 1306 1307 /* 1308 * The following call to vdrop() must come after the above call 1309 * to vm_page_insert() in case both affect the same object and 1310 * vnode. Otherwise, the affected vnode's hold count could 1311 * temporarily become zero. 1312 */ 1313 if (vp != NULL) 1314 vdrop(vp); 1315 1316 /* 1317 * Don't wakeup too often - wakeup the pageout daemon when 1318 * we would be nearly out of memory. 1319 */ 1320 if (vm_paging_needed()) 1321 pagedaemon_wakeup(); 1322 1323 return (m); 1324} 1325 1326/* 1327 * vm_wait: (also see VM_WAIT macro) 1328 * 1329 * Block until free pages are available for allocation 1330 * - Called in various places before memory allocations. 1331 */ 1332void 1333vm_wait(void) 1334{ 1335 1336 mtx_lock(&vm_page_queue_free_mtx); 1337 if (curproc == pageproc) { 1338 vm_pageout_pages_needed = 1; 1339 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1340 PDROP | PSWP, "VMWait", 0); 1341 } else { 1342 if (!vm_pages_needed) { 1343 vm_pages_needed = 1; 1344 wakeup(&vm_pages_needed); 1345 } 1346 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1347 "vmwait", 0); 1348 } 1349} 1350 1351/* 1352 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1353 * 1354 * Block until free pages are available for allocation 1355 * - Called only in vm_fault so that processes page faulting 1356 * can be easily tracked. 1357 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1358 * processes will be able to grab memory first. Do not change 1359 * this balance without careful testing first. 1360 */ 1361void 1362vm_waitpfault(void) 1363{ 1364 1365 mtx_lock(&vm_page_queue_free_mtx); 1366 if (!vm_pages_needed) { 1367 vm_pages_needed = 1; 1368 wakeup(&vm_pages_needed); 1369 } 1370 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1371 "pfault", 0); 1372} 1373 1374/* 1375 * vm_page_requeue: 1376 * 1377 * Move the given page to the tail of its present page queue. 1378 * 1379 * The page queues must be locked. 1380 */ 1381void 1382vm_page_requeue(vm_page_t m) 1383{ 1384 int queue = VM_PAGE_GETQUEUE(m); 1385 struct vpgqueues *vpq; 1386 1387 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1388 KASSERT(queue != PQ_NONE, 1389 ("vm_page_requeue: page %p is not queued", m)); 1390 vpq = &vm_page_queues[queue]; 1391 TAILQ_REMOVE(&vpq->pl, m, pageq); 1392 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1393} 1394 1395/* 1396 * vm_page_queue_remove: 1397 * 1398 * Remove the given page from the specified queue. 1399 * 1400 * The page and page queues must be locked. 1401 */ 1402static __inline void 1403vm_page_queue_remove(int queue, vm_page_t m) 1404{ 1405 struct vpgqueues *pq; 1406 1407 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1408 vm_page_lock_assert(m, MA_OWNED); 1409 pq = &vm_page_queues[queue]; 1410 TAILQ_REMOVE(&pq->pl, m, pageq); 1411 (*pq->cnt)--; 1412} 1413 1414/* 1415 * vm_pageq_remove: 1416 * 1417 * Remove a page from its queue. 1418 * 1419 * The given page must be locked. 1420 * This routine may not block. 1421 */ 1422void 1423vm_pageq_remove(vm_page_t m) 1424{ 1425 int queue = VM_PAGE_GETQUEUE(m); 1426 1427 vm_page_lock_assert(m, MA_OWNED); 1428 if (queue != PQ_NONE) { 1429 vm_page_lock_queues(); 1430 VM_PAGE_SETQUEUE2(m, PQ_NONE); 1431 vm_page_queue_remove(queue, m); 1432 vm_page_unlock_queues(); 1433 } 1434} 1435 1436/* 1437 * vm_page_enqueue: 1438 * 1439 * Add the given page to the specified queue. 1440 * 1441 * The page queues must be locked. 1442 */ 1443static void 1444vm_page_enqueue(int queue, vm_page_t m) 1445{ 1446 struct vpgqueues *vpq; 1447 1448 vpq = &vm_page_queues[queue]; 1449 VM_PAGE_SETQUEUE2(m, queue); 1450 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1451 ++*vpq->cnt; 1452} 1453 1454/* 1455 * vm_page_activate: 1456 * 1457 * Put the specified page on the active list (if appropriate). 1458 * Ensure that act_count is at least ACT_INIT but do not otherwise 1459 * mess with it. 1460 * 1461 * The page must be locked. 1462 * This routine may not block. 1463 */ 1464void 1465vm_page_activate(vm_page_t m) 1466{ 1467 int queue; 1468 1469 vm_page_lock_assert(m, MA_OWNED); 1470 if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) != PQ_ACTIVE) { 1471 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1472 if (m->act_count < ACT_INIT) 1473 m->act_count = ACT_INIT; 1474 vm_page_lock_queues(); 1475 if (queue != PQ_NONE) 1476 vm_page_queue_remove(queue, m); 1477 vm_page_enqueue(PQ_ACTIVE, m); 1478 vm_page_unlock_queues(); 1479 } else 1480 KASSERT(queue == PQ_NONE, 1481 ("vm_page_activate: wired page %p is queued", m)); 1482 } else { 1483 if (m->act_count < ACT_INIT) 1484 m->act_count = ACT_INIT; 1485 } 1486} 1487 1488/* 1489 * vm_page_free_wakeup: 1490 * 1491 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1492 * routine is called when a page has been added to the cache or free 1493 * queues. 1494 * 1495 * The page queues must be locked. 1496 * This routine may not block. 1497 */ 1498static inline void 1499vm_page_free_wakeup(void) 1500{ 1501 1502 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1503 /* 1504 * if pageout daemon needs pages, then tell it that there are 1505 * some free. 1506 */ 1507 if (vm_pageout_pages_needed && 1508 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1509 wakeup(&vm_pageout_pages_needed); 1510 vm_pageout_pages_needed = 0; 1511 } 1512 /* 1513 * wakeup processes that are waiting on memory if we hit a 1514 * high water mark. And wakeup scheduler process if we have 1515 * lots of memory. this process will swapin processes. 1516 */ 1517 if (vm_pages_needed && !vm_page_count_min()) { 1518 vm_pages_needed = 0; 1519 wakeup(&cnt.v_free_count); 1520 } 1521} 1522 1523/* 1524 * vm_page_free_toq: 1525 * 1526 * Returns the given page to the free list, 1527 * disassociating it with any VM object. 1528 * 1529 * Object and page must be locked prior to entry. 1530 * This routine may not block. 1531 */ 1532 1533void 1534vm_page_free_toq(vm_page_t m) 1535{ 1536 1537 if ((m->flags & PG_UNMANAGED) == 0) { 1538 vm_page_lock_assert(m, MA_OWNED); 1539 KASSERT(!pmap_page_is_mapped(m), 1540 ("vm_page_free_toq: freeing mapped page %p", m)); 1541 } 1542 PCPU_INC(cnt.v_tfree); 1543 1544 if (m->busy || VM_PAGE_IS_FREE(m)) { 1545 printf( 1546 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1547 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1548 m->hold_count); 1549 if (VM_PAGE_IS_FREE(m)) 1550 panic("vm_page_free: freeing free page"); 1551 else 1552 panic("vm_page_free: freeing busy page"); 1553 } 1554 1555 /* 1556 * unqueue, then remove page. Note that we cannot destroy 1557 * the page here because we do not want to call the pager's 1558 * callback routine until after we've put the page on the 1559 * appropriate free queue. 1560 */ 1561 if ((m->flags & PG_UNMANAGED) == 0) 1562 vm_pageq_remove(m); 1563 vm_page_remove(m); 1564 1565 /* 1566 * If fictitious remove object association and 1567 * return, otherwise delay object association removal. 1568 */ 1569 if ((m->flags & PG_FICTITIOUS) != 0) { 1570 return; 1571 } 1572 1573 m->valid = 0; 1574 vm_page_undirty(m); 1575 1576 if (m->wire_count != 0) { 1577 if (m->wire_count > 1) { 1578 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1579 m->wire_count, (long)m->pindex); 1580 } 1581 panic("vm_page_free: freeing wired page"); 1582 } 1583 if (m->hold_count != 0) { 1584 m->flags &= ~PG_ZERO; 1585 vm_page_lock_queues(); 1586 vm_page_enqueue(PQ_HOLD, m); 1587 vm_page_unlock_queues(); 1588 } else { 1589 /* 1590 * Restore the default memory attribute to the page. 1591 */ 1592 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1593 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1594 1595 /* 1596 * Insert the page into the physical memory allocator's 1597 * cache/free page queues. 1598 */ 1599 mtx_lock(&vm_page_queue_free_mtx); 1600 m->flags |= PG_FREE; 1601 cnt.v_free_count++; 1602#if VM_NRESERVLEVEL > 0 1603 if (!vm_reserv_free_page(m)) 1604#else 1605 if (TRUE) 1606#endif 1607 vm_phys_free_pages(m, 0); 1608 if ((m->flags & PG_ZERO) != 0) 1609 ++vm_page_zero_count; 1610 else 1611 vm_page_zero_idle_wakeup(); 1612 vm_page_free_wakeup(); 1613 mtx_unlock(&vm_page_queue_free_mtx); 1614 } 1615} 1616 1617/* 1618 * vm_page_wire: 1619 * 1620 * Mark this page as wired down by yet 1621 * another map, removing it from paging queues 1622 * as necessary. 1623 * 1624 * If the page is fictitious, then its wire count must remain one. 1625 * 1626 * The page must be locked. 1627 * This routine may not block. 1628 */ 1629void 1630vm_page_wire(vm_page_t m) 1631{ 1632 1633 /* 1634 * Only bump the wire statistics if the page is not already wired, 1635 * and only unqueue the page if it is on some queue (if it is unmanaged 1636 * it is already off the queues). 1637 */ 1638 vm_page_lock_assert(m, MA_OWNED); 1639 if ((m->flags & PG_FICTITIOUS) != 0) { 1640 KASSERT(m->wire_count == 1, 1641 ("vm_page_wire: fictitious page %p's wire count isn't one", 1642 m)); 1643 return; 1644 } 1645 if (m->wire_count == 0) { 1646 if ((m->flags & PG_UNMANAGED) == 0) 1647 vm_pageq_remove(m); 1648 atomic_add_int(&cnt.v_wire_count, 1); 1649 } 1650 m->wire_count++; 1651 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1652} 1653 1654/* 1655 * vm_page_unwire: 1656 * 1657 * Release one wiring of the specified page, potentially enabling it to be 1658 * paged again. If paging is enabled, then the value of the parameter 1659 * "activate" determines to which queue the page is added. If "activate" is 1660 * non-zero, then the page is added to the active queue. Otherwise, it is 1661 * added to the inactive queue. 1662 * 1663 * However, unless the page belongs to an object, it is not enqueued because 1664 * it cannot be paged out. 1665 * 1666 * If a page is fictitious, then its wire count must alway be one. 1667 * 1668 * A managed page must be locked. 1669 */ 1670void 1671vm_page_unwire(vm_page_t m, int activate) 1672{ 1673 1674 if ((m->flags & PG_UNMANAGED) == 0) 1675 vm_page_lock_assert(m, MA_OWNED); 1676 if ((m->flags & PG_FICTITIOUS) != 0) { 1677 KASSERT(m->wire_count == 1, 1678 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 1679 return; 1680 } 1681 if (m->wire_count > 0) { 1682 m->wire_count--; 1683 if (m->wire_count == 0) { 1684 atomic_subtract_int(&cnt.v_wire_count, 1); 1685 if ((m->flags & PG_UNMANAGED) != 0 || 1686 m->object == NULL) 1687 return; 1688 vm_page_lock_queues(); 1689 if (activate) 1690 vm_page_enqueue(PQ_ACTIVE, m); 1691 else { 1692 vm_page_flag_clear(m, PG_WINATCFLS); 1693 vm_page_enqueue(PQ_INACTIVE, m); 1694 } 1695 vm_page_unlock_queues(); 1696 } 1697 } else 1698 panic("vm_page_unwire: page %p's wire count is zero", m); 1699} 1700 1701/* 1702 * Move the specified page to the inactive queue. 1703 * 1704 * Many pages placed on the inactive queue should actually go 1705 * into the cache, but it is difficult to figure out which. What 1706 * we do instead, if the inactive target is well met, is to put 1707 * clean pages at the head of the inactive queue instead of the tail. 1708 * This will cause them to be moved to the cache more quickly and 1709 * if not actively re-referenced, reclaimed more quickly. If we just 1710 * stick these pages at the end of the inactive queue, heavy filesystem 1711 * meta-data accesses can cause an unnecessary paging load on memory bound 1712 * processes. This optimization causes one-time-use metadata to be 1713 * reused more quickly. 1714 * 1715 * Normally athead is 0 resulting in LRU operation. athead is set 1716 * to 1 if we want this page to be 'as if it were placed in the cache', 1717 * except without unmapping it from the process address space. 1718 * 1719 * This routine may not block. 1720 */ 1721static inline void 1722_vm_page_deactivate(vm_page_t m, int athead) 1723{ 1724 int queue; 1725 1726 vm_page_lock_assert(m, MA_OWNED); 1727 1728 /* 1729 * Ignore if already inactive. 1730 */ 1731 if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) == PQ_INACTIVE) 1732 return; 1733 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1734 vm_page_lock_queues(); 1735 vm_page_flag_clear(m, PG_WINATCFLS); 1736 if (queue != PQ_NONE) 1737 vm_page_queue_remove(queue, m); 1738 if (athead) 1739 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, 1740 pageq); 1741 else 1742 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, 1743 pageq); 1744 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1745 cnt.v_inactive_count++; 1746 vm_page_unlock_queues(); 1747 } 1748} 1749 1750/* 1751 * Move the specified page to the inactive queue. 1752 * 1753 * The page must be locked. 1754 */ 1755void 1756vm_page_deactivate(vm_page_t m) 1757{ 1758 1759 _vm_page_deactivate(m, 0); 1760} 1761 1762/* 1763 * vm_page_try_to_cache: 1764 * 1765 * Returns 0 on failure, 1 on success 1766 */ 1767int 1768vm_page_try_to_cache(vm_page_t m) 1769{ 1770 1771 vm_page_lock_assert(m, MA_OWNED); 1772 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1773 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1774 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1775 return (0); 1776 pmap_remove_all(m); 1777 if (m->dirty) 1778 return (0); 1779 vm_page_cache(m); 1780 return (1); 1781} 1782 1783/* 1784 * vm_page_try_to_free() 1785 * 1786 * Attempt to free the page. If we cannot free it, we do nothing. 1787 * 1 is returned on success, 0 on failure. 1788 */ 1789int 1790vm_page_try_to_free(vm_page_t m) 1791{ 1792 1793 vm_page_lock_assert(m, MA_OWNED); 1794 if (m->object != NULL) 1795 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1796 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1797 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1798 return (0); 1799 pmap_remove_all(m); 1800 if (m->dirty) 1801 return (0); 1802 vm_page_free(m); 1803 return (1); 1804} 1805 1806/* 1807 * vm_page_cache 1808 * 1809 * Put the specified page onto the page cache queue (if appropriate). 1810 * 1811 * This routine may not block. 1812 */ 1813void 1814vm_page_cache(vm_page_t m) 1815{ 1816 vm_object_t object; 1817 vm_page_t root; 1818 1819 vm_page_lock_assert(m, MA_OWNED); 1820 object = m->object; 1821 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1822 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1823 m->hold_count || m->wire_count) 1824 panic("vm_page_cache: attempting to cache busy page"); 1825 pmap_remove_all(m); 1826 if (m->dirty != 0) 1827 panic("vm_page_cache: page %p is dirty", m); 1828 if (m->valid == 0 || object->type == OBJT_DEFAULT || 1829 (object->type == OBJT_SWAP && 1830 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 1831 /* 1832 * Hypothesis: A cache-elgible page belonging to a 1833 * default object or swap object but without a backing 1834 * store must be zero filled. 1835 */ 1836 vm_page_free(m); 1837 return; 1838 } 1839 KASSERT((m->flags & PG_CACHED) == 0, 1840 ("vm_page_cache: page %p is already cached", m)); 1841 PCPU_INC(cnt.v_tcached); 1842 1843 /* 1844 * Remove the page from the paging queues. 1845 */ 1846 vm_pageq_remove(m); 1847 1848 /* 1849 * Remove the page from the object's collection of resident 1850 * pages. 1851 */ 1852 if (m != object->root) 1853 vm_page_splay(m->pindex, object->root); 1854 if (m->left == NULL) 1855 root = m->right; 1856 else { 1857 root = vm_page_splay(m->pindex, m->left); 1858 root->right = m->right; 1859 } 1860 object->root = root; 1861 TAILQ_REMOVE(&object->memq, m, listq); 1862 object->resident_page_count--; 1863 object->generation++; 1864 1865 /* 1866 * Restore the default memory attribute to the page. 1867 */ 1868 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1869 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1870 1871 /* 1872 * Insert the page into the object's collection of cached pages 1873 * and the physical memory allocator's cache/free page queues. 1874 */ 1875 m->flags &= ~PG_ZERO; 1876 mtx_lock(&vm_page_queue_free_mtx); 1877 m->flags |= PG_CACHED; 1878 cnt.v_cache_count++; 1879 root = object->cache; 1880 if (root == NULL) { 1881 m->left = NULL; 1882 m->right = NULL; 1883 } else { 1884 root = vm_page_splay(m->pindex, root); 1885 if (m->pindex < root->pindex) { 1886 m->left = root->left; 1887 m->right = root; 1888 root->left = NULL; 1889 } else if (__predict_false(m->pindex == root->pindex)) 1890 panic("vm_page_cache: offset already cached"); 1891 else { 1892 m->right = root->right; 1893 m->left = root; 1894 root->right = NULL; 1895 } 1896 } 1897 object->cache = m; 1898#if VM_NRESERVLEVEL > 0 1899 if (!vm_reserv_free_page(m)) { 1900#else 1901 if (TRUE) { 1902#endif 1903 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 1904 vm_phys_free_pages(m, 0); 1905 } 1906 vm_page_free_wakeup(); 1907 mtx_unlock(&vm_page_queue_free_mtx); 1908 1909 /* 1910 * Increment the vnode's hold count if this is the object's only 1911 * cached page. Decrement the vnode's hold count if this was 1912 * the object's only resident page. 1913 */ 1914 if (object->type == OBJT_VNODE) { 1915 if (root == NULL && object->resident_page_count != 0) 1916 vhold(object->handle); 1917 else if (root != NULL && object->resident_page_count == 0) 1918 vdrop(object->handle); 1919 } 1920} 1921 1922/* 1923 * vm_page_dontneed 1924 * 1925 * Cache, deactivate, or do nothing as appropriate. This routine 1926 * is typically used by madvise() MADV_DONTNEED. 1927 * 1928 * Generally speaking we want to move the page into the cache so 1929 * it gets reused quickly. However, this can result in a silly syndrome 1930 * due to the page recycling too quickly. Small objects will not be 1931 * fully cached. On the otherhand, if we move the page to the inactive 1932 * queue we wind up with a problem whereby very large objects 1933 * unnecessarily blow away our inactive and cache queues. 1934 * 1935 * The solution is to move the pages based on a fixed weighting. We 1936 * either leave them alone, deactivate them, or move them to the cache, 1937 * where moving them to the cache has the highest weighting. 1938 * By forcing some pages into other queues we eventually force the 1939 * system to balance the queues, potentially recovering other unrelated 1940 * space from active. The idea is to not force this to happen too 1941 * often. 1942 */ 1943void 1944vm_page_dontneed(vm_page_t m) 1945{ 1946 int dnw; 1947 int head; 1948 1949 vm_page_lock_assert(m, MA_OWNED); 1950 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1951 dnw = PCPU_GET(dnweight); 1952 PCPU_INC(dnweight); 1953 1954 /* 1955 * Occasionally leave the page alone. 1956 */ 1957 if ((dnw & 0x01F0) == 0 || 1958 VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) { 1959 if (m->act_count >= ACT_INIT) 1960 --m->act_count; 1961 return; 1962 } 1963 1964 /* 1965 * Clear any references to the page. Otherwise, the page daemon will 1966 * immediately reactivate the page. 1967 * 1968 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 1969 * pmap operation, such as pmap_remove(), could clear a reference in 1970 * the pmap and set PG_REFERENCED on the page before the 1971 * pmap_clear_reference() had completed. Consequently, the page would 1972 * appear referenced based upon an old reference that occurred before 1973 * this function ran. 1974 */ 1975 pmap_clear_reference(m); 1976 vm_page_lock_queues(); 1977 vm_page_flag_clear(m, PG_REFERENCED); 1978 vm_page_unlock_queues(); 1979 1980 if (m->dirty == 0 && pmap_is_modified(m)) 1981 vm_page_dirty(m); 1982 1983 if (m->dirty || (dnw & 0x0070) == 0) { 1984 /* 1985 * Deactivate the page 3 times out of 32. 1986 */ 1987 head = 0; 1988 } else { 1989 /* 1990 * Cache the page 28 times out of every 32. Note that 1991 * the page is deactivated instead of cached, but placed 1992 * at the head of the queue instead of the tail. 1993 */ 1994 head = 1; 1995 } 1996 _vm_page_deactivate(m, head); 1997} 1998 1999/* 2000 * Grab a page, waiting until we are waken up due to the page 2001 * changing state. We keep on waiting, if the page continues 2002 * to be in the object. If the page doesn't exist, first allocate it 2003 * and then conditionally zero it. 2004 * 2005 * This routine may block. 2006 */ 2007vm_page_t 2008vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2009{ 2010 vm_page_t m; 2011 2012 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2013retrylookup: 2014 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2015 if ((m->oflags & VPO_BUSY) != 0 || m->busy != 0) { 2016 if ((allocflags & VM_ALLOC_RETRY) != 0) { 2017 /* 2018 * Reference the page before unlocking and 2019 * sleeping so that the page daemon is less 2020 * likely to reclaim it. 2021 */ 2022 vm_page_lock_queues(); 2023 vm_page_flag_set(m, PG_REFERENCED); 2024 } 2025 vm_page_sleep(m, "pgrbwt"); 2026 if ((allocflags & VM_ALLOC_RETRY) == 0) 2027 return (NULL); 2028 goto retrylookup; 2029 } else { 2030 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2031 vm_page_lock(m); 2032 vm_page_wire(m); 2033 vm_page_unlock(m); 2034 } 2035 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 2036 vm_page_busy(m); 2037 return (m); 2038 } 2039 } 2040 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 2041 if (m == NULL) { 2042 VM_OBJECT_UNLOCK(object); 2043 VM_WAIT; 2044 VM_OBJECT_LOCK(object); 2045 if ((allocflags & VM_ALLOC_RETRY) == 0) 2046 return (NULL); 2047 goto retrylookup; 2048 } else if (m->valid != 0) 2049 return (m); 2050 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2051 pmap_zero_page(m); 2052 return (m); 2053} 2054 2055/* 2056 * Mapping function for valid bits or for dirty bits in 2057 * a page. May not block. 2058 * 2059 * Inputs are required to range within a page. 2060 */ 2061int 2062vm_page_bits(int base, int size) 2063{ 2064 int first_bit; 2065 int last_bit; 2066 2067 KASSERT( 2068 base + size <= PAGE_SIZE, 2069 ("vm_page_bits: illegal base/size %d/%d", base, size) 2070 ); 2071 2072 if (size == 0) /* handle degenerate case */ 2073 return (0); 2074 2075 first_bit = base >> DEV_BSHIFT; 2076 last_bit = (base + size - 1) >> DEV_BSHIFT; 2077 2078 return ((2 << last_bit) - (1 << first_bit)); 2079} 2080 2081/* 2082 * vm_page_set_valid: 2083 * 2084 * Sets portions of a page valid. The arguments are expected 2085 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2086 * of any partial chunks touched by the range. The invalid portion of 2087 * such chunks will be zeroed. 2088 * 2089 * (base + size) must be less then or equal to PAGE_SIZE. 2090 */ 2091void 2092vm_page_set_valid(vm_page_t m, int base, int size) 2093{ 2094 int endoff, frag; 2095 2096 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2097 if (size == 0) /* handle degenerate case */ 2098 return; 2099 2100 /* 2101 * If the base is not DEV_BSIZE aligned and the valid 2102 * bit is clear, we have to zero out a portion of the 2103 * first block. 2104 */ 2105 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2106 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2107 pmap_zero_page_area(m, frag, base - frag); 2108 2109 /* 2110 * If the ending offset is not DEV_BSIZE aligned and the 2111 * valid bit is clear, we have to zero out a portion of 2112 * the last block. 2113 */ 2114 endoff = base + size; 2115 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2116 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2117 pmap_zero_page_area(m, endoff, 2118 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2119 2120 /* 2121 * Assert that no previously invalid block that is now being validated 2122 * is already dirty. 2123 */ 2124 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2125 ("vm_page_set_valid: page %p is dirty", m)); 2126 2127 /* 2128 * Set valid bits inclusive of any overlap. 2129 */ 2130 m->valid |= vm_page_bits(base, size); 2131} 2132 2133/* 2134 * Clear the given bits from the specified page's dirty field. 2135 */ 2136static __inline void 2137vm_page_clear_dirty_mask(vm_page_t m, int pagebits) 2138{ 2139 2140 /* 2141 * If the object is locked and the page is neither VPO_BUSY nor 2142 * PG_WRITEABLE, then the page's dirty field cannot possibly be 2143 * modified by a concurrent pmap operation. 2144 */ 2145 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2146 if ((m->oflags & VPO_BUSY) == 0 && (m->flags & PG_WRITEABLE) == 0) 2147 m->dirty &= ~pagebits; 2148 else { 2149 vm_page_lock_queues(); 2150 m->dirty &= ~pagebits; 2151 vm_page_unlock_queues(); 2152 } 2153} 2154 2155/* 2156 * vm_page_set_validclean: 2157 * 2158 * Sets portions of a page valid and clean. The arguments are expected 2159 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2160 * of any partial chunks touched by the range. The invalid portion of 2161 * such chunks will be zero'd. 2162 * 2163 * This routine may not block. 2164 * 2165 * (base + size) must be less then or equal to PAGE_SIZE. 2166 */ 2167void 2168vm_page_set_validclean(vm_page_t m, int base, int size) 2169{ 2170 u_long oldvalid; 2171 int endoff, frag, pagebits; 2172 2173 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2174 if (size == 0) /* handle degenerate case */ 2175 return; 2176 2177 /* 2178 * If the base is not DEV_BSIZE aligned and the valid 2179 * bit is clear, we have to zero out a portion of the 2180 * first block. 2181 */ 2182 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2183 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2184 pmap_zero_page_area(m, frag, base - frag); 2185 2186 /* 2187 * If the ending offset is not DEV_BSIZE aligned and the 2188 * valid bit is clear, we have to zero out a portion of 2189 * the last block. 2190 */ 2191 endoff = base + size; 2192 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2193 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2194 pmap_zero_page_area(m, endoff, 2195 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2196 2197 /* 2198 * Set valid, clear dirty bits. If validating the entire 2199 * page we can safely clear the pmap modify bit. We also 2200 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2201 * takes a write fault on a MAP_NOSYNC memory area the flag will 2202 * be set again. 2203 * 2204 * We set valid bits inclusive of any overlap, but we can only 2205 * clear dirty bits for DEV_BSIZE chunks that are fully within 2206 * the range. 2207 */ 2208 oldvalid = m->valid; 2209 pagebits = vm_page_bits(base, size); 2210 m->valid |= pagebits; 2211#if 0 /* NOT YET */ 2212 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2213 frag = DEV_BSIZE - frag; 2214 base += frag; 2215 size -= frag; 2216 if (size < 0) 2217 size = 0; 2218 } 2219 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2220#endif 2221 if (base == 0 && size == PAGE_SIZE) { 2222 /* 2223 * The page can only be modified within the pmap if it is 2224 * mapped, and it can only be mapped if it was previously 2225 * fully valid. 2226 */ 2227 if (oldvalid == VM_PAGE_BITS_ALL) 2228 /* 2229 * Perform the pmap_clear_modify() first. Otherwise, 2230 * a concurrent pmap operation, such as 2231 * pmap_protect(), could clear a modification in the 2232 * pmap and set the dirty field on the page before 2233 * pmap_clear_modify() had begun and after the dirty 2234 * field was cleared here. 2235 */ 2236 pmap_clear_modify(m); 2237 m->dirty = 0; 2238 m->oflags &= ~VPO_NOSYNC; 2239 } else if (oldvalid != VM_PAGE_BITS_ALL) 2240 m->dirty &= ~pagebits; 2241 else 2242 vm_page_clear_dirty_mask(m, pagebits); 2243} 2244 2245void 2246vm_page_clear_dirty(vm_page_t m, int base, int size) 2247{ 2248 2249 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2250} 2251 2252/* 2253 * vm_page_set_invalid: 2254 * 2255 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2256 * valid and dirty bits for the effected areas are cleared. 2257 * 2258 * May not block. 2259 */ 2260void 2261vm_page_set_invalid(vm_page_t m, int base, int size) 2262{ 2263 int bits; 2264 2265 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2266 KASSERT((m->oflags & VPO_BUSY) == 0, 2267 ("vm_page_set_invalid: page %p is busy", m)); 2268 bits = vm_page_bits(base, size); 2269 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2270 pmap_remove_all(m); 2271 KASSERT(!pmap_page_is_mapped(m), 2272 ("vm_page_set_invalid: page %p is mapped", m)); 2273 m->valid &= ~bits; 2274 m->dirty &= ~bits; 2275 m->object->generation++; 2276} 2277 2278/* 2279 * vm_page_zero_invalid() 2280 * 2281 * The kernel assumes that the invalid portions of a page contain 2282 * garbage, but such pages can be mapped into memory by user code. 2283 * When this occurs, we must zero out the non-valid portions of the 2284 * page so user code sees what it expects. 2285 * 2286 * Pages are most often semi-valid when the end of a file is mapped 2287 * into memory and the file's size is not page aligned. 2288 */ 2289void 2290vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2291{ 2292 int b; 2293 int i; 2294 2295 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2296 /* 2297 * Scan the valid bits looking for invalid sections that 2298 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2299 * valid bit may be set ) have already been zerod by 2300 * vm_page_set_validclean(). 2301 */ 2302 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2303 if (i == (PAGE_SIZE / DEV_BSIZE) || 2304 (m->valid & (1 << i)) 2305 ) { 2306 if (i > b) { 2307 pmap_zero_page_area(m, 2308 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2309 } 2310 b = i + 1; 2311 } 2312 } 2313 2314 /* 2315 * setvalid is TRUE when we can safely set the zero'd areas 2316 * as being valid. We can do this if there are no cache consistancy 2317 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2318 */ 2319 if (setvalid) 2320 m->valid = VM_PAGE_BITS_ALL; 2321} 2322 2323/* 2324 * vm_page_is_valid: 2325 * 2326 * Is (partial) page valid? Note that the case where size == 0 2327 * will return FALSE in the degenerate case where the page is 2328 * entirely invalid, and TRUE otherwise. 2329 * 2330 * May not block. 2331 */ 2332int 2333vm_page_is_valid(vm_page_t m, int base, int size) 2334{ 2335 int bits = vm_page_bits(base, size); 2336 2337 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2338 if (m->valid && ((m->valid & bits) == bits)) 2339 return 1; 2340 else 2341 return 0; 2342} 2343 2344/* 2345 * update dirty bits from pmap/mmu. May not block. 2346 */ 2347void 2348vm_page_test_dirty(vm_page_t m) 2349{ 2350 2351 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2352 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2353 vm_page_dirty(m); 2354} 2355 2356int so_zerocp_fullpage = 0; 2357 2358/* 2359 * Replace the given page with a copy. The copied page assumes 2360 * the portion of the given page's "wire_count" that is not the 2361 * responsibility of this copy-on-write mechanism. 2362 * 2363 * The object containing the given page must have a non-zero 2364 * paging-in-progress count and be locked. 2365 */ 2366void 2367vm_page_cowfault(vm_page_t m) 2368{ 2369 vm_page_t mnew; 2370 vm_object_t object; 2371 vm_pindex_t pindex; 2372 2373 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 2374 vm_page_lock_assert(m, MA_OWNED); 2375 object = m->object; 2376 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2377 KASSERT(object->paging_in_progress != 0, 2378 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2379 object)); 2380 pindex = m->pindex; 2381 2382 retry_alloc: 2383 pmap_remove_all(m); 2384 vm_page_remove(m); 2385 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2386 if (mnew == NULL) { 2387 vm_page_insert(m, object, pindex); 2388 vm_page_unlock(m); 2389 VM_OBJECT_UNLOCK(object); 2390 VM_WAIT; 2391 VM_OBJECT_LOCK(object); 2392 if (m == vm_page_lookup(object, pindex)) { 2393 vm_page_lock(m); 2394 goto retry_alloc; 2395 } else { 2396 /* 2397 * Page disappeared during the wait. 2398 */ 2399 return; 2400 } 2401 } 2402 2403 if (m->cow == 0) { 2404 /* 2405 * check to see if we raced with an xmit complete when 2406 * waiting to allocate a page. If so, put things back 2407 * the way they were 2408 */ 2409 vm_page_unlock(m); 2410 vm_page_lock(mnew); 2411 vm_page_free(mnew); 2412 vm_page_unlock(mnew); 2413 vm_page_insert(m, object, pindex); 2414 } else { /* clear COW & copy page */ 2415 if (!so_zerocp_fullpage) 2416 pmap_copy_page(m, mnew); 2417 mnew->valid = VM_PAGE_BITS_ALL; 2418 vm_page_dirty(mnew); 2419 mnew->wire_count = m->wire_count - m->cow; 2420 m->wire_count = m->cow; 2421 vm_page_unlock(m); 2422 } 2423} 2424 2425void 2426vm_page_cowclear(vm_page_t m) 2427{ 2428 2429 vm_page_lock_assert(m, MA_OWNED); 2430 if (m->cow) { 2431 m->cow--; 2432 /* 2433 * let vm_fault add back write permission lazily 2434 */ 2435 } 2436 /* 2437 * sf_buf_free() will free the page, so we needn't do it here 2438 */ 2439} 2440 2441int 2442vm_page_cowsetup(vm_page_t m) 2443{ 2444 2445 vm_page_lock_assert(m, MA_OWNED); 2446 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 || 2447 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 2448 return (EBUSY); 2449 m->cow++; 2450 pmap_remove_write(m); 2451 VM_OBJECT_UNLOCK(m->object); 2452 return (0); 2453} 2454 2455#include "opt_ddb.h" 2456#ifdef DDB 2457#include <sys/kernel.h> 2458 2459#include <ddb/ddb.h> 2460 2461DB_SHOW_COMMAND(page, vm_page_print_page_info) 2462{ 2463 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 2464 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 2465 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 2466 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 2467 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 2468 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 2469 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 2470 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 2471 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 2472 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 2473} 2474 2475DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2476{ 2477 2478 db_printf("PQ_FREE:"); 2479 db_printf(" %d", cnt.v_free_count); 2480 db_printf("\n"); 2481 2482 db_printf("PQ_CACHE:"); 2483 db_printf(" %d", cnt.v_cache_count); 2484 db_printf("\n"); 2485 2486 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2487 *vm_page_queues[PQ_ACTIVE].cnt, 2488 *vm_page_queues[PQ_INACTIVE].cnt); 2489} 2490#endif /* DDB */ 2491