vm_page.c revision 209173
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 209173 2010-06-14 19:54:19Z 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 * vm_page_rename: 883 * 884 * Move the given memory entry from its 885 * current object to the specified target object/offset. 886 * 887 * The object must be locked. 888 * This routine may not block. 889 * 890 * Note: swap associated with the page must be invalidated by the move. We 891 * have to do this for several reasons: (1) we aren't freeing the 892 * page, (2) we are dirtying the page, (3) the VM system is probably 893 * moving the page from object A to B, and will then later move 894 * the backing store from A to B and we can't have a conflict. 895 * 896 * Note: we *always* dirty the page. It is necessary both for the 897 * fact that we moved it, and because we may be invalidating 898 * swap. If the page is on the cache, we have to deactivate it 899 * or vm_page_dirty() will panic. Dirty pages are not allowed 900 * on the cache. 901 */ 902void 903vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 904{ 905 906 vm_page_remove(m); 907 vm_page_insert(m, new_object, new_pindex); 908 vm_page_dirty(m); 909} 910 911/* 912 * Convert all of the given object's cached pages that have a 913 * pindex within the given range into free pages. If the value 914 * zero is given for "end", then the range's upper bound is 915 * infinity. If the given object is backed by a vnode and it 916 * transitions from having one or more cached pages to none, the 917 * vnode's hold count is reduced. 918 */ 919void 920vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 921{ 922 vm_page_t m, m_next; 923 boolean_t empty; 924 925 mtx_lock(&vm_page_queue_free_mtx); 926 if (__predict_false(object->cache == NULL)) { 927 mtx_unlock(&vm_page_queue_free_mtx); 928 return; 929 } 930 m = object->cache = vm_page_splay(start, object->cache); 931 if (m->pindex < start) { 932 if (m->right == NULL) 933 m = NULL; 934 else { 935 m_next = vm_page_splay(start, m->right); 936 m_next->left = m; 937 m->right = NULL; 938 m = object->cache = m_next; 939 } 940 } 941 942 /* 943 * At this point, "m" is either (1) a reference to the page 944 * with the least pindex that is greater than or equal to 945 * "start" or (2) NULL. 946 */ 947 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 948 /* 949 * Find "m"'s successor and remove "m" from the 950 * object's cache. 951 */ 952 if (m->right == NULL) { 953 object->cache = m->left; 954 m_next = NULL; 955 } else { 956 m_next = vm_page_splay(start, m->right); 957 m_next->left = m->left; 958 object->cache = m_next; 959 } 960 /* Convert "m" to a free page. */ 961 m->object = NULL; 962 m->valid = 0; 963 /* Clear PG_CACHED and set PG_FREE. */ 964 m->flags ^= PG_CACHED | PG_FREE; 965 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 966 ("vm_page_cache_free: page %p has inconsistent flags", m)); 967 cnt.v_cache_count--; 968 cnt.v_free_count++; 969 } 970 empty = object->cache == NULL; 971 mtx_unlock(&vm_page_queue_free_mtx); 972 if (object->type == OBJT_VNODE && empty) 973 vdrop(object->handle); 974} 975 976/* 977 * Returns the cached page that is associated with the given 978 * object and offset. If, however, none exists, returns NULL. 979 * 980 * The free page queue must be locked. 981 */ 982static inline vm_page_t 983vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 984{ 985 vm_page_t m; 986 987 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 988 if ((m = object->cache) != NULL && m->pindex != pindex) { 989 m = vm_page_splay(pindex, m); 990 if ((object->cache = m)->pindex != pindex) 991 m = NULL; 992 } 993 return (m); 994} 995 996/* 997 * Remove the given cached page from its containing object's 998 * collection of cached pages. 999 * 1000 * The free page queue must be locked. 1001 */ 1002void 1003vm_page_cache_remove(vm_page_t m) 1004{ 1005 vm_object_t object; 1006 vm_page_t root; 1007 1008 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1009 KASSERT((m->flags & PG_CACHED) != 0, 1010 ("vm_page_cache_remove: page %p is not cached", m)); 1011 object = m->object; 1012 if (m != object->cache) { 1013 root = vm_page_splay(m->pindex, object->cache); 1014 KASSERT(root == m, 1015 ("vm_page_cache_remove: page %p is not cached in object %p", 1016 m, object)); 1017 } 1018 if (m->left == NULL) 1019 root = m->right; 1020 else if (m->right == NULL) 1021 root = m->left; 1022 else { 1023 root = vm_page_splay(m->pindex, m->left); 1024 root->right = m->right; 1025 } 1026 object->cache = root; 1027 m->object = NULL; 1028 cnt.v_cache_count--; 1029} 1030 1031/* 1032 * Transfer all of the cached pages with offset greater than or 1033 * equal to 'offidxstart' from the original object's cache to the 1034 * new object's cache. However, any cached pages with offset 1035 * greater than or equal to the new object's size are kept in the 1036 * original object. Initially, the new object's cache must be 1037 * empty. Offset 'offidxstart' in the original object must 1038 * correspond to offset zero in the new object. 1039 * 1040 * The new object must be locked. 1041 */ 1042void 1043vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1044 vm_object_t new_object) 1045{ 1046 vm_page_t m, m_next; 1047 1048 /* 1049 * Insertion into an object's collection of cached pages 1050 * requires the object to be locked. In contrast, removal does 1051 * not. 1052 */ 1053 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 1054 KASSERT(new_object->cache == NULL, 1055 ("vm_page_cache_transfer: object %p has cached pages", 1056 new_object)); 1057 mtx_lock(&vm_page_queue_free_mtx); 1058 if ((m = orig_object->cache) != NULL) { 1059 /* 1060 * Transfer all of the pages with offset greater than or 1061 * equal to 'offidxstart' from the original object's 1062 * cache to the new object's cache. 1063 */ 1064 m = vm_page_splay(offidxstart, m); 1065 if (m->pindex < offidxstart) { 1066 orig_object->cache = m; 1067 new_object->cache = m->right; 1068 m->right = NULL; 1069 } else { 1070 orig_object->cache = m->left; 1071 new_object->cache = m; 1072 m->left = NULL; 1073 } 1074 while ((m = new_object->cache) != NULL) { 1075 if ((m->pindex - offidxstart) >= new_object->size) { 1076 /* 1077 * Return all of the cached pages with 1078 * offset greater than or equal to the 1079 * new object's size to the original 1080 * object's cache. 1081 */ 1082 new_object->cache = m->left; 1083 m->left = orig_object->cache; 1084 orig_object->cache = m; 1085 break; 1086 } 1087 m_next = vm_page_splay(m->pindex, m->right); 1088 /* Update the page's object and offset. */ 1089 m->object = new_object; 1090 m->pindex -= offidxstart; 1091 if (m_next == NULL) 1092 break; 1093 m->right = NULL; 1094 m_next->left = m; 1095 new_object->cache = m_next; 1096 } 1097 KASSERT(new_object->cache == NULL || 1098 new_object->type == OBJT_SWAP, 1099 ("vm_page_cache_transfer: object %p's type is incompatible" 1100 " with cached pages", new_object)); 1101 } 1102 mtx_unlock(&vm_page_queue_free_mtx); 1103} 1104 1105/* 1106 * vm_page_alloc: 1107 * 1108 * Allocate and return a memory cell associated 1109 * with this VM object/offset pair. 1110 * 1111 * page_req classes: 1112 * VM_ALLOC_NORMAL normal process request 1113 * VM_ALLOC_SYSTEM system *really* needs a page 1114 * VM_ALLOC_INTERRUPT interrupt time request 1115 * VM_ALLOC_ZERO zero page 1116 * VM_ALLOC_WIRED wire the allocated page 1117 * VM_ALLOC_NOOBJ page is not associated with a vm object 1118 * VM_ALLOC_NOBUSY do not set the page busy 1119 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1120 * is cached 1121 * 1122 * This routine may not sleep. 1123 */ 1124vm_page_t 1125vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1126{ 1127 struct vnode *vp = NULL; 1128 vm_object_t m_object; 1129 vm_page_t m; 1130 int flags, page_req; 1131 1132 page_req = req & VM_ALLOC_CLASS_MASK; 1133 KASSERT(curthread->td_intr_nesting_level == 0 || 1134 page_req == VM_ALLOC_INTERRUPT, 1135 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 1136 1137 if ((req & VM_ALLOC_NOOBJ) == 0) { 1138 KASSERT(object != NULL, 1139 ("vm_page_alloc: NULL object.")); 1140 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1141 } 1142 1143 /* 1144 * The pager is allowed to eat deeper into the free page list. 1145 */ 1146 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 1147 page_req = VM_ALLOC_SYSTEM; 1148 }; 1149 1150 mtx_lock(&vm_page_queue_free_mtx); 1151 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1152 (page_req == VM_ALLOC_SYSTEM && 1153 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1154 (page_req == VM_ALLOC_INTERRUPT && 1155 cnt.v_free_count + cnt.v_cache_count > 0)) { 1156 /* 1157 * Allocate from the free queue if the number of free pages 1158 * exceeds the minimum for the request class. 1159 */ 1160 if (object != NULL && 1161 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1162 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1163 mtx_unlock(&vm_page_queue_free_mtx); 1164 return (NULL); 1165 } 1166 if (vm_phys_unfree_page(m)) 1167 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1168#if VM_NRESERVLEVEL > 0 1169 else if (!vm_reserv_reactivate_page(m)) 1170#else 1171 else 1172#endif 1173 panic("vm_page_alloc: cache page %p is missing" 1174 " from the free queue", m); 1175 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1176 mtx_unlock(&vm_page_queue_free_mtx); 1177 return (NULL); 1178#if VM_NRESERVLEVEL > 0 1179 } else if (object == NULL || object->type == OBJT_DEVICE || 1180 object->type == OBJT_SG || 1181 (object->flags & OBJ_COLORED) == 0 || 1182 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1183#else 1184 } else { 1185#endif 1186 m = vm_phys_alloc_pages(object != NULL ? 1187 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1188#if VM_NRESERVLEVEL > 0 1189 if (m == NULL && vm_reserv_reclaim_inactive()) { 1190 m = vm_phys_alloc_pages(object != NULL ? 1191 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1192 0); 1193 } 1194#endif 1195 } 1196 } else { 1197 /* 1198 * Not allocatable, give up. 1199 */ 1200 mtx_unlock(&vm_page_queue_free_mtx); 1201 atomic_add_int(&vm_pageout_deficit, 1); 1202 pagedaemon_wakeup(); 1203 return (NULL); 1204 } 1205 1206 /* 1207 * At this point we had better have found a good page. 1208 */ 1209 1210 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1211 KASSERT(m->queue == PQ_NONE, 1212 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1213 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1214 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1215 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1216 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1217 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1218 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1219 pmap_page_get_memattr(m))); 1220 if ((m->flags & PG_CACHED) != 0) { 1221 KASSERT(m->valid != 0, 1222 ("vm_page_alloc: cached page %p is invalid", m)); 1223 if (m->object == object && m->pindex == pindex) 1224 cnt.v_reactivated++; 1225 else 1226 m->valid = 0; 1227 m_object = m->object; 1228 vm_page_cache_remove(m); 1229 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1230 vp = m_object->handle; 1231 } else { 1232 KASSERT(VM_PAGE_IS_FREE(m), 1233 ("vm_page_alloc: page %p is not free", m)); 1234 KASSERT(m->valid == 0, 1235 ("vm_page_alloc: free page %p is valid", m)); 1236 cnt.v_free_count--; 1237 } 1238 1239 /* 1240 * Initialize structure. Only the PG_ZERO flag is inherited. 1241 */ 1242 flags = 0; 1243 if (m->flags & PG_ZERO) { 1244 vm_page_zero_count--; 1245 if (req & VM_ALLOC_ZERO) 1246 flags = PG_ZERO; 1247 } 1248 if (object == NULL || object->type == OBJT_PHYS) 1249 flags |= PG_UNMANAGED; 1250 m->flags = flags; 1251 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 1252 m->oflags = 0; 1253 else 1254 m->oflags = VPO_BUSY; 1255 if (req & VM_ALLOC_WIRED) { 1256 atomic_add_int(&cnt.v_wire_count, 1); 1257 m->wire_count = 1; 1258 } 1259 m->act_count = 0; 1260 mtx_unlock(&vm_page_queue_free_mtx); 1261 1262 if (object != NULL) { 1263 /* Ignore device objects; the pager sets "memattr" for them. */ 1264 if (object->memattr != VM_MEMATTR_DEFAULT && 1265 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1266 pmap_page_set_memattr(m, object->memattr); 1267 vm_page_insert(m, object, pindex); 1268 } else 1269 m->pindex = pindex; 1270 1271 /* 1272 * The following call to vdrop() must come after the above call 1273 * to vm_page_insert() in case both affect the same object and 1274 * vnode. Otherwise, the affected vnode's hold count could 1275 * temporarily become zero. 1276 */ 1277 if (vp != NULL) 1278 vdrop(vp); 1279 1280 /* 1281 * Don't wakeup too often - wakeup the pageout daemon when 1282 * we would be nearly out of memory. 1283 */ 1284 if (vm_paging_needed()) 1285 pagedaemon_wakeup(); 1286 1287 return (m); 1288} 1289 1290/* 1291 * vm_wait: (also see VM_WAIT macro) 1292 * 1293 * Block until free pages are available for allocation 1294 * - Called in various places before memory allocations. 1295 */ 1296void 1297vm_wait(void) 1298{ 1299 1300 mtx_lock(&vm_page_queue_free_mtx); 1301 if (curproc == pageproc) { 1302 vm_pageout_pages_needed = 1; 1303 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1304 PDROP | PSWP, "VMWait", 0); 1305 } else { 1306 if (!vm_pages_needed) { 1307 vm_pages_needed = 1; 1308 wakeup(&vm_pages_needed); 1309 } 1310 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1311 "vmwait", 0); 1312 } 1313} 1314 1315/* 1316 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1317 * 1318 * Block until free pages are available for allocation 1319 * - Called only in vm_fault so that processes page faulting 1320 * can be easily tracked. 1321 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1322 * processes will be able to grab memory first. Do not change 1323 * this balance without careful testing first. 1324 */ 1325void 1326vm_waitpfault(void) 1327{ 1328 1329 mtx_lock(&vm_page_queue_free_mtx); 1330 if (!vm_pages_needed) { 1331 vm_pages_needed = 1; 1332 wakeup(&vm_pages_needed); 1333 } 1334 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1335 "pfault", 0); 1336} 1337 1338/* 1339 * vm_page_requeue: 1340 * 1341 * Move the given page to the tail of its present page queue. 1342 * 1343 * The page queues must be locked. 1344 */ 1345void 1346vm_page_requeue(vm_page_t m) 1347{ 1348 int queue = VM_PAGE_GETQUEUE(m); 1349 struct vpgqueues *vpq; 1350 1351 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1352 KASSERT(queue != PQ_NONE, 1353 ("vm_page_requeue: page %p is not queued", m)); 1354 vpq = &vm_page_queues[queue]; 1355 TAILQ_REMOVE(&vpq->pl, m, pageq); 1356 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1357} 1358 1359/* 1360 * vm_page_queue_remove: 1361 * 1362 * Remove the given page from the specified queue. 1363 * 1364 * The page and page queues must be locked. 1365 */ 1366static __inline void 1367vm_page_queue_remove(int queue, vm_page_t m) 1368{ 1369 struct vpgqueues *pq; 1370 1371 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1372 vm_page_lock_assert(m, MA_OWNED); 1373 pq = &vm_page_queues[queue]; 1374 TAILQ_REMOVE(&pq->pl, m, pageq); 1375 (*pq->cnt)--; 1376} 1377 1378/* 1379 * vm_pageq_remove: 1380 * 1381 * Remove a page from its queue. 1382 * 1383 * The given page must be locked. 1384 * This routine may not block. 1385 */ 1386void 1387vm_pageq_remove(vm_page_t m) 1388{ 1389 int queue = VM_PAGE_GETQUEUE(m); 1390 1391 vm_page_lock_assert(m, MA_OWNED); 1392 if (queue != PQ_NONE) { 1393 vm_page_lock_queues(); 1394 VM_PAGE_SETQUEUE2(m, PQ_NONE); 1395 vm_page_queue_remove(queue, m); 1396 vm_page_unlock_queues(); 1397 } 1398} 1399 1400/* 1401 * vm_page_enqueue: 1402 * 1403 * Add the given page to the specified queue. 1404 * 1405 * The page queues must be locked. 1406 */ 1407static void 1408vm_page_enqueue(int queue, vm_page_t m) 1409{ 1410 struct vpgqueues *vpq; 1411 1412 vpq = &vm_page_queues[queue]; 1413 VM_PAGE_SETQUEUE2(m, queue); 1414 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1415 ++*vpq->cnt; 1416} 1417 1418/* 1419 * vm_page_activate: 1420 * 1421 * Put the specified page on the active list (if appropriate). 1422 * Ensure that act_count is at least ACT_INIT but do not otherwise 1423 * mess with it. 1424 * 1425 * The page must be locked. 1426 * This routine may not block. 1427 */ 1428void 1429vm_page_activate(vm_page_t m) 1430{ 1431 int queue; 1432 1433 vm_page_lock_assert(m, MA_OWNED); 1434 if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) != PQ_ACTIVE) { 1435 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1436 if (m->act_count < ACT_INIT) 1437 m->act_count = ACT_INIT; 1438 vm_page_lock_queues(); 1439 if (queue != PQ_NONE) 1440 vm_page_queue_remove(queue, m); 1441 vm_page_enqueue(PQ_ACTIVE, m); 1442 vm_page_unlock_queues(); 1443 } else 1444 KASSERT(queue == PQ_NONE, 1445 ("vm_page_activate: wired page %p is queued", m)); 1446 } else { 1447 if (m->act_count < ACT_INIT) 1448 m->act_count = ACT_INIT; 1449 } 1450} 1451 1452/* 1453 * vm_page_free_wakeup: 1454 * 1455 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1456 * routine is called when a page has been added to the cache or free 1457 * queues. 1458 * 1459 * The page queues must be locked. 1460 * This routine may not block. 1461 */ 1462static inline void 1463vm_page_free_wakeup(void) 1464{ 1465 1466 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1467 /* 1468 * if pageout daemon needs pages, then tell it that there are 1469 * some free. 1470 */ 1471 if (vm_pageout_pages_needed && 1472 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1473 wakeup(&vm_pageout_pages_needed); 1474 vm_pageout_pages_needed = 0; 1475 } 1476 /* 1477 * wakeup processes that are waiting on memory if we hit a 1478 * high water mark. And wakeup scheduler process if we have 1479 * lots of memory. this process will swapin processes. 1480 */ 1481 if (vm_pages_needed && !vm_page_count_min()) { 1482 vm_pages_needed = 0; 1483 wakeup(&cnt.v_free_count); 1484 } 1485} 1486 1487/* 1488 * vm_page_free_toq: 1489 * 1490 * Returns the given page to the free list, 1491 * disassociating it with any VM object. 1492 * 1493 * Object and page must be locked prior to entry. 1494 * This routine may not block. 1495 */ 1496 1497void 1498vm_page_free_toq(vm_page_t m) 1499{ 1500 1501 if ((m->flags & PG_UNMANAGED) == 0) { 1502 vm_page_lock_assert(m, MA_OWNED); 1503 KASSERT(!pmap_page_is_mapped(m), 1504 ("vm_page_free_toq: freeing mapped page %p", m)); 1505 } 1506 PCPU_INC(cnt.v_tfree); 1507 1508 if (m->busy || VM_PAGE_IS_FREE(m)) { 1509 printf( 1510 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1511 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1512 m->hold_count); 1513 if (VM_PAGE_IS_FREE(m)) 1514 panic("vm_page_free: freeing free page"); 1515 else 1516 panic("vm_page_free: freeing busy page"); 1517 } 1518 1519 /* 1520 * unqueue, then remove page. Note that we cannot destroy 1521 * the page here because we do not want to call the pager's 1522 * callback routine until after we've put the page on the 1523 * appropriate free queue. 1524 */ 1525 if ((m->flags & PG_UNMANAGED) == 0) 1526 vm_pageq_remove(m); 1527 vm_page_remove(m); 1528 1529 /* 1530 * If fictitious remove object association and 1531 * return, otherwise delay object association removal. 1532 */ 1533 if ((m->flags & PG_FICTITIOUS) != 0) { 1534 return; 1535 } 1536 1537 m->valid = 0; 1538 vm_page_undirty(m); 1539 1540 if (m->wire_count != 0) { 1541 if (m->wire_count > 1) { 1542 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1543 m->wire_count, (long)m->pindex); 1544 } 1545 panic("vm_page_free: freeing wired page"); 1546 } 1547 if (m->hold_count != 0) { 1548 m->flags &= ~PG_ZERO; 1549 vm_page_lock_queues(); 1550 vm_page_enqueue(PQ_HOLD, m); 1551 vm_page_unlock_queues(); 1552 } else { 1553 /* 1554 * Restore the default memory attribute to the page. 1555 */ 1556 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1557 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1558 1559 /* 1560 * Insert the page into the physical memory allocator's 1561 * cache/free page queues. 1562 */ 1563 mtx_lock(&vm_page_queue_free_mtx); 1564 m->flags |= PG_FREE; 1565 cnt.v_free_count++; 1566#if VM_NRESERVLEVEL > 0 1567 if (!vm_reserv_free_page(m)) 1568#else 1569 if (TRUE) 1570#endif 1571 vm_phys_free_pages(m, 0); 1572 if ((m->flags & PG_ZERO) != 0) 1573 ++vm_page_zero_count; 1574 else 1575 vm_page_zero_idle_wakeup(); 1576 vm_page_free_wakeup(); 1577 mtx_unlock(&vm_page_queue_free_mtx); 1578 } 1579} 1580 1581/* 1582 * vm_page_wire: 1583 * 1584 * Mark this page as wired down by yet 1585 * another map, removing it from paging queues 1586 * as necessary. 1587 * 1588 * If the page is fictitious, then its wire count must remain one. 1589 * 1590 * The page must be locked. 1591 * This routine may not block. 1592 */ 1593void 1594vm_page_wire(vm_page_t m) 1595{ 1596 1597 /* 1598 * Only bump the wire statistics if the page is not already wired, 1599 * and only unqueue the page if it is on some queue (if it is unmanaged 1600 * it is already off the queues). 1601 */ 1602 vm_page_lock_assert(m, MA_OWNED); 1603 if ((m->flags & PG_FICTITIOUS) != 0) { 1604 KASSERT(m->wire_count == 1, 1605 ("vm_page_wire: fictitious page %p's wire count isn't one", 1606 m)); 1607 return; 1608 } 1609 if (m->wire_count == 0) { 1610 if ((m->flags & PG_UNMANAGED) == 0) 1611 vm_pageq_remove(m); 1612 atomic_add_int(&cnt.v_wire_count, 1); 1613 } 1614 m->wire_count++; 1615 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1616} 1617 1618/* 1619 * vm_page_unwire: 1620 * 1621 * Release one wiring of the specified page, potentially enabling it to be 1622 * paged again. If paging is enabled, then the value of the parameter 1623 * "activate" determines to which queue the page is added. If "activate" is 1624 * non-zero, then the page is added to the active queue. Otherwise, it is 1625 * added to the inactive queue. 1626 * 1627 * However, unless the page belongs to an object, it is not enqueued because 1628 * it cannot be paged out. 1629 * 1630 * If a page is fictitious, then its wire count must alway be one. 1631 * 1632 * A managed page must be locked. 1633 */ 1634void 1635vm_page_unwire(vm_page_t m, int activate) 1636{ 1637 1638 if ((m->flags & PG_UNMANAGED) == 0) 1639 vm_page_lock_assert(m, MA_OWNED); 1640 if ((m->flags & PG_FICTITIOUS) != 0) { 1641 KASSERT(m->wire_count == 1, 1642 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 1643 return; 1644 } 1645 if (m->wire_count > 0) { 1646 m->wire_count--; 1647 if (m->wire_count == 0) { 1648 atomic_subtract_int(&cnt.v_wire_count, 1); 1649 if ((m->flags & PG_UNMANAGED) != 0 || 1650 m->object == NULL) 1651 return; 1652 vm_page_lock_queues(); 1653 if (activate) 1654 vm_page_enqueue(PQ_ACTIVE, m); 1655 else { 1656 vm_page_flag_clear(m, PG_WINATCFLS); 1657 vm_page_enqueue(PQ_INACTIVE, m); 1658 } 1659 vm_page_unlock_queues(); 1660 } 1661 } else 1662 panic("vm_page_unwire: page %p's wire count is zero", m); 1663} 1664 1665/* 1666 * Move the specified page to the inactive queue. 1667 * 1668 * Many pages placed on the inactive queue should actually go 1669 * into the cache, but it is difficult to figure out which. What 1670 * we do instead, if the inactive target is well met, is to put 1671 * clean pages at the head of the inactive queue instead of the tail. 1672 * This will cause them to be moved to the cache more quickly and 1673 * if not actively re-referenced, reclaimed more quickly. If we just 1674 * stick these pages at the end of the inactive queue, heavy filesystem 1675 * meta-data accesses can cause an unnecessary paging load on memory bound 1676 * processes. This optimization causes one-time-use metadata to be 1677 * reused more quickly. 1678 * 1679 * Normally athead is 0 resulting in LRU operation. athead is set 1680 * to 1 if we want this page to be 'as if it were placed in the cache', 1681 * except without unmapping it from the process address space. 1682 * 1683 * This routine may not block. 1684 */ 1685static inline void 1686_vm_page_deactivate(vm_page_t m, int athead) 1687{ 1688 int queue; 1689 1690 vm_page_lock_assert(m, MA_OWNED); 1691 1692 /* 1693 * Ignore if already inactive. 1694 */ 1695 if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) == PQ_INACTIVE) 1696 return; 1697 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1698 vm_page_lock_queues(); 1699 vm_page_flag_clear(m, PG_WINATCFLS); 1700 if (queue != PQ_NONE) 1701 vm_page_queue_remove(queue, m); 1702 if (athead) 1703 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, 1704 pageq); 1705 else 1706 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, 1707 pageq); 1708 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1709 cnt.v_inactive_count++; 1710 vm_page_unlock_queues(); 1711 } 1712} 1713 1714/* 1715 * Move the specified page to the inactive queue. 1716 * 1717 * The page must be locked. 1718 */ 1719void 1720vm_page_deactivate(vm_page_t m) 1721{ 1722 1723 _vm_page_deactivate(m, 0); 1724} 1725 1726/* 1727 * vm_page_try_to_cache: 1728 * 1729 * Returns 0 on failure, 1 on success 1730 */ 1731int 1732vm_page_try_to_cache(vm_page_t m) 1733{ 1734 1735 vm_page_lock_assert(m, MA_OWNED); 1736 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1737 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1738 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1739 return (0); 1740 pmap_remove_all(m); 1741 if (m->dirty) 1742 return (0); 1743 vm_page_cache(m); 1744 return (1); 1745} 1746 1747/* 1748 * vm_page_try_to_free() 1749 * 1750 * Attempt to free the page. If we cannot free it, we do nothing. 1751 * 1 is returned on success, 0 on failure. 1752 */ 1753int 1754vm_page_try_to_free(vm_page_t m) 1755{ 1756 1757 vm_page_lock_assert(m, MA_OWNED); 1758 if (m->object != NULL) 1759 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1760 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1761 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1762 return (0); 1763 pmap_remove_all(m); 1764 if (m->dirty) 1765 return (0); 1766 vm_page_free(m); 1767 return (1); 1768} 1769 1770/* 1771 * vm_page_cache 1772 * 1773 * Put the specified page onto the page cache queue (if appropriate). 1774 * 1775 * This routine may not block. 1776 */ 1777void 1778vm_page_cache(vm_page_t m) 1779{ 1780 vm_object_t object; 1781 vm_page_t root; 1782 1783 vm_page_lock_assert(m, MA_OWNED); 1784 object = m->object; 1785 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1786 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1787 m->hold_count || m->wire_count) 1788 panic("vm_page_cache: attempting to cache busy page"); 1789 pmap_remove_all(m); 1790 if (m->dirty != 0) 1791 panic("vm_page_cache: page %p is dirty", m); 1792 if (m->valid == 0 || object->type == OBJT_DEFAULT || 1793 (object->type == OBJT_SWAP && 1794 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 1795 /* 1796 * Hypothesis: A cache-elgible page belonging to a 1797 * default object or swap object but without a backing 1798 * store must be zero filled. 1799 */ 1800 vm_page_free(m); 1801 return; 1802 } 1803 KASSERT((m->flags & PG_CACHED) == 0, 1804 ("vm_page_cache: page %p is already cached", m)); 1805 PCPU_INC(cnt.v_tcached); 1806 1807 /* 1808 * Remove the page from the paging queues. 1809 */ 1810 vm_pageq_remove(m); 1811 1812 /* 1813 * Remove the page from the object's collection of resident 1814 * pages. 1815 */ 1816 if (m != object->root) 1817 vm_page_splay(m->pindex, object->root); 1818 if (m->left == NULL) 1819 root = m->right; 1820 else { 1821 root = vm_page_splay(m->pindex, m->left); 1822 root->right = m->right; 1823 } 1824 object->root = root; 1825 TAILQ_REMOVE(&object->memq, m, listq); 1826 object->resident_page_count--; 1827 object->generation++; 1828 1829 /* 1830 * Restore the default memory attribute to the page. 1831 */ 1832 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1833 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1834 1835 /* 1836 * Insert the page into the object's collection of cached pages 1837 * and the physical memory allocator's cache/free page queues. 1838 */ 1839 m->flags &= ~PG_ZERO; 1840 mtx_lock(&vm_page_queue_free_mtx); 1841 m->flags |= PG_CACHED; 1842 cnt.v_cache_count++; 1843 root = object->cache; 1844 if (root == NULL) { 1845 m->left = NULL; 1846 m->right = NULL; 1847 } else { 1848 root = vm_page_splay(m->pindex, root); 1849 if (m->pindex < root->pindex) { 1850 m->left = root->left; 1851 m->right = root; 1852 root->left = NULL; 1853 } else if (__predict_false(m->pindex == root->pindex)) 1854 panic("vm_page_cache: offset already cached"); 1855 else { 1856 m->right = root->right; 1857 m->left = root; 1858 root->right = NULL; 1859 } 1860 } 1861 object->cache = m; 1862#if VM_NRESERVLEVEL > 0 1863 if (!vm_reserv_free_page(m)) { 1864#else 1865 if (TRUE) { 1866#endif 1867 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 1868 vm_phys_free_pages(m, 0); 1869 } 1870 vm_page_free_wakeup(); 1871 mtx_unlock(&vm_page_queue_free_mtx); 1872 1873 /* 1874 * Increment the vnode's hold count if this is the object's only 1875 * cached page. Decrement the vnode's hold count if this was 1876 * the object's only resident page. 1877 */ 1878 if (object->type == OBJT_VNODE) { 1879 if (root == NULL && object->resident_page_count != 0) 1880 vhold(object->handle); 1881 else if (root != NULL && object->resident_page_count == 0) 1882 vdrop(object->handle); 1883 } 1884} 1885 1886/* 1887 * vm_page_dontneed 1888 * 1889 * Cache, deactivate, or do nothing as appropriate. This routine 1890 * is typically used by madvise() MADV_DONTNEED. 1891 * 1892 * Generally speaking we want to move the page into the cache so 1893 * it gets reused quickly. However, this can result in a silly syndrome 1894 * due to the page recycling too quickly. Small objects will not be 1895 * fully cached. On the otherhand, if we move the page to the inactive 1896 * queue we wind up with a problem whereby very large objects 1897 * unnecessarily blow away our inactive and cache queues. 1898 * 1899 * The solution is to move the pages based on a fixed weighting. We 1900 * either leave them alone, deactivate them, or move them to the cache, 1901 * where moving them to the cache has the highest weighting. 1902 * By forcing some pages into other queues we eventually force the 1903 * system to balance the queues, potentially recovering other unrelated 1904 * space from active. The idea is to not force this to happen too 1905 * often. 1906 */ 1907void 1908vm_page_dontneed(vm_page_t m) 1909{ 1910 int dnw; 1911 int head; 1912 1913 vm_page_lock_assert(m, MA_OWNED); 1914 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1915 dnw = PCPU_GET(dnweight); 1916 PCPU_INC(dnweight); 1917 1918 /* 1919 * Occasionally leave the page alone. 1920 */ 1921 if ((dnw & 0x01F0) == 0 || 1922 VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) { 1923 if (m->act_count >= ACT_INIT) 1924 --m->act_count; 1925 return; 1926 } 1927 1928 /* 1929 * Clear any references to the page. Otherwise, the page daemon will 1930 * immediately reactivate the page. 1931 * 1932 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 1933 * pmap operation, such as pmap_remove(), could clear a reference in 1934 * the pmap and set PG_REFERENCED on the page before the 1935 * pmap_clear_reference() had completed. Consequently, the page would 1936 * appear referenced based upon an old reference that occurred before 1937 * this function ran. 1938 */ 1939 pmap_clear_reference(m); 1940 vm_page_lock_queues(); 1941 vm_page_flag_clear(m, PG_REFERENCED); 1942 vm_page_unlock_queues(); 1943 1944 if (m->dirty == 0 && pmap_is_modified(m)) 1945 vm_page_dirty(m); 1946 1947 if (m->dirty || (dnw & 0x0070) == 0) { 1948 /* 1949 * Deactivate the page 3 times out of 32. 1950 */ 1951 head = 0; 1952 } else { 1953 /* 1954 * Cache the page 28 times out of every 32. Note that 1955 * the page is deactivated instead of cached, but placed 1956 * at the head of the queue instead of the tail. 1957 */ 1958 head = 1; 1959 } 1960 _vm_page_deactivate(m, head); 1961} 1962 1963/* 1964 * Grab a page, waiting until we are waken up due to the page 1965 * changing state. We keep on waiting, if the page continues 1966 * to be in the object. If the page doesn't exist, first allocate it 1967 * and then conditionally zero it. 1968 * 1969 * This routine may block. 1970 */ 1971vm_page_t 1972vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1973{ 1974 vm_page_t m; 1975 1976 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1977retrylookup: 1978 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1979 if ((m->oflags & VPO_BUSY) != 0 || m->busy != 0) { 1980 if ((allocflags & VM_ALLOC_RETRY) != 0) { 1981 /* 1982 * Reference the page before unlocking and 1983 * sleeping so that the page daemon is less 1984 * likely to reclaim it. 1985 */ 1986 vm_page_lock_queues(); 1987 vm_page_flag_set(m, PG_REFERENCED); 1988 } 1989 vm_page_sleep(m, "pgrbwt"); 1990 if ((allocflags & VM_ALLOC_RETRY) == 0) 1991 return (NULL); 1992 goto retrylookup; 1993 } else { 1994 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1995 vm_page_lock(m); 1996 vm_page_wire(m); 1997 vm_page_unlock(m); 1998 } 1999 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 2000 vm_page_busy(m); 2001 return (m); 2002 } 2003 } 2004 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 2005 if (m == NULL) { 2006 VM_OBJECT_UNLOCK(object); 2007 VM_WAIT; 2008 VM_OBJECT_LOCK(object); 2009 if ((allocflags & VM_ALLOC_RETRY) == 0) 2010 return (NULL); 2011 goto retrylookup; 2012 } else if (m->valid != 0) 2013 return (m); 2014 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2015 pmap_zero_page(m); 2016 return (m); 2017} 2018 2019/* 2020 * Mapping function for valid bits or for dirty bits in 2021 * a page. May not block. 2022 * 2023 * Inputs are required to range within a page. 2024 */ 2025int 2026vm_page_bits(int base, int size) 2027{ 2028 int first_bit; 2029 int last_bit; 2030 2031 KASSERT( 2032 base + size <= PAGE_SIZE, 2033 ("vm_page_bits: illegal base/size %d/%d", base, size) 2034 ); 2035 2036 if (size == 0) /* handle degenerate case */ 2037 return (0); 2038 2039 first_bit = base >> DEV_BSHIFT; 2040 last_bit = (base + size - 1) >> DEV_BSHIFT; 2041 2042 return ((2 << last_bit) - (1 << first_bit)); 2043} 2044 2045/* 2046 * vm_page_set_valid: 2047 * 2048 * Sets portions of a page valid. The arguments are expected 2049 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2050 * of any partial chunks touched by the range. The invalid portion of 2051 * such chunks will be zeroed. 2052 * 2053 * (base + size) must be less then or equal to PAGE_SIZE. 2054 */ 2055void 2056vm_page_set_valid(vm_page_t m, int base, int size) 2057{ 2058 int endoff, frag; 2059 2060 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2061 if (size == 0) /* handle degenerate case */ 2062 return; 2063 2064 /* 2065 * If the base is not DEV_BSIZE aligned and the valid 2066 * bit is clear, we have to zero out a portion of the 2067 * first block. 2068 */ 2069 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2070 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2071 pmap_zero_page_area(m, frag, base - frag); 2072 2073 /* 2074 * If the ending offset is not DEV_BSIZE aligned and the 2075 * valid bit is clear, we have to zero out a portion of 2076 * the last block. 2077 */ 2078 endoff = base + size; 2079 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2080 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2081 pmap_zero_page_area(m, endoff, 2082 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2083 2084 /* 2085 * Assert that no previously invalid block that is now being validated 2086 * is already dirty. 2087 */ 2088 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2089 ("vm_page_set_valid: page %p is dirty", m)); 2090 2091 /* 2092 * Set valid bits inclusive of any overlap. 2093 */ 2094 m->valid |= vm_page_bits(base, size); 2095} 2096 2097/* 2098 * Clear the given bits from the specified page's dirty field. 2099 */ 2100static __inline void 2101vm_page_clear_dirty_mask(vm_page_t m, int pagebits) 2102{ 2103 2104 /* 2105 * If the object is locked and the page is neither VPO_BUSY nor 2106 * PG_WRITEABLE, then the page's dirty field cannot possibly be 2107 * modified by a concurrent pmap operation. 2108 */ 2109 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2110 if ((m->oflags & VPO_BUSY) == 0 && (m->flags & PG_WRITEABLE) == 0) 2111 m->dirty &= ~pagebits; 2112 else { 2113 vm_page_lock_queues(); 2114 m->dirty &= ~pagebits; 2115 vm_page_unlock_queues(); 2116 } 2117} 2118 2119/* 2120 * vm_page_set_validclean: 2121 * 2122 * Sets portions of a page valid and clean. The arguments are expected 2123 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2124 * of any partial chunks touched by the range. The invalid portion of 2125 * such chunks will be zero'd. 2126 * 2127 * This routine may not block. 2128 * 2129 * (base + size) must be less then or equal to PAGE_SIZE. 2130 */ 2131void 2132vm_page_set_validclean(vm_page_t m, int base, int size) 2133{ 2134 u_long oldvalid; 2135 int endoff, frag, pagebits; 2136 2137 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2138 if (size == 0) /* handle degenerate case */ 2139 return; 2140 2141 /* 2142 * If the base is not DEV_BSIZE aligned and the valid 2143 * bit is clear, we have to zero out a portion of the 2144 * first block. 2145 */ 2146 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2147 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2148 pmap_zero_page_area(m, frag, base - frag); 2149 2150 /* 2151 * If the ending offset is not DEV_BSIZE aligned and the 2152 * valid bit is clear, we have to zero out a portion of 2153 * the last block. 2154 */ 2155 endoff = base + size; 2156 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2157 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2158 pmap_zero_page_area(m, endoff, 2159 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2160 2161 /* 2162 * Set valid, clear dirty bits. If validating the entire 2163 * page we can safely clear the pmap modify bit. We also 2164 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2165 * takes a write fault on a MAP_NOSYNC memory area the flag will 2166 * be set again. 2167 * 2168 * We set valid bits inclusive of any overlap, but we can only 2169 * clear dirty bits for DEV_BSIZE chunks that are fully within 2170 * the range. 2171 */ 2172 oldvalid = m->valid; 2173 pagebits = vm_page_bits(base, size); 2174 m->valid |= pagebits; 2175#if 0 /* NOT YET */ 2176 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2177 frag = DEV_BSIZE - frag; 2178 base += frag; 2179 size -= frag; 2180 if (size < 0) 2181 size = 0; 2182 } 2183 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2184#endif 2185 if (base == 0 && size == PAGE_SIZE) { 2186 /* 2187 * The page can only be modified within the pmap if it is 2188 * mapped, and it can only be mapped if it was previously 2189 * fully valid. 2190 */ 2191 if (oldvalid == VM_PAGE_BITS_ALL) 2192 /* 2193 * Perform the pmap_clear_modify() first. Otherwise, 2194 * a concurrent pmap operation, such as 2195 * pmap_protect(), could clear a modification in the 2196 * pmap and set the dirty field on the page before 2197 * pmap_clear_modify() had begun and after the dirty 2198 * field was cleared here. 2199 */ 2200 pmap_clear_modify(m); 2201 m->dirty = 0; 2202 m->oflags &= ~VPO_NOSYNC; 2203 } else if (oldvalid != VM_PAGE_BITS_ALL) 2204 m->dirty &= ~pagebits; 2205 else 2206 vm_page_clear_dirty_mask(m, pagebits); 2207} 2208 2209void 2210vm_page_clear_dirty(vm_page_t m, int base, int size) 2211{ 2212 2213 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2214} 2215 2216/* 2217 * vm_page_set_invalid: 2218 * 2219 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2220 * valid and dirty bits for the effected areas are cleared. 2221 * 2222 * May not block. 2223 */ 2224void 2225vm_page_set_invalid(vm_page_t m, int base, int size) 2226{ 2227 int bits; 2228 2229 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2230 KASSERT((m->oflags & VPO_BUSY) == 0, 2231 ("vm_page_set_invalid: page %p is busy", m)); 2232 bits = vm_page_bits(base, size); 2233 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2234 pmap_remove_all(m); 2235 KASSERT(!pmap_page_is_mapped(m), 2236 ("vm_page_set_invalid: page %p is mapped", m)); 2237 m->valid &= ~bits; 2238 m->dirty &= ~bits; 2239 m->object->generation++; 2240} 2241 2242/* 2243 * vm_page_zero_invalid() 2244 * 2245 * The kernel assumes that the invalid portions of a page contain 2246 * garbage, but such pages can be mapped into memory by user code. 2247 * When this occurs, we must zero out the non-valid portions of the 2248 * page so user code sees what it expects. 2249 * 2250 * Pages are most often semi-valid when the end of a file is mapped 2251 * into memory and the file's size is not page aligned. 2252 */ 2253void 2254vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2255{ 2256 int b; 2257 int i; 2258 2259 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2260 /* 2261 * Scan the valid bits looking for invalid sections that 2262 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2263 * valid bit may be set ) have already been zerod by 2264 * vm_page_set_validclean(). 2265 */ 2266 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2267 if (i == (PAGE_SIZE / DEV_BSIZE) || 2268 (m->valid & (1 << i)) 2269 ) { 2270 if (i > b) { 2271 pmap_zero_page_area(m, 2272 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2273 } 2274 b = i + 1; 2275 } 2276 } 2277 2278 /* 2279 * setvalid is TRUE when we can safely set the zero'd areas 2280 * as being valid. We can do this if there are no cache consistancy 2281 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2282 */ 2283 if (setvalid) 2284 m->valid = VM_PAGE_BITS_ALL; 2285} 2286 2287/* 2288 * vm_page_is_valid: 2289 * 2290 * Is (partial) page valid? Note that the case where size == 0 2291 * will return FALSE in the degenerate case where the page is 2292 * entirely invalid, and TRUE otherwise. 2293 * 2294 * May not block. 2295 */ 2296int 2297vm_page_is_valid(vm_page_t m, int base, int size) 2298{ 2299 int bits = vm_page_bits(base, size); 2300 2301 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2302 if (m->valid && ((m->valid & bits) == bits)) 2303 return 1; 2304 else 2305 return 0; 2306} 2307 2308/* 2309 * update dirty bits from pmap/mmu. May not block. 2310 */ 2311void 2312vm_page_test_dirty(vm_page_t m) 2313{ 2314 2315 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2316 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2317 vm_page_dirty(m); 2318} 2319 2320int so_zerocp_fullpage = 0; 2321 2322/* 2323 * Replace the given page with a copy. The copied page assumes 2324 * the portion of the given page's "wire_count" that is not the 2325 * responsibility of this copy-on-write mechanism. 2326 * 2327 * The object containing the given page must have a non-zero 2328 * paging-in-progress count and be locked. 2329 */ 2330void 2331vm_page_cowfault(vm_page_t m) 2332{ 2333 vm_page_t mnew; 2334 vm_object_t object; 2335 vm_pindex_t pindex; 2336 2337 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 2338 vm_page_lock_assert(m, MA_OWNED); 2339 object = m->object; 2340 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2341 KASSERT(object->paging_in_progress != 0, 2342 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2343 object)); 2344 pindex = m->pindex; 2345 2346 retry_alloc: 2347 pmap_remove_all(m); 2348 vm_page_remove(m); 2349 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2350 if (mnew == NULL) { 2351 vm_page_insert(m, object, pindex); 2352 vm_page_unlock(m); 2353 VM_OBJECT_UNLOCK(object); 2354 VM_WAIT; 2355 VM_OBJECT_LOCK(object); 2356 if (m == vm_page_lookup(object, pindex)) { 2357 vm_page_lock(m); 2358 goto retry_alloc; 2359 } else { 2360 /* 2361 * Page disappeared during the wait. 2362 */ 2363 return; 2364 } 2365 } 2366 2367 if (m->cow == 0) { 2368 /* 2369 * check to see if we raced with an xmit complete when 2370 * waiting to allocate a page. If so, put things back 2371 * the way they were 2372 */ 2373 vm_page_unlock(m); 2374 vm_page_lock(mnew); 2375 vm_page_free(mnew); 2376 vm_page_unlock(mnew); 2377 vm_page_insert(m, object, pindex); 2378 } else { /* clear COW & copy page */ 2379 if (!so_zerocp_fullpage) 2380 pmap_copy_page(m, mnew); 2381 mnew->valid = VM_PAGE_BITS_ALL; 2382 vm_page_dirty(mnew); 2383 mnew->wire_count = m->wire_count - m->cow; 2384 m->wire_count = m->cow; 2385 vm_page_unlock(m); 2386 } 2387} 2388 2389void 2390vm_page_cowclear(vm_page_t m) 2391{ 2392 2393 vm_page_lock_assert(m, MA_OWNED); 2394 if (m->cow) { 2395 m->cow--; 2396 /* 2397 * let vm_fault add back write permission lazily 2398 */ 2399 } 2400 /* 2401 * sf_buf_free() will free the page, so we needn't do it here 2402 */ 2403} 2404 2405int 2406vm_page_cowsetup(vm_page_t m) 2407{ 2408 2409 vm_page_lock_assert(m, MA_OWNED); 2410 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 || 2411 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 2412 return (EBUSY); 2413 m->cow++; 2414 pmap_remove_write(m); 2415 VM_OBJECT_UNLOCK(m->object); 2416 return (0); 2417} 2418 2419#include "opt_ddb.h" 2420#ifdef DDB 2421#include <sys/kernel.h> 2422 2423#include <ddb/ddb.h> 2424 2425DB_SHOW_COMMAND(page, vm_page_print_page_info) 2426{ 2427 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 2428 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 2429 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 2430 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 2431 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 2432 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 2433 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 2434 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 2435 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 2436 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 2437} 2438 2439DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2440{ 2441 2442 db_printf("PQ_FREE:"); 2443 db_printf(" %d", cnt.v_free_count); 2444 db_printf("\n"); 2445 2446 db_printf("PQ_CACHE:"); 2447 db_printf(" %d", cnt.v_cache_count); 2448 db_printf("\n"); 2449 2450 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2451 *vm_page_queues[PQ_ACTIVE].cnt, 2452 *vm_page_queues[PQ_INACTIVE].cnt); 2453} 2454#endif /* DDB */ 2455