vm_page.c revision 208990
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 208990 2010-06-10 16:56:35Z 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 * For a managed page, the PG_WRITEABLE flag can be set only if 491 * the page is VPO_BUSY. Currently this flag is only set by 492 * pmap_enter(). 493 */ 494 KASSERT((bits & PG_WRITEABLE) == 0 || 495 (m->flags & (PG_UNMANAGED | PG_FICTITIOUS)) != 0 || 496 (m->oflags & VPO_BUSY) != 0, ("PG_WRITEABLE and !VPO_BUSY")); 497 m->flags |= bits; 498} 499 500void 501vm_page_flag_clear(vm_page_t m, unsigned short bits) 502{ 503 504 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 505 KASSERT((bits & PG_REFERENCED) == 0 || VM_OBJECT_LOCKED(m->object), 506 ("PG_REFERENCED and !VM_OBJECT_LOCKED")); 507 m->flags &= ~bits; 508} 509 510void 511vm_page_busy(vm_page_t m) 512{ 513 514 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 515 KASSERT((m->oflags & VPO_BUSY) == 0, 516 ("vm_page_busy: page already busy!!!")); 517 m->oflags |= VPO_BUSY; 518} 519 520/* 521 * vm_page_flash: 522 * 523 * wakeup anyone waiting for the page. 524 */ 525void 526vm_page_flash(vm_page_t m) 527{ 528 529 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 530 if (m->oflags & VPO_WANTED) { 531 m->oflags &= ~VPO_WANTED; 532 wakeup(m); 533 } 534} 535 536/* 537 * vm_page_wakeup: 538 * 539 * clear the VPO_BUSY flag and wakeup anyone waiting for the 540 * page. 541 * 542 */ 543void 544vm_page_wakeup(vm_page_t m) 545{ 546 547 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 548 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 549 m->oflags &= ~VPO_BUSY; 550 vm_page_flash(m); 551} 552 553void 554vm_page_io_start(vm_page_t m) 555{ 556 557 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 558 m->busy++; 559} 560 561void 562vm_page_io_finish(vm_page_t m) 563{ 564 565 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 566 m->busy--; 567 if (m->busy == 0) 568 vm_page_flash(m); 569} 570 571/* 572 * Keep page from being freed by the page daemon 573 * much of the same effect as wiring, except much lower 574 * overhead and should be used only for *very* temporary 575 * holding ("wiring"). 576 */ 577void 578vm_page_hold(vm_page_t mem) 579{ 580 581 vm_page_lock_assert(mem, MA_OWNED); 582 mem->hold_count++; 583} 584 585void 586vm_page_unhold(vm_page_t mem) 587{ 588 589 vm_page_lock_assert(mem, MA_OWNED); 590 --mem->hold_count; 591 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 592 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD)) 593 vm_page_free_toq(mem); 594} 595 596/* 597 * vm_page_free: 598 * 599 * Free a page. 600 */ 601void 602vm_page_free(vm_page_t m) 603{ 604 605 m->flags &= ~PG_ZERO; 606 vm_page_free_toq(m); 607} 608 609/* 610 * vm_page_free_zero: 611 * 612 * Free a page to the zerod-pages queue 613 */ 614void 615vm_page_free_zero(vm_page_t m) 616{ 617 618 m->flags |= PG_ZERO; 619 vm_page_free_toq(m); 620} 621 622/* 623 * vm_page_sleep: 624 * 625 * Sleep and release the page and page queues locks. 626 * 627 * The object containing the given page must be locked. 628 */ 629void 630vm_page_sleep(vm_page_t m, const char *msg) 631{ 632 633 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 634 if (mtx_owned(&vm_page_queue_mtx)) 635 vm_page_unlock_queues(); 636 if (mtx_owned(vm_page_lockptr(m))) 637 vm_page_unlock(m); 638 639 /* 640 * It's possible that while we sleep, the page will get 641 * unbusied and freed. If we are holding the object 642 * lock, we will assume we hold a reference to the object 643 * such that even if m->object changes, we can re-lock 644 * it. 645 */ 646 m->oflags |= VPO_WANTED; 647 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 648} 649 650/* 651 * vm_page_dirty: 652 * 653 * make page all dirty 654 */ 655void 656vm_page_dirty(vm_page_t m) 657{ 658 659 KASSERT((m->flags & PG_CACHED) == 0, 660 ("vm_page_dirty: page in cache!")); 661 KASSERT(!VM_PAGE_IS_FREE(m), 662 ("vm_page_dirty: page is free!")); 663 KASSERT(m->valid == VM_PAGE_BITS_ALL, 664 ("vm_page_dirty: page is invalid!")); 665 m->dirty = VM_PAGE_BITS_ALL; 666} 667 668/* 669 * vm_page_splay: 670 * 671 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 672 * the vm_page containing the given pindex. If, however, that 673 * pindex is not found in the vm_object, returns a vm_page that is 674 * adjacent to the pindex, coming before or after it. 675 */ 676vm_page_t 677vm_page_splay(vm_pindex_t pindex, vm_page_t root) 678{ 679 struct vm_page dummy; 680 vm_page_t lefttreemax, righttreemin, y; 681 682 if (root == NULL) 683 return (root); 684 lefttreemax = righttreemin = &dummy; 685 for (;; root = y) { 686 if (pindex < root->pindex) { 687 if ((y = root->left) == NULL) 688 break; 689 if (pindex < y->pindex) { 690 /* Rotate right. */ 691 root->left = y->right; 692 y->right = root; 693 root = y; 694 if ((y = root->left) == NULL) 695 break; 696 } 697 /* Link into the new root's right tree. */ 698 righttreemin->left = root; 699 righttreemin = root; 700 } else if (pindex > root->pindex) { 701 if ((y = root->right) == NULL) 702 break; 703 if (pindex > y->pindex) { 704 /* Rotate left. */ 705 root->right = y->left; 706 y->left = root; 707 root = y; 708 if ((y = root->right) == NULL) 709 break; 710 } 711 /* Link into the new root's left tree. */ 712 lefttreemax->right = root; 713 lefttreemax = root; 714 } else 715 break; 716 } 717 /* Assemble the new root. */ 718 lefttreemax->right = root->left; 719 righttreemin->left = root->right; 720 root->left = dummy.right; 721 root->right = dummy.left; 722 return (root); 723} 724 725/* 726 * vm_page_insert: [ internal use only ] 727 * 728 * Inserts the given mem entry into the object and object list. 729 * 730 * The pagetables are not updated but will presumably fault the page 731 * in if necessary, or if a kernel page the caller will at some point 732 * enter the page into the kernel's pmap. We are not allowed to block 733 * here so we *can't* do this anyway. 734 * 735 * The object and page must be locked. 736 * This routine may not block. 737 */ 738void 739vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 740{ 741 vm_page_t root; 742 743 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 744 if (m->object != NULL) 745 panic("vm_page_insert: page already inserted"); 746 747 /* 748 * Record the object/offset pair in this page 749 */ 750 m->object = object; 751 m->pindex = pindex; 752 753 /* 754 * Now link into the object's ordered list of backed pages. 755 */ 756 root = object->root; 757 if (root == NULL) { 758 m->left = NULL; 759 m->right = NULL; 760 TAILQ_INSERT_TAIL(&object->memq, m, listq); 761 } else { 762 root = vm_page_splay(pindex, root); 763 if (pindex < root->pindex) { 764 m->left = root->left; 765 m->right = root; 766 root->left = NULL; 767 TAILQ_INSERT_BEFORE(root, m, listq); 768 } else if (pindex == root->pindex) 769 panic("vm_page_insert: offset already allocated"); 770 else { 771 m->right = root->right; 772 m->left = root; 773 root->right = NULL; 774 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 775 } 776 } 777 object->root = m; 778 object->generation++; 779 780 /* 781 * show that the object has one more resident page. 782 */ 783 object->resident_page_count++; 784 /* 785 * Hold the vnode until the last page is released. 786 */ 787 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 788 vhold((struct vnode *)object->handle); 789 790 /* 791 * Since we are inserting a new and possibly dirty page, 792 * update the object's OBJ_MIGHTBEDIRTY flag. 793 */ 794 if (m->flags & PG_WRITEABLE) 795 vm_object_set_writeable_dirty(object); 796} 797 798/* 799 * vm_page_remove: 800 * NOTE: used by device pager as well -wfj 801 * 802 * Removes the given mem entry from the object/offset-page 803 * table and the object page list, but do not invalidate/terminate 804 * the backing store. 805 * 806 * The object and page must be locked. 807 * The underlying pmap entry (if any) is NOT removed here. 808 * This routine may not block. 809 */ 810void 811vm_page_remove(vm_page_t m) 812{ 813 vm_object_t object; 814 vm_page_t root; 815 816 if ((m->flags & PG_UNMANAGED) == 0) 817 vm_page_lock_assert(m, MA_OWNED); 818 if ((object = m->object) == NULL) 819 return; 820 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 821 if (m->oflags & VPO_BUSY) { 822 m->oflags &= ~VPO_BUSY; 823 vm_page_flash(m); 824 } 825 826 /* 827 * Now remove from the object's list of backed pages. 828 */ 829 if (m != object->root) 830 vm_page_splay(m->pindex, object->root); 831 if (m->left == NULL) 832 root = m->right; 833 else { 834 root = vm_page_splay(m->pindex, m->left); 835 root->right = m->right; 836 } 837 object->root = root; 838 TAILQ_REMOVE(&object->memq, m, listq); 839 840 /* 841 * And show that the object has one fewer resident page. 842 */ 843 object->resident_page_count--; 844 object->generation++; 845 /* 846 * The vnode may now be recycled. 847 */ 848 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 849 vdrop((struct vnode *)object->handle); 850 851 m->object = NULL; 852} 853 854/* 855 * vm_page_lookup: 856 * 857 * Returns the page associated with the object/offset 858 * pair specified; if none is found, NULL is returned. 859 * 860 * The object must be locked. 861 * This routine may not block. 862 * This is a critical path routine 863 */ 864vm_page_t 865vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 866{ 867 vm_page_t m; 868 869 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 870 if ((m = object->root) != NULL && m->pindex != pindex) { 871 m = vm_page_splay(pindex, m); 872 if ((object->root = m)->pindex != pindex) 873 m = NULL; 874 } 875 return (m); 876} 877 878/* 879 * vm_page_rename: 880 * 881 * Move the given memory entry from its 882 * current object to the specified target object/offset. 883 * 884 * The object must be locked. 885 * This routine may not block. 886 * 887 * Note: swap associated with the page must be invalidated by the move. We 888 * have to do this for several reasons: (1) we aren't freeing the 889 * page, (2) we are dirtying the page, (3) the VM system is probably 890 * moving the page from object A to B, and will then later move 891 * the backing store from A to B and we can't have a conflict. 892 * 893 * Note: we *always* dirty the page. It is necessary both for the 894 * fact that we moved it, and because we may be invalidating 895 * swap. If the page is on the cache, we have to deactivate it 896 * or vm_page_dirty() will panic. Dirty pages are not allowed 897 * on the cache. 898 */ 899void 900vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 901{ 902 903 vm_page_remove(m); 904 vm_page_insert(m, new_object, new_pindex); 905 vm_page_dirty(m); 906} 907 908/* 909 * Convert all of the given object's cached pages that have a 910 * pindex within the given range into free pages. If the value 911 * zero is given for "end", then the range's upper bound is 912 * infinity. If the given object is backed by a vnode and it 913 * transitions from having one or more cached pages to none, the 914 * vnode's hold count is reduced. 915 */ 916void 917vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 918{ 919 vm_page_t m, m_next; 920 boolean_t empty; 921 922 mtx_lock(&vm_page_queue_free_mtx); 923 if (__predict_false(object->cache == NULL)) { 924 mtx_unlock(&vm_page_queue_free_mtx); 925 return; 926 } 927 m = object->cache = vm_page_splay(start, object->cache); 928 if (m->pindex < start) { 929 if (m->right == NULL) 930 m = NULL; 931 else { 932 m_next = vm_page_splay(start, m->right); 933 m_next->left = m; 934 m->right = NULL; 935 m = object->cache = m_next; 936 } 937 } 938 939 /* 940 * At this point, "m" is either (1) a reference to the page 941 * with the least pindex that is greater than or equal to 942 * "start" or (2) NULL. 943 */ 944 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 945 /* 946 * Find "m"'s successor and remove "m" from the 947 * object's cache. 948 */ 949 if (m->right == NULL) { 950 object->cache = m->left; 951 m_next = NULL; 952 } else { 953 m_next = vm_page_splay(start, m->right); 954 m_next->left = m->left; 955 object->cache = m_next; 956 } 957 /* Convert "m" to a free page. */ 958 m->object = NULL; 959 m->valid = 0; 960 /* Clear PG_CACHED and set PG_FREE. */ 961 m->flags ^= PG_CACHED | PG_FREE; 962 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 963 ("vm_page_cache_free: page %p has inconsistent flags", m)); 964 cnt.v_cache_count--; 965 cnt.v_free_count++; 966 } 967 empty = object->cache == NULL; 968 mtx_unlock(&vm_page_queue_free_mtx); 969 if (object->type == OBJT_VNODE && empty) 970 vdrop(object->handle); 971} 972 973/* 974 * Returns the cached page that is associated with the given 975 * object and offset. If, however, none exists, returns NULL. 976 * 977 * The free page queue must be locked. 978 */ 979static inline vm_page_t 980vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 981{ 982 vm_page_t m; 983 984 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 985 if ((m = object->cache) != NULL && m->pindex != pindex) { 986 m = vm_page_splay(pindex, m); 987 if ((object->cache = m)->pindex != pindex) 988 m = NULL; 989 } 990 return (m); 991} 992 993/* 994 * Remove the given cached page from its containing object's 995 * collection of cached pages. 996 * 997 * The free page queue must be locked. 998 */ 999void 1000vm_page_cache_remove(vm_page_t m) 1001{ 1002 vm_object_t object; 1003 vm_page_t root; 1004 1005 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1006 KASSERT((m->flags & PG_CACHED) != 0, 1007 ("vm_page_cache_remove: page %p is not cached", m)); 1008 object = m->object; 1009 if (m != object->cache) { 1010 root = vm_page_splay(m->pindex, object->cache); 1011 KASSERT(root == m, 1012 ("vm_page_cache_remove: page %p is not cached in object %p", 1013 m, object)); 1014 } 1015 if (m->left == NULL) 1016 root = m->right; 1017 else if (m->right == NULL) 1018 root = m->left; 1019 else { 1020 root = vm_page_splay(m->pindex, m->left); 1021 root->right = m->right; 1022 } 1023 object->cache = root; 1024 m->object = NULL; 1025 cnt.v_cache_count--; 1026} 1027 1028/* 1029 * Transfer all of the cached pages with offset greater than or 1030 * equal to 'offidxstart' from the original object's cache to the 1031 * new object's cache. However, any cached pages with offset 1032 * greater than or equal to the new object's size are kept in the 1033 * original object. Initially, the new object's cache must be 1034 * empty. Offset 'offidxstart' in the original object must 1035 * correspond to offset zero in the new object. 1036 * 1037 * The new object must be locked. 1038 */ 1039void 1040vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1041 vm_object_t new_object) 1042{ 1043 vm_page_t m, m_next; 1044 1045 /* 1046 * Insertion into an object's collection of cached pages 1047 * requires the object to be locked. In contrast, removal does 1048 * not. 1049 */ 1050 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 1051 KASSERT(new_object->cache == NULL, 1052 ("vm_page_cache_transfer: object %p has cached pages", 1053 new_object)); 1054 mtx_lock(&vm_page_queue_free_mtx); 1055 if ((m = orig_object->cache) != NULL) { 1056 /* 1057 * Transfer all of the pages with offset greater than or 1058 * equal to 'offidxstart' from the original object's 1059 * cache to the new object's cache. 1060 */ 1061 m = vm_page_splay(offidxstart, m); 1062 if (m->pindex < offidxstart) { 1063 orig_object->cache = m; 1064 new_object->cache = m->right; 1065 m->right = NULL; 1066 } else { 1067 orig_object->cache = m->left; 1068 new_object->cache = m; 1069 m->left = NULL; 1070 } 1071 while ((m = new_object->cache) != NULL) { 1072 if ((m->pindex - offidxstart) >= new_object->size) { 1073 /* 1074 * Return all of the cached pages with 1075 * offset greater than or equal to the 1076 * new object's size to the original 1077 * object's cache. 1078 */ 1079 new_object->cache = m->left; 1080 m->left = orig_object->cache; 1081 orig_object->cache = m; 1082 break; 1083 } 1084 m_next = vm_page_splay(m->pindex, m->right); 1085 /* Update the page's object and offset. */ 1086 m->object = new_object; 1087 m->pindex -= offidxstart; 1088 if (m_next == NULL) 1089 break; 1090 m->right = NULL; 1091 m_next->left = m; 1092 new_object->cache = m_next; 1093 } 1094 KASSERT(new_object->cache == NULL || 1095 new_object->type == OBJT_SWAP, 1096 ("vm_page_cache_transfer: object %p's type is incompatible" 1097 " with cached pages", new_object)); 1098 } 1099 mtx_unlock(&vm_page_queue_free_mtx); 1100} 1101 1102/* 1103 * vm_page_alloc: 1104 * 1105 * Allocate and return a memory cell associated 1106 * with this VM object/offset pair. 1107 * 1108 * page_req classes: 1109 * VM_ALLOC_NORMAL normal process request 1110 * VM_ALLOC_SYSTEM system *really* needs a page 1111 * VM_ALLOC_INTERRUPT interrupt time request 1112 * VM_ALLOC_ZERO zero page 1113 * VM_ALLOC_WIRED wire the allocated page 1114 * VM_ALLOC_NOOBJ page is not associated with a vm object 1115 * VM_ALLOC_NOBUSY do not set the page busy 1116 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1117 * is cached 1118 * 1119 * This routine may not sleep. 1120 */ 1121vm_page_t 1122vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1123{ 1124 struct vnode *vp = NULL; 1125 vm_object_t m_object; 1126 vm_page_t m; 1127 int flags, page_req; 1128 1129 page_req = req & VM_ALLOC_CLASS_MASK; 1130 KASSERT(curthread->td_intr_nesting_level == 0 || 1131 page_req == VM_ALLOC_INTERRUPT, 1132 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 1133 1134 if ((req & VM_ALLOC_NOOBJ) == 0) { 1135 KASSERT(object != NULL, 1136 ("vm_page_alloc: NULL object.")); 1137 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1138 } 1139 1140 /* 1141 * The pager is allowed to eat deeper into the free page list. 1142 */ 1143 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 1144 page_req = VM_ALLOC_SYSTEM; 1145 }; 1146 1147 mtx_lock(&vm_page_queue_free_mtx); 1148 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1149 (page_req == VM_ALLOC_SYSTEM && 1150 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1151 (page_req == VM_ALLOC_INTERRUPT && 1152 cnt.v_free_count + cnt.v_cache_count > 0)) { 1153 /* 1154 * Allocate from the free queue if the number of free pages 1155 * exceeds the minimum for the request class. 1156 */ 1157 if (object != NULL && 1158 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1159 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1160 mtx_unlock(&vm_page_queue_free_mtx); 1161 return (NULL); 1162 } 1163 if (vm_phys_unfree_page(m)) 1164 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1165#if VM_NRESERVLEVEL > 0 1166 else if (!vm_reserv_reactivate_page(m)) 1167#else 1168 else 1169#endif 1170 panic("vm_page_alloc: cache page %p is missing" 1171 " from the free queue", m); 1172 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1173 mtx_unlock(&vm_page_queue_free_mtx); 1174 return (NULL); 1175#if VM_NRESERVLEVEL > 0 1176 } else if (object == NULL || object->type == OBJT_DEVICE || 1177 object->type == OBJT_SG || 1178 (object->flags & OBJ_COLORED) == 0 || 1179 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1180#else 1181 } else { 1182#endif 1183 m = vm_phys_alloc_pages(object != NULL ? 1184 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1185#if VM_NRESERVLEVEL > 0 1186 if (m == NULL && vm_reserv_reclaim_inactive()) { 1187 m = vm_phys_alloc_pages(object != NULL ? 1188 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1189 0); 1190 } 1191#endif 1192 } 1193 } else { 1194 /* 1195 * Not allocatable, give up. 1196 */ 1197 mtx_unlock(&vm_page_queue_free_mtx); 1198 atomic_add_int(&vm_pageout_deficit, 1); 1199 pagedaemon_wakeup(); 1200 return (NULL); 1201 } 1202 1203 /* 1204 * At this point we had better have found a good page. 1205 */ 1206 1207 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1208 KASSERT(m->queue == PQ_NONE, 1209 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1210 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1211 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1212 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1213 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1214 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1215 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1216 pmap_page_get_memattr(m))); 1217 if ((m->flags & PG_CACHED) != 0) { 1218 KASSERT(m->valid != 0, 1219 ("vm_page_alloc: cached page %p is invalid", m)); 1220 if (m->object == object && m->pindex == pindex) 1221 cnt.v_reactivated++; 1222 else 1223 m->valid = 0; 1224 m_object = m->object; 1225 vm_page_cache_remove(m); 1226 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1227 vp = m_object->handle; 1228 } else { 1229 KASSERT(VM_PAGE_IS_FREE(m), 1230 ("vm_page_alloc: page %p is not free", m)); 1231 KASSERT(m->valid == 0, 1232 ("vm_page_alloc: free page %p is valid", m)); 1233 cnt.v_free_count--; 1234 } 1235 1236 /* 1237 * Initialize structure. Only the PG_ZERO flag is inherited. 1238 */ 1239 flags = 0; 1240 if (m->flags & PG_ZERO) { 1241 vm_page_zero_count--; 1242 if (req & VM_ALLOC_ZERO) 1243 flags = PG_ZERO; 1244 } 1245 if (object == NULL || object->type == OBJT_PHYS) 1246 flags |= PG_UNMANAGED; 1247 m->flags = flags; 1248 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 1249 m->oflags = 0; 1250 else 1251 m->oflags = VPO_BUSY; 1252 if (req & VM_ALLOC_WIRED) { 1253 atomic_add_int(&cnt.v_wire_count, 1); 1254 m->wire_count = 1; 1255 } 1256 m->act_count = 0; 1257 mtx_unlock(&vm_page_queue_free_mtx); 1258 1259 if (object != NULL) { 1260 /* Ignore device objects; the pager sets "memattr" for them. */ 1261 if (object->memattr != VM_MEMATTR_DEFAULT && 1262 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1263 pmap_page_set_memattr(m, object->memattr); 1264 vm_page_insert(m, object, pindex); 1265 } else 1266 m->pindex = pindex; 1267 1268 /* 1269 * The following call to vdrop() must come after the above call 1270 * to vm_page_insert() in case both affect the same object and 1271 * vnode. Otherwise, the affected vnode's hold count could 1272 * temporarily become zero. 1273 */ 1274 if (vp != NULL) 1275 vdrop(vp); 1276 1277 /* 1278 * Don't wakeup too often - wakeup the pageout daemon when 1279 * we would be nearly out of memory. 1280 */ 1281 if (vm_paging_needed()) 1282 pagedaemon_wakeup(); 1283 1284 return (m); 1285} 1286 1287/* 1288 * vm_wait: (also see VM_WAIT macro) 1289 * 1290 * Block until free pages are available for allocation 1291 * - Called in various places before memory allocations. 1292 */ 1293void 1294vm_wait(void) 1295{ 1296 1297 mtx_lock(&vm_page_queue_free_mtx); 1298 if (curproc == pageproc) { 1299 vm_pageout_pages_needed = 1; 1300 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1301 PDROP | PSWP, "VMWait", 0); 1302 } else { 1303 if (!vm_pages_needed) { 1304 vm_pages_needed = 1; 1305 wakeup(&vm_pages_needed); 1306 } 1307 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1308 "vmwait", 0); 1309 } 1310} 1311 1312/* 1313 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1314 * 1315 * Block until free pages are available for allocation 1316 * - Called only in vm_fault so that processes page faulting 1317 * can be easily tracked. 1318 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1319 * processes will be able to grab memory first. Do not change 1320 * this balance without careful testing first. 1321 */ 1322void 1323vm_waitpfault(void) 1324{ 1325 1326 mtx_lock(&vm_page_queue_free_mtx); 1327 if (!vm_pages_needed) { 1328 vm_pages_needed = 1; 1329 wakeup(&vm_pages_needed); 1330 } 1331 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1332 "pfault", 0); 1333} 1334 1335/* 1336 * vm_page_requeue: 1337 * 1338 * Move the given page to the tail of its present page queue. 1339 * 1340 * The page queues must be locked. 1341 */ 1342void 1343vm_page_requeue(vm_page_t m) 1344{ 1345 int queue = VM_PAGE_GETQUEUE(m); 1346 struct vpgqueues *vpq; 1347 1348 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1349 KASSERT(queue != PQ_NONE, 1350 ("vm_page_requeue: page %p is not queued", m)); 1351 vpq = &vm_page_queues[queue]; 1352 TAILQ_REMOVE(&vpq->pl, m, pageq); 1353 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1354} 1355 1356/* 1357 * vm_page_queue_remove: 1358 * 1359 * Remove the given page from the specified queue. 1360 * 1361 * The page and page queues must be locked. 1362 */ 1363static __inline void 1364vm_page_queue_remove(int queue, vm_page_t m) 1365{ 1366 struct vpgqueues *pq; 1367 1368 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1369 vm_page_lock_assert(m, MA_OWNED); 1370 pq = &vm_page_queues[queue]; 1371 TAILQ_REMOVE(&pq->pl, m, pageq); 1372 (*pq->cnt)--; 1373} 1374 1375/* 1376 * vm_pageq_remove: 1377 * 1378 * Remove a page from its queue. 1379 * 1380 * The given page must be locked. 1381 * This routine may not block. 1382 */ 1383void 1384vm_pageq_remove(vm_page_t m) 1385{ 1386 int queue = VM_PAGE_GETQUEUE(m); 1387 1388 vm_page_lock_assert(m, MA_OWNED); 1389 if (queue != PQ_NONE) { 1390 vm_page_lock_queues(); 1391 VM_PAGE_SETQUEUE2(m, PQ_NONE); 1392 vm_page_queue_remove(queue, m); 1393 vm_page_unlock_queues(); 1394 } 1395} 1396 1397/* 1398 * vm_page_enqueue: 1399 * 1400 * Add the given page to the specified queue. 1401 * 1402 * The page queues must be locked. 1403 */ 1404static void 1405vm_page_enqueue(int queue, vm_page_t m) 1406{ 1407 struct vpgqueues *vpq; 1408 1409 vpq = &vm_page_queues[queue]; 1410 VM_PAGE_SETQUEUE2(m, queue); 1411 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1412 ++*vpq->cnt; 1413} 1414 1415/* 1416 * vm_page_activate: 1417 * 1418 * Put the specified page on the active list (if appropriate). 1419 * Ensure that act_count is at least ACT_INIT but do not otherwise 1420 * mess with it. 1421 * 1422 * The page must be locked. 1423 * This routine may not block. 1424 */ 1425void 1426vm_page_activate(vm_page_t m) 1427{ 1428 int queue; 1429 1430 vm_page_lock_assert(m, MA_OWNED); 1431 if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) != PQ_ACTIVE) { 1432 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1433 if (m->act_count < ACT_INIT) 1434 m->act_count = ACT_INIT; 1435 vm_page_lock_queues(); 1436 if (queue != PQ_NONE) 1437 vm_page_queue_remove(queue, m); 1438 vm_page_enqueue(PQ_ACTIVE, m); 1439 vm_page_unlock_queues(); 1440 } else 1441 KASSERT(queue == PQ_NONE, 1442 ("vm_page_activate: wired page %p is queued", m)); 1443 } else { 1444 if (m->act_count < ACT_INIT) 1445 m->act_count = ACT_INIT; 1446 } 1447} 1448 1449/* 1450 * vm_page_free_wakeup: 1451 * 1452 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1453 * routine is called when a page has been added to the cache or free 1454 * queues. 1455 * 1456 * The page queues must be locked. 1457 * This routine may not block. 1458 */ 1459static inline void 1460vm_page_free_wakeup(void) 1461{ 1462 1463 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1464 /* 1465 * if pageout daemon needs pages, then tell it that there are 1466 * some free. 1467 */ 1468 if (vm_pageout_pages_needed && 1469 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1470 wakeup(&vm_pageout_pages_needed); 1471 vm_pageout_pages_needed = 0; 1472 } 1473 /* 1474 * wakeup processes that are waiting on memory if we hit a 1475 * high water mark. And wakeup scheduler process if we have 1476 * lots of memory. this process will swapin processes. 1477 */ 1478 if (vm_pages_needed && !vm_page_count_min()) { 1479 vm_pages_needed = 0; 1480 wakeup(&cnt.v_free_count); 1481 } 1482} 1483 1484/* 1485 * vm_page_free_toq: 1486 * 1487 * Returns the given page to the free list, 1488 * disassociating it with any VM object. 1489 * 1490 * Object and page must be locked prior to entry. 1491 * This routine may not block. 1492 */ 1493 1494void 1495vm_page_free_toq(vm_page_t m) 1496{ 1497 1498 if ((m->flags & PG_UNMANAGED) == 0) { 1499 vm_page_lock_assert(m, MA_OWNED); 1500 KASSERT(!pmap_page_is_mapped(m), 1501 ("vm_page_free_toq: freeing mapped page %p", m)); 1502 } 1503 PCPU_INC(cnt.v_tfree); 1504 1505 if (m->busy || VM_PAGE_IS_FREE(m)) { 1506 printf( 1507 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1508 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1509 m->hold_count); 1510 if (VM_PAGE_IS_FREE(m)) 1511 panic("vm_page_free: freeing free page"); 1512 else 1513 panic("vm_page_free: freeing busy page"); 1514 } 1515 1516 /* 1517 * unqueue, then remove page. Note that we cannot destroy 1518 * the page here because we do not want to call the pager's 1519 * callback routine until after we've put the page on the 1520 * appropriate free queue. 1521 */ 1522 if ((m->flags & PG_UNMANAGED) == 0) 1523 vm_pageq_remove(m); 1524 vm_page_remove(m); 1525 1526 /* 1527 * If fictitious remove object association and 1528 * return, otherwise delay object association removal. 1529 */ 1530 if ((m->flags & PG_FICTITIOUS) != 0) { 1531 return; 1532 } 1533 1534 m->valid = 0; 1535 vm_page_undirty(m); 1536 1537 if (m->wire_count != 0) { 1538 if (m->wire_count > 1) { 1539 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1540 m->wire_count, (long)m->pindex); 1541 } 1542 panic("vm_page_free: freeing wired page"); 1543 } 1544 if (m->hold_count != 0) { 1545 m->flags &= ~PG_ZERO; 1546 vm_page_lock_queues(); 1547 vm_page_enqueue(PQ_HOLD, m); 1548 vm_page_unlock_queues(); 1549 } else { 1550 /* 1551 * Restore the default memory attribute to the page. 1552 */ 1553 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1554 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1555 1556 /* 1557 * Insert the page into the physical memory allocator's 1558 * cache/free page queues. 1559 */ 1560 mtx_lock(&vm_page_queue_free_mtx); 1561 m->flags |= PG_FREE; 1562 cnt.v_free_count++; 1563#if VM_NRESERVLEVEL > 0 1564 if (!vm_reserv_free_page(m)) 1565#else 1566 if (TRUE) 1567#endif 1568 vm_phys_free_pages(m, 0); 1569 if ((m->flags & PG_ZERO) != 0) 1570 ++vm_page_zero_count; 1571 else 1572 vm_page_zero_idle_wakeup(); 1573 vm_page_free_wakeup(); 1574 mtx_unlock(&vm_page_queue_free_mtx); 1575 } 1576} 1577 1578/* 1579 * vm_page_wire: 1580 * 1581 * Mark this page as wired down by yet 1582 * another map, removing it from paging queues 1583 * as necessary. 1584 * 1585 * The page must be locked. 1586 * This routine may not block. 1587 */ 1588void 1589vm_page_wire(vm_page_t m) 1590{ 1591 1592 /* 1593 * Only bump the wire statistics if the page is not already wired, 1594 * and only unqueue the page if it is on some queue (if it is unmanaged 1595 * it is already off the queues). 1596 */ 1597 vm_page_lock_assert(m, MA_OWNED); 1598 if (m->flags & PG_FICTITIOUS) 1599 return; 1600 if (m->wire_count == 0) { 1601 if ((m->flags & PG_UNMANAGED) == 0) 1602 vm_pageq_remove(m); 1603 atomic_add_int(&cnt.v_wire_count, 1); 1604 } 1605 m->wire_count++; 1606 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1607} 1608 1609/* 1610 * vm_page_unwire: 1611 * 1612 * Release one wiring of this page, potentially 1613 * enabling it to be paged again. 1614 * 1615 * Many pages placed on the inactive queue should actually go 1616 * into the cache, but it is difficult to figure out which. What 1617 * we do instead, if the inactive target is well met, is to put 1618 * clean pages at the head of the inactive queue instead of the tail. 1619 * This will cause them to be moved to the cache more quickly and 1620 * if not actively re-referenced, freed more quickly. If we just 1621 * stick these pages at the end of the inactive queue, heavy filesystem 1622 * meta-data accesses can cause an unnecessary paging load on memory bound 1623 * processes. This optimization causes one-time-use metadata to be 1624 * reused more quickly. 1625 * 1626 * BUT, if we are in a low-memory situation we have no choice but to 1627 * put clean pages on the cache queue. 1628 * 1629 * A number of routines use vm_page_unwire() to guarantee that the page 1630 * will go into either the inactive or active queues, and will NEVER 1631 * be placed in the cache - for example, just after dirtying a page. 1632 * dirty pages in the cache are not allowed. 1633 * 1634 * The page must be locked. 1635 * This routine may not block. 1636 */ 1637void 1638vm_page_unwire(vm_page_t m, int activate) 1639{ 1640 1641 if ((m->flags & PG_UNMANAGED) == 0) 1642 vm_page_lock_assert(m, MA_OWNED); 1643 if (m->flags & PG_FICTITIOUS) 1644 return; 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 return; 1651 vm_page_lock_queues(); 1652 if (activate) 1653 vm_page_enqueue(PQ_ACTIVE, m); 1654 else { 1655 vm_page_flag_clear(m, PG_WINATCFLS); 1656 vm_page_enqueue(PQ_INACTIVE, m); 1657 } 1658 vm_page_unlock_queues(); 1659 } 1660 } else { 1661 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1662 } 1663} 1664 1665/* 1666 * Move the specified page to the inactive queue. 1667 * 1668 * Normally athead is 0 resulting in LRU operation. athead is set 1669 * to 1 if we want this page to be 'as if it were placed in the cache', 1670 * except without unmapping it from the process address space. 1671 * 1672 * This routine may not block. 1673 */ 1674static inline void 1675_vm_page_deactivate(vm_page_t m, int athead) 1676{ 1677 int queue; 1678 1679 vm_page_lock_assert(m, MA_OWNED); 1680 1681 /* 1682 * Ignore if already inactive. 1683 */ 1684 if ((queue = VM_PAGE_GETKNOWNQUEUE2(m)) == PQ_INACTIVE) 1685 return; 1686 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1687 vm_page_lock_queues(); 1688 vm_page_flag_clear(m, PG_WINATCFLS); 1689 if (queue != PQ_NONE) 1690 vm_page_queue_remove(queue, m); 1691 if (athead) 1692 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, 1693 pageq); 1694 else 1695 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, 1696 pageq); 1697 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1698 cnt.v_inactive_count++; 1699 vm_page_unlock_queues(); 1700 } 1701} 1702 1703/* 1704 * Move the specified page to the inactive queue. 1705 * 1706 * The page must be locked. 1707 */ 1708void 1709vm_page_deactivate(vm_page_t m) 1710{ 1711 1712 _vm_page_deactivate(m, 0); 1713} 1714 1715/* 1716 * vm_page_try_to_cache: 1717 * 1718 * Returns 0 on failure, 1 on success 1719 */ 1720int 1721vm_page_try_to_cache(vm_page_t m) 1722{ 1723 1724 vm_page_lock_assert(m, MA_OWNED); 1725 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1726 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1727 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1728 return (0); 1729 pmap_remove_all(m); 1730 if (m->dirty) 1731 return (0); 1732 vm_page_cache(m); 1733 return (1); 1734} 1735 1736/* 1737 * vm_page_try_to_free() 1738 * 1739 * Attempt to free the page. If we cannot free it, we do nothing. 1740 * 1 is returned on success, 0 on failure. 1741 */ 1742int 1743vm_page_try_to_free(vm_page_t m) 1744{ 1745 1746 vm_page_lock_assert(m, MA_OWNED); 1747 if (m->object != NULL) 1748 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1749 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1750 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1751 return (0); 1752 pmap_remove_all(m); 1753 if (m->dirty) 1754 return (0); 1755 vm_page_free(m); 1756 return (1); 1757} 1758 1759/* 1760 * vm_page_cache 1761 * 1762 * Put the specified page onto the page cache queue (if appropriate). 1763 * 1764 * This routine may not block. 1765 */ 1766void 1767vm_page_cache(vm_page_t m) 1768{ 1769 vm_object_t object; 1770 vm_page_t root; 1771 1772 vm_page_lock_assert(m, MA_OWNED); 1773 object = m->object; 1774 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1775 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1776 m->hold_count || m->wire_count) 1777 panic("vm_page_cache: attempting to cache busy page"); 1778 pmap_remove_all(m); 1779 if (m->dirty != 0) 1780 panic("vm_page_cache: page %p is dirty", m); 1781 if (m->valid == 0 || object->type == OBJT_DEFAULT || 1782 (object->type == OBJT_SWAP && 1783 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 1784 /* 1785 * Hypothesis: A cache-elgible page belonging to a 1786 * default object or swap object but without a backing 1787 * store must be zero filled. 1788 */ 1789 vm_page_free(m); 1790 return; 1791 } 1792 KASSERT((m->flags & PG_CACHED) == 0, 1793 ("vm_page_cache: page %p is already cached", m)); 1794 PCPU_INC(cnt.v_tcached); 1795 1796 /* 1797 * Remove the page from the paging queues. 1798 */ 1799 vm_pageq_remove(m); 1800 1801 /* 1802 * Remove the page from the object's collection of resident 1803 * pages. 1804 */ 1805 if (m != object->root) 1806 vm_page_splay(m->pindex, object->root); 1807 if (m->left == NULL) 1808 root = m->right; 1809 else { 1810 root = vm_page_splay(m->pindex, m->left); 1811 root->right = m->right; 1812 } 1813 object->root = root; 1814 TAILQ_REMOVE(&object->memq, m, listq); 1815 object->resident_page_count--; 1816 object->generation++; 1817 1818 /* 1819 * Restore the default memory attribute to the page. 1820 */ 1821 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1822 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1823 1824 /* 1825 * Insert the page into the object's collection of cached pages 1826 * and the physical memory allocator's cache/free page queues. 1827 */ 1828 m->flags &= ~PG_ZERO; 1829 mtx_lock(&vm_page_queue_free_mtx); 1830 m->flags |= PG_CACHED; 1831 cnt.v_cache_count++; 1832 root = object->cache; 1833 if (root == NULL) { 1834 m->left = NULL; 1835 m->right = NULL; 1836 } else { 1837 root = vm_page_splay(m->pindex, root); 1838 if (m->pindex < root->pindex) { 1839 m->left = root->left; 1840 m->right = root; 1841 root->left = NULL; 1842 } else if (__predict_false(m->pindex == root->pindex)) 1843 panic("vm_page_cache: offset already cached"); 1844 else { 1845 m->right = root->right; 1846 m->left = root; 1847 root->right = NULL; 1848 } 1849 } 1850 object->cache = m; 1851#if VM_NRESERVLEVEL > 0 1852 if (!vm_reserv_free_page(m)) { 1853#else 1854 if (TRUE) { 1855#endif 1856 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 1857 vm_phys_free_pages(m, 0); 1858 } 1859 vm_page_free_wakeup(); 1860 mtx_unlock(&vm_page_queue_free_mtx); 1861 1862 /* 1863 * Increment the vnode's hold count if this is the object's only 1864 * cached page. Decrement the vnode's hold count if this was 1865 * the object's only resident page. 1866 */ 1867 if (object->type == OBJT_VNODE) { 1868 if (root == NULL && object->resident_page_count != 0) 1869 vhold(object->handle); 1870 else if (root != NULL && object->resident_page_count == 0) 1871 vdrop(object->handle); 1872 } 1873} 1874 1875/* 1876 * vm_page_dontneed 1877 * 1878 * Cache, deactivate, or do nothing as appropriate. This routine 1879 * is typically used by madvise() MADV_DONTNEED. 1880 * 1881 * Generally speaking we want to move the page into the cache so 1882 * it gets reused quickly. However, this can result in a silly syndrome 1883 * due to the page recycling too quickly. Small objects will not be 1884 * fully cached. On the otherhand, if we move the page to the inactive 1885 * queue we wind up with a problem whereby very large objects 1886 * unnecessarily blow away our inactive and cache queues. 1887 * 1888 * The solution is to move the pages based on a fixed weighting. We 1889 * either leave them alone, deactivate them, or move them to the cache, 1890 * where moving them to the cache has the highest weighting. 1891 * By forcing some pages into other queues we eventually force the 1892 * system to balance the queues, potentially recovering other unrelated 1893 * space from active. The idea is to not force this to happen too 1894 * often. 1895 */ 1896void 1897vm_page_dontneed(vm_page_t m) 1898{ 1899 int dnw; 1900 int head; 1901 1902 vm_page_lock_assert(m, MA_OWNED); 1903 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1904 dnw = PCPU_GET(dnweight); 1905 PCPU_INC(dnweight); 1906 1907 /* 1908 * Occasionally leave the page alone. 1909 */ 1910 if ((dnw & 0x01F0) == 0 || 1911 VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) { 1912 if (m->act_count >= ACT_INIT) 1913 --m->act_count; 1914 return; 1915 } 1916 1917 /* 1918 * Clear any references to the page. Otherwise, the page daemon will 1919 * immediately reactivate the page. 1920 * 1921 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 1922 * pmap operation, such as pmap_remove(), could clear a reference in 1923 * the pmap and set PG_REFERENCED on the page before the 1924 * pmap_clear_reference() had completed. Consequently, the page would 1925 * appear referenced based upon an old reference that occurred before 1926 * this function ran. 1927 */ 1928 pmap_clear_reference(m); 1929 vm_page_lock_queues(); 1930 vm_page_flag_clear(m, PG_REFERENCED); 1931 vm_page_unlock_queues(); 1932 1933 if (m->dirty == 0 && pmap_is_modified(m)) 1934 vm_page_dirty(m); 1935 1936 if (m->dirty || (dnw & 0x0070) == 0) { 1937 /* 1938 * Deactivate the page 3 times out of 32. 1939 */ 1940 head = 0; 1941 } else { 1942 /* 1943 * Cache the page 28 times out of every 32. Note that 1944 * the page is deactivated instead of cached, but placed 1945 * at the head of the queue instead of the tail. 1946 */ 1947 head = 1; 1948 } 1949 _vm_page_deactivate(m, head); 1950} 1951 1952/* 1953 * Grab a page, waiting until we are waken up due to the page 1954 * changing state. We keep on waiting, if the page continues 1955 * to be in the object. If the page doesn't exist, first allocate it 1956 * and then conditionally zero it. 1957 * 1958 * This routine may block. 1959 */ 1960vm_page_t 1961vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1962{ 1963 vm_page_t m; 1964 1965 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1966retrylookup: 1967 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1968 if ((m->oflags & VPO_BUSY) != 0 || m->busy != 0) { 1969 if ((allocflags & VM_ALLOC_RETRY) != 0) { 1970 /* 1971 * Reference the page before unlocking and 1972 * sleeping so that the page daemon is less 1973 * likely to reclaim it. 1974 */ 1975 vm_page_lock_queues(); 1976 vm_page_flag_set(m, PG_REFERENCED); 1977 } 1978 vm_page_sleep(m, "pgrbwt"); 1979 if ((allocflags & VM_ALLOC_RETRY) == 0) 1980 return (NULL); 1981 goto retrylookup; 1982 } else { 1983 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1984 vm_page_lock(m); 1985 vm_page_wire(m); 1986 vm_page_unlock(m); 1987 } 1988 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1989 vm_page_busy(m); 1990 return (m); 1991 } 1992 } 1993 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1994 if (m == NULL) { 1995 VM_OBJECT_UNLOCK(object); 1996 VM_WAIT; 1997 VM_OBJECT_LOCK(object); 1998 if ((allocflags & VM_ALLOC_RETRY) == 0) 1999 return (NULL); 2000 goto retrylookup; 2001 } else if (m->valid != 0) 2002 return (m); 2003 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2004 pmap_zero_page(m); 2005 return (m); 2006} 2007 2008/* 2009 * Mapping function for valid bits or for dirty bits in 2010 * a page. May not block. 2011 * 2012 * Inputs are required to range within a page. 2013 */ 2014int 2015vm_page_bits(int base, int size) 2016{ 2017 int first_bit; 2018 int last_bit; 2019 2020 KASSERT( 2021 base + size <= PAGE_SIZE, 2022 ("vm_page_bits: illegal base/size %d/%d", base, size) 2023 ); 2024 2025 if (size == 0) /* handle degenerate case */ 2026 return (0); 2027 2028 first_bit = base >> DEV_BSHIFT; 2029 last_bit = (base + size - 1) >> DEV_BSHIFT; 2030 2031 return ((2 << last_bit) - (1 << first_bit)); 2032} 2033 2034/* 2035 * vm_page_set_valid: 2036 * 2037 * Sets portions of a page valid. The arguments are expected 2038 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2039 * of any partial chunks touched by the range. The invalid portion of 2040 * such chunks will be zeroed. 2041 * 2042 * (base + size) must be less then or equal to PAGE_SIZE. 2043 */ 2044void 2045vm_page_set_valid(vm_page_t m, int base, int size) 2046{ 2047 int endoff, frag; 2048 2049 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2050 if (size == 0) /* handle degenerate case */ 2051 return; 2052 2053 /* 2054 * If the base is not DEV_BSIZE aligned and the valid 2055 * bit is clear, we have to zero out a portion of the 2056 * first block. 2057 */ 2058 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2059 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2060 pmap_zero_page_area(m, frag, base - frag); 2061 2062 /* 2063 * If the ending offset is not DEV_BSIZE aligned and the 2064 * valid bit is clear, we have to zero out a portion of 2065 * the last block. 2066 */ 2067 endoff = base + size; 2068 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2069 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2070 pmap_zero_page_area(m, endoff, 2071 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2072 2073 /* 2074 * Assert that no previously invalid block that is now being validated 2075 * is already dirty. 2076 */ 2077 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2078 ("vm_page_set_valid: page %p is dirty", m)); 2079 2080 /* 2081 * Set valid bits inclusive of any overlap. 2082 */ 2083 m->valid |= vm_page_bits(base, size); 2084} 2085 2086/* 2087 * Clear the given bits from the specified page's dirty field. 2088 */ 2089static __inline void 2090vm_page_clear_dirty_mask(vm_page_t m, int pagebits) 2091{ 2092 2093 /* 2094 * If the object is locked and the page is neither VPO_BUSY nor 2095 * PG_WRITEABLE, then the page's dirty field cannot possibly be 2096 * modified by a concurrent pmap operation. 2097 */ 2098 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2099 if ((m->oflags & VPO_BUSY) == 0 && (m->flags & PG_WRITEABLE) == 0) 2100 m->dirty &= ~pagebits; 2101 else { 2102 vm_page_lock_queues(); 2103 m->dirty &= ~pagebits; 2104 vm_page_unlock_queues(); 2105 } 2106} 2107 2108/* 2109 * vm_page_set_validclean: 2110 * 2111 * Sets portions of a page valid and clean. The arguments are expected 2112 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2113 * of any partial chunks touched by the range. The invalid portion of 2114 * such chunks will be zero'd. 2115 * 2116 * This routine may not block. 2117 * 2118 * (base + size) must be less then or equal to PAGE_SIZE. 2119 */ 2120void 2121vm_page_set_validclean(vm_page_t m, int base, int size) 2122{ 2123 u_long oldvalid; 2124 int endoff, frag, pagebits; 2125 2126 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2127 if (size == 0) /* handle degenerate case */ 2128 return; 2129 2130 /* 2131 * If the base is not DEV_BSIZE aligned and the valid 2132 * bit is clear, we have to zero out a portion of the 2133 * first block. 2134 */ 2135 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2136 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2137 pmap_zero_page_area(m, frag, base - frag); 2138 2139 /* 2140 * If the ending offset is not DEV_BSIZE aligned and the 2141 * valid bit is clear, we have to zero out a portion of 2142 * the last block. 2143 */ 2144 endoff = base + size; 2145 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2146 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2147 pmap_zero_page_area(m, endoff, 2148 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2149 2150 /* 2151 * Set valid, clear dirty bits. If validating the entire 2152 * page we can safely clear the pmap modify bit. We also 2153 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2154 * takes a write fault on a MAP_NOSYNC memory area the flag will 2155 * be set again. 2156 * 2157 * We set valid bits inclusive of any overlap, but we can only 2158 * clear dirty bits for DEV_BSIZE chunks that are fully within 2159 * the range. 2160 */ 2161 oldvalid = m->valid; 2162 pagebits = vm_page_bits(base, size); 2163 m->valid |= pagebits; 2164#if 0 /* NOT YET */ 2165 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2166 frag = DEV_BSIZE - frag; 2167 base += frag; 2168 size -= frag; 2169 if (size < 0) 2170 size = 0; 2171 } 2172 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2173#endif 2174 if (base == 0 && size == PAGE_SIZE) { 2175 /* 2176 * The page can only be modified within the pmap if it is 2177 * mapped, and it can only be mapped if it was previously 2178 * fully valid. 2179 */ 2180 if (oldvalid == VM_PAGE_BITS_ALL) 2181 /* 2182 * Perform the pmap_clear_modify() first. Otherwise, 2183 * a concurrent pmap operation, such as 2184 * pmap_protect(), could clear a modification in the 2185 * pmap and set the dirty field on the page before 2186 * pmap_clear_modify() had begun and after the dirty 2187 * field was cleared here. 2188 */ 2189 pmap_clear_modify(m); 2190 m->dirty = 0; 2191 m->oflags &= ~VPO_NOSYNC; 2192 } else if (oldvalid != VM_PAGE_BITS_ALL) 2193 m->dirty &= ~pagebits; 2194 else 2195 vm_page_clear_dirty_mask(m, pagebits); 2196} 2197 2198void 2199vm_page_clear_dirty(vm_page_t m, int base, int size) 2200{ 2201 2202 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2203} 2204 2205/* 2206 * vm_page_set_invalid: 2207 * 2208 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2209 * valid and dirty bits for the effected areas are cleared. 2210 * 2211 * May not block. 2212 */ 2213void 2214vm_page_set_invalid(vm_page_t m, int base, int size) 2215{ 2216 int bits; 2217 2218 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2219 KASSERT((m->oflags & VPO_BUSY) == 0, 2220 ("vm_page_set_invalid: page %p is busy", m)); 2221 bits = vm_page_bits(base, size); 2222 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2223 pmap_remove_all(m); 2224 KASSERT(!pmap_page_is_mapped(m), 2225 ("vm_page_set_invalid: page %p is mapped", m)); 2226 m->valid &= ~bits; 2227 m->dirty &= ~bits; 2228 m->object->generation++; 2229} 2230 2231/* 2232 * vm_page_zero_invalid() 2233 * 2234 * The kernel assumes that the invalid portions of a page contain 2235 * garbage, but such pages can be mapped into memory by user code. 2236 * When this occurs, we must zero out the non-valid portions of the 2237 * page so user code sees what it expects. 2238 * 2239 * Pages are most often semi-valid when the end of a file is mapped 2240 * into memory and the file's size is not page aligned. 2241 */ 2242void 2243vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2244{ 2245 int b; 2246 int i; 2247 2248 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2249 /* 2250 * Scan the valid bits looking for invalid sections that 2251 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2252 * valid bit may be set ) have already been zerod by 2253 * vm_page_set_validclean(). 2254 */ 2255 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2256 if (i == (PAGE_SIZE / DEV_BSIZE) || 2257 (m->valid & (1 << i)) 2258 ) { 2259 if (i > b) { 2260 pmap_zero_page_area(m, 2261 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2262 } 2263 b = i + 1; 2264 } 2265 } 2266 2267 /* 2268 * setvalid is TRUE when we can safely set the zero'd areas 2269 * as being valid. We can do this if there are no cache consistancy 2270 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2271 */ 2272 if (setvalid) 2273 m->valid = VM_PAGE_BITS_ALL; 2274} 2275 2276/* 2277 * vm_page_is_valid: 2278 * 2279 * Is (partial) page valid? Note that the case where size == 0 2280 * will return FALSE in the degenerate case where the page is 2281 * entirely invalid, and TRUE otherwise. 2282 * 2283 * May not block. 2284 */ 2285int 2286vm_page_is_valid(vm_page_t m, int base, int size) 2287{ 2288 int bits = vm_page_bits(base, size); 2289 2290 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2291 if (m->valid && ((m->valid & bits) == bits)) 2292 return 1; 2293 else 2294 return 0; 2295} 2296 2297/* 2298 * update dirty bits from pmap/mmu. May not block. 2299 */ 2300void 2301vm_page_test_dirty(vm_page_t m) 2302{ 2303 2304 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2305 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2306 vm_page_dirty(m); 2307} 2308 2309int so_zerocp_fullpage = 0; 2310 2311/* 2312 * Replace the given page with a copy. The copied page assumes 2313 * the portion of the given page's "wire_count" that is not the 2314 * responsibility of this copy-on-write mechanism. 2315 * 2316 * The object containing the given page must have a non-zero 2317 * paging-in-progress count and be locked. 2318 */ 2319void 2320vm_page_cowfault(vm_page_t m) 2321{ 2322 vm_page_t mnew; 2323 vm_object_t object; 2324 vm_pindex_t pindex; 2325 2326 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 2327 vm_page_lock_assert(m, MA_OWNED); 2328 object = m->object; 2329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2330 KASSERT(object->paging_in_progress != 0, 2331 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2332 object)); 2333 pindex = m->pindex; 2334 2335 retry_alloc: 2336 pmap_remove_all(m); 2337 vm_page_remove(m); 2338 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2339 if (mnew == NULL) { 2340 vm_page_insert(m, object, pindex); 2341 vm_page_unlock(m); 2342 VM_OBJECT_UNLOCK(object); 2343 VM_WAIT; 2344 VM_OBJECT_LOCK(object); 2345 if (m == vm_page_lookup(object, pindex)) { 2346 vm_page_lock(m); 2347 goto retry_alloc; 2348 } else { 2349 /* 2350 * Page disappeared during the wait. 2351 */ 2352 return; 2353 } 2354 } 2355 2356 if (m->cow == 0) { 2357 /* 2358 * check to see if we raced with an xmit complete when 2359 * waiting to allocate a page. If so, put things back 2360 * the way they were 2361 */ 2362 vm_page_unlock(m); 2363 vm_page_lock(mnew); 2364 vm_page_free(mnew); 2365 vm_page_unlock(mnew); 2366 vm_page_insert(m, object, pindex); 2367 } else { /* clear COW & copy page */ 2368 if (!so_zerocp_fullpage) 2369 pmap_copy_page(m, mnew); 2370 mnew->valid = VM_PAGE_BITS_ALL; 2371 vm_page_dirty(mnew); 2372 mnew->wire_count = m->wire_count - m->cow; 2373 m->wire_count = m->cow; 2374 vm_page_unlock(m); 2375 } 2376} 2377 2378void 2379vm_page_cowclear(vm_page_t m) 2380{ 2381 2382 vm_page_lock_assert(m, MA_OWNED); 2383 if (m->cow) { 2384 m->cow--; 2385 /* 2386 * let vm_fault add back write permission lazily 2387 */ 2388 } 2389 /* 2390 * sf_buf_free() will free the page, so we needn't do it here 2391 */ 2392} 2393 2394int 2395vm_page_cowsetup(vm_page_t m) 2396{ 2397 2398 vm_page_lock_assert(m, MA_OWNED); 2399 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 || 2400 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 2401 return (EBUSY); 2402 m->cow++; 2403 pmap_remove_write(m); 2404 VM_OBJECT_UNLOCK(m->object); 2405 return (0); 2406} 2407 2408#include "opt_ddb.h" 2409#ifdef DDB 2410#include <sys/kernel.h> 2411 2412#include <ddb/ddb.h> 2413 2414DB_SHOW_COMMAND(page, vm_page_print_page_info) 2415{ 2416 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 2417 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 2418 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 2419 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 2420 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 2421 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 2422 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 2423 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 2424 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 2425 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 2426} 2427 2428DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2429{ 2430 2431 db_printf("PQ_FREE:"); 2432 db_printf(" %d", cnt.v_free_count); 2433 db_printf("\n"); 2434 2435 db_printf("PQ_CACHE:"); 2436 db_printf(" %d", cnt.v_cache_count); 2437 db_printf("\n"); 2438 2439 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2440 *vm_page_queues[PQ_ACTIVE].cnt, 2441 *vm_page_queues[PQ_INACTIVE].cnt); 2442} 2443#endif /* DDB */ 2444