vm_page.c revision 164100
1/*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 33 */ 34 35/*- 36 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 37 * All rights reserved. 38 * 39 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 40 * 41 * Permission to use, copy, modify and distribute this software and 42 * its documentation is hereby granted, provided that both the copyright 43 * notice and this permission notice appear in all copies of the 44 * software, derivative works or modified versions, and any portions 45 * thereof, and that both notices appear in supporting documentation. 46 * 47 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 48 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 49 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 50 * 51 * Carnegie Mellon requests users of this software to return to 52 * 53 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 54 * School of Computer Science 55 * Carnegie Mellon University 56 * Pittsburgh PA 15213-3890 57 * 58 * any improvements or extensions that they make and grant Carnegie the 59 * rights to redistribute these changes. 60 */ 61 62/* 63 * GENERAL RULES ON VM_PAGE MANIPULATION 64 * 65 * - a pageq mutex is required when adding or removing a page from a 66 * page queue (vm_page_queue[]), regardless of other mutexes or the 67 * busy state of a page. 68 * 69 * - a hash chain mutex is required when associating or disassociating 70 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 71 * regardless of other mutexes or the busy state of a page. 72 * 73 * - either a hash chain mutex OR a busied page is required in order 74 * to modify the page flags. A hash chain mutex must be obtained in 75 * order to busy a page. A page's flags cannot be modified by a 76 * hash chain mutex if the page is marked busy. 77 * 78 * - The object memq mutex is held when inserting or removing 79 * pages from an object (vm_page_insert() or vm_page_remove()). This 80 * is different from the object's main mutex. 81 * 82 * Generally speaking, you have to be aware of side effects when running 83 * vm_page ops. A vm_page_lookup() will return with the hash chain 84 * locked, whether it was able to lookup the page or not. vm_page_free(), 85 * vm_page_cache(), vm_page_activate(), and a number of other routines 86 * will release the hash chain mutex for you. Intermediate manipulation 87 * routines such as vm_page_flag_set() expect the hash chain to be held 88 * on entry and the hash chain will remain held on return. 89 * 90 * pageq scanning can only occur with the pageq in question locked. 91 * We have a known bottleneck with the active queue, but the cache 92 * and free queues are actually arrays already. 93 */ 94 95/* 96 * Resident memory management module. 97 */ 98 99#include <sys/cdefs.h> 100__FBSDID("$FreeBSD: head/sys/vm/vm_page.c 164100 2006-11-08 18:43:47Z alc $"); 101 102#include <sys/param.h> 103#include <sys/systm.h> 104#include <sys/lock.h> 105#include <sys/kernel.h> 106#include <sys/malloc.h> 107#include <sys/mutex.h> 108#include <sys/proc.h> 109#include <sys/sysctl.h> 110#include <sys/vmmeter.h> 111#include <sys/vnode.h> 112 113#include <vm/vm.h> 114#include <vm/vm_param.h> 115#include <vm/vm_kern.h> 116#include <vm/vm_object.h> 117#include <vm/vm_page.h> 118#include <vm/vm_pageout.h> 119#include <vm/vm_pager.h> 120#include <vm/vm_extern.h> 121#include <vm/uma.h> 122#include <vm/uma_int.h> 123 124#include <machine/md_var.h> 125 126/* 127 * Associated with page of user-allocatable memory is a 128 * page structure. 129 */ 130 131struct mtx vm_page_queue_mtx; 132struct mtx vm_page_queue_free_mtx; 133 134vm_page_t vm_page_array = 0; 135int vm_page_array_size = 0; 136long first_page = 0; 137int vm_page_zero_count = 0; 138 139static int boot_pages = UMA_BOOT_PAGES; 140TUNABLE_INT("vm.boot_pages", &boot_pages); 141SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 142 "number of pages allocated for bootstrapping the VM system"); 143 144/* 145 * vm_set_page_size: 146 * 147 * Sets the page size, perhaps based upon the memory 148 * size. Must be called before any use of page-size 149 * dependent functions. 150 */ 151void 152vm_set_page_size(void) 153{ 154 if (cnt.v_page_size == 0) 155 cnt.v_page_size = PAGE_SIZE; 156 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 157 panic("vm_set_page_size: page size not a power of two"); 158} 159 160/* 161 * vm_page_blacklist_lookup: 162 * 163 * See if a physical address in this page has been listed 164 * in the blacklist tunable. Entries in the tunable are 165 * separated by spaces or commas. If an invalid integer is 166 * encountered then the rest of the string is skipped. 167 */ 168static int 169vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 170{ 171 vm_paddr_t bad; 172 char *cp, *pos; 173 174 for (pos = list; *pos != '\0'; pos = cp) { 175 bad = strtoq(pos, &cp, 0); 176 if (*cp != '\0') { 177 if (*cp == ' ' || *cp == ',') { 178 cp++; 179 if (cp == pos) 180 continue; 181 } else 182 break; 183 } 184 if (pa == trunc_page(bad)) 185 return (1); 186 } 187 return (0); 188} 189 190/* 191 * vm_page_startup: 192 * 193 * Initializes the resident memory module. 194 * 195 * Allocates memory for the page cells, and 196 * for the object/offset-to-page hash table headers. 197 * Each page cell is initialized and placed on the free list. 198 */ 199vm_offset_t 200vm_page_startup(vm_offset_t vaddr) 201{ 202 vm_offset_t mapped; 203 vm_size_t npages; 204 vm_paddr_t page_range; 205 vm_paddr_t new_end; 206 int i; 207 vm_paddr_t pa; 208 int nblocks; 209 vm_paddr_t last_pa; 210 char *list; 211 212 /* the biggest memory array is the second group of pages */ 213 vm_paddr_t end; 214 vm_paddr_t biggestsize; 215 int biggestone; 216 217 vm_paddr_t total; 218 219 total = 0; 220 biggestsize = 0; 221 biggestone = 0; 222 nblocks = 0; 223 vaddr = round_page(vaddr); 224 225 for (i = 0; phys_avail[i + 1]; i += 2) { 226 phys_avail[i] = round_page(phys_avail[i]); 227 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 228 } 229 230 for (i = 0; phys_avail[i + 1]; i += 2) { 231 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 232 233 if (size > biggestsize) { 234 biggestone = i; 235 biggestsize = size; 236 } 237 ++nblocks; 238 total += size; 239 } 240 241 end = phys_avail[biggestone+1]; 242 243 /* 244 * Initialize the locks. 245 */ 246 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 247 MTX_RECURSE); 248 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 249 MTX_SPIN); 250 251 /* 252 * Initialize the queue headers for the free queue, the active queue 253 * and the inactive queue. 254 */ 255 vm_pageq_init(); 256 257 /* 258 * Allocate memory for use when boot strapping the kernel memory 259 * allocator. 260 */ 261 new_end = end - (boot_pages * UMA_SLAB_SIZE); 262 new_end = trunc_page(new_end); 263 mapped = pmap_map(&vaddr, new_end, end, 264 VM_PROT_READ | VM_PROT_WRITE); 265 bzero((void *)mapped, end - new_end); 266 uma_startup((void *)mapped, boot_pages); 267 268#if defined(__amd64__) || defined(__i386__) 269 /* 270 * Allocate a bitmap to indicate that a random physical page 271 * needs to be included in a minidump. 272 * 273 * The amd64 port needs this to indicate which direct map pages 274 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 275 * 276 * However, i386 still needs this workspace internally within the 277 * minidump code. In theory, they are not needed on i386, but are 278 * included should the sf_buf code decide to use them. 279 */ 280 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 281 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 282 new_end -= vm_page_dump_size; 283 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 284 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 285 bzero((void *)vm_page_dump, vm_page_dump_size); 286#endif 287 /* 288 * Compute the number of pages of memory that will be available for 289 * use (taking into account the overhead of a page structure per 290 * page). 291 */ 292 first_page = phys_avail[0] / PAGE_SIZE; 293 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 294 npages = (total - (page_range * sizeof(struct vm_page)) - 295 (end - new_end)) / PAGE_SIZE; 296 end = new_end; 297 298 /* 299 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 300 */ 301 vaddr += PAGE_SIZE; 302 303 /* 304 * Initialize the mem entry structures now, and put them in the free 305 * queue. 306 */ 307 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 308 mapped = pmap_map(&vaddr, new_end, end, 309 VM_PROT_READ | VM_PROT_WRITE); 310 vm_page_array = (vm_page_t) mapped; 311#ifdef __amd64__ 312 /* 313 * pmap_map on amd64 comes out of the direct-map, not kvm like i386, 314 * so the pages must be tracked for a crashdump to include this data. 315 * This includes the vm_page_array and the early UMA bootstrap pages. 316 */ 317 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 318 dump_add_page(pa); 319#endif 320 phys_avail[biggestone + 1] = new_end; 321 322 /* 323 * This assertion tests the hypothesis that npages and total are 324 * redundant. XXX 325 */ 326 page_range = 0; 327 for (i = 0; phys_avail[i + 1] != 0; i += 2) 328 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 329 KASSERT(page_range == npages, 330 ("vm_page_startup: inconsistent page counts")); 331 332 /* 333 * Clear all of the page structures 334 */ 335 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 336 vm_page_array_size = page_range; 337 338 /* 339 * Construct the free queue(s) in descending order (by physical 340 * address) so that the first 16MB of physical memory is allocated 341 * last rather than first. On large-memory machines, this avoids 342 * the exhaustion of low physical memory before isa_dma_init has run. 343 */ 344 cnt.v_page_count = 0; 345 cnt.v_free_count = 0; 346 list = getenv("vm.blacklist"); 347 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 348 pa = phys_avail[i]; 349 last_pa = phys_avail[i + 1]; 350 while (pa < last_pa) { 351 if (list != NULL && 352 vm_page_blacklist_lookup(list, pa)) 353 printf("Skipping page with pa 0x%jx\n", 354 (uintmax_t)pa); 355 else 356 vm_pageq_add_new_page(pa); 357 pa += PAGE_SIZE; 358 } 359 } 360 freeenv(list); 361 return (vaddr); 362} 363 364void 365vm_page_flag_set(vm_page_t m, unsigned short bits) 366{ 367 368 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 369 m->flags |= bits; 370} 371 372void 373vm_page_flag_clear(vm_page_t m, unsigned short bits) 374{ 375 376 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 377 m->flags &= ~bits; 378} 379 380void 381vm_page_busy(vm_page_t m) 382{ 383 384 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 385 KASSERT((m->oflags & VPO_BUSY) == 0, 386 ("vm_page_busy: page already busy!!!")); 387 m->oflags |= VPO_BUSY; 388} 389 390/* 391 * vm_page_flash: 392 * 393 * wakeup anyone waiting for the page. 394 */ 395void 396vm_page_flash(vm_page_t m) 397{ 398 399 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 400 if (m->oflags & VPO_WANTED) { 401 m->oflags &= ~VPO_WANTED; 402 wakeup(m); 403 } 404} 405 406/* 407 * vm_page_wakeup: 408 * 409 * clear the VPO_BUSY flag and wakeup anyone waiting for the 410 * page. 411 * 412 */ 413void 414vm_page_wakeup(vm_page_t m) 415{ 416 417 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 418 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 419 m->oflags &= ~VPO_BUSY; 420 vm_page_flash(m); 421} 422 423void 424vm_page_io_start(vm_page_t m) 425{ 426 427 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 428 m->busy++; 429} 430 431void 432vm_page_io_finish(vm_page_t m) 433{ 434 435 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 436 m->busy--; 437 if (m->busy == 0) 438 vm_page_flash(m); 439} 440 441/* 442 * Keep page from being freed by the page daemon 443 * much of the same effect as wiring, except much lower 444 * overhead and should be used only for *very* temporary 445 * holding ("wiring"). 446 */ 447void 448vm_page_hold(vm_page_t mem) 449{ 450 451 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 452 mem->hold_count++; 453} 454 455void 456vm_page_unhold(vm_page_t mem) 457{ 458 459 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 460 --mem->hold_count; 461 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 462 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD)) 463 vm_page_free_toq(mem); 464} 465 466/* 467 * vm_page_free: 468 * 469 * Free a page 470 * 471 * The clearing of PG_ZERO is a temporary safety until the code can be 472 * reviewed to determine that PG_ZERO is being properly cleared on 473 * write faults or maps. PG_ZERO was previously cleared in 474 * vm_page_alloc(). 475 */ 476void 477vm_page_free(vm_page_t m) 478{ 479 vm_page_flag_clear(m, PG_ZERO); 480 vm_page_free_toq(m); 481 vm_page_zero_idle_wakeup(); 482} 483 484/* 485 * vm_page_free_zero: 486 * 487 * Free a page to the zerod-pages queue 488 */ 489void 490vm_page_free_zero(vm_page_t m) 491{ 492 vm_page_flag_set(m, PG_ZERO); 493 vm_page_free_toq(m); 494} 495 496/* 497 * vm_page_sleep: 498 * 499 * Sleep and release the page queues lock. 500 * 501 * The object containing the given page must be locked. 502 */ 503void 504vm_page_sleep(vm_page_t m, const char *msg) 505{ 506 507 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 508 if (!mtx_owned(&vm_page_queue_mtx)) 509 vm_page_lock_queues(); 510 vm_page_flag_set(m, PG_REFERENCED); 511 vm_page_unlock_queues(); 512 513 /* 514 * It's possible that while we sleep, the page will get 515 * unbusied and freed. If we are holding the object 516 * lock, we will assume we hold a reference to the object 517 * such that even if m->object changes, we can re-lock 518 * it. 519 */ 520 m->oflags |= VPO_WANTED; 521 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 522} 523 524/* 525 * vm_page_dirty: 526 * 527 * make page all dirty 528 */ 529void 530vm_page_dirty(vm_page_t m) 531{ 532 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_CACHE, 533 ("vm_page_dirty: page in cache!")); 534 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_FREE, 535 ("vm_page_dirty: page is free!")); 536 m->dirty = VM_PAGE_BITS_ALL; 537} 538 539/* 540 * vm_page_splay: 541 * 542 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 543 * the vm_page containing the given pindex. If, however, that 544 * pindex is not found in the vm_object, returns a vm_page that is 545 * adjacent to the pindex, coming before or after it. 546 */ 547vm_page_t 548vm_page_splay(vm_pindex_t pindex, vm_page_t root) 549{ 550 struct vm_page dummy; 551 vm_page_t lefttreemax, righttreemin, y; 552 553 if (root == NULL) 554 return (root); 555 lefttreemax = righttreemin = &dummy; 556 for (;; root = y) { 557 if (pindex < root->pindex) { 558 if ((y = root->left) == NULL) 559 break; 560 if (pindex < y->pindex) { 561 /* Rotate right. */ 562 root->left = y->right; 563 y->right = root; 564 root = y; 565 if ((y = root->left) == NULL) 566 break; 567 } 568 /* Link into the new root's right tree. */ 569 righttreemin->left = root; 570 righttreemin = root; 571 } else if (pindex > root->pindex) { 572 if ((y = root->right) == NULL) 573 break; 574 if (pindex > y->pindex) { 575 /* Rotate left. */ 576 root->right = y->left; 577 y->left = root; 578 root = y; 579 if ((y = root->right) == NULL) 580 break; 581 } 582 /* Link into the new root's left tree. */ 583 lefttreemax->right = root; 584 lefttreemax = root; 585 } else 586 break; 587 } 588 /* Assemble the new root. */ 589 lefttreemax->right = root->left; 590 righttreemin->left = root->right; 591 root->left = dummy.right; 592 root->right = dummy.left; 593 return (root); 594} 595 596/* 597 * vm_page_insert: [ internal use only ] 598 * 599 * Inserts the given mem entry into the object and object list. 600 * 601 * The pagetables are not updated but will presumably fault the page 602 * in if necessary, or if a kernel page the caller will at some point 603 * enter the page into the kernel's pmap. We are not allowed to block 604 * here so we *can't* do this anyway. 605 * 606 * The object and page must be locked. 607 * This routine may not block. 608 */ 609void 610vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 611{ 612 vm_page_t root; 613 614 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 615 if (m->object != NULL) 616 panic("vm_page_insert: page already inserted"); 617 618 /* 619 * Record the object/offset pair in this page 620 */ 621 m->object = object; 622 m->pindex = pindex; 623 624 /* 625 * Now link into the object's ordered list of backed pages. 626 */ 627 root = object->root; 628 if (root == NULL) { 629 m->left = NULL; 630 m->right = NULL; 631 TAILQ_INSERT_TAIL(&object->memq, m, listq); 632 } else { 633 root = vm_page_splay(pindex, root); 634 if (pindex < root->pindex) { 635 m->left = root->left; 636 m->right = root; 637 root->left = NULL; 638 TAILQ_INSERT_BEFORE(root, m, listq); 639 } else if (pindex == root->pindex) 640 panic("vm_page_insert: offset already allocated"); 641 else { 642 m->right = root->right; 643 m->left = root; 644 root->right = NULL; 645 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 646 } 647 } 648 object->root = m; 649 object->generation++; 650 651 /* 652 * show that the object has one more resident page. 653 */ 654 object->resident_page_count++; 655 /* 656 * Hold the vnode until the last page is released. 657 */ 658 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 659 vhold((struct vnode *)object->handle); 660 661 /* 662 * Since we are inserting a new and possibly dirty page, 663 * update the object's OBJ_MIGHTBEDIRTY flag. 664 */ 665 if (m->flags & PG_WRITEABLE) 666 vm_object_set_writeable_dirty(object); 667} 668 669/* 670 * vm_page_remove: 671 * NOTE: used by device pager as well -wfj 672 * 673 * Removes the given mem entry from the object/offset-page 674 * table and the object page list, but do not invalidate/terminate 675 * the backing store. 676 * 677 * The object and page must be locked. 678 * The underlying pmap entry (if any) is NOT removed here. 679 * This routine may not block. 680 */ 681void 682vm_page_remove(vm_page_t m) 683{ 684 vm_object_t object; 685 vm_page_t root; 686 687 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 688 if ((object = m->object) == NULL) 689 return; 690 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 691 if (m->oflags & VPO_BUSY) { 692 m->oflags &= ~VPO_BUSY; 693 vm_page_flash(m); 694 } 695 696 /* 697 * Now remove from the object's list of backed pages. 698 */ 699 if (m != object->root) 700 vm_page_splay(m->pindex, object->root); 701 if (m->left == NULL) 702 root = m->right; 703 else { 704 root = vm_page_splay(m->pindex, m->left); 705 root->right = m->right; 706 } 707 object->root = root; 708 TAILQ_REMOVE(&object->memq, m, listq); 709 710 /* 711 * And show that the object has one fewer resident page. 712 */ 713 object->resident_page_count--; 714 object->generation++; 715 /* 716 * The vnode may now be recycled. 717 */ 718 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 719 vdrop((struct vnode *)object->handle); 720 721 m->object = NULL; 722} 723 724/* 725 * vm_page_lookup: 726 * 727 * Returns the page associated with the object/offset 728 * pair specified; if none is found, NULL is returned. 729 * 730 * The object must be locked. 731 * This routine may not block. 732 * This is a critical path routine 733 */ 734vm_page_t 735vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 736{ 737 vm_page_t m; 738 739 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 740 if ((m = object->root) != NULL && m->pindex != pindex) { 741 m = vm_page_splay(pindex, m); 742 if ((object->root = m)->pindex != pindex) 743 m = NULL; 744 } 745 return (m); 746} 747 748/* 749 * vm_page_rename: 750 * 751 * Move the given memory entry from its 752 * current object to the specified target object/offset. 753 * 754 * The object must be locked. 755 * This routine may not block. 756 * 757 * Note: swap associated with the page must be invalidated by the move. We 758 * have to do this for several reasons: (1) we aren't freeing the 759 * page, (2) we are dirtying the page, (3) the VM system is probably 760 * moving the page from object A to B, and will then later move 761 * the backing store from A to B and we can't have a conflict. 762 * 763 * Note: we *always* dirty the page. It is necessary both for the 764 * fact that we moved it, and because we may be invalidating 765 * swap. If the page is on the cache, we have to deactivate it 766 * or vm_page_dirty() will panic. Dirty pages are not allowed 767 * on the cache. 768 */ 769void 770vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 771{ 772 773 vm_page_remove(m); 774 vm_page_insert(m, new_object, new_pindex); 775 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 776 vm_page_deactivate(m); 777 vm_page_dirty(m); 778} 779 780/* 781 * vm_page_select_cache: 782 * 783 * Move a page of the given color from the cache queue to the free 784 * queue. As pages might be found, but are not applicable, they are 785 * deactivated. 786 * 787 * This routine may not block. 788 */ 789vm_page_t 790vm_page_select_cache(int color) 791{ 792 vm_object_t object; 793 vm_page_t m; 794 boolean_t was_trylocked; 795 796 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 797 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 798 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 799 KASSERT(!pmap_page_is_mapped(m), 800 ("Found mapped cache page %p", m)); 801 KASSERT((m->flags & PG_UNMANAGED) == 0, 802 ("Found unmanaged cache page %p", m)); 803 KASSERT(m->wire_count == 0, ("Found wired cache page %p", m)); 804 if (m->hold_count == 0 && (object = m->object, 805 (was_trylocked = VM_OBJECT_TRYLOCK(object)) || 806 VM_OBJECT_LOCKED(object))) { 807 KASSERT((m->oflags & VPO_BUSY) == 0 && m->busy == 0, 808 ("Found busy cache page %p", m)); 809 vm_page_free(m); 810 if (was_trylocked) 811 VM_OBJECT_UNLOCK(object); 812 break; 813 } 814 vm_page_deactivate(m); 815 } 816 return (m); 817} 818 819/* 820 * vm_page_alloc: 821 * 822 * Allocate and return a memory cell associated 823 * with this VM object/offset pair. 824 * 825 * page_req classes: 826 * VM_ALLOC_NORMAL normal process request 827 * VM_ALLOC_SYSTEM system *really* needs a page 828 * VM_ALLOC_INTERRUPT interrupt time request 829 * VM_ALLOC_ZERO zero page 830 * 831 * This routine may not block. 832 * 833 * Additional special handling is required when called from an 834 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 835 * the page cache in this case. 836 */ 837vm_page_t 838vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 839{ 840 vm_page_t m = NULL; 841 int color, flags, page_req; 842 843 page_req = req & VM_ALLOC_CLASS_MASK; 844 KASSERT(curthread->td_intr_nesting_level == 0 || 845 page_req == VM_ALLOC_INTERRUPT, 846 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 847 848 if ((req & VM_ALLOC_NOOBJ) == 0) { 849 KASSERT(object != NULL, 850 ("vm_page_alloc: NULL object.")); 851 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 852 color = (pindex + object->pg_color) & PQ_COLORMASK; 853 } else 854 color = pindex & PQ_COLORMASK; 855 856 /* 857 * The pager is allowed to eat deeper into the free page list. 858 */ 859 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 860 page_req = VM_ALLOC_SYSTEM; 861 }; 862 863loop: 864 mtx_lock_spin(&vm_page_queue_free_mtx); 865 if (cnt.v_free_count > cnt.v_free_reserved || 866 (page_req == VM_ALLOC_SYSTEM && 867 cnt.v_cache_count == 0 && 868 cnt.v_free_count > cnt.v_interrupt_free_min) || 869 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 870 /* 871 * Allocate from the free queue if the number of free pages 872 * exceeds the minimum for the request class. 873 */ 874 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 875 } else if (page_req != VM_ALLOC_INTERRUPT) { 876 mtx_unlock_spin(&vm_page_queue_free_mtx); 877 /* 878 * Allocatable from cache (non-interrupt only). On success, 879 * we must free the page and try again, thus ensuring that 880 * cnt.v_*_free_min counters are replenished. 881 */ 882 vm_page_lock_queues(); 883 if ((m = vm_page_select_cache(color)) == NULL) { 884 KASSERT(cnt.v_cache_count == 0, 885 ("vm_page_alloc: cache queue is missing %d pages", 886 cnt.v_cache_count)); 887 vm_page_unlock_queues(); 888 atomic_add_int(&vm_pageout_deficit, 1); 889 pagedaemon_wakeup(); 890 891 if (page_req != VM_ALLOC_SYSTEM) 892 return (NULL); 893 894 mtx_lock_spin(&vm_page_queue_free_mtx); 895 if (cnt.v_free_count <= cnt.v_interrupt_free_min) { 896 mtx_unlock_spin(&vm_page_queue_free_mtx); 897 return (NULL); 898 } 899 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 900 } else { 901 vm_page_unlock_queues(); 902 goto loop; 903 } 904 } else { 905 /* 906 * Not allocatable from cache from interrupt, give up. 907 */ 908 mtx_unlock_spin(&vm_page_queue_free_mtx); 909 atomic_add_int(&vm_pageout_deficit, 1); 910 pagedaemon_wakeup(); 911 return (NULL); 912 } 913 914 /* 915 * At this point we had better have found a good page. 916 */ 917 918 KASSERT( 919 m != NULL, 920 ("vm_page_alloc(): missing page on free queue") 921 ); 922 923 /* 924 * Remove from free queue 925 */ 926 vm_pageq_remove_nowakeup(m); 927 928 /* 929 * Initialize structure. Only the PG_ZERO flag is inherited. 930 */ 931 flags = 0; 932 if (m->flags & PG_ZERO) { 933 vm_page_zero_count--; 934 if (req & VM_ALLOC_ZERO) 935 flags = PG_ZERO; 936 } 937 m->flags = flags; 938 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 939 m->oflags = 0; 940 else 941 m->oflags = VPO_BUSY; 942 if (req & VM_ALLOC_WIRED) { 943 atomic_add_int(&cnt.v_wire_count, 1); 944 m->wire_count = 1; 945 } else 946 m->wire_count = 0; 947 m->hold_count = 0; 948 m->act_count = 0; 949 m->busy = 0; 950 m->valid = 0; 951 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 952 mtx_unlock_spin(&vm_page_queue_free_mtx); 953 954 if ((req & VM_ALLOC_NOOBJ) == 0) 955 vm_page_insert(m, object, pindex); 956 else 957 m->pindex = pindex; 958 959 /* 960 * Don't wakeup too often - wakeup the pageout daemon when 961 * we would be nearly out of memory. 962 */ 963 if (vm_paging_needed()) 964 pagedaemon_wakeup(); 965 966 return (m); 967} 968 969/* 970 * vm_wait: (also see VM_WAIT macro) 971 * 972 * Block until free pages are available for allocation 973 * - Called in various places before memory allocations. 974 */ 975void 976vm_wait(void) 977{ 978 979 vm_page_lock_queues(); 980 if (curproc == pageproc) { 981 vm_pageout_pages_needed = 1; 982 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 983 PDROP | PSWP, "VMWait", 0); 984 } else { 985 if (!vm_pages_needed) { 986 vm_pages_needed = 1; 987 wakeup(&vm_pages_needed); 988 } 989 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 990 "vmwait", 0); 991 } 992} 993 994/* 995 * vm_waitpfault: (also see VM_WAITPFAULT macro) 996 * 997 * Block until free pages are available for allocation 998 * - Called only in vm_fault so that processes page faulting 999 * can be easily tracked. 1000 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1001 * processes will be able to grab memory first. Do not change 1002 * this balance without careful testing first. 1003 */ 1004void 1005vm_waitpfault(void) 1006{ 1007 1008 vm_page_lock_queues(); 1009 if (!vm_pages_needed) { 1010 vm_pages_needed = 1; 1011 wakeup(&vm_pages_needed); 1012 } 1013 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 1014 "pfault", 0); 1015} 1016 1017/* 1018 * vm_page_activate: 1019 * 1020 * Put the specified page on the active list (if appropriate). 1021 * Ensure that act_count is at least ACT_INIT but do not otherwise 1022 * mess with it. 1023 * 1024 * The page queues must be locked. 1025 * This routine may not block. 1026 */ 1027void 1028vm_page_activate(vm_page_t m) 1029{ 1030 1031 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1032 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) { 1033 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1034 cnt.v_reactivated++; 1035 vm_pageq_remove(m); 1036 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1037 if (m->act_count < ACT_INIT) 1038 m->act_count = ACT_INIT; 1039 vm_pageq_enqueue(PQ_ACTIVE, m); 1040 } 1041 } else { 1042 if (m->act_count < ACT_INIT) 1043 m->act_count = ACT_INIT; 1044 } 1045} 1046 1047/* 1048 * vm_page_free_wakeup: 1049 * 1050 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1051 * routine is called when a page has been added to the cache or free 1052 * queues. 1053 * 1054 * The page queues must be locked. 1055 * This routine may not block. 1056 */ 1057static inline void 1058vm_page_free_wakeup(void) 1059{ 1060 1061 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1062 /* 1063 * if pageout daemon needs pages, then tell it that there are 1064 * some free. 1065 */ 1066 if (vm_pageout_pages_needed && 1067 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1068 wakeup(&vm_pageout_pages_needed); 1069 vm_pageout_pages_needed = 0; 1070 } 1071 /* 1072 * wakeup processes that are waiting on memory if we hit a 1073 * high water mark. And wakeup scheduler process if we have 1074 * lots of memory. this process will swapin processes. 1075 */ 1076 if (vm_pages_needed && !vm_page_count_min()) { 1077 vm_pages_needed = 0; 1078 wakeup(&cnt.v_free_count); 1079 } 1080} 1081 1082/* 1083 * vm_page_free_toq: 1084 * 1085 * Returns the given page to the PQ_FREE list, 1086 * disassociating it with any VM object. 1087 * 1088 * Object and page must be locked prior to entry. 1089 * This routine may not block. 1090 */ 1091 1092void 1093vm_page_free_toq(vm_page_t m) 1094{ 1095 struct vpgqueues *pq; 1096 1097 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1098 KASSERT(!pmap_page_is_mapped(m), 1099 ("vm_page_free_toq: freeing mapped page %p", m)); 1100 cnt.v_tfree++; 1101 1102 if (m->busy || VM_PAGE_INQUEUE1(m, PQ_FREE)) { 1103 printf( 1104 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1105 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1106 m->hold_count); 1107 if (VM_PAGE_INQUEUE1(m, PQ_FREE)) 1108 panic("vm_page_free: freeing free page"); 1109 else 1110 panic("vm_page_free: freeing busy page"); 1111 } 1112 1113 /* 1114 * unqueue, then remove page. Note that we cannot destroy 1115 * the page here because we do not want to call the pager's 1116 * callback routine until after we've put the page on the 1117 * appropriate free queue. 1118 */ 1119 vm_pageq_remove_nowakeup(m); 1120 vm_page_remove(m); 1121 1122 /* 1123 * If fictitious remove object association and 1124 * return, otherwise delay object association removal. 1125 */ 1126 if ((m->flags & PG_FICTITIOUS) != 0) { 1127 return; 1128 } 1129 1130 m->valid = 0; 1131 vm_page_undirty(m); 1132 1133 if (m->wire_count != 0) { 1134 if (m->wire_count > 1) { 1135 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1136 m->wire_count, (long)m->pindex); 1137 } 1138 panic("vm_page_free: freeing wired page"); 1139 } 1140 if (m->hold_count != 0) { 1141 m->flags &= ~PG_ZERO; 1142 VM_PAGE_SETQUEUE2(m, PQ_HOLD); 1143 } else 1144 VM_PAGE_SETQUEUE1(m, PQ_FREE); 1145 pq = &vm_page_queues[VM_PAGE_GETQUEUE(m)]; 1146 mtx_lock_spin(&vm_page_queue_free_mtx); 1147 pq->lcnt++; 1148 ++(*pq->cnt); 1149 1150 /* 1151 * Put zero'd pages on the end ( where we look for zero'd pages 1152 * first ) and non-zerod pages at the head. 1153 */ 1154 if (m->flags & PG_ZERO) { 1155 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1156 ++vm_page_zero_count; 1157 } else { 1158 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1159 } 1160 mtx_unlock_spin(&vm_page_queue_free_mtx); 1161 vm_page_free_wakeup(); 1162} 1163 1164/* 1165 * vm_page_unmanage: 1166 * 1167 * Prevent PV management from being done on the page. The page is 1168 * removed from the paging queues as if it were wired, and as a 1169 * consequence of no longer being managed the pageout daemon will not 1170 * touch it (since there is no way to locate the pte mappings for the 1171 * page). madvise() calls that mess with the pmap will also no longer 1172 * operate on the page. 1173 * 1174 * Beyond that the page is still reasonably 'normal'. Freeing the page 1175 * will clear the flag. 1176 * 1177 * This routine is used by OBJT_PHYS objects - objects using unswappable 1178 * physical memory as backing store rather then swap-backed memory and 1179 * will eventually be extended to support 4MB unmanaged physical 1180 * mappings. 1181 */ 1182void 1183vm_page_unmanage(vm_page_t m) 1184{ 1185 1186 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1187 if ((m->flags & PG_UNMANAGED) == 0) { 1188 if (m->wire_count == 0) 1189 vm_pageq_remove(m); 1190 } 1191 vm_page_flag_set(m, PG_UNMANAGED); 1192} 1193 1194/* 1195 * vm_page_wire: 1196 * 1197 * Mark this page as wired down by yet 1198 * another map, removing it from paging queues 1199 * as necessary. 1200 * 1201 * The page queues must be locked. 1202 * This routine may not block. 1203 */ 1204void 1205vm_page_wire(vm_page_t m) 1206{ 1207 1208 /* 1209 * Only bump the wire statistics if the page is not already wired, 1210 * and only unqueue the page if it is on some queue (if it is unmanaged 1211 * it is already off the queues). 1212 */ 1213 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1214 if (m->flags & PG_FICTITIOUS) 1215 return; 1216 if (m->wire_count == 0) { 1217 if ((m->flags & PG_UNMANAGED) == 0) 1218 vm_pageq_remove(m); 1219 atomic_add_int(&cnt.v_wire_count, 1); 1220 } 1221 m->wire_count++; 1222 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1223} 1224 1225/* 1226 * vm_page_unwire: 1227 * 1228 * Release one wiring of this page, potentially 1229 * enabling it to be paged again. 1230 * 1231 * Many pages placed on the inactive queue should actually go 1232 * into the cache, but it is difficult to figure out which. What 1233 * we do instead, if the inactive target is well met, is to put 1234 * clean pages at the head of the inactive queue instead of the tail. 1235 * This will cause them to be moved to the cache more quickly and 1236 * if not actively re-referenced, freed more quickly. If we just 1237 * stick these pages at the end of the inactive queue, heavy filesystem 1238 * meta-data accesses can cause an unnecessary paging load on memory bound 1239 * processes. This optimization causes one-time-use metadata to be 1240 * reused more quickly. 1241 * 1242 * BUT, if we are in a low-memory situation we have no choice but to 1243 * put clean pages on the cache queue. 1244 * 1245 * A number of routines use vm_page_unwire() to guarantee that the page 1246 * will go into either the inactive or active queues, and will NEVER 1247 * be placed in the cache - for example, just after dirtying a page. 1248 * dirty pages in the cache are not allowed. 1249 * 1250 * The page queues must be locked. 1251 * This routine may not block. 1252 */ 1253void 1254vm_page_unwire(vm_page_t m, int activate) 1255{ 1256 1257 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1258 if (m->flags & PG_FICTITIOUS) 1259 return; 1260 if (m->wire_count > 0) { 1261 m->wire_count--; 1262 if (m->wire_count == 0) { 1263 atomic_subtract_int(&cnt.v_wire_count, 1); 1264 if (m->flags & PG_UNMANAGED) { 1265 ; 1266 } else if (activate) 1267 vm_pageq_enqueue(PQ_ACTIVE, m); 1268 else { 1269 vm_page_flag_clear(m, PG_WINATCFLS); 1270 vm_pageq_enqueue(PQ_INACTIVE, m); 1271 } 1272 } 1273 } else { 1274 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1275 } 1276} 1277 1278 1279/* 1280 * Move the specified page to the inactive queue. If the page has 1281 * any associated swap, the swap is deallocated. 1282 * 1283 * Normally athead is 0 resulting in LRU operation. athead is set 1284 * to 1 if we want this page to be 'as if it were placed in the cache', 1285 * except without unmapping it from the process address space. 1286 * 1287 * This routine may not block. 1288 */ 1289static inline void 1290_vm_page_deactivate(vm_page_t m, int athead) 1291{ 1292 1293 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1294 1295 /* 1296 * Ignore if already inactive. 1297 */ 1298 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) 1299 return; 1300 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1301 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1302 cnt.v_reactivated++; 1303 vm_page_flag_clear(m, PG_WINATCFLS); 1304 vm_pageq_remove(m); 1305 if (athead) 1306 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1307 else 1308 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1309 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1310 vm_page_queues[PQ_INACTIVE].lcnt++; 1311 cnt.v_inactive_count++; 1312 } 1313} 1314 1315void 1316vm_page_deactivate(vm_page_t m) 1317{ 1318 _vm_page_deactivate(m, 0); 1319} 1320 1321/* 1322 * vm_page_try_to_cache: 1323 * 1324 * Returns 0 on failure, 1 on success 1325 */ 1326int 1327vm_page_try_to_cache(vm_page_t m) 1328{ 1329 1330 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1331 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1332 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1333 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1334 return (0); 1335 } 1336 pmap_remove_all(m); 1337 if (m->dirty) 1338 return (0); 1339 vm_page_cache(m); 1340 return (1); 1341} 1342 1343/* 1344 * vm_page_try_to_free() 1345 * 1346 * Attempt to free the page. If we cannot free it, we do nothing. 1347 * 1 is returned on success, 0 on failure. 1348 */ 1349int 1350vm_page_try_to_free(vm_page_t m) 1351{ 1352 1353 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1354 if (m->object != NULL) 1355 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1356 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1357 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1358 return (0); 1359 } 1360 pmap_remove_all(m); 1361 if (m->dirty) 1362 return (0); 1363 vm_page_free(m); 1364 return (1); 1365} 1366 1367/* 1368 * vm_page_cache 1369 * 1370 * Put the specified page onto the page cache queue (if appropriate). 1371 * 1372 * This routine may not block. 1373 */ 1374void 1375vm_page_cache(vm_page_t m) 1376{ 1377 1378 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1379 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1380 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1381 m->hold_count || m->wire_count) { 1382 printf("vm_page_cache: attempting to cache busy page\n"); 1383 return; 1384 } 1385 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1386 return; 1387 1388 /* 1389 * Remove all pmaps and indicate that the page is not 1390 * writeable or mapped. 1391 */ 1392 pmap_remove_all(m); 1393 if (m->dirty != 0) { 1394 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1395 (long)m->pindex); 1396 } 1397 vm_pageq_remove_nowakeup(m); 1398 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1399 vm_page_free_wakeup(); 1400} 1401 1402/* 1403 * vm_page_dontneed 1404 * 1405 * Cache, deactivate, or do nothing as appropriate. This routine 1406 * is typically used by madvise() MADV_DONTNEED. 1407 * 1408 * Generally speaking we want to move the page into the cache so 1409 * it gets reused quickly. However, this can result in a silly syndrome 1410 * due to the page recycling too quickly. Small objects will not be 1411 * fully cached. On the otherhand, if we move the page to the inactive 1412 * queue we wind up with a problem whereby very large objects 1413 * unnecessarily blow away our inactive and cache queues. 1414 * 1415 * The solution is to move the pages based on a fixed weighting. We 1416 * either leave them alone, deactivate them, or move them to the cache, 1417 * where moving them to the cache has the highest weighting. 1418 * By forcing some pages into other queues we eventually force the 1419 * system to balance the queues, potentially recovering other unrelated 1420 * space from active. The idea is to not force this to happen too 1421 * often. 1422 */ 1423void 1424vm_page_dontneed(vm_page_t m) 1425{ 1426 static int dnweight; 1427 int dnw; 1428 int head; 1429 1430 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1431 dnw = ++dnweight; 1432 1433 /* 1434 * occassionally leave the page alone 1435 */ 1436 if ((dnw & 0x01F0) == 0 || 1437 VM_PAGE_INQUEUE2(m, PQ_INACTIVE) || 1438 VM_PAGE_INQUEUE1(m, PQ_CACHE) 1439 ) { 1440 if (m->act_count >= ACT_INIT) 1441 --m->act_count; 1442 return; 1443 } 1444 1445 if (m->dirty == 0 && pmap_is_modified(m)) 1446 vm_page_dirty(m); 1447 1448 if (m->dirty || (dnw & 0x0070) == 0) { 1449 /* 1450 * Deactivate the page 3 times out of 32. 1451 */ 1452 head = 0; 1453 } else { 1454 /* 1455 * Cache the page 28 times out of every 32. Note that 1456 * the page is deactivated instead of cached, but placed 1457 * at the head of the queue instead of the tail. 1458 */ 1459 head = 1; 1460 } 1461 _vm_page_deactivate(m, head); 1462} 1463 1464/* 1465 * Grab a page, waiting until we are waken up due to the page 1466 * changing state. We keep on waiting, if the page continues 1467 * to be in the object. If the page doesn't exist, first allocate it 1468 * and then conditionally zero it. 1469 * 1470 * This routine may block. 1471 */ 1472vm_page_t 1473vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1474{ 1475 vm_page_t m; 1476 1477 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1478retrylookup: 1479 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1480 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) { 1481 if ((allocflags & VM_ALLOC_RETRY) == 0) 1482 return (NULL); 1483 goto retrylookup; 1484 } else { 1485 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1486 vm_page_lock_queues(); 1487 vm_page_wire(m); 1488 vm_page_unlock_queues(); 1489 } 1490 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1491 vm_page_busy(m); 1492 return (m); 1493 } 1494 } 1495 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1496 if (m == NULL) { 1497 VM_OBJECT_UNLOCK(object); 1498 VM_WAIT; 1499 VM_OBJECT_LOCK(object); 1500 if ((allocflags & VM_ALLOC_RETRY) == 0) 1501 return (NULL); 1502 goto retrylookup; 1503 } 1504 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1505 pmap_zero_page(m); 1506 return (m); 1507} 1508 1509/* 1510 * Mapping function for valid bits or for dirty bits in 1511 * a page. May not block. 1512 * 1513 * Inputs are required to range within a page. 1514 */ 1515inline int 1516vm_page_bits(int base, int size) 1517{ 1518 int first_bit; 1519 int last_bit; 1520 1521 KASSERT( 1522 base + size <= PAGE_SIZE, 1523 ("vm_page_bits: illegal base/size %d/%d", base, size) 1524 ); 1525 1526 if (size == 0) /* handle degenerate case */ 1527 return (0); 1528 1529 first_bit = base >> DEV_BSHIFT; 1530 last_bit = (base + size - 1) >> DEV_BSHIFT; 1531 1532 return ((2 << last_bit) - (1 << first_bit)); 1533} 1534 1535/* 1536 * vm_page_set_validclean: 1537 * 1538 * Sets portions of a page valid and clean. The arguments are expected 1539 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1540 * of any partial chunks touched by the range. The invalid portion of 1541 * such chunks will be zero'd. 1542 * 1543 * This routine may not block. 1544 * 1545 * (base + size) must be less then or equal to PAGE_SIZE. 1546 */ 1547void 1548vm_page_set_validclean(vm_page_t m, int base, int size) 1549{ 1550 int pagebits; 1551 int frag; 1552 int endoff; 1553 1554 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1555 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1556 if (size == 0) /* handle degenerate case */ 1557 return; 1558 1559 /* 1560 * If the base is not DEV_BSIZE aligned and the valid 1561 * bit is clear, we have to zero out a portion of the 1562 * first block. 1563 */ 1564 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1565 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1566 pmap_zero_page_area(m, frag, base - frag); 1567 1568 /* 1569 * If the ending offset is not DEV_BSIZE aligned and the 1570 * valid bit is clear, we have to zero out a portion of 1571 * the last block. 1572 */ 1573 endoff = base + size; 1574 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1575 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1576 pmap_zero_page_area(m, endoff, 1577 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1578 1579 /* 1580 * Set valid, clear dirty bits. If validating the entire 1581 * page we can safely clear the pmap modify bit. We also 1582 * use this opportunity to clear the VPO_NOSYNC flag. If a process 1583 * takes a write fault on a MAP_NOSYNC memory area the flag will 1584 * be set again. 1585 * 1586 * We set valid bits inclusive of any overlap, but we can only 1587 * clear dirty bits for DEV_BSIZE chunks that are fully within 1588 * the range. 1589 */ 1590 pagebits = vm_page_bits(base, size); 1591 m->valid |= pagebits; 1592#if 0 /* NOT YET */ 1593 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1594 frag = DEV_BSIZE - frag; 1595 base += frag; 1596 size -= frag; 1597 if (size < 0) 1598 size = 0; 1599 } 1600 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1601#endif 1602 m->dirty &= ~pagebits; 1603 if (base == 0 && size == PAGE_SIZE) { 1604 pmap_clear_modify(m); 1605 m->oflags &= ~VPO_NOSYNC; 1606 } 1607} 1608 1609void 1610vm_page_clear_dirty(vm_page_t m, int base, int size) 1611{ 1612 1613 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1614 m->dirty &= ~vm_page_bits(base, size); 1615} 1616 1617/* 1618 * vm_page_set_invalid: 1619 * 1620 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1621 * valid and dirty bits for the effected areas are cleared. 1622 * 1623 * May not block. 1624 */ 1625void 1626vm_page_set_invalid(vm_page_t m, int base, int size) 1627{ 1628 int bits; 1629 1630 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1631 bits = vm_page_bits(base, size); 1632 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1633 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 1634 pmap_remove_all(m); 1635 m->valid &= ~bits; 1636 m->dirty &= ~bits; 1637 m->object->generation++; 1638} 1639 1640/* 1641 * vm_page_zero_invalid() 1642 * 1643 * The kernel assumes that the invalid portions of a page contain 1644 * garbage, but such pages can be mapped into memory by user code. 1645 * When this occurs, we must zero out the non-valid portions of the 1646 * page so user code sees what it expects. 1647 * 1648 * Pages are most often semi-valid when the end of a file is mapped 1649 * into memory and the file's size is not page aligned. 1650 */ 1651void 1652vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1653{ 1654 int b; 1655 int i; 1656 1657 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1658 /* 1659 * Scan the valid bits looking for invalid sections that 1660 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1661 * valid bit may be set ) have already been zerod by 1662 * vm_page_set_validclean(). 1663 */ 1664 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1665 if (i == (PAGE_SIZE / DEV_BSIZE) || 1666 (m->valid & (1 << i)) 1667 ) { 1668 if (i > b) { 1669 pmap_zero_page_area(m, 1670 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1671 } 1672 b = i + 1; 1673 } 1674 } 1675 1676 /* 1677 * setvalid is TRUE when we can safely set the zero'd areas 1678 * as being valid. We can do this if there are no cache consistancy 1679 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1680 */ 1681 if (setvalid) 1682 m->valid = VM_PAGE_BITS_ALL; 1683} 1684 1685/* 1686 * vm_page_is_valid: 1687 * 1688 * Is (partial) page valid? Note that the case where size == 0 1689 * will return FALSE in the degenerate case where the page is 1690 * entirely invalid, and TRUE otherwise. 1691 * 1692 * May not block. 1693 */ 1694int 1695vm_page_is_valid(vm_page_t m, int base, int size) 1696{ 1697 int bits = vm_page_bits(base, size); 1698 1699 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1700 if (m->valid && ((m->valid & bits) == bits)) 1701 return 1; 1702 else 1703 return 0; 1704} 1705 1706/* 1707 * update dirty bits from pmap/mmu. May not block. 1708 */ 1709void 1710vm_page_test_dirty(vm_page_t m) 1711{ 1712 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1713 vm_page_dirty(m); 1714 } 1715} 1716 1717int so_zerocp_fullpage = 0; 1718 1719void 1720vm_page_cowfault(vm_page_t m) 1721{ 1722 vm_page_t mnew; 1723 vm_object_t object; 1724 vm_pindex_t pindex; 1725 1726 object = m->object; 1727 pindex = m->pindex; 1728 1729 retry_alloc: 1730 pmap_remove_all(m); 1731 vm_page_remove(m); 1732 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 1733 if (mnew == NULL) { 1734 vm_page_insert(m, object, pindex); 1735 vm_page_unlock_queues(); 1736 VM_OBJECT_UNLOCK(object); 1737 VM_WAIT; 1738 VM_OBJECT_LOCK(object); 1739 vm_page_lock_queues(); 1740 goto retry_alloc; 1741 } 1742 1743 if (m->cow == 0) { 1744 /* 1745 * check to see if we raced with an xmit complete when 1746 * waiting to allocate a page. If so, put things back 1747 * the way they were 1748 */ 1749 vm_page_free(mnew); 1750 vm_page_insert(m, object, pindex); 1751 } else { /* clear COW & copy page */ 1752 if (!so_zerocp_fullpage) 1753 pmap_copy_page(m, mnew); 1754 mnew->valid = VM_PAGE_BITS_ALL; 1755 vm_page_dirty(mnew); 1756 mnew->wire_count = m->wire_count - m->cow; 1757 m->wire_count = m->cow; 1758 } 1759} 1760 1761void 1762vm_page_cowclear(vm_page_t m) 1763{ 1764 1765 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1766 if (m->cow) { 1767 m->cow--; 1768 /* 1769 * let vm_fault add back write permission lazily 1770 */ 1771 } 1772 /* 1773 * sf_buf_free() will free the page, so we needn't do it here 1774 */ 1775} 1776 1777void 1778vm_page_cowsetup(vm_page_t m) 1779{ 1780 1781 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1782 m->cow++; 1783 pmap_remove_write(m); 1784} 1785 1786#include "opt_ddb.h" 1787#ifdef DDB 1788#include <sys/kernel.h> 1789 1790#include <ddb/ddb.h> 1791 1792DB_SHOW_COMMAND(page, vm_page_print_page_info) 1793{ 1794 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1795 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1796 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1797 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1798 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1799 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1800 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1801 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1802 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1803 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1804} 1805 1806DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1807{ 1808 int i; 1809 db_printf("PQ_FREE:"); 1810 for (i = 0; i < PQ_NUMCOLORS; i++) { 1811 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1812 } 1813 db_printf("\n"); 1814 1815 db_printf("PQ_CACHE:"); 1816 for (i = 0; i < PQ_NUMCOLORS; i++) { 1817 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1818 } 1819 db_printf("\n"); 1820 1821 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1822 vm_page_queues[PQ_ACTIVE].lcnt, 1823 vm_page_queues[PQ_INACTIVE].lcnt); 1824} 1825#endif /* DDB */ 1826