vm_page.c revision 170292
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 170292 2007-06-04 21:45:18Z attilio $"); 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 vm_paddr_t low_water, high_water; 216 int biggestone; 217 218 vm_paddr_t total; 219 220 total = 0; 221 biggestsize = 0; 222 biggestone = 0; 223 nblocks = 0; 224 vaddr = round_page(vaddr); 225 226 for (i = 0; phys_avail[i + 1]; i += 2) { 227 phys_avail[i] = round_page(phys_avail[i]); 228 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 229 } 230 231 low_water = phys_avail[0]; 232 high_water = phys_avail[1]; 233 234 for (i = 0; phys_avail[i + 1]; i += 2) { 235 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 236 237 if (size > biggestsize) { 238 biggestone = i; 239 biggestsize = size; 240 } 241 if (phys_avail[i] < low_water) 242 low_water = phys_avail[i]; 243 if (phys_avail[i + 1] > high_water) 244 high_water = phys_avail[i + 1]; 245 ++nblocks; 246 total += size; 247 } 248 249 end = phys_avail[biggestone+1]; 250 251 /* 252 * Initialize the locks. 253 */ 254 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 255 MTX_RECURSE); 256 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 257 MTX_DEF); 258 259 /* 260 * Initialize the queue headers for the free queue, the active queue 261 * and the inactive queue. 262 */ 263 vm_pageq_init(); 264 265 /* 266 * Allocate memory for use when boot strapping the kernel memory 267 * allocator. 268 */ 269 new_end = end - (boot_pages * UMA_SLAB_SIZE); 270 new_end = trunc_page(new_end); 271 mapped = pmap_map(&vaddr, new_end, end, 272 VM_PROT_READ | VM_PROT_WRITE); 273 bzero((void *)mapped, end - new_end); 274 uma_startup((void *)mapped, boot_pages); 275 276#if defined(__amd64__) || defined(__i386__) 277 /* 278 * Allocate a bitmap to indicate that a random physical page 279 * needs to be included in a minidump. 280 * 281 * The amd64 port needs this to indicate which direct map pages 282 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 283 * 284 * However, i386 still needs this workspace internally within the 285 * minidump code. In theory, they are not needed on i386, but are 286 * included should the sf_buf code decide to use them. 287 */ 288 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 289 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 290 new_end -= vm_page_dump_size; 291 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 292 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 293 bzero((void *)vm_page_dump, vm_page_dump_size); 294#endif 295 /* 296 * Compute the number of pages of memory that will be available for 297 * use (taking into account the overhead of a page structure per 298 * page). 299 */ 300 first_page = low_water / PAGE_SIZE; 301#ifdef VM_PHYSSEG_SPARSE 302 page_range = 0; 303 for (i = 0; phys_avail[i + 1] != 0; i += 2) 304 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 305#elif defined(VM_PHYSSEG_DENSE) 306 page_range = high_water / PAGE_SIZE - first_page; 307#else 308#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 309#endif 310 npages = (total - (page_range * sizeof(struct vm_page)) - 311 (end - new_end)) / PAGE_SIZE; 312 end = new_end; 313 314 /* 315 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 316 */ 317 vaddr += PAGE_SIZE; 318 319 /* 320 * Initialize the mem entry structures now, and put them in the free 321 * queue. 322 */ 323 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 324 mapped = pmap_map(&vaddr, new_end, end, 325 VM_PROT_READ | VM_PROT_WRITE); 326 vm_page_array = (vm_page_t) mapped; 327#ifdef __amd64__ 328 /* 329 * pmap_map on amd64 comes out of the direct-map, not kvm like i386, 330 * so the pages must be tracked for a crashdump to include this data. 331 * This includes the vm_page_array and the early UMA bootstrap pages. 332 */ 333 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 334 dump_add_page(pa); 335#endif 336 phys_avail[biggestone + 1] = new_end; 337 338 /* 339 * Clear all of the page structures 340 */ 341 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 342 vm_page_array_size = page_range; 343 344 /* 345 * This assertion tests the hypothesis that npages and total are 346 * redundant. XXX 347 */ 348 page_range = 0; 349 for (i = 0; phys_avail[i + 1] != 0; i += 2) 350 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 351 KASSERT(page_range == npages, 352 ("vm_page_startup: inconsistent page counts")); 353 354 /* 355 * Construct the free queue(s) in descending order (by physical 356 * address) so that the first 16MB of physical memory is allocated 357 * last rather than first. On large-memory machines, this avoids 358 * the exhaustion of low physical memory before isa_dma_init has run. 359 */ 360 cnt.v_page_count = 0; 361 cnt.v_free_count = 0; 362 list = getenv("vm.blacklist"); 363 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 364 pa = phys_avail[i]; 365 last_pa = phys_avail[i + 1]; 366 while (pa < last_pa) { 367 if (list != NULL && 368 vm_page_blacklist_lookup(list, pa)) 369 printf("Skipping page with pa 0x%jx\n", 370 (uintmax_t)pa); 371 else 372 vm_pageq_add_new_page(pa); 373 pa += PAGE_SIZE; 374 } 375 } 376 freeenv(list); 377 return (vaddr); 378} 379 380void 381vm_page_flag_set(vm_page_t m, unsigned short bits) 382{ 383 384 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 385 m->flags |= bits; 386} 387 388void 389vm_page_flag_clear(vm_page_t m, unsigned short bits) 390{ 391 392 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 393 m->flags &= ~bits; 394} 395 396void 397vm_page_busy(vm_page_t m) 398{ 399 400 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 401 KASSERT((m->oflags & VPO_BUSY) == 0, 402 ("vm_page_busy: page already busy!!!")); 403 m->oflags |= VPO_BUSY; 404} 405 406/* 407 * vm_page_flash: 408 * 409 * wakeup anyone waiting for the page. 410 */ 411void 412vm_page_flash(vm_page_t m) 413{ 414 415 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 416 if (m->oflags & VPO_WANTED) { 417 m->oflags &= ~VPO_WANTED; 418 wakeup(m); 419 } 420} 421 422/* 423 * vm_page_wakeup: 424 * 425 * clear the VPO_BUSY flag and wakeup anyone waiting for the 426 * page. 427 * 428 */ 429void 430vm_page_wakeup(vm_page_t m) 431{ 432 433 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 434 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 435 m->oflags &= ~VPO_BUSY; 436 vm_page_flash(m); 437} 438 439void 440vm_page_io_start(vm_page_t m) 441{ 442 443 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 444 m->busy++; 445} 446 447void 448vm_page_io_finish(vm_page_t m) 449{ 450 451 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 452 m->busy--; 453 if (m->busy == 0) 454 vm_page_flash(m); 455} 456 457/* 458 * Keep page from being freed by the page daemon 459 * much of the same effect as wiring, except much lower 460 * overhead and should be used only for *very* temporary 461 * holding ("wiring"). 462 */ 463void 464vm_page_hold(vm_page_t mem) 465{ 466 467 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 468 mem->hold_count++; 469} 470 471void 472vm_page_unhold(vm_page_t mem) 473{ 474 475 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 476 --mem->hold_count; 477 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 478 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD)) 479 vm_page_free_toq(mem); 480} 481 482/* 483 * vm_page_free: 484 * 485 * Free a page. 486 */ 487void 488vm_page_free(vm_page_t m) 489{ 490 491 m->flags &= ~PG_ZERO; 492 vm_page_free_toq(m); 493} 494 495/* 496 * vm_page_free_zero: 497 * 498 * Free a page to the zerod-pages queue 499 */ 500void 501vm_page_free_zero(vm_page_t m) 502{ 503 504 m->flags |= PG_ZERO; 505 vm_page_free_toq(m); 506} 507 508/* 509 * vm_page_sleep: 510 * 511 * Sleep and release the page queues lock. 512 * 513 * The object containing the given page must be locked. 514 */ 515void 516vm_page_sleep(vm_page_t m, const char *msg) 517{ 518 519 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 520 if (!mtx_owned(&vm_page_queue_mtx)) 521 vm_page_lock_queues(); 522 vm_page_flag_set(m, PG_REFERENCED); 523 vm_page_unlock_queues(); 524 525 /* 526 * It's possible that while we sleep, the page will get 527 * unbusied and freed. If we are holding the object 528 * lock, we will assume we hold a reference to the object 529 * such that even if m->object changes, we can re-lock 530 * it. 531 */ 532 m->oflags |= VPO_WANTED; 533 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 534} 535 536/* 537 * vm_page_dirty: 538 * 539 * make page all dirty 540 */ 541void 542vm_page_dirty(vm_page_t m) 543{ 544 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_CACHE, 545 ("vm_page_dirty: page in cache!")); 546 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_FREE, 547 ("vm_page_dirty: page is free!")); 548 m->dirty = VM_PAGE_BITS_ALL; 549} 550 551/* 552 * vm_page_splay: 553 * 554 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 555 * the vm_page containing the given pindex. If, however, that 556 * pindex is not found in the vm_object, returns a vm_page that is 557 * adjacent to the pindex, coming before or after it. 558 */ 559vm_page_t 560vm_page_splay(vm_pindex_t pindex, vm_page_t root) 561{ 562 struct vm_page dummy; 563 vm_page_t lefttreemax, righttreemin, y; 564 565 if (root == NULL) 566 return (root); 567 lefttreemax = righttreemin = &dummy; 568 for (;; root = y) { 569 if (pindex < root->pindex) { 570 if ((y = root->left) == NULL) 571 break; 572 if (pindex < y->pindex) { 573 /* Rotate right. */ 574 root->left = y->right; 575 y->right = root; 576 root = y; 577 if ((y = root->left) == NULL) 578 break; 579 } 580 /* Link into the new root's right tree. */ 581 righttreemin->left = root; 582 righttreemin = root; 583 } else if (pindex > root->pindex) { 584 if ((y = root->right) == NULL) 585 break; 586 if (pindex > y->pindex) { 587 /* Rotate left. */ 588 root->right = y->left; 589 y->left = root; 590 root = y; 591 if ((y = root->right) == NULL) 592 break; 593 } 594 /* Link into the new root's left tree. */ 595 lefttreemax->right = root; 596 lefttreemax = root; 597 } else 598 break; 599 } 600 /* Assemble the new root. */ 601 lefttreemax->right = root->left; 602 righttreemin->left = root->right; 603 root->left = dummy.right; 604 root->right = dummy.left; 605 return (root); 606} 607 608/* 609 * vm_page_insert: [ internal use only ] 610 * 611 * Inserts the given mem entry into the object and object list. 612 * 613 * The pagetables are not updated but will presumably fault the page 614 * in if necessary, or if a kernel page the caller will at some point 615 * enter the page into the kernel's pmap. We are not allowed to block 616 * here so we *can't* do this anyway. 617 * 618 * The object and page must be locked. 619 * This routine may not block. 620 */ 621void 622vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 623{ 624 vm_page_t root; 625 626 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 627 if (m->object != NULL) 628 panic("vm_page_insert: page already inserted"); 629 630 /* 631 * Record the object/offset pair in this page 632 */ 633 m->object = object; 634 m->pindex = pindex; 635 636 /* 637 * Now link into the object's ordered list of backed pages. 638 */ 639 root = object->root; 640 if (root == NULL) { 641 m->left = NULL; 642 m->right = NULL; 643 TAILQ_INSERT_TAIL(&object->memq, m, listq); 644 } else { 645 root = vm_page_splay(pindex, root); 646 if (pindex < root->pindex) { 647 m->left = root->left; 648 m->right = root; 649 root->left = NULL; 650 TAILQ_INSERT_BEFORE(root, m, listq); 651 } else if (pindex == root->pindex) 652 panic("vm_page_insert: offset already allocated"); 653 else { 654 m->right = root->right; 655 m->left = root; 656 root->right = NULL; 657 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 658 } 659 } 660 object->root = m; 661 object->generation++; 662 663 /* 664 * show that the object has one more resident page. 665 */ 666 object->resident_page_count++; 667 /* 668 * Hold the vnode until the last page is released. 669 */ 670 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 671 vhold((struct vnode *)object->handle); 672 673 /* 674 * Since we are inserting a new and possibly dirty page, 675 * update the object's OBJ_MIGHTBEDIRTY flag. 676 */ 677 if (m->flags & PG_WRITEABLE) 678 vm_object_set_writeable_dirty(object); 679} 680 681/* 682 * vm_page_remove: 683 * NOTE: used by device pager as well -wfj 684 * 685 * Removes the given mem entry from the object/offset-page 686 * table and the object page list, but do not invalidate/terminate 687 * the backing store. 688 * 689 * The object and page must be locked. 690 * The underlying pmap entry (if any) is NOT removed here. 691 * This routine may not block. 692 */ 693void 694vm_page_remove(vm_page_t m) 695{ 696 vm_object_t object; 697 vm_page_t root; 698 699 if ((object = m->object) == NULL) 700 return; 701 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 702 if (m->oflags & VPO_BUSY) { 703 m->oflags &= ~VPO_BUSY; 704 vm_page_flash(m); 705 } 706 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 707 708 /* 709 * Now remove from the object's list of backed pages. 710 */ 711 if (m != object->root) 712 vm_page_splay(m->pindex, object->root); 713 if (m->left == NULL) 714 root = m->right; 715 else { 716 root = vm_page_splay(m->pindex, m->left); 717 root->right = m->right; 718 } 719 object->root = root; 720 TAILQ_REMOVE(&object->memq, m, listq); 721 722 /* 723 * And show that the object has one fewer resident page. 724 */ 725 object->resident_page_count--; 726 object->generation++; 727 /* 728 * The vnode may now be recycled. 729 */ 730 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 731 vdrop((struct vnode *)object->handle); 732 733 m->object = NULL; 734} 735 736/* 737 * vm_page_lookup: 738 * 739 * Returns the page associated with the object/offset 740 * pair specified; if none is found, NULL is returned. 741 * 742 * The object must be locked. 743 * This routine may not block. 744 * This is a critical path routine 745 */ 746vm_page_t 747vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 748{ 749 vm_page_t m; 750 751 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 752 if ((m = object->root) != NULL && m->pindex != pindex) { 753 m = vm_page_splay(pindex, m); 754 if ((object->root = m)->pindex != pindex) 755 m = NULL; 756 } 757 return (m); 758} 759 760/* 761 * vm_page_rename: 762 * 763 * Move the given memory entry from its 764 * current object to the specified target object/offset. 765 * 766 * The object must be locked. 767 * This routine may not block. 768 * 769 * Note: swap associated with the page must be invalidated by the move. We 770 * have to do this for several reasons: (1) we aren't freeing the 771 * page, (2) we are dirtying the page, (3) the VM system is probably 772 * moving the page from object A to B, and will then later move 773 * the backing store from A to B and we can't have a conflict. 774 * 775 * Note: we *always* dirty the page. It is necessary both for the 776 * fact that we moved it, and because we may be invalidating 777 * swap. If the page is on the cache, we have to deactivate it 778 * or vm_page_dirty() will panic. Dirty pages are not allowed 779 * on the cache. 780 */ 781void 782vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 783{ 784 785 vm_page_remove(m); 786 vm_page_insert(m, new_object, new_pindex); 787 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 788 vm_page_deactivate(m); 789 vm_page_dirty(m); 790} 791 792/* 793 * vm_page_select_cache: 794 * 795 * Move a page of the given color from the cache queue to the free 796 * queue. As pages might be found, but are not applicable, they are 797 * deactivated. 798 * 799 * This routine may not block. 800 */ 801vm_page_t 802vm_page_select_cache(int color) 803{ 804 vm_object_t object; 805 vm_page_t m; 806 boolean_t was_trylocked; 807 808 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 809 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 810 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 811 KASSERT(!pmap_page_is_mapped(m), 812 ("Found mapped cache page %p", m)); 813 KASSERT((m->flags & PG_UNMANAGED) == 0, 814 ("Found unmanaged cache page %p", m)); 815 KASSERT(m->wire_count == 0, ("Found wired cache page %p", m)); 816 if (m->hold_count == 0 && (object = m->object, 817 (was_trylocked = VM_OBJECT_TRYLOCK(object)) || 818 VM_OBJECT_LOCKED(object))) { 819 KASSERT((m->oflags & VPO_BUSY) == 0 && m->busy == 0, 820 ("Found busy cache page %p", m)); 821 vm_page_free(m); 822 if (was_trylocked) 823 VM_OBJECT_UNLOCK(object); 824 break; 825 } 826 vm_page_deactivate(m); 827 } 828 return (m); 829} 830 831/* 832 * vm_page_alloc: 833 * 834 * Allocate and return a memory cell associated 835 * with this VM object/offset pair. 836 * 837 * page_req classes: 838 * VM_ALLOC_NORMAL normal process request 839 * VM_ALLOC_SYSTEM system *really* needs a page 840 * VM_ALLOC_INTERRUPT interrupt time request 841 * VM_ALLOC_ZERO zero page 842 * 843 * This routine may not block. 844 * 845 * Additional special handling is required when called from an 846 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 847 * the page cache in this case. 848 */ 849vm_page_t 850vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 851{ 852 vm_page_t m = NULL; 853 int color, flags, page_req; 854 855 page_req = req & VM_ALLOC_CLASS_MASK; 856 KASSERT(curthread->td_intr_nesting_level == 0 || 857 page_req == VM_ALLOC_INTERRUPT, 858 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 859 860 if ((req & VM_ALLOC_NOOBJ) == 0) { 861 KASSERT(object != NULL, 862 ("vm_page_alloc: NULL object.")); 863 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 864 color = (pindex + object->pg_color) & PQ_COLORMASK; 865 } else 866 color = pindex & PQ_COLORMASK; 867 868 /* 869 * The pager is allowed to eat deeper into the free page list. 870 */ 871 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 872 page_req = VM_ALLOC_SYSTEM; 873 }; 874 875loop: 876 mtx_lock(&vm_page_queue_free_mtx); 877 if (cnt.v_free_count > cnt.v_free_reserved || 878 (page_req == VM_ALLOC_SYSTEM && 879 cnt.v_cache_count == 0 && 880 cnt.v_free_count > cnt.v_interrupt_free_min) || 881 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 882 /* 883 * Allocate from the free queue if the number of free pages 884 * exceeds the minimum for the request class. 885 */ 886 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 887 } else if (page_req != VM_ALLOC_INTERRUPT) { 888 mtx_unlock(&vm_page_queue_free_mtx); 889 /* 890 * Allocatable from cache (non-interrupt only). On success, 891 * we must free the page and try again, thus ensuring that 892 * cnt.v_*_free_min counters are replenished. 893 */ 894 vm_page_lock_queues(); 895 if ((m = vm_page_select_cache(color)) == NULL) { 896 KASSERT(cnt.v_cache_count == 0, 897 ("vm_page_alloc: cache queue is missing %d pages", 898 cnt.v_cache_count)); 899 vm_page_unlock_queues(); 900 atomic_add_int(&vm_pageout_deficit, 1); 901 pagedaemon_wakeup(); 902 903 if (page_req != VM_ALLOC_SYSTEM) 904 return (NULL); 905 906 mtx_lock(&vm_page_queue_free_mtx); 907 if (cnt.v_free_count <= cnt.v_interrupt_free_min) { 908 mtx_unlock(&vm_page_queue_free_mtx); 909 return (NULL); 910 } 911 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 912 } else { 913 vm_page_unlock_queues(); 914 goto loop; 915 } 916 } else { 917 /* 918 * Not allocatable from cache from interrupt, give up. 919 */ 920 mtx_unlock(&vm_page_queue_free_mtx); 921 atomic_add_int(&vm_pageout_deficit, 1); 922 pagedaemon_wakeup(); 923 return (NULL); 924 } 925 926 /* 927 * At this point we had better have found a good page. 928 */ 929 930 KASSERT( 931 m != NULL, 932 ("vm_page_alloc(): missing page on free queue") 933 ); 934 935 /* 936 * Remove from free queue 937 */ 938 vm_pageq_remove_nowakeup(m); 939 940 /* 941 * Initialize structure. Only the PG_ZERO flag is inherited. 942 */ 943 flags = 0; 944 if (m->flags & PG_ZERO) { 945 vm_page_zero_count--; 946 if (req & VM_ALLOC_ZERO) 947 flags = PG_ZERO; 948 } 949 if (object != NULL && object->type == OBJT_PHYS) 950 flags |= PG_UNMANAGED; 951 m->flags = flags; 952 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 953 m->oflags = 0; 954 else 955 m->oflags = VPO_BUSY; 956 if (req & VM_ALLOC_WIRED) { 957 atomic_add_int(&cnt.v_wire_count, 1); 958 m->wire_count = 1; 959 } else 960 m->wire_count = 0; 961 m->hold_count = 0; 962 m->act_count = 0; 963 m->busy = 0; 964 m->valid = 0; 965 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 966 mtx_unlock(&vm_page_queue_free_mtx); 967 968 if ((req & VM_ALLOC_NOOBJ) == 0) 969 vm_page_insert(m, object, pindex); 970 else 971 m->pindex = pindex; 972 973 /* 974 * Don't wakeup too often - wakeup the pageout daemon when 975 * we would be nearly out of memory. 976 */ 977 if (vm_paging_needed()) 978 pagedaemon_wakeup(); 979 980 return (m); 981} 982 983/* 984 * vm_wait: (also see VM_WAIT macro) 985 * 986 * Block until free pages are available for allocation 987 * - Called in various places before memory allocations. 988 */ 989void 990vm_wait(void) 991{ 992 993 mtx_lock(&vm_page_queue_free_mtx); 994 if (curproc == pageproc) { 995 vm_pageout_pages_needed = 1; 996 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 997 PDROP | PSWP, "VMWait", 0); 998 } else { 999 if (!vm_pages_needed) { 1000 vm_pages_needed = 1; 1001 wakeup(&vm_pages_needed); 1002 } 1003 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1004 "vmwait", 0); 1005 } 1006} 1007 1008/* 1009 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1010 * 1011 * Block until free pages are available for allocation 1012 * - Called only in vm_fault so that processes page faulting 1013 * can be easily tracked. 1014 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1015 * processes will be able to grab memory first. Do not change 1016 * this balance without careful testing first. 1017 */ 1018void 1019vm_waitpfault(void) 1020{ 1021 1022 mtx_lock(&vm_page_queue_free_mtx); 1023 if (!vm_pages_needed) { 1024 vm_pages_needed = 1; 1025 wakeup(&vm_pages_needed); 1026 } 1027 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1028 "pfault", 0); 1029} 1030 1031/* 1032 * vm_page_activate: 1033 * 1034 * Put the specified page on the active list (if appropriate). 1035 * Ensure that act_count is at least ACT_INIT but do not otherwise 1036 * mess with it. 1037 * 1038 * The page queues must be locked. 1039 * This routine may not block. 1040 */ 1041void 1042vm_page_activate(vm_page_t m) 1043{ 1044 1045 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1046 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) { 1047 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1048 PCPU_INC(cnt.v_reactivated); 1049 vm_pageq_remove(m); 1050 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1051 if (m->act_count < ACT_INIT) 1052 m->act_count = ACT_INIT; 1053 vm_pageq_enqueue(PQ_ACTIVE, m); 1054 } 1055 } else { 1056 if (m->act_count < ACT_INIT) 1057 m->act_count = ACT_INIT; 1058 } 1059} 1060 1061/* 1062 * vm_page_free_wakeup: 1063 * 1064 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1065 * routine is called when a page has been added to the cache or free 1066 * queues. 1067 * 1068 * The page queues must be locked. 1069 * This routine may not block. 1070 */ 1071static inline void 1072vm_page_free_wakeup(void) 1073{ 1074 1075 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1076 /* 1077 * if pageout daemon needs pages, then tell it that there are 1078 * some free. 1079 */ 1080 if (vm_pageout_pages_needed && 1081 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1082 wakeup(&vm_pageout_pages_needed); 1083 vm_pageout_pages_needed = 0; 1084 } 1085 /* 1086 * wakeup processes that are waiting on memory if we hit a 1087 * high water mark. And wakeup scheduler process if we have 1088 * lots of memory. this process will swapin processes. 1089 */ 1090 if (vm_pages_needed && !vm_page_count_min()) { 1091 vm_pages_needed = 0; 1092 wakeup(&cnt.v_free_count); 1093 } 1094} 1095 1096/* 1097 * vm_page_free_toq: 1098 * 1099 * Returns the given page to the PQ_FREE list, 1100 * disassociating it with any VM object. 1101 * 1102 * Object and page must be locked prior to entry. 1103 * This routine may not block. 1104 */ 1105 1106void 1107vm_page_free_toq(vm_page_t m) 1108{ 1109 struct vpgqueues *pq; 1110 1111 if (VM_PAGE_GETQUEUE(m) != PQ_NONE) 1112 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1113 KASSERT(!pmap_page_is_mapped(m), 1114 ("vm_page_free_toq: freeing mapped page %p", m)); 1115 PCPU_INC(cnt.v_tfree); 1116 1117 if (m->busy || VM_PAGE_INQUEUE1(m, PQ_FREE)) { 1118 printf( 1119 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1120 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1121 m->hold_count); 1122 if (VM_PAGE_INQUEUE1(m, PQ_FREE)) 1123 panic("vm_page_free: freeing free page"); 1124 else 1125 panic("vm_page_free: freeing busy page"); 1126 } 1127 1128 /* 1129 * unqueue, then remove page. Note that we cannot destroy 1130 * the page here because we do not want to call the pager's 1131 * callback routine until after we've put the page on the 1132 * appropriate free queue. 1133 */ 1134 vm_pageq_remove_nowakeup(m); 1135 vm_page_remove(m); 1136 1137 /* 1138 * If fictitious remove object association and 1139 * return, otherwise delay object association removal. 1140 */ 1141 if ((m->flags & PG_FICTITIOUS) != 0) { 1142 return; 1143 } 1144 1145 m->valid = 0; 1146 vm_page_undirty(m); 1147 1148 if (m->wire_count != 0) { 1149 if (m->wire_count > 1) { 1150 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1151 m->wire_count, (long)m->pindex); 1152 } 1153 panic("vm_page_free: freeing wired page"); 1154 } 1155 if (m->hold_count != 0) { 1156 m->flags &= ~PG_ZERO; 1157 vm_pageq_enqueue(PQ_HOLD, m); 1158 return; 1159 } 1160 VM_PAGE_SETQUEUE1(m, PQ_FREE); 1161 mtx_lock(&vm_page_queue_free_mtx); 1162 pq = &vm_page_queues[VM_PAGE_GETQUEUE(m)]; 1163 pq->lcnt++; 1164 ++(*pq->cnt); 1165 1166 /* 1167 * Put zero'd pages on the end ( where we look for zero'd pages 1168 * first ) and non-zerod pages at the head. 1169 */ 1170 if (m->flags & PG_ZERO) { 1171 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1172 ++vm_page_zero_count; 1173 } else { 1174 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1175 vm_page_zero_idle_wakeup(); 1176 } 1177 vm_page_free_wakeup(); 1178 mtx_unlock(&vm_page_queue_free_mtx); 1179} 1180 1181/* 1182 * vm_page_wire: 1183 * 1184 * Mark this page as wired down by yet 1185 * another map, removing it from paging queues 1186 * as necessary. 1187 * 1188 * The page queues must be locked. 1189 * This routine may not block. 1190 */ 1191void 1192vm_page_wire(vm_page_t m) 1193{ 1194 1195 /* 1196 * Only bump the wire statistics if the page is not already wired, 1197 * and only unqueue the page if it is on some queue (if it is unmanaged 1198 * it is already off the queues). 1199 */ 1200 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1201 if (m->flags & PG_FICTITIOUS) 1202 return; 1203 if (m->wire_count == 0) { 1204 if ((m->flags & PG_UNMANAGED) == 0) 1205 vm_pageq_remove(m); 1206 atomic_add_int(&cnt.v_wire_count, 1); 1207 } 1208 m->wire_count++; 1209 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1210} 1211 1212/* 1213 * vm_page_unwire: 1214 * 1215 * Release one wiring of this page, potentially 1216 * enabling it to be paged again. 1217 * 1218 * Many pages placed on the inactive queue should actually go 1219 * into the cache, but it is difficult to figure out which. What 1220 * we do instead, if the inactive target is well met, is to put 1221 * clean pages at the head of the inactive queue instead of the tail. 1222 * This will cause them to be moved to the cache more quickly and 1223 * if not actively re-referenced, freed more quickly. If we just 1224 * stick these pages at the end of the inactive queue, heavy filesystem 1225 * meta-data accesses can cause an unnecessary paging load on memory bound 1226 * processes. This optimization causes one-time-use metadata to be 1227 * reused more quickly. 1228 * 1229 * BUT, if we are in a low-memory situation we have no choice but to 1230 * put clean pages on the cache queue. 1231 * 1232 * A number of routines use vm_page_unwire() to guarantee that the page 1233 * will go into either the inactive or active queues, and will NEVER 1234 * be placed in the cache - for example, just after dirtying a page. 1235 * dirty pages in the cache are not allowed. 1236 * 1237 * The page queues must be locked. 1238 * This routine may not block. 1239 */ 1240void 1241vm_page_unwire(vm_page_t m, int activate) 1242{ 1243 1244 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1245 if (m->flags & PG_FICTITIOUS) 1246 return; 1247 if (m->wire_count > 0) { 1248 m->wire_count--; 1249 if (m->wire_count == 0) { 1250 atomic_subtract_int(&cnt.v_wire_count, 1); 1251 if (m->flags & PG_UNMANAGED) { 1252 ; 1253 } else if (activate) 1254 vm_pageq_enqueue(PQ_ACTIVE, m); 1255 else { 1256 vm_page_flag_clear(m, PG_WINATCFLS); 1257 vm_pageq_enqueue(PQ_INACTIVE, m); 1258 } 1259 } 1260 } else { 1261 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1262 } 1263} 1264 1265 1266/* 1267 * Move the specified page to the inactive queue. If the page has 1268 * any associated swap, the swap is deallocated. 1269 * 1270 * Normally athead is 0 resulting in LRU operation. athead is set 1271 * to 1 if we want this page to be 'as if it were placed in the cache', 1272 * except without unmapping it from the process address space. 1273 * 1274 * This routine may not block. 1275 */ 1276static inline void 1277_vm_page_deactivate(vm_page_t m, int athead) 1278{ 1279 1280 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1281 1282 /* 1283 * Ignore if already inactive. 1284 */ 1285 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) 1286 return; 1287 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1288 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1289 PCPU_INC(cnt.v_reactivated); 1290 vm_page_flag_clear(m, PG_WINATCFLS); 1291 vm_pageq_remove(m); 1292 if (athead) 1293 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1294 else 1295 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1296 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1297 vm_page_queues[PQ_INACTIVE].lcnt++; 1298 1299 /* 1300 * Just not use an atomic here since vm_page_queues_lock 1301 * alredy protects this field. 1302 */ 1303 cnt.v_inactive_count++; 1304 } 1305} 1306 1307void 1308vm_page_deactivate(vm_page_t m) 1309{ 1310 _vm_page_deactivate(m, 0); 1311} 1312 1313/* 1314 * vm_page_try_to_cache: 1315 * 1316 * Returns 0 on failure, 1 on success 1317 */ 1318int 1319vm_page_try_to_cache(vm_page_t m) 1320{ 1321 1322 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1323 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1324 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1325 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1326 return (0); 1327 } 1328 pmap_remove_all(m); 1329 if (m->dirty) 1330 return (0); 1331 vm_page_cache(m); 1332 return (1); 1333} 1334 1335/* 1336 * vm_page_try_to_free() 1337 * 1338 * Attempt to free the page. If we cannot free it, we do nothing. 1339 * 1 is returned on success, 0 on failure. 1340 */ 1341int 1342vm_page_try_to_free(vm_page_t m) 1343{ 1344 1345 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1346 if (m->object != NULL) 1347 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1348 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1349 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1350 return (0); 1351 } 1352 pmap_remove_all(m); 1353 if (m->dirty) 1354 return (0); 1355 vm_page_free(m); 1356 return (1); 1357} 1358 1359/* 1360 * vm_page_cache 1361 * 1362 * Put the specified page onto the page cache queue (if appropriate). 1363 * 1364 * This routine may not block. 1365 */ 1366void 1367vm_page_cache(vm_page_t m) 1368{ 1369 1370 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1371 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1372 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1373 m->hold_count || m->wire_count) { 1374 printf("vm_page_cache: attempting to cache busy page\n"); 1375 return; 1376 } 1377 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1378 return; 1379 1380 /* 1381 * Remove all pmaps and indicate that the page is not 1382 * writeable or mapped. 1383 */ 1384 pmap_remove_all(m); 1385 if (m->dirty != 0) { 1386 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1387 (long)m->pindex); 1388 } 1389 vm_pageq_remove_nowakeup(m); 1390 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1391 mtx_lock(&vm_page_queue_free_mtx); 1392 vm_page_free_wakeup(); 1393 mtx_unlock(&vm_page_queue_free_mtx); 1394} 1395 1396/* 1397 * vm_page_dontneed 1398 * 1399 * Cache, deactivate, or do nothing as appropriate. This routine 1400 * is typically used by madvise() MADV_DONTNEED. 1401 * 1402 * Generally speaking we want to move the page into the cache so 1403 * it gets reused quickly. However, this can result in a silly syndrome 1404 * due to the page recycling too quickly. Small objects will not be 1405 * fully cached. On the otherhand, if we move the page to the inactive 1406 * queue we wind up with a problem whereby very large objects 1407 * unnecessarily blow away our inactive and cache queues. 1408 * 1409 * The solution is to move the pages based on a fixed weighting. We 1410 * either leave them alone, deactivate them, or move them to the cache, 1411 * where moving them to the cache has the highest weighting. 1412 * By forcing some pages into other queues we eventually force the 1413 * system to balance the queues, potentially recovering other unrelated 1414 * space from active. The idea is to not force this to happen too 1415 * often. 1416 */ 1417void 1418vm_page_dontneed(vm_page_t m) 1419{ 1420 static int dnweight; 1421 int dnw; 1422 int head; 1423 1424 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1425 dnw = ++dnweight; 1426 1427 /* 1428 * occassionally leave the page alone 1429 */ 1430 if ((dnw & 0x01F0) == 0 || 1431 VM_PAGE_INQUEUE2(m, PQ_INACTIVE) || 1432 VM_PAGE_INQUEUE1(m, PQ_CACHE) 1433 ) { 1434 if (m->act_count >= ACT_INIT) 1435 --m->act_count; 1436 return; 1437 } 1438 1439 if (m->dirty == 0 && pmap_is_modified(m)) 1440 vm_page_dirty(m); 1441 1442 if (m->dirty || (dnw & 0x0070) == 0) { 1443 /* 1444 * Deactivate the page 3 times out of 32. 1445 */ 1446 head = 0; 1447 } else { 1448 /* 1449 * Cache the page 28 times out of every 32. Note that 1450 * the page is deactivated instead of cached, but placed 1451 * at the head of the queue instead of the tail. 1452 */ 1453 head = 1; 1454 } 1455 _vm_page_deactivate(m, head); 1456} 1457 1458/* 1459 * Grab a page, waiting until we are waken up due to the page 1460 * changing state. We keep on waiting, if the page continues 1461 * to be in the object. If the page doesn't exist, first allocate it 1462 * and then conditionally zero it. 1463 * 1464 * This routine may block. 1465 */ 1466vm_page_t 1467vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1468{ 1469 vm_page_t m; 1470 1471 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1472retrylookup: 1473 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1474 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) { 1475 if ((allocflags & VM_ALLOC_RETRY) == 0) 1476 return (NULL); 1477 goto retrylookup; 1478 } else { 1479 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1480 vm_page_lock_queues(); 1481 vm_page_wire(m); 1482 vm_page_unlock_queues(); 1483 } 1484 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1485 vm_page_busy(m); 1486 return (m); 1487 } 1488 } 1489 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1490 if (m == NULL) { 1491 VM_OBJECT_UNLOCK(object); 1492 VM_WAIT; 1493 VM_OBJECT_LOCK(object); 1494 if ((allocflags & VM_ALLOC_RETRY) == 0) 1495 return (NULL); 1496 goto retrylookup; 1497 } 1498 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1499 pmap_zero_page(m); 1500 return (m); 1501} 1502 1503/* 1504 * Mapping function for valid bits or for dirty bits in 1505 * a page. May not block. 1506 * 1507 * Inputs are required to range within a page. 1508 */ 1509inline int 1510vm_page_bits(int base, int size) 1511{ 1512 int first_bit; 1513 int last_bit; 1514 1515 KASSERT( 1516 base + size <= PAGE_SIZE, 1517 ("vm_page_bits: illegal base/size %d/%d", base, size) 1518 ); 1519 1520 if (size == 0) /* handle degenerate case */ 1521 return (0); 1522 1523 first_bit = base >> DEV_BSHIFT; 1524 last_bit = (base + size - 1) >> DEV_BSHIFT; 1525 1526 return ((2 << last_bit) - (1 << first_bit)); 1527} 1528 1529/* 1530 * vm_page_set_validclean: 1531 * 1532 * Sets portions of a page valid and clean. The arguments are expected 1533 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1534 * of any partial chunks touched by the range. The invalid portion of 1535 * such chunks will be zero'd. 1536 * 1537 * This routine may not block. 1538 * 1539 * (base + size) must be less then or equal to PAGE_SIZE. 1540 */ 1541void 1542vm_page_set_validclean(vm_page_t m, int base, int size) 1543{ 1544 int pagebits; 1545 int frag; 1546 int endoff; 1547 1548 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1549 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1550 if (size == 0) /* handle degenerate case */ 1551 return; 1552 1553 /* 1554 * If the base is not DEV_BSIZE aligned and the valid 1555 * bit is clear, we have to zero out a portion of the 1556 * first block. 1557 */ 1558 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1559 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1560 pmap_zero_page_area(m, frag, base - frag); 1561 1562 /* 1563 * If the ending offset is not DEV_BSIZE aligned and the 1564 * valid bit is clear, we have to zero out a portion of 1565 * the last block. 1566 */ 1567 endoff = base + size; 1568 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1569 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1570 pmap_zero_page_area(m, endoff, 1571 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1572 1573 /* 1574 * Set valid, clear dirty bits. If validating the entire 1575 * page we can safely clear the pmap modify bit. We also 1576 * use this opportunity to clear the VPO_NOSYNC flag. If a process 1577 * takes a write fault on a MAP_NOSYNC memory area the flag will 1578 * be set again. 1579 * 1580 * We set valid bits inclusive of any overlap, but we can only 1581 * clear dirty bits for DEV_BSIZE chunks that are fully within 1582 * the range. 1583 */ 1584 pagebits = vm_page_bits(base, size); 1585 m->valid |= pagebits; 1586#if 0 /* NOT YET */ 1587 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1588 frag = DEV_BSIZE - frag; 1589 base += frag; 1590 size -= frag; 1591 if (size < 0) 1592 size = 0; 1593 } 1594 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1595#endif 1596 m->dirty &= ~pagebits; 1597 if (base == 0 && size == PAGE_SIZE) { 1598 pmap_clear_modify(m); 1599 m->oflags &= ~VPO_NOSYNC; 1600 } 1601} 1602 1603void 1604vm_page_clear_dirty(vm_page_t m, int base, int size) 1605{ 1606 1607 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1608 m->dirty &= ~vm_page_bits(base, size); 1609} 1610 1611/* 1612 * vm_page_set_invalid: 1613 * 1614 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1615 * valid and dirty bits for the effected areas are cleared. 1616 * 1617 * May not block. 1618 */ 1619void 1620vm_page_set_invalid(vm_page_t m, int base, int size) 1621{ 1622 int bits; 1623 1624 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1625 bits = vm_page_bits(base, size); 1626 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1627 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 1628 pmap_remove_all(m); 1629 m->valid &= ~bits; 1630 m->dirty &= ~bits; 1631 m->object->generation++; 1632} 1633 1634/* 1635 * vm_page_zero_invalid() 1636 * 1637 * The kernel assumes that the invalid portions of a page contain 1638 * garbage, but such pages can be mapped into memory by user code. 1639 * When this occurs, we must zero out the non-valid portions of the 1640 * page so user code sees what it expects. 1641 * 1642 * Pages are most often semi-valid when the end of a file is mapped 1643 * into memory and the file's size is not page aligned. 1644 */ 1645void 1646vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1647{ 1648 int b; 1649 int i; 1650 1651 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1652 /* 1653 * Scan the valid bits looking for invalid sections that 1654 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1655 * valid bit may be set ) have already been zerod by 1656 * vm_page_set_validclean(). 1657 */ 1658 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1659 if (i == (PAGE_SIZE / DEV_BSIZE) || 1660 (m->valid & (1 << i)) 1661 ) { 1662 if (i > b) { 1663 pmap_zero_page_area(m, 1664 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1665 } 1666 b = i + 1; 1667 } 1668 } 1669 1670 /* 1671 * setvalid is TRUE when we can safely set the zero'd areas 1672 * as being valid. We can do this if there are no cache consistancy 1673 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1674 */ 1675 if (setvalid) 1676 m->valid = VM_PAGE_BITS_ALL; 1677} 1678 1679/* 1680 * vm_page_is_valid: 1681 * 1682 * Is (partial) page valid? Note that the case where size == 0 1683 * will return FALSE in the degenerate case where the page is 1684 * entirely invalid, and TRUE otherwise. 1685 * 1686 * May not block. 1687 */ 1688int 1689vm_page_is_valid(vm_page_t m, int base, int size) 1690{ 1691 int bits = vm_page_bits(base, size); 1692 1693 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1694 if (m->valid && ((m->valid & bits) == bits)) 1695 return 1; 1696 else 1697 return 0; 1698} 1699 1700/* 1701 * update dirty bits from pmap/mmu. May not block. 1702 */ 1703void 1704vm_page_test_dirty(vm_page_t m) 1705{ 1706 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1707 vm_page_dirty(m); 1708 } 1709} 1710 1711int so_zerocp_fullpage = 0; 1712 1713void 1714vm_page_cowfault(vm_page_t m) 1715{ 1716 vm_page_t mnew; 1717 vm_object_t object; 1718 vm_pindex_t pindex; 1719 1720 object = m->object; 1721 pindex = m->pindex; 1722 1723 retry_alloc: 1724 pmap_remove_all(m); 1725 vm_page_remove(m); 1726 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 1727 if (mnew == NULL) { 1728 vm_page_insert(m, object, pindex); 1729 vm_page_unlock_queues(); 1730 VM_OBJECT_UNLOCK(object); 1731 VM_WAIT; 1732 VM_OBJECT_LOCK(object); 1733 vm_page_lock_queues(); 1734 goto retry_alloc; 1735 } 1736 1737 if (m->cow == 0) { 1738 /* 1739 * check to see if we raced with an xmit complete when 1740 * waiting to allocate a page. If so, put things back 1741 * the way they were 1742 */ 1743 vm_page_free(mnew); 1744 vm_page_insert(m, object, pindex); 1745 } else { /* clear COW & copy page */ 1746 if (!so_zerocp_fullpage) 1747 pmap_copy_page(m, mnew); 1748 mnew->valid = VM_PAGE_BITS_ALL; 1749 vm_page_dirty(mnew); 1750 mnew->wire_count = m->wire_count - m->cow; 1751 m->wire_count = m->cow; 1752 } 1753} 1754 1755void 1756vm_page_cowclear(vm_page_t m) 1757{ 1758 1759 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1760 if (m->cow) { 1761 m->cow--; 1762 /* 1763 * let vm_fault add back write permission lazily 1764 */ 1765 } 1766 /* 1767 * sf_buf_free() will free the page, so we needn't do it here 1768 */ 1769} 1770 1771void 1772vm_page_cowsetup(vm_page_t m) 1773{ 1774 1775 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1776 m->cow++; 1777 pmap_remove_write(m); 1778} 1779 1780#include "opt_ddb.h" 1781#ifdef DDB 1782#include <sys/kernel.h> 1783 1784#include <ddb/ddb.h> 1785 1786DB_SHOW_COMMAND(page, vm_page_print_page_info) 1787{ 1788 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1789 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1790 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1791 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1792 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1793 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1794 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1795 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1796 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1797 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1798} 1799 1800DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1801{ 1802 int i; 1803 db_printf("PQ_FREE:"); 1804 for (i = 0; i < PQ_NUMCOLORS; i++) { 1805 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1806 } 1807 db_printf("\n"); 1808 1809 db_printf("PQ_CACHE:"); 1810 for (i = 0; i < PQ_NUMCOLORS; i++) { 1811 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1812 } 1813 db_printf("\n"); 1814 1815 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1816 vm_page_queues[PQ_ACTIVE].lcnt, 1817 vm_page_queues[PQ_INACTIVE].lcnt); 1818} 1819#endif /* DDB */ 1820