vm_page.c revision 166736
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 166736 2007-02-15 05:43:38Z 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 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 page_range = high_water / PAGE_SIZE - first_page; 302 npages = (total - (page_range * sizeof(struct vm_page)) - 303 (end - new_end)) / PAGE_SIZE; 304 end = new_end; 305 306 /* 307 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 308 */ 309 vaddr += PAGE_SIZE; 310 311 /* 312 * Initialize the mem entry structures now, and put them in the free 313 * queue. 314 */ 315 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 316 mapped = pmap_map(&vaddr, new_end, end, 317 VM_PROT_READ | VM_PROT_WRITE); 318 vm_page_array = (vm_page_t) mapped; 319#ifdef __amd64__ 320 /* 321 * pmap_map on amd64 comes out of the direct-map, not kvm like i386, 322 * so the pages must be tracked for a crashdump to include this data. 323 * This includes the vm_page_array and the early UMA bootstrap pages. 324 */ 325 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 326 dump_add_page(pa); 327#endif 328 phys_avail[biggestone + 1] = new_end; 329 330 /* 331 * Clear all of the page structures 332 */ 333 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 334 vm_page_array_size = page_range; 335 336 /* 337 * This assertion tests the hypothesis that npages and total are 338 * redundant. XXX 339 */ 340 page_range = 0; 341 for (i = 0; phys_avail[i + 1] != 0; i += 2) 342 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 343 KASSERT(page_range == npages, 344 ("vm_page_startup: inconsistent page counts")); 345 346 /* 347 * Construct the free queue(s) in descending order (by physical 348 * address) so that the first 16MB of physical memory is allocated 349 * last rather than first. On large-memory machines, this avoids 350 * the exhaustion of low physical memory before isa_dma_init has run. 351 */ 352 cnt.v_page_count = 0; 353 cnt.v_free_count = 0; 354 list = getenv("vm.blacklist"); 355 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 356 pa = phys_avail[i]; 357 last_pa = phys_avail[i + 1]; 358 while (pa < last_pa) { 359 if (list != NULL && 360 vm_page_blacklist_lookup(list, pa)) 361 printf("Skipping page with pa 0x%jx\n", 362 (uintmax_t)pa); 363 else 364 vm_pageq_add_new_page(pa); 365 pa += PAGE_SIZE; 366 } 367 } 368 freeenv(list); 369 return (vaddr); 370} 371 372void 373vm_page_flag_set(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_flag_clear(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_busy(vm_page_t m) 390{ 391 392 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 393 KASSERT((m->oflags & VPO_BUSY) == 0, 394 ("vm_page_busy: page already busy!!!")); 395 m->oflags |= VPO_BUSY; 396} 397 398/* 399 * vm_page_flash: 400 * 401 * wakeup anyone waiting for the page. 402 */ 403void 404vm_page_flash(vm_page_t m) 405{ 406 407 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 408 if (m->oflags & VPO_WANTED) { 409 m->oflags &= ~VPO_WANTED; 410 wakeup(m); 411 } 412} 413 414/* 415 * vm_page_wakeup: 416 * 417 * clear the VPO_BUSY flag and wakeup anyone waiting for the 418 * page. 419 * 420 */ 421void 422vm_page_wakeup(vm_page_t m) 423{ 424 425 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 426 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 427 m->oflags &= ~VPO_BUSY; 428 vm_page_flash(m); 429} 430 431void 432vm_page_io_start(vm_page_t m) 433{ 434 435 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 436 m->busy++; 437} 438 439void 440vm_page_io_finish(vm_page_t m) 441{ 442 443 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 444 m->busy--; 445 if (m->busy == 0) 446 vm_page_flash(m); 447} 448 449/* 450 * Keep page from being freed by the page daemon 451 * much of the same effect as wiring, except much lower 452 * overhead and should be used only for *very* temporary 453 * holding ("wiring"). 454 */ 455void 456vm_page_hold(vm_page_t mem) 457{ 458 459 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 460 mem->hold_count++; 461} 462 463void 464vm_page_unhold(vm_page_t mem) 465{ 466 467 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 468 --mem->hold_count; 469 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 470 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD)) 471 vm_page_free_toq(mem); 472} 473 474/* 475 * vm_page_free: 476 * 477 * Free a page 478 * 479 * The clearing of PG_ZERO is a temporary safety until the code can be 480 * reviewed to determine that PG_ZERO is being properly cleared on 481 * write faults or maps. PG_ZERO was previously cleared in 482 * vm_page_alloc(). 483 */ 484void 485vm_page_free(vm_page_t m) 486{ 487 vm_page_flag_clear(m, PG_ZERO); 488 vm_page_free_toq(m); 489} 490 491/* 492 * vm_page_free_zero: 493 * 494 * Free a page to the zerod-pages queue 495 */ 496void 497vm_page_free_zero(vm_page_t m) 498{ 499 vm_page_flag_set(m, PG_ZERO); 500 vm_page_free_toq(m); 501} 502 503/* 504 * vm_page_sleep: 505 * 506 * Sleep and release the page queues lock. 507 * 508 * The object containing the given page must be locked. 509 */ 510void 511vm_page_sleep(vm_page_t m, const char *msg) 512{ 513 514 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 515 if (!mtx_owned(&vm_page_queue_mtx)) 516 vm_page_lock_queues(); 517 vm_page_flag_set(m, PG_REFERENCED); 518 vm_page_unlock_queues(); 519 520 /* 521 * It's possible that while we sleep, the page will get 522 * unbusied and freed. If we are holding the object 523 * lock, we will assume we hold a reference to the object 524 * such that even if m->object changes, we can re-lock 525 * it. 526 */ 527 m->oflags |= VPO_WANTED; 528 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 529} 530 531/* 532 * vm_page_dirty: 533 * 534 * make page all dirty 535 */ 536void 537vm_page_dirty(vm_page_t m) 538{ 539 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_CACHE, 540 ("vm_page_dirty: page in cache!")); 541 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_FREE, 542 ("vm_page_dirty: page is free!")); 543 m->dirty = VM_PAGE_BITS_ALL; 544} 545 546/* 547 * vm_page_splay: 548 * 549 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 550 * the vm_page containing the given pindex. If, however, that 551 * pindex is not found in the vm_object, returns a vm_page that is 552 * adjacent to the pindex, coming before or after it. 553 */ 554vm_page_t 555vm_page_splay(vm_pindex_t pindex, vm_page_t root) 556{ 557 struct vm_page dummy; 558 vm_page_t lefttreemax, righttreemin, y; 559 560 if (root == NULL) 561 return (root); 562 lefttreemax = righttreemin = &dummy; 563 for (;; root = y) { 564 if (pindex < root->pindex) { 565 if ((y = root->left) == NULL) 566 break; 567 if (pindex < y->pindex) { 568 /* Rotate right. */ 569 root->left = y->right; 570 y->right = root; 571 root = y; 572 if ((y = root->left) == NULL) 573 break; 574 } 575 /* Link into the new root's right tree. */ 576 righttreemin->left = root; 577 righttreemin = root; 578 } else if (pindex > root->pindex) { 579 if ((y = root->right) == NULL) 580 break; 581 if (pindex > y->pindex) { 582 /* Rotate left. */ 583 root->right = y->left; 584 y->left = root; 585 root = y; 586 if ((y = root->right) == NULL) 587 break; 588 } 589 /* Link into the new root's left tree. */ 590 lefttreemax->right = root; 591 lefttreemax = root; 592 } else 593 break; 594 } 595 /* Assemble the new root. */ 596 lefttreemax->right = root->left; 597 righttreemin->left = root->right; 598 root->left = dummy.right; 599 root->right = dummy.left; 600 return (root); 601} 602 603/* 604 * vm_page_insert: [ internal use only ] 605 * 606 * Inserts the given mem entry into the object and object list. 607 * 608 * The pagetables are not updated but will presumably fault the page 609 * in if necessary, or if a kernel page the caller will at some point 610 * enter the page into the kernel's pmap. We are not allowed to block 611 * here so we *can't* do this anyway. 612 * 613 * The object and page must be locked. 614 * This routine may not block. 615 */ 616void 617vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 618{ 619 vm_page_t root; 620 621 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 622 if (m->object != NULL) 623 panic("vm_page_insert: page already inserted"); 624 625 /* 626 * Record the object/offset pair in this page 627 */ 628 m->object = object; 629 m->pindex = pindex; 630 631 /* 632 * Now link into the object's ordered list of backed pages. 633 */ 634 root = object->root; 635 if (root == NULL) { 636 m->left = NULL; 637 m->right = NULL; 638 TAILQ_INSERT_TAIL(&object->memq, m, listq); 639 } else { 640 root = vm_page_splay(pindex, root); 641 if (pindex < root->pindex) { 642 m->left = root->left; 643 m->right = root; 644 root->left = NULL; 645 TAILQ_INSERT_BEFORE(root, m, listq); 646 } else if (pindex == root->pindex) 647 panic("vm_page_insert: offset already allocated"); 648 else { 649 m->right = root->right; 650 m->left = root; 651 root->right = NULL; 652 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 653 } 654 } 655 object->root = m; 656 object->generation++; 657 658 /* 659 * show that the object has one more resident page. 660 */ 661 object->resident_page_count++; 662 /* 663 * Hold the vnode until the last page is released. 664 */ 665 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 666 vhold((struct vnode *)object->handle); 667 668 /* 669 * Since we are inserting a new and possibly dirty page, 670 * update the object's OBJ_MIGHTBEDIRTY flag. 671 */ 672 if (m->flags & PG_WRITEABLE) 673 vm_object_set_writeable_dirty(object); 674} 675 676/* 677 * vm_page_remove: 678 * NOTE: used by device pager as well -wfj 679 * 680 * Removes the given mem entry from the object/offset-page 681 * table and the object page list, but do not invalidate/terminate 682 * the backing store. 683 * 684 * The object and page must be locked. 685 * The underlying pmap entry (if any) is NOT removed here. 686 * This routine may not block. 687 */ 688void 689vm_page_remove(vm_page_t m) 690{ 691 vm_object_t object; 692 vm_page_t root; 693 694 if ((object = m->object) == NULL) 695 return; 696 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 697 if (m->oflags & VPO_BUSY) { 698 m->oflags &= ~VPO_BUSY; 699 vm_page_flash(m); 700 } 701 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 702 703 /* 704 * Now remove from the object's list of backed pages. 705 */ 706 if (m != object->root) 707 vm_page_splay(m->pindex, object->root); 708 if (m->left == NULL) 709 root = m->right; 710 else { 711 root = vm_page_splay(m->pindex, m->left); 712 root->right = m->right; 713 } 714 object->root = root; 715 TAILQ_REMOVE(&object->memq, m, listq); 716 717 /* 718 * And show that the object has one fewer resident page. 719 */ 720 object->resident_page_count--; 721 object->generation++; 722 /* 723 * The vnode may now be recycled. 724 */ 725 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 726 vdrop((struct vnode *)object->handle); 727 728 m->object = NULL; 729} 730 731/* 732 * vm_page_lookup: 733 * 734 * Returns the page associated with the object/offset 735 * pair specified; if none is found, NULL is returned. 736 * 737 * The object must be locked. 738 * This routine may not block. 739 * This is a critical path routine 740 */ 741vm_page_t 742vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 743{ 744 vm_page_t m; 745 746 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 747 if ((m = object->root) != NULL && m->pindex != pindex) { 748 m = vm_page_splay(pindex, m); 749 if ((object->root = m)->pindex != pindex) 750 m = NULL; 751 } 752 return (m); 753} 754 755/* 756 * vm_page_rename: 757 * 758 * Move the given memory entry from its 759 * current object to the specified target object/offset. 760 * 761 * The object must be locked. 762 * This routine may not block. 763 * 764 * Note: swap associated with the page must be invalidated by the move. We 765 * have to do this for several reasons: (1) we aren't freeing the 766 * page, (2) we are dirtying the page, (3) the VM system is probably 767 * moving the page from object A to B, and will then later move 768 * the backing store from A to B and we can't have a conflict. 769 * 770 * Note: we *always* dirty the page. It is necessary both for the 771 * fact that we moved it, and because we may be invalidating 772 * swap. If the page is on the cache, we have to deactivate it 773 * or vm_page_dirty() will panic. Dirty pages are not allowed 774 * on the cache. 775 */ 776void 777vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 778{ 779 780 vm_page_remove(m); 781 vm_page_insert(m, new_object, new_pindex); 782 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 783 vm_page_deactivate(m); 784 vm_page_dirty(m); 785} 786 787/* 788 * vm_page_select_cache: 789 * 790 * Move a page of the given color from the cache queue to the free 791 * queue. As pages might be found, but are not applicable, they are 792 * deactivated. 793 * 794 * This routine may not block. 795 */ 796vm_page_t 797vm_page_select_cache(int color) 798{ 799 vm_object_t object; 800 vm_page_t m; 801 boolean_t was_trylocked; 802 803 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 804 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 805 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 806 KASSERT(!pmap_page_is_mapped(m), 807 ("Found mapped cache page %p", m)); 808 KASSERT((m->flags & PG_UNMANAGED) == 0, 809 ("Found unmanaged cache page %p", m)); 810 KASSERT(m->wire_count == 0, ("Found wired cache page %p", m)); 811 if (m->hold_count == 0 && (object = m->object, 812 (was_trylocked = VM_OBJECT_TRYLOCK(object)) || 813 VM_OBJECT_LOCKED(object))) { 814 KASSERT((m->oflags & VPO_BUSY) == 0 && m->busy == 0, 815 ("Found busy cache page %p", m)); 816 vm_page_free(m); 817 if (was_trylocked) 818 VM_OBJECT_UNLOCK(object); 819 break; 820 } 821 vm_page_deactivate(m); 822 } 823 return (m); 824} 825 826/* 827 * vm_page_alloc: 828 * 829 * Allocate and return a memory cell associated 830 * with this VM object/offset pair. 831 * 832 * page_req classes: 833 * VM_ALLOC_NORMAL normal process request 834 * VM_ALLOC_SYSTEM system *really* needs a page 835 * VM_ALLOC_INTERRUPT interrupt time request 836 * VM_ALLOC_ZERO zero page 837 * 838 * This routine may not block. 839 * 840 * Additional special handling is required when called from an 841 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 842 * the page cache in this case. 843 */ 844vm_page_t 845vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 846{ 847 vm_page_t m = NULL; 848 int color, flags, page_req; 849 850 page_req = req & VM_ALLOC_CLASS_MASK; 851 KASSERT(curthread->td_intr_nesting_level == 0 || 852 page_req == VM_ALLOC_INTERRUPT, 853 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 854 855 if ((req & VM_ALLOC_NOOBJ) == 0) { 856 KASSERT(object != NULL, 857 ("vm_page_alloc: NULL object.")); 858 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 859 color = (pindex + object->pg_color) & PQ_COLORMASK; 860 } else 861 color = pindex & PQ_COLORMASK; 862 863 /* 864 * The pager is allowed to eat deeper into the free page list. 865 */ 866 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 867 page_req = VM_ALLOC_SYSTEM; 868 }; 869 870loop: 871 mtx_lock(&vm_page_queue_free_mtx); 872 if (cnt.v_free_count > cnt.v_free_reserved || 873 (page_req == VM_ALLOC_SYSTEM && 874 cnt.v_cache_count == 0 && 875 cnt.v_free_count > cnt.v_interrupt_free_min) || 876 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 877 /* 878 * Allocate from the free queue if the number of free pages 879 * exceeds the minimum for the request class. 880 */ 881 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 882 } else if (page_req != VM_ALLOC_INTERRUPT) { 883 mtx_unlock(&vm_page_queue_free_mtx); 884 /* 885 * Allocatable from cache (non-interrupt only). On success, 886 * we must free the page and try again, thus ensuring that 887 * cnt.v_*_free_min counters are replenished. 888 */ 889 vm_page_lock_queues(); 890 if ((m = vm_page_select_cache(color)) == NULL) { 891 KASSERT(cnt.v_cache_count == 0, 892 ("vm_page_alloc: cache queue is missing %d pages", 893 cnt.v_cache_count)); 894 vm_page_unlock_queues(); 895 atomic_add_int(&vm_pageout_deficit, 1); 896 pagedaemon_wakeup(); 897 898 if (page_req != VM_ALLOC_SYSTEM) 899 return (NULL); 900 901 mtx_lock(&vm_page_queue_free_mtx); 902 if (cnt.v_free_count <= cnt.v_interrupt_free_min) { 903 mtx_unlock(&vm_page_queue_free_mtx); 904 return (NULL); 905 } 906 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 907 } else { 908 vm_page_unlock_queues(); 909 goto loop; 910 } 911 } else { 912 /* 913 * Not allocatable from cache from interrupt, give up. 914 */ 915 mtx_unlock(&vm_page_queue_free_mtx); 916 atomic_add_int(&vm_pageout_deficit, 1); 917 pagedaemon_wakeup(); 918 return (NULL); 919 } 920 921 /* 922 * At this point we had better have found a good page. 923 */ 924 925 KASSERT( 926 m != NULL, 927 ("vm_page_alloc(): missing page on free queue") 928 ); 929 930 /* 931 * Remove from free queue 932 */ 933 vm_pageq_remove_nowakeup(m); 934 935 /* 936 * Initialize structure. Only the PG_ZERO flag is inherited. 937 */ 938 flags = 0; 939 if (m->flags & PG_ZERO) { 940 vm_page_zero_count--; 941 if (req & VM_ALLOC_ZERO) 942 flags = PG_ZERO; 943 } 944 m->flags = flags; 945 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 946 m->oflags = 0; 947 else 948 m->oflags = VPO_BUSY; 949 if (req & VM_ALLOC_WIRED) { 950 atomic_add_int(&cnt.v_wire_count, 1); 951 m->wire_count = 1; 952 } else 953 m->wire_count = 0; 954 m->hold_count = 0; 955 m->act_count = 0; 956 m->busy = 0; 957 m->valid = 0; 958 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 959 mtx_unlock(&vm_page_queue_free_mtx); 960 961 if ((req & VM_ALLOC_NOOBJ) == 0) 962 vm_page_insert(m, object, pindex); 963 else 964 m->pindex = pindex; 965 966 /* 967 * Don't wakeup too often - wakeup the pageout daemon when 968 * we would be nearly out of memory. 969 */ 970 if (vm_paging_needed()) 971 pagedaemon_wakeup(); 972 973 return (m); 974} 975 976/* 977 * vm_wait: (also see VM_WAIT macro) 978 * 979 * Block until free pages are available for allocation 980 * - Called in various places before memory allocations. 981 */ 982void 983vm_wait(void) 984{ 985 986 mtx_lock(&vm_page_queue_free_mtx); 987 if (curproc == pageproc) { 988 vm_pageout_pages_needed = 1; 989 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 990 PDROP | PSWP, "VMWait", 0); 991 } else { 992 if (!vm_pages_needed) { 993 vm_pages_needed = 1; 994 wakeup(&vm_pages_needed); 995 } 996 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 997 "vmwait", 0); 998 } 999} 1000 1001/* 1002 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1003 * 1004 * Block until free pages are available for allocation 1005 * - Called only in vm_fault so that processes page faulting 1006 * can be easily tracked. 1007 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1008 * processes will be able to grab memory first. Do not change 1009 * this balance without careful testing first. 1010 */ 1011void 1012vm_waitpfault(void) 1013{ 1014 1015 mtx_lock(&vm_page_queue_free_mtx); 1016 if (!vm_pages_needed) { 1017 vm_pages_needed = 1; 1018 wakeup(&vm_pages_needed); 1019 } 1020 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1021 "pfault", 0); 1022} 1023 1024/* 1025 * vm_page_activate: 1026 * 1027 * Put the specified page on the active list (if appropriate). 1028 * Ensure that act_count is at least ACT_INIT but do not otherwise 1029 * mess with it. 1030 * 1031 * The page queues must be locked. 1032 * This routine may not block. 1033 */ 1034void 1035vm_page_activate(vm_page_t m) 1036{ 1037 1038 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1039 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) { 1040 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1041 cnt.v_reactivated++; 1042 vm_pageq_remove(m); 1043 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1044 if (m->act_count < ACT_INIT) 1045 m->act_count = ACT_INIT; 1046 vm_pageq_enqueue(PQ_ACTIVE, m); 1047 } 1048 } else { 1049 if (m->act_count < ACT_INIT) 1050 m->act_count = ACT_INIT; 1051 } 1052} 1053 1054/* 1055 * vm_page_free_wakeup: 1056 * 1057 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1058 * routine is called when a page has been added to the cache or free 1059 * queues. 1060 * 1061 * The page queues must be locked. 1062 * This routine may not block. 1063 */ 1064static inline void 1065vm_page_free_wakeup(void) 1066{ 1067 1068 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1069 /* 1070 * if pageout daemon needs pages, then tell it that there are 1071 * some free. 1072 */ 1073 if (vm_pageout_pages_needed && 1074 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1075 wakeup(&vm_pageout_pages_needed); 1076 vm_pageout_pages_needed = 0; 1077 } 1078 /* 1079 * wakeup processes that are waiting on memory if we hit a 1080 * high water mark. And wakeup scheduler process if we have 1081 * lots of memory. this process will swapin processes. 1082 */ 1083 if (vm_pages_needed && !vm_page_count_min()) { 1084 vm_pages_needed = 0; 1085 wakeup(&cnt.v_free_count); 1086 } 1087} 1088 1089/* 1090 * vm_page_free_toq: 1091 * 1092 * Returns the given page to the PQ_FREE list, 1093 * disassociating it with any VM object. 1094 * 1095 * Object and page must be locked prior to entry. 1096 * This routine may not block. 1097 */ 1098 1099void 1100vm_page_free_toq(vm_page_t m) 1101{ 1102 struct vpgqueues *pq; 1103 1104 if (VM_PAGE_GETQUEUE(m) != PQ_NONE) 1105 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1106 KASSERT(!pmap_page_is_mapped(m), 1107 ("vm_page_free_toq: freeing mapped page %p", m)); 1108 cnt.v_tfree++; 1109 1110 if (m->busy || VM_PAGE_INQUEUE1(m, PQ_FREE)) { 1111 printf( 1112 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1113 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1114 m->hold_count); 1115 if (VM_PAGE_INQUEUE1(m, PQ_FREE)) 1116 panic("vm_page_free: freeing free page"); 1117 else 1118 panic("vm_page_free: freeing busy page"); 1119 } 1120 1121 /* 1122 * unqueue, then remove page. Note that we cannot destroy 1123 * the page here because we do not want to call the pager's 1124 * callback routine until after we've put the page on the 1125 * appropriate free queue. 1126 */ 1127 vm_pageq_remove_nowakeup(m); 1128 vm_page_remove(m); 1129 1130 /* 1131 * If fictitious remove object association and 1132 * return, otherwise delay object association removal. 1133 */ 1134 if ((m->flags & PG_FICTITIOUS) != 0) { 1135 return; 1136 } 1137 1138 m->valid = 0; 1139 vm_page_undirty(m); 1140 1141 if (m->wire_count != 0) { 1142 if (m->wire_count > 1) { 1143 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1144 m->wire_count, (long)m->pindex); 1145 } 1146 panic("vm_page_free: freeing wired page"); 1147 } 1148 if (m->hold_count != 0) { 1149 m->flags &= ~PG_ZERO; 1150 vm_pageq_enqueue(PQ_HOLD, m); 1151 return; 1152 } 1153 VM_PAGE_SETQUEUE1(m, PQ_FREE); 1154 mtx_lock(&vm_page_queue_free_mtx); 1155 pq = &vm_page_queues[VM_PAGE_GETQUEUE(m)]; 1156 pq->lcnt++; 1157 ++(*pq->cnt); 1158 1159 /* 1160 * Put zero'd pages on the end ( where we look for zero'd pages 1161 * first ) and non-zerod pages at the head. 1162 */ 1163 if (m->flags & PG_ZERO) { 1164 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1165 ++vm_page_zero_count; 1166 } else { 1167 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1168 vm_page_zero_idle_wakeup(); 1169 } 1170 vm_page_free_wakeup(); 1171 mtx_unlock(&vm_page_queue_free_mtx); 1172} 1173 1174/* 1175 * vm_page_unmanage: 1176 * 1177 * Prevent PV management from being done on the page. The page is 1178 * removed from the paging queues as if it were wired, and as a 1179 * consequence of no longer being managed the pageout daemon will not 1180 * touch it (since there is no way to locate the pte mappings for the 1181 * page). madvise() calls that mess with the pmap will also no longer 1182 * operate on the page. 1183 * 1184 * Beyond that the page is still reasonably 'normal'. Freeing the page 1185 * will clear the flag. 1186 * 1187 * This routine is used by OBJT_PHYS objects - objects using unswappable 1188 * physical memory as backing store rather then swap-backed memory and 1189 * will eventually be extended to support 4MB unmanaged physical 1190 * mappings. 1191 */ 1192void 1193vm_page_unmanage(vm_page_t m) 1194{ 1195 1196 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1197 if ((m->flags & PG_UNMANAGED) == 0) { 1198 if (m->wire_count == 0) 1199 vm_pageq_remove(m); 1200 } 1201 vm_page_flag_set(m, PG_UNMANAGED); 1202} 1203 1204/* 1205 * vm_page_wire: 1206 * 1207 * Mark this page as wired down by yet 1208 * another map, removing it from paging queues 1209 * as necessary. 1210 * 1211 * The page queues must be locked. 1212 * This routine may not block. 1213 */ 1214void 1215vm_page_wire(vm_page_t m) 1216{ 1217 1218 /* 1219 * Only bump the wire statistics if the page is not already wired, 1220 * and only unqueue the page if it is on some queue (if it is unmanaged 1221 * it is already off the queues). 1222 */ 1223 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1224 if (m->flags & PG_FICTITIOUS) 1225 return; 1226 if (m->wire_count == 0) { 1227 if ((m->flags & PG_UNMANAGED) == 0) 1228 vm_pageq_remove(m); 1229 atomic_add_int(&cnt.v_wire_count, 1); 1230 } 1231 m->wire_count++; 1232 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1233} 1234 1235/* 1236 * vm_page_unwire: 1237 * 1238 * Release one wiring of this page, potentially 1239 * enabling it to be paged again. 1240 * 1241 * Many pages placed on the inactive queue should actually go 1242 * into the cache, but it is difficult to figure out which. What 1243 * we do instead, if the inactive target is well met, is to put 1244 * clean pages at the head of the inactive queue instead of the tail. 1245 * This will cause them to be moved to the cache more quickly and 1246 * if not actively re-referenced, freed more quickly. If we just 1247 * stick these pages at the end of the inactive queue, heavy filesystem 1248 * meta-data accesses can cause an unnecessary paging load on memory bound 1249 * processes. This optimization causes one-time-use metadata to be 1250 * reused more quickly. 1251 * 1252 * BUT, if we are in a low-memory situation we have no choice but to 1253 * put clean pages on the cache queue. 1254 * 1255 * A number of routines use vm_page_unwire() to guarantee that the page 1256 * will go into either the inactive or active queues, and will NEVER 1257 * be placed in the cache - for example, just after dirtying a page. 1258 * dirty pages in the cache are not allowed. 1259 * 1260 * The page queues must be locked. 1261 * This routine may not block. 1262 */ 1263void 1264vm_page_unwire(vm_page_t m, int activate) 1265{ 1266 1267 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1268 if (m->flags & PG_FICTITIOUS) 1269 return; 1270 if (m->wire_count > 0) { 1271 m->wire_count--; 1272 if (m->wire_count == 0) { 1273 atomic_subtract_int(&cnt.v_wire_count, 1); 1274 if (m->flags & PG_UNMANAGED) { 1275 ; 1276 } else if (activate) 1277 vm_pageq_enqueue(PQ_ACTIVE, m); 1278 else { 1279 vm_page_flag_clear(m, PG_WINATCFLS); 1280 vm_pageq_enqueue(PQ_INACTIVE, m); 1281 } 1282 } 1283 } else { 1284 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1285 } 1286} 1287 1288 1289/* 1290 * Move the specified page to the inactive queue. If the page has 1291 * any associated swap, the swap is deallocated. 1292 * 1293 * Normally athead is 0 resulting in LRU operation. athead is set 1294 * to 1 if we want this page to be 'as if it were placed in the cache', 1295 * except without unmapping it from the process address space. 1296 * 1297 * This routine may not block. 1298 */ 1299static inline void 1300_vm_page_deactivate(vm_page_t m, int athead) 1301{ 1302 1303 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1304 1305 /* 1306 * Ignore if already inactive. 1307 */ 1308 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) 1309 return; 1310 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1311 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1312 cnt.v_reactivated++; 1313 vm_page_flag_clear(m, PG_WINATCFLS); 1314 vm_pageq_remove(m); 1315 if (athead) 1316 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1317 else 1318 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1319 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1320 vm_page_queues[PQ_INACTIVE].lcnt++; 1321 cnt.v_inactive_count++; 1322 } 1323} 1324 1325void 1326vm_page_deactivate(vm_page_t m) 1327{ 1328 _vm_page_deactivate(m, 0); 1329} 1330 1331/* 1332 * vm_page_try_to_cache: 1333 * 1334 * Returns 0 on failure, 1 on success 1335 */ 1336int 1337vm_page_try_to_cache(vm_page_t m) 1338{ 1339 1340 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1341 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1342 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1343 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1344 return (0); 1345 } 1346 pmap_remove_all(m); 1347 if (m->dirty) 1348 return (0); 1349 vm_page_cache(m); 1350 return (1); 1351} 1352 1353/* 1354 * vm_page_try_to_free() 1355 * 1356 * Attempt to free the page. If we cannot free it, we do nothing. 1357 * 1 is returned on success, 0 on failure. 1358 */ 1359int 1360vm_page_try_to_free(vm_page_t m) 1361{ 1362 1363 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1364 if (m->object != NULL) 1365 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1366 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1367 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1368 return (0); 1369 } 1370 pmap_remove_all(m); 1371 if (m->dirty) 1372 return (0); 1373 vm_page_free(m); 1374 return (1); 1375} 1376 1377/* 1378 * vm_page_cache 1379 * 1380 * Put the specified page onto the page cache queue (if appropriate). 1381 * 1382 * This routine may not block. 1383 */ 1384void 1385vm_page_cache(vm_page_t m) 1386{ 1387 1388 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1389 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1390 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1391 m->hold_count || m->wire_count) { 1392 printf("vm_page_cache: attempting to cache busy page\n"); 1393 return; 1394 } 1395 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1396 return; 1397 1398 /* 1399 * Remove all pmaps and indicate that the page is not 1400 * writeable or mapped. 1401 */ 1402 pmap_remove_all(m); 1403 if (m->dirty != 0) { 1404 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1405 (long)m->pindex); 1406 } 1407 vm_pageq_remove_nowakeup(m); 1408 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1409 mtx_lock(&vm_page_queue_free_mtx); 1410 vm_page_free_wakeup(); 1411 mtx_unlock(&vm_page_queue_free_mtx); 1412} 1413 1414/* 1415 * vm_page_dontneed 1416 * 1417 * Cache, deactivate, or do nothing as appropriate. This routine 1418 * is typically used by madvise() MADV_DONTNEED. 1419 * 1420 * Generally speaking we want to move the page into the cache so 1421 * it gets reused quickly. However, this can result in a silly syndrome 1422 * due to the page recycling too quickly. Small objects will not be 1423 * fully cached. On the otherhand, if we move the page to the inactive 1424 * queue we wind up with a problem whereby very large objects 1425 * unnecessarily blow away our inactive and cache queues. 1426 * 1427 * The solution is to move the pages based on a fixed weighting. We 1428 * either leave them alone, deactivate them, or move them to the cache, 1429 * where moving them to the cache has the highest weighting. 1430 * By forcing some pages into other queues we eventually force the 1431 * system to balance the queues, potentially recovering other unrelated 1432 * space from active. The idea is to not force this to happen too 1433 * often. 1434 */ 1435void 1436vm_page_dontneed(vm_page_t m) 1437{ 1438 static int dnweight; 1439 int dnw; 1440 int head; 1441 1442 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1443 dnw = ++dnweight; 1444 1445 /* 1446 * occassionally leave the page alone 1447 */ 1448 if ((dnw & 0x01F0) == 0 || 1449 VM_PAGE_INQUEUE2(m, PQ_INACTIVE) || 1450 VM_PAGE_INQUEUE1(m, PQ_CACHE) 1451 ) { 1452 if (m->act_count >= ACT_INIT) 1453 --m->act_count; 1454 return; 1455 } 1456 1457 if (m->dirty == 0 && pmap_is_modified(m)) 1458 vm_page_dirty(m); 1459 1460 if (m->dirty || (dnw & 0x0070) == 0) { 1461 /* 1462 * Deactivate the page 3 times out of 32. 1463 */ 1464 head = 0; 1465 } else { 1466 /* 1467 * Cache the page 28 times out of every 32. Note that 1468 * the page is deactivated instead of cached, but placed 1469 * at the head of the queue instead of the tail. 1470 */ 1471 head = 1; 1472 } 1473 _vm_page_deactivate(m, head); 1474} 1475 1476/* 1477 * Grab a page, waiting until we are waken up due to the page 1478 * changing state. We keep on waiting, if the page continues 1479 * to be in the object. If the page doesn't exist, first allocate it 1480 * and then conditionally zero it. 1481 * 1482 * This routine may block. 1483 */ 1484vm_page_t 1485vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1486{ 1487 vm_page_t m; 1488 1489 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1490retrylookup: 1491 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1492 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) { 1493 if ((allocflags & VM_ALLOC_RETRY) == 0) 1494 return (NULL); 1495 goto retrylookup; 1496 } else { 1497 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1498 vm_page_lock_queues(); 1499 vm_page_wire(m); 1500 vm_page_unlock_queues(); 1501 } 1502 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1503 vm_page_busy(m); 1504 return (m); 1505 } 1506 } 1507 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1508 if (m == NULL) { 1509 VM_OBJECT_UNLOCK(object); 1510 VM_WAIT; 1511 VM_OBJECT_LOCK(object); 1512 if ((allocflags & VM_ALLOC_RETRY) == 0) 1513 return (NULL); 1514 goto retrylookup; 1515 } 1516 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1517 pmap_zero_page(m); 1518 return (m); 1519} 1520 1521/* 1522 * Mapping function for valid bits or for dirty bits in 1523 * a page. May not block. 1524 * 1525 * Inputs are required to range within a page. 1526 */ 1527inline int 1528vm_page_bits(int base, int size) 1529{ 1530 int first_bit; 1531 int last_bit; 1532 1533 KASSERT( 1534 base + size <= PAGE_SIZE, 1535 ("vm_page_bits: illegal base/size %d/%d", base, size) 1536 ); 1537 1538 if (size == 0) /* handle degenerate case */ 1539 return (0); 1540 1541 first_bit = base >> DEV_BSHIFT; 1542 last_bit = (base + size - 1) >> DEV_BSHIFT; 1543 1544 return ((2 << last_bit) - (1 << first_bit)); 1545} 1546 1547/* 1548 * vm_page_set_validclean: 1549 * 1550 * Sets portions of a page valid and clean. The arguments are expected 1551 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1552 * of any partial chunks touched by the range. The invalid portion of 1553 * such chunks will be zero'd. 1554 * 1555 * This routine may not block. 1556 * 1557 * (base + size) must be less then or equal to PAGE_SIZE. 1558 */ 1559void 1560vm_page_set_validclean(vm_page_t m, int base, int size) 1561{ 1562 int pagebits; 1563 int frag; 1564 int endoff; 1565 1566 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1567 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1568 if (size == 0) /* handle degenerate case */ 1569 return; 1570 1571 /* 1572 * If the base is not DEV_BSIZE aligned and the valid 1573 * bit is clear, we have to zero out a portion of the 1574 * first block. 1575 */ 1576 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1577 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1578 pmap_zero_page_area(m, frag, base - frag); 1579 1580 /* 1581 * If the ending offset is not DEV_BSIZE aligned and the 1582 * valid bit is clear, we have to zero out a portion of 1583 * the last block. 1584 */ 1585 endoff = base + size; 1586 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1587 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1588 pmap_zero_page_area(m, endoff, 1589 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1590 1591 /* 1592 * Set valid, clear dirty bits. If validating the entire 1593 * page we can safely clear the pmap modify bit. We also 1594 * use this opportunity to clear the VPO_NOSYNC flag. If a process 1595 * takes a write fault on a MAP_NOSYNC memory area the flag will 1596 * be set again. 1597 * 1598 * We set valid bits inclusive of any overlap, but we can only 1599 * clear dirty bits for DEV_BSIZE chunks that are fully within 1600 * the range. 1601 */ 1602 pagebits = vm_page_bits(base, size); 1603 m->valid |= pagebits; 1604#if 0 /* NOT YET */ 1605 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1606 frag = DEV_BSIZE - frag; 1607 base += frag; 1608 size -= frag; 1609 if (size < 0) 1610 size = 0; 1611 } 1612 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1613#endif 1614 m->dirty &= ~pagebits; 1615 if (base == 0 && size == PAGE_SIZE) { 1616 pmap_clear_modify(m); 1617 m->oflags &= ~VPO_NOSYNC; 1618 } 1619} 1620 1621void 1622vm_page_clear_dirty(vm_page_t m, int base, int size) 1623{ 1624 1625 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1626 m->dirty &= ~vm_page_bits(base, size); 1627} 1628 1629/* 1630 * vm_page_set_invalid: 1631 * 1632 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1633 * valid and dirty bits for the effected areas are cleared. 1634 * 1635 * May not block. 1636 */ 1637void 1638vm_page_set_invalid(vm_page_t m, int base, int size) 1639{ 1640 int bits; 1641 1642 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1643 bits = vm_page_bits(base, size); 1644 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1645 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 1646 pmap_remove_all(m); 1647 m->valid &= ~bits; 1648 m->dirty &= ~bits; 1649 m->object->generation++; 1650} 1651 1652/* 1653 * vm_page_zero_invalid() 1654 * 1655 * The kernel assumes that the invalid portions of a page contain 1656 * garbage, but such pages can be mapped into memory by user code. 1657 * When this occurs, we must zero out the non-valid portions of the 1658 * page so user code sees what it expects. 1659 * 1660 * Pages are most often semi-valid when the end of a file is mapped 1661 * into memory and the file's size is not page aligned. 1662 */ 1663void 1664vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1665{ 1666 int b; 1667 int i; 1668 1669 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1670 /* 1671 * Scan the valid bits looking for invalid sections that 1672 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1673 * valid bit may be set ) have already been zerod by 1674 * vm_page_set_validclean(). 1675 */ 1676 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1677 if (i == (PAGE_SIZE / DEV_BSIZE) || 1678 (m->valid & (1 << i)) 1679 ) { 1680 if (i > b) { 1681 pmap_zero_page_area(m, 1682 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1683 } 1684 b = i + 1; 1685 } 1686 } 1687 1688 /* 1689 * setvalid is TRUE when we can safely set the zero'd areas 1690 * as being valid. We can do this if there are no cache consistancy 1691 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1692 */ 1693 if (setvalid) 1694 m->valid = VM_PAGE_BITS_ALL; 1695} 1696 1697/* 1698 * vm_page_is_valid: 1699 * 1700 * Is (partial) page valid? Note that the case where size == 0 1701 * will return FALSE in the degenerate case where the page is 1702 * entirely invalid, and TRUE otherwise. 1703 * 1704 * May not block. 1705 */ 1706int 1707vm_page_is_valid(vm_page_t m, int base, int size) 1708{ 1709 int bits = vm_page_bits(base, size); 1710 1711 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1712 if (m->valid && ((m->valid & bits) == bits)) 1713 return 1; 1714 else 1715 return 0; 1716} 1717 1718/* 1719 * update dirty bits from pmap/mmu. May not block. 1720 */ 1721void 1722vm_page_test_dirty(vm_page_t m) 1723{ 1724 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1725 vm_page_dirty(m); 1726 } 1727} 1728 1729int so_zerocp_fullpage = 0; 1730 1731void 1732vm_page_cowfault(vm_page_t m) 1733{ 1734 vm_page_t mnew; 1735 vm_object_t object; 1736 vm_pindex_t pindex; 1737 1738 object = m->object; 1739 pindex = m->pindex; 1740 1741 retry_alloc: 1742 pmap_remove_all(m); 1743 vm_page_remove(m); 1744 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 1745 if (mnew == NULL) { 1746 vm_page_insert(m, object, pindex); 1747 vm_page_unlock_queues(); 1748 VM_OBJECT_UNLOCK(object); 1749 VM_WAIT; 1750 VM_OBJECT_LOCK(object); 1751 vm_page_lock_queues(); 1752 goto retry_alloc; 1753 } 1754 1755 if (m->cow == 0) { 1756 /* 1757 * check to see if we raced with an xmit complete when 1758 * waiting to allocate a page. If so, put things back 1759 * the way they were 1760 */ 1761 vm_page_free(mnew); 1762 vm_page_insert(m, object, pindex); 1763 } else { /* clear COW & copy page */ 1764 if (!so_zerocp_fullpage) 1765 pmap_copy_page(m, mnew); 1766 mnew->valid = VM_PAGE_BITS_ALL; 1767 vm_page_dirty(mnew); 1768 mnew->wire_count = m->wire_count - m->cow; 1769 m->wire_count = m->cow; 1770 } 1771} 1772 1773void 1774vm_page_cowclear(vm_page_t m) 1775{ 1776 1777 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1778 if (m->cow) { 1779 m->cow--; 1780 /* 1781 * let vm_fault add back write permission lazily 1782 */ 1783 } 1784 /* 1785 * sf_buf_free() will free the page, so we needn't do it here 1786 */ 1787} 1788 1789void 1790vm_page_cowsetup(vm_page_t m) 1791{ 1792 1793 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1794 m->cow++; 1795 pmap_remove_write(m); 1796} 1797 1798#include "opt_ddb.h" 1799#ifdef DDB 1800#include <sys/kernel.h> 1801 1802#include <ddb/ddb.h> 1803 1804DB_SHOW_COMMAND(page, vm_page_print_page_info) 1805{ 1806 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1807 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1808 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1809 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1810 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1811 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1812 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1813 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1814 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1815 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1816} 1817 1818DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1819{ 1820 int i; 1821 db_printf("PQ_FREE:"); 1822 for (i = 0; i < PQ_NUMCOLORS; i++) { 1823 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1824 } 1825 db_printf("\n"); 1826 1827 db_printf("PQ_CACHE:"); 1828 for (i = 0; i < PQ_NUMCOLORS; i++) { 1829 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1830 } 1831 db_printf("\n"); 1832 1833 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1834 vm_page_queues[PQ_ACTIVE].lcnt, 1835 vm_page_queues[PQ_INACTIVE].lcnt); 1836} 1837#endif /* DDB */ 1838