vm_page.c revision 136931
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 136931 2004-10-24 23:53:47Z alc $"); 101 102#include <sys/param.h> 103#include <sys/systm.h> 104#include <sys/lock.h> 105#include <sys/malloc.h> 106#include <sys/mutex.h> 107#include <sys/proc.h> 108#include <sys/vmmeter.h> 109#include <sys/vnode.h> 110 111#include <vm/vm.h> 112#include <vm/vm_param.h> 113#include <vm/vm_kern.h> 114#include <vm/vm_object.h> 115#include <vm/vm_page.h> 116#include <vm/vm_pageout.h> 117#include <vm/vm_pager.h> 118#include <vm/vm_extern.h> 119#include <vm/uma.h> 120#include <vm/uma_int.h> 121 122/* 123 * Associated with page of user-allocatable memory is a 124 * page structure. 125 */ 126 127struct mtx vm_page_queue_mtx; 128struct mtx vm_page_queue_free_mtx; 129 130vm_page_t vm_page_array = 0; 131int vm_page_array_size = 0; 132long first_page = 0; 133int vm_page_zero_count = 0; 134 135/* 136 * vm_set_page_size: 137 * 138 * Sets the page size, perhaps based upon the memory 139 * size. Must be called before any use of page-size 140 * dependent functions. 141 */ 142void 143vm_set_page_size(void) 144{ 145 if (cnt.v_page_size == 0) 146 cnt.v_page_size = PAGE_SIZE; 147 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 148 panic("vm_set_page_size: page size not a power of two"); 149} 150 151/* 152 * vm_page_startup: 153 * 154 * Initializes the resident memory module. 155 * 156 * Allocates memory for the page cells, and 157 * for the object/offset-to-page hash table headers. 158 * Each page cell is initialized and placed on the free list. 159 */ 160vm_offset_t 161vm_page_startup(vm_offset_t vaddr) 162{ 163 vm_offset_t mapped; 164 vm_size_t npages; 165 vm_paddr_t page_range; 166 vm_paddr_t new_end; 167 int i; 168 vm_paddr_t pa; 169 int nblocks; 170 vm_paddr_t last_pa; 171 172 /* the biggest memory array is the second group of pages */ 173 vm_paddr_t end; 174 vm_paddr_t biggestsize; 175 int biggestone; 176 177 vm_paddr_t total; 178 vm_size_t bootpages; 179 180 total = 0; 181 biggestsize = 0; 182 biggestone = 0; 183 nblocks = 0; 184 vaddr = round_page(vaddr); 185 186 for (i = 0; phys_avail[i + 1]; i += 2) { 187 phys_avail[i] = round_page(phys_avail[i]); 188 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 189 } 190 191 for (i = 0; phys_avail[i + 1]; i += 2) { 192 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 193 194 if (size > biggestsize) { 195 biggestone = i; 196 biggestsize = size; 197 } 198 ++nblocks; 199 total += size; 200 } 201 202 end = phys_avail[biggestone+1]; 203 204 /* 205 * Initialize the locks. 206 */ 207 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 208 MTX_RECURSE); 209 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 210 MTX_SPIN); 211 212 /* 213 * Initialize the queue headers for the free queue, the active queue 214 * and the inactive queue. 215 */ 216 vm_pageq_init(); 217 218 /* 219 * Allocate memory for use when boot strapping the kernel memory 220 * allocator. 221 */ 222 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 223 new_end = end - bootpages; 224 new_end = trunc_page(new_end); 225 mapped = pmap_map(&vaddr, new_end, end, 226 VM_PROT_READ | VM_PROT_WRITE); 227 bzero((caddr_t) mapped, end - new_end); 228 uma_startup((caddr_t)mapped); 229 230 /* 231 * Compute the number of pages of memory that will be available for 232 * use (taking into account the overhead of a page structure per 233 * page). 234 */ 235 first_page = phys_avail[0] / PAGE_SIZE; 236 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 237 npages = (total - (page_range * sizeof(struct vm_page)) - 238 (end - new_end)) / PAGE_SIZE; 239 end = new_end; 240 241 /* 242 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 243 */ 244 vaddr += PAGE_SIZE; 245 246 /* 247 * Initialize the mem entry structures now, and put them in the free 248 * queue. 249 */ 250 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 251 mapped = pmap_map(&vaddr, new_end, end, 252 VM_PROT_READ | VM_PROT_WRITE); 253 vm_page_array = (vm_page_t) mapped; 254 phys_avail[biggestone + 1] = new_end; 255 256 /* 257 * Clear all of the page structures 258 */ 259 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 260 vm_page_array_size = page_range; 261 262 /* 263 * Construct the free queue(s) in descending order (by physical 264 * address) so that the first 16MB of physical memory is allocated 265 * last rather than first. On large-memory machines, this avoids 266 * the exhaustion of low physical memory before isa_dma_init has run. 267 */ 268 cnt.v_page_count = 0; 269 cnt.v_free_count = 0; 270 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 271 pa = phys_avail[i]; 272 last_pa = phys_avail[i + 1]; 273 while (pa < last_pa && npages-- > 0) { 274 vm_pageq_add_new_page(pa); 275 pa += PAGE_SIZE; 276 } 277 } 278 return (vaddr); 279} 280 281void 282vm_page_flag_set(vm_page_t m, unsigned short bits) 283{ 284 285 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 286 m->flags |= bits; 287} 288 289void 290vm_page_flag_clear(vm_page_t m, unsigned short bits) 291{ 292 293 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 294 m->flags &= ~bits; 295} 296 297void 298vm_page_busy(vm_page_t m) 299{ 300 301 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 302 KASSERT((m->flags & PG_BUSY) == 0, 303 ("vm_page_busy: page already busy!!!")); 304 vm_page_flag_set(m, PG_BUSY); 305} 306 307/* 308 * vm_page_flash: 309 * 310 * wakeup anyone waiting for the page. 311 */ 312void 313vm_page_flash(vm_page_t m) 314{ 315 if (m->flags & PG_WANTED) { 316 vm_page_flag_clear(m, PG_WANTED); 317 wakeup(m); 318 } 319} 320 321/* 322 * vm_page_wakeup: 323 * 324 * clear the PG_BUSY flag and wakeup anyone waiting for the 325 * page. 326 * 327 */ 328void 329vm_page_wakeup(vm_page_t m) 330{ 331 332 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 333 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 334 vm_page_flag_clear(m, PG_BUSY); 335 vm_page_flash(m); 336} 337 338void 339vm_page_io_start(vm_page_t m) 340{ 341 342 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 343 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 344 m->busy++; 345} 346 347void 348vm_page_io_finish(vm_page_t m) 349{ 350 351 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 352 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 353 m->busy--; 354 if (m->busy == 0) 355 vm_page_flash(m); 356} 357 358/* 359 * Keep page from being freed by the page daemon 360 * much of the same effect as wiring, except much lower 361 * overhead and should be used only for *very* temporary 362 * holding ("wiring"). 363 */ 364void 365vm_page_hold(vm_page_t mem) 366{ 367 368 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 369 mem->hold_count++; 370} 371 372void 373vm_page_unhold(vm_page_t mem) 374{ 375 376 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 377 --mem->hold_count; 378 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 379 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 380 vm_page_free_toq(mem); 381} 382 383/* 384 * vm_page_free: 385 * 386 * Free a page 387 * 388 * The clearing of PG_ZERO is a temporary safety until the code can be 389 * reviewed to determine that PG_ZERO is being properly cleared on 390 * write faults or maps. PG_ZERO was previously cleared in 391 * vm_page_alloc(). 392 */ 393void 394vm_page_free(vm_page_t m) 395{ 396 vm_page_flag_clear(m, PG_ZERO); 397 vm_page_free_toq(m); 398 vm_page_zero_idle_wakeup(); 399} 400 401/* 402 * vm_page_free_zero: 403 * 404 * Free a page to the zerod-pages queue 405 */ 406void 407vm_page_free_zero(vm_page_t m) 408{ 409 vm_page_flag_set(m, PG_ZERO); 410 vm_page_free_toq(m); 411} 412 413/* 414 * vm_page_sleep_if_busy: 415 * 416 * Sleep and release the page queues lock if PG_BUSY is set or, 417 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 418 * thread slept and the page queues lock was released. 419 * Otherwise, retains the page queues lock and returns FALSE. 420 */ 421int 422vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 423{ 424 vm_object_t object; 425 int is_object_locked; 426 427 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 428 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 429 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 430 /* 431 * It's possible that while we sleep, the page will get 432 * unbusied and freed. If we are holding the object 433 * lock, we will assume we hold a reference to the object 434 * such that even if m->object changes, we can re-lock 435 * it. 436 * 437 * Remove mtx_owned() after vm_object locking is finished. 438 */ 439 object = m->object; 440 if ((is_object_locked = object != NULL && 441 mtx_owned(&object->mtx))) 442 mtx_unlock(&object->mtx); 443 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 444 if (is_object_locked) 445 mtx_lock(&object->mtx); 446 return (TRUE); 447 } 448 return (FALSE); 449} 450 451/* 452 * vm_page_dirty: 453 * 454 * make page all dirty 455 */ 456void 457vm_page_dirty(vm_page_t m) 458{ 459 KASSERT(m->queue - m->pc != PQ_CACHE, 460 ("vm_page_dirty: page in cache!")); 461 KASSERT(m->queue - m->pc != PQ_FREE, 462 ("vm_page_dirty: page is free!")); 463 m->dirty = VM_PAGE_BITS_ALL; 464} 465 466/* 467 * vm_page_splay: 468 * 469 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 470 * the vm_page containing the given pindex. If, however, that 471 * pindex is not found in the vm_object, returns a vm_page that is 472 * adjacent to the pindex, coming before or after it. 473 */ 474vm_page_t 475vm_page_splay(vm_pindex_t pindex, vm_page_t root) 476{ 477 struct vm_page dummy; 478 vm_page_t lefttreemax, righttreemin, y; 479 480 if (root == NULL) 481 return (root); 482 lefttreemax = righttreemin = &dummy; 483 for (;; root = y) { 484 if (pindex < root->pindex) { 485 if ((y = root->left) == NULL) 486 break; 487 if (pindex < y->pindex) { 488 /* Rotate right. */ 489 root->left = y->right; 490 y->right = root; 491 root = y; 492 if ((y = root->left) == NULL) 493 break; 494 } 495 /* Link into the new root's right tree. */ 496 righttreemin->left = root; 497 righttreemin = root; 498 } else if (pindex > root->pindex) { 499 if ((y = root->right) == NULL) 500 break; 501 if (pindex > y->pindex) { 502 /* Rotate left. */ 503 root->right = y->left; 504 y->left = root; 505 root = y; 506 if ((y = root->right) == NULL) 507 break; 508 } 509 /* Link into the new root's left tree. */ 510 lefttreemax->right = root; 511 lefttreemax = root; 512 } else 513 break; 514 } 515 /* Assemble the new root. */ 516 lefttreemax->right = root->left; 517 righttreemin->left = root->right; 518 root->left = dummy.right; 519 root->right = dummy.left; 520 return (root); 521} 522 523/* 524 * vm_page_insert: [ internal use only ] 525 * 526 * Inserts the given mem entry into the object and object list. 527 * 528 * The pagetables are not updated but will presumably fault the page 529 * in if necessary, or if a kernel page the caller will at some point 530 * enter the page into the kernel's pmap. We are not allowed to block 531 * here so we *can't* do this anyway. 532 * 533 * The object and page must be locked. 534 * This routine may not block. 535 */ 536void 537vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 538{ 539 vm_page_t root; 540 541 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 542 if (m->object != NULL) 543 panic("vm_page_insert: page already inserted"); 544 545 /* 546 * Record the object/offset pair in this page 547 */ 548 m->object = object; 549 m->pindex = pindex; 550 551 /* 552 * Now link into the object's ordered list of backed pages. 553 */ 554 root = object->root; 555 if (root == NULL) { 556 m->left = NULL; 557 m->right = NULL; 558 TAILQ_INSERT_TAIL(&object->memq, m, listq); 559 } else { 560 root = vm_page_splay(pindex, root); 561 if (pindex < root->pindex) { 562 m->left = root->left; 563 m->right = root; 564 root->left = NULL; 565 TAILQ_INSERT_BEFORE(root, m, listq); 566 } else if (pindex == root->pindex) 567 panic("vm_page_insert: offset already allocated"); 568 else { 569 m->right = root->right; 570 m->left = root; 571 root->right = NULL; 572 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 573 } 574 } 575 object->root = m; 576 object->generation++; 577 578 /* 579 * show that the object has one more resident page. 580 */ 581 object->resident_page_count++; 582 583 /* 584 * Since we are inserting a new and possibly dirty page, 585 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 586 */ 587 if (m->flags & PG_WRITEABLE) 588 vm_object_set_writeable_dirty(object); 589} 590 591/* 592 * vm_page_remove: 593 * NOTE: used by device pager as well -wfj 594 * 595 * Removes the given mem entry from the object/offset-page 596 * table and the object page list, but do not invalidate/terminate 597 * the backing store. 598 * 599 * The object and page must be locked. 600 * The underlying pmap entry (if any) is NOT removed here. 601 * This routine may not block. 602 */ 603void 604vm_page_remove(vm_page_t m) 605{ 606 vm_object_t object; 607 vm_page_t root; 608 609 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 610 if (m->object == NULL) 611 return; 612 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 613 if ((m->flags & PG_BUSY) == 0) { 614 panic("vm_page_remove: page not busy"); 615 } 616 617 /* 618 * Basically destroy the page. 619 */ 620 vm_page_wakeup(m); 621 622 object = m->object; 623 624 /* 625 * Now remove from the object's list of backed pages. 626 */ 627 if (m != object->root) 628 vm_page_splay(m->pindex, object->root); 629 if (m->left == NULL) 630 root = m->right; 631 else { 632 root = vm_page_splay(m->pindex, m->left); 633 root->right = m->right; 634 } 635 object->root = root; 636 TAILQ_REMOVE(&object->memq, m, listq); 637 638 /* 639 * And show that the object has one fewer resident page. 640 */ 641 object->resident_page_count--; 642 object->generation++; 643 644 m->object = NULL; 645} 646 647/* 648 * vm_page_lookup: 649 * 650 * Returns the page associated with the object/offset 651 * pair specified; if none is found, NULL is returned. 652 * 653 * The object must be locked. 654 * This routine may not block. 655 * This is a critical path routine 656 */ 657vm_page_t 658vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 659{ 660 vm_page_t m; 661 662 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 663 if ((m = object->root) != NULL && m->pindex != pindex) { 664 m = vm_page_splay(pindex, m); 665 if ((object->root = m)->pindex != pindex) 666 m = NULL; 667 } 668 return (m); 669} 670 671/* 672 * vm_page_rename: 673 * 674 * Move the given memory entry from its 675 * current object to the specified target object/offset. 676 * 677 * The object must be locked. 678 * This routine may not block. 679 * 680 * Note: swap associated with the page must be invalidated by the move. We 681 * have to do this for several reasons: (1) we aren't freeing the 682 * page, (2) we are dirtying the page, (3) the VM system is probably 683 * moving the page from object A to B, and will then later move 684 * the backing store from A to B and we can't have a conflict. 685 * 686 * Note: we *always* dirty the page. It is necessary both for the 687 * fact that we moved it, and because we may be invalidating 688 * swap. If the page is on the cache, we have to deactivate it 689 * or vm_page_dirty() will panic. Dirty pages are not allowed 690 * on the cache. 691 */ 692void 693vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 694{ 695 696 vm_page_remove(m); 697 vm_page_insert(m, new_object, new_pindex); 698 if (m->queue - m->pc == PQ_CACHE) 699 vm_page_deactivate(m); 700 vm_page_dirty(m); 701} 702 703/* 704 * vm_page_select_cache: 705 * 706 * Find a page on the cache queue with color optimization. As pages 707 * might be found, but not applicable, they are deactivated. This 708 * keeps us from using potentially busy cached pages. 709 * 710 * This routine may not block. 711 */ 712vm_page_t 713vm_page_select_cache(int color) 714{ 715 vm_page_t m; 716 717 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 718 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 719 if ((m->flags & PG_BUSY) == 0 && m->busy == 0 && 720 m->hold_count == 0 && (VM_OBJECT_TRYLOCK(m->object) || 721 VM_OBJECT_LOCKED(m->object))) { 722 KASSERT(m->dirty == 0, 723 ("Found dirty cache page %p", m)); 724 KASSERT(!pmap_page_is_mapped(m), 725 ("Found mapped cache page %p", m)); 726 KASSERT((m->flags & PG_UNMANAGED) == 0, 727 ("Found unmanaged cache page %p", m)); 728 KASSERT(m->wire_count == 0, 729 ("Found wired cache page %p", m)); 730 break; 731 } 732 vm_page_deactivate(m); 733 } 734 return (m); 735} 736 737/* 738 * vm_page_alloc: 739 * 740 * Allocate and return a memory cell associated 741 * with this VM object/offset pair. 742 * 743 * page_req classes: 744 * VM_ALLOC_NORMAL normal process request 745 * VM_ALLOC_SYSTEM system *really* needs a page 746 * VM_ALLOC_INTERRUPT interrupt time request 747 * VM_ALLOC_ZERO zero page 748 * 749 * This routine may not block. 750 * 751 * Additional special handling is required when called from an 752 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 753 * the page cache in this case. 754 */ 755vm_page_t 756vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 757{ 758 vm_object_t m_object; 759 vm_page_t m = NULL; 760 int color, flags, page_req; 761 762 page_req = req & VM_ALLOC_CLASS_MASK; 763 764 if ((req & VM_ALLOC_NOOBJ) == 0) { 765 KASSERT(object != NULL, 766 ("vm_page_alloc: NULL object.")); 767 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 768 color = (pindex + object->pg_color) & PQ_L2_MASK; 769 } else 770 color = pindex & PQ_L2_MASK; 771 772 /* 773 * The pager is allowed to eat deeper into the free page list. 774 */ 775 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 776 page_req = VM_ALLOC_SYSTEM; 777 }; 778 779loop: 780 mtx_lock_spin(&vm_page_queue_free_mtx); 781 if (cnt.v_free_count > cnt.v_free_reserved || 782 (page_req == VM_ALLOC_SYSTEM && 783 cnt.v_cache_count == 0 && 784 cnt.v_free_count > cnt.v_interrupt_free_min) || 785 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 786 /* 787 * Allocate from the free queue if the number of free pages 788 * exceeds the minimum for the request class. 789 */ 790 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 791 } else if (page_req != VM_ALLOC_INTERRUPT) { 792 mtx_unlock_spin(&vm_page_queue_free_mtx); 793 /* 794 * Allocatable from cache (non-interrupt only). On success, 795 * we must free the page and try again, thus ensuring that 796 * cnt.v_*_free_min counters are replenished. 797 */ 798 vm_page_lock_queues(); 799 if ((m = vm_page_select_cache(color)) == NULL) { 800#if defined(DIAGNOSTIC) 801 if (cnt.v_cache_count > 0) 802 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 803#endif 804 vm_page_unlock_queues(); 805 atomic_add_int(&vm_pageout_deficit, 1); 806 pagedaemon_wakeup(); 807 return (NULL); 808 } 809 m_object = m->object; 810 VM_OBJECT_LOCK_ASSERT(m_object, MA_OWNED); 811 vm_page_busy(m); 812 vm_page_free(m); 813 vm_page_unlock_queues(); 814 if (m_object != object) 815 VM_OBJECT_UNLOCK(m_object); 816 goto loop; 817 } else { 818 /* 819 * Not allocatable from cache from interrupt, give up. 820 */ 821 mtx_unlock_spin(&vm_page_queue_free_mtx); 822 atomic_add_int(&vm_pageout_deficit, 1); 823 pagedaemon_wakeup(); 824 return (NULL); 825 } 826 827 /* 828 * At this point we had better have found a good page. 829 */ 830 831 KASSERT( 832 m != NULL, 833 ("vm_page_alloc(): missing page on free queue") 834 ); 835 836 /* 837 * Remove from free queue 838 */ 839 vm_pageq_remove_nowakeup(m); 840 841 /* 842 * Initialize structure. Only the PG_ZERO flag is inherited. 843 */ 844 flags = PG_BUSY; 845 if (m->flags & PG_ZERO) { 846 vm_page_zero_count--; 847 if (req & VM_ALLOC_ZERO) 848 flags = PG_ZERO | PG_BUSY; 849 } 850 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 851 flags &= ~PG_BUSY; 852 m->flags = flags; 853 if (req & VM_ALLOC_WIRED) { 854 atomic_add_int(&cnt.v_wire_count, 1); 855 m->wire_count = 1; 856 } else 857 m->wire_count = 0; 858 m->hold_count = 0; 859 m->act_count = 0; 860 m->busy = 0; 861 m->valid = 0; 862 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 863 mtx_unlock_spin(&vm_page_queue_free_mtx); 864 865 if ((req & VM_ALLOC_NOOBJ) == 0) 866 vm_page_insert(m, object, pindex); 867 else 868 m->pindex = pindex; 869 870 /* 871 * Don't wakeup too often - wakeup the pageout daemon when 872 * we would be nearly out of memory. 873 */ 874 if (vm_paging_needed()) 875 pagedaemon_wakeup(); 876 877 return (m); 878} 879 880/* 881 * vm_wait: (also see VM_WAIT macro) 882 * 883 * Block until free pages are available for allocation 884 * - Called in various places before memory allocations. 885 */ 886void 887vm_wait(void) 888{ 889 890 vm_page_lock_queues(); 891 if (curproc == pageproc) { 892 vm_pageout_pages_needed = 1; 893 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 894 PDROP | PSWP, "VMWait", 0); 895 } else { 896 if (!vm_pages_needed) { 897 vm_pages_needed = 1; 898 wakeup(&vm_pages_needed); 899 } 900 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 901 "vmwait", 0); 902 } 903} 904 905/* 906 * vm_waitpfault: (also see VM_WAITPFAULT macro) 907 * 908 * Block until free pages are available for allocation 909 * - Called only in vm_fault so that processes page faulting 910 * can be easily tracked. 911 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 912 * processes will be able to grab memory first. Do not change 913 * this balance without careful testing first. 914 */ 915void 916vm_waitpfault(void) 917{ 918 919 vm_page_lock_queues(); 920 if (!vm_pages_needed) { 921 vm_pages_needed = 1; 922 wakeup(&vm_pages_needed); 923 } 924 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 925 "pfault", 0); 926} 927 928/* 929 * vm_page_activate: 930 * 931 * Put the specified page on the active list (if appropriate). 932 * Ensure that act_count is at least ACT_INIT but do not otherwise 933 * mess with it. 934 * 935 * The page queues must be locked. 936 * This routine may not block. 937 */ 938void 939vm_page_activate(vm_page_t m) 940{ 941 942 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 943 if (m->queue != PQ_ACTIVE) { 944 if ((m->queue - m->pc) == PQ_CACHE) 945 cnt.v_reactivated++; 946 vm_pageq_remove(m); 947 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 948 if (m->act_count < ACT_INIT) 949 m->act_count = ACT_INIT; 950 vm_pageq_enqueue(PQ_ACTIVE, m); 951 } 952 } else { 953 if (m->act_count < ACT_INIT) 954 m->act_count = ACT_INIT; 955 } 956} 957 958/* 959 * vm_page_free_wakeup: 960 * 961 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 962 * routine is called when a page has been added to the cache or free 963 * queues. 964 * 965 * The page queues must be locked. 966 * This routine may not block. 967 */ 968static __inline void 969vm_page_free_wakeup(void) 970{ 971 972 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 973 /* 974 * if pageout daemon needs pages, then tell it that there are 975 * some free. 976 */ 977 if (vm_pageout_pages_needed && 978 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 979 wakeup(&vm_pageout_pages_needed); 980 vm_pageout_pages_needed = 0; 981 } 982 /* 983 * wakeup processes that are waiting on memory if we hit a 984 * high water mark. And wakeup scheduler process if we have 985 * lots of memory. this process will swapin processes. 986 */ 987 if (vm_pages_needed && !vm_page_count_min()) { 988 vm_pages_needed = 0; 989 wakeup(&cnt.v_free_count); 990 } 991} 992 993/* 994 * vm_page_free_toq: 995 * 996 * Returns the given page to the PQ_FREE list, 997 * disassociating it with any VM object. 998 * 999 * Object and page must be locked prior to entry. 1000 * This routine may not block. 1001 */ 1002 1003void 1004vm_page_free_toq(vm_page_t m) 1005{ 1006 struct vpgqueues *pq; 1007 vm_object_t object = m->object; 1008 1009 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1010 cnt.v_tfree++; 1011 1012 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1013 printf( 1014 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1015 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1016 m->hold_count); 1017 if ((m->queue - m->pc) == PQ_FREE) 1018 panic("vm_page_free: freeing free page"); 1019 else 1020 panic("vm_page_free: freeing busy page"); 1021 } 1022 1023 /* 1024 * unqueue, then remove page. Note that we cannot destroy 1025 * the page here because we do not want to call the pager's 1026 * callback routine until after we've put the page on the 1027 * appropriate free queue. 1028 */ 1029 vm_pageq_remove_nowakeup(m); 1030 vm_page_remove(m); 1031 1032 /* 1033 * If fictitious remove object association and 1034 * return, otherwise delay object association removal. 1035 */ 1036 if ((m->flags & PG_FICTITIOUS) != 0) { 1037 return; 1038 } 1039 1040 m->valid = 0; 1041 vm_page_undirty(m); 1042 1043 if (m->wire_count != 0) { 1044 if (m->wire_count > 1) { 1045 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1046 m->wire_count, (long)m->pindex); 1047 } 1048 panic("vm_page_free: freeing wired page"); 1049 } 1050 1051 /* 1052 * If we've exhausted the object's resident pages we want to free 1053 * it up. 1054 */ 1055 if (object && 1056 (object->type == OBJT_VNODE) && 1057 ((object->flags & OBJ_DEAD) == 0) 1058 ) { 1059 struct vnode *vp = (struct vnode *)object->handle; 1060 1061 if (vp) { 1062 VI_LOCK(vp); 1063 if (VSHOULDFREE(vp)) 1064 vfree(vp); 1065 VI_UNLOCK(vp); 1066 } 1067 } 1068 1069 /* 1070 * Clear the UNMANAGED flag when freeing an unmanaged page. 1071 */ 1072 if (m->flags & PG_UNMANAGED) { 1073 m->flags &= ~PG_UNMANAGED; 1074 } 1075 1076 if (m->hold_count != 0) { 1077 m->flags &= ~PG_ZERO; 1078 m->queue = PQ_HOLD; 1079 } else 1080 m->queue = PQ_FREE + m->pc; 1081 pq = &vm_page_queues[m->queue]; 1082 mtx_lock_spin(&vm_page_queue_free_mtx); 1083 pq->lcnt++; 1084 ++(*pq->cnt); 1085 1086 /* 1087 * Put zero'd pages on the end ( where we look for zero'd pages 1088 * first ) and non-zerod pages at the head. 1089 */ 1090 if (m->flags & PG_ZERO) { 1091 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1092 ++vm_page_zero_count; 1093 } else { 1094 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1095 } 1096 mtx_unlock_spin(&vm_page_queue_free_mtx); 1097 vm_page_free_wakeup(); 1098} 1099 1100/* 1101 * vm_page_unmanage: 1102 * 1103 * Prevent PV management from being done on the page. The page is 1104 * removed from the paging queues as if it were wired, and as a 1105 * consequence of no longer being managed the pageout daemon will not 1106 * touch it (since there is no way to locate the pte mappings for the 1107 * page). madvise() calls that mess with the pmap will also no longer 1108 * operate on the page. 1109 * 1110 * Beyond that the page is still reasonably 'normal'. Freeing the page 1111 * will clear the flag. 1112 * 1113 * This routine is used by OBJT_PHYS objects - objects using unswappable 1114 * physical memory as backing store rather then swap-backed memory and 1115 * will eventually be extended to support 4MB unmanaged physical 1116 * mappings. 1117 */ 1118void 1119vm_page_unmanage(vm_page_t m) 1120{ 1121 1122 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1123 if ((m->flags & PG_UNMANAGED) == 0) { 1124 if (m->wire_count == 0) 1125 vm_pageq_remove(m); 1126 } 1127 vm_page_flag_set(m, PG_UNMANAGED); 1128} 1129 1130/* 1131 * vm_page_wire: 1132 * 1133 * Mark this page as wired down by yet 1134 * another map, removing it from paging queues 1135 * as necessary. 1136 * 1137 * The page queues must be locked. 1138 * This routine may not block. 1139 */ 1140void 1141vm_page_wire(vm_page_t m) 1142{ 1143 1144 /* 1145 * Only bump the wire statistics if the page is not already wired, 1146 * and only unqueue the page if it is on some queue (if it is unmanaged 1147 * it is already off the queues). 1148 */ 1149 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1150 if (m->flags & PG_FICTITIOUS) 1151 return; 1152 if (m->wire_count == 0) { 1153 if ((m->flags & PG_UNMANAGED) == 0) 1154 vm_pageq_remove(m); 1155 atomic_add_int(&cnt.v_wire_count, 1); 1156 } 1157 m->wire_count++; 1158 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1159} 1160 1161/* 1162 * vm_page_unwire: 1163 * 1164 * Release one wiring of this page, potentially 1165 * enabling it to be paged again. 1166 * 1167 * Many pages placed on the inactive queue should actually go 1168 * into the cache, but it is difficult to figure out which. What 1169 * we do instead, if the inactive target is well met, is to put 1170 * clean pages at the head of the inactive queue instead of the tail. 1171 * This will cause them to be moved to the cache more quickly and 1172 * if not actively re-referenced, freed more quickly. If we just 1173 * stick these pages at the end of the inactive queue, heavy filesystem 1174 * meta-data accesses can cause an unnecessary paging load on memory bound 1175 * processes. This optimization causes one-time-use metadata to be 1176 * reused more quickly. 1177 * 1178 * BUT, if we are in a low-memory situation we have no choice but to 1179 * put clean pages on the cache queue. 1180 * 1181 * A number of routines use vm_page_unwire() to guarantee that the page 1182 * will go into either the inactive or active queues, and will NEVER 1183 * be placed in the cache - for example, just after dirtying a page. 1184 * dirty pages in the cache are not allowed. 1185 * 1186 * The page queues must be locked. 1187 * This routine may not block. 1188 */ 1189void 1190vm_page_unwire(vm_page_t m, int activate) 1191{ 1192 1193 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1194 if (m->flags & PG_FICTITIOUS) 1195 return; 1196 if (m->wire_count > 0) { 1197 m->wire_count--; 1198 if (m->wire_count == 0) { 1199 atomic_subtract_int(&cnt.v_wire_count, 1); 1200 if (m->flags & PG_UNMANAGED) { 1201 ; 1202 } else if (activate) 1203 vm_pageq_enqueue(PQ_ACTIVE, m); 1204 else { 1205 vm_page_flag_clear(m, PG_WINATCFLS); 1206 vm_pageq_enqueue(PQ_INACTIVE, m); 1207 } 1208 } 1209 } else { 1210 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1211 } 1212} 1213 1214 1215/* 1216 * Move the specified page to the inactive queue. If the page has 1217 * any associated swap, the swap is deallocated. 1218 * 1219 * Normally athead is 0 resulting in LRU operation. athead is set 1220 * to 1 if we want this page to be 'as if it were placed in the cache', 1221 * except without unmapping it from the process address space. 1222 * 1223 * This routine may not block. 1224 */ 1225static __inline void 1226_vm_page_deactivate(vm_page_t m, int athead) 1227{ 1228 1229 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1230 1231 /* 1232 * Ignore if already inactive. 1233 */ 1234 if (m->queue == PQ_INACTIVE) 1235 return; 1236 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1237 if ((m->queue - m->pc) == PQ_CACHE) 1238 cnt.v_reactivated++; 1239 vm_page_flag_clear(m, PG_WINATCFLS); 1240 vm_pageq_remove(m); 1241 if (athead) 1242 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1243 else 1244 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1245 m->queue = PQ_INACTIVE; 1246 vm_page_queues[PQ_INACTIVE].lcnt++; 1247 cnt.v_inactive_count++; 1248 } 1249} 1250 1251void 1252vm_page_deactivate(vm_page_t m) 1253{ 1254 _vm_page_deactivate(m, 0); 1255} 1256 1257/* 1258 * vm_page_try_to_cache: 1259 * 1260 * Returns 0 on failure, 1 on success 1261 */ 1262int 1263vm_page_try_to_cache(vm_page_t m) 1264{ 1265 1266 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1267 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1268 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1269 return (0); 1270 } 1271 pmap_remove_all(m); 1272 if (m->dirty) 1273 return (0); 1274 vm_page_cache(m); 1275 return (1); 1276} 1277 1278/* 1279 * vm_page_try_to_free() 1280 * 1281 * Attempt to free the page. If we cannot free it, we do nothing. 1282 * 1 is returned on success, 0 on failure. 1283 */ 1284int 1285vm_page_try_to_free(vm_page_t m) 1286{ 1287 1288 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1289 if (m->object != NULL) 1290 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1291 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1292 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1293 return (0); 1294 } 1295 pmap_remove_all(m); 1296 if (m->dirty) 1297 return (0); 1298 vm_page_busy(m); 1299 vm_page_free(m); 1300 return (1); 1301} 1302 1303/* 1304 * vm_page_cache 1305 * 1306 * Put the specified page onto the page cache queue (if appropriate). 1307 * 1308 * This routine may not block. 1309 */ 1310void 1311vm_page_cache(vm_page_t m) 1312{ 1313 1314 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1315 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1316 m->hold_count || m->wire_count) { 1317 printf("vm_page_cache: attempting to cache busy page\n"); 1318 return; 1319 } 1320 if ((m->queue - m->pc) == PQ_CACHE) 1321 return; 1322 1323 /* 1324 * Remove all pmaps and indicate that the page is not 1325 * writeable or mapped. 1326 */ 1327 pmap_remove_all(m); 1328 if (m->dirty != 0) { 1329 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1330 (long)m->pindex); 1331 } 1332 vm_pageq_remove_nowakeup(m); 1333 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1334 vm_page_free_wakeup(); 1335} 1336 1337/* 1338 * vm_page_dontneed 1339 * 1340 * Cache, deactivate, or do nothing as appropriate. This routine 1341 * is typically used by madvise() MADV_DONTNEED. 1342 * 1343 * Generally speaking we want to move the page into the cache so 1344 * it gets reused quickly. However, this can result in a silly syndrome 1345 * due to the page recycling too quickly. Small objects will not be 1346 * fully cached. On the otherhand, if we move the page to the inactive 1347 * queue we wind up with a problem whereby very large objects 1348 * unnecessarily blow away our inactive and cache queues. 1349 * 1350 * The solution is to move the pages based on a fixed weighting. We 1351 * either leave them alone, deactivate them, or move them to the cache, 1352 * where moving them to the cache has the highest weighting. 1353 * By forcing some pages into other queues we eventually force the 1354 * system to balance the queues, potentially recovering other unrelated 1355 * space from active. The idea is to not force this to happen too 1356 * often. 1357 */ 1358void 1359vm_page_dontneed(vm_page_t m) 1360{ 1361 static int dnweight; 1362 int dnw; 1363 int head; 1364 1365 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1366 dnw = ++dnweight; 1367 1368 /* 1369 * occassionally leave the page alone 1370 */ 1371 if ((dnw & 0x01F0) == 0 || 1372 m->queue == PQ_INACTIVE || 1373 m->queue - m->pc == PQ_CACHE 1374 ) { 1375 if (m->act_count >= ACT_INIT) 1376 --m->act_count; 1377 return; 1378 } 1379 1380 if (m->dirty == 0 && pmap_is_modified(m)) 1381 vm_page_dirty(m); 1382 1383 if (m->dirty || (dnw & 0x0070) == 0) { 1384 /* 1385 * Deactivate the page 3 times out of 32. 1386 */ 1387 head = 0; 1388 } else { 1389 /* 1390 * Cache the page 28 times out of every 32. Note that 1391 * the page is deactivated instead of cached, but placed 1392 * at the head of the queue instead of the tail. 1393 */ 1394 head = 1; 1395 } 1396 _vm_page_deactivate(m, head); 1397} 1398 1399/* 1400 * Grab a page, waiting until we are waken up due to the page 1401 * changing state. We keep on waiting, if the page continues 1402 * to be in the object. If the page doesn't exist, first allocate it 1403 * and then conditionally zero it. 1404 * 1405 * This routine may block. 1406 */ 1407vm_page_t 1408vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1409{ 1410 vm_page_t m; 1411 1412 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1413retrylookup: 1414 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1415 vm_page_lock_queues(); 1416 if (m->busy || (m->flags & PG_BUSY)) { 1417 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1418 VM_OBJECT_UNLOCK(object); 1419 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0); 1420 VM_OBJECT_LOCK(object); 1421 if ((allocflags & VM_ALLOC_RETRY) == 0) 1422 return (NULL); 1423 goto retrylookup; 1424 } else { 1425 if (allocflags & VM_ALLOC_WIRED) 1426 vm_page_wire(m); 1427 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1428 vm_page_busy(m); 1429 vm_page_unlock_queues(); 1430 return (m); 1431 } 1432 } 1433 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1434 if (m == NULL) { 1435 VM_OBJECT_UNLOCK(object); 1436 VM_WAIT; 1437 VM_OBJECT_LOCK(object); 1438 if ((allocflags & VM_ALLOC_RETRY) == 0) 1439 return (NULL); 1440 goto retrylookup; 1441 } 1442 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1443 pmap_zero_page(m); 1444 return (m); 1445} 1446 1447/* 1448 * Mapping function for valid bits or for dirty bits in 1449 * a page. May not block. 1450 * 1451 * Inputs are required to range within a page. 1452 */ 1453__inline int 1454vm_page_bits(int base, int size) 1455{ 1456 int first_bit; 1457 int last_bit; 1458 1459 KASSERT( 1460 base + size <= PAGE_SIZE, 1461 ("vm_page_bits: illegal base/size %d/%d", base, size) 1462 ); 1463 1464 if (size == 0) /* handle degenerate case */ 1465 return (0); 1466 1467 first_bit = base >> DEV_BSHIFT; 1468 last_bit = (base + size - 1) >> DEV_BSHIFT; 1469 1470 return ((2 << last_bit) - (1 << first_bit)); 1471} 1472 1473/* 1474 * vm_page_set_validclean: 1475 * 1476 * Sets portions of a page valid and clean. The arguments are expected 1477 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1478 * of any partial chunks touched by the range. The invalid portion of 1479 * such chunks will be zero'd. 1480 * 1481 * This routine may not block. 1482 * 1483 * (base + size) must be less then or equal to PAGE_SIZE. 1484 */ 1485void 1486vm_page_set_validclean(vm_page_t m, int base, int size) 1487{ 1488 int pagebits; 1489 int frag; 1490 int endoff; 1491 1492 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1493 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1494 if (size == 0) /* handle degenerate case */ 1495 return; 1496 1497 /* 1498 * If the base is not DEV_BSIZE aligned and the valid 1499 * bit is clear, we have to zero out a portion of the 1500 * first block. 1501 */ 1502 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1503 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1504 pmap_zero_page_area(m, frag, base - frag); 1505 1506 /* 1507 * If the ending offset is not DEV_BSIZE aligned and the 1508 * valid bit is clear, we have to zero out a portion of 1509 * the last block. 1510 */ 1511 endoff = base + size; 1512 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1513 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1514 pmap_zero_page_area(m, endoff, 1515 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1516 1517 /* 1518 * Set valid, clear dirty bits. If validating the entire 1519 * page we can safely clear the pmap modify bit. We also 1520 * use this opportunity to clear the PG_NOSYNC flag. If a process 1521 * takes a write fault on a MAP_NOSYNC memory area the flag will 1522 * be set again. 1523 * 1524 * We set valid bits inclusive of any overlap, but we can only 1525 * clear dirty bits for DEV_BSIZE chunks that are fully within 1526 * the range. 1527 */ 1528 pagebits = vm_page_bits(base, size); 1529 m->valid |= pagebits; 1530#if 0 /* NOT YET */ 1531 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1532 frag = DEV_BSIZE - frag; 1533 base += frag; 1534 size -= frag; 1535 if (size < 0) 1536 size = 0; 1537 } 1538 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1539#endif 1540 m->dirty &= ~pagebits; 1541 if (base == 0 && size == PAGE_SIZE) { 1542 pmap_clear_modify(m); 1543 vm_page_flag_clear(m, PG_NOSYNC); 1544 } 1545} 1546 1547void 1548vm_page_clear_dirty(vm_page_t m, int base, int size) 1549{ 1550 1551 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1552 m->dirty &= ~vm_page_bits(base, size); 1553} 1554 1555/* 1556 * vm_page_set_invalid: 1557 * 1558 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1559 * valid and dirty bits for the effected areas are cleared. 1560 * 1561 * May not block. 1562 */ 1563void 1564vm_page_set_invalid(vm_page_t m, int base, int size) 1565{ 1566 int bits; 1567 1568 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1569 bits = vm_page_bits(base, size); 1570 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1571 m->valid &= ~bits; 1572 m->dirty &= ~bits; 1573 m->object->generation++; 1574} 1575 1576/* 1577 * vm_page_zero_invalid() 1578 * 1579 * The kernel assumes that the invalid portions of a page contain 1580 * garbage, but such pages can be mapped into memory by user code. 1581 * When this occurs, we must zero out the non-valid portions of the 1582 * page so user code sees what it expects. 1583 * 1584 * Pages are most often semi-valid when the end of a file is mapped 1585 * into memory and the file's size is not page aligned. 1586 */ 1587void 1588vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1589{ 1590 int b; 1591 int i; 1592 1593 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1594 /* 1595 * Scan the valid bits looking for invalid sections that 1596 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1597 * valid bit may be set ) have already been zerod by 1598 * vm_page_set_validclean(). 1599 */ 1600 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1601 if (i == (PAGE_SIZE / DEV_BSIZE) || 1602 (m->valid & (1 << i)) 1603 ) { 1604 if (i > b) { 1605 pmap_zero_page_area(m, 1606 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1607 } 1608 b = i + 1; 1609 } 1610 } 1611 1612 /* 1613 * setvalid is TRUE when we can safely set the zero'd areas 1614 * as being valid. We can do this if there are no cache consistancy 1615 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1616 */ 1617 if (setvalid) 1618 m->valid = VM_PAGE_BITS_ALL; 1619} 1620 1621/* 1622 * vm_page_is_valid: 1623 * 1624 * Is (partial) page valid? Note that the case where size == 0 1625 * will return FALSE in the degenerate case where the page is 1626 * entirely invalid, and TRUE otherwise. 1627 * 1628 * May not block. 1629 */ 1630int 1631vm_page_is_valid(vm_page_t m, int base, int size) 1632{ 1633 int bits = vm_page_bits(base, size); 1634 1635 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1636 if (m->valid && ((m->valid & bits) == bits)) 1637 return 1; 1638 else 1639 return 0; 1640} 1641 1642/* 1643 * update dirty bits from pmap/mmu. May not block. 1644 */ 1645void 1646vm_page_test_dirty(vm_page_t m) 1647{ 1648 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1649 vm_page_dirty(m); 1650 } 1651} 1652 1653int so_zerocp_fullpage = 0; 1654 1655void 1656vm_page_cowfault(vm_page_t m) 1657{ 1658 vm_page_t mnew; 1659 vm_object_t object; 1660 vm_pindex_t pindex; 1661 1662 object = m->object; 1663 pindex = m->pindex; 1664 1665 retry_alloc: 1666 vm_page_busy(m); 1667 vm_page_remove(m); 1668 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 1669 if (mnew == NULL) { 1670 vm_page_insert(m, object, pindex); 1671 vm_page_unlock_queues(); 1672 VM_OBJECT_UNLOCK(object); 1673 VM_WAIT; 1674 VM_OBJECT_LOCK(object); 1675 vm_page_lock_queues(); 1676 goto retry_alloc; 1677 } 1678 1679 if (m->cow == 0) { 1680 /* 1681 * check to see if we raced with an xmit complete when 1682 * waiting to allocate a page. If so, put things back 1683 * the way they were 1684 */ 1685 vm_page_free(mnew); 1686 vm_page_insert(m, object, pindex); 1687 } else { /* clear COW & copy page */ 1688 if (!so_zerocp_fullpage) 1689 pmap_copy_page(m, mnew); 1690 mnew->valid = VM_PAGE_BITS_ALL; 1691 vm_page_dirty(mnew); 1692 vm_page_flag_clear(mnew, PG_BUSY); 1693 } 1694} 1695 1696void 1697vm_page_cowclear(vm_page_t m) 1698{ 1699 1700 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1701 if (m->cow) { 1702 m->cow--; 1703 /* 1704 * let vm_fault add back write permission lazily 1705 */ 1706 } 1707 /* 1708 * sf_buf_free() will free the page, so we needn't do it here 1709 */ 1710} 1711 1712void 1713vm_page_cowsetup(vm_page_t m) 1714{ 1715 1716 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1717 m->cow++; 1718 pmap_page_protect(m, VM_PROT_READ); 1719} 1720 1721#include "opt_ddb.h" 1722#ifdef DDB 1723#include <sys/kernel.h> 1724 1725#include <ddb/ddb.h> 1726 1727DB_SHOW_COMMAND(page, vm_page_print_page_info) 1728{ 1729 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1730 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1731 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1732 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1733 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1734 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1735 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1736 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1737 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1738 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1739} 1740 1741DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1742{ 1743 int i; 1744 db_printf("PQ_FREE:"); 1745 for (i = 0; i < PQ_L2_SIZE; i++) { 1746 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1747 } 1748 db_printf("\n"); 1749 1750 db_printf("PQ_CACHE:"); 1751 for (i = 0; i < PQ_L2_SIZE; i++) { 1752 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1753 } 1754 db_printf("\n"); 1755 1756 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1757 vm_page_queues[PQ_ACTIVE].lcnt, 1758 vm_page_queues[PQ_INACTIVE].lcnt); 1759} 1760#endif /* DDB */ 1761