vm_page.c revision 129057
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 129057 2004-05-09 01:00:15Z 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_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 209 MTX_SPIN); 210 211 /* 212 * Initialize the queue headers for the free queue, the active queue 213 * and the inactive queue. 214 */ 215 vm_pageq_init(); 216 217 /* 218 * Allocate memory for use when boot strapping the kernel memory 219 * allocator. 220 */ 221 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 222 new_end = end - bootpages; 223 new_end = trunc_page(new_end); 224 mapped = pmap_map(&vaddr, new_end, end, 225 VM_PROT_READ | VM_PROT_WRITE); 226 bzero((caddr_t) mapped, end - new_end); 227 uma_startup((caddr_t)mapped); 228 229 /* 230 * Compute the number of pages of memory that will be available for 231 * use (taking into account the overhead of a page structure per 232 * page). 233 */ 234 first_page = phys_avail[0] / PAGE_SIZE; 235 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 236 npages = (total - (page_range * sizeof(struct vm_page)) - 237 (end - new_end)) / PAGE_SIZE; 238 end = new_end; 239 240 /* 241 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 242 */ 243 vaddr += PAGE_SIZE; 244 245 /* 246 * Initialize the mem entry structures now, and put them in the free 247 * queue. 248 */ 249 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 250 mapped = pmap_map(&vaddr, new_end, end, 251 VM_PROT_READ | VM_PROT_WRITE); 252 vm_page_array = (vm_page_t) mapped; 253 phys_avail[biggestone + 1] = new_end; 254 255 /* 256 * Clear all of the page structures 257 */ 258 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 259 vm_page_array_size = page_range; 260 261 /* 262 * Construct the free queue(s) in descending order (by physical 263 * address) so that the first 16MB of physical memory is allocated 264 * last rather than first. On large-memory machines, this avoids 265 * the exhaustion of low physical memory before isa_dmainit has run. 266 */ 267 cnt.v_page_count = 0; 268 cnt.v_free_count = 0; 269 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 270 pa = phys_avail[i]; 271 last_pa = phys_avail[i + 1]; 272 while (pa < last_pa && npages-- > 0) { 273 vm_pageq_add_new_page(pa); 274 pa += PAGE_SIZE; 275 } 276 } 277 return (vaddr); 278} 279 280void 281vm_page_flag_set(vm_page_t m, unsigned short bits) 282{ 283 284 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 285 m->flags |= bits; 286} 287 288void 289vm_page_flag_clear(vm_page_t m, unsigned short bits) 290{ 291 292 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 293 m->flags &= ~bits; 294} 295 296void 297vm_page_busy(vm_page_t m) 298{ 299 KASSERT((m->flags & PG_BUSY) == 0, 300 ("vm_page_busy: page already busy!!!")); 301 vm_page_flag_set(m, PG_BUSY); 302} 303 304/* 305 * vm_page_flash: 306 * 307 * wakeup anyone waiting for the page. 308 */ 309void 310vm_page_flash(vm_page_t m) 311{ 312 if (m->flags & PG_WANTED) { 313 vm_page_flag_clear(m, PG_WANTED); 314 wakeup(m); 315 } 316} 317 318/* 319 * vm_page_wakeup: 320 * 321 * clear the PG_BUSY flag and wakeup anyone waiting for the 322 * page. 323 * 324 */ 325void 326vm_page_wakeup(vm_page_t m) 327{ 328 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 329 vm_page_flag_clear(m, PG_BUSY); 330 vm_page_flash(m); 331} 332 333void 334vm_page_io_start(vm_page_t m) 335{ 336 337 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 338 m->busy++; 339} 340 341void 342vm_page_io_finish(vm_page_t m) 343{ 344 345 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 346 m->busy--; 347 if (m->busy == 0) 348 vm_page_flash(m); 349} 350 351/* 352 * Keep page from being freed by the page daemon 353 * much of the same effect as wiring, except much lower 354 * overhead and should be used only for *very* temporary 355 * holding ("wiring"). 356 */ 357void 358vm_page_hold(vm_page_t mem) 359{ 360 361 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 362 mem->hold_count++; 363} 364 365void 366vm_page_unhold(vm_page_t mem) 367{ 368 369 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 370 --mem->hold_count; 371 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 372 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 373 vm_page_free_toq(mem); 374} 375 376/* 377 * vm_page_free: 378 * 379 * Free a page 380 * 381 * The clearing of PG_ZERO is a temporary safety until the code can be 382 * reviewed to determine that PG_ZERO is being properly cleared on 383 * write faults or maps. PG_ZERO was previously cleared in 384 * vm_page_alloc(). 385 */ 386void 387vm_page_free(vm_page_t m) 388{ 389 vm_page_flag_clear(m, PG_ZERO); 390 vm_page_free_toq(m); 391 vm_page_zero_idle_wakeup(); 392} 393 394/* 395 * vm_page_free_zero: 396 * 397 * Free a page to the zerod-pages queue 398 */ 399void 400vm_page_free_zero(vm_page_t m) 401{ 402 vm_page_flag_set(m, PG_ZERO); 403 vm_page_free_toq(m); 404} 405 406/* 407 * vm_page_sleep_if_busy: 408 * 409 * Sleep and release the page queues lock if PG_BUSY is set or, 410 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 411 * thread slept and the page queues lock was released. 412 * Otherwise, retains the page queues lock and returns FALSE. 413 */ 414int 415vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 416{ 417 int is_object_locked; 418 419 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 420 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 421 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 422 /* 423 * Remove mtx_owned() after vm_object locking is finished. 424 */ 425 if ((is_object_locked = m->object != NULL && 426 mtx_owned(&m->object->mtx))) 427 mtx_unlock(&m->object->mtx); 428 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 429 if (is_object_locked) 430 mtx_lock(&m->object->mtx); 431 return (TRUE); 432 } 433 return (FALSE); 434} 435 436/* 437 * vm_page_dirty: 438 * 439 * make page all dirty 440 */ 441void 442vm_page_dirty(vm_page_t m) 443{ 444 KASSERT(m->queue - m->pc != PQ_CACHE, 445 ("vm_page_dirty: page in cache!")); 446 KASSERT(m->queue - m->pc != PQ_FREE, 447 ("vm_page_dirty: page is free!")); 448 m->dirty = VM_PAGE_BITS_ALL; 449} 450 451/* 452 * vm_page_splay: 453 * 454 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 455 * the vm_page containing the given pindex. If, however, that 456 * pindex is not found in the vm_object, returns a vm_page that is 457 * adjacent to the pindex, coming before or after it. 458 */ 459vm_page_t 460vm_page_splay(vm_pindex_t pindex, vm_page_t root) 461{ 462 struct vm_page dummy; 463 vm_page_t lefttreemax, righttreemin, y; 464 465 if (root == NULL) 466 return (root); 467 lefttreemax = righttreemin = &dummy; 468 for (;; root = y) { 469 if (pindex < root->pindex) { 470 if ((y = root->left) == NULL) 471 break; 472 if (pindex < y->pindex) { 473 /* Rotate right. */ 474 root->left = y->right; 475 y->right = root; 476 root = y; 477 if ((y = root->left) == NULL) 478 break; 479 } 480 /* Link into the new root's right tree. */ 481 righttreemin->left = root; 482 righttreemin = root; 483 } else if (pindex > root->pindex) { 484 if ((y = root->right) == NULL) 485 break; 486 if (pindex > y->pindex) { 487 /* Rotate left. */ 488 root->right = y->left; 489 y->left = root; 490 root = y; 491 if ((y = root->right) == NULL) 492 break; 493 } 494 /* Link into the new root's left tree. */ 495 lefttreemax->right = root; 496 lefttreemax = root; 497 } else 498 break; 499 } 500 /* Assemble the new root. */ 501 lefttreemax->right = root->left; 502 righttreemin->left = root->right; 503 root->left = dummy.right; 504 root->right = dummy.left; 505 return (root); 506} 507 508/* 509 * vm_page_insert: [ internal use only ] 510 * 511 * Inserts the given mem entry into the object and object list. 512 * 513 * The pagetables are not updated but will presumably fault the page 514 * in if necessary, or if a kernel page the caller will at some point 515 * enter the page into the kernel's pmap. We are not allowed to block 516 * here so we *can't* do this anyway. 517 * 518 * The object and page must be locked, and must be splhigh. 519 * This routine may not block. 520 */ 521void 522vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 523{ 524 vm_page_t root; 525 526 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 527 if (m->object != NULL) 528 panic("vm_page_insert: page already inserted"); 529 530 /* 531 * Record the object/offset pair in this page 532 */ 533 m->object = object; 534 m->pindex = pindex; 535 536 /* 537 * Now link into the object's ordered list of backed pages. 538 */ 539 root = object->root; 540 if (root == NULL) { 541 m->left = NULL; 542 m->right = NULL; 543 TAILQ_INSERT_TAIL(&object->memq, m, listq); 544 } else { 545 root = vm_page_splay(pindex, root); 546 if (pindex < root->pindex) { 547 m->left = root->left; 548 m->right = root; 549 root->left = NULL; 550 TAILQ_INSERT_BEFORE(root, m, listq); 551 } else if (pindex == root->pindex) 552 panic("vm_page_insert: offset already allocated"); 553 else { 554 m->right = root->right; 555 m->left = root; 556 root->right = NULL; 557 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 558 } 559 } 560 object->root = m; 561 object->generation++; 562 563 /* 564 * show that the object has one more resident page. 565 */ 566 object->resident_page_count++; 567 568 /* 569 * Since we are inserting a new and possibly dirty page, 570 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 571 */ 572 if (m->flags & PG_WRITEABLE) 573 vm_object_set_writeable_dirty(object); 574} 575 576/* 577 * vm_page_remove: 578 * NOTE: used by device pager as well -wfj 579 * 580 * Removes the given mem entry from the object/offset-page 581 * table and the object page list, but do not invalidate/terminate 582 * the backing store. 583 * 584 * The object and page must be locked, and at splhigh. 585 * The underlying pmap entry (if any) is NOT removed here. 586 * This routine may not block. 587 */ 588void 589vm_page_remove(vm_page_t m) 590{ 591 vm_object_t object; 592 vm_page_t root; 593 594 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 595 if (m->object == NULL) 596 return; 597#ifndef __alpha__ 598 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 599#endif 600 if ((m->flags & PG_BUSY) == 0) { 601 panic("vm_page_remove: page not busy"); 602 } 603 604 /* 605 * Basically destroy the page. 606 */ 607 vm_page_wakeup(m); 608 609 object = m->object; 610 611 /* 612 * Now remove from the object's list of backed pages. 613 */ 614 if (m != object->root) 615 vm_page_splay(m->pindex, object->root); 616 if (m->left == NULL) 617 root = m->right; 618 else { 619 root = vm_page_splay(m->pindex, m->left); 620 root->right = m->right; 621 } 622 object->root = root; 623 TAILQ_REMOVE(&object->memq, m, listq); 624 625 /* 626 * And show that the object has one fewer resident page. 627 */ 628 object->resident_page_count--; 629 object->generation++; 630 631 m->object = NULL; 632} 633 634/* 635 * vm_page_lookup: 636 * 637 * Returns the page associated with the object/offset 638 * pair specified; if none is found, NULL is returned. 639 * 640 * The object must be locked. 641 * This routine may not block. 642 * This is a critical path routine 643 */ 644vm_page_t 645vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 646{ 647 vm_page_t m; 648 649 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 650 if ((m = object->root) != NULL && m->pindex != pindex) { 651 m = vm_page_splay(pindex, m); 652 if ((object->root = m)->pindex != pindex) 653 m = NULL; 654 } 655 return (m); 656} 657 658/* 659 * vm_page_rename: 660 * 661 * Move the given memory entry from its 662 * current object to the specified target object/offset. 663 * 664 * The object must be locked. 665 * This routine may not block. 666 * 667 * Note: this routine will raise itself to splvm(), the caller need not. 668 * 669 * Note: swap associated with the page must be invalidated by the move. We 670 * have to do this for several reasons: (1) we aren't freeing the 671 * page, (2) we are dirtying the page, (3) the VM system is probably 672 * moving the page from object A to B, and will then later move 673 * the backing store from A to B and we can't have a conflict. 674 * 675 * Note: we *always* dirty the page. It is necessary both for the 676 * fact that we moved it, and because we may be invalidating 677 * swap. If the page is on the cache, we have to deactivate it 678 * or vm_page_dirty() will panic. Dirty pages are not allowed 679 * on the cache. 680 */ 681void 682vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 683{ 684 int s; 685 686 s = splvm(); 687 vm_page_remove(m); 688 vm_page_insert(m, new_object, new_pindex); 689 if (m->queue - m->pc == PQ_CACHE) 690 vm_page_deactivate(m); 691 vm_page_dirty(m); 692 splx(s); 693} 694 695/* 696 * vm_page_select_cache: 697 * 698 * Find a page on the cache queue with color optimization. As pages 699 * might be found, but not applicable, they are deactivated. This 700 * keeps us from using potentially busy cached pages. 701 * 702 * This routine must be called at splvm(). 703 * This routine may not block. 704 */ 705vm_page_t 706vm_page_select_cache(int color) 707{ 708 vm_page_t m; 709 710 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 711 while (TRUE) { 712 m = vm_pageq_find(PQ_CACHE, color, FALSE); 713 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 714 m->hold_count || m->wire_count || 715 (!VM_OBJECT_TRYLOCK(m->object) && 716 !VM_OBJECT_LOCKED(m->object)))) { 717 vm_page_deactivate(m); 718 continue; 719 } 720 return m; 721 } 722} 723 724/* 725 * vm_page_alloc: 726 * 727 * Allocate and return a memory cell associated 728 * with this VM object/offset pair. 729 * 730 * page_req classes: 731 * VM_ALLOC_NORMAL normal process request 732 * VM_ALLOC_SYSTEM system *really* needs a page 733 * VM_ALLOC_INTERRUPT interrupt time request 734 * VM_ALLOC_ZERO zero page 735 * 736 * This routine may not block. 737 * 738 * Additional special handling is required when called from an 739 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 740 * the page cache in this case. 741 */ 742vm_page_t 743vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 744{ 745 vm_object_t m_object; 746 vm_page_t m = NULL; 747 int color, flags, page_req, s; 748 749 page_req = req & VM_ALLOC_CLASS_MASK; 750 751 if ((req & VM_ALLOC_NOOBJ) == 0) { 752 KASSERT(object != NULL, 753 ("vm_page_alloc: NULL object.")); 754 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 755 color = (pindex + object->pg_color) & PQ_L2_MASK; 756 } else 757 color = pindex & PQ_L2_MASK; 758 759 /* 760 * The pager is allowed to eat deeper into the free page list. 761 */ 762 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 763 page_req = VM_ALLOC_SYSTEM; 764 }; 765 766 s = splvm(); 767loop: 768 mtx_lock_spin(&vm_page_queue_free_mtx); 769 if (cnt.v_free_count > cnt.v_free_reserved || 770 (page_req == VM_ALLOC_SYSTEM && 771 cnt.v_cache_count == 0 && 772 cnt.v_free_count > cnt.v_interrupt_free_min) || 773 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 774 /* 775 * Allocate from the free queue if the number of free pages 776 * exceeds the minimum for the request class. 777 */ 778 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 779 } else if (page_req != VM_ALLOC_INTERRUPT) { 780 mtx_unlock_spin(&vm_page_queue_free_mtx); 781 /* 782 * Allocatable from cache (non-interrupt only). On success, 783 * we must free the page and try again, thus ensuring that 784 * cnt.v_*_free_min counters are replenished. 785 */ 786 vm_page_lock_queues(); 787 if ((m = vm_page_select_cache(color)) == NULL) { 788 vm_page_unlock_queues(); 789 splx(s); 790#if defined(DIAGNOSTIC) 791 if (cnt.v_cache_count > 0) 792 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 793#endif 794 atomic_add_int(&vm_pageout_deficit, 1); 795 pagedaemon_wakeup(); 796 return (NULL); 797 } 798 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 799 m_object = m->object; 800 VM_OBJECT_LOCK_ASSERT(m_object, MA_OWNED); 801 vm_page_busy(m); 802 vm_page_free(m); 803 vm_page_unlock_queues(); 804 if (m_object != object) 805 VM_OBJECT_UNLOCK(m_object); 806 goto loop; 807 } else { 808 /* 809 * Not allocatable from cache from interrupt, give up. 810 */ 811 mtx_unlock_spin(&vm_page_queue_free_mtx); 812 splx(s); 813 atomic_add_int(&vm_pageout_deficit, 1); 814 pagedaemon_wakeup(); 815 return (NULL); 816 } 817 818 /* 819 * At this point we had better have found a good page. 820 */ 821 822 KASSERT( 823 m != NULL, 824 ("vm_page_alloc(): missing page on free queue\n") 825 ); 826 827 /* 828 * Remove from free queue 829 */ 830 831 vm_pageq_remove_nowakeup(m); 832 833 /* 834 * Initialize structure. Only the PG_ZERO flag is inherited. 835 */ 836 flags = PG_BUSY; 837 if (m->flags & PG_ZERO) { 838 vm_page_zero_count--; 839 if (req & VM_ALLOC_ZERO) 840 flags = PG_ZERO | PG_BUSY; 841 } 842 m->flags = flags; 843 if (req & VM_ALLOC_WIRED) { 844 atomic_add_int(&cnt.v_wire_count, 1); 845 m->wire_count = 1; 846 } else 847 m->wire_count = 0; 848 m->hold_count = 0; 849 m->act_count = 0; 850 m->busy = 0; 851 m->valid = 0; 852 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 853 mtx_unlock_spin(&vm_page_queue_free_mtx); 854 855 /* 856 * vm_page_insert() is safe prior to the splx(). Note also that 857 * inserting a page here does not insert it into the pmap (which 858 * could cause us to block allocating memory). We cannot block 859 * anywhere. 860 */ 861 if ((req & VM_ALLOC_NOOBJ) == 0) 862 vm_page_insert(m, object, pindex); 863 else 864 m->pindex = pindex; 865 866 /* 867 * Don't wakeup too often - wakeup the pageout daemon when 868 * we would be nearly out of memory. 869 */ 870 if (vm_paging_needed()) 871 pagedaemon_wakeup(); 872 873 splx(s); 874 return (m); 875} 876 877/* 878 * vm_wait: (also see VM_WAIT macro) 879 * 880 * Block until free pages are available for allocation 881 * - Called in various places before memory allocations. 882 */ 883void 884vm_wait(void) 885{ 886 int s; 887 888 s = splvm(); 889 vm_page_lock_queues(); 890 if (curproc == pageproc) { 891 vm_pageout_pages_needed = 1; 892 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 893 PDROP | PSWP, "VMWait", 0); 894 } else { 895 if (!vm_pages_needed) { 896 vm_pages_needed = 1; 897 wakeup(&vm_pages_needed); 898 } 899 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 900 "vmwait", 0); 901 } 902 splx(s); 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 int s; 919 920 s = splvm(); 921 vm_page_lock_queues(); 922 if (!vm_pages_needed) { 923 vm_pages_needed = 1; 924 wakeup(&vm_pages_needed); 925 } 926 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 927 "pfault", 0); 928 splx(s); 929} 930 931/* 932 * vm_page_activate: 933 * 934 * Put the specified page on the active list (if appropriate). 935 * Ensure that act_count is at least ACT_INIT but do not otherwise 936 * mess with it. 937 * 938 * The page queues must be locked. 939 * This routine may not block. 940 */ 941void 942vm_page_activate(vm_page_t m) 943{ 944 int s; 945 946 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 947 s = splvm(); 948 if (m->queue != PQ_ACTIVE) { 949 if ((m->queue - m->pc) == PQ_CACHE) 950 cnt.v_reactivated++; 951 vm_pageq_remove(m); 952 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 953 if (m->act_count < ACT_INIT) 954 m->act_count = ACT_INIT; 955 vm_pageq_enqueue(PQ_ACTIVE, m); 956 } 957 } else { 958 if (m->act_count < ACT_INIT) 959 m->act_count = ACT_INIT; 960 } 961 splx(s); 962} 963 964/* 965 * vm_page_free_wakeup: 966 * 967 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 968 * routine is called when a page has been added to the cache or free 969 * queues. 970 * 971 * This routine may not block. 972 * This routine must be called at splvm() 973 */ 974static __inline void 975vm_page_free_wakeup(void) 976{ 977 978 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 979 /* 980 * if pageout daemon needs pages, then tell it that there are 981 * some free. 982 */ 983 if (vm_pageout_pages_needed && 984 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 985 wakeup(&vm_pageout_pages_needed); 986 vm_pageout_pages_needed = 0; 987 } 988 /* 989 * wakeup processes that are waiting on memory if we hit a 990 * high water mark. And wakeup scheduler process if we have 991 * lots of memory. this process will swapin processes. 992 */ 993 if (vm_pages_needed && !vm_page_count_min()) { 994 vm_pages_needed = 0; 995 wakeup(&cnt.v_free_count); 996 } 997} 998 999/* 1000 * vm_page_free_toq: 1001 * 1002 * Returns the given page to the PQ_FREE list, 1003 * disassociating it with any VM object. 1004 * 1005 * Object and page must be locked prior to entry. 1006 * This routine may not block. 1007 */ 1008 1009void 1010vm_page_free_toq(vm_page_t m) 1011{ 1012 int s; 1013 struct vpgqueues *pq; 1014 vm_object_t object = m->object; 1015 1016 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1017 s = splvm(); 1018 cnt.v_tfree++; 1019 1020 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1021 printf( 1022 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1023 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1024 m->hold_count); 1025 if ((m->queue - m->pc) == PQ_FREE) 1026 panic("vm_page_free: freeing free page"); 1027 else 1028 panic("vm_page_free: freeing busy page"); 1029 } 1030 1031 /* 1032 * unqueue, then remove page. Note that we cannot destroy 1033 * the page here because we do not want to call the pager's 1034 * callback routine until after we've put the page on the 1035 * appropriate free queue. 1036 */ 1037 vm_pageq_remove_nowakeup(m); 1038 vm_page_remove(m); 1039 1040 /* 1041 * If fictitious remove object association and 1042 * return, otherwise delay object association removal. 1043 */ 1044 if ((m->flags & PG_FICTITIOUS) != 0) { 1045 splx(s); 1046 return; 1047 } 1048 1049 m->valid = 0; 1050 vm_page_undirty(m); 1051 1052 if (m->wire_count != 0) { 1053 if (m->wire_count > 1) { 1054 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1055 m->wire_count, (long)m->pindex); 1056 } 1057 panic("vm_page_free: freeing wired page\n"); 1058 } 1059 1060 /* 1061 * If we've exhausted the object's resident pages we want to free 1062 * it up. 1063 */ 1064 if (object && 1065 (object->type == OBJT_VNODE) && 1066 ((object->flags & OBJ_DEAD) == 0) 1067 ) { 1068 struct vnode *vp = (struct vnode *)object->handle; 1069 1070 if (vp) { 1071 VI_LOCK(vp); 1072 if (VSHOULDFREE(vp)) 1073 vfree(vp); 1074 VI_UNLOCK(vp); 1075 } 1076 } 1077 1078 /* 1079 * Clear the UNMANAGED flag when freeing an unmanaged page. 1080 */ 1081 if (m->flags & PG_UNMANAGED) { 1082 m->flags &= ~PG_UNMANAGED; 1083 } 1084 1085 if (m->hold_count != 0) { 1086 m->flags &= ~PG_ZERO; 1087 m->queue = PQ_HOLD; 1088 } else 1089 m->queue = PQ_FREE + m->pc; 1090 pq = &vm_page_queues[m->queue]; 1091 mtx_lock_spin(&vm_page_queue_free_mtx); 1092 pq->lcnt++; 1093 ++(*pq->cnt); 1094 1095 /* 1096 * Put zero'd pages on the end ( where we look for zero'd pages 1097 * first ) and non-zerod pages at the head. 1098 */ 1099 if (m->flags & PG_ZERO) { 1100 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1101 ++vm_page_zero_count; 1102 } else { 1103 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1104 } 1105 mtx_unlock_spin(&vm_page_queue_free_mtx); 1106 vm_page_free_wakeup(); 1107 splx(s); 1108} 1109 1110/* 1111 * vm_page_unmanage: 1112 * 1113 * Prevent PV management from being done on the page. The page is 1114 * removed from the paging queues as if it were wired, and as a 1115 * consequence of no longer being managed the pageout daemon will not 1116 * touch it (since there is no way to locate the pte mappings for the 1117 * page). madvise() calls that mess with the pmap will also no longer 1118 * operate on the page. 1119 * 1120 * Beyond that the page is still reasonably 'normal'. Freeing the page 1121 * will clear the flag. 1122 * 1123 * This routine is used by OBJT_PHYS objects - objects using unswappable 1124 * physical memory as backing store rather then swap-backed memory and 1125 * will eventually be extended to support 4MB unmanaged physical 1126 * mappings. 1127 */ 1128void 1129vm_page_unmanage(vm_page_t m) 1130{ 1131 int s; 1132 1133 s = splvm(); 1134 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1135 if ((m->flags & PG_UNMANAGED) == 0) { 1136 if (m->wire_count == 0) 1137 vm_pageq_remove(m); 1138 } 1139 vm_page_flag_set(m, PG_UNMANAGED); 1140 splx(s); 1141} 1142 1143/* 1144 * vm_page_wire: 1145 * 1146 * Mark this page as wired down by yet 1147 * another map, removing it from paging queues 1148 * as necessary. 1149 * 1150 * The page queues must be locked. 1151 * This routine may not block. 1152 */ 1153void 1154vm_page_wire(vm_page_t m) 1155{ 1156 int s; 1157 1158 /* 1159 * Only bump the wire statistics if the page is not already wired, 1160 * and only unqueue the page if it is on some queue (if it is unmanaged 1161 * it is already off the queues). 1162 */ 1163 s = splvm(); 1164 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1165 if (m->wire_count == 0) { 1166 if ((m->flags & PG_UNMANAGED) == 0) 1167 vm_pageq_remove(m); 1168 atomic_add_int(&cnt.v_wire_count, 1); 1169 } 1170 m->wire_count++; 1171 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1172 splx(s); 1173} 1174 1175/* 1176 * vm_page_unwire: 1177 * 1178 * Release one wiring of this page, potentially 1179 * enabling it to be paged again. 1180 * 1181 * Many pages placed on the inactive queue should actually go 1182 * into the cache, but it is difficult to figure out which. What 1183 * we do instead, if the inactive target is well met, is to put 1184 * clean pages at the head of the inactive queue instead of the tail. 1185 * This will cause them to be moved to the cache more quickly and 1186 * if not actively re-referenced, freed more quickly. If we just 1187 * stick these pages at the end of the inactive queue, heavy filesystem 1188 * meta-data accesses can cause an unnecessary paging load on memory bound 1189 * processes. This optimization causes one-time-use metadata to be 1190 * reused more quickly. 1191 * 1192 * BUT, if we are in a low-memory situation we have no choice but to 1193 * put clean pages on the cache queue. 1194 * 1195 * A number of routines use vm_page_unwire() to guarantee that the page 1196 * will go into either the inactive or active queues, and will NEVER 1197 * be placed in the cache - for example, just after dirtying a page. 1198 * dirty pages in the cache are not allowed. 1199 * 1200 * The page queues must be locked. 1201 * This routine may not block. 1202 */ 1203void 1204vm_page_unwire(vm_page_t m, int activate) 1205{ 1206 int s; 1207 1208 s = splvm(); 1209 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1210 if (m->wire_count > 0) { 1211 m->wire_count--; 1212 if (m->wire_count == 0) { 1213 atomic_subtract_int(&cnt.v_wire_count, 1); 1214 if (m->flags & PG_UNMANAGED) { 1215 ; 1216 } else if (activate) 1217 vm_pageq_enqueue(PQ_ACTIVE, m); 1218 else { 1219 vm_page_flag_clear(m, PG_WINATCFLS); 1220 vm_pageq_enqueue(PQ_INACTIVE, m); 1221 } 1222 } 1223 } else { 1224 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1225 } 1226 splx(s); 1227} 1228 1229 1230/* 1231 * Move the specified page to the inactive queue. If the page has 1232 * any associated swap, the swap is deallocated. 1233 * 1234 * Normally athead is 0 resulting in LRU operation. athead is set 1235 * to 1 if we want this page to be 'as if it were placed in the cache', 1236 * except without unmapping it from the process address space. 1237 * 1238 * This routine may not block. 1239 */ 1240static __inline void 1241_vm_page_deactivate(vm_page_t m, int athead) 1242{ 1243 int s; 1244 1245 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1246 /* 1247 * Ignore if already inactive. 1248 */ 1249 if (m->queue == PQ_INACTIVE) 1250 return; 1251 1252 s = splvm(); 1253 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1254 if ((m->queue - m->pc) == PQ_CACHE) 1255 cnt.v_reactivated++; 1256 vm_page_flag_clear(m, PG_WINATCFLS); 1257 vm_pageq_remove(m); 1258 if (athead) 1259 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1260 else 1261 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1262 m->queue = PQ_INACTIVE; 1263 vm_page_queues[PQ_INACTIVE].lcnt++; 1264 cnt.v_inactive_count++; 1265 } 1266 splx(s); 1267} 1268 1269void 1270vm_page_deactivate(vm_page_t m) 1271{ 1272 _vm_page_deactivate(m, 0); 1273} 1274 1275/* 1276 * vm_page_try_to_cache: 1277 * 1278 * Returns 0 on failure, 1 on success 1279 */ 1280int 1281vm_page_try_to_cache(vm_page_t m) 1282{ 1283 1284 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1285 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1286 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1287 return (0); 1288 } 1289 pmap_remove_all(m); 1290 if (m->dirty) 1291 return (0); 1292 vm_page_cache(m); 1293 return (1); 1294} 1295 1296/* 1297 * vm_page_try_to_free() 1298 * 1299 * Attempt to free the page. If we cannot free it, we do nothing. 1300 * 1 is returned on success, 0 on failure. 1301 */ 1302int 1303vm_page_try_to_free(vm_page_t m) 1304{ 1305 1306 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1307 if (m->object != NULL) 1308 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1309 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1310 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1311 return (0); 1312 } 1313 pmap_remove_all(m); 1314 if (m->dirty) 1315 return (0); 1316 vm_page_busy(m); 1317 vm_page_free(m); 1318 return (1); 1319} 1320 1321/* 1322 * vm_page_cache 1323 * 1324 * Put the specified page onto the page cache queue (if appropriate). 1325 * 1326 * This routine may not block. 1327 */ 1328void 1329vm_page_cache(vm_page_t m) 1330{ 1331 int s; 1332 1333 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1334 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1335 m->hold_count || m->wire_count) { 1336 printf("vm_page_cache: attempting to cache busy page\n"); 1337 return; 1338 } 1339 if ((m->queue - m->pc) == PQ_CACHE) 1340 return; 1341 1342 /* 1343 * Remove all pmaps and indicate that the page is not 1344 * writeable or mapped. 1345 */ 1346 pmap_remove_all(m); 1347 if (m->dirty != 0) { 1348 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1349 (long)m->pindex); 1350 } 1351 s = splvm(); 1352 vm_pageq_remove_nowakeup(m); 1353 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1354 vm_page_free_wakeup(); 1355 splx(s); 1356} 1357 1358/* 1359 * vm_page_dontneed 1360 * 1361 * Cache, deactivate, or do nothing as appropriate. This routine 1362 * is typically used by madvise() MADV_DONTNEED. 1363 * 1364 * Generally speaking we want to move the page into the cache so 1365 * it gets reused quickly. However, this can result in a silly syndrome 1366 * due to the page recycling too quickly. Small objects will not be 1367 * fully cached. On the otherhand, if we move the page to the inactive 1368 * queue we wind up with a problem whereby very large objects 1369 * unnecessarily blow away our inactive and cache queues. 1370 * 1371 * The solution is to move the pages based on a fixed weighting. We 1372 * either leave them alone, deactivate them, or move them to the cache, 1373 * where moving them to the cache has the highest weighting. 1374 * By forcing some pages into other queues we eventually force the 1375 * system to balance the queues, potentially recovering other unrelated 1376 * space from active. The idea is to not force this to happen too 1377 * often. 1378 */ 1379void 1380vm_page_dontneed(vm_page_t m) 1381{ 1382 static int dnweight; 1383 int dnw; 1384 int head; 1385 1386 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1387 dnw = ++dnweight; 1388 1389 /* 1390 * occassionally leave the page alone 1391 */ 1392 if ((dnw & 0x01F0) == 0 || 1393 m->queue == PQ_INACTIVE || 1394 m->queue - m->pc == PQ_CACHE 1395 ) { 1396 if (m->act_count >= ACT_INIT) 1397 --m->act_count; 1398 return; 1399 } 1400 1401 if (m->dirty == 0 && pmap_is_modified(m)) 1402 vm_page_dirty(m); 1403 1404 if (m->dirty || (dnw & 0x0070) == 0) { 1405 /* 1406 * Deactivate the page 3 times out of 32. 1407 */ 1408 head = 0; 1409 } else { 1410 /* 1411 * Cache the page 28 times out of every 32. Note that 1412 * the page is deactivated instead of cached, but placed 1413 * at the head of the queue instead of the tail. 1414 */ 1415 head = 1; 1416 } 1417 _vm_page_deactivate(m, head); 1418} 1419 1420/* 1421 * Grab a page, waiting until we are waken up due to the page 1422 * changing state. We keep on waiting, if the page continues 1423 * to be in the object. If the page doesn't exist, first allocate it 1424 * and then conditionally zero it. 1425 * 1426 * This routine may block. 1427 */ 1428vm_page_t 1429vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1430{ 1431 vm_page_t m; 1432 1433 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1434retrylookup: 1435 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1436 vm_page_lock_queues(); 1437 if (m->busy || (m->flags & PG_BUSY)) { 1438 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1439 VM_OBJECT_UNLOCK(object); 1440 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0); 1441 VM_OBJECT_LOCK(object); 1442 if ((allocflags & VM_ALLOC_RETRY) == 0) 1443 return (NULL); 1444 goto retrylookup; 1445 } else { 1446 if (allocflags & VM_ALLOC_WIRED) 1447 vm_page_wire(m); 1448 vm_page_busy(m); 1449 vm_page_unlock_queues(); 1450 return (m); 1451 } 1452 } 1453 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1454 if (m == NULL) { 1455 VM_OBJECT_UNLOCK(object); 1456 VM_WAIT; 1457 VM_OBJECT_LOCK(object); 1458 if ((allocflags & VM_ALLOC_RETRY) == 0) 1459 return (NULL); 1460 goto retrylookup; 1461 } 1462 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1463 pmap_zero_page(m); 1464 return (m); 1465} 1466 1467/* 1468 * Mapping function for valid bits or for dirty bits in 1469 * a page. May not block. 1470 * 1471 * Inputs are required to range within a page. 1472 */ 1473__inline int 1474vm_page_bits(int base, int size) 1475{ 1476 int first_bit; 1477 int last_bit; 1478 1479 KASSERT( 1480 base + size <= PAGE_SIZE, 1481 ("vm_page_bits: illegal base/size %d/%d", base, size) 1482 ); 1483 1484 if (size == 0) /* handle degenerate case */ 1485 return (0); 1486 1487 first_bit = base >> DEV_BSHIFT; 1488 last_bit = (base + size - 1) >> DEV_BSHIFT; 1489 1490 return ((2 << last_bit) - (1 << first_bit)); 1491} 1492 1493/* 1494 * vm_page_set_validclean: 1495 * 1496 * Sets portions of a page valid and clean. The arguments are expected 1497 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1498 * of any partial chunks touched by the range. The invalid portion of 1499 * such chunks will be zero'd. 1500 * 1501 * This routine may not block. 1502 * 1503 * (base + size) must be less then or equal to PAGE_SIZE. 1504 */ 1505void 1506vm_page_set_validclean(vm_page_t m, int base, int size) 1507{ 1508 int pagebits; 1509 int frag; 1510 int endoff; 1511 1512 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1513 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1514 if (size == 0) /* handle degenerate case */ 1515 return; 1516 1517 /* 1518 * If the base is not DEV_BSIZE aligned and the valid 1519 * bit is clear, we have to zero out a portion of the 1520 * first block. 1521 */ 1522 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1523 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1524 pmap_zero_page_area(m, frag, base - frag); 1525 1526 /* 1527 * If the ending offset is not DEV_BSIZE aligned and the 1528 * valid bit is clear, we have to zero out a portion of 1529 * the last block. 1530 */ 1531 endoff = base + size; 1532 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1533 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1534 pmap_zero_page_area(m, endoff, 1535 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1536 1537 /* 1538 * Set valid, clear dirty bits. If validating the entire 1539 * page we can safely clear the pmap modify bit. We also 1540 * use this opportunity to clear the PG_NOSYNC flag. If a process 1541 * takes a write fault on a MAP_NOSYNC memory area the flag will 1542 * be set again. 1543 * 1544 * We set valid bits inclusive of any overlap, but we can only 1545 * clear dirty bits for DEV_BSIZE chunks that are fully within 1546 * the range. 1547 */ 1548 pagebits = vm_page_bits(base, size); 1549 m->valid |= pagebits; 1550#if 0 /* NOT YET */ 1551 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1552 frag = DEV_BSIZE - frag; 1553 base += frag; 1554 size -= frag; 1555 if (size < 0) 1556 size = 0; 1557 } 1558 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1559#endif 1560 m->dirty &= ~pagebits; 1561 if (base == 0 && size == PAGE_SIZE) { 1562 pmap_clear_modify(m); 1563 vm_page_flag_clear(m, PG_NOSYNC); 1564 } 1565} 1566 1567void 1568vm_page_clear_dirty(vm_page_t m, int base, int size) 1569{ 1570 1571 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1572 m->dirty &= ~vm_page_bits(base, size); 1573} 1574 1575/* 1576 * vm_page_set_invalid: 1577 * 1578 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1579 * valid and dirty bits for the effected areas are cleared. 1580 * 1581 * May not block. 1582 */ 1583void 1584vm_page_set_invalid(vm_page_t m, int base, int size) 1585{ 1586 int bits; 1587 1588 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1589 bits = vm_page_bits(base, size); 1590 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1591 m->valid &= ~bits; 1592 m->dirty &= ~bits; 1593 m->object->generation++; 1594} 1595 1596/* 1597 * vm_page_zero_invalid() 1598 * 1599 * The kernel assumes that the invalid portions of a page contain 1600 * garbage, but such pages can be mapped into memory by user code. 1601 * When this occurs, we must zero out the non-valid portions of the 1602 * page so user code sees what it expects. 1603 * 1604 * Pages are most often semi-valid when the end of a file is mapped 1605 * into memory and the file's size is not page aligned. 1606 */ 1607void 1608vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1609{ 1610 int b; 1611 int i; 1612 1613 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1614 /* 1615 * Scan the valid bits looking for invalid sections that 1616 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1617 * valid bit may be set ) have already been zerod by 1618 * vm_page_set_validclean(). 1619 */ 1620 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1621 if (i == (PAGE_SIZE / DEV_BSIZE) || 1622 (m->valid & (1 << i)) 1623 ) { 1624 if (i > b) { 1625 pmap_zero_page_area(m, 1626 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1627 } 1628 b = i + 1; 1629 } 1630 } 1631 1632 /* 1633 * setvalid is TRUE when we can safely set the zero'd areas 1634 * as being valid. We can do this if there are no cache consistancy 1635 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1636 */ 1637 if (setvalid) 1638 m->valid = VM_PAGE_BITS_ALL; 1639} 1640 1641/* 1642 * vm_page_is_valid: 1643 * 1644 * Is (partial) page valid? Note that the case where size == 0 1645 * will return FALSE in the degenerate case where the page is 1646 * entirely invalid, and TRUE otherwise. 1647 * 1648 * May not block. 1649 */ 1650int 1651vm_page_is_valid(vm_page_t m, int base, int size) 1652{ 1653 int bits = vm_page_bits(base, size); 1654 1655 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1656 if (m->valid && ((m->valid & bits) == bits)) 1657 return 1; 1658 else 1659 return 0; 1660} 1661 1662/* 1663 * update dirty bits from pmap/mmu. May not block. 1664 */ 1665void 1666vm_page_test_dirty(vm_page_t m) 1667{ 1668 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1669 vm_page_dirty(m); 1670 } 1671} 1672 1673int so_zerocp_fullpage = 0; 1674 1675void 1676vm_page_cowfault(vm_page_t m) 1677{ 1678 vm_page_t mnew; 1679 vm_object_t object; 1680 vm_pindex_t pindex; 1681 1682 object = m->object; 1683 pindex = m->pindex; 1684 vm_page_busy(m); 1685 1686 retry_alloc: 1687 vm_page_remove(m); 1688 /* 1689 * An interrupt allocation is requested because the page 1690 * queues lock is held. 1691 */ 1692 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1693 if (mnew == NULL) { 1694 vm_page_insert(m, object, pindex); 1695 vm_page_unlock_queues(); 1696 VM_OBJECT_UNLOCK(object); 1697 VM_WAIT; 1698 VM_OBJECT_LOCK(object); 1699 vm_page_lock_queues(); 1700 goto retry_alloc; 1701 } 1702 1703 if (m->cow == 0) { 1704 /* 1705 * check to see if we raced with an xmit complete when 1706 * waiting to allocate a page. If so, put things back 1707 * the way they were 1708 */ 1709 vm_page_busy(mnew); 1710 vm_page_free(mnew); 1711 vm_page_insert(m, object, pindex); 1712 } else { /* clear COW & copy page */ 1713 if (!so_zerocp_fullpage) 1714 pmap_copy_page(m, mnew); 1715 mnew->valid = VM_PAGE_BITS_ALL; 1716 vm_page_dirty(mnew); 1717 vm_page_flag_clear(mnew, PG_BUSY); 1718 } 1719} 1720 1721void 1722vm_page_cowclear(vm_page_t m) 1723{ 1724 1725 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1726 if (m->cow) { 1727 m->cow--; 1728 /* 1729 * let vm_fault add back write permission lazily 1730 */ 1731 } 1732 /* 1733 * sf_buf_free() will free the page, so we needn't do it here 1734 */ 1735} 1736 1737void 1738vm_page_cowsetup(vm_page_t m) 1739{ 1740 1741 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1742 m->cow++; 1743 pmap_page_protect(m, VM_PROT_READ); 1744} 1745 1746#include "opt_ddb.h" 1747#ifdef DDB 1748#include <sys/kernel.h> 1749 1750#include <ddb/ddb.h> 1751 1752DB_SHOW_COMMAND(page, vm_page_print_page_info) 1753{ 1754 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1755 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1756 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1757 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1758 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1759 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1760 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1761 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1762 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1763 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1764} 1765 1766DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1767{ 1768 int i; 1769 db_printf("PQ_FREE:"); 1770 for (i = 0; i < PQ_L2_SIZE; i++) { 1771 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1772 } 1773 db_printf("\n"); 1774 1775 db_printf("PQ_CACHE:"); 1776 for (i = 0; i < PQ_L2_SIZE; i++) { 1777 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1778 } 1779 db_printf("\n"); 1780 1781 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1782 vm_page_queues[PQ_ACTIVE].lcnt, 1783 vm_page_queues[PQ_INACTIVE].lcnt); 1784} 1785#endif /* DDB */ 1786