vm_page.c revision 139825
190075Sobrien/*- 290075Sobrien * Copyright (c) 1991 Regents of the University of California. 390075Sobrien * All rights reserved. 490075Sobrien * 5169689Skan * This code is derived from software contributed to Berkeley by 690075Sobrien * The Mach Operating System project at Carnegie-Mellon University. 790075Sobrien * 890075Sobrien * Redistribution and use in source and binary forms, with or without 990075Sobrien * modification, are permitted provided that the following conditions 1090075Sobrien * are met: 1190075Sobrien * 1. Redistributions of source code must retain the above copyright 1290075Sobrien * notice, this list of conditions and the following disclaimer. 1390075Sobrien * 2. Redistributions in binary form must reproduce the above copyright 1490075Sobrien * notice, this list of conditions and the following disclaimer in the 15146895Skan * documentation and/or other materials provided with the distribution. 1690075Sobrien * 4. Neither the name of the University nor the names of its contributors 17146895Skan * may be used to endorse or promote products derived from this software 18146895Skan * without specific prior written permission. 19146895Skan * 20146895Skan * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 2190075Sobrien * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 2290075Sobrien * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 2390075Sobrien * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24169689Skan * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25169689Skan * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26169689Skan * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27110611Skan * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28110611Skan * 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 139825 2005-01-07 02:29:27Z imp $"); 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{ 315 316 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 317 if (m->flags & PG_WANTED) { 318 vm_page_flag_clear(m, PG_WANTED); 319 wakeup(m); 320 } 321} 322 323/* 324 * vm_page_wakeup: 325 * 326 * clear the PG_BUSY flag and wakeup anyone waiting for the 327 * page. 328 * 329 */ 330void 331vm_page_wakeup(vm_page_t m) 332{ 333 334 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 335 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 336 vm_page_flag_clear(m, PG_BUSY); 337 vm_page_flash(m); 338} 339 340void 341vm_page_io_start(vm_page_t m) 342{ 343 344 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 345 m->busy++; 346} 347 348void 349vm_page_io_finish(vm_page_t m) 350{ 351 352 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 353 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 354 m->busy--; 355 if (m->busy == 0) 356 vm_page_flash(m); 357} 358 359/* 360 * Keep page from being freed by the page daemon 361 * much of the same effect as wiring, except much lower 362 * overhead and should be used only for *very* temporary 363 * holding ("wiring"). 364 */ 365void 366vm_page_hold(vm_page_t mem) 367{ 368 369 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 370 mem->hold_count++; 371} 372 373void 374vm_page_unhold(vm_page_t mem) 375{ 376 377 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 378 --mem->hold_count; 379 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 380 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 381 vm_page_free_toq(mem); 382} 383 384/* 385 * vm_page_free: 386 * 387 * Free a page 388 * 389 * The clearing of PG_ZERO is a temporary safety until the code can be 390 * reviewed to determine that PG_ZERO is being properly cleared on 391 * write faults or maps. PG_ZERO was previously cleared in 392 * vm_page_alloc(). 393 */ 394void 395vm_page_free(vm_page_t m) 396{ 397 vm_page_flag_clear(m, PG_ZERO); 398 vm_page_free_toq(m); 399 vm_page_zero_idle_wakeup(); 400} 401 402/* 403 * vm_page_free_zero: 404 * 405 * Free a page to the zerod-pages queue 406 */ 407void 408vm_page_free_zero(vm_page_t m) 409{ 410 vm_page_flag_set(m, PG_ZERO); 411 vm_page_free_toq(m); 412} 413 414/* 415 * vm_page_sleep_if_busy: 416 * 417 * Sleep and release the page queues lock if PG_BUSY is set or, 418 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 419 * thread slept and the page queues lock was released. 420 * Otherwise, retains the page queues lock and returns FALSE. 421 */ 422int 423vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 424{ 425 vm_object_t object; 426 427 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 428 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 429 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 430 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 431 /* 432 * It's possible that while we sleep, the page will get 433 * unbusied and freed. If we are holding the object 434 * lock, we will assume we hold a reference to the object 435 * such that even if m->object changes, we can re-lock 436 * it. 437 */ 438 object = m->object; 439 VM_OBJECT_UNLOCK(object); 440 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 441 VM_OBJECT_LOCK(object); 442 return (TRUE); 443 } 444 return (FALSE); 445} 446 447/* 448 * vm_page_dirty: 449 * 450 * make page all dirty 451 */ 452void 453vm_page_dirty(vm_page_t m) 454{ 455 KASSERT(m->queue - m->pc != PQ_CACHE, 456 ("vm_page_dirty: page in cache!")); 457 KASSERT(m->queue - m->pc != PQ_FREE, 458 ("vm_page_dirty: page is free!")); 459 m->dirty = VM_PAGE_BITS_ALL; 460} 461 462/* 463 * vm_page_splay: 464 * 465 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 466 * the vm_page containing the given pindex. If, however, that 467 * pindex is not found in the vm_object, returns a vm_page that is 468 * adjacent to the pindex, coming before or after it. 469 */ 470vm_page_t 471vm_page_splay(vm_pindex_t pindex, vm_page_t root) 472{ 473 struct vm_page dummy; 474 vm_page_t lefttreemax, righttreemin, y; 475 476 if (root == NULL) 477 return (root); 478 lefttreemax = righttreemin = &dummy; 479 for (;; root = y) { 480 if (pindex < root->pindex) { 481 if ((y = root->left) == NULL) 482 break; 483 if (pindex < y->pindex) { 484 /* Rotate right. */ 485 root->left = y->right; 486 y->right = root; 487 root = y; 488 if ((y = root->left) == NULL) 489 break; 490 } 491 /* Link into the new root's right tree. */ 492 righttreemin->left = root; 493 righttreemin = root; 494 } else if (pindex > root->pindex) { 495 if ((y = root->right) == NULL) 496 break; 497 if (pindex > y->pindex) { 498 /* Rotate left. */ 499 root->right = y->left; 500 y->left = root; 501 root = y; 502 if ((y = root->right) == NULL) 503 break; 504 } 505 /* Link into the new root's left tree. */ 506 lefttreemax->right = root; 507 lefttreemax = root; 508 } else 509 break; 510 } 511 /* Assemble the new root. */ 512 lefttreemax->right = root->left; 513 righttreemin->left = root->right; 514 root->left = dummy.right; 515 root->right = dummy.left; 516 return (root); 517} 518 519/* 520 * vm_page_insert: [ internal use only ] 521 * 522 * Inserts the given mem entry into the object and object list. 523 * 524 * The pagetables are not updated but will presumably fault the page 525 * in if necessary, or if a kernel page the caller will at some point 526 * enter the page into the kernel's pmap. We are not allowed to block 527 * here so we *can't* do this anyway. 528 * 529 * The object and page must be locked. 530 * This routine may not block. 531 */ 532void 533vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 534{ 535 vm_page_t root; 536 537 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 538 if (m->object != NULL) 539 panic("vm_page_insert: page already inserted"); 540 541 /* 542 * Record the object/offset pair in this page 543 */ 544 m->object = object; 545 m->pindex = pindex; 546 547 /* 548 * Now link into the object's ordered list of backed pages. 549 */ 550 root = object->root; 551 if (root == NULL) { 552 m->left = NULL; 553 m->right = NULL; 554 TAILQ_INSERT_TAIL(&object->memq, m, listq); 555 } else { 556 root = vm_page_splay(pindex, root); 557 if (pindex < root->pindex) { 558 m->left = root->left; 559 m->right = root; 560 root->left = NULL; 561 TAILQ_INSERT_BEFORE(root, m, listq); 562 } else if (pindex == root->pindex) 563 panic("vm_page_insert: offset already allocated"); 564 else { 565 m->right = root->right; 566 m->left = root; 567 root->right = NULL; 568 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 569 } 570 } 571 object->root = m; 572 object->generation++; 573 574 /* 575 * show that the object has one more resident page. 576 */ 577 object->resident_page_count++; 578 579 /* 580 * Since we are inserting a new and possibly dirty page, 581 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 582 */ 583 if (m->flags & PG_WRITEABLE) 584 vm_object_set_writeable_dirty(object); 585} 586 587/* 588 * vm_page_remove: 589 * NOTE: used by device pager as well -wfj 590 * 591 * Removes the given mem entry from the object/offset-page 592 * table and the object page list, but do not invalidate/terminate 593 * the backing store. 594 * 595 * The object and page must be locked. 596 * The underlying pmap entry (if any) is NOT removed here. 597 * This routine may not block. 598 */ 599void 600vm_page_remove(vm_page_t m) 601{ 602 vm_object_t object; 603 vm_page_t root; 604 605 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 606 if ((object = m->object) == NULL) 607 return; 608 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 609 if (m->flags & PG_BUSY) { 610 vm_page_flag_clear(m, PG_BUSY); 611 vm_page_flash(m); 612 } 613 614 /* 615 * Now remove from the object's list of backed pages. 616 */ 617 if (m != object->root) 618 vm_page_splay(m->pindex, object->root); 619 if (m->left == NULL) 620 root = m->right; 621 else { 622 root = vm_page_splay(m->pindex, m->left); 623 root->right = m->right; 624 } 625 object->root = root; 626 TAILQ_REMOVE(&object->memq, m, listq); 627 628 /* 629 * And show that the object has one fewer resident page. 630 */ 631 object->resident_page_count--; 632 object->generation++; 633 634 m->object = NULL; 635} 636 637/* 638 * vm_page_lookup: 639 * 640 * Returns the page associated with the object/offset 641 * pair specified; if none is found, NULL is returned. 642 * 643 * The object must be locked. 644 * This routine may not block. 645 * This is a critical path routine 646 */ 647vm_page_t 648vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 649{ 650 vm_page_t m; 651 652 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 653 if ((m = object->root) != NULL && m->pindex != pindex) { 654 m = vm_page_splay(pindex, m); 655 if ((object->root = m)->pindex != pindex) 656 m = NULL; 657 } 658 return (m); 659} 660 661/* 662 * vm_page_rename: 663 * 664 * Move the given memory entry from its 665 * current object to the specified target object/offset. 666 * 667 * The object must be locked. 668 * This routine may not block. 669 * 670 * Note: swap associated with the page must be invalidated by the move. We 671 * have to do this for several reasons: (1) we aren't freeing the 672 * page, (2) we are dirtying the page, (3) the VM system is probably 673 * moving the page from object A to B, and will then later move 674 * the backing store from A to B and we can't have a conflict. 675 * 676 * Note: we *always* dirty the page. It is necessary both for the 677 * fact that we moved it, and because we may be invalidating 678 * swap. If the page is on the cache, we have to deactivate it 679 * or vm_page_dirty() will panic. Dirty pages are not allowed 680 * on the cache. 681 */ 682void 683vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 684{ 685 686 vm_page_remove(m); 687 vm_page_insert(m, new_object, new_pindex); 688 if (m->queue - m->pc == PQ_CACHE) 689 vm_page_deactivate(m); 690 vm_page_dirty(m); 691} 692 693/* 694 * vm_page_select_cache: 695 * 696 * Find a page on the cache queue with color optimization. As pages 697 * might be found, but not applicable, they are deactivated. This 698 * keeps us from using potentially busy cached pages. 699 * 700 * This routine may not block. 701 */ 702vm_page_t 703vm_page_select_cache(int color) 704{ 705 vm_page_t m; 706 707 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 708 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 709 if (m->hold_count == 0 && (VM_OBJECT_TRYLOCK(m->object) || 710 VM_OBJECT_LOCKED(m->object))) { 711 KASSERT((m->flags & PG_BUSY) == 0 && m->busy == 0, 712 ("Found busy cache page %p", m)); 713 KASSERT(m->dirty == 0, 714 ("Found dirty cache page %p", m)); 715 KASSERT(!pmap_page_is_mapped(m), 716 ("Found mapped cache page %p", m)); 717 KASSERT((m->flags & PG_UNMANAGED) == 0, 718 ("Found unmanaged cache page %p", m)); 719 KASSERT(m->wire_count == 0, 720 ("Found wired cache page %p", m)); 721 break; 722 } 723 vm_page_deactivate(m); 724 } 725 return (m); 726} 727 728/* 729 * vm_page_alloc: 730 * 731 * Allocate and return a memory cell associated 732 * with this VM object/offset pair. 733 * 734 * page_req classes: 735 * VM_ALLOC_NORMAL normal process request 736 * VM_ALLOC_SYSTEM system *really* needs a page 737 * VM_ALLOC_INTERRUPT interrupt time request 738 * VM_ALLOC_ZERO zero page 739 * 740 * This routine may not block. 741 * 742 * Additional special handling is required when called from an 743 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 744 * the page cache in this case. 745 */ 746vm_page_t 747vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 748{ 749 vm_object_t m_object; 750 vm_page_t m = NULL; 751 int color, flags, page_req; 752 753 page_req = req & VM_ALLOC_CLASS_MASK; 754 KASSERT(curthread->td_intr_nesting_level == 0 || 755 page_req == VM_ALLOC_INTERRUPT, 756 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 757 758 if ((req & VM_ALLOC_NOOBJ) == 0) { 759 KASSERT(object != NULL, 760 ("vm_page_alloc: NULL object.")); 761 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 762 color = (pindex + object->pg_color) & PQ_L2_MASK; 763 } else 764 color = pindex & PQ_L2_MASK; 765 766 /* 767 * The pager is allowed to eat deeper into the free page list. 768 */ 769 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 770 page_req = VM_ALLOC_SYSTEM; 771 }; 772 773loop: 774 mtx_lock_spin(&vm_page_queue_free_mtx); 775 if (cnt.v_free_count > cnt.v_free_reserved || 776 (page_req == VM_ALLOC_SYSTEM && 777 cnt.v_cache_count == 0 && 778 cnt.v_free_count > cnt.v_interrupt_free_min) || 779 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 780 /* 781 * Allocate from the free queue if the number of free pages 782 * exceeds the minimum for the request class. 783 */ 784 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 785 } else if (page_req != VM_ALLOC_INTERRUPT) { 786 mtx_unlock_spin(&vm_page_queue_free_mtx); 787 /* 788 * Allocatable from cache (non-interrupt only). On success, 789 * we must free the page and try again, thus ensuring that 790 * cnt.v_*_free_min counters are replenished. 791 */ 792 vm_page_lock_queues(); 793 if ((m = vm_page_select_cache(color)) == NULL) { 794#if defined(DIAGNOSTIC) 795 if (cnt.v_cache_count > 0) 796 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 797#endif 798 vm_page_unlock_queues(); 799 atomic_add_int(&vm_pageout_deficit, 1); 800 pagedaemon_wakeup(); 801 return (NULL); 802 } 803 m_object = m->object; 804 VM_OBJECT_LOCK_ASSERT(m_object, MA_OWNED); 805 vm_page_free(m); 806 vm_page_unlock_queues(); 807 if (m_object != object) 808 VM_OBJECT_UNLOCK(m_object); 809 goto loop; 810 } else { 811 /* 812 * Not allocatable from cache from interrupt, give up. 813 */ 814 mtx_unlock_spin(&vm_page_queue_free_mtx); 815 atomic_add_int(&vm_pageout_deficit, 1); 816 pagedaemon_wakeup(); 817 return (NULL); 818 } 819 820 /* 821 * At this point we had better have found a good page. 822 */ 823 824 KASSERT( 825 m != NULL, 826 ("vm_page_alloc(): missing page on free queue") 827 ); 828 829 /* 830 * Remove from free queue 831 */ 832 vm_pageq_remove_nowakeup(m); 833 834 /* 835 * Initialize structure. Only the PG_ZERO flag is inherited. 836 */ 837 flags = PG_BUSY; 838 if (m->flags & PG_ZERO) { 839 vm_page_zero_count--; 840 if (req & VM_ALLOC_ZERO) 841 flags = PG_ZERO | PG_BUSY; 842 } 843 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 844 flags &= ~PG_BUSY; 845 m->flags = flags; 846 if (req & VM_ALLOC_WIRED) { 847 atomic_add_int(&cnt.v_wire_count, 1); 848 m->wire_count = 1; 849 } else 850 m->wire_count = 0; 851 m->hold_count = 0; 852 m->act_count = 0; 853 m->busy = 0; 854 m->valid = 0; 855 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 856 mtx_unlock_spin(&vm_page_queue_free_mtx); 857 858 if ((req & VM_ALLOC_NOOBJ) == 0) 859 vm_page_insert(m, object, pindex); 860 else 861 m->pindex = pindex; 862 863 /* 864 * Don't wakeup too often - wakeup the pageout daemon when 865 * we would be nearly out of memory. 866 */ 867 if (vm_paging_needed()) 868 pagedaemon_wakeup(); 869 870 return (m); 871} 872 873/* 874 * vm_wait: (also see VM_WAIT macro) 875 * 876 * Block until free pages are available for allocation 877 * - Called in various places before memory allocations. 878 */ 879void 880vm_wait(void) 881{ 882 883 vm_page_lock_queues(); 884 if (curproc == pageproc) { 885 vm_pageout_pages_needed = 1; 886 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 887 PDROP | PSWP, "VMWait", 0); 888 } else { 889 if (!vm_pages_needed) { 890 vm_pages_needed = 1; 891 wakeup(&vm_pages_needed); 892 } 893 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 894 "vmwait", 0); 895 } 896} 897 898/* 899 * vm_waitpfault: (also see VM_WAITPFAULT macro) 900 * 901 * Block until free pages are available for allocation 902 * - Called only in vm_fault so that processes page faulting 903 * can be easily tracked. 904 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 905 * processes will be able to grab memory first. Do not change 906 * this balance without careful testing first. 907 */ 908void 909vm_waitpfault(void) 910{ 911 912 vm_page_lock_queues(); 913 if (!vm_pages_needed) { 914 vm_pages_needed = 1; 915 wakeup(&vm_pages_needed); 916 } 917 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 918 "pfault", 0); 919} 920 921/* 922 * vm_page_activate: 923 * 924 * Put the specified page on the active list (if appropriate). 925 * Ensure that act_count is at least ACT_INIT but do not otherwise 926 * mess with it. 927 * 928 * The page queues must be locked. 929 * This routine may not block. 930 */ 931void 932vm_page_activate(vm_page_t m) 933{ 934 935 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 936 if (m->queue != PQ_ACTIVE) { 937 if ((m->queue - m->pc) == PQ_CACHE) 938 cnt.v_reactivated++; 939 vm_pageq_remove(m); 940 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 941 if (m->act_count < ACT_INIT) 942 m->act_count = ACT_INIT; 943 vm_pageq_enqueue(PQ_ACTIVE, m); 944 } 945 } else { 946 if (m->act_count < ACT_INIT) 947 m->act_count = ACT_INIT; 948 } 949} 950 951/* 952 * vm_page_free_wakeup: 953 * 954 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 955 * routine is called when a page has been added to the cache or free 956 * queues. 957 * 958 * The page queues must be locked. 959 * This routine may not block. 960 */ 961static __inline void 962vm_page_free_wakeup(void) 963{ 964 965 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 966 /* 967 * if pageout daemon needs pages, then tell it that there are 968 * some free. 969 */ 970 if (vm_pageout_pages_needed && 971 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 972 wakeup(&vm_pageout_pages_needed); 973 vm_pageout_pages_needed = 0; 974 } 975 /* 976 * wakeup processes that are waiting on memory if we hit a 977 * high water mark. And wakeup scheduler process if we have 978 * lots of memory. this process will swapin processes. 979 */ 980 if (vm_pages_needed && !vm_page_count_min()) { 981 vm_pages_needed = 0; 982 wakeup(&cnt.v_free_count); 983 } 984} 985 986/* 987 * vm_page_free_toq: 988 * 989 * Returns the given page to the PQ_FREE list, 990 * disassociating it with any VM object. 991 * 992 * Object and page must be locked prior to entry. 993 * This routine may not block. 994 */ 995 996void 997vm_page_free_toq(vm_page_t m) 998{ 999 struct vpgqueues *pq; 1000 vm_object_t object = m->object; 1001 1002 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1003 cnt.v_tfree++; 1004 1005 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1006 printf( 1007 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1008 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1009 m->hold_count); 1010 if ((m->queue - m->pc) == PQ_FREE) 1011 panic("vm_page_free: freeing free page"); 1012 else 1013 panic("vm_page_free: freeing busy page"); 1014 } 1015 1016 /* 1017 * unqueue, then remove page. Note that we cannot destroy 1018 * the page here because we do not want to call the pager's 1019 * callback routine until after we've put the page on the 1020 * appropriate free queue. 1021 */ 1022 vm_pageq_remove_nowakeup(m); 1023 vm_page_remove(m); 1024 1025 /* 1026 * If fictitious remove object association and 1027 * return, otherwise delay object association removal. 1028 */ 1029 if ((m->flags & PG_FICTITIOUS) != 0) { 1030 return; 1031 } 1032 1033 m->valid = 0; 1034 vm_page_undirty(m); 1035 1036 if (m->wire_count != 0) { 1037 if (m->wire_count > 1) { 1038 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1039 m->wire_count, (long)m->pindex); 1040 } 1041 panic("vm_page_free: freeing wired page"); 1042 } 1043 1044 /* 1045 * If we've exhausted the object's resident pages we want to free 1046 * it up. 1047 */ 1048 if (object && 1049 (object->type == OBJT_VNODE) && 1050 ((object->flags & OBJ_DEAD) == 0) 1051 ) { 1052 struct vnode *vp = (struct vnode *)object->handle; 1053 1054 if (vp) { 1055 VI_LOCK(vp); 1056 if (VSHOULDFREE(vp)) 1057 vfree(vp); 1058 VI_UNLOCK(vp); 1059 } 1060 } 1061 1062 /* 1063 * Clear the UNMANAGED flag when freeing an unmanaged page. 1064 */ 1065 if (m->flags & PG_UNMANAGED) { 1066 m->flags &= ~PG_UNMANAGED; 1067 } 1068 1069 if (m->hold_count != 0) { 1070 m->flags &= ~PG_ZERO; 1071 m->queue = PQ_HOLD; 1072 } else 1073 m->queue = PQ_FREE + m->pc; 1074 pq = &vm_page_queues[m->queue]; 1075 mtx_lock_spin(&vm_page_queue_free_mtx); 1076 pq->lcnt++; 1077 ++(*pq->cnt); 1078 1079 /* 1080 * Put zero'd pages on the end ( where we look for zero'd pages 1081 * first ) and non-zerod pages at the head. 1082 */ 1083 if (m->flags & PG_ZERO) { 1084 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1085 ++vm_page_zero_count; 1086 } else { 1087 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1088 } 1089 mtx_unlock_spin(&vm_page_queue_free_mtx); 1090 vm_page_free_wakeup(); 1091} 1092 1093/* 1094 * vm_page_unmanage: 1095 * 1096 * Prevent PV management from being done on the page. The page is 1097 * removed from the paging queues as if it were wired, and as a 1098 * consequence of no longer being managed the pageout daemon will not 1099 * touch it (since there is no way to locate the pte mappings for the 1100 * page). madvise() calls that mess with the pmap will also no longer 1101 * operate on the page. 1102 * 1103 * Beyond that the page is still reasonably 'normal'. Freeing the page 1104 * will clear the flag. 1105 * 1106 * This routine is used by OBJT_PHYS objects - objects using unswappable 1107 * physical memory as backing store rather then swap-backed memory and 1108 * will eventually be extended to support 4MB unmanaged physical 1109 * mappings. 1110 */ 1111void 1112vm_page_unmanage(vm_page_t m) 1113{ 1114 1115 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1116 if ((m->flags & PG_UNMANAGED) == 0) { 1117 if (m->wire_count == 0) 1118 vm_pageq_remove(m); 1119 } 1120 vm_page_flag_set(m, PG_UNMANAGED); 1121} 1122 1123/* 1124 * vm_page_wire: 1125 * 1126 * Mark this page as wired down by yet 1127 * another map, removing it from paging queues 1128 * as necessary. 1129 * 1130 * The page queues must be locked. 1131 * This routine may not block. 1132 */ 1133void 1134vm_page_wire(vm_page_t m) 1135{ 1136 1137 /* 1138 * Only bump the wire statistics if the page is not already wired, 1139 * and only unqueue the page if it is on some queue (if it is unmanaged 1140 * it is already off the queues). 1141 */ 1142 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1143 if (m->flags & PG_FICTITIOUS) 1144 return; 1145 if (m->wire_count == 0) { 1146 if ((m->flags & PG_UNMANAGED) == 0) 1147 vm_pageq_remove(m); 1148 atomic_add_int(&cnt.v_wire_count, 1); 1149 } 1150 m->wire_count++; 1151 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1152} 1153 1154/* 1155 * vm_page_unwire: 1156 * 1157 * Release one wiring of this page, potentially 1158 * enabling it to be paged again. 1159 * 1160 * Many pages placed on the inactive queue should actually go 1161 * into the cache, but it is difficult to figure out which. What 1162 * we do instead, if the inactive target is well met, is to put 1163 * clean pages at the head of the inactive queue instead of the tail. 1164 * This will cause them to be moved to the cache more quickly and 1165 * if not actively re-referenced, freed more quickly. If we just 1166 * stick these pages at the end of the inactive queue, heavy filesystem 1167 * meta-data accesses can cause an unnecessary paging load on memory bound 1168 * processes. This optimization causes one-time-use metadata to be 1169 * reused more quickly. 1170 * 1171 * BUT, if we are in a low-memory situation we have no choice but to 1172 * put clean pages on the cache queue. 1173 * 1174 * A number of routines use vm_page_unwire() to guarantee that the page 1175 * will go into either the inactive or active queues, and will NEVER 1176 * be placed in the cache - for example, just after dirtying a page. 1177 * dirty pages in the cache are not allowed. 1178 * 1179 * The page queues must be locked. 1180 * This routine may not block. 1181 */ 1182void 1183vm_page_unwire(vm_page_t m, int activate) 1184{ 1185 1186 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1187 if (m->flags & PG_FICTITIOUS) 1188 return; 1189 if (m->wire_count > 0) { 1190 m->wire_count--; 1191 if (m->wire_count == 0) { 1192 atomic_subtract_int(&cnt.v_wire_count, 1); 1193 if (m->flags & PG_UNMANAGED) { 1194 ; 1195 } else if (activate) 1196 vm_pageq_enqueue(PQ_ACTIVE, m); 1197 else { 1198 vm_page_flag_clear(m, PG_WINATCFLS); 1199 vm_pageq_enqueue(PQ_INACTIVE, m); 1200 } 1201 } 1202 } else { 1203 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1204 } 1205} 1206 1207 1208/* 1209 * Move the specified page to the inactive queue. If the page has 1210 * any associated swap, the swap is deallocated. 1211 * 1212 * Normally athead is 0 resulting in LRU operation. athead is set 1213 * to 1 if we want this page to be 'as if it were placed in the cache', 1214 * except without unmapping it from the process address space. 1215 * 1216 * This routine may not block. 1217 */ 1218static __inline void 1219_vm_page_deactivate(vm_page_t m, int athead) 1220{ 1221 1222 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1223 1224 /* 1225 * Ignore if already inactive. 1226 */ 1227 if (m->queue == PQ_INACTIVE) 1228 return; 1229 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1230 if ((m->queue - m->pc) == PQ_CACHE) 1231 cnt.v_reactivated++; 1232 vm_page_flag_clear(m, PG_WINATCFLS); 1233 vm_pageq_remove(m); 1234 if (athead) 1235 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1236 else 1237 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1238 m->queue = PQ_INACTIVE; 1239 vm_page_queues[PQ_INACTIVE].lcnt++; 1240 cnt.v_inactive_count++; 1241 } 1242} 1243 1244void 1245vm_page_deactivate(vm_page_t m) 1246{ 1247 _vm_page_deactivate(m, 0); 1248} 1249 1250/* 1251 * vm_page_try_to_cache: 1252 * 1253 * Returns 0 on failure, 1 on success 1254 */ 1255int 1256vm_page_try_to_cache(vm_page_t m) 1257{ 1258 1259 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1260 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1261 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1262 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1263 return (0); 1264 } 1265 pmap_remove_all(m); 1266 if (m->dirty) 1267 return (0); 1268 vm_page_cache(m); 1269 return (1); 1270} 1271 1272/* 1273 * vm_page_try_to_free() 1274 * 1275 * Attempt to free the page. If we cannot free it, we do nothing. 1276 * 1 is returned on success, 0 on failure. 1277 */ 1278int 1279vm_page_try_to_free(vm_page_t m) 1280{ 1281 1282 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1283 if (m->object != NULL) 1284 VM_OBJECT_LOCK_ASSERT(m->object, 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_free(m); 1293 return (1); 1294} 1295 1296/* 1297 * vm_page_cache 1298 * 1299 * Put the specified page onto the page cache queue (if appropriate). 1300 * 1301 * This routine may not block. 1302 */ 1303void 1304vm_page_cache(vm_page_t m) 1305{ 1306 1307 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1308 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1309 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1310 m->hold_count || m->wire_count) { 1311 printf("vm_page_cache: attempting to cache busy page\n"); 1312 return; 1313 } 1314 if ((m->queue - m->pc) == PQ_CACHE) 1315 return; 1316 1317 /* 1318 * Remove all pmaps and indicate that the page is not 1319 * writeable or mapped. 1320 */ 1321 pmap_remove_all(m); 1322 if (m->dirty != 0) { 1323 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1324 (long)m->pindex); 1325 } 1326 vm_pageq_remove_nowakeup(m); 1327 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1328 vm_page_free_wakeup(); 1329} 1330 1331/* 1332 * vm_page_dontneed 1333 * 1334 * Cache, deactivate, or do nothing as appropriate. This routine 1335 * is typically used by madvise() MADV_DONTNEED. 1336 * 1337 * Generally speaking we want to move the page into the cache so 1338 * it gets reused quickly. However, this can result in a silly syndrome 1339 * due to the page recycling too quickly. Small objects will not be 1340 * fully cached. On the otherhand, if we move the page to the inactive 1341 * queue we wind up with a problem whereby very large objects 1342 * unnecessarily blow away our inactive and cache queues. 1343 * 1344 * The solution is to move the pages based on a fixed weighting. We 1345 * either leave them alone, deactivate them, or move them to the cache, 1346 * where moving them to the cache has the highest weighting. 1347 * By forcing some pages into other queues we eventually force the 1348 * system to balance the queues, potentially recovering other unrelated 1349 * space from active. The idea is to not force this to happen too 1350 * often. 1351 */ 1352void 1353vm_page_dontneed(vm_page_t m) 1354{ 1355 static int dnweight; 1356 int dnw; 1357 int head; 1358 1359 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1360 dnw = ++dnweight; 1361 1362 /* 1363 * occassionally leave the page alone 1364 */ 1365 if ((dnw & 0x01F0) == 0 || 1366 m->queue == PQ_INACTIVE || 1367 m->queue - m->pc == PQ_CACHE 1368 ) { 1369 if (m->act_count >= ACT_INIT) 1370 --m->act_count; 1371 return; 1372 } 1373 1374 if (m->dirty == 0 && pmap_is_modified(m)) 1375 vm_page_dirty(m); 1376 1377 if (m->dirty || (dnw & 0x0070) == 0) { 1378 /* 1379 * Deactivate the page 3 times out of 32. 1380 */ 1381 head = 0; 1382 } else { 1383 /* 1384 * Cache the page 28 times out of every 32. Note that 1385 * the page is deactivated instead of cached, but placed 1386 * at the head of the queue instead of the tail. 1387 */ 1388 head = 1; 1389 } 1390 _vm_page_deactivate(m, head); 1391} 1392 1393/* 1394 * Grab a page, waiting until we are waken up due to the page 1395 * changing state. We keep on waiting, if the page continues 1396 * to be in the object. If the page doesn't exist, first allocate it 1397 * and then conditionally zero it. 1398 * 1399 * This routine may block. 1400 */ 1401vm_page_t 1402vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1403{ 1404 vm_page_t m; 1405 1406 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1407retrylookup: 1408 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1409 vm_page_lock_queues(); 1410 if (m->busy || (m->flags & PG_BUSY)) { 1411 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1412 VM_OBJECT_UNLOCK(object); 1413 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0); 1414 VM_OBJECT_LOCK(object); 1415 if ((allocflags & VM_ALLOC_RETRY) == 0) 1416 return (NULL); 1417 goto retrylookup; 1418 } else { 1419 if (allocflags & VM_ALLOC_WIRED) 1420 vm_page_wire(m); 1421 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1422 vm_page_busy(m); 1423 vm_page_unlock_queues(); 1424 return (m); 1425 } 1426 } 1427 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1428 if (m == NULL) { 1429 VM_OBJECT_UNLOCK(object); 1430 VM_WAIT; 1431 VM_OBJECT_LOCK(object); 1432 if ((allocflags & VM_ALLOC_RETRY) == 0) 1433 return (NULL); 1434 goto retrylookup; 1435 } 1436 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1437 pmap_zero_page(m); 1438 return (m); 1439} 1440 1441/* 1442 * Mapping function for valid bits or for dirty bits in 1443 * a page. May not block. 1444 * 1445 * Inputs are required to range within a page. 1446 */ 1447__inline int 1448vm_page_bits(int base, int size) 1449{ 1450 int first_bit; 1451 int last_bit; 1452 1453 KASSERT( 1454 base + size <= PAGE_SIZE, 1455 ("vm_page_bits: illegal base/size %d/%d", base, size) 1456 ); 1457 1458 if (size == 0) /* handle degenerate case */ 1459 return (0); 1460 1461 first_bit = base >> DEV_BSHIFT; 1462 last_bit = (base + size - 1) >> DEV_BSHIFT; 1463 1464 return ((2 << last_bit) - (1 << first_bit)); 1465} 1466 1467/* 1468 * vm_page_set_validclean: 1469 * 1470 * Sets portions of a page valid and clean. The arguments are expected 1471 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1472 * of any partial chunks touched by the range. The invalid portion of 1473 * such chunks will be zero'd. 1474 * 1475 * This routine may not block. 1476 * 1477 * (base + size) must be less then or equal to PAGE_SIZE. 1478 */ 1479void 1480vm_page_set_validclean(vm_page_t m, int base, int size) 1481{ 1482 int pagebits; 1483 int frag; 1484 int endoff; 1485 1486 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1487 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1488 if (size == 0) /* handle degenerate case */ 1489 return; 1490 1491 /* 1492 * If the base is not DEV_BSIZE aligned and the valid 1493 * bit is clear, we have to zero out a portion of the 1494 * first block. 1495 */ 1496 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1497 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1498 pmap_zero_page_area(m, frag, base - frag); 1499 1500 /* 1501 * If the ending offset is not DEV_BSIZE aligned and the 1502 * valid bit is clear, we have to zero out a portion of 1503 * the last block. 1504 */ 1505 endoff = base + size; 1506 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1507 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1508 pmap_zero_page_area(m, endoff, 1509 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1510 1511 /* 1512 * Set valid, clear dirty bits. If validating the entire 1513 * page we can safely clear the pmap modify bit. We also 1514 * use this opportunity to clear the PG_NOSYNC flag. If a process 1515 * takes a write fault on a MAP_NOSYNC memory area the flag will 1516 * be set again. 1517 * 1518 * We set valid bits inclusive of any overlap, but we can only 1519 * clear dirty bits for DEV_BSIZE chunks that are fully within 1520 * the range. 1521 */ 1522 pagebits = vm_page_bits(base, size); 1523 m->valid |= pagebits; 1524#if 0 /* NOT YET */ 1525 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1526 frag = DEV_BSIZE - frag; 1527 base += frag; 1528 size -= frag; 1529 if (size < 0) 1530 size = 0; 1531 } 1532 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1533#endif 1534 m->dirty &= ~pagebits; 1535 if (base == 0 && size == PAGE_SIZE) { 1536 pmap_clear_modify(m); 1537 vm_page_flag_clear(m, PG_NOSYNC); 1538 } 1539} 1540 1541void 1542vm_page_clear_dirty(vm_page_t m, int base, int size) 1543{ 1544 1545 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1546 m->dirty &= ~vm_page_bits(base, size); 1547} 1548 1549/* 1550 * vm_page_set_invalid: 1551 * 1552 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1553 * valid and dirty bits for the effected areas are cleared. 1554 * 1555 * May not block. 1556 */ 1557void 1558vm_page_set_invalid(vm_page_t m, int base, int size) 1559{ 1560 int bits; 1561 1562 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1563 bits = vm_page_bits(base, size); 1564 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1565 m->valid &= ~bits; 1566 m->dirty &= ~bits; 1567 m->object->generation++; 1568} 1569 1570/* 1571 * vm_page_zero_invalid() 1572 * 1573 * The kernel assumes that the invalid portions of a page contain 1574 * garbage, but such pages can be mapped into memory by user code. 1575 * When this occurs, we must zero out the non-valid portions of the 1576 * page so user code sees what it expects. 1577 * 1578 * Pages are most often semi-valid when the end of a file is mapped 1579 * into memory and the file's size is not page aligned. 1580 */ 1581void 1582vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1583{ 1584 int b; 1585 int i; 1586 1587 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1588 /* 1589 * Scan the valid bits looking for invalid sections that 1590 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1591 * valid bit may be set ) have already been zerod by 1592 * vm_page_set_validclean(). 1593 */ 1594 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1595 if (i == (PAGE_SIZE / DEV_BSIZE) || 1596 (m->valid & (1 << i)) 1597 ) { 1598 if (i > b) { 1599 pmap_zero_page_area(m, 1600 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1601 } 1602 b = i + 1; 1603 } 1604 } 1605 1606 /* 1607 * setvalid is TRUE when we can safely set the zero'd areas 1608 * as being valid. We can do this if there are no cache consistancy 1609 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1610 */ 1611 if (setvalid) 1612 m->valid = VM_PAGE_BITS_ALL; 1613} 1614 1615/* 1616 * vm_page_is_valid: 1617 * 1618 * Is (partial) page valid? Note that the case where size == 0 1619 * will return FALSE in the degenerate case where the page is 1620 * entirely invalid, and TRUE otherwise. 1621 * 1622 * May not block. 1623 */ 1624int 1625vm_page_is_valid(vm_page_t m, int base, int size) 1626{ 1627 int bits = vm_page_bits(base, size); 1628 1629 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1630 if (m->valid && ((m->valid & bits) == bits)) 1631 return 1; 1632 else 1633 return 0; 1634} 1635 1636/* 1637 * update dirty bits from pmap/mmu. May not block. 1638 */ 1639void 1640vm_page_test_dirty(vm_page_t m) 1641{ 1642 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1643 vm_page_dirty(m); 1644 } 1645} 1646 1647int so_zerocp_fullpage = 0; 1648 1649void 1650vm_page_cowfault(vm_page_t m) 1651{ 1652 vm_page_t mnew; 1653 vm_object_t object; 1654 vm_pindex_t pindex; 1655 1656 object = m->object; 1657 pindex = m->pindex; 1658 1659 retry_alloc: 1660 vm_page_remove(m); 1661 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 1662 if (mnew == NULL) { 1663 vm_page_insert(m, object, pindex); 1664 vm_page_unlock_queues(); 1665 VM_OBJECT_UNLOCK(object); 1666 VM_WAIT; 1667 VM_OBJECT_LOCK(object); 1668 vm_page_lock_queues(); 1669 goto retry_alloc; 1670 } 1671 1672 if (m->cow == 0) { 1673 /* 1674 * check to see if we raced with an xmit complete when 1675 * waiting to allocate a page. If so, put things back 1676 * the way they were 1677 */ 1678 vm_page_free(mnew); 1679 vm_page_insert(m, object, pindex); 1680 } else { /* clear COW & copy page */ 1681 if (!so_zerocp_fullpage) 1682 pmap_copy_page(m, mnew); 1683 mnew->valid = VM_PAGE_BITS_ALL; 1684 vm_page_dirty(mnew); 1685 vm_page_flag_clear(mnew, PG_BUSY); 1686 } 1687} 1688 1689void 1690vm_page_cowclear(vm_page_t m) 1691{ 1692 1693 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1694 if (m->cow) { 1695 m->cow--; 1696 /* 1697 * let vm_fault add back write permission lazily 1698 */ 1699 } 1700 /* 1701 * sf_buf_free() will free the page, so we needn't do it here 1702 */ 1703} 1704 1705void 1706vm_page_cowsetup(vm_page_t m) 1707{ 1708 1709 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1710 m->cow++; 1711 pmap_page_protect(m, VM_PROT_READ); 1712} 1713 1714#include "opt_ddb.h" 1715#ifdef DDB 1716#include <sys/kernel.h> 1717 1718#include <ddb/ddb.h> 1719 1720DB_SHOW_COMMAND(page, vm_page_print_page_info) 1721{ 1722 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1723 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1724 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1725 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1726 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1727 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1728 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1729 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1730 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1731 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1732} 1733 1734DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1735{ 1736 int i; 1737 db_printf("PQ_FREE:"); 1738 for (i = 0; i < PQ_L2_SIZE; i++) { 1739 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1740 } 1741 db_printf("\n"); 1742 1743 db_printf("PQ_CACHE:"); 1744 for (i = 0; i < PQ_L2_SIZE; i++) { 1745 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1746 } 1747 db_printf("\n"); 1748 1749 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1750 vm_page_queues[PQ_ACTIVE].lcnt, 1751 vm_page_queues[PQ_INACTIVE].lcnt); 1752} 1753#endif /* DDB */ 1754