vm_page.c revision 113856
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 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 37 * $FreeBSD: head/sys/vm/vm_page.c 113856 2003-04-22 14:26:02Z alc $ 38 */ 39 40/* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 67/* 68 * GENERAL RULES ON VM_PAGE MANIPULATION 69 * 70 * - a pageq mutex is required when adding or removing a page from a 71 * page queue (vm_page_queue[]), regardless of other mutexes or the 72 * busy state of a page. 73 * 74 * - a hash chain mutex is required when associating or disassociating 75 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 76 * regardless of other mutexes or the busy state of a page. 77 * 78 * - either a hash chain mutex OR a busied page is required in order 79 * to modify the page flags. A hash chain mutex must be obtained in 80 * order to busy a page. A page's flags cannot be modified by a 81 * hash chain mutex if the page is marked busy. 82 * 83 * - The object memq mutex is held when inserting or removing 84 * pages from an object (vm_page_insert() or vm_page_remove()). This 85 * is different from the object's main mutex. 86 * 87 * Generally speaking, you have to be aware of side effects when running 88 * vm_page ops. A vm_page_lookup() will return with the hash chain 89 * locked, whether it was able to lookup the page or not. vm_page_free(), 90 * vm_page_cache(), vm_page_activate(), and a number of other routines 91 * will release the hash chain mutex for you. Intermediate manipulation 92 * routines such as vm_page_flag_set() expect the hash chain to be held 93 * on entry and the hash chain will remain held on return. 94 * 95 * pageq scanning can only occur with the pageq in question locked. 96 * We have a known bottleneck with the active queue, but the cache 97 * and free queues are actually arrays already. 98 */ 99 100/* 101 * Resident memory management module. 102 */ 103 104#include <sys/param.h> 105#include <sys/systm.h> 106#include <sys/lock.h> 107#include <sys/malloc.h> 108#include <sys/mutex.h> 109#include <sys/proc.h> 110#include <sys/vmmeter.h> 111#include <sys/vnode.h> 112 113#include <vm/vm.h> 114#include <vm/vm_param.h> 115#include <vm/vm_kern.h> 116#include <vm/vm_object.h> 117#include <vm/vm_page.h> 118#include <vm/vm_pageout.h> 119#include <vm/vm_pager.h> 120#include <vm/vm_extern.h> 121#include <vm/uma.h> 122#include <vm/uma_int.h> 123 124/* 125 * Associated with page of user-allocatable memory is a 126 * page structure. 127 */ 128 129struct mtx vm_page_queue_mtx; 130struct mtx vm_page_queue_free_mtx; 131 132vm_page_t vm_page_array = 0; 133int vm_page_array_size = 0; 134long first_page = 0; 135int vm_page_zero_count = 0; 136 137/* 138 * vm_set_page_size: 139 * 140 * Sets the page size, perhaps based upon the memory 141 * size. Must be called before any use of page-size 142 * dependent functions. 143 */ 144void 145vm_set_page_size(void) 146{ 147 if (cnt.v_page_size == 0) 148 cnt.v_page_size = PAGE_SIZE; 149 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 150 panic("vm_set_page_size: page size not a power of two"); 151} 152 153/* 154 * vm_page_startup: 155 * 156 * Initializes the resident memory module. 157 * 158 * Allocates memory for the page cells, and 159 * for the object/offset-to-page hash table headers. 160 * Each page cell is initialized and placed on the free list. 161 */ 162vm_offset_t 163vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 164{ 165 vm_offset_t mapped; 166 vm_size_t npages; 167 vm_paddr_t page_range; 168 vm_paddr_t new_end; 169 int i; 170 vm_paddr_t pa; 171 int nblocks; 172 vm_paddr_t last_pa; 173 174 /* the biggest memory array is the second group of pages */ 175 vm_paddr_t end; 176 vm_paddr_t biggestsize; 177 int biggestone; 178 179 vm_paddr_t total; 180 vm_size_t bootpages; 181 182 total = 0; 183 biggestsize = 0; 184 biggestone = 0; 185 nblocks = 0; 186 vaddr = round_page(vaddr); 187 188 for (i = 0; phys_avail[i + 1]; i += 2) { 189 phys_avail[i] = round_page(phys_avail[i]); 190 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 191 } 192 193 for (i = 0; phys_avail[i + 1]; i += 2) { 194 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 195 196 if (size > biggestsize) { 197 biggestone = i; 198 biggestsize = size; 199 } 200 ++nblocks; 201 total += size; 202 } 203 204 end = phys_avail[biggestone+1]; 205 206 /* 207 * Initialize the locks. 208 */ 209 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF); 210 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 211 MTX_SPIN); 212 213 /* 214 * Initialize the queue headers for the free queue, the active queue 215 * and the inactive queue. 216 */ 217 vm_pageq_init(); 218 219 /* 220 * Allocate memory for use when boot strapping the kernel memory 221 * allocator. 222 */ 223 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 224 new_end = end - bootpages; 225 new_end = trunc_page(new_end); 226 mapped = pmap_map(&vaddr, new_end, end, 227 VM_PROT_READ | VM_PROT_WRITE); 228 bzero((caddr_t) mapped, end - new_end); 229 uma_startup((caddr_t)mapped); 230 231 /* 232 * Compute the number of pages of memory that will be available for 233 * use (taking into account the overhead of a page structure per 234 * page). 235 */ 236 first_page = phys_avail[0] / PAGE_SIZE; 237 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 238 npages = (total - (page_range * sizeof(struct vm_page)) - 239 (end - new_end)) / PAGE_SIZE; 240 end = new_end; 241 242 /* 243 * Initialize the mem entry structures now, and put them in the free 244 * queue. 245 */ 246 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 247 mapped = pmap_map(&vaddr, new_end, end, 248 VM_PROT_READ | VM_PROT_WRITE); 249 vm_page_array = (vm_page_t) mapped; 250 phys_avail[biggestone + 1] = new_end; 251 252 /* 253 * Clear all of the page structures 254 */ 255 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 256 vm_page_array_size = page_range; 257 258 /* 259 * Construct the free queue(s) in descending order (by physical 260 * address) so that the first 16MB of physical memory is allocated 261 * last rather than first. On large-memory machines, this avoids 262 * the exhaustion of low physical memory before isa_dmainit has run. 263 */ 264 cnt.v_page_count = 0; 265 cnt.v_free_count = 0; 266 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 267 pa = phys_avail[i]; 268 last_pa = phys_avail[i + 1]; 269 while (pa < last_pa && npages-- > 0) { 270 vm_pageq_add_new_page(pa); 271 pa += PAGE_SIZE; 272 } 273 } 274 return (vaddr); 275} 276 277void 278vm_page_flag_set(vm_page_t m, unsigned short bits) 279{ 280 281 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 282 m->flags |= bits; 283} 284 285void 286vm_page_flag_clear(vm_page_t m, unsigned short bits) 287{ 288 289 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 290 m->flags &= ~bits; 291} 292 293void 294vm_page_busy(vm_page_t m) 295{ 296 KASSERT((m->flags & PG_BUSY) == 0, 297 ("vm_page_busy: page already busy!!!")); 298 vm_page_flag_set(m, PG_BUSY); 299} 300 301/* 302 * vm_page_flash: 303 * 304 * wakeup anyone waiting for the page. 305 */ 306void 307vm_page_flash(vm_page_t m) 308{ 309 if (m->flags & PG_WANTED) { 310 vm_page_flag_clear(m, PG_WANTED); 311 wakeup(m); 312 } 313} 314 315/* 316 * vm_page_wakeup: 317 * 318 * clear the PG_BUSY flag and wakeup anyone waiting for the 319 * page. 320 * 321 */ 322void 323vm_page_wakeup(vm_page_t m) 324{ 325 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 326 vm_page_flag_clear(m, PG_BUSY); 327 vm_page_flash(m); 328} 329 330void 331vm_page_io_start(vm_page_t m) 332{ 333 334 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 335 m->busy++; 336} 337 338void 339vm_page_io_finish(vm_page_t m) 340{ 341 342 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 343 m->busy--; 344 if (m->busy == 0) 345 vm_page_flash(m); 346} 347 348/* 349 * Keep page from being freed by the page daemon 350 * much of the same effect as wiring, except much lower 351 * overhead and should be used only for *very* temporary 352 * holding ("wiring"). 353 */ 354void 355vm_page_hold(vm_page_t mem) 356{ 357 358 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 359 mem->hold_count++; 360} 361 362void 363vm_page_unhold(vm_page_t mem) 364{ 365 366 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 367 --mem->hold_count; 368 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 369 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 370 vm_page_free_toq(mem); 371} 372 373/* 374 * vm_page_copy: 375 * 376 * Copy one page to another 377 */ 378void 379vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 380{ 381 pmap_copy_page(src_m, dest_m); 382 dest_m->valid = VM_PAGE_BITS_ALL; 383} 384 385/* 386 * vm_page_free: 387 * 388 * Free a page 389 * 390 * The clearing of PG_ZERO is a temporary safety until the code can be 391 * reviewed to determine that PG_ZERO is being properly cleared on 392 * write faults or maps. PG_ZERO was previously cleared in 393 * vm_page_alloc(). 394 */ 395void 396vm_page_free(vm_page_t m) 397{ 398 vm_page_flag_clear(m, PG_ZERO); 399 vm_page_free_toq(m); 400 vm_page_zero_idle_wakeup(); 401} 402 403/* 404 * vm_page_free_zero: 405 * 406 * Free a page to the zerod-pages queue 407 */ 408void 409vm_page_free_zero(vm_page_t m) 410{ 411 vm_page_flag_set(m, PG_ZERO); 412 vm_page_free_toq(m); 413} 414 415/* 416 * vm_page_sleep_if_busy: 417 * 418 * Sleep and release the page queues lock if PG_BUSY is set or, 419 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 420 * thread slept and the page queues lock was released. 421 * Otherwise, retains the page queues lock and returns FALSE. 422 */ 423int 424vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 425{ 426 int is_object_locked; 427 428 mtx_assert(&vm_page_queue_mtx, 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 * Remove mtx_owned() after vm_object locking is finished. 433 */ 434 if ((is_object_locked = m->object != NULL && 435 mtx_owned(&m->object->mtx))) 436 mtx_unlock(&m->object->mtx); 437 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 438 if (is_object_locked) 439 mtx_lock(&m->object->mtx); 440 return (TRUE); 441 } 442 return (FALSE); 443} 444 445/* 446 * vm_page_dirty: 447 * 448 * make page all dirty 449 */ 450void 451vm_page_dirty(vm_page_t m) 452{ 453 KASSERT(m->queue - m->pc != PQ_CACHE, 454 ("vm_page_dirty: page in cache!")); 455 KASSERT(m->queue - m->pc != PQ_FREE, 456 ("vm_page_dirty: page is free!")); 457 m->dirty = VM_PAGE_BITS_ALL; 458} 459 460/* 461 * vm_page_splay: 462 * 463 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 464 * the vm_page containing the given pindex. If, however, that 465 * pindex is not found in the vm_object, returns a vm_page that is 466 * adjacent to the pindex, coming before or after it. 467 */ 468vm_page_t 469vm_page_splay(vm_pindex_t pindex, vm_page_t root) 470{ 471 struct vm_page dummy; 472 vm_page_t lefttreemax, righttreemin, y; 473 474 if (root == NULL) 475 return (root); 476 lefttreemax = righttreemin = &dummy; 477 for (;; root = y) { 478 if (pindex < root->pindex) { 479 if ((y = root->left) == NULL) 480 break; 481 if (pindex < y->pindex) { 482 /* Rotate right. */ 483 root->left = y->right; 484 y->right = root; 485 root = y; 486 if ((y = root->left) == NULL) 487 break; 488 } 489 /* Link into the new root's right tree. */ 490 righttreemin->left = root; 491 righttreemin = root; 492 } else if (pindex > root->pindex) { 493 if ((y = root->right) == NULL) 494 break; 495 if (pindex > y->pindex) { 496 /* Rotate left. */ 497 root->right = y->left; 498 y->left = root; 499 root = y; 500 if ((y = root->right) == NULL) 501 break; 502 } 503 /* Link into the new root's left tree. */ 504 lefttreemax->right = root; 505 lefttreemax = root; 506 } else 507 break; 508 } 509 /* Assemble the new root. */ 510 lefttreemax->right = root->left; 511 righttreemin->left = root->right; 512 root->left = dummy.right; 513 root->right = dummy.left; 514 return (root); 515} 516 517/* 518 * vm_page_insert: [ internal use only ] 519 * 520 * Inserts the given mem entry into the object and object list. 521 * 522 * The pagetables are not updated but will presumably fault the page 523 * in if necessary, or if a kernel page the caller will at some point 524 * enter the page into the kernel's pmap. We are not allowed to block 525 * here so we *can't* do this anyway. 526 * 527 * The object and page must be locked, and must be splhigh. 528 * This routine may not block. 529 */ 530void 531vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 532{ 533 vm_page_t root; 534 535 if (!VM_OBJECT_LOCKED(object)) 536 GIANT_REQUIRED; 537 if (m->object != NULL) 538 panic("vm_page_insert: already inserted"); 539 540 /* 541 * Record the object/offset pair in this page 542 */ 543 m->object = object; 544 m->pindex = pindex; 545 546 /* 547 * Now link into the object's ordered list of backed pages. 548 */ 549 root = object->root; 550 if (root == NULL) { 551 m->left = NULL; 552 m->right = NULL; 553 TAILQ_INSERT_TAIL(&object->memq, m, listq); 554 } else { 555 root = vm_page_splay(pindex, root); 556 if (pindex < root->pindex) { 557 m->left = root->left; 558 m->right = root; 559 root->left = NULL; 560 TAILQ_INSERT_BEFORE(root, m, listq); 561 } else { 562 m->right = root->right; 563 m->left = root; 564 root->right = NULL; 565 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 566 } 567 } 568 object->root = m; 569 object->generation++; 570 571 /* 572 * show that the object has one more resident page. 573 */ 574 object->resident_page_count++; 575 576 /* 577 * Since we are inserting a new and possibly dirty page, 578 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 579 */ 580 if (m->flags & PG_WRITEABLE) 581 vm_object_set_writeable_dirty(object); 582} 583 584/* 585 * vm_page_remove: 586 * NOTE: used by device pager as well -wfj 587 * 588 * Removes the given mem entry from the object/offset-page 589 * table and the object page list, but do not invalidate/terminate 590 * the backing store. 591 * 592 * The object and page must be locked, and at splhigh. 593 * The underlying pmap entry (if any) is NOT removed here. 594 * This routine may not block. 595 */ 596void 597vm_page_remove(vm_page_t m) 598{ 599 vm_object_t object; 600 vm_page_t root; 601 602 GIANT_REQUIRED; 603 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 604 if (m->object == NULL) 605 return; 606 607 if ((m->flags & PG_BUSY) == 0) { 608 panic("vm_page_remove: page not busy"); 609 } 610 611 /* 612 * Basically destroy the page. 613 */ 614 vm_page_wakeup(m); 615 616 object = m->object; 617 618 /* 619 * Now remove from the object's list of backed pages. 620 */ 621 if (m != object->root) 622 vm_page_splay(m->pindex, object->root); 623 if (m->left == NULL) 624 root = m->right; 625 else { 626 root = vm_page_splay(m->pindex, m->left); 627 root->right = m->right; 628 } 629 object->root = root; 630 TAILQ_REMOVE(&object->memq, m, listq); 631 632 /* 633 * And show that the object has one fewer resident page. 634 */ 635 object->resident_page_count--; 636 object->generation++; 637 638 m->object = NULL; 639} 640 641/* 642 * vm_page_lookup: 643 * 644 * Returns the page associated with the object/offset 645 * pair specified; if none is found, NULL is returned. 646 * 647 * The object must be locked. 648 * This routine may not block. 649 * This is a critical path routine 650 */ 651vm_page_t 652vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 653{ 654 vm_page_t m; 655 656 if (!VM_OBJECT_LOCKED(object)) 657 GIANT_REQUIRED; 658 m = vm_page_splay(pindex, object->root); 659 if ((object->root = m) != NULL && m->pindex != pindex) 660 m = NULL; 661 return (m); 662} 663 664/* 665 * vm_page_rename: 666 * 667 * Move the given memory entry from its 668 * current object to the specified target object/offset. 669 * 670 * The object must be locked. 671 * This routine may not block. 672 * 673 * Note: this routine will raise itself to splvm(), the caller need not. 674 * 675 * Note: swap associated with the page must be invalidated by the move. We 676 * have to do this for several reasons: (1) we aren't freeing the 677 * page, (2) we are dirtying the page, (3) the VM system is probably 678 * moving the page from object A to B, and will then later move 679 * the backing store from A to B and we can't have a conflict. 680 * 681 * Note: we *always* dirty the page. It is necessary both for the 682 * fact that we moved it, and because we may be invalidating 683 * swap. If the page is on the cache, we have to deactivate it 684 * or vm_page_dirty() will panic. Dirty pages are not allowed 685 * on the cache. 686 */ 687void 688vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 689{ 690 int s; 691 692 s = splvm(); 693 vm_page_remove(m); 694 vm_page_insert(m, new_object, new_pindex); 695 if (m->queue - m->pc == PQ_CACHE) 696 vm_page_deactivate(m); 697 vm_page_dirty(m); 698 splx(s); 699} 700 701/* 702 * vm_page_select_cache: 703 * 704 * Find a page on the cache queue with color optimization. As pages 705 * might be found, but not applicable, they are deactivated. This 706 * keeps us from using potentially busy cached pages. 707 * 708 * This routine must be called at splvm(). 709 * This routine may not block. 710 */ 711static vm_page_t 712vm_page_select_cache(vm_pindex_t color) 713{ 714 vm_page_t m; 715 716 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 717 while (TRUE) { 718 m = vm_pageq_find(PQ_CACHE, color & PQ_L2_MASK, FALSE); 719 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 720 m->hold_count || m->wire_count)) { 721 vm_page_deactivate(m); 722 continue; 723 } 724 return m; 725 } 726} 727 728/* 729 * vm_page_select_free: 730 * 731 * Find a free or zero page, with specified preference. 732 * 733 * This routine must be called at splvm(). 734 * This routine may not block. 735 */ 736static __inline vm_page_t 737vm_page_select_free(vm_pindex_t color, boolean_t prefer_zero) 738{ 739 vm_page_t m; 740 741 m = vm_pageq_find(PQ_FREE, color & PQ_L2_MASK, prefer_zero); 742 return (m); 743} 744 745/* 746 * vm_page_alloc: 747 * 748 * Allocate and return a memory cell associated 749 * with this VM object/offset pair. 750 * 751 * page_req classes: 752 * VM_ALLOC_NORMAL normal process request 753 * VM_ALLOC_SYSTEM system *really* needs a page 754 * VM_ALLOC_INTERRUPT interrupt time request 755 * VM_ALLOC_ZERO zero page 756 * 757 * This routine may not block. 758 * 759 * Additional special handling is required when called from an 760 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 761 * the page cache in this case. 762 */ 763vm_page_t 764vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 765{ 766 vm_page_t m = NULL; 767 vm_pindex_t color; 768 int flags, page_req, s; 769 770 page_req = req & VM_ALLOC_CLASS_MASK; 771 772 if ((req & VM_ALLOC_NOOBJ) == 0) { 773 KASSERT(object != NULL, 774 ("vm_page_alloc: NULL object.")); 775 KASSERT(!vm_page_lookup(object, pindex), 776 ("vm_page_alloc: page already allocated")); 777 color = pindex + object->pg_color; 778 } else 779 color = pindex; 780 781 /* 782 * The pager is allowed to eat deeper into the free page list. 783 */ 784 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 785 page_req = VM_ALLOC_SYSTEM; 786 }; 787 788 s = splvm(); 789loop: 790 mtx_lock_spin(&vm_page_queue_free_mtx); 791 if (cnt.v_free_count > cnt.v_free_reserved || 792 (page_req == VM_ALLOC_SYSTEM && 793 cnt.v_cache_count == 0 && 794 cnt.v_free_count > cnt.v_interrupt_free_min) || 795 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 796 /* 797 * Allocate from the free queue if the number of free pages 798 * exceeds the minimum for the request class. 799 */ 800 m = vm_page_select_free(color, (req & VM_ALLOC_ZERO) != 0); 801 } else if (page_req != VM_ALLOC_INTERRUPT) { 802 mtx_unlock_spin(&vm_page_queue_free_mtx); 803 /* 804 * Allocatable from cache (non-interrupt only). On success, 805 * we must free the page and try again, thus ensuring that 806 * cnt.v_*_free_min counters are replenished. 807 */ 808 vm_page_lock_queues(); 809 if ((m = vm_page_select_cache(color)) == NULL) { 810 vm_page_unlock_queues(); 811 splx(s); 812#if defined(DIAGNOSTIC) 813 if (cnt.v_cache_count > 0) 814 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 815#endif 816 atomic_add_int(&vm_pageout_deficit, 1); 817 pagedaemon_wakeup(); 818 return (NULL); 819 } 820 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 821 vm_page_busy(m); 822 pmap_remove_all(m); 823 vm_page_free(m); 824 vm_page_unlock_queues(); 825 goto loop; 826 } else { 827 /* 828 * Not allocatable from cache from interrupt, give up. 829 */ 830 mtx_unlock_spin(&vm_page_queue_free_mtx); 831 splx(s); 832 atomic_add_int(&vm_pageout_deficit, 1); 833 pagedaemon_wakeup(); 834 return (NULL); 835 } 836 837 /* 838 * At this point we had better have found a good page. 839 */ 840 841 KASSERT( 842 m != NULL, 843 ("vm_page_alloc(): missing page on free queue\n") 844 ); 845 846 /* 847 * Remove from free queue 848 */ 849 850 vm_pageq_remove_nowakeup(m); 851 852 /* 853 * Initialize structure. Only the PG_ZERO flag is inherited. 854 */ 855 flags = PG_BUSY; 856 if (m->flags & PG_ZERO) { 857 vm_page_zero_count--; 858 if (req & VM_ALLOC_ZERO) 859 flags = PG_ZERO | PG_BUSY; 860 } 861 m->flags = flags; 862 if (req & VM_ALLOC_WIRED) { 863 atomic_add_int(&cnt.v_wire_count, 1); 864 m->wire_count = 1; 865 } else 866 m->wire_count = 0; 867 m->hold_count = 0; 868 m->act_count = 0; 869 m->busy = 0; 870 m->valid = 0; 871 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 872 mtx_unlock_spin(&vm_page_queue_free_mtx); 873 874 /* 875 * vm_page_insert() is safe prior to the splx(). Note also that 876 * inserting a page here does not insert it into the pmap (which 877 * could cause us to block allocating memory). We cannot block 878 * anywhere. 879 */ 880 if ((req & VM_ALLOC_NOOBJ) == 0) 881 vm_page_insert(m, object, pindex); 882 883 /* 884 * Don't wakeup too often - wakeup the pageout daemon when 885 * we would be nearly out of memory. 886 */ 887 if (vm_paging_needed()) 888 pagedaemon_wakeup(); 889 890 splx(s); 891 return (m); 892} 893 894/* 895 * vm_wait: (also see VM_WAIT macro) 896 * 897 * Block until free pages are available for allocation 898 * - Called in various places before memory allocations. 899 */ 900void 901vm_wait(void) 902{ 903 int s; 904 905 s = splvm(); 906 vm_page_lock_queues(); 907 if (curproc == pageproc) { 908 vm_pageout_pages_needed = 1; 909 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 910 PDROP | PSWP, "VMWait", 0); 911 } else { 912 if (!vm_pages_needed) { 913 vm_pages_needed = 1; 914 wakeup(&vm_pages_needed); 915 } 916 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 917 "vmwait", 0); 918 } 919 splx(s); 920} 921 922/* 923 * vm_waitpfault: (also see VM_WAITPFAULT macro) 924 * 925 * Block until free pages are available for allocation 926 * - Called only in vm_fault so that processes page faulting 927 * can be easily tracked. 928 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 929 * processes will be able to grab memory first. Do not change 930 * this balance without careful testing first. 931 */ 932void 933vm_waitpfault(void) 934{ 935 int s; 936 937 s = splvm(); 938 vm_page_lock_queues(); 939 if (!vm_pages_needed) { 940 vm_pages_needed = 1; 941 wakeup(&vm_pages_needed); 942 } 943 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 944 "pfault", 0); 945 splx(s); 946} 947 948/* 949 * vm_page_activate: 950 * 951 * Put the specified page on the active list (if appropriate). 952 * Ensure that act_count is at least ACT_INIT but do not otherwise 953 * mess with it. 954 * 955 * The page queues must be locked. 956 * This routine may not block. 957 */ 958void 959vm_page_activate(vm_page_t m) 960{ 961 int s; 962 963 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 964 s = splvm(); 965 if (m->queue != PQ_ACTIVE) { 966 if ((m->queue - m->pc) == PQ_CACHE) 967 cnt.v_reactivated++; 968 vm_pageq_remove(m); 969 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 970 if (m->act_count < ACT_INIT) 971 m->act_count = ACT_INIT; 972 vm_pageq_enqueue(PQ_ACTIVE, m); 973 } 974 } else { 975 if (m->act_count < ACT_INIT) 976 m->act_count = ACT_INIT; 977 } 978 splx(s); 979} 980 981/* 982 * vm_page_free_wakeup: 983 * 984 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 985 * routine is called when a page has been added to the cache or free 986 * queues. 987 * 988 * This routine may not block. 989 * This routine must be called at splvm() 990 */ 991static __inline void 992vm_page_free_wakeup(void) 993{ 994 995 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 996 /* 997 * if pageout daemon needs pages, then tell it that there are 998 * some free. 999 */ 1000 if (vm_pageout_pages_needed && 1001 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1002 wakeup(&vm_pageout_pages_needed); 1003 vm_pageout_pages_needed = 0; 1004 } 1005 /* 1006 * wakeup processes that are waiting on memory if we hit a 1007 * high water mark. And wakeup scheduler process if we have 1008 * lots of memory. this process will swapin processes. 1009 */ 1010 if (vm_pages_needed && !vm_page_count_min()) { 1011 vm_pages_needed = 0; 1012 wakeup(&cnt.v_free_count); 1013 } 1014} 1015 1016/* 1017 * vm_page_free_toq: 1018 * 1019 * Returns the given page to the PQ_FREE list, 1020 * disassociating it with any VM object. 1021 * 1022 * Object and page must be locked prior to entry. 1023 * This routine may not block. 1024 */ 1025 1026void 1027vm_page_free_toq(vm_page_t m) 1028{ 1029 int s; 1030 struct vpgqueues *pq; 1031 vm_object_t object = m->object; 1032 1033 GIANT_REQUIRED; 1034 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1035 s = splvm(); 1036 cnt.v_tfree++; 1037 1038 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1039 printf( 1040 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1041 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1042 m->hold_count); 1043 if ((m->queue - m->pc) == PQ_FREE) 1044 panic("vm_page_free: freeing free page"); 1045 else 1046 panic("vm_page_free: freeing busy page"); 1047 } 1048 1049 /* 1050 * unqueue, then remove page. Note that we cannot destroy 1051 * the page here because we do not want to call the pager's 1052 * callback routine until after we've put the page on the 1053 * appropriate free queue. 1054 */ 1055 vm_pageq_remove_nowakeup(m); 1056 vm_page_remove(m); 1057 1058 /* 1059 * If fictitious remove object association and 1060 * return, otherwise delay object association removal. 1061 */ 1062 if ((m->flags & PG_FICTITIOUS) != 0) { 1063 splx(s); 1064 return; 1065 } 1066 1067 m->valid = 0; 1068 vm_page_undirty(m); 1069 1070 if (m->wire_count != 0) { 1071 if (m->wire_count > 1) { 1072 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1073 m->wire_count, (long)m->pindex); 1074 } 1075 panic("vm_page_free: freeing wired page\n"); 1076 } 1077 1078 /* 1079 * If we've exhausted the object's resident pages we want to free 1080 * it up. 1081 */ 1082 if (object && 1083 (object->type == OBJT_VNODE) && 1084 ((object->flags & OBJ_DEAD) == 0) 1085 ) { 1086 struct vnode *vp = (struct vnode *)object->handle; 1087 1088 if (vp) { 1089 VI_LOCK(vp); 1090 if (VSHOULDFREE(vp)) 1091 vfree(vp); 1092 VI_UNLOCK(vp); 1093 } 1094 } 1095 1096 /* 1097 * Clear the UNMANAGED flag when freeing an unmanaged page. 1098 */ 1099 if (m->flags & PG_UNMANAGED) { 1100 m->flags &= ~PG_UNMANAGED; 1101 } 1102 1103 if (m->hold_count != 0) { 1104 m->flags &= ~PG_ZERO; 1105 m->queue = PQ_HOLD; 1106 } else 1107 m->queue = PQ_FREE + m->pc; 1108 pq = &vm_page_queues[m->queue]; 1109 mtx_lock_spin(&vm_page_queue_free_mtx); 1110 pq->lcnt++; 1111 ++(*pq->cnt); 1112 1113 /* 1114 * Put zero'd pages on the end ( where we look for zero'd pages 1115 * first ) and non-zerod pages at the head. 1116 */ 1117 if (m->flags & PG_ZERO) { 1118 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1119 ++vm_page_zero_count; 1120 } else { 1121 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1122 } 1123 mtx_unlock_spin(&vm_page_queue_free_mtx); 1124 vm_page_free_wakeup(); 1125 splx(s); 1126} 1127 1128/* 1129 * vm_page_unmanage: 1130 * 1131 * Prevent PV management from being done on the page. The page is 1132 * removed from the paging queues as if it were wired, and as a 1133 * consequence of no longer being managed the pageout daemon will not 1134 * touch it (since there is no way to locate the pte mappings for the 1135 * page). madvise() calls that mess with the pmap will also no longer 1136 * operate on the page. 1137 * 1138 * Beyond that the page is still reasonably 'normal'. Freeing the page 1139 * will clear the flag. 1140 * 1141 * This routine is used by OBJT_PHYS objects - objects using unswappable 1142 * physical memory as backing store rather then swap-backed memory and 1143 * will eventually be extended to support 4MB unmanaged physical 1144 * mappings. 1145 */ 1146void 1147vm_page_unmanage(vm_page_t m) 1148{ 1149 int s; 1150 1151 s = splvm(); 1152 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1153 if ((m->flags & PG_UNMANAGED) == 0) { 1154 if (m->wire_count == 0) 1155 vm_pageq_remove(m); 1156 } 1157 vm_page_flag_set(m, PG_UNMANAGED); 1158 splx(s); 1159} 1160 1161/* 1162 * vm_page_wire: 1163 * 1164 * Mark this page as wired down by yet 1165 * another map, removing it from paging queues 1166 * as necessary. 1167 * 1168 * The page queues must be locked. 1169 * This routine may not block. 1170 */ 1171void 1172vm_page_wire(vm_page_t m) 1173{ 1174 int s; 1175 1176 /* 1177 * Only bump the wire statistics if the page is not already wired, 1178 * and only unqueue the page if it is on some queue (if it is unmanaged 1179 * it is already off the queues). 1180 */ 1181 s = splvm(); 1182 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1183 if (m->wire_count == 0) { 1184 if ((m->flags & PG_UNMANAGED) == 0) 1185 vm_pageq_remove(m); 1186 atomic_add_int(&cnt.v_wire_count, 1); 1187 } 1188 m->wire_count++; 1189 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1190 splx(s); 1191} 1192 1193/* 1194 * vm_page_unwire: 1195 * 1196 * Release one wiring of this page, potentially 1197 * enabling it to be paged again. 1198 * 1199 * Many pages placed on the inactive queue should actually go 1200 * into the cache, but it is difficult to figure out which. What 1201 * we do instead, if the inactive target is well met, is to put 1202 * clean pages at the head of the inactive queue instead of the tail. 1203 * This will cause them to be moved to the cache more quickly and 1204 * if not actively re-referenced, freed more quickly. If we just 1205 * stick these pages at the end of the inactive queue, heavy filesystem 1206 * meta-data accesses can cause an unnecessary paging load on memory bound 1207 * processes. This optimization causes one-time-use metadata to be 1208 * reused more quickly. 1209 * 1210 * BUT, if we are in a low-memory situation we have no choice but to 1211 * put clean pages on the cache queue. 1212 * 1213 * A number of routines use vm_page_unwire() to guarantee that the page 1214 * will go into either the inactive or active queues, and will NEVER 1215 * be placed in the cache - for example, just after dirtying a page. 1216 * dirty pages in the cache are not allowed. 1217 * 1218 * The page queues must be locked. 1219 * This routine may not block. 1220 */ 1221void 1222vm_page_unwire(vm_page_t m, int activate) 1223{ 1224 int s; 1225 1226 s = splvm(); 1227 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1228 if (m->wire_count > 0) { 1229 m->wire_count--; 1230 if (m->wire_count == 0) { 1231 atomic_subtract_int(&cnt.v_wire_count, 1); 1232 if (m->flags & PG_UNMANAGED) { 1233 ; 1234 } else if (activate) 1235 vm_pageq_enqueue(PQ_ACTIVE, m); 1236 else { 1237 vm_page_flag_clear(m, PG_WINATCFLS); 1238 vm_pageq_enqueue(PQ_INACTIVE, m); 1239 } 1240 } 1241 } else { 1242 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1243 } 1244 splx(s); 1245} 1246 1247 1248/* 1249 * Move the specified page to the inactive queue. If the page has 1250 * any associated swap, the swap is deallocated. 1251 * 1252 * Normally athead is 0 resulting in LRU operation. athead is set 1253 * to 1 if we want this page to be 'as if it were placed in the cache', 1254 * except without unmapping it from the process address space. 1255 * 1256 * This routine may not block. 1257 */ 1258static __inline void 1259_vm_page_deactivate(vm_page_t m, int athead) 1260{ 1261 int s; 1262 1263 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1264 /* 1265 * Ignore if already inactive. 1266 */ 1267 if (m->queue == PQ_INACTIVE) 1268 return; 1269 1270 s = splvm(); 1271 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1272 if ((m->queue - m->pc) == PQ_CACHE) 1273 cnt.v_reactivated++; 1274 vm_page_flag_clear(m, PG_WINATCFLS); 1275 vm_pageq_remove(m); 1276 if (athead) 1277 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1278 else 1279 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1280 m->queue = PQ_INACTIVE; 1281 vm_page_queues[PQ_INACTIVE].lcnt++; 1282 cnt.v_inactive_count++; 1283 } 1284 splx(s); 1285} 1286 1287void 1288vm_page_deactivate(vm_page_t m) 1289{ 1290 _vm_page_deactivate(m, 0); 1291} 1292 1293/* 1294 * vm_page_try_to_cache: 1295 * 1296 * Returns 0 on failure, 1 on success 1297 */ 1298int 1299vm_page_try_to_cache(vm_page_t m) 1300{ 1301 1302 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1303 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1304 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1305 return (0); 1306 } 1307 vm_page_test_dirty(m); 1308 if (m->dirty) 1309 return (0); 1310 vm_page_cache(m); 1311 return (1); 1312} 1313 1314/* 1315 * vm_page_try_to_free() 1316 * 1317 * Attempt to free the page. If we cannot free it, we do nothing. 1318 * 1 is returned on success, 0 on failure. 1319 */ 1320int 1321vm_page_try_to_free(vm_page_t m) 1322{ 1323 1324 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1325 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1326 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1327 return (0); 1328 } 1329 vm_page_test_dirty(m); 1330 if (m->dirty) 1331 return (0); 1332 vm_page_busy(m); 1333 pmap_remove_all(m); 1334 vm_page_free(m); 1335 return (1); 1336} 1337 1338/* 1339 * vm_page_cache 1340 * 1341 * Put the specified page onto the page cache queue (if appropriate). 1342 * 1343 * This routine may not block. 1344 */ 1345void 1346vm_page_cache(vm_page_t m) 1347{ 1348 int s; 1349 1350 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1351 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1352 printf("vm_page_cache: attempting to cache busy page\n"); 1353 return; 1354 } 1355 if ((m->queue - m->pc) == PQ_CACHE) 1356 return; 1357 1358 /* 1359 * Remove all pmaps and indicate that the page is not 1360 * writeable or mapped. 1361 */ 1362 pmap_remove_all(m); 1363 if (m->dirty != 0) { 1364 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1365 (long)m->pindex); 1366 } 1367 s = splvm(); 1368 vm_pageq_remove_nowakeup(m); 1369 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1370 vm_page_free_wakeup(); 1371 splx(s); 1372} 1373 1374/* 1375 * vm_page_dontneed 1376 * 1377 * Cache, deactivate, or do nothing as appropriate. This routine 1378 * is typically used by madvise() MADV_DONTNEED. 1379 * 1380 * Generally speaking we want to move the page into the cache so 1381 * it gets reused quickly. However, this can result in a silly syndrome 1382 * due to the page recycling too quickly. Small objects will not be 1383 * fully cached. On the otherhand, if we move the page to the inactive 1384 * queue we wind up with a problem whereby very large objects 1385 * unnecessarily blow away our inactive and cache queues. 1386 * 1387 * The solution is to move the pages based on a fixed weighting. We 1388 * either leave them alone, deactivate them, or move them to the cache, 1389 * where moving them to the cache has the highest weighting. 1390 * By forcing some pages into other queues we eventually force the 1391 * system to balance the queues, potentially recovering other unrelated 1392 * space from active. The idea is to not force this to happen too 1393 * often. 1394 */ 1395void 1396vm_page_dontneed(vm_page_t m) 1397{ 1398 static int dnweight; 1399 int dnw; 1400 int head; 1401 1402 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1403 dnw = ++dnweight; 1404 1405 /* 1406 * occassionally leave the page alone 1407 */ 1408 if ((dnw & 0x01F0) == 0 || 1409 m->queue == PQ_INACTIVE || 1410 m->queue - m->pc == PQ_CACHE 1411 ) { 1412 if (m->act_count >= ACT_INIT) 1413 --m->act_count; 1414 return; 1415 } 1416 1417 if (m->dirty == 0) 1418 vm_page_test_dirty(m); 1419 1420 if (m->dirty || (dnw & 0x0070) == 0) { 1421 /* 1422 * Deactivate the page 3 times out of 32. 1423 */ 1424 head = 0; 1425 } else { 1426 /* 1427 * Cache the page 28 times out of every 32. Note that 1428 * the page is deactivated instead of cached, but placed 1429 * at the head of the queue instead of the tail. 1430 */ 1431 head = 1; 1432 } 1433 _vm_page_deactivate(m, head); 1434} 1435 1436/* 1437 * Grab a page, waiting until we are waken up due to the page 1438 * changing state. We keep on waiting, if the page continues 1439 * to be in the object. If the page doesn't exist, allocate it. 1440 * 1441 * This routine may block. 1442 */ 1443vm_page_t 1444vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1445{ 1446 vm_page_t m; 1447 int s, generation; 1448 1449 GIANT_REQUIRED; 1450retrylookup: 1451 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1452 vm_page_lock_queues(); 1453 if (m->busy || (m->flags & PG_BUSY)) { 1454 generation = object->generation; 1455 1456 s = splvm(); 1457 while ((object->generation == generation) && 1458 (m->busy || (m->flags & PG_BUSY))) { 1459 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1460 msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0); 1461 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1462 vm_page_unlock_queues(); 1463 splx(s); 1464 return NULL; 1465 } 1466 } 1467 vm_page_unlock_queues(); 1468 splx(s); 1469 goto retrylookup; 1470 } else { 1471 if (allocflags & VM_ALLOC_WIRED) 1472 vm_page_wire(m); 1473 vm_page_busy(m); 1474 vm_page_unlock_queues(); 1475 return m; 1476 } 1477 } 1478 1479 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1480 if (m == NULL) { 1481 VM_WAIT; 1482 if ((allocflags & VM_ALLOC_RETRY) == 0) 1483 return NULL; 1484 goto retrylookup; 1485 } 1486 1487 return m; 1488} 1489 1490/* 1491 * Mapping function for valid bits or for dirty bits in 1492 * a page. May not block. 1493 * 1494 * Inputs are required to range within a page. 1495 */ 1496__inline int 1497vm_page_bits(int base, int size) 1498{ 1499 int first_bit; 1500 int last_bit; 1501 1502 KASSERT( 1503 base + size <= PAGE_SIZE, 1504 ("vm_page_bits: illegal base/size %d/%d", base, size) 1505 ); 1506 1507 if (size == 0) /* handle degenerate case */ 1508 return (0); 1509 1510 first_bit = base >> DEV_BSHIFT; 1511 last_bit = (base + size - 1) >> DEV_BSHIFT; 1512 1513 return ((2 << last_bit) - (1 << first_bit)); 1514} 1515 1516/* 1517 * vm_page_set_validclean: 1518 * 1519 * Sets portions of a page valid and clean. The arguments are expected 1520 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1521 * of any partial chunks touched by the range. The invalid portion of 1522 * such chunks will be zero'd. 1523 * 1524 * This routine may not block. 1525 * 1526 * (base + size) must be less then or equal to PAGE_SIZE. 1527 */ 1528void 1529vm_page_set_validclean(vm_page_t m, int base, int size) 1530{ 1531 int pagebits; 1532 int frag; 1533 int endoff; 1534 1535 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1536 if (size == 0) /* handle degenerate case */ 1537 return; 1538 1539 /* 1540 * If the base is not DEV_BSIZE aligned and the valid 1541 * bit is clear, we have to zero out a portion of the 1542 * first block. 1543 */ 1544 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1545 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1546 pmap_zero_page_area(m, frag, base - frag); 1547 1548 /* 1549 * If the ending offset is not DEV_BSIZE aligned and the 1550 * valid bit is clear, we have to zero out a portion of 1551 * the last block. 1552 */ 1553 endoff = base + size; 1554 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1555 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1556 pmap_zero_page_area(m, endoff, 1557 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1558 1559 /* 1560 * Set valid, clear dirty bits. If validating the entire 1561 * page we can safely clear the pmap modify bit. We also 1562 * use this opportunity to clear the PG_NOSYNC flag. If a process 1563 * takes a write fault on a MAP_NOSYNC memory area the flag will 1564 * be set again. 1565 * 1566 * We set valid bits inclusive of any overlap, but we can only 1567 * clear dirty bits for DEV_BSIZE chunks that are fully within 1568 * the range. 1569 */ 1570 pagebits = vm_page_bits(base, size); 1571 m->valid |= pagebits; 1572#if 0 /* NOT YET */ 1573 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1574 frag = DEV_BSIZE - frag; 1575 base += frag; 1576 size -= frag; 1577 if (size < 0) 1578 size = 0; 1579 } 1580 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1581#endif 1582 m->dirty &= ~pagebits; 1583 if (base == 0 && size == PAGE_SIZE) { 1584 pmap_clear_modify(m); 1585 vm_page_flag_clear(m, PG_NOSYNC); 1586 } 1587} 1588 1589#if 0 1590 1591void 1592vm_page_set_dirty(vm_page_t m, int base, int size) 1593{ 1594 m->dirty |= vm_page_bits(base, size); 1595} 1596 1597#endif 1598 1599void 1600vm_page_clear_dirty(vm_page_t m, int base, int size) 1601{ 1602 GIANT_REQUIRED; 1603 m->dirty &= ~vm_page_bits(base, size); 1604} 1605 1606/* 1607 * vm_page_set_invalid: 1608 * 1609 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1610 * valid and dirty bits for the effected areas are cleared. 1611 * 1612 * May not block. 1613 */ 1614void 1615vm_page_set_invalid(vm_page_t m, int base, int size) 1616{ 1617 int bits; 1618 1619 GIANT_REQUIRED; 1620 bits = vm_page_bits(base, size); 1621 m->valid &= ~bits; 1622 m->dirty &= ~bits; 1623 m->object->generation++; 1624} 1625 1626/* 1627 * vm_page_zero_invalid() 1628 * 1629 * The kernel assumes that the invalid portions of a page contain 1630 * garbage, but such pages can be mapped into memory by user code. 1631 * When this occurs, we must zero out the non-valid portions of the 1632 * page so user code sees what it expects. 1633 * 1634 * Pages are most often semi-valid when the end of a file is mapped 1635 * into memory and the file's size is not page aligned. 1636 */ 1637void 1638vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1639{ 1640 int b; 1641 int i; 1642 1643 /* 1644 * Scan the valid bits looking for invalid sections that 1645 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1646 * valid bit may be set ) have already been zerod by 1647 * vm_page_set_validclean(). 1648 */ 1649 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1650 if (i == (PAGE_SIZE / DEV_BSIZE) || 1651 (m->valid & (1 << i)) 1652 ) { 1653 if (i > b) { 1654 pmap_zero_page_area(m, 1655 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1656 } 1657 b = i + 1; 1658 } 1659 } 1660 1661 /* 1662 * setvalid is TRUE when we can safely set the zero'd areas 1663 * as being valid. We can do this if there are no cache consistancy 1664 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1665 */ 1666 if (setvalid) 1667 m->valid = VM_PAGE_BITS_ALL; 1668} 1669 1670/* 1671 * vm_page_is_valid: 1672 * 1673 * Is (partial) page valid? Note that the case where size == 0 1674 * will return FALSE in the degenerate case where the page is 1675 * entirely invalid, and TRUE otherwise. 1676 * 1677 * May not block. 1678 */ 1679int 1680vm_page_is_valid(vm_page_t m, int base, int size) 1681{ 1682 int bits = vm_page_bits(base, size); 1683 1684 if (m->valid && ((m->valid & bits) == bits)) 1685 return 1; 1686 else 1687 return 0; 1688} 1689 1690/* 1691 * update dirty bits from pmap/mmu. May not block. 1692 */ 1693void 1694vm_page_test_dirty(vm_page_t m) 1695{ 1696 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1697 vm_page_dirty(m); 1698 } 1699} 1700 1701int so_zerocp_fullpage = 0; 1702 1703void 1704vm_page_cowfault(vm_page_t m) 1705{ 1706 vm_page_t mnew; 1707 vm_object_t object; 1708 vm_pindex_t pindex; 1709 1710 object = m->object; 1711 pindex = m->pindex; 1712 vm_page_busy(m); 1713 1714 retry_alloc: 1715 vm_page_remove(m); 1716 /* 1717 * An interrupt allocation is requested because the page 1718 * queues lock is held. 1719 */ 1720 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1721 if (mnew == NULL) { 1722 vm_page_insert(m, object, pindex); 1723 vm_page_unlock_queues(); 1724 VM_WAIT; 1725 vm_page_lock_queues(); 1726 goto retry_alloc; 1727 } 1728 1729 if (m->cow == 0) { 1730 /* 1731 * check to see if we raced with an xmit complete when 1732 * waiting to allocate a page. If so, put things back 1733 * the way they were 1734 */ 1735 vm_page_busy(mnew); 1736 vm_page_free(mnew); 1737 vm_page_insert(m, object, pindex); 1738 } else { /* clear COW & copy page */ 1739 if (so_zerocp_fullpage) { 1740 mnew->valid = VM_PAGE_BITS_ALL; 1741 } else { 1742 vm_page_copy(m, mnew); 1743 } 1744 vm_page_dirty(mnew); 1745 vm_page_flag_clear(mnew, PG_BUSY); 1746 } 1747} 1748 1749void 1750vm_page_cowclear(vm_page_t m) 1751{ 1752 1753 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1754 if (m->cow) { 1755 m->cow--; 1756 /* 1757 * let vm_fault add back write permission lazily 1758 */ 1759 } 1760 /* 1761 * sf_buf_free() will free the page, so we needn't do it here 1762 */ 1763} 1764 1765void 1766vm_page_cowsetup(vm_page_t m) 1767{ 1768 1769 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1770 m->cow++; 1771 pmap_page_protect(m, VM_PROT_READ); 1772} 1773 1774#include "opt_ddb.h" 1775#ifdef DDB 1776#include <sys/kernel.h> 1777 1778#include <ddb/ddb.h> 1779 1780DB_SHOW_COMMAND(page, vm_page_print_page_info) 1781{ 1782 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1783 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1784 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1785 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1786 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1787 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1788 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1789 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1790 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1791 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1792} 1793 1794DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1795{ 1796 int i; 1797 db_printf("PQ_FREE:"); 1798 for (i = 0; i < PQ_L2_SIZE; i++) { 1799 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1800 } 1801 db_printf("\n"); 1802 1803 db_printf("PQ_CACHE:"); 1804 for (i = 0; i < PQ_L2_SIZE; i++) { 1805 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1806 } 1807 db_printf("\n"); 1808 1809 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1810 vm_page_queues[PQ_ACTIVE].lcnt, 1811 vm_page_queues[PQ_INACTIVE].lcnt); 1812} 1813#endif /* DDB */ 1814