vm_page.c revision 113841
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 113841 2003-04-22 05:36:14Z 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 (m->object != NULL) 536 panic("vm_page_insert: already inserted"); 537 538 /* 539 * Record the object/offset pair in this page 540 */ 541 m->object = object; 542 m->pindex = pindex; 543 544 mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED); 545 /* 546 * Now link into the object's ordered list of backed pages. 547 */ 548 root = object->root; 549 if (root == NULL) { 550 m->left = NULL; 551 m->right = NULL; 552 TAILQ_INSERT_TAIL(&object->memq, m, listq); 553 } else { 554 root = vm_page_splay(pindex, root); 555 if (pindex < root->pindex) { 556 m->left = root->left; 557 m->right = root; 558 root->left = NULL; 559 TAILQ_INSERT_BEFORE(root, m, listq); 560 } else { 561 m->right = root->right; 562 m->left = root; 563 root->right = NULL; 564 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 565 } 566 } 567 object->root = m; 568 object->generation++; 569 570 /* 571 * show that the object has one more resident page. 572 */ 573 object->resident_page_count++; 574 575 /* 576 * Since we are inserting a new and possibly dirty page, 577 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 578 */ 579 if (m->flags & PG_WRITEABLE) 580 vm_object_set_writeable_dirty(object); 581} 582 583/* 584 * vm_page_remove: 585 * NOTE: used by device pager as well -wfj 586 * 587 * Removes the given mem entry from the object/offset-page 588 * table and the object page list, but do not invalidate/terminate 589 * the backing store. 590 * 591 * The object and page must be locked, and at splhigh. 592 * The underlying pmap entry (if any) is NOT removed here. 593 * This routine may not block. 594 */ 595void 596vm_page_remove(vm_page_t m) 597{ 598 vm_object_t object; 599 vm_page_t root; 600 601 GIANT_REQUIRED; 602 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 603 if (m->object == NULL) 604 return; 605 606 if ((m->flags & PG_BUSY) == 0) { 607 panic("vm_page_remove: page not busy"); 608 } 609 610 /* 611 * Basically destroy the page. 612 */ 613 vm_page_wakeup(m); 614 615 object = m->object; 616 617 /* 618 * Now remove from the object's list of backed pages. 619 */ 620 if (m != object->root) 621 vm_page_splay(m->pindex, object->root); 622 if (m->left == NULL) 623 root = m->right; 624 else { 625 root = vm_page_splay(m->pindex, m->left); 626 root->right = m->right; 627 } 628 object->root = root; 629 TAILQ_REMOVE(&object->memq, m, listq); 630 631 /* 632 * And show that the object has one fewer resident page. 633 */ 634 object->resident_page_count--; 635 object->generation++; 636 637 m->object = NULL; 638} 639 640/* 641 * vm_page_lookup: 642 * 643 * Returns the page associated with the object/offset 644 * pair specified; if none is found, NULL is returned. 645 * 646 * The object must be locked. 647 * This routine may not block. 648 * This is a critical path routine 649 */ 650vm_page_t 651vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 652{ 653 vm_page_t m; 654 655 if (!VM_OBJECT_LOCKED(object)) 656 GIANT_REQUIRED; 657 m = vm_page_splay(pindex, object->root); 658 if ((object->root = m) != NULL && m->pindex != pindex) 659 m = NULL; 660 return (m); 661} 662 663/* 664 * vm_page_rename: 665 * 666 * Move the given memory entry from its 667 * current object to the specified target object/offset. 668 * 669 * The object must be locked. 670 * This routine may not block. 671 * 672 * Note: this routine will raise itself to splvm(), the caller need not. 673 * 674 * Note: swap associated with the page must be invalidated by the move. We 675 * have to do this for several reasons: (1) we aren't freeing the 676 * page, (2) we are dirtying the page, (3) the VM system is probably 677 * moving the page from object A to B, and will then later move 678 * the backing store from A to B and we can't have a conflict. 679 * 680 * Note: we *always* dirty the page. It is necessary both for the 681 * fact that we moved it, and because we may be invalidating 682 * swap. If the page is on the cache, we have to deactivate it 683 * or vm_page_dirty() will panic. Dirty pages are not allowed 684 * on the cache. 685 */ 686void 687vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 688{ 689 int s; 690 691 s = splvm(); 692 vm_page_remove(m); 693 vm_page_insert(m, new_object, new_pindex); 694 if (m->queue - m->pc == PQ_CACHE) 695 vm_page_deactivate(m); 696 vm_page_dirty(m); 697 splx(s); 698} 699 700/* 701 * vm_page_select_cache: 702 * 703 * Find a page on the cache queue with color optimization. As pages 704 * might be found, but not applicable, they are deactivated. This 705 * keeps us from using potentially busy cached pages. 706 * 707 * This routine must be called at splvm(). 708 * This routine may not block. 709 */ 710static vm_page_t 711vm_page_select_cache(vm_pindex_t color) 712{ 713 vm_page_t m; 714 715 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 716 while (TRUE) { 717 m = vm_pageq_find(PQ_CACHE, color & PQ_L2_MASK, FALSE); 718 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 719 m->hold_count || m->wire_count)) { 720 vm_page_deactivate(m); 721 continue; 722 } 723 return m; 724 } 725} 726 727/* 728 * vm_page_select_free: 729 * 730 * Find a free or zero page, with specified preference. 731 * 732 * This routine must be called at splvm(). 733 * This routine may not block. 734 */ 735static __inline vm_page_t 736vm_page_select_free(vm_pindex_t color, boolean_t prefer_zero) 737{ 738 vm_page_t m; 739 740 m = vm_pageq_find(PQ_FREE, color & PQ_L2_MASK, prefer_zero); 741 return (m); 742} 743 744/* 745 * vm_page_alloc: 746 * 747 * Allocate and return a memory cell associated 748 * with this VM object/offset pair. 749 * 750 * page_req classes: 751 * VM_ALLOC_NORMAL normal process request 752 * VM_ALLOC_SYSTEM system *really* needs a page 753 * VM_ALLOC_INTERRUPT interrupt time request 754 * VM_ALLOC_ZERO zero page 755 * 756 * This routine may not block. 757 * 758 * Additional special handling is required when called from an 759 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 760 * the page cache in this case. 761 */ 762vm_page_t 763vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 764{ 765 vm_page_t m = NULL; 766 vm_pindex_t color; 767 int flags, page_req, s; 768 769 page_req = req & VM_ALLOC_CLASS_MASK; 770 771 if ((req & VM_ALLOC_NOOBJ) == 0) { 772 KASSERT(object != NULL, 773 ("vm_page_alloc: NULL object.")); 774 KASSERT(!vm_page_lookup(object, pindex), 775 ("vm_page_alloc: page already allocated")); 776 color = pindex + object->pg_color; 777 } else 778 color = pindex; 779 780 /* 781 * The pager is allowed to eat deeper into the free page list. 782 */ 783 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 784 page_req = VM_ALLOC_SYSTEM; 785 }; 786 787 s = splvm(); 788loop: 789 mtx_lock_spin(&vm_page_queue_free_mtx); 790 if (cnt.v_free_count > cnt.v_free_reserved || 791 (page_req == VM_ALLOC_SYSTEM && 792 cnt.v_cache_count == 0 && 793 cnt.v_free_count > cnt.v_interrupt_free_min) || 794 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 795 /* 796 * Allocate from the free queue if the number of free pages 797 * exceeds the minimum for the request class. 798 */ 799 m = vm_page_select_free(color, (req & VM_ALLOC_ZERO) != 0); 800 } else if (page_req != VM_ALLOC_INTERRUPT) { 801 mtx_unlock_spin(&vm_page_queue_free_mtx); 802 /* 803 * Allocatable from cache (non-interrupt only). On success, 804 * we must free the page and try again, thus ensuring that 805 * cnt.v_*_free_min counters are replenished. 806 */ 807 vm_page_lock_queues(); 808 if ((m = vm_page_select_cache(color)) == NULL) { 809 vm_page_unlock_queues(); 810 splx(s); 811#if defined(DIAGNOSTIC) 812 if (cnt.v_cache_count > 0) 813 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 814#endif 815 atomic_add_int(&vm_pageout_deficit, 1); 816 pagedaemon_wakeup(); 817 return (NULL); 818 } 819 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 820 vm_page_busy(m); 821 pmap_remove_all(m); 822 vm_page_free(m); 823 vm_page_unlock_queues(); 824 goto loop; 825 } else { 826 /* 827 * Not allocatable from cache from interrupt, give up. 828 */ 829 mtx_unlock_spin(&vm_page_queue_free_mtx); 830 splx(s); 831 atomic_add_int(&vm_pageout_deficit, 1); 832 pagedaemon_wakeup(); 833 return (NULL); 834 } 835 836 /* 837 * At this point we had better have found a good page. 838 */ 839 840 KASSERT( 841 m != NULL, 842 ("vm_page_alloc(): missing page on free queue\n") 843 ); 844 845 /* 846 * Remove from free queue 847 */ 848 849 vm_pageq_remove_nowakeup(m); 850 851 /* 852 * Initialize structure. Only the PG_ZERO flag is inherited. 853 */ 854 flags = PG_BUSY; 855 if (m->flags & PG_ZERO) { 856 vm_page_zero_count--; 857 if (req & VM_ALLOC_ZERO) 858 flags = PG_ZERO | PG_BUSY; 859 } 860 m->flags = flags; 861 if (req & VM_ALLOC_WIRED) { 862 atomic_add_int(&cnt.v_wire_count, 1); 863 m->wire_count = 1; 864 } else 865 m->wire_count = 0; 866 m->hold_count = 0; 867 m->act_count = 0; 868 m->busy = 0; 869 m->valid = 0; 870 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 871 mtx_unlock_spin(&vm_page_queue_free_mtx); 872 873 /* 874 * vm_page_insert() is safe prior to the splx(). Note also that 875 * inserting a page here does not insert it into the pmap (which 876 * could cause us to block allocating memory). We cannot block 877 * anywhere. 878 */ 879 if ((req & VM_ALLOC_NOOBJ) == 0) 880 vm_page_insert(m, object, pindex); 881 882 /* 883 * Don't wakeup too often - wakeup the pageout daemon when 884 * we would be nearly out of memory. 885 */ 886 if (vm_paging_needed()) 887 pagedaemon_wakeup(); 888 889 splx(s); 890 return (m); 891} 892 893/* 894 * vm_wait: (also see VM_WAIT macro) 895 * 896 * Block until free pages are available for allocation 897 * - Called in various places before memory allocations. 898 */ 899void 900vm_wait(void) 901{ 902 int s; 903 904 s = splvm(); 905 vm_page_lock_queues(); 906 if (curproc == pageproc) { 907 vm_pageout_pages_needed = 1; 908 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 909 PDROP | PSWP, "VMWait", 0); 910 } else { 911 if (!vm_pages_needed) { 912 vm_pages_needed = 1; 913 wakeup(&vm_pages_needed); 914 } 915 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 916 "vmwait", 0); 917 } 918 splx(s); 919} 920 921/* 922 * vm_waitpfault: (also see VM_WAITPFAULT macro) 923 * 924 * Block until free pages are available for allocation 925 * - Called only in vm_fault so that processes page faulting 926 * can be easily tracked. 927 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 928 * processes will be able to grab memory first. Do not change 929 * this balance without careful testing first. 930 */ 931void 932vm_waitpfault(void) 933{ 934 int s; 935 936 s = splvm(); 937 vm_page_lock_queues(); 938 if (!vm_pages_needed) { 939 vm_pages_needed = 1; 940 wakeup(&vm_pages_needed); 941 } 942 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 943 "pfault", 0); 944 splx(s); 945} 946 947/* 948 * vm_page_activate: 949 * 950 * Put the specified page on the active list (if appropriate). 951 * Ensure that act_count is at least ACT_INIT but do not otherwise 952 * mess with it. 953 * 954 * The page queues must be locked. 955 * This routine may not block. 956 */ 957void 958vm_page_activate(vm_page_t m) 959{ 960 int s; 961 962 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 963 s = splvm(); 964 if (m->queue != PQ_ACTIVE) { 965 if ((m->queue - m->pc) == PQ_CACHE) 966 cnt.v_reactivated++; 967 vm_pageq_remove(m); 968 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 969 if (m->act_count < ACT_INIT) 970 m->act_count = ACT_INIT; 971 vm_pageq_enqueue(PQ_ACTIVE, m); 972 } 973 } else { 974 if (m->act_count < ACT_INIT) 975 m->act_count = ACT_INIT; 976 } 977 splx(s); 978} 979 980/* 981 * vm_page_free_wakeup: 982 * 983 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 984 * routine is called when a page has been added to the cache or free 985 * queues. 986 * 987 * This routine may not block. 988 * This routine must be called at splvm() 989 */ 990static __inline void 991vm_page_free_wakeup(void) 992{ 993 994 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 995 /* 996 * if pageout daemon needs pages, then tell it that there are 997 * some free. 998 */ 999 if (vm_pageout_pages_needed && 1000 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1001 wakeup(&vm_pageout_pages_needed); 1002 vm_pageout_pages_needed = 0; 1003 } 1004 /* 1005 * wakeup processes that are waiting on memory if we hit a 1006 * high water mark. And wakeup scheduler process if we have 1007 * lots of memory. this process will swapin processes. 1008 */ 1009 if (vm_pages_needed && !vm_page_count_min()) { 1010 vm_pages_needed = 0; 1011 wakeup(&cnt.v_free_count); 1012 } 1013} 1014 1015/* 1016 * vm_page_free_toq: 1017 * 1018 * Returns the given page to the PQ_FREE list, 1019 * disassociating it with any VM object. 1020 * 1021 * Object and page must be locked prior to entry. 1022 * This routine may not block. 1023 */ 1024 1025void 1026vm_page_free_toq(vm_page_t m) 1027{ 1028 int s; 1029 struct vpgqueues *pq; 1030 vm_object_t object = m->object; 1031 1032 GIANT_REQUIRED; 1033 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1034 s = splvm(); 1035 cnt.v_tfree++; 1036 1037 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1038 printf( 1039 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1040 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1041 m->hold_count); 1042 if ((m->queue - m->pc) == PQ_FREE) 1043 panic("vm_page_free: freeing free page"); 1044 else 1045 panic("vm_page_free: freeing busy page"); 1046 } 1047 1048 /* 1049 * unqueue, then remove page. Note that we cannot destroy 1050 * the page here because we do not want to call the pager's 1051 * callback routine until after we've put the page on the 1052 * appropriate free queue. 1053 */ 1054 vm_pageq_remove_nowakeup(m); 1055 vm_page_remove(m); 1056 1057 /* 1058 * If fictitious remove object association and 1059 * return, otherwise delay object association removal. 1060 */ 1061 if ((m->flags & PG_FICTITIOUS) != 0) { 1062 splx(s); 1063 return; 1064 } 1065 1066 m->valid = 0; 1067 vm_page_undirty(m); 1068 1069 if (m->wire_count != 0) { 1070 if (m->wire_count > 1) { 1071 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1072 m->wire_count, (long)m->pindex); 1073 } 1074 panic("vm_page_free: freeing wired page\n"); 1075 } 1076 1077 /* 1078 * If we've exhausted the object's resident pages we want to free 1079 * it up. 1080 */ 1081 if (object && 1082 (object->type == OBJT_VNODE) && 1083 ((object->flags & OBJ_DEAD) == 0) 1084 ) { 1085 struct vnode *vp = (struct vnode *)object->handle; 1086 1087 if (vp) { 1088 VI_LOCK(vp); 1089 if (VSHOULDFREE(vp)) 1090 vfree(vp); 1091 VI_UNLOCK(vp); 1092 } 1093 } 1094 1095 /* 1096 * Clear the UNMANAGED flag when freeing an unmanaged page. 1097 */ 1098 if (m->flags & PG_UNMANAGED) { 1099 m->flags &= ~PG_UNMANAGED; 1100 } 1101 1102 if (m->hold_count != 0) { 1103 m->flags &= ~PG_ZERO; 1104 m->queue = PQ_HOLD; 1105 } else 1106 m->queue = PQ_FREE + m->pc; 1107 pq = &vm_page_queues[m->queue]; 1108 mtx_lock_spin(&vm_page_queue_free_mtx); 1109 pq->lcnt++; 1110 ++(*pq->cnt); 1111 1112 /* 1113 * Put zero'd pages on the end ( where we look for zero'd pages 1114 * first ) and non-zerod pages at the head. 1115 */ 1116 if (m->flags & PG_ZERO) { 1117 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1118 ++vm_page_zero_count; 1119 } else { 1120 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1121 } 1122 mtx_unlock_spin(&vm_page_queue_free_mtx); 1123 vm_page_free_wakeup(); 1124 splx(s); 1125} 1126 1127/* 1128 * vm_page_unmanage: 1129 * 1130 * Prevent PV management from being done on the page. The page is 1131 * removed from the paging queues as if it were wired, and as a 1132 * consequence of no longer being managed the pageout daemon will not 1133 * touch it (since there is no way to locate the pte mappings for the 1134 * page). madvise() calls that mess with the pmap will also no longer 1135 * operate on the page. 1136 * 1137 * Beyond that the page is still reasonably 'normal'. Freeing the page 1138 * will clear the flag. 1139 * 1140 * This routine is used by OBJT_PHYS objects - objects using unswappable 1141 * physical memory as backing store rather then swap-backed memory and 1142 * will eventually be extended to support 4MB unmanaged physical 1143 * mappings. 1144 */ 1145void 1146vm_page_unmanage(vm_page_t m) 1147{ 1148 int s; 1149 1150 s = splvm(); 1151 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1152 if ((m->flags & PG_UNMANAGED) == 0) { 1153 if (m->wire_count == 0) 1154 vm_pageq_remove(m); 1155 } 1156 vm_page_flag_set(m, PG_UNMANAGED); 1157 splx(s); 1158} 1159 1160/* 1161 * vm_page_wire: 1162 * 1163 * Mark this page as wired down by yet 1164 * another map, removing it from paging queues 1165 * as necessary. 1166 * 1167 * The page queues must be locked. 1168 * This routine may not block. 1169 */ 1170void 1171vm_page_wire(vm_page_t m) 1172{ 1173 int s; 1174 1175 /* 1176 * Only bump the wire statistics if the page is not already wired, 1177 * and only unqueue the page if it is on some queue (if it is unmanaged 1178 * it is already off the queues). 1179 */ 1180 s = splvm(); 1181 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1182 if (m->wire_count == 0) { 1183 if ((m->flags & PG_UNMANAGED) == 0) 1184 vm_pageq_remove(m); 1185 atomic_add_int(&cnt.v_wire_count, 1); 1186 } 1187 m->wire_count++; 1188 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1189 splx(s); 1190} 1191 1192/* 1193 * vm_page_unwire: 1194 * 1195 * Release one wiring of this page, potentially 1196 * enabling it to be paged again. 1197 * 1198 * Many pages placed on the inactive queue should actually go 1199 * into the cache, but it is difficult to figure out which. What 1200 * we do instead, if the inactive target is well met, is to put 1201 * clean pages at the head of the inactive queue instead of the tail. 1202 * This will cause them to be moved to the cache more quickly and 1203 * if not actively re-referenced, freed more quickly. If we just 1204 * stick these pages at the end of the inactive queue, heavy filesystem 1205 * meta-data accesses can cause an unnecessary paging load on memory bound 1206 * processes. This optimization causes one-time-use metadata to be 1207 * reused more quickly. 1208 * 1209 * BUT, if we are in a low-memory situation we have no choice but to 1210 * put clean pages on the cache queue. 1211 * 1212 * A number of routines use vm_page_unwire() to guarantee that the page 1213 * will go into either the inactive or active queues, and will NEVER 1214 * be placed in the cache - for example, just after dirtying a page. 1215 * dirty pages in the cache are not allowed. 1216 * 1217 * The page queues must be locked. 1218 * This routine may not block. 1219 */ 1220void 1221vm_page_unwire(vm_page_t m, int activate) 1222{ 1223 int s; 1224 1225 s = splvm(); 1226 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1227 if (m->wire_count > 0) { 1228 m->wire_count--; 1229 if (m->wire_count == 0) { 1230 atomic_subtract_int(&cnt.v_wire_count, 1); 1231 if (m->flags & PG_UNMANAGED) { 1232 ; 1233 } else if (activate) 1234 vm_pageq_enqueue(PQ_ACTIVE, m); 1235 else { 1236 vm_page_flag_clear(m, PG_WINATCFLS); 1237 vm_pageq_enqueue(PQ_INACTIVE, m); 1238 } 1239 } 1240 } else { 1241 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1242 } 1243 splx(s); 1244} 1245 1246 1247/* 1248 * Move the specified page to the inactive queue. If the page has 1249 * any associated swap, the swap is deallocated. 1250 * 1251 * Normally athead is 0 resulting in LRU operation. athead is set 1252 * to 1 if we want this page to be 'as if it were placed in the cache', 1253 * except without unmapping it from the process address space. 1254 * 1255 * This routine may not block. 1256 */ 1257static __inline void 1258_vm_page_deactivate(vm_page_t m, int athead) 1259{ 1260 int s; 1261 1262 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1263 /* 1264 * Ignore if already inactive. 1265 */ 1266 if (m->queue == PQ_INACTIVE) 1267 return; 1268 1269 s = splvm(); 1270 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1271 if ((m->queue - m->pc) == PQ_CACHE) 1272 cnt.v_reactivated++; 1273 vm_page_flag_clear(m, PG_WINATCFLS); 1274 vm_pageq_remove(m); 1275 if (athead) 1276 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1277 else 1278 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1279 m->queue = PQ_INACTIVE; 1280 vm_page_queues[PQ_INACTIVE].lcnt++; 1281 cnt.v_inactive_count++; 1282 } 1283 splx(s); 1284} 1285 1286void 1287vm_page_deactivate(vm_page_t m) 1288{ 1289 _vm_page_deactivate(m, 0); 1290} 1291 1292/* 1293 * vm_page_try_to_cache: 1294 * 1295 * Returns 0 on failure, 1 on success 1296 */ 1297int 1298vm_page_try_to_cache(vm_page_t m) 1299{ 1300 1301 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1302 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1303 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1304 return (0); 1305 } 1306 vm_page_test_dirty(m); 1307 if (m->dirty) 1308 return (0); 1309 vm_page_cache(m); 1310 return (1); 1311} 1312 1313/* 1314 * vm_page_try_to_free() 1315 * 1316 * Attempt to free the page. If we cannot free it, we do nothing. 1317 * 1 is returned on success, 0 on failure. 1318 */ 1319int 1320vm_page_try_to_free(vm_page_t m) 1321{ 1322 1323 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1324 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1325 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1326 return (0); 1327 } 1328 vm_page_test_dirty(m); 1329 if (m->dirty) 1330 return (0); 1331 vm_page_busy(m); 1332 pmap_remove_all(m); 1333 vm_page_free(m); 1334 return (1); 1335} 1336 1337/* 1338 * vm_page_cache 1339 * 1340 * Put the specified page onto the page cache queue (if appropriate). 1341 * 1342 * This routine may not block. 1343 */ 1344void 1345vm_page_cache(vm_page_t m) 1346{ 1347 int s; 1348 1349 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1350 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1351 printf("vm_page_cache: attempting to cache busy page\n"); 1352 return; 1353 } 1354 if ((m->queue - m->pc) == PQ_CACHE) 1355 return; 1356 1357 /* 1358 * Remove all pmaps and indicate that the page is not 1359 * writeable or mapped. 1360 */ 1361 pmap_remove_all(m); 1362 if (m->dirty != 0) { 1363 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1364 (long)m->pindex); 1365 } 1366 s = splvm(); 1367 vm_pageq_remove_nowakeup(m); 1368 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1369 vm_page_free_wakeup(); 1370 splx(s); 1371} 1372 1373/* 1374 * vm_page_dontneed 1375 * 1376 * Cache, deactivate, or do nothing as appropriate. This routine 1377 * is typically used by madvise() MADV_DONTNEED. 1378 * 1379 * Generally speaking we want to move the page into the cache so 1380 * it gets reused quickly. However, this can result in a silly syndrome 1381 * due to the page recycling too quickly. Small objects will not be 1382 * fully cached. On the otherhand, if we move the page to the inactive 1383 * queue we wind up with a problem whereby very large objects 1384 * unnecessarily blow away our inactive and cache queues. 1385 * 1386 * The solution is to move the pages based on a fixed weighting. We 1387 * either leave them alone, deactivate them, or move them to the cache, 1388 * where moving them to the cache has the highest weighting. 1389 * By forcing some pages into other queues we eventually force the 1390 * system to balance the queues, potentially recovering other unrelated 1391 * space from active. The idea is to not force this to happen too 1392 * often. 1393 */ 1394void 1395vm_page_dontneed(vm_page_t m) 1396{ 1397 static int dnweight; 1398 int dnw; 1399 int head; 1400 1401 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1402 dnw = ++dnweight; 1403 1404 /* 1405 * occassionally leave the page alone 1406 */ 1407 if ((dnw & 0x01F0) == 0 || 1408 m->queue == PQ_INACTIVE || 1409 m->queue - m->pc == PQ_CACHE 1410 ) { 1411 if (m->act_count >= ACT_INIT) 1412 --m->act_count; 1413 return; 1414 } 1415 1416 if (m->dirty == 0) 1417 vm_page_test_dirty(m); 1418 1419 if (m->dirty || (dnw & 0x0070) == 0) { 1420 /* 1421 * Deactivate the page 3 times out of 32. 1422 */ 1423 head = 0; 1424 } else { 1425 /* 1426 * Cache the page 28 times out of every 32. Note that 1427 * the page is deactivated instead of cached, but placed 1428 * at the head of the queue instead of the tail. 1429 */ 1430 head = 1; 1431 } 1432 _vm_page_deactivate(m, head); 1433} 1434 1435/* 1436 * Grab a page, waiting until we are waken up due to the page 1437 * changing state. We keep on waiting, if the page continues 1438 * to be in the object. If the page doesn't exist, allocate it. 1439 * 1440 * This routine may block. 1441 */ 1442vm_page_t 1443vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1444{ 1445 vm_page_t m; 1446 int s, generation; 1447 1448 GIANT_REQUIRED; 1449retrylookup: 1450 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1451 vm_page_lock_queues(); 1452 if (m->busy || (m->flags & PG_BUSY)) { 1453 generation = object->generation; 1454 1455 s = splvm(); 1456 while ((object->generation == generation) && 1457 (m->busy || (m->flags & PG_BUSY))) { 1458 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1459 msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0); 1460 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1461 vm_page_unlock_queues(); 1462 splx(s); 1463 return NULL; 1464 } 1465 } 1466 vm_page_unlock_queues(); 1467 splx(s); 1468 goto retrylookup; 1469 } else { 1470 if (allocflags & VM_ALLOC_WIRED) 1471 vm_page_wire(m); 1472 vm_page_busy(m); 1473 vm_page_unlock_queues(); 1474 return m; 1475 } 1476 } 1477 1478 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1479 if (m == NULL) { 1480 VM_WAIT; 1481 if ((allocflags & VM_ALLOC_RETRY) == 0) 1482 return NULL; 1483 goto retrylookup; 1484 } 1485 1486 return m; 1487} 1488 1489/* 1490 * Mapping function for valid bits or for dirty bits in 1491 * a page. May not block. 1492 * 1493 * Inputs are required to range within a page. 1494 */ 1495__inline int 1496vm_page_bits(int base, int size) 1497{ 1498 int first_bit; 1499 int last_bit; 1500 1501 KASSERT( 1502 base + size <= PAGE_SIZE, 1503 ("vm_page_bits: illegal base/size %d/%d", base, size) 1504 ); 1505 1506 if (size == 0) /* handle degenerate case */ 1507 return (0); 1508 1509 first_bit = base >> DEV_BSHIFT; 1510 last_bit = (base + size - 1) >> DEV_BSHIFT; 1511 1512 return ((2 << last_bit) - (1 << first_bit)); 1513} 1514 1515/* 1516 * vm_page_set_validclean: 1517 * 1518 * Sets portions of a page valid and clean. The arguments are expected 1519 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1520 * of any partial chunks touched by the range. The invalid portion of 1521 * such chunks will be zero'd. 1522 * 1523 * This routine may not block. 1524 * 1525 * (base + size) must be less then or equal to PAGE_SIZE. 1526 */ 1527void 1528vm_page_set_validclean(vm_page_t m, int base, int size) 1529{ 1530 int pagebits; 1531 int frag; 1532 int endoff; 1533 1534 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1535 if (size == 0) /* handle degenerate case */ 1536 return; 1537 1538 /* 1539 * If the base is not DEV_BSIZE aligned and the valid 1540 * bit is clear, we have to zero out a portion of the 1541 * first block. 1542 */ 1543 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1544 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1545 pmap_zero_page_area(m, frag, base - frag); 1546 1547 /* 1548 * If the ending offset is not DEV_BSIZE aligned and the 1549 * valid bit is clear, we have to zero out a portion of 1550 * the last block. 1551 */ 1552 endoff = base + size; 1553 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1554 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1555 pmap_zero_page_area(m, endoff, 1556 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1557 1558 /* 1559 * Set valid, clear dirty bits. If validating the entire 1560 * page we can safely clear the pmap modify bit. We also 1561 * use this opportunity to clear the PG_NOSYNC flag. If a process 1562 * takes a write fault on a MAP_NOSYNC memory area the flag will 1563 * be set again. 1564 * 1565 * We set valid bits inclusive of any overlap, but we can only 1566 * clear dirty bits for DEV_BSIZE chunks that are fully within 1567 * the range. 1568 */ 1569 pagebits = vm_page_bits(base, size); 1570 m->valid |= pagebits; 1571#if 0 /* NOT YET */ 1572 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1573 frag = DEV_BSIZE - frag; 1574 base += frag; 1575 size -= frag; 1576 if (size < 0) 1577 size = 0; 1578 } 1579 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1580#endif 1581 m->dirty &= ~pagebits; 1582 if (base == 0 && size == PAGE_SIZE) { 1583 pmap_clear_modify(m); 1584 vm_page_flag_clear(m, PG_NOSYNC); 1585 } 1586} 1587 1588#if 0 1589 1590void 1591vm_page_set_dirty(vm_page_t m, int base, int size) 1592{ 1593 m->dirty |= vm_page_bits(base, size); 1594} 1595 1596#endif 1597 1598void 1599vm_page_clear_dirty(vm_page_t m, int base, int size) 1600{ 1601 GIANT_REQUIRED; 1602 m->dirty &= ~vm_page_bits(base, size); 1603} 1604 1605/* 1606 * vm_page_set_invalid: 1607 * 1608 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1609 * valid and dirty bits for the effected areas are cleared. 1610 * 1611 * May not block. 1612 */ 1613void 1614vm_page_set_invalid(vm_page_t m, int base, int size) 1615{ 1616 int bits; 1617 1618 GIANT_REQUIRED; 1619 bits = vm_page_bits(base, size); 1620 m->valid &= ~bits; 1621 m->dirty &= ~bits; 1622 m->object->generation++; 1623} 1624 1625/* 1626 * vm_page_zero_invalid() 1627 * 1628 * The kernel assumes that the invalid portions of a page contain 1629 * garbage, but such pages can be mapped into memory by user code. 1630 * When this occurs, we must zero out the non-valid portions of the 1631 * page so user code sees what it expects. 1632 * 1633 * Pages are most often semi-valid when the end of a file is mapped 1634 * into memory and the file's size is not page aligned. 1635 */ 1636void 1637vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1638{ 1639 int b; 1640 int i; 1641 1642 /* 1643 * Scan the valid bits looking for invalid sections that 1644 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1645 * valid bit may be set ) have already been zerod by 1646 * vm_page_set_validclean(). 1647 */ 1648 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1649 if (i == (PAGE_SIZE / DEV_BSIZE) || 1650 (m->valid & (1 << i)) 1651 ) { 1652 if (i > b) { 1653 pmap_zero_page_area(m, 1654 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1655 } 1656 b = i + 1; 1657 } 1658 } 1659 1660 /* 1661 * setvalid is TRUE when we can safely set the zero'd areas 1662 * as being valid. We can do this if there are no cache consistancy 1663 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1664 */ 1665 if (setvalid) 1666 m->valid = VM_PAGE_BITS_ALL; 1667} 1668 1669/* 1670 * vm_page_is_valid: 1671 * 1672 * Is (partial) page valid? Note that the case where size == 0 1673 * will return FALSE in the degenerate case where the page is 1674 * entirely invalid, and TRUE otherwise. 1675 * 1676 * May not block. 1677 */ 1678int 1679vm_page_is_valid(vm_page_t m, int base, int size) 1680{ 1681 int bits = vm_page_bits(base, size); 1682 1683 if (m->valid && ((m->valid & bits) == bits)) 1684 return 1; 1685 else 1686 return 0; 1687} 1688 1689/* 1690 * update dirty bits from pmap/mmu. May not block. 1691 */ 1692void 1693vm_page_test_dirty(vm_page_t m) 1694{ 1695 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1696 vm_page_dirty(m); 1697 } 1698} 1699 1700int so_zerocp_fullpage = 0; 1701 1702void 1703vm_page_cowfault(vm_page_t m) 1704{ 1705 vm_page_t mnew; 1706 vm_object_t object; 1707 vm_pindex_t pindex; 1708 1709 object = m->object; 1710 pindex = m->pindex; 1711 vm_page_busy(m); 1712 1713 retry_alloc: 1714 vm_page_remove(m); 1715 /* 1716 * An interrupt allocation is requested because the page 1717 * queues lock is held. 1718 */ 1719 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1720 if (mnew == NULL) { 1721 vm_page_insert(m, object, pindex); 1722 vm_page_unlock_queues(); 1723 VM_WAIT; 1724 vm_page_lock_queues(); 1725 goto retry_alloc; 1726 } 1727 1728 if (m->cow == 0) { 1729 /* 1730 * check to see if we raced with an xmit complete when 1731 * waiting to allocate a page. If so, put things back 1732 * the way they were 1733 */ 1734 vm_page_busy(mnew); 1735 vm_page_free(mnew); 1736 vm_page_insert(m, object, pindex); 1737 } else { /* clear COW & copy page */ 1738 if (so_zerocp_fullpage) { 1739 mnew->valid = VM_PAGE_BITS_ALL; 1740 } else { 1741 vm_page_copy(m, mnew); 1742 } 1743 vm_page_dirty(mnew); 1744 vm_page_flag_clear(mnew, PG_BUSY); 1745 } 1746} 1747 1748void 1749vm_page_cowclear(vm_page_t m) 1750{ 1751 1752 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1753 if (m->cow) { 1754 m->cow--; 1755 /* 1756 * let vm_fault add back write permission lazily 1757 */ 1758 } 1759 /* 1760 * sf_buf_free() will free the page, so we needn't do it here 1761 */ 1762} 1763 1764void 1765vm_page_cowsetup(vm_page_t m) 1766{ 1767 1768 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1769 m->cow++; 1770 pmap_page_protect(m, VM_PROT_READ); 1771} 1772 1773#include "opt_ddb.h" 1774#ifdef DDB 1775#include <sys/kernel.h> 1776 1777#include <ddb/ddb.h> 1778 1779DB_SHOW_COMMAND(page, vm_page_print_page_info) 1780{ 1781 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1782 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1783 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1784 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1785 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1786 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1787 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1788 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1789 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1790 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1791} 1792 1793DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1794{ 1795 int i; 1796 db_printf("PQ_FREE:"); 1797 for (i = 0; i < PQ_L2_SIZE; i++) { 1798 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1799 } 1800 db_printf("\n"); 1801 1802 db_printf("PQ_CACHE:"); 1803 for (i = 0; i < PQ_L2_SIZE; i++) { 1804 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1805 } 1806 db_printf("\n"); 1807 1808 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1809 vm_page_queues[PQ_ACTIVE].lcnt, 1810 vm_page_queues[PQ_INACTIVE].lcnt); 1811} 1812#endif /* DDB */ 1813