vm_page.c revision 124028
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 */ 38 39/* 40 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 41 * All rights reserved. 42 * 43 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 44 * 45 * Permission to use, copy, modify and distribute this software and 46 * its documentation is hereby granted, provided that both the copyright 47 * notice and this permission notice appear in all copies of the 48 * software, derivative works or modified versions, and any portions 49 * thereof, and that both notices appear in supporting documentation. 50 * 51 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 52 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 53 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 54 * 55 * Carnegie Mellon requests users of this software to return to 56 * 57 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 58 * School of Computer Science 59 * Carnegie Mellon University 60 * Pittsburgh PA 15213-3890 61 * 62 * any improvements or extensions that they make and grant Carnegie the 63 * rights to redistribute these changes. 64 */ 65 66/* 67 * GENERAL RULES ON VM_PAGE MANIPULATION 68 * 69 * - a pageq mutex is required when adding or removing a page from a 70 * page queue (vm_page_queue[]), regardless of other mutexes or the 71 * busy state of a page. 72 * 73 * - a hash chain mutex is required when associating or disassociating 74 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 75 * regardless of other mutexes or the busy state of a page. 76 * 77 * - either a hash chain mutex OR a busied page is required in order 78 * to modify the page flags. A hash chain mutex must be obtained in 79 * order to busy a page. A page's flags cannot be modified by a 80 * hash chain mutex if the page is marked busy. 81 * 82 * - The object memq mutex is held when inserting or removing 83 * pages from an object (vm_page_insert() or vm_page_remove()). This 84 * is different from the object's main mutex. 85 * 86 * Generally speaking, you have to be aware of side effects when running 87 * vm_page ops. A vm_page_lookup() will return with the hash chain 88 * locked, whether it was able to lookup the page or not. vm_page_free(), 89 * vm_page_cache(), vm_page_activate(), and a number of other routines 90 * will release the hash chain mutex for you. Intermediate manipulation 91 * routines such as vm_page_flag_set() expect the hash chain to be held 92 * on entry and the hash chain will remain held on return. 93 * 94 * pageq scanning can only occur with the pageq in question locked. 95 * We have a known bottleneck with the active queue, but the cache 96 * and free queues are actually arrays already. 97 */ 98 99/* 100 * Resident memory management module. 101 */ 102 103#include <sys/cdefs.h> 104__FBSDID("$FreeBSD: head/sys/vm/vm_page.c 124028 2003-12-31 19:02:01Z alc $"); 105 106#include <sys/param.h> 107#include <sys/systm.h> 108#include <sys/lock.h> 109#include <sys/malloc.h> 110#include <sys/mutex.h> 111#include <sys/proc.h> 112#include <sys/vmmeter.h> 113#include <sys/vnode.h> 114 115#include <vm/vm.h> 116#include <vm/vm_param.h> 117#include <vm/vm_kern.h> 118#include <vm/vm_object.h> 119#include <vm/vm_page.h> 120#include <vm/vm_pageout.h> 121#include <vm/vm_pager.h> 122#include <vm/vm_extern.h> 123#include <vm/uma.h> 124#include <vm/uma_int.h> 125 126/* 127 * Associated with page of user-allocatable memory is a 128 * page structure. 129 */ 130 131struct mtx vm_page_queue_mtx; 132struct mtx vm_page_queue_free_mtx; 133 134vm_page_t vm_page_array = 0; 135int vm_page_array_size = 0; 136long first_page = 0; 137int vm_page_zero_count = 0; 138 139/* 140 * vm_set_page_size: 141 * 142 * Sets the page size, perhaps based upon the memory 143 * size. Must be called before any use of page-size 144 * dependent functions. 145 */ 146void 147vm_set_page_size(void) 148{ 149 if (cnt.v_page_size == 0) 150 cnt.v_page_size = PAGE_SIZE; 151 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 152 panic("vm_set_page_size: page size not a power of two"); 153} 154 155/* 156 * vm_page_startup: 157 * 158 * Initializes the resident memory module. 159 * 160 * Allocates memory for the page cells, and 161 * for the object/offset-to-page hash table headers. 162 * Each page cell is initialized and placed on the free list. 163 */ 164vm_offset_t 165vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 166{ 167 vm_offset_t mapped; 168 vm_size_t npages; 169 vm_paddr_t page_range; 170 vm_paddr_t new_end; 171 int i; 172 vm_paddr_t pa; 173 int nblocks; 174 vm_paddr_t last_pa; 175 176 /* the biggest memory array is the second group of pages */ 177 vm_paddr_t end; 178 vm_paddr_t biggestsize; 179 int biggestone; 180 181 vm_paddr_t total; 182 vm_size_t bootpages; 183 184 total = 0; 185 biggestsize = 0; 186 biggestone = 0; 187 nblocks = 0; 188 vaddr = round_page(vaddr); 189 190 for (i = 0; phys_avail[i + 1]; i += 2) { 191 phys_avail[i] = round_page(phys_avail[i]); 192 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 193 } 194 195 for (i = 0; phys_avail[i + 1]; i += 2) { 196 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 197 198 if (size > biggestsize) { 199 biggestone = i; 200 biggestsize = size; 201 } 202 ++nblocks; 203 total += size; 204 } 205 206 end = phys_avail[biggestone+1]; 207 208 /* 209 * Initialize the locks. 210 */ 211 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF); 212 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 213 MTX_SPIN); 214 215 /* 216 * Initialize the queue headers for the free queue, the active queue 217 * and the inactive queue. 218 */ 219 vm_pageq_init(); 220 221 /* 222 * Allocate memory for use when boot strapping the kernel memory 223 * allocator. 224 */ 225 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 226 new_end = end - bootpages; 227 new_end = trunc_page(new_end); 228 mapped = pmap_map(&vaddr, new_end, end, 229 VM_PROT_READ | VM_PROT_WRITE); 230 bzero((caddr_t) mapped, end - new_end); 231 uma_startup((caddr_t)mapped); 232 233 /* 234 * Compute the number of pages of memory that will be available for 235 * use (taking into account the overhead of a page structure per 236 * page). 237 */ 238 first_page = phys_avail[0] / PAGE_SIZE; 239 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 240 npages = (total - (page_range * sizeof(struct vm_page)) - 241 (end - new_end)) / PAGE_SIZE; 242 end = new_end; 243 244 /* 245 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 246 */ 247 vaddr += PAGE_SIZE; 248 249 /* 250 * Initialize the mem entry structures now, and put them in the free 251 * queue. 252 */ 253 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 254 mapped = pmap_map(&vaddr, new_end, end, 255 VM_PROT_READ | VM_PROT_WRITE); 256 vm_page_array = (vm_page_t) mapped; 257 phys_avail[biggestone + 1] = new_end; 258 259 /* 260 * Clear all of the page structures 261 */ 262 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 263 vm_page_array_size = page_range; 264 265 /* 266 * Construct the free queue(s) in descending order (by physical 267 * address) so that the first 16MB of physical memory is allocated 268 * last rather than first. On large-memory machines, this avoids 269 * the exhaustion of low physical memory before isa_dmainit has run. 270 */ 271 cnt.v_page_count = 0; 272 cnt.v_free_count = 0; 273 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 274 pa = phys_avail[i]; 275 last_pa = phys_avail[i + 1]; 276 while (pa < last_pa && npages-- > 0) { 277 vm_pageq_add_new_page(pa); 278 pa += PAGE_SIZE; 279 } 280 } 281 return (vaddr); 282} 283 284void 285vm_page_flag_set(vm_page_t m, unsigned short bits) 286{ 287 288 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 289 m->flags |= bits; 290} 291 292void 293vm_page_flag_clear(vm_page_t m, unsigned short bits) 294{ 295 296 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 297 m->flags &= ~bits; 298} 299 300void 301vm_page_busy(vm_page_t m) 302{ 303 KASSERT((m->flags & PG_BUSY) == 0, 304 ("vm_page_busy: page already busy!!!")); 305 vm_page_flag_set(m, PG_BUSY); 306} 307 308/* 309 * vm_page_flash: 310 * 311 * wakeup anyone waiting for the page. 312 */ 313void 314vm_page_flash(vm_page_t m) 315{ 316 if (m->flags & PG_WANTED) { 317 vm_page_flag_clear(m, PG_WANTED); 318 wakeup(m); 319 } 320} 321 322/* 323 * vm_page_wakeup: 324 * 325 * clear the PG_BUSY flag and wakeup anyone waiting for the 326 * page. 327 * 328 */ 329void 330vm_page_wakeup(vm_page_t m) 331{ 332 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 333 vm_page_flag_clear(m, PG_BUSY); 334 vm_page_flash(m); 335} 336 337void 338vm_page_io_start(vm_page_t m) 339{ 340 341 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 342 m->busy++; 343} 344 345void 346vm_page_io_finish(vm_page_t m) 347{ 348 349 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 350 m->busy--; 351 if (m->busy == 0) 352 vm_page_flash(m); 353} 354 355/* 356 * Keep page from being freed by the page daemon 357 * much of the same effect as wiring, except much lower 358 * overhead and should be used only for *very* temporary 359 * holding ("wiring"). 360 */ 361void 362vm_page_hold(vm_page_t mem) 363{ 364 365 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 366 mem->hold_count++; 367} 368 369void 370vm_page_unhold(vm_page_t mem) 371{ 372 373 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 374 --mem->hold_count; 375 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 376 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 377 vm_page_free_toq(mem); 378} 379 380/* 381 * vm_page_free: 382 * 383 * Free a page 384 * 385 * The clearing of PG_ZERO is a temporary safety until the code can be 386 * reviewed to determine that PG_ZERO is being properly cleared on 387 * write faults or maps. PG_ZERO was previously cleared in 388 * vm_page_alloc(). 389 */ 390void 391vm_page_free(vm_page_t m) 392{ 393 vm_page_flag_clear(m, PG_ZERO); 394 vm_page_free_toq(m); 395 vm_page_zero_idle_wakeup(); 396} 397 398/* 399 * vm_page_free_zero: 400 * 401 * Free a page to the zerod-pages queue 402 */ 403void 404vm_page_free_zero(vm_page_t m) 405{ 406 vm_page_flag_set(m, PG_ZERO); 407 vm_page_free_toq(m); 408} 409 410/* 411 * vm_page_sleep_if_busy: 412 * 413 * Sleep and release the page queues lock if PG_BUSY is set or, 414 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 415 * thread slept and the page queues lock was released. 416 * Otherwise, retains the page queues lock and returns FALSE. 417 */ 418int 419vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 420{ 421 int is_object_locked; 422 423 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 424 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 425 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 426 /* 427 * Remove mtx_owned() after vm_object locking is finished. 428 */ 429 if ((is_object_locked = m->object != NULL && 430 mtx_owned(&m->object->mtx))) 431 mtx_unlock(&m->object->mtx); 432 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 433 if (is_object_locked) 434 mtx_lock(&m->object->mtx); 435 return (TRUE); 436 } 437 return (FALSE); 438} 439 440/* 441 * vm_page_dirty: 442 * 443 * make page all dirty 444 */ 445void 446vm_page_dirty(vm_page_t m) 447{ 448 KASSERT(m->queue - m->pc != PQ_CACHE, 449 ("vm_page_dirty: page in cache!")); 450 KASSERT(m->queue - m->pc != PQ_FREE, 451 ("vm_page_dirty: page is free!")); 452 m->dirty = VM_PAGE_BITS_ALL; 453} 454 455/* 456 * vm_page_splay: 457 * 458 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 459 * the vm_page containing the given pindex. If, however, that 460 * pindex is not found in the vm_object, returns a vm_page that is 461 * adjacent to the pindex, coming before or after it. 462 */ 463vm_page_t 464vm_page_splay(vm_pindex_t pindex, vm_page_t root) 465{ 466 struct vm_page dummy; 467 vm_page_t lefttreemax, righttreemin, y; 468 469 if (root == NULL) 470 return (root); 471 lefttreemax = righttreemin = &dummy; 472 for (;; root = y) { 473 if (pindex < root->pindex) { 474 if ((y = root->left) == NULL) 475 break; 476 if (pindex < y->pindex) { 477 /* Rotate right. */ 478 root->left = y->right; 479 y->right = root; 480 root = y; 481 if ((y = root->left) == NULL) 482 break; 483 } 484 /* Link into the new root's right tree. */ 485 righttreemin->left = root; 486 righttreemin = root; 487 } else if (pindex > root->pindex) { 488 if ((y = root->right) == NULL) 489 break; 490 if (pindex > y->pindex) { 491 /* Rotate left. */ 492 root->right = y->left; 493 y->left = root; 494 root = y; 495 if ((y = root->right) == NULL) 496 break; 497 } 498 /* Link into the new root's left tree. */ 499 lefttreemax->right = root; 500 lefttreemax = root; 501 } else 502 break; 503 } 504 /* Assemble the new root. */ 505 lefttreemax->right = root->left; 506 righttreemin->left = root->right; 507 root->left = dummy.right; 508 root->right = dummy.left; 509 return (root); 510} 511 512/* 513 * vm_page_insert: [ internal use only ] 514 * 515 * Inserts the given mem entry into the object and object list. 516 * 517 * The pagetables are not updated but will presumably fault the page 518 * in if necessary, or if a kernel page the caller will at some point 519 * enter the page into the kernel's pmap. We are not allowed to block 520 * here so we *can't* do this anyway. 521 * 522 * The object and page must be locked, and must be splhigh. 523 * This routine may not block. 524 */ 525void 526vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 527{ 528 vm_page_t root; 529 530 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 531 if (m->object != NULL) 532 panic("vm_page_insert: already inserted"); 533 534 /* 535 * Record the object/offset pair in this page 536 */ 537 m->object = object; 538 m->pindex = pindex; 539 540 /* 541 * Now link into the object's ordered list of backed pages. 542 */ 543 root = object->root; 544 if (root == NULL) { 545 m->left = NULL; 546 m->right = NULL; 547 TAILQ_INSERT_TAIL(&object->memq, m, listq); 548 } else { 549 root = vm_page_splay(pindex, root); 550 if (pindex < root->pindex) { 551 m->left = root->left; 552 m->right = root; 553 root->left = NULL; 554 TAILQ_INSERT_BEFORE(root, m, listq); 555 } else { 556 m->right = root->right; 557 m->left = root; 558 root->right = NULL; 559 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 560 } 561 } 562 object->root = m; 563 object->generation++; 564 565 /* 566 * show that the object has one more resident page. 567 */ 568 object->resident_page_count++; 569 570 /* 571 * Since we are inserting a new and possibly dirty page, 572 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 573 */ 574 if (m->flags & PG_WRITEABLE) 575 vm_object_set_writeable_dirty(object); 576} 577 578/* 579 * vm_page_remove: 580 * NOTE: used by device pager as well -wfj 581 * 582 * Removes the given mem entry from the object/offset-page 583 * table and the object page list, but do not invalidate/terminate 584 * the backing store. 585 * 586 * The object and page must be locked, and at splhigh. 587 * The underlying pmap entry (if any) is NOT removed here. 588 * This routine may not block. 589 */ 590void 591vm_page_remove(vm_page_t m) 592{ 593 vm_object_t object; 594 vm_page_t root; 595 596 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 597 if (m->object == NULL) 598 return; 599#ifndef __alpha__ 600 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 601#endif 602 if ((m->flags & PG_BUSY) == 0) { 603 panic("vm_page_remove: page not busy"); 604 } 605 606 /* 607 * Basically destroy the page. 608 */ 609 vm_page_wakeup(m); 610 611 object = m->object; 612 613 /* 614 * Now remove from the object's list of backed pages. 615 */ 616 if (m != object->root) 617 vm_page_splay(m->pindex, object->root); 618 if (m->left == NULL) 619 root = m->right; 620 else { 621 root = vm_page_splay(m->pindex, m->left); 622 root->right = m->right; 623 } 624 object->root = root; 625 TAILQ_REMOVE(&object->memq, m, listq); 626 627 /* 628 * And show that the object has one fewer resident page. 629 */ 630 object->resident_page_count--; 631 object->generation++; 632 633 m->object = NULL; 634} 635 636/* 637 * vm_page_lookup: 638 * 639 * Returns the page associated with the object/offset 640 * pair specified; if none is found, NULL is returned. 641 * 642 * The object must be locked. 643 * This routine may not block. 644 * This is a critical path routine 645 */ 646vm_page_t 647vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 648{ 649 vm_page_t m; 650 651 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 652 if ((m = object->root) != NULL && m->pindex != pindex) { 653 m = vm_page_splay(pindex, m); 654 if ((object->root = m)->pindex != pindex) 655 m = NULL; 656 } 657 return (m); 658} 659 660/* 661 * vm_page_rename: 662 * 663 * Move the given memory entry from its 664 * current object to the specified target object/offset. 665 * 666 * The object must be locked. 667 * This routine may not block. 668 * 669 * Note: this routine will raise itself to splvm(), the caller need not. 670 * 671 * Note: swap associated with the page must be invalidated by the move. We 672 * have to do this for several reasons: (1) we aren't freeing the 673 * page, (2) we are dirtying the page, (3) the VM system is probably 674 * moving the page from object A to B, and will then later move 675 * the backing store from A to B and we can't have a conflict. 676 * 677 * Note: we *always* dirty the page. It is necessary both for the 678 * fact that we moved it, and because we may be invalidating 679 * swap. If the page is on the cache, we have to deactivate it 680 * or vm_page_dirty() will panic. Dirty pages are not allowed 681 * on the cache. 682 */ 683void 684vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 685{ 686 int s; 687 688 s = splvm(); 689 vm_page_remove(m); 690 vm_page_insert(m, new_object, new_pindex); 691 if (m->queue - m->pc == PQ_CACHE) 692 vm_page_deactivate(m); 693 vm_page_dirty(m); 694 splx(s); 695} 696 697/* 698 * vm_page_select_cache: 699 * 700 * Find a page on the cache queue with color optimization. As pages 701 * might be found, but not applicable, they are deactivated. This 702 * keeps us from using potentially busy cached pages. 703 * 704 * This routine must be called at splvm(). 705 * This routine may not block. 706 */ 707vm_page_t 708vm_page_select_cache(int color) 709{ 710 vm_page_t m; 711 712 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 713 while (TRUE) { 714 m = vm_pageq_find(PQ_CACHE, color, FALSE); 715 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 716 m->hold_count || m->wire_count || 717 (!VM_OBJECT_TRYLOCK(m->object) && 718 !VM_OBJECT_LOCKED(m->object)))) { 719 vm_page_deactivate(m); 720 continue; 721 } 722 return m; 723 } 724} 725 726/* 727 * vm_page_alloc: 728 * 729 * Allocate and return a memory cell associated 730 * with this VM object/offset pair. 731 * 732 * page_req classes: 733 * VM_ALLOC_NORMAL normal process request 734 * VM_ALLOC_SYSTEM system *really* needs a page 735 * VM_ALLOC_INTERRUPT interrupt time request 736 * VM_ALLOC_ZERO zero page 737 * 738 * This routine may not block. 739 * 740 * Additional special handling is required when called from an 741 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 742 * the page cache in this case. 743 */ 744vm_page_t 745vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 746{ 747 vm_object_t m_object; 748 vm_page_t m = NULL; 749 int color, flags, page_req, s; 750 751 page_req = req & VM_ALLOC_CLASS_MASK; 752 753 if ((req & VM_ALLOC_NOOBJ) == 0) { 754 KASSERT(object != NULL, 755 ("vm_page_alloc: NULL object.")); 756 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 757 KASSERT(!vm_page_lookup(object, pindex), 758 ("vm_page_alloc: page already allocated")); 759 color = (pindex + object->pg_color) & PQ_L2_MASK; 760 } else 761 color = pindex & PQ_L2_MASK; 762 763 /* 764 * The pager is allowed to eat deeper into the free page list. 765 */ 766 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 767 page_req = VM_ALLOC_SYSTEM; 768 }; 769 770 s = splvm(); 771loop: 772 mtx_lock_spin(&vm_page_queue_free_mtx); 773 if (cnt.v_free_count > cnt.v_free_reserved || 774 (page_req == VM_ALLOC_SYSTEM && 775 cnt.v_cache_count == 0 && 776 cnt.v_free_count > cnt.v_interrupt_free_min) || 777 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 778 /* 779 * Allocate from the free queue if the number of free pages 780 * exceeds the minimum for the request class. 781 */ 782 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 783 } else if (page_req != VM_ALLOC_INTERRUPT) { 784 mtx_unlock_spin(&vm_page_queue_free_mtx); 785 /* 786 * Allocatable from cache (non-interrupt only). On success, 787 * we must free the page and try again, thus ensuring that 788 * cnt.v_*_free_min counters are replenished. 789 */ 790 vm_page_lock_queues(); 791 if ((m = vm_page_select_cache(color)) == NULL) { 792 vm_page_unlock_queues(); 793 splx(s); 794#if defined(DIAGNOSTIC) 795 if (cnt.v_cache_count > 0) 796 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 797#endif 798 atomic_add_int(&vm_pageout_deficit, 1); 799 pagedaemon_wakeup(); 800 return (NULL); 801 } 802 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 803 m_object = m->object; 804 VM_OBJECT_LOCK_ASSERT(m_object, MA_OWNED); 805 vm_page_busy(m); 806 pmap_remove_all(m); 807 vm_page_free(m); 808 vm_page_unlock_queues(); 809 if (m_object != object) 810 VM_OBJECT_UNLOCK(m_object); 811 goto loop; 812 } else { 813 /* 814 * Not allocatable from cache from interrupt, give up. 815 */ 816 mtx_unlock_spin(&vm_page_queue_free_mtx); 817 splx(s); 818 atomic_add_int(&vm_pageout_deficit, 1); 819 pagedaemon_wakeup(); 820 return (NULL); 821 } 822 823 /* 824 * At this point we had better have found a good page. 825 */ 826 827 KASSERT( 828 m != NULL, 829 ("vm_page_alloc(): missing page on free queue\n") 830 ); 831 832 /* 833 * Remove from free queue 834 */ 835 836 vm_pageq_remove_nowakeup(m); 837 838 /* 839 * Initialize structure. Only the PG_ZERO flag is inherited. 840 */ 841 flags = PG_BUSY; 842 if (m->flags & PG_ZERO) { 843 vm_page_zero_count--; 844 if (req & VM_ALLOC_ZERO) 845 flags = PG_ZERO | PG_BUSY; 846 } 847 m->flags = flags; 848 if (req & VM_ALLOC_WIRED) { 849 atomic_add_int(&cnt.v_wire_count, 1); 850 m->wire_count = 1; 851 } else 852 m->wire_count = 0; 853 m->hold_count = 0; 854 m->act_count = 0; 855 m->busy = 0; 856 m->valid = 0; 857 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 858 mtx_unlock_spin(&vm_page_queue_free_mtx); 859 860 /* 861 * vm_page_insert() is safe prior to the splx(). Note also that 862 * inserting a page here does not insert it into the pmap (which 863 * could cause us to block allocating memory). We cannot block 864 * anywhere. 865 */ 866 if ((req & VM_ALLOC_NOOBJ) == 0) 867 vm_page_insert(m, object, pindex); 868 else 869 m->pindex = pindex; 870 871 /* 872 * Don't wakeup too often - wakeup the pageout daemon when 873 * we would be nearly out of memory. 874 */ 875 if (vm_paging_needed()) 876 pagedaemon_wakeup(); 877 878 splx(s); 879 return (m); 880} 881 882/* 883 * vm_wait: (also see VM_WAIT macro) 884 * 885 * Block until free pages are available for allocation 886 * - Called in various places before memory allocations. 887 */ 888void 889vm_wait(void) 890{ 891 int s; 892 893 s = splvm(); 894 vm_page_lock_queues(); 895 if (curproc == pageproc) { 896 vm_pageout_pages_needed = 1; 897 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 898 PDROP | PSWP, "VMWait", 0); 899 } else { 900 if (!vm_pages_needed) { 901 vm_pages_needed = 1; 902 wakeup(&vm_pages_needed); 903 } 904 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 905 "vmwait", 0); 906 } 907 splx(s); 908} 909 910/* 911 * vm_waitpfault: (also see VM_WAITPFAULT macro) 912 * 913 * Block until free pages are available for allocation 914 * - Called only in vm_fault so that processes page faulting 915 * can be easily tracked. 916 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 917 * processes will be able to grab memory first. Do not change 918 * this balance without careful testing first. 919 */ 920void 921vm_waitpfault(void) 922{ 923 int s; 924 925 s = splvm(); 926 vm_page_lock_queues(); 927 if (!vm_pages_needed) { 928 vm_pages_needed = 1; 929 wakeup(&vm_pages_needed); 930 } 931 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 932 "pfault", 0); 933 splx(s); 934} 935 936/* 937 * vm_page_activate: 938 * 939 * Put the specified page on the active list (if appropriate). 940 * Ensure that act_count is at least ACT_INIT but do not otherwise 941 * mess with it. 942 * 943 * The page queues must be locked. 944 * This routine may not block. 945 */ 946void 947vm_page_activate(vm_page_t m) 948{ 949 int s; 950 951 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 952 s = splvm(); 953 if (m->queue != PQ_ACTIVE) { 954 if ((m->queue - m->pc) == PQ_CACHE) 955 cnt.v_reactivated++; 956 vm_pageq_remove(m); 957 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 958 if (m->act_count < ACT_INIT) 959 m->act_count = ACT_INIT; 960 vm_pageq_enqueue(PQ_ACTIVE, m); 961 } 962 } else { 963 if (m->act_count < ACT_INIT) 964 m->act_count = ACT_INIT; 965 } 966 splx(s); 967} 968 969/* 970 * vm_page_free_wakeup: 971 * 972 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 973 * routine is called when a page has been added to the cache or free 974 * queues. 975 * 976 * This routine may not block. 977 * This routine must be called at splvm() 978 */ 979static __inline void 980vm_page_free_wakeup(void) 981{ 982 983 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 984 /* 985 * if pageout daemon needs pages, then tell it that there are 986 * some free. 987 */ 988 if (vm_pageout_pages_needed && 989 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 990 wakeup(&vm_pageout_pages_needed); 991 vm_pageout_pages_needed = 0; 992 } 993 /* 994 * wakeup processes that are waiting on memory if we hit a 995 * high water mark. And wakeup scheduler process if we have 996 * lots of memory. this process will swapin processes. 997 */ 998 if (vm_pages_needed && !vm_page_count_min()) { 999 vm_pages_needed = 0; 1000 wakeup(&cnt.v_free_count); 1001 } 1002} 1003 1004/* 1005 * vm_page_free_toq: 1006 * 1007 * Returns the given page to the PQ_FREE list, 1008 * disassociating it with any VM object. 1009 * 1010 * Object and page must be locked prior to entry. 1011 * This routine may not block. 1012 */ 1013 1014void 1015vm_page_free_toq(vm_page_t m) 1016{ 1017 int s; 1018 struct vpgqueues *pq; 1019 vm_object_t object = m->object; 1020 1021 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1022 s = splvm(); 1023 cnt.v_tfree++; 1024 1025 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1026 printf( 1027 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1028 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1029 m->hold_count); 1030 if ((m->queue - m->pc) == PQ_FREE) 1031 panic("vm_page_free: freeing free page"); 1032 else 1033 panic("vm_page_free: freeing busy page"); 1034 } 1035 1036 /* 1037 * unqueue, then remove page. Note that we cannot destroy 1038 * the page here because we do not want to call the pager's 1039 * callback routine until after we've put the page on the 1040 * appropriate free queue. 1041 */ 1042 vm_pageq_remove_nowakeup(m); 1043 vm_page_remove(m); 1044 1045 /* 1046 * If fictitious remove object association and 1047 * return, otherwise delay object association removal. 1048 */ 1049 if ((m->flags & PG_FICTITIOUS) != 0) { 1050 splx(s); 1051 return; 1052 } 1053 1054 m->valid = 0; 1055 vm_page_undirty(m); 1056 1057 if (m->wire_count != 0) { 1058 if (m->wire_count > 1) { 1059 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1060 m->wire_count, (long)m->pindex); 1061 } 1062 panic("vm_page_free: freeing wired page\n"); 1063 } 1064 1065 /* 1066 * If we've exhausted the object's resident pages we want to free 1067 * it up. 1068 */ 1069 if (object && 1070 (object->type == OBJT_VNODE) && 1071 ((object->flags & OBJ_DEAD) == 0) 1072 ) { 1073 struct vnode *vp = (struct vnode *)object->handle; 1074 1075 if (vp) { 1076 VI_LOCK(vp); 1077 if (VSHOULDFREE(vp)) 1078 vfree(vp); 1079 VI_UNLOCK(vp); 1080 } 1081 } 1082 1083 /* 1084 * Clear the UNMANAGED flag when freeing an unmanaged page. 1085 */ 1086 if (m->flags & PG_UNMANAGED) { 1087 m->flags &= ~PG_UNMANAGED; 1088 } 1089 1090 if (m->hold_count != 0) { 1091 m->flags &= ~PG_ZERO; 1092 m->queue = PQ_HOLD; 1093 } else 1094 m->queue = PQ_FREE + m->pc; 1095 pq = &vm_page_queues[m->queue]; 1096 mtx_lock_spin(&vm_page_queue_free_mtx); 1097 pq->lcnt++; 1098 ++(*pq->cnt); 1099 1100 /* 1101 * Put zero'd pages on the end ( where we look for zero'd pages 1102 * first ) and non-zerod pages at the head. 1103 */ 1104 if (m->flags & PG_ZERO) { 1105 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1106 ++vm_page_zero_count; 1107 } else { 1108 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1109 } 1110 mtx_unlock_spin(&vm_page_queue_free_mtx); 1111 vm_page_free_wakeup(); 1112 splx(s); 1113} 1114 1115/* 1116 * vm_page_unmanage: 1117 * 1118 * Prevent PV management from being done on the page. The page is 1119 * removed from the paging queues as if it were wired, and as a 1120 * consequence of no longer being managed the pageout daemon will not 1121 * touch it (since there is no way to locate the pte mappings for the 1122 * page). madvise() calls that mess with the pmap will also no longer 1123 * operate on the page. 1124 * 1125 * Beyond that the page is still reasonably 'normal'. Freeing the page 1126 * will clear the flag. 1127 * 1128 * This routine is used by OBJT_PHYS objects - objects using unswappable 1129 * physical memory as backing store rather then swap-backed memory and 1130 * will eventually be extended to support 4MB unmanaged physical 1131 * mappings. 1132 */ 1133void 1134vm_page_unmanage(vm_page_t m) 1135{ 1136 int s; 1137 1138 s = splvm(); 1139 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1140 if ((m->flags & PG_UNMANAGED) == 0) { 1141 if (m->wire_count == 0) 1142 vm_pageq_remove(m); 1143 } 1144 vm_page_flag_set(m, PG_UNMANAGED); 1145 splx(s); 1146} 1147 1148/* 1149 * vm_page_wire: 1150 * 1151 * Mark this page as wired down by yet 1152 * another map, removing it from paging queues 1153 * as necessary. 1154 * 1155 * The page queues must be locked. 1156 * This routine may not block. 1157 */ 1158void 1159vm_page_wire(vm_page_t m) 1160{ 1161 int s; 1162 1163 /* 1164 * Only bump the wire statistics if the page is not already wired, 1165 * and only unqueue the page if it is on some queue (if it is unmanaged 1166 * it is already off the queues). 1167 */ 1168 s = splvm(); 1169 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1170 if (m->wire_count == 0) { 1171 if ((m->flags & PG_UNMANAGED) == 0) 1172 vm_pageq_remove(m); 1173 atomic_add_int(&cnt.v_wire_count, 1); 1174 } 1175 m->wire_count++; 1176 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1177 splx(s); 1178} 1179 1180/* 1181 * vm_page_unwire: 1182 * 1183 * Release one wiring of this page, potentially 1184 * enabling it to be paged again. 1185 * 1186 * Many pages placed on the inactive queue should actually go 1187 * into the cache, but it is difficult to figure out which. What 1188 * we do instead, if the inactive target is well met, is to put 1189 * clean pages at the head of the inactive queue instead of the tail. 1190 * This will cause them to be moved to the cache more quickly and 1191 * if not actively re-referenced, freed more quickly. If we just 1192 * stick these pages at the end of the inactive queue, heavy filesystem 1193 * meta-data accesses can cause an unnecessary paging load on memory bound 1194 * processes. This optimization causes one-time-use metadata to be 1195 * reused more quickly. 1196 * 1197 * BUT, if we are in a low-memory situation we have no choice but to 1198 * put clean pages on the cache queue. 1199 * 1200 * A number of routines use vm_page_unwire() to guarantee that the page 1201 * will go into either the inactive or active queues, and will NEVER 1202 * be placed in the cache - for example, just after dirtying a page. 1203 * dirty pages in the cache are not allowed. 1204 * 1205 * The page queues must be locked. 1206 * This routine may not block. 1207 */ 1208void 1209vm_page_unwire(vm_page_t m, int activate) 1210{ 1211 int s; 1212 1213 s = splvm(); 1214 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1215 if (m->wire_count > 0) { 1216 m->wire_count--; 1217 if (m->wire_count == 0) { 1218 atomic_subtract_int(&cnt.v_wire_count, 1); 1219 if (m->flags & PG_UNMANAGED) { 1220 ; 1221 } else if (activate) 1222 vm_pageq_enqueue(PQ_ACTIVE, m); 1223 else { 1224 vm_page_flag_clear(m, PG_WINATCFLS); 1225 vm_pageq_enqueue(PQ_INACTIVE, m); 1226 } 1227 } 1228 } else { 1229 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1230 } 1231 splx(s); 1232} 1233 1234 1235/* 1236 * Move the specified page to the inactive queue. If the page has 1237 * any associated swap, the swap is deallocated. 1238 * 1239 * Normally athead is 0 resulting in LRU operation. athead is set 1240 * to 1 if we want this page to be 'as if it were placed in the cache', 1241 * except without unmapping it from the process address space. 1242 * 1243 * This routine may not block. 1244 */ 1245static __inline void 1246_vm_page_deactivate(vm_page_t m, int athead) 1247{ 1248 int s; 1249 1250 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1251 /* 1252 * Ignore if already inactive. 1253 */ 1254 if (m->queue == PQ_INACTIVE) 1255 return; 1256 1257 s = splvm(); 1258 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1259 if ((m->queue - m->pc) == PQ_CACHE) 1260 cnt.v_reactivated++; 1261 vm_page_flag_clear(m, PG_WINATCFLS); 1262 vm_pageq_remove(m); 1263 if (athead) 1264 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1265 else 1266 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1267 m->queue = PQ_INACTIVE; 1268 vm_page_queues[PQ_INACTIVE].lcnt++; 1269 cnt.v_inactive_count++; 1270 } 1271 splx(s); 1272} 1273 1274void 1275vm_page_deactivate(vm_page_t m) 1276{ 1277 _vm_page_deactivate(m, 0); 1278} 1279 1280/* 1281 * vm_page_try_to_cache: 1282 * 1283 * Returns 0 on failure, 1 on success 1284 */ 1285int 1286vm_page_try_to_cache(vm_page_t m) 1287{ 1288 1289 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1290 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1291 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1292 return (0); 1293 } 1294 vm_page_test_dirty(m); 1295 if (m->dirty) 1296 return (0); 1297 vm_page_cache(m); 1298 return (1); 1299} 1300 1301/* 1302 * vm_page_try_to_free() 1303 * 1304 * Attempt to free the page. If we cannot free it, we do nothing. 1305 * 1 is returned on success, 0 on failure. 1306 */ 1307int 1308vm_page_try_to_free(vm_page_t m) 1309{ 1310 1311 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1312 if (m->object != NULL) 1313 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1314 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1315 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1316 return (0); 1317 } 1318 vm_page_test_dirty(m); 1319 if (m->dirty) 1320 return (0); 1321 vm_page_busy(m); 1322 pmap_remove_all(m); 1323 vm_page_free(m); 1324 return (1); 1325} 1326 1327/* 1328 * vm_page_cache 1329 * 1330 * Put the specified page onto the page cache queue (if appropriate). 1331 * 1332 * This routine may not block. 1333 */ 1334void 1335vm_page_cache(vm_page_t m) 1336{ 1337 int s; 1338 1339 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1340 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1341 m->hold_count || m->wire_count) { 1342 printf("vm_page_cache: attempting to cache busy page\n"); 1343 return; 1344 } 1345 if ((m->queue - m->pc) == PQ_CACHE) 1346 return; 1347 1348 /* 1349 * Remove all pmaps and indicate that the page is not 1350 * writeable or mapped. 1351 */ 1352 pmap_remove_all(m); 1353 if (m->dirty != 0) { 1354 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1355 (long)m->pindex); 1356 } 1357 s = splvm(); 1358 vm_pageq_remove_nowakeup(m); 1359 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1360 vm_page_free_wakeup(); 1361 splx(s); 1362} 1363 1364/* 1365 * vm_page_dontneed 1366 * 1367 * Cache, deactivate, or do nothing as appropriate. This routine 1368 * is typically used by madvise() MADV_DONTNEED. 1369 * 1370 * Generally speaking we want to move the page into the cache so 1371 * it gets reused quickly. However, this can result in a silly syndrome 1372 * due to the page recycling too quickly. Small objects will not be 1373 * fully cached. On the otherhand, if we move the page to the inactive 1374 * queue we wind up with a problem whereby very large objects 1375 * unnecessarily blow away our inactive and cache queues. 1376 * 1377 * The solution is to move the pages based on a fixed weighting. We 1378 * either leave them alone, deactivate them, or move them to the cache, 1379 * where moving them to the cache has the highest weighting. 1380 * By forcing some pages into other queues we eventually force the 1381 * system to balance the queues, potentially recovering other unrelated 1382 * space from active. The idea is to not force this to happen too 1383 * often. 1384 */ 1385void 1386vm_page_dontneed(vm_page_t m) 1387{ 1388 static int dnweight; 1389 int dnw; 1390 int head; 1391 1392 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1393 dnw = ++dnweight; 1394 1395 /* 1396 * occassionally leave the page alone 1397 */ 1398 if ((dnw & 0x01F0) == 0 || 1399 m->queue == PQ_INACTIVE || 1400 m->queue - m->pc == PQ_CACHE 1401 ) { 1402 if (m->act_count >= ACT_INIT) 1403 --m->act_count; 1404 return; 1405 } 1406 1407 if (m->dirty == 0) 1408 vm_page_test_dirty(m); 1409 1410 if (m->dirty || (dnw & 0x0070) == 0) { 1411 /* 1412 * Deactivate the page 3 times out of 32. 1413 */ 1414 head = 0; 1415 } else { 1416 /* 1417 * Cache the page 28 times out of every 32. Note that 1418 * the page is deactivated instead of cached, but placed 1419 * at the head of the queue instead of the tail. 1420 */ 1421 head = 1; 1422 } 1423 _vm_page_deactivate(m, head); 1424} 1425 1426/* 1427 * Grab a page, waiting until we are waken up due to the page 1428 * changing state. We keep on waiting, if the page continues 1429 * to be in the object. If the page doesn't exist, allocate it. 1430 * 1431 * This routine may block. 1432 */ 1433vm_page_t 1434vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1435{ 1436 vm_page_t m; 1437 1438 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1439retrylookup: 1440 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1441 vm_page_lock_queues(); 1442 if (m->busy || (m->flags & PG_BUSY)) { 1443 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1444 VM_OBJECT_UNLOCK(object); 1445 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0); 1446 VM_OBJECT_LOCK(object); 1447 if ((allocflags & VM_ALLOC_RETRY) == 0) 1448 return (NULL); 1449 goto retrylookup; 1450 } else { 1451 if (allocflags & VM_ALLOC_WIRED) 1452 vm_page_wire(m); 1453 vm_page_busy(m); 1454 vm_page_unlock_queues(); 1455 return m; 1456 } 1457 } 1458 1459 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1460 if (m == NULL) { 1461 VM_OBJECT_UNLOCK(object); 1462 VM_WAIT; 1463 VM_OBJECT_LOCK(object); 1464 if ((allocflags & VM_ALLOC_RETRY) == 0) 1465 return NULL; 1466 goto retrylookup; 1467 } 1468 1469 return m; 1470} 1471 1472/* 1473 * Mapping function for valid bits or for dirty bits in 1474 * a page. May not block. 1475 * 1476 * Inputs are required to range within a page. 1477 */ 1478__inline int 1479vm_page_bits(int base, int size) 1480{ 1481 int first_bit; 1482 int last_bit; 1483 1484 KASSERT( 1485 base + size <= PAGE_SIZE, 1486 ("vm_page_bits: illegal base/size %d/%d", base, size) 1487 ); 1488 1489 if (size == 0) /* handle degenerate case */ 1490 return (0); 1491 1492 first_bit = base >> DEV_BSHIFT; 1493 last_bit = (base + size - 1) >> DEV_BSHIFT; 1494 1495 return ((2 << last_bit) - (1 << first_bit)); 1496} 1497 1498/* 1499 * vm_page_set_validclean: 1500 * 1501 * Sets portions of a page valid and clean. The arguments are expected 1502 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1503 * of any partial chunks touched by the range. The invalid portion of 1504 * such chunks will be zero'd. 1505 * 1506 * This routine may not block. 1507 * 1508 * (base + size) must be less then or equal to PAGE_SIZE. 1509 */ 1510void 1511vm_page_set_validclean(vm_page_t m, int base, int size) 1512{ 1513 int pagebits; 1514 int frag; 1515 int endoff; 1516 1517 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1518 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1519 if (size == 0) /* handle degenerate case */ 1520 return; 1521 1522 /* 1523 * If the base is not DEV_BSIZE aligned and the valid 1524 * bit is clear, we have to zero out a portion of the 1525 * first block. 1526 */ 1527 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1528 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1529 pmap_zero_page_area(m, frag, base - frag); 1530 1531 /* 1532 * If the ending offset is not DEV_BSIZE aligned and the 1533 * valid bit is clear, we have to zero out a portion of 1534 * the last block. 1535 */ 1536 endoff = base + size; 1537 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1538 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1539 pmap_zero_page_area(m, endoff, 1540 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1541 1542 /* 1543 * Set valid, clear dirty bits. If validating the entire 1544 * page we can safely clear the pmap modify bit. We also 1545 * use this opportunity to clear the PG_NOSYNC flag. If a process 1546 * takes a write fault on a MAP_NOSYNC memory area the flag will 1547 * be set again. 1548 * 1549 * We set valid bits inclusive of any overlap, but we can only 1550 * clear dirty bits for DEV_BSIZE chunks that are fully within 1551 * the range. 1552 */ 1553 pagebits = vm_page_bits(base, size); 1554 m->valid |= pagebits; 1555#if 0 /* NOT YET */ 1556 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1557 frag = DEV_BSIZE - frag; 1558 base += frag; 1559 size -= frag; 1560 if (size < 0) 1561 size = 0; 1562 } 1563 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1564#endif 1565 m->dirty &= ~pagebits; 1566 if (base == 0 && size == PAGE_SIZE) { 1567 pmap_clear_modify(m); 1568 vm_page_flag_clear(m, PG_NOSYNC); 1569 } 1570} 1571 1572void 1573vm_page_clear_dirty(vm_page_t m, int base, int size) 1574{ 1575 1576 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1577 m->dirty &= ~vm_page_bits(base, size); 1578} 1579 1580/* 1581 * vm_page_set_invalid: 1582 * 1583 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1584 * valid and dirty bits for the effected areas are cleared. 1585 * 1586 * May not block. 1587 */ 1588void 1589vm_page_set_invalid(vm_page_t m, int base, int size) 1590{ 1591 int bits; 1592 1593 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1594 bits = vm_page_bits(base, size); 1595 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1596 m->valid &= ~bits; 1597 m->dirty &= ~bits; 1598 m->object->generation++; 1599} 1600 1601/* 1602 * vm_page_zero_invalid() 1603 * 1604 * The kernel assumes that the invalid portions of a page contain 1605 * garbage, but such pages can be mapped into memory by user code. 1606 * When this occurs, we must zero out the non-valid portions of the 1607 * page so user code sees what it expects. 1608 * 1609 * Pages are most often semi-valid when the end of a file is mapped 1610 * into memory and the file's size is not page aligned. 1611 */ 1612void 1613vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1614{ 1615 int b; 1616 int i; 1617 1618 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1619 /* 1620 * Scan the valid bits looking for invalid sections that 1621 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1622 * valid bit may be set ) have already been zerod by 1623 * vm_page_set_validclean(). 1624 */ 1625 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1626 if (i == (PAGE_SIZE / DEV_BSIZE) || 1627 (m->valid & (1 << i)) 1628 ) { 1629 if (i > b) { 1630 pmap_zero_page_area(m, 1631 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1632 } 1633 b = i + 1; 1634 } 1635 } 1636 1637 /* 1638 * setvalid is TRUE when we can safely set the zero'd areas 1639 * as being valid. We can do this if there are no cache consistancy 1640 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1641 */ 1642 if (setvalid) 1643 m->valid = VM_PAGE_BITS_ALL; 1644} 1645 1646/* 1647 * vm_page_is_valid: 1648 * 1649 * Is (partial) page valid? Note that the case where size == 0 1650 * will return FALSE in the degenerate case where the page is 1651 * entirely invalid, and TRUE otherwise. 1652 * 1653 * May not block. 1654 */ 1655int 1656vm_page_is_valid(vm_page_t m, int base, int size) 1657{ 1658 int bits = vm_page_bits(base, size); 1659 1660 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1661 if (m->valid && ((m->valid & bits) == bits)) 1662 return 1; 1663 else 1664 return 0; 1665} 1666 1667/* 1668 * update dirty bits from pmap/mmu. May not block. 1669 */ 1670void 1671vm_page_test_dirty(vm_page_t m) 1672{ 1673 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1674 vm_page_dirty(m); 1675 } 1676} 1677 1678int so_zerocp_fullpage = 0; 1679 1680void 1681vm_page_cowfault(vm_page_t m) 1682{ 1683 vm_page_t mnew; 1684 vm_object_t object; 1685 vm_pindex_t pindex; 1686 1687 object = m->object; 1688 pindex = m->pindex; 1689 vm_page_busy(m); 1690 1691 retry_alloc: 1692 vm_page_remove(m); 1693 /* 1694 * An interrupt allocation is requested because the page 1695 * queues lock is held. 1696 */ 1697 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1698 if (mnew == NULL) { 1699 vm_page_insert(m, object, pindex); 1700 vm_page_unlock_queues(); 1701 VM_OBJECT_UNLOCK(object); 1702 VM_WAIT; 1703 VM_OBJECT_LOCK(object); 1704 vm_page_lock_queues(); 1705 goto retry_alloc; 1706 } 1707 1708 if (m->cow == 0) { 1709 /* 1710 * check to see if we raced with an xmit complete when 1711 * waiting to allocate a page. If so, put things back 1712 * the way they were 1713 */ 1714 vm_page_busy(mnew); 1715 vm_page_free(mnew); 1716 vm_page_insert(m, object, pindex); 1717 } else { /* clear COW & copy page */ 1718 if (!so_zerocp_fullpage) 1719 pmap_copy_page(m, mnew); 1720 mnew->valid = VM_PAGE_BITS_ALL; 1721 vm_page_dirty(mnew); 1722 vm_page_flag_clear(mnew, PG_BUSY); 1723 } 1724} 1725 1726void 1727vm_page_cowclear(vm_page_t m) 1728{ 1729 1730 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1731 if (m->cow) { 1732 m->cow--; 1733 /* 1734 * let vm_fault add back write permission lazily 1735 */ 1736 } 1737 /* 1738 * sf_buf_free() will free the page, so we needn't do it here 1739 */ 1740} 1741 1742void 1743vm_page_cowsetup(vm_page_t m) 1744{ 1745 1746 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1747 m->cow++; 1748 pmap_page_protect(m, VM_PROT_READ); 1749} 1750 1751#include "opt_ddb.h" 1752#ifdef DDB 1753#include <sys/kernel.h> 1754 1755#include <ddb/ddb.h> 1756 1757DB_SHOW_COMMAND(page, vm_page_print_page_info) 1758{ 1759 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1760 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1761 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1762 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1763 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1764 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1765 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1766 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1767 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1768 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1769} 1770 1771DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1772{ 1773 int i; 1774 db_printf("PQ_FREE:"); 1775 for (i = 0; i < PQ_L2_SIZE; i++) { 1776 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1777 } 1778 db_printf("\n"); 1779 1780 db_printf("PQ_CACHE:"); 1781 for (i = 0; i < PQ_L2_SIZE; i++) { 1782 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1783 } 1784 db_printf("\n"); 1785 1786 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1787 vm_page_queues[PQ_ACTIVE].lcnt, 1788 vm_page_queues[PQ_INACTIVE].lcnt); 1789} 1790#endif /* DDB */ 1791