vm_page.c revision 106691
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 106691 2002-11-09 08:27:42Z 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, page_range; 167 vm_offset_t new_end; 168 int i; 169 vm_offset_t pa; 170 int nblocks; 171 vm_offset_t last_pa; 172 173 /* the biggest memory array is the second group of pages */ 174 vm_offset_t end; 175 vm_offset_t biggestone, biggestsize; 176 177 vm_offset_t total; 178 vm_size_t bootpages; 179 180 total = 0; 181 biggestsize = 0; 182 biggestone = 0; 183 nblocks = 0; 184 vaddr = round_page(vaddr); 185 186 for (i = 0; phys_avail[i + 1]; i += 2) { 187 phys_avail[i] = round_page(phys_avail[i]); 188 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 189 } 190 191 for (i = 0; phys_avail[i + 1]; i += 2) { 192 vm_size_t size = phys_avail[i + 1] - phys_avail[i]; 193 194 if (size > biggestsize) { 195 biggestone = i; 196 biggestsize = size; 197 } 198 ++nblocks; 199 total += size; 200 } 201 202 end = phys_avail[biggestone+1]; 203 204 /* 205 * Initialize the locks. 206 */ 207 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF); 208 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 209 MTX_SPIN); 210 211 /* 212 * Initialize the queue headers for the free queue, the active queue 213 * and the inactive queue. 214 */ 215 vm_pageq_init(); 216 217 /* 218 * Allocate memory for use when boot strapping the kernel memory 219 * allocator. 220 */ 221 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 222 new_end = end - bootpages; 223 new_end = trunc_page(new_end); 224 mapped = pmap_map(&vaddr, new_end, end, 225 VM_PROT_READ | VM_PROT_WRITE); 226 bzero((caddr_t) mapped, end - new_end); 227 uma_startup((caddr_t)mapped); 228 229 /* 230 * Compute the number of pages of memory that will be available for 231 * use (taking into account the overhead of a page structure per 232 * page). 233 */ 234 first_page = phys_avail[0] / PAGE_SIZE; 235 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 236 npages = (total - (page_range * sizeof(struct vm_page)) - 237 (end - new_end)) / PAGE_SIZE; 238 end = new_end; 239 240 /* 241 * Initialize the mem entry structures now, and put them in the free 242 * queue. 243 */ 244 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 245 mapped = pmap_map(&vaddr, new_end, end, 246 VM_PROT_READ | VM_PROT_WRITE); 247 vm_page_array = (vm_page_t) mapped; 248 249 /* 250 * Clear all of the page structures 251 */ 252 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 253 vm_page_array_size = page_range; 254 255 /* 256 * Construct the free queue(s) in descending order (by physical 257 * address) so that the first 16MB of physical memory is allocated 258 * last rather than first. On large-memory machines, this avoids 259 * the exhaustion of low physical memory before isa_dmainit has run. 260 */ 261 cnt.v_page_count = 0; 262 cnt.v_free_count = 0; 263 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 264 pa = phys_avail[i]; 265 if (i == biggestone) 266 last_pa = new_end; 267 else 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 GIANT_REQUIRED; 281 m->flags |= bits; 282} 283 284void 285vm_page_flag_clear(vm_page_t m, unsigned short bits) 286{ 287 GIANT_REQUIRED; 288 m->flags &= ~bits; 289} 290 291void 292vm_page_busy(vm_page_t m) 293{ 294 KASSERT((m->flags & PG_BUSY) == 0, 295 ("vm_page_busy: page already busy!!!")); 296 vm_page_flag_set(m, PG_BUSY); 297} 298 299/* 300 * vm_page_flash: 301 * 302 * wakeup anyone waiting for the page. 303 */ 304void 305vm_page_flash(vm_page_t m) 306{ 307 if (m->flags & PG_WANTED) { 308 vm_page_flag_clear(m, PG_WANTED); 309 wakeup(m); 310 } 311} 312 313/* 314 * vm_page_wakeup: 315 * 316 * clear the PG_BUSY flag and wakeup anyone waiting for the 317 * page. 318 * 319 */ 320void 321vm_page_wakeup(vm_page_t m) 322{ 323 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 324 vm_page_flag_clear(m, PG_BUSY); 325 vm_page_flash(m); 326} 327 328/* 329 * 330 * 331 */ 332void 333vm_page_io_start(vm_page_t m) 334{ 335 336 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 337 m->busy++; 338} 339 340void 341vm_page_io_finish(vm_page_t m) 342{ 343 344 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 345 m->busy--; 346 if (m->busy == 0) 347 vm_page_flash(m); 348} 349 350/* 351 * Keep page from being freed by the page daemon 352 * much of the same effect as wiring, except much lower 353 * overhead and should be used only for *very* temporary 354 * holding ("wiring"). 355 */ 356void 357vm_page_hold(vm_page_t mem) 358{ 359 GIANT_REQUIRED; 360 mem->hold_count++; 361} 362 363void 364vm_page_unhold(vm_page_t mem) 365{ 366 GIANT_REQUIRED; 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_protect: 375 * 376 * Reduce the protection of a page. This routine never raises the 377 * protection and therefore can be safely called if the page is already 378 * at VM_PROT_NONE (it will be a NOP effectively ). 379 */ 380void 381vm_page_protect(vm_page_t mem, int prot) 382{ 383 if (prot == VM_PROT_NONE) { 384 if (pmap_page_is_mapped(mem) || (mem->flags & PG_WRITEABLE)) { 385 pmap_page_protect(mem, VM_PROT_NONE); 386 vm_page_flag_clear(mem, PG_WRITEABLE); 387 } 388 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 389 pmap_page_protect(mem, VM_PROT_READ); 390 vm_page_flag_clear(mem, PG_WRITEABLE); 391 } 392} 393 394/* 395 * vm_page_copy: 396 * 397 * Copy one page to another 398 */ 399void 400vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 401{ 402 pmap_copy_page(src_m, dest_m); 403 dest_m->valid = VM_PAGE_BITS_ALL; 404} 405 406/* 407 * vm_page_free: 408 * 409 * Free a page 410 * 411 * The clearing of PG_ZERO is a temporary safety until the code can be 412 * reviewed to determine that PG_ZERO is being properly cleared on 413 * write faults or maps. PG_ZERO was previously cleared in 414 * vm_page_alloc(). 415 */ 416void 417vm_page_free(vm_page_t m) 418{ 419 vm_page_flag_clear(m, PG_ZERO); 420 vm_page_free_toq(m); 421 vm_page_zero_idle_wakeup(); 422} 423 424/* 425 * vm_page_free_zero: 426 * 427 * Free a page to the zerod-pages queue 428 */ 429void 430vm_page_free_zero(vm_page_t m) 431{ 432 vm_page_flag_set(m, PG_ZERO); 433 vm_page_free_toq(m); 434} 435 436/* 437 * vm_page_sleep_busy: 438 * 439 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 440 * m->busy is zero. Returns TRUE if it had to sleep ( including if 441 * it almost had to sleep and made temporary spl*() mods), FALSE 442 * otherwise. 443 * 444 * This routine assumes that interrupts can only remove the busy 445 * status from a page, not set the busy status or change it from 446 * PG_BUSY to m->busy or vise versa (which would create a timing 447 * window). 448 */ 449int 450vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 451{ 452 GIANT_REQUIRED; 453 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 454 int s = splvm(); 455 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 456 /* 457 * Page is busy. Wait and retry. 458 */ 459 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 460 tsleep(m, PVM, msg, 0); 461 } 462 splx(s); 463 return (TRUE); 464 /* not reached */ 465 } 466 return (FALSE); 467} 468 469/* 470 * vm_page_sleep_if_busy: 471 * 472 * Sleep and release the page queues lock if PG_BUSY is set or, 473 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 474 * thread slept and the page queues lock was released. 475 * Otherwise, retains the page queues lock and returns FALSE. 476 */ 477int 478vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 479{ 480 481 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 482 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 483 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 484 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 485 return (TRUE); 486 } 487 return (FALSE); 488} 489 490/* 491 * vm_page_dirty: 492 * 493 * make page all dirty 494 */ 495void 496vm_page_dirty(vm_page_t m) 497{ 498 KASSERT(m->queue - m->pc != PQ_CACHE, 499 ("vm_page_dirty: page in cache!")); 500 m->dirty = VM_PAGE_BITS_ALL; 501} 502 503/* 504 * vm_page_splay: 505 * 506 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 507 * the vm_page containing the given pindex. If, however, that 508 * pindex is not found in the vm_object, returns a vm_page that is 509 * adjacent to the pindex, coming before or after it. 510 */ 511vm_page_t 512vm_page_splay(vm_pindex_t pindex, vm_page_t root) 513{ 514 struct vm_page dummy; 515 vm_page_t lefttreemax, righttreemin, y; 516 517 if (root == NULL) 518 return (root); 519 lefttreemax = righttreemin = &dummy; 520 for (;; root = y) { 521 if (pindex < root->pindex) { 522 if ((y = root->left) == NULL) 523 break; 524 if (pindex < y->pindex) { 525 /* Rotate right. */ 526 root->left = y->right; 527 y->right = root; 528 root = y; 529 if ((y = root->left) == NULL) 530 break; 531 } 532 /* Link into the new root's right tree. */ 533 righttreemin->left = root; 534 righttreemin = root; 535 } else if (pindex > root->pindex) { 536 if ((y = root->right) == NULL) 537 break; 538 if (pindex > y->pindex) { 539 /* Rotate left. */ 540 root->right = y->left; 541 y->left = root; 542 root = y; 543 if ((y = root->right) == NULL) 544 break; 545 } 546 /* Link into the new root's left tree. */ 547 lefttreemax->right = root; 548 lefttreemax = root; 549 } else 550 break; 551 } 552 /* Assemble the new root. */ 553 lefttreemax->right = root->left; 554 righttreemin->left = root->right; 555 root->left = dummy.right; 556 root->right = dummy.left; 557 return (root); 558} 559 560/* 561 * vm_page_insert: [ internal use only ] 562 * 563 * Inserts the given mem entry into the object and object list. 564 * 565 * The pagetables are not updated but will presumably fault the page 566 * in if necessary, or if a kernel page the caller will at some point 567 * enter the page into the kernel's pmap. We are not allowed to block 568 * here so we *can't* do this anyway. 569 * 570 * The object and page must be locked, and must be splhigh. 571 * This routine may not block. 572 */ 573void 574vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 575{ 576 vm_page_t root; 577 578 GIANT_REQUIRED; 579 580 if (m->object != NULL) 581 panic("vm_page_insert: already inserted"); 582 583 /* 584 * Record the object/offset pair in this page 585 */ 586 m->object = object; 587 m->pindex = pindex; 588 589 /* 590 * Now link into the object's ordered list of backed pages. 591 */ 592 root = object->root; 593 if (root == NULL) { 594 m->left = NULL; 595 m->right = NULL; 596 TAILQ_INSERT_TAIL(&object->memq, m, listq); 597 } else { 598 root = vm_page_splay(pindex, root); 599 if (pindex < root->pindex) { 600 m->left = root->left; 601 m->right = root; 602 root->left = NULL; 603 TAILQ_INSERT_BEFORE(root, m, listq); 604 } else { 605 m->right = root->right; 606 m->left = root; 607 root->right = NULL; 608 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 609 } 610 } 611 object->root = m; 612 object->generation++; 613 614 /* 615 * show that the object has one more resident page. 616 */ 617 object->resident_page_count++; 618 619 /* 620 * Since we are inserting a new and possibly dirty page, 621 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 622 */ 623 if (m->flags & PG_WRITEABLE) 624 vm_object_set_writeable_dirty(object); 625} 626 627/* 628 * vm_page_remove: 629 * NOTE: used by device pager as well -wfj 630 * 631 * Removes the given mem entry from the object/offset-page 632 * table and the object page list, but do not invalidate/terminate 633 * the backing store. 634 * 635 * The object and page must be locked, and at splhigh. 636 * The underlying pmap entry (if any) is NOT removed here. 637 * This routine may not block. 638 */ 639void 640vm_page_remove(vm_page_t m) 641{ 642 vm_object_t object; 643 vm_page_t root; 644 645 GIANT_REQUIRED; 646 647 if (m->object == NULL) 648 return; 649 650 if ((m->flags & PG_BUSY) == 0) { 651 panic("vm_page_remove: page not busy"); 652 } 653 654 /* 655 * Basically destroy the page. 656 */ 657 vm_page_wakeup(m); 658 659 object = m->object; 660 661 /* 662 * Now remove from the object's list of backed pages. 663 */ 664 if (m != object->root) 665 vm_page_splay(m->pindex, object->root); 666 if (m->left == NULL) 667 root = m->right; 668 else { 669 root = vm_page_splay(m->pindex, m->left); 670 root->right = m->right; 671 } 672 object->root = root; 673 TAILQ_REMOVE(&object->memq, m, listq); 674 675 /* 676 * And show that the object has one fewer resident page. 677 */ 678 object->resident_page_count--; 679 object->generation++; 680 681 m->object = NULL; 682} 683 684/* 685 * vm_page_lookup: 686 * 687 * Returns the page associated with the object/offset 688 * pair specified; if none is found, NULL is returned. 689 * 690 * The object must be locked. 691 * This routine may not block. 692 * This is a critical path routine 693 */ 694vm_page_t 695vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 696{ 697 vm_page_t m; 698 699 GIANT_REQUIRED; 700 701 m = vm_page_splay(pindex, object->root); 702 if ((object->root = m) != NULL && m->pindex != pindex) 703 m = NULL; 704 return (m); 705} 706 707/* 708 * vm_page_rename: 709 * 710 * Move the given memory entry from its 711 * current object to the specified target object/offset. 712 * 713 * The object must be locked. 714 * This routine may not block. 715 * 716 * Note: this routine will raise itself to splvm(), the caller need not. 717 * 718 * Note: swap associated with the page must be invalidated by the move. We 719 * have to do this for several reasons: (1) we aren't freeing the 720 * page, (2) we are dirtying the page, (3) the VM system is probably 721 * moving the page from object A to B, and will then later move 722 * the backing store from A to B and we can't have a conflict. 723 * 724 * Note: we *always* dirty the page. It is necessary both for the 725 * fact that we moved it, and because we may be invalidating 726 * swap. If the page is on the cache, we have to deactivate it 727 * or vm_page_dirty() will panic. Dirty pages are not allowed 728 * on the cache. 729 */ 730void 731vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 732{ 733 int s; 734 735 s = splvm(); 736 vm_page_lock_queues(); 737 vm_page_remove(m); 738 vm_page_insert(m, new_object, new_pindex); 739 if (m->queue - m->pc == PQ_CACHE) 740 vm_page_deactivate(m); 741 vm_page_dirty(m); 742 vm_page_unlock_queues(); 743 splx(s); 744} 745 746/* 747 * vm_page_select_cache: 748 * 749 * Find a page on the cache queue with color optimization. As pages 750 * might be found, but not applicable, they are deactivated. This 751 * keeps us from using potentially busy cached pages. 752 * 753 * This routine must be called at splvm(). 754 * This routine may not block. 755 */ 756static vm_page_t 757vm_page_select_cache(vm_pindex_t color) 758{ 759 vm_page_t m; 760 761 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 762 while (TRUE) { 763 m = vm_pageq_find(PQ_CACHE, color & PQ_L2_MASK, FALSE); 764 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 765 m->hold_count || m->wire_count)) { 766 vm_page_deactivate(m); 767 continue; 768 } 769 return m; 770 } 771} 772 773/* 774 * vm_page_select_free: 775 * 776 * Find a free or zero page, with specified preference. 777 * 778 * This routine must be called at splvm(). 779 * This routine may not block. 780 */ 781static __inline vm_page_t 782vm_page_select_free(vm_pindex_t color, boolean_t prefer_zero) 783{ 784 vm_page_t m; 785 786 m = vm_pageq_find(PQ_FREE, color & PQ_L2_MASK, prefer_zero); 787 return (m); 788} 789 790/* 791 * vm_page_alloc: 792 * 793 * Allocate and return a memory cell associated 794 * with this VM object/offset pair. 795 * 796 * page_req classes: 797 * VM_ALLOC_NORMAL normal process request 798 * VM_ALLOC_SYSTEM system *really* needs a page 799 * VM_ALLOC_INTERRUPT interrupt time request 800 * VM_ALLOC_ZERO zero page 801 * 802 * This routine may not block. 803 * 804 * Additional special handling is required when called from an 805 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 806 * the page cache in this case. 807 */ 808vm_page_t 809vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 810{ 811 vm_page_t m = NULL; 812 vm_pindex_t color; 813 int page_req, s; 814 815 GIANT_REQUIRED; 816 817#ifdef INVARIANTS 818 if ((req & VM_ALLOC_NOOBJ) == 0) { 819 KASSERT(object != NULL, 820 ("vm_page_alloc: NULL object.")); 821 KASSERT(!vm_page_lookup(object, pindex), 822 ("vm_page_alloc: page already allocated")); 823 } 824#endif 825 826 page_req = req & VM_ALLOC_CLASS_MASK; 827 828 if ((req & VM_ALLOC_NOOBJ) == 0) 829 color = pindex + object->pg_color; 830 else 831 color = pindex; 832 833 /* 834 * The pager is allowed to eat deeper into the free page list. 835 */ 836 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 837 page_req = VM_ALLOC_SYSTEM; 838 }; 839 840 s = splvm(); 841loop: 842 mtx_lock_spin(&vm_page_queue_free_mtx); 843 if (cnt.v_free_count > cnt.v_free_reserved) { 844 /* 845 * Allocate from the free queue if there are plenty of pages 846 * in it. 847 */ 848 m = vm_page_select_free(color, (req & VM_ALLOC_ZERO) != 0); 849 850 } else if ( 851 (page_req == VM_ALLOC_SYSTEM && 852 cnt.v_cache_count == 0 && 853 cnt.v_free_count > cnt.v_interrupt_free_min) || 854 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 855 ) { 856 /* 857 * Interrupt or system, dig deeper into the free list. 858 */ 859 m = vm_page_select_free(color, FALSE); 860 } else if (page_req != VM_ALLOC_INTERRUPT) { 861 mtx_unlock_spin(&vm_page_queue_free_mtx); 862 /* 863 * Allocatable from cache (non-interrupt only). On success, 864 * we must free the page and try again, thus ensuring that 865 * cnt.v_*_free_min counters are replenished. 866 */ 867 vm_page_lock_queues(); 868 if ((m = vm_page_select_cache(color)) == NULL) { 869 vm_page_unlock_queues(); 870 splx(s); 871#if defined(DIAGNOSTIC) 872 if (cnt.v_cache_count > 0) 873 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 874#endif 875 vm_pageout_deficit++; 876 pagedaemon_wakeup(); 877 return (NULL); 878 } 879 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 880 vm_page_busy(m); 881 vm_page_protect(m, VM_PROT_NONE); 882 vm_page_free(m); 883 vm_page_unlock_queues(); 884 goto loop; 885 } else { 886 /* 887 * Not allocatable from cache from interrupt, give up. 888 */ 889 mtx_unlock_spin(&vm_page_queue_free_mtx); 890 splx(s); 891 vm_pageout_deficit++; 892 pagedaemon_wakeup(); 893 return (NULL); 894 } 895 896 /* 897 * At this point we had better have found a good page. 898 */ 899 900 KASSERT( 901 m != NULL, 902 ("vm_page_alloc(): missing page on free queue\n") 903 ); 904 905 /* 906 * Remove from free queue 907 */ 908 909 vm_pageq_remove_nowakeup(m); 910 911 /* 912 * Initialize structure. Only the PG_ZERO flag is inherited. 913 */ 914 if (m->flags & PG_ZERO) { 915 vm_page_zero_count--; 916 m->flags = PG_ZERO | PG_BUSY; 917 } else { 918 m->flags = PG_BUSY; 919 } 920 if (req & VM_ALLOC_WIRED) { 921 cnt.v_wire_count++; 922 m->wire_count = 1; 923 } else 924 m->wire_count = 0; 925 m->hold_count = 0; 926 m->act_count = 0; 927 m->busy = 0; 928 m->valid = 0; 929 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 930 mtx_unlock_spin(&vm_page_queue_free_mtx); 931 932 /* 933 * vm_page_insert() is safe prior to the splx(). Note also that 934 * inserting a page here does not insert it into the pmap (which 935 * could cause us to block allocating memory). We cannot block 936 * anywhere. 937 */ 938 if ((req & VM_ALLOC_NOOBJ) == 0) 939 vm_page_insert(m, object, pindex); 940 941 /* 942 * Don't wakeup too often - wakeup the pageout daemon when 943 * we would be nearly out of memory. 944 */ 945 if (vm_paging_needed()) 946 pagedaemon_wakeup(); 947 948 splx(s); 949 return (m); 950} 951 952/* 953 * vm_wait: (also see VM_WAIT macro) 954 * 955 * Block until free pages are available for allocation 956 * - Called in various places before memory allocations. 957 */ 958void 959vm_wait(void) 960{ 961 int s; 962 963 s = splvm(); 964 if (curproc == pageproc) { 965 vm_pageout_pages_needed = 1; 966 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 967 } else { 968 if (!vm_pages_needed) { 969 vm_pages_needed = 1; 970 wakeup(&vm_pages_needed); 971 } 972 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 973 } 974 splx(s); 975} 976 977/* 978 * vm_waitpfault: (also see VM_WAITPFAULT macro) 979 * 980 * Block until free pages are available for allocation 981 * - Called only in vm_fault so that processes page faulting 982 * can be easily tracked. 983 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 984 * processes will be able to grab memory first. Do not change 985 * this balance without careful testing first. 986 */ 987void 988vm_waitpfault(void) 989{ 990 int s; 991 992 s = splvm(); 993 if (!vm_pages_needed) { 994 vm_pages_needed = 1; 995 wakeup(&vm_pages_needed); 996 } 997 tsleep(&cnt.v_free_count, PUSER, "pfault", 0); 998 splx(s); 999} 1000 1001/* 1002 * vm_page_activate: 1003 * 1004 * Put the specified page on the active list (if appropriate). 1005 * Ensure that act_count is at least ACT_INIT but do not otherwise 1006 * mess with it. 1007 * 1008 * The page queues must be locked. 1009 * This routine may not block. 1010 */ 1011void 1012vm_page_activate(vm_page_t m) 1013{ 1014 int s; 1015 1016 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1017 s = splvm(); 1018 if (m->queue != PQ_ACTIVE) { 1019 if ((m->queue - m->pc) == PQ_CACHE) 1020 cnt.v_reactivated++; 1021 vm_pageq_remove(m); 1022 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1023 if (m->act_count < ACT_INIT) 1024 m->act_count = ACT_INIT; 1025 vm_pageq_enqueue(PQ_ACTIVE, m); 1026 } 1027 } else { 1028 if (m->act_count < ACT_INIT) 1029 m->act_count = ACT_INIT; 1030 } 1031 splx(s); 1032} 1033 1034/* 1035 * vm_page_free_wakeup: 1036 * 1037 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1038 * routine is called when a page has been added to the cache or free 1039 * queues. 1040 * 1041 * This routine may not block. 1042 * This routine must be called at splvm() 1043 */ 1044static __inline void 1045vm_page_free_wakeup(void) 1046{ 1047 /* 1048 * if pageout daemon needs pages, then tell it that there are 1049 * some free. 1050 */ 1051 if (vm_pageout_pages_needed && 1052 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1053 wakeup(&vm_pageout_pages_needed); 1054 vm_pageout_pages_needed = 0; 1055 } 1056 /* 1057 * wakeup processes that are waiting on memory if we hit a 1058 * high water mark. And wakeup scheduler process if we have 1059 * lots of memory. this process will swapin processes. 1060 */ 1061 if (vm_pages_needed && !vm_page_count_min()) { 1062 vm_pages_needed = 0; 1063 wakeup(&cnt.v_free_count); 1064 } 1065} 1066 1067/* 1068 * vm_page_free_toq: 1069 * 1070 * Returns the given page to the PQ_FREE list, 1071 * disassociating it with any VM object. 1072 * 1073 * Object and page must be locked prior to entry. 1074 * This routine may not block. 1075 */ 1076 1077void 1078vm_page_free_toq(vm_page_t m) 1079{ 1080 int s; 1081 struct vpgqueues *pq; 1082 vm_object_t object = m->object; 1083 1084 GIANT_REQUIRED; 1085 s = splvm(); 1086 cnt.v_tfree++; 1087 1088 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1089 printf( 1090 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1091 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1092 m->hold_count); 1093 if ((m->queue - m->pc) == PQ_FREE) 1094 panic("vm_page_free: freeing free page"); 1095 else 1096 panic("vm_page_free: freeing busy page"); 1097 } 1098 1099 /* 1100 * unqueue, then remove page. Note that we cannot destroy 1101 * the page here because we do not want to call the pager's 1102 * callback routine until after we've put the page on the 1103 * appropriate free queue. 1104 */ 1105 vm_pageq_remove_nowakeup(m); 1106 vm_page_remove(m); 1107 1108 /* 1109 * If fictitious remove object association and 1110 * return, otherwise delay object association removal. 1111 */ 1112 if ((m->flags & PG_FICTITIOUS) != 0) { 1113 splx(s); 1114 return; 1115 } 1116 1117 m->valid = 0; 1118 vm_page_undirty(m); 1119 1120 if (m->wire_count != 0) { 1121 if (m->wire_count > 1) { 1122 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1123 m->wire_count, (long)m->pindex); 1124 } 1125 panic("vm_page_free: freeing wired page\n"); 1126 } 1127 1128 /* 1129 * If we've exhausted the object's resident pages we want to free 1130 * it up. 1131 */ 1132 if (object && 1133 (object->type == OBJT_VNODE) && 1134 ((object->flags & OBJ_DEAD) == 0) 1135 ) { 1136 struct vnode *vp = (struct vnode *)object->handle; 1137 1138 if (vp) { 1139 VI_LOCK(vp); 1140 if (VSHOULDFREE(vp)) 1141 vfree(vp); 1142 VI_UNLOCK(vp); 1143 } 1144 } 1145 1146 /* 1147 * Clear the UNMANAGED flag when freeing an unmanaged page. 1148 */ 1149 if (m->flags & PG_UNMANAGED) { 1150 m->flags &= ~PG_UNMANAGED; 1151 } else { 1152#ifdef __alpha__ 1153 pmap_page_is_free(m); 1154#endif 1155 } 1156 1157 if (m->hold_count != 0) { 1158 m->flags &= ~PG_ZERO; 1159 m->queue = PQ_HOLD; 1160 } else 1161 m->queue = PQ_FREE + m->pc; 1162 pq = &vm_page_queues[m->queue]; 1163 mtx_lock_spin(&vm_page_queue_free_mtx); 1164 pq->lcnt++; 1165 ++(*pq->cnt); 1166 1167 /* 1168 * Put zero'd pages on the end ( where we look for zero'd pages 1169 * first ) and non-zerod pages at the head. 1170 */ 1171 if (m->flags & PG_ZERO) { 1172 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1173 ++vm_page_zero_count; 1174 } else { 1175 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1176 } 1177 mtx_unlock_spin(&vm_page_queue_free_mtx); 1178 vm_page_free_wakeup(); 1179 splx(s); 1180} 1181 1182/* 1183 * vm_page_unmanage: 1184 * 1185 * Prevent PV management from being done on the page. The page is 1186 * removed from the paging queues as if it were wired, and as a 1187 * consequence of no longer being managed the pageout daemon will not 1188 * touch it (since there is no way to locate the pte mappings for the 1189 * page). madvise() calls that mess with the pmap will also no longer 1190 * operate on the page. 1191 * 1192 * Beyond that the page is still reasonably 'normal'. Freeing the page 1193 * will clear the flag. 1194 * 1195 * This routine is used by OBJT_PHYS objects - objects using unswappable 1196 * physical memory as backing store rather then swap-backed memory and 1197 * will eventually be extended to support 4MB unmanaged physical 1198 * mappings. 1199 */ 1200void 1201vm_page_unmanage(vm_page_t m) 1202{ 1203 int s; 1204 1205 s = splvm(); 1206 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1207 if ((m->flags & PG_UNMANAGED) == 0) { 1208 if (m->wire_count == 0) 1209 vm_pageq_remove(m); 1210 } 1211 vm_page_flag_set(m, PG_UNMANAGED); 1212 splx(s); 1213} 1214 1215/* 1216 * vm_page_wire: 1217 * 1218 * Mark this page as wired down by yet 1219 * another map, removing it from paging queues 1220 * as necessary. 1221 * 1222 * The page queues must be locked. 1223 * This routine may not block. 1224 */ 1225void 1226vm_page_wire(vm_page_t m) 1227{ 1228 int s; 1229 1230 /* 1231 * Only bump the wire statistics if the page is not already wired, 1232 * and only unqueue the page if it is on some queue (if it is unmanaged 1233 * it is already off the queues). 1234 */ 1235 s = splvm(); 1236 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1237 if (m->wire_count == 0) { 1238 if ((m->flags & PG_UNMANAGED) == 0) 1239 vm_pageq_remove(m); 1240 cnt.v_wire_count++; 1241 } 1242 m->wire_count++; 1243 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1244 splx(s); 1245} 1246 1247/* 1248 * vm_page_unwire: 1249 * 1250 * Release one wiring of this page, potentially 1251 * enabling it to be paged again. 1252 * 1253 * Many pages placed on the inactive queue should actually go 1254 * into the cache, but it is difficult to figure out which. What 1255 * we do instead, if the inactive target is well met, is to put 1256 * clean pages at the head of the inactive queue instead of the tail. 1257 * This will cause them to be moved to the cache more quickly and 1258 * if not actively re-referenced, freed more quickly. If we just 1259 * stick these pages at the end of the inactive queue, heavy filesystem 1260 * meta-data accesses can cause an unnecessary paging load on memory bound 1261 * processes. This optimization causes one-time-use metadata to be 1262 * reused more quickly. 1263 * 1264 * BUT, if we are in a low-memory situation we have no choice but to 1265 * put clean pages on the cache queue. 1266 * 1267 * A number of routines use vm_page_unwire() to guarantee that the page 1268 * will go into either the inactive or active queues, and will NEVER 1269 * be placed in the cache - for example, just after dirtying a page. 1270 * dirty pages in the cache are not allowed. 1271 * 1272 * The page queues must be locked. 1273 * This routine may not block. 1274 */ 1275void 1276vm_page_unwire(vm_page_t m, int activate) 1277{ 1278 int s; 1279 1280 s = splvm(); 1281 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1282 if (m->wire_count > 0) { 1283 m->wire_count--; 1284 if (m->wire_count == 0) { 1285 cnt.v_wire_count--; 1286 if (m->flags & PG_UNMANAGED) { 1287 ; 1288 } else if (activate) 1289 vm_pageq_enqueue(PQ_ACTIVE, m); 1290 else { 1291 vm_page_flag_clear(m, PG_WINATCFLS); 1292 vm_pageq_enqueue(PQ_INACTIVE, m); 1293 } 1294 } 1295 } else { 1296 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1297 } 1298 splx(s); 1299} 1300 1301 1302/* 1303 * Move the specified page to the inactive queue. If the page has 1304 * any associated swap, the swap is deallocated. 1305 * 1306 * Normally athead is 0 resulting in LRU operation. athead is set 1307 * to 1 if we want this page to be 'as if it were placed in the cache', 1308 * except without unmapping it from the process address space. 1309 * 1310 * This routine may not block. 1311 */ 1312static __inline void 1313_vm_page_deactivate(vm_page_t m, int athead) 1314{ 1315 int s; 1316 1317 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1318 /* 1319 * Ignore if already inactive. 1320 */ 1321 if (m->queue == PQ_INACTIVE) 1322 return; 1323 1324 s = splvm(); 1325 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1326 if ((m->queue - m->pc) == PQ_CACHE) 1327 cnt.v_reactivated++; 1328 vm_page_flag_clear(m, PG_WINATCFLS); 1329 vm_pageq_remove(m); 1330 if (athead) 1331 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1332 else 1333 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1334 m->queue = PQ_INACTIVE; 1335 vm_page_queues[PQ_INACTIVE].lcnt++; 1336 cnt.v_inactive_count++; 1337 } 1338 splx(s); 1339} 1340 1341void 1342vm_page_deactivate(vm_page_t m) 1343{ 1344 _vm_page_deactivate(m, 0); 1345} 1346 1347/* 1348 * vm_page_try_to_cache: 1349 * 1350 * Returns 0 on failure, 1 on success 1351 */ 1352int 1353vm_page_try_to_cache(vm_page_t m) 1354{ 1355 1356 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1357 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1358 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1359 return (0); 1360 } 1361 vm_page_test_dirty(m); 1362 if (m->dirty) 1363 return (0); 1364 vm_page_cache(m); 1365 return (1); 1366} 1367 1368/* 1369 * vm_page_try_to_free() 1370 * 1371 * Attempt to free the page. If we cannot free it, we do nothing. 1372 * 1 is returned on success, 0 on failure. 1373 */ 1374int 1375vm_page_try_to_free(vm_page_t m) 1376{ 1377 1378 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1379 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1380 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1381 return (0); 1382 } 1383 vm_page_test_dirty(m); 1384 if (m->dirty) 1385 return (0); 1386 vm_page_busy(m); 1387 vm_page_protect(m, VM_PROT_NONE); 1388 vm_page_free(m); 1389 return (1); 1390} 1391 1392/* 1393 * vm_page_cache 1394 * 1395 * Put the specified page onto the page cache queue (if appropriate). 1396 * 1397 * This routine may not block. 1398 */ 1399void 1400vm_page_cache(vm_page_t m) 1401{ 1402 int s; 1403 1404 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1405 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1406 printf("vm_page_cache: attempting to cache busy page\n"); 1407 return; 1408 } 1409 if ((m->queue - m->pc) == PQ_CACHE) 1410 return; 1411 1412 /* 1413 * Remove all pmaps and indicate that the page is not 1414 * writeable or mapped. 1415 */ 1416 vm_page_protect(m, VM_PROT_NONE); 1417 if (m->dirty != 0) { 1418 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1419 (long)m->pindex); 1420 } 1421 s = splvm(); 1422 vm_pageq_remove_nowakeup(m); 1423 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1424 vm_page_free_wakeup(); 1425 splx(s); 1426} 1427 1428/* 1429 * vm_page_dontneed 1430 * 1431 * Cache, deactivate, or do nothing as appropriate. This routine 1432 * is typically used by madvise() MADV_DONTNEED. 1433 * 1434 * Generally speaking we want to move the page into the cache so 1435 * it gets reused quickly. However, this can result in a silly syndrome 1436 * due to the page recycling too quickly. Small objects will not be 1437 * fully cached. On the otherhand, if we move the page to the inactive 1438 * queue we wind up with a problem whereby very large objects 1439 * unnecessarily blow away our inactive and cache queues. 1440 * 1441 * The solution is to move the pages based on a fixed weighting. We 1442 * either leave them alone, deactivate them, or move them to the cache, 1443 * where moving them to the cache has the highest weighting. 1444 * By forcing some pages into other queues we eventually force the 1445 * system to balance the queues, potentially recovering other unrelated 1446 * space from active. The idea is to not force this to happen too 1447 * often. 1448 */ 1449void 1450vm_page_dontneed(vm_page_t m) 1451{ 1452 static int dnweight; 1453 int dnw; 1454 int head; 1455 1456 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1457 dnw = ++dnweight; 1458 1459 /* 1460 * occassionally leave the page alone 1461 */ 1462 if ((dnw & 0x01F0) == 0 || 1463 m->queue == PQ_INACTIVE || 1464 m->queue - m->pc == PQ_CACHE 1465 ) { 1466 if (m->act_count >= ACT_INIT) 1467 --m->act_count; 1468 return; 1469 } 1470 1471 if (m->dirty == 0) 1472 vm_page_test_dirty(m); 1473 1474 if (m->dirty || (dnw & 0x0070) == 0) { 1475 /* 1476 * Deactivate the page 3 times out of 32. 1477 */ 1478 head = 0; 1479 } else { 1480 /* 1481 * Cache the page 28 times out of every 32. Note that 1482 * the page is deactivated instead of cached, but placed 1483 * at the head of the queue instead of the tail. 1484 */ 1485 head = 1; 1486 } 1487 _vm_page_deactivate(m, head); 1488} 1489 1490/* 1491 * Grab a page, waiting until we are waken up due to the page 1492 * changing state. We keep on waiting, if the page continues 1493 * to be in the object. If the page doesn't exist, allocate it. 1494 * 1495 * This routine may block. 1496 */ 1497vm_page_t 1498vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1499{ 1500 vm_page_t m; 1501 int s, generation; 1502 1503 GIANT_REQUIRED; 1504retrylookup: 1505 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1506 vm_page_lock_queues(); 1507 if (m->busy || (m->flags & PG_BUSY)) { 1508 generation = object->generation; 1509 1510 s = splvm(); 1511 while ((object->generation == generation) && 1512 (m->busy || (m->flags & PG_BUSY))) { 1513 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1514 msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0); 1515 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1516 vm_page_unlock_queues(); 1517 splx(s); 1518 return NULL; 1519 } 1520 } 1521 vm_page_unlock_queues(); 1522 splx(s); 1523 goto retrylookup; 1524 } else { 1525 if (allocflags & VM_ALLOC_WIRED) 1526 vm_page_wire(m); 1527 vm_page_busy(m); 1528 vm_page_unlock_queues(); 1529 return m; 1530 } 1531 } 1532 1533 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1534 if (m == NULL) { 1535 VM_WAIT; 1536 if ((allocflags & VM_ALLOC_RETRY) == 0) 1537 return NULL; 1538 goto retrylookup; 1539 } 1540 1541 return m; 1542} 1543 1544/* 1545 * Mapping function for valid bits or for dirty bits in 1546 * a page. May not block. 1547 * 1548 * Inputs are required to range within a page. 1549 */ 1550__inline int 1551vm_page_bits(int base, int size) 1552{ 1553 int first_bit; 1554 int last_bit; 1555 1556 KASSERT( 1557 base + size <= PAGE_SIZE, 1558 ("vm_page_bits: illegal base/size %d/%d", base, size) 1559 ); 1560 1561 if (size == 0) /* handle degenerate case */ 1562 return (0); 1563 1564 first_bit = base >> DEV_BSHIFT; 1565 last_bit = (base + size - 1) >> DEV_BSHIFT; 1566 1567 return ((2 << last_bit) - (1 << first_bit)); 1568} 1569 1570/* 1571 * vm_page_set_validclean: 1572 * 1573 * Sets portions of a page valid and clean. The arguments are expected 1574 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1575 * of any partial chunks touched by the range. The invalid portion of 1576 * such chunks will be zero'd. 1577 * 1578 * This routine may not block. 1579 * 1580 * (base + size) must be less then or equal to PAGE_SIZE. 1581 */ 1582void 1583vm_page_set_validclean(vm_page_t m, int base, int size) 1584{ 1585 int pagebits; 1586 int frag; 1587 int endoff; 1588 1589 GIANT_REQUIRED; 1590 if (size == 0) /* handle degenerate case */ 1591 return; 1592 1593 /* 1594 * If the base is not DEV_BSIZE aligned and the valid 1595 * bit is clear, we have to zero out a portion of the 1596 * first block. 1597 */ 1598 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1599 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1600 pmap_zero_page_area(m, frag, base - frag); 1601 1602 /* 1603 * If the ending offset is not DEV_BSIZE aligned and the 1604 * valid bit is clear, we have to zero out a portion of 1605 * the last block. 1606 */ 1607 endoff = base + size; 1608 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1609 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1610 pmap_zero_page_area(m, endoff, 1611 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1612 1613 /* 1614 * Set valid, clear dirty bits. If validating the entire 1615 * page we can safely clear the pmap modify bit. We also 1616 * use this opportunity to clear the PG_NOSYNC flag. If a process 1617 * takes a write fault on a MAP_NOSYNC memory area the flag will 1618 * be set again. 1619 * 1620 * We set valid bits inclusive of any overlap, but we can only 1621 * clear dirty bits for DEV_BSIZE chunks that are fully within 1622 * the range. 1623 */ 1624 pagebits = vm_page_bits(base, size); 1625 m->valid |= pagebits; 1626#if 0 /* NOT YET */ 1627 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1628 frag = DEV_BSIZE - frag; 1629 base += frag; 1630 size -= frag; 1631 if (size < 0) 1632 size = 0; 1633 } 1634 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1635#endif 1636 m->dirty &= ~pagebits; 1637 if (base == 0 && size == PAGE_SIZE) { 1638 pmap_clear_modify(m); 1639 vm_page_flag_clear(m, PG_NOSYNC); 1640 } 1641} 1642 1643#if 0 1644 1645void 1646vm_page_set_dirty(vm_page_t m, int base, int size) 1647{ 1648 m->dirty |= vm_page_bits(base, size); 1649} 1650 1651#endif 1652 1653void 1654vm_page_clear_dirty(vm_page_t m, int base, int size) 1655{ 1656 GIANT_REQUIRED; 1657 m->dirty &= ~vm_page_bits(base, size); 1658} 1659 1660/* 1661 * vm_page_set_invalid: 1662 * 1663 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1664 * valid and dirty bits for the effected areas are cleared. 1665 * 1666 * May not block. 1667 */ 1668void 1669vm_page_set_invalid(vm_page_t m, int base, int size) 1670{ 1671 int bits; 1672 1673 GIANT_REQUIRED; 1674 bits = vm_page_bits(base, size); 1675 m->valid &= ~bits; 1676 m->dirty &= ~bits; 1677 m->object->generation++; 1678} 1679 1680/* 1681 * vm_page_zero_invalid() 1682 * 1683 * The kernel assumes that the invalid portions of a page contain 1684 * garbage, but such pages can be mapped into memory by user code. 1685 * When this occurs, we must zero out the non-valid portions of the 1686 * page so user code sees what it expects. 1687 * 1688 * Pages are most often semi-valid when the end of a file is mapped 1689 * into memory and the file's size is not page aligned. 1690 */ 1691void 1692vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1693{ 1694 int b; 1695 int i; 1696 1697 /* 1698 * Scan the valid bits looking for invalid sections that 1699 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1700 * valid bit may be set ) have already been zerod by 1701 * vm_page_set_validclean(). 1702 */ 1703 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1704 if (i == (PAGE_SIZE / DEV_BSIZE) || 1705 (m->valid & (1 << i)) 1706 ) { 1707 if (i > b) { 1708 pmap_zero_page_area(m, 1709 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1710 } 1711 b = i + 1; 1712 } 1713 } 1714 1715 /* 1716 * setvalid is TRUE when we can safely set the zero'd areas 1717 * as being valid. We can do this if there are no cache consistancy 1718 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1719 */ 1720 if (setvalid) 1721 m->valid = VM_PAGE_BITS_ALL; 1722} 1723 1724/* 1725 * vm_page_is_valid: 1726 * 1727 * Is (partial) page valid? Note that the case where size == 0 1728 * will return FALSE in the degenerate case where the page is 1729 * entirely invalid, and TRUE otherwise. 1730 * 1731 * May not block. 1732 */ 1733int 1734vm_page_is_valid(vm_page_t m, int base, int size) 1735{ 1736 int bits = vm_page_bits(base, size); 1737 1738 if (m->valid && ((m->valid & bits) == bits)) 1739 return 1; 1740 else 1741 return 0; 1742} 1743 1744/* 1745 * update dirty bits from pmap/mmu. May not block. 1746 */ 1747void 1748vm_page_test_dirty(vm_page_t m) 1749{ 1750 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1751 vm_page_dirty(m); 1752 } 1753} 1754 1755int so_zerocp_fullpage = 0; 1756 1757void 1758vm_page_cowfault(vm_page_t m) 1759{ 1760 vm_page_t mnew; 1761 vm_object_t object; 1762 vm_pindex_t pindex; 1763 1764 object = m->object; 1765 pindex = m->pindex; 1766 vm_page_busy(m); 1767 1768 retry_alloc: 1769 vm_page_remove(m); 1770 /* 1771 * An interrupt allocation is requested because the page 1772 * queues lock is held. 1773 */ 1774 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1775 if (mnew == NULL) { 1776 vm_page_insert(m, object, pindex); 1777 vm_page_unlock_queues(); 1778 VM_WAIT; 1779 vm_page_lock_queues(); 1780 goto retry_alloc; 1781 } 1782 1783 if (m->cow == 0) { 1784 /* 1785 * check to see if we raced with an xmit complete when 1786 * waiting to allocate a page. If so, put things back 1787 * the way they were 1788 */ 1789 vm_page_busy(mnew); 1790 vm_page_free(mnew); 1791 vm_page_insert(m, object, pindex); 1792 } else { /* clear COW & copy page */ 1793 if (so_zerocp_fullpage) { 1794 mnew->valid = VM_PAGE_BITS_ALL; 1795 } else { 1796 vm_page_copy(m, mnew); 1797 } 1798 vm_page_dirty(mnew); 1799 vm_page_flag_clear(mnew, PG_BUSY); 1800 } 1801} 1802 1803void 1804vm_page_cowclear(vm_page_t m) 1805{ 1806 1807 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1808 if (m->cow) { 1809 m->cow--; 1810 /* 1811 * let vm_fault add back write permission lazily 1812 */ 1813 } 1814 /* 1815 * sf_buf_free() will free the page, so we needn't do it here 1816 */ 1817} 1818 1819void 1820vm_page_cowsetup(vm_page_t m) 1821{ 1822 1823 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1824 m->cow++; 1825 vm_page_protect(m, VM_PROT_READ); 1826} 1827 1828#include "opt_ddb.h" 1829#ifdef DDB 1830#include <sys/kernel.h> 1831 1832#include <ddb/ddb.h> 1833 1834DB_SHOW_COMMAND(page, vm_page_print_page_info) 1835{ 1836 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1837 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1838 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1839 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1840 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1841 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1842 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1843 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1844 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1845 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1846} 1847 1848DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1849{ 1850 int i; 1851 db_printf("PQ_FREE:"); 1852 for (i = 0; i < PQ_L2_SIZE; i++) { 1853 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1854 } 1855 db_printf("\n"); 1856 1857 db_printf("PQ_CACHE:"); 1858 for (i = 0; i < PQ_L2_SIZE; i++) { 1859 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1860 } 1861 db_printf("\n"); 1862 1863 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1864 vm_page_queues[PQ_ACTIVE].lcnt, 1865 vm_page_queues[PQ_INACTIVE].lcnt); 1866} 1867#endif /* DDB */ 1868