vm_page.c revision 106359
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 106359 2002-11-02 22:39:30Z 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 */ 128static struct vm_page **vm_page_buckets; /* Array of buckets */ 129static int vm_page_bucket_count; /* How big is array? */ 130static int vm_page_hash_mask; /* Mask for hash function */ 131 132struct mtx vm_page_queue_mtx; 133struct mtx vm_page_queue_free_mtx; 134 135vm_page_t vm_page_array = 0; 136int vm_page_array_size = 0; 137long first_page = 0; 138int vm_page_zero_count = 0; 139 140/* 141 * vm_set_page_size: 142 * 143 * Sets the page size, perhaps based upon the memory 144 * size. Must be called before any use of page-size 145 * dependent functions. 146 */ 147void 148vm_set_page_size(void) 149{ 150 if (cnt.v_page_size == 0) 151 cnt.v_page_size = PAGE_SIZE; 152 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 153 panic("vm_set_page_size: page size not a power of two"); 154} 155 156/* 157 * vm_page_startup: 158 * 159 * Initializes the resident memory module. 160 * 161 * Allocates memory for the page cells, and 162 * for the object/offset-to-page hash table headers. 163 * Each page cell is initialized and placed on the free list. 164 */ 165vm_offset_t 166vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 167{ 168 vm_offset_t mapped; 169 struct vm_page **bucket; 170 vm_size_t npages, page_range; 171 vm_offset_t new_end; 172 int i; 173 vm_offset_t pa; 174 int nblocks; 175 vm_offset_t last_pa; 176 177 /* the biggest memory array is the second group of pages */ 178 vm_offset_t end; 179 vm_offset_t biggestone, biggestsize; 180 181 vm_offset_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_size_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 allocator 223 */ 224 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 225 new_end = end - bootpages; 226 new_end = trunc_page(new_end); 227 mapped = pmap_map(&vaddr, new_end, end, 228 VM_PROT_READ | VM_PROT_WRITE); 229 bzero((caddr_t) mapped, end - new_end); 230 uma_startup((caddr_t)mapped); 231 232 end = new_end; 233 234 /* 235 * Allocate (and initialize) the hash table buckets. 236 * 237 * The number of buckets MUST BE a power of 2, and the actual value is 238 * the next power of 2 greater than the number of physical pages in 239 * the system. 240 * 241 * We make the hash table approximately 2x the number of pages to 242 * reduce the chain length. This is about the same size using the 243 * singly-linked list as the 1x hash table we were using before 244 * using TAILQ but the chain length will be smaller. 245 * 246 * Note: This computation can be tweaked if desired. 247 */ 248 if (vm_page_bucket_count == 0) { 249 vm_page_bucket_count = 1; 250 while (vm_page_bucket_count < atop(total)) 251 vm_page_bucket_count <<= 1; 252 } 253 vm_page_bucket_count <<= 1; 254 vm_page_hash_mask = vm_page_bucket_count - 1; 255 256 /* 257 * Validate these addresses. 258 */ 259 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 260 new_end = trunc_page(new_end); 261 mapped = pmap_map(&vaddr, new_end, end, 262 VM_PROT_READ | VM_PROT_WRITE); 263 bzero((caddr_t) mapped, end - new_end); 264 265 vm_page_buckets = (struct vm_page **)mapped; 266 bucket = vm_page_buckets; 267 for (i = 0; i < vm_page_bucket_count; i++) { 268 *bucket = NULL; 269 bucket++; 270 } 271 272 /* 273 * Compute the number of pages of memory that will be available for 274 * use (taking into account the overhead of a page structure per 275 * page). 276 */ 277 first_page = phys_avail[0] / PAGE_SIZE; 278 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 279 npages = (total - (page_range * sizeof(struct vm_page)) - 280 (end - new_end)) / PAGE_SIZE; 281 end = new_end; 282 283 /* 284 * Initialize the mem entry structures now, and put them in the free 285 * queue. 286 */ 287 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 288 mapped = pmap_map(&vaddr, new_end, end, 289 VM_PROT_READ | VM_PROT_WRITE); 290 vm_page_array = (vm_page_t) mapped; 291 292 /* 293 * Clear all of the page structures 294 */ 295 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 296 vm_page_array_size = page_range; 297 298 /* 299 * Construct the free queue(s) in descending order (by physical 300 * address) so that the first 16MB of physical memory is allocated 301 * last rather than first. On large-memory machines, this avoids 302 * the exhaustion of low physical memory before isa_dmainit has run. 303 */ 304 cnt.v_page_count = 0; 305 cnt.v_free_count = 0; 306 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 307 pa = phys_avail[i]; 308 if (i == biggestone) 309 last_pa = new_end; 310 else 311 last_pa = phys_avail[i + 1]; 312 while (pa < last_pa && npages-- > 0) { 313 vm_pageq_add_new_page(pa); 314 pa += PAGE_SIZE; 315 } 316 } 317 return (vaddr); 318} 319 320void 321vm_page_flag_set(vm_page_t m, unsigned short bits) 322{ 323 GIANT_REQUIRED; 324 m->flags |= bits; 325} 326 327void 328vm_page_flag_clear(vm_page_t m, unsigned short bits) 329{ 330 GIANT_REQUIRED; 331 m->flags &= ~bits; 332} 333 334void 335vm_page_busy(vm_page_t m) 336{ 337 KASSERT((m->flags & PG_BUSY) == 0, 338 ("vm_page_busy: page already busy!!!")); 339 vm_page_flag_set(m, PG_BUSY); 340} 341 342/* 343 * vm_page_flash: 344 * 345 * wakeup anyone waiting for the page. 346 */ 347void 348vm_page_flash(vm_page_t m) 349{ 350 if (m->flags & PG_WANTED) { 351 vm_page_flag_clear(m, PG_WANTED); 352 wakeup(m); 353 } 354} 355 356/* 357 * vm_page_wakeup: 358 * 359 * clear the PG_BUSY flag and wakeup anyone waiting for the 360 * page. 361 * 362 */ 363void 364vm_page_wakeup(vm_page_t m) 365{ 366 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 367 vm_page_flag_clear(m, PG_BUSY); 368 vm_page_flash(m); 369} 370 371/* 372 * 373 * 374 */ 375void 376vm_page_io_start(vm_page_t m) 377{ 378 379 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 380 m->busy++; 381} 382 383void 384vm_page_io_finish(vm_page_t m) 385{ 386 387 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 388 m->busy--; 389 if (m->busy == 0) 390 vm_page_flash(m); 391} 392 393/* 394 * Keep page from being freed by the page daemon 395 * much of the same effect as wiring, except much lower 396 * overhead and should be used only for *very* temporary 397 * holding ("wiring"). 398 */ 399void 400vm_page_hold(vm_page_t mem) 401{ 402 GIANT_REQUIRED; 403 mem->hold_count++; 404} 405 406void 407vm_page_unhold(vm_page_t mem) 408{ 409 GIANT_REQUIRED; 410 --mem->hold_count; 411 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 412 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 413 vm_page_free_toq(mem); 414} 415 416/* 417 * vm_page_protect: 418 * 419 * Reduce the protection of a page. This routine never raises the 420 * protection and therefore can be safely called if the page is already 421 * at VM_PROT_NONE (it will be a NOP effectively ). 422 */ 423void 424vm_page_protect(vm_page_t mem, int prot) 425{ 426 if (prot == VM_PROT_NONE) { 427 if (pmap_page_is_mapped(mem) || (mem->flags & PG_WRITEABLE)) { 428 pmap_page_protect(mem, VM_PROT_NONE); 429 vm_page_flag_clear(mem, PG_WRITEABLE); 430 } 431 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 432 pmap_page_protect(mem, VM_PROT_READ); 433 vm_page_flag_clear(mem, PG_WRITEABLE); 434 } 435} 436 437/* 438 * vm_page_copy: 439 * 440 * Copy one page to another 441 */ 442void 443vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 444{ 445 pmap_copy_page(src_m, dest_m); 446 dest_m->valid = VM_PAGE_BITS_ALL; 447} 448 449/* 450 * vm_page_free: 451 * 452 * Free a page 453 * 454 * The clearing of PG_ZERO is a temporary safety until the code can be 455 * reviewed to determine that PG_ZERO is being properly cleared on 456 * write faults or maps. PG_ZERO was previously cleared in 457 * vm_page_alloc(). 458 */ 459void 460vm_page_free(vm_page_t m) 461{ 462 vm_page_flag_clear(m, PG_ZERO); 463 vm_page_free_toq(m); 464 vm_page_zero_idle_wakeup(); 465} 466 467/* 468 * vm_page_free_zero: 469 * 470 * Free a page to the zerod-pages queue 471 */ 472void 473vm_page_free_zero(vm_page_t m) 474{ 475 vm_page_flag_set(m, PG_ZERO); 476 vm_page_free_toq(m); 477} 478 479/* 480 * vm_page_sleep_busy: 481 * 482 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 483 * m->busy is zero. Returns TRUE if it had to sleep ( including if 484 * it almost had to sleep and made temporary spl*() mods), FALSE 485 * otherwise. 486 * 487 * This routine assumes that interrupts can only remove the busy 488 * status from a page, not set the busy status or change it from 489 * PG_BUSY to m->busy or vise versa (which would create a timing 490 * window). 491 */ 492int 493vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 494{ 495 GIANT_REQUIRED; 496 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 497 int s = splvm(); 498 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 499 /* 500 * Page is busy. Wait and retry. 501 */ 502 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 503 tsleep(m, PVM, msg, 0); 504 } 505 splx(s); 506 return (TRUE); 507 /* not reached */ 508 } 509 return (FALSE); 510} 511 512/* 513 * vm_page_sleep_if_busy: 514 * 515 * Sleep and release the page queues lock if PG_BUSY is set or, 516 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 517 * thread slept and the page queues lock was released. 518 * Otherwise, retains the page queues lock and returns FALSE. 519 */ 520int 521vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 522{ 523 524 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 525 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 526 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 527 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 528 return (TRUE); 529 } 530 return (FALSE); 531} 532 533/* 534 * vm_page_dirty: 535 * 536 * make page all dirty 537 */ 538void 539vm_page_dirty(vm_page_t m) 540{ 541 KASSERT(m->queue - m->pc != PQ_CACHE, 542 ("vm_page_dirty: page in cache!")); 543 m->dirty = VM_PAGE_BITS_ALL; 544} 545 546/* 547 * vm_page_splay: 548 * 549 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 550 * the vm_page containing the given pindex. If, however, that 551 * pindex is not found in the vm_object, returns a vm_page that is 552 * adjacent to the pindex, coming before or after it. 553 */ 554static vm_page_t 555vm_page_splay(vm_pindex_t pindex, vm_page_t root) 556{ 557 struct vm_page dummy; 558 vm_page_t lefttreemax, righttreemin, y; 559 560 if (root == NULL) 561 return (root); 562 lefttreemax = righttreemin = &dummy; 563 for (;; root = y) { 564 if (pindex < root->pindex) { 565 if ((y = root->left) == NULL) 566 break; 567 if (pindex < y->pindex) { 568 /* Rotate right. */ 569 root->left = y->right; 570 y->right = root; 571 root = y; 572 if ((y = root->left) == NULL) 573 break; 574 } 575 /* Link into the new root's right tree. */ 576 righttreemin->left = root; 577 righttreemin = root; 578 } else if (pindex > root->pindex) { 579 if ((y = root->right) == NULL) 580 break; 581 if (pindex > y->pindex) { 582 /* Rotate left. */ 583 root->right = y->left; 584 y->left = root; 585 root = y; 586 if ((y = root->right) == NULL) 587 break; 588 } 589 /* Link into the new root's left tree. */ 590 lefttreemax->right = root; 591 lefttreemax = root; 592 } else 593 break; 594 } 595 /* Assemble the new root. */ 596 lefttreemax->right = root->left; 597 righttreemin->left = root->right; 598 root->left = dummy.right; 599 root->right = dummy.left; 600 return (root); 601} 602 603/* 604 * vm_page_insert: [ internal use only ] 605 * 606 * Inserts the given mem entry into the object and object list. 607 * 608 * The pagetables are not updated but will presumably fault the page 609 * in if necessary, or if a kernel page the caller will at some point 610 * enter the page into the kernel's pmap. We are not allowed to block 611 * here so we *can't* do this anyway. 612 * 613 * The object and page must be locked, and must be splhigh. 614 * This routine may not block. 615 */ 616void 617vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 618{ 619 vm_page_t root; 620 621 GIANT_REQUIRED; 622 623 if (m->object != NULL) 624 panic("vm_page_insert: already inserted"); 625 626 /* 627 * Record the object/offset pair in this page 628 */ 629 m->object = object; 630 m->pindex = pindex; 631 632 /* 633 * Now link into the object's ordered list of backed pages. 634 */ 635 root = vm_page_splay(pindex, object->root); 636 if (root == NULL) { 637 m->left = NULL; 638 m->right = NULL; 639 TAILQ_INSERT_TAIL(&object->memq, m, listq); 640 } else if (pindex < root->pindex) { 641 m->left = root->left; 642 m->right = root; 643 root->left = NULL; 644 TAILQ_INSERT_BEFORE(root, m, listq); 645 } else { 646 m->right = root->right; 647 m->left = root; 648 root->right = NULL; 649 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 650 } 651 object->root = m; 652 object->generation++; 653 654 /* 655 * show that the object has one more resident page. 656 */ 657 object->resident_page_count++; 658 659 /* 660 * Since we are inserting a new and possibly dirty page, 661 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 662 */ 663 if (m->flags & PG_WRITEABLE) 664 vm_object_set_writeable_dirty(object); 665} 666 667/* 668 * vm_page_remove: 669 * NOTE: used by device pager as well -wfj 670 * 671 * Removes the given mem entry from the object/offset-page 672 * table and the object page list, but do not invalidate/terminate 673 * the backing store. 674 * 675 * The object and page must be locked, and at splhigh. 676 * The underlying pmap entry (if any) is NOT removed here. 677 * This routine may not block. 678 */ 679void 680vm_page_remove(vm_page_t m) 681{ 682 vm_object_t object; 683 vm_page_t root; 684 685 GIANT_REQUIRED; 686 687 if (m->object == NULL) 688 return; 689 690 if ((m->flags & PG_BUSY) == 0) { 691 panic("vm_page_remove: page not busy"); 692 } 693 694 /* 695 * Basically destroy the page. 696 */ 697 vm_page_wakeup(m); 698 699 object = m->object; 700 701 /* 702 * Now remove from the object's list of backed pages. 703 */ 704 if (m != object->root) 705 vm_page_splay(m->pindex, object->root); 706 if (m->left == NULL) 707 root = m->right; 708 else { 709 root = vm_page_splay(m->pindex, m->left); 710 root->right = m->right; 711 } 712 object->root = root; 713 TAILQ_REMOVE(&object->memq, m, listq); 714 715 /* 716 * And show that the object has one fewer resident page. 717 */ 718 object->resident_page_count--; 719 object->generation++; 720 721 m->object = NULL; 722} 723 724/* 725 * vm_page_lookup: 726 * 727 * Returns the page associated with the object/offset 728 * pair specified; if none is found, NULL is returned. 729 * 730 * The object must be locked. 731 * This routine may not block. 732 * This is a critical path routine 733 */ 734vm_page_t 735vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 736{ 737 vm_page_t m; 738 739 GIANT_REQUIRED; 740 741 m = vm_page_splay(pindex, object->root); 742 if ((object->root = m) != NULL && m->pindex != pindex) 743 m = NULL; 744 return (m); 745} 746 747/* 748 * vm_page_rename: 749 * 750 * Move the given memory entry from its 751 * current object to the specified target object/offset. 752 * 753 * The object must be locked. 754 * This routine may not block. 755 * 756 * Note: this routine will raise itself to splvm(), the caller need not. 757 * 758 * Note: swap associated with the page must be invalidated by the move. We 759 * have to do this for several reasons: (1) we aren't freeing the 760 * page, (2) we are dirtying the page, (3) the VM system is probably 761 * moving the page from object A to B, and will then later move 762 * the backing store from A to B and we can't have a conflict. 763 * 764 * Note: we *always* dirty the page. It is necessary both for the 765 * fact that we moved it, and because we may be invalidating 766 * swap. If the page is on the cache, we have to deactivate it 767 * or vm_page_dirty() will panic. Dirty pages are not allowed 768 * on the cache. 769 */ 770void 771vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 772{ 773 int s; 774 775 s = splvm(); 776 vm_page_lock_queues(); 777 vm_page_remove(m); 778 vm_page_insert(m, new_object, new_pindex); 779 if (m->queue - m->pc == PQ_CACHE) 780 vm_page_deactivate(m); 781 vm_page_dirty(m); 782 vm_page_unlock_queues(); 783 splx(s); 784} 785 786/* 787 * vm_page_select_cache: 788 * 789 * Find a page on the cache queue with color optimization. As pages 790 * might be found, but not applicable, they are deactivated. This 791 * keeps us from using potentially busy cached pages. 792 * 793 * This routine must be called at splvm(). 794 * This routine may not block. 795 */ 796static vm_page_t 797vm_page_select_cache(vm_pindex_t color) 798{ 799 vm_page_t m; 800 801 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 802 while (TRUE) { 803 m = vm_pageq_find(PQ_CACHE, color & PQ_L2_MASK, FALSE); 804 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 805 m->hold_count || m->wire_count)) { 806 vm_page_deactivate(m); 807 continue; 808 } 809 return m; 810 } 811} 812 813/* 814 * vm_page_select_free: 815 * 816 * Find a free or zero page, with specified preference. 817 * 818 * This routine must be called at splvm(). 819 * This routine may not block. 820 */ 821static __inline vm_page_t 822vm_page_select_free(vm_pindex_t color, boolean_t prefer_zero) 823{ 824 vm_page_t m; 825 826 m = vm_pageq_find(PQ_FREE, color & PQ_L2_MASK, prefer_zero); 827 return (m); 828} 829 830/* 831 * vm_page_alloc: 832 * 833 * Allocate and return a memory cell associated 834 * with this VM object/offset pair. 835 * 836 * page_req classes: 837 * VM_ALLOC_NORMAL normal process request 838 * VM_ALLOC_SYSTEM system *really* needs a page 839 * VM_ALLOC_INTERRUPT interrupt time request 840 * VM_ALLOC_ZERO zero page 841 * 842 * This routine may not block. 843 * 844 * Additional special handling is required when called from an 845 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 846 * the page cache in this case. 847 */ 848vm_page_t 849vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 850{ 851 vm_page_t m = NULL; 852 vm_pindex_t color; 853 int page_req, s; 854 855 GIANT_REQUIRED; 856 857#ifdef INVARIANTS 858 if ((req & VM_ALLOC_NOOBJ) == 0) { 859 KASSERT(object != NULL, 860 ("vm_page_alloc: NULL object.")); 861 KASSERT(!vm_page_lookup(object, pindex), 862 ("vm_page_alloc: page already allocated")); 863 } 864#endif 865 866 page_req = req & VM_ALLOC_CLASS_MASK; 867 868 if ((req & VM_ALLOC_NOOBJ) == 0) 869 color = pindex + object->pg_color; 870 else 871 color = pindex; 872 873 /* 874 * The pager is allowed to eat deeper into the free page list. 875 */ 876 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 877 page_req = VM_ALLOC_SYSTEM; 878 }; 879 880 s = splvm(); 881loop: 882 mtx_lock_spin(&vm_page_queue_free_mtx); 883 if (cnt.v_free_count > cnt.v_free_reserved) { 884 /* 885 * Allocate from the free queue if there are plenty of pages 886 * in it. 887 */ 888 m = vm_page_select_free(color, (req & VM_ALLOC_ZERO) != 0); 889 890 } else if ( 891 (page_req == VM_ALLOC_SYSTEM && 892 cnt.v_cache_count == 0 && 893 cnt.v_free_count > cnt.v_interrupt_free_min) || 894 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 895 ) { 896 /* 897 * Interrupt or system, dig deeper into the free list. 898 */ 899 m = vm_page_select_free(color, FALSE); 900 } else if (page_req != VM_ALLOC_INTERRUPT) { 901 mtx_unlock_spin(&vm_page_queue_free_mtx); 902 /* 903 * Allocatable from cache (non-interrupt only). On success, 904 * we must free the page and try again, thus ensuring that 905 * cnt.v_*_free_min counters are replenished. 906 */ 907 vm_page_lock_queues(); 908 if ((m = vm_page_select_cache(color)) == NULL) { 909 vm_page_unlock_queues(); 910 splx(s); 911#if defined(DIAGNOSTIC) 912 if (cnt.v_cache_count > 0) 913 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 914#endif 915 vm_pageout_deficit++; 916 pagedaemon_wakeup(); 917 return (NULL); 918 } 919 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 920 vm_page_busy(m); 921 vm_page_protect(m, VM_PROT_NONE); 922 vm_page_free(m); 923 vm_page_unlock_queues(); 924 goto loop; 925 } else { 926 /* 927 * Not allocatable from cache from interrupt, give up. 928 */ 929 mtx_unlock_spin(&vm_page_queue_free_mtx); 930 splx(s); 931 vm_pageout_deficit++; 932 pagedaemon_wakeup(); 933 return (NULL); 934 } 935 936 /* 937 * At this point we had better have found a good page. 938 */ 939 940 KASSERT( 941 m != NULL, 942 ("vm_page_alloc(): missing page on free queue\n") 943 ); 944 945 /* 946 * Remove from free queue 947 */ 948 949 vm_pageq_remove_nowakeup(m); 950 951 /* 952 * Initialize structure. Only the PG_ZERO flag is inherited. 953 */ 954 if (m->flags & PG_ZERO) { 955 vm_page_zero_count--; 956 m->flags = PG_ZERO | PG_BUSY; 957 } else { 958 m->flags = PG_BUSY; 959 } 960 if (req & VM_ALLOC_WIRED) { 961 cnt.v_wire_count++; 962 m->wire_count = 1; 963 } else 964 m->wire_count = 0; 965 m->hold_count = 0; 966 m->act_count = 0; 967 m->busy = 0; 968 m->valid = 0; 969 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 970 mtx_unlock_spin(&vm_page_queue_free_mtx); 971 972 /* 973 * vm_page_insert() is safe prior to the splx(). Note also that 974 * inserting a page here does not insert it into the pmap (which 975 * could cause us to block allocating memory). We cannot block 976 * anywhere. 977 */ 978 if ((req & VM_ALLOC_NOOBJ) == 0) 979 vm_page_insert(m, object, pindex); 980 981 /* 982 * Don't wakeup too often - wakeup the pageout daemon when 983 * we would be nearly out of memory. 984 */ 985 if (vm_paging_needed()) 986 pagedaemon_wakeup(); 987 988 splx(s); 989 return (m); 990} 991 992/* 993 * vm_wait: (also see VM_WAIT macro) 994 * 995 * Block until free pages are available for allocation 996 * - Called in various places before memory allocations. 997 */ 998void 999vm_wait(void) 1000{ 1001 int s; 1002 1003 s = splvm(); 1004 if (curproc == pageproc) { 1005 vm_pageout_pages_needed = 1; 1006 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 1007 } else { 1008 if (!vm_pages_needed) { 1009 vm_pages_needed = 1; 1010 wakeup(&vm_pages_needed); 1011 } 1012 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 1013 } 1014 splx(s); 1015} 1016 1017/* 1018 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1019 * 1020 * Block until free pages are available for allocation 1021 * - Called only in vm_fault so that processes page faulting 1022 * can be easily tracked. 1023 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1024 * processes will be able to grab memory first. Do not change 1025 * this balance without careful testing first. 1026 */ 1027void 1028vm_waitpfault(void) 1029{ 1030 int s; 1031 1032 s = splvm(); 1033 if (!vm_pages_needed) { 1034 vm_pages_needed = 1; 1035 wakeup(&vm_pages_needed); 1036 } 1037 tsleep(&cnt.v_free_count, PUSER, "pfault", 0); 1038 splx(s); 1039} 1040 1041/* 1042 * vm_page_activate: 1043 * 1044 * Put the specified page on the active list (if appropriate). 1045 * Ensure that act_count is at least ACT_INIT but do not otherwise 1046 * mess with it. 1047 * 1048 * The page queues must be locked. 1049 * This routine may not block. 1050 */ 1051void 1052vm_page_activate(vm_page_t m) 1053{ 1054 int s; 1055 1056 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1057 s = splvm(); 1058 if (m->queue != PQ_ACTIVE) { 1059 if ((m->queue - m->pc) == PQ_CACHE) 1060 cnt.v_reactivated++; 1061 vm_pageq_remove(m); 1062 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1063 if (m->act_count < ACT_INIT) 1064 m->act_count = ACT_INIT; 1065 vm_pageq_enqueue(PQ_ACTIVE, m); 1066 } 1067 } else { 1068 if (m->act_count < ACT_INIT) 1069 m->act_count = ACT_INIT; 1070 } 1071 splx(s); 1072} 1073 1074/* 1075 * vm_page_free_wakeup: 1076 * 1077 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1078 * routine is called when a page has been added to the cache or free 1079 * queues. 1080 * 1081 * This routine may not block. 1082 * This routine must be called at splvm() 1083 */ 1084static __inline void 1085vm_page_free_wakeup(void) 1086{ 1087 /* 1088 * if pageout daemon needs pages, then tell it that there are 1089 * some free. 1090 */ 1091 if (vm_pageout_pages_needed && 1092 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1093 wakeup(&vm_pageout_pages_needed); 1094 vm_pageout_pages_needed = 0; 1095 } 1096 /* 1097 * wakeup processes that are waiting on memory if we hit a 1098 * high water mark. And wakeup scheduler process if we have 1099 * lots of memory. this process will swapin processes. 1100 */ 1101 if (vm_pages_needed && !vm_page_count_min()) { 1102 vm_pages_needed = 0; 1103 wakeup(&cnt.v_free_count); 1104 } 1105} 1106 1107/* 1108 * vm_page_free_toq: 1109 * 1110 * Returns the given page to the PQ_FREE list, 1111 * disassociating it with any VM object. 1112 * 1113 * Object and page must be locked prior to entry. 1114 * This routine may not block. 1115 */ 1116 1117void 1118vm_page_free_toq(vm_page_t m) 1119{ 1120 int s; 1121 struct vpgqueues *pq; 1122 vm_object_t object = m->object; 1123 1124 GIANT_REQUIRED; 1125 s = splvm(); 1126 cnt.v_tfree++; 1127 1128 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1129 printf( 1130 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1131 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1132 m->hold_count); 1133 if ((m->queue - m->pc) == PQ_FREE) 1134 panic("vm_page_free: freeing free page"); 1135 else 1136 panic("vm_page_free: freeing busy page"); 1137 } 1138 1139 /* 1140 * unqueue, then remove page. Note that we cannot destroy 1141 * the page here because we do not want to call the pager's 1142 * callback routine until after we've put the page on the 1143 * appropriate free queue. 1144 */ 1145 vm_pageq_remove_nowakeup(m); 1146 vm_page_remove(m); 1147 1148 /* 1149 * If fictitious remove object association and 1150 * return, otherwise delay object association removal. 1151 */ 1152 if ((m->flags & PG_FICTITIOUS) != 0) { 1153 splx(s); 1154 return; 1155 } 1156 1157 m->valid = 0; 1158 vm_page_undirty(m); 1159 1160 if (m->wire_count != 0) { 1161 if (m->wire_count > 1) { 1162 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1163 m->wire_count, (long)m->pindex); 1164 } 1165 panic("vm_page_free: freeing wired page\n"); 1166 } 1167 1168 /* 1169 * If we've exhausted the object's resident pages we want to free 1170 * it up. 1171 */ 1172 if (object && 1173 (object->type == OBJT_VNODE) && 1174 ((object->flags & OBJ_DEAD) == 0) 1175 ) { 1176 struct vnode *vp = (struct vnode *)object->handle; 1177 1178 if (vp) { 1179 VI_LOCK(vp); 1180 if (VSHOULDFREE(vp)) 1181 vfree(vp); 1182 VI_UNLOCK(vp); 1183 } 1184 } 1185 1186 /* 1187 * Clear the UNMANAGED flag when freeing an unmanaged page. 1188 */ 1189 if (m->flags & PG_UNMANAGED) { 1190 m->flags &= ~PG_UNMANAGED; 1191 } else { 1192#ifdef __alpha__ 1193 pmap_page_is_free(m); 1194#endif 1195 } 1196 1197 if (m->hold_count != 0) { 1198 m->flags &= ~PG_ZERO; 1199 m->queue = PQ_HOLD; 1200 } else 1201 m->queue = PQ_FREE + m->pc; 1202 pq = &vm_page_queues[m->queue]; 1203 mtx_lock_spin(&vm_page_queue_free_mtx); 1204 pq->lcnt++; 1205 ++(*pq->cnt); 1206 1207 /* 1208 * Put zero'd pages on the end ( where we look for zero'd pages 1209 * first ) and non-zerod pages at the head. 1210 */ 1211 if (m->flags & PG_ZERO) { 1212 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1213 ++vm_page_zero_count; 1214 } else { 1215 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1216 } 1217 mtx_unlock_spin(&vm_page_queue_free_mtx); 1218 vm_page_free_wakeup(); 1219 splx(s); 1220} 1221 1222/* 1223 * vm_page_unmanage: 1224 * 1225 * Prevent PV management from being done on the page. The page is 1226 * removed from the paging queues as if it were wired, and as a 1227 * consequence of no longer being managed the pageout daemon will not 1228 * touch it (since there is no way to locate the pte mappings for the 1229 * page). madvise() calls that mess with the pmap will also no longer 1230 * operate on the page. 1231 * 1232 * Beyond that the page is still reasonably 'normal'. Freeing the page 1233 * will clear the flag. 1234 * 1235 * This routine is used by OBJT_PHYS objects - objects using unswappable 1236 * physical memory as backing store rather then swap-backed memory and 1237 * will eventually be extended to support 4MB unmanaged physical 1238 * mappings. 1239 */ 1240void 1241vm_page_unmanage(vm_page_t m) 1242{ 1243 int s; 1244 1245 s = splvm(); 1246 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1247 if ((m->flags & PG_UNMANAGED) == 0) { 1248 if (m->wire_count == 0) 1249 vm_pageq_remove(m); 1250 } 1251 vm_page_flag_set(m, PG_UNMANAGED); 1252 splx(s); 1253} 1254 1255/* 1256 * vm_page_wire: 1257 * 1258 * Mark this page as wired down by yet 1259 * another map, removing it from paging queues 1260 * as necessary. 1261 * 1262 * The page queues must be locked. 1263 * This routine may not block. 1264 */ 1265void 1266vm_page_wire(vm_page_t m) 1267{ 1268 int s; 1269 1270 /* 1271 * Only bump the wire statistics if the page is not already wired, 1272 * and only unqueue the page if it is on some queue (if it is unmanaged 1273 * it is already off the queues). 1274 */ 1275 s = splvm(); 1276 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1277 if (m->wire_count == 0) { 1278 if ((m->flags & PG_UNMANAGED) == 0) 1279 vm_pageq_remove(m); 1280 cnt.v_wire_count++; 1281 } 1282 m->wire_count++; 1283 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1284 splx(s); 1285} 1286 1287/* 1288 * vm_page_unwire: 1289 * 1290 * Release one wiring of this page, potentially 1291 * enabling it to be paged again. 1292 * 1293 * Many pages placed on the inactive queue should actually go 1294 * into the cache, but it is difficult to figure out which. What 1295 * we do instead, if the inactive target is well met, is to put 1296 * clean pages at the head of the inactive queue instead of the tail. 1297 * This will cause them to be moved to the cache more quickly and 1298 * if not actively re-referenced, freed more quickly. If we just 1299 * stick these pages at the end of the inactive queue, heavy filesystem 1300 * meta-data accesses can cause an unnecessary paging load on memory bound 1301 * processes. This optimization causes one-time-use metadata to be 1302 * reused more quickly. 1303 * 1304 * BUT, if we are in a low-memory situation we have no choice but to 1305 * put clean pages on the cache queue. 1306 * 1307 * A number of routines use vm_page_unwire() to guarantee that the page 1308 * will go into either the inactive or active queues, and will NEVER 1309 * be placed in the cache - for example, just after dirtying a page. 1310 * dirty pages in the cache are not allowed. 1311 * 1312 * The page queues must be locked. 1313 * This routine may not block. 1314 */ 1315void 1316vm_page_unwire(vm_page_t m, int activate) 1317{ 1318 int s; 1319 1320 s = splvm(); 1321 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1322 if (m->wire_count > 0) { 1323 m->wire_count--; 1324 if (m->wire_count == 0) { 1325 cnt.v_wire_count--; 1326 if (m->flags & PG_UNMANAGED) { 1327 ; 1328 } else if (activate) 1329 vm_pageq_enqueue(PQ_ACTIVE, m); 1330 else { 1331 vm_page_flag_clear(m, PG_WINATCFLS); 1332 vm_pageq_enqueue(PQ_INACTIVE, m); 1333 } 1334 } 1335 } else { 1336 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1337 } 1338 splx(s); 1339} 1340 1341 1342/* 1343 * Move the specified page to the inactive queue. If the page has 1344 * any associated swap, the swap is deallocated. 1345 * 1346 * Normally athead is 0 resulting in LRU operation. athead is set 1347 * to 1 if we want this page to be 'as if it were placed in the cache', 1348 * except without unmapping it from the process address space. 1349 * 1350 * This routine may not block. 1351 */ 1352static __inline void 1353_vm_page_deactivate(vm_page_t m, int athead) 1354{ 1355 int s; 1356 1357 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1358 /* 1359 * Ignore if already inactive. 1360 */ 1361 if (m->queue == PQ_INACTIVE) 1362 return; 1363 1364 s = splvm(); 1365 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1366 if ((m->queue - m->pc) == PQ_CACHE) 1367 cnt.v_reactivated++; 1368 vm_page_flag_clear(m, PG_WINATCFLS); 1369 vm_pageq_remove(m); 1370 if (athead) 1371 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1372 else 1373 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1374 m->queue = PQ_INACTIVE; 1375 vm_page_queues[PQ_INACTIVE].lcnt++; 1376 cnt.v_inactive_count++; 1377 } 1378 splx(s); 1379} 1380 1381void 1382vm_page_deactivate(vm_page_t m) 1383{ 1384 _vm_page_deactivate(m, 0); 1385} 1386 1387/* 1388 * vm_page_try_to_cache: 1389 * 1390 * Returns 0 on failure, 1 on success 1391 */ 1392int 1393vm_page_try_to_cache(vm_page_t m) 1394{ 1395 1396 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1397 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1398 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1399 return (0); 1400 } 1401 vm_page_test_dirty(m); 1402 if (m->dirty) 1403 return (0); 1404 vm_page_cache(m); 1405 return (1); 1406} 1407 1408/* 1409 * vm_page_try_to_free() 1410 * 1411 * Attempt to free the page. If we cannot free it, we do nothing. 1412 * 1 is returned on success, 0 on failure. 1413 */ 1414int 1415vm_page_try_to_free(vm_page_t m) 1416{ 1417 1418 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1419 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1420 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1421 return (0); 1422 } 1423 vm_page_test_dirty(m); 1424 if (m->dirty) 1425 return (0); 1426 vm_page_busy(m); 1427 vm_page_protect(m, VM_PROT_NONE); 1428 vm_page_free(m); 1429 return (1); 1430} 1431 1432/* 1433 * vm_page_cache 1434 * 1435 * Put the specified page onto the page cache queue (if appropriate). 1436 * 1437 * This routine may not block. 1438 */ 1439void 1440vm_page_cache(vm_page_t m) 1441{ 1442 int s; 1443 1444 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1445 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1446 printf("vm_page_cache: attempting to cache busy page\n"); 1447 return; 1448 } 1449 if ((m->queue - m->pc) == PQ_CACHE) 1450 return; 1451 1452 /* 1453 * Remove all pmaps and indicate that the page is not 1454 * writeable or mapped. 1455 */ 1456 vm_page_protect(m, VM_PROT_NONE); 1457 if (m->dirty != 0) { 1458 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1459 (long)m->pindex); 1460 } 1461 s = splvm(); 1462 vm_pageq_remove_nowakeup(m); 1463 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1464 vm_page_free_wakeup(); 1465 splx(s); 1466} 1467 1468/* 1469 * vm_page_dontneed 1470 * 1471 * Cache, deactivate, or do nothing as appropriate. This routine 1472 * is typically used by madvise() MADV_DONTNEED. 1473 * 1474 * Generally speaking we want to move the page into the cache so 1475 * it gets reused quickly. However, this can result in a silly syndrome 1476 * due to the page recycling too quickly. Small objects will not be 1477 * fully cached. On the otherhand, if we move the page to the inactive 1478 * queue we wind up with a problem whereby very large objects 1479 * unnecessarily blow away our inactive and cache queues. 1480 * 1481 * The solution is to move the pages based on a fixed weighting. We 1482 * either leave them alone, deactivate them, or move them to the cache, 1483 * where moving them to the cache has the highest weighting. 1484 * By forcing some pages into other queues we eventually force the 1485 * system to balance the queues, potentially recovering other unrelated 1486 * space from active. The idea is to not force this to happen too 1487 * often. 1488 */ 1489void 1490vm_page_dontneed(vm_page_t m) 1491{ 1492 static int dnweight; 1493 int dnw; 1494 int head; 1495 1496 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1497 dnw = ++dnweight; 1498 1499 /* 1500 * occassionally leave the page alone 1501 */ 1502 if ((dnw & 0x01F0) == 0 || 1503 m->queue == PQ_INACTIVE || 1504 m->queue - m->pc == PQ_CACHE 1505 ) { 1506 if (m->act_count >= ACT_INIT) 1507 --m->act_count; 1508 return; 1509 } 1510 1511 if (m->dirty == 0) 1512 vm_page_test_dirty(m); 1513 1514 if (m->dirty || (dnw & 0x0070) == 0) { 1515 /* 1516 * Deactivate the page 3 times out of 32. 1517 */ 1518 head = 0; 1519 } else { 1520 /* 1521 * Cache the page 28 times out of every 32. Note that 1522 * the page is deactivated instead of cached, but placed 1523 * at the head of the queue instead of the tail. 1524 */ 1525 head = 1; 1526 } 1527 _vm_page_deactivate(m, head); 1528} 1529 1530/* 1531 * Grab a page, waiting until we are waken up due to the page 1532 * changing state. We keep on waiting, if the page continues 1533 * to be in the object. If the page doesn't exist, allocate it. 1534 * 1535 * This routine may block. 1536 */ 1537vm_page_t 1538vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1539{ 1540 vm_page_t m; 1541 int s, generation; 1542 1543 GIANT_REQUIRED; 1544retrylookup: 1545 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1546 vm_page_lock_queues(); 1547 if (m->busy || (m->flags & PG_BUSY)) { 1548 generation = object->generation; 1549 1550 s = splvm(); 1551 while ((object->generation == generation) && 1552 (m->busy || (m->flags & PG_BUSY))) { 1553 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1554 msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0); 1555 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1556 vm_page_unlock_queues(); 1557 splx(s); 1558 return NULL; 1559 } 1560 } 1561 vm_page_unlock_queues(); 1562 splx(s); 1563 goto retrylookup; 1564 } else { 1565 if (allocflags & VM_ALLOC_WIRED) 1566 vm_page_wire(m); 1567 vm_page_busy(m); 1568 vm_page_unlock_queues(); 1569 return m; 1570 } 1571 } 1572 1573 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1574 if (m == NULL) { 1575 VM_WAIT; 1576 if ((allocflags & VM_ALLOC_RETRY) == 0) 1577 return NULL; 1578 goto retrylookup; 1579 } 1580 1581 return m; 1582} 1583 1584/* 1585 * Mapping function for valid bits or for dirty bits in 1586 * a page. May not block. 1587 * 1588 * Inputs are required to range within a page. 1589 */ 1590__inline int 1591vm_page_bits(int base, int size) 1592{ 1593 int first_bit; 1594 int last_bit; 1595 1596 KASSERT( 1597 base + size <= PAGE_SIZE, 1598 ("vm_page_bits: illegal base/size %d/%d", base, size) 1599 ); 1600 1601 if (size == 0) /* handle degenerate case */ 1602 return (0); 1603 1604 first_bit = base >> DEV_BSHIFT; 1605 last_bit = (base + size - 1) >> DEV_BSHIFT; 1606 1607 return ((2 << last_bit) - (1 << first_bit)); 1608} 1609 1610/* 1611 * vm_page_set_validclean: 1612 * 1613 * Sets portions of a page valid and clean. The arguments are expected 1614 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1615 * of any partial chunks touched by the range. The invalid portion of 1616 * such chunks will be zero'd. 1617 * 1618 * This routine may not block. 1619 * 1620 * (base + size) must be less then or equal to PAGE_SIZE. 1621 */ 1622void 1623vm_page_set_validclean(vm_page_t m, int base, int size) 1624{ 1625 int pagebits; 1626 int frag; 1627 int endoff; 1628 1629 GIANT_REQUIRED; 1630 if (size == 0) /* handle degenerate case */ 1631 return; 1632 1633 /* 1634 * If the base is not DEV_BSIZE aligned and the valid 1635 * bit is clear, we have to zero out a portion of the 1636 * first block. 1637 */ 1638 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1639 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1640 pmap_zero_page_area(m, frag, base - frag); 1641 1642 /* 1643 * If the ending offset is not DEV_BSIZE aligned and the 1644 * valid bit is clear, we have to zero out a portion of 1645 * the last block. 1646 */ 1647 endoff = base + size; 1648 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1649 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1650 pmap_zero_page_area(m, endoff, 1651 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1652 1653 /* 1654 * Set valid, clear dirty bits. If validating the entire 1655 * page we can safely clear the pmap modify bit. We also 1656 * use this opportunity to clear the PG_NOSYNC flag. If a process 1657 * takes a write fault on a MAP_NOSYNC memory area the flag will 1658 * be set again. 1659 * 1660 * We set valid bits inclusive of any overlap, but we can only 1661 * clear dirty bits for DEV_BSIZE chunks that are fully within 1662 * the range. 1663 */ 1664 pagebits = vm_page_bits(base, size); 1665 m->valid |= pagebits; 1666#if 0 /* NOT YET */ 1667 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1668 frag = DEV_BSIZE - frag; 1669 base += frag; 1670 size -= frag; 1671 if (size < 0) 1672 size = 0; 1673 } 1674 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1675#endif 1676 m->dirty &= ~pagebits; 1677 if (base == 0 && size == PAGE_SIZE) { 1678 pmap_clear_modify(m); 1679 vm_page_flag_clear(m, PG_NOSYNC); 1680 } 1681} 1682 1683#if 0 1684 1685void 1686vm_page_set_dirty(vm_page_t m, int base, int size) 1687{ 1688 m->dirty |= vm_page_bits(base, size); 1689} 1690 1691#endif 1692 1693void 1694vm_page_clear_dirty(vm_page_t m, int base, int size) 1695{ 1696 GIANT_REQUIRED; 1697 m->dirty &= ~vm_page_bits(base, size); 1698} 1699 1700/* 1701 * vm_page_set_invalid: 1702 * 1703 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1704 * valid and dirty bits for the effected areas are cleared. 1705 * 1706 * May not block. 1707 */ 1708void 1709vm_page_set_invalid(vm_page_t m, int base, int size) 1710{ 1711 int bits; 1712 1713 GIANT_REQUIRED; 1714 bits = vm_page_bits(base, size); 1715 m->valid &= ~bits; 1716 m->dirty &= ~bits; 1717 m->object->generation++; 1718} 1719 1720/* 1721 * vm_page_zero_invalid() 1722 * 1723 * The kernel assumes that the invalid portions of a page contain 1724 * garbage, but such pages can be mapped into memory by user code. 1725 * When this occurs, we must zero out the non-valid portions of the 1726 * page so user code sees what it expects. 1727 * 1728 * Pages are most often semi-valid when the end of a file is mapped 1729 * into memory and the file's size is not page aligned. 1730 */ 1731void 1732vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1733{ 1734 int b; 1735 int i; 1736 1737 /* 1738 * Scan the valid bits looking for invalid sections that 1739 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1740 * valid bit may be set ) have already been zerod by 1741 * vm_page_set_validclean(). 1742 */ 1743 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1744 if (i == (PAGE_SIZE / DEV_BSIZE) || 1745 (m->valid & (1 << i)) 1746 ) { 1747 if (i > b) { 1748 pmap_zero_page_area(m, 1749 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1750 } 1751 b = i + 1; 1752 } 1753 } 1754 1755 /* 1756 * setvalid is TRUE when we can safely set the zero'd areas 1757 * as being valid. We can do this if there are no cache consistancy 1758 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1759 */ 1760 if (setvalid) 1761 m->valid = VM_PAGE_BITS_ALL; 1762} 1763 1764/* 1765 * vm_page_is_valid: 1766 * 1767 * Is (partial) page valid? Note that the case where size == 0 1768 * will return FALSE in the degenerate case where the page is 1769 * entirely invalid, and TRUE otherwise. 1770 * 1771 * May not block. 1772 */ 1773int 1774vm_page_is_valid(vm_page_t m, int base, int size) 1775{ 1776 int bits = vm_page_bits(base, size); 1777 1778 if (m->valid && ((m->valid & bits) == bits)) 1779 return 1; 1780 else 1781 return 0; 1782} 1783 1784/* 1785 * update dirty bits from pmap/mmu. May not block. 1786 */ 1787void 1788vm_page_test_dirty(vm_page_t m) 1789{ 1790 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1791 vm_page_dirty(m); 1792 } 1793} 1794 1795int so_zerocp_fullpage = 0; 1796 1797void 1798vm_page_cowfault(vm_page_t m) 1799{ 1800 vm_page_t mnew; 1801 vm_object_t object; 1802 vm_pindex_t pindex; 1803 1804 object = m->object; 1805 pindex = m->pindex; 1806 vm_page_busy(m); 1807 1808 retry_alloc: 1809 vm_page_remove(m); 1810 /* 1811 * An interrupt allocation is requested because the page 1812 * queues lock is held. 1813 */ 1814 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1815 if (mnew == NULL) { 1816 vm_page_insert(m, object, pindex); 1817 vm_page_unlock_queues(); 1818 VM_WAIT; 1819 vm_page_lock_queues(); 1820 goto retry_alloc; 1821 } 1822 1823 if (m->cow == 0) { 1824 /* 1825 * check to see if we raced with an xmit complete when 1826 * waiting to allocate a page. If so, put things back 1827 * the way they were 1828 */ 1829 vm_page_busy(mnew); 1830 vm_page_free(mnew); 1831 vm_page_insert(m, object, pindex); 1832 } else { /* clear COW & copy page */ 1833 if (so_zerocp_fullpage) { 1834 mnew->valid = VM_PAGE_BITS_ALL; 1835 } else { 1836 vm_page_copy(m, mnew); 1837 } 1838 vm_page_dirty(mnew); 1839 vm_page_flag_clear(mnew, PG_BUSY); 1840 } 1841} 1842 1843void 1844vm_page_cowclear(vm_page_t m) 1845{ 1846 1847 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1848 if (m->cow) { 1849 m->cow--; 1850 /* 1851 * let vm_fault add back write permission lazily 1852 */ 1853 } 1854 /* 1855 * sf_buf_free() will free the page, so we needn't do it here 1856 */ 1857} 1858 1859void 1860vm_page_cowsetup(vm_page_t m) 1861{ 1862 1863 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1864 m->cow++; 1865 vm_page_protect(m, VM_PROT_READ); 1866} 1867 1868#include "opt_ddb.h" 1869#ifdef DDB 1870#include <sys/kernel.h> 1871 1872#include <ddb/ddb.h> 1873 1874DB_SHOW_COMMAND(page, vm_page_print_page_info) 1875{ 1876 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1877 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1878 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1879 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1880 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1881 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1882 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1883 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1884 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1885 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1886} 1887 1888DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1889{ 1890 int i; 1891 db_printf("PQ_FREE:"); 1892 for (i = 0; i < PQ_L2_SIZE; i++) { 1893 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1894 } 1895 db_printf("\n"); 1896 1897 db_printf("PQ_CACHE:"); 1898 for (i = 0; i < PQ_L2_SIZE; i++) { 1899 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1900 } 1901 db_printf("\n"); 1902 1903 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1904 vm_page_queues[PQ_ACTIVE].lcnt, 1905 vm_page_queues[PQ_INACTIVE].lcnt); 1906} 1907#endif /* DDB */ 1908