vm_page.c revision 95532
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 95532 2002-04-26 22:44:15Z 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 mtx vm_page_buckets_mtx; 129static struct vm_page **vm_page_buckets; /* Array of buckets */ 130static int vm_page_bucket_count; /* How big is array? */ 131static int vm_page_hash_mask; /* Mask for hash function */ 132 133vm_page_t vm_page_array = 0; 134int vm_page_array_size = 0; 135long first_page = 0; 136int vm_page_zero_count = 0; 137 138/* 139 * vm_set_page_size: 140 * 141 * Sets the page size, perhaps based upon the memory 142 * size. Must be called before any use of page-size 143 * dependent functions. 144 */ 145void 146vm_set_page_size(void) 147{ 148 if (cnt.v_page_size == 0) 149 cnt.v_page_size = PAGE_SIZE; 150 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 151 panic("vm_set_page_size: page size not a power of two"); 152} 153 154/* 155 * vm_page_startup: 156 * 157 * Initializes the resident memory module. 158 * 159 * Allocates memory for the page cells, and 160 * for the object/offset-to-page hash table headers. 161 * Each page cell is initialized and placed on the free list. 162 */ 163vm_offset_t 164vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 165{ 166 vm_offset_t mapped; 167 struct vm_page **bucket; 168 vm_size_t npages, page_range; 169 vm_offset_t new_end; 170 int i; 171 vm_offset_t pa; 172 int nblocks; 173 vm_offset_t last_pa; 174 175 /* the biggest memory array is the second group of pages */ 176 vm_offset_t end; 177 vm_offset_t biggestone, biggestsize; 178 179 vm_offset_t total; 180 vm_size_t bootpages; 181 182 total = 0; 183 biggestsize = 0; 184 biggestone = 0; 185 nblocks = 0; 186 vaddr = round_page(vaddr); 187 188 for (i = 0; phys_avail[i + 1]; i += 2) { 189 phys_avail[i] = round_page(phys_avail[i]); 190 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 191 } 192 193 for (i = 0; phys_avail[i + 1]; i += 2) { 194 vm_size_t size = phys_avail[i + 1] - phys_avail[i]; 195 196 if (size > biggestsize) { 197 biggestone = i; 198 biggestsize = size; 199 } 200 ++nblocks; 201 total += size; 202 } 203 204 end = phys_avail[biggestone+1]; 205 206 /* 207 * Initialize the queue headers for the free queue, the active queue 208 * and the inactive queue. 209 */ 210 vm_pageq_init(); 211 212 /* 213 * Allocate memory for use when boot strapping the kernel memory allocator 214 */ 215 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 216 new_end = end - bootpages; 217 new_end = trunc_page(new_end); 218 mapped = pmap_map(&vaddr, new_end, end, 219 VM_PROT_READ | VM_PROT_WRITE); 220 bzero((caddr_t) mapped, end - new_end); 221 uma_startup((caddr_t)mapped); 222 223 end = new_end; 224 225 /* 226 * Allocate (and initialize) the hash table buckets. 227 * 228 * The number of buckets MUST BE a power of 2, and the actual value is 229 * the next power of 2 greater than the number of physical pages in 230 * the system. 231 * 232 * We make the hash table approximately 2x the number of pages to 233 * reduce the chain length. This is about the same size using the 234 * singly-linked list as the 1x hash table we were using before 235 * using TAILQ but the chain length will be smaller. 236 * 237 * Note: This computation can be tweaked if desired. 238 */ 239 if (vm_page_bucket_count == 0) { 240 vm_page_bucket_count = 1; 241 while (vm_page_bucket_count < atop(total)) 242 vm_page_bucket_count <<= 1; 243 } 244 vm_page_bucket_count <<= 1; 245 vm_page_hash_mask = vm_page_bucket_count - 1; 246 247 /* 248 * Validate these addresses. 249 */ 250 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 251 new_end = trunc_page(new_end); 252 mapped = pmap_map(&vaddr, new_end, end, 253 VM_PROT_READ | VM_PROT_WRITE); 254 bzero((caddr_t) mapped, end - new_end); 255 256 mtx_init(&vm_page_buckets_mtx, "vm page buckets mutex", NULL, MTX_DEF); 257 vm_page_buckets = (struct vm_page **)mapped; 258 bucket = vm_page_buckets; 259 for (i = 0; i < vm_page_bucket_count; i++) { 260 *bucket = NULL; 261 bucket++; 262 } 263 264 /* 265 * Compute the number of pages of memory that will be available for 266 * use (taking into account the overhead of a page structure per 267 * page). 268 */ 269 first_page = phys_avail[0] / PAGE_SIZE; 270 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 271 npages = (total - (page_range * sizeof(struct vm_page)) - 272 (end - new_end)) / PAGE_SIZE; 273 end = new_end; 274 275 /* 276 * Initialize the mem entry structures now, and put them in the free 277 * queue. 278 */ 279 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 280 mapped = pmap_map(&vaddr, new_end, end, 281 VM_PROT_READ | VM_PROT_WRITE); 282 vm_page_array = (vm_page_t) mapped; 283 284 /* 285 * Clear all of the page structures 286 */ 287 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 288 vm_page_array_size = page_range; 289 290 /* 291 * Construct the free queue(s) in descending order (by physical 292 * address) so that the first 16MB of physical memory is allocated 293 * last rather than first. On large-memory machines, this avoids 294 * the exhaustion of low physical memory before isa_dmainit has run. 295 */ 296 cnt.v_page_count = 0; 297 cnt.v_free_count = 0; 298 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 299 pa = phys_avail[i]; 300 if (i == biggestone) 301 last_pa = new_end; 302 else 303 last_pa = phys_avail[i + 1]; 304 while (pa < last_pa && npages-- > 0) { 305 vm_pageq_add_new_page(pa); 306 pa += PAGE_SIZE; 307 } 308 } 309 return (vaddr); 310} 311 312/* 313 * vm_page_hash: 314 * 315 * Distributes the object/offset key pair among hash buckets. 316 * 317 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 318 * This routine may not block. 319 * 320 * We try to randomize the hash based on the object to spread the pages 321 * out in the hash table without it costing us too much. 322 */ 323static __inline int 324vm_page_hash(vm_object_t object, vm_pindex_t pindex) 325{ 326 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 327 328 return (i & vm_page_hash_mask); 329} 330 331void 332vm_page_flag_set(vm_page_t m, unsigned short bits) 333{ 334 GIANT_REQUIRED; 335 m->flags |= bits; 336} 337 338void 339vm_page_flag_clear(vm_page_t m, unsigned short bits) 340{ 341 GIANT_REQUIRED; 342 m->flags &= ~bits; 343} 344 345void 346vm_page_busy(vm_page_t m) 347{ 348 KASSERT((m->flags & PG_BUSY) == 0, 349 ("vm_page_busy: page already busy!!!")); 350 vm_page_flag_set(m, PG_BUSY); 351} 352 353/* 354 * vm_page_flash: 355 * 356 * wakeup anyone waiting for the page. 357 */ 358void 359vm_page_flash(vm_page_t m) 360{ 361 if (m->flags & PG_WANTED) { 362 vm_page_flag_clear(m, PG_WANTED); 363 wakeup(m); 364 } 365} 366 367/* 368 * vm_page_wakeup: 369 * 370 * clear the PG_BUSY flag and wakeup anyone waiting for the 371 * page. 372 * 373 */ 374void 375vm_page_wakeup(vm_page_t m) 376{ 377 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 378 vm_page_flag_clear(m, PG_BUSY); 379 vm_page_flash(m); 380} 381 382/* 383 * 384 * 385 */ 386void 387vm_page_io_start(vm_page_t m) 388{ 389 GIANT_REQUIRED; 390 m->busy++; 391} 392 393void 394vm_page_io_finish(vm_page_t m) 395{ 396 GIANT_REQUIRED; 397 m->busy--; 398 if (m->busy == 0) 399 vm_page_flash(m); 400} 401 402/* 403 * Keep page from being freed by the page daemon 404 * much of the same effect as wiring, except much lower 405 * overhead and should be used only for *very* temporary 406 * holding ("wiring"). 407 */ 408void 409vm_page_hold(vm_page_t mem) 410{ 411 GIANT_REQUIRED; 412 mem->hold_count++; 413} 414 415void 416vm_page_unhold(vm_page_t mem) 417{ 418 GIANT_REQUIRED; 419 --mem->hold_count; 420 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 421 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 422 vm_page_free_toq(mem); 423} 424 425/* 426 * vm_page_protect: 427 * 428 * Reduce the protection of a page. This routine never raises the 429 * protection and therefore can be safely called if the page is already 430 * at VM_PROT_NONE (it will be a NOP effectively ). 431 */ 432void 433vm_page_protect(vm_page_t mem, int prot) 434{ 435 if (prot == VM_PROT_NONE) { 436 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) { 437 pmap_page_protect(mem, VM_PROT_NONE); 438 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED); 439 } 440 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 441 pmap_page_protect(mem, VM_PROT_READ); 442 vm_page_flag_clear(mem, PG_WRITEABLE); 443 } 444} 445/* 446 * vm_page_zero_fill: 447 * 448 * Zero-fill the specified page. 449 * Written as a standard pagein routine, to 450 * be used by the zero-fill object. 451 */ 452boolean_t 453vm_page_zero_fill(vm_page_t m) 454{ 455 pmap_zero_page(m); 456 return (TRUE); 457} 458 459/* 460 * vm_page_copy: 461 * 462 * Copy one page to another 463 */ 464void 465vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 466{ 467 pmap_copy_page(src_m, dest_m); 468 dest_m->valid = VM_PAGE_BITS_ALL; 469} 470 471/* 472 * vm_page_free: 473 * 474 * Free a page 475 * 476 * The clearing of PG_ZERO is a temporary safety until the code can be 477 * reviewed to determine that PG_ZERO is being properly cleared on 478 * write faults or maps. PG_ZERO was previously cleared in 479 * vm_page_alloc(). 480 */ 481void 482vm_page_free(vm_page_t m) 483{ 484 vm_page_flag_clear(m, PG_ZERO); 485 vm_page_free_toq(m); 486 vm_page_zero_idle_wakeup(); 487} 488 489/* 490 * vm_page_free_zero: 491 * 492 * Free a page to the zerod-pages queue 493 */ 494void 495vm_page_free_zero(vm_page_t m) 496{ 497 vm_page_flag_set(m, PG_ZERO); 498 vm_page_free_toq(m); 499} 500 501/* 502 * vm_page_sleep_busy: 503 * 504 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 505 * m->busy is zero. Returns TRUE if it had to sleep ( including if 506 * it almost had to sleep and made temporary spl*() mods), FALSE 507 * otherwise. 508 * 509 * This routine assumes that interrupts can only remove the busy 510 * status from a page, not set the busy status or change it from 511 * PG_BUSY to m->busy or vise versa (which would create a timing 512 * window). 513 */ 514int 515vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 516{ 517 GIANT_REQUIRED; 518 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 519 int s = splvm(); 520 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 521 /* 522 * Page is busy. Wait and retry. 523 */ 524 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 525 tsleep(m, PVM, msg, 0); 526 } 527 splx(s); 528 return (TRUE); 529 /* not reached */ 530 } 531 return (FALSE); 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_undirty: 548 * 549 * Set page to not be dirty. Note: does not clear pmap modify bits 550 */ 551void 552vm_page_undirty(vm_page_t m) 553{ 554 m->dirty = 0; 555} 556 557/* 558 * vm_page_insert: [ internal use only ] 559 * 560 * Inserts the given mem entry into the object and object list. 561 * 562 * The pagetables are not updated but will presumably fault the page 563 * in if necessary, or if a kernel page the caller will at some point 564 * enter the page into the kernel's pmap. We are not allowed to block 565 * here so we *can't* do this anyway. 566 * 567 * The object and page must be locked, and must be splhigh. 568 * This routine may not block. 569 */ 570void 571vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 572{ 573 struct vm_page **bucket; 574 575 GIANT_REQUIRED; 576 577 if (m->object != NULL) 578 panic("vm_page_insert: already inserted"); 579 580 /* 581 * Record the object/offset pair in this page 582 */ 583 m->object = object; 584 m->pindex = pindex; 585 586 /* 587 * Insert it into the object_object/offset hash table 588 */ 589 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 590 mtx_lock(&vm_page_buckets_mtx); 591 m->hnext = *bucket; 592 *bucket = m; 593 mtx_unlock(&vm_page_buckets_mtx); 594 595 /* 596 * Now link into the object's list of backed pages. 597 */ 598 TAILQ_INSERT_TAIL(&object->memq, m, listq); 599 object->generation++; 600 601 /* 602 * show that the object has one more resident page. 603 */ 604 object->resident_page_count++; 605 606 /* 607 * Since we are inserting a new and possibly dirty page, 608 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 609 */ 610 if (m->flags & PG_WRITEABLE) 611 vm_object_set_writeable_dirty(object); 612} 613 614/* 615 * vm_page_remove: 616 * NOTE: used by device pager as well -wfj 617 * 618 * Removes the given mem entry from the object/offset-page 619 * table and the object page list, but do not invalidate/terminate 620 * the backing store. 621 * 622 * The object and page must be locked, and at splhigh. 623 * The underlying pmap entry (if any) is NOT removed here. 624 * This routine may not block. 625 */ 626void 627vm_page_remove(vm_page_t m) 628{ 629 vm_object_t object; 630 vm_page_t *bucket; 631 632 GIANT_REQUIRED; 633 634 if (m->object == NULL) 635 return; 636 637 if ((m->flags & PG_BUSY) == 0) { 638 panic("vm_page_remove: page not busy"); 639 } 640 641 /* 642 * Basically destroy the page. 643 */ 644 vm_page_wakeup(m); 645 646 object = m->object; 647 648 /* 649 * Remove from the object_object/offset hash table. The object 650 * must be on the hash queue, we will panic if it isn't 651 */ 652 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 653 mtx_lock(&vm_page_buckets_mtx); 654 while (*bucket != m) { 655 if (*bucket == NULL) 656 panic("vm_page_remove(): page not found in hash"); 657 bucket = &(*bucket)->hnext; 658 } 659 *bucket = m->hnext; 660 m->hnext = NULL; 661 mtx_unlock(&vm_page_buckets_mtx); 662 663 /* 664 * Now remove from the object's list of backed pages. 665 */ 666 TAILQ_REMOVE(&object->memq, m, listq); 667 668 /* 669 * And show that the object has one fewer resident page. 670 */ 671 object->resident_page_count--; 672 object->generation++; 673 674 m->object = NULL; 675} 676 677/* 678 * vm_page_lookup: 679 * 680 * Returns the page associated with the object/offset 681 * pair specified; if none is found, NULL is returned. 682 * 683 * The object must be locked. No side effects. 684 * This routine may not block. 685 * This is a critical path routine 686 */ 687vm_page_t 688vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 689{ 690 vm_page_t m; 691 struct vm_page **bucket; 692 693 /* 694 * Search the hash table for this object/offset pair 695 */ 696 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 697 mtx_lock(&vm_page_buckets_mtx); 698 for (m = *bucket; m != NULL; m = m->hnext) { 699 if ((m->object == object) && (m->pindex == pindex)) { 700 mtx_unlock(&vm_page_buckets_mtx); 701 return (m); 702 } 703 } 704 mtx_unlock(&vm_page_buckets_mtx); 705 return (NULL); 706} 707 708/* 709 * vm_page_rename: 710 * 711 * Move the given memory entry from its 712 * current object to the specified target object/offset. 713 * 714 * The object must be locked. 715 * This routine may not block. 716 * 717 * Note: this routine will raise itself to splvm(), the caller need not. 718 * 719 * Note: swap associated with the page must be invalidated by the move. We 720 * have to do this for several reasons: (1) we aren't freeing the 721 * page, (2) we are dirtying the page, (3) the VM system is probably 722 * moving the page from object A to B, and will then later move 723 * the backing store from A to B and we can't have a conflict. 724 * 725 * Note: we *always* dirty the page. It is necessary both for the 726 * fact that we moved it, and because we may be invalidating 727 * swap. If the page is on the cache, we have to deactivate it 728 * or vm_page_dirty() will panic. Dirty pages are not allowed 729 * on the cache. 730 */ 731void 732vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 733{ 734 int s; 735 736 s = splvm(); 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 splx(s); 743} 744 745/* 746 * vm_page_select_cache: 747 * 748 * Find a page on the cache queue with color optimization. As pages 749 * might be found, but not applicable, they are deactivated. This 750 * keeps us from using potentially busy cached pages. 751 * 752 * This routine must be called at splvm(). 753 * This routine may not block. 754 */ 755static vm_page_t 756vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 757{ 758 vm_page_t m; 759 760 GIANT_REQUIRED; 761 while (TRUE) { 762 m = vm_pageq_find( 763 PQ_CACHE, 764 (pindex + object->pg_color) & PQ_L2_MASK, 765 FALSE 766 ); 767 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 768 m->hold_count || m->wire_count)) { 769 vm_page_deactivate(m); 770 continue; 771 } 772 return m; 773 } 774} 775 776/* 777 * vm_page_select_free: 778 * 779 * Find a free or zero page, with specified preference. 780 * 781 * This routine must be called at splvm(). 782 * This routine may not block. 783 */ 784static __inline vm_page_t 785vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 786{ 787 vm_page_t m; 788 789 m = vm_pageq_find( 790 PQ_FREE, 791 (pindex + object->pg_color) & PQ_L2_MASK, 792 prefer_zero 793 ); 794 return (m); 795} 796 797/* 798 * vm_page_alloc: 799 * 800 * Allocate and return a memory cell associated 801 * with this VM object/offset pair. 802 * 803 * page_req classes: 804 * VM_ALLOC_NORMAL normal process request 805 * VM_ALLOC_SYSTEM system *really* needs a page 806 * VM_ALLOC_INTERRUPT interrupt time request 807 * VM_ALLOC_ZERO zero page 808 * 809 * This routine may not block. 810 * 811 * Additional special handling is required when called from an 812 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 813 * the page cache in this case. 814 */ 815vm_page_t 816vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 817{ 818 vm_page_t m = NULL; 819 int s; 820 821 GIANT_REQUIRED; 822 823 KASSERT(!vm_page_lookup(object, pindex), 824 ("vm_page_alloc: page already allocated")); 825 826 /* 827 * The pager is allowed to eat deeper into the free page list. 828 */ 829 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 830 page_req = VM_ALLOC_SYSTEM; 831 }; 832 833 s = splvm(); 834 835loop: 836 if (cnt.v_free_count > cnt.v_free_reserved) { 837 /* 838 * Allocate from the free queue if there are plenty of pages 839 * in it. 840 */ 841 if (page_req == VM_ALLOC_ZERO) 842 m = vm_page_select_free(object, pindex, TRUE); 843 else 844 m = vm_page_select_free(object, pindex, FALSE); 845 } else if ( 846 (page_req == VM_ALLOC_SYSTEM && 847 cnt.v_cache_count == 0 && 848 cnt.v_free_count > cnt.v_interrupt_free_min) || 849 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 850 ) { 851 /* 852 * Interrupt or system, dig deeper into the free list. 853 */ 854 m = vm_page_select_free(object, pindex, FALSE); 855 } else if (page_req != VM_ALLOC_INTERRUPT) { 856 /* 857 * Allocatable from cache (non-interrupt only). On success, 858 * we must free the page and try again, thus ensuring that 859 * cnt.v_*_free_min counters are replenished. 860 */ 861 m = vm_page_select_cache(object, pindex); 862 if (m == NULL) { 863 splx(s); 864#if defined(DIAGNOSTIC) 865 if (cnt.v_cache_count > 0) 866 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 867#endif 868 vm_pageout_deficit++; 869 pagedaemon_wakeup(); 870 return (NULL); 871 } 872 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 873 vm_page_busy(m); 874 vm_page_protect(m, VM_PROT_NONE); 875 vm_page_free(m); 876 goto loop; 877 } else { 878 /* 879 * Not allocatable from cache from interrupt, give up. 880 */ 881 splx(s); 882 vm_pageout_deficit++; 883 pagedaemon_wakeup(); 884 return (NULL); 885 } 886 887 /* 888 * At this point we had better have found a good page. 889 */ 890 891 KASSERT( 892 m != NULL, 893 ("vm_page_alloc(): missing page on free queue\n") 894 ); 895 896 /* 897 * Remove from free queue 898 */ 899 900 vm_pageq_remove_nowakeup(m); 901 902 /* 903 * Initialize structure. Only the PG_ZERO flag is inherited. 904 */ 905 if (m->flags & PG_ZERO) { 906 vm_page_zero_count--; 907 m->flags = PG_ZERO | PG_BUSY; 908 } else { 909 m->flags = PG_BUSY; 910 } 911 m->wire_count = 0; 912 m->hold_count = 0; 913 m->act_count = 0; 914 m->busy = 0; 915 m->valid = 0; 916 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 917 918 /* 919 * vm_page_insert() is safe prior to the splx(). Note also that 920 * inserting a page here does not insert it into the pmap (which 921 * could cause us to block allocating memory). We cannot block 922 * anywhere. 923 */ 924 vm_page_insert(m, object, pindex); 925 926 /* 927 * Don't wakeup too often - wakeup the pageout daemon when 928 * we would be nearly out of memory. 929 */ 930 if (vm_paging_needed()) 931 pagedaemon_wakeup(); 932 933 splx(s); 934 return (m); 935} 936 937/* 938 * vm_wait: (also see VM_WAIT macro) 939 * 940 * Block until free pages are available for allocation 941 * - Called in various places before memory allocations. 942 */ 943void 944vm_wait(void) 945{ 946 int s; 947 948 s = splvm(); 949 if (curproc == pageproc) { 950 vm_pageout_pages_needed = 1; 951 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 952 } else { 953 if (!vm_pages_needed) { 954 vm_pages_needed = 1; 955 wakeup(&vm_pages_needed); 956 } 957 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 958 } 959 splx(s); 960} 961 962/* 963 * vm_waitpfault: (also see VM_WAITPFAULT macro) 964 * 965 * Block until free pages are available for allocation 966 * - Called only in vm_fault so that processes page faulting 967 * can be easily tracked. 968 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 969 * processes will be able to grab memory first. Do not change 970 * this balance without careful testing first. 971 */ 972void 973vm_waitpfault(void) 974{ 975 int s; 976 977 s = splvm(); 978 if (!vm_pages_needed) { 979 vm_pages_needed = 1; 980 wakeup(&vm_pages_needed); 981 } 982 tsleep(&cnt.v_free_count, PUSER, "pfault", 0); 983 splx(s); 984} 985 986/* 987 * vm_page_activate: 988 * 989 * Put the specified page on the active list (if appropriate). 990 * Ensure that act_count is at least ACT_INIT but do not otherwise 991 * mess with it. 992 * 993 * The page queues must be locked. 994 * This routine may not block. 995 */ 996void 997vm_page_activate(vm_page_t m) 998{ 999 int s; 1000 1001 GIANT_REQUIRED; 1002 s = splvm(); 1003 if (m->queue != PQ_ACTIVE) { 1004 if ((m->queue - m->pc) == PQ_CACHE) 1005 cnt.v_reactivated++; 1006 vm_pageq_remove(m); 1007 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1008 if (m->act_count < ACT_INIT) 1009 m->act_count = ACT_INIT; 1010 vm_pageq_enqueue(PQ_ACTIVE, m); 1011 } 1012 } else { 1013 if (m->act_count < ACT_INIT) 1014 m->act_count = ACT_INIT; 1015 } 1016 splx(s); 1017} 1018 1019/* 1020 * vm_page_free_wakeup: 1021 * 1022 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1023 * routine is called when a page has been added to the cache or free 1024 * queues. 1025 * 1026 * This routine may not block. 1027 * This routine must be called at splvm() 1028 */ 1029static __inline void 1030vm_page_free_wakeup(void) 1031{ 1032 /* 1033 * if pageout daemon needs pages, then tell it that there are 1034 * some free. 1035 */ 1036 if (vm_pageout_pages_needed && 1037 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1038 wakeup(&vm_pageout_pages_needed); 1039 vm_pageout_pages_needed = 0; 1040 } 1041 /* 1042 * wakeup processes that are waiting on memory if we hit a 1043 * high water mark. And wakeup scheduler process if we have 1044 * lots of memory. this process will swapin processes. 1045 */ 1046 if (vm_pages_needed && !vm_page_count_min()) { 1047 vm_pages_needed = 0; 1048 wakeup(&cnt.v_free_count); 1049 } 1050} 1051 1052/* 1053 * vm_page_free_toq: 1054 * 1055 * Returns the given page to the PQ_FREE list, 1056 * disassociating it with any VM object. 1057 * 1058 * Object and page must be locked prior to entry. 1059 * This routine may not block. 1060 */ 1061 1062void 1063vm_page_free_toq(vm_page_t m) 1064{ 1065 int s; 1066 struct vpgqueues *pq; 1067 vm_object_t object = m->object; 1068 1069 GIANT_REQUIRED; 1070 s = splvm(); 1071 cnt.v_tfree++; 1072 1073 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1074 printf( 1075 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1076 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1077 m->hold_count); 1078 if ((m->queue - m->pc) == PQ_FREE) 1079 panic("vm_page_free: freeing free page"); 1080 else 1081 panic("vm_page_free: freeing busy page"); 1082 } 1083 1084 /* 1085 * unqueue, then remove page. Note that we cannot destroy 1086 * the page here because we do not want to call the pager's 1087 * callback routine until after we've put the page on the 1088 * appropriate free queue. 1089 */ 1090 vm_pageq_remove_nowakeup(m); 1091 vm_page_remove(m); 1092 1093 /* 1094 * If fictitious remove object association and 1095 * return, otherwise delay object association removal. 1096 */ 1097 if ((m->flags & PG_FICTITIOUS) != 0) { 1098 splx(s); 1099 return; 1100 } 1101 1102 m->valid = 0; 1103 vm_page_undirty(m); 1104 1105 if (m->wire_count != 0) { 1106 if (m->wire_count > 1) { 1107 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1108 m->wire_count, (long)m->pindex); 1109 } 1110 panic("vm_page_free: freeing wired page\n"); 1111 } 1112 1113 /* 1114 * If we've exhausted the object's resident pages we want to free 1115 * it up. 1116 */ 1117 if (object && 1118 (object->type == OBJT_VNODE) && 1119 ((object->flags & OBJ_DEAD) == 0) 1120 ) { 1121 struct vnode *vp = (struct vnode *)object->handle; 1122 1123 if (vp && VSHOULDFREE(vp)) 1124 vfree(vp); 1125 } 1126 1127 /* 1128 * Clear the UNMANAGED flag when freeing an unmanaged page. 1129 */ 1130 if (m->flags & PG_UNMANAGED) { 1131 m->flags &= ~PG_UNMANAGED; 1132 } else { 1133#ifdef __alpha__ 1134 pmap_page_is_free(m); 1135#endif 1136 } 1137 1138 if (m->hold_count != 0) { 1139 m->flags &= ~PG_ZERO; 1140 m->queue = PQ_HOLD; 1141 } else 1142 m->queue = PQ_FREE + m->pc; 1143 pq = &vm_page_queues[m->queue]; 1144 pq->lcnt++; 1145 ++(*pq->cnt); 1146 1147 /* 1148 * Put zero'd pages on the end ( where we look for zero'd pages 1149 * first ) and non-zerod pages at the head. 1150 */ 1151 if (m->flags & PG_ZERO) { 1152 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1153 ++vm_page_zero_count; 1154 } else { 1155 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1156 } 1157 vm_page_free_wakeup(); 1158 splx(s); 1159} 1160 1161/* 1162 * vm_page_unmanage: 1163 * 1164 * Prevent PV management from being done on the page. The page is 1165 * removed from the paging queues as if it were wired, and as a 1166 * consequence of no longer being managed the pageout daemon will not 1167 * touch it (since there is no way to locate the pte mappings for the 1168 * page). madvise() calls that mess with the pmap will also no longer 1169 * operate on the page. 1170 * 1171 * Beyond that the page is still reasonably 'normal'. Freeing the page 1172 * will clear the flag. 1173 * 1174 * This routine is used by OBJT_PHYS objects - objects using unswappable 1175 * physical memory as backing store rather then swap-backed memory and 1176 * will eventually be extended to support 4MB unmanaged physical 1177 * mappings. 1178 */ 1179void 1180vm_page_unmanage(vm_page_t m) 1181{ 1182 int s; 1183 1184 s = splvm(); 1185 if ((m->flags & PG_UNMANAGED) == 0) { 1186 if (m->wire_count == 0) 1187 vm_pageq_remove(m); 1188 } 1189 vm_page_flag_set(m, PG_UNMANAGED); 1190 splx(s); 1191} 1192 1193/* 1194 * vm_page_wire: 1195 * 1196 * Mark this page as wired down by yet 1197 * another map, removing it from paging queues 1198 * as necessary. 1199 * 1200 * The page queues must be locked. 1201 * This routine may not block. 1202 */ 1203void 1204vm_page_wire(vm_page_t m) 1205{ 1206 int s; 1207 1208 /* 1209 * Only bump the wire statistics if the page is not already wired, 1210 * and only unqueue the page if it is on some queue (if it is unmanaged 1211 * it is already off the queues). 1212 */ 1213 s = splvm(); 1214 if (m->wire_count == 0) { 1215 if ((m->flags & PG_UNMANAGED) == 0) 1216 vm_pageq_remove(m); 1217 cnt.v_wire_count++; 1218 } 1219 m->wire_count++; 1220 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1221 splx(s); 1222 vm_page_flag_set(m, PG_MAPPED); 1223} 1224 1225/* 1226 * vm_page_unwire: 1227 * 1228 * Release one wiring of this page, potentially 1229 * enabling it to be paged again. 1230 * 1231 * Many pages placed on the inactive queue should actually go 1232 * into the cache, but it is difficult to figure out which. What 1233 * we do instead, if the inactive target is well met, is to put 1234 * clean pages at the head of the inactive queue instead of the tail. 1235 * This will cause them to be moved to the cache more quickly and 1236 * if not actively re-referenced, freed more quickly. If we just 1237 * stick these pages at the end of the inactive queue, heavy filesystem 1238 * meta-data accesses can cause an unnecessary paging load on memory bound 1239 * processes. This optimization causes one-time-use metadata to be 1240 * reused more quickly. 1241 * 1242 * BUT, if we are in a low-memory situation we have no choice but to 1243 * put clean pages on the cache queue. 1244 * 1245 * A number of routines use vm_page_unwire() to guarantee that the page 1246 * will go into either the inactive or active queues, and will NEVER 1247 * be placed in the cache - for example, just after dirtying a page. 1248 * dirty pages in the cache are not allowed. 1249 * 1250 * The page queues must be locked. 1251 * This routine may not block. 1252 */ 1253void 1254vm_page_unwire(vm_page_t m, int activate) 1255{ 1256 int s; 1257 1258 s = splvm(); 1259 1260 if (m->wire_count > 0) { 1261 m->wire_count--; 1262 if (m->wire_count == 0) { 1263 cnt.v_wire_count--; 1264 if (m->flags & PG_UNMANAGED) { 1265 ; 1266 } else if (activate) 1267 vm_pageq_enqueue(PQ_ACTIVE, m); 1268 else { 1269 vm_page_flag_clear(m, PG_WINATCFLS); 1270 vm_pageq_enqueue(PQ_INACTIVE, m); 1271 } 1272 } 1273 } else { 1274 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1275 } 1276 splx(s); 1277} 1278 1279 1280/* 1281 * Move the specified page to the inactive queue. If the page has 1282 * any associated swap, the swap is deallocated. 1283 * 1284 * Normally athead is 0 resulting in LRU operation. athead is set 1285 * to 1 if we want this page to be 'as if it were placed in the cache', 1286 * except without unmapping it from the process address space. 1287 * 1288 * This routine may not block. 1289 */ 1290static __inline void 1291_vm_page_deactivate(vm_page_t m, int athead) 1292{ 1293 int s; 1294 1295 GIANT_REQUIRED; 1296 /* 1297 * Ignore if already inactive. 1298 */ 1299 if (m->queue == PQ_INACTIVE) 1300 return; 1301 1302 s = splvm(); 1303 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1304 if ((m->queue - m->pc) == PQ_CACHE) 1305 cnt.v_reactivated++; 1306 vm_page_flag_clear(m, PG_WINATCFLS); 1307 vm_pageq_remove(m); 1308 if (athead) 1309 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1310 else 1311 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1312 m->queue = PQ_INACTIVE; 1313 vm_page_queues[PQ_INACTIVE].lcnt++; 1314 cnt.v_inactive_count++; 1315 } 1316 splx(s); 1317} 1318 1319void 1320vm_page_deactivate(vm_page_t m) 1321{ 1322 _vm_page_deactivate(m, 0); 1323} 1324 1325/* 1326 * vm_page_try_to_cache: 1327 * 1328 * Returns 0 on failure, 1 on success 1329 */ 1330int 1331vm_page_try_to_cache(vm_page_t m) 1332{ 1333 GIANT_REQUIRED; 1334 1335 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1336 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1337 return (0); 1338 } 1339 vm_page_test_dirty(m); 1340 if (m->dirty) 1341 return (0); 1342 vm_page_cache(m); 1343 return (1); 1344} 1345 1346/* 1347 * vm_page_try_to_free() 1348 * 1349 * Attempt to free the page. If we cannot free it, we do nothing. 1350 * 1 is returned on success, 0 on failure. 1351 */ 1352int 1353vm_page_try_to_free(vm_page_t m) 1354{ 1355 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1356 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1357 return (0); 1358 } 1359 vm_page_test_dirty(m); 1360 if (m->dirty) 1361 return (0); 1362 vm_page_busy(m); 1363 vm_page_protect(m, VM_PROT_NONE); 1364 vm_page_free(m); 1365 return (1); 1366} 1367 1368/* 1369 * vm_page_cache 1370 * 1371 * Put the specified page onto the page cache queue (if appropriate). 1372 * 1373 * This routine may not block. 1374 */ 1375void 1376vm_page_cache(vm_page_t m) 1377{ 1378 int s; 1379 1380 GIANT_REQUIRED; 1381 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1382 printf("vm_page_cache: attempting to cache busy page\n"); 1383 return; 1384 } 1385 if ((m->queue - m->pc) == PQ_CACHE) 1386 return; 1387 1388 /* 1389 * Remove all pmaps and indicate that the page is not 1390 * writeable or mapped. 1391 */ 1392 vm_page_protect(m, VM_PROT_NONE); 1393 if (m->dirty != 0) { 1394 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1395 (long)m->pindex); 1396 } 1397 s = splvm(); 1398 vm_pageq_remove_nowakeup(m); 1399 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1400 vm_page_free_wakeup(); 1401 splx(s); 1402} 1403 1404/* 1405 * vm_page_dontneed 1406 * 1407 * Cache, deactivate, or do nothing as appropriate. This routine 1408 * is typically used by madvise() MADV_DONTNEED. 1409 * 1410 * Generally speaking we want to move the page into the cache so 1411 * it gets reused quickly. However, this can result in a silly syndrome 1412 * due to the page recycling too quickly. Small objects will not be 1413 * fully cached. On the otherhand, if we move the page to the inactive 1414 * queue we wind up with a problem whereby very large objects 1415 * unnecessarily blow away our inactive and cache queues. 1416 * 1417 * The solution is to move the pages based on a fixed weighting. We 1418 * either leave them alone, deactivate them, or move them to the cache, 1419 * where moving them to the cache has the highest weighting. 1420 * By forcing some pages into other queues we eventually force the 1421 * system to balance the queues, potentially recovering other unrelated 1422 * space from active. The idea is to not force this to happen too 1423 * often. 1424 */ 1425void 1426vm_page_dontneed(vm_page_t m) 1427{ 1428 static int dnweight; 1429 int dnw; 1430 int head; 1431 1432 GIANT_REQUIRED; 1433 dnw = ++dnweight; 1434 1435 /* 1436 * occassionally leave the page alone 1437 */ 1438 if ((dnw & 0x01F0) == 0 || 1439 m->queue == PQ_INACTIVE || 1440 m->queue - m->pc == PQ_CACHE 1441 ) { 1442 if (m->act_count >= ACT_INIT) 1443 --m->act_count; 1444 return; 1445 } 1446 1447 if (m->dirty == 0) 1448 vm_page_test_dirty(m); 1449 1450 if (m->dirty || (dnw & 0x0070) == 0) { 1451 /* 1452 * Deactivate the page 3 times out of 32. 1453 */ 1454 head = 0; 1455 } else { 1456 /* 1457 * Cache the page 28 times out of every 32. Note that 1458 * the page is deactivated instead of cached, but placed 1459 * at the head of the queue instead of the tail. 1460 */ 1461 head = 1; 1462 } 1463 _vm_page_deactivate(m, head); 1464} 1465 1466/* 1467 * Grab a page, waiting until we are waken up due to the page 1468 * changing state. We keep on waiting, if the page continues 1469 * to be in the object. If the page doesn't exist, allocate it. 1470 * 1471 * This routine may block. 1472 */ 1473vm_page_t 1474vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1475{ 1476 vm_page_t m; 1477 int s, generation; 1478 1479 GIANT_REQUIRED; 1480retrylookup: 1481 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1482 if (m->busy || (m->flags & PG_BUSY)) { 1483 generation = object->generation; 1484 1485 s = splvm(); 1486 while ((object->generation == generation) && 1487 (m->busy || (m->flags & PG_BUSY))) { 1488 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1489 tsleep(m, PVM, "pgrbwt", 0); 1490 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1491 splx(s); 1492 return NULL; 1493 } 1494 } 1495 splx(s); 1496 goto retrylookup; 1497 } else { 1498 vm_page_busy(m); 1499 return m; 1500 } 1501 } 1502 1503 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1504 if (m == NULL) { 1505 VM_WAIT; 1506 if ((allocflags & VM_ALLOC_RETRY) == 0) 1507 return NULL; 1508 goto retrylookup; 1509 } 1510 1511 return m; 1512} 1513 1514/* 1515 * Mapping function for valid bits or for dirty bits in 1516 * a page. May not block. 1517 * 1518 * Inputs are required to range within a page. 1519 */ 1520__inline int 1521vm_page_bits(int base, int size) 1522{ 1523 int first_bit; 1524 int last_bit; 1525 1526 KASSERT( 1527 base + size <= PAGE_SIZE, 1528 ("vm_page_bits: illegal base/size %d/%d", base, size) 1529 ); 1530 1531 if (size == 0) /* handle degenerate case */ 1532 return (0); 1533 1534 first_bit = base >> DEV_BSHIFT; 1535 last_bit = (base + size - 1) >> DEV_BSHIFT; 1536 1537 return ((2 << last_bit) - (1 << first_bit)); 1538} 1539 1540/* 1541 * vm_page_set_validclean: 1542 * 1543 * Sets portions of a page valid and clean. The arguments are expected 1544 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1545 * of any partial chunks touched by the range. The invalid portion of 1546 * such chunks will be zero'd. 1547 * 1548 * This routine may not block. 1549 * 1550 * (base + size) must be less then or equal to PAGE_SIZE. 1551 */ 1552void 1553vm_page_set_validclean(vm_page_t m, int base, int size) 1554{ 1555 int pagebits; 1556 int frag; 1557 int endoff; 1558 1559 GIANT_REQUIRED; 1560 if (size == 0) /* handle degenerate case */ 1561 return; 1562 1563 /* 1564 * If the base is not DEV_BSIZE aligned and the valid 1565 * bit is clear, we have to zero out a portion of the 1566 * first block. 1567 */ 1568 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1569 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1570 pmap_zero_page_area(m, frag, base - frag); 1571 1572 /* 1573 * If the ending offset is not DEV_BSIZE aligned and the 1574 * valid bit is clear, we have to zero out a portion of 1575 * the last block. 1576 */ 1577 endoff = base + size; 1578 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1579 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1580 pmap_zero_page_area(m, endoff, 1581 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1582 1583 /* 1584 * Set valid, clear dirty bits. If validating the entire 1585 * page we can safely clear the pmap modify bit. We also 1586 * use this opportunity to clear the PG_NOSYNC flag. If a process 1587 * takes a write fault on a MAP_NOSYNC memory area the flag will 1588 * be set again. 1589 * 1590 * We set valid bits inclusive of any overlap, but we can only 1591 * clear dirty bits for DEV_BSIZE chunks that are fully within 1592 * the range. 1593 */ 1594 pagebits = vm_page_bits(base, size); 1595 m->valid |= pagebits; 1596#if 0 /* NOT YET */ 1597 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1598 frag = DEV_BSIZE - frag; 1599 base += frag; 1600 size -= frag; 1601 if (size < 0) 1602 size = 0; 1603 } 1604 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1605#endif 1606 m->dirty &= ~pagebits; 1607 if (base == 0 && size == PAGE_SIZE) { 1608 pmap_clear_modify(m); 1609 vm_page_flag_clear(m, PG_NOSYNC); 1610 } 1611} 1612 1613#if 0 1614 1615void 1616vm_page_set_dirty(vm_page_t m, int base, int size) 1617{ 1618 m->dirty |= vm_page_bits(base, size); 1619} 1620 1621#endif 1622 1623void 1624vm_page_clear_dirty(vm_page_t m, int base, int size) 1625{ 1626 GIANT_REQUIRED; 1627 m->dirty &= ~vm_page_bits(base, size); 1628} 1629 1630/* 1631 * vm_page_set_invalid: 1632 * 1633 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1634 * valid and dirty bits for the effected areas are cleared. 1635 * 1636 * May not block. 1637 */ 1638void 1639vm_page_set_invalid(vm_page_t m, int base, int size) 1640{ 1641 int bits; 1642 1643 GIANT_REQUIRED; 1644 bits = vm_page_bits(base, size); 1645 m->valid &= ~bits; 1646 m->dirty &= ~bits; 1647 m->object->generation++; 1648} 1649 1650/* 1651 * vm_page_zero_invalid() 1652 * 1653 * The kernel assumes that the invalid portions of a page contain 1654 * garbage, but such pages can be mapped into memory by user code. 1655 * When this occurs, we must zero out the non-valid portions of the 1656 * page so user code sees what it expects. 1657 * 1658 * Pages are most often semi-valid when the end of a file is mapped 1659 * into memory and the file's size is not page aligned. 1660 */ 1661void 1662vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1663{ 1664 int b; 1665 int i; 1666 1667 /* 1668 * Scan the valid bits looking for invalid sections that 1669 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1670 * valid bit may be set ) have already been zerod by 1671 * vm_page_set_validclean(). 1672 */ 1673 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1674 if (i == (PAGE_SIZE / DEV_BSIZE) || 1675 (m->valid & (1 << i)) 1676 ) { 1677 if (i > b) { 1678 pmap_zero_page_area(m, 1679 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1680 } 1681 b = i + 1; 1682 } 1683 } 1684 1685 /* 1686 * setvalid is TRUE when we can safely set the zero'd areas 1687 * as being valid. We can do this if there are no cache consistancy 1688 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1689 */ 1690 if (setvalid) 1691 m->valid = VM_PAGE_BITS_ALL; 1692} 1693 1694/* 1695 * vm_page_is_valid: 1696 * 1697 * Is (partial) page valid? Note that the case where size == 0 1698 * will return FALSE in the degenerate case where the page is 1699 * entirely invalid, and TRUE otherwise. 1700 * 1701 * May not block. 1702 */ 1703int 1704vm_page_is_valid(vm_page_t m, int base, int size) 1705{ 1706 int bits = vm_page_bits(base, size); 1707 1708 if (m->valid && ((m->valid & bits) == bits)) 1709 return 1; 1710 else 1711 return 0; 1712} 1713 1714/* 1715 * update dirty bits from pmap/mmu. May not block. 1716 */ 1717void 1718vm_page_test_dirty(vm_page_t m) 1719{ 1720 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1721 vm_page_dirty(m); 1722 } 1723} 1724 1725#include "opt_ddb.h" 1726#ifdef DDB 1727#include <sys/kernel.h> 1728 1729#include <ddb/ddb.h> 1730 1731DB_SHOW_COMMAND(page, vm_page_print_page_info) 1732{ 1733 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1734 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1735 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1736 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1737 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1738 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1739 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1740 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1741 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1742 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1743} 1744 1745DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1746{ 1747 int i; 1748 db_printf("PQ_FREE:"); 1749 for (i = 0; i < PQ_L2_SIZE; i++) { 1750 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1751 } 1752 db_printf("\n"); 1753 1754 db_printf("PQ_CACHE:"); 1755 for (i = 0; i < PQ_L2_SIZE; i++) { 1756 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1757 } 1758 db_printf("\n"); 1759 1760 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1761 vm_page_queues[PQ_ACTIVE].lcnt, 1762 vm_page_queues[PQ_INACTIVE].lcnt); 1763} 1764#endif /* DDB */ 1765