vm_page.c revision 100005
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 100005 2002-07-14 23:51:55Z 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 133struct mtx vm_page_queue_mtx; 134struct mtx vm_page_queue_free_mtx; 135 136vm_page_t vm_page_array = 0; 137int vm_page_array_size = 0; 138long first_page = 0; 139int vm_page_zero_count = 0; 140 141/* 142 * vm_set_page_size: 143 * 144 * Sets the page size, perhaps based upon the memory 145 * size. Must be called before any use of page-size 146 * dependent functions. 147 */ 148void 149vm_set_page_size(void) 150{ 151 if (cnt.v_page_size == 0) 152 cnt.v_page_size = PAGE_SIZE; 153 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 154 panic("vm_set_page_size: page size not a power of two"); 155} 156 157/* 158 * vm_page_startup: 159 * 160 * Initializes the resident memory module. 161 * 162 * Allocates memory for the page cells, and 163 * for the object/offset-to-page hash table headers. 164 * Each page cell is initialized and placed on the free list. 165 */ 166vm_offset_t 167vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 168{ 169 vm_offset_t mapped; 170 struct vm_page **bucket; 171 vm_size_t npages, page_range; 172 vm_offset_t new_end; 173 int i; 174 vm_offset_t pa; 175 int nblocks; 176 vm_offset_t last_pa; 177 178 /* the biggest memory array is the second group of pages */ 179 vm_offset_t end; 180 vm_offset_t biggestone, biggestsize; 181 182 vm_offset_t total; 183 vm_size_t bootpages; 184 185 total = 0; 186 biggestsize = 0; 187 biggestone = 0; 188 nblocks = 0; 189 vaddr = round_page(vaddr); 190 191 for (i = 0; phys_avail[i + 1]; i += 2) { 192 phys_avail[i] = round_page(phys_avail[i]); 193 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 194 } 195 196 for (i = 0; phys_avail[i + 1]; i += 2) { 197 vm_size_t size = phys_avail[i + 1] - phys_avail[i]; 198 199 if (size > biggestsize) { 200 biggestone = i; 201 biggestsize = size; 202 } 203 ++nblocks; 204 total += size; 205 } 206 207 end = phys_avail[biggestone+1]; 208 209 /* 210 * Initialize the locks. 211 */ 212 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF); 213 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 214 MTX_SPIN); 215 216 /* 217 * Initialize the queue headers for the free queue, the active queue 218 * and the inactive queue. 219 */ 220 vm_pageq_init(); 221 222 /* 223 * Allocate memory for use when boot strapping the kernel memory allocator 224 */ 225 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 226 new_end = end - bootpages; 227 new_end = trunc_page(new_end); 228 mapped = pmap_map(&vaddr, new_end, end, 229 VM_PROT_READ | VM_PROT_WRITE); 230 bzero((caddr_t) mapped, end - new_end); 231 uma_startup((caddr_t)mapped); 232 233 end = new_end; 234 235 /* 236 * Allocate (and initialize) the hash table buckets. 237 * 238 * The number of buckets MUST BE a power of 2, and the actual value is 239 * the next power of 2 greater than the number of physical pages in 240 * the system. 241 * 242 * We make the hash table approximately 2x the number of pages to 243 * reduce the chain length. This is about the same size using the 244 * singly-linked list as the 1x hash table we were using before 245 * using TAILQ but the chain length will be smaller. 246 * 247 * Note: This computation can be tweaked if desired. 248 */ 249 if (vm_page_bucket_count == 0) { 250 vm_page_bucket_count = 1; 251 while (vm_page_bucket_count < atop(total)) 252 vm_page_bucket_count <<= 1; 253 } 254 vm_page_bucket_count <<= 1; 255 vm_page_hash_mask = vm_page_bucket_count - 1; 256 257 /* 258 * Validate these addresses. 259 */ 260 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 261 new_end = trunc_page(new_end); 262 mapped = pmap_map(&vaddr, new_end, end, 263 VM_PROT_READ | VM_PROT_WRITE); 264 bzero((caddr_t) mapped, end - new_end); 265 266 mtx_init(&vm_page_buckets_mtx, "vm page buckets mutex", NULL, MTX_SPIN); 267 vm_page_buckets = (struct vm_page **)mapped; 268 bucket = vm_page_buckets; 269 for (i = 0; i < vm_page_bucket_count; i++) { 270 *bucket = NULL; 271 bucket++; 272 } 273 274 /* 275 * Compute the number of pages of memory that will be available for 276 * use (taking into account the overhead of a page structure per 277 * page). 278 */ 279 first_page = phys_avail[0] / PAGE_SIZE; 280 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 281 npages = (total - (page_range * sizeof(struct vm_page)) - 282 (end - new_end)) / PAGE_SIZE; 283 end = new_end; 284 285 /* 286 * Initialize the mem entry structures now, and put them in the free 287 * queue. 288 */ 289 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 290 mapped = pmap_map(&vaddr, new_end, end, 291 VM_PROT_READ | VM_PROT_WRITE); 292 vm_page_array = (vm_page_t) mapped; 293 294 /* 295 * Clear all of the page structures 296 */ 297 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 298 vm_page_array_size = page_range; 299 300 /* 301 * Construct the free queue(s) in descending order (by physical 302 * address) so that the first 16MB of physical memory is allocated 303 * last rather than first. On large-memory machines, this avoids 304 * the exhaustion of low physical memory before isa_dmainit has run. 305 */ 306 cnt.v_page_count = 0; 307 cnt.v_free_count = 0; 308 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 309 pa = phys_avail[i]; 310 if (i == biggestone) 311 last_pa = new_end; 312 else 313 last_pa = phys_avail[i + 1]; 314 while (pa < last_pa && npages-- > 0) { 315 vm_pageq_add_new_page(pa); 316 pa += PAGE_SIZE; 317 } 318 } 319 return (vaddr); 320} 321 322/* 323 * vm_page_hash: 324 * 325 * Distributes the object/offset key pair among hash buckets. 326 * 327 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 328 * This routine may not block. 329 * 330 * We try to randomize the hash based on the object to spread the pages 331 * out in the hash table without it costing us too much. 332 */ 333static __inline int 334vm_page_hash(vm_object_t object, vm_pindex_t pindex) 335{ 336 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 337 338 return (i & vm_page_hash_mask); 339} 340 341void 342vm_page_flag_set(vm_page_t m, unsigned short bits) 343{ 344 GIANT_REQUIRED; 345 m->flags |= bits; 346} 347 348void 349vm_page_flag_clear(vm_page_t m, unsigned short bits) 350{ 351 GIANT_REQUIRED; 352 m->flags &= ~bits; 353} 354 355void 356vm_page_busy(vm_page_t m) 357{ 358 KASSERT((m->flags & PG_BUSY) == 0, 359 ("vm_page_busy: page already busy!!!")); 360 vm_page_flag_set(m, PG_BUSY); 361} 362 363/* 364 * vm_page_flash: 365 * 366 * wakeup anyone waiting for the page. 367 */ 368void 369vm_page_flash(vm_page_t m) 370{ 371 if (m->flags & PG_WANTED) { 372 vm_page_flag_clear(m, PG_WANTED); 373 wakeup(m); 374 } 375} 376 377/* 378 * vm_page_wakeup: 379 * 380 * clear the PG_BUSY flag and wakeup anyone waiting for the 381 * page. 382 * 383 */ 384void 385vm_page_wakeup(vm_page_t m) 386{ 387 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 388 vm_page_flag_clear(m, PG_BUSY); 389 vm_page_flash(m); 390} 391 392/* 393 * 394 * 395 */ 396void 397vm_page_io_start(vm_page_t m) 398{ 399 GIANT_REQUIRED; 400 m->busy++; 401} 402 403void 404vm_page_io_finish(vm_page_t m) 405{ 406 GIANT_REQUIRED; 407 m->busy--; 408 if (m->busy == 0) 409 vm_page_flash(m); 410} 411 412/* 413 * Keep page from being freed by the page daemon 414 * much of the same effect as wiring, except much lower 415 * overhead and should be used only for *very* temporary 416 * holding ("wiring"). 417 */ 418void 419vm_page_hold(vm_page_t mem) 420{ 421 GIANT_REQUIRED; 422 mem->hold_count++; 423} 424 425void 426vm_page_unhold(vm_page_t mem) 427{ 428 GIANT_REQUIRED; 429 --mem->hold_count; 430 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 431 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 432 vm_page_free_toq(mem); 433} 434 435/* 436 * vm_page_protect: 437 * 438 * Reduce the protection of a page. This routine never raises the 439 * protection and therefore can be safely called if the page is already 440 * at VM_PROT_NONE (it will be a NOP effectively ). 441 */ 442void 443vm_page_protect(vm_page_t mem, int prot) 444{ 445 if (prot == VM_PROT_NONE) { 446 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) { 447 pmap_page_protect(mem, VM_PROT_NONE); 448 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED); 449 } 450 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 451 pmap_page_protect(mem, VM_PROT_READ); 452 vm_page_flag_clear(mem, PG_WRITEABLE); 453 } 454} 455/* 456 * vm_page_zero_fill: 457 * 458 * Zero-fill the specified page. 459 * Written as a standard pagein routine, to 460 * be used by the zero-fill object. 461 */ 462boolean_t 463vm_page_zero_fill(vm_page_t m) 464{ 465 pmap_zero_page(m); 466 return (TRUE); 467} 468 469/* 470 * vm_page_zero_fill_area: 471 * 472 * Like vm_page_zero_fill but only fill the specified area. 473 */ 474boolean_t 475vm_page_zero_fill_area(vm_page_t m, int off, int size) 476{ 477 pmap_zero_page_area(m, off, size); 478 return (TRUE); 479} 480 481/* 482 * vm_page_copy: 483 * 484 * Copy one page to another 485 */ 486void 487vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 488{ 489 pmap_copy_page(src_m, dest_m); 490 dest_m->valid = VM_PAGE_BITS_ALL; 491} 492 493/* 494 * vm_page_free: 495 * 496 * Free a page 497 * 498 * The clearing of PG_ZERO is a temporary safety until the code can be 499 * reviewed to determine that PG_ZERO is being properly cleared on 500 * write faults or maps. PG_ZERO was previously cleared in 501 * vm_page_alloc(). 502 */ 503void 504vm_page_free(vm_page_t m) 505{ 506 vm_page_flag_clear(m, PG_ZERO); 507 vm_page_free_toq(m); 508 vm_page_zero_idle_wakeup(); 509} 510 511/* 512 * vm_page_free_zero: 513 * 514 * Free a page to the zerod-pages queue 515 */ 516void 517vm_page_free_zero(vm_page_t m) 518{ 519 vm_page_flag_set(m, PG_ZERO); 520 vm_page_free_toq(m); 521} 522 523/* 524 * vm_page_sleep_busy: 525 * 526 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 527 * m->busy is zero. Returns TRUE if it had to sleep ( including if 528 * it almost had to sleep and made temporary spl*() mods), FALSE 529 * otherwise. 530 * 531 * This routine assumes that interrupts can only remove the busy 532 * status from a page, not set the busy status or change it from 533 * PG_BUSY to m->busy or vise versa (which would create a timing 534 * window). 535 */ 536int 537vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 538{ 539 GIANT_REQUIRED; 540 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 541 int s = splvm(); 542 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 543 /* 544 * Page is busy. Wait and retry. 545 */ 546 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 547 tsleep(m, PVM, msg, 0); 548 } 549 splx(s); 550 return (TRUE); 551 /* not reached */ 552 } 553 return (FALSE); 554} 555/* 556 * vm_page_dirty: 557 * 558 * make page all dirty 559 */ 560void 561vm_page_dirty(vm_page_t m) 562{ 563 KASSERT(m->queue - m->pc != PQ_CACHE, 564 ("vm_page_dirty: page in cache!")); 565 m->dirty = VM_PAGE_BITS_ALL; 566} 567 568/* 569 * vm_page_undirty: 570 * 571 * Set page to not be dirty. Note: does not clear pmap modify bits 572 */ 573void 574vm_page_undirty(vm_page_t m) 575{ 576 m->dirty = 0; 577} 578 579/* 580 * vm_page_insert: [ internal use only ] 581 * 582 * Inserts the given mem entry into the object and object list. 583 * 584 * The pagetables are not updated but will presumably fault the page 585 * in if necessary, or if a kernel page the caller will at some point 586 * enter the page into the kernel's pmap. We are not allowed to block 587 * here so we *can't* do this anyway. 588 * 589 * The object and page must be locked, and must be splhigh. 590 * This routine may not block. 591 */ 592void 593vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 594{ 595 struct vm_page **bucket; 596 597 GIANT_REQUIRED; 598 599 if (m->object != NULL) 600 panic("vm_page_insert: already inserted"); 601 602 /* 603 * Record the object/offset pair in this page 604 */ 605 m->object = object; 606 m->pindex = pindex; 607 608 /* 609 * Insert it into the object_object/offset hash table 610 */ 611 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 612 mtx_lock_spin(&vm_page_buckets_mtx); 613 m->hnext = *bucket; 614 *bucket = m; 615 mtx_unlock_spin(&vm_page_buckets_mtx); 616 617 /* 618 * Now link into the object's list of backed pages. 619 */ 620 TAILQ_INSERT_TAIL(&object->memq, m, listq); 621 object->generation++; 622 623 /* 624 * show that the object has one more resident page. 625 */ 626 object->resident_page_count++; 627 628 /* 629 * Since we are inserting a new and possibly dirty page, 630 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 631 */ 632 if (m->flags & PG_WRITEABLE) 633 vm_object_set_writeable_dirty(object); 634} 635 636/* 637 * vm_page_remove: 638 * NOTE: used by device pager as well -wfj 639 * 640 * Removes the given mem entry from the object/offset-page 641 * table and the object page list, but do not invalidate/terminate 642 * the backing store. 643 * 644 * The object and page must be locked, and at splhigh. 645 * The underlying pmap entry (if any) is NOT removed here. 646 * This routine may not block. 647 */ 648void 649vm_page_remove(vm_page_t m) 650{ 651 vm_object_t object; 652 vm_page_t *bucket; 653 654 GIANT_REQUIRED; 655 656 if (m->object == NULL) 657 return; 658 659 if ((m->flags & PG_BUSY) == 0) { 660 panic("vm_page_remove: page not busy"); 661 } 662 663 /* 664 * Basically destroy the page. 665 */ 666 vm_page_wakeup(m); 667 668 object = m->object; 669 670 /* 671 * Remove from the object_object/offset hash table. The object 672 * must be on the hash queue, we will panic if it isn't 673 */ 674 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 675 mtx_lock_spin(&vm_page_buckets_mtx); 676 while (*bucket != m) { 677 if (*bucket == NULL) 678 panic("vm_page_remove(): page not found in hash"); 679 bucket = &(*bucket)->hnext; 680 } 681 *bucket = m->hnext; 682 m->hnext = NULL; 683 mtx_unlock_spin(&vm_page_buckets_mtx); 684 685 /* 686 * Now remove from the object's list of backed pages. 687 */ 688 TAILQ_REMOVE(&object->memq, m, listq); 689 690 /* 691 * And show that the object has one fewer resident page. 692 */ 693 object->resident_page_count--; 694 object->generation++; 695 696 m->object = NULL; 697} 698 699/* 700 * vm_page_lookup: 701 * 702 * Returns the page associated with the object/offset 703 * pair specified; if none is found, NULL is returned. 704 * 705 * The object must be locked. No side effects. 706 * This routine may not block. 707 * This is a critical path routine 708 */ 709vm_page_t 710vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 711{ 712 vm_page_t m; 713 struct vm_page **bucket; 714 715 /* 716 * Search the hash table for this object/offset pair 717 */ 718 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 719 mtx_lock_spin(&vm_page_buckets_mtx); 720 for (m = *bucket; m != NULL; m = m->hnext) 721 if (m->object == object && m->pindex == pindex) 722 break; 723 mtx_unlock_spin(&vm_page_buckets_mtx); 724 return (m); 725} 726 727/* 728 * vm_page_rename: 729 * 730 * Move the given memory entry from its 731 * current object to the specified target object/offset. 732 * 733 * The object must be locked. 734 * This routine may not block. 735 * 736 * Note: this routine will raise itself to splvm(), the caller need not. 737 * 738 * Note: swap associated with the page must be invalidated by the move. We 739 * have to do this for several reasons: (1) we aren't freeing the 740 * page, (2) we are dirtying the page, (3) the VM system is probably 741 * moving the page from object A to B, and will then later move 742 * the backing store from A to B and we can't have a conflict. 743 * 744 * Note: we *always* dirty the page. It is necessary both for the 745 * fact that we moved it, and because we may be invalidating 746 * swap. If the page is on the cache, we have to deactivate it 747 * or vm_page_dirty() will panic. Dirty pages are not allowed 748 * on the cache. 749 */ 750void 751vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 752{ 753 int s; 754 755 s = splvm(); 756 vm_page_lock_queues(); 757 vm_page_remove(m); 758 vm_page_insert(m, new_object, new_pindex); 759 if (m->queue - m->pc == PQ_CACHE) 760 vm_page_deactivate(m); 761 vm_page_dirty(m); 762 vm_page_unlock_queues(); 763 splx(s); 764} 765 766/* 767 * vm_page_select_cache: 768 * 769 * Find a page on the cache queue with color optimization. As pages 770 * might be found, but not applicable, they are deactivated. This 771 * keeps us from using potentially busy cached pages. 772 * 773 * This routine must be called at splvm(). 774 * This routine may not block. 775 */ 776static vm_page_t 777vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 778{ 779 vm_page_t m; 780 781 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 782 while (TRUE) { 783 m = vm_pageq_find( 784 PQ_CACHE, 785 (pindex + object->pg_color) & PQ_L2_MASK, 786 FALSE 787 ); 788 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 789 m->hold_count || m->wire_count)) { 790 vm_page_deactivate(m); 791 continue; 792 } 793 return m; 794 } 795} 796 797/* 798 * vm_page_select_free: 799 * 800 * Find a free or zero page, with specified preference. 801 * 802 * This routine must be called at splvm(). 803 * This routine may not block. 804 */ 805static __inline vm_page_t 806vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 807{ 808 vm_page_t m; 809 810 m = vm_pageq_find( 811 PQ_FREE, 812 (pindex + object->pg_color) & PQ_L2_MASK, 813 prefer_zero 814 ); 815 return (m); 816} 817 818/* 819 * vm_page_alloc: 820 * 821 * Allocate and return a memory cell associated 822 * with this VM object/offset pair. 823 * 824 * page_req classes: 825 * VM_ALLOC_NORMAL normal process request 826 * VM_ALLOC_SYSTEM system *really* needs a page 827 * VM_ALLOC_INTERRUPT interrupt time request 828 * VM_ALLOC_ZERO zero page 829 * 830 * This routine may not block. 831 * 832 * Additional special handling is required when called from an 833 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 834 * the page cache in this case. 835 */ 836vm_page_t 837vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 838{ 839 vm_page_t m = NULL; 840 boolean_t prefer_zero; 841 int s; 842 843 GIANT_REQUIRED; 844 845 KASSERT(!vm_page_lookup(object, pindex), 846 ("vm_page_alloc: page already allocated")); 847 848 prefer_zero = (page_req & VM_ALLOC_ZERO) != 0 ? TRUE : FALSE; 849 page_req &= ~VM_ALLOC_ZERO; 850 851 /* 852 * The pager is allowed to eat deeper into the free page list. 853 */ 854 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 855 page_req = VM_ALLOC_SYSTEM; 856 }; 857 858 s = splvm(); 859loop: 860 mtx_lock_spin(&vm_page_queue_free_mtx); 861 if (cnt.v_free_count > cnt.v_free_reserved) { 862 /* 863 * Allocate from the free queue if there are plenty of pages 864 * in it. 865 */ 866 m = vm_page_select_free(object, pindex, prefer_zero); 867 } else if ( 868 (page_req == VM_ALLOC_SYSTEM && 869 cnt.v_cache_count == 0 && 870 cnt.v_free_count > cnt.v_interrupt_free_min) || 871 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 872 ) { 873 /* 874 * Interrupt or system, dig deeper into the free list. 875 */ 876 m = vm_page_select_free(object, pindex, FALSE); 877 } else if (page_req != VM_ALLOC_INTERRUPT) { 878 mtx_unlock_spin(&vm_page_queue_free_mtx); 879 /* 880 * Allocatable from cache (non-interrupt only). On success, 881 * we must free the page and try again, thus ensuring that 882 * cnt.v_*_free_min counters are replenished. 883 */ 884 vm_page_lock_queues(); 885 if ((m = vm_page_select_cache(object, pindex)) == NULL) { 886 vm_page_unlock_queues(); 887 splx(s); 888#if defined(DIAGNOSTIC) 889 if (cnt.v_cache_count > 0) 890 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 891#endif 892 vm_pageout_deficit++; 893 pagedaemon_wakeup(); 894 return (NULL); 895 } 896 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 897 vm_page_busy(m); 898 vm_page_protect(m, VM_PROT_NONE); 899 vm_page_free(m); 900 vm_page_unlock_queues(); 901 goto loop; 902 } else { 903 /* 904 * Not allocatable from cache from interrupt, give up. 905 */ 906 mtx_unlock_spin(&vm_page_queue_free_mtx); 907 splx(s); 908 vm_pageout_deficit++; 909 pagedaemon_wakeup(); 910 return (NULL); 911 } 912 913 /* 914 * At this point we had better have found a good page. 915 */ 916 917 KASSERT( 918 m != NULL, 919 ("vm_page_alloc(): missing page on free queue\n") 920 ); 921 922 /* 923 * Remove from free queue 924 */ 925 926 vm_pageq_remove_nowakeup(m); 927 928 /* 929 * Initialize structure. Only the PG_ZERO flag is inherited. 930 */ 931 if (m->flags & PG_ZERO) { 932 vm_page_zero_count--; 933 m->flags = PG_ZERO | PG_BUSY; 934 } else { 935 m->flags = PG_BUSY; 936 } 937 m->wire_count = 0; 938 m->hold_count = 0; 939 m->act_count = 0; 940 m->busy = 0; 941 m->valid = 0; 942 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 943 mtx_unlock_spin(&vm_page_queue_free_mtx); 944 945 /* 946 * vm_page_insert() is safe prior to the splx(). Note also that 947 * inserting a page here does not insert it into the pmap (which 948 * could cause us to block allocating memory). We cannot block 949 * anywhere. 950 */ 951 vm_page_insert(m, object, pindex); 952 953 /* 954 * Don't wakeup too often - wakeup the pageout daemon when 955 * we would be nearly out of memory. 956 */ 957 if (vm_paging_needed()) 958 pagedaemon_wakeup(); 959 960 splx(s); 961 return (m); 962} 963 964/* 965 * vm_wait: (also see VM_WAIT macro) 966 * 967 * Block until free pages are available for allocation 968 * - Called in various places before memory allocations. 969 */ 970void 971vm_wait(void) 972{ 973 int s; 974 975 s = splvm(); 976 if (curproc == pageproc) { 977 vm_pageout_pages_needed = 1; 978 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 979 } else { 980 if (!vm_pages_needed) { 981 vm_pages_needed = 1; 982 wakeup(&vm_pages_needed); 983 } 984 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 985 } 986 splx(s); 987} 988 989/* 990 * vm_waitpfault: (also see VM_WAITPFAULT macro) 991 * 992 * Block until free pages are available for allocation 993 * - Called only in vm_fault so that processes page faulting 994 * can be easily tracked. 995 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 996 * processes will be able to grab memory first. Do not change 997 * this balance without careful testing first. 998 */ 999void 1000vm_waitpfault(void) 1001{ 1002 int s; 1003 1004 s = splvm(); 1005 if (!vm_pages_needed) { 1006 vm_pages_needed = 1; 1007 wakeup(&vm_pages_needed); 1008 } 1009 tsleep(&cnt.v_free_count, PUSER, "pfault", 0); 1010 splx(s); 1011} 1012 1013/* 1014 * vm_page_activate: 1015 * 1016 * Put the specified page on the active list (if appropriate). 1017 * Ensure that act_count is at least ACT_INIT but do not otherwise 1018 * mess with it. 1019 * 1020 * The page queues must be locked. 1021 * This routine may not block. 1022 */ 1023void 1024vm_page_activate(vm_page_t m) 1025{ 1026 int s; 1027 1028 GIANT_REQUIRED; 1029 s = splvm(); 1030 if (m->queue != PQ_ACTIVE) { 1031 if ((m->queue - m->pc) == PQ_CACHE) 1032 cnt.v_reactivated++; 1033 vm_pageq_remove(m); 1034 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1035 if (m->act_count < ACT_INIT) 1036 m->act_count = ACT_INIT; 1037 vm_pageq_enqueue(PQ_ACTIVE, m); 1038 } 1039 } else { 1040 if (m->act_count < ACT_INIT) 1041 m->act_count = ACT_INIT; 1042 } 1043 splx(s); 1044} 1045 1046/* 1047 * vm_page_free_wakeup: 1048 * 1049 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1050 * routine is called when a page has been added to the cache or free 1051 * queues. 1052 * 1053 * This routine may not block. 1054 * This routine must be called at splvm() 1055 */ 1056static __inline void 1057vm_page_free_wakeup(void) 1058{ 1059 /* 1060 * if pageout daemon needs pages, then tell it that there are 1061 * some free. 1062 */ 1063 if (vm_pageout_pages_needed && 1064 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1065 wakeup(&vm_pageout_pages_needed); 1066 vm_pageout_pages_needed = 0; 1067 } 1068 /* 1069 * wakeup processes that are waiting on memory if we hit a 1070 * high water mark. And wakeup scheduler process if we have 1071 * lots of memory. this process will swapin processes. 1072 */ 1073 if (vm_pages_needed && !vm_page_count_min()) { 1074 vm_pages_needed = 0; 1075 wakeup(&cnt.v_free_count); 1076 } 1077} 1078 1079/* 1080 * vm_page_free_toq: 1081 * 1082 * Returns the given page to the PQ_FREE list, 1083 * disassociating it with any VM object. 1084 * 1085 * Object and page must be locked prior to entry. 1086 * This routine may not block. 1087 */ 1088 1089void 1090vm_page_free_toq(vm_page_t m) 1091{ 1092 int s; 1093 struct vpgqueues *pq; 1094 vm_object_t object = m->object; 1095 1096 GIANT_REQUIRED; 1097 s = splvm(); 1098 cnt.v_tfree++; 1099 1100 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1101 printf( 1102 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1103 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1104 m->hold_count); 1105 if ((m->queue - m->pc) == PQ_FREE) 1106 panic("vm_page_free: freeing free page"); 1107 else 1108 panic("vm_page_free: freeing busy page"); 1109 } 1110 1111 /* 1112 * unqueue, then remove page. Note that we cannot destroy 1113 * the page here because we do not want to call the pager's 1114 * callback routine until after we've put the page on the 1115 * appropriate free queue. 1116 */ 1117 vm_pageq_remove_nowakeup(m); 1118 vm_page_remove(m); 1119 1120 /* 1121 * If fictitious remove object association and 1122 * return, otherwise delay object association removal. 1123 */ 1124 if ((m->flags & PG_FICTITIOUS) != 0) { 1125 splx(s); 1126 return; 1127 } 1128 1129 m->valid = 0; 1130 vm_page_undirty(m); 1131 1132 if (m->wire_count != 0) { 1133 if (m->wire_count > 1) { 1134 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1135 m->wire_count, (long)m->pindex); 1136 } 1137 panic("vm_page_free: freeing wired page\n"); 1138 } 1139 1140 /* 1141 * If we've exhausted the object's resident pages we want to free 1142 * it up. 1143 */ 1144 if (object && 1145 (object->type == OBJT_VNODE) && 1146 ((object->flags & OBJ_DEAD) == 0) 1147 ) { 1148 struct vnode *vp = (struct vnode *)object->handle; 1149 1150 if (vp && VSHOULDFREE(vp)) 1151 vfree(vp); 1152 } 1153 1154 /* 1155 * Clear the UNMANAGED flag when freeing an unmanaged page. 1156 */ 1157 if (m->flags & PG_UNMANAGED) { 1158 m->flags &= ~PG_UNMANAGED; 1159 } else { 1160#ifdef __alpha__ 1161 pmap_page_is_free(m); 1162#endif 1163 } 1164 1165 if (m->hold_count != 0) { 1166 m->flags &= ~PG_ZERO; 1167 m->queue = PQ_HOLD; 1168 } else 1169 m->queue = PQ_FREE + m->pc; 1170 pq = &vm_page_queues[m->queue]; 1171 mtx_lock_spin(&vm_page_queue_free_mtx); 1172 pq->lcnt++; 1173 ++(*pq->cnt); 1174 1175 /* 1176 * Put zero'd pages on the end ( where we look for zero'd pages 1177 * first ) and non-zerod pages at the head. 1178 */ 1179 if (m->flags & PG_ZERO) { 1180 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1181 ++vm_page_zero_count; 1182 } else { 1183 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1184 } 1185 mtx_unlock_spin(&vm_page_queue_free_mtx); 1186 vm_page_free_wakeup(); 1187 splx(s); 1188} 1189 1190/* 1191 * vm_page_unmanage: 1192 * 1193 * Prevent PV management from being done on the page. The page is 1194 * removed from the paging queues as if it were wired, and as a 1195 * consequence of no longer being managed the pageout daemon will not 1196 * touch it (since there is no way to locate the pte mappings for the 1197 * page). madvise() calls that mess with the pmap will also no longer 1198 * operate on the page. 1199 * 1200 * Beyond that the page is still reasonably 'normal'. Freeing the page 1201 * will clear the flag. 1202 * 1203 * This routine is used by OBJT_PHYS objects - objects using unswappable 1204 * physical memory as backing store rather then swap-backed memory and 1205 * will eventually be extended to support 4MB unmanaged physical 1206 * mappings. 1207 */ 1208void 1209vm_page_unmanage(vm_page_t m) 1210{ 1211 int s; 1212 1213 s = splvm(); 1214 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1215 if ((m->flags & PG_UNMANAGED) == 0) { 1216 if (m->wire_count == 0) 1217 vm_pageq_remove(m); 1218 } 1219 vm_page_flag_set(m, PG_UNMANAGED); 1220 splx(s); 1221} 1222 1223/* 1224 * vm_page_wire: 1225 * 1226 * Mark this page as wired down by yet 1227 * another map, removing it from paging queues 1228 * as necessary. 1229 * 1230 * The page queues must be locked. 1231 * This routine may not block. 1232 */ 1233void 1234vm_page_wire(vm_page_t m) 1235{ 1236 int s; 1237 1238 /* 1239 * Only bump the wire statistics if the page is not already wired, 1240 * and only unqueue the page if it is on some queue (if it is unmanaged 1241 * it is already off the queues). 1242 */ 1243 s = splvm(); 1244#ifndef __alpha__ 1245 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1246#endif 1247 if (m->wire_count == 0) { 1248 if ((m->flags & PG_UNMANAGED) == 0) 1249 vm_pageq_remove(m); 1250 cnt.v_wire_count++; 1251 } 1252 m->wire_count++; 1253 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1254 splx(s); 1255 vm_page_flag_set(m, PG_MAPPED); 1256} 1257 1258/* 1259 * vm_page_unwire: 1260 * 1261 * Release one wiring of this page, potentially 1262 * enabling it to be paged again. 1263 * 1264 * Many pages placed on the inactive queue should actually go 1265 * into the cache, but it is difficult to figure out which. What 1266 * we do instead, if the inactive target is well met, is to put 1267 * clean pages at the head of the inactive queue instead of the tail. 1268 * This will cause them to be moved to the cache more quickly and 1269 * if not actively re-referenced, freed more quickly. If we just 1270 * stick these pages at the end of the inactive queue, heavy filesystem 1271 * meta-data accesses can cause an unnecessary paging load on memory bound 1272 * processes. This optimization causes one-time-use metadata to be 1273 * reused more quickly. 1274 * 1275 * BUT, if we are in a low-memory situation we have no choice but to 1276 * put clean pages on the cache queue. 1277 * 1278 * A number of routines use vm_page_unwire() to guarantee that the page 1279 * will go into either the inactive or active queues, and will NEVER 1280 * be placed in the cache - for example, just after dirtying a page. 1281 * dirty pages in the cache are not allowed. 1282 * 1283 * The page queues must be locked. 1284 * This routine may not block. 1285 */ 1286void 1287vm_page_unwire(vm_page_t m, int activate) 1288{ 1289 int s; 1290 1291 s = splvm(); 1292 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1293 if (m->wire_count > 0) { 1294 m->wire_count--; 1295 if (m->wire_count == 0) { 1296 cnt.v_wire_count--; 1297 if (m->flags & PG_UNMANAGED) { 1298 ; 1299 } else if (activate) 1300 vm_pageq_enqueue(PQ_ACTIVE, m); 1301 else { 1302 vm_page_flag_clear(m, PG_WINATCFLS); 1303 vm_pageq_enqueue(PQ_INACTIVE, m); 1304 } 1305 } 1306 } else { 1307 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1308 } 1309 splx(s); 1310} 1311 1312 1313/* 1314 * Move the specified page to the inactive queue. If the page has 1315 * any associated swap, the swap is deallocated. 1316 * 1317 * Normally athead is 0 resulting in LRU operation. athead is set 1318 * to 1 if we want this page to be 'as if it were placed in the cache', 1319 * except without unmapping it from the process address space. 1320 * 1321 * This routine may not block. 1322 */ 1323static __inline void 1324_vm_page_deactivate(vm_page_t m, int athead) 1325{ 1326 int s; 1327 1328 GIANT_REQUIRED; 1329 /* 1330 * Ignore if already inactive. 1331 */ 1332 if (m->queue == PQ_INACTIVE) 1333 return; 1334 1335 s = splvm(); 1336 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1337 if ((m->queue - m->pc) == PQ_CACHE) 1338 cnt.v_reactivated++; 1339 vm_page_flag_clear(m, PG_WINATCFLS); 1340 vm_pageq_remove(m); 1341 if (athead) 1342 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1343 else 1344 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1345 m->queue = PQ_INACTIVE; 1346 vm_page_queues[PQ_INACTIVE].lcnt++; 1347 cnt.v_inactive_count++; 1348 } 1349 splx(s); 1350} 1351 1352void 1353vm_page_deactivate(vm_page_t m) 1354{ 1355 _vm_page_deactivate(m, 0); 1356} 1357 1358/* 1359 * vm_page_try_to_cache: 1360 * 1361 * Returns 0 on failure, 1 on success 1362 */ 1363int 1364vm_page_try_to_cache(vm_page_t m) 1365{ 1366 GIANT_REQUIRED; 1367 1368 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1369 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1370 return (0); 1371 } 1372 vm_page_test_dirty(m); 1373 if (m->dirty) 1374 return (0); 1375 vm_page_cache(m); 1376 return (1); 1377} 1378 1379/* 1380 * vm_page_try_to_free() 1381 * 1382 * Attempt to free the page. If we cannot free it, we do nothing. 1383 * 1 is returned on success, 0 on failure. 1384 */ 1385int 1386vm_page_try_to_free(vm_page_t m) 1387{ 1388 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1389 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1390 return (0); 1391 } 1392 vm_page_test_dirty(m); 1393 if (m->dirty) 1394 return (0); 1395 vm_page_busy(m); 1396 vm_page_protect(m, VM_PROT_NONE); 1397 vm_page_free(m); 1398 return (1); 1399} 1400 1401/* 1402 * vm_page_cache 1403 * 1404 * Put the specified page onto the page cache queue (if appropriate). 1405 * 1406 * This routine may not block. 1407 */ 1408void 1409vm_page_cache(vm_page_t m) 1410{ 1411 int s; 1412 1413 GIANT_REQUIRED; 1414 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1415 printf("vm_page_cache: attempting to cache busy page\n"); 1416 return; 1417 } 1418 if ((m->queue - m->pc) == PQ_CACHE) 1419 return; 1420 1421 /* 1422 * Remove all pmaps and indicate that the page is not 1423 * writeable or mapped. 1424 */ 1425 vm_page_protect(m, VM_PROT_NONE); 1426 if (m->dirty != 0) { 1427 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1428 (long)m->pindex); 1429 } 1430 s = splvm(); 1431 vm_pageq_remove_nowakeup(m); 1432 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1433 vm_page_free_wakeup(); 1434 splx(s); 1435} 1436 1437/* 1438 * vm_page_dontneed 1439 * 1440 * Cache, deactivate, or do nothing as appropriate. This routine 1441 * is typically used by madvise() MADV_DONTNEED. 1442 * 1443 * Generally speaking we want to move the page into the cache so 1444 * it gets reused quickly. However, this can result in a silly syndrome 1445 * due to the page recycling too quickly. Small objects will not be 1446 * fully cached. On the otherhand, if we move the page to the inactive 1447 * queue we wind up with a problem whereby very large objects 1448 * unnecessarily blow away our inactive and cache queues. 1449 * 1450 * The solution is to move the pages based on a fixed weighting. We 1451 * either leave them alone, deactivate them, or move them to the cache, 1452 * where moving them to the cache has the highest weighting. 1453 * By forcing some pages into other queues we eventually force the 1454 * system to balance the queues, potentially recovering other unrelated 1455 * space from active. The idea is to not force this to happen too 1456 * often. 1457 */ 1458void 1459vm_page_dontneed(vm_page_t m) 1460{ 1461 static int dnweight; 1462 int dnw; 1463 int head; 1464 1465 GIANT_REQUIRED; 1466 dnw = ++dnweight; 1467 1468 /* 1469 * occassionally leave the page alone 1470 */ 1471 if ((dnw & 0x01F0) == 0 || 1472 m->queue == PQ_INACTIVE || 1473 m->queue - m->pc == PQ_CACHE 1474 ) { 1475 if (m->act_count >= ACT_INIT) 1476 --m->act_count; 1477 return; 1478 } 1479 1480 if (m->dirty == 0) 1481 vm_page_test_dirty(m); 1482 1483 if (m->dirty || (dnw & 0x0070) == 0) { 1484 /* 1485 * Deactivate the page 3 times out of 32. 1486 */ 1487 head = 0; 1488 } else { 1489 /* 1490 * Cache the page 28 times out of every 32. Note that 1491 * the page is deactivated instead of cached, but placed 1492 * at the head of the queue instead of the tail. 1493 */ 1494 head = 1; 1495 } 1496 _vm_page_deactivate(m, head); 1497} 1498 1499/* 1500 * Grab a page, waiting until we are waken up due to the page 1501 * changing state. We keep on waiting, if the page continues 1502 * to be in the object. If the page doesn't exist, allocate it. 1503 * 1504 * This routine may block. 1505 */ 1506vm_page_t 1507vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1508{ 1509 vm_page_t m; 1510 int s, generation; 1511 1512 GIANT_REQUIRED; 1513retrylookup: 1514 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1515 if (m->busy || (m->flags & PG_BUSY)) { 1516 generation = object->generation; 1517 1518 s = splvm(); 1519 while ((object->generation == generation) && 1520 (m->busy || (m->flags & PG_BUSY))) { 1521 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1522 tsleep(m, PVM, "pgrbwt", 0); 1523 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1524 splx(s); 1525 return NULL; 1526 } 1527 } 1528 splx(s); 1529 goto retrylookup; 1530 } else { 1531 vm_page_busy(m); 1532 return m; 1533 } 1534 } 1535 1536 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1537 if (m == NULL) { 1538 VM_WAIT; 1539 if ((allocflags & VM_ALLOC_RETRY) == 0) 1540 return NULL; 1541 goto retrylookup; 1542 } 1543 1544 return m; 1545} 1546 1547/* 1548 * Mapping function for valid bits or for dirty bits in 1549 * a page. May not block. 1550 * 1551 * Inputs are required to range within a page. 1552 */ 1553__inline int 1554vm_page_bits(int base, int size) 1555{ 1556 int first_bit; 1557 int last_bit; 1558 1559 KASSERT( 1560 base + size <= PAGE_SIZE, 1561 ("vm_page_bits: illegal base/size %d/%d", base, size) 1562 ); 1563 1564 if (size == 0) /* handle degenerate case */ 1565 return (0); 1566 1567 first_bit = base >> DEV_BSHIFT; 1568 last_bit = (base + size - 1) >> DEV_BSHIFT; 1569 1570 return ((2 << last_bit) - (1 << first_bit)); 1571} 1572 1573/* 1574 * vm_page_set_validclean: 1575 * 1576 * Sets portions of a page valid and clean. The arguments are expected 1577 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1578 * of any partial chunks touched by the range. The invalid portion of 1579 * such chunks will be zero'd. 1580 * 1581 * This routine may not block. 1582 * 1583 * (base + size) must be less then or equal to PAGE_SIZE. 1584 */ 1585void 1586vm_page_set_validclean(vm_page_t m, int base, int size) 1587{ 1588 int pagebits; 1589 int frag; 1590 int endoff; 1591 1592 GIANT_REQUIRED; 1593 if (size == 0) /* handle degenerate case */ 1594 return; 1595 1596 /* 1597 * If the base is not DEV_BSIZE aligned and the valid 1598 * bit is clear, we have to zero out a portion of the 1599 * first block. 1600 */ 1601 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1602 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1603 pmap_zero_page_area(m, frag, base - frag); 1604 1605 /* 1606 * If the ending offset is not DEV_BSIZE aligned and the 1607 * valid bit is clear, we have to zero out a portion of 1608 * the last block. 1609 */ 1610 endoff = base + size; 1611 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1612 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1613 pmap_zero_page_area(m, endoff, 1614 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1615 1616 /* 1617 * Set valid, clear dirty bits. If validating the entire 1618 * page we can safely clear the pmap modify bit. We also 1619 * use this opportunity to clear the PG_NOSYNC flag. If a process 1620 * takes a write fault on a MAP_NOSYNC memory area the flag will 1621 * be set again. 1622 * 1623 * We set valid bits inclusive of any overlap, but we can only 1624 * clear dirty bits for DEV_BSIZE chunks that are fully within 1625 * the range. 1626 */ 1627 pagebits = vm_page_bits(base, size); 1628 m->valid |= pagebits; 1629#if 0 /* NOT YET */ 1630 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1631 frag = DEV_BSIZE - frag; 1632 base += frag; 1633 size -= frag; 1634 if (size < 0) 1635 size = 0; 1636 } 1637 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1638#endif 1639 m->dirty &= ~pagebits; 1640 if (base == 0 && size == PAGE_SIZE) { 1641 pmap_clear_modify(m); 1642 vm_page_flag_clear(m, PG_NOSYNC); 1643 } 1644} 1645 1646#if 0 1647 1648void 1649vm_page_set_dirty(vm_page_t m, int base, int size) 1650{ 1651 m->dirty |= vm_page_bits(base, size); 1652} 1653 1654#endif 1655 1656void 1657vm_page_clear_dirty(vm_page_t m, int base, int size) 1658{ 1659 GIANT_REQUIRED; 1660 m->dirty &= ~vm_page_bits(base, size); 1661} 1662 1663/* 1664 * vm_page_set_invalid: 1665 * 1666 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1667 * valid and dirty bits for the effected areas are cleared. 1668 * 1669 * May not block. 1670 */ 1671void 1672vm_page_set_invalid(vm_page_t m, int base, int size) 1673{ 1674 int bits; 1675 1676 GIANT_REQUIRED; 1677 bits = vm_page_bits(base, size); 1678 m->valid &= ~bits; 1679 m->dirty &= ~bits; 1680 m->object->generation++; 1681} 1682 1683/* 1684 * vm_page_zero_invalid() 1685 * 1686 * The kernel assumes that the invalid portions of a page contain 1687 * garbage, but such pages can be mapped into memory by user code. 1688 * When this occurs, we must zero out the non-valid portions of the 1689 * page so user code sees what it expects. 1690 * 1691 * Pages are most often semi-valid when the end of a file is mapped 1692 * into memory and the file's size is not page aligned. 1693 */ 1694void 1695vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1696{ 1697 int b; 1698 int i; 1699 1700 /* 1701 * Scan the valid bits looking for invalid sections that 1702 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1703 * valid bit may be set ) have already been zerod by 1704 * vm_page_set_validclean(). 1705 */ 1706 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1707 if (i == (PAGE_SIZE / DEV_BSIZE) || 1708 (m->valid & (1 << i)) 1709 ) { 1710 if (i > b) { 1711 pmap_zero_page_area(m, 1712 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1713 } 1714 b = i + 1; 1715 } 1716 } 1717 1718 /* 1719 * setvalid is TRUE when we can safely set the zero'd areas 1720 * as being valid. We can do this if there are no cache consistancy 1721 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1722 */ 1723 if (setvalid) 1724 m->valid = VM_PAGE_BITS_ALL; 1725} 1726 1727/* 1728 * vm_page_is_valid: 1729 * 1730 * Is (partial) page valid? Note that the case where size == 0 1731 * will return FALSE in the degenerate case where the page is 1732 * entirely invalid, and TRUE otherwise. 1733 * 1734 * May not block. 1735 */ 1736int 1737vm_page_is_valid(vm_page_t m, int base, int size) 1738{ 1739 int bits = vm_page_bits(base, size); 1740 1741 if (m->valid && ((m->valid & bits) == bits)) 1742 return 1; 1743 else 1744 return 0; 1745} 1746 1747/* 1748 * update dirty bits from pmap/mmu. May not block. 1749 */ 1750void 1751vm_page_test_dirty(vm_page_t m) 1752{ 1753 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1754 vm_page_dirty(m); 1755 } 1756} 1757 1758int so_zerocp_fullpage = 0; 1759 1760void 1761vm_page_cowfault(vm_page_t m) 1762{ 1763 vm_page_t mnew; 1764 vm_object_t object; 1765 vm_pindex_t pindex; 1766 1767 object = m->object; 1768 pindex = m->pindex; 1769 vm_page_busy(m); 1770 1771 retry_alloc: 1772 vm_page_remove(m); 1773 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 1774 if (mnew == NULL) { 1775 vm_page_insert(m, object, pindex); 1776 VM_WAIT; 1777 goto retry_alloc; 1778 } 1779 1780 if (m->cow == 0) { 1781 /* 1782 * check to see if we raced with an xmit complete when 1783 * waiting to allocate a page. If so, put things back 1784 * the way they were 1785 */ 1786 vm_page_busy(mnew); 1787 vm_page_free(mnew); 1788 vm_page_insert(m, object, pindex); 1789 } else { /* clear COW & copy page */ 1790 if (so_zerocp_fullpage) { 1791 mnew->valid = VM_PAGE_BITS_ALL; 1792 } else { 1793 vm_page_copy(m, mnew); 1794 } 1795 vm_page_dirty(mnew); 1796 vm_page_flag_clear(mnew, PG_BUSY); 1797 } 1798} 1799 1800void 1801vm_page_cowclear(vm_page_t m) 1802{ 1803 1804 /* XXX KDM find out if giant is required here. */ 1805 GIANT_REQUIRED; 1806 if (m->cow) { 1807 atomic_subtract_int(&m->cow, 1); 1808 /* 1809 * let vm_fault add back write permission lazily 1810 */ 1811 } 1812 /* 1813 * sf_buf_free() will free the page, so we needn't do it here 1814 */ 1815} 1816 1817void 1818vm_page_cowsetup(vm_page_t m) 1819{ 1820 /* XXX KDM find out if giant is required here */ 1821 GIANT_REQUIRED; 1822 atomic_add_int(&m->cow, 1); 1823 vm_page_protect(m, VM_PROT_READ); 1824} 1825 1826#include "opt_ddb.h" 1827#ifdef DDB 1828#include <sys/kernel.h> 1829 1830#include <ddb/ddb.h> 1831 1832DB_SHOW_COMMAND(page, vm_page_print_page_info) 1833{ 1834 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1835 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1836 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1837 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1838 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1839 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1840 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1841 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1842 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1843 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1844} 1845 1846DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1847{ 1848 int i; 1849 db_printf("PQ_FREE:"); 1850 for (i = 0; i < PQ_L2_SIZE; i++) { 1851 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1852 } 1853 db_printf("\n"); 1854 1855 db_printf("PQ_CACHE:"); 1856 for (i = 0; i < PQ_L2_SIZE; i++) { 1857 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1858 } 1859 db_printf("\n"); 1860 1861 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1862 vm_page_queues[PQ_ACTIVE].lcnt, 1863 vm_page_queues[PQ_INACTIVE].lcnt); 1864} 1865#endif /* DDB */ 1866