vm_page.c revision 100415
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 100415 2002-07-20 20:58:46Z 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 req) 838{ 839 vm_page_t m = NULL; 840 int page_req, s; 841 842 GIANT_REQUIRED; 843 844 KASSERT(!vm_page_lookup(object, pindex), 845 ("vm_page_alloc: page already allocated")); 846 847 page_req = req & VM_ALLOC_CLASS_MASK; 848 849 /* 850 * The pager is allowed to eat deeper into the free page list. 851 */ 852 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 853 page_req = VM_ALLOC_SYSTEM; 854 }; 855 856 s = splvm(); 857loop: 858 mtx_lock_spin(&vm_page_queue_free_mtx); 859 if (cnt.v_free_count > cnt.v_free_reserved) { 860 /* 861 * Allocate from the free queue if there are plenty of pages 862 * in it. 863 */ 864 m = vm_page_select_free(object, pindex, 865 (req & VM_ALLOC_ZERO) != 0); 866 } else if ( 867 (page_req == VM_ALLOC_SYSTEM && 868 cnt.v_cache_count == 0 && 869 cnt.v_free_count > cnt.v_interrupt_free_min) || 870 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 871 ) { 872 /* 873 * Interrupt or system, dig deeper into the free list. 874 */ 875 m = vm_page_select_free(object, pindex, FALSE); 876 } else if (page_req != VM_ALLOC_INTERRUPT) { 877 mtx_unlock_spin(&vm_page_queue_free_mtx); 878 /* 879 * Allocatable from cache (non-interrupt only). On success, 880 * we must free the page and try again, thus ensuring that 881 * cnt.v_*_free_min counters are replenished. 882 */ 883 vm_page_lock_queues(); 884 if ((m = vm_page_select_cache(object, pindex)) == NULL) { 885 vm_page_unlock_queues(); 886 splx(s); 887#if defined(DIAGNOSTIC) 888 if (cnt.v_cache_count > 0) 889 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 890#endif 891 vm_pageout_deficit++; 892 pagedaemon_wakeup(); 893 return (NULL); 894 } 895 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 896 vm_page_busy(m); 897 vm_page_protect(m, VM_PROT_NONE); 898 vm_page_free(m); 899 vm_page_unlock_queues(); 900 goto loop; 901 } else { 902 /* 903 * Not allocatable from cache from interrupt, give up. 904 */ 905 mtx_unlock_spin(&vm_page_queue_free_mtx); 906 splx(s); 907 vm_pageout_deficit++; 908 pagedaemon_wakeup(); 909 return (NULL); 910 } 911 912 /* 913 * At this point we had better have found a good page. 914 */ 915 916 KASSERT( 917 m != NULL, 918 ("vm_page_alloc(): missing page on free queue\n") 919 ); 920 921 /* 922 * Remove from free queue 923 */ 924 925 vm_pageq_remove_nowakeup(m); 926 927 /* 928 * Initialize structure. Only the PG_ZERO flag is inherited. 929 */ 930 if (m->flags & PG_ZERO) { 931 vm_page_zero_count--; 932 m->flags = PG_ZERO | PG_BUSY; 933 } else { 934 m->flags = PG_BUSY; 935 } 936 if (req & VM_ALLOC_WIRED) { 937 cnt.v_wire_count++; 938 m->flags |= PG_MAPPED; /* XXX this does not belong here */ 939 m->wire_count = 1; 940 } else 941 m->wire_count = 0; 942 m->hold_count = 0; 943 m->act_count = 0; 944 m->busy = 0; 945 m->valid = 0; 946 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 947 mtx_unlock_spin(&vm_page_queue_free_mtx); 948 949 /* 950 * vm_page_insert() is safe prior to the splx(). Note also that 951 * inserting a page here does not insert it into the pmap (which 952 * could cause us to block allocating memory). We cannot block 953 * anywhere. 954 */ 955 vm_page_insert(m, object, pindex); 956 957 /* 958 * Don't wakeup too often - wakeup the pageout daemon when 959 * we would be nearly out of memory. 960 */ 961 if (vm_paging_needed()) 962 pagedaemon_wakeup(); 963 964 splx(s); 965 return (m); 966} 967 968/* 969 * vm_wait: (also see VM_WAIT macro) 970 * 971 * Block until free pages are available for allocation 972 * - Called in various places before memory allocations. 973 */ 974void 975vm_wait(void) 976{ 977 int s; 978 979 s = splvm(); 980 if (curproc == pageproc) { 981 vm_pageout_pages_needed = 1; 982 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 983 } else { 984 if (!vm_pages_needed) { 985 vm_pages_needed = 1; 986 wakeup(&vm_pages_needed); 987 } 988 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 989 } 990 splx(s); 991} 992 993/* 994 * vm_waitpfault: (also see VM_WAITPFAULT macro) 995 * 996 * Block until free pages are available for allocation 997 * - Called only in vm_fault so that processes page faulting 998 * can be easily tracked. 999 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1000 * processes will be able to grab memory first. Do not change 1001 * this balance without careful testing first. 1002 */ 1003void 1004vm_waitpfault(void) 1005{ 1006 int s; 1007 1008 s = splvm(); 1009 if (!vm_pages_needed) { 1010 vm_pages_needed = 1; 1011 wakeup(&vm_pages_needed); 1012 } 1013 tsleep(&cnt.v_free_count, PUSER, "pfault", 0); 1014 splx(s); 1015} 1016 1017/* 1018 * vm_page_activate: 1019 * 1020 * Put the specified page on the active list (if appropriate). 1021 * Ensure that act_count is at least ACT_INIT but do not otherwise 1022 * mess with it. 1023 * 1024 * The page queues must be locked. 1025 * This routine may not block. 1026 */ 1027void 1028vm_page_activate(vm_page_t m) 1029{ 1030 int s; 1031 1032 GIANT_REQUIRED; 1033 s = splvm(); 1034 if (m->queue != PQ_ACTIVE) { 1035 if ((m->queue - m->pc) == PQ_CACHE) 1036 cnt.v_reactivated++; 1037 vm_pageq_remove(m); 1038 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1039 if (m->act_count < ACT_INIT) 1040 m->act_count = ACT_INIT; 1041 vm_pageq_enqueue(PQ_ACTIVE, m); 1042 } 1043 } else { 1044 if (m->act_count < ACT_INIT) 1045 m->act_count = ACT_INIT; 1046 } 1047 splx(s); 1048} 1049 1050/* 1051 * vm_page_free_wakeup: 1052 * 1053 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1054 * routine is called when a page has been added to the cache or free 1055 * queues. 1056 * 1057 * This routine may not block. 1058 * This routine must be called at splvm() 1059 */ 1060static __inline void 1061vm_page_free_wakeup(void) 1062{ 1063 /* 1064 * if pageout daemon needs pages, then tell it that there are 1065 * some free. 1066 */ 1067 if (vm_pageout_pages_needed && 1068 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1069 wakeup(&vm_pageout_pages_needed); 1070 vm_pageout_pages_needed = 0; 1071 } 1072 /* 1073 * wakeup processes that are waiting on memory if we hit a 1074 * high water mark. And wakeup scheduler process if we have 1075 * lots of memory. this process will swapin processes. 1076 */ 1077 if (vm_pages_needed && !vm_page_count_min()) { 1078 vm_pages_needed = 0; 1079 wakeup(&cnt.v_free_count); 1080 } 1081} 1082 1083/* 1084 * vm_page_free_toq: 1085 * 1086 * Returns the given page to the PQ_FREE list, 1087 * disassociating it with any VM object. 1088 * 1089 * Object and page must be locked prior to entry. 1090 * This routine may not block. 1091 */ 1092 1093void 1094vm_page_free_toq(vm_page_t m) 1095{ 1096 int s; 1097 struct vpgqueues *pq; 1098 vm_object_t object = m->object; 1099 1100 GIANT_REQUIRED; 1101 s = splvm(); 1102 cnt.v_tfree++; 1103 1104 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1105 printf( 1106 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1107 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1108 m->hold_count); 1109 if ((m->queue - m->pc) == PQ_FREE) 1110 panic("vm_page_free: freeing free page"); 1111 else 1112 panic("vm_page_free: freeing busy page"); 1113 } 1114 1115 /* 1116 * unqueue, then remove page. Note that we cannot destroy 1117 * the page here because we do not want to call the pager's 1118 * callback routine until after we've put the page on the 1119 * appropriate free queue. 1120 */ 1121 vm_pageq_remove_nowakeup(m); 1122 vm_page_remove(m); 1123 1124 /* 1125 * If fictitious remove object association and 1126 * return, otherwise delay object association removal. 1127 */ 1128 if ((m->flags & PG_FICTITIOUS) != 0) { 1129 splx(s); 1130 return; 1131 } 1132 1133 m->valid = 0; 1134 vm_page_undirty(m); 1135 1136 if (m->wire_count != 0) { 1137 if (m->wire_count > 1) { 1138 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1139 m->wire_count, (long)m->pindex); 1140 } 1141 panic("vm_page_free: freeing wired page\n"); 1142 } 1143 1144 /* 1145 * If we've exhausted the object's resident pages we want to free 1146 * it up. 1147 */ 1148 if (object && 1149 (object->type == OBJT_VNODE) && 1150 ((object->flags & OBJ_DEAD) == 0) 1151 ) { 1152 struct vnode *vp = (struct vnode *)object->handle; 1153 1154 if (vp && VSHOULDFREE(vp)) 1155 vfree(vp); 1156 } 1157 1158 /* 1159 * Clear the UNMANAGED flag when freeing an unmanaged page. 1160 */ 1161 if (m->flags & PG_UNMANAGED) { 1162 m->flags &= ~PG_UNMANAGED; 1163 } else { 1164#ifdef __alpha__ 1165 pmap_page_is_free(m); 1166#endif 1167 } 1168 1169 if (m->hold_count != 0) { 1170 m->flags &= ~PG_ZERO; 1171 m->queue = PQ_HOLD; 1172 } else 1173 m->queue = PQ_FREE + m->pc; 1174 pq = &vm_page_queues[m->queue]; 1175 mtx_lock_spin(&vm_page_queue_free_mtx); 1176 pq->lcnt++; 1177 ++(*pq->cnt); 1178 1179 /* 1180 * Put zero'd pages on the end ( where we look for zero'd pages 1181 * first ) and non-zerod pages at the head. 1182 */ 1183 if (m->flags & PG_ZERO) { 1184 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1185 ++vm_page_zero_count; 1186 } else { 1187 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1188 } 1189 mtx_unlock_spin(&vm_page_queue_free_mtx); 1190 vm_page_free_wakeup(); 1191 splx(s); 1192} 1193 1194/* 1195 * vm_page_unmanage: 1196 * 1197 * Prevent PV management from being done on the page. The page is 1198 * removed from the paging queues as if it were wired, and as a 1199 * consequence of no longer being managed the pageout daemon will not 1200 * touch it (since there is no way to locate the pte mappings for the 1201 * page). madvise() calls that mess with the pmap will also no longer 1202 * operate on the page. 1203 * 1204 * Beyond that the page is still reasonably 'normal'. Freeing the page 1205 * will clear the flag. 1206 * 1207 * This routine is used by OBJT_PHYS objects - objects using unswappable 1208 * physical memory as backing store rather then swap-backed memory and 1209 * will eventually be extended to support 4MB unmanaged physical 1210 * mappings. 1211 */ 1212void 1213vm_page_unmanage(vm_page_t m) 1214{ 1215 int s; 1216 1217 s = splvm(); 1218 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1219 if ((m->flags & PG_UNMANAGED) == 0) { 1220 if (m->wire_count == 0) 1221 vm_pageq_remove(m); 1222 } 1223 vm_page_flag_set(m, PG_UNMANAGED); 1224 splx(s); 1225} 1226 1227/* 1228 * vm_page_wire: 1229 * 1230 * Mark this page as wired down by yet 1231 * another map, removing it from paging queues 1232 * as necessary. 1233 * 1234 * The page queues must be locked. 1235 * This routine may not block. 1236 */ 1237void 1238vm_page_wire(vm_page_t m) 1239{ 1240 int s; 1241 1242 /* 1243 * Only bump the wire statistics if the page is not already wired, 1244 * and only unqueue the page if it is on some queue (if it is unmanaged 1245 * it is already off the queues). 1246 */ 1247 s = splvm(); 1248 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1249 if (m->wire_count == 0) { 1250 if ((m->flags & PG_UNMANAGED) == 0) 1251 vm_pageq_remove(m); 1252 cnt.v_wire_count++; 1253 } 1254 m->wire_count++; 1255 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1256 splx(s); 1257 vm_page_flag_set(m, PG_MAPPED); /* XXX this does not belong here */ 1258} 1259 1260/* 1261 * vm_page_unwire: 1262 * 1263 * Release one wiring of this page, potentially 1264 * enabling it to be paged again. 1265 * 1266 * Many pages placed on the inactive queue should actually go 1267 * into the cache, but it is difficult to figure out which. What 1268 * we do instead, if the inactive target is well met, is to put 1269 * clean pages at the head of the inactive queue instead of the tail. 1270 * This will cause them to be moved to the cache more quickly and 1271 * if not actively re-referenced, freed more quickly. If we just 1272 * stick these pages at the end of the inactive queue, heavy filesystem 1273 * meta-data accesses can cause an unnecessary paging load on memory bound 1274 * processes. This optimization causes one-time-use metadata to be 1275 * reused more quickly. 1276 * 1277 * BUT, if we are in a low-memory situation we have no choice but to 1278 * put clean pages on the cache queue. 1279 * 1280 * A number of routines use vm_page_unwire() to guarantee that the page 1281 * will go into either the inactive or active queues, and will NEVER 1282 * be placed in the cache - for example, just after dirtying a page. 1283 * dirty pages in the cache are not allowed. 1284 * 1285 * The page queues must be locked. 1286 * This routine may not block. 1287 */ 1288void 1289vm_page_unwire(vm_page_t m, int activate) 1290{ 1291 int s; 1292 1293 s = splvm(); 1294 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1295 if (m->wire_count > 0) { 1296 m->wire_count--; 1297 if (m->wire_count == 0) { 1298 cnt.v_wire_count--; 1299 if (m->flags & PG_UNMANAGED) { 1300 ; 1301 } else if (activate) 1302 vm_pageq_enqueue(PQ_ACTIVE, m); 1303 else { 1304 vm_page_flag_clear(m, PG_WINATCFLS); 1305 vm_pageq_enqueue(PQ_INACTIVE, m); 1306 } 1307 } 1308 } else { 1309 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1310 } 1311 splx(s); 1312} 1313 1314 1315/* 1316 * Move the specified page to the inactive queue. If the page has 1317 * any associated swap, the swap is deallocated. 1318 * 1319 * Normally athead is 0 resulting in LRU operation. athead is set 1320 * to 1 if we want this page to be 'as if it were placed in the cache', 1321 * except without unmapping it from the process address space. 1322 * 1323 * This routine may not block. 1324 */ 1325static __inline void 1326_vm_page_deactivate(vm_page_t m, int athead) 1327{ 1328 int s; 1329 1330 GIANT_REQUIRED; 1331 /* 1332 * Ignore if already inactive. 1333 */ 1334 if (m->queue == PQ_INACTIVE) 1335 return; 1336 1337 s = splvm(); 1338 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1339 if ((m->queue - m->pc) == PQ_CACHE) 1340 cnt.v_reactivated++; 1341 vm_page_flag_clear(m, PG_WINATCFLS); 1342 vm_pageq_remove(m); 1343 if (athead) 1344 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1345 else 1346 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1347 m->queue = PQ_INACTIVE; 1348 vm_page_queues[PQ_INACTIVE].lcnt++; 1349 cnt.v_inactive_count++; 1350 } 1351 splx(s); 1352} 1353 1354void 1355vm_page_deactivate(vm_page_t m) 1356{ 1357 _vm_page_deactivate(m, 0); 1358} 1359 1360/* 1361 * vm_page_try_to_cache: 1362 * 1363 * Returns 0 on failure, 1 on success 1364 */ 1365int 1366vm_page_try_to_cache(vm_page_t m) 1367{ 1368 1369 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1370 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1371 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1372 return (0); 1373 } 1374 vm_page_test_dirty(m); 1375 if (m->dirty) 1376 return (0); 1377 vm_page_cache(m); 1378 return (1); 1379} 1380 1381/* 1382 * vm_page_try_to_free() 1383 * 1384 * Attempt to free the page. If we cannot free it, we do nothing. 1385 * 1 is returned on success, 0 on failure. 1386 */ 1387int 1388vm_page_try_to_free(vm_page_t m) 1389{ 1390 1391 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1392 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1393 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1394 return (0); 1395 } 1396 vm_page_test_dirty(m); 1397 if (m->dirty) 1398 return (0); 1399 vm_page_busy(m); 1400 vm_page_protect(m, VM_PROT_NONE); 1401 vm_page_free(m); 1402 return (1); 1403} 1404 1405/* 1406 * vm_page_cache 1407 * 1408 * Put the specified page onto the page cache queue (if appropriate). 1409 * 1410 * This routine may not block. 1411 */ 1412void 1413vm_page_cache(vm_page_t m) 1414{ 1415 int s; 1416 1417 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1418 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1419 printf("vm_page_cache: attempting to cache busy page\n"); 1420 return; 1421 } 1422 if ((m->queue - m->pc) == PQ_CACHE) 1423 return; 1424 1425 /* 1426 * Remove all pmaps and indicate that the page is not 1427 * writeable or mapped. 1428 */ 1429 vm_page_protect(m, VM_PROT_NONE); 1430 if (m->dirty != 0) { 1431 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1432 (long)m->pindex); 1433 } 1434 s = splvm(); 1435 vm_pageq_remove_nowakeup(m); 1436 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1437 vm_page_free_wakeup(); 1438 splx(s); 1439} 1440 1441/* 1442 * vm_page_dontneed 1443 * 1444 * Cache, deactivate, or do nothing as appropriate. This routine 1445 * is typically used by madvise() MADV_DONTNEED. 1446 * 1447 * Generally speaking we want to move the page into the cache so 1448 * it gets reused quickly. However, this can result in a silly syndrome 1449 * due to the page recycling too quickly. Small objects will not be 1450 * fully cached. On the otherhand, if we move the page to the inactive 1451 * queue we wind up with a problem whereby very large objects 1452 * unnecessarily blow away our inactive and cache queues. 1453 * 1454 * The solution is to move the pages based on a fixed weighting. We 1455 * either leave them alone, deactivate them, or move them to the cache, 1456 * where moving them to the cache has the highest weighting. 1457 * By forcing some pages into other queues we eventually force the 1458 * system to balance the queues, potentially recovering other unrelated 1459 * space from active. The idea is to not force this to happen too 1460 * often. 1461 */ 1462void 1463vm_page_dontneed(vm_page_t m) 1464{ 1465 static int dnweight; 1466 int dnw; 1467 int head; 1468 1469 GIANT_REQUIRED; 1470 dnw = ++dnweight; 1471 1472 /* 1473 * occassionally leave the page alone 1474 */ 1475 if ((dnw & 0x01F0) == 0 || 1476 m->queue == PQ_INACTIVE || 1477 m->queue - m->pc == PQ_CACHE 1478 ) { 1479 if (m->act_count >= ACT_INIT) 1480 --m->act_count; 1481 return; 1482 } 1483 1484 if (m->dirty == 0) 1485 vm_page_test_dirty(m); 1486 1487 if (m->dirty || (dnw & 0x0070) == 0) { 1488 /* 1489 * Deactivate the page 3 times out of 32. 1490 */ 1491 head = 0; 1492 } else { 1493 /* 1494 * Cache the page 28 times out of every 32. Note that 1495 * the page is deactivated instead of cached, but placed 1496 * at the head of the queue instead of the tail. 1497 */ 1498 head = 1; 1499 } 1500 _vm_page_deactivate(m, head); 1501} 1502 1503/* 1504 * Grab a page, waiting until we are waken up due to the page 1505 * changing state. We keep on waiting, if the page continues 1506 * to be in the object. If the page doesn't exist, allocate it. 1507 * 1508 * This routine may block. 1509 */ 1510vm_page_t 1511vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1512{ 1513 vm_page_t m; 1514 int s, generation; 1515 1516 GIANT_REQUIRED; 1517retrylookup: 1518 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1519 if (m->busy || (m->flags & PG_BUSY)) { 1520 generation = object->generation; 1521 1522 s = splvm(); 1523 while ((object->generation == generation) && 1524 (m->busy || (m->flags & PG_BUSY))) { 1525 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1526 tsleep(m, PVM, "pgrbwt", 0); 1527 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1528 splx(s); 1529 return NULL; 1530 } 1531 } 1532 splx(s); 1533 goto retrylookup; 1534 } else { 1535 vm_page_busy(m); 1536 return m; 1537 } 1538 } 1539 1540 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1541 if (m == NULL) { 1542 VM_WAIT; 1543 if ((allocflags & VM_ALLOC_RETRY) == 0) 1544 return NULL; 1545 goto retrylookup; 1546 } 1547 1548 return m; 1549} 1550 1551/* 1552 * Mapping function for valid bits or for dirty bits in 1553 * a page. May not block. 1554 * 1555 * Inputs are required to range within a page. 1556 */ 1557__inline int 1558vm_page_bits(int base, int size) 1559{ 1560 int first_bit; 1561 int last_bit; 1562 1563 KASSERT( 1564 base + size <= PAGE_SIZE, 1565 ("vm_page_bits: illegal base/size %d/%d", base, size) 1566 ); 1567 1568 if (size == 0) /* handle degenerate case */ 1569 return (0); 1570 1571 first_bit = base >> DEV_BSHIFT; 1572 last_bit = (base + size - 1) >> DEV_BSHIFT; 1573 1574 return ((2 << last_bit) - (1 << first_bit)); 1575} 1576 1577/* 1578 * vm_page_set_validclean: 1579 * 1580 * Sets portions of a page valid and clean. The arguments are expected 1581 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1582 * of any partial chunks touched by the range. The invalid portion of 1583 * such chunks will be zero'd. 1584 * 1585 * This routine may not block. 1586 * 1587 * (base + size) must be less then or equal to PAGE_SIZE. 1588 */ 1589void 1590vm_page_set_validclean(vm_page_t m, int base, int size) 1591{ 1592 int pagebits; 1593 int frag; 1594 int endoff; 1595 1596 GIANT_REQUIRED; 1597 if (size == 0) /* handle degenerate case */ 1598 return; 1599 1600 /* 1601 * If the base is not DEV_BSIZE aligned and the valid 1602 * bit is clear, we have to zero out a portion of the 1603 * first block. 1604 */ 1605 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1606 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1607 pmap_zero_page_area(m, frag, base - frag); 1608 1609 /* 1610 * If the ending offset is not DEV_BSIZE aligned and the 1611 * valid bit is clear, we have to zero out a portion of 1612 * the last block. 1613 */ 1614 endoff = base + size; 1615 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1616 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1617 pmap_zero_page_area(m, endoff, 1618 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1619 1620 /* 1621 * Set valid, clear dirty bits. If validating the entire 1622 * page we can safely clear the pmap modify bit. We also 1623 * use this opportunity to clear the PG_NOSYNC flag. If a process 1624 * takes a write fault on a MAP_NOSYNC memory area the flag will 1625 * be set again. 1626 * 1627 * We set valid bits inclusive of any overlap, but we can only 1628 * clear dirty bits for DEV_BSIZE chunks that are fully within 1629 * the range. 1630 */ 1631 pagebits = vm_page_bits(base, size); 1632 m->valid |= pagebits; 1633#if 0 /* NOT YET */ 1634 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1635 frag = DEV_BSIZE - frag; 1636 base += frag; 1637 size -= frag; 1638 if (size < 0) 1639 size = 0; 1640 } 1641 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1642#endif 1643 m->dirty &= ~pagebits; 1644 if (base == 0 && size == PAGE_SIZE) { 1645 pmap_clear_modify(m); 1646 vm_page_flag_clear(m, PG_NOSYNC); 1647 } 1648} 1649 1650#if 0 1651 1652void 1653vm_page_set_dirty(vm_page_t m, int base, int size) 1654{ 1655 m->dirty |= vm_page_bits(base, size); 1656} 1657 1658#endif 1659 1660void 1661vm_page_clear_dirty(vm_page_t m, int base, int size) 1662{ 1663 GIANT_REQUIRED; 1664 m->dirty &= ~vm_page_bits(base, size); 1665} 1666 1667/* 1668 * vm_page_set_invalid: 1669 * 1670 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1671 * valid and dirty bits for the effected areas are cleared. 1672 * 1673 * May not block. 1674 */ 1675void 1676vm_page_set_invalid(vm_page_t m, int base, int size) 1677{ 1678 int bits; 1679 1680 GIANT_REQUIRED; 1681 bits = vm_page_bits(base, size); 1682 m->valid &= ~bits; 1683 m->dirty &= ~bits; 1684 m->object->generation++; 1685} 1686 1687/* 1688 * vm_page_zero_invalid() 1689 * 1690 * The kernel assumes that the invalid portions of a page contain 1691 * garbage, but such pages can be mapped into memory by user code. 1692 * When this occurs, we must zero out the non-valid portions of the 1693 * page so user code sees what it expects. 1694 * 1695 * Pages are most often semi-valid when the end of a file is mapped 1696 * into memory and the file's size is not page aligned. 1697 */ 1698void 1699vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1700{ 1701 int b; 1702 int i; 1703 1704 /* 1705 * Scan the valid bits looking for invalid sections that 1706 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1707 * valid bit may be set ) have already been zerod by 1708 * vm_page_set_validclean(). 1709 */ 1710 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1711 if (i == (PAGE_SIZE / DEV_BSIZE) || 1712 (m->valid & (1 << i)) 1713 ) { 1714 if (i > b) { 1715 pmap_zero_page_area(m, 1716 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1717 } 1718 b = i + 1; 1719 } 1720 } 1721 1722 /* 1723 * setvalid is TRUE when we can safely set the zero'd areas 1724 * as being valid. We can do this if there are no cache consistancy 1725 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1726 */ 1727 if (setvalid) 1728 m->valid = VM_PAGE_BITS_ALL; 1729} 1730 1731/* 1732 * vm_page_is_valid: 1733 * 1734 * Is (partial) page valid? Note that the case where size == 0 1735 * will return FALSE in the degenerate case where the page is 1736 * entirely invalid, and TRUE otherwise. 1737 * 1738 * May not block. 1739 */ 1740int 1741vm_page_is_valid(vm_page_t m, int base, int size) 1742{ 1743 int bits = vm_page_bits(base, size); 1744 1745 if (m->valid && ((m->valid & bits) == bits)) 1746 return 1; 1747 else 1748 return 0; 1749} 1750 1751/* 1752 * update dirty bits from pmap/mmu. May not block. 1753 */ 1754void 1755vm_page_test_dirty(vm_page_t m) 1756{ 1757 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1758 vm_page_dirty(m); 1759 } 1760} 1761 1762int so_zerocp_fullpage = 0; 1763 1764void 1765vm_page_cowfault(vm_page_t m) 1766{ 1767 vm_page_t mnew; 1768 vm_object_t object; 1769 vm_pindex_t pindex; 1770 1771 object = m->object; 1772 pindex = m->pindex; 1773 vm_page_busy(m); 1774 1775 retry_alloc: 1776 vm_page_remove(m); 1777 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 1778 if (mnew == NULL) { 1779 vm_page_insert(m, object, pindex); 1780 VM_WAIT; 1781 goto retry_alloc; 1782 } 1783 1784 if (m->cow == 0) { 1785 /* 1786 * check to see if we raced with an xmit complete when 1787 * waiting to allocate a page. If so, put things back 1788 * the way they were 1789 */ 1790 vm_page_busy(mnew); 1791 vm_page_free(mnew); 1792 vm_page_insert(m, object, pindex); 1793 } else { /* clear COW & copy page */ 1794 if (so_zerocp_fullpage) { 1795 mnew->valid = VM_PAGE_BITS_ALL; 1796 } else { 1797 vm_page_copy(m, mnew); 1798 } 1799 vm_page_dirty(mnew); 1800 vm_page_flag_clear(mnew, PG_BUSY); 1801 } 1802} 1803 1804void 1805vm_page_cowclear(vm_page_t m) 1806{ 1807 1808 /* XXX KDM find out if giant is required here. */ 1809 GIANT_REQUIRED; 1810 if (m->cow) { 1811 atomic_subtract_int(&m->cow, 1); 1812 /* 1813 * let vm_fault add back write permission lazily 1814 */ 1815 } 1816 /* 1817 * sf_buf_free() will free the page, so we needn't do it here 1818 */ 1819} 1820 1821void 1822vm_page_cowsetup(vm_page_t m) 1823{ 1824 /* XXX KDM find out if giant is required here */ 1825 GIANT_REQUIRED; 1826 atomic_add_int(&m->cow, 1); 1827 vm_page_protect(m, VM_PROT_READ); 1828} 1829 1830#include "opt_ddb.h" 1831#ifdef DDB 1832#include <sys/kernel.h> 1833 1834#include <ddb/ddb.h> 1835 1836DB_SHOW_COMMAND(page, vm_page_print_page_info) 1837{ 1838 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1839 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1840 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1841 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1842 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1843 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1844 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1845 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1846 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1847 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1848} 1849 1850DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1851{ 1852 int i; 1853 db_printf("PQ_FREE:"); 1854 for (i = 0; i < PQ_L2_SIZE; i++) { 1855 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1856 } 1857 db_printf("\n"); 1858 1859 db_printf("PQ_CACHE:"); 1860 for (i = 0; i < PQ_L2_SIZE; i++) { 1861 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1862 } 1863 db_printf("\n"); 1864 1865 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1866 vm_page_queues[PQ_ACTIVE].lcnt, 1867 vm_page_queues[PQ_INACTIVE].lcnt); 1868} 1869#endif /* DDB */ 1870