vm_page.c revision 102835
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 102835 2002-09-02 04:04:12Z 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{ 399 400 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 401 m->busy++; 402} 403 404void 405vm_page_io_finish(vm_page_t m) 406{ 407 408 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 409 m->busy--; 410 if (m->busy == 0) 411 vm_page_flash(m); 412} 413 414/* 415 * Keep page from being freed by the page daemon 416 * much of the same effect as wiring, except much lower 417 * overhead and should be used only for *very* temporary 418 * holding ("wiring"). 419 */ 420void 421vm_page_hold(vm_page_t mem) 422{ 423 GIANT_REQUIRED; 424 mem->hold_count++; 425} 426 427void 428vm_page_unhold(vm_page_t mem) 429{ 430 GIANT_REQUIRED; 431 --mem->hold_count; 432 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 433 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 434 vm_page_free_toq(mem); 435} 436 437/* 438 * vm_page_protect: 439 * 440 * Reduce the protection of a page. This routine never raises the 441 * protection and therefore can be safely called if the page is already 442 * at VM_PROT_NONE (it will be a NOP effectively ). 443 */ 444void 445vm_page_protect(vm_page_t mem, int prot) 446{ 447 if (prot == VM_PROT_NONE) { 448 if (pmap_page_is_mapped(mem) || (mem->flags & PG_WRITEABLE)) { 449 pmap_page_protect(mem, VM_PROT_NONE); 450 vm_page_flag_clear(mem, PG_WRITEABLE); 451 } 452 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 453 pmap_page_protect(mem, VM_PROT_READ); 454 vm_page_flag_clear(mem, PG_WRITEABLE); 455 } 456} 457 458/* 459 * vm_page_copy: 460 * 461 * Copy one page to another 462 */ 463void 464vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 465{ 466 pmap_copy_page(src_m, dest_m); 467 dest_m->valid = VM_PAGE_BITS_ALL; 468} 469 470/* 471 * vm_page_free: 472 * 473 * Free a page 474 * 475 * The clearing of PG_ZERO is a temporary safety until the code can be 476 * reviewed to determine that PG_ZERO is being properly cleared on 477 * write faults or maps. PG_ZERO was previously cleared in 478 * vm_page_alloc(). 479 */ 480void 481vm_page_free(vm_page_t m) 482{ 483 vm_page_flag_clear(m, PG_ZERO); 484 vm_page_free_toq(m); 485 vm_page_zero_idle_wakeup(); 486} 487 488/* 489 * vm_page_free_zero: 490 * 491 * Free a page to the zerod-pages queue 492 */ 493void 494vm_page_free_zero(vm_page_t m) 495{ 496 vm_page_flag_set(m, PG_ZERO); 497 vm_page_free_toq(m); 498} 499 500/* 501 * vm_page_sleep_busy: 502 * 503 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 504 * m->busy is zero. Returns TRUE if it had to sleep ( including if 505 * it almost had to sleep and made temporary spl*() mods), FALSE 506 * otherwise. 507 * 508 * This routine assumes that interrupts can only remove the busy 509 * status from a page, not set the busy status or change it from 510 * PG_BUSY to m->busy or vise versa (which would create a timing 511 * window). 512 */ 513int 514vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 515{ 516 GIANT_REQUIRED; 517 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 518 int s = splvm(); 519 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 520 /* 521 * Page is busy. Wait and retry. 522 */ 523 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 524 tsleep(m, PVM, msg, 0); 525 } 526 splx(s); 527 return (TRUE); 528 /* not reached */ 529 } 530 return (FALSE); 531} 532 533/* 534 * vm_page_sleep_if_busy: 535 * 536 * Sleep and release the page queues lock if PG_BUSY is set or, 537 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 538 * thread slept and the page queues lock was released. 539 * Otherwise, retains the page queues lock and returns FALSE. 540 */ 541int 542vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 543{ 544 545 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 546 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 547 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 548 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 549 return (TRUE); 550 } 551 return (FALSE); 552} 553 554/* 555 * vm_page_dirty: 556 * 557 * make page all dirty 558 */ 559void 560vm_page_dirty(vm_page_t m) 561{ 562 KASSERT(m->queue - m->pc != PQ_CACHE, 563 ("vm_page_dirty: page in cache!")); 564 m->dirty = VM_PAGE_BITS_ALL; 565} 566 567/* 568 * vm_page_undirty: 569 * 570 * Set page to not be dirty. Note: does not clear pmap modify bits 571 */ 572void 573vm_page_undirty(vm_page_t m) 574{ 575 m->dirty = 0; 576} 577 578/* 579 * vm_page_insert: [ internal use only ] 580 * 581 * Inserts the given mem entry into the object and object list. 582 * 583 * The pagetables are not updated but will presumably fault the page 584 * in if necessary, or if a kernel page the caller will at some point 585 * enter the page into the kernel's pmap. We are not allowed to block 586 * here so we *can't* do this anyway. 587 * 588 * The object and page must be locked, and must be splhigh. 589 * This routine may not block. 590 */ 591void 592vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 593{ 594 struct vm_page **bucket; 595 596 GIANT_REQUIRED; 597 598 if (m->object != NULL) 599 panic("vm_page_insert: already inserted"); 600 601 /* 602 * Record the object/offset pair in this page 603 */ 604 m->object = object; 605 m->pindex = pindex; 606 607 /* 608 * Insert it into the object_object/offset hash table 609 */ 610 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 611 mtx_lock_spin(&vm_page_buckets_mtx); 612 m->hnext = *bucket; 613 *bucket = m; 614 mtx_unlock_spin(&vm_page_buckets_mtx); 615 616 /* 617 * Now link into the object's list of backed pages. 618 */ 619 TAILQ_INSERT_TAIL(&object->memq, m, listq); 620 object->generation++; 621 622 /* 623 * show that the object has one more resident page. 624 */ 625 object->resident_page_count++; 626 627 /* 628 * Since we are inserting a new and possibly dirty page, 629 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 630 */ 631 if (m->flags & PG_WRITEABLE) 632 vm_object_set_writeable_dirty(object); 633} 634 635/* 636 * vm_page_remove: 637 * NOTE: used by device pager as well -wfj 638 * 639 * Removes the given mem entry from the object/offset-page 640 * table and the object page list, but do not invalidate/terminate 641 * the backing store. 642 * 643 * The object and page must be locked, and at splhigh. 644 * The underlying pmap entry (if any) is NOT removed here. 645 * This routine may not block. 646 */ 647void 648vm_page_remove(vm_page_t m) 649{ 650 vm_object_t object; 651 vm_page_t *bucket; 652 653 GIANT_REQUIRED; 654 655 if (m->object == NULL) 656 return; 657 658 if ((m->flags & PG_BUSY) == 0) { 659 panic("vm_page_remove: page not busy"); 660 } 661 662 /* 663 * Basically destroy the page. 664 */ 665 vm_page_wakeup(m); 666 667 object = m->object; 668 669 /* 670 * Remove from the object_object/offset hash table. The object 671 * must be on the hash queue, we will panic if it isn't 672 */ 673 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 674 mtx_lock_spin(&vm_page_buckets_mtx); 675 while (*bucket != m) { 676 if (*bucket == NULL) 677 panic("vm_page_remove(): page not found in hash"); 678 bucket = &(*bucket)->hnext; 679 } 680 *bucket = m->hnext; 681 m->hnext = NULL; 682 mtx_unlock_spin(&vm_page_buckets_mtx); 683 684 /* 685 * Now remove from the object's list of backed pages. 686 */ 687 TAILQ_REMOVE(&object->memq, m, listq); 688 689 /* 690 * And show that the object has one fewer resident page. 691 */ 692 object->resident_page_count--; 693 object->generation++; 694 695 m->object = NULL; 696} 697 698/* 699 * vm_page_lookup: 700 * 701 * Returns the page associated with the object/offset 702 * pair specified; if none is found, NULL is returned. 703 * 704 * The object must be locked. No side effects. 705 * This routine may not block. 706 * This is a critical path routine 707 */ 708vm_page_t 709vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 710{ 711 vm_page_t m; 712 struct vm_page **bucket; 713 714 /* 715 * Search the hash table for this object/offset pair 716 */ 717 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 718 mtx_lock_spin(&vm_page_buckets_mtx); 719 for (m = *bucket; m != NULL; m = m->hnext) 720 if (m->object == object && m->pindex == pindex) 721 break; 722 mtx_unlock_spin(&vm_page_buckets_mtx); 723 return (m); 724} 725 726/* 727 * vm_page_rename: 728 * 729 * Move the given memory entry from its 730 * current object to the specified target object/offset. 731 * 732 * The object must be locked. 733 * This routine may not block. 734 * 735 * Note: this routine will raise itself to splvm(), the caller need not. 736 * 737 * Note: swap associated with the page must be invalidated by the move. We 738 * have to do this for several reasons: (1) we aren't freeing the 739 * page, (2) we are dirtying the page, (3) the VM system is probably 740 * moving the page from object A to B, and will then later move 741 * the backing store from A to B and we can't have a conflict. 742 * 743 * Note: we *always* dirty the page. It is necessary both for the 744 * fact that we moved it, and because we may be invalidating 745 * swap. If the page is on the cache, we have to deactivate it 746 * or vm_page_dirty() will panic. Dirty pages are not allowed 747 * on the cache. 748 */ 749void 750vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 751{ 752 int s; 753 754 s = splvm(); 755 vm_page_lock_queues(); 756 vm_page_remove(m); 757 vm_page_insert(m, new_object, new_pindex); 758 if (m->queue - m->pc == PQ_CACHE) 759 vm_page_deactivate(m); 760 vm_page_dirty(m); 761 vm_page_unlock_queues(); 762 splx(s); 763} 764 765/* 766 * vm_page_select_cache: 767 * 768 * Find a page on the cache queue with color optimization. As pages 769 * might be found, but not applicable, they are deactivated. This 770 * keeps us from using potentially busy cached pages. 771 * 772 * This routine must be called at splvm(). 773 * This routine may not block. 774 */ 775static vm_page_t 776vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 777{ 778 vm_page_t m; 779 780 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 781 while (TRUE) { 782 m = vm_pageq_find( 783 PQ_CACHE, 784 (pindex + object->pg_color) & PQ_L2_MASK, 785 FALSE 786 ); 787 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 788 m->hold_count || m->wire_count)) { 789 vm_page_deactivate(m); 790 continue; 791 } 792 return m; 793 } 794} 795 796/* 797 * vm_page_select_free: 798 * 799 * Find a free or zero page, with specified preference. 800 * 801 * This routine must be called at splvm(). 802 * This routine may not block. 803 */ 804static __inline vm_page_t 805vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 806{ 807 vm_page_t m; 808 809 m = vm_pageq_find( 810 PQ_FREE, 811 (pindex + object->pg_color) & PQ_L2_MASK, 812 prefer_zero 813 ); 814 return (m); 815} 816 817/* 818 * vm_page_alloc: 819 * 820 * Allocate and return a memory cell associated 821 * with this VM object/offset pair. 822 * 823 * page_req classes: 824 * VM_ALLOC_NORMAL normal process request 825 * VM_ALLOC_SYSTEM system *really* needs a page 826 * VM_ALLOC_INTERRUPT interrupt time request 827 * VM_ALLOC_ZERO zero page 828 * 829 * This routine may not block. 830 * 831 * Additional special handling is required when called from an 832 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 833 * the page cache in this case. 834 */ 835vm_page_t 836vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 837{ 838 vm_page_t m = NULL; 839 int page_req, s; 840 841 GIANT_REQUIRED; 842 843 KASSERT(!vm_page_lookup(object, pindex), 844 ("vm_page_alloc: page already allocated")); 845 846 page_req = req & VM_ALLOC_CLASS_MASK; 847 848 /* 849 * The pager is allowed to eat deeper into the free page list. 850 */ 851 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 852 page_req = VM_ALLOC_SYSTEM; 853 }; 854 855 s = splvm(); 856loop: 857 mtx_lock_spin(&vm_page_queue_free_mtx); 858 if (cnt.v_free_count > cnt.v_free_reserved) { 859 /* 860 * Allocate from the free queue if there are plenty of pages 861 * in it. 862 */ 863 m = vm_page_select_free(object, pindex, 864 (req & VM_ALLOC_ZERO) != 0); 865 } else if ( 866 (page_req == VM_ALLOC_SYSTEM && 867 cnt.v_cache_count == 0 && 868 cnt.v_free_count > cnt.v_interrupt_free_min) || 869 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 870 ) { 871 /* 872 * Interrupt or system, dig deeper into the free list. 873 */ 874 m = vm_page_select_free(object, pindex, FALSE); 875 } else if (page_req != VM_ALLOC_INTERRUPT) { 876 mtx_unlock_spin(&vm_page_queue_free_mtx); 877 /* 878 * Allocatable from cache (non-interrupt only). On success, 879 * we must free the page and try again, thus ensuring that 880 * cnt.v_*_free_min counters are replenished. 881 */ 882 vm_page_lock_queues(); 883 if ((m = vm_page_select_cache(object, pindex)) == NULL) { 884 vm_page_unlock_queues(); 885 splx(s); 886#if defined(DIAGNOSTIC) 887 if (cnt.v_cache_count > 0) 888 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 889#endif 890 vm_pageout_deficit++; 891 pagedaemon_wakeup(); 892 return (NULL); 893 } 894 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 895 vm_page_busy(m); 896 vm_page_protect(m, VM_PROT_NONE); 897 vm_page_free(m); 898 vm_page_unlock_queues(); 899 goto loop; 900 } else { 901 /* 902 * Not allocatable from cache from interrupt, give up. 903 */ 904 mtx_unlock_spin(&vm_page_queue_free_mtx); 905 splx(s); 906 vm_pageout_deficit++; 907 pagedaemon_wakeup(); 908 return (NULL); 909 } 910 911 /* 912 * At this point we had better have found a good page. 913 */ 914 915 KASSERT( 916 m != NULL, 917 ("vm_page_alloc(): missing page on free queue\n") 918 ); 919 920 /* 921 * Remove from free queue 922 */ 923 924 vm_pageq_remove_nowakeup(m); 925 926 /* 927 * Initialize structure. Only the PG_ZERO flag is inherited. 928 */ 929 if (m->flags & PG_ZERO) { 930 vm_page_zero_count--; 931 m->flags = PG_ZERO | PG_BUSY; 932 } else { 933 m->flags = PG_BUSY; 934 } 935 if (req & VM_ALLOC_WIRED) { 936 cnt.v_wire_count++; 937 m->wire_count = 1; 938 } else 939 m->wire_count = 0; 940 m->hold_count = 0; 941 m->act_count = 0; 942 m->busy = 0; 943 m->valid = 0; 944 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 945 mtx_unlock_spin(&vm_page_queue_free_mtx); 946 947 /* 948 * vm_page_insert() is safe prior to the splx(). Note also that 949 * inserting a page here does not insert it into the pmap (which 950 * could cause us to block allocating memory). We cannot block 951 * anywhere. 952 */ 953 vm_page_insert(m, object, pindex); 954 955 /* 956 * Don't wakeup too often - wakeup the pageout daemon when 957 * we would be nearly out of memory. 958 */ 959 if (vm_paging_needed()) 960 pagedaemon_wakeup(); 961 962 splx(s); 963 return (m); 964} 965 966/* 967 * vm_wait: (also see VM_WAIT macro) 968 * 969 * Block until free pages are available for allocation 970 * - Called in various places before memory allocations. 971 */ 972void 973vm_wait(void) 974{ 975 int s; 976 977 s = splvm(); 978 if (curproc == pageproc) { 979 vm_pageout_pages_needed = 1; 980 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 981 } else { 982 if (!vm_pages_needed) { 983 vm_pages_needed = 1; 984 wakeup(&vm_pages_needed); 985 } 986 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 987 } 988 splx(s); 989} 990 991/* 992 * vm_waitpfault: (also see VM_WAITPFAULT macro) 993 * 994 * Block until free pages are available for allocation 995 * - Called only in vm_fault so that processes page faulting 996 * can be easily tracked. 997 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 998 * processes will be able to grab memory first. Do not change 999 * this balance without careful testing first. 1000 */ 1001void 1002vm_waitpfault(void) 1003{ 1004 int s; 1005 1006 s = splvm(); 1007 if (!vm_pages_needed) { 1008 vm_pages_needed = 1; 1009 wakeup(&vm_pages_needed); 1010 } 1011 tsleep(&cnt.v_free_count, PUSER, "pfault", 0); 1012 splx(s); 1013} 1014 1015/* 1016 * vm_page_activate: 1017 * 1018 * Put the specified page on the active list (if appropriate). 1019 * Ensure that act_count is at least ACT_INIT but do not otherwise 1020 * mess with it. 1021 * 1022 * The page queues must be locked. 1023 * This routine may not block. 1024 */ 1025void 1026vm_page_activate(vm_page_t m) 1027{ 1028 int s; 1029 1030 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1031 s = splvm(); 1032 if (m->queue != PQ_ACTIVE) { 1033 if ((m->queue - m->pc) == PQ_CACHE) 1034 cnt.v_reactivated++; 1035 vm_pageq_remove(m); 1036 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1037 if (m->act_count < ACT_INIT) 1038 m->act_count = ACT_INIT; 1039 vm_pageq_enqueue(PQ_ACTIVE, m); 1040 } 1041 } else { 1042 if (m->act_count < ACT_INIT) 1043 m->act_count = ACT_INIT; 1044 } 1045 splx(s); 1046} 1047 1048/* 1049 * vm_page_free_wakeup: 1050 * 1051 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1052 * routine is called when a page has been added to the cache or free 1053 * queues. 1054 * 1055 * This routine may not block. 1056 * This routine must be called at splvm() 1057 */ 1058static __inline void 1059vm_page_free_wakeup(void) 1060{ 1061 /* 1062 * if pageout daemon needs pages, then tell it that there are 1063 * some free. 1064 */ 1065 if (vm_pageout_pages_needed && 1066 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1067 wakeup(&vm_pageout_pages_needed); 1068 vm_pageout_pages_needed = 0; 1069 } 1070 /* 1071 * wakeup processes that are waiting on memory if we hit a 1072 * high water mark. And wakeup scheduler process if we have 1073 * lots of memory. this process will swapin processes. 1074 */ 1075 if (vm_pages_needed && !vm_page_count_min()) { 1076 vm_pages_needed = 0; 1077 wakeup(&cnt.v_free_count); 1078 } 1079} 1080 1081/* 1082 * vm_page_free_toq: 1083 * 1084 * Returns the given page to the PQ_FREE list, 1085 * disassociating it with any VM object. 1086 * 1087 * Object and page must be locked prior to entry. 1088 * This routine may not block. 1089 */ 1090 1091void 1092vm_page_free_toq(vm_page_t m) 1093{ 1094 int s; 1095 struct vpgqueues *pq; 1096 vm_object_t object = m->object; 1097 1098 GIANT_REQUIRED; 1099 s = splvm(); 1100 cnt.v_tfree++; 1101 1102 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1103 printf( 1104 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1105 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1106 m->hold_count); 1107 if ((m->queue - m->pc) == PQ_FREE) 1108 panic("vm_page_free: freeing free page"); 1109 else 1110 panic("vm_page_free: freeing busy page"); 1111 } 1112 1113 /* 1114 * unqueue, then remove page. Note that we cannot destroy 1115 * the page here because we do not want to call the pager's 1116 * callback routine until after we've put the page on the 1117 * appropriate free queue. 1118 */ 1119 vm_pageq_remove_nowakeup(m); 1120 vm_page_remove(m); 1121 1122 /* 1123 * If fictitious remove object association and 1124 * return, otherwise delay object association removal. 1125 */ 1126 if ((m->flags & PG_FICTITIOUS) != 0) { 1127 splx(s); 1128 return; 1129 } 1130 1131 m->valid = 0; 1132 vm_page_undirty(m); 1133 1134 if (m->wire_count != 0) { 1135 if (m->wire_count > 1) { 1136 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1137 m->wire_count, (long)m->pindex); 1138 } 1139 panic("vm_page_free: freeing wired page\n"); 1140 } 1141 1142 /* 1143 * If we've exhausted the object's resident pages we want to free 1144 * it up. 1145 */ 1146 if (object && 1147 (object->type == OBJT_VNODE) && 1148 ((object->flags & OBJ_DEAD) == 0) 1149 ) { 1150 struct vnode *vp = (struct vnode *)object->handle; 1151 1152 if (vp) { 1153 VI_LOCK(vp); 1154 if (VSHOULDFREE(vp)) 1155 vfree(vp); 1156 VI_UNLOCK(vp); 1157 } 1158 } 1159 1160 /* 1161 * Clear the UNMANAGED flag when freeing an unmanaged page. 1162 */ 1163 if (m->flags & PG_UNMANAGED) { 1164 m->flags &= ~PG_UNMANAGED; 1165 } else { 1166#ifdef __alpha__ 1167 pmap_page_is_free(m); 1168#endif 1169 } 1170 1171 if (m->hold_count != 0) { 1172 m->flags &= ~PG_ZERO; 1173 m->queue = PQ_HOLD; 1174 } else 1175 m->queue = PQ_FREE + m->pc; 1176 pq = &vm_page_queues[m->queue]; 1177 mtx_lock_spin(&vm_page_queue_free_mtx); 1178 pq->lcnt++; 1179 ++(*pq->cnt); 1180 1181 /* 1182 * Put zero'd pages on the end ( where we look for zero'd pages 1183 * first ) and non-zerod pages at the head. 1184 */ 1185 if (m->flags & PG_ZERO) { 1186 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1187 ++vm_page_zero_count; 1188 } else { 1189 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1190 } 1191 mtx_unlock_spin(&vm_page_queue_free_mtx); 1192 vm_page_free_wakeup(); 1193 splx(s); 1194} 1195 1196/* 1197 * vm_page_unmanage: 1198 * 1199 * Prevent PV management from being done on the page. The page is 1200 * removed from the paging queues as if it were wired, and as a 1201 * consequence of no longer being managed the pageout daemon will not 1202 * touch it (since there is no way to locate the pte mappings for the 1203 * page). madvise() calls that mess with the pmap will also no longer 1204 * operate on the page. 1205 * 1206 * Beyond that the page is still reasonably 'normal'. Freeing the page 1207 * will clear the flag. 1208 * 1209 * This routine is used by OBJT_PHYS objects - objects using unswappable 1210 * physical memory as backing store rather then swap-backed memory and 1211 * will eventually be extended to support 4MB unmanaged physical 1212 * mappings. 1213 */ 1214void 1215vm_page_unmanage(vm_page_t m) 1216{ 1217 int s; 1218 1219 s = splvm(); 1220 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1221 if ((m->flags & PG_UNMANAGED) == 0) { 1222 if (m->wire_count == 0) 1223 vm_pageq_remove(m); 1224 } 1225 vm_page_flag_set(m, PG_UNMANAGED); 1226 splx(s); 1227} 1228 1229/* 1230 * vm_page_wire: 1231 * 1232 * Mark this page as wired down by yet 1233 * another map, removing it from paging queues 1234 * as necessary. 1235 * 1236 * The page queues must be locked. 1237 * This routine may not block. 1238 */ 1239void 1240vm_page_wire(vm_page_t m) 1241{ 1242 int s; 1243 1244 /* 1245 * Only bump the wire statistics if the page is not already wired, 1246 * and only unqueue the page if it is on some queue (if it is unmanaged 1247 * it is already off the queues). 1248 */ 1249 s = splvm(); 1250 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1251 if (m->wire_count == 0) { 1252 if ((m->flags & PG_UNMANAGED) == 0) 1253 vm_pageq_remove(m); 1254 cnt.v_wire_count++; 1255 } 1256 m->wire_count++; 1257 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1258 splx(s); 1259} 1260 1261/* 1262 * vm_page_unwire: 1263 * 1264 * Release one wiring of this page, potentially 1265 * enabling it to be paged again. 1266 * 1267 * Many pages placed on the inactive queue should actually go 1268 * into the cache, but it is difficult to figure out which. What 1269 * we do instead, if the inactive target is well met, is to put 1270 * clean pages at the head of the inactive queue instead of the tail. 1271 * This will cause them to be moved to the cache more quickly and 1272 * if not actively re-referenced, freed more quickly. If we just 1273 * stick these pages at the end of the inactive queue, heavy filesystem 1274 * meta-data accesses can cause an unnecessary paging load on memory bound 1275 * processes. This optimization causes one-time-use metadata to be 1276 * reused more quickly. 1277 * 1278 * BUT, if we are in a low-memory situation we have no choice but to 1279 * put clean pages on the cache queue. 1280 * 1281 * A number of routines use vm_page_unwire() to guarantee that the page 1282 * will go into either the inactive or active queues, and will NEVER 1283 * be placed in the cache - for example, just after dirtying a page. 1284 * dirty pages in the cache are not allowed. 1285 * 1286 * The page queues must be locked. 1287 * This routine may not block. 1288 */ 1289void 1290vm_page_unwire(vm_page_t m, int activate) 1291{ 1292 int s; 1293 1294 s = splvm(); 1295 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1296 if (m->wire_count > 0) { 1297 m->wire_count--; 1298 if (m->wire_count == 0) { 1299 cnt.v_wire_count--; 1300 if (m->flags & PG_UNMANAGED) { 1301 ; 1302 } else if (activate) 1303 vm_pageq_enqueue(PQ_ACTIVE, m); 1304 else { 1305 vm_page_flag_clear(m, PG_WINATCFLS); 1306 vm_pageq_enqueue(PQ_INACTIVE, m); 1307 } 1308 } 1309 } else { 1310 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1311 } 1312 splx(s); 1313} 1314 1315 1316/* 1317 * Move the specified page to the inactive queue. If the page has 1318 * any associated swap, the swap is deallocated. 1319 * 1320 * Normally athead is 0 resulting in LRU operation. athead is set 1321 * to 1 if we want this page to be 'as if it were placed in the cache', 1322 * except without unmapping it from the process address space. 1323 * 1324 * This routine may not block. 1325 */ 1326static __inline void 1327_vm_page_deactivate(vm_page_t m, int athead) 1328{ 1329 int s; 1330 1331 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1332 /* 1333 * Ignore if already inactive. 1334 */ 1335 if (m->queue == PQ_INACTIVE) 1336 return; 1337 1338 s = splvm(); 1339 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1340 if ((m->queue - m->pc) == PQ_CACHE) 1341 cnt.v_reactivated++; 1342 vm_page_flag_clear(m, PG_WINATCFLS); 1343 vm_pageq_remove(m); 1344 if (athead) 1345 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1346 else 1347 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1348 m->queue = PQ_INACTIVE; 1349 vm_page_queues[PQ_INACTIVE].lcnt++; 1350 cnt.v_inactive_count++; 1351 } 1352 splx(s); 1353} 1354 1355void 1356vm_page_deactivate(vm_page_t m) 1357{ 1358 _vm_page_deactivate(m, 0); 1359} 1360 1361/* 1362 * vm_page_try_to_cache: 1363 * 1364 * Returns 0 on failure, 1 on success 1365 */ 1366int 1367vm_page_try_to_cache(vm_page_t m) 1368{ 1369 1370 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1371 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1372 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1373 return (0); 1374 } 1375 vm_page_test_dirty(m); 1376 if (m->dirty) 1377 return (0); 1378 vm_page_cache(m); 1379 return (1); 1380} 1381 1382/* 1383 * vm_page_try_to_free() 1384 * 1385 * Attempt to free the page. If we cannot free it, we do nothing. 1386 * 1 is returned on success, 0 on failure. 1387 */ 1388int 1389vm_page_try_to_free(vm_page_t m) 1390{ 1391 1392 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1393 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1394 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1395 return (0); 1396 } 1397 vm_page_test_dirty(m); 1398 if (m->dirty) 1399 return (0); 1400 vm_page_busy(m); 1401 vm_page_protect(m, VM_PROT_NONE); 1402 vm_page_free(m); 1403 return (1); 1404} 1405 1406/* 1407 * vm_page_cache 1408 * 1409 * Put the specified page onto the page cache queue (if appropriate). 1410 * 1411 * This routine may not block. 1412 */ 1413void 1414vm_page_cache(vm_page_t m) 1415{ 1416 int s; 1417 1418 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1419 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1420 printf("vm_page_cache: attempting to cache busy page\n"); 1421 return; 1422 } 1423 if ((m->queue - m->pc) == PQ_CACHE) 1424 return; 1425 1426 /* 1427 * Remove all pmaps and indicate that the page is not 1428 * writeable or mapped. 1429 */ 1430 vm_page_protect(m, VM_PROT_NONE); 1431 if (m->dirty != 0) { 1432 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1433 (long)m->pindex); 1434 } 1435 s = splvm(); 1436 vm_pageq_remove_nowakeup(m); 1437 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1438 vm_page_free_wakeup(); 1439 splx(s); 1440} 1441 1442/* 1443 * vm_page_dontneed 1444 * 1445 * Cache, deactivate, or do nothing as appropriate. This routine 1446 * is typically used by madvise() MADV_DONTNEED. 1447 * 1448 * Generally speaking we want to move the page into the cache so 1449 * it gets reused quickly. However, this can result in a silly syndrome 1450 * due to the page recycling too quickly. Small objects will not be 1451 * fully cached. On the otherhand, if we move the page to the inactive 1452 * queue we wind up with a problem whereby very large objects 1453 * unnecessarily blow away our inactive and cache queues. 1454 * 1455 * The solution is to move the pages based on a fixed weighting. We 1456 * either leave them alone, deactivate them, or move them to the cache, 1457 * where moving them to the cache has the highest weighting. 1458 * By forcing some pages into other queues we eventually force the 1459 * system to balance the queues, potentially recovering other unrelated 1460 * space from active. The idea is to not force this to happen too 1461 * often. 1462 */ 1463void 1464vm_page_dontneed(vm_page_t m) 1465{ 1466 static int dnweight; 1467 int dnw; 1468 int head; 1469 1470 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1471 dnw = ++dnweight; 1472 1473 /* 1474 * occassionally leave the page alone 1475 */ 1476 if ((dnw & 0x01F0) == 0 || 1477 m->queue == PQ_INACTIVE || 1478 m->queue - m->pc == PQ_CACHE 1479 ) { 1480 if (m->act_count >= ACT_INIT) 1481 --m->act_count; 1482 return; 1483 } 1484 1485 if (m->dirty == 0) 1486 vm_page_test_dirty(m); 1487 1488 if (m->dirty || (dnw & 0x0070) == 0) { 1489 /* 1490 * Deactivate the page 3 times out of 32. 1491 */ 1492 head = 0; 1493 } else { 1494 /* 1495 * Cache the page 28 times out of every 32. Note that 1496 * the page is deactivated instead of cached, but placed 1497 * at the head of the queue instead of the tail. 1498 */ 1499 head = 1; 1500 } 1501 _vm_page_deactivate(m, head); 1502} 1503 1504/* 1505 * Grab a page, waiting until we are waken up due to the page 1506 * changing state. We keep on waiting, if the page continues 1507 * to be in the object. If the page doesn't exist, allocate it. 1508 * 1509 * This routine may block. 1510 */ 1511vm_page_t 1512vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1513{ 1514 vm_page_t m; 1515 int s, generation; 1516 1517 GIANT_REQUIRED; 1518retrylookup: 1519 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1520 vm_page_lock_queues(); 1521 if (m->busy || (m->flags & PG_BUSY)) { 1522 generation = object->generation; 1523 1524 s = splvm(); 1525 while ((object->generation == generation) && 1526 (m->busy || (m->flags & PG_BUSY))) { 1527 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1528 msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0); 1529 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1530 vm_page_unlock_queues(); 1531 splx(s); 1532 return NULL; 1533 } 1534 } 1535 vm_page_unlock_queues(); 1536 splx(s); 1537 goto retrylookup; 1538 } else { 1539 if (allocflags & VM_ALLOC_WIRED) 1540 vm_page_wire(m); 1541 vm_page_busy(m); 1542 vm_page_unlock_queues(); 1543 return m; 1544 } 1545 } 1546 1547 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1548 if (m == NULL) { 1549 VM_WAIT; 1550 if ((allocflags & VM_ALLOC_RETRY) == 0) 1551 return NULL; 1552 goto retrylookup; 1553 } 1554 1555 return m; 1556} 1557 1558/* 1559 * Mapping function for valid bits or for dirty bits in 1560 * a page. May not block. 1561 * 1562 * Inputs are required to range within a page. 1563 */ 1564__inline int 1565vm_page_bits(int base, int size) 1566{ 1567 int first_bit; 1568 int last_bit; 1569 1570 KASSERT( 1571 base + size <= PAGE_SIZE, 1572 ("vm_page_bits: illegal base/size %d/%d", base, size) 1573 ); 1574 1575 if (size == 0) /* handle degenerate case */ 1576 return (0); 1577 1578 first_bit = base >> DEV_BSHIFT; 1579 last_bit = (base + size - 1) >> DEV_BSHIFT; 1580 1581 return ((2 << last_bit) - (1 << first_bit)); 1582} 1583 1584/* 1585 * vm_page_set_validclean: 1586 * 1587 * Sets portions of a page valid and clean. The arguments are expected 1588 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1589 * of any partial chunks touched by the range. The invalid portion of 1590 * such chunks will be zero'd. 1591 * 1592 * This routine may not block. 1593 * 1594 * (base + size) must be less then or equal to PAGE_SIZE. 1595 */ 1596void 1597vm_page_set_validclean(vm_page_t m, int base, int size) 1598{ 1599 int pagebits; 1600 int frag; 1601 int endoff; 1602 1603 GIANT_REQUIRED; 1604 if (size == 0) /* handle degenerate case */ 1605 return; 1606 1607 /* 1608 * If the base is not DEV_BSIZE aligned and the valid 1609 * bit is clear, we have to zero out a portion of the 1610 * first block. 1611 */ 1612 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1613 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1614 pmap_zero_page_area(m, frag, base - frag); 1615 1616 /* 1617 * If the ending offset is not DEV_BSIZE aligned and the 1618 * valid bit is clear, we have to zero out a portion of 1619 * the last block. 1620 */ 1621 endoff = base + size; 1622 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1623 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1624 pmap_zero_page_area(m, endoff, 1625 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1626 1627 /* 1628 * Set valid, clear dirty bits. If validating the entire 1629 * page we can safely clear the pmap modify bit. We also 1630 * use this opportunity to clear the PG_NOSYNC flag. If a process 1631 * takes a write fault on a MAP_NOSYNC memory area the flag will 1632 * be set again. 1633 * 1634 * We set valid bits inclusive of any overlap, but we can only 1635 * clear dirty bits for DEV_BSIZE chunks that are fully within 1636 * the range. 1637 */ 1638 pagebits = vm_page_bits(base, size); 1639 m->valid |= pagebits; 1640#if 0 /* NOT YET */ 1641 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1642 frag = DEV_BSIZE - frag; 1643 base += frag; 1644 size -= frag; 1645 if (size < 0) 1646 size = 0; 1647 } 1648 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1649#endif 1650 m->dirty &= ~pagebits; 1651 if (base == 0 && size == PAGE_SIZE) { 1652 pmap_clear_modify(m); 1653 vm_page_flag_clear(m, PG_NOSYNC); 1654 } 1655} 1656 1657#if 0 1658 1659void 1660vm_page_set_dirty(vm_page_t m, int base, int size) 1661{ 1662 m->dirty |= vm_page_bits(base, size); 1663} 1664 1665#endif 1666 1667void 1668vm_page_clear_dirty(vm_page_t m, int base, int size) 1669{ 1670 GIANT_REQUIRED; 1671 m->dirty &= ~vm_page_bits(base, size); 1672} 1673 1674/* 1675 * vm_page_set_invalid: 1676 * 1677 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1678 * valid and dirty bits for the effected areas are cleared. 1679 * 1680 * May not block. 1681 */ 1682void 1683vm_page_set_invalid(vm_page_t m, int base, int size) 1684{ 1685 int bits; 1686 1687 GIANT_REQUIRED; 1688 bits = vm_page_bits(base, size); 1689 m->valid &= ~bits; 1690 m->dirty &= ~bits; 1691 m->object->generation++; 1692} 1693 1694/* 1695 * vm_page_zero_invalid() 1696 * 1697 * The kernel assumes that the invalid portions of a page contain 1698 * garbage, but such pages can be mapped into memory by user code. 1699 * When this occurs, we must zero out the non-valid portions of the 1700 * page so user code sees what it expects. 1701 * 1702 * Pages are most often semi-valid when the end of a file is mapped 1703 * into memory and the file's size is not page aligned. 1704 */ 1705void 1706vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1707{ 1708 int b; 1709 int i; 1710 1711 /* 1712 * Scan the valid bits looking for invalid sections that 1713 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1714 * valid bit may be set ) have already been zerod by 1715 * vm_page_set_validclean(). 1716 */ 1717 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1718 if (i == (PAGE_SIZE / DEV_BSIZE) || 1719 (m->valid & (1 << i)) 1720 ) { 1721 if (i > b) { 1722 pmap_zero_page_area(m, 1723 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1724 } 1725 b = i + 1; 1726 } 1727 } 1728 1729 /* 1730 * setvalid is TRUE when we can safely set the zero'd areas 1731 * as being valid. We can do this if there are no cache consistancy 1732 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1733 */ 1734 if (setvalid) 1735 m->valid = VM_PAGE_BITS_ALL; 1736} 1737 1738/* 1739 * vm_page_is_valid: 1740 * 1741 * Is (partial) page valid? Note that the case where size == 0 1742 * will return FALSE in the degenerate case where the page is 1743 * entirely invalid, and TRUE otherwise. 1744 * 1745 * May not block. 1746 */ 1747int 1748vm_page_is_valid(vm_page_t m, int base, int size) 1749{ 1750 int bits = vm_page_bits(base, size); 1751 1752 if (m->valid && ((m->valid & bits) == bits)) 1753 return 1; 1754 else 1755 return 0; 1756} 1757 1758/* 1759 * update dirty bits from pmap/mmu. May not block. 1760 */ 1761void 1762vm_page_test_dirty(vm_page_t m) 1763{ 1764 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1765 vm_page_dirty(m); 1766 } 1767} 1768 1769int so_zerocp_fullpage = 0; 1770 1771void 1772vm_page_cowfault(vm_page_t m) 1773{ 1774 vm_page_t mnew; 1775 vm_object_t object; 1776 vm_pindex_t pindex; 1777 1778 object = m->object; 1779 pindex = m->pindex; 1780 vm_page_busy(m); 1781 1782 retry_alloc: 1783 vm_page_remove(m); 1784 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 1785 if (mnew == NULL) { 1786 vm_page_insert(m, object, pindex); 1787 VM_WAIT; 1788 goto retry_alloc; 1789 } 1790 1791 if (m->cow == 0) { 1792 /* 1793 * check to see if we raced with an xmit complete when 1794 * waiting to allocate a page. If so, put things back 1795 * the way they were 1796 */ 1797 vm_page_busy(mnew); 1798 vm_page_free(mnew); 1799 vm_page_insert(m, object, pindex); 1800 } else { /* clear COW & copy page */ 1801 if (so_zerocp_fullpage) { 1802 mnew->valid = VM_PAGE_BITS_ALL; 1803 } else { 1804 vm_page_copy(m, mnew); 1805 } 1806 vm_page_dirty(mnew); 1807 vm_page_flag_clear(mnew, PG_BUSY); 1808 } 1809} 1810 1811void 1812vm_page_cowclear(vm_page_t m) 1813{ 1814 1815 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1816 if (m->cow) { 1817 m->cow--; 1818 /* 1819 * let vm_fault add back write permission lazily 1820 */ 1821 } 1822 /* 1823 * sf_buf_free() will free the page, so we needn't do it here 1824 */ 1825} 1826 1827void 1828vm_page_cowsetup(vm_page_t m) 1829{ 1830 1831 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1832 m->cow++; 1833 vm_page_protect(m, VM_PROT_READ); 1834} 1835 1836#include "opt_ddb.h" 1837#ifdef DDB 1838#include <sys/kernel.h> 1839 1840#include <ddb/ddb.h> 1841 1842DB_SHOW_COMMAND(page, vm_page_print_page_info) 1843{ 1844 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1845 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1846 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1847 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1848 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1849 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1850 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1851 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1852 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1853 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1854} 1855 1856DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1857{ 1858 int i; 1859 db_printf("PQ_FREE:"); 1860 for (i = 0; i < PQ_L2_SIZE; i++) { 1861 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1862 } 1863 db_printf("\n"); 1864 1865 db_printf("PQ_CACHE:"); 1866 for (i = 0; i < PQ_L2_SIZE; i++) { 1867 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1868 } 1869 db_printf("\n"); 1870 1871 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1872 vm_page_queues[PQ_ACTIVE].lcnt, 1873 vm_page_queues[PQ_INACTIVE].lcnt); 1874} 1875#endif /* DDB */ 1876